2 * Copyright (c) 2012 Mellanox Technologies. All rights reserved.
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:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
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.
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
33 #include <rdma/ib_mad.h>
34 #include <rdma/ib_smi.h>
35 #include <rdma/ib_cache.h>
36 #include <rdma/ib_sa.h>
38 #include <linux/mlx4/cmd.h>
39 #include <linux/rbtree.h>
40 #include <linux/delay.h>
45 #define MAX_PEND_REQS_PER_FUNC 4
46 #define MAD_TIMEOUT_MS 2000
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)
54 #define mcg_error_group(group, format, arg...) \
55 pr_err(" %16s: " format, (group)->name, ## arg)
58 static union ib_gid mgid0;
60 static struct workqueue_struct *clean_wq;
67 enum mcast_group_state {
75 enum mcast_state state;
78 struct list_head pending;
81 struct ib_sa_mcmember_data {
83 union ib_gid port_gid;
91 __be32 sl_flowlabel_hoplimit;
98 struct ib_sa_mcmember_data rec;
100 struct list_head mgid0_list;
101 struct mlx4_ib_demux_ctx *demux;
102 struct mcast_member func[MAX_VFS];
104 struct work_struct work;
105 struct list_head pending_list;
107 enum mcast_group_state state;
108 enum mcast_group_state prev_state;
109 struct ib_sa_mad response_sa_mad;
112 char name[33]; /* MGID string */
113 struct device_attribute dentry;
115 /* refcount is the reference count for the following:
116 1. Each queued request
117 2. Each invocation of the worker thread
118 3. Membership of the port at the SA
122 /* delayed work to clean pending SM request */
123 struct delayed_work timeout_work;
124 struct list_head cleanup_list;
129 struct ib_sa_mad sa_mad;
130 struct list_head group_list;
131 struct list_head func_list;
132 struct mcast_group *group;
137 #define safe_atomic_dec(ref) \
139 if (atomic_dec_and_test(ref)) \
140 mcg_warn_group(group, "did not expect to reach zero\n"); \
143 static const char *get_state_string(enum mcast_group_state state)
148 case MCAST_JOIN_SENT:
149 return "MCAST_JOIN_SENT";
150 case MCAST_LEAVE_SENT:
151 return "MCAST_LEAVE_SENT";
152 case MCAST_RESP_READY:
153 return "MCAST_RESP_READY";
155 return "Invalid State";
158 static struct mcast_group *mcast_find(struct mlx4_ib_demux_ctx *ctx,
161 struct rb_node *node = ctx->mcg_table.rb_node;
162 struct mcast_group *group;
166 group = rb_entry(node, struct mcast_group, node);
167 ret = memcmp(mgid->raw, group->rec.mgid.raw, sizeof *mgid);
172 node = node->rb_left;
174 node = node->rb_right;
179 static struct mcast_group *mcast_insert(struct mlx4_ib_demux_ctx *ctx,
180 struct mcast_group *group)
182 struct rb_node **link = &ctx->mcg_table.rb_node;
183 struct rb_node *parent = NULL;
184 struct mcast_group *cur_group;
189 cur_group = rb_entry(parent, struct mcast_group, node);
191 ret = memcmp(group->rec.mgid.raw, cur_group->rec.mgid.raw,
192 sizeof group->rec.mgid);
194 link = &(*link)->rb_left;
196 link = &(*link)->rb_right;
200 rb_link_node(&group->node, parent, link);
201 rb_insert_color(&group->node, &ctx->mcg_table);
205 static int send_mad_to_wire(struct mlx4_ib_demux_ctx *ctx, struct ib_mad *mad)
207 struct mlx4_ib_dev *dev = ctx->dev;
208 struct ib_ah_attr ah_attr;
210 spin_lock(&dev->sm_lock);
211 if (!dev->sm_ah[ctx->port - 1]) {
212 /* port is not yet Active, sm_ah not ready */
213 spin_unlock(&dev->sm_lock);
216 mlx4_ib_query_ah(dev->sm_ah[ctx->port - 1], &ah_attr);
217 spin_unlock(&dev->sm_lock);
218 return mlx4_ib_send_to_wire(dev, mlx4_master_func_num(dev->dev), ctx->port,
219 IB_QPT_GSI, 0, 1, IB_QP1_QKEY, &ah_attr, mad);
222 static int send_mad_to_slave(int slave, struct mlx4_ib_demux_ctx *ctx,
225 struct mlx4_ib_dev *dev = ctx->dev;
226 struct ib_mad_agent *agent = dev->send_agent[ctx->port - 1][1];
228 struct ib_ah_attr ah_attr;
230 /* Our agent might not yet be registered when mads start to arrive */
234 ib_query_ah(dev->sm_ah[ctx->port - 1], &ah_attr);
