2 * Copyright (c) 2004 Topspin Communications. All rights reserved.
3 * Copyright (c) 2005 Intel Corporation. All rights reserved.
4 * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
5 * Copyright (c) 2005 Voltaire, Inc. All rights reserved.
7 * This software is available to you under a choice of one of two
8 * licenses. You may choose to be licensed under the terms of the GNU
9 * General Public License (GPL) Version 2, available from the file
10 * COPYING in the main directory of this source tree, or the
11 * OpenIB.org BSD license below:
13 * Redistribution and use in source and binary forms, with or
14 * without modification, are permitted provided that the following
17 * - Redistributions of source code must retain the above
18 * copyright notice, this list of conditions and the following
21 * - Redistributions in binary form must reproduce the above
22 * copyright notice, this list of conditions and the following
23 * disclaimer in the documentation and/or other materials
24 * provided with the distribution.
26 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
27 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
28 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
29 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
30 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
31 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
32 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
36 #include <linux/module.h>
37 #include <linux/errno.h>
38 #include <linux/slab.h>
39 #include <linux/workqueue.h>
40 #include <linux/netdevice.h>
41 #include <net/addrconf.h>
43 #include <rdma/ib_cache.h>
45 #include "core_priv.h"
47 struct ib_pkey_cache {
52 struct ib_update_work {
53 struct work_struct work;
54 struct ib_device *device;
56 bool enforce_security;
62 enum gid_attr_find_mask {
63 GID_ATTR_FIND_MASK_GID = 1UL << 0,
64 GID_ATTR_FIND_MASK_NETDEV = 1UL << 1,
65 GID_ATTR_FIND_MASK_DEFAULT = 1UL << 2,
66 GID_ATTR_FIND_MASK_GID_TYPE = 1UL << 3,
69 enum gid_table_entry_state {
70 GID_TABLE_ENTRY_INVALID = 1,
71 GID_TABLE_ENTRY_VALID = 2,
73 * Indicates that entry is pending to be removed, there may
74 * be active users of this GID entry.
75 * When last user of the GID entry releases reference to it,
76 * GID entry is detached from the table.
78 GID_TABLE_ENTRY_PENDING_DEL = 3,
81 struct roce_gid_ndev_storage {
82 struct rcu_head rcu_head;
83 struct net_device *ndev;
86 struct ib_gid_table_entry {
88 struct work_struct del_work;
89 struct ib_gid_attr attr;
91 /* Store the ndev pointer to release reference later on in
92 * call_rcu context because by that time gid_table_entry
93 * and attr might be already freed. So keep a copy of it.
94 * ndev_storage is freed by rcu callback.
96 struct roce_gid_ndev_storage *ndev_storage;
97 enum gid_table_entry_state state;
100 struct ib_gid_table {
102 /* In RoCE, adding a GID to the table requires:
103 * (a) Find if this GID is already exists.
104 * (b) Find a free space.
105 * (c) Write the new GID
107 * Delete requires different set of operations:
112 /* Any writer to data_vec must hold this lock and the write side of
113 * rwlock. Readers must hold only rwlock. All writers must be in a
117 /* rwlock protects data_vec[ix]->state and entry pointer.
120 struct ib_gid_table_entry **data_vec;
121 /* bit field, each bit indicates the index of default GID */
122 u32 default_gid_indices;
125 static void dispatch_gid_change_event(struct ib_device *ib_dev, u8 port)
127 struct ib_event event;
129 event.device = ib_dev;
130 event.element.port_num = port;
131 event.event = IB_EVENT_GID_CHANGE;
133 ib_dispatch_event(&event);
136 static const char * const gid_type_str[] = {
137 [IB_GID_TYPE_IB] = "IB/RoCE v1",
138 [IB_GID_TYPE_ROCE_UDP_ENCAP] = "RoCE v2",
141 const char *ib_cache_gid_type_str(enum ib_gid_type gid_type)
143 if (gid_type < ARRAY_SIZE(gid_type_str) && gid_type_str[gid_type])
144 return gid_type_str[gid_type];
146 return "Invalid GID type";
148 EXPORT_SYMBOL(ib_cache_gid_type_str);
150 /** rdma_is_zero_gid - Check if given GID is zero or not.
152 * Returns true if given GID is zero, returns false otherwise.
154 bool rdma_is_zero_gid(const union ib_gid *gid)
156 return !memcmp(gid, &zgid, sizeof(*gid));
158 EXPORT_SYMBOL(rdma_is_zero_gid);
160 /** is_gid_index_default - Check if a given index belongs to
161 * reserved default GIDs or not.
