2 * Copyright (c) 2004 Topspin Communications. All rights reserved.
3 * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
5 * This software is available to you under a choice of one of two
6 * licenses. You may choose to be licensed under the terms of the GNU
7 * General Public License (GPL) Version 2, available from the file
8 * COPYING in the main directory of this source tree, or the
9 * OpenIB.org BSD license below:
11 * Redistribution and use in source and binary forms, with or
12 * without modification, are permitted provided that the following
15 * - Redistributions of source code must retain the above
16 * copyright notice, this list of conditions and the following
19 * - Redistributions in binary form must reproduce the above
20 * copyright notice, this list of conditions and the following
21 * disclaimer in the documentation and/or other materials
22 * provided with the distribution.
24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
34 #include <linux/module.h>
35 #include <linux/string.h>
36 #include <linux/errno.h>
37 #include <linux/kernel.h>
38 #include <linux/slab.h>
39 #include <linux/init.h>
40 #include <linux/netdevice.h>
41 #include <net/net_namespace.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/hashtable.h>
45 #include <rdma/rdma_netlink.h>
46 #include <rdma/ib_addr.h>
47 #include <rdma/ib_cache.h>
48 #include <rdma/rdma_counter.h>
50 #include "core_priv.h"
53 MODULE_AUTHOR("Roland Dreier");
54 MODULE_DESCRIPTION("core kernel InfiniBand API");
55 MODULE_LICENSE("Dual BSD/GPL");
57 struct workqueue_struct *ib_comp_wq;
58 struct workqueue_struct *ib_comp_unbound_wq;
59 struct workqueue_struct *ib_wq;
60 EXPORT_SYMBOL_GPL(ib_wq);
61 static struct workqueue_struct *ib_unreg_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
114 unsigned int rdma_dev_net_id;
117 * A list of net namespaces is maintained in an xarray. This is necessary
118 * because we can't get the locking right using the existing net ns list. We
119 * would require a init_net callback after the list is updated.
121 static DEFINE_XARRAY_FLAGS(rdma_nets, XA_FLAGS_ALLOC);
123 * rwsem to protect accessing the rdma_nets xarray entries.
125 static DECLARE_RWSEM(rdma_nets_rwsem);
127 bool ib_devices_shared_netns = true;
128 module_param_named(netns_mode, ib_devices_shared_netns, bool, 0444);
129 MODULE_PARM_DESC(netns_mode,
130 "Share device among net namespaces; default=1 (shared)");
132 * rdma_dev_access_netns() - Return whether an rdma device can be accessed
133 * from a specified net namespace or not.
134 * @dev: Pointer to rdma device which needs to be checked
135 * @net: Pointer to net namesapce for which access to be checked
137 * When the rdma device is in shared mode, it ignores the net namespace.
138 * When the rdma device is exclusive to a net namespace, rdma device net
139 * namespace is checked against the specified one.
141 bool rdma_dev_access_netns(const struct ib_device *dev, const struct net *net)
143 return (ib_devices_shared_netns ||
144 net_eq(read_pnet(&dev->coredev.rdma_net), net));
146 EXPORT_SYMBOL(rdma_dev_access_netns);
149 * xarray has this behavior where it won't iterate over NULL values stored in
150 * allocated arrays. So we need our own iterator to see all values stored in
151 * the array. This does the same thing as xa_for_each except that it also
152 * returns NULL valued entries if the array is allocating. Simplified to only
153 * work on simple xarrays.
155 static void *xan_find_marked(struct xarray *xa, unsigned long *indexp,
158 XA_STATE(xas, xa, *indexp);
163 entry = xas_find_marked(&xas, ULONG_MAX, filter);
164 if (xa_is_zero(entry))
166 } while (xas_retry(&xas, entry));
170 *indexp = xas.xa_index;
171 if (xa_is_zero(entry))
175 return XA_ERROR(-ENOENT);
177 #define xan_for_each_marked(xa, index, entry, filter) \
178 for (index = 0, entry = xan_find_marked(xa, &(index), filter); \
180 (index)++, entry = xan_find_marked(xa, &(index), filter))
182 /* RCU hash table mapping netdevice pointers to struct ib_port_data */
183 static DEFINE_SPINLOCK(ndev_hash_lock);
184 static DECLARE_HASHTABLE(ndev_hash, 5);
186 static void free_netdevs(struct ib_device *ib_dev);
187 static void ib_unregister_work(struct work_struct *work);
188 static void __ib_unregister_device(struct ib_device *device);
189 static int ib_security_change(struct notifier_block *nb, unsigned long event,
191 static void ib_policy_change_task(struct work_struct *work);
192 static DECLARE_WORK(ib_policy_change_work, ib_policy_change_task);
194 static void __ibdev_printk(const char *level, const struct ib_device *ibdev,
195 struct va_format *vaf)
197 if (ibdev && ibdev->dev.parent)
198 dev_printk_emit(level[1] - '0',
201 dev_driver_string(ibdev->dev.parent),
202 dev_name(ibdev->dev.parent),
203 dev_name(&ibdev->dev),
207 level, dev_name(&ibdev->dev), vaf);
209 printk("%s(NULL ib_device): %pV", level, vaf);
212 void ibdev_printk(const char *level, const struct ib_device *ibdev,
213 const char *format, ...)
215 struct va_format vaf;
218 va_start(args, format);
223 __ibdev_printk(level, ibdev, &vaf);
227 EXPORT_SYMBOL(ibdev_printk);
229 #define define_ibdev_printk_level(func, level) \
230 void func(const struct ib_device *ibdev, const char *fmt, ...) \
232 struct va_format vaf; \
235 va_start(args, fmt); \
240 __ibdev_printk(level, ibdev, &vaf); \
246 define_ibdev_printk_level(ibdev_emerg, KERN_EMERG);
247 define_ibdev_printk_level(ibdev_alert, KERN_ALERT);
248 define_ibdev_printk_level(ibdev_crit, KERN_CRIT);
249 define_ibdev_printk_level(ibdev_err, KERN_ERR);
250 define_ibdev_printk_level(ibdev_warn, KERN_WARNING);
251 define_ibdev_printk_level(ibdev_notice, KERN_NOTICE);
252 define_ibdev_printk_level(ibdev_info, KERN_INFO);
254 static struct notifier_block ibdev_lsm_nb = {
255 .notifier_call = ib_security_change,
258 static int rdma_dev_change_netns(struct ib_device *device, struct net *cur_net,
261 /* Pointer to the RCU head at the start of the ib_port_data array */
262 struct ib_port_data_rcu {
263 struct rcu_head rcu_head;
264 struct ib_port_data pdata[];
267 static void ib_device_check_mandatory(struct ib_device *device)
269 #define IB_MANDATORY_FUNC(x) { offsetof(struct ib_device_ops, x), #x }
270 static const struct {
273 } mandatory_table[] = {
274 IB_MANDATORY_FUNC(query_device),
275 IB_MANDATORY_FUNC(query_port),
276 IB_MANDATORY_FUNC(alloc_pd),
277 IB_MANDATORY_FUNC(dealloc_pd),
278 IB_MANDATORY_FUNC(create_qp),
279 IB_MANDATORY_FUNC(modify_qp),
280 IB_MANDATORY_FUNC(destroy_qp),
281 IB_MANDATORY_FUNC(post_send),
282 IB_MANDATORY_FUNC(post_recv),
283 IB_MANDATORY_FUNC(create_cq),
284 IB_MANDATORY_FUNC(destroy_cq),
285 IB_MANDATORY_FUNC(poll_cq),
286 IB_MANDATORY_FUNC(req_notify_cq),
287 IB_MANDATORY_FUNC(get_dma_mr),
288 IB_MANDATORY_FUNC(reg_user_mr),
289 IB_MANDATORY_FUNC(dereg_mr),
290 IB_MANDATORY_FUNC(get_port_immutable)
294 device->kverbs_provider = true;
295 for (i = 0; i < ARRAY_SIZE(mandatory_table); ++i) {
296 if (!*(void **) ((void *) &device->ops +
297 mandatory_table[i].offset)) {
298 device->kverbs_provider = false;
305 * Caller must perform ib_device_put() to return the device reference count
306 * when ib_device_get_by_index() returns valid device pointer.
308 struct ib_device *ib_device_get_by_index(const struct net *net, u32 index)
310 struct ib_device *device;
312 down_read(&devices_rwsem);
313 device = xa_load(&devices, index);
315 if (!rdma_dev_access_netns(device, net)) {
320 if (!ib_device_try_get(device))
324 up_read(&devices_rwsem);
329 * ib_device_put - Release IB device reference
330 * @device: device whose reference to be released
332 * ib_device_put() releases reference to the IB device to allow it to be
333 * unregistered and eventually free.
335 void ib_device_put(struct ib_device *device)
337 if (refcount_dec_and_test(&device->refcount))
338 complete(&device->unreg_completion);
340 EXPORT_SYMBOL(ib_device_put);
342 static struct ib_device *__ib_device_get_by_name(const char *name)
344 struct ib_device *device;
347 xa_for_each (&devices, index, device)
348 if (!strcmp(name, dev_name(&device->dev)))
355 * ib_device_get_by_name - Find an IB device by name
356 * @name: The name to look for
357 * @driver_id: The driver ID that must match (RDMA_DRIVER_UNKNOWN matches all)
359 * Find and hold an ib_device by its name. The caller must call
360 * ib_device_put() on the returned pointer.
