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),
219 ctx->port, IB_QPT_GSI, 0, 1, IB_QP1_QKEY,
220 &ah_attr, NULL, mad);
223 static int send_mad_to_slave(int slave, struct mlx4_ib_demux_ctx *ctx,
226 struct mlx4_ib_dev *dev = ctx->dev;
227 struct ib_mad_agent *agent = dev->send_agent[ctx->port - 1][1];
229 struct ib_ah_attr ah_attr;
231 /* Our agent might not yet be registered when mads start to arrive */
235 ib_query_ah(dev->sm_ah[ctx->port - 1], &ah_attr);
237 if (ib_find_cached_pkey(&dev->ib_dev, ctx->port, IB_DEFAULT_PKEY_FULL, &wc.pkey_index))
240 wc.dlid_path_bits = 0;
241 wc.port_num = ctx->port;
242 wc.slid = ah_attr.dlid; /* opensm lid */
244 return mlx4_ib_send_to_slave(dev, slave, ctx->port, IB_QPT_GSI, &wc, NULL, mad);
247 static int send_join_to_wire(struct mcast_group *group, struct ib_sa_mad *sa_mad)
249 struct ib_sa_mad mad;
250 struct ib_sa_mcmember_data *sa_mad_data = (struct ib_sa_mcmember_data *)&mad.data;
253 /* we rely on a mad request as arrived from a VF */
254 memcpy(&mad, sa_mad, sizeof mad);
256 /* fix port GID to be the real one (slave 0) */
257 sa_mad_data->port_gid.global.interface_id = group->demux->guid_cache[0];
259 /* assign our own TID */
260 mad.mad_hdr.tid = mlx4_ib_get_new_demux_tid(group->demux);
261 group->last_req_tid = mad.mad_hdr.tid; /* keep it for later validation */
263 ret = send_mad_to_wire(group->demux, (struct ib_mad *)&mad);
264 /* set timeout handler */
266 /* calls mlx4_ib_mcg_timeout_handler */
267 queue_delayed_work(group->demux->mcg_wq, &group->timeout_work,
268 msecs_to_jiffies(MAD_TIMEOUT_MS));
274 static int send_leave_to_wire(struct mcast_group *group, u8 join_state)
276 struct ib_sa_mad mad;
277 struct ib_sa_mcmember_data *sa_data = (struct ib_sa_mcmember_data *)&mad.data;
280 memset(&mad, 0, sizeof mad);
281 mad.mad_hdr.base_version = 1;
282 mad.mad_hdr.mgmt_class = IB_MGMT_CLASS_SUBN_ADM;
283 mad.mad_hdr.class_version = 2;
284 mad.mad_hdr.method = IB_SA_METHOD_DELETE;
285 mad.mad_hdr.status = cpu_to_be16(0);
286 mad.mad_hdr.class_specific = cpu_to_be16(0);
287 mad.mad_hdr.tid = mlx4_ib_get_new_demux_tid(group->demux);
288 group->last_req_tid = mad.mad_hdr.tid; /* keep it for later validation */
289 mad.mad_hdr.attr_id = cpu_to_be16(IB_SA_ATTR_MC_MEMBER_REC);
290 mad.mad_hdr.attr_mod = cpu_to_be32(0);
291 mad.sa_hdr.sm_key = 0x0;
292 mad.sa_hdr.attr_offset = cpu_to_be16(7);
293 mad.sa_hdr.comp_mask = IB_SA_MCMEMBER_REC_MGID |
294 IB_SA_MCMEMBER_REC_PORT_GID | IB_SA_MCMEMBER_REC_JOIN_STATE;
296 *sa_data = group->rec;
297 sa_data->scope_join_state = join_state;
299 ret = send_mad_to_wire(group->demux, (struct ib_mad *)&mad);
301 group->state = MCAST_IDLE;
303 /* set timeout handler */
305 /* calls mlx4_ib_mcg_timeout_handler */
306 queue_delayed_work(group->demux->mcg_wq, &group->timeout_work,
307 msecs_to_jiffies(MAD_TIMEOUT_MS));
313 static int send_reply_to_slave(int slave, struct mcast_group *group,
314 struct ib_sa_mad *req_sa_mad, u16 status)
316 struct ib_sa_mad mad;
317 struct ib_sa_mcmember_data *sa_data = (struct ib_sa_mcmember_data *)&mad.data;
