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/if_vlan.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_event event;
55 bool enforce_security;
61 enum gid_attr_find_mask {
62 GID_ATTR_FIND_MASK_GID = 1UL << 0,
63 GID_ATTR_FIND_MASK_NETDEV = 1UL << 1,
64 GID_ATTR_FIND_MASK_DEFAULT = 1UL << 2,
65 GID_ATTR_FIND_MASK_GID_TYPE = 1UL << 3,
68 enum gid_table_entry_state {
69 GID_TABLE_ENTRY_INVALID = 1,
70 GID_TABLE_ENTRY_VALID = 2,
72 * Indicates that entry is pending to be removed, there may
73 * be active users of this GID entry.
74 * When last user of the GID entry releases reference to it,
75 * GID entry is detached from the table.
77 GID_TABLE_ENTRY_PENDING_DEL = 3,
80 struct roce_gid_ndev_storage {
81 struct rcu_head rcu_head;
82 struct net_device *ndev;
85 struct ib_gid_table_entry {
87 struct work_struct del_work;
88 struct ib_gid_attr attr;
90 /* Store the ndev pointer to release reference later on in
91 * call_rcu context because by that time gid_table_entry
92 * and attr might be already freed. So keep a copy of it.
93 * ndev_storage is freed by rcu callback.
95 struct roce_gid_ndev_storage *ndev_storage;
96 enum gid_table_entry_state state;
101 /* In RoCE, adding a GID to the table requires:
102 * (a) Find if this GID is already exists.
103 * (b) Find a free space.
104 * (c) Write the new GID
106 * Delete requires different set of operations:
111 /* Any writer to data_vec must hold this lock and the write side of
112 * rwlock. Readers must hold only rwlock. All writers must be in a
116 /* rwlock protects data_vec[ix]->state and entry pointer.
119 struct ib_gid_table_entry **data_vec;
120 /* bit field, each bit indicates the index of default GID */
121 u32 default_gid_indices;
124 static void dispatch_gid_change_event(struct ib_device *ib_dev, u32 port)
126 struct ib_event event;
128 event.device = ib_dev;
129 event.element.port_num = port;
130 event.event = IB_EVENT_GID_CHANGE;
132 ib_dispatch_event_clients(&event);
135 static const char * const gid_type_str[] = {
136 /* IB/RoCE v1 value is set for IB_GID_TYPE_IB and IB_GID_TYPE_ROCE for
137 * user space compatibility reasons.
139 [IB_GID_TYPE_IB] = "IB/RoCE v1",
140 [IB_GID_TYPE_ROCE] = "IB/RoCE v1",
141 [IB_GID_TYPE_ROCE_UDP_ENCAP] = "RoCE v2",
144 const char *ib_cache_gid_type_str(enum ib_gid_type gid_type)
146 if (gid_type < ARRAY_SIZE(gid_type_str) && gid_type_str[gid_type])
147 return gid_type_str[gid_type];
149 return "Invalid GID type";
151 EXPORT_SYMBOL(ib_cache_gid_type_str);
153 /** rdma_is_zero_gid - Check if given GID is zero or not.
155 * Returns true if given GID is zero, returns false otherwise.
157 bool rdma_is_zero_gid(const union ib_gid *gid)
159 return !memcmp(gid, &zgid, sizeof(*gid));
161 EXPORT_SYMBOL(rdma_is_zero_gid);
163 /** is_gid_index_default - Check if a given index belongs to
164 * reserved default GIDs or not.
165 * @table: GID table pointer
166 * @index: Index to check in GID table
167 * Returns true if index is one of the reserved default GID index otherwise
170 static bool is_gid_index_default(const struct ib_gid_table *table,
173 return index < 32 && (BIT(index) & table->default_gid_indices);
176 int ib_cache_gid_parse_type_str(const char *buf)
186 if (buf[len - 1] == '\n')
189 for (i = 0; i < ARRAY_SIZE(gid_type_str); ++i)
190 if (gid_type_str[i] && !strncmp(buf, gid_type_str[i], len) &&
191 len == strlen(gid_type_str[i])) {
198 EXPORT_SYMBOL(ib_cache_gid_parse_type_str);
200 static struct ib_gid_table *rdma_gid_table(struct ib_device *device, u32 port)
202 return device->port_data[port].cache.gid;
205 static bool is_gid_entry_free(const struct ib_gid_table_entry *entry)
210 static bool is_gid_entry_valid(const struct ib_gid_table_entry *entry)
212 return entry && entry->state == GID_TABLE_ENTRY_VALID;
215 static void schedule_free_gid(struct kref *kref)
217 struct ib_gid_table_entry *entry =
218 container_of(kref, struct ib_gid_table_entry, kref);
220 queue_work(ib_wq, &entry->del_work);
223 static void put_gid_ndev(struct rcu_head *head)
225 struct roce_gid_ndev_storage *storage =
226 container_of(head, struct roce_gid_ndev_storage, rcu_head);
228 WARN_ON(!storage->ndev);
229 /* At this point its safe to release netdev reference,
230 * as all callers working on gid_attr->ndev are done
233 dev_put(storage->ndev);
237 static void free_gid_entry_locked(struct ib_gid_table_entry *entry)
239 struct ib_device *device = entry->attr.device;
240 u32 port_num = entry->attr.port_num;
241 struct ib_gid_table *table = rdma_gid_table(device, port_num);
243 dev_dbg(&device->dev, "%s port=%u index=%u gid %pI6\n", __func__,
244 port_num, entry->attr.index, entry->attr.gid.raw);
246 write_lock_irq(&table->rwlock);
249 * The only way to avoid overwriting NULL in table is
250 * by comparing if it is same entry in table or not!
251 * If new entry in table is added by the time we free here,
252 * don't overwrite the table entry.