162 * @table: GID table pointer
163 * @index: Index to check in GID table
164 * Returns true if index is one of the reserved default GID index otherwise
167 static bool is_gid_index_default(const struct ib_gid_table *table,
170 return index < 32 && (BIT(index) & table->default_gid_indices);
173 int ib_cache_gid_parse_type_str(const char *buf)
183 if (buf[len - 1] == '\n')
186 for (i = 0; i < ARRAY_SIZE(gid_type_str); ++i)
187 if (gid_type_str[i] && !strncmp(buf, gid_type_str[i], len) &&
188 len == strlen(gid_type_str[i])) {
195 EXPORT_SYMBOL(ib_cache_gid_parse_type_str);
197 static struct ib_gid_table *rdma_gid_table(struct ib_device *device, u8 port)
199 return device->port_data[port].cache.gid;
202 static bool is_gid_entry_free(const struct ib_gid_table_entry *entry)
207 static bool is_gid_entry_valid(const struct ib_gid_table_entry *entry)
209 return entry && entry->state == GID_TABLE_ENTRY_VALID;
212 static void schedule_free_gid(struct kref *kref)
214 struct ib_gid_table_entry *entry =
215 container_of(kref, struct ib_gid_table_entry, kref);
217 queue_work(ib_wq, &entry->del_work);
220 static void put_gid_ndev(struct rcu_head *head)
222 struct roce_gid_ndev_storage *storage =
223 container_of(head, struct roce_gid_ndev_storage, rcu_head);
225 WARN_ON(!storage->ndev);
226 /* At this point its safe to release netdev reference,
227 * as all callers working on gid_attr->ndev are done
230 dev_put(storage->ndev);
234 static void free_gid_entry_locked(struct ib_gid_table_entry *entry)
236 struct ib_device *device = entry->attr.device;
237 u8 port_num = entry->attr.port_num;
238 struct ib_gid_table *table = rdma_gid_table(device, port_num);
240 dev_dbg(&device->dev, "%s port=%d index=%d gid %pI6\n", __func__,
241 port_num, entry->attr.index, entry->attr.gid.raw);
243 write_lock_irq(&table->rwlock);
246 * The only way to avoid overwriting NULL in table is
247 * by comparing if it is same entry in table or not!
248 * If new entry in table is added by the time we free here,
249 * don't overwrite the table entry.
251 if (entry == table->data_vec[entry->attr.index])
252 table->data_vec[entry->attr.index] = NULL;
253 /* Now this index is ready to be allocated */
254 write_unlock_irq(&table->rwlock);
256 if (entry->ndev_storage)
257 call_rcu(&entry->ndev_storage->rcu_head, put_gid_ndev);
261 static void free_gid_entry(struct kref *kref)
263 struct ib_gid_table_entry *entry =
264 container_of(kref, struct ib_gid_table_entry, kref);
266 free_gid_entry_locked(entry);
270 * free_gid_work - Release reference to the GID entry
271 * @work: Work structure to refer to GID entry which needs to be
274 * free_gid_work() frees the entry from the HCA's hardware table
275 * if provider supports it. It releases reference to netdevice.
277 static void free_gid_work(struct work_struct *work)
279 struct ib_gid_table_entry *entry =
280 container_of(work, struct ib_gid_table_entry, del_work);
281 struct ib_device *device = entry->attr.device;
282 u8 port_num = entry->attr.port_num;
283 struct ib_gid_table *table = rdma_gid_table(device, port_num);
285 mutex_lock(&table->lock);
286 free_gid_entry_locked(entry);
287 mutex_unlock(&table->lock);
290 static struct ib_gid_table_entry *
291 alloc_gid_entry(const struct ib_gid_attr *attr)
293 struct ib_gid_table_entry *entry;
294 struct net_device *ndev;
296 entry = kzalloc(sizeof(*entry), GFP_KERNEL);
300 ndev = rcu_dereference_protected(attr->ndev, 1);
302 entry->ndev_storage = kzalloc(sizeof(*entry->ndev_storage),
304 if (!entry->ndev_storage) {
309 entry->ndev_storage->ndev = ndev;
311 kref_init(&entry->kref);
312 memcpy(&entry->attr, attr, sizeof(*attr));
313 INIT_WORK(&entry->del_work, free_gid_work);
314 entry->state = GID_TABLE_ENTRY_INVALID;
318 static void store_gid_entry(struct ib_gid_table *table,
319 struct ib_gid_table_entry *entry)
321 entry->state = GID_TABLE_ENTRY_VALID;
323 dev_dbg(&entry->attr.device->dev, "%s port=%d index=%d gid %pI6\n",
324 __func__, entry->attr.port_num, entry->attr.index,
325 entry->attr.gid.raw);
327 lockdep_assert_held(&table->lock);
328 write_lock_irq(&table->rwlock);
329 table->data_vec[entry->attr.index] = entry;
330 write_unlock_irq(&table->rwlock);
333 static void get_gid_entry(struct ib_gid_table_entry *entry)
335 kref_get(&entry->kref);
338 static void put_gid_entry(struct ib_gid_table_entry *entry)
340 kref_put(&entry->kref, schedule_free_gid);
343 static void put_gid_entry_locked(struct ib_gid_table_entry *entry)
345 kref_put(&entry->kref, free_gid_entry);
348 static int add_roce_gid(struct ib_gid_table_entry *entry)
350 const struct ib_gid_attr *attr = &entry->attr;
354 dev_err(&attr->device->dev, "%s NULL netdev port=%d index=%d\n",
355 __func__, attr->port_num, attr->index);
358 if (rdma_cap_roce_gid_table(attr->device, attr->port_num)) {
359 ret = attr->device->ops.add_gid(attr, &entry->context);
361 dev_err(&attr->device->dev,
362 "%s GID add failed port=%d index=%d\n",
363 __func__, attr->port_num, attr->index);
371 * del_gid - Delete GID table entry
373 * @ib_dev: IB device whose GID entry to be deleted
374 * @port: Port number of the IB device
375 * @table: GID table of the IB device for a port
376 * @ix: GID entry index to delete
379 static void del_gid(struct ib_device *ib_dev, u8 port,
380 struct ib_gid_table *table, int ix)
382 struct roce_gid_ndev_storage *ndev_storage;
383 struct ib_gid_table_entry *entry;
385 lockdep_assert_held(&table->lock);
387 dev_dbg(&ib_dev->dev, "%s port=%d index=%d gid %pI6\n", __func__, port,
388 ix, table->data_vec[ix]->attr.gid.raw);
390 write_lock_irq(&table->rwlock);
391 entry = table->data_vec[ix];
392 entry->state = GID_TABLE_ENTRY_PENDING_DEL;
394 * For non RoCE protocol, GID entry slot is ready to use.