238 wc.dlid_path_bits = 0;
239 wc.port_num = ctx->port;
240 wc.slid = ah_attr.dlid; /* opensm lid */
242 return mlx4_ib_send_to_slave(dev, slave, ctx->port, IB_QPT_GSI, &wc, NULL, mad);
245 static int send_join_to_wire(struct mcast_group *group, struct ib_sa_mad *sa_mad)
247 struct ib_sa_mad mad;
248 struct ib_sa_mcmember_data *sa_mad_data = (struct ib_sa_mcmember_data *)&mad.data;
251 /* we rely on a mad request as arrived from a VF */
252 memcpy(&mad, sa_mad, sizeof mad);
254 /* fix port GID to be the real one (slave 0) */
255 sa_mad_data->port_gid.global.interface_id = group->demux->guid_cache[0];
257 /* assign our own TID */
258 mad.mad_hdr.tid = mlx4_ib_get_new_demux_tid(group->demux);
259 group->last_req_tid = mad.mad_hdr.tid; /* keep it for later validation */
261 ret = send_mad_to_wire(group->demux, (struct ib_mad *)&mad);
262 /* set timeout handler */
264 /* calls mlx4_ib_mcg_timeout_handler */
265 queue_delayed_work(group->demux->mcg_wq, &group->timeout_work,
266 msecs_to_jiffies(MAD_TIMEOUT_MS));
272 static int send_leave_to_wire(struct mcast_group *group, u8 join_state)
274 struct ib_sa_mad mad;
275 struct ib_sa_mcmember_data *sa_data = (struct ib_sa_mcmember_data *)&mad.data;
278 memset(&mad, 0, sizeof mad);
279 mad.mad_hdr.base_version = 1;
280 mad.mad_hdr.mgmt_class = IB_MGMT_CLASS_SUBN_ADM;
281 mad.mad_hdr.class_version = 2;
282 mad.mad_hdr.method = IB_SA_METHOD_DELETE;
283 mad.mad_hdr.status = cpu_to_be16(0);
284 mad.mad_hdr.class_specific = cpu_to_be16(0);
285 mad.mad_hdr.tid = mlx4_ib_get_new_demux_tid(group->demux);
286 group->last_req_tid = mad.mad_hdr.tid; /* keep it for later validation */
287 mad.mad_hdr.attr_id = cpu_to_be16(IB_SA_ATTR_MC_MEMBER_REC);
288 mad.mad_hdr.attr_mod = cpu_to_be32(0);
289 mad.sa_hdr.sm_key = 0x0;
290 mad.sa_hdr.attr_offset = cpu_to_be16(7);
291 mad.sa_hdr.comp_mask = IB_SA_MCMEMBER_REC_MGID |
292 IB_SA_MCMEMBER_REC_PORT_GID | IB_SA_MCMEMBER_REC_JOIN_STATE;
294 *sa_data = group->rec;
295 sa_data->scope_join_state = join_state;
297 ret = send_mad_to_wire(group->demux, (struct ib_mad *)&mad);
299 group->state = MCAST_IDLE;
301 /* set timeout handler */
303 /* calls mlx4_ib_mcg_timeout_handler */
304 queue_delayed_work(group->demux->mcg_wq, &group->timeout_work,
305 msecs_to_jiffies(MAD_TIMEOUT_MS));
311 static int send_reply_to_slave(int slave, struct mcast_group *group,
312 struct ib_sa_mad *req_sa_mad, u16 status)
314 struct ib_sa_mad mad;
315 struct ib_sa_mcmember_data *sa_data = (struct ib_sa_mcmember_data *)&mad.data;
316 struct ib_sa_mcmember_data *req_sa_data = (struct ib_sa_mcmember_data *)&req_sa_mad->data;
319 memset(&mad, 0, sizeof mad);
320 mad.mad_hdr.base_version = 1;
321 mad.mad_hdr.mgmt_class = IB_MGMT_CLASS_SUBN_ADM;
322 mad.mad_hdr.class_version = 2;
323 mad.mad_hdr.method = IB_MGMT_METHOD_GET_RESP;
324 mad.mad_hdr.status = cpu_to_be16(status);
325 mad.mad_hdr.class_specific = cpu_to_be16(0);
326 mad.mad_hdr.tid = req_sa_mad->mad_hdr.tid;
327 *(u8 *)&mad.mad_hdr.tid = 0; /* resetting tid to 0 */
328 mad.mad_hdr.attr_id = cpu_to_be16(IB_SA_ATTR_MC_MEMBER_REC);
329 mad.mad_hdr.attr_mod = cpu_to_be32(0);
330 mad.sa_hdr.sm_key = req_sa_mad->sa_hdr.sm_key;
331 mad.sa_hdr.attr_offset = cpu_to_be16(7);
332 mad.sa_hdr.comp_mask = 0; /* ignored on responses, see IBTA spec */
334 *sa_data = group->rec;
336 /* reconstruct VF's requested join_state and port_gid */
337 sa_data->scope_join_state &= 0xf0;
338 sa_data->scope_join_state |= (group->func[slave].join_state & 0x0f);
339 memcpy(&sa_data->port_gid, &req_sa_data->port_gid, sizeof req_sa_data->port_gid);
341 ret = send_mad_to_slave(slave, group->demux, (struct ib_mad *)&mad);
345 static int check_selector(ib_sa_comp_mask comp_mask,
346 ib_sa_comp_mask selector_mask,
347 ib_sa_comp_mask value_mask,
348 u8 src_value, u8 dst_value)
351 u8 selector = dst_value >> 6;
355 if (!(comp_mask & selector_mask) || !(comp_mask & value_mask))
