xref: /openbmc/linux/drivers/infiniband/hw/mlx4/cm.c (revision b830f94f)
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  (30 * 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 *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 
176 	spin_lock(&sriov->id_map_lock);
177 	if (!xa_erase(&sriov->pv_id_table, ent->pv_cm_id))
178 		goto out;
179 	found_ent = id_map_find_by_sl_id(&dev->ib_dev, ent->slave_id, ent->sl_cm_id);
180 	if (found_ent && found_ent == ent)
181 		rb_erase(&found_ent->node, sl_id_map);
182 
183 out:
184 	list_del(&ent->list);
185 	spin_unlock(&sriov->id_map_lock);
186 	kfree(ent);
187 }
188 
189 static void id_map_find_del(struct ib_device *ibdev, int pv_cm_id)
190 {
191 	struct mlx4_ib_sriov *sriov = &to_mdev(ibdev)->sriov;
192 	struct rb_root *sl_id_map = &sriov->sl_id_map;
193 	struct id_map_entry *ent, *found_ent;
194 
195 	spin_lock(&sriov->id_map_lock);
196 	ent = xa_erase(&sriov->pv_id_table, pv_cm_id);
197 	if (!ent)
198 		goto out;
199 	found_ent = id_map_find_by_sl_id(ibdev, ent->slave_id, ent->sl_cm_id);
200 	if (found_ent && found_ent == ent)
201 		rb_erase(&found_ent->node, sl_id_map);
202 out:
203 	spin_unlock(&sriov->id_map_lock);
204 }
205 
206 static void sl_id_map_add(struct ib_device *ibdev, struct id_map_entry *new)
207 {
208 	struct rb_root *sl_id_map = &to_mdev(ibdev)->sriov.sl_id_map;
209 	struct rb_node **link = &sl_id_map->rb_node, *parent = NULL;
210 	struct id_map_entry *ent;
211 	int slave_id = new->slave_id;
212 	int sl_cm_id = new->sl_cm_id;
213 
214 	ent = id_map_find_by_sl_id(ibdev, slave_id, sl_cm_id);
215 	if (ent) {
216 		pr_debug("overriding existing sl_id_map entry (cm_id = %x)\n",
217 			 sl_cm_id);
218 
219 		rb_replace_node(&ent->node, &new->node, sl_id_map);
220 		return;
221 	}
222 
223 	/* Go to the bottom of the tree */
224 	while (*link) {
225 		parent = *link;
226 		ent = rb_entry(parent, struct id_map_entry, node);
227 
228 		if (ent->sl_cm_id > sl_cm_id || (ent->sl_cm_id == sl_cm_id && ent->slave_id > slave_id))
229 			link = &(*link)->rb_left;
230 		else
231 			link = &(*link)->rb_right;
232 	}
233 
234 	rb_link_node(&new->node, parent, link);
235 	rb_insert_color(&new->node, sl_id_map);
236 }
237 
238 static struct id_map_entry *
239 id_map_alloc(struct ib_device *ibdev, int slave_id, u32 sl_cm_id)
240 {
241 	int ret;
242 	struct id_map_entry *ent;
243 	struct mlx4_ib_sriov *sriov = &to_mdev(ibdev)->sriov;
244 
245 	ent = kmalloc(sizeof (struct id_map_entry), GFP_KERNEL);
246 	if (!ent)
247 		return ERR_PTR(-ENOMEM);
248 
249 	ent->sl_cm_id = sl_cm_id;
250 	ent->slave_id = slave_id;
251 	ent->scheduled_delete = 0;
252 	ent->dev = to_mdev(ibdev);
253 	INIT_DELAYED_WORK(&ent->timeout, id_map_ent_timeout);
254 
255 	ret = xa_alloc_cyclic(&sriov->pv_id_table, &ent->pv_cm_id, ent,
256 			xa_limit_32b, &sriov->pv_id_next, GFP_KERNEL);
257 	if (ret >= 0) {
258 		spin_lock(&sriov->id_map_lock);
259 		sl_id_map_add(ibdev, ent);
260 		list_add_tail(&ent->list, &sriov->cm_list);
261 		spin_unlock(&sriov->id_map_lock);
262 		return ent;
263 	}
264 
265 	/*error flow*/
266 	kfree(ent);
267 	mlx4_ib_warn(ibdev, "Allocation failed (err:0x%x)\n", ret);
268 	return ERR_PTR(-ENOMEM);