396 if (!rdma_protocol_roce(ib_dev, port))
397 table->data_vec[ix] = NULL;
398 write_unlock_irq(&table->rwlock);
400 ndev_storage = entry->ndev_storage;
402 entry->ndev_storage = NULL;
403 rcu_assign_pointer(entry->attr.ndev, NULL);
404 call_rcu(&ndev_storage->rcu_head, put_gid_ndev);
407 if (rdma_cap_roce_gid_table(ib_dev, port))
408 ib_dev->ops.del_gid(&entry->attr, &entry->context);
410 put_gid_entry_locked(entry);
414 * add_modify_gid - Add or modify GID table entry
416 * @table: GID table in which GID to be added or modified
417 * @attr: Attributes of the GID
419 * Returns 0 on success or appropriate error code. It accepts zero
420 * GID addition for non RoCE ports for HCA's who report them as valid
421 * GID. However such zero GIDs are not added to the cache.
423 static int add_modify_gid(struct ib_gid_table *table,
424 const struct ib_gid_attr *attr)
426 struct ib_gid_table_entry *entry;
430 * Invalidate any old entry in the table to make it safe to write to
433 if (is_gid_entry_valid(table->data_vec[attr->index]))
434 del_gid(attr->device, attr->port_num, table, attr->index);
437 * Some HCA's report multiple GID entries with only one valid GID, and
438 * leave other unused entries as the zero GID. Convert zero GIDs to
439 * empty table entries instead of storing them.
441 if (rdma_is_zero_gid(&attr->gid))
444 entry = alloc_gid_entry(attr);
448 if (rdma_protocol_roce(attr->device, attr->port_num)) {
449 ret = add_roce_gid(entry);
454 store_gid_entry(table, entry);
458 put_gid_entry(entry);
462 /* rwlock should be read locked, or lock should be held */
463 static int find_gid(struct ib_gid_table *table, const union ib_gid *gid,
464 const struct ib_gid_attr *val, bool default_gid,
465 unsigned long mask, int *pempty)
469 int empty = pempty ? -1 : 0;
471 while (i < table->sz && (found < 0 || empty < 0)) {
472 struct ib_gid_table_entry *data = table->data_vec[i];
473 struct ib_gid_attr *attr;
478 /* find_gid() is used during GID addition where it is expected
479 * to return a free entry slot which is not duplicate.
480 * Free entry slot is requested and returned if pempty is set,
481 * so lookup free slot only if requested.
483 if (pempty && empty < 0) {
484 if (is_gid_entry_free(data) &&
486 is_gid_index_default(table, curr_index)) {
488 * Found an invalid (free) entry; allocate it.
489 * If default GID is requested, then our
490 * found slot must be one of the DEFAULT
491 * reserved slots or we fail.
492 * This ensures that only DEFAULT reserved
493 * slots are used for default property GIDs.
500 * Additionally find_gid() is used to find valid entry during
501 * lookup operation; so ignore the entries which are marked as
502 * pending for removal and the entries which are marked as
505 if (!is_gid_entry_valid(data))
512 if (mask & GID_ATTR_FIND_MASK_GID_TYPE &&
513 attr->gid_type != val->gid_type)
516 if (mask & GID_ATTR_FIND_MASK_GID &&
517 memcmp(gid, &data->attr.gid, sizeof(*gid)))
520 if (mask & GID_ATTR_FIND_MASK_NETDEV &&
521 attr->ndev != val->ndev)
524 if (mask & GID_ATTR_FIND_MASK_DEFAULT &&
525 is_gid_index_default(table, curr_index) != default_gid)
537 static void make_default_gid(struct net_device *dev, union ib_gid *gid)
539 gid->global.subnet_prefix = cpu_to_be64(0xfe80000000000000LL);
540 addrconf_ifid_eui48(&gid->raw[8], dev);
543 static int __ib_cache_gid_add(struct ib_device *ib_dev, u8 port,
544 union ib_gid *gid, struct ib_gid_attr *attr,
545 unsigned long mask, bool default_gid)
547 struct ib_gid_table *table;
552 /* Do not allow adding zero GID in support of
553 * IB spec version 1.3 section 4.1.1 point (6) and
554 * section 12.7.10 and section 12.7.20
556 if (rdma_is_zero_gid(gid))
559 table = rdma_gid_table(ib_dev, port);
561 mutex_lock(&table->lock);
563 ix = find_gid(table, gid, attr, default_gid, mask, &empty);
571 attr->device = ib_dev;
573 attr->port_num = port;
575 ret = add_modify_gid(table, attr);
577 dispatch_gid_change_event(ib_dev, port);
580 mutex_unlock(&table->lock);
582 pr_warn("%s: unable to add gid %pI6 error=%d\n",