254 if (entry == table->data_vec[entry->attr.index])
255 table->data_vec[entry->attr.index] = NULL;
256 /* Now this index is ready to be allocated */
257 write_unlock_irq(&table->rwlock);
259 if (entry->ndev_storage)
260 call_rcu(&entry->ndev_storage->rcu_head, put_gid_ndev);
264 static void free_gid_entry(struct kref *kref)
266 struct ib_gid_table_entry *entry =
267 container_of(kref, struct ib_gid_table_entry, kref);
269 free_gid_entry_locked(entry);
273 * free_gid_work - Release reference to the GID entry
274 * @work: Work structure to refer to GID entry which needs to be
277 * free_gid_work() frees the entry from the HCA's hardware table
278 * if provider supports it. It releases reference to netdevice.
280 static void free_gid_work(struct work_struct *work)
282 struct ib_gid_table_entry *entry =
283 container_of(work, struct ib_gid_table_entry, del_work);
284 struct ib_device *device = entry->attr.device;
285 u32 port_num = entry->attr.port_num;
286 struct ib_gid_table *table = rdma_gid_table(device, port_num);
288 mutex_lock(&table->lock);
289 free_gid_entry_locked(entry);
290 mutex_unlock(&table->lock);
293 static struct ib_gid_table_entry *
294 alloc_gid_entry(const struct ib_gid_attr *attr)
296 struct ib_gid_table_entry *entry;
297 struct net_device *ndev;
299 entry = kzalloc(sizeof(*entry), GFP_KERNEL);
303 ndev = rcu_dereference_protected(attr->ndev, 1);
305 entry->ndev_storage = kzalloc(sizeof(*entry->ndev_storage),
307 if (!entry->ndev_storage) {
312 entry->ndev_storage->ndev = ndev;
314 kref_init(&entry->kref);
315 memcpy(&entry->attr, attr, sizeof(*attr));
316 INIT_WORK(&entry->del_work, free_gid_work);
317 entry->state = GID_TABLE_ENTRY_INVALID;
321 static void store_gid_entry(struct ib_gid_table *table,
322 struct ib_gid_table_entry *entry)
324 entry->state = GID_TABLE_ENTRY_VALID;
326 dev_dbg(&entry->attr.device->dev, "%s port=%u index=%u gid %pI6\n",
327 __func__, entry->attr.port_num, entry->attr.index,
328 entry->attr.gid.raw);
330 lockdep_assert_held(&table->lock);
331 write_lock_irq(&table->rwlock);
332 table->data_vec[entry->attr.index] = entry;
333 write_unlock_irq(&table->rwlock);
336 static void get_gid_entry(struct ib_gid_table_entry *entry)
338 kref_get(&entry->kref);
341 static void put_gid_entry(struct ib_gid_table_entry *entry)
343 kref_put(&entry->kref, schedule_free_gid);
346 static void put_gid_entry_locked(struct ib_gid_table_entry *entry)
348 kref_put(&entry->kref, free_gid_entry);
351 static int add_roce_gid(struct ib_gid_table_entry *entry)
353 const struct ib_gid_attr *attr = &entry->attr;
357 dev_err(&attr->device->dev, "%s NULL netdev port=%u index=%u\n",
358 __func__, attr->port_num, attr->index);
361 if (rdma_cap_roce_gid_table(attr->device, attr->port_num)) {
362 ret = attr->device->ops.add_gid(attr, &entry->context);
364 dev_err(&attr->device->dev,
365 "%s GID add failed port=%u index=%u\n",
366 __func__, attr->port_num, attr->index);
374 * del_gid - Delete GID table entry
376 * @ib_dev: IB device whose GID entry to be deleted
377 * @port: Port number of the IB device
378 * @table: GID table of the IB device for a port
379 * @ix: GID entry index to delete
382 static void del_gid(struct ib_device *ib_dev, u32 port,
383 struct ib_gid_table *table, int ix)
385 struct roce_gid_ndev_storage *ndev_storage;
386 struct ib_gid_table_entry *entry;
388 lockdep_assert_held(&table->lock);
390 dev_dbg(&ib_dev->dev, "%s port=%u index=%d gid %pI6\n", __func__, port,
391 ix, table->data_vec[ix]->attr.gid.raw);
393 write_lock_irq(&table->rwlock);
394 entry = table->data_vec[ix];
395 entry->state = GID_TABLE_ENTRY_PENDING_DEL;
397 * For non RoCE protocol, GID entry slot is ready to use.
399 if (!rdma_protocol_roce(ib_dev, port))
400 table->data_vec[ix] = NULL;
401 write_unlock_irq(&table->rwlock);
403 ndev_storage = entry->ndev_storage;
405 entry->ndev_storage = NULL;
406 rcu_assign_pointer(entry->attr.ndev, NULL);
407 call_rcu(&ndev_storage->rcu_head, put_gid_ndev);
410 if (rdma_cap_roce_gid_table(ib_dev, port))
411 ib_dev->ops.del_gid(&entry->attr, &entry->context);
413 put_gid_entry_locked(entry);
417 * add_modify_gid - Add or modify GID table entry
419 * @table: GID table in which GID to be added or modified
420 * @attr: Attributes of the GID
422 * Returns 0 on success or appropriate error code. It accepts zero
423 * GID addition for non RoCE ports for HCA's who report them as valid
424 * GID. However such zero GIDs are not added to the cache.
426 static int add_modify_gid(struct ib_gid_table *table,
427 const struct ib_gid_attr *attr)
429 struct ib_gid_table_entry *entry;
433 * Invalidate any old entry in the table to make it safe to write to
436 if (is_gid_entry_valid(table->data_vec[attr->index]))
437 del_gid(attr->device, attr->port_num, table, attr->index);
440 * Some HCA's report multiple GID entries with only one valid GID, and
441 * leave other unused entries as the zero GID. Convert zero GIDs to
442 * empty table entries instead of storing them.
