projects/linux·blob master · refs

1657 lines · 42.5 KiB

1/*
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.
6 *
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:
12 *
13 * Redistribution and use in source and binary forms, with or
14 * without modification, are permitted provided that the following
15 * conditions are met:
16 *
17 * - Redistributions of source code must retain the above
18 * copyright notice, this list of conditions and the following
19 * disclaimer.
20 *
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.
25 *
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
33 * SOFTWARE.
34 */
35
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>
42
43#include <rdma/ib_cache.h>
44
45#include "core_priv.h"
46
47struct ib_pkey_cache {
48 int table_len;
49 u16 table[] __counted_by(table_len);
50};
51
52struct ib_update_work {
53 struct work_struct work;
54 struct ib_event event;
55 bool enforce_security;
56};
57
58union ib_gid zgid;
59EXPORT_SYMBOL(zgid);
60
61enum 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,
66};
67
68enum gid_table_entry_state {
69 GID_TABLE_ENTRY_INVALID = 1,
70 GID_TABLE_ENTRY_VALID = 2,
71 /*
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.
76 */
77 GID_TABLE_ENTRY_PENDING_DEL = 3,
78};
79
80struct roce_gid_ndev_storage {
81 struct rcu_head rcu_head;
82 struct net_device *ndev;
83};
84
85struct ib_gid_table_entry {
86 struct kref kref;
87 struct work_struct del_work;
88 struct ib_gid_attr attr;
89 void *context;
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.
94 */
95 struct roce_gid_ndev_storage *ndev_storage;
96 enum gid_table_entry_state state;
97};
98
99struct ib_gid_table {
100 int sz;
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
105 *
106 * Delete requires different set of operations:
107 * (a) Find the GID
108 * (b) Delete it.
109 *
110 **/
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
113 * sleepable context.
114 */
115 struct mutex lock;
116 /* rwlock protects data_vec[ix]->state and entry pointer.
117 */
118 rwlock_t rwlock;
119 /* bit field, each bit indicates the index of default GID */
120 u32 default_gid_indices;
121 struct ib_gid_table_entry *data_vec[] __counted_by(sz);
122};
123
124static void dispatch_gid_change_event(struct ib_device *ib_dev, u32 port)
125{
126 struct ib_event event;
127
128 event.device = ib_dev;
129 event.element.port_num = port;
130 event.event = IB_EVENT_GID_CHANGE;
131
132 ib_dispatch_event_clients(&event);
133}
134
135static 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.
138 */
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",
142};
143
144const char *ib_cache_gid_type_str(enum ib_gid_type gid_type)
145{
146 if (gid_type < ARRAY_SIZE(gid_type_str) && gid_type_str[gid_type])
147 return gid_type_str[gid_type];
148
149 return "Invalid GID type";
150}
151EXPORT_SYMBOL(ib_cache_gid_type_str);
152
153/** rdma_is_zero_gid - Check if given GID is zero or not.
154 * @gid: GID to check
155 * Returns true if given GID is zero, returns false otherwise.
156 */
157bool rdma_is_zero_gid(const union ib_gid *gid)
158{
159 return !memcmp(gid, &zgid, sizeof(*gid));
160}
161EXPORT_SYMBOL(rdma_is_zero_gid);
162
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
168 * returns false.
169 */
170static bool is_gid_index_default(const struct ib_gid_table *table,
171 unsigned int index)
172{
173 return index < 32 && (BIT(index) & table->default_gid_indices);
174}
175
176int ib_cache_gid_parse_type_str(const char *buf)
177{
178 unsigned int i;
179 size_t len;
180 int err = -EINVAL;
181
182 len = strlen(buf);
183 if (len == 0)
184 return -EINVAL;
185
186 if (buf[len - 1] == '\n')
187 len--;
188
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])) {
192 err = i;
193 break;
194 }
195
196 return err;
197}
198EXPORT_SYMBOL(ib_cache_gid_parse_type_str);
199
200static struct ib_gid_table *rdma_gid_table(struct ib_device *device, u32 port)
201{
202 return device->port_data[port].cache.gid;
203}
204
205static bool is_gid_entry_free(const struct ib_gid_table_entry *entry)
206{
207 return !entry;
208}
209
210static bool is_gid_entry_valid(const struct ib_gid_table_entry *entry)
211{
212 return entry && entry->state == GID_TABLE_ENTRY_VALID;
213}
214
215static void schedule_free_gid(struct kref *kref)
216{
217 struct ib_gid_table_entry *entry =
218 container_of(kref, struct ib_gid_table_entry, kref);
219
220 queue_work(ib_wq, &entry->del_work);
221}
222
223static void put_gid_ndev(struct rcu_head *head)
224{
225 struct roce_gid_ndev_storage *storage =
226 container_of(head, struct roce_gid_ndev_storage, rcu_head);
227
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
231 * using this netdev.
232 */
233 dev_put(storage->ndev);
234 kfree(storage);
235}
236
237static void free_gid_entry_locked(struct ib_gid_table_entry *entry)
238{
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);
242
243 dev_dbg(&device->dev, "%s port=%u index=%u gid %pI6\n", __func__,
244 port_num, entry->attr.index, entry->attr.gid.raw);
245
246 write_lock_irq(&table->rwlock);
247
248 /*
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.
253 */
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);
258
259 if (entry->ndev_storage)
260 call_rcu(&entry->ndev_storage->rcu_head, put_gid_ndev);
261 kfree(entry);
262}
263
264static void free_gid_entry(struct kref *kref)
265{
266 struct ib_gid_table_entry *entry =
267 container_of(kref, struct ib_gid_table_entry, kref);
268
269 free_gid_entry_locked(entry);
270}
271
272/**
273 * free_gid_work - Release reference to the GID entry
274 * @work: Work structure to refer to GID entry which needs to be
275 * deleted.
276 *
277 * free_gid_work() frees the entry from the HCA's hardware table
278 * if provider supports it. It releases reference to netdevice.
