xref: /openbmc/linux/drivers/infiniband/hw/mlx4/mcg.c (revision 8730046c)
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 #include <rdma/ib_smi.h>
35 #include <rdma/ib_cache.h>
36 #include <rdma/ib_sa.h>
37 
38 #include <linux/mlx4/cmd.h>
39 #include <linux/rbtree.h>
40 #include <linux/delay.h>
41 
42 #include "mlx4_ib.h"
43 
44 #define MAX_VFS		80
45 #define MAX_PEND_REQS_PER_FUNC 4
46 #define MAD_TIMEOUT_MS	2000
47 
48 #define mcg_warn(fmt, arg...)	pr_warn("MCG WARNING: " fmt, ##arg)
49 #define mcg_error(fmt, arg...)	pr_err(fmt, ##arg)
50 #define mcg_warn_group(group, format, arg...) \
51 	pr_warn("%s-%d: %16s (port %d): WARNING: " format, __func__, __LINE__,\
52 	(group)->name, group->demux->port, ## arg)
53 
54 #define mcg_debug_group(group, format, arg...) \
55 	pr_debug("%s-%d: %16s (port %d): WARNING: " format, __func__, __LINE__,\
56 		 (group)->name, (group)->demux->port, ## arg)
57 
58 #define mcg_error_group(group, format, arg...) \
59 	pr_err("  %16s: " format, (group)->name, ## arg)
60 
61 
62 static union ib_gid mgid0;
63 
64 static struct workqueue_struct *clean_wq;
65 
66 enum mcast_state {
67 	MCAST_NOT_MEMBER = 0,
68 	MCAST_MEMBER,
69 };
70 
71 enum mcast_group_state {
72 	MCAST_IDLE,
73 	MCAST_JOIN_SENT,
74 	MCAST_LEAVE_SENT,
75 	MCAST_RESP_READY
76 };
77 
78 struct mcast_member {
79 	enum mcast_state state;
80 	uint8_t			join_state;
81 	int			num_pend_reqs;
82 	struct list_head	pending;
83 };
84 
85 struct ib_sa_mcmember_data {
86 	union ib_gid	mgid;
87 	union ib_gid	port_gid;
88 	__be32		qkey;
89 	__be16		mlid;
90 	u8		mtusel_mtu;
91 	u8		tclass;
92 	__be16		pkey;
93 	u8		ratesel_rate;
94 	u8		lifetmsel_lifetm;
95 	__be32		sl_flowlabel_hoplimit;
96 	u8		scope_join_state;
97 	u8		proxy_join;
98 	u8		reserved[2];
99 } __packed __aligned(4);
100 
101 struct mcast_group {
102 	struct ib_sa_mcmember_data rec;
103 	struct rb_node		node;
104 	struct list_head	mgid0_list;
105 	struct mlx4_ib_demux_ctx *demux;
106 	struct mcast_member	func[MAX_VFS];
107 	struct mutex		lock;
108 	struct work_struct	work;
109 	struct list_head	pending_list;
110 	int			members[3];
111 	enum mcast_group_state	state;
112 	enum mcast_group_state	prev_state;
113 	struct ib_sa_mad	response_sa_mad;
114 	__be64			last_req_tid;
115 
116 	char			name[33]; /* MGID string */
117 	struct device_attribute	dentry;
118 
119 	/* refcount is the reference count for the following:
120 	   1. Each queued request
121 	   2. Each invocation of the worker thread
122 	   3. Membership of the port at the SA
123 	*/
124 	atomic_t		refcount;
125 
126 	/* delayed work to clean pending SM request */
127 	struct delayed_work	timeout_work;
128 	struct list_head	cleanup_list;
129 };
130 
131 struct mcast_req {
132 	int			func;
133 	struct ib_sa_mad	sa_mad;
134 	struct list_head	group_list;
135 	struct list_head	func_list;
136 	struct mcast_group	*group;
137 	int			clean;
138 };
139 
140 
141 #define safe_atomic_dec(ref) \
142 	do {\
143 		if (atomic_dec_and_test(ref)) \
144 			mcg_warn_group(group, "did not expect to reach zero\n"); \
145 	} while (0)
146 
147 static const char *get_state_string(enum mcast_group_state state)
148 {
149 	switch (state) {
150 	case MCAST_IDLE:
151 		return "MCAST_IDLE";
152 	case MCAST_JOIN_SENT:
153 		return "MCAST_JOIN_SENT";
154 	case MCAST_LEAVE_SENT:
155 		return "MCAST_LEAVE_SENT";
156 	case MCAST_RESP_READY:
157 		return "MCAST_RESP_READY";
158 	}
159 	return "Invalid State";
160 }
161 
162 static struct mcast_group *mcast_find(struct mlx4_ib_demux_ctx *ctx,
163 				      union ib_gid *mgid)
164 {
165 	struct rb_node *node = ctx->mcg_table.rb_node;
166 	struct mcast_group *group;
167 	int ret;
168 
169 	while (node) {
170 		group = rb_entry(node, struct mcast_group, node);
171 		ret = memcmp(mgid->raw, group->rec.mgid.raw, sizeof *mgid);
172 		if (!ret)
173 			return group;
174 
175 		if (ret < 0)
176 			node = node->rb_left;
177 		else
178 			node = node->rb_right;
179 	}
180 	return NULL;
181 }
182 
183 static struct mcast_group *mcast_insert(struct mlx4_ib_demux_ctx *ctx,
184 					struct mcast_group *group)
185 {
186 	struct rb_node **link = &ctx->mcg_table.rb_node;
187 	struct rb_node *parent = NULL;
188 	struct mcast_group *cur_group;
189 	int ret;
190 
191 	while (*link) {
192 		parent = *link;
193 		cur_group = rb_entry(parent, struct mcast_group, node);
194 
195 		ret = memcmp(group->rec.mgid.raw, cur_group->rec.mgid.raw,
196 			     sizeof group->rec.mgid);
197 		if (ret < 0)
198 			link = &(*link)->rb_left;
199 		else if (ret > 0)
200 			link = &(*link)->rb_right;
201 		else
202 			return cur_group;
203 	}
204 	rb_link_node(&group->node, parent, link);
205 	rb_insert_color(&group->node, &ctx->mcg_table);
206 	return NULL;
207 }
208 
209 static int send_mad_to_wire(struct mlx4_ib_demux_ctx *ctx, struct ib_mad *mad)
210 {
211 	struct mlx4_ib_dev *dev = ctx->dev;
212 	struct ib_ah_attr	ah_attr;
213 	unsigned long flags;
214 
215 	spin_lock_irqsave(&dev->sm_lock, flags);