362 struct ib_device *ib_device_get_by_name(const char *name,
363 enum rdma_driver_id driver_id)
365 struct ib_device *device;
367 down_read(&devices_rwsem);
368 device = __ib_device_get_by_name(name);
369 if (device && driver_id != RDMA_DRIVER_UNKNOWN &&
370 device->ops.driver_id != driver_id)
374 if (!ib_device_try_get(device))
377 up_read(&devices_rwsem);
380 EXPORT_SYMBOL(ib_device_get_by_name);
382 static int rename_compat_devs(struct ib_device *device)
384 struct ib_core_device *cdev;
388 mutex_lock(&device->compat_devs_mutex);
389 xa_for_each (&device->compat_devs, index, cdev) {
390 ret = device_rename(&cdev->dev, dev_name(&device->dev));
393 "Fail to rename compatdev to new name %s\n",
394 dev_name(&device->dev));
398 mutex_unlock(&device->compat_devs_mutex);
402 int ib_device_rename(struct ib_device *ibdev, const char *name)
408 down_write(&devices_rwsem);
409 if (!strcmp(name, dev_name(&ibdev->dev))) {
410 up_write(&devices_rwsem);
414 if (__ib_device_get_by_name(name)) {
415 up_write(&devices_rwsem);
419 ret = device_rename(&ibdev->dev, name);
421 up_write(&devices_rwsem);
425 strscpy(ibdev->name, name, IB_DEVICE_NAME_MAX);
426 ret = rename_compat_devs(ibdev);
428 downgrade_write(&devices_rwsem);
429 down_read(&ibdev->client_data_rwsem);
430 xan_for_each_marked(&ibdev->client_data, index, client_data,
431 CLIENT_DATA_REGISTERED) {
432 struct ib_client *client = xa_load(&clients, index);
434 if (!client || !client->rename)
437 client->rename(ibdev, client_data);
439 up_read(&ibdev->client_data_rwsem);
440 up_read(&devices_rwsem);
444 int ib_device_set_dim(struct ib_device *ibdev, u8 use_dim)
448 ibdev->use_cq_dim = use_dim;
453 static int alloc_name(struct ib_device *ibdev, const char *name)
455 struct ib_device *device;
461 lockdep_assert_held_write(&devices_rwsem);
463 xa_for_each (&devices, index, device) {
464 char buf[IB_DEVICE_NAME_MAX];
466 if (sscanf(dev_name(&device->dev), name, &i) != 1)
468 if (i < 0 || i >= INT_MAX)
470 snprintf(buf, sizeof buf, name, i);
471 if (strcmp(buf, dev_name(&device->dev)) != 0)
474 rc = ida_alloc_range(&inuse, i, i, GFP_KERNEL);
479 rc = ida_alloc(&inuse, GFP_KERNEL);
483 rc = dev_set_name(&ibdev->dev, name, rc);
489 static void ib_device_release(struct device *device)
491 struct ib_device *dev = container_of(device, struct ib_device, dev);
494 WARN_ON(refcount_read(&dev->refcount));
495 if (dev->hw_stats_data)
496 ib_device_release_hw_stats(dev->hw_stats_data);
497 if (dev->port_data) {
498 ib_cache_release_one(dev);
499 ib_security_release_port_pkey_list(dev);
500 rdma_counter_release(dev);
501 kfree_rcu(container_of(dev->port_data, struct ib_port_data_rcu,
506 mutex_destroy(&dev->subdev_lock);
507 mutex_destroy(&dev->unregistration_lock);
508 mutex_destroy(&dev->compat_devs_mutex);
510 xa_destroy(&dev->compat_devs);
511 xa_destroy(&dev->client_data);
512 kfree_rcu(dev, rcu_head);
515 static int ib_device_uevent(const struct device *device,
516 struct kobj_uevent_env *env)
518 if (add_uevent_var(env, "NAME=%s", dev_name(device)))
522 * It would be nice to pass the node GUID with the event...
528 static const void *net_namespace(const struct device *d)
530 const struct ib_core_device *coredev =
531 container_of(d, struct ib_core_device, dev);
533 return read_pnet(&coredev->rdma_net);
536 static struct class ib_class = {
537 .name = "infiniband",
538 .dev_release = ib_device_release,
539 .dev_uevent = ib_device_uevent,
540 .ns_type = &net_ns_type_operations,
541 .namespace = net_namespace,
544 static void rdma_init_coredev(struct ib_core_device *coredev,
545 struct ib_device *dev, struct net *net)
547 /* This BUILD_BUG_ON is intended to catch layout change
548 * of union of ib_core_device and device.
549 * dev must be the first element as ib_core and providers
550 * driver uses it. Adding anything in ib_core_device before
551 * device will break this assumption.
553 BUILD_BUG_ON(offsetof(struct ib_device, coredev.dev) !=
554 offsetof(struct ib_device, dev));
556 coredev->dev.class = &ib_class;
557 coredev->dev.groups = dev->groups;
558 device_initialize(&coredev->dev);
559 coredev->owner = dev;
560 INIT_LIST_HEAD(&coredev->port_list);
561 write_pnet(&coredev->rdma_net, net);
565 * _ib_alloc_device - allocate an IB device struct
566 * @size:size of structure to allocate
568 * Low-level drivers should use ib_alloc_device() to allocate &struct
569 * ib_device. @size is the size of the structure to be allocated,
570 * including any private data used by the low-level driver.
571 * ib_dealloc_device() must be used to free structures allocated with
574 struct ib_device *_ib_alloc_device(size_t size)
576 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 rdma_init_coredev(&device->coredev, device, &init_net);
593 INIT_LIST_HEAD(&device->event_handler_list);
594 spin_lock_init(&device->qp_open_list_lock);
595 init_rwsem(&device->event_handler_rwsem);
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);
608 spin_lock_init(&device->cq_pools_lock);
609 for (i = 0; i < ARRAY_SIZE(device->cq_pools); i++)
610 INIT_LIST_HEAD(&device->cq_pools[i]);
612 rwlock_init(&device->cache_lock);
614 device->uverbs_cmd_mask =
615 BIT_ULL(IB_USER_VERBS_CMD_ALLOC_MW) |
616 BIT_ULL(IB_USER_VERBS_CMD_ALLOC_PD) |
617 BIT_ULL(IB_USER_VERBS_CMD_ATTACH_MCAST) |
618 BIT_ULL(IB_USER_VERBS_CMD_CLOSE_XRCD) |
619 BIT_ULL(IB_USER_VERBS_CMD_CREATE_AH) |
620 BIT_ULL(IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
621 BIT_ULL(IB_USER_VERBS_CMD_CREATE_CQ) |
622 BIT_ULL(IB_USER_VERBS_CMD_CREATE_QP) |
623 BIT_ULL(IB_USER_VERBS_CMD_CREATE_SRQ) |
624 BIT_ULL(IB_USER_VERBS_CMD_CREATE_XSRQ) |
625 BIT_ULL(IB_USER_VERBS_CMD_DEALLOC_MW) |
626 BIT_ULL(IB_USER_VERBS_CMD_DEALLOC_PD) |
627 BIT_ULL(IB_USER_VERBS_CMD_DEREG_MR) |
628 BIT_ULL(IB_USER_VERBS_CMD_DESTROY_AH) |
629 BIT_ULL(IB_USER_VERBS_CMD_DESTROY_CQ) |
630 BIT_ULL(IB_USER_VERBS_CMD_DESTROY_QP) |
631 BIT_ULL(IB_USER_VERBS_CMD_DESTROY_SRQ) |
632 BIT_ULL(IB_USER_VERBS_CMD_DETACH_MCAST) |
633 BIT_ULL(IB_USER_VERBS_CMD_GET_CONTEXT) |
634 BIT_ULL(IB_USER_VERBS_CMD_MODIFY_QP) |
635 BIT_ULL(IB_USER_VERBS_CMD_MODIFY_SRQ) |
636 BIT_ULL(IB_USER_VERBS_CMD_OPEN_QP) |
637 BIT_ULL(IB_USER_VERBS_CMD_OPEN_XRCD) |
638 BIT_ULL(IB_USER_VERBS_CMD_QUERY_DEVICE) |
639 BIT_ULL(IB_USER_VERBS_CMD_QUERY_PORT) |
640 BIT_ULL(IB_USER_VERBS_CMD_QUERY_QP) |
641 BIT_ULL(IB_USER_VERBS_CMD_QUERY_SRQ) |
642 BIT_ULL(IB_USER_VERBS_CMD_REG_MR) |
643 BIT_ULL(IB_USER_VERBS_CMD_REREG_MR) |
644 BIT_ULL(IB_USER_VERBS_CMD_RESIZE_CQ);
646 mutex_init(&device->subdev_lock);
647 INIT_LIST_HEAD(&device->subdev_list_head);
648 INIT_LIST_HEAD(&device->subdev_list);
652 EXPORT_SYMBOL(_ib_alloc_device);
655 * ib_dealloc_device - free an IB device struct
656 * @device:structure to free
658 * Free a structure allocated with ib_alloc_device().
660 void ib_dealloc_device(struct ib_device *device)
662 if (device->ops.dealloc_driver)
663 device->ops.dealloc_driver(device);
666 * ib_unregister_driver() requires all devices to remain in the xarray
667 * while their ops are callable. The last op we call is dealloc_driver
668 * above. This is needed to create a fence on op callbacks prior to
669 * allowing the driver module to unload.
671 down_write(&devices_rwsem);
672 if (xa_load(&devices, device->index) == device)
673 xa_erase(&devices, device->index);
674 up_write(&devices_rwsem);
676 /* Expedite releasing netdev references */
677 free_netdevs(device);
679 WARN_ON(!xa_empty(&device->compat_devs));
680 WARN_ON(!xa_empty(&device->client_data));
681 WARN_ON(refcount_read(&device->refcount));
682 rdma_restrack_clean(device);
683 /* Balances with device_initialize */
684 put_device(&device->dev);
686 EXPORT_SYMBOL(ib_dealloc_device);
689 * add_client_context() and remove_client_context() must be safe against
690 * parallel calls on the same device - registration/unregistration of both the
691 * device and client can be occurring in parallel.
693 * The routines need to be a fence, any caller must not return until the add
694 * or remove is fully completed.
696 static int add_client_context(struct ib_device *device,
697 struct ib_client *client)
701 if (!device->kverbs_provider && !client->no_kverbs_req)
704 down_write(&device->client_data_rwsem);
706 * So long as the client is registered hold both the client and device
707 * unregistration locks.
709 if (!refcount_inc_not_zero(&client->uses))
711 refcount_inc(&device->refcount);
714 * Another caller to add_client_context got here first and has already
715 * completely initialized context.
717 if (xa_get_mark(&device->client_data, client->client_id,
718 CLIENT_DATA_REGISTERED))
721 ret = xa_err(xa_store(&device->client_data, client->client_id, NULL,
725 downgrade_write(&device->client_data_rwsem);
727 if (client->add(device)) {
729 * If a client fails to add then the error code is
730 * ignored, but we won't call any more ops on this
733 xa_erase(&device->client_data, client->client_id);
734 up_read(&device->client_data_rwsem);
735 ib_device_put(device);
736 ib_client_put(client);
741 /* Readers shall not see a client until add has been completed */
742 xa_set_mark(&device->client_data, client->client_id,
743 CLIENT_DATA_REGISTERED);
744 up_read(&device->client_data_rwsem);
748 ib_device_put(device);
749 ib_client_put(client);
751 up_write(&device->client_data_rwsem);
755 static void remove_client_context(struct ib_device *device,
756 unsigned int client_id)
758 struct ib_client *client;
761 down_write(&device->client_data_rwsem);
762 if (!xa_get_mark(&device->client_data, client_id,
763 CLIENT_DATA_REGISTERED)) {
764 up_write(&device->client_data_rwsem);
767 client_data = xa_load(&device->client_data, client_id);
768 xa_clear_mark(&device->client_data, client_id, CLIENT_DATA_REGISTERED);
769 client = xa_load(&clients, client_id);
770 up_write(&device->client_data_rwsem);
773 * Notice we cannot be holding any exclusive locks when calling the
774 * remove callback as the remove callback can recurse back into any
775 * public functions in this module and thus try for any locks those
778 * For this reason clients and drivers should not call the
779 * unregistration functions will holdling any locks.
