xref: /linux/drivers/infiniband/core/cache.c (revision ca55b2fef3a9373fcfc30f82fd26bc7fccbda732)
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/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>
42 
43 #include <rdma/ib_cache.h>
44 
45 #include "core_priv.h"
46 
47 struct ib_pkey_cache {
48 	int             table_len;
49 	u16             table[0];
50 };
51 
52 struct ib_update_work {
53 	struct work_struct work;
54 	struct ib_device  *device;
55 	u8                 port_num;
56 };
57 
58 union ib_gid zgid;
59 EXPORT_SYMBOL(zgid);
60 
61 static const struct ib_gid_attr zattr;
62 
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 };
68 
69 enum gid_table_entry_props {
70 	GID_TABLE_ENTRY_INVALID		= 1UL << 0,
71 	GID_TABLE_ENTRY_DEFAULT		= 1UL << 1,
72 };
73 
74 enum gid_table_write_action {
75 	GID_TABLE_WRITE_ACTION_ADD,
76 	GID_TABLE_WRITE_ACTION_DEL,
77 	/* MODIFY only updates the GID table. Currently only used by
78 	 * ib_cache_update.
79 	 */
80 	GID_TABLE_WRITE_ACTION_MODIFY
81 };
82 
83 struct ib_gid_table_entry {
84 	/* This lock protects an entry from being
85 	 * read and written simultaneously.
86 	 */
87 	rwlock_t	    lock;
88 	unsigned long	    props;
89 	union ib_gid        gid;
90 	struct ib_gid_attr  attr;
91 	void		   *context;
92 };
93 
94 struct ib_gid_table {
95 	int                  sz;
96 	/* In RoCE, adding a GID to the table requires:
97 	 * (a) Find if this GID is already exists.
98 	 * (b) Find a free space.
99 	 * (c) Write the new GID
100 	 *
101 	 * Delete requires different set of operations:
102 	 * (a) Find the GID
103 	 * (b) Delete it.
104 	 *
105 	 * Add/delete should be carried out atomically.
106 	 * This is done by locking this mutex from multiple
107 	 * writers. We don't need this lock for IB, as the MAD
108 	 * layer replaces all entries. All data_vec entries
109 	 * are locked by this lock.
110 	 **/
111 	struct mutex         lock;
112 	struct ib_gid_table_entry *data_vec;
113 };
114 
115 static int write_gid(struct ib_device *ib_dev, u8 port,
116 		     struct ib_gid_table *table, int ix,
117 		     const union ib_gid *gid,
118 		     const struct ib_gid_attr *attr,
119 		     enum gid_table_write_action action,
120 		     bool  default_gid)
121 {
122 	int ret = 0;
123 	struct net_device *old_net_dev;
124 	unsigned long flags;
125 
126 	/* in rdma_cap_roce_gid_table, this funciton should be protected by a
127 	 * sleep-able lock.
128 	 */
129 	write_lock_irqsave(&table->data_vec[ix].lock, flags);
130 
131 	if (rdma_cap_roce_gid_table(ib_dev, port)) {
132 		table->data_vec[ix].props |= GID_TABLE_ENTRY_INVALID;
133 		write_unlock_irqrestore(&table->data_vec[ix].lock, flags);
134 		/* GID_TABLE_WRITE_ACTION_MODIFY currently isn't supported by
135 		 * RoCE providers and thus only updates the cache.
