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;
182 /* in rdma_cap_roce_gid_table, this funciton should be protected by a
186 if (rdma_cap_roce_gid_table(ib_dev, port)) {
187 table->data_vec[ix].props |= GID_TABLE_ENTRY_INVALID;
188 write_unlock_irq(&table->rwlock);
189 /* GID_TABLE_WRITE_ACTION_MODIFY currently isn't supported by
190 * RoCE providers and thus only updates the cache.
192 if (action == GID_TABLE_WRITE_ACTION_ADD)
193 ret = ib_dev->add_gid(ib_dev, port, ix, gid, attr,
194 &table->data_vec[ix].context);
195 else if (action == GID_TABLE_WRITE_ACTION_DEL)
196 ret = ib_dev->del_gid(ib_dev, port, ix,
197 &table->data_vec[ix].context);
198 write_lock_irq(&table->rwlock);
201 old_net_dev = table->data_vec[ix].attr.ndev;
202 if (old_net_dev && old_net_dev != attr->ndev)
203 dev_put(old_net_dev);
204 /* if modify_gid failed, just delete the old gid */
205 if (ret || action == GID_TABLE_WRITE_ACTION_DEL) {
208 table->data_vec[ix].context = NULL;
211 table->data_vec[ix].props |= GID_TABLE_ENTRY_DEFAULT;
212 memcpy(&table->data_vec[ix].gid, gid, sizeof(*gid));
213 memcpy(&table->data_vec[ix].attr, attr, sizeof(*attr));
214 if (table->data_vec[ix].attr.ndev &&
215 table->data_vec[ix].attr.ndev != old_net_dev)
216 dev_hold(table->data_vec[ix].attr.ndev);
218 table->data_vec[ix].props &= ~GID_TABLE_ENTRY_INVALID;
223 static int add_gid(struct ib_device *ib_dev, u8 port,
224 struct ib_gid_table *table, int ix,
225 const union ib_gid *gid,
226 const struct ib_gid_attr *attr,
228 return write_gid(ib_dev, port, table, ix, gid, attr,
229 GID_TABLE_WRITE_ACTION_ADD, default_gid);
232 static int modify_gid(struct ib_device *ib_dev, u8 port,
233 struct ib_gid_table *table, int ix,
234 const union ib_gid *gid,
235 const struct ib_gid_attr *attr,
237 return write_gid(ib_dev, port, table, ix, gid, attr,
238 GID_TABLE_WRITE_ACTION_MODIFY, default_gid);
241 static int del_gid(struct ib_device *ib_dev, u8 port,
242 struct ib_gid_table *table, int ix,
244 return write_gid(ib_dev, port, table, ix, &zgid, &zattr,
245 GID_TABLE_WRITE_ACTION_DEL, default_gid);
248 /* rwlock should be read locked */
249 static int find_gid(struct ib_gid_table *table, const union ib_gid *gid,
250 const struct ib_gid_attr *val, bool default_gid,
251 unsigned long mask, int *pempty)
255 int empty = pempty ? -1 : 0;
257 while (i < table->sz && (found < 0 || empty < 0)) {
258 struct ib_gid_table_entry *data = &table->data_vec[i];
259 struct ib_gid_attr *attr = &data->attr;
264 if (data->props & GID_TABLE_ENTRY_INVALID)
268 if (!memcmp(&data->gid, &zgid, sizeof(*gid)) &&
269 !memcmp(attr, &zattr, sizeof(*attr)) &&
276 if (mask & GID_ATTR_FIND_MASK_GID_TYPE &&
277 attr->gid_type != val->gid_type)
280 if (mask & GID_ATTR_FIND_MASK_GID &&
281 memcmp(gid, &data->gid, sizeof(*gid)))
284 if (mask & GID_ATTR_FIND_MASK_NETDEV &&
285 attr->ndev != val->ndev)
288 if (mask & GID_ATTR_FIND_MASK_DEFAULT &&
289 !!(data->props & GID_TABLE_ENTRY_DEFAULT) !=
