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;
61 static const struct ib_gid_attr zattr;
63 enum gid_attr_find_mask {
64 GID_ATTR_FIND_MASK_GID = 1UL << 0,
65 GID_ATTR_FIND_MASK_NETDEV = 1UL << 1,
66 GID_ATTR_FIND_MASK_DEFAULT = 1UL << 2,
67 GID_ATTR_FIND_MASK_GID_TYPE = 1UL << 3,
70 enum gid_table_entry_props {
71 GID_TABLE_ENTRY_INVALID = 1UL << 0,
72 GID_TABLE_ENTRY_DEFAULT = 1UL << 1,
75 enum gid_table_write_action {
76 GID_TABLE_WRITE_ACTION_ADD,
77 GID_TABLE_WRITE_ACTION_DEL,
78 /* MODIFY only updates the GID table. Currently only used by
81 GID_TABLE_WRITE_ACTION_MODIFY
84 struct ib_gid_table_entry {
87 struct ib_gid_attr attr;
93 /* In RoCE, adding a GID to the table requires:
94 * (a) Find if this GID is already exists.
95 * (b) Find a free space.
96 * (c) Write the new GID
98 * Delete requires different set of operations:
102 * Add/delete should be carried out atomically.
103 * This is done by locking this mutex from multiple
104 * writers. We don't need this lock for IB, as the MAD
105 * layer replaces all entries. All data_vec entries
106 * are locked by this lock.
109 /* This lock protects the table entries from being
110 * read and written simultaneously.
113 struct ib_gid_table_entry *data_vec;
116 static void dispatch_gid_change_event(struct ib_device *ib_dev, u8 port)
118 if (rdma_cap_roce_gid_table(ib_dev, port)) {
119 struct ib_event event;
121 event.device = ib_dev;
122 event.element.port_num = port;
123 event.event = IB_EVENT_GID_CHANGE;
125 ib_dispatch_event(&event);
129 static const char * const gid_type_str[] = {
130 [IB_GID_TYPE_IB] = "IB/RoCE v1",
131 [IB_GID_TYPE_ROCE_UDP_ENCAP] = "RoCE v2",
134 const char *ib_cache_gid_type_str(enum ib_gid_type gid_type)
136 if (gid_type < ARRAY_SIZE(gid_type_str) && gid_type_str[gid_type])
137 return gid_type_str[gid_type];
139 return "Invalid GID type";
141 EXPORT_SYMBOL(ib_cache_gid_type_str);
143 int ib_cache_gid_parse_type_str(const char *buf)
153 if (buf[len - 1] == '\n')
156 for (i = 0; i < ARRAY_SIZE(gid_type_str); ++i)
157 if (gid_type_str[i] && !strncmp(buf, gid_type_str[i], len) &&
158 len == strlen(gid_type_str[i])) {
165 EXPORT_SYMBOL(ib_cache_gid_parse_type_str);
167 /* This function expects that rwlock will be write locked in all
168 * scenarios and that lock will be locked in sleep-able (RoCE)
171 static int write_gid(struct ib_device *ib_dev, u8 port,
172 struct ib_gid_table *table, int ix,
173 const union ib_gid *gid,
174 const struct ib_gid_attr *attr,
175 enum gid_table_write_action action,
177 __releases(&table->rwlock) __acquires(&table->rwlock)
180 struct net_device *old_net_dev;
181 enum ib_gid_type old_gid_type;
183 /* in rdma_cap_roce_gid_table, this funciton should be protected by a
187 if (rdma_cap_roce_gid_table(ib_dev, port)) {
188 table->data_vec[ix].props |= GID_TABLE_ENTRY_INVALID;
189 write_unlock_irq(&table->rwlock);
190 /* GID_TABLE_WRITE_ACTION_MODIFY currently isn't supported by
191 * RoCE providers and thus only updates the cache.