782 client->remove(device, client_data);
784 xa_erase(&device->client_data, client_id);
785 ib_device_put(device);
786 ib_client_put(client);
789 static int alloc_port_data(struct ib_device *device)
791 struct ib_port_data_rcu *pdata_rcu;
794 if (device->port_data)
797 /* This can only be called once the physical port range is defined */
798 if (WARN_ON(!device->phys_port_cnt))
801 /* Reserve U32_MAX so the logic to go over all the ports is sane */
802 if (WARN_ON(device->phys_port_cnt == U32_MAX))
806 * device->port_data is indexed directly by the port number to make
807 * access to this data as efficient as possible.
809 * Therefore port_data is declared as a 1 based array with potential
810 * empty slots at the beginning.
812 pdata_rcu = kzalloc(struct_size(pdata_rcu, pdata,
813 size_add(rdma_end_port(device), 1)),
818 * The rcu_head is put in front of the port data array and the stored
819 * pointer is adjusted since we never need to see that member until
822 device->port_data = pdata_rcu->pdata;
824 rdma_for_each_port (device, port) {
825 struct ib_port_data *pdata = &device->port_data[port];
827 pdata->ib_dev = device;
828 spin_lock_init(&pdata->pkey_list_lock);
829 INIT_LIST_HEAD(&pdata->pkey_list);
830 spin_lock_init(&pdata->netdev_lock);
831 INIT_HLIST_NODE(&pdata->ndev_hash_link);
836 static int verify_immutable(const struct ib_device *dev, u32 port)
838 return WARN_ON(!rdma_cap_ib_mad(dev, port) &&
839 rdma_max_mad_size(dev, port) != 0);
842 static int setup_port_data(struct ib_device *device)
847 ret = alloc_port_data(device);
851 rdma_for_each_port (device, port) {
852 struct ib_port_data *pdata = &device->port_data[port];
854 ret = device->ops.get_port_immutable(device, port,
859 if (verify_immutable(device, port))
866 * ib_port_immutable_read() - Read rdma port's immutable data
868 * @port: port number whose immutable data to read. It starts with index 1 and
869 * valid upto including rdma_end_port().
871 const struct ib_port_immutable*
872 ib_port_immutable_read(struct ib_device *dev, unsigned int port)
874 WARN_ON(!rdma_is_port_valid(dev, port));
875 return &dev->port_data[port].immutable;
877 EXPORT_SYMBOL(ib_port_immutable_read);
879 void ib_get_device_fw_str(struct ib_device *dev, char *str)
881 if (dev->ops.get_dev_fw_str)
882 dev->ops.get_dev_fw_str(dev, str);
886 EXPORT_SYMBOL(ib_get_device_fw_str);
888 static void ib_policy_change_task(struct work_struct *work)
890 struct ib_device *dev;
893 down_read(&devices_rwsem);
894 xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
897 rdma_for_each_port (dev, i) {
899 ib_get_cached_subnet_prefix(dev, i, &sp);
900 ib_security_cache_change(dev, i, sp);
903 up_read(&devices_rwsem);
906 static int ib_security_change(struct notifier_block *nb, unsigned long event,
909 if (event != LSM_POLICY_CHANGE)
912 schedule_work(&ib_policy_change_work);
913 ib_mad_agent_security_change();
918 static void compatdev_release(struct device *dev)
920 struct ib_core_device *cdev =
921 container_of(dev, struct ib_core_device, dev);
926 static int add_one_compat_dev(struct ib_device *device,
927 struct rdma_dev_net *rnet)
929 struct ib_core_device *cdev;
932 lockdep_assert_held(&rdma_nets_rwsem);
933 if (!ib_devices_shared_netns)
937 * Create and add compat device in all namespaces other than where it
938 * is currently bound to.
940 if (net_eq(read_pnet(&rnet->net),
941 read_pnet(&device->coredev.rdma_net)))
945 * The first of init_net() or ib_register_device() to take the
946 * compat_devs_mutex wins and gets to add the device. Others will wait
947 * for completion here.
949 mutex_lock(&device->compat_devs_mutex);
950 cdev = xa_load(&device->compat_devs, rnet->id);
955 ret = xa_reserve(&device->compat_devs, rnet->id, GFP_KERNEL);
959 cdev = kzalloc(sizeof(*cdev), GFP_KERNEL);
965 cdev->dev.parent = device->dev.parent;
966 rdma_init_coredev(cdev, device, read_pnet(&rnet->net));
967 cdev->dev.release = compatdev_release;
968 ret = dev_set_name(&cdev->dev, "%s", dev_name(&device->dev));
972 ret = device_add(&cdev->dev);
975 ret = ib_setup_port_attrs(cdev);
979 ret = xa_err(xa_store(&device->compat_devs, rnet->id,
984 mutex_unlock(&device->compat_devs_mutex);
988 ib_free_port_attrs(cdev);
990 device_del(&cdev->dev);
992 put_device(&cdev->dev);
994 xa_release(&device->compat_devs, rnet->id);
996 mutex_unlock(&device->compat_devs_mutex);
1000 static void remove_one_compat_dev(struct ib_device *device, u32 id)
1002 struct ib_core_device *cdev;
1004 mutex_lock(&device->compat_devs_mutex);
1005 cdev = xa_erase(&device->compat_devs, id);
1006 mutex_unlock(&device->compat_devs_mutex);
1008 ib_free_port_attrs(cdev);
1009 device_del(&cdev->dev);
1010 put_device(&cdev->dev);
1014 static void remove_compat_devs(struct ib_device *device)
1016 struct ib_core_device *cdev;
1017 unsigned long index;
1019 xa_for_each (&device->compat_devs, index, cdev)
1020 remove_one_compat_dev(device, index);
1023 static int add_compat_devs(struct ib_device *device)
1025 struct rdma_dev_net *rnet;
1026 unsigned long index;
1029 lockdep_assert_held(&devices_rwsem);
1031 down_read(&rdma_nets_rwsem);
1032 xa_for_each (&rdma_nets, index, rnet) {
1033 ret = add_one_compat_dev(device, rnet);
1037 up_read(&rdma_nets_rwsem);
1041 static void remove_all_compat_devs(void)
1043 struct ib_compat_device *cdev;
1044 struct ib_device *dev;
1045 unsigned long index;
1047 down_read(&devices_rwsem);
1048 xa_for_each (&devices, index, dev) {
1049 unsigned long c_index = 0;
1051 /* Hold nets_rwsem so that any other thread modifying this
1052 * system param can sync with this thread.
1054 down_read(&rdma_nets_rwsem);
1055 xa_for_each (&dev->compat_devs, c_index, cdev)
1056 remove_one_compat_dev(dev, c_index);
1057 up_read(&rdma_nets_rwsem);
1059 up_read(&devices_rwsem);
1062 static int add_all_compat_devs(void)
1064 struct rdma_dev_net *rnet;
1065 struct ib_device *dev;
1066 unsigned long index;
1069 down_read(&devices_rwsem);
1070 xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
1071 unsigned long net_index = 0;
1073 /* Hold nets_rwsem so that any other thread modifying this
1074 * system param can sync with this thread.
1076 down_read(&rdma_nets_rwsem);
1077 xa_for_each (&rdma_nets, net_index, rnet) {
1078 ret = add_one_compat_dev(dev, rnet);
1082 up_read(&rdma_nets_rwsem);
1084 up_read(&devices_rwsem);
1086 remove_all_compat_devs();
1090 int rdma_compatdev_set(u8 enable)
1092 struct rdma_dev_net *rnet;
1093 unsigned long index;
1096 down_write(&rdma_nets_rwsem);
1097 if (ib_devices_shared_netns == enable) {
1098 up_write(&rdma_nets_rwsem);
1102 /* enable/disable of compat devices is not supported
1103 * when more than default init_net exists.
1105 xa_for_each (&rdma_nets, index, rnet) {
1110 ib_devices_shared_netns = enable;
1111 up_write(&rdma_nets_rwsem);
1116 ret = add_all_compat_devs();
1118 remove_all_compat_devs();
1122 static void rdma_dev_exit_net(struct net *net)
1124 struct rdma_dev_net *rnet = rdma_net_to_dev_net(net);
1125 struct ib_device *dev;
1126 unsigned long index;
1129 down_write(&rdma_nets_rwsem);
1131 * Prevent the ID from being re-used and hide the id from xa_for_each.
1133 ret = xa_err(xa_store(&rdma_nets, rnet->id, NULL, GFP_KERNEL));
1135 up_write(&rdma_nets_rwsem);
1137 down_read(&devices_rwsem);
1138 xa_for_each (&devices, index, dev) {
1139 get_device(&dev->dev);
1141 * Release the devices_rwsem so that pontentially blocking
1142 * device_del, doesn't hold the devices_rwsem for too long.
1144 up_read(&devices_rwsem);
1146 remove_one_compat_dev(dev, rnet->id);
1149 * If the real device is in the NS then move it back to init.
1151 rdma_dev_change_netns(dev, net, &init_net);
1153 put_device(&dev->dev);
1154 down_read(&devices_rwsem);
1156 up_read(&devices_rwsem);
1158 rdma_nl_net_exit(rnet);
1159 xa_erase(&rdma_nets, rnet->id);
1162 static __net_init int rdma_dev_init_net(struct net *net)
1164 struct rdma_dev_net *rnet = rdma_net_to_dev_net(net);
1165 unsigned long index;
1166 struct ib_device *dev;
1169 write_pnet(&rnet->net, net);
1171 ret = rdma_nl_net_init(rnet);
1175 /* No need to create any compat devices in default init_net. */
1176 if (net_eq(net, &init_net))
1179 ret = xa_alloc(&rdma_nets, &rnet->id, rnet, xa_limit_32b, GFP_KERNEL);
1181 rdma_nl_net_exit(rnet);
1185 down_read(&devices_rwsem);
1186 xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
1187 /* Hold nets_rwsem so that netlink command cannot change
1188 * system configuration for device sharing mode.
1190 down_read(&rdma_nets_rwsem);
1191 ret = add_one_compat_dev(dev, rnet);
1192 up_read(&rdma_nets_rwsem);
1196 up_read(&devices_rwsem);
1199 rdma_dev_exit_net(net);
1205 * Assign the unique string device name and the unique device index. This is
1206 * undone by ib_dealloc_device.
1208 static int assign_name(struct ib_device *device, const char *name)
1213 down_write(&devices_rwsem);
1214 /* Assign a unique name to the device */
1215 if (strchr(name, '%'))
1216 ret = alloc_name(device, name);
1218 ret = dev_set_name(&device->dev, name);
1222 if (__ib_device_get_by_name(dev_name(&device->dev))) {
1226 strscpy(device->name, dev_name(&device->dev), IB_DEVICE_NAME_MAX);
1228 ret = xa_alloc_cyclic(&devices, &device->index, device, xa_limit_31b,
1229 &last_id, GFP_KERNEL);
1234 up_write(&devices_rwsem);
1239 * setup_device() allocates memory and sets up data that requires calling the
1240 * device ops, this is the only reason these actions are not done during
1241 * ib_alloc_device. It is undone by ib_dealloc_device().