318 struct ib_sa_mcmember_data *req_sa_data = (struct ib_sa_mcmember_data *)&req_sa_mad->data;
321 memset(&mad, 0, sizeof mad);
322 mad.mad_hdr.base_version = 1;
323 mad.mad_hdr.mgmt_class = IB_MGMT_CLASS_SUBN_ADM;
324 mad.mad_hdr.class_version = 2;
325 mad.mad_hdr.method = IB_MGMT_METHOD_GET_RESP;
326 mad.mad_hdr.status = cpu_to_be16(status);
327 mad.mad_hdr.class_specific = cpu_to_be16(0);
328 mad.mad_hdr.tid = req_sa_mad->mad_hdr.tid;
329 *(u8 *)&mad.mad_hdr.tid = 0; /* resetting tid to 0 */
330 mad.mad_hdr.attr_id = cpu_to_be16(IB_SA_ATTR_MC_MEMBER_REC);
331 mad.mad_hdr.attr_mod = cpu_to_be32(0);
332 mad.sa_hdr.sm_key = req_sa_mad->sa_hdr.sm_key;
333 mad.sa_hdr.attr_offset = cpu_to_be16(7);
334 mad.sa_hdr.comp_mask = 0; /* ignored on responses, see IBTA spec */
336 *sa_data = group->rec;
338 /* reconstruct VF's requested join_state and port_gid */
339 sa_data->scope_join_state &= 0xf0;
340 sa_data->scope_join_state |= (group->func[slave].join_state & 0x0f);
341 memcpy(&sa_data->port_gid, &req_sa_data->port_gid, sizeof req_sa_data->port_gid);
343 ret = send_mad_to_slave(slave, group->demux, (struct ib_mad *)&mad);
347 static int check_selector(ib_sa_comp_mask comp_mask,
348 ib_sa_comp_mask selector_mask,
349 ib_sa_comp_mask value_mask,
350 u8 src_value, u8 dst_value)
353 u8 selector = dst_value >> 6;
357 if (!(comp_mask & selector_mask) || !(comp_mask & value_mask))
362 err = (src_value <= dst_value);
365 err = (src_value >= dst_value);
368 err = (src_value != dst_value);
378 static u16 cmp_rec(struct ib_sa_mcmember_data *src,
379 struct ib_sa_mcmember_data *dst, ib_sa_comp_mask comp_mask)
381 /* src is group record, dst is request record */
382 /* MGID must already match */
383 /* Port_GID we always replace to our Port_GID, so it is a match */
385 #define MAD_STATUS_REQ_INVALID 0x0200
386 if (comp_mask & IB_SA_MCMEMBER_REC_QKEY && src->qkey != dst->qkey)
387 return MAD_STATUS_REQ_INVALID;
388 if (comp_mask & IB_SA_MCMEMBER_REC_MLID && src->mlid != dst->mlid)
389 return MAD_STATUS_REQ_INVALID;
390 if (check_selector(comp_mask, IB_SA_MCMEMBER_REC_MTU_SELECTOR,
391 IB_SA_MCMEMBER_REC_MTU,
392 src->mtusel_mtu, dst->mtusel_mtu))
393 return MAD_STATUS_REQ_INVALID;
394 if (comp_mask & IB_SA_MCMEMBER_REC_TRAFFIC_CLASS &&
395 src->tclass != dst->tclass)
396 return MAD_STATUS_REQ_INVALID;
397 if (comp_mask & IB_SA_MCMEMBER_REC_PKEY && src->pkey != dst->pkey)
398 return MAD_STATUS_REQ_INVALID;
399 if (check_selector(comp_mask, IB_SA_MCMEMBER_REC_RATE_SELECTOR,
400 IB_SA_MCMEMBER_REC_RATE,
401 src->ratesel_rate, dst->ratesel_rate))
402 return MAD_STATUS_REQ_INVALID;
403 if (check_selector(comp_mask,
404 IB_SA_MCMEMBER_REC_PACKET_LIFE_TIME_SELECTOR,
405 IB_SA_MCMEMBER_REC_PACKET_LIFE_TIME,
406 src->lifetmsel_lifetm, dst->lifetmsel_lifetm))
407 return MAD_STATUS_REQ_INVALID;
408 if (comp_mask & IB_SA_MCMEMBER_REC_SL &&
409 (be32_to_cpu(src->sl_flowlabel_hoplimit) & 0xf0000000) !=
410 (be32_to_cpu(dst->sl_flowlabel_hoplimit) & 0xf0000000))
411 return MAD_STATUS_REQ_INVALID;
412 if (comp_mask & IB_SA_MCMEMBER_REC_FLOW_LABEL &&