360 err = (src_value <= dst_value);
363 err = (src_value >= dst_value);
366 err = (src_value != dst_value);
376 static u16 cmp_rec(struct ib_sa_mcmember_data *src,
377 struct ib_sa_mcmember_data *dst, ib_sa_comp_mask comp_mask)
379 /* src is group record, dst is request record */
380 /* MGID must already match */
381 /* Port_GID we always replace to our Port_GID, so it is a match */
383 #define MAD_STATUS_REQ_INVALID 0x0200
384 if (comp_mask & IB_SA_MCMEMBER_REC_QKEY && src->qkey != dst->qkey)
385 return MAD_STATUS_REQ_INVALID;
386 if (comp_mask & IB_SA_MCMEMBER_REC_MLID && src->mlid != dst->mlid)
387 return MAD_STATUS_REQ_INVALID;
388 if (check_selector(comp_mask, IB_SA_MCMEMBER_REC_MTU_SELECTOR,
389 IB_SA_MCMEMBER_REC_MTU,
390 src->mtusel_mtu, dst->mtusel_mtu))
391 return MAD_STATUS_REQ_INVALID;
392 if (comp_mask & IB_SA_MCMEMBER_REC_TRAFFIC_CLASS &&
393 src->tclass != dst->tclass)
394 return MAD_STATUS_REQ_INVALID;
395 if (comp_mask & IB_SA_MCMEMBER_REC_PKEY && src->pkey != dst->pkey)
396 return MAD_STATUS_REQ_INVALID;
397 if (check_selector(comp_mask, IB_SA_MCMEMBER_REC_RATE_SELECTOR,
398 IB_SA_MCMEMBER_REC_RATE,
399 src->ratesel_rate, dst->ratesel_rate))
400 return MAD_STATUS_REQ_INVALID;
401 if (check_selector(comp_mask,
402 IB_SA_MCMEMBER_REC_PACKET_LIFE_TIME_SELECTOR,
403 IB_SA_MCMEMBER_REC_PACKET_LIFE_TIME,
404 src->lifetmsel_lifetm, dst->lifetmsel_lifetm))
405 return MAD_STATUS_REQ_INVALID;
406 if (comp_mask & IB_SA_MCMEMBER_REC_SL &&
407 (be32_to_cpu(src->sl_flowlabel_hoplimit) & 0xf0000000) !=
408 (be32_to_cpu(dst->sl_flowlabel_hoplimit) & 0xf0000000))
409 return MAD_STATUS_REQ_INVALID;
410 if (comp_mask & IB_SA_MCMEMBER_REC_FLOW_LABEL &&
411 (be32_to_cpu(src->sl_flowlabel_hoplimit) & 0x0fffff00) !=
412 (be32_to_cpu(dst->sl_flowlabel_hoplimit) & 0x0fffff00))
413 return MAD_STATUS_REQ_INVALID;
414 if (comp_mask & IB_SA_MCMEMBER_REC_HOP_LIMIT &&
415 (be32_to_cpu(src->sl_flowlabel_hoplimit) & 0x000000ff) !=
416 (be32_to_cpu(dst->sl_flowlabel_hoplimit) & 0x000000ff))
417 return MAD_STATUS_REQ_INVALID;
418 if (comp_mask & IB_SA_MCMEMBER_REC_SCOPE &&
419 (src->scope_join_state & 0xf0) !=
420 (dst->scope_join_state & 0xf0))
421 return MAD_STATUS_REQ_INVALID;
423 /* join_state checked separately, proxy_join ignored */
428 /* release group, return 1 if this was last release and group is destroyed
429 * timout work is canceled sync */
430 static int release_group(struct mcast_group *group, int from_timeout_handler)
432 struct mlx4_ib_demux_ctx *ctx = group->demux;
435 mutex_lock(&ctx->mcg_table_lock);
436 mutex_lock(&group->lock);
437 if (atomic_dec_and_test(&group->refcount)) {
438 if (!from_timeout_handler) {
439 if (group->state != MCAST_IDLE &&
440 !cancel_delayed_work(&group->timeout_work)) {
441 atomic_inc(&group->refcount);
442 mutex_unlock(&group->lock);
443 mutex_unlock(&ctx->mcg_table_lock);
448 nzgroup = memcmp(&group->rec.mgid, &mgid0, sizeof mgid0);
450 del_sysfs_port_mcg_attr(ctx->dev, ctx->port, &group->dentry.attr);
451 if (!list_empty(&group->pending_list))
452 mcg_warn_group(group, "releasing a group with non empty pending list\n");
454 rb_erase(&group->node, &ctx->mcg_table);
455 list_del_init(&group->mgid0_list);
456 mutex_unlock(&group->lock);
457 mutex_unlock(&ctx->mcg_table_lock);
461 mutex_unlock(&group->lock);
462 mutex_unlock(&ctx->mcg_table_lock);
467 static void adjust_membership(struct mcast_group *group, u8 join_state, int inc)
471 for (i = 0; i < 3; i++, join_state >>= 1)
472 if (join_state & 0x1)
473 group->members[i] += inc;
476 static u8 get_leave_state(struct mcast_group *group)
481 for (i = 0; i < 3; i++)
482 if (!group->members[i])
483 leave_state |= (1 << i);
485 return leave_state & (group->rec.scope_join_state & 7);
488 static int join_group(struct mcast_group *group, int slave, u8 join_mask)
493 /* remove bits that slave is already member of, and adjust */
494 join_state = join_mask & (~group->func[slave].join_state);