1243 static int setup_device(struct ib_device *device)
1245 struct ib_udata uhw = {.outlen = 0, .inlen = 0};
1248 ib_device_check_mandatory(device);
1250 ret = setup_port_data(device);
1252 dev_warn(&device->dev, "Couldn't create per-port data\n");
1256 memset(&device->attrs, 0, sizeof(device->attrs));
1257 ret = device->ops.query_device(device, &device->attrs, &uhw);
1259 dev_warn(&device->dev,
1260 "Couldn't query the device attributes\n");
1267 static void disable_device(struct ib_device *device)
1271 WARN_ON(!refcount_read(&device->refcount));
1273 down_write(&devices_rwsem);
1274 xa_clear_mark(&devices, device->index, DEVICE_REGISTERED);
1275 up_write(&devices_rwsem);
1278 * Remove clients in LIFO order, see assign_client_id. This could be
1279 * more efficient if xarray learns to reverse iterate. Since no new
1280 * clients can be added to this ib_device past this point we only need
1281 * the maximum possible client_id value here.
1283 down_read(&clients_rwsem);
1284 cid = highest_client_id;
1285 up_read(&clients_rwsem);
1288 remove_client_context(device, cid);
1291 ib_cq_pool_cleanup(device);
1293 /* Pairs with refcount_set in enable_device */
1294 ib_device_put(device);
1295 wait_for_completion(&device->unreg_completion);
1298 * compat devices must be removed after device refcount drops to zero.
1299 * Otherwise init_net() may add more compatdevs after removing compat
1300 * devices and before device is disabled.
1302 remove_compat_devs(device);
1306 * An enabled device is visible to all clients and to all the public facing
1307 * APIs that return a device pointer. This always returns with a new get, even
1310 static int enable_device_and_get(struct ib_device *device)
1312 struct ib_client *client;
1313 unsigned long index;
1317 * One ref belongs to the xa and the other belongs to this
1318 * thread. This is needed to guard against parallel unregistration.
1320 refcount_set(&device->refcount, 2);
1321 down_write(&devices_rwsem);
1322 xa_set_mark(&devices, device->index, DEVICE_REGISTERED);
1325 * By using downgrade_write() we ensure that no other thread can clear
1326 * DEVICE_REGISTERED while we are completing the client setup.
1328 downgrade_write(&devices_rwsem);
1330 if (device->ops.enable_driver) {
1331 ret = device->ops.enable_driver(device);
1336 down_read(&clients_rwsem);
1337 xa_for_each_marked (&clients, index, client, CLIENT_REGISTERED) {
1338 ret = add_client_context(device, client);
1342 up_read(&clients_rwsem);
1344 ret = add_compat_devs(device);
1346 up_read(&devices_rwsem);
1350 static void prevent_dealloc_device(struct ib_device *ib_dev)
1355 * ib_register_device - Register an IB device with IB core
1356 * @device: Device to register
1357 * @name: unique string device name. This may include a '%' which will
1358 * cause a unique index to be added to the passed device name.
1359 * @dma_device: pointer to a DMA-capable device. If %NULL, then the IB
1360 * device will be used. In this case the caller should fully
1361 * setup the ibdev for DMA. This usually means using dma_virt_ops.
1363 * Low-level drivers use ib_register_device() to register their
1364 * devices with the IB core. All registered clients will receive a
1365 * callback for each device that is added. @device must be allocated
1366 * with ib_alloc_device().
1368 * If the driver uses ops.dealloc_driver and calls any ib_unregister_device()
1369 * asynchronously then the device pointer may become freed as soon as this
1372 int ib_register_device(struct ib_device *device, const char *name,
1373 struct device *dma_device)
1377 ret = assign_name(device, name);
1382 * If the caller does not provide a DMA capable device then the IB core
1383 * will set up ib_sge and scatterlist structures that stash the kernel
1384 * virtual address into the address field.
1386 WARN_ON(dma_device && !dma_device->dma_parms);
1387 device->dma_device = dma_device;
1389 ret = setup_device(device);
1393 ret = ib_cache_setup_one(device);
1395 dev_warn(&device->dev,
1396 "Couldn't set up InfiniBand P_Key/GID cache\n");
1400 device->groups[0] = &ib_dev_attr_group;
1401 device->groups[1] = device->ops.device_group;
1402 ret = ib_setup_device_attrs(device);
1406 ib_device_register_rdmacg(device);
1408 rdma_counter_init(device);
1411 * Ensure that ADD uevent is not fired because it
1412 * is too early amd device is not initialized yet.
1414 dev_set_uevent_suppress(&device->dev, true);
1415 ret = device_add(&device->dev);
1419 ret = ib_setup_port_attrs(&device->coredev);
1421 dev_warn(&device->dev,
1422 "Couldn't register device with driver model\n");
1426 ret = enable_device_and_get(device);
1428 void (*dealloc_fn)(struct ib_device *);
1431 * If we hit this error flow then we don't want to
1432 * automatically dealloc the device since the caller is
1433 * expected to call ib_dealloc_device() after
1434 * ib_register_device() fails. This is tricky due to the
1435 * possibility for a parallel unregistration along with this
1436 * error flow. Since we have a refcount here we know any
1437 * parallel flow is stopped in disable_device and will see the
1438 * special dealloc_driver pointer, causing the responsibility to
1439 * ib_dealloc_device() to revert back to this thread.
1441 dealloc_fn = device->ops.dealloc_driver;
1442 device->ops.dealloc_driver = prevent_dealloc_device;
1443 ib_device_put(device);
1444 __ib_unregister_device(device);
1445 device->ops.dealloc_driver = dealloc_fn;
1446 dev_set_uevent_suppress(&device->dev, false);
1449 dev_set_uevent_suppress(&device->dev, false);
1450 /* Mark for userspace that device is ready */
1451 kobject_uevent(&device->dev.kobj, KOBJ_ADD);
1452 ib_device_put(device);
1457 device_del(&device->dev);
1459 dev_set_uevent_suppress(&device->dev, false);
1460 ib_device_unregister_rdmacg(device);
1462 ib_cache_cleanup_one(device);
1465 EXPORT_SYMBOL(ib_register_device);
1467 /* Callers must hold a get on the device. */
1468 static void __ib_unregister_device(struct ib_device *ib_dev)
1470 struct ib_device *sub, *tmp;
1472 mutex_lock(&ib_dev->subdev_lock);
1473 list_for_each_entry_safe_reverse(sub, tmp,
1474 &ib_dev->subdev_list_head,
1476 list_del(&sub->subdev_list);
1477 ib_dev->ops.del_sub_dev(sub);
1478 ib_device_put(ib_dev);
1480 mutex_unlock(&ib_dev->subdev_lock);
1483 * We have a registration lock so that all the calls to unregister are
1484 * fully fenced, once any unregister returns the device is truely
1485 * unregistered even if multiple callers are unregistering it at the
1486 * same time. This also interacts with the registration flow and
1487 * provides sane semantics if register and unregister are racing.
1489 mutex_lock(&ib_dev->unregistration_lock);
1490 if (!refcount_read(&ib_dev->refcount))
1493 disable_device(ib_dev);
1495 /* Expedite removing unregistered pointers from the hash table */
1496 free_netdevs(ib_dev);
1498 ib_free_port_attrs(&ib_dev->coredev);
1499 device_del(&ib_dev->dev);
1500 ib_device_unregister_rdmacg(ib_dev);
1501 ib_cache_cleanup_one(ib_dev);
1504 * Drivers using the new flow may not call ib_dealloc_device except
1505 * in error unwind prior to registration success.
1507 if (ib_dev->ops.dealloc_driver &&
1508 ib_dev->ops.dealloc_driver != prevent_dealloc_device) {
1509 WARN_ON(kref_read(&ib_dev->dev.kobj.kref) <= 1);
1510 ib_dealloc_device(ib_dev);
1513 mutex_unlock(&ib_dev->unregistration_lock);
1517 * ib_unregister_device - Unregister an IB device
1518 * @ib_dev: The device to unregister
1520 * Unregister an IB device. All clients will receive a remove callback.
1522 * Callers should call this routine only once, and protect against races with
1523 * registration. Typically it should only be called as part of a remove
1524 * callback in an implementation of driver core's struct device_driver and
1527 * If ops.dealloc_driver is used then ib_dev will be freed upon return from
1530 void ib_unregister_device(struct ib_device *ib_dev)
1532 get_device(&ib_dev->dev);
1533 __ib_unregister_device(ib_dev);
1534 put_device(&ib_dev->dev);
1536 EXPORT_SYMBOL(ib_unregister_device);
1539 * ib_unregister_device_and_put - Unregister a device while holding a 'get'
1540 * @ib_dev: The device to unregister
1542 * This is the same as ib_unregister_device(), except it includes an internal
1543 * ib_device_put() that should match a 'get' obtained by the caller.
1545 * It is safe to call this routine concurrently from multiple threads while
1546 * holding the 'get'. When the function returns the device is fully
1549 * Drivers using this flow MUST use the driver_unregister callback to clean up
1550 * their resources associated with the device and dealloc it.
1552 void ib_unregister_device_and_put(struct ib_device *ib_dev)
1554 WARN_ON(!ib_dev->ops.dealloc_driver);
1555 get_device(&ib_dev->dev);
1556 ib_device_put(ib_dev);
1557 __ib_unregister_device(ib_dev);
1558 put_device(&ib_dev->dev);
1560 EXPORT_SYMBOL(ib_unregister_device_and_put);
1563 * ib_unregister_driver - Unregister all IB devices for a driver
1564 * @driver_id: The driver to unregister
1566 * This implements a fence for device unregistration. It only returns once all
1567 * devices associated with the driver_id have fully completed their
1568 * unregistration and returned from ib_unregister_device*().
1570 * If device's are not yet unregistered it goes ahead and starts unregistering
1573 * This does not block creation of new devices with the given driver_id, that
1574 * is the responsibility of the caller.
1576 void ib_unregister_driver(enum rdma_driver_id driver_id)
1578 struct ib_device *ib_dev;
1579 unsigned long index;
1581 down_read(&devices_rwsem);
1582 xa_for_each (&devices, index, ib_dev) {
1583 if (ib_dev->ops.driver_id != driver_id)
1586 get_device(&ib_dev->dev);
1587 up_read(&devices_rwsem);
1589 WARN_ON(!ib_dev->ops.dealloc_driver);
1590 __ib_unregister_device(ib_dev);
1592 put_device(&ib_dev->dev);
1593 down_read(&devices_rwsem);
1595 up_read(&devices_rwsem);
1597 EXPORT_SYMBOL(ib_unregister_driver);
1599 static void ib_unregister_work(struct work_struct *work)
1601 struct ib_device *ib_dev =
1602 container_of(work, struct ib_device, unregistration_work);
1604 __ib_unregister_device(ib_dev);
1605 put_device(&ib_dev->dev);
1609 * ib_unregister_device_queued - Unregister a device using a work queue
1610 * @ib_dev: The device to unregister
1612 * This schedules an asynchronous unregistration using a WQ for the device. A
1613 * driver should use this to avoid holding locks while doing unregistration,
1614 * such as holding the RTNL lock.