136 		 */
137 		if (action == GID_TABLE_WRITE_ACTION_ADD)
138 			ret = ib_dev->add_gid(ib_dev, port, ix, gid, attr,
139 					      &table->data_vec[ix].context);
140 		else if (action == GID_TABLE_WRITE_ACTION_DEL)
141 			ret = ib_dev->del_gid(ib_dev, port, ix,
142 					      &table->data_vec[ix].context);
143 		write_lock_irqsave(&table->data_vec[ix].lock, flags);
144 	}
145 
146 	old_net_dev = table->data_vec[ix].attr.ndev;
147 	if (old_net_dev && old_net_dev != attr->ndev)
148 		dev_put(old_net_dev);
149 	/* if modify_gid failed, just delete the old gid */
150 	if (ret || action == GID_TABLE_WRITE_ACTION_DEL) {
151 		gid = &zgid;
152 		attr = &zattr;
153 		table->data_vec[ix].context = NULL;
154 	}
155 	if (default_gid)
156 		table->data_vec[ix].props |= GID_TABLE_ENTRY_DEFAULT;
157 	memcpy(&table->data_vec[ix].gid, gid, sizeof(*gid));
158 	memcpy(&table->data_vec[ix].attr, attr, sizeof(*attr));
159 	if (table->data_vec[ix].attr.ndev &&
160 	    table->data_vec[ix].attr.ndev != old_net_dev)
161 		dev_hold(table->data_vec[ix].attr.ndev);
162 
163 	table->data_vec[ix].props &= ~GID_TABLE_ENTRY_INVALID;
164 
165 	write_unlock_irqrestore(&table->data_vec[ix].lock, flags);
166 
167 	if (!ret && rdma_cap_roce_gid_table(ib_dev, port)) {
168 		struct ib_event event;
169 
170 		event.device		= ib_dev;
171 		event.element.port_num	= port;
172 		event.event		= IB_EVENT_GID_CHANGE;
173 
174 		ib_dispatch_event(&event);
175 	}
176 	return ret;
177 }
178 
179 static int add_gid(struct ib_device *ib_dev, u8 port,
180 		   struct ib_gid_table *table, int ix,
181 		   const union ib_gid *gid,
182 		   const struct ib_gid_attr *attr,
183 		   bool  default_gid) {
184 	return write_gid(ib_dev, port, table, ix, gid, attr,
185 			 GID_TABLE_WRITE_ACTION_ADD, default_gid);
186 }
187 
188 static int modify_gid(struct ib_device *ib_dev, u8 port,
189 		      struct ib_gid_table *table, int ix,
190 		      const union ib_gid *gid,
191 		      const struct ib_gid_attr *attr,
192 		      bool  default_gid) {
193 	return write_gid(ib_dev, port, table, ix, gid, attr,
194 			 GID_TABLE_WRITE_ACTION_MODIFY, default_gid);
195 }
196 
197 static int del_gid(struct ib_device *ib_dev, u8 port,
198 		   struct ib_gid_table *table, int ix,
199 		   bool  default_gid) {
200 	return write_gid(ib_dev, port, table, ix, &zgid, &zattr,
201 			 GID_TABLE_WRITE_ACTION_DEL, default_gid);
202 }
203 
204 static int find_gid(struct ib_gid_table *table, const union ib_gid *gid,
205 		    const struct ib_gid_attr *val, bool default_gid,
206 		    unsigned long mask)
207 {
208 	int i;
209 
210 	for (i = 0; i < table->sz; i++) {
211 		unsigned long flags;
212 		struct ib_gid_attr *attr = &table->data_vec[i].attr;
213 
214 		read_lock_irqsave(&table->data_vec[i].lock, flags);
215 
216 		if (table->data_vec[i].props & GID_TABLE_ENTRY_INVALID)
217 			goto next;
218 
219 		if (mask & GID_ATTR_FIND_MASK_GID &&
220 		    memcmp(gid, &table->data_vec[i].gid, sizeof(*gid)))
221 			goto next;
222 
223 		if (mask & GID_ATTR_FIND_MASK_NETDEV &&
224 		    attr->ndev != val->ndev)
225 			goto next;
226 
227 		if (mask & GID_ATTR_FIND_MASK_DEFAULT &&
228 		    !!(table->data_vec[i].props & GID_TABLE_ENTRY_DEFAULT) !=
229 		    default_gid)
230 			goto next;
231 
232 		read_unlock_irqrestore(&table->data_vec[i].lock, flags);
233 		return i;
234 next:
235 		read_unlock_irqrestore(&table->data_vec[i].lock, flags);
236 	}
237 
238 	return -1;
239 }
240 
241 static void make_default_gid(struct  net_device *dev, union ib_gid *gid)
242 {