413 (be32_to_cpu(src->sl_flowlabel_hoplimit) & 0x0fffff00) !=
414 (be32_to_cpu(dst->sl_flowlabel_hoplimit) & 0x0fffff00))
415 return MAD_STATUS_REQ_INVALID;
416 if (comp_mask & IB_SA_MCMEMBER_REC_HOP_LIMIT &&
417 (be32_to_cpu(src->sl_flowlabel_hoplimit) & 0x000000ff) !=
418 (be32_to_cpu(dst->sl_flowlabel_hoplimit) & 0x000000ff))
419 return MAD_STATUS_REQ_INVALID;
420 if (comp_mask & IB_SA_MCMEMBER_REC_SCOPE &&
421 (src->scope_join_state & 0xf0) !=
422 (dst->scope_join_state & 0xf0))
423 return MAD_STATUS_REQ_INVALID;
425 /* join_state checked separately, proxy_join ignored */
430 /* release group, return 1 if this was last release and group is destroyed
431 * timout work is canceled sync */
432 static int release_group(struct mcast_group *group, int from_timeout_handler)
434 struct mlx4_ib_demux_ctx *ctx = group->demux;
437 mutex_lock(&ctx->mcg_table_lock);
438 mutex_lock(&group->lock);
439 if (atomic_dec_and_test(&group->refcount)) {
440 if (!from_timeout_handler) {
441 if (group->state != MCAST_IDLE &&
442 !cancel_delayed_work(&group->timeout_work)) {
443 atomic_inc(&group->refcount);
444 mutex_unlock(&group->lock);
445 mutex_unlock(&ctx->mcg_table_lock);
450 nzgroup = memcmp(&group->rec.mgid, &mgid0, sizeof mgid0);
452 del_sysfs_port_mcg_attr(ctx->dev, ctx->port, &group->dentry.attr);
453 if (!list_empty(&group->pending_list))
454 mcg_warn_group(group, "releasing a group with non empty pending list\n");
456 rb_erase(&group->node, &ctx->mcg_table);
457 list_del_init(&group->mgid0_list);
458 mutex_unlock(&group->lock);
459 mutex_unlock(&ctx->mcg_table_lock);
463 mutex_unlock(&group->lock);
464 mutex_unlock(&ctx->mcg_table_lock);
469 static void adjust_membership(struct mcast_group *group, u8 join_state, int inc)
473 for (i = 0; i < 3; i++, join_state >>= 1)
474 if (join_state & 0x1)
475 group->members[i] += inc;
478 static u8 get_leave_state(struct mcast_group *group)
483 for (i = 0; i < 3; i++)
484 if (!group->members[i])
485 leave_state |= (1 << i);
487 return leave_state & (group->rec.scope_join_state & 7);
490 static int join_group(struct mcast_group *group, int slave, u8 join_mask)
495 /* remove bits that slave is already member of, and adjust */
496 join_state = join_mask & (~group->func[slave].join_state);
497 adjust_membership(group, join_state, 1);
498 group->func[slave].join_state |= join_state;
499 if (group->func[slave].state != MCAST_MEMBER && join_state) {
500 group->func[slave].state = MCAST_MEMBER;
506 static int leave_group(struct mcast_group *group, int slave, u8 leave_state)
510 adjust_membership(group, leave_state, -1);
511 group->func[slave].join_state &= ~leave_state;
512 if (!group->func[slave].join_state) {
513 group->func[slave].state = MCAST_NOT_MEMBER;
519 static int check_leave(struct mcast_group *group, int slave, u8 leave_mask)
521 if (group->func[slave].state != MCAST_MEMBER)
522 return MAD_STATUS_REQ_INVALID;
524 /* make sure we're not deleting unset bits */
525 if (~group->func[slave].join_state & leave_mask)
526 return MAD_STATUS_REQ_INVALID;
529 return MAD_STATUS_REQ_INVALID;