279 */
280static void free_gid_work(struct work_struct *work)
281{
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);
287
288 mutex_lock(&table->lock);
289 free_gid_entry_locked(entry);
290 mutex_unlock(&table->lock);
291}
292
293static struct ib_gid_table_entry *
294alloc_gid_entry(const struct ib_gid_attr *attr)
295{
296 struct ib_gid_table_entry *entry;
297 struct net_device *ndev;
298
299 entry = kzalloc_obj(*entry);
300 if (!entry)
301 return NULL;
302
303 ndev = rcu_dereference_protected(attr->ndev, 1);
304 if (ndev) {
305 entry->ndev_storage = kzalloc_obj(*entry->ndev_storage);
306 if (!entry->ndev_storage) {
307 kfree(entry);
308 return NULL;
309 }
310 dev_hold(ndev);
311 entry->ndev_storage->ndev = ndev;
312 }
313 kref_init(&entry->kref);
314 memcpy(&entry->attr, attr, sizeof(*attr));
315 INIT_WORK(&entry->del_work, free_gid_work);
316 entry->state = GID_TABLE_ENTRY_INVALID;
317 return entry;
318}
319
320static void store_gid_entry(struct ib_gid_table *table,
321 struct ib_gid_table_entry *entry)
322{
323 entry->state = GID_TABLE_ENTRY_VALID;
324
325 dev_dbg(&entry->attr.device->dev, "%s port=%u index=%u gid %pI6\n",
326 __func__, entry->attr.port_num, entry->attr.index,
327 entry->attr.gid.raw);
328
329 lockdep_assert_held(&table->lock);
330 write_lock_irq(&table->rwlock);
331 table->data_vec[entry->attr.index] = entry;
332 write_unlock_irq(&table->rwlock);
333}
334
335static void get_gid_entry(struct ib_gid_table_entry *entry)
336{
337 kref_get(&entry->kref);
338}
339
340static void put_gid_entry(struct ib_gid_table_entry *entry)
341{
342 kref_put(&entry->kref, schedule_free_gid);
343}
344
345static void put_gid_entry_locked(struct ib_gid_table_entry *entry)
346{
347 kref_put(&entry->kref, free_gid_entry);
348}
349
350static int add_roce_gid(struct ib_gid_table_entry *entry)
351{
352 const struct ib_gid_attr *attr = &entry->attr;
353 int ret;
354
355 if (!attr->ndev) {
356 dev_err(&attr->device->dev, "%s NULL netdev port=%u index=%u\n",
357 __func__, attr->port_num, attr->index);
358 return -EINVAL;
359 }
360 if (rdma_cap_roce_gid_table(attr->device, attr->port_num)) {
361 ret = attr->device->ops.add_gid(attr, &entry->context);
362 if (ret) {
363 dev_err(&attr->device->dev,
364 "%s GID add failed port=%u index=%u\n",
365 __func__, attr->port_num, attr->index);
366 return ret;
367 }
368 }
369 return 0;
370}
371
372/**
373 * del_gid - Delete GID table entry
374 *
375 * @ib_dev: IB device whose GID entry to be deleted
376 * @port: Port number of the IB device
377 * @table: GID table of the IB device for a port
378 * @ix: GID entry index to delete
379 *
380 */
381static void del_gid(struct ib_device *ib_dev, u32 port,
382 struct ib_gid_table *table, int ix)
383{
384 struct roce_gid_ndev_storage *ndev_storage;
385 struct ib_gid_table_entry *entry;
386
387 lockdep_assert_held(&table->lock);
388
389 dev_dbg(&ib_dev->dev, "%s port=%u index=%d gid %pI6\n", __func__, port,
390 ix, table->data_vec[ix]->attr.gid.raw);
391
392 write_lock_irq(&table->rwlock);
393 entry = table->data_vec[ix];
394 entry->state = GID_TABLE_ENTRY_PENDING_DEL;
395 /*
396 * For non RoCE protocol, GID entry slot is ready to use.
397 */
398 if (!rdma_protocol_roce(ib_dev, port))
399 table->data_vec[ix] = NULL;
400 write_unlock_irq(&table->rwlock);
401
402 if (rdma_cap_roce_gid_table(ib_dev, port))
403 ib_dev->ops.del_gid(&entry->attr, &entry->context);
404
405 ndev_storage = entry->ndev_storage;
406 if (ndev_storage) {
407 entry->ndev_storage = NULL;
408 rcu_assign_pointer(entry->attr.ndev, NULL);
409 call_rcu(&ndev_storage->rcu_head, put_gid_ndev);
410 }
411
412 put_gid_entry_locked(entry);
413}
414
415/**
416 * add_modify_gid - Add or modify GID table entry
417 *
418 * @table: GID table in which GID to be added or modified
419 * @attr: Attributes of the GID
420 *
421 * Returns 0 on success or appropriate error code. It accepts zero
422 * GID addition for non RoCE ports for HCA's who report them as valid
423 * GID. However such zero GIDs are not added to the cache.
424 */
425static int add_modify_gid(struct ib_gid_table *table,
426 const struct ib_gid_attr *attr)
427{
428 struct ib_gid_table_entry *entry;
429 int ret = 0;
430
431 /*
432 * Invalidate any old entry in the table to make it safe to write to
433 * this index.
434 */
435 if (is_gid_entry_valid(table->data_vec[attr->index]))
436 del_gid(attr->device, attr->port_num, table, attr->index);
437
438 /*
439 * Some HCA's report multiple GID entries with only one valid GID, and
440 * leave other unused entries as the zero GID. Convert zero GIDs to
441 * empty table entries instead of storing them.
442 */
443 if (rdma_is_zero_gid(&attr->gid))
444 return 0;
445
446 entry = alloc_gid_entry(attr);
447 if (!entry)
448 return -ENOMEM;
449
450 if (rdma_protocol_roce(attr->device, attr->port_num)) {
451 ret = add_roce_gid(entry);
452 if (ret)
453 goto done;
454 }
455
456 store_gid_entry(table, entry);
457 return 0;
458
459done:
460 put_gid_entry(entry);
461 return ret;
462}
463
464/* rwlock should be read locked, or lock should be held */
465static int find_gid(struct ib_gid_table *table, const union ib_gid *gid,
466 const struct ib_gid_attr *val, bool default_gid,
467 unsigned long mask, int *pempty)
468{
469 int i = 0;
470 int found = -1;
471 int empty = pempty ? -1 : 0;
472
473 while (i < table->sz && (found < 0 || empty < 0)) {
474 struct ib_gid_table_entry *data = table->data_vec[i];
475 struct ib_gid_attr *attr;
476 int curr_index = i;
477
478 i++;
479
480 /* find_gid() is used during GID addition where it is expected
481 * to return a free entry slot which is not duplicate.
482 * Free entry slot is requested and returned if pempty is set,
483 * so lookup free slot only if requested.
484 */
485 if (pempty && empty < 0) {
486 if (is_gid_entry_free(data) &&
487 default_gid ==
488 is_gid_index_default(table, curr_index)) {
489 /*
490 * Found an invalid (free) entry; allocate it.
491 * If default GID is requested, then our
492 * found slot must be one of the DEFAULT
493 * reserved slots or we fail.
494 * This ensures that only DEFAULT reserved
495 * slots are used for default property GIDs.
496 */
497 empty = curr_index;
498 }
499 }
500
501 /*
502 * Additionally find_gid() is used to find valid entry during
503 * lookup operation; so ignore the entries which are marked as
504 * pending for removal and the entries which are marked as
505 * invalid.