1616 * Drivers using this API must use ib_unregister_driver before module unload
1617 * to ensure that all scheduled unregistrations have completed.
1619 void ib_unregister_device_queued(struct ib_device *ib_dev)
1621 WARN_ON(!refcount_read(&ib_dev->refcount));
1622 WARN_ON(!ib_dev->ops.dealloc_driver);
1623 get_device(&ib_dev->dev);
1624 if (!queue_work(ib_unreg_wq, &ib_dev->unregistration_work))
1625 put_device(&ib_dev->dev);
1627 EXPORT_SYMBOL(ib_unregister_device_queued);
1630 * The caller must pass in a device that has the kref held and the refcount
1631 * released. If the device is in cur_net and still registered then it is moved
1634 static int rdma_dev_change_netns(struct ib_device *device, struct net *cur_net,
1640 mutex_lock(&device->unregistration_lock);
1643 * If a device not under ib_device_get() or if the unregistration_lock
1644 * is not held, the namespace can be changed, or it can be unregistered.
1645 * Check again under the lock.
1647 if (refcount_read(&device->refcount) == 0 ||
1648 !net_eq(cur_net, read_pnet(&device->coredev.rdma_net))) {
1653 kobject_uevent(&device->dev.kobj, KOBJ_REMOVE);
1654 disable_device(device);
1657 * At this point no one can be using the device, so it is safe to
1658 * change the namespace.
1660 write_pnet(&device->coredev.rdma_net, net);
1662 down_read(&devices_rwsem);
1664 * Currently rdma devices are system wide unique. So the device name
1665 * is guaranteed free in the new namespace. Publish the new namespace
1666 * at the sysfs level.
1668 ret = device_rename(&device->dev, dev_name(&device->dev));
1669 up_read(&devices_rwsem);
1671 dev_warn(&device->dev,
1672 "%s: Couldn't rename device after namespace change\n",
1674 /* Try and put things back and re-enable the device */
1675 write_pnet(&device->coredev.rdma_net, cur_net);
1678 ret2 = enable_device_and_get(device);
1681 * This shouldn't really happen, but if it does, let the user
1682 * retry at later point. So don't disable the device.
1684 dev_warn(&device->dev,
1685 "%s: Couldn't re-enable device after namespace change\n",
1688 kobject_uevent(&device->dev.kobj, KOBJ_ADD);
1690 ib_device_put(device);
1692 mutex_unlock(&device->unregistration_lock);
1698 int ib_device_set_netns_put(struct sk_buff *skb,
1699 struct ib_device *dev, u32 ns_fd)
1704 net = get_net_ns_by_fd(ns_fd);
1710 if (!netlink_ns_capable(skb, net->user_ns, CAP_NET_ADMIN)) {
1716 * All the ib_clients, including uverbs, are reset when the namespace is
1717 * changed and this cannot be blocked waiting for userspace to do
1718 * something, so disassociation is mandatory.
1720 if (!dev->ops.disassociate_ucontext || ib_devices_shared_netns) {
1725 get_device(&dev->dev);
1727 ret = rdma_dev_change_netns(dev, current->nsproxy->net_ns, net);
1728 put_device(&dev->dev);
1740 static struct pernet_operations rdma_dev_net_ops = {
1741 .init = rdma_dev_init_net,
1742 .exit = rdma_dev_exit_net,
1743 .id = &rdma_dev_net_id,
1744 .size = sizeof(struct rdma_dev_net),
1747 static int assign_client_id(struct ib_client *client)
1751 lockdep_assert_held(&clients_rwsem);
1753 * The add/remove callbacks must be called in FIFO/LIFO order. To
1754 * achieve this we assign client_ids so they are sorted in
1755 * registration order.
1757 client->client_id = highest_client_id;
1758 ret = xa_insert(&clients, client->client_id, client, GFP_KERNEL);
1762 highest_client_id++;
1763 xa_set_mark(&clients, client->client_id, CLIENT_REGISTERED);
1767 static void remove_client_id(struct ib_client *client)
1769 down_write(&clients_rwsem);
1770 xa_erase(&clients, client->client_id);
1771 for (; highest_client_id; highest_client_id--)
1772 if (xa_load(&clients, highest_client_id - 1))
1774 up_write(&clients_rwsem);
1778 * ib_register_client - Register an IB client
1779 * @client:Client to register
1781 * Upper level users of the IB drivers can use ib_register_client() to
1782 * register callbacks for IB device addition and removal. When an IB
1783 * device is added, each registered client's add method will be called
1784 * (in the order the clients were registered), and when a device is
1785 * removed, each client's remove method will be called (in the reverse
1786 * order that clients were registered). In addition, when
1787 * ib_register_client() is called, the client will receive an add
1788 * callback for all devices already registered.
1790 int ib_register_client(struct ib_client *client)
1792 struct ib_device *device;
1793 unsigned long index;
1794 bool need_unreg = false;
1797 refcount_set(&client->uses, 1);
1798 init_completion(&client->uses_zero);
1801 * The devices_rwsem is held in write mode to ensure that a racing
1802 * ib_register_device() sees a consisent view of clients and devices.
1804 down_write(&devices_rwsem);
1805 down_write(&clients_rwsem);
1806 ret = assign_client_id(client);
1811 xa_for_each_marked (&devices, index, device, DEVICE_REGISTERED) {
1812 ret = add_client_context(device, client);
1818 up_write(&clients_rwsem);
1819 up_write(&devices_rwsem);
1820 if (need_unreg && ret)
1821 ib_unregister_client(client);
1824 EXPORT_SYMBOL(ib_register_client);
1827 * ib_unregister_client - Unregister an IB client
1828 * @client:Client to unregister
1830 * Upper level users use ib_unregister_client() to remove their client
1831 * registration. When ib_unregister_client() is called, the client
1832 * will receive a remove callback for each IB device still registered.
1834 * This is a full fence, once it returns no client callbacks will be called,
1835 * or are running in another thread.
1837 void ib_unregister_client(struct ib_client *client)
1839 struct ib_device *device;
1840 unsigned long index;
1842 down_write(&clients_rwsem);
1843 ib_client_put(client);
1844 xa_clear_mark(&clients, client->client_id, CLIENT_REGISTERED);
1845 up_write(&clients_rwsem);
1847 /* We do not want to have locks while calling client->remove() */
1849 xa_for_each (&devices, index, device) {
1850 if (!ib_device_try_get(device))
1854 remove_client_context(device, client->client_id);
1856 ib_device_put(device);
1862 * remove_client_context() is not a fence, it can return even though a
1863 * removal is ongoing. Wait until all removals are completed.
1865 wait_for_completion(&client->uses_zero);
1866 remove_client_id(client);
1868 EXPORT_SYMBOL(ib_unregister_client);
1870 static int __ib_get_global_client_nl_info(const char *client_name,
1871 struct ib_client_nl_info *res)
1873 struct ib_client *client;
1874 unsigned long index;
1877 down_read(&clients_rwsem);
1878 xa_for_each_marked (&clients, index, client, CLIENT_REGISTERED) {
1879 if (strcmp(client->name, client_name) != 0)
1881 if (!client->get_global_nl_info) {
1885 ret = client->get_global_nl_info(res);
1886 if (WARN_ON(ret == -ENOENT))
1888 if (!ret && res->cdev)
1889 get_device(res->cdev);
1892 up_read(&clients_rwsem);
1896 static int __ib_get_client_nl_info(struct ib_device *ibdev,
1897 const char *client_name,
1898 struct ib_client_nl_info *res)
1900 unsigned long index;
1904 down_read(&ibdev->client_data_rwsem);
1905 xan_for_each_marked (&ibdev->client_data, index, client_data,
1906 CLIENT_DATA_REGISTERED) {
1907 struct ib_client *client = xa_load(&clients, index);
1909 if (!client || strcmp(client->name, client_name) != 0)
1911 if (!client->get_nl_info) {
1915 ret = client->get_nl_info(ibdev, client_data, res);
1916 if (WARN_ON(ret == -ENOENT))
1920 * The cdev is guaranteed valid as long as we are inside the
1921 * client_data_rwsem as remove_one can't be called. Keep it
1922 * valid for the caller.
1924 if (!ret && res->cdev)
1925 get_device(res->cdev);
1928 up_read(&ibdev->client_data_rwsem);
1934 * ib_get_client_nl_info - Fetch the nl_info from a client
1936 * @client_name: Name of the client
1937 * @res: Result of the query
1939 int ib_get_client_nl_info(struct ib_device *ibdev, const char *client_name,
1940 struct ib_client_nl_info *res)
1945 ret = __ib_get_client_nl_info(ibdev, client_name, res);
1947 ret = __ib_get_global_client_nl_info(client_name, res);
1948 #ifdef CONFIG_MODULES
1949 if (ret == -ENOENT) {
1950 request_module("rdma-client-%s", client_name);
1952 ret = __ib_get_client_nl_info(ibdev, client_name, res);
1954 ret = __ib_get_global_client_nl_info(client_name, res);
1963 if (WARN_ON(!res->cdev))
1969 * ib_set_client_data - Set IB client context
1970 * @device:Device to set context for
1971 * @client:Client to set context for
1972 * @data:Context to set
1974 * ib_set_client_data() sets client context data that can be retrieved with
1975 * ib_get_client_data(). This can only be called while the client is
1976 * registered to the device, once the ib_client remove() callback returns this
1979 void ib_set_client_data(struct ib_device *device, struct ib_client *client,
1984 if (WARN_ON(IS_ERR(data)))
1987 rc = xa_store(&device->client_data, client->client_id, data,
1989 WARN_ON(xa_is_err(rc));
1991 EXPORT_SYMBOL(ib_set_client_data);
1994 * ib_register_event_handler - Register an IB event handler
1995 * @event_handler:Handler to register
1997 * ib_register_event_handler() registers an event handler that will be
1998 * called back when asynchronous IB events occur (as defined in
1999 * chapter 11 of the InfiniBand Architecture Specification). This
2000 * callback occurs in workqueue context.
2002 void ib_register_event_handler(struct ib_event_handler *event_handler)
2004 down_write(&event_handler->device->event_handler_rwsem);
2005 list_add_tail(&event_handler->list,
2006 &event_handler->device->event_handler_list);
2007 up_write(&event_handler->device->event_handler_rwsem);
2009 EXPORT_SYMBOL(ib_register_event_handler);
2012 * ib_unregister_event_handler - Unregister an event handler
2013 * @event_handler:Handler to unregister
2015 * Unregister an event handler registered with
2016 * ib_register_event_handler().