534 static void mlx4_ib_mcg_timeout_handler(struct work_struct *work)
536 struct delayed_work *delay = to_delayed_work(work);
537 struct mcast_group *group;
538 struct mcast_req *req = NULL;
540 group = container_of(delay, typeof(*group), timeout_work);
542 mutex_lock(&group->lock);
543 if (group->state == MCAST_JOIN_SENT) {
544 if (!list_empty(&group->pending_list)) {
545 req = list_first_entry(&group->pending_list, struct mcast_req, group_list);
546 list_del(&req->group_list);
547 list_del(&req->func_list);
548 --group->func[req->func].num_pend_reqs;
549 mutex_unlock(&group->lock);
551 if (memcmp(&group->rec.mgid, &mgid0, sizeof mgid0)) {
552 if (release_group(group, 1))
558 mutex_lock(&group->lock);
560 mcg_warn_group(group, "DRIVER BUG\n");
561 } else if (group->state == MCAST_LEAVE_SENT) {
562 if (group->rec.scope_join_state & 7)
563 group->rec.scope_join_state &= 0xf8;
564 group->state = MCAST_IDLE;
565 mutex_unlock(&group->lock);
566 if (release_group(group, 1))
568 mutex_lock(&group->lock);
570 mcg_warn_group(group, "invalid state %s\n", get_state_string(group->state));
571 group->state = MCAST_IDLE;
572 atomic_inc(&group->refcount);
573 if (!queue_work(group->demux->mcg_wq, &group->work))
574 safe_atomic_dec(&group->refcount);
576 mutex_unlock(&group->lock);
579 static int handle_leave_req(struct mcast_group *group, u8 leave_mask,
580 struct mcast_req *req)
585 leave_mask = group->func[req->func].join_state;
587 status = check_leave(group, req->func, leave_mask);
589 leave_group(group, req->func, leave_mask);
592 send_reply_to_slave(req->func, group, &req->sa_mad, status);
593 --group->func[req->func].num_pend_reqs;
594 list_del(&req->group_list);
595 list_del(&req->func_list);
600 static int handle_join_req(struct mcast_group *group, u8 join_mask,
601 struct mcast_req *req)
603 u8 group_join_state = group->rec.scope_join_state & 7;
606 struct ib_sa_mcmember_data *sa_data = (struct ib_sa_mcmember_data *)req->sa_mad.data;
608 if (join_mask == (group_join_state & join_mask)) {
609 /* port's membership need not change */
610 status = cmp_rec(&group->rec, sa_data, req->sa_mad.sa_hdr.comp_mask);
612 join_group(group, req->func, join_mask);
614 --group->func[req->func].num_pend_reqs;
615 send_reply_to_slave(req->func, group, &req->sa_mad, status);
616 list_del(&req->group_list);
617 list_del(&req->func_list);
621 /* port's membership needs to be updated */
622 group->prev_state = group->state;
623 if (send_join_to_wire(group, &req->sa_mad)) {
624 --group->func[req->func].num_pend_reqs;
625 list_del(&req->group_list);
626 list_del(&req->func_list);
629 group->state = group->prev_state;
631 group->state = MCAST_JOIN_SENT;
637 static void mlx4_ib_mcg_work_handler(struct work_struct *work)
639 struct mcast_group *group;
640 struct mcast_req *req = NULL;
641 struct ib_sa_mcmember_data *sa_data;
643 int rc = 1; /* release_count - this is for the scheduled work */
647 group = container_of(work, typeof(*group), work);
649 mutex_lock(&group->lock);
651 /* First, let's see if a response from SM is waiting regarding this group.