506 */
507 if (!is_gid_entry_valid(data))
508 continue;
509
510 if (found >= 0)
511 continue;
512
513 attr = &data->attr;
514 if (mask & GID_ATTR_FIND_MASK_GID_TYPE &&
515 attr->gid_type != val->gid_type)
516 continue;
517
518 if (mask & GID_ATTR_FIND_MASK_GID &&
519 memcmp(gid, &data->attr.gid, sizeof(*gid)))
520 continue;
521
522 if (mask & GID_ATTR_FIND_MASK_NETDEV &&
523 attr->ndev != val->ndev)
524 continue;
525
526 if (mask & GID_ATTR_FIND_MASK_DEFAULT &&
527 is_gid_index_default(table, curr_index) != default_gid)
528 continue;
529
530 found = curr_index;
531 }
532
533 if (pempty)
534 *pempty = empty;
535
536 return found;
537}
538
539static void make_default_gid(struct net_device *dev, union ib_gid *gid)
540{
541 gid->global.subnet_prefix = cpu_to_be64(0xfe80000000000000LL);
542 addrconf_ifid_eui48(&gid->raw[8], dev);
543}
544
545static int __ib_cache_gid_add(struct ib_device *ib_dev, u32 port,
546 union ib_gid *gid, struct ib_gid_attr *attr,
547 unsigned long mask, bool default_gid)
548{
549 struct ib_gid_table *table;
550 int ret = 0;
551 int empty;
552 int ix;
553
554 /* Do not allow adding zero GID in support of
555 * IB spec version 1.3 section 4.1.1 point (6) and
556 * section 12.7.10 and section 12.7.20
557 */
558 if (rdma_is_zero_gid(gid))
559 return -EINVAL;
560
561 table = rdma_gid_table(ib_dev, port);
562
563 mutex_lock(&table->lock);
564
565 ix = find_gid(table, gid, attr, default_gid, mask, &empty);
566 if (ix >= 0)
567 goto out_unlock;
568
569 if (empty < 0) {
570 ret = -ENOSPC;
571 goto out_unlock;
572 }
573 attr->device = ib_dev;
574 attr->index = empty;
575 attr->port_num = port;
576 attr->gid = *gid;
577 ret = add_modify_gid(table, attr);
578 if (!ret)
579 dispatch_gid_change_event(ib_dev, port);
580
581out_unlock:
582 mutex_unlock(&table->lock);
583 if (ret)
584 pr_warn_ratelimited("%s: unable to add gid %pI6 error=%d\n",
585 __func__, gid->raw, ret);
586 return ret;
587}
588
589int ib_cache_gid_add(struct ib_device *ib_dev, u32 port,
590 union ib_gid *gid, struct ib_gid_attr *attr)
591{
592 unsigned long mask = GID_ATTR_FIND_MASK_GID |
593 GID_ATTR_FIND_MASK_GID_TYPE |
594 GID_ATTR_FIND_MASK_NETDEV;
595
596 return __ib_cache_gid_add(ib_dev, port, gid, attr, mask, false);
597}
598
599static int
600_ib_cache_gid_del(struct ib_device *ib_dev, u32 port,
601 union ib_gid *gid, struct ib_gid_attr *attr,
602 unsigned long mask, bool default_gid)
603{
604 struct ib_gid_table *table;
605 int ret = 0;
606 int ix;
607
608 table = rdma_gid_table(ib_dev, port);
609
610 mutex_lock(&table->lock);
611
612 ix = find_gid(table, gid, attr, default_gid, mask, NULL);
613 if (ix < 0) {
614 ret = -EINVAL;
615 goto out_unlock;
616 }
617
618 del_gid(ib_dev, port, table, ix);
619 dispatch_gid_change_event(ib_dev, port);
620
621out_unlock:
622 mutex_unlock(&table->lock);
623 if (ret)
624 pr_debug("%s: can't delete gid %pI6 error=%d\n",
625 __func__, gid->raw, ret);
626 return ret;
627}
628
629int ib_cache_gid_del(struct ib_device *ib_dev, u32 port,
630 union ib_gid *gid, struct ib_gid_attr *attr)
631{
632 unsigned long mask = GID_ATTR_FIND_MASK_GID |
633 GID_ATTR_FIND_MASK_GID_TYPE |
634 GID_ATTR_FIND_MASK_DEFAULT |
635 GID_ATTR_FIND_MASK_NETDEV;
636
637 return _ib_cache_gid_del(ib_dev, port, gid, attr, mask, false);
638}
639
640int ib_cache_gid_del_all_netdev_gids(struct ib_device *ib_dev, u32 port,
641 struct net_device *ndev)
642{
643 struct ib_gid_table *table;
644 int ix;
645 bool deleted = false;
646
647 table = rdma_gid_table(ib_dev, port);
648
649 mutex_lock(&table->lock);
650
651 for (ix = 0; ix < table->sz; ix++) {
652 if (is_gid_entry_valid(table->data_vec[ix]) &&
653 table->data_vec[ix]->attr.ndev == ndev) {
654 del_gid(ib_dev, port, table, ix);
655 deleted = true;
656 }
657 }
658
659 mutex_unlock(&table->lock);
660
661 if (deleted)
662 dispatch_gid_change_event(ib_dev, port);
663
664 return 0;
665}
666
667/**
668 * rdma_find_gid_by_port - Returns the GID entry attributes when it finds
669 * a valid GID entry for given search parameters. It searches for the specified
670 * GID value in the local software cache.
671 * @ib_dev: The device to query.
672 * @gid: The GID value to search for.
673 * @gid_type: The GID type to search for.
674 * @port: The port number of the device where the GID value should be searched.
675 * @ndev: In RoCE, the net device of the device. NULL means ignore.
676 *
677 * Returns sgid attributes if the GID is found with valid reference or
678 * returns ERR_PTR for the error.
679 * The caller must invoke rdma_put_gid_attr() to release the reference.
680 */
681const struct ib_gid_attr *
682rdma_find_gid_by_port(struct ib_device *ib_dev,
683 const union ib_gid *gid,
684 enum ib_gid_type gid_type,
685 u32 port, struct net_device *ndev)
686{
687 int local_index;
688 struct ib_gid_table *table;
689 unsigned long mask = GID_ATTR_FIND_MASK_GID |
690 GID_ATTR_FIND_MASK_GID_TYPE;
691 struct ib_gid_attr val = {.ndev = ndev, .gid_type = gid_type};
692 const struct ib_gid_attr *attr;
693 unsigned long flags;
694
695 if (!rdma_is_port_valid(ib_dev, port))
696 return ERR_PTR(-ENOENT);
697
698 table = rdma_gid_table(ib_dev, port);
699
700 if (ndev)
701 mask |= GID_ATTR_FIND_MASK_NETDEV;
702
703 read_lock_irqsave(&table->rwlock, flags);
704 local_index = find_gid(table, gid, &val, false, mask, NULL);
705 if (local_index >= 0) {
706 get_gid_entry(table->data_vec[local_index]);
707 attr = &table->data_vec[local_index]->attr;
708 read_unlock_irqrestore(&table->rwlock, flags);
709 return attr;
710 }
711
712 read_unlock_irqrestore(&table->rwlock, flags);
713 return ERR_PTR(-ENOENT);
714}
715EXPORT_SYMBOL(rdma_find_gid_by_port);
716
717/**
718 * rdma_find_gid_by_filter - Returns the GID table attribute where a
719 * specified GID value occurs
720 * @ib_dev: The device to query.