583 __func__, gid->raw, ret);
587 int ib_cache_gid_add(struct ib_device *ib_dev, u8 port,
588 union ib_gid *gid, struct ib_gid_attr *attr)
590 unsigned long mask = GID_ATTR_FIND_MASK_GID |
591 GID_ATTR_FIND_MASK_GID_TYPE |
592 GID_ATTR_FIND_MASK_NETDEV;
594 return __ib_cache_gid_add(ib_dev, port, gid, attr, mask, false);
598 _ib_cache_gid_del(struct ib_device *ib_dev, u8 port,
599 union ib_gid *gid, struct ib_gid_attr *attr,
600 unsigned long mask, bool default_gid)
602 struct ib_gid_table *table;
606 table = rdma_gid_table(ib_dev, port);
608 mutex_lock(&table->lock);
610 ix = find_gid(table, gid, attr, default_gid, mask, NULL);
616 del_gid(ib_dev, port, table, ix);
617 dispatch_gid_change_event(ib_dev, port);
620 mutex_unlock(&table->lock);
622 pr_debug("%s: can't delete gid %pI6 error=%d\n",
623 __func__, gid->raw, ret);
627 int ib_cache_gid_del(struct ib_device *ib_dev, u8 port,
628 union ib_gid *gid, struct ib_gid_attr *attr)
630 unsigned long mask = GID_ATTR_FIND_MASK_GID |
631 GID_ATTR_FIND_MASK_GID_TYPE |
632 GID_ATTR_FIND_MASK_DEFAULT |
633 GID_ATTR_FIND_MASK_NETDEV;
635 return _ib_cache_gid_del(ib_dev, port, gid, attr, mask, false);
638 int ib_cache_gid_del_all_netdev_gids(struct ib_device *ib_dev, u8 port,
639 struct net_device *ndev)
641 struct ib_gid_table *table;
643 bool deleted = false;
645 table = rdma_gid_table(ib_dev, port);
647 mutex_lock(&table->lock);
649 for (ix = 0; ix < table->sz; ix++) {
650 if (is_gid_entry_valid(table->data_vec[ix]) &&
651 table->data_vec[ix]->attr.ndev == ndev) {
652 del_gid(ib_dev, port, table, ix);
657 mutex_unlock(&table->lock);
660 dispatch_gid_change_event(ib_dev, port);
666 * rdma_find_gid_by_port - Returns the GID entry attributes when it finds
667 * a valid GID entry for given search parameters. It searches for the specified
668 * GID value in the local software cache.
669 * @device: The device to query.
670 * @gid: The GID value to search for.
671 * @gid_type: The GID type to search for.
672 * @port_num: The port number of the device where the GID value should be
674 * @ndev: In RoCE, the net device of the device. NULL means ignore.
676 * Returns sgid attributes if the GID is found with valid reference or
677 * returns ERR_PTR for the error.
678 * The caller must invoke rdma_put_gid_attr() to release the reference.
680 const struct ib_gid_attr *
681 rdma_find_gid_by_port(struct ib_device *ib_dev,
682 const union ib_gid *gid,
683 enum ib_gid_type gid_type,
684 u8 port, struct net_device *ndev)
687 struct ib_gid_table *table;
688 unsigned long mask = GID_ATTR_FIND_MASK_GID |
689 GID_ATTR_FIND_MASK_GID_TYPE;
690 struct ib_gid_attr val = {.ndev = ndev, .gid_type = gid_type};
691 const struct ib_gid_attr *attr;
694 if (!rdma_is_port_valid(ib_dev, port))
695 return ERR_PTR(-ENOENT);
697 table = rdma_gid_table(ib_dev, port);
700 mask |= GID_ATTR_FIND_MASK_NETDEV;
702 read_lock_irqsave(&table->rwlock, flags);
703 local_index = find_gid(table, gid, &val, false, mask, NULL);
704 if (local_index >= 0) {
705 get_gid_entry(table->data_vec[local_index]);
706 attr = &table->data_vec[local_index]->attr;
707 read_unlock_irqrestore(&table->rwlock, flags);
711 read_unlock_irqrestore(&table->rwlock, flags);
712 return ERR_PTR(-ENOENT);
714 EXPORT_SYMBOL(rdma_find_gid_by_port);
717 * rdma_find_gid_by_filter - Returns the GID table attribute where a
718 * specified GID value occurs
719 * @device: The device to query.
720 * @gid: The GID value to search for.
721 * @port: The port number of the device where the GID value could be
723 * @filter: The filter function is executed on any matching GID in the table.
724 * If the filter function returns true, the corresponding index is returned,
725 * otherwise, we continue searching the GID table. It's guaranteed that
726 * while filter is executed, ndev field is valid and the structure won't
727 * change. filter is executed in an atomic context. filter must not be NULL.
729 * rdma_find_gid_by_filter() searches for the specified GID value
730 * of which the filter function returns true in the port's GID table.