216 	if (!dev->sm_ah[ctx->port - 1]) {
217 		/* port is not yet Active, sm_ah not ready */
218 		spin_unlock_irqrestore(&dev->sm_lock, flags);
219 		return -EAGAIN;
220 	}
221 	mlx4_ib_query_ah(dev->sm_ah[ctx->port - 1], &ah_attr);
222 	spin_unlock_irqrestore(&dev->sm_lock, flags);
223 	return mlx4_ib_send_to_wire(dev, mlx4_master_func_num(dev->dev),
224 				    ctx->port, IB_QPT_GSI, 0, 1, IB_QP1_QKEY,
225 				    &ah_attr, NULL, 0xffff, mad);
226 }
227 
228 static int send_mad_to_slave(int slave, struct mlx4_ib_demux_ctx *ctx,
229 			     struct ib_mad *mad)
230 {
231 	struct mlx4_ib_dev *dev = ctx->dev;
232 	struct ib_mad_agent *agent = dev->send_agent[ctx->port - 1][1];
233 	struct ib_wc wc;
234 	struct ib_ah_attr ah_attr;
235 
236 	/* Our agent might not yet be registered when mads start to arrive */
237 	if (!agent)
238 		return -EAGAIN;
239 
240 	ib_query_ah(dev->sm_ah[ctx->port - 1], &ah_attr);
241 
242 	if (ib_find_cached_pkey(&dev->ib_dev, ctx->port, IB_DEFAULT_PKEY_FULL, &wc.pkey_index))
243 		return -EINVAL;
244 	wc.sl = 0;
245 	wc.dlid_path_bits = 0;
246 	wc.port_num = ctx->port;
247 	wc.slid = ah_attr.dlid;  /* opensm lid */
248 	wc.src_qp = 1;
249 	return mlx4_ib_send_to_slave(dev, slave, ctx->port, IB_QPT_GSI, &wc, NULL, mad);
250 }
251 
252 static int send_join_to_wire(struct mcast_group *group, struct ib_sa_mad *sa_mad)
253 {
254 	struct ib_sa_mad mad;
255 	struct ib_sa_mcmember_data *sa_mad_data = (struct ib_sa_mcmember_data *)&mad.data;
256 	int ret;
257 
258 	/* we rely on a mad request as arrived from a VF */
259 	memcpy(&mad, sa_mad, sizeof mad);
260 
261 	/* fix port GID to be the real one (slave 0) */
262 	sa_mad_data->port_gid.global.interface_id = group->demux->guid_cache[0];
263 
264 	/* assign our own TID */
265 	mad.mad_hdr.tid = mlx4_ib_get_new_demux_tid(group->demux);
266 	group->last_req_tid = mad.mad_hdr.tid; /* keep it for later validation */
267 
268 	ret = send_mad_to_wire(group->demux, (struct ib_mad *)&mad);
269 	/* set timeout handler */
270 	if (!ret) {
271 		/* calls mlx4_ib_mcg_timeout_handler */
272 		queue_delayed_work(group->demux->mcg_wq, &group->timeout_work,
273 				msecs_to_jiffies(MAD_TIMEOUT_MS));
274 	}
275 
276 	return ret;
277 }
278 
279 static int send_leave_to_wire(struct mcast_group *group, u8 join_state)
280 {
281 	struct ib_sa_mad mad;
282 	struct ib_sa_mcmember_data *sa_data = (struct ib_sa_mcmember_data *)&mad.data;
283 	int ret;
284 
285 	memset(&mad, 0, sizeof mad);
286 	mad.mad_hdr.base_version = 1;
287 	mad.mad_hdr.mgmt_class = IB_MGMT_CLASS_SUBN_ADM;
288 	mad.mad_hdr.class_version = 2;
289 	mad.mad_hdr.method = IB_SA_METHOD_DELETE;
290 	mad.mad_hdr.status = cpu_to_be16(0);
291 	mad.mad_hdr.class_specific = cpu_to_be16(0);
292 	mad.mad_hdr.tid = mlx4_ib_get_new_demux_tid(group->demux);
293 	group->last_req_tid = mad.mad_hdr.tid; /* keep it for later validation */
294 	mad.mad_hdr.attr_id = cpu_to_be16(IB_SA_ATTR_MC_MEMBER_REC);
295 	mad.mad_hdr.attr_mod = cpu_to_be32(0);
296 	mad.sa_hdr.sm_key = 0x0;
297 	mad.sa_hdr.attr_offset = cpu_to_be16(7);
298 	mad.sa_hdr.comp_mask = IB_SA_MCMEMBER_REC_MGID |
299 		IB_SA_MCMEMBER_REC_PORT_GID | IB_SA_MCMEMBER_REC_JOIN_STATE;
300 
301 	*sa_data = group->rec;
302 	sa_data->scope_join_state = join_state;
303 
304 	ret = send_mad_to_wire(group->demux, (struct ib_mad *)&mad);
305 	if (ret)
306 		group->state = MCAST_IDLE;
307 
308 	/* set timeout handler */
309 	if (!ret) {
310 		/* calls mlx4_ib_mcg_timeout_handler */
311 		queue_delayed_work(group->demux->mcg_wq, &group->timeout_work,
312 				msecs_to_jiffies(MAD_TIMEOUT_MS));
313 	}
314 
315 	return ret;
316 }
317 
318 static int send_reply_to_slave(int slave, struct mcast_group *group,
319 		struct ib_sa_mad *req_sa_mad, u16 status)
320 {
321 	struct ib_sa_mad mad;
322 	struct ib_sa_mcmember_data *sa_data = (struct ib_sa_mcmember_data *)&mad.data;
323 	struct ib_sa_mcmember_data *req_sa_data = (struct ib_sa_mcmember_data *)&req_sa_mad->data;
324 	int ret;
325 
326 	memset(&mad, 0, sizeof mad);
327 	mad.mad_hdr.base_version = 1;
328 	mad.mad_hdr.mgmt_class = IB_MGMT_CLASS_SUBN_ADM;
329 	mad.mad_hdr.class_version = 2;
330 	mad.mad_hdr.method = IB_MGMT_METHOD_GET_RESP;
331 	mad.mad_hdr.status = cpu_to_be16(status);
332 	mad.mad_hdr.class_specific = cpu_to_be16(0);
333 	mad.mad_hdr.tid = req_sa_mad->mad_hdr.tid;
334 	*(u8 *)&mad.mad_hdr.tid = 0; /* resetting tid to 0 */
335 	mad.mad_hdr.attr_id = cpu_to_be16(IB_SA_ATTR_MC_MEMBER_REC);
336 	mad.mad_hdr.attr_mod = cpu_to_be32(0);
337 	mad.sa_hdr.sm_key = req_sa_mad->sa_hdr.sm_key;
338 	mad.sa_hdr.attr_offset = cpu_to_be16(7);
339 	mad.sa_hdr.comp_mask = 0; /* ignored on responses, see IBTA spec */
340 
341 	*sa_data = group->rec;
342 
343 	/* reconstruct VF's requested join_state and port_gid */
344 	sa_data->scope_join_state &= 0xf0;
345 	sa_data->scope_join_state |= (group->func[slave].join_state & 0x0f);
346 	memcpy(&sa_data->port_gid, &req_sa_data->port_gid, sizeof req_sa_data->port_gid);
347 