721 * @gid: The GID value to search for.
722 * @port: The port number of the device where the GID value could be
723 * searched.
724 * @filter: The filter function is executed on any matching GID in the table.
725 * If the filter function returns true, the corresponding index is returned,
726 * otherwise, we continue searching the GID table. It's guaranteed that
727 * while filter is executed, ndev field is valid and the structure won't
728 * change. filter is executed in an atomic context. filter must not be NULL.
729 * @context: Private data to pass into the call-back.
730 *
731 * rdma_find_gid_by_filter() searches for the specified GID value
732 * of which the filter function returns true in the port's GID table.
733 *
734 */
735const struct ib_gid_attr *rdma_find_gid_by_filter(
736 struct ib_device *ib_dev, const union ib_gid *gid, u32 port,
737 bool (*filter)(const union ib_gid *gid, const struct ib_gid_attr *,
738 void *),
739 void *context)
740{
741 const struct ib_gid_attr *res = ERR_PTR(-ENOENT);
742 struct ib_gid_table *table;
743 unsigned long flags;
744 unsigned int i;
745
746 if (!rdma_is_port_valid(ib_dev, port))
747 return ERR_PTR(-EINVAL);
748
749 table = rdma_gid_table(ib_dev, port);
750
751 read_lock_irqsave(&table->rwlock, flags);
752 for (i = 0; i < table->sz; i++) {
753 struct ib_gid_table_entry *entry = table->data_vec[i];
754
755 if (!is_gid_entry_valid(entry))
756 continue;
757
758 if (memcmp(gid, &entry->attr.gid, sizeof(*gid)))
759 continue;
760
761 if (filter(gid, &entry->attr, context)) {
762 get_gid_entry(entry);
763 res = &entry->attr;
764 break;
765 }
766 }
767 read_unlock_irqrestore(&table->rwlock, flags);
768 return res;
769}
770
771static struct ib_gid_table *alloc_gid_table(int sz)
772{
773 struct ib_gid_table *table = kzalloc_flex(*table, data_vec, sz);
774
775 if (!table)
776 return NULL;
777
778 table->sz = sz;
779
780 mutex_init(&table->lock);
781 rwlock_init(&table->rwlock);
782 return table;
783}
784
785static void release_gid_table(struct ib_device *device,
786 struct ib_gid_table *table)
787{
788 int i;
789
790 if (!table)
791 return;
792
793 for (i = 0; i < table->sz; i++) {
794 if (is_gid_entry_free(table->data_vec[i]))
795 continue;
796
797 WARN_ONCE(true,
798 "GID entry ref leak for dev %s index %d ref=%u\n",
799 dev_name(&device->dev), i,
800 kref_read(&table->data_vec[i]->kref));
801 }
802
803 mutex_destroy(&table->lock);
804 kfree(table);
805}
806
807static void cleanup_gid_table_port(struct ib_device *ib_dev, u32 port,
808 struct ib_gid_table *table)
809{
810 int i;
811
812 if (!table)
813 return;
814
815 mutex_lock(&table->lock);
816 for (i = 0; i < table->sz; ++i) {
817 if (is_gid_entry_valid(table->data_vec[i]))
818 del_gid(ib_dev, port, table, i);
819 }
820 mutex_unlock(&table->lock);
821}
822
823void ib_cache_gid_set_default_gid(struct ib_device *ib_dev, u32 port,
824 struct net_device *ndev,
825 unsigned long gid_type_mask,
826 enum ib_cache_gid_default_mode mode)
827{
828 union ib_gid gid = { };
829 struct ib_gid_attr gid_attr;
830 unsigned int gid_type;
831 unsigned long mask;
832
833 mask = GID_ATTR_FIND_MASK_GID_TYPE |
834 GID_ATTR_FIND_MASK_DEFAULT |
835 GID_ATTR_FIND_MASK_NETDEV;
836 memset(&gid_attr, 0, sizeof(gid_attr));
837 gid_attr.ndev = ndev;
838
839 for (gid_type = 0; gid_type < IB_GID_TYPE_SIZE; ++gid_type) {
840 if (1UL << gid_type & ~gid_type_mask)
841 continue;
842
843 gid_attr.gid_type = gid_type;
844
845 if (mode == IB_CACHE_GID_DEFAULT_MODE_SET) {
846 make_default_gid(ndev, &gid);
847 __ib_cache_gid_add(ib_dev, port, &gid,
848 &gid_attr, mask, true);
849 } else if (mode == IB_CACHE_GID_DEFAULT_MODE_DELETE) {
850 _ib_cache_gid_del(ib_dev, port, &gid,
851 &gid_attr, mask, true);
852 }
853 }
854}
855
856static void gid_table_reserve_default(struct ib_device *ib_dev, u32 port,
857 struct ib_gid_table *table)
858{
859 unsigned int i;
860 unsigned long roce_gid_type_mask;
861 unsigned int num_default_gids;
862
863 roce_gid_type_mask = roce_gid_type_mask_support(ib_dev, port);
864 num_default_gids = hweight_long(roce_gid_type_mask);
865 /* Reserve starting indices for default GIDs */
866 for (i = 0; i < num_default_gids && i < table->sz; i++)
867 table->default_gid_indices |= BIT(i);
868}
869
870
871static void gid_table_release_one(struct ib_device *ib_dev)
872{
873 u32 p;
874
875 rdma_for_each_port (ib_dev, p) {
876 release_gid_table(ib_dev, ib_dev->port_data[p].cache.gid);
877 ib_dev->port_data[p].cache.gid = NULL;
878 }
879}
880
881static int _gid_table_setup_one(struct ib_device *ib_dev)
882{
883 struct ib_gid_table *table;
884 u32 rdma_port;
885
886 rdma_for_each_port (ib_dev, rdma_port) {
887 table = alloc_gid_table(
888 ib_dev->port_data[rdma_port].immutable.gid_tbl_len);
889 if (!table)
890 goto rollback_table_setup;
891
892 gid_table_reserve_default(ib_dev, rdma_port, table);
893 ib_dev->port_data[rdma_port].cache.gid = table;
894 }
895 return 0;
896
897rollback_table_setup:
898 gid_table_release_one(ib_dev);
899 return -ENOMEM;
900}
901
902static void gid_table_cleanup_one(struct ib_device *ib_dev)
903{
904 u32 p;
905
906 rdma_for_each_port (ib_dev, p)
907 cleanup_gid_table_port(ib_dev, p,
908 ib_dev->port_data[p].cache.gid);
909}
910
911static int gid_table_setup_one(struct ib_device *ib_dev)
912{
913 int err;
914
915 err = _gid_table_setup_one(ib_dev);
916
917 if (err)
918 return err;
919
920 /*
921 * Mark the device as ready for GID cache updates. This allows netdev
922 * event handlers to update the GID cache even before the device is
923 * fully registered.