243 	gid->global.subnet_prefix = cpu_to_be64(0xfe80000000000000LL);
244 	addrconf_ifid_eui48(&gid->raw[8], dev);
245 }
246 
247 int ib_cache_gid_add(struct ib_device *ib_dev, u8 port,
248 		     union ib_gid *gid, struct ib_gid_attr *attr)
249 {
250 	struct ib_gid_table **ports_table = ib_dev->cache.gid_cache;
251 	struct ib_gid_table *table;
252 	int ix;
253 	int ret = 0;
254 	struct net_device *idev;
255 
256 	table = ports_table[port - rdma_start_port(ib_dev)];
257 
258 	if (!memcmp(gid, &zgid, sizeof(*gid)))
259 		return -EINVAL;
260 
261 	if (ib_dev->get_netdev) {
262 		idev = ib_dev->get_netdev(ib_dev, port);
263 		if (idev && attr->ndev != idev) {
264 			union ib_gid default_gid;
265 
266 			/* Adding default GIDs in not permitted */
267 			make_default_gid(idev, &default_gid);
268 			if (!memcmp(gid, &default_gid, sizeof(*gid))) {
269 				dev_put(idev);
270 				return -EPERM;
271 			}
272 		}
273 		if (idev)
274 			dev_put(idev);
275 	}
276 
277 	mutex_lock(&table->lock);
278 
279 	ix = find_gid(table, gid, attr, false, GID_ATTR_FIND_MASK_GID |
280 		      GID_ATTR_FIND_MASK_NETDEV);
281 	if (ix >= 0)
282 		goto out_unlock;
283 
284 	ix = find_gid(table, &zgid, NULL, false, GID_ATTR_FIND_MASK_GID |
285 		      GID_ATTR_FIND_MASK_DEFAULT);
286 	if (ix < 0) {
287 		ret = -ENOSPC;
288 		goto out_unlock;
289 	}
290 
291 	add_gid(ib_dev, port, table, ix, gid, attr, false);
292 
293 out_unlock:
294 	mutex_unlock(&table->lock);
295 	return ret;
296 }
297 
298 int ib_cache_gid_del(struct ib_device *ib_dev, u8 port,
299 		     union ib_gid *gid, struct ib_gid_attr *attr)
300 {
301 	struct ib_gid_table **ports_table = ib_dev->cache.gid_cache;
302 	struct ib_gid_table *table;
303 	int ix;
304 
305 	table = ports_table[port - rdma_start_port(ib_dev)];
306 
307 	mutex_lock(&table->lock);
308 
309 	ix = find_gid(table, gid, attr, false,
310 		      GID_ATTR_FIND_MASK_GID	  |
311 		      GID_ATTR_FIND_MASK_NETDEV	  |
312 		      GID_ATTR_FIND_MASK_DEFAULT);
313 	if (ix < 0)
314 		goto out_unlock;
315 
316 	del_gid(ib_dev, port, table, ix, false);
317 
318 out_unlock:
319 	mutex_unlock(&table->lock);
320 	return 0;
321 }
322 
323 int ib_cache_gid_del_all_netdev_gids(struct ib_device *ib_dev, u8 port,
324 				     struct net_device *ndev)
325 {
326 	struct ib_gid_table **ports_table = ib_dev->cache.gid_cache;
327 	struct ib_gid_table *table;
328 	int ix;
329 
330 	table  = ports_table[port - rdma_start_port(ib_dev)];
331 
332 	mutex_lock(&table->lock);
333 
334 	for (ix = 0; ix < table->sz; ix++)
335 		if (table->data_vec[ix].attr.ndev == ndev)
336 			del_gid(ib_dev, port, table, ix, false);
337 
338 	mutex_unlock(&table->lock);
339 	return 0;
340 }
341 
342 static int __ib_cache_gid_get(struct ib_device *ib_dev, u8 port, int index,
343 			      union ib_gid *gid, struct ib_gid_attr *attr)
344 {
345 	struct ib_gid_table **ports_table = ib_dev->cache.gid_cache;
346 	struct ib_gid_table *table;
347 	unsigned long flags;
348 
349 	table = ports_table[port - rdma_start_port(ib_dev)];
350 
351 	if (index < 0 || index >= table->sz)
352 		return -EINVAL;
353 
354 	read_lock_irqsave(&table->data_vec[index].lock, flags);
355 	if (table->data_vec[index].props & GID_TABLE_ENTRY_INVALID) {
356 		read_unlock_irqrestore(&table->data_vec[index].lock, flags);
357 		return -EAGAIN;
358 	}
359 
360 	memcpy(gid, &table->data_vec[index].gid, sizeof(*gid));
361 	if (attr) {
362 		memcpy(attr, &table->data_vec[index].attr, sizeof(*attr));
363 		if (attr->ndev)
364 			dev_hold(attr->ndev);
365 	}
366 