302 static void make_default_gid(struct net_device *dev, union ib_gid *gid)
304 gid->global.subnet_prefix = cpu_to_be64(0xfe80000000000000LL);
305 addrconf_ifid_eui48(&gid->raw[8], dev);
308 int ib_cache_gid_add(struct ib_device *ib_dev, u8 port,
309 union ib_gid *gid, struct ib_gid_attr *attr)
311 struct ib_gid_table **ports_table = ib_dev->cache.gid_cache;
312 struct ib_gid_table *table;
315 struct net_device *idev;
318 table = ports_table[port - rdma_start_port(ib_dev)];
320 if (!memcmp(gid, &zgid, sizeof(*gid)))
323 if (ib_dev->get_netdev) {
324 idev = ib_dev->get_netdev(ib_dev, port);
325 if (idev && attr->ndev != idev) {
326 union ib_gid default_gid;
328 /* Adding default GIDs in not permitted */
329 make_default_gid(idev, &default_gid);
330 if (!memcmp(gid, &default_gid, sizeof(*gid))) {
339 mutex_lock(&table->lock);
340 write_lock_irq(&table->rwlock);
342 ix = find_gid(table, gid, attr, false, GID_ATTR_FIND_MASK_GID |
343 GID_ATTR_FIND_MASK_GID_TYPE |
344 GID_ATTR_FIND_MASK_NETDEV, &empty);
353 ret = add_gid(ib_dev, port, table, empty, gid, attr, false);
355 dispatch_gid_change_event(ib_dev, port);
358 write_unlock_irq(&table->rwlock);
359 mutex_unlock(&table->lock);
363 int ib_cache_gid_del(struct ib_device *ib_dev, u8 port,
364 union ib_gid *gid, struct ib_gid_attr *attr)
366 struct ib_gid_table **ports_table = ib_dev->cache.gid_cache;
367 struct ib_gid_table *table;
370 table = ports_table[port - rdma_start_port(ib_dev)];
372 mutex_lock(&table->lock);
373 write_lock_irq(&table->rwlock);
375 ix = find_gid(table, gid, attr, false,
376 GID_ATTR_FIND_MASK_GID |
377 GID_ATTR_FIND_MASK_GID_TYPE |
378 GID_ATTR_FIND_MASK_NETDEV |
379 GID_ATTR_FIND_MASK_DEFAULT,
384 if (!del_gid(ib_dev, port, table, ix, false))
385 dispatch_gid_change_event(ib_dev, port);
388 write_unlock_irq(&table->rwlock);
389 mutex_unlock(&table->lock);
393 int ib_cache_gid_del_all_netdev_gids(struct ib_device *ib_dev, u8 port,
394 struct net_device *ndev)
396 struct ib_gid_table **ports_table = ib_dev->cache.gid_cache;
397 struct ib_gid_table *table;
399 bool deleted = false;
401 table = ports_table[port - rdma_start_port(ib_dev)];
403 mutex_lock(&table->lock);
404 write_lock_irq(&table->rwlock);
406 for (ix = 0; ix < table->sz; ix++)
407 if (table->data_vec[ix].attr.ndev == ndev)
408 if (!del_gid(ib_dev, port, table, ix, false))
411 write_unlock_irq(&table->rwlock);
412 mutex_unlock(&table->lock);
415 dispatch_gid_change_event(ib_dev, port);
420 static int __ib_cache_gid_get(struct ib_device *ib_dev, u8 port, int index,
421 union ib_gid *gid, struct ib_gid_attr *attr)
423 struct ib_gid_table **ports_table = ib_dev->cache.gid_cache;
424 struct ib_gid_table *table;
426 table = ports_table[port - rdma_start_port(ib_dev)];
428 if (index < 0 || index >= table->sz)
431 if (table->data_vec[index].props & GID_TABLE_ENTRY_INVALID)
434 memcpy(gid, &table->data_vec[index].gid, sizeof(*gid));