2018 void ib_unregister_event_handler(struct ib_event_handler *event_handler)
2020 down_write(&event_handler->device->event_handler_rwsem);
2021 list_del(&event_handler->list);
2022 up_write(&event_handler->device->event_handler_rwsem);
2024 EXPORT_SYMBOL(ib_unregister_event_handler);
2026 void ib_dispatch_event_clients(struct ib_event *event)
2028 struct ib_event_handler *handler;
2030 down_read(&event->device->event_handler_rwsem);
2032 list_for_each_entry(handler, &event->device->event_handler_list, list)
2033 handler->handler(handler, event);
2035 up_read(&event->device->event_handler_rwsem);
2038 static int iw_query_port(struct ib_device *device,
2040 struct ib_port_attr *port_attr)
2042 struct in_device *inetdev;
2043 struct net_device *netdev;
2045 memset(port_attr, 0, sizeof(*port_attr));
2047 netdev = ib_device_get_netdev(device, port_num);
2051 port_attr->max_mtu = IB_MTU_4096;
2052 port_attr->active_mtu = ib_mtu_int_to_enum(netdev->mtu);
2054 if (!netif_carrier_ok(netdev)) {
2055 port_attr->state = IB_PORT_DOWN;
2056 port_attr->phys_state = IB_PORT_PHYS_STATE_DISABLED;
2059 inetdev = __in_dev_get_rcu(netdev);
2061 if (inetdev && inetdev->ifa_list) {
2062 port_attr->state = IB_PORT_ACTIVE;
2063 port_attr->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
2065 port_attr->state = IB_PORT_INIT;
2066 port_attr->phys_state =
2067 IB_PORT_PHYS_STATE_PORT_CONFIGURATION_TRAINING;
2074 return device->ops.query_port(device, port_num, port_attr);
2077 static int __ib_query_port(struct ib_device *device,
2079 struct ib_port_attr *port_attr)
2083 memset(port_attr, 0, sizeof(*port_attr));
2085 err = device->ops.query_port(device, port_num, port_attr);
2086 if (err || port_attr->subnet_prefix)
2089 if (rdma_port_get_link_layer(device, port_num) !=
2090 IB_LINK_LAYER_INFINIBAND)
2093 ib_get_cached_subnet_prefix(device, port_num,
2094 &port_attr->subnet_prefix);
2099 * ib_query_port - Query IB port attributes
2100 * @device:Device to query
2101 * @port_num:Port number to query
2102 * @port_attr:Port attributes
2104 * ib_query_port() returns the attributes of a port through the
2105 * @port_attr pointer.
2107 int ib_query_port(struct ib_device *device,
2109 struct ib_port_attr *port_attr)
2111 if (!rdma_is_port_valid(device, port_num))
2114 if (rdma_protocol_iwarp(device, port_num))
2115 return iw_query_port(device, port_num, port_attr);
2117 return __ib_query_port(device, port_num, port_attr);
2119 EXPORT_SYMBOL(ib_query_port);
2121 static void add_ndev_hash(struct ib_port_data *pdata)
2123 unsigned long flags;
2127 spin_lock_irqsave(&ndev_hash_lock, flags);
2128 if (hash_hashed(&pdata->ndev_hash_link)) {
2129 hash_del_rcu(&pdata->ndev_hash_link);
2130 spin_unlock_irqrestore(&ndev_hash_lock, flags);
2132 * We cannot do hash_add_rcu after a hash_del_rcu until the
2136 spin_lock_irqsave(&ndev_hash_lock, flags);
2139 hash_add_rcu(ndev_hash, &pdata->ndev_hash_link,
2140 (uintptr_t)pdata->netdev);
2141 spin_unlock_irqrestore(&ndev_hash_lock, flags);
2145 * ib_device_set_netdev - Associate the ib_dev with an underlying net_device
2146 * @ib_dev: Device to modify
2147 * @ndev: net_device to affiliate, may be NULL
2148 * @port: IB port the net_device is connected to
2150 * Drivers should use this to link the ib_device to a netdev so the netdev
2151 * shows up in interfaces like ib_enum_roce_netdev. Only one netdev may be
2152 * affiliated with any port.
2154 * The caller must ensure that the given ndev is not unregistered or
2155 * unregistering, and that either the ib_device is unregistered or
2156 * ib_device_set_netdev() is called with NULL when the ndev sends a
2157 * NETDEV_UNREGISTER event.
2159 int ib_device_set_netdev(struct ib_device *ib_dev, struct net_device *ndev,
2162 struct net_device *old_ndev;
2163 struct ib_port_data *pdata;
2164 unsigned long flags;
2167 if (!rdma_is_port_valid(ib_dev, port))
2171 * Drivers wish to call this before ib_register_driver, so we have to
2172 * setup the port data early.
2174 ret = alloc_port_data(ib_dev);
2178 pdata = &ib_dev->port_data[port];
2179 spin_lock_irqsave(&pdata->netdev_lock, flags);
2180 old_ndev = rcu_dereference_protected(
2181 pdata->netdev, lockdep_is_held(&pdata->netdev_lock));
2182 if (old_ndev == ndev) {
2183 spin_unlock_irqrestore(&pdata->netdev_lock, flags);
2187 rcu_assign_pointer(pdata->netdev, ndev);
2188 netdev_put(old_ndev, &pdata->netdev_tracker);
2189 netdev_hold(ndev, &pdata->netdev_tracker, GFP_ATOMIC);
2190 spin_unlock_irqrestore(&pdata->netdev_lock, flags);
2192 add_ndev_hash(pdata);
2195 EXPORT_SYMBOL(ib_device_set_netdev);
2197 static void free_netdevs(struct ib_device *ib_dev)
2199 unsigned long flags;
2202 if (!ib_dev->port_data)
2205 rdma_for_each_port (ib_dev, port) {
2206 struct ib_port_data *pdata = &ib_dev->port_data[port];
2207 struct net_device *ndev;
2209 spin_lock_irqsave(&pdata->netdev_lock, flags);
2210 ndev = rcu_dereference_protected(
2211 pdata->netdev, lockdep_is_held(&pdata->netdev_lock));
2213 spin_lock(&ndev_hash_lock);
2214 hash_del_rcu(&pdata->ndev_hash_link);
2215 spin_unlock(&ndev_hash_lock);
2218 * If this is the last dev_put there is still a
2219 * synchronize_rcu before the netdev is kfreed, so we
2220 * can continue to rely on unlocked pointer
2221 * comparisons after the put
2223 rcu_assign_pointer(pdata->netdev, NULL);
2224 netdev_put(ndev, &pdata->netdev_tracker);
2226 spin_unlock_irqrestore(&pdata->netdev_lock, flags);
2230 struct net_device *ib_device_get_netdev(struct ib_device *ib_dev,
2233 struct ib_port_data *pdata;
2234 struct net_device *res;
2236 if (!rdma_is_port_valid(ib_dev, port))
2239 pdata = &ib_dev->port_data[port];
2242 * New drivers should use ib_device_set_netdev() not the legacy
2245 if (ib_dev->ops.get_netdev)
2246 res = ib_dev->ops.get_netdev(ib_dev, port);
2248 spin_lock(&pdata->netdev_lock);
2249 res = rcu_dereference_protected(
2250 pdata->netdev, lockdep_is_held(&pdata->netdev_lock));
2252 spin_unlock(&pdata->netdev_lock);
2256 * If we are starting to unregister expedite things by preventing
2257 * propagation of an unregistering netdev.
2259 if (res && res->reg_state != NETREG_REGISTERED) {
2268 * ib_device_get_by_netdev - Find an IB device associated with a netdev
2269 * @ndev: netdev to locate
2270 * @driver_id: The driver ID that must match (RDMA_DRIVER_UNKNOWN matches all)
2272 * Find and hold an ib_device that is associated with a netdev via
2273 * ib_device_set_netdev(). The caller must call ib_device_put() on the
2276 struct ib_device *ib_device_get_by_netdev(struct net_device *ndev,
2277 enum rdma_driver_id driver_id)
2279 struct ib_device *res = NULL;
2280 struct ib_port_data *cur;
2283 hash_for_each_possible_rcu (ndev_hash, cur, ndev_hash_link,
2285 if (rcu_access_pointer(cur->netdev) == ndev &&
2286 (driver_id == RDMA_DRIVER_UNKNOWN ||
2287 cur->ib_dev->ops.driver_id == driver_id) &&
2288 ib_device_try_get(cur->ib_dev)) {
2297 EXPORT_SYMBOL(ib_device_get_by_netdev);
2300 * ib_enum_roce_netdev - enumerate all RoCE ports
2301 * @ib_dev : IB device we want to query
2302 * @filter: Should we call the callback?
2303 * @filter_cookie: Cookie passed to filter
2304 * @cb: Callback to call for each found RoCE ports
2305 * @cookie: Cookie passed back to the callback
2307 * Enumerates all of the physical RoCE ports of ib_dev
2308 * which are related to netdevice and calls callback() on each
2309 * device for which filter() function returns non zero.
2311 void ib_enum_roce_netdev(struct ib_device *ib_dev,
2312 roce_netdev_filter filter,
2313 void *filter_cookie,
2314 roce_netdev_callback cb,
2319 rdma_for_each_port (ib_dev, port)
2320 if (rdma_protocol_roce(ib_dev, port)) {
2321 struct net_device *idev =
2322 ib_device_get_netdev(ib_dev, port);
2324 if (filter(ib_dev, port, idev, filter_cookie))
2325 cb(ib_dev, port, idev, cookie);
2331 * ib_enum_all_roce_netdevs - enumerate all RoCE devices
2332 * @filter: Should we call the callback?
2333 * @filter_cookie: Cookie passed to filter
2334 * @cb: Callback to call for each found RoCE ports
2335 * @cookie: Cookie passed back to the callback
2337 * Enumerates all RoCE devices' physical ports which are related
2338 * to netdevices and calls callback() on each device for which
2339 * filter() function returns non zero.
2341 void ib_enum_all_roce_netdevs(roce_netdev_filter filter,
2342 void *filter_cookie,
2343 roce_netdev_callback cb,
2346 struct ib_device *dev;
2347 unsigned long index;
2349 down_read(&devices_rwsem);
2350 xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED)
2351 ib_enum_roce_netdev(dev, filter, filter_cookie, cb, cookie);
2352 up_read(&devices_rwsem);
2356 * ib_enum_all_devs - enumerate all ib_devices
2357 * @cb: Callback to call for each found ib_device
2359 * Enumerates all ib_devices and calls callback() on each device.
2361 int ib_enum_all_devs(nldev_callback nldev_cb, struct sk_buff *skb,
2362 struct netlink_callback *cb)
2364 unsigned long index;
2365 struct ib_device *dev;
2366 unsigned int idx = 0;
2369 down_read(&devices_rwsem);
2370 xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
2371 if (!rdma_dev_access_netns(dev, sock_net(skb->sk)))
2374 ret = nldev_cb(dev, skb, cb, idx);
2379 up_read(&devices_rwsem);
2384 * ib_query_pkey - Get P_Key table entry
2385 * @device:Device to query
2386 * @port_num:Port number to query
2387 * @index:P_Key table index to query
2388 * @pkey:Returned P_Key
2390 * ib_query_pkey() fetches the specified P_Key table entry.