444 if (rdma_is_zero_gid(&attr->gid))
447 entry = alloc_gid_entry(attr);
451 if (rdma_protocol_roce(attr->device, attr->port_num)) {
452 ret = add_roce_gid(entry);
457 store_gid_entry(table, entry);
461 put_gid_entry(entry);
465 /* rwlock should be read locked, or lock should be held */
466 static int find_gid(struct ib_gid_table *table, const union ib_gid *gid,
467 const struct ib_gid_attr *val, bool default_gid,
468 unsigned long mask, int *pempty)
472 int empty = pempty ? -1 : 0;
474 while (i < table->sz && (found < 0 || empty < 0)) {
475 struct ib_gid_table_entry *data = table->data_vec[i];
476 struct ib_gid_attr *attr;
481 /* find_gid() is used during GID addition where it is expected
482 * to return a free entry slot which is not duplicate.
483 * Free entry slot is requested and returned if pempty is set,
484 * so lookup free slot only if requested.
486 if (pempty && empty < 0) {
487 if (is_gid_entry_free(data) &&
489 is_gid_index_default(table, curr_index)) {
491 * Found an invalid (free) entry; allocate it.
492 * If default GID is requested, then our
493 * found slot must be one of the DEFAULT
494 * reserved slots or we fail.
495 * This ensures that only DEFAULT reserved
496 * slots are used for default property GIDs.
503 * Additionally find_gid() is used to find valid entry during
504 * lookup operation; so ignore the entries which are marked as
505 * pending for removal and the entries which are marked as
508 if (!is_gid_entry_valid(data))
515 if (mask & GID_ATTR_FIND_MASK_GID_TYPE &&
516 attr->gid_type != val->gid_type)
519 if (mask & GID_ATTR_FIND_MASK_GID &&
520 memcmp(gid, &data->attr.gid, sizeof(*gid)))
523 if (mask & GID_ATTR_FIND_MASK_NETDEV &&
524 attr->ndev != val->ndev)
527 if (mask & GID_ATTR_FIND_MASK_DEFAULT &&
528 is_gid_index_default(table, curr_index) != default_gid)
540 static void make_default_gid(struct net_device *dev, union ib_gid *gid)
542 gid->global.subnet_prefix = cpu_to_be64(0xfe80000000000000LL);
543 addrconf_ifid_eui48(&gid->raw[8], dev);
546 static int __ib_cache_gid_add(struct ib_device *ib_dev, u32 port,
547 union ib_gid *gid, struct ib_gid_attr *attr,
548 unsigned long mask, bool default_gid)
550 struct ib_gid_table *table;
555 /* Do not allow adding zero GID in support of
556 * IB spec version 1.3 section 4.1.1 point (6) and
557 * section 12.7.10 and section 12.7.20
559 if (rdma_is_zero_gid(gid))
562 table = rdma_gid_table(ib_dev, port);
564 mutex_lock(&table->lock);
566 ix = find_gid(table, gid, attr, default_gid, mask, &empty);
574 attr->device = ib_dev;
576 attr->port_num = port;
578 ret = add_modify_gid(table, attr);
580 dispatch_gid_change_event(ib_dev, port);
583 mutex_unlock(&table->lock);
585 pr_warn("%s: unable to add gid %pI6 error=%d\n",
586 __func__, gid->raw, ret);
590 int ib_cache_gid_add(struct ib_device *ib_dev, u32 port,
591 union ib_gid *gid, struct ib_gid_attr *attr)
593 unsigned long mask = GID_ATTR_FIND_MASK_GID |
594 GID_ATTR_FIND_MASK_GID_TYPE |
595 GID_ATTR_FIND_MASK_NETDEV;
597 return __ib_cache_gid_add(ib_dev, port, gid, attr, mask, false);
601 _ib_cache_gid_del(struct ib_device *ib_dev, u32 port,
602 union ib_gid *gid, struct ib_gid_attr *attr,
603 unsigned long mask, bool default_gid)
605 struct ib_gid_table *table;
609 table = rdma_gid_table(ib_dev, port);
611 mutex_lock(&table->lock);
613 ix = find_gid(table, gid, attr, default_gid, mask, NULL);
619 del_gid(ib_dev, port, table, ix);
620 dispatch_gid_change_event(ib_dev, port);
623 mutex_unlock(&table->lock);
625 pr_debug("%s: can't delete gid %pI6 error=%d\n",
626 __func__, gid->raw, ret);
630 int ib_cache_gid_del(struct ib_device *ib_dev, u32 port,
631 union ib_gid *gid, struct ib_gid_attr *attr)
633 unsigned long mask = GID_ATTR_FIND_MASK_GID |
634 GID_ATTR_FIND_MASK_GID_TYPE |
635 GID_ATTR_FIND_MASK_DEFAULT |
636 GID_ATTR_FIND_MASK_NETDEV;
638 return _ib_cache_gid_del(ib_dev, port, gid, attr, mask, false);
641 int ib_cache_gid_del_all_netdev_gids(struct ib_device *ib_dev, u32 port,
642 struct net_device *ndev)
644 struct ib_gid_table *table;
646 bool deleted = false;
648 table = rdma_gid_table(ib_dev, port);
650 mutex_lock(&table->lock);
652 for (ix = 0; ix < table->sz; ix++) {
653 if (is_gid_entry_valid(table->data_vec[ix]) &&
654 table->data_vec[ix]->attr.ndev == ndev) {
655 del_gid(ib_dev, port, table, ix);
660 mutex_unlock(&table->lock);
663 dispatch_gid_change_event(ib_dev, port);
669 * rdma_find_gid_by_port - Returns the GID entry attributes when it finds
670 * a valid GID entry for given search parameters. It searches for the specified
671 * GID value in the local software cache.