436 memcpy(attr, &table->data_vec[index].attr, sizeof(*attr));
438 dev_hold(attr->ndev);
444 static int _ib_cache_gid_table_find(struct ib_device *ib_dev,
445 const union ib_gid *gid,
446 const struct ib_gid_attr *val,
448 u8 *port, u16 *index)
450 struct ib_gid_table **ports_table = ib_dev->cache.gid_cache;
451 struct ib_gid_table *table;
456 for (p = 0; p < ib_dev->phys_port_cnt; p++) {
457 table = ports_table[p];
458 read_lock_irqsave(&table->rwlock, flags);
459 local_index = find_gid(table, gid, val, false, mask, NULL);
460 if (local_index >= 0) {
462 *index = local_index;
464 *port = p + rdma_start_port(ib_dev);
465 read_unlock_irqrestore(&table->rwlock, flags);
468 read_unlock_irqrestore(&table->rwlock, flags);
474 static int ib_cache_gid_find(struct ib_device *ib_dev,
475 const union ib_gid *gid,
476 enum ib_gid_type gid_type,
477 struct net_device *ndev, u8 *port,
480 unsigned long mask = GID_ATTR_FIND_MASK_GID |
481 GID_ATTR_FIND_MASK_GID_TYPE;
482 struct ib_gid_attr gid_attr_val = {.ndev = ndev, .gid_type = gid_type};
485 mask |= GID_ATTR_FIND_MASK_NETDEV;
487 return _ib_cache_gid_table_find(ib_dev, gid, &gid_attr_val,
491 int ib_find_cached_gid_by_port(struct ib_device *ib_dev,
492 const union ib_gid *gid,
493 enum ib_gid_type gid_type,
494 u8 port, struct net_device *ndev,
498 struct ib_gid_table **ports_table = ib_dev->cache.gid_cache;
499 struct ib_gid_table *table;
500 unsigned long mask = GID_ATTR_FIND_MASK_GID |
501 GID_ATTR_FIND_MASK_GID_TYPE;
502 struct ib_gid_attr val = {.ndev = ndev, .gid_type = gid_type};
505 if (port < rdma_start_port(ib_dev) ||
506 port > rdma_end_port(ib_dev))
509 table = ports_table[port - rdma_start_port(ib_dev)];
512 mask |= GID_ATTR_FIND_MASK_NETDEV;
514 read_lock_irqsave(&table->rwlock, flags);
515 local_index = find_gid(table, gid, &val, false, mask, NULL);
516 if (local_index >= 0) {
518 *index = local_index;
519 read_unlock_irqrestore(&table->rwlock, flags);
523 read_unlock_irqrestore(&table->rwlock, flags);
526 EXPORT_SYMBOL(ib_find_cached_gid_by_port);
529 * ib_find_gid_by_filter - Returns the GID table index where a specified
531 * @device: The device to query.
532 * @gid: The GID value to search for.
533 * @port_num: The port number of the device where the GID value could be
535 * @filter: The filter function is executed on any matching GID in the table.
536 * If the filter function returns true, the corresponding index is returned,
537 * otherwise, we continue searching the GID table. It's guaranteed that
538 * while filter is executed, ndev field is valid and the structure won't
539 * change. filter is executed in an atomic context. filter must not be NULL.
540 * @index: The index into the cached GID table where the GID was found. This
541 * parameter may be NULL.
543 * ib_cache_gid_find_by_filter() searches for the specified GID value
544 * of which the filter function returns true in the port's GID table.
545 * This function is only supported on RoCE ports.