924 */
925 ib_device_enable_gid_updates(ib_dev);
926
927 rdma_roce_rescan_device(ib_dev);
928
929 return err;
930}
931
932/**
933 * rdma_query_gid - Read the GID content from the GID software cache
934 * @device: Device to query the GID
935 * @port_num: Port number of the device
936 * @index: Index of the GID table entry to read
937 * @gid: Pointer to GID where to store the entry's GID
938 *
939 * rdma_query_gid() only reads the GID entry content for requested device,
940 * port and index. It reads for IB, RoCE and iWarp link layers. It doesn't
941 * hold any reference to the GID table entry in the HCA or software cache.
942 *
943 * Returns 0 on success or appropriate error code.
944 *
945 */
946int rdma_query_gid(struct ib_device *device, u32 port_num,
947 int index, union ib_gid *gid)
948{
949 struct ib_gid_table *table;
950 unsigned long flags;
951 int res;
952
953 if (!rdma_is_port_valid(device, port_num))
954 return -EINVAL;
955
956 table = rdma_gid_table(device, port_num);
957 read_lock_irqsave(&table->rwlock, flags);
958
959 if (index < 0 || index >= table->sz) {
960 res = -EINVAL;
961 goto done;
962 }
963
964 if (!is_gid_entry_valid(table->data_vec[index])) {
965 res = -ENOENT;
966 goto done;
967 }
968
969 memcpy(gid, &table->data_vec[index]->attr.gid, sizeof(*gid));
970 res = 0;
971
972done:
973 read_unlock_irqrestore(&table->rwlock, flags);
974 return res;
975}
976EXPORT_SYMBOL(rdma_query_gid);
977
978/**
979 * rdma_read_gid_hw_context - Read the HW GID context from GID attribute
980 * @attr: Potinter to the GID attribute
981 *
982 * rdma_read_gid_hw_context() reads the drivers GID HW context corresponding
983 * to the SGID attr. Callers are required to already be holding the reference
984 * to an existing GID entry.
985 *
986 * Returns the HW GID context
987 *
988 */
989void *rdma_read_gid_hw_context(const struct ib_gid_attr *attr)
990{
991 return container_of(attr, struct ib_gid_table_entry, attr)->context;
992}
993EXPORT_SYMBOL(rdma_read_gid_hw_context);
994
995/**
996 * rdma_find_gid - Returns SGID attributes if the matching GID is found.
997 * @device: The device to query.
998 * @gid: The GID value to search for.
999 * @gid_type: The GID type to search for.
1000 * @ndev: In RoCE, the net device of the device. NULL means ignore.
1001 *
1002 * rdma_find_gid() searches for the specified GID value in the software cache.
1003 *
1004 * Returns GID attributes if a valid GID is found or returns ERR_PTR for the
1005 * error. The caller must invoke rdma_put_gid_attr() to release the reference.
1006 *
1007 */
1008const struct ib_gid_attr *rdma_find_gid(struct ib_device *device,
1009 const union ib_gid *gid,
1010 enum ib_gid_type gid_type,
1011 struct net_device *ndev)
1012{
1013 unsigned long mask = GID_ATTR_FIND_MASK_GID |
1014 GID_ATTR_FIND_MASK_GID_TYPE;
1015 struct ib_gid_attr gid_attr_val = {.ndev = ndev, .gid_type = gid_type};
1016 u32 p;
1017
1018 if (ndev)
1019 mask |= GID_ATTR_FIND_MASK_NETDEV;
1020
1021 rdma_for_each_port(device, p) {
1022 struct ib_gid_table *table;
1023 unsigned long flags;
1024 int index;
1025
1026 table = device->port_data[p].cache.gid;
1027 read_lock_irqsave(&table->rwlock, flags);
1028 index = find_gid(table, gid, &gid_attr_val, false, mask, NULL);
1029 if (index >= 0) {
1030 const struct ib_gid_attr *attr;
1031
1032 get_gid_entry(table->data_vec[index]);
1033 attr = &table->data_vec[index]->attr;
1034 read_unlock_irqrestore(&table->rwlock, flags);
1035 return attr;
1036 }
1037 read_unlock_irqrestore(&table->rwlock, flags);
1038 }
1039
1040 return ERR_PTR(-ENOENT);
1041}
1042EXPORT_SYMBOL(rdma_find_gid);
1043
1044int ib_get_cached_pkey(struct ib_device *device,
1045 u32 port_num,
1046 int index,
1047 u16 *pkey)
1048{
1049 struct ib_pkey_cache *cache;
1050 unsigned long flags;
1051 int ret = 0;
1052
1053 if (!rdma_is_port_valid(device, port_num))
1054 return -EINVAL;
1055
1056 read_lock_irqsave(&device->cache_lock, flags);
1057
1058 cache = device->port_data[port_num].cache.pkey;
1059
1060 if (!cache || index < 0 || index >= cache->table_len)
1061 ret = -EINVAL;
1062 else
1063 *pkey = cache->table[index];
1064
1065 read_unlock_irqrestore(&device->cache_lock, flags);
1066
1067 return ret;
1068}
1069EXPORT_SYMBOL(ib_get_cached_pkey);
1070
1071void ib_get_cached_subnet_prefix(struct ib_device *device, u32 port_num,
1072 u64 *sn_pfx)
1073{
1074 unsigned long flags;
1075
1076 read_lock_irqsave(&device->cache_lock, flags);
1077 *sn_pfx = device->port_data[port_num].cache.subnet_prefix;
1078 read_unlock_irqrestore(&device->cache_lock, flags);
1079}
1080EXPORT_SYMBOL(ib_get_cached_subnet_prefix);
1081
1082int ib_find_cached_pkey(struct ib_device *device, u32 port_num,
1083 u16 pkey, u16 *index)
1084{
1085 struct ib_pkey_cache *cache;
1086 unsigned long flags;
1087 int i;
1088 int ret = -ENOENT;
1089 int partial_ix = -1;
1090
1091 if (!rdma_is_port_valid(device, port_num))
1092 return -EINVAL;
1093
1094 read_lock_irqsave(&device->cache_lock, flags);
1095
1096 cache = device->port_data[port_num].cache.pkey;
1097 if (!cache) {
1098 ret = -EINVAL;
1099 goto err;
1100 }
1101
1102 *index = -1;
1103
1104 for (i = 0; i < cache->table_len; ++i)
1105 if ((cache->table[i] & 0x7fff) == (pkey & 0x7fff)) {
1106 if (cache->table[i] & 0x8000) {
1107 *index = i;
1108 ret = 0;