348 	ret = send_mad_to_slave(slave, group->demux, (struct ib_mad *)&mad);
349 	return ret;
350 }
351 
352 static int check_selector(ib_sa_comp_mask comp_mask,
353 			  ib_sa_comp_mask selector_mask,
354 			  ib_sa_comp_mask value_mask,
355 			  u8 src_value, u8 dst_value)
356 {
357 	int err;
358 	u8 selector = dst_value >> 6;
359 	dst_value &= 0x3f;
360 	src_value &= 0x3f;
361 
362 	if (!(comp_mask & selector_mask) || !(comp_mask & value_mask))
363 		return 0;
364 
365 	switch (selector) {
366 	case IB_SA_GT:
367 		err = (src_value <= dst_value);
368 		break;
369 	case IB_SA_LT:
370 		err = (src_value >= dst_value);
371 		break;
372 	case IB_SA_EQ:
373 		err = (src_value != dst_value);
374 		break;
375 	default:
376 		err = 0;
377 		break;
378 	}
379 
380 	return err;
381 }
382 
383 static u16 cmp_rec(struct ib_sa_mcmember_data *src,
384 		   struct ib_sa_mcmember_data *dst, ib_sa_comp_mask comp_mask)
385 {
386 	/* src is group record, dst is request record */
387 	/* MGID must already match */
388 	/* Port_GID we always replace to our Port_GID, so it is a match */
389 
390 #define MAD_STATUS_REQ_INVALID 0x0200
391 	if (comp_mask & IB_SA_MCMEMBER_REC_QKEY && src->qkey != dst->qkey)
392 		return MAD_STATUS_REQ_INVALID;
393 	if (comp_mask & IB_SA_MCMEMBER_REC_MLID && src->mlid != dst->mlid)
394 		return MAD_STATUS_REQ_INVALID;
395 	if (check_selector(comp_mask, IB_SA_MCMEMBER_REC_MTU_SELECTOR,
396 				 IB_SA_MCMEMBER_REC_MTU,
397 				 src->mtusel_mtu, dst->mtusel_mtu))
398 		return MAD_STATUS_REQ_INVALID;
399 	if (comp_mask & IB_SA_MCMEMBER_REC_TRAFFIC_CLASS &&
400 	    src->tclass != dst->tclass)
401 		return MAD_STATUS_REQ_INVALID;
402 	if (comp_mask & IB_SA_MCMEMBER_REC_PKEY && src->pkey != dst->pkey)
403 		return MAD_STATUS_REQ_INVALID;
404 	if (check_selector(comp_mask, IB_SA_MCMEMBER_REC_RATE_SELECTOR,
405 				 IB_SA_MCMEMBER_REC_RATE,
406 				 src->ratesel_rate, dst->ratesel_rate))
407 		return MAD_STATUS_REQ_INVALID;
408 	if (check_selector(comp_mask,
409 				 IB_SA_MCMEMBER_REC_PACKET_LIFE_TIME_SELECTOR,
410 				 IB_SA_MCMEMBER_REC_PACKET_LIFE_TIME,
411 				 src->lifetmsel_lifetm, dst->lifetmsel_lifetm))
412 		return MAD_STATUS_REQ_INVALID;
413 	if (comp_mask & IB_SA_MCMEMBER_REC_SL &&
414 			(be32_to_cpu(src->sl_flowlabel_hoplimit) & 0xf0000000) !=
415 			(be32_to_cpu(dst->sl_flowlabel_hoplimit) & 0xf0000000))
416 		return MAD_STATUS_REQ_INVALID;
417 	if (comp_mask & IB_SA_MCMEMBER_REC_FLOW_LABEL &&
418 			(be32_to_cpu(src->sl_flowlabel_hoplimit) & 0x0fffff00) !=
419 			(be32_to_cpu(dst->sl_flowlabel_hoplimit) & 0x0fffff00))
420 		return MAD_STATUS_REQ_INVALID;
421 	if (comp_mask & IB_SA_MCMEMBER_REC_HOP_LIMIT &&
422 			(be32_to_cpu(src->sl_flowlabel_hoplimit) & 0x000000ff) !=
423 			(be32_to_cpu(dst->sl_flowlabel_hoplimit) & 0x000000ff))
424 		return MAD_STATUS_REQ_INVALID;
425 	if (comp_mask & IB_SA_MCMEMBER_REC_SCOPE &&
426 			(src->scope_join_state & 0xf0) !=
427 			(dst->scope_join_state & 0xf0))
428 		return MAD_STATUS_REQ_INVALID;
429 
430 	/* join_state checked separately, proxy_join ignored */
431 
432 	return 0;
433 }
434 
435 /* release group, return 1 if this was last release and group is destroyed
436  * timout work is canceled sync */
437 static int release_group(struct mcast_group *group, int from_timeout_handler)
438 {
439 	struct mlx4_ib_demux_ctx *ctx = group->demux;
440 	int nzgroup;
441 
442 	mutex_lock(&ctx->mcg_table_lock);
443 	mutex_lock(&group->lock);
444 	if (atomic_dec_and_test(&group->refcount)) {
445 		if (!from_timeout_handler) {
446 			if (group->state != MCAST_IDLE &&
447 			    !cancel_delayed_work(&group->timeout_work)) {
448 				atomic_inc(&group->refcount);
449 				mutex_unlock(&group->lock);
450 				mutex_unlock(&ctx->mcg_table_lock);
451 				return 0;
452 			}
453 		}
454 
455 		nzgroup = memcmp(&group->rec.mgid, &mgid0, sizeof mgid0);
456 		if (nzgroup)
457 			del_sysfs_port_mcg_attr(ctx->dev, ctx->port, &group->dentry.attr);
458 		if (!list_empty(&group->pending_list))
459 			mcg_warn_group(group, "releasing a group with non empty pending list\n");
460 		if (nzgroup)
461 			rb_erase(&group->node, &ctx->mcg_table);
462 		list_del_init(&group->mgid0_list);
463 		mutex_unlock(&group->lock);
464 		mutex_unlock(&ctx->mcg_table_lock);
465 		kfree(group);
466 		return 1;
467 	} else {
468 		mutex_unlock(&group->lock);
469 		mutex_unlock(&ctx->mcg_table_lock);
470 	}
471 	return 0;
472 }
473 
474 static void adjust_membership(struct mcast_group *group, u8 join_state, int inc)
475 {
476 	int i;
477 
478 	for (i = 0; i < 3; i++, join_state >>= 1)
479 		if (join_state & 0x1)
480 			group->members[i] += inc;
481 }
482 
483 static u8 get_leave_state(struct mcast_group *group)
484 {
485 	u8 leave_state = 0;
486 	int i;
487 
488 	for (i = 0; i < 3; i++)
489 		if (!group->members[i])
490 			leave_state |= (1 << i);
491 
492 	return leave_state & (group->rec.scope_join_state & 0xf);
493 }
494 
495 static int join_group(struct mcast_group *group, int slave, u8 join_mask)
496 {
497 	int ret = 0;