367 	read_unlock_irqrestore(&table->data_vec[index].lock, flags);
368 	return 0;
369 }
370 
371 static int _ib_cache_gid_table_find(struct ib_device *ib_dev,
372 				    const union ib_gid *gid,
373 				    const struct ib_gid_attr *val,
374 				    unsigned long mask,
375 				    u8 *port, u16 *index)
376 {
377 	struct ib_gid_table **ports_table = ib_dev->cache.gid_cache;
378 	struct ib_gid_table *table;
379 	u8 p;
380 	int local_index;
381 
382 	for (p = 0; p < ib_dev->phys_port_cnt; p++) {
383 		table = ports_table[p];
384 		local_index = find_gid(table, gid, val, false, mask);
385 		if (local_index >= 0) {
386 			if (index)
387 				*index = local_index;
388 			if (port)
389 				*port = p + rdma_start_port(ib_dev);
390 			return 0;
391 		}
392 	}
393 
394 	return -ENOENT;
395 }
396 
397 static int ib_cache_gid_find(struct ib_device *ib_dev,
398 			     const union ib_gid *gid,
399 			     struct net_device *ndev, u8 *port,
400 			     u16 *index)
401 {
402 	unsigned long mask = GID_ATTR_FIND_MASK_GID;
403 	struct ib_gid_attr gid_attr_val = {.ndev = ndev};
404 
405 	if (ndev)
406 		mask |= GID_ATTR_FIND_MASK_NETDEV;
407 
408 	return _ib_cache_gid_table_find(ib_dev, gid, &gid_attr_val,
409 					mask, port, index);
410 }
411 
412 int ib_cache_gid_find_by_port(struct ib_device *ib_dev,
413 			      const union ib_gid *gid,
414 			      u8 port, struct net_device *ndev,
415 			      u16 *index)
416 {
417 	int local_index;
418 	struct ib_gid_table **ports_table = ib_dev->cache.gid_cache;
419 	struct ib_gid_table *table;
420 	unsigned long mask = GID_ATTR_FIND_MASK_GID;
421 	struct ib_gid_attr val = {.ndev = ndev};
422 
423 	if (port < rdma_start_port(ib_dev) ||
424 	    port > rdma_end_port(ib_dev))
425 		return -ENOENT;
426 
427 	table = ports_table[port - rdma_start_port(ib_dev)];
428 
429 	if (ndev)
430 		mask |= GID_ATTR_FIND_MASK_NETDEV;
431 
432 	local_index = find_gid(table, gid, &val, false, mask);
433 	if (local_index >= 0) {
434 		if (index)
435 			*index = local_index;
436 		return 0;
437 	}
438 
439 	return -ENOENT;
440 }
441 
442 static struct ib_gid_table *alloc_gid_table(int sz)
443 {
444 	unsigned int i;
445 	struct ib_gid_table *table =
446 		kzalloc(sizeof(struct ib_gid_table), GFP_KERNEL);
447 	if (!table)
448 		return NULL;
449 
450 	table->data_vec = kcalloc(sz, sizeof(*table->data_vec), GFP_KERNEL);
451 	if (!table->data_vec)
452 		goto err_free_table;
453 
454 	mutex_init(&table->lock);
455 
456 	table->sz = sz;
457 
458 	for (i = 0; i < sz; i++)
459 		rwlock_init(&table->data_vec[i].lock);
460 
461 	return table;
462 
463 err_free_table:
464 	kfree(table);
465 	return NULL;
466 }
467 
468 static void release_gid_table(struct ib_gid_table *table)
469 {
470 	if (table) {
471 		kfree(table->data_vec);
472 		kfree(table);
473 	}
474 }
475 
476 static void cleanup_gid_table_port(struct ib_device *ib_dev, u8 port,
477 				   struct ib_gid_table *table)
478 {
479 	int i;
480 
481 	if (!table)
482 		return;
483 
484 	for (i = 0; i < table->sz; ++i) {
485 		if (memcmp(&table->data_vec[i].gid, &zgid,
486 			   sizeof(table->data_vec[i].gid)))
487 			del_gid(ib_dev, port, table, i,
488 				table->data_vec[i].props &
489 				GID_ATTR_FIND_MASK_DEFAULT);
490 	}
491 }
492 
493 void ib_cache_gid_set_default_gid(struct ib_device *ib_dev, u8 port,
494 				  struct net_device *ndev,
495 				  enum ib_cache_gid_default_mode mode)
496 {
497 	struct ib_gid_table **ports_table = ib_dev->cache.gid_cache;
498 	union ib_gid gid;
499 	struct ib_gid_attr gid_attr;
500 	struct ib_gid_table *table;
501 	int ix;