495 adjust_membership(group, join_state, 1);
496 group->func[slave].join_state |= join_state;
497 if (group->func[slave].state != MCAST_MEMBER && join_state) {
498 group->func[slave].state = MCAST_MEMBER;
504 static int leave_group(struct mcast_group *group, int slave, u8 leave_state)
508 adjust_membership(group, leave_state, -1);
509 group->func[slave].join_state &= ~leave_state;
510 if (!group->func[slave].join_state) {
511 group->func[slave].state = MCAST_NOT_MEMBER;
517 static int check_leave(struct mcast_group *group, int slave, u8 leave_mask)
519 if (group->func[slave].state != MCAST_MEMBER)
520 return MAD_STATUS_REQ_INVALID;
522 /* make sure we're not deleting unset bits */
523 if (~group->func[slave].join_state & leave_mask)
524 return MAD_STATUS_REQ_INVALID;
527 return MAD_STATUS_REQ_INVALID;
532 static void mlx4_ib_mcg_timeout_handler(struct work_struct *work)
534 struct delayed_work *delay = to_delayed_work(work);
535 struct mcast_group *group;
536 struct mcast_req *req = NULL;
538 group = container_of(delay, typeof(*group), timeout_work);
540 mutex_lock(&group->lock);
541 if (group->state == MCAST_JOIN_SENT) {
542 if (!list_empty(&group->pending_list)) {
543 req = list_first_entry(&group->pending_list, struct mcast_req, group_list);
544 list_del(&req->group_list);
545 list_del(&req->func_list);
546 --group->func[req->func].num_pend_reqs;
547 mutex_unlock(&group->lock);
549 if (memcmp(&group->rec.mgid, &mgid0, sizeof mgid0)) {
550 if (release_group(group, 1))
556 mutex_lock(&group->lock);
558 mcg_warn_group(group, "DRIVER BUG\n");
559 } else if (group->state == MCAST_LEAVE_SENT) {
560 if (group->rec.scope_join_state & 7)
561 group->rec.scope_join_state &= 0xf8;
562 group->state = MCAST_IDLE;
563 mutex_unlock(&group->lock);
564 if (release_group(group, 1))
566 mutex_lock(&group->lock);
568 mcg_warn_group(group, "invalid state %s\n", get_state_string(group->state));
569 group->state = MCAST_IDLE;
570 atomic_inc(&group->refcount);
571 if (!queue_work(group->demux->mcg_wq, &group->work))
572 safe_atomic_dec(&group->refcount);
574 mutex_unlock(&group->lock);
577 static int handle_leave_req(struct mcast_group *group, u8 leave_mask,
578 struct mcast_req *req)
583 leave_mask = group->func[req->func].join_state;
585 status = check_leave(group, req->func, leave_mask);
587 leave_group(group, req->func, leave_mask);
590 send_reply_to_slave(req->func, group, &req->sa_mad, status);
591 --group->func[req->func].num_pend_reqs;
592 list_del(&req->group_list);
593 list_del(&req->func_list);
598 static int handle_join_req(struct mcast_group *group, u8 join_mask,
599 struct mcast_req *req)
601 u8 group_join_state = group->rec.scope_join_state & 7;
604 struct ib_sa_mcmember_data *sa_data = (struct ib_sa_mcmember_data *)req->sa_mad.data;
606 if (join_mask == (group_join_state & join_mask)) {
607 /* port's membership need not change */
608 status = cmp_rec(&group->rec, sa_data, req->sa_mad.sa_hdr.comp_mask);
610 join_group(group, req->func, join_mask);
612 --group->func[req->func].num_pend_reqs;
613 send_reply_to_slave(req->func, group, &req->sa_mad, status);
614 list_del(&req->group_list);
615 list_del(&req->func_list);
619 /* port's membership needs to be updated */
620 group->prev_state = group->state;
621 if (send_join_to_wire(group, &req->sa_mad)) {
622 --group->func[req->func].num_pend_reqs;
623 list_del(&req->group_list);
624 list_del(&req->func_list);
627 group->state = group->prev_state;
629 group->state = MCAST_JOIN_SENT;
635 static void mlx4_ib_mcg_work_handler(struct work_struct *work)
637 struct mcast_group *group;
638 struct mcast_req *req = NULL;
639 struct ib_sa_mcmember_data *sa_data;
641 int rc = 1; /* release_count - this is for the scheduled work */
645 group = container_of(work, typeof(*group), work);
647 mutex_lock(&group->lock);
649 /* First, let's see if a response from SM is waiting regarding this group.