498 	u8 join_state;
499 
500 	/* remove bits that slave is already member of, and adjust */
501 	join_state = join_mask & (~group->func[slave].join_state);
502 	adjust_membership(group, join_state, 1);
503 	group->func[slave].join_state |= join_state;
504 	if (group->func[slave].state != MCAST_MEMBER && join_state) {
505 		group->func[slave].state = MCAST_MEMBER;
506 		ret = 1;
507 	}
508 	return ret;
509 }
510 
511 static int leave_group(struct mcast_group *group, int slave, u8 leave_state)
512 {
513 	int ret = 0;
514 
515 	adjust_membership(group, leave_state, -1);
516 	group->func[slave].join_state &= ~leave_state;
517 	if (!group->func[slave].join_state) {
518 		group->func[slave].state = MCAST_NOT_MEMBER;
519 		ret = 1;
520 	}
521 	return ret;
522 }
523 
524 static int check_leave(struct mcast_group *group, int slave, u8 leave_mask)
525 {
526 	if (group->func[slave].state != MCAST_MEMBER)
527 		return MAD_STATUS_REQ_INVALID;
528 
529 	/* make sure we're not deleting unset bits */
530 	if (~group->func[slave].join_state & leave_mask)
531 		return MAD_STATUS_REQ_INVALID;
532 
533 	if (!leave_mask)
534 		return MAD_STATUS_REQ_INVALID;
535 
536 	return 0;
537 }
538 
539 static void mlx4_ib_mcg_timeout_handler(struct work_struct *work)
540 {
541 	struct delayed_work *delay = to_delayed_work(work);
542 	struct mcast_group *group;
543 	struct mcast_req *req = NULL;
544 
545 	group = container_of(delay, typeof(*group), timeout_work);
546 
547 	mutex_lock(&group->lock);
548 	if (group->state == MCAST_JOIN_SENT) {
549 		if (!list_empty(&group->pending_list)) {
550 			req = list_first_entry(&group->pending_list, struct mcast_req, group_list);
551 			list_del(&req->group_list);
552 			list_del(&req->func_list);
553 			--group->func[req->func].num_pend_reqs;
554 			mutex_unlock(&group->lock);
555 			kfree(req);
556 			if (memcmp(&group->rec.mgid, &mgid0, sizeof mgid0)) {
557 				if (release_group(group, 1))
558 					return;
559 			} else {
560 				kfree(group);
561 				return;
562 			}
563 			mutex_lock(&group->lock);
564 		} else
565 			mcg_warn_group(group, "DRIVER BUG\n");
566 	} else if (group->state == MCAST_LEAVE_SENT) {
567 		if (group->rec.scope_join_state & 0xf)
568 			group->rec.scope_join_state &= 0xf0;
569 		group->state = MCAST_IDLE;
570 		mutex_unlock(&group->lock);
571 		if (release_group(group, 1))
572 			return;
573 		mutex_lock(&group->lock);
574 	} else
575 		mcg_warn_group(group, "invalid state %s\n", get_state_string(group->state));
576 	group->state = MCAST_IDLE;
577 	atomic_inc(&group->refcount);
578 	if (!queue_work(group->demux->mcg_wq, &group->work))
579 		safe_atomic_dec(&group->refcount);
580 
581 	mutex_unlock(&group->lock);
582 }
583 
584 static int handle_leave_req(struct mcast_group *group, u8 leave_mask,
585 			    struct mcast_req *req)
586 {
587 	u16 status;
588 
589 	if (req->clean)
590 		leave_mask = group->func[req->func].join_state;
591 
592 	status = check_leave(group, req->func, leave_mask);
593 	if (!status)
594 		leave_group(group, req->func, leave_mask);
595 
596 	if (!req->clean)
597 		send_reply_to_slave(req->func, group, &req->sa_mad, status);
598 	--group->func[req->func].num_pend_reqs;
599 	list_del(&req->group_list);
600 	list_del(&req->func_list);
601 	kfree(req);
602 	return 1;
603 }
604 
605 static int handle_join_req(struct mcast_group *group, u8 join_mask,
606 			   struct mcast_req *req)
607 {
608 	u8 group_join_state = group->rec.scope_join_state & 0xf;
609 	int ref = 0;
610 	u16 status;
611 	struct ib_sa_mcmember_data *sa_data = (struct ib_sa_mcmember_data *)req->sa_mad.data;
612 
613 	if (join_mask == (group_join_state & join_mask)) {
614 		/* port's membership need not change */
615 		status = cmp_rec(&group->rec, sa_data, req->sa_mad.sa_hdr.comp_mask);
616 		if (!status)
617 			join_group(group, req->func, join_mask);
618 
619 		--group->func[req->func].num_pend_reqs;
620 		send_reply_to_slave(req->func, group, &req->sa_mad, status);
621 		list_del(&req->group_list);
622 		list_del(&req->func_list);
623 		kfree(req);
624 		++ref;
625 	} else {
626 		/* port's membership needs to be updated */
627 		group->prev_state = group->state;
628 		if (send_join_to_wire(group, &req->sa_mad)) {
629 			--group->func[req->func].num_pend_reqs;
630 			list_del(&req->group_list);
631 			list_del(&req->func_list);
632 			kfree(req);
633 			ref = 1;
634 			group->state = group->prev_state;
635 		} else
636 			group->state = MCAST_JOIN_SENT;
637 	}
638 
639 	return ref;
640 }
641 
642 static void mlx4_ib_mcg_work_handler(struct work_struct *work)
643 {
644 	struct mcast_group *group;
645 	struct mcast_req *req = NULL;
646 	struct ib_sa_mcmember_data *sa_data;
647 	u8 req_join_state;
648 	int rc = 1; /* release_count - this is for the scheduled work */
649 	u16 status;
650 	u8 method;
651 
652 	group = container_of(work, typeof(*group), work);
653 
654 	mutex_lock(&group->lock);
655 
656 	/* First, let's see if a response from SM is waiting regarding this group.