502 	union ib_gid current_gid;
503 	struct ib_gid_attr current_gid_attr = {};
504 
505 	table  = ports_table[port - rdma_start_port(ib_dev)];
506 
507 	make_default_gid(ndev, &gid);
508 	memset(&gid_attr, 0, sizeof(gid_attr));
509 	gid_attr.ndev = ndev;
510 
511 	mutex_lock(&table->lock);
512 	ix = find_gid(table, NULL, NULL, true, GID_ATTR_FIND_MASK_DEFAULT);
513 
514 	/* Coudn't find default GID location */
515 	WARN_ON(ix < 0);
516 
517 	if (!__ib_cache_gid_get(ib_dev, port, ix,
518 				&current_gid, &current_gid_attr) &&
519 	    mode == IB_CACHE_GID_DEFAULT_MODE_SET &&
520 	    !memcmp(&gid, &current_gid, sizeof(gid)) &&
521 	    !memcmp(&gid_attr, &current_gid_attr, sizeof(gid_attr)))
522 		goto unlock;
523 
524 	if ((memcmp(&current_gid, &zgid, sizeof(current_gid)) ||
525 	     memcmp(&current_gid_attr, &zattr,
526 		    sizeof(current_gid_attr))) &&
527 	    del_gid(ib_dev, port, table, ix, true)) {
528 		pr_warn("ib_cache_gid: can't delete index %d for default gid %pI6\n",
529 			ix, gid.raw);
530 		goto unlock;
531 	}
532 
533 	if (mode == IB_CACHE_GID_DEFAULT_MODE_SET)
534 		if (add_gid(ib_dev, port, table, ix, &gid, &gid_attr, true))
535 			pr_warn("ib_cache_gid: unable to add default gid %pI6\n",
536 				gid.raw);
537 
538 unlock:
539 	if (current_gid_attr.ndev)
540 		dev_put(current_gid_attr.ndev);
541 	mutex_unlock(&table->lock);
542 }
543 
544 static int gid_table_reserve_default(struct ib_device *ib_dev, u8 port,
545 				     struct ib_gid_table *table)
546 {
547 	if (rdma_protocol_roce(ib_dev, port)) {
548 		struct ib_gid_table_entry *entry = &table->data_vec[0];
549 
550 		entry->props |= GID_TABLE_ENTRY_DEFAULT;
551 	}
552 
553 	return 0;
554 }
555 
556 static int _gid_table_setup_one(struct ib_device *ib_dev)
557 {
558 	u8 port;
559 	struct ib_gid_table **table;
560 	int err = 0;
561 
562 	table = kcalloc(ib_dev->phys_port_cnt, sizeof(*table), GFP_KERNEL);
563 
564 	if (!table) {
565 		pr_warn("failed to allocate ib gid cache for %s\n",
566 			ib_dev->name);
567 		return -ENOMEM;
568 	}
569 
570 	for (port = 0; port < ib_dev->phys_port_cnt; port++) {
571 		u8 rdma_port = port + rdma_start_port(ib_dev);
572 
573 		table[port] =
574 			alloc_gid_table(
575 				ib_dev->port_immutable[rdma_port].gid_tbl_len);
576 		if (!table[port]) {
577 			err = -ENOMEM;
578 			goto rollback_table_setup;
579 		}
580 
581 		err = gid_table_reserve_default(ib_dev,
582 						port + rdma_start_port(ib_dev),
583 						table[port]);
584 		if (err)
585 			goto rollback_table_setup;
586 	}
587 
588 	ib_dev->cache.gid_cache = table;
589 	return 0;
590 
591 rollback_table_setup:
592 	for (port = 0; port < ib_dev->phys_port_cnt; port++) {
593 		cleanup_gid_table_port(ib_dev, port + rdma_start_port(ib_dev),
594 				       table[port]);
595 		release_gid_table(table[port]);
596 	}
597 
598 	kfree(table);
599 	return err;
600 }
601 
602 static void gid_table_release_one(struct ib_device *ib_dev)
603 {
604 	struct ib_gid_table **table = ib_dev->cache.gid_cache;
605 	u8 port;
606 
607 	if (!table)
608 		return;
609 
610 	for (port = 0; port < ib_dev->phys_port_cnt; port++)
611 		release_gid_table(table[port]);
612 
613 	kfree(table);
614 	ib_dev->cache.gid_cache = NULL;
615 }
616 
617 static void gid_table_cleanup_one(struct ib_device *ib_dev)
618 {
619 	struct ib_gid_table **table = ib_dev->cache.gid_cache;
620 	u8 port;
621 
622 	if (!table)
623 		return;
624 
625 	for (port = 0; port < ib_dev->phys_port_cnt; port++)
626 		cleanup_gid_table_port(ib_dev, port + rdma_start_port(ib_dev),