733 const struct ib_gid_attr *rdma_find_gid_by_filter(
734 struct ib_device *ib_dev, const union ib_gid *gid, u8 port,
735 bool (*filter)(const union ib_gid *gid, const struct ib_gid_attr *,
739 const struct ib_gid_attr *res = ERR_PTR(-ENOENT);
740 struct ib_gid_table *table;
744 if (!rdma_is_port_valid(ib_dev, port))
745 return ERR_PTR(-EINVAL);
747 table = rdma_gid_table(ib_dev, port);
749 read_lock_irqsave(&table->rwlock, flags);
750 for (i = 0; i < table->sz; i++) {
751 struct ib_gid_table_entry *entry = table->data_vec[i];
753 if (!is_gid_entry_valid(entry))
756 if (memcmp(gid, &entry->attr.gid, sizeof(*gid)))
759 if (filter(gid, &entry->attr, context)) {
760 get_gid_entry(entry);
765 read_unlock_irqrestore(&table->rwlock, flags);
769 static struct ib_gid_table *alloc_gid_table(int sz)
771 struct ib_gid_table *table = kzalloc(sizeof(*table), GFP_KERNEL);
776 table->data_vec = kcalloc(sz, sizeof(*table->data_vec), GFP_KERNEL);
777 if (!table->data_vec)
780 mutex_init(&table->lock);
783 rwlock_init(&table->rwlock);
791 static void release_gid_table(struct ib_device *device,
792 struct ib_gid_table *table)
800 for (i = 0; i < table->sz; i++) {
801 if (is_gid_entry_free(table->data_vec[i]))
803 if (kref_read(&table->data_vec[i]->kref) > 1) {
804 dev_err(&device->dev,
805 "GID entry ref leak for index %d ref=%d\n", i,
806 kref_read(&table->data_vec[i]->kref));
813 kfree(table->data_vec);
817 static void cleanup_gid_table_port(struct ib_device *ib_dev, u8 port,
818 struct ib_gid_table *table)
821 bool deleted = false;
826 mutex_lock(&table->lock);
827 for (i = 0; i < table->sz; ++i) {
828 if (is_gid_entry_valid(table->data_vec[i])) {
829 del_gid(ib_dev, port, table, i);
833 mutex_unlock(&table->lock);
836 dispatch_gid_change_event(ib_dev, port);
839 void ib_cache_gid_set_default_gid(struct ib_device *ib_dev, u8 port,
840 struct net_device *ndev,
841 unsigned long gid_type_mask,
842 enum ib_cache_gid_default_mode mode)
844 union ib_gid gid = { };
845 struct ib_gid_attr gid_attr;
846 unsigned int gid_type;
849 mask = GID_ATTR_FIND_MASK_GID_TYPE |
850 GID_ATTR_FIND_MASK_DEFAULT |
851 GID_ATTR_FIND_MASK_NETDEV;
852 memset(&gid_attr, 0, sizeof(gid_attr));
853 gid_attr.ndev = ndev;
855 for (gid_type = 0; gid_type < IB_GID_TYPE_SIZE; ++gid_type) {
856 if (1UL << gid_type & ~gid_type_mask)
859 gid_attr.gid_type = gid_type;
861 if (mode == IB_CACHE_GID_DEFAULT_MODE_SET) {
862 make_default_gid(ndev, &gid);
863 __ib_cache_gid_add(ib_dev, port, &gid,
864 &gid_attr, mask, true);
865 } else if (mode == IB_CACHE_GID_DEFAULT_MODE_DELETE) {
866 _ib_cache_gid_del(ib_dev, port, &gid,
867 &gid_attr, mask, true);
872 static void gid_table_reserve_default(struct ib_device *ib_dev, u8 port,
873 struct ib_gid_table *table)
876 unsigned long roce_gid_type_mask;
877 unsigned int num_default_gids;
879 roce_gid_type_mask = roce_gid_type_mask_support(ib_dev, port);
880 num_default_gids = hweight_long(roce_gid_type_mask);
881 /* Reserve starting indices for default GIDs */
882 for (i = 0; i < num_default_gids && i < table->sz; i++)
883 table->default_gid_indices |= BIT(i);
887 static void gid_table_release_one(struct ib_device *ib_dev)
891 rdma_for_each_port (ib_dev, p) {
892 release_gid_table(ib_dev, ib_dev->port_data[p].cache.gid);
893 ib_dev->port_data[p].cache.gid = NULL;
897 static int _gid_table_setup_one(struct ib_device *ib_dev)
899 struct ib_gid_table *table;
900 unsigned int rdma_port;
902 rdma_for_each_port (ib_dev, rdma_port) {
903 table = alloc_gid_table(
904 ib_dev->port_data[rdma_port].immutable.gid_tbl_len);
906 goto rollback_table_setup;
908 gid_table_reserve_default(ib_dev, rdma_port, table);
909 ib_dev->port_data[rdma_port].cache.gid = table;
913 rollback_table_setup:
914 gid_table_release_one(ib_dev);
918 static void gid_table_cleanup_one(struct ib_device *ib_dev)
922 rdma_for_each_port (ib_dev, p)
923 cleanup_gid_table_port(ib_dev, p,
924 ib_dev->port_data[p].cache.gid);
927 static int gid_table_setup_one(struct ib_device *ib_dev)
931 err = _gid_table_setup_one(ib_dev);
936 rdma_roce_rescan_device(ib_dev);
942 * rdma_query_gid - Read the GID content from the GID software cache
943 * @device: Device to query the GID
944 * @port_num: Port number of the device
945 * @index: Index of the GID table entry to read
946 * @gid: Pointer to GID where to store the entry's GID
948 * rdma_query_gid() only reads the GID entry content for requested device,
949 * port and index. It reads for IB, RoCE and iWarp link layers. It doesn't
950 * hold any reference to the GID table entry in the HCA or software cache.
952 * Returns 0 on success or appropriate error code.
955 int rdma_query_gid(struct ib_device *device, u8 port_num,
956 int index, union ib_gid *gid)
958 struct ib_gid_table *table;
962 if (!rdma_is_port_valid(device, port_num))
965 table = rdma_gid_table(device, port_num);
966 read_lock_irqsave(&table->rwlock, flags);
968 if (index < 0 || index >= table->sz ||
969 !is_gid_entry_valid(table->data_vec[index]))
972 memcpy(gid, &table->data_vec[index]->attr.gid, sizeof(*gid));
976 read_unlock_irqrestore(&table->rwlock, flags);
979 EXPORT_SYMBOL(rdma_query_gid);
982 * rdma_find_gid - Returns SGID attributes if the matching GID is found.