193 if (action == GID_TABLE_WRITE_ACTION_ADD)
194 ret = ib_dev->add_gid(ib_dev, port, ix, gid, attr,
195 &table->data_vec[ix].context);
196 else if (action == GID_TABLE_WRITE_ACTION_DEL)
197 ret = ib_dev->del_gid(ib_dev, port, ix,
198 &table->data_vec[ix].context);
199 write_lock_irq(&table->rwlock);
202 old_net_dev = table->data_vec[ix].attr.ndev;
203 old_gid_type = table->data_vec[ix].attr.gid_type;
204 if (old_net_dev && old_net_dev != attr->ndev)
205 dev_put(old_net_dev);
206 /* if modify_gid failed, just delete the old gid */
207 if (ret || action == GID_TABLE_WRITE_ACTION_DEL) {
210 table->data_vec[ix].context = NULL;
213 memcpy(&table->data_vec[ix].gid, gid, sizeof(*gid));
214 memcpy(&table->data_vec[ix].attr, attr, sizeof(*attr));
216 table->data_vec[ix].props |= GID_TABLE_ENTRY_DEFAULT;
217 if (action == GID_TABLE_WRITE_ACTION_DEL)
218 table->data_vec[ix].attr.gid_type = old_gid_type;
220 if (table->data_vec[ix].attr.ndev &&
221 table->data_vec[ix].attr.ndev != old_net_dev)
222 dev_hold(table->data_vec[ix].attr.ndev);
224 table->data_vec[ix].props &= ~GID_TABLE_ENTRY_INVALID;
229 static int add_gid(struct ib_device *ib_dev, u8 port,
230 struct ib_gid_table *table, int ix,
231 const union ib_gid *gid,
232 const struct ib_gid_attr *attr,
234 return write_gid(ib_dev, port, table, ix, gid, attr,
235 GID_TABLE_WRITE_ACTION_ADD, default_gid);
238 static int modify_gid(struct ib_device *ib_dev, u8 port,
239 struct ib_gid_table *table, int ix,
240 const union ib_gid *gid,
241 const struct ib_gid_attr *attr,
243 return write_gid(ib_dev, port, table, ix, gid, attr,
244 GID_TABLE_WRITE_ACTION_MODIFY, default_gid);
247 static int del_gid(struct ib_device *ib_dev, u8 port,
248 struct ib_gid_table *table, int ix,
250 return write_gid(ib_dev, port, table, ix, &zgid, &zattr,
251 GID_TABLE_WRITE_ACTION_DEL, default_gid);
254 /* rwlock should be read locked */
255 static int find_gid(struct ib_gid_table *table, const union ib_gid *gid,
256 const struct ib_gid_attr *val, bool default_gid,
257 unsigned long mask, int *pempty)
261 int empty = pempty ? -1 : 0;
263 while (i < table->sz && (found < 0 || empty < 0)) {
264 struct ib_gid_table_entry *data = &table->data_vec[i];
265 struct ib_gid_attr *attr = &data->attr;
270 if (data->props & GID_TABLE_ENTRY_INVALID)
274 if (!memcmp(&data->gid, &zgid, sizeof(*gid)) &&
275 !memcmp(attr, &zattr, sizeof(*attr)) &&
282 if (mask & GID_ATTR_FIND_MASK_GID_TYPE &&
283 attr->gid_type != val->gid_type)
286 if (mask & GID_ATTR_FIND_MASK_GID &&
287 memcmp(gid, &data->gid, sizeof(*gid)))
290 if (mask & GID_ATTR_FIND_MASK_NETDEV &&
291 attr->ndev != val->ndev)
294 if (mask & GID_ATTR_FIND_MASK_DEFAULT &&
295 !!(data->props & GID_TABLE_ENTRY_DEFAULT) !=
308 static void make_default_gid(struct net_device *dev, union ib_gid *gid)
310 gid->global.subnet_prefix = cpu_to_be64(0xfe80000000000000LL);
311 addrconf_ifid_eui48(&gid->raw[8], dev);
314 int ib_cache_gid_add(struct ib_device *ib_dev, u8 port,
315 union ib_gid *gid, struct ib_gid_attr *attr)
317 struct ib_gid_table **ports_table = ib_dev->cache.gid_cache;
318 struct ib_gid_table *table;
321 struct net_device *idev;
324 table = ports_table[port - rdma_start_port(ib_dev)];