650 * If so, we need to update the group's REC. If this is a bad response, we
651 * may need to send a bad response to a VF waiting for it. If VF is waiting
652 * and this is a good response, the VF will be answered later in this func. */
653 if (group->state == MCAST_RESP_READY) {
654 /* cancels mlx4_ib_mcg_timeout_handler */
655 cancel_delayed_work(&group->timeout_work);
656 status = be16_to_cpu(group->response_sa_mad.mad_hdr.status);
657 method = group->response_sa_mad.mad_hdr.method;
658 if (group->last_req_tid != group->response_sa_mad.mad_hdr.tid) {
659 mcg_warn_group(group, "Got MAD response to existing MGID but wrong TID, dropping. Resp TID=%llx, group TID=%llx\n",
660 be64_to_cpu(group->response_sa_mad.mad_hdr.tid),
661 be64_to_cpu(group->last_req_tid));
662 group->state = group->prev_state;
663 goto process_requests;
666 if (!list_empty(&group->pending_list))
667 req = list_first_entry(&group->pending_list,
668 struct mcast_req, group_list);
669 if ((method == IB_MGMT_METHOD_GET_RESP)) {
671 send_reply_to_slave(req->func, group, &req->sa_mad, status);
672 --group->func[req->func].num_pend_reqs;
673 list_del(&req->group_list);
674 list_del(&req->func_list);
678 mcg_warn_group(group, "no request for failed join\n");
679 } else if (method == IB_SA_METHOD_DELETE_RESP && group->demux->flushing)
685 resp_join_state = ((struct ib_sa_mcmember_data *)
686 group->response_sa_mad.data)->scope_join_state & 7;
687 cur_join_state = group->rec.scope_join_state & 7;
689 if (method == IB_MGMT_METHOD_GET_RESP) {
690 /* successfull join */
691 if (!cur_join_state && resp_join_state)
693 } else if (!resp_join_state)
695 memcpy(&group->rec, group->response_sa_mad.data, sizeof group->rec);
697 group->state = MCAST_IDLE;
701 /* We should now go over pending join/leave requests, as long as we are idle. */
702 while (!list_empty(&group->pending_list) && group->state == MCAST_IDLE) {
703 req = list_first_entry(&group->pending_list, struct mcast_req,
705 sa_data = (struct ib_sa_mcmember_data *)req->sa_mad.data;
706 req_join_state = sa_data->scope_join_state & 0x7;
708 /* For a leave request, we will immediately answer the VF, and
709 * update our internal counters. The actual leave will be sent
710 * to SM later, if at all needed. We dequeue the request now. */
711 if (req->sa_mad.mad_hdr.method == IB_SA_METHOD_DELETE)
712 rc += handle_leave_req(group, req_join_state, req);
714 rc += handle_join_req(group, req_join_state, req);
718 if (group->state == MCAST_IDLE) {
719 req_join_state = get_leave_state(group);
720 if (req_join_state) {
721 group->rec.scope_join_state &= ~req_join_state;
722 group->prev_state = group->state;
723 if (send_leave_to_wire(group, req_join_state)) {
724 group->state = group->prev_state;
727 group->state = MCAST_LEAVE_SENT;
731 if (!list_empty(&group->pending_list) && group->state == MCAST_IDLE)
732 goto process_requests;
733 mutex_unlock(&group->lock);
736 release_group(group, 0);
739 static struct mcast_group *search_relocate_mgid0_group(struct mlx4_ib_demux_ctx *ctx,
741 union ib_gid *new_mgid)
743 struct mcast_group *group = NULL, *cur_group;
744 struct mcast_req *req;
745 struct list_head *pos;
748 mutex_lock(&ctx->mcg_table_lock);
749 list_for_each_safe(pos, n, &ctx->mcg_mgid0_list) {
750 group = list_entry(pos, struct mcast_group, mgid0_list);
751 mutex_lock(&group->lock);
752 if (group->last_req_tid == tid) {
753 if (memcmp(new_mgid, &mgid0, sizeof mgid0)) {
754 group->rec.mgid = *new_mgid;
755 sprintf(group->name, "%016llx%016llx",
756 be64_to_cpu(group->rec.mgid.global.subnet_prefix),
757 be64_to_cpu(group->rec.mgid.global.interface_id));
758 list_del_init(&group->mgid0_list);
759 cur_group = mcast_insert(ctx, group);
761 /* A race between our code and SM. Silently cleaning the new one */
762 req = list_first_entry(&group->pending_list,
763 struct mcast_req, group_list);