983 * @device: The device to query.
984 * @gid: The GID value to search for.
985 * @gid_type: The GID type to search for.
986 * @ndev: In RoCE, the net device of the device. NULL means ignore.
988 * rdma_find_gid() searches for the specified GID value in the software cache.
990 * Returns GID attributes if a valid GID is found or returns ERR_PTR for the
991 * error. The caller must invoke rdma_put_gid_attr() to release the reference.
994 const struct ib_gid_attr *rdma_find_gid(struct ib_device *device,
995 const union ib_gid *gid,
996 enum ib_gid_type gid_type,
997 struct net_device *ndev)
999 unsigned long mask = GID_ATTR_FIND_MASK_GID |
1000 GID_ATTR_FIND_MASK_GID_TYPE;
1001 struct ib_gid_attr gid_attr_val = {.ndev = ndev, .gid_type = gid_type};
1005 mask |= GID_ATTR_FIND_MASK_NETDEV;
1007 rdma_for_each_port(device, p) {
1008 struct ib_gid_table *table;
1009 unsigned long flags;
1012 table = device->port_data[p].cache.gid;
1013 read_lock_irqsave(&table->rwlock, flags);
1014 index = find_gid(table, gid, &gid_attr_val, false, mask, NULL);
1016 const struct ib_gid_attr *attr;
1018 get_gid_entry(table->data_vec[index]);
1019 attr = &table->data_vec[index]->attr;
1020 read_unlock_irqrestore(&table->rwlock, flags);
1023 read_unlock_irqrestore(&table->rwlock, flags);
1026 return ERR_PTR(-ENOENT);
1028 EXPORT_SYMBOL(rdma_find_gid);
1030 int ib_get_cached_pkey(struct ib_device *device,
1035 struct ib_pkey_cache *cache;
1036 unsigned long flags;
1039 if (!rdma_is_port_valid(device, port_num))
1042 read_lock_irqsave(&device->cache.lock, flags);
1044 cache = device->port_data[port_num].cache.pkey;
1046 if (index < 0 || index >= cache->table_len)
1049 *pkey = cache->table[index];
1051 read_unlock_irqrestore(&device->cache.lock, flags);
1055 EXPORT_SYMBOL(ib_get_cached_pkey);
1057 int ib_get_cached_subnet_prefix(struct ib_device *device,
1061 unsigned long flags;
1063 if (!rdma_is_port_valid(device, port_num))
1066 read_lock_irqsave(&device->cache.lock, flags);
1067 *sn_pfx = device->port_data[port_num].cache.subnet_prefix;
1068 read_unlock_irqrestore(&device->cache.lock, flags);
1072 EXPORT_SYMBOL(ib_get_cached_subnet_prefix);
1074 int ib_find_cached_pkey(struct ib_device *device,
1079 struct ib_pkey_cache *cache;
1080 unsigned long flags;
1083 int partial_ix = -1;
1085 if (!rdma_is_port_valid(device, port_num))
1088 read_lock_irqsave(&device->cache.lock, flags);
1090 cache = device->port_data[port_num].cache.pkey;
1094 for (i = 0; i < cache->table_len; ++i)
1095 if ((cache->table[i] & 0x7fff) == (pkey & 0x7fff)) {
1096 if (cache->table[i] & 0x8000) {
1104 if (ret && partial_ix >= 0) {
1105 *index = partial_ix;
1109 read_unlock_irqrestore(&device->cache.lock, flags);
1113 EXPORT_SYMBOL(ib_find_cached_pkey);
1115 int ib_find_exact_cached_pkey(struct ib_device *device,
1120 struct ib_pkey_cache *cache;
1121 unsigned long flags;
1125 if (!rdma_is_port_valid(device, port_num))
1128 read_lock_irqsave(&device->cache.lock, flags);
1130 cache = device->port_data[port_num].cache.pkey;
1134 for (i = 0; i < cache->table_len; ++i)
1135 if (cache->table[i] == pkey) {
1141 read_unlock_irqrestore(&device->cache.lock, flags);
1145 EXPORT_SYMBOL(ib_find_exact_cached_pkey);
1147 int ib_get_cached_lmc(struct ib_device *device,
1151 unsigned long flags;
1154 if (!rdma_is_port_valid(device, port_num))
1157 read_lock_irqsave(&device->cache.lock, flags);
1158 *lmc = device->port_data[port_num].cache.lmc;
1159 read_unlock_irqrestore(&device->cache.lock, flags);
1163 EXPORT_SYMBOL(ib_get_cached_lmc);
1165 int ib_get_cached_port_state(struct ib_device *device,
1167 enum ib_port_state *port_state)
1169 unsigned long flags;
1172 if (!rdma_is_port_valid(device, port_num))
1175 read_lock_irqsave(&device->cache.lock, flags);
1176 *port_state = device->port_data[port_num].cache.port_state;
1177 read_unlock_irqrestore(&device->cache.lock, flags);
1181 EXPORT_SYMBOL(ib_get_cached_port_state);
1184 * rdma_get_gid_attr - Returns GID attributes for a port of a device
1185 * at a requested gid_index, if a valid GID entry exists.
1186 * @device: The device to query.