627 				       table[port]);
628 }
629 
630 static int gid_table_setup_one(struct ib_device *ib_dev)
631 {
632 	int err;
633 
634 	err = _gid_table_setup_one(ib_dev);
635 
636 	if (err)
637 		return err;
638 
639 	err = roce_rescan_device(ib_dev);
640 
641 	if (err) {
642 		gid_table_cleanup_one(ib_dev);
643 		gid_table_release_one(ib_dev);
644 	}
645 
646 	return err;
647 }
648 
649 int ib_get_cached_gid(struct ib_device *device,
650 		      u8                port_num,
651 		      int               index,
652 		      union ib_gid     *gid)
653 {
654 	if (port_num < rdma_start_port(device) || port_num > rdma_end_port(device))
655 		return -EINVAL;
656 
657 	return __ib_cache_gid_get(device, port_num, index, gid, NULL);
658 }
659 EXPORT_SYMBOL(ib_get_cached_gid);
660 
661 int ib_find_cached_gid(struct ib_device *device,
662 		       const union ib_gid *gid,
663 		       u8               *port_num,
664 		       u16              *index)
665 {
666 	return ib_cache_gid_find(device, gid, NULL, port_num, index);
667 }
668 EXPORT_SYMBOL(ib_find_cached_gid);
669 
670 int ib_get_cached_pkey(struct ib_device *device,
671 		       u8                port_num,
672 		       int               index,
673 		       u16              *pkey)
674 {
675 	struct ib_pkey_cache *cache;
676 	unsigned long flags;
677 	int ret = 0;
678 
679 	if (port_num < rdma_start_port(device) || port_num > rdma_end_port(device))
680 		return -EINVAL;
681 
682 	read_lock_irqsave(&device->cache.lock, flags);
683 
684 	cache = device->cache.pkey_cache[port_num - rdma_start_port(device)];
685 
686 	if (index < 0 || index >= cache->table_len)
687 		ret = -EINVAL;
688 	else
689 		*pkey = cache->table[index];
690 
691 	read_unlock_irqrestore(&device->cache.lock, flags);
692 
693 	return ret;
694 }
695 EXPORT_SYMBOL(ib_get_cached_pkey);
696 
697 int ib_find_cached_pkey(struct ib_device *device,
698 			u8                port_num,
699 			u16               pkey,
700 			u16              *index)
701 {
702 	struct ib_pkey_cache *cache;
703 	unsigned long flags;
704 	int i;
705 	int ret = -ENOENT;
706 	int partial_ix = -1;
707 
708 	if (port_num < rdma_start_port(device) || port_num > rdma_end_port(device))
709 		return -EINVAL;
710 
711 	read_lock_irqsave(&device->cache.lock, flags);
712 
713 	cache = device->cache.pkey_cache[port_num - rdma_start_port(device)];
714 
715 	*index = -1;
716 
717 	for (i = 0; i < cache->table_len; ++i)
718 		if ((cache->table[i] & 0x7fff) == (pkey & 0x7fff)) {
719 			if (cache->table[i] & 0x8000) {
720 				*index = i;
721 				ret = 0;
722 				break;
723 			} else
724 				partial_ix = i;
725 		}
726 
727 	if (ret && partial_ix >= 0) {
728 		*index = partial_ix;
729 		ret = 0;
730 	}
731 
732 	read_unlock_irqrestore(&device->cache.lock, flags);
733 
734 	return ret;
735 }
736 EXPORT_SYMBOL(ib_find_cached_pkey);
737 
738 int ib_find_exact_cached_pkey(struct ib_device *device,
739 			      u8                port_num,
740 			      u16               pkey,
741 			      u16              *index)
742 {
743 	struct ib_pkey_cache *cache;
744 	unsigned long flags;
745 	int i;
746 	int ret = -ENOENT;
747 
748 	if (port_num < rdma_start_port(device) || port_num > rdma_end_port(device))
749 		return -EINVAL;
750 
751 	read_lock_irqsave(&device->cache.lock, flags);
752 
753 	cache = device->cache.pkey_cache[port_num - rdma_start_port(device)];
754 
755 	*index = -1;
756 
757 	for (i = 0; i < cache->table_len; ++i)
758 		if (cache->table[i] == pkey) {
759 			*index = i;
760 			ret = 0;
761 			break;
762 		}
763 