326 if (!memcmp(gid, &zgid, sizeof(*gid)))
329 if (ib_dev->get_netdev) {
330 idev = ib_dev->get_netdev(ib_dev, port);
331 if (idev && attr->ndev != idev) {
332 union ib_gid default_gid;
334 /* Adding default GIDs in not permitted */
335 make_default_gid(idev, &default_gid);
336 if (!memcmp(gid, &default_gid, sizeof(*gid))) {
345 mutex_lock(&table->lock);
346 write_lock_irq(&table->rwlock);
348 ix = find_gid(table, gid, attr, false, GID_ATTR_FIND_MASK_GID |
349 GID_ATTR_FIND_MASK_GID_TYPE |
350 GID_ATTR_FIND_MASK_NETDEV, &empty);
359 ret = add_gid(ib_dev, port, table, empty, gid, attr, false);
361 dispatch_gid_change_event(ib_dev, port);
364 write_unlock_irq(&table->rwlock);
365 mutex_unlock(&table->lock);
369 int ib_cache_gid_del(struct ib_device *ib_dev, u8 port,
370 union ib_gid *gid, struct ib_gid_attr *attr)
372 struct ib_gid_table **ports_table = ib_dev->cache.gid_cache;
373 struct ib_gid_table *table;
376 table = ports_table[port - rdma_start_port(ib_dev)];
378 mutex_lock(&table->lock);
379 write_lock_irq(&table->rwlock);
381 ix = find_gid(table, gid, attr, false,
382 GID_ATTR_FIND_MASK_GID |
383 GID_ATTR_FIND_MASK_GID_TYPE |
384 GID_ATTR_FIND_MASK_NETDEV |
385 GID_ATTR_FIND_MASK_DEFAULT,
390 if (!del_gid(ib_dev, port, table, ix, false))
391 dispatch_gid_change_event(ib_dev, port);
394 write_unlock_irq(&table->rwlock);
395 mutex_unlock(&table->lock);
399 int ib_cache_gid_del_all_netdev_gids(struct ib_device *ib_dev, u8 port,
400 struct net_device *ndev)
402 struct ib_gid_table **ports_table = ib_dev->cache.gid_cache;
403 struct ib_gid_table *table;
405 bool deleted = false;
407 table = ports_table[port - rdma_start_port(ib_dev)];
409 mutex_lock(&table->lock);
410 write_lock_irq(&table->rwlock);
412 for (ix = 0; ix < table->sz; ix++)
413 if (table->data_vec[ix].attr.ndev == ndev)
414 if (!del_gid(ib_dev, port, table, ix,
415 !!(table->data_vec[ix].props &
416 GID_TABLE_ENTRY_DEFAULT)))
419 write_unlock_irq(&table->rwlock);
420 mutex_unlock(&table->lock);
423 dispatch_gid_change_event(ib_dev, port);
428 static int __ib_cache_gid_get(struct ib_device *ib_dev, u8 port, int index,
429 union ib_gid *gid, struct ib_gid_attr *attr)
431 struct ib_gid_table **ports_table = ib_dev->cache.gid_cache;
432 struct ib_gid_table *table;
434 table = ports_table[port - rdma_start_port(ib_dev)];
436 if (index < 0 || index >= table->sz)
439 if (table->data_vec[index].props & GID_TABLE_ENTRY_INVALID)
442 memcpy(gid, &table->data_vec[index].gid, sizeof(*gid));
444 memcpy(attr, &table->data_vec[index].attr, sizeof(*attr));
446 dev_hold(attr->ndev);
452 static int _ib_cache_gid_table_find(struct ib_device *ib_dev,
453 const union ib_gid *gid,
454 const struct ib_gid_attr *val,
456 u8 *port, u16 *index)
458 struct ib_gid_table **ports_table = ib_dev->cache.gid_cache;
459 struct ib_gid_table *table;
464 for (p = 0; p < ib_dev->phys_port_cnt; p++) {
465 table = ports_table[p];
466 read_lock_irqsave(&table->rwlock, flags);
467 local_index = find_gid(table, gid, val, false, mask, NULL);
468 if (local_index >= 0) {
470 *index = local_index;
472 *port = p + rdma_start_port(ib_dev);
473 read_unlock_irqrestore(&table->rwlock, flags);
476 read_unlock_irqrestore(&table->rwlock, flags);