1187 * @port_num: The port number on the device where the GID value
1189 * @index: Index of the GID table entry whose attributes are to
1192 * rdma_get_gid_attr() acquires reference count of gid attributes from the
1193 * cached GID table. Caller must invoke rdma_put_gid_attr() to release
1194 * reference to gid attribute regardless of link layer.
1196 * Returns pointer to valid gid attribute or ERR_PTR for the appropriate error
1199 const struct ib_gid_attr *
1200 rdma_get_gid_attr(struct ib_device *device, u8 port_num, int index)
1202 const struct ib_gid_attr *attr = ERR_PTR(-EINVAL);
1203 struct ib_gid_table *table;
1204 unsigned long flags;
1206 if (!rdma_is_port_valid(device, port_num))
1207 return ERR_PTR(-EINVAL);
1209 table = rdma_gid_table(device, port_num);
1210 if (index < 0 || index >= table->sz)
1211 return ERR_PTR(-EINVAL);
1213 read_lock_irqsave(&table->rwlock, flags);
1214 if (!is_gid_entry_valid(table->data_vec[index]))
1217 get_gid_entry(table->data_vec[index]);
1218 attr = &table->data_vec[index]->attr;
1220 read_unlock_irqrestore(&table->rwlock, flags);
1223 EXPORT_SYMBOL(rdma_get_gid_attr);
1226 * rdma_put_gid_attr - Release reference to the GID attribute
1227 * @attr: Pointer to the GID attribute whose reference
1228 * needs to be released.
1230 * rdma_put_gid_attr() must be used to release reference whose
1231 * reference is acquired using rdma_get_gid_attr() or any APIs
1232 * which returns a pointer to the ib_gid_attr regardless of link layer
1236 void rdma_put_gid_attr(const struct ib_gid_attr *attr)
1238 struct ib_gid_table_entry *entry =
1239 container_of(attr, struct ib_gid_table_entry, attr);
1241 put_gid_entry(entry);
1243 EXPORT_SYMBOL(rdma_put_gid_attr);
1246 * rdma_hold_gid_attr - Get reference to existing GID attribute
1248 * @attr: Pointer to the GID attribute whose reference
1249 * needs to be taken.
1251 * Increase the reference count to a GID attribute to keep it from being
1252 * freed. Callers are required to already be holding a reference to attribute.
1255 void rdma_hold_gid_attr(const struct ib_gid_attr *attr)
1257 struct ib_gid_table_entry *entry =
1258 container_of(attr, struct ib_gid_table_entry, attr);
1260 get_gid_entry(entry);
1262 EXPORT_SYMBOL(rdma_hold_gid_attr);
1265 * rdma_read_gid_attr_ndev_rcu - Read GID attribute netdevice
1266 * which must be in UP state.
1268 * @attr:Pointer to the GID attribute
1270 * Returns pointer to netdevice if the netdevice was attached to GID and
1271 * netdevice is in UP state. Caller must hold RCU lock as this API
1272 * reads the netdev flags which can change while netdevice migrates to
1273 * different net namespace. Returns ERR_PTR with error code otherwise.
1276 struct net_device *rdma_read_gid_attr_ndev_rcu(const struct ib_gid_attr *attr)
1278 struct ib_gid_table_entry *entry =
1279 container_of(attr, struct ib_gid_table_entry, attr);
1280 struct ib_device *device = entry->attr.device;
1281 struct net_device *ndev = ERR_PTR(-ENODEV);
1282 u8 port_num = entry->attr.port_num;
1283 struct ib_gid_table *table;
1284 unsigned long flags;
1287 table = rdma_gid_table(device, port_num);
1289 read_lock_irqsave(&table->rwlock, flags);
1290 valid = is_gid_entry_valid(table->data_vec[attr->index]);
1292 ndev = rcu_dereference(attr->ndev);
1294 (ndev && ((READ_ONCE(ndev->flags) & IFF_UP) == 0)))
1295 ndev = ERR_PTR(-ENODEV);
1297 read_unlock_irqrestore(&table->rwlock, flags);
1300 EXPORT_SYMBOL(rdma_read_gid_attr_ndev_rcu);
1302 static int get_lower_dev_vlan(struct net_device *lower_dev, void *data)
1304 u16 *vlan_id = data;
1306 if (is_vlan_dev(lower_dev))
1307 *vlan_id = vlan_dev_vlan_id(lower_dev);
1309 /* We are interested only in first level vlan device, so
1310 * always return 1 to stop iterating over next level devices.
1316 * rdma_read_gid_l2_fields - Read the vlan ID and source MAC address
1319 * @attr: GID attribute pointer whose L2 fields to be read
1320 * @vlan_id: Pointer to vlan id to fill up if the GID entry has
1321 * vlan id. It is optional.
1322 * @smac: Pointer to smac to fill up for a GID entry. It is optional.
1324 * rdma_read_gid_l2_fields() returns 0 on success and returns vlan id
1325 * (if gid entry has vlan) and source MAC, or returns error.
1327 int rdma_read_gid_l2_fields(const struct ib_gid_attr *attr,
1328 u16 *vlan_id, u8 *smac)
1330 struct net_device *ndev;
1333 ndev = rcu_dereference(attr->ndev);
1339 ether_addr_copy(smac, ndev->dev_addr);
1342 if (is_vlan_dev(ndev)) {
1343 *vlan_id = vlan_dev_vlan_id(ndev);
1345 /* If the netdev is upper device and if it's lower
1346 * device is vlan device, consider vlan id of the
1347 * the lower vlan device for this gid entry.