1109 break;
1110 } else {
1111 partial_ix = i;
1112 }
1113 }
1114
1115 if (ret && partial_ix >= 0) {
1116 *index = partial_ix;
1117 ret = 0;
1118 }
1119
1120err:
1121 read_unlock_irqrestore(&device->cache_lock, flags);
1122
1123 return ret;
1124}
1125EXPORT_SYMBOL(ib_find_cached_pkey);
1126
1127int ib_get_cached_lmc(struct ib_device *device, u32 port_num, u8 *lmc)
1128{
1129 unsigned long flags;
1130 int ret = 0;
1131
1132 if (!rdma_is_port_valid(device, port_num))
1133 return -EINVAL;
1134
1135 read_lock_irqsave(&device->cache_lock, flags);
1136 *lmc = device->port_data[port_num].cache.lmc;
1137 read_unlock_irqrestore(&device->cache_lock, flags);
1138
1139 return ret;
1140}
1141EXPORT_SYMBOL(ib_get_cached_lmc);
1142
1143int ib_get_cached_port_state(struct ib_device *device, u32 port_num,
1144 enum ib_port_state *port_state)
1145{
1146 unsigned long flags;
1147 int ret = 0;
1148
1149 if (!rdma_is_port_valid(device, port_num))
1150 return -EINVAL;
1151
1152 read_lock_irqsave(&device->cache_lock, flags);
1153 *port_state = device->port_data[port_num].cache.port_state;
1154 read_unlock_irqrestore(&device->cache_lock, flags);
1155
1156 return ret;
1157}
1158EXPORT_SYMBOL(ib_get_cached_port_state);
1159
1160/**
1161 * rdma_get_gid_attr - Returns GID attributes for a port of a device
1162 * at a requested gid_index, if a valid GID entry exists.
1163 * @device: The device to query.
1164 * @port_num: The port number on the device where the GID value
1165 * is to be queried.
1166 * @index: Index of the GID table entry whose attributes are to
1167 * be queried.
1168 *
1169 * rdma_get_gid_attr() acquires reference count of gid attributes from the
1170 * cached GID table. Caller must invoke rdma_put_gid_attr() to release
1171 * reference to gid attribute regardless of link layer.
1172 *
1173 * Returns pointer to valid gid attribute or ERR_PTR for the appropriate error
1174 * code.
1175 */
1176const struct ib_gid_attr *
1177rdma_get_gid_attr(struct ib_device *device, u32 port_num, int index)
1178{
1179 const struct ib_gid_attr *attr = ERR_PTR(-ENODATA);
1180 struct ib_gid_table *table;
1181 unsigned long flags;
1182
1183 if (!rdma_is_port_valid(device, port_num))
1184 return ERR_PTR(-EINVAL);
1185
1186 table = rdma_gid_table(device, port_num);
1187 if (index < 0 || index >= table->sz)
1188 return ERR_PTR(-EINVAL);
1189
1190 read_lock_irqsave(&table->rwlock, flags);
1191 if (!is_gid_entry_valid(table->data_vec[index]))
1192 goto done;
1193
1194 get_gid_entry(table->data_vec[index]);
1195 attr = &table->data_vec[index]->attr;
1196done:
1197 read_unlock_irqrestore(&table->rwlock, flags);
1198 return attr;
1199}
1200EXPORT_SYMBOL(rdma_get_gid_attr);
1201
1202/**
1203 * rdma_query_gid_table - Reads GID table entries of all the ports of a device up to max_entries.
1204 * @device: The device to query.
1205 * @entries: Entries where GID entries are returned.
1206 * @max_entries: Maximum number of entries that can be returned.
1207 * Entries array must be allocated to hold max_entries number of entries.
1208 *
1209 * Returns number of entries on success or appropriate error code.
1210 */
1211ssize_t rdma_query_gid_table(struct ib_device *device,
1212 struct ib_uverbs_gid_entry *entries,
1213 size_t max_entries)
1214{
1215 const struct ib_gid_attr *gid_attr;
1216 ssize_t num_entries = 0, ret;
1217 struct ib_gid_table *table;
1218 u32 port_num, i;
1219 struct net_device *ndev;
1220 unsigned long flags;
1221
1222 rdma_for_each_port(device, port_num) {
1223 table = rdma_gid_table(device, port_num);
1224 read_lock_irqsave(&table->rwlock, flags);
1225 for (i = 0; i < table->sz; i++) {
1226 if (!is_gid_entry_valid(table->data_vec[i]))
1227 continue;
1228 if (num_entries >= max_entries) {
1229 ret = -EINVAL;
1230 goto err;
1231 }
1232
1233 gid_attr = &table->data_vec[i]->attr;
1234
1235 memcpy(&entries->gid, &gid_attr->gid,
1236 sizeof(gid_attr->gid));
1237 entries->gid_index = gid_attr->index;
1238 entries->port_num = gid_attr->port_num;
1239 entries->gid_type = gid_attr->gid_type;
1240 ndev = rcu_dereference_protected(
1241 gid_attr->ndev,
1242 lockdep_is_held(&table->rwlock));
1243 if (ndev)
1244 entries->netdev_ifindex = ndev->ifindex;
1245
1246 num_entries++;
1247 entries++;
1248 }
1249 read_unlock_irqrestore(&table->rwlock, flags);
1250 }
1251
1252 return num_entries;
1253err:
1254 read_unlock_irqrestore(&table->rwlock, flags);
1255 return ret;
1256}
1257EXPORT_SYMBOL(rdma_query_gid_table);
1258
1259/**
1260 * rdma_put_gid_attr - Release reference to the GID attribute
1261 * @attr: Pointer to the GID attribute whose reference
1262 * needs to be released.
1263 *
1264 * rdma_put_gid_attr() must be used to release reference whose
1265 * reference is acquired using rdma_get_gid_attr() or any APIs
1266 * which returns a pointer to the ib_gid_attr regardless of link layer
1267 * of IB or RoCE.
1268 *
1269 */
1270void rdma_put_gid_attr(const struct ib_gid_attr *attr)
1271{
1272 struct ib_gid_table_entry *entry =
1273 container_of(attr, struct ib_gid_table_entry, attr);
1274
1275 put_gid_entry(entry);
1276}
1277EXPORT_SYMBOL(rdma_put_gid_attr);
1278
1279/**
1280 * rdma_hold_gid_attr - Get reference to existing GID attribute
1281 *
1282 * @attr: Pointer to the GID attribute whose reference
1283 * needs to be taken.
