xref: /openbmc/linux/drivers/infiniband/hw/mlx4/cm.c (revision 9cfc5c90)
1 /*
2  * Copyright (c) 2012 Mellanox Technologies. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 
33 #include <rdma/ib_mad.h>
34 
35 #include <linux/mlx4/cmd.h>
36 #include <linux/rbtree.h>
37 #include <linux/idr.h>
38 #include <rdma/ib_cm.h>
39 
40 #include "mlx4_ib.h"
41 
42 #define CM_CLEANUP_CACHE_TIMEOUT  (5 * HZ)
43 
44 struct id_map_entry {
45 	struct rb_node node;
46 
47 	u32 sl_cm_id;
48 	u32 pv_cm_id;
49 	int slave_id;
50 	int scheduled_delete;
51 	struct mlx4_ib_dev *dev;
52 
53 	struct list_head list;
54 	struct delayed_work timeout;
55 };
56 
57 struct cm_generic_msg {
58 	struct ib_mad_hdr hdr;
59 
60 	__be32 local_comm_id;
61 	__be32 remote_comm_id;
62 };
63 
64 struct cm_sidr_generic_msg {
65 	struct ib_mad_hdr hdr;
66 	__be32 request_id;
67 };
68 
69 struct cm_req_msg {
70 	unsigned char unused[0x60];
71 	union ib_gid primary_path_sgid;
72 };
73 
74 
75 static void set_local_comm_id(struct ib_mad *mad, u32 cm_id)
76 {
77 	if (mad->mad_hdr.attr_id == CM_SIDR_REQ_ATTR_ID) {
78 		struct cm_sidr_generic_msg *msg =
79 			(struct cm_sidr_generic_msg *)mad;
80 		msg->request_id = cpu_to_be32(cm_id);
81 	} else if (mad->mad_hdr.attr_id == CM_SIDR_REP_ATTR_ID) {
82 		pr_err("trying to set local_comm_id in SIDR_REP\n");
83 		return;
84 	} else {
85 		struct cm_generic_msg *msg = (struct cm_generic_msg *)mad;
86 		msg->local_comm_id = cpu_to_be32(cm_id);
87 	}
88 }
89 
90 static u32 get_local_comm_id(struct ib_mad *mad)
91 {
92 	if (mad->mad_hdr.attr_id == CM_SIDR_REQ_ATTR_ID) {
93 		struct cm_sidr_generic_msg *msg =
94 			(struct cm_sidr_generic_msg *)mad;
95 		return be32_to_cpu(msg->request_id);
96 	} else if (mad->mad_hdr.attr_id == CM_SIDR_REP_ATTR_ID) {
97 		pr_err("trying to set local_comm_id in SIDR_REP\n");
98 		return -1;
99 	} else {
100 		struct cm_generic_msg *msg = (struct cm_generic_msg *)mad;
101 		return be32_to_cpu(msg->local_comm_id);
102 	}
103 }
104 
105 static void set_remote_comm_id(struct ib_mad *mad, u32 cm_id)
106 {
107 	if (mad->mad_hdr.attr_id == CM_SIDR_REP_ATTR_ID) {
108 		struct cm_sidr_generic_msg *msg =
109 			(struct cm_sidr_generic_msg *)mad;
110 		msg->request_id = cpu_to_be32(cm_id);
111 	} else if (mad->mad_hdr.attr_id == CM_SIDR_REQ_ATTR_ID) {
112 		pr_err("trying to set remote_comm_id in SIDR_REQ\n");
113 		return;
114 	} else {
115 		struct cm_generic_msg *msg = (struct cm_generic_msg *)mad;
116 		msg->remote_comm_id = cpu_to_be32(cm_id);
117 	}
118 }
119 
120 static u32 get_remote_comm_id(struct ib_mad *mad)
121 {
122 	if (mad->mad_hdr.attr_id == CM_SIDR_REP_ATTR_ID) {
123 		struct cm_sidr_generic_msg *msg =
124 			(struct cm_sidr_generic_msg *)mad;
125 		return be32_to_cpu(msg->request_id);
126 	} else if (mad->mad_hdr.attr_id == CM_SIDR_REQ_ATTR_ID) {