652 * If so, we need to update the group's REC. If this is a bad response, we
653 * may need to send a bad response to a VF waiting for it. If VF is waiting
654 * and this is a good response, the VF will be answered later in this func. */
655 if (group->state == MCAST_RESP_READY) {
656 /* cancels mlx4_ib_mcg_timeout_handler */
657 cancel_delayed_work(&group->timeout_work);
658 status = be16_to_cpu(group->response_sa_mad.mad_hdr.status);
659 method = group->response_sa_mad.mad_hdr.method;
660 if (group->last_req_tid != group->response_sa_mad.mad_hdr.tid) {
661 mcg_warn_group(group, "Got MAD response to existing MGID but wrong TID, dropping. Resp TID=%llx, group TID=%llx\n",
662 be64_to_cpu(group->response_sa_mad.mad_hdr.tid),
663 be64_to_cpu(group->last_req_tid));
664 group->state = group->prev_state;
665 goto process_requests;
668 if (!list_empty(&group->pending_list))
669 req = list_first_entry(&group->pending_list,
670 struct mcast_req, group_list);
671 if ((method == IB_MGMT_METHOD_GET_RESP)) {
673 send_reply_to_slave(req->func, group, &req->sa_mad, status);
674 --group->func[req->func].num_pend_reqs;
675 list_del(&req->group_list);
676 list_del(&req->func_list);
680 mcg_warn_group(group, "no request for failed join\n");
681 } else if (method == IB_SA_METHOD_DELETE_RESP && group->demux->flushing)
687 resp_join_state = ((struct ib_sa_mcmember_data *)
688 group->response_sa_mad.data)->scope_join_state & 7;
689 cur_join_state = group->rec.scope_join_state & 7;
691 if (method == IB_MGMT_METHOD_GET_RESP) {
692 /* successfull join */
693 if (!cur_join_state && resp_join_state)
695 } else if (!resp_join_state)
697 memcpy(&group->rec, group->response_sa_mad.data, sizeof group->rec);
699 group->state = MCAST_IDLE;
703 /* We should now go over pending join/leave requests, as long as we are idle. */
704 while (!list_empty(&group->pending_list) && group->state == MCAST_IDLE) {
705 req = list_first_entry(&group->pending_list, struct mcast_req,
707 sa_data = (struct ib_sa_mcmember_data *)req->sa_mad.data;
708 req_join_state = sa_data->scope_join_state & 0x7;
710 /* For a leave request, we will immediately answer the VF, and
711 * update our internal counters. The actual leave will be sent
712 * to SM later, if at all needed. We dequeue the request now. */
713 if (req->sa_mad.mad_hdr.method == IB_SA_METHOD_DELETE)
714 rc += handle_leave_req(group, req_join_state, req);
716 rc += handle_join_req(group, req_join_state, req);
720 if (group->state == MCAST_IDLE) {
721 req_join_state = get_leave_state(group);
722 if (req_join_state) {
723 group->rec.scope_join_state &= ~req_join_state;
724 group->prev_state = group->state;
725 if (send_leave_to_wire(group, req_join_state)) {
726 group->state = group->prev_state;
729 group->state = MCAST_LEAVE_SENT;
733 if (!list_empty(&group->pending_list) && group->state == MCAST_IDLE)
734 goto process_requests;
735 mutex_unlock(&group->lock);
738 release_group(group, 0);
741 static struct mcast_group *search_relocate_mgid0_group(struct mlx4_ib_demux_ctx *ctx,