482 static int ib_cache_gid_find(struct ib_device *ib_dev,
483 const union ib_gid *gid,
484 enum ib_gid_type gid_type,
485 struct net_device *ndev, u8 *port,
488 unsigned long mask = GID_ATTR_FIND_MASK_GID |
489 GID_ATTR_FIND_MASK_GID_TYPE;
490 struct ib_gid_attr gid_attr_val = {.ndev = ndev, .gid_type = gid_type};
493 mask |= GID_ATTR_FIND_MASK_NETDEV;
495 return _ib_cache_gid_table_find(ib_dev, gid, &gid_attr_val,
499 int ib_find_cached_gid_by_port(struct ib_device *ib_dev,
500 const union ib_gid *gid,
501 enum ib_gid_type gid_type,
502 u8 port, struct net_device *ndev,
506 struct ib_gid_table **ports_table = ib_dev->cache.gid_cache;
507 struct ib_gid_table *table;
508 unsigned long mask = GID_ATTR_FIND_MASK_GID |
509 GID_ATTR_FIND_MASK_GID_TYPE;
510 struct ib_gid_attr val = {.ndev = ndev, .gid_type = gid_type};
513 if (port < rdma_start_port(ib_dev) ||
514 port > rdma_end_port(ib_dev))
517 table = ports_table[port - rdma_start_port(ib_dev)];
520 mask |= GID_ATTR_FIND_MASK_NETDEV;
522 read_lock_irqsave(&table->rwlock, flags);
523 local_index = find_gid(table, gid, &val, false, mask, NULL);
524 if (local_index >= 0) {
526 *index = local_index;
527 read_unlock_irqrestore(&table->rwlock, flags);
531 read_unlock_irqrestore(&table->rwlock, flags);
534 EXPORT_SYMBOL(ib_find_cached_gid_by_port);
537 * ib_find_gid_by_filter - Returns the GID table index where a specified
539 * @device: The device to query.
540 * @gid: The GID value to search for.
541 * @port_num: The port number of the device where the GID value could be
543 * @filter: The filter function is executed on any matching GID in the table.
544 * If the filter function returns true, the corresponding index is returned,
545 * otherwise, we continue searching the GID table. It's guaranteed that
546 * while filter is executed, ndev field is valid and the structure won't
547 * change. filter is executed in an atomic context. filter must not be NULL.
548 * @index: The index into the cached GID table where the GID was found. This
549 * parameter may be NULL.
551 * ib_cache_gid_find_by_filter() searches for the specified GID value
552 * of which the filter function returns true in the port's GID table.
553 * This function is only supported on RoCE ports.
556 static int ib_cache_gid_find_by_filter(struct ib_device *ib_dev,
557 const union ib_gid *gid,
559 bool (*filter)(const union ib_gid *,
560 const struct ib_gid_attr *,
565 struct ib_gid_table **ports_table = ib_dev->cache.gid_cache;
566 struct ib_gid_table *table;
574 if (port < rdma_start_port(ib_dev) ||
575 port > rdma_end_port(ib_dev) ||
576 !rdma_protocol_roce(ib_dev, port))
577 return -EPROTONOSUPPORT;
579 table = ports_table[port - rdma_start_port(ib_dev)];
581 read_lock_irqsave(&table->rwlock, flags);
582 for (i = 0; i < table->sz; i++) {
583 struct ib_gid_attr attr;
585 if (table->data_vec[i].props & GID_TABLE_ENTRY_INVALID)
588 if (memcmp(gid, &table->data_vec[i].gid, sizeof(*gid)))
591 memcpy(&attr, &table->data_vec[i].attr, sizeof(attr));
593 if (filter(gid, &attr, context))
600 read_unlock_irqrestore(&table->rwlock, flags);
610 static struct ib_gid_table *alloc_gid_table(int sz)
612 struct ib_gid_table *table =
613 kzalloc(sizeof(struct ib_gid_table), GFP_KERNEL);