548 static int ib_cache_gid_find_by_filter(struct ib_device *ib_dev,
549 const union ib_gid *gid,
551 bool (*filter)(const union ib_gid *,
552 const struct ib_gid_attr *,
557 struct ib_gid_table **ports_table = ib_dev->cache.gid_cache;
558 struct ib_gid_table *table;
566 if (port < rdma_start_port(ib_dev) ||
567 port > rdma_end_port(ib_dev) ||
568 !rdma_protocol_roce(ib_dev, port))
569 return -EPROTONOSUPPORT;
571 table = ports_table[port - rdma_start_port(ib_dev)];
573 read_lock_irqsave(&table->rwlock, flags);
574 for (i = 0; i < table->sz; i++) {
575 struct ib_gid_attr attr;
577 if (table->data_vec[i].props & GID_TABLE_ENTRY_INVALID)
580 if (memcmp(gid, &table->data_vec[i].gid, sizeof(*gid)))
583 memcpy(&attr, &table->data_vec[i].attr, sizeof(attr));
585 if (filter(gid, &attr, context))
592 read_unlock_irqrestore(&table->rwlock, flags);
602 static struct ib_gid_table *alloc_gid_table(int sz)
604 struct ib_gid_table *table =
605 kzalloc(sizeof(struct ib_gid_table), GFP_KERNEL);
610 table->data_vec = kcalloc(sz, sizeof(*table->data_vec), GFP_KERNEL);
611 if (!table->data_vec)
614 mutex_init(&table->lock);
617 rwlock_init(&table->rwlock);
626 static void release_gid_table(struct ib_gid_table *table)
629 kfree(table->data_vec);
634 static void cleanup_gid_table_port(struct ib_device *ib_dev, u8 port,
635 struct ib_gid_table *table)
638 bool deleted = false;
643 write_lock_irq(&table->rwlock);
644 for (i = 0; i < table->sz; ++i) {
645 if (memcmp(&table->data_vec[i].gid, &zgid,
646 sizeof(table->data_vec[i].gid)))
647 if (!del_gid(ib_dev, port, table, i,
648 table->data_vec[i].props &
649 GID_ATTR_FIND_MASK_DEFAULT))
652 write_unlock_irq(&table->rwlock);
655 dispatch_gid_change_event(ib_dev, port);
658 void ib_cache_gid_set_default_gid(struct ib_device *ib_dev, u8 port,
659 struct net_device *ndev,
660 unsigned long gid_type_mask,
661 enum ib_cache_gid_default_mode mode)
663 struct ib_gid_table **ports_table = ib_dev->cache.gid_cache;
665 struct ib_gid_attr gid_attr;
666 struct ib_gid_attr zattr_type = zattr;
667 struct ib_gid_table *table;
668 unsigned int gid_type;
670 table = ports_table[port - rdma_start_port(ib_dev)];
672 make_default_gid(ndev, &gid);
673 memset(&gid_attr, 0, sizeof(gid_attr));
674 gid_attr.ndev = ndev;
676 for (gid_type = 0; gid_type < IB_GID_TYPE_SIZE; ++gid_type) {
678 union ib_gid current_gid;
679 struct ib_gid_attr current_gid_attr = {};
681 if (1UL << gid_type & ~gid_type_mask)
684 gid_attr.gid_type = gid_type;
686 mutex_lock(&table->lock);
687 write_lock_irq(&table->rwlock);
688 ix = find_gid(table, NULL, &gid_attr, true,
689 GID_ATTR_FIND_MASK_GID_TYPE |
690 GID_ATTR_FIND_MASK_DEFAULT,
693 /* Coudn't find default GID location */
696 zattr_type.gid_type = gid_type;
698 if (!__ib_cache_gid_get(ib_dev, port, ix,
699 ¤t_gid, ¤t_gid_attr) &&
700 mode == IB_CACHE_GID_DEFAULT_MODE_SET &&
701 !memcmp(&gid, ¤t_gid, sizeof(gid)) &&
702 !memcmp(&gid_attr, ¤t_gid_attr, sizeof(gid_attr)))
705 if (memcmp(¤t_gid, &zgid, sizeof(current_gid)) ||
706 memcmp(¤t_gid_attr, &zattr_type,
707 sizeof(current_gid_attr))) {
708 if (del_gid(ib_dev, port, table, ix, true)) {
709 pr_warn("ib_cache_gid: can't delete index %d for default gid %pI6\n",
713 dispatch_gid_change_event(ib_dev, port);
717 if (mode == IB_CACHE_GID_DEFAULT_MODE_SET) {
718 if (add_gid(ib_dev, port, table, ix, &gid, &gid_attr, true))
719 pr_warn("ib_cache_gid: unable to add default gid %pI6\n",
722 dispatch_gid_change_event(ib_dev, port);