764 	read_unlock_irqrestore(&device->cache.lock, flags);
765 
766 	return ret;
767 }
768 EXPORT_SYMBOL(ib_find_exact_cached_pkey);
769 
770 int ib_get_cached_lmc(struct ib_device *device,
771 		      u8                port_num,
772 		      u8                *lmc)
773 {
774 	unsigned long flags;
775 	int ret = 0;
776 
777 	if (port_num < rdma_start_port(device) || port_num > rdma_end_port(device))
778 		return -EINVAL;
779 
780 	read_lock_irqsave(&device->cache.lock, flags);
781 	*lmc = device->cache.lmc_cache[port_num - rdma_start_port(device)];
782 	read_unlock_irqrestore(&device->cache.lock, flags);
783 
784 	return ret;
785 }
786 EXPORT_SYMBOL(ib_get_cached_lmc);
787 
788 static void ib_cache_update(struct ib_device *device,
789 			    u8                port)
790 {
791 	struct ib_port_attr       *tprops = NULL;
792 	struct ib_pkey_cache      *pkey_cache = NULL, *old_pkey_cache;
793 	struct ib_gid_cache {
794 		int             table_len;
795 		union ib_gid    table[0];
796 	}			  *gid_cache = NULL;
797 	int                        i;
798 	int                        ret;
799 	struct ib_gid_table	  *table;
800 	struct ib_gid_table	 **ports_table = device->cache.gid_cache;
801 	bool			   use_roce_gid_table =
802 					rdma_cap_roce_gid_table(device, port);
803 
804 	if (port < rdma_start_port(device) || port > rdma_end_port(device))
805 		return;
806 
807 	table = ports_table[port - rdma_start_port(device)];
808 
809 	tprops = kmalloc(sizeof *tprops, GFP_KERNEL);
810 	if (!tprops)
811 		return;
812 
813 	ret = ib_query_port(device, port, tprops);
814 	if (ret) {
815 		printk(KERN_WARNING "ib_query_port failed (%d) for %s\n",
816 		       ret, device->name);
817 		goto err;
818 	}
819 
820 	pkey_cache = kmalloc(sizeof *pkey_cache + tprops->pkey_tbl_len *
821 			     sizeof *pkey_cache->table, GFP_KERNEL);
822 	if (!pkey_cache)
823 		goto err;
824 
825 	pkey_cache->table_len = tprops->pkey_tbl_len;
826 
827 	if (!use_roce_gid_table) {
828 		gid_cache = kmalloc(sizeof(*gid_cache) + tprops->gid_tbl_len *
829 			    sizeof(*gid_cache->table), GFP_KERNEL);
830 		if (!gid_cache)
831 			goto err;
832 
833 		gid_cache->table_len = tprops->gid_tbl_len;
834 	}
835 
836 	for (i = 0; i < pkey_cache->table_len; ++i) {
837 		ret = ib_query_pkey(device, port, i, pkey_cache->table + i);
838 		if (ret) {
839 			printk(KERN_WARNING "ib_query_pkey failed (%d) for %s (index %d)\n",
840 			       ret, device->name, i);
841 			goto err;
842 		}
843 	}
844 
845 	if (!use_roce_gid_table) {
846 		for (i = 0;  i < gid_cache->table_len; ++i) {
847 			ret = ib_query_gid(device, port, i,
848 					   gid_cache->table + i);
849 			if (ret) {
850 				printk(KERN_WARNING "ib_query_gid failed (%d) for %s (index %d)\n",
851 				       ret, device->name, i);
852 				goto err;
853 			}
854 		}
855 	}
856 
857 	write_lock_irq(&device->cache.lock);
858 
859 	old_pkey_cache = device->cache.pkey_cache[port - rdma_start_port(device)];
860 
861 	device->cache.pkey_cache[port - rdma_start_port(device)] = pkey_cache;
862 	if (!use_roce_gid_table) {
863 		for (i = 0; i < gid_cache->table_len; i++) {
864 			modify_gid(device, port, table, i, gid_cache->table + i,
865 				   &zattr, false);
866 		}
867 	}
868 
869 	device->cache.lmc_cache[port - rdma_start_port(device)] = tprops->lmc;
870 
871 	write_unlock_irq(&device->cache.lock);
872 
873 	kfree(gid_cache);
874 	kfree(old_pkey_cache);
875 	kfree(tprops);
876 	return;
877 
878 err:
879 	kfree(pkey_cache);
880 	kfree(gid_cache);
881 	kfree(tprops);
882 }
883 
884 static void ib_cache_task(struct work_struct *_work)