269 }
270 
271 static struct id_map_entry *
272 id_map_get(struct ib_device *ibdev, int *pv_cm_id, int slave_id, int sl_cm_id)
273 {
274 	struct id_map_entry *ent;
275 	struct mlx4_ib_sriov *sriov = &to_mdev(ibdev)->sriov;
276 
277 	spin_lock(&sriov->id_map_lock);
278 	if (*pv_cm_id == -1) {
279 		ent = id_map_find_by_sl_id(ibdev, slave_id, sl_cm_id);
280 		if (ent)
281 			*pv_cm_id = (int) ent->pv_cm_id;
282 	} else
283 		ent = xa_load(&sriov->pv_id_table, *pv_cm_id);
284 	spin_unlock(&sriov->id_map_lock);
285 
286 	return ent;
287 }
288 
289 static void schedule_delayed(struct ib_device *ibdev, struct id_map_entry *id)
290 {
291 	struct mlx4_ib_sriov *sriov = &to_mdev(ibdev)->sriov;
292 	unsigned long flags;
293 
294 	spin_lock(&sriov->id_map_lock);
295 	spin_lock_irqsave(&sriov->going_down_lock, flags);
296 	/*make sure that there is no schedule inside the scheduled work.*/
297 	if (!sriov->is_going_down) {
298 		id->scheduled_delete = 1;
299 		schedule_delayed_work(&id->timeout, CM_CLEANUP_CACHE_TIMEOUT);
300 	}
301 	spin_unlock_irqrestore(&sriov->going_down_lock, flags);
302 	spin_unlock(&sriov->id_map_lock);
303 }
304 
305 int mlx4_ib_multiplex_cm_handler(struct ib_device *ibdev, int port, int slave_id,
306 		struct ib_mad *mad)
307 {
308 	struct id_map_entry *id;
309 	u32 sl_cm_id;
310 	int pv_cm_id = -1;
311 
312 	if (mad->mad_hdr.attr_id == CM_REQ_ATTR_ID ||
313 			mad->mad_hdr.attr_id == CM_REP_ATTR_ID ||
314 			mad->mad_hdr.attr_id == CM_SIDR_REQ_ATTR_ID) {
315 		sl_cm_id = get_local_comm_id(mad);
316 		id = id_map_get(ibdev, &pv_cm_id, slave_id, sl_cm_id);
317 		if (id)
318 			goto cont;
319 		id = id_map_alloc(ibdev, slave_id, sl_cm_id);
320 		if (IS_ERR(id)) {
321 			mlx4_ib_warn(ibdev, "%s: id{slave: %d, sl_cm_id: 0x%x} Failed to id_map_alloc\n",
322 				__func__, slave_id, sl_cm_id);
323 			return PTR_ERR(id);
324 		}
325 	} else if (mad->mad_hdr.attr_id == CM_REJ_ATTR_ID ||
326 		   mad->mad_hdr.attr_id == CM_SIDR_REP_ATTR_ID) {
327 		return 0;
328 	} else {
329 		sl_cm_id = get_local_comm_id(mad);
330 		id = id_map_get(ibdev, &pv_cm_id, slave_id, sl_cm_id);
331 	}
332 
333 	if (!id) {
334 		pr_debug("id{slave: %d, sl_cm_id: 0x%x} is NULL!\n",
335 			 slave_id, sl_cm_id);
336 		return -EINVAL;
337 	}
338 
339 cont:
340 	set_local_comm_id(mad, id->pv_cm_id);
341 
342 	if (mad->mad_hdr.attr_id == CM_DREQ_ATTR_ID)
343 		schedule_delayed(ibdev, id);
344 	else if (mad->mad_hdr.attr_id == CM_DREP_ATTR_ID)
345 		id_map_find_del(ibdev, pv_cm_id);
346 
347 	return 0;
348 }
349 
350 int mlx4_ib_demux_cm_handler(struct ib_device *ibdev, int port, int *slave,
351 			     struct ib_mad *mad)
352 {
353 	u32 pv_cm_id;
354 	struct id_map_entry *id;
355 
356 	if (mad->mad_hdr.attr_id == CM_REQ_ATTR_ID ||
357 	    mad->mad_hdr.attr_id == CM_SIDR_REQ_ATTR_ID) {
358 		union ib_gid gid;
359 
360 		if (!slave)
361 			return 0;
362 
363 		gid = gid_from_req_msg(ibdev, mad);
364 		*slave = mlx4_ib_find_real_gid(ibdev, port, gid.global.interface_id);
365 		if (*slave < 0) {
366 			mlx4_ib_warn(ibdev, "failed matching slave_id by gid (0x%llx)\n",
367 				     be64_to_cpu(gid.global.interface_id));
368 			return -ENOENT;
369 		}