127 		pr_err("trying to set remote_comm_id in SIDR_REQ\n");
128 		return -1;
129 	} else {
130 		struct cm_generic_msg *msg = (struct cm_generic_msg *)mad;
131 		return be32_to_cpu(msg->remote_comm_id);
132 	}
133 }
134 
135 static union ib_gid gid_from_req_msg(struct ib_device *ibdev, struct ib_mad *mad)
136 {
137 	struct cm_req_msg *msg = (struct cm_req_msg *)mad;
138 
139 	return msg->primary_path_sgid;
140 }
141 
142 /* Lock should be taken before called */
143 static struct id_map_entry *
144 id_map_find_by_sl_id(struct ib_device *ibdev, u32 slave_id, u32 sl_cm_id)
145 {
146 	struct rb_root *sl_id_map = &to_mdev(ibdev)->sriov.sl_id_map;
147 	struct rb_node *node = sl_id_map->rb_node;
148 
149 	while (node) {
150 		struct id_map_entry *id_map_entry =
151 			rb_entry(node, struct id_map_entry, node);
152 
153 		if (id_map_entry->sl_cm_id > sl_cm_id)
154 			node = node->rb_left;
155 		else if (id_map_entry->sl_cm_id < sl_cm_id)
156 			node = node->rb_right;
157 		else if (id_map_entry->slave_id > slave_id)
158 			node = node->rb_left;
159 		else if (id_map_entry->slave_id < slave_id)
160 			node = node->rb_right;
161 		else
162 			return id_map_entry;
163 	}
164 	return NULL;
165 }
166 
167 static void id_map_ent_timeout(struct work_struct *work)
168 {
169 	struct delayed_work *delay = to_delayed_work(work);
170 	struct id_map_entry *ent = container_of(delay, struct id_map_entry, timeout);
171 	struct id_map_entry *db_ent, *found_ent;
172 	struct mlx4_ib_dev *dev = ent->dev;
173 	struct mlx4_ib_sriov *sriov = &dev->sriov;
174 	struct rb_root *sl_id_map = &sriov->sl_id_map;
175 	int pv_id = (int) ent->pv_cm_id;
176 
177 	spin_lock(&sriov->id_map_lock);
178 	db_ent = (struct id_map_entry *)idr_find(&sriov->pv_id_table, pv_id);
179 	if (!db_ent)
180 		goto out;
181 	found_ent = id_map_find_by_sl_id(&dev->ib_dev, ent->slave_id, ent->sl_cm_id);
182 	if (found_ent && found_ent == ent)
183 		rb_erase(&found_ent->node, sl_id_map);
184 	idr_remove(&sriov->pv_id_table, pv_id);
185 
186 out:
187 	list_del(&ent->list);
188 	spin_unlock(&sriov->id_map_lock);
189 	kfree(ent);
190 }
191 
192 static void id_map_find_del(struct ib_device *ibdev, int pv_cm_id)
193 {
194 	struct mlx4_ib_sriov *sriov = &to_mdev(ibdev)->sriov;
195 	struct rb_root *sl_id_map = &sriov->sl_id_map;
196 	struct id_map_entry *ent, *found_ent;
197 
198 	spin_lock(&sriov->id_map_lock);
199 	ent = (struct id_map_entry *)idr_find(&sriov->pv_id_table, pv_cm_id);
200 	if (!ent)
201 		goto out;
202 	found_ent = id_map_find_by_sl_id(ibdev, ent->slave_id, ent->sl_cm_id);
203 	if (found_ent && found_ent == ent)
204 		rb_erase(&found_ent->node, sl_id_map);
205 	idr_remove(&sriov->pv_id_table, pv_cm_id);
206 out:
207 	spin_unlock(&sriov->id_map_lock);
208 }
209 