2392 int ib_query_pkey(struct ib_device *device,
2393 u32 port_num, u16 index, u16 *pkey)
2395 if (!rdma_is_port_valid(device, port_num))
2398 if (!device->ops.query_pkey)
2401 return device->ops.query_pkey(device, port_num, index, pkey);
2403 EXPORT_SYMBOL(ib_query_pkey);
2406 * ib_modify_device - Change IB device attributes
2407 * @device:Device to modify
2408 * @device_modify_mask:Mask of attributes to change
2409 * @device_modify:New attribute values
2411 * ib_modify_device() changes a device's attributes as specified by
2412 * the @device_modify_mask and @device_modify structure.
2414 int ib_modify_device(struct ib_device *device,
2415 int device_modify_mask,
2416 struct ib_device_modify *device_modify)
2418 if (!device->ops.modify_device)
2421 return device->ops.modify_device(device, device_modify_mask,
2424 EXPORT_SYMBOL(ib_modify_device);
2427 * ib_modify_port - Modifies the attributes for the specified port.
2428 * @device: The device to modify.
2429 * @port_num: The number of the port to modify.
2430 * @port_modify_mask: Mask used to specify which attributes of the port
2432 * @port_modify: New attribute values for the port.
2434 * ib_modify_port() changes a port's attributes as specified by the
2435 * @port_modify_mask and @port_modify structure.
2437 int ib_modify_port(struct ib_device *device,
2438 u32 port_num, int port_modify_mask,
2439 struct ib_port_modify *port_modify)
2443 if (!rdma_is_port_valid(device, port_num))
2446 if (device->ops.modify_port)
2447 rc = device->ops.modify_port(device, port_num,
2450 else if (rdma_protocol_roce(device, port_num) &&
2451 ((port_modify->set_port_cap_mask & ~IB_PORT_CM_SUP) == 0 ||
2452 (port_modify->clr_port_cap_mask & ~IB_PORT_CM_SUP) == 0))
2458 EXPORT_SYMBOL(ib_modify_port);
2461 * ib_find_gid - Returns the port number and GID table index where
2462 * a specified GID value occurs. Its searches only for IB link layer.
2463 * @device: The device to query.
2464 * @gid: The GID value to search for.
2465 * @port_num: The port number of the device where the GID value was found.
2466 * @index: The index into the GID table where the GID was found. This
2467 * parameter may be NULL.
2469 int ib_find_gid(struct ib_device *device, union ib_gid *gid,
2470 u32 *port_num, u16 *index)
2472 union ib_gid tmp_gid;
2476 rdma_for_each_port (device, port) {
2477 if (!rdma_protocol_ib(device, port))
2480 for (i = 0; i < device->port_data[port].immutable.gid_tbl_len;
2482 ret = rdma_query_gid(device, port, i, &tmp_gid);
2486 if (!memcmp(&tmp_gid, gid, sizeof *gid)) {
2497 EXPORT_SYMBOL(ib_find_gid);
2500 * ib_find_pkey - Returns the PKey table index where a specified
2501 * PKey value occurs.
2502 * @device: The device to query.
2503 * @port_num: The port number of the device to search for the PKey.
2504 * @pkey: The PKey value to search for.
2505 * @index: The index into the PKey table where the PKey was found.
2507 int ib_find_pkey(struct ib_device *device,
2508 u32 port_num, u16 pkey, u16 *index)
2512 int partial_ix = -1;
2514 for (i = 0; i < device->port_data[port_num].immutable.pkey_tbl_len;
2516 ret = ib_query_pkey(device, port_num, i, &tmp_pkey);
2519 if ((pkey & 0x7fff) == (tmp_pkey & 0x7fff)) {
2520 /* if there is full-member pkey take it.*/
2521 if (tmp_pkey & 0x8000) {
2530 /*no full-member, if exists take the limited*/
2531 if (partial_ix >= 0) {
2532 *index = partial_ix;
2537 EXPORT_SYMBOL(ib_find_pkey);
2540 * ib_get_net_dev_by_params() - Return the appropriate net_dev
2541 * for a received CM request
2542 * @dev: An RDMA device on which the request has been received.
2543 * @port: Port number on the RDMA device.
2544 * @pkey: The Pkey the request came on.
2545 * @gid: A GID that the net_dev uses to communicate.
2546 * @addr: Contains the IP address that the request specified as its
2550 struct net_device *ib_get_net_dev_by_params(struct ib_device *dev,
2553 const union ib_gid *gid,
2554 const struct sockaddr *addr)
2556 struct net_device *net_dev = NULL;
2557 unsigned long index;
2560 if (!rdma_protocol_ib(dev, port))
2564 * Holding the read side guarantees that the client will not become
2565 * unregistered while we are calling get_net_dev_by_params()
2567 down_read(&dev->client_data_rwsem);
2568 xan_for_each_marked (&dev->client_data, index, client_data,
2569 CLIENT_DATA_REGISTERED) {
2570 struct ib_client *client = xa_load(&clients, index);
2572 if (!client || !client->get_net_dev_by_params)
2575 net_dev = client->get_net_dev_by_params(dev, port, pkey, gid,
2580 up_read(&dev->client_data_rwsem);
2584 EXPORT_SYMBOL(ib_get_net_dev_by_params);
2586 void ib_set_device_ops(struct ib_device *dev, const struct ib_device_ops *ops)
2588 struct ib_device_ops *dev_ops = &dev->ops;
2589 #define SET_DEVICE_OP(ptr, name) \
2592 if (!((ptr)->name)) \
2593 (ptr)->name = ops->name; \
2596 #define SET_OBJ_SIZE(ptr, name) SET_DEVICE_OP(ptr, size_##name)
2598 if (ops->driver_id != RDMA_DRIVER_UNKNOWN) {
2599 WARN_ON(dev_ops->driver_id != RDMA_DRIVER_UNKNOWN &&
2600 dev_ops->driver_id != ops->driver_id);
2601 dev_ops->driver_id = ops->driver_id;
2604 WARN_ON(dev_ops->owner && dev_ops->owner != ops->owner);
2605 dev_ops->owner = ops->owner;
2607 if (ops->uverbs_abi_ver)
2608 dev_ops->uverbs_abi_ver = ops->uverbs_abi_ver;
2610 dev_ops->uverbs_no_driver_id_binding |=
2611 ops->uverbs_no_driver_id_binding;
2613 SET_DEVICE_OP(dev_ops, add_gid);
2614 SET_DEVICE_OP(dev_ops, add_sub_dev);
2615 SET_DEVICE_OP(dev_ops, advise_mr);
2616 SET_DEVICE_OP(dev_ops, alloc_dm);
2617 SET_DEVICE_OP(dev_ops, alloc_hw_device_stats);
2618 SET_DEVICE_OP(dev_ops, alloc_hw_port_stats);
2619 SET_DEVICE_OP(dev_ops, alloc_mr);
2620 SET_DEVICE_OP(dev_ops, alloc_mr_integrity);
2621 SET_DEVICE_OP(dev_ops, alloc_mw);
2622 SET_DEVICE_OP(dev_ops, alloc_pd);
2623 SET_DEVICE_OP(dev_ops, alloc_rdma_netdev);
2624 SET_DEVICE_OP(dev_ops, alloc_ucontext);
2625 SET_DEVICE_OP(dev_ops, alloc_xrcd);
2626 SET_DEVICE_OP(dev_ops, attach_mcast);
2627 SET_DEVICE_OP(dev_ops, check_mr_status);
2628 SET_DEVICE_OP(dev_ops, counter_alloc_stats);
2629 SET_DEVICE_OP(dev_ops, counter_bind_qp);
2630 SET_DEVICE_OP(dev_ops, counter_dealloc);
2631 SET_DEVICE_OP(dev_ops, counter_unbind_qp);
2632 SET_DEVICE_OP(dev_ops, counter_update_stats);
2633 SET_DEVICE_OP(dev_ops, create_ah);
2634 SET_DEVICE_OP(dev_ops, create_counters);
2635 SET_DEVICE_OP(dev_ops, create_cq);
2636 SET_DEVICE_OP(dev_ops, create_flow);
2637 SET_DEVICE_OP(dev_ops, create_qp);
2638 SET_DEVICE_OP(dev_ops, create_rwq_ind_table);
2639 SET_DEVICE_OP(dev_ops, create_srq);
2640 SET_DEVICE_OP(dev_ops, create_user_ah);
2641 SET_DEVICE_OP(dev_ops, create_wq);
2642 SET_DEVICE_OP(dev_ops, dealloc_dm);
2643 SET_DEVICE_OP(dev_ops, dealloc_driver);
2644 SET_DEVICE_OP(dev_ops, dealloc_mw);
2645 SET_DEVICE_OP(dev_ops, dealloc_pd);
2646 SET_DEVICE_OP(dev_ops, dealloc_ucontext);
2647 SET_DEVICE_OP(dev_ops, dealloc_xrcd);
2648 SET_DEVICE_OP(dev_ops, del_gid);
2649 SET_DEVICE_OP(dev_ops, del_sub_dev);
2650 SET_DEVICE_OP(dev_ops, dereg_mr);
2651 SET_DEVICE_OP(dev_ops, destroy_ah);
2652 SET_DEVICE_OP(dev_ops, destroy_counters);
2653 SET_DEVICE_OP(dev_ops, destroy_cq);
2654 SET_DEVICE_OP(dev_ops, destroy_flow);
2655 SET_DEVICE_OP(dev_ops, destroy_flow_action);
2656 SET_DEVICE_OP(dev_ops, destroy_qp);
2657 SET_DEVICE_OP(dev_ops, destroy_rwq_ind_table);
2658 SET_DEVICE_OP(dev_ops, destroy_srq);
2659 SET_DEVICE_OP(dev_ops, destroy_wq);