657 	 * If so, we need to update the group's REC. If this is a bad response, we
658 	 * may need to send a bad response to a VF waiting for it. If VF is waiting
659 	 * and this is a good response, the VF will be answered later in this func. */
660 	if (group->state == MCAST_RESP_READY) {
661 		/* cancels mlx4_ib_mcg_timeout_handler */
662 		cancel_delayed_work(&group->timeout_work);
663 		status = be16_to_cpu(group->response_sa_mad.mad_hdr.status);
664 		method = group->response_sa_mad.mad_hdr.method;
665 		if (group->last_req_tid != group->response_sa_mad.mad_hdr.tid) {
666 			mcg_warn_group(group, "Got MAD response to existing MGID but wrong TID, dropping. Resp TID=%llx, group TID=%llx\n",
667 				be64_to_cpu(group->response_sa_mad.mad_hdr.tid),
668 				be64_to_cpu(group->last_req_tid));
669 			group->state = group->prev_state;
670 			goto process_requests;
671 		}
672 		if (status) {
673 			if (!list_empty(&group->pending_list))
674 				req = list_first_entry(&group->pending_list,
675 						struct mcast_req, group_list);
676 			if ((method == IB_MGMT_METHOD_GET_RESP)) {
677 					if (req) {
678 						send_reply_to_slave(req->func, group, &req->sa_mad, status);
679 						--group->func[req->func].num_pend_reqs;
680 						list_del(&req->group_list);
681 						list_del(&req->func_list);
682 						kfree(req);
683 						++rc;
684 					} else
685 						mcg_warn_group(group, "no request for failed join\n");
686 			} else if (method == IB_SA_METHOD_DELETE_RESP && group->demux->flushing)
687 				++rc;
688 		} else {
689 			u8 resp_join_state;
690 			u8 cur_join_state;
691 
692 			resp_join_state = ((struct ib_sa_mcmember_data *)
693 						group->response_sa_mad.data)->scope_join_state & 0xf;
694 			cur_join_state = group->rec.scope_join_state & 0xf;
695 
696 			if (method == IB_MGMT_METHOD_GET_RESP) {
697 				/* successfull join */
698 				if (!cur_join_state && resp_join_state)
699 					--rc;
700 			} else if (!resp_join_state)
701 					++rc;
702 			memcpy(&group->rec, group->response_sa_mad.data, sizeof group->rec);
703 		}
704 		group->state = MCAST_IDLE;
705 	}
706 
707 process_requests:
708 	/* We should now go over pending join/leave requests, as long as we are idle. */
709 	while (!list_empty(&group->pending_list) && group->state == MCAST_IDLE) {
710 		req = list_first_entry(&group->pending_list, struct mcast_req,
711 				       group_list);
712 		sa_data = (struct ib_sa_mcmember_data *)req->sa_mad.data;
713 		req_join_state = sa_data->scope_join_state & 0xf;
714 
715 		/* For a leave request, we will immediately answer the VF, and
716 		 * update our internal counters. The actual leave will be sent
717 		 * to SM later, if at all needed. We dequeue the request now. */
718 		if (req->sa_mad.mad_hdr.method == IB_SA_METHOD_DELETE)
719 			rc += handle_leave_req(group, req_join_state, req);
720 		else
721 			rc += handle_join_req(group, req_join_state, req);
722 	}
723 
724 	/* Handle leaves */
725 	if (group->state == MCAST_IDLE) {
726 		req_join_state = get_leave_state(group);
727 		if (req_join_state) {
728 			group->rec.scope_join_state &= ~req_join_state;
729 			group->prev_state = group->state;
730 			if (send_leave_to_wire(group, req_join_state)) {
731 				group->state = group->prev_state;
732 				++rc;
733 			} else
734 				group->state = MCAST_LEAVE_SENT;
735 		}
736 	}
737 
738 	if (!list_empty(&group->pending_list) && group->state == MCAST_IDLE)
739 		goto process_requests;
740 	mutex_unlock(&group->lock);
741 
742 	while (rc--)
743 		release_group(group, 0);
744 }
745 
746 static struct mcast_group *search_relocate_mgid0_group(struct mlx4_ib_demux_ctx *ctx,
747 						       __be64 tid,
748 						       union ib_gid *new_mgid)
749 {
750 	struct mcast_group *group = NULL, *cur_group, *n;
751 	struct mcast_req *req;
752 
753 	mutex_lock(&ctx->mcg_table_lock);
754 	list_for_each_entry_safe(group, n, &ctx->mcg_mgid0_list, mgid0_list) {
755 		mutex_lock(&group->lock);
756 		if (group->last_req_tid == tid) {
757 			if (memcmp(new_mgid, &mgid0, sizeof mgid0)) {
758 				group->rec.mgid = *new_mgid;
759 				sprintf(group->name, "%016llx%016llx",
760 						be64_to_cpu(group->rec.mgid.global.subnet_prefix),
761 						be64_to_cpu(group->rec.mgid.global.interface_id));
762 				list_del_init(&group->mgid0_list);
763 				cur_group = mcast_insert(ctx, group);
764 				if (cur_group) {
765 					/* A race between our code and SM. Silently cleaning the new one */
766 					req = list_first_entry(&group->pending_list,
767 							       struct mcast_req, group_list);
768 					--group->func[req->func].num_pend_reqs;
769 					list_del(&req->group_list);
770 					list_del(&req->func_list);
771 					kfree(req);
772 					mutex_unlock(&group->lock);
773 					mutex_unlock(&ctx->mcg_table_lock);
774 					release_group(group, 0);
775 					return NULL;
776 				}
777 
778 				atomic_inc(&group->refcount);
779 				add_sysfs_port_mcg_attr(ctx->dev, ctx->port, &group->dentry.attr);