743 union ib_gid *new_mgid)
745 struct mcast_group *group = NULL, *cur_group;
746 struct mcast_req *req;
747 struct list_head *pos;
750 mutex_lock(&ctx->mcg_table_lock);
751 list_for_each_safe(pos, n, &ctx->mcg_mgid0_list) {
752 group = list_entry(pos, struct mcast_group, mgid0_list);
753 mutex_lock(&group->lock);
754 if (group->last_req_tid == tid) {
755 if (memcmp(new_mgid, &mgid0, sizeof mgid0)) {
756 group->rec.mgid = *new_mgid;
757 sprintf(group->name, "%016llx%016llx",
758 be64_to_cpu(group->rec.mgid.global.subnet_prefix),
759 be64_to_cpu(group->rec.mgid.global.interface_id));
760 list_del_init(&group->mgid0_list);
761 cur_group = mcast_insert(ctx, group);
763 /* A race between our code and SM. Silently cleaning the new one */
764 req = list_first_entry(&group->pending_list,
765 struct mcast_req, group_list);
766 --group->func[req->func].num_pend_reqs;
767 list_del(&req->group_list);
768 list_del(&req->func_list);
770 mutex_unlock(&group->lock);
771 mutex_unlock(&ctx->mcg_table_lock);
772 release_group(group, 0);
776 atomic_inc(&group->refcount);
777 add_sysfs_port_mcg_attr(ctx->dev, ctx->port, &group->dentry.attr);
778 mutex_unlock(&group->lock);
779 mutex_unlock(&ctx->mcg_table_lock);
782 struct mcast_req *tmp1, *tmp2;
784 list_del(&group->mgid0_list);
785 if (!list_empty(&group->pending_list) && group->state != MCAST_IDLE)
786 cancel_delayed_work_sync(&group->timeout_work);
788 list_for_each_entry_safe(tmp1, tmp2, &group->pending_list, group_list) {
789 list_del(&tmp1->group_list);
792 mutex_unlock(&group->lock);
793 mutex_unlock(&ctx->mcg_table_lock);
798 mutex_unlock(&group->lock);
800 mutex_unlock(&ctx->mcg_table_lock);
805 static ssize_t sysfs_show_group(struct device *dev,
806 struct device_attribute *attr, char *buf);
808 static struct mcast_group *acquire_group(struct mlx4_ib_demux_ctx *ctx,
809 union ib_gid *mgid, int create,
812 struct mcast_group *group, *cur_group;
816 is_mgid0 = !memcmp(&mgid0, mgid, sizeof mgid0);
818 group = mcast_find(ctx, mgid);
824 return ERR_PTR(-ENOENT);
826 group = kzalloc(sizeof *group, gfp_mask);
828 return ERR_PTR(-ENOMEM);
831 group->rec.mgid = *mgid;
832 INIT_LIST_HEAD(&group->pending_list);
833 INIT_LIST_HEAD(&group->mgid0_list);
834 for (i = 0; i < MAX_VFS; ++i)
835 INIT_LIST_HEAD(&group->func[i].pending);
836 INIT_WORK(&group->work, mlx4_ib_mcg_work_handler);
837 INIT_DELAYED_WORK(&group->timeout_work, mlx4_ib_mcg_timeout_handler);
838 mutex_init(&group->lock);
839 sprintf(group->name, "%016llx%016llx",
840 be64_to_cpu(group->rec.mgid.global.subnet_prefix),
841 be64_to_cpu(group->rec.mgid.global.interface_id));
842 sysfs_attr_init(&group->dentry.attr);
843 group->dentry.show = sysfs_show_group;
844 group->dentry.store = NULL;
845 group->dentry.attr.name = group->name;
846 group->dentry.attr.mode = 0400;
847 group->state = MCAST_IDLE;
850 list_add(&group->mgid0_list, &ctx->mcg_mgid0_list);
854 cur_group = mcast_insert(ctx, group);
856 mcg_warn("group just showed up %s - confused\n", cur_group->name);
858 return ERR_PTR(-EINVAL);