210 static void sl_id_map_add(struct ib_device *ibdev, struct id_map_entry *new)
211 {
212 	struct rb_root *sl_id_map = &to_mdev(ibdev)->sriov.sl_id_map;
213 	struct rb_node **link = &sl_id_map->rb_node, *parent = NULL;
214 	struct id_map_entry *ent;
215 	int slave_id = new->slave_id;
216 	int sl_cm_id = new->sl_cm_id;
217 
218 	ent = id_map_find_by_sl_id(ibdev, slave_id, sl_cm_id);
219 	if (ent) {
220 		pr_debug("overriding existing sl_id_map entry (cm_id = %x)\n",
221 			 sl_cm_id);
222 
223 		rb_replace_node(&ent->node, &new->node, sl_id_map);
224 		return;
225 	}
226 
227 	/* Go to the bottom of the tree */
228 	while (*link) {
229 		parent = *link;
230 		ent = rb_entry(parent, struct id_map_entry, node);
231 
232 		if (ent->sl_cm_id > sl_cm_id || (ent->sl_cm_id == sl_cm_id && ent->slave_id > slave_id))
233 			link = &(*link)->rb_left;
234 		else
235 			link = &(*link)->rb_right;
236 	}
237 
238 	rb_link_node(&new->node, parent, link);
239 	rb_insert_color(&new->node, sl_id_map);
240 }
241 
242 static struct id_map_entry *
243 id_map_alloc(struct ib_device *ibdev, int slave_id, u32 sl_cm_id)
244 {
245 	int ret;
246 	struct id_map_entry *ent;
247 	struct mlx4_ib_sriov *sriov = &to_mdev(ibdev)->sriov;
248 
249 	ent = kmalloc(sizeof (struct id_map_entry), GFP_KERNEL);
250 	if (!ent) {
251 		mlx4_ib_warn(ibdev, "Couldn't allocate id cache entry - out of memory\n");
252 		return ERR_PTR(-ENOMEM);
253 	}
254 
255 	ent->sl_cm_id = sl_cm_id;
256 	ent->slave_id = slave_id;
257 	ent->scheduled_delete = 0;
258 	ent->dev = to_mdev(ibdev);
259 	INIT_DELAYED_WORK(&ent->timeout, id_map_ent_timeout);
260 
261 	idr_preload(GFP_KERNEL);
262 	spin_lock(&to_mdev(ibdev)->sriov.id_map_lock);
263 
264 	ret = idr_alloc_cyclic(&sriov->pv_id_table, ent, 0, 0, GFP_NOWAIT);
265 	if (ret >= 0) {
266 		ent->pv_cm_id = (u32)ret;
267 		sl_id_map_add(ibdev, ent);
268 		list_add_tail(&ent->list, &sriov->cm_list);
269 	}
270 
271 	spin_unlock(&sriov->id_map_lock);
272 	idr_preload_end();
273 
274 	if (ret >= 0)
275 		return ent;
276 
277 	/*error flow*/
278 	kfree(ent);
279 	mlx4_ib_warn(ibdev, "No more space in the idr (err:0x%x)\n", ret);
280 	return ERR_PTR(-ENOMEM);
281 }
282 
283 static struct id_map_entry *
284 id_map_get(struct ib_device *ibdev, int *pv_cm_id, int sl_cm_id, int slave_id)
285 {
286 	struct id_map_entry *ent;
287 	struct mlx4_ib_sriov *sriov = &to_mdev(ibdev)->sriov;
288 
289 	spin_lock(&sriov->id_map_lock);
290 	if (*pv_cm_id == -1) {
291 		ent = id_map_find_by_sl_id(ibdev, sl_cm_id, slave_id);
292 		if (ent)
293 			*pv_cm_id = (int) ent->pv_cm_id;
294 	} else
295 		ent = (struct id_map_entry *)idr_find(&sriov->pv_id_table, *pv_cm_id);
296 	spin_unlock(&sriov->id_map_lock);
297 
298 	return ent;
299 }
300 
301 static void schedule_delayed(struct ib_device *ibdev, struct id_map_entry *id)
302 {