1284 *
1285 * Increase the reference count to a GID attribute to keep it from being
1286 * freed. Callers are required to already be holding a reference to attribute.
1287 *
1288 */
1289void rdma_hold_gid_attr(const struct ib_gid_attr *attr)
1290{
1291 struct ib_gid_table_entry *entry =
1292 container_of(attr, struct ib_gid_table_entry, attr);
1293
1294 get_gid_entry(entry);
1295}
1296EXPORT_SYMBOL(rdma_hold_gid_attr);
1297
1298/**
1299 * rdma_read_gid_attr_ndev_rcu - Read GID attribute netdevice
1300 * which must be in UP state.
1301 *
1302 * @attr:Pointer to the GID attribute
1303 *
1304 * Returns pointer to netdevice if the netdevice was attached to GID and
1305 * netdevice is in UP state. Caller must hold RCU lock as this API
1306 * reads the netdev flags which can change while netdevice migrates to
1307 * different net namespace. Returns ERR_PTR with error code otherwise.
1308 *
1309 */
1310struct net_device *rdma_read_gid_attr_ndev_rcu(const struct ib_gid_attr *attr)
1311{
1312 struct ib_gid_table_entry *entry =
1313 container_of(attr, struct ib_gid_table_entry, attr);
1314 struct ib_device *device = entry->attr.device;
1315 struct net_device *ndev = ERR_PTR(-EINVAL);
1316 u32 port_num = entry->attr.port_num;
1317 struct ib_gid_table *table;
1318 unsigned long flags;
1319 bool valid;
1320
1321 table = rdma_gid_table(device, port_num);
1322
1323 read_lock_irqsave(&table->rwlock, flags);
1324 valid = is_gid_entry_valid(table->data_vec[attr->index]);
1325 if (valid) {
1326 ndev = rcu_dereference(attr->ndev);
1327 if (!ndev)
1328 ndev = ERR_PTR(-ENODEV);
1329 }
1330 read_unlock_irqrestore(&table->rwlock, flags);
1331 return ndev;
1332}
1333EXPORT_SYMBOL(rdma_read_gid_attr_ndev_rcu);
1334
1335static int get_lower_dev_vlan(struct net_device *lower_dev,
1336 struct netdev_nested_priv *priv)
1337{
1338 u16 *vlan_id = (u16 *)priv->data;
1339
1340 if (is_vlan_dev(lower_dev))
1341 *vlan_id = vlan_dev_vlan_id(lower_dev);
1342
1343 /* We are interested only in first level vlan device, so
1344 * always return 1 to stop iterating over next level devices.
1345 */
1346 return 1;
1347}
1348
1349/**
1350 * rdma_read_gid_l2_fields - Read the vlan ID and source MAC address
1351 * of a GID entry.
1352 *
1353 * @attr: GID attribute pointer whose L2 fields to be read
1354 * @vlan_id: Pointer to vlan id to fill up if the GID entry has
1355 * vlan id. It is optional.
1356 * @smac: Pointer to smac to fill up for a GID entry. It is optional.
1357 *
1358 * rdma_read_gid_l2_fields() returns 0 on success and returns vlan id
1359 * (if gid entry has vlan) and source MAC, or returns error.
1360 */
1361int rdma_read_gid_l2_fields(const struct ib_gid_attr *attr,
1362 u16 *vlan_id, u8 *smac)
1363{
1364 struct netdev_nested_priv priv = {
1365 .data = (void *)vlan_id,
1366 };
1367 struct net_device *ndev;
1368
1369 rcu_read_lock();
1370 ndev = rcu_dereference(attr->ndev);
1371 if (!ndev) {
1372 rcu_read_unlock();
1373 return -ENODEV;
1374 }
1375 if (smac)
1376 ether_addr_copy(smac, ndev->dev_addr);
1377 if (vlan_id) {
1378 *vlan_id = 0xffff;
1379 if (is_vlan_dev(ndev)) {
1380 *vlan_id = vlan_dev_vlan_id(ndev);
1381 } else {
1382 /* If the netdev is upper device and if it's lower
1383 * device is vlan device, consider vlan id of
1384 * the lower vlan device for this gid entry.
1385 */
1386 netdev_walk_all_lower_dev_rcu(attr->ndev,
1387 get_lower_dev_vlan, &priv);
1388 }
1389 }
1390 rcu_read_unlock();
1391 return 0;
1392}
1393EXPORT_SYMBOL(rdma_read_gid_l2_fields);
1394
1395static int config_non_roce_gid_cache(struct ib_device *device,
1396 u32 port, struct ib_port_attr *tprops)
1397{
1398 struct ib_gid_attr gid_attr = {};
1399 struct ib_gid_table *table;
1400 int ret = 0;
1401 int i;
1402
1403 gid_attr.device = device;
1404 gid_attr.port_num = port;
1405 table = rdma_gid_table(device, port);
1406
1407 mutex_lock(&table->lock);
1408 for (i = 0; i < tprops->gid_tbl_len; ++i) {
1409 if (!device->ops.query_gid)
1410 continue;
1411 ret = device->ops.query_gid(device, port, i, &gid_attr.gid);
1412 if (ret) {
1413 dev_warn(&device->dev,
1414 "query_gid failed (%d) for index %d\n", ret,
1415 i);
1416 goto err;
1417 }
1418
1419 if (rdma_protocol_iwarp(device, port)) {
1420 struct net_device *ndev;
1421
1422 ndev = ib_device_get_netdev(device, port);
1423 if (!ndev)
1424 continue;
1425 RCU_INIT_POINTER(gid_attr.ndev, ndev);
1426 dev_put(ndev);
1427 }
1428
1429 gid_attr.index = i;
1430 tprops->subnet_prefix =
1431 be64_to_cpu(gid_attr.gid.global.subnet_prefix);
1432 add_modify_gid(table, &gid_attr);
1433 }
1434err:
1435 mutex_unlock(&table->lock);
1436 return ret;
1437}
1438
1439static int
1440ib_cache_update(struct ib_device *device, u32 port, bool update_gids,
1441 bool update_pkeys, bool enforce_security)
1442{
1443 struct ib_port_attr *tprops = NULL;
1444 struct ib_pkey_cache *pkey_cache = NULL;
1445 struct ib_pkey_cache *old_pkey_cache = NULL;
1446 int i;
1447 int ret;
1448
1449 if (!rdma_is_port_valid(device, port))
1450 return -EINVAL;
1451
1452 tprops = kmalloc_obj(*tprops);
1453 if (!tprops)
1454 return -ENOMEM;
1455
1456 ret = ib_query_port(device, port, tprops);
1457 if (ret) {
1458 dev_warn(&device->dev, "ib_query_port failed (%d)\n", ret);
1459 goto err;
1460 }
1461
1462 if (!rdma_protocol_roce(device, port) && update_gids) {
1463 ret = config_non_roce_gid_cache(device, port,
1464 tprops);
1465 if (ret)