885 {
886 	struct ib_update_work *work =
887 		container_of(_work, struct ib_update_work, work);
888 
889 	ib_cache_update(work->device, work->port_num);
890 	kfree(work);
891 }
892 
893 static void ib_cache_event(struct ib_event_handler *handler,
894 			   struct ib_event *event)
895 {
896 	struct ib_update_work *work;
897 
898 	if (event->event == IB_EVENT_PORT_ERR    ||
899 	    event->event == IB_EVENT_PORT_ACTIVE ||
900 	    event->event == IB_EVENT_LID_CHANGE  ||
901 	    event->event == IB_EVENT_PKEY_CHANGE ||
902 	    event->event == IB_EVENT_SM_CHANGE   ||
903 	    event->event == IB_EVENT_CLIENT_REREGISTER ||
904 	    event->event == IB_EVENT_GID_CHANGE) {
905 		work = kmalloc(sizeof *work, GFP_ATOMIC);
906 		if (work) {
907 			INIT_WORK(&work->work, ib_cache_task);
908 			work->device   = event->device;
909 			work->port_num = event->element.port_num;
910 			queue_work(ib_wq, &work->work);
911 		}
912 	}
913 }
914 
915 int ib_cache_setup_one(struct ib_device *device)
916 {
917 	int p;
918 	int err;
919 
920 	rwlock_init(&device->cache.lock);
921 
922 	device->cache.pkey_cache =
923 		kzalloc(sizeof *device->cache.pkey_cache *
924 			(rdma_end_port(device) - rdma_start_port(device) + 1), GFP_KERNEL);
925 	device->cache.lmc_cache = kmalloc(sizeof *device->cache.lmc_cache *
926 					  (rdma_end_port(device) -
927 					   rdma_start_port(device) + 1),
928 					  GFP_KERNEL);
929 	if (!device->cache.pkey_cache ||
930 	    !device->cache.lmc_cache) {
931 		printk(KERN_WARNING "Couldn't allocate cache "
932 		       "for %s\n", device->name);
933 		return -ENOMEM;
934 	}
935 
936 	err = gid_table_setup_one(device);
937 	if (err)
938 		/* Allocated memory will be cleaned in the release function */
939 		return err;
940 
941 	for (p = 0; p <= rdma_end_port(device) - rdma_start_port(device); ++p)
942 		ib_cache_update(device, p + rdma_start_port(device));
943 
944 	INIT_IB_EVENT_HANDLER(&device->cache.event_handler,
945 			      device, ib_cache_event);
946 	err = ib_register_event_handler(&device->cache.event_handler);
947 	if (err)
948 		goto err;
949 
950 	return 0;
951 
952 err:
953 	gid_table_cleanup_one(device);
954 	return err;
955 }
956 
957 void ib_cache_release_one(struct ib_device *device)
958 {
959 	int p;
960 
961 	/*
962 	 * The release function frees all the cache elements.
963 	 * This function should be called as part of freeing
964 	 * all the device's resources when the cache could no
965 	 * longer be accessed.
966 	 */
967 	if (device->cache.pkey_cache)
968 		for (p = 0;
969 		     p <= rdma_end_port(device) - rdma_start_port(device); ++p)
970 			kfree(device->cache.pkey_cache[p]);
971 
972 	gid_table_release_one(device);
973 	kfree(device->cache.pkey_cache);
974 	kfree(device->cache.lmc_cache);
975 }
976 
977 void ib_cache_cleanup_one(struct ib_device *device)
978 {
979 	/* The cleanup function unregisters the event handler,
980 	 * waits for all in-progress workqueue elements and cleans
981 	 * up the GID cache. This function should be called after
982 	 * the device was removed from the devices list and all
983 	 * clients were removed, so the cache exists but is
984 	 * non-functional and shouldn't be updated anymore.
985 	 */
986 	ib_unregister_event_handler(&device->cache.event_handler);
987 	flush_workqueue(ib_wq);
988 	gid_table_cleanup_one(device);
989 }
990 
991 void __init ib_cache_setup(void)
992 {
993 	roce_gid_mgmt_init();
994 }
995 
996 void __exit ib_cache_cleanup(void)
997 {
998 	roce_gid_mgmt_cleanup();
999 }
1000