618 table->data_vec = kcalloc(sz, sizeof(*table->data_vec), GFP_KERNEL);
619 if (!table->data_vec)
622 mutex_init(&table->lock);
625 rwlock_init(&table->rwlock);
634 static void release_gid_table(struct ib_gid_table *table)
637 kfree(table->data_vec);
642 static void cleanup_gid_table_port(struct ib_device *ib_dev, u8 port,
643 struct ib_gid_table *table)
646 bool deleted = false;
651 write_lock_irq(&table->rwlock);
652 for (i = 0; i < table->sz; ++i) {
653 if (memcmp(&table->data_vec[i].gid, &zgid,
654 sizeof(table->data_vec[i].gid)))
655 if (!del_gid(ib_dev, port, table, i,
656 table->data_vec[i].props &
657 GID_ATTR_FIND_MASK_DEFAULT))
660 write_unlock_irq(&table->rwlock);
663 dispatch_gid_change_event(ib_dev, port);
666 void ib_cache_gid_set_default_gid(struct ib_device *ib_dev, u8 port,
667 struct net_device *ndev,
668 unsigned long gid_type_mask,
669 enum ib_cache_gid_default_mode mode)
671 struct ib_gid_table **ports_table = ib_dev->cache.gid_cache;
673 struct ib_gid_attr gid_attr;
674 struct ib_gid_attr zattr_type = zattr;
675 struct ib_gid_table *table;
676 unsigned int gid_type;
678 table = ports_table[port - rdma_start_port(ib_dev)];
680 make_default_gid(ndev, &gid);
681 memset(&gid_attr, 0, sizeof(gid_attr));
682 gid_attr.ndev = ndev;
684 for (gid_type = 0; gid_type < IB_GID_TYPE_SIZE; ++gid_type) {
686 union ib_gid current_gid;
687 struct ib_gid_attr current_gid_attr = {};
689 if (1UL << gid_type & ~gid_type_mask)
692 gid_attr.gid_type = gid_type;
694 mutex_lock(&table->lock);
695 write_lock_irq(&table->rwlock);
696 ix = find_gid(table, NULL, &gid_attr, true,
697 GID_ATTR_FIND_MASK_GID_TYPE |
698 GID_ATTR_FIND_MASK_DEFAULT,
701 /* Coudn't find default GID location */
705 zattr_type.gid_type = gid_type;
707 if (!__ib_cache_gid_get(ib_dev, port, ix,
708 ¤t_gid, ¤t_gid_attr) &&
709 mode == IB_CACHE_GID_DEFAULT_MODE_SET &&
710 !memcmp(&gid, ¤t_gid, sizeof(gid)) &&
711 !memcmp(&gid_attr, ¤t_gid_attr, sizeof(gid_attr)))
714 if (memcmp(¤t_gid, &zgid, sizeof(current_gid)) ||
715 memcmp(¤t_gid_attr, &zattr_type,
716 sizeof(current_gid_attr))) {
717 if (del_gid(ib_dev, port, table, ix, true)) {
718 pr_warn("ib_cache_gid: can't delete index %d for default gid %pI6\n",
722 dispatch_gid_change_event(ib_dev, port);
726 if (mode == IB_CACHE_GID_DEFAULT_MODE_SET) {
727 if (add_gid(ib_dev, port, table, ix, &gid, &gid_attr, true))
728 pr_warn("ib_cache_gid: unable to add default gid %pI6\n",
731 dispatch_gid_change_event(ib_dev, port);
735 if (current_gid_attr.ndev)
736 dev_put(current_gid_attr.ndev);
737 write_unlock_irq(&table->rwlock);
738 mutex_unlock(&table->lock);
742 static int gid_table_reserve_default(struct ib_device *ib_dev, u8 port,
743 struct ib_gid_table *table)
746 unsigned long roce_gid_type_mask;
747 unsigned int num_default_gids;
748 unsigned int current_gid = 0;
750 roce_gid_type_mask = roce_gid_type_mask_support(ib_dev, port);
751 num_default_gids = hweight_long(roce_gid_type_mask);
752 for (i = 0; i < num_default_gids && i < table->sz; i++) {
753 struct ib_gid_table_entry *entry =
756 entry->props |= GID_TABLE_ENTRY_DEFAULT;
757 current_gid = find_next_bit(&roce_gid_type_mask,
760 entry->attr.gid_type = current_gid++;