780 				mutex_unlock(&group->lock);
781 				mutex_unlock(&ctx->mcg_table_lock);
782 				return group;
783 			} else {
784 				struct mcast_req *tmp1, *tmp2;
785 
786 				list_del(&group->mgid0_list);
787 				if (!list_empty(&group->pending_list) && group->state != MCAST_IDLE)
788 					cancel_delayed_work_sync(&group->timeout_work);
789 
790 				list_for_each_entry_safe(tmp1, tmp2, &group->pending_list, group_list) {
791 					list_del(&tmp1->group_list);
792 					kfree(tmp1);
793 				}
794 				mutex_unlock(&group->lock);
795 				mutex_unlock(&ctx->mcg_table_lock);
796 				kfree(group);
797 				return NULL;
798 			}
799 		}
800 		mutex_unlock(&group->lock);
801 	}
802 	mutex_unlock(&ctx->mcg_table_lock);
803 
804 	return NULL;
805 }
806 
807 static ssize_t sysfs_show_group(struct device *dev,
808 		struct device_attribute *attr, char *buf);
809 
810 static struct mcast_group *acquire_group(struct mlx4_ib_demux_ctx *ctx,
811 					 union ib_gid *mgid, int create,
812 					 gfp_t gfp_mask)
813 {
814 	struct mcast_group *group, *cur_group;
815 	int is_mgid0;
816 	int i;
817 
818 	is_mgid0 = !memcmp(&mgid0, mgid, sizeof mgid0);
819 	if (!is_mgid0) {
820 		group = mcast_find(ctx, mgid);
821 		if (group)
822 			goto found;
823 	}
824 
825 	if (!create)
826 		return ERR_PTR(-ENOENT);
827 
828 	group = kzalloc(sizeof *group, gfp_mask);
829 	if (!group)
830 		return ERR_PTR(-ENOMEM);
831 
832 	group->demux = ctx;
833 	group->rec.mgid = *mgid;
834 	INIT_LIST_HEAD(&group->pending_list);
835 	INIT_LIST_HEAD(&group->mgid0_list);
836 	for (i = 0; i < MAX_VFS; ++i)
837 		INIT_LIST_HEAD(&group->func[i].pending);
838 	INIT_WORK(&group->work, mlx4_ib_mcg_work_handler);
839 	INIT_DELAYED_WORK(&group->timeout_work, mlx4_ib_mcg_timeout_handler);
840 	mutex_init(&group->lock);
841 	sprintf(group->name, "%016llx%016llx",
842 			be64_to_cpu(group->rec.mgid.global.subnet_prefix),
843 			be64_to_cpu(group->rec.mgid.global.interface_id));
844 	sysfs_attr_init(&group->dentry.attr);
845 	group->dentry.show = sysfs_show_group;
846 	group->dentry.store = NULL;
847 	group->dentry.attr.name = group->name;
848 	group->dentry.attr.mode = 0400;
849 	group->state = MCAST_IDLE;
850 
851 	if (is_mgid0) {
852 		list_add(&group->mgid0_list, &ctx->mcg_mgid0_list);
853 		goto found;
854 	}
855 
856 	cur_group = mcast_insert(ctx, group);
857 	if (cur_group) {
858 		mcg_warn("group just showed up %s - confused\n", cur_group->name);
859 		kfree(group);
860 		return ERR_PTR(-EINVAL);
861 	}
862 
863 	add_sysfs_port_mcg_attr(ctx->dev, ctx->port, &group->dentry.attr);
864 
865 found:
866 	atomic_inc(&group->refcount);
867 	return group;
868 }
869 
870 static void queue_req(struct mcast_req *req)
871 {
872 	struct mcast_group *group = req->group;
873 
874 	atomic_inc(&group->refcount); /* for the request */
875 	atomic_inc(&group->refcount); /* for scheduling the work */
876 	list_add_tail(&req->group_list, &group->pending_list);
877 	list_add_tail(&req->func_list, &group->func[req->func].pending);
878 	/* calls mlx4_ib_mcg_work_handler */
879 	if (!queue_work(group->demux->mcg_wq, &group->work))
880 		safe_atomic_dec(&group->refcount);
881 }
882 
883 int mlx4_ib_mcg_demux_handler(struct ib_device *ibdev, int port, int slave,
884 			      struct ib_sa_mad *mad)
885 {
886 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
887 	struct ib_sa_mcmember_data *rec = (struct ib_sa_mcmember_data *)mad->data;
888 	struct mlx4_ib_demux_ctx *ctx = &dev->sriov.demux[port - 1];
889 	struct mcast_group *group;
890 
891 	switch (mad->mad_hdr.method) {
892 	case IB_MGMT_METHOD_GET_RESP:
893 	case IB_SA_METHOD_DELETE_RESP:
894 		mutex_lock(&ctx->mcg_table_lock);
895 		group = acquire_group(ctx, &rec->mgid, 0, GFP_KERNEL);
896 		mutex_unlock(&ctx->mcg_table_lock);
897 		if (IS_ERR(group)) {
898 			if (mad->mad_hdr.method == IB_MGMT_METHOD_GET_RESP) {
899 				__be64 tid = mad->mad_hdr.tid;
900 				*(u8 *)(&tid) = (u8)slave; /* in group we kept the modified TID */
901 				group = search_relocate_mgid0_group(ctx, tid, &rec->mgid);
902 			} else
903 				group = NULL;
904 		}
905 
906 		if (!group)
907 			return 1;
908 
909 		mutex_lock(&group->lock);
910 		group->response_sa_mad = *mad;
911 		group->prev_state = group->state;
912 		group->state = MCAST_RESP_READY;
913 		/* calls mlx4_ib_mcg_work_handler */
914 		atomic_inc(&group->refcount);
915 		if (!queue_work(ctx->mcg_wq, &group->work))
916 			safe_atomic_dec(&group->refcount);
917 		mutex_unlock(&group->lock);
918 		release_group(group, 0);
919 		return 1; /* consumed */
920 	case IB_MGMT_METHOD_SET:
921 	case IB_SA_METHOD_GET_TABLE:
922 	case IB_SA_METHOD_GET_TABLE_RESP:
923 	case IB_SA_METHOD_DELETE:
924 		return 0; /* not consumed, pass-through to guest over tunnel */
925 	default:
926 		mcg_warn("In demux, port %d: unexpected MCMember method: 0x%x, dropping\n",
927 			port, mad->mad_hdr.method);
928 		return 1; /* consumed */
929 	}
930 }
931 