764 --group->func[req->func].num_pend_reqs;
765 list_del(&req->group_list);
766 list_del(&req->func_list);
768 mutex_unlock(&group->lock);
769 mutex_unlock(&ctx->mcg_table_lock);
770 release_group(group, 0);
774 atomic_inc(&group->refcount);
775 add_sysfs_port_mcg_attr(ctx->dev, ctx->port, &group->dentry.attr);
776 mutex_unlock(&group->lock);
777 mutex_unlock(&ctx->mcg_table_lock);
780 struct mcast_req *tmp1, *tmp2;
782 list_del(&group->mgid0_list);
783 if (!list_empty(&group->pending_list) && group->state != MCAST_IDLE)
784 cancel_delayed_work_sync(&group->timeout_work);
786 list_for_each_entry_safe(tmp1, tmp2, &group->pending_list, group_list) {
787 list_del(&tmp1->group_list);
790 mutex_unlock(&group->lock);
791 mutex_unlock(&ctx->mcg_table_lock);
796 mutex_unlock(&group->lock);
798 mutex_unlock(&ctx->mcg_table_lock);
803 static ssize_t sysfs_show_group(struct device *dev,
804 struct device_attribute *attr, char *buf);
806 static struct mcast_group *acquire_group(struct mlx4_ib_demux_ctx *ctx,
807 union ib_gid *mgid, int create,
810 struct mcast_group *group, *cur_group;
814 is_mgid0 = !memcmp(&mgid0, mgid, sizeof mgid0);
816 group = mcast_find(ctx, mgid);
822 return ERR_PTR(-ENOENT);
824 group = kzalloc(sizeof *group, gfp_mask);
826 return ERR_PTR(-ENOMEM);
829 group->rec.mgid = *mgid;
830 INIT_LIST_HEAD(&group->pending_list);
831 INIT_LIST_HEAD(&group->mgid0_list);
832 for (i = 0; i < MAX_VFS; ++i)
833 INIT_LIST_HEAD(&group->func[i].pending);
834 INIT_WORK(&group->work, mlx4_ib_mcg_work_handler);
835 INIT_DELAYED_WORK(&group->timeout_work, mlx4_ib_mcg_timeout_handler);
836 mutex_init(&group->lock);
837 sprintf(group->name, "%016llx%016llx",
838 be64_to_cpu(group->rec.mgid.global.subnet_prefix),
839 be64_to_cpu(group->rec.mgid.global.interface_id));
840 sysfs_attr_init(&group->dentry.attr);
841 group->dentry.show = sysfs_show_group;
842 group->dentry.store = NULL;
843 group->dentry.attr.name = group->name;
844 group->dentry.attr.mode = 0400;
845 group->state = MCAST_IDLE;
848 list_add(&group->mgid0_list, &ctx->mcg_mgid0_list);
852 cur_group = mcast_insert(ctx, group);
854 mcg_warn("group just showed up %s - confused\n", cur_group->name);
856 return ERR_PTR(-EINVAL);
859 add_sysfs_port_mcg_attr(ctx->dev, ctx->port, &group->dentry.attr);
862 atomic_inc(&group->refcount);
866 static void queue_req(struct mcast_req *req)
868 struct mcast_group *group = req->group;
870 atomic_inc(&group->refcount); /* for the request */
871 atomic_inc(&group->refcount); /* for scheduling the work */
872 list_add_tail(&req->group_list, &group->pending_list);
873 list_add_tail(&req->func_list, &group->func[req->func].pending);
874 /* calls mlx4_ib_mcg_work_handler */
875 if (!queue_work(group->demux->mcg_wq, &group->work))
876 safe_atomic_dec(&group->refcount);
879 int mlx4_ib_mcg_demux_handler(struct ib_device *ibdev, int port, int slave,
880 struct ib_sa_mad *mad)
882 struct mlx4_ib_dev *dev = to_mdev(ibdev);
883 struct ib_sa_mcmember_data *rec = (struct ib_sa_mcmember_data *)mad->data;
884 struct mlx4_ib_demux_ctx *ctx = &dev->sriov.demux[port - 1];
885 struct mcast_group *group;
887 switch (mad->mad_hdr.method) {
888 case IB_MGMT_METHOD_GET_RESP:
889 case IB_SA_METHOD_DELETE_RESP:
890 mutex_lock(&ctx->mcg_table_lock);
891 group = acquire_group(ctx, &rec->mgid, 0, GFP_KERNEL);
892 mutex_unlock(&ctx->mcg_table_lock);
894 if (mad->mad_hdr.method == IB_MGMT_METHOD_GET_RESP) {
895 __be64 tid = mad->mad_hdr.tid;
896 *(u8 *)(&tid) = (u8)slave; /* in group we kept the modified TID */
897 group = search_relocate_mgid0_group(ctx, tid, &rec->mgid);
905 mutex_lock(&group->lock);
906 group->response_sa_mad = *mad;
907 group->prev_state = group->state;
908 group->state = MCAST_RESP_READY;
909 /* calls mlx4_ib_mcg_work_handler */
910 atomic_inc(&group->refcount);
911 if (!queue_work(ctx->mcg_wq, &group->work))
912 safe_atomic_dec(&group->refcount);