726 if (current_gid_attr.ndev)
727 dev_put(current_gid_attr.ndev);
728 write_unlock_irq(&table->rwlock);
729 mutex_unlock(&table->lock);
733 static int gid_table_reserve_default(struct ib_device *ib_dev, u8 port,
734 struct ib_gid_table *table)
737 unsigned long roce_gid_type_mask;
738 unsigned int num_default_gids;
739 unsigned int current_gid = 0;
741 roce_gid_type_mask = roce_gid_type_mask_support(ib_dev, port);
742 num_default_gids = hweight_long(roce_gid_type_mask);
743 for (i = 0; i < num_default_gids && i < table->sz; i++) {
744 struct ib_gid_table_entry *entry =
747 entry->props |= GID_TABLE_ENTRY_DEFAULT;
748 current_gid = find_next_bit(&roce_gid_type_mask,
751 entry->attr.gid_type = current_gid++;
757 static int _gid_table_setup_one(struct ib_device *ib_dev)
760 struct ib_gid_table **table;
763 table = kcalloc(ib_dev->phys_port_cnt, sizeof(*table), GFP_KERNEL);
766 pr_warn("failed to allocate ib gid cache for %s\n",
771 for (port = 0; port < ib_dev->phys_port_cnt; port++) {
772 u8 rdma_port = port + rdma_start_port(ib_dev);
776 ib_dev->port_immutable[rdma_port].gid_tbl_len);
779 goto rollback_table_setup;
782 err = gid_table_reserve_default(ib_dev,
783 port + rdma_start_port(ib_dev),
786 goto rollback_table_setup;
789 ib_dev->cache.gid_cache = table;
792 rollback_table_setup:
793 for (port = 0; port < ib_dev->phys_port_cnt; port++) {
794 cleanup_gid_table_port(ib_dev, port + rdma_start_port(ib_dev),
796 release_gid_table(table[port]);
803 static void gid_table_release_one(struct ib_device *ib_dev)
805 struct ib_gid_table **table = ib_dev->cache.gid_cache;
811 for (port = 0; port < ib_dev->phys_port_cnt; port++)
812 release_gid_table(table[port]);
815 ib_dev->cache.gid_cache = NULL;
818 static void gid_table_cleanup_one(struct ib_device *ib_dev)
820 struct ib_gid_table **table = ib_dev->cache.gid_cache;
826 for (port = 0; port < ib_dev->phys_port_cnt; port++)
827 cleanup_gid_table_port(ib_dev, port + rdma_start_port(ib_dev),
831 static int gid_table_setup_one(struct ib_device *ib_dev)
835 err = _gid_table_setup_one(ib_dev);
840 err = roce_rescan_device(ib_dev);
843 gid_table_cleanup_one(ib_dev);
844 gid_table_release_one(ib_dev);
850 int ib_get_cached_gid(struct ib_device *device,
854 struct ib_gid_attr *gid_attr)
858 struct ib_gid_table **ports_table = device->cache.gid_cache;
859 struct ib_gid_table *table = ports_table[port_num - rdma_start_port(device)];
861 if (port_num < rdma_start_port(device) || port_num > rdma_end_port(device))
864 read_lock_irqsave(&table->rwlock, flags);
865 res = __ib_cache_gid_get(device, port_num, index, gid, gid_attr);
866 read_unlock_irqrestore(&table->rwlock, flags);
870 EXPORT_SYMBOL(ib_get_cached_gid);
872 int ib_find_cached_gid(struct ib_device *device,
873 const union ib_gid *gid,
874 enum ib_gid_type gid_type,
875 struct net_device *ndev,
879 return ib_cache_gid_find(device, gid, gid_type, ndev, port_num, index);
881 EXPORT_SYMBOL(ib_find_cached_gid);
883 int ib_find_gid_by_filter(struct ib_device *device,
884 const union ib_gid *gid,
886 bool (*filter)(const union ib_gid *gid,
887 const struct ib_gid_attr *,
889 void *context, u16 *index)
891 /* Only RoCE GID table supports filter function */
892 if (!rdma_cap_roce_gid_table(device, port_num) && filter)
893 return -EPROTONOSUPPORT;
895 return ib_cache_gid_find_by_filter(device, gid,
899 EXPORT_SYMBOL(ib_find_gid_by_filter);
901 int ib_get_cached_pkey(struct ib_device *device,