303 	struct mlx4_ib_sriov *sriov = &to_mdev(ibdev)->sriov;
304 	unsigned long flags;
305 
306 	spin_lock(&sriov->id_map_lock);
307 	spin_lock_irqsave(&sriov->going_down_lock, flags);
308 	/*make sure that there is no schedule inside the scheduled work.*/
309 	if (!sriov->is_going_down) {
310 		id->scheduled_delete = 1;
311 		schedule_delayed_work(&id->timeout, CM_CLEANUP_CACHE_TIMEOUT);
312 	}
313 	spin_unlock_irqrestore(&sriov->going_down_lock, flags);
314 	spin_unlock(&sriov->id_map_lock);
315 }
316 
317 int mlx4_ib_multiplex_cm_handler(struct ib_device *ibdev, int port, int slave_id,
318 		struct ib_mad *mad)
319 {
320 	struct id_map_entry *id;
321 	u32 sl_cm_id;
322 	int pv_cm_id = -1;
323 
324 	if (mad->mad_hdr.attr_id == CM_REQ_ATTR_ID ||
325 			mad->mad_hdr.attr_id == CM_REP_ATTR_ID ||
326 			mad->mad_hdr.attr_id == CM_SIDR_REQ_ATTR_ID) {
327 		sl_cm_id = get_local_comm_id(mad);
328 		id = id_map_alloc(ibdev, slave_id, sl_cm_id);
329 		if (IS_ERR(id)) {
330 			mlx4_ib_warn(ibdev, "%s: id{slave: %d, sl_cm_id: 0x%x} Failed to id_map_alloc\n",
331 				__func__, slave_id, sl_cm_id);
332 			return PTR_ERR(id);
333 		}
334 	} else if (mad->mad_hdr.attr_id == CM_REJ_ATTR_ID ||
335 		   mad->mad_hdr.attr_id == CM_SIDR_REP_ATTR_ID) {
336 		return 0;
337 	} else {
338 		sl_cm_id = get_local_comm_id(mad);
339 		id = id_map_get(ibdev, &pv_cm_id, slave_id, sl_cm_id);
340 	}
341 
342 	if (!id) {
343 		pr_debug("id{slave: %d, sl_cm_id: 0x%x} is NULL!\n",
344 			 slave_id, sl_cm_id);
345 		return -EINVAL;
346 	}
347 
348 	set_local_comm_id(mad, id->pv_cm_id);
349 
350 	if (mad->mad_hdr.attr_id == CM_DREQ_ATTR_ID)
351 		schedule_delayed(ibdev, id);
352 	else if (mad->mad_hdr.attr_id == CM_DREP_ATTR_ID)
353 		id_map_find_del(ibdev, pv_cm_id);
354 
355 	return 0;
356 }
357 
358 int mlx4_ib_demux_cm_handler(struct ib_device *ibdev, int port, int *slave,
359 			     struct ib_mad *mad)
360 {
361 	u32 pv_cm_id;
362 	struct id_map_entry *id;
363 
364 	if (mad->mad_hdr.attr_id == CM_REQ_ATTR_ID ||
365 	    mad->mad_hdr.attr_id == CM_SIDR_REQ_ATTR_ID) {
366 		union ib_gid gid;
367 
368 		if (!slave)
369 			return 0;
370 
371 		gid = gid_from_req_msg(ibdev, mad);
372 		*slave = mlx4_ib_find_real_gid(ibdev, port, gid.global.interface_id);
373 		if (*slave < 0) {
374 			mlx4_ib_warn(ibdev, "failed matching slave_id by gid (0x%llx)\n",
375 				     be64_to_cpu(gid.global.interface_id));
376 			return -ENOENT;
377 		}
378 		return 0;
379 	}
380 
381 	pv_cm_id = get_remote_comm_id(mad);
382 	id = id_map_get(ibdev, (int *)&pv_cm_id, -1, -1);
383 
384 	if (!id) {
385 		pr_debug("Couldn't find an entry for pv_cm_id 0x%x\n", pv_cm_id);
386 		return -ENOENT;
387 	}
388 
389 	if (slave)
390 		*slave = id->slave_id;
391 	set_remote_comm_id(mad, id->sl_cm_id);
392 
393 	if (mad->mad_hdr.attr_id == CM_DREQ_ATTR_ID)