672 * @ib_dev: The device to query.
673 * @gid: The GID value to search for.
674 * @gid_type: The GID type to search for.
675 * @port: The port number of the device where the GID value should be searched.
676 * @ndev: In RoCE, the net device of the device. NULL means ignore.
678 * Returns sgid attributes if the GID is found with valid reference or
679 * returns ERR_PTR for the error.
680 * The caller must invoke rdma_put_gid_attr() to release the reference.
682 const struct ib_gid_attr *
683 rdma_find_gid_by_port(struct ib_device *ib_dev,
684 const union ib_gid *gid,
685 enum ib_gid_type gid_type,
686 u32 port, struct net_device *ndev)
689 struct ib_gid_table *table;
690 unsigned long mask = GID_ATTR_FIND_MASK_GID |
691 GID_ATTR_FIND_MASK_GID_TYPE;
692 struct ib_gid_attr val = {.ndev = ndev, .gid_type = gid_type};
693 const struct ib_gid_attr *attr;
696 if (!rdma_is_port_valid(ib_dev, port))
697 return ERR_PTR(-ENOENT);
699 table = rdma_gid_table(ib_dev, port);
702 mask |= GID_ATTR_FIND_MASK_NETDEV;
704 read_lock_irqsave(&table->rwlock, flags);
705 local_index = find_gid(table, gid, &val, false, mask, NULL);
706 if (local_index >= 0) {
707 get_gid_entry(table->data_vec[local_index]);
708 attr = &table->data_vec[local_index]->attr;
709 read_unlock_irqrestore(&table->rwlock, flags);
713 read_unlock_irqrestore(&table->rwlock, flags);
714 return ERR_PTR(-ENOENT);
716 EXPORT_SYMBOL(rdma_find_gid_by_port);
719 * rdma_find_gid_by_filter - Returns the GID table attribute where a
720 * specified GID value occurs
721 * @ib_dev: The device to query.
722 * @gid: The GID value to search for.
723 * @port: The port number of the device where the GID value could be
725 * @filter: The filter function is executed on any matching GID in the table.
726 * If the filter function returns true, the corresponding index is returned,
727 * otherwise, we continue searching the GID table. It's guaranteed that
728 * while filter is executed, ndev field is valid and the structure won't
729 * change. filter is executed in an atomic context. filter must not be NULL.
730 * @context: Private data to pass into the call-back.
732 * rdma_find_gid_by_filter() searches for the specified GID value
733 * of which the filter function returns true in the port's GID table.
736 const struct ib_gid_attr *rdma_find_gid_by_filter(
737 struct ib_device *ib_dev, const union ib_gid *gid, u32 port,
738 bool (*filter)(const union ib_gid *gid, const struct ib_gid_attr *,
742 const struct ib_gid_attr *res = ERR_PTR(-ENOENT);
743 struct ib_gid_table *table;
747 if (!rdma_is_port_valid(ib_dev, port))
748 return ERR_PTR(-EINVAL);
750 table = rdma_gid_table(ib_dev, port);
752 read_lock_irqsave(&table->rwlock, flags);
753 for (i = 0; i < table->sz; i++) {
754 struct ib_gid_table_entry *entry = table->data_vec[i];
756 if (!is_gid_entry_valid(entry))
759 if (memcmp(gid, &entry->attr.gid, sizeof(*gid)))
762 if (filter(gid, &entry->attr, context)) {
763 get_gid_entry(entry);
768 read_unlock_irqrestore(&table->rwlock, flags);
772 static struct ib_gid_table *alloc_gid_table(int sz)
774 struct ib_gid_table *table = kzalloc(sizeof(*table), GFP_KERNEL);
779 table->data_vec = kcalloc(sz, sizeof(*table->data_vec), GFP_KERNEL);
780 if (!table->data_vec)
783 mutex_init(&table->lock);
786 rwlock_init(&table->rwlock);
794 static void release_gid_table(struct ib_device *device,
795 struct ib_gid_table *table)
803 for (i = 0; i < table->sz; i++) {
804 if (is_gid_entry_free(table->data_vec[i]))
806 if (kref_read(&table->data_vec[i]->kref) > 1) {
807 dev_err(&device->dev,
808 "GID entry ref leak for index %d ref=%u\n", i,
809 kref_read(&table->data_vec[i]->kref));
816 mutex_destroy(&table->lock);
817 kfree(table->data_vec);
821 static void cleanup_gid_table_port(struct ib_device *ib_dev, u32 port,
822 struct ib_gid_table *table)
829 mutex_lock(&table->lock);
830 for (i = 0; i < table->sz; ++i) {
831 if (is_gid_entry_valid(table->data_vec[i]))
832 del_gid(ib_dev, port, table, i);
834 mutex_unlock(&table->lock);
837 void ib_cache_gid_set_default_gid(struct ib_device *ib_dev, u32 port,
838 struct net_device *ndev,
839 unsigned long gid_type_mask,
840 enum ib_cache_gid_default_mode mode)
842 union ib_gid gid = { };