766 static int _gid_table_setup_one(struct ib_device *ib_dev)
769 struct ib_gid_table **table;
772 table = kcalloc(ib_dev->phys_port_cnt, sizeof(*table), GFP_KERNEL);
775 pr_warn("failed to allocate ib gid cache for %s\n",
780 for (port = 0; port < ib_dev->phys_port_cnt; port++) {
781 u8 rdma_port = port + rdma_start_port(ib_dev);
785 ib_dev->port_immutable[rdma_port].gid_tbl_len);
788 goto rollback_table_setup;
791 err = gid_table_reserve_default(ib_dev,
792 port + rdma_start_port(ib_dev),
795 goto rollback_table_setup;
798 ib_dev->cache.gid_cache = table;
801 rollback_table_setup:
802 for (port = 0; port < ib_dev->phys_port_cnt; port++) {
803 cleanup_gid_table_port(ib_dev, port + rdma_start_port(ib_dev),
805 release_gid_table(table[port]);
812 static void gid_table_release_one(struct ib_device *ib_dev)
814 struct ib_gid_table **table = ib_dev->cache.gid_cache;
820 for (port = 0; port < ib_dev->phys_port_cnt; port++)
821 release_gid_table(table[port]);
824 ib_dev->cache.gid_cache = NULL;
827 static void gid_table_cleanup_one(struct ib_device *ib_dev)
829 struct ib_gid_table **table = ib_dev->cache.gid_cache;
835 for (port = 0; port < ib_dev->phys_port_cnt; port++)
836 cleanup_gid_table_port(ib_dev, port + rdma_start_port(ib_dev),
840 static int gid_table_setup_one(struct ib_device *ib_dev)
844 err = _gid_table_setup_one(ib_dev);
849 err = roce_rescan_device(ib_dev);
852 gid_table_cleanup_one(ib_dev);
853 gid_table_release_one(ib_dev);
859 int ib_get_cached_gid(struct ib_device *device,
863 struct ib_gid_attr *gid_attr)
867 struct ib_gid_table **ports_table = device->cache.gid_cache;
868 struct ib_gid_table *table = ports_table[port_num - rdma_start_port(device)];
870 if (port_num < rdma_start_port(device) || port_num > rdma_end_port(device))
873 read_lock_irqsave(&table->rwlock, flags);
874 res = __ib_cache_gid_get(device, port_num, index, gid, gid_attr);
875 read_unlock_irqrestore(&table->rwlock, flags);
879 EXPORT_SYMBOL(ib_get_cached_gid);
881 int ib_find_cached_gid(struct ib_device *device,
882 const union ib_gid *gid,
883 enum ib_gid_type gid_type,
884 struct net_device *ndev,
888 return ib_cache_gid_find(device, gid, gid_type, ndev, port_num, index);
890 EXPORT_SYMBOL(ib_find_cached_gid);
892 int ib_find_gid_by_filter(struct ib_device *device,
893 const union ib_gid *gid,
895 bool (*filter)(const union ib_gid *gid,
896 const struct ib_gid_attr *,
898 void *context, u16 *index)
900 /* Only RoCE GID table supports filter function */
901 if (!rdma_cap_roce_gid_table(device, port_num) && filter)
902 return -EPROTONOSUPPORT;
904 return ib_cache_gid_find_by_filter(device, gid,
908 EXPORT_SYMBOL(ib_find_gid_by_filter);
910 int ib_get_cached_pkey(struct ib_device *device,
915 struct ib_pkey_cache *cache;
919 if (port_num < rdma_start_port(device) || port_num > rdma_end_port(device))
922 read_lock_irqsave(&device->cache.lock, flags);
924 cache = device->cache.pkey_cache[port_num - rdma_start_port(device)];
926 if (index < 0 || index >= cache->table_len)
929 *pkey = cache->table[index];
931 read_unlock_irqrestore(&device->cache.lock, flags);
935 EXPORT_SYMBOL(ib_get_cached_pkey);
937 int ib_find_cached_pkey(struct ib_device *device,