1466 goto err;
1467 }
1468
1469 update_pkeys &= !!tprops->pkey_tbl_len;
1470
1471 if (update_pkeys) {
1472 pkey_cache = kmalloc_flex(*pkey_cache, table,
1473 tprops->pkey_tbl_len);
1474 if (!pkey_cache) {
1475 ret = -ENOMEM;
1476 goto err;
1477 }
1478
1479 pkey_cache->table_len = tprops->pkey_tbl_len;
1480
1481 for (i = 0; i < pkey_cache->table_len; ++i) {
1482 ret = ib_query_pkey(device, port, i,
1483 pkey_cache->table + i);
1484 if (ret) {
1485 dev_warn(&device->dev,
1486 "ib_query_pkey failed (%d) for index %d\n",
1487 ret, i);
1488 goto err;
1489 }
1490 }
1491 }
1492
1493 write_lock_irq(&device->cache_lock);
1494
1495 if (update_pkeys) {
1496 old_pkey_cache = device->port_data[port].cache.pkey;
1497 device->port_data[port].cache.pkey = pkey_cache;
1498 }
1499 device->port_data[port].cache.lmc = tprops->lmc;
1500
1501 if (device->port_data[port].cache.port_state != IB_PORT_NOP &&
1502 device->port_data[port].cache.port_state != tprops->state)
1503 ibdev_info(device, "Port: %d Link %s\n", port,
1504 ib_port_state_to_str(tprops->state));
1505
1506 device->port_data[port].cache.port_state = tprops->state;
1507
1508 device->port_data[port].cache.subnet_prefix = tprops->subnet_prefix;
1509 write_unlock_irq(&device->cache_lock);
1510
1511 if (enforce_security)
1512 ib_security_cache_change(device,
1513 port,
1514 tprops->subnet_prefix);
1515
1516 kfree(old_pkey_cache);
1517 kfree(tprops);
1518 return 0;
1519
1520err:
1521 kfree(pkey_cache);
1522 kfree(tprops);
1523 return ret;
1524}
1525
1526static void ib_cache_event_task(struct work_struct *_work)
1527{
1528 struct ib_update_work *work =
1529 container_of(_work, struct ib_update_work, work);
1530 int ret;
1531
1532 /* Before distributing the cache update event, first sync
1533 * the cache.
1534 */
1535 ret = ib_cache_update(work->event.device, work->event.element.port_num,
1536 work->event.event == IB_EVENT_GID_CHANGE ||
1537 work->event.event == IB_EVENT_CLIENT_REREGISTER,
1538 work->event.event == IB_EVENT_PKEY_CHANGE,
1539 work->enforce_security);
1540
1541 /* GID event is notified already for individual GID entries by
1542 * dispatch_gid_change_event(). Hence, notifiy for rest of the
1543 * events.
1544 */
1545 if (!ret && work->event.event != IB_EVENT_GID_CHANGE)
1546 ib_dispatch_event_clients(&work->event);
1547
1548 kfree(work);
1549}
1550
1551static void ib_generic_event_task(struct work_struct *_work)
1552{
1553 struct ib_update_work *work =
1554 container_of(_work, struct ib_update_work, work);
1555
1556 ib_dispatch_event_clients(&work->event);
1557 kfree(work);
1558}
1559
1560static bool is_cache_update_event(const struct ib_event *event)
1561{
1562 return (event->event == IB_EVENT_PORT_ERR ||
1563 event->event == IB_EVENT_PORT_ACTIVE ||
1564 event->event == IB_EVENT_LID_CHANGE ||
1565 event->event == IB_EVENT_PKEY_CHANGE ||
1566 event->event == IB_EVENT_CLIENT_REREGISTER ||
1567 event->event == IB_EVENT_GID_CHANGE);
1568}
1569
1570/**
1571 * ib_dispatch_event - Dispatch an asynchronous event
1572 * @event:Event to dispatch
1573 *
1574 * Low-level drivers must call ib_dispatch_event() to dispatch the
1575 * event to all registered event handlers when an asynchronous event
1576 * occurs.
1577 */
1578void ib_dispatch_event(const struct ib_event *event)
1579{
1580 struct ib_update_work *work;
1581
1582 work = kzalloc_obj(*work, GFP_ATOMIC);
1583 if (!work)
1584 return;
1585
1586 if (is_cache_update_event(event))
1587 INIT_WORK(&work->work, ib_cache_event_task);
1588 else
1589 INIT_WORK(&work->work, ib_generic_event_task);
1590
1591 work->event = *event;
1592 if (event->event == IB_EVENT_PKEY_CHANGE ||
1593 event->event == IB_EVENT_GID_CHANGE)
1594 work->enforce_security = true;
1595
1596 queue_work(ib_wq, &work->work);
1597}
1598EXPORT_SYMBOL(ib_dispatch_event);
1599
1600int ib_cache_setup_one(struct ib_device *device)
1601{
1602 u32 p;
1603 int err;
1604
1605 err = gid_table_setup_one(device);
1606 if (err)
1607 return err;
1608
1609 rdma_for_each_port (device, p) {
1610 err = ib_cache_update(device, p, true, true, true);
1611 if (err) {
1612 gid_table_cleanup_one(device);
1613 return err;
1614 }
1615 }
1616
1617 return 0;
1618}
1619
1620void ib_cache_release_one(struct ib_device *device)
1621{
1622 u32 p;
1623
1624 /*
1625 * The release function frees all the cache elements.
1626 * This function should be called as part of freeing
1627 * all the device's resources when the cache could no
1628 * longer be accessed.
1629 */
1630 rdma_for_each_port (device, p)
1631 kfree(device->port_data[p].cache.pkey);
1632
1633 gid_table_release_one(device);
1634}
1635
1636void ib_cache_cleanup_one(struct ib_device *device)
1637{
1638 /*
1639 * Clear the GID updates mark first to prevent event handlers from
1640 * accessing the device while it's being torn down.
1641 */
1642 ib_device_disable_gid_updates(device);
1643
1644 /* The cleanup function waits for all in-progress workqueue
1645 * elements and cleans up the GID cache. This function should be
1646 * called after the device was removed from the devices list and
1647 * all clients were removed, so the cache exists but is
1648 * non-functional and shouldn't be updated anymore.
1649 */
1650 flush_workqueue(ib_wq);
1651 gid_table_cleanup_one(device);
1652
1653 /*
1654 * Flush the wq second time for any pending GID delete work.
1655 */
1656 flush_workqueue(ib_wq);
1657}