843 struct ib_gid_attr gid_attr;
844 unsigned int gid_type;
847 mask = GID_ATTR_FIND_MASK_GID_TYPE |
848 GID_ATTR_FIND_MASK_DEFAULT |
849 GID_ATTR_FIND_MASK_NETDEV;
850 memset(&gid_attr, 0, sizeof(gid_attr));
851 gid_attr.ndev = ndev;
853 for (gid_type = 0; gid_type < IB_GID_TYPE_SIZE; ++gid_type) {
854 if (1UL << gid_type & ~gid_type_mask)
857 gid_attr.gid_type = gid_type;
859 if (mode == IB_CACHE_GID_DEFAULT_MODE_SET) {
860 make_default_gid(ndev, &gid);
861 __ib_cache_gid_add(ib_dev, port, &gid,
862 &gid_attr, mask, true);
863 } else if (mode == IB_CACHE_GID_DEFAULT_MODE_DELETE) {
864 _ib_cache_gid_del(ib_dev, port, &gid,
865 &gid_attr, mask, true);
870 static void gid_table_reserve_default(struct ib_device *ib_dev, u32 port,
871 struct ib_gid_table *table)
874 unsigned long roce_gid_type_mask;
875 unsigned int num_default_gids;
877 roce_gid_type_mask = roce_gid_type_mask_support(ib_dev, port);
878 num_default_gids = hweight_long(roce_gid_type_mask);
879 /* Reserve starting indices for default GIDs */
880 for (i = 0; i < num_default_gids && i < table->sz; i++)
881 table->default_gid_indices |= BIT(i);
885 static void gid_table_release_one(struct ib_device *ib_dev)
889 rdma_for_each_port (ib_dev, p) {
890 release_gid_table(ib_dev, ib_dev->port_data[p].cache.gid);
891 ib_dev->port_data[p].cache.gid = NULL;
895 static int _gid_table_setup_one(struct ib_device *ib_dev)
897 struct ib_gid_table *table;
900 rdma_for_each_port (ib_dev, rdma_port) {
901 table = alloc_gid_table(
902 ib_dev->port_data[rdma_port].immutable.gid_tbl_len);
904 goto rollback_table_setup;
906 gid_table_reserve_default(ib_dev, rdma_port, table);
907 ib_dev->port_data[rdma_port].cache.gid = table;
911 rollback_table_setup:
912 gid_table_release_one(ib_dev);
916 static void gid_table_cleanup_one(struct ib_device *ib_dev)
920 rdma_for_each_port (ib_dev, p)
921 cleanup_gid_table_port(ib_dev, p,
922 ib_dev->port_data[p].cache.gid);
925 static int gid_table_setup_one(struct ib_device *ib_dev)
929 err = _gid_table_setup_one(ib_dev);
934 rdma_roce_rescan_device(ib_dev);
940 * rdma_query_gid - Read the GID content from the GID software cache
941 * @device: Device to query the GID
942 * @port_num: Port number of the device
943 * @index: Index of the GID table entry to read
944 * @gid: Pointer to GID where to store the entry's GID
946 * rdma_query_gid() only reads the GID entry content for requested device,
947 * port and index. It reads for IB, RoCE and iWarp link layers. It doesn't
948 * hold any reference to the GID table entry in the HCA or software cache.
950 * Returns 0 on success or appropriate error code.
953 int rdma_query_gid(struct ib_device *device, u32 port_num,
954 int index, union ib_gid *gid)
956 struct ib_gid_table *table;
960 if (!rdma_is_port_valid(device, port_num))
963 table = rdma_gid_table(device, port_num);
964 read_lock_irqsave(&table->rwlock, flags);
966 if (index < 0 || index >= table->sz) {
971 if (!is_gid_entry_valid(table->data_vec[index])) {
976 memcpy(gid, &table->data_vec[index]->attr.gid, sizeof(*gid));
980 read_unlock_irqrestore(&table->rwlock, flags);
983 EXPORT_SYMBOL(rdma_query_gid);
986 * rdma_read_gid_hw_context - Read the HW GID context from GID attribute
987 * @attr: Potinter to the GID attribute
989 * rdma_read_gid_hw_context() reads the drivers GID HW context corresponding
990 * to the SGID attr. Callers are required to already be holding the reference
991 * to an existing GID entry.
993 * Returns the HW GID context
996 void *rdma_read_gid_hw_context(const struct ib_gid_attr *attr)
998 return container_of(attr, struct ib_gid_table_entry, attr)->context;
1000 EXPORT_SYMBOL(rdma_read_gid_hw_context);
1003 * rdma_find_gid - Returns SGID attributes if the matching GID is found.
1004 * @device: The device to query.
1005 * @gid: The GID value to search for.
1006 * @gid_type: The GID type to search for.
1007 * @ndev: In RoCE, the net device of the device. NULL means ignore.
1009 * rdma_find_gid() searches for the specified GID value in the software cache.
1011 * Returns GID attributes if a valid GID is found or returns ERR_PTR for the
1012 * error. The caller must invoke rdma_put_gid_attr() to release the reference.