932 int mlx4_ib_mcg_multiplex_handler(struct ib_device *ibdev, int port,
933 				  int slave, struct ib_sa_mad *sa_mad)
934 {
935 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
936 	struct ib_sa_mcmember_data *rec = (struct ib_sa_mcmember_data *)sa_mad->data;
937 	struct mlx4_ib_demux_ctx *ctx = &dev->sriov.demux[port - 1];
938 	struct mcast_group *group;
939 	struct mcast_req *req;
940 	int may_create = 0;
941 
942 	if (ctx->flushing)
943 		return -EAGAIN;
944 
945 	switch (sa_mad->mad_hdr.method) {
946 	case IB_MGMT_METHOD_SET:
947 		may_create = 1;
948 	case IB_SA_METHOD_DELETE:
949 		req = kzalloc(sizeof *req, GFP_KERNEL);
950 		if (!req)
951 			return -ENOMEM;
952 
953 		req->func = slave;
954 		req->sa_mad = *sa_mad;
955 
956 		mutex_lock(&ctx->mcg_table_lock);
957 		group = acquire_group(ctx, &rec->mgid, may_create, GFP_KERNEL);
958 		mutex_unlock(&ctx->mcg_table_lock);
959 		if (IS_ERR(group)) {
960 			kfree(req);
961 			return PTR_ERR(group);
962 		}
963 		mutex_lock(&group->lock);
964 		if (group->func[slave].num_pend_reqs > MAX_PEND_REQS_PER_FUNC) {
965 			mutex_unlock(&group->lock);
966 			mcg_debug_group(group, "Port %d, Func %d has too many pending requests (%d), dropping\n",
967 					port, slave, MAX_PEND_REQS_PER_FUNC);
968 			release_group(group, 0);
969 			kfree(req);
970 			return -ENOMEM;
971 		}
972 		++group->func[slave].num_pend_reqs;
973 		req->group = group;
974 		queue_req(req);
975 		mutex_unlock(&group->lock);
976 		release_group(group, 0);
977 		return 1; /* consumed */
978 	case IB_SA_METHOD_GET_TABLE:
979 	case IB_MGMT_METHOD_GET_RESP:
980 	case IB_SA_METHOD_GET_TABLE_RESP:
981 	case IB_SA_METHOD_DELETE_RESP:
982 		return 0; /* not consumed, pass-through */
983 	default:
984 		mcg_warn("In multiplex, port %d, func %d: unexpected MCMember method: 0x%x, dropping\n",
985 			port, slave, sa_mad->mad_hdr.method);
986 		return 1; /* consumed */
987 	}
988 }
989 
990 static ssize_t sysfs_show_group(struct device *dev,
991 		struct device_attribute *attr, char *buf)
992 {
993 	struct mcast_group *group =
994 		container_of(attr, struct mcast_group, dentry);
995 	struct mcast_req *req = NULL;
996 	char pending_str[40];
997 	char state_str[40];
998 	ssize_t len = 0;
999 	int f;
1000 
1001 	if (group->state == MCAST_IDLE)
1002 		sprintf(state_str, "%s", get_state_string(group->state));
1003 	else
1004 		sprintf(state_str, "%s(TID=0x%llx)",
1005 				get_state_string(group->state),
1006 				be64_to_cpu(group->last_req_tid));
1007 	if (list_empty(&group->pending_list)) {
1008 		sprintf(pending_str, "No");
1009 	} else {
1010 		req = list_first_entry(&group->pending_list, struct mcast_req, group_list);
1011 		sprintf(pending_str, "Yes(TID=0x%llx)",
1012 				be64_to_cpu(req->sa_mad.mad_hdr.tid));
1013 	}
1014 	len += sprintf(buf + len, "%1d [%02d,%02d,%02d] %4d %4s %5s     ",
1015 			group->rec.scope_join_state & 0xf,
1016 			group->members[2], group->members[1], group->members[0],
1017 			atomic_read(&group->refcount),
1018 			pending_str,
1019 			state_str);
1020 	for (f = 0; f < MAX_VFS; ++f)
1021 		if (group->func[f].state == MCAST_MEMBER)
1022 			len += sprintf(buf + len, "%d[%1x] ",
1023 					f, group->func[f].join_state);
1024 
1025 	len += sprintf(buf + len, "\t\t(%4hx %4x %2x %2x %2x %2x %2x "
1026 		"%4x %4x %2x %2x)\n",
1027 		be16_to_cpu(group->rec.pkey),
1028 		be32_to_cpu(group->rec.qkey),
1029 		(group->rec.mtusel_mtu & 0xc0) >> 6,
1030 		group->rec.mtusel_mtu & 0x3f,
1031 		group->rec.tclass,
1032 		(group->rec.ratesel_rate & 0xc0) >> 6,
1033 		group->rec.ratesel_rate & 0x3f,
1034 		(be32_to_cpu(group->rec.sl_flowlabel_hoplimit) & 0xf0000000) >> 28,
1035 		(be32_to_cpu(group->rec.sl_flowlabel_hoplimit) & 0x0fffff00) >> 8,
1036 		be32_to_cpu(group->rec.sl_flowlabel_hoplimit) & 0x000000ff,
1037 		group->rec.proxy_join);
1038 
1039 	return len;
1040 }
1041 
1042 int mlx4_ib_mcg_port_init(struct mlx4_ib_demux_ctx *ctx)
1043 {
1044 	char name[20];
1045 
1046 	atomic_set(&ctx->tid, 0);
1047 	sprintf(name, "mlx4_ib_mcg%d", ctx->port);
1048 	ctx->mcg_wq = alloc_ordered_workqueue(name, WQ_MEM_RECLAIM);
1049 	if (!ctx->mcg_wq)
1050 		return -ENOMEM;
1051 
1052 	mutex_init(&ctx->mcg_table_lock);
1053 	ctx->mcg_table = RB_ROOT;
1054 	INIT_LIST_HEAD(&ctx->mcg_mgid0_list);
1055 	ctx->flushing = 0;
1056 
1057 	return 0;
1058 }
1059 
1060 static void force_clean_group(struct mcast_group *group)
1061 {
1062 	struct mcast_req *req, *tmp
1063 		;
1064 	list_for_each_entry_safe(req, tmp, &group->pending_list, group_list) {
1065 		list_del(&req->group_list);
1066 		kfree(req);
1067 	}
1068 	del_sysfs_port_mcg_attr(group->demux->dev, group->demux->port, &group->dentry.attr);
1069 	rb_erase(&group->node, &group->demux->mcg_table);
1070 	kfree(group);
1071 }
1072 
1073 static void _mlx4_ib_mcg_port_cleanup(struct mlx4_ib_demux_ctx *ctx, int destroy_wq)
1074 {
1075 	int i;
1076 	struct rb_node *p;
1077 	struct mcast_group *group;
1078 	unsigned long end;