370 		return 0;
371 	}
372 
373 	pv_cm_id = get_remote_comm_id(mad);
374 	id = id_map_get(ibdev, (int *)&pv_cm_id, -1, -1);
375 
376 	if (!id) {
377 		pr_debug("Couldn't find an entry for pv_cm_id 0x%x\n", pv_cm_id);
378 		return -ENOENT;
379 	}
380 
381 	if (slave)
382 		*slave = id->slave_id;
383 	set_remote_comm_id(mad, id->sl_cm_id);
384 
385 	if (mad->mad_hdr.attr_id == CM_DREQ_ATTR_ID)
386 		schedule_delayed(ibdev, id);
387 	else if (mad->mad_hdr.attr_id == CM_REJ_ATTR_ID ||
388 			mad->mad_hdr.attr_id == CM_DREP_ATTR_ID) {
389 		id_map_find_del(ibdev, (int) pv_cm_id);
390 	}
391 
392 	return 0;
393 }
394 
395 void mlx4_ib_cm_paravirt_init(struct mlx4_ib_dev *dev)
396 {
397 	spin_lock_init(&dev->sriov.id_map_lock);
398 	INIT_LIST_HEAD(&dev->sriov.cm_list);
399 	dev->sriov.sl_id_map = RB_ROOT;
400 	xa_init_flags(&dev->sriov.pv_id_table, XA_FLAGS_ALLOC);
401 }
402 
403 /* slave = -1 ==> all slaves */
404 /* TBD -- call paravirt clean for single slave.  Need for slave RESET event */
405 void mlx4_ib_cm_paravirt_clean(struct mlx4_ib_dev *dev, int slave)
406 {
407 	struct mlx4_ib_sriov *sriov = &dev->sriov;
408 	struct rb_root *sl_id_map = &sriov->sl_id_map;
409 	struct list_head lh;
410 	struct rb_node *nd;
411 	int need_flush = 0;
412 	struct id_map_entry *map, *tmp_map;
413 	/* cancel all delayed work queue entries */
414 	INIT_LIST_HEAD(&lh);
415 	spin_lock(&sriov->id_map_lock);
416 	list_for_each_entry_safe(map, tmp_map, &dev->sriov.cm_list, list) {
417 		if (slave < 0 || slave == map->slave_id) {
418 			if (map->scheduled_delete)
419 				need_flush |= !cancel_delayed_work(&map->timeout);
420 		}
421 	}
422 
423 	spin_unlock(&sriov->id_map_lock);
424 
425 	if (need_flush)
426 		flush_scheduled_work(); /* make sure all timers were flushed */
427 
428 	/* now, remove all leftover entries from databases*/
429 	spin_lock(&sriov->id_map_lock);
430 	if (slave < 0) {
431 		while (rb_first(sl_id_map)) {
432 			struct id_map_entry *ent =
433 				rb_entry(rb_first(sl_id_map),
434 					 struct id_map_entry, node);
435 
436 			rb_erase(&ent->node, sl_id_map);
437 			xa_erase(&sriov->pv_id_table, ent->pv_cm_id);
438 		}
439 		list_splice_init(&dev->sriov.cm_list, &lh);
440 	} else {
441 		/* first, move nodes belonging to slave to db remove list */
442 		nd = rb_first(sl_id_map);
443 		while (nd) {
444 			struct id_map_entry *ent =
445 				rb_entry(nd, struct id_map_entry, node);
446 			nd = rb_next(nd);
447 			if (ent->slave_id == slave)
448 				list_move_tail(&ent->list, &lh);
449 		}
450 		/* remove those nodes from databases */
451 		list_for_each_entry_safe(map, tmp_map, &lh, list) {
452 			rb_erase(&map->node, sl_id_map);
453 			xa_erase(&sriov->pv_id_table, map->pv_cm_id);
454 		}
455 
456 		/* add remaining nodes from cm_list */
457 		list_for_each_entry_safe(map, tmp_map, &dev->sriov.cm_list, list) {
458 			if (slave == map->slave_id)
459 				list_move_tail(&map->list, &lh);
460 		}
461 	}
462 
463 	spin_unlock(&sriov->id_map_lock);
464 
465 	/* free any map entries left behind due to cancel_delayed_work above */
466 	list_for_each_entry_safe(map, tmp_map, &lh, list) {
467 		list_del(&map->list);
468 		kfree(map);
469 	}
470 }
471