1015 const struct ib_gid_attr *rdma_find_gid(struct ib_device *device,
1016 const union ib_gid *gid,
1017 enum ib_gid_type gid_type,
1018 struct net_device *ndev)
1020 unsigned long mask = GID_ATTR_FIND_MASK_GID |
1021 GID_ATTR_FIND_MASK_GID_TYPE;
1022 struct ib_gid_attr gid_attr_val = {.ndev = ndev, .gid_type = gid_type};
1026 mask |= GID_ATTR_FIND_MASK_NETDEV;
1028 rdma_for_each_port(device, p) {
1029 struct ib_gid_table *table;
1030 unsigned long flags;
1033 table = device->port_data[p].cache.gid;
1034 read_lock_irqsave(&table->rwlock, flags);
1035 index = find_gid(table, gid, &gid_attr_val, false, mask, NULL);
1037 const struct ib_gid_attr *attr;
1039 get_gid_entry(table->data_vec[index]);
1040 attr = &table->data_vec[index]->attr;
1041 read_unlock_irqrestore(&table->rwlock, flags);
1044 read_unlock_irqrestore(&table->rwlock, flags);
1047 return ERR_PTR(-ENOENT);
1049 EXPORT_SYMBOL(rdma_find_gid);
1051 int ib_get_cached_pkey(struct ib_device *device,
1056 struct ib_pkey_cache *cache;
1057 unsigned long flags;
1060 if (!rdma_is_port_valid(device, port_num))
1063 read_lock_irqsave(&device->cache_lock, flags);
1065 cache = device->port_data[port_num].cache.pkey;
1067 if (!cache || index < 0 || index >= cache->table_len)
1070 *pkey = cache->table[index];
1072 read_unlock_irqrestore(&device->cache_lock, flags);
1076 EXPORT_SYMBOL(ib_get_cached_pkey);
1078 void ib_get_cached_subnet_prefix(struct ib_device *device, u32 port_num,
1081 unsigned long flags;
1083 read_lock_irqsave(&device->cache_lock, flags);
1084 *sn_pfx = device->port_data[port_num].cache.subnet_prefix;
1085 read_unlock_irqrestore(&device->cache_lock, flags);
1087 EXPORT_SYMBOL(ib_get_cached_subnet_prefix);
1089 int ib_find_cached_pkey(struct ib_device *device, u32 port_num,
1090 u16 pkey, u16 *index)
1092 struct ib_pkey_cache *cache;
1093 unsigned long flags;
1096 int partial_ix = -1;
1098 if (!rdma_is_port_valid(device, port_num))
1101 read_lock_irqsave(&device->cache_lock, flags);
1103 cache = device->port_data[port_num].cache.pkey;
1111 for (i = 0; i < cache->table_len; ++i)
1112 if ((cache->table[i] & 0x7fff) == (pkey & 0x7fff)) {
1113 if (cache->table[i] & 0x8000) {
1122 if (ret && partial_ix >= 0) {
1123 *index = partial_ix;
1128 read_unlock_irqrestore(&device->cache_lock, flags);
1132 EXPORT_SYMBOL(ib_find_cached_pkey);
1134 int ib_find_exact_cached_pkey(struct ib_device *device, u32 port_num,
1135 u16 pkey, u16 *index)
1137 struct ib_pkey_cache *cache;
1138 unsigned long flags;
1142 if (!rdma_is_port_valid(device, port_num))
1145 read_lock_irqsave(&device->cache_lock, flags);
1147 cache = device->port_data[port_num].cache.pkey;
1155 for (i = 0; i < cache->table_len; ++i)
1156 if (cache->table[i] == pkey) {
1163 read_unlock_irqrestore(&device->cache_lock, flags);
1167 EXPORT_SYMBOL(ib_find_exact_cached_pkey);
1169 int ib_get_cached_lmc(struct ib_device *device, u32 port_num, u8 *lmc)
1171 unsigned long flags;
1174 if (!rdma_is_port_valid(device, port_num))
1177 read_lock_irqsave(&device->cache_lock, flags);
1178 *lmc = device->port_data[port_num].cache.lmc;
1179 read_unlock_irqrestore(&device->cache_lock, flags);
1183 EXPORT_SYMBOL(ib_get_cached_lmc);
1185 int ib_get_cached_port_state(struct ib_device *device, u32 port_num,
1186 enum ib_port_state *port_state)
1188 unsigned long flags;
1191 if (!rdma_is_port_valid(device, port_num))
1194 read_lock_irqsave(&device->cache_lock, flags);
1195 *port_state = device->port_data[port_num].cache.port_state;
1196 read_unlock_irqrestore(&device->cache_lock, flags);
1200 EXPORT_SYMBOL(ib_get_cached_port_state);
1203 * rdma_get_gid_attr - Returns GID attributes for a port of a device
1204 * at a requested gid_index, if a valid GID entry exists.
1205 * @device: The device to query.
1206 * @port_num: The port number on the device where the GID value
1208 * @index: Index of the GID table entry whose attributes are to
1211 * rdma_get_gid_attr() acquires reference count of gid attributes from the
1212 * cached GID table. Caller must invoke rdma_put_gid_attr() to release
1213 * reference to gid attribute regardless of link layer.
1215 * Returns pointer to valid gid attribute or ERR_PTR for the appropriate error
1218 const struct ib_gid_attr *
1219 rdma_get_gid_attr(struct ib_device *device, u32 port_num, int index)
1221 const struct ib_gid_attr *attr = ERR_PTR(-ENODATA);
1222 struct ib_gid_table *table;
1223 unsigned long flags;
1225 if (!rdma_is_port_valid(device, port_num))
1226 return ERR_PTR(-EINVAL);
1228 table = rdma_gid_table(device, port_num);
1229 if (index < 0 || index >= table->sz)
1230 return ERR_PTR(-EINVAL);
1232 read_lock_irqsave(&table->rwlock, flags);
1233 if (!is_gid_entry_valid(table->data_vec[index]))
1236 get_gid_entry(table->data_vec[index]);
1237 attr = &table->data_vec[index]->attr;
1239 read_unlock_irqrestore(&table->rwlock, flags);
1242 EXPORT_SYMBOL(rdma_get_gid_attr);
1245 * rdma_query_gid_table - Reads GID table entries of all the ports of a device up to max_entries.
1246 * @device: The device to query.
1247 * @entries: Entries where GID entries are returned.
1248 * @max_entries: Maximum number of entries that can be returned.
1249 * Entries array must be allocated to hold max_entries number of entries.
1251 * Returns number of entries on success or appropriate error code.