394 		schedule_delayed(ibdev, id);
395 	else if (mad->mad_hdr.attr_id == CM_REJ_ATTR_ID ||
396 			mad->mad_hdr.attr_id == CM_DREP_ATTR_ID) {
397 		id_map_find_del(ibdev, (int) pv_cm_id);
398 	}
399 
400 	return 0;
401 }
402 
403 void mlx4_ib_cm_paravirt_init(struct mlx4_ib_dev *dev)
404 {
405 	spin_lock_init(&dev->sriov.id_map_lock);
406 	INIT_LIST_HEAD(&dev->sriov.cm_list);
407 	dev->sriov.sl_id_map = RB_ROOT;
408 	idr_init(&dev->sriov.pv_id_table);
409 }
410 
411 /* slave = -1 ==> all slaves */
412 /* TBD -- call paravirt clean for single slave.  Need for slave RESET event */
413 void mlx4_ib_cm_paravirt_clean(struct mlx4_ib_dev *dev, int slave)
414 {
415 	struct mlx4_ib_sriov *sriov = &dev->sriov;
416 	struct rb_root *sl_id_map = &sriov->sl_id_map;
417 	struct list_head lh;
418 	struct rb_node *nd;
419 	int need_flush = 1;
420 	struct id_map_entry *map, *tmp_map;
421 	/* cancel all delayed work queue entries */
422 	INIT_LIST_HEAD(&lh);
423 	spin_lock(&sriov->id_map_lock);
424 	list_for_each_entry_safe(map, tmp_map, &dev->sriov.cm_list, list) {
425 		if (slave < 0 || slave == map->slave_id) {
426 			if (map->scheduled_delete)
427 				need_flush &= !!cancel_delayed_work(&map->timeout);
428 		}
429 	}
430 
431 	spin_unlock(&sriov->id_map_lock);
432 
433 	if (!need_flush)
434 		flush_scheduled_work(); /* make sure all timers were flushed */
435 
436 	/* now, remove all leftover entries from databases*/
437 	spin_lock(&sriov->id_map_lock);
438 	if (slave < 0) {
439 		while (rb_first(sl_id_map)) {
440 			struct id_map_entry *ent =
441 				rb_entry(rb_first(sl_id_map),
442 					 struct id_map_entry, node);
443 
444 			rb_erase(&ent->node, sl_id_map);
445 			idr_remove(&sriov->pv_id_table, (int) ent->pv_cm_id);
446 		}
447 		list_splice_init(&dev->sriov.cm_list, &lh);
448 	} else {
449 		/* first, move nodes belonging to slave to db remove list */
450 		nd = rb_first(sl_id_map);
451 		while (nd) {
452 			struct id_map_entry *ent =
453 				rb_entry(nd, struct id_map_entry, node);
454 			nd = rb_next(nd);
455 			if (ent->slave_id == slave)
456 				list_move_tail(&ent->list, &lh);
457 		}
458 		/* remove those nodes from databases */
459 		list_for_each_entry_safe(map, tmp_map, &lh, list) {
460 			rb_erase(&map->node, sl_id_map);
461 			idr_remove(&sriov->pv_id_table, (int) map->pv_cm_id);
462 		}
463 
464 		/* add remaining nodes from cm_list */
465 		list_for_each_entry_safe(map, tmp_map, &dev->sriov.cm_list, list) {
466 			if (slave == map->slave_id)
467 				list_move_tail(&map->list, &lh);
468 		}
469 	}
470 
471 	spin_unlock(&sriov->id_map_lock);
472 
473 	/* free any map entries left behind due to cancel_delayed_work above */
474 	list_for_each_entry_safe(map, tmp_map, &lh, list) {
475 		list_del(&map->list);
476 		kfree(map);
477 	}
478 }
479