942 struct ib_pkey_cache *cache;
948 if (port_num < rdma_start_port(device) || port_num > rdma_end_port(device))
951 read_lock_irqsave(&device->cache.lock, flags);
953 cache = device->cache.pkey_cache[port_num - rdma_start_port(device)];
957 for (i = 0; i < cache->table_len; ++i)
958 if ((cache->table[i] & 0x7fff) == (pkey & 0x7fff)) {
959 if (cache->table[i] & 0x8000) {
967 if (ret && partial_ix >= 0) {
972 read_unlock_irqrestore(&device->cache.lock, flags);
976 EXPORT_SYMBOL(ib_find_cached_pkey);
978 int ib_find_exact_cached_pkey(struct ib_device *device,
983 struct ib_pkey_cache *cache;
988 if (port_num < rdma_start_port(device) || port_num > rdma_end_port(device))
991 read_lock_irqsave(&device->cache.lock, flags);
993 cache = device->cache.pkey_cache[port_num - rdma_start_port(device)];
997 for (i = 0; i < cache->table_len; ++i)
998 if (cache->table[i] == pkey) {
1004 read_unlock_irqrestore(&device->cache.lock, flags);
1008 EXPORT_SYMBOL(ib_find_exact_cached_pkey);
1010 int ib_get_cached_lmc(struct ib_device *device,
1014 unsigned long flags;
1017 if (port_num < rdma_start_port(device) || port_num > rdma_end_port(device))
1020 read_lock_irqsave(&device->cache.lock, flags);
1021 *lmc = device->cache.lmc_cache[port_num - rdma_start_port(device)];
1022 read_unlock_irqrestore(&device->cache.lock, flags);
1026 EXPORT_SYMBOL(ib_get_cached_lmc);
1028 static void ib_cache_update(struct ib_device *device,
1031 struct ib_port_attr *tprops = NULL;
1032 struct ib_pkey_cache *pkey_cache = NULL, *old_pkey_cache;
1033 struct ib_gid_cache {
1035 union ib_gid table[0];
1036 } *gid_cache = NULL;
1039 struct ib_gid_table *table;
1040 struct ib_gid_table **ports_table = device->cache.gid_cache;
1041 bool use_roce_gid_table =
1042 rdma_cap_roce_gid_table(device, port);
1044 if (port < rdma_start_port(device) || port > rdma_end_port(device))
1047 table = ports_table[port - rdma_start_port(device)];
1049 tprops = kmalloc(sizeof *tprops, GFP_KERNEL);
1053 ret = ib_query_port(device, port, tprops);
1055 pr_warn("ib_query_port failed (%d) for %s\n",
1060 pkey_cache = kmalloc(sizeof *pkey_cache + tprops->pkey_tbl_len *
1061 sizeof *pkey_cache->table, GFP_KERNEL);
1065 pkey_cache->table_len = tprops->pkey_tbl_len;
1067 if (!use_roce_gid_table) {
1068 gid_cache = kmalloc(sizeof(*gid_cache) + tprops->gid_tbl_len *
1069 sizeof(*gid_cache->table), GFP_KERNEL);
1073 gid_cache->table_len = tprops->gid_tbl_len;
1076 for (i = 0; i < pkey_cache->table_len; ++i) {
1077 ret = ib_query_pkey(device, port, i, pkey_cache->table + i);
1079 pr_warn("ib_query_pkey failed (%d) for %s (index %d)\n",
1080 ret, device->name, i);
1085 if (!use_roce_gid_table) {
1086 for (i = 0; i < gid_cache->table_len; ++i) {
1087 ret = ib_query_gid(device, port, i,
1088 gid_cache->table + i, NULL);
1090 pr_warn("ib_query_gid failed (%d) for %s (index %d)\n",
1091 ret, device->name, i);
1097 write_lock_irq(&device->cache.lock);
1099 old_pkey_cache = device->cache.pkey_cache[port - rdma_start_port(device)];
1101 device->cache.pkey_cache[port - rdma_start_port(device)] = pkey_cache;
1102 if (!use_roce_gid_table) {
1103 write_lock(&table->rwlock);
1104 for (i = 0; i < gid_cache->table_len; i++) {
1105 modify_gid(device, port, table, i, gid_cache->table + i,