913 mutex_unlock(&group->lock);
914 release_group(group, 0);
915 return 1; /* consumed */
916 case IB_MGMT_METHOD_SET:
917 case IB_SA_METHOD_GET_TABLE:
918 case IB_SA_METHOD_GET_TABLE_RESP:
919 case IB_SA_METHOD_DELETE:
920 return 0; /* not consumed, pass-through to guest over tunnel */
922 mcg_warn("In demux, port %d: unexpected MCMember method: 0x%x, dropping\n",
923 port, mad->mad_hdr.method);
924 return 1; /* consumed */
928 int mlx4_ib_mcg_multiplex_handler(struct ib_device *ibdev, int port,
929 int slave, struct ib_sa_mad *sa_mad)
931 struct mlx4_ib_dev *dev = to_mdev(ibdev);
932 struct ib_sa_mcmember_data *rec = (struct ib_sa_mcmember_data *)sa_mad->data;
933 struct mlx4_ib_demux_ctx *ctx = &dev->sriov.demux[port - 1];
934 struct mcast_group *group;
935 struct mcast_req *req;
941 switch (sa_mad->mad_hdr.method) {
942 case IB_MGMT_METHOD_SET:
944 case IB_SA_METHOD_DELETE:
945 req = kzalloc(sizeof *req, GFP_KERNEL);
950 req->sa_mad = *sa_mad;
952 mutex_lock(&ctx->mcg_table_lock);
953 group = acquire_group(ctx, &rec->mgid, may_create, GFP_KERNEL);
954 mutex_unlock(&ctx->mcg_table_lock);
957 return PTR_ERR(group);
959 mutex_lock(&group->lock);
960 if (group->func[slave].num_pend_reqs > MAX_PEND_REQS_PER_FUNC) {
961 mutex_unlock(&group->lock);
962 mcg_warn_group(group, "Port %d, Func %d has too many pending requests (%d), dropping\n",
963 port, slave, MAX_PEND_REQS_PER_FUNC);
964 release_group(group, 0);
968 ++group->func[slave].num_pend_reqs;
971 mutex_unlock(&group->lock);
972 release_group(group, 0);
973 return 1; /* consumed */
974 case IB_SA_METHOD_GET_TABLE:
975 case IB_MGMT_METHOD_GET_RESP:
976 case IB_SA_METHOD_GET_TABLE_RESP:
977 case IB_SA_METHOD_DELETE_RESP:
978 return 0; /* not consumed, pass-through */
980 mcg_warn("In multiplex, port %d, func %d: unexpected MCMember method: 0x%x, dropping\n",
981 port, slave, sa_mad->mad_hdr.method);
982 return 1; /* consumed */
986 static ssize_t sysfs_show_group(struct device *dev,
987 struct device_attribute *attr, char *buf)
989 struct mcast_group *group =
990 container_of(attr, struct mcast_group, dentry);
991 struct mcast_req *req = NULL;
992 char pending_str[40];
997 if (group->state == MCAST_IDLE)
998 sprintf(state_str, "%s", get_state_string(group->state));
1000 sprintf(state_str, "%s(TID=0x%llx)",
1001 get_state_string(group->state),
1002 be64_to_cpu(group->last_req_tid));
1003 if (list_empty(&group->pending_list)) {
1004 sprintf(pending_str, "No");
1006 req = list_first_entry(&group->pending_list, struct mcast_req, group_list);
1007 sprintf(pending_str, "Yes(TID=0x%llx)",
1008 be64_to_cpu(req->sa_mad.mad_hdr.tid));
1010 len += sprintf(buf + len, "%1d [%02d,%02d,%02d] %4d %4s %5s ",
1011 group->rec.scope_join_state & 0xf,
1012 group->members[2], group->members[1], group->members[0],
1013 atomic_read(&group->refcount),
1016 for (f = 0; f < MAX_VFS; ++f)
1017 if (group->func[f].state == MCAST_MEMBER)
1018 len += sprintf(buf + len, "%d[%1x] ",
1019 f, group->func[f].join_state);
1021 len += sprintf(buf + len, "\t\t(%4hx %4x %2x %2x %2x %2x %2x "
1022 "%4x %4x %2x %2x)\n",
1023 be16_to_cpu(group->rec.pkey),
1024 be32_to_cpu(group->rec.qkey),
1025 (group->rec.mtusel_mtu & 0xc0) >> 6,
1026 group->rec.mtusel_mtu & 0x3f,
1028 (group->rec.ratesel_rate & 0xc0) >> 6,
1029 group->rec.ratesel_rate & 0x3f,
1030 (be32_to_cpu(group->rec.sl_flowlabel_hoplimit) & 0xf0000000) >> 28,
1031 (be32_to_cpu(group->rec.sl_flowlabel_hoplimit) & 0x0fffff00) >> 8,
1032 be32_to_cpu(group->rec.sl_flowlabel_hoplimit) & 0x000000ff,
1033 group->rec.proxy_join);
1038 int mlx4_ib_mcg_port_init(struct mlx4_ib_demux_ctx *ctx)
1042 atomic_set(&ctx->tid, 0);
1043 sprintf(name, "mlx4_ib_mcg%d", ctx->port);
1044 ctx->mcg_wq = create_singlethread_workqueue(name);
1048 mutex_init(&ctx->mcg_table_lock);
1049 ctx->mcg_table = RB_ROOT;
1050 INIT_LIST_HEAD(&ctx->mcg_mgid0_list);