1253 ssize_t rdma_query_gid_table(struct ib_device *device,
1254 struct ib_uverbs_gid_entry *entries,
1257 const struct ib_gid_attr *gid_attr;
1258 ssize_t num_entries = 0, ret;
1259 struct ib_gid_table *table;
1261 struct net_device *ndev;
1262 unsigned long flags;
1264 rdma_for_each_port(device, port_num) {
1265 table = rdma_gid_table(device, port_num);
1266 read_lock_irqsave(&table->rwlock, flags);
1267 for (i = 0; i < table->sz; i++) {
1268 if (!is_gid_entry_valid(table->data_vec[i]))
1270 if (num_entries >= max_entries) {
1275 gid_attr = &table->data_vec[i]->attr;
1277 memcpy(&entries->gid, &gid_attr->gid,
1278 sizeof(gid_attr->gid));
1279 entries->gid_index = gid_attr->index;
1280 entries->port_num = gid_attr->port_num;
1281 entries->gid_type = gid_attr->gid_type;
1282 ndev = rcu_dereference_protected(
1284 lockdep_is_held(&table->rwlock));
1286 entries->netdev_ifindex = ndev->ifindex;
1291 read_unlock_irqrestore(&table->rwlock, flags);
1296 read_unlock_irqrestore(&table->rwlock, flags);
1299 EXPORT_SYMBOL(rdma_query_gid_table);
1302 * rdma_put_gid_attr - Release reference to the GID attribute
1303 * @attr: Pointer to the GID attribute whose reference
1304 * needs to be released.
1306 * rdma_put_gid_attr() must be used to release reference whose
1307 * reference is acquired using rdma_get_gid_attr() or any APIs
1308 * which returns a pointer to the ib_gid_attr regardless of link layer
1312 void rdma_put_gid_attr(const struct ib_gid_attr *attr)
1314 struct ib_gid_table_entry *entry =
1315 container_of(attr, struct ib_gid_table_entry, attr);
1317 put_gid_entry(entry);
1319 EXPORT_SYMBOL(rdma_put_gid_attr);
1322 * rdma_hold_gid_attr - Get reference to existing GID attribute
1324 * @attr: Pointer to the GID attribute whose reference
1325 * needs to be taken.
1327 * Increase the reference count to a GID attribute to keep it from being
1328 * freed. Callers are required to already be holding a reference to attribute.
1331 void rdma_hold_gid_attr(const struct ib_gid_attr *attr)
1333 struct ib_gid_table_entry *entry =
1334 container_of(attr, struct ib_gid_table_entry, attr);
1336 get_gid_entry(entry);
1338 EXPORT_SYMBOL(rdma_hold_gid_attr);
1341 * rdma_read_gid_attr_ndev_rcu - Read GID attribute netdevice
1342 * which must be in UP state.
1344 * @attr:Pointer to the GID attribute
1346 * Returns pointer to netdevice if the netdevice was attached to GID and
1347 * netdevice is in UP state. Caller must hold RCU lock as this API
1348 * reads the netdev flags which can change while netdevice migrates to
1349 * different net namespace. Returns ERR_PTR with error code otherwise.
1352 struct net_device *rdma_read_gid_attr_ndev_rcu(const struct ib_gid_attr *attr)
1354 struct ib_gid_table_entry *entry =
1355 container_of(attr, struct ib_gid_table_entry, attr);
1356 struct ib_device *device = entry->attr.device;
1357 struct net_device *ndev = ERR_PTR(-EINVAL);
1358 u32 port_num = entry->attr.port_num;
1359 struct ib_gid_table *table;
1360 unsigned long flags;
1363 table = rdma_gid_table(device, port_num);
1365 read_lock_irqsave(&table->rwlock, flags);
1366 valid = is_gid_entry_valid(table->data_vec[attr->index]);
1368 ndev = rcu_dereference(attr->ndev);
1370 ndev = ERR_PTR(-ENODEV);
1372 read_unlock_irqrestore(&table->rwlock, flags);
1375 EXPORT_SYMBOL(rdma_read_gid_attr_ndev_rcu);
1377 static int get_lower_dev_vlan(struct net_device *lower_dev,
1378 struct netdev_nested_priv *priv)
1380 u16 *vlan_id = (u16 *)priv->data;
1382 if (is_vlan_dev(lower_dev))
1383 *vlan_id = vlan_dev_vlan_id(lower_dev);
1385 /* We are interested only in first level vlan device, so
1386 * always return 1 to stop iterating over next level devices.
1392 * rdma_read_gid_l2_fields - Read the vlan ID and source MAC address
1395 * @attr: GID attribute pointer whose L2 fields to be read
1396 * @vlan_id: Pointer to vlan id to fill up if the GID entry has
1397 * vlan id. It is optional.
1398 * @smac: Pointer to smac to fill up for a GID entry. It is optional.
1400 * rdma_read_gid_l2_fields() returns 0 on success and returns vlan id
1401 * (if gid entry has vlan) and source MAC, or returns error.
1403 int rdma_read_gid_l2_fields(const struct ib_gid_attr *attr,
1404 u16 *vlan_id, u8 *smac)
1406 struct netdev_nested_priv priv = {
1407 .data = (void *)vlan_id,
1409 struct net_device *ndev;
1412 ndev = rcu_dereference(attr->ndev);
1418 ether_addr_copy(smac, ndev->dev_addr);
1421 if (is_vlan_dev(ndev)) {
1422 *vlan_id = vlan_dev_vlan_id(ndev);
1424 /* If the netdev is upper device and if it's lower
1425 * device is vlan device, consider vlan id of
1426 * the lower vlan device for this gid entry.