1349 netdev_walk_all_lower_dev_rcu(attr->ndev,
1350 get_lower_dev_vlan, vlan_id);
1356 EXPORT_SYMBOL(rdma_read_gid_l2_fields);
1358 static int config_non_roce_gid_cache(struct ib_device *device,
1359 u8 port, int gid_tbl_len)
1361 struct ib_gid_attr gid_attr = {};
1362 struct ib_gid_table *table;
1366 gid_attr.device = device;
1367 gid_attr.port_num = port;
1368 table = rdma_gid_table(device, port);
1370 mutex_lock(&table->lock);
1371 for (i = 0; i < gid_tbl_len; ++i) {
1372 if (!device->ops.query_gid)
1374 ret = device->ops.query_gid(device, port, i, &gid_attr.gid);
1376 dev_warn(&device->dev,
1377 "query_gid failed (%d) for index %d\n", ret,
1382 add_modify_gid(table, &gid_attr);
1385 mutex_unlock(&table->lock);
1389 static void ib_cache_update(struct ib_device *device,
1391 bool enforce_security)
1393 struct ib_port_attr *tprops = NULL;
1394 struct ib_pkey_cache *pkey_cache = NULL, *old_pkey_cache;
1398 if (!rdma_is_port_valid(device, port))
1401 tprops = kmalloc(sizeof *tprops, GFP_KERNEL);
1405 ret = ib_query_port(device, port, tprops);
1407 dev_warn(&device->dev, "ib_query_port failed (%d)\n", ret);
1411 if (!rdma_protocol_roce(device, port)) {
1412 ret = config_non_roce_gid_cache(device, port,
1413 tprops->gid_tbl_len);
1418 pkey_cache = kmalloc(struct_size(pkey_cache, table,
1419 tprops->pkey_tbl_len),
1424 pkey_cache->table_len = tprops->pkey_tbl_len;
1426 for (i = 0; i < pkey_cache->table_len; ++i) {
1427 ret = ib_query_pkey(device, port, i, pkey_cache->table + i);
1429 dev_warn(&device->dev,
1430 "ib_query_pkey failed (%d) for index %d\n",
1436 write_lock_irq(&device->cache.lock);
1438 old_pkey_cache = device->port_data[port].cache.pkey;
1440 device->port_data[port].cache.pkey = pkey_cache;
1441 device->port_data[port].cache.lmc = tprops->lmc;
1442 device->port_data[port].cache.port_state = tprops->state;
1444 device->port_data[port].cache.subnet_prefix = tprops->subnet_prefix;
1445 write_unlock_irq(&device->cache.lock);
1447 if (enforce_security)
1448 ib_security_cache_change(device,
1450 tprops->subnet_prefix);
1452 kfree(old_pkey_cache);
1461 static void ib_cache_task(struct work_struct *_work)
1463 struct ib_update_work *work =
1464 container_of(_work, struct ib_update_work, work);
1466 ib_cache_update(work->device,
1468 work->enforce_security);
1472 static void ib_cache_event(struct ib_event_handler *handler,
1473 struct ib_event *event)
1475 struct ib_update_work *work;
1477 if (event->event == IB_EVENT_PORT_ERR ||
1478 event->event == IB_EVENT_PORT_ACTIVE ||
1479 event->event == IB_EVENT_LID_CHANGE ||
1480 event->event == IB_EVENT_PKEY_CHANGE ||
1481 event->event == IB_EVENT_CLIENT_REREGISTER ||
1482 event->event == IB_EVENT_GID_CHANGE) {
1483 work = kmalloc(sizeof *work, GFP_ATOMIC);
1485 INIT_WORK(&work->work, ib_cache_task);
1486 work->device = event->device;
1487 work->port_num = event->element.port_num;
1488 if (event->event == IB_EVENT_PKEY_CHANGE ||
1489 event->event == IB_EVENT_GID_CHANGE)
1490 work->enforce_security = true;
1492 work->enforce_security = false;
1494 queue_work(ib_wq, &work->work);
1499 int ib_cache_setup_one(struct ib_device *device)
1504 rwlock_init(&device->cache.lock);
1506 err = gid_table_setup_one(device);
1510 rdma_for_each_port (device, p)
1511 ib_cache_update(device, p, true);
1513 INIT_IB_EVENT_HANDLER(&device->cache.event_handler,
1514 device, ib_cache_event);
1515 ib_register_event_handler(&device->cache.event_handler);
1519 void ib_cache_release_one(struct ib_device *device)
1524 * The release function frees all the cache elements.
1525 * This function should be called as part of freeing
1526 * all the device's resources when the cache could no
1527 * longer be accessed.
1529 rdma_for_each_port (device, p)
1530 kfree(device->port_data[p].cache.pkey);
1532 gid_table_release_one(device);
1535 void ib_cache_cleanup_one(struct ib_device *device)
1537 /* The cleanup function unregisters the event handler,
1538 * waits for all in-progress workqueue elements and cleans
1539 * up the GID cache. This function should be called after
1540 * the device was removed from the devices list and all
1541 * clients were removed, so the cache exists but is
1542 * non-functional and shouldn't be updated anymore.
1544 ib_unregister_event_handler(&device->cache.event_handler);
1545 flush_workqueue(ib_wq);
1546 gid_table_cleanup_one(device);
1549 * Flush the wq second time for any pending GID delete work.
1551 flush_workqueue(ib_wq);