1056 static void force_clean_group(struct mcast_group *group)
1058 struct mcast_req *req, *tmp
1060 list_for_each_entry_safe(req, tmp, &group->pending_list, group_list) {
1061 list_del(&req->group_list);
1064 del_sysfs_port_mcg_attr(group->demux->dev, group->demux->port, &group->dentry.attr);
1065 rb_erase(&group->node, &group->demux->mcg_table);
1069 static void _mlx4_ib_mcg_port_cleanup(struct mlx4_ib_demux_ctx *ctx, int destroy_wq)
1073 struct mcast_group *group;
1081 for (i = 0; i < MAX_VFS; ++i)
1082 clean_vf_mcast(ctx, i);
1084 end = jiffies + msecs_to_jiffies(MAD_TIMEOUT_MS + 3000);
1087 mutex_lock(&ctx->mcg_table_lock);
1088 for (p = rb_first(&ctx->mcg_table); p; p = rb_next(p))
1090 mutex_unlock(&ctx->mcg_table_lock);
1095 } while (time_after(end, jiffies));
1097 flush_workqueue(ctx->mcg_wq);
1099 destroy_workqueue(ctx->mcg_wq);
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);
1107 force_clean_group(group);
1109 mutex_unlock(&ctx->mcg_table_lock);
1116 struct work_struct work;
1117 struct mlx4_ib_demux_ctx *ctx;
1121 static void mcg_clean_task(struct work_struct *work)
1123 struct clean_work *cw = container_of(work, struct clean_work, work);
1125 _mlx4_ib_mcg_port_cleanup(cw->ctx, cw->destroy_wq);
1129 void mlx4_ib_mcg_port_cleanup(struct mlx4_ib_demux_ctx *ctx, int destroy_wq)
1131 struct clean_work *work;
1134 _mlx4_ib_mcg_port_cleanup(ctx, destroy_wq);
1138 work = kmalloc(sizeof *work, GFP_KERNEL);
1140 mcg_warn("failed allocating work for cleanup\n");
1145 work->destroy_wq = destroy_wq;
1146 INIT_WORK(&work->work, mcg_clean_task);
1147 queue_work(clean_wq, &work->work);
1150 static void build_leave_mad(struct mcast_req *req)
1152 struct ib_sa_mad *mad = &req->sa_mad;
1154 mad->mad_hdr.method = IB_SA_METHOD_DELETE;
1158 static void clear_pending_reqs(struct mcast_group *group, int vf)
1160 struct mcast_req *req, *tmp, *group_first = NULL;
1164 if (!list_empty(&group->pending_list))
1165 group_first = list_first_entry(&group->pending_list, struct mcast_req, group_list);
1167 list_for_each_entry_safe(req, tmp, &group->func[vf].pending, func_list) {
1169 if (group_first == req &&
1170 (group->state == MCAST_JOIN_SENT ||
1171 group->state == MCAST_LEAVE_SENT)) {
1172 clear = cancel_delayed_work(&group->timeout_work);
1174 group->state = MCAST_IDLE;
1177 --group->func[vf].num_pend_reqs;
1178 list_del(&req->group_list);
1179 list_del(&req->func_list);
1181 atomic_dec(&group->refcount);
1185 if (!pend && (!list_empty(&group->func[vf].pending) || group->func[vf].num_pend_reqs)) {
1186 mcg_warn_group(group, "DRIVER BUG: list_empty %d, num_pend_reqs %d\n",
1187 list_empty(&group->func[vf].pending), group->func[vf].num_pend_reqs);
1191 static int push_deleteing_req(struct mcast_group *group, int slave)
1193 struct mcast_req *req;
1194 struct mcast_req *pend_req;
1196 if (!group->func[slave].join_state)
1199 req = kzalloc(sizeof *req, GFP_KERNEL);
1201 mcg_warn_group(group, "failed allocation - may leave stall groups\n");
1205 if (!list_empty(&group->func[slave].pending)) {
1206 pend_req = list_entry(group->func[slave].pending.prev, struct mcast_req, group_list);
1207 if (pend_req->clean) {
1216 ++group->func[slave].num_pend_reqs;
1217 build_leave_mad(req);
1222 void clean_vf_mcast(struct mlx4_ib_demux_ctx *ctx, int slave)
1224 struct mcast_group *group;
1227 mutex_lock(&ctx->mcg_table_lock);
1228 for (p = rb_first(&ctx->mcg_table); p; p = rb_next(p)) {
1229 group = rb_entry(p, struct mcast_group, node);
1230 mutex_lock(&group->lock);
1231 if (atomic_read(&group->refcount)) {
1232 /* clear pending requests of this VF */
1233 clear_pending_reqs(group, slave);
1234 push_deleteing_req(group, slave);
1236 mutex_unlock(&group->lock);
1238 mutex_unlock(&ctx->mcg_table_lock);
1242 int mlx4_ib_mcg_init(void)
1244 clean_wq = create_singlethread_workqueue("mlx4_ib_mcg");
1251 void mlx4_ib_mcg_destroy(void)
1253 destroy_workqueue(clean_wq);