1428 netdev_walk_all_lower_dev_rcu(attr->ndev,
1429 get_lower_dev_vlan, &priv);
1435 EXPORT_SYMBOL(rdma_read_gid_l2_fields);
1437 static int config_non_roce_gid_cache(struct ib_device *device,
1438 u32 port, struct ib_port_attr *tprops)
1440 struct ib_gid_attr gid_attr = {};
1441 struct ib_gid_table *table;
1445 gid_attr.device = device;
1446 gid_attr.port_num = port;
1447 table = rdma_gid_table(device, port);
1449 mutex_lock(&table->lock);
1450 for (i = 0; i < tprops->gid_tbl_len; ++i) {
1451 if (!device->ops.query_gid)
1453 ret = device->ops.query_gid(device, port, i, &gid_attr.gid);
1455 dev_warn(&device->dev,
1456 "query_gid failed (%d) for index %d\n", ret,
1461 tprops->subnet_prefix =
1462 be64_to_cpu(gid_attr.gid.global.subnet_prefix);
1463 add_modify_gid(table, &gid_attr);
1466 mutex_unlock(&table->lock);
1471 ib_cache_update(struct ib_device *device, u32 port, bool update_gids,
1472 bool update_pkeys, bool enforce_security)
1474 struct ib_port_attr *tprops = NULL;
1475 struct ib_pkey_cache *pkey_cache = NULL;
1476 struct ib_pkey_cache *old_pkey_cache = NULL;
1480 if (!rdma_is_port_valid(device, port))
1483 tprops = kmalloc(sizeof *tprops, GFP_KERNEL);
1487 ret = ib_query_port(device, port, tprops);
1489 dev_warn(&device->dev, "ib_query_port failed (%d)\n", ret);
1493 if (!rdma_protocol_roce(device, port) && update_gids) {
1494 ret = config_non_roce_gid_cache(device, port,
1500 update_pkeys &= !!tprops->pkey_tbl_len;
1503 pkey_cache = kmalloc(struct_size(pkey_cache, table,
1504 tprops->pkey_tbl_len),
1511 pkey_cache->table_len = tprops->pkey_tbl_len;
1513 for (i = 0; i < pkey_cache->table_len; ++i) {
1514 ret = ib_query_pkey(device, port, i,
1515 pkey_cache->table + i);
1517 dev_warn(&device->dev,
1518 "ib_query_pkey failed (%d) for index %d\n",
1525 write_lock_irq(&device->cache_lock);
1528 old_pkey_cache = device->port_data[port].cache.pkey;
1529 device->port_data[port].cache.pkey = pkey_cache;
1531 device->port_data[port].cache.lmc = tprops->lmc;
1532 device->port_data[port].cache.port_state = tprops->state;
1534 device->port_data[port].cache.subnet_prefix = tprops->subnet_prefix;
1535 write_unlock_irq(&device->cache_lock);
1537 if (enforce_security)
1538 ib_security_cache_change(device,
1540 tprops->subnet_prefix);
1542 kfree(old_pkey_cache);
1552 static void ib_cache_event_task(struct work_struct *_work)
1554 struct ib_update_work *work =
1555 container_of(_work, struct ib_update_work, work);
1558 /* Before distributing the cache update event, first sync
1561 ret = ib_cache_update(work->event.device, work->event.element.port_num,
1562 work->event.event == IB_EVENT_GID_CHANGE,
1563 work->event.event == IB_EVENT_PKEY_CHANGE,
1564 work->enforce_security);
1566 /* GID event is notified already for individual GID entries by
1567 * dispatch_gid_change_event(). Hence, notifiy for rest of the
1570 if (!ret && work->event.event != IB_EVENT_GID_CHANGE)
1571 ib_dispatch_event_clients(&work->event);
1576 static void ib_generic_event_task(struct work_struct *_work)
1578 struct ib_update_work *work =
1579 container_of(_work, struct ib_update_work, work);
1581 ib_dispatch_event_clients(&work->event);
1585 static bool is_cache_update_event(const struct ib_event *event)
1587 return (event->event == IB_EVENT_PORT_ERR ||
1588 event->event == IB_EVENT_PORT_ACTIVE ||
1589 event->event == IB_EVENT_LID_CHANGE ||
1590 event->event == IB_EVENT_PKEY_CHANGE ||
1591 event->event == IB_EVENT_CLIENT_REREGISTER ||
1592 event->event == IB_EVENT_GID_CHANGE);
1596 * ib_dispatch_event - Dispatch an asynchronous event
1597 * @event:Event to dispatch
1599 * Low-level drivers must call ib_dispatch_event() to dispatch the
1600 * event to all registered event handlers when an asynchronous event
1603 void ib_dispatch_event(const struct ib_event *event)
1605 struct ib_update_work *work;
1607 work = kzalloc(sizeof(*work), GFP_ATOMIC);
1611 if (is_cache_update_event(event))
1612 INIT_WORK(&work->work, ib_cache_event_task);
1614 INIT_WORK(&work->work, ib_generic_event_task);
1616 work->event = *event;
1617 if (event->event == IB_EVENT_PKEY_CHANGE ||
1618 event->event == IB_EVENT_GID_CHANGE)
1619 work->enforce_security = true;
1621 queue_work(ib_wq, &work->work);
1623 EXPORT_SYMBOL(ib_dispatch_event);
1625 int ib_cache_setup_one(struct ib_device *device)
1630 err = gid_table_setup_one(device);
1634 rdma_for_each_port (device, p) {
1635 err = ib_cache_update(device, p, true, true, true);
1643 void ib_cache_release_one(struct ib_device *device)
1648 * The release function frees all the cache elements.
1649 * This function should be called as part of freeing
1650 * all the device's resources when the cache could no
1651 * longer be accessed.
1653 rdma_for_each_port (device, p)
1654 kfree(device->port_data[p].cache.pkey);
1656 gid_table_release_one(device);
1659 void ib_cache_cleanup_one(struct ib_device *device)
1661 /* The cleanup function waits for all in-progress workqueue
1662 * elements and cleans up the GID cache. This function should be
1663 * called after the device was removed from the devices list and
1664 * all clients were removed, so the cache exists but is
1665 * non-functional and shouldn't be updated anymore.
1667 flush_workqueue(ib_wq);
1668 gid_table_cleanup_one(device);
1671 * Flush the wq second time for any pending GID delete work.
1673 flush_workqueue(ib_wq);