1079 	int count;
1080 
1081 	for (i = 0; i < MAX_VFS; ++i)
1082 		clean_vf_mcast(ctx, i);
1083 
1084 	end = jiffies + msecs_to_jiffies(MAD_TIMEOUT_MS + 3000);
1085 	do {
1086 		count = 0;
1087 		mutex_lock(&ctx->mcg_table_lock);
1088 		for (p = rb_first(&ctx->mcg_table); p; p = rb_next(p))
1089 			++count;
1090 		mutex_unlock(&ctx->mcg_table_lock);
1091 		if (!count)
1092 			break;
1093 
1094 		msleep(1);
1095 	} while (time_after(end, jiffies));
1096 
1097 	flush_workqueue(ctx->mcg_wq);
1098 	if (destroy_wq)
1099 		destroy_workqueue(ctx->mcg_wq);
1100 
1101 	mutex_lock(&ctx->mcg_table_lock);
1102 	while ((p = rb_first(&ctx->mcg_table)) != NULL) {
1103 		group = rb_entry(p, struct mcast_group, node);
1104 		if (atomic_read(&group->refcount))
1105 			mcg_warn_group(group, "group refcount %d!!! (pointer %p)\n", atomic_read(&group->refcount), group);
1106 
1107 		force_clean_group(group);
1108 	}
1109 	mutex_unlock(&ctx->mcg_table_lock);
1110 }
1111 
1112 struct clean_work {
1113 	struct work_struct work;
1114 	struct mlx4_ib_demux_ctx *ctx;
1115 	int destroy_wq;
1116 };
1117 
1118 static void mcg_clean_task(struct work_struct *work)
1119 {
1120 	struct clean_work *cw = container_of(work, struct clean_work, work);
1121 
1122 	_mlx4_ib_mcg_port_cleanup(cw->ctx, cw->destroy_wq);
1123 	cw->ctx->flushing = 0;
1124 	kfree(cw);
1125 }
1126 
1127 void mlx4_ib_mcg_port_cleanup(struct mlx4_ib_demux_ctx *ctx, int destroy_wq)
1128 {
1129 	struct clean_work *work;
1130 
1131 	if (ctx->flushing)
1132 		return;
1133 
1134 	ctx->flushing = 1;
1135 
1136 	if (destroy_wq) {
1137 		_mlx4_ib_mcg_port_cleanup(ctx, destroy_wq);
1138 		ctx->flushing = 0;
1139 		return;
1140 	}
1141 
1142 	work = kmalloc(sizeof *work, GFP_KERNEL);
1143 	if (!work) {
1144 		ctx->flushing = 0;
1145 		return;
1146 	}
1147 
1148 	work->ctx = ctx;
1149 	work->destroy_wq = destroy_wq;
1150 	INIT_WORK(&work->work, mcg_clean_task);
1151 	queue_work(clean_wq, &work->work);
1152 }
1153 
1154 static void build_leave_mad(struct mcast_req *req)
1155 {
1156 	struct ib_sa_mad *mad = &req->sa_mad;
1157 
1158 	mad->mad_hdr.method = IB_SA_METHOD_DELETE;
1159 }
1160 
1161 
1162 static void clear_pending_reqs(struct mcast_group *group, int vf)
1163 {
1164 	struct mcast_req *req, *tmp, *group_first = NULL;
1165 	int clear;
1166 	int pend = 0;
1167 
1168 	if (!list_empty(&group->pending_list))
1169 		group_first = list_first_entry(&group->pending_list, struct mcast_req, group_list);
1170 
1171 	list_for_each_entry_safe(req, tmp, &group->func[vf].pending, func_list) {
1172 		clear = 1;
1173 		if (group_first == req &&
1174 		    (group->state == MCAST_JOIN_SENT ||
1175 		     group->state == MCAST_LEAVE_SENT)) {
1176 			clear = cancel_delayed_work(&group->timeout_work);
1177 			pend = !clear;
1178 			group->state = MCAST_IDLE;
1179 		}
1180 		if (clear) {
1181 			--group->func[vf].num_pend_reqs;
1182 			list_del(&req->group_list);
1183 			list_del(&req->func_list);
1184 			kfree(req);
1185 			atomic_dec(&group->refcount);
1186 		}
1187 	}
1188 
1189 	if (!pend && (!list_empty(&group->func[vf].pending) || group->func[vf].num_pend_reqs)) {
1190 		mcg_warn_group(group, "DRIVER BUG: list_empty %d, num_pend_reqs %d\n",
1191 			       list_empty(&group->func[vf].pending), group->func[vf].num_pend_reqs);
1192 	}
1193 }
1194 
1195 static int push_deleteing_req(struct mcast_group *group, int slave)
1196 {
1197 	struct mcast_req *req;
1198 	struct mcast_req *pend_req;
1199 
1200 	if (!group->func[slave].join_state)
1201 		return 0;
1202 
1203 	req = kzalloc(sizeof *req, GFP_KERNEL);
1204 	if (!req)
1205 		return -ENOMEM;
1206 
1207 	if (!list_empty(&group->func[slave].pending)) {
1208 		pend_req = list_entry(group->func[slave].pending.prev, struct mcast_req, group_list);
1209 		if (pend_req->clean) {
1210 			kfree(req);
1211 			return 0;
1212 		}
1213 	}
1214 
1215 	req->clean = 1;
1216 	req->func = slave;
1217 	req->group = group;
1218 	++group->func[slave].num_pend_reqs;
1219 	build_leave_mad(req);
1220 	queue_req(req);
1221 	return 0;
1222 }
1223 
1224 void clean_vf_mcast(struct mlx4_ib_demux_ctx *ctx, int slave)
1225 {
1226 	struct mcast_group *group;
1227 	struct rb_node *p;
1228 
1229 	mutex_lock(&ctx->mcg_table_lock);
1230 	for (p = rb_first(&ctx->mcg_table); p; p = rb_next(p)) {
1231 		group = rb_entry(p, struct mcast_group, node);
1232 		mutex_lock(&group->lock);
1233 		if (atomic_read(&group->refcount)) {
1234 			/* clear pending requests of this VF */
1235 			clear_pending_reqs(group, slave);
1236 			push_deleteing_req(group, slave);
1237 		}
1238 		mutex_unlock(&group->lock);
1239 	}
1240 	mutex_unlock(&ctx->mcg_table_lock);
1241 }
1242 
1243 
1244 int mlx4_ib_mcg_init(void)
1245 {
1246 	clean_wq = alloc_ordered_workqueue("mlx4_ib_mcg", WQ_MEM_RECLAIM);
1247 	if (!clean_wq)
1248 		return -ENOMEM;
1249 
1250 	return 0;
1251 }
1252 
1253 void mlx4_ib_mcg_destroy(void)
1254 {
1255 	destroy_workqueue(clean_wq);
1256 }
1257