2660 SET_DEVICE_OP(dev_ops, device_group);
2661 SET_DEVICE_OP(dev_ops, detach_mcast);
2662 SET_DEVICE_OP(dev_ops, disassociate_ucontext);
2663 SET_DEVICE_OP(dev_ops, drain_rq);
2664 SET_DEVICE_OP(dev_ops, drain_sq);
2665 SET_DEVICE_OP(dev_ops, enable_driver);
2666 SET_DEVICE_OP(dev_ops, fill_res_cm_id_entry);
2667 SET_DEVICE_OP(dev_ops, fill_res_cq_entry);
2668 SET_DEVICE_OP(dev_ops, fill_res_cq_entry_raw);
2669 SET_DEVICE_OP(dev_ops, fill_res_mr_entry);
2670 SET_DEVICE_OP(dev_ops, fill_res_mr_entry_raw);
2671 SET_DEVICE_OP(dev_ops, fill_res_qp_entry);
2672 SET_DEVICE_OP(dev_ops, fill_res_qp_entry_raw);
2673 SET_DEVICE_OP(dev_ops, fill_res_srq_entry);
2674 SET_DEVICE_OP(dev_ops, fill_res_srq_entry_raw);
2675 SET_DEVICE_OP(dev_ops, fill_stat_mr_entry);
2676 SET_DEVICE_OP(dev_ops, get_dev_fw_str);
2677 SET_DEVICE_OP(dev_ops, get_dma_mr);
2678 SET_DEVICE_OP(dev_ops, get_hw_stats);
2679 SET_DEVICE_OP(dev_ops, get_link_layer);
2680 SET_DEVICE_OP(dev_ops, get_netdev);
2681 SET_DEVICE_OP(dev_ops, get_numa_node);
2682 SET_DEVICE_OP(dev_ops, get_port_immutable);
2683 SET_DEVICE_OP(dev_ops, get_vector_affinity);
2684 SET_DEVICE_OP(dev_ops, get_vf_config);
2685 SET_DEVICE_OP(dev_ops, get_vf_guid);
2686 SET_DEVICE_OP(dev_ops, get_vf_stats);
2687 SET_DEVICE_OP(dev_ops, iw_accept);
2688 SET_DEVICE_OP(dev_ops, iw_add_ref);
2689 SET_DEVICE_OP(dev_ops, iw_connect);
2690 SET_DEVICE_OP(dev_ops, iw_create_listen);
2691 SET_DEVICE_OP(dev_ops, iw_destroy_listen);
2692 SET_DEVICE_OP(dev_ops, iw_get_qp);
2693 SET_DEVICE_OP(dev_ops, iw_reject);
2694 SET_DEVICE_OP(dev_ops, iw_rem_ref);
2695 SET_DEVICE_OP(dev_ops, map_mr_sg);
2696 SET_DEVICE_OP(dev_ops, map_mr_sg_pi);
2697 SET_DEVICE_OP(dev_ops, mmap);
2698 SET_DEVICE_OP(dev_ops, mmap_free);
2699 SET_DEVICE_OP(dev_ops, modify_ah);
2700 SET_DEVICE_OP(dev_ops, modify_cq);
2701 SET_DEVICE_OP(dev_ops, modify_device);
2702 SET_DEVICE_OP(dev_ops, modify_hw_stat);
2703 SET_DEVICE_OP(dev_ops, modify_port);
2704 SET_DEVICE_OP(dev_ops, modify_qp);
2705 SET_DEVICE_OP(dev_ops, modify_srq);
2706 SET_DEVICE_OP(dev_ops, modify_wq);
2707 SET_DEVICE_OP(dev_ops, peek_cq);
2708 SET_DEVICE_OP(dev_ops, poll_cq);
2709 SET_DEVICE_OP(dev_ops, port_groups);
2710 SET_DEVICE_OP(dev_ops, post_recv);
2711 SET_DEVICE_OP(dev_ops, post_send);
2712 SET_DEVICE_OP(dev_ops, post_srq_recv);
2713 SET_DEVICE_OP(dev_ops, process_mad);
2714 SET_DEVICE_OP(dev_ops, query_ah);
2715 SET_DEVICE_OP(dev_ops, query_device);
2716 SET_DEVICE_OP(dev_ops, query_gid);
2717 SET_DEVICE_OP(dev_ops, query_pkey);
2718 SET_DEVICE_OP(dev_ops, query_port);
2719 SET_DEVICE_OP(dev_ops, query_qp);
2720 SET_DEVICE_OP(dev_ops, query_srq);
2721 SET_DEVICE_OP(dev_ops, query_ucontext);
2722 SET_DEVICE_OP(dev_ops, rdma_netdev_get_params);
2723 SET_DEVICE_OP(dev_ops, read_counters);
2724 SET_DEVICE_OP(dev_ops, reg_dm_mr);
2725 SET_DEVICE_OP(dev_ops, reg_user_mr);
2726 SET_DEVICE_OP(dev_ops, reg_user_mr_dmabuf);
2727 SET_DEVICE_OP(dev_ops, req_notify_cq);
2728 SET_DEVICE_OP(dev_ops, rereg_user_mr);
2729 SET_DEVICE_OP(dev_ops, resize_cq);
2730 SET_DEVICE_OP(dev_ops, set_vf_guid);
2731 SET_DEVICE_OP(dev_ops, set_vf_link_state);
2733 SET_OBJ_SIZE(dev_ops, ib_ah);
2734 SET_OBJ_SIZE(dev_ops, ib_counters);
2735 SET_OBJ_SIZE(dev_ops, ib_cq);
2736 SET_OBJ_SIZE(dev_ops, ib_mw);
2737 SET_OBJ_SIZE(dev_ops, ib_pd);
2738 SET_OBJ_SIZE(dev_ops, ib_qp);
2739 SET_OBJ_SIZE(dev_ops, ib_rwq_ind_table);
2740 SET_OBJ_SIZE(dev_ops, ib_srq);
2741 SET_OBJ_SIZE(dev_ops, ib_ucontext);
2742 SET_OBJ_SIZE(dev_ops, ib_xrcd);
2744 EXPORT_SYMBOL(ib_set_device_ops);
2746 int ib_add_sub_device(struct ib_device *parent,
2747 enum rdma_nl_dev_type type,
2750 struct ib_device *sub;
2753 if (!parent->ops.add_sub_dev || !parent->ops.del_sub_dev)
2756 if (!ib_device_try_get(parent))
2759 sub = parent->ops.add_sub_dev(parent, type, name);
2761 ib_device_put(parent);
2762 return PTR_ERR(sub);
2766 sub->parent = parent;
2768 mutex_lock(&parent->subdev_lock);
2769 list_add_tail(&parent->subdev_list_head, &sub->subdev_list);
2770 mutex_unlock(&parent->subdev_lock);
2774 EXPORT_SYMBOL(ib_add_sub_device);
2776 int ib_del_sub_device_and_put(struct ib_device *sub)
2778 struct ib_device *parent = sub->parent;
2783 mutex_lock(&parent->subdev_lock);
2784 list_del(&sub->subdev_list);
2785 mutex_unlock(&parent->subdev_lock);
2788 parent->ops.del_sub_dev(sub);
2789 ib_device_put(parent);
2793 EXPORT_SYMBOL(ib_del_sub_device_and_put);
2795 #ifdef CONFIG_INFINIBAND_VIRT_DMA
2796 int ib_dma_virt_map_sg(struct ib_device *dev, struct scatterlist *sg, int nents)
2798 struct scatterlist *s;
2801 for_each_sg(sg, s, nents, i) {
2802 sg_dma_address(s) = (uintptr_t)sg_virt(s);
2803 sg_dma_len(s) = s->length;
2807 EXPORT_SYMBOL(ib_dma_virt_map_sg);
2808 #endif /* CONFIG_INFINIBAND_VIRT_DMA */
2810 static const struct rdma_nl_cbs ibnl_ls_cb_table[RDMA_NL_LS_NUM_OPS] = {
2811 [RDMA_NL_LS_OP_RESOLVE] = {
2812 .doit = ib_nl_handle_resolve_resp,
2813 .flags = RDMA_NL_ADMIN_PERM,
2815 [RDMA_NL_LS_OP_SET_TIMEOUT] = {
2816 .doit = ib_nl_handle_set_timeout,
2817 .flags = RDMA_NL_ADMIN_PERM,
2819 [RDMA_NL_LS_OP_IP_RESOLVE] = {
2820 .doit = ib_nl_handle_ip_res_resp,
2821 .flags = RDMA_NL_ADMIN_PERM,
2825 static int __init ib_core_init(void)
2829 ib_wq = alloc_workqueue("infiniband", 0, 0);
2833 ib_unreg_wq = alloc_workqueue("ib-unreg-wq", WQ_UNBOUND,
2834 WQ_UNBOUND_MAX_ACTIVE);
2838 ib_comp_wq = alloc_workqueue("ib-comp-wq",
2839 WQ_HIGHPRI | WQ_MEM_RECLAIM | WQ_SYSFS, 0);
2843 ib_comp_unbound_wq =
2844 alloc_workqueue("ib-comp-unb-wq",
2845 WQ_UNBOUND | WQ_HIGHPRI | WQ_MEM_RECLAIM |
2846 WQ_SYSFS, WQ_UNBOUND_MAX_ACTIVE);
2847 if (!ib_comp_unbound_wq)
2850 ret = class_register(&ib_class);
2852 pr_warn("Couldn't create InfiniBand device class\n");
2853 goto err_comp_unbound;
2860 pr_warn("Couldn't init IB address resolution\n");
2864 ret = ib_mad_init();
2866 pr_warn("Couldn't init IB MAD\n");
2872 pr_warn("Couldn't init SA\n");
2876 ret = register_blocking_lsm_notifier(&ibdev_lsm_nb);
2878 pr_warn("Couldn't register LSM notifier. ret %d\n", ret);
2882 ret = register_pernet_device(&rdma_dev_net_ops);
2884 pr_warn("Couldn't init compat dev. ret %d\n", ret);
2889 rdma_nl_register(RDMA_NL_LS, ibnl_ls_cb_table);
2890 ret = roce_gid_mgmt_init();
2892 pr_warn("Couldn't init RoCE GID management\n");
2899 rdma_nl_unregister(RDMA_NL_LS);
2901 unregister_pernet_device(&rdma_dev_net_ops);
2903 unregister_blocking_lsm_notifier(&ibdev_lsm_nb);
2911 class_unregister(&ib_class);
2913 destroy_workqueue(ib_comp_unbound_wq);
2915 destroy_workqueue(ib_comp_wq);
2917 destroy_workqueue(ib_unreg_wq);
2919 destroy_workqueue(ib_wq);
2923 static void __exit ib_core_cleanup(void)
2925 roce_gid_mgmt_cleanup();
2926 rdma_nl_unregister(RDMA_NL_LS);
2928 unregister_pernet_device(&rdma_dev_net_ops);
2929 unregister_blocking_lsm_notifier(&ibdev_lsm_nb);
2934 class_unregister(&ib_class);
2935 destroy_workqueue(ib_comp_unbound_wq);
2936 destroy_workqueue(ib_comp_wq);
2937 /* Make sure that any pending umem accounting work is done. */
2938 destroy_workqueue(ib_wq);
2939 destroy_workqueue(ib_unreg_wq);
2940 WARN_ON(!xa_empty(&clients));
2941 WARN_ON(!xa_empty(&devices));
2944 MODULE_ALIAS_RDMA_NETLINK(RDMA_NL_LS, 4);
2946 /* ib core relies on netdev stack to first register net_ns_type_operations
2947 * ns kobject type before ib_core initialization.
2949 fs_initcall(ib_core_init);
2950 module_exit(ib_core_cleanup);