906 struct ib_pkey_cache *cache;
910 if (port_num < rdma_start_port(device) || port_num > rdma_end_port(device))
913 read_lock_irqsave(&device->cache.lock, flags);
915 cache = device->cache.pkey_cache[port_num - rdma_start_port(device)];
917 if (index < 0 || index >= cache->table_len)
920 *pkey = cache->table[index];
922 read_unlock_irqrestore(&device->cache.lock, flags);
926 EXPORT_SYMBOL(ib_get_cached_pkey);
928 int ib_find_cached_pkey(struct ib_device *device,
933 struct ib_pkey_cache *cache;
939 if (port_num < rdma_start_port(device) || port_num > rdma_end_port(device))
942 read_lock_irqsave(&device->cache.lock, flags);
944 cache = device->cache.pkey_cache[port_num - rdma_start_port(device)];
948 for (i = 0; i < cache->table_len; ++i)
949 if ((cache->table[i] & 0x7fff) == (pkey & 0x7fff)) {
950 if (cache->table[i] & 0x8000) {
958 if (ret && partial_ix >= 0) {
963 read_unlock_irqrestore(&device->cache.lock, flags);
967 EXPORT_SYMBOL(ib_find_cached_pkey);
969 int ib_find_exact_cached_pkey(struct ib_device *device,
974 struct ib_pkey_cache *cache;
979 if (port_num < rdma_start_port(device) || port_num > rdma_end_port(device))
982 read_lock_irqsave(&device->cache.lock, flags);
984 cache = device->cache.pkey_cache[port_num - rdma_start_port(device)];
988 for (i = 0; i < cache->table_len; ++i)
989 if (cache->table[i] == pkey) {
995 read_unlock_irqrestore(&device->cache.lock, flags);
999 EXPORT_SYMBOL(ib_find_exact_cached_pkey);
1001 int ib_get_cached_lmc(struct ib_device *device,
1005 unsigned long flags;
1008 if (port_num < rdma_start_port(device) || port_num > rdma_end_port(device))
1011 read_lock_irqsave(&device->cache.lock, flags);
1012 *lmc = device->cache.lmc_cache[port_num - rdma_start_port(device)];
1013 read_unlock_irqrestore(&device->cache.lock, flags);
1017 EXPORT_SYMBOL(ib_get_cached_lmc);
1019 static void ib_cache_update(struct ib_device *device,
1022 struct ib_port_attr *tprops = NULL;
1023 struct ib_pkey_cache *pkey_cache = NULL, *old_pkey_cache;
1024 struct ib_gid_cache {
1026 union ib_gid table[0];
1027 } *gid_cache = NULL;
1030 struct ib_gid_table *table;
1031 struct ib_gid_table **ports_table = device->cache.gid_cache;
1032 bool use_roce_gid_table =
1033 rdma_cap_roce_gid_table(device, port);
1035 if (port < rdma_start_port(device) || port > rdma_end_port(device))
1038 table = ports_table[port - rdma_start_port(device)];
1040 tprops = kmalloc(sizeof *tprops, GFP_KERNEL);
1044 ret = ib_query_port(device, port, tprops);
1046 pr_warn("ib_query_port failed (%d) for %s\n",
1051 pkey_cache = kmalloc(sizeof *pkey_cache + tprops->pkey_tbl_len *
1052 sizeof *pkey_cache->table, GFP_KERNEL);
1056 pkey_cache->table_len = tprops->pkey_tbl_len;
1058 if (!use_roce_gid_table) {
1059 gid_cache = kmalloc(sizeof(*gid_cache) + tprops->gid_tbl_len *
1060 sizeof(*gid_cache->table), GFP_KERNEL);
1064 gid_cache->table_len = tprops->gid_tbl_len;
1067 for (i = 0; i < pkey_cache->table_len; ++i) {
1068 ret = ib_query_pkey(device, port, i, pkey_cache->table + i);
1070 pr_warn("ib_query_pkey failed (%d) for %s (index %d)\n",
1071 ret, device->name, i);
1076 if (!use_roce_gid_table) {
1077 for (i = 0; i < gid_cache->table_len; ++i) {
1078 ret = ib_query_gid(device, port, i,
1079 gid_cache->table + i, NULL);
1081 pr_warn("ib_query_gid failed (%d) for %s (index %d)\n",
1082 ret, device->name, i);
1088 write_lock_irq(&device->cache.lock);
1090 old_pkey_cache = device->cache.pkey_cache[port - rdma_start_port(device)];