1108 write_unlock(&table->rwlock);
1111 device->cache.lmc_cache[port - rdma_start_port(device)] = tprops->lmc;
1113 write_unlock_irq(&device->cache.lock);
1116 kfree(old_pkey_cache);
1126 static void ib_cache_task(struct work_struct *_work)
1128 struct ib_update_work *work =
1129 container_of(_work, struct ib_update_work, work);
1131 ib_cache_update(work->device, work->port_num);
1135 static void ib_cache_event(struct ib_event_handler *handler,
1136 struct ib_event *event)
1138 struct ib_update_work *work;
1140 if (event->event == IB_EVENT_PORT_ERR ||
1141 event->event == IB_EVENT_PORT_ACTIVE ||
1142 event->event == IB_EVENT_LID_CHANGE ||
1143 event->event == IB_EVENT_PKEY_CHANGE ||
1144 event->event == IB_EVENT_SM_CHANGE ||
1145 event->event == IB_EVENT_CLIENT_REREGISTER ||
1146 event->event == IB_EVENT_GID_CHANGE) {
1147 work = kmalloc(sizeof *work, GFP_ATOMIC);
1149 INIT_WORK(&work->work, ib_cache_task);
1150 work->device = event->device;
1151 work->port_num = event->element.port_num;
1152 queue_work(ib_wq, &work->work);
1157 int ib_cache_setup_one(struct ib_device *device)
1162 rwlock_init(&device->cache.lock);
1164 device->cache.pkey_cache =
1165 kzalloc(sizeof *device->cache.pkey_cache *
1166 (rdma_end_port(device) - rdma_start_port(device) + 1), GFP_KERNEL);
1167 device->cache.lmc_cache = kmalloc(sizeof *device->cache.lmc_cache *
1168 (rdma_end_port(device) -
1169 rdma_start_port(device) + 1),
1171 if (!device->cache.pkey_cache ||
1172 !device->cache.lmc_cache) {
1173 pr_warn("Couldn't allocate cache for %s\n", device->name);
1177 err = gid_table_setup_one(device);
1179 /* Allocated memory will be cleaned in the release function */
1182 for (p = 0; p <= rdma_end_port(device) - rdma_start_port(device); ++p)
1183 ib_cache_update(device, p + rdma_start_port(device));
1185 INIT_IB_EVENT_HANDLER(&device->cache.event_handler,
1186 device, ib_cache_event);
1187 err = ib_register_event_handler(&device->cache.event_handler);
1194 gid_table_cleanup_one(device);
1198 void ib_cache_release_one(struct ib_device *device)
1203 * The release function frees all the cache elements.
1204 * This function should be called as part of freeing
1205 * all the device's resources when the cache could no
1206 * longer be accessed.
1208 if (device->cache.pkey_cache)
1210 p <= rdma_end_port(device) - rdma_start_port(device); ++p)
1211 kfree(device->cache.pkey_cache[p]);
1213 gid_table_release_one(device);
1214 kfree(device->cache.pkey_cache);
1215 kfree(device->cache.lmc_cache);
1218 void ib_cache_cleanup_one(struct ib_device *device)
1220 /* The cleanup function unregisters the event handler,
1221 * waits for all in-progress workqueue elements and cleans
1222 * up the GID cache. This function should be called after
1223 * the device was removed from the devices list and all
1224 * clients were removed, so the cache exists but is
1225 * non-functional and shouldn't be updated anymore.
1227 ib_unregister_event_handler(&device->cache.event_handler);
1228 flush_workqueue(ib_wq);
1229 gid_table_cleanup_one(device);
1232 void __init ib_cache_setup(void)
1234 roce_gid_mgmt_init();
1237 void __exit ib_cache_cleanup(void)
1239 roce_gid_mgmt_cleanup();