1092 device->cache.pkey_cache[port - rdma_start_port(device)] = pkey_cache;
1093 if (!use_roce_gid_table) {
1094 write_lock(&table->rwlock);
1095 for (i = 0; i < gid_cache->table_len; i++) {
1096 modify_gid(device, port, table, i, gid_cache->table + i,
1099 write_unlock(&table->rwlock);
1102 device->cache.lmc_cache[port - rdma_start_port(device)] = tprops->lmc;
1104 write_unlock_irq(&device->cache.lock);
1107 kfree(old_pkey_cache);
1117 static void ib_cache_task(struct work_struct *_work)
1119 struct ib_update_work *work =
1120 container_of(_work, struct ib_update_work, work);
1122 ib_cache_update(work->device, work->port_num);
1126 static void ib_cache_event(struct ib_event_handler *handler,
1127 struct ib_event *event)
1129 struct ib_update_work *work;
1131 if (event->event == IB_EVENT_PORT_ERR ||
1132 event->event == IB_EVENT_PORT_ACTIVE ||
1133 event->event == IB_EVENT_LID_CHANGE ||
1134 event->event == IB_EVENT_PKEY_CHANGE ||
1135 event->event == IB_EVENT_SM_CHANGE ||
1136 event->event == IB_EVENT_CLIENT_REREGISTER ||
1137 event->event == IB_EVENT_GID_CHANGE) {
1138 work = kmalloc(sizeof *work, GFP_ATOMIC);
1140 INIT_WORK(&work->work, ib_cache_task);
1141 work->device = event->device;
1142 work->port_num = event->element.port_num;
1143 queue_work(ib_wq, &work->work);
1148 int ib_cache_setup_one(struct ib_device *device)
1153 rwlock_init(&device->cache.lock);
1155 device->cache.pkey_cache =
1156 kzalloc(sizeof *device->cache.pkey_cache *
1157 (rdma_end_port(device) - rdma_start_port(device) + 1), GFP_KERNEL);
1158 device->cache.lmc_cache = kmalloc(sizeof *device->cache.lmc_cache *
1159 (rdma_end_port(device) -
1160 rdma_start_port(device) + 1),
1162 if (!device->cache.pkey_cache ||
1163 !device->cache.lmc_cache) {
1164 pr_warn("Couldn't allocate cache for %s\n", device->name);
1168 err = gid_table_setup_one(device);
1170 /* Allocated memory will be cleaned in the release function */
1173 for (p = 0; p <= rdma_end_port(device) - rdma_start_port(device); ++p)
1174 ib_cache_update(device, p + rdma_start_port(device));
1176 INIT_IB_EVENT_HANDLER(&device->cache.event_handler,
1177 device, ib_cache_event);
1178 err = ib_register_event_handler(&device->cache.event_handler);
1185 gid_table_cleanup_one(device);
1189 void ib_cache_release_one(struct ib_device *device)
1194 * The release function frees all the cache elements.
1195 * This function should be called as part of freeing
1196 * all the device's resources when the cache could no
1197 * longer be accessed.
1199 if (device->cache.pkey_cache)
1201 p <= rdma_end_port(device) - rdma_start_port(device); ++p)
1202 kfree(device->cache.pkey_cache[p]);
1204 gid_table_release_one(device);
1205 kfree(device->cache.pkey_cache);
1206 kfree(device->cache.lmc_cache);
1209 void ib_cache_cleanup_one(struct ib_device *device)
1211 /* The cleanup function unregisters the event handler,
1212 * waits for all in-progress workqueue elements and cleans
1213 * up the GID cache. This function should be called after
1214 * the device was removed from the devices list and all
1215 * clients were removed, so the cache exists but is
1216 * non-functional and shouldn't be updated anymore.
1218 ib_unregister_event_handler(&device->cache.event_handler);
1219 flush_workqueue(ib_wq);
1220 gid_table_cleanup_one(device);
1223 void __init ib_cache_setup(void)
1225 roce_gid_mgmt_init();
1228 void __exit ib_cache_cleanup(void)
1230 roce_gid_mgmt_cleanup();