861 add_sysfs_port_mcg_attr(ctx->dev, ctx->port, &group->dentry.attr);
864 atomic_inc(&group->refcount);
868 static void queue_req(struct mcast_req *req)
870 struct mcast_group *group = req->group;
872 atomic_inc(&group->refcount); /* for the request */
873 atomic_inc(&group->refcount); /* for scheduling the work */
874 list_add_tail(&req->group_list, &group->pending_list);
875 list_add_tail(&req->func_list, &group->func[req->func].pending);
876 /* calls mlx4_ib_mcg_work_handler */
877 if (!queue_work(group->demux->mcg_wq, &group->work))
878 safe_atomic_dec(&group->refcount);
881 int mlx4_ib_mcg_demux_handler(struct ib_device *ibdev, int port, int slave,
882 struct ib_sa_mad *mad)
884 struct mlx4_ib_dev *dev = to_mdev(ibdev);
885 struct ib_sa_mcmember_data *rec = (struct ib_sa_mcmember_data *)mad->data;
886 struct mlx4_ib_demux_ctx *ctx = &dev->sriov.demux[port - 1];
887 struct mcast_group *group;
889 switch (mad->mad_hdr.method) {
890 case IB_MGMT_METHOD_GET_RESP:
891 case IB_SA_METHOD_DELETE_RESP:
892 mutex_lock(&ctx->mcg_table_lock);
893 group = acquire_group(ctx, &rec->mgid, 0, GFP_KERNEL);
894 mutex_unlock(&ctx->mcg_table_lock);
896 if (mad->mad_hdr.method == IB_MGMT_METHOD_GET_RESP) {
897 __be64 tid = mad->mad_hdr.tid;
898 *(u8 *)(&tid) = (u8)slave; /* in group we kept the modified TID */
899 group = search_relocate_mgid0_group(ctx, tid, &rec->mgid);
907 mutex_lock(&group->lock);
908 group->response_sa_mad = *mad;
909 group->prev_state = group->state;
910 group->state = MCAST_RESP_READY;
911 /* calls mlx4_ib_mcg_work_handler */
912 atomic_inc(&group->refcount);
913 if (!queue_work(ctx->mcg_wq, &group->work))
914 safe_atomic_dec(&group->refcount);
915 mutex_unlock(&group->lock);
916 release_group(group, 0);
917 return 1; /* consumed */
918 case IB_MGMT_METHOD_SET:
919 case IB_SA_METHOD_GET_TABLE:
920 case IB_SA_METHOD_GET_TABLE_RESP:
921 case IB_SA_METHOD_DELETE:
922 return 0; /* not consumed, pass-through to guest over tunnel */
924 mcg_warn("In demux, port %d: unexpected MCMember method: 0x%x, dropping\n",
925 port, mad->mad_hdr.method);
926 return 1; /* consumed */
930 int mlx4_ib_mcg_multiplex_handler(struct ib_device *ibdev, int port,
931 int slave, struct ib_sa_mad *sa_mad)
933 struct mlx4_ib_dev *dev = to_mdev(ibdev);
934 struct ib_sa_mcmember_data *rec = (struct ib_sa_mcmember_data *)sa_mad->data;
935 struct mlx4_ib_demux_ctx *ctx = &dev->sriov.demux[port - 1];
936 struct mcast_group *group;
937 struct mcast_req *req;
943 switch (sa_mad->mad_hdr.method) {
944 case IB_MGMT_METHOD_SET:
946 case IB_SA_METHOD_DELETE:
947 req = kzalloc(sizeof *req, GFP_KERNEL);
952 req->sa_mad = *sa_mad;
954 mutex_lock(&ctx->mcg_table_lock);
955 group = acquire_group(ctx, &rec->mgid, may_create, GFP_KERNEL);
956 mutex_unlock(&ctx->mcg_table_lock);
959 return PTR_ERR(group);
961 mutex_lock(&group->lock);
962 if (group->func[slave].num_pend_reqs > MAX_PEND_REQS_PER_FUNC) {
963 mutex_unlock(&group->lock);
964 mcg_warn_group(group, "Port %d, Func %d has too many pending requests (%d), dropping\n",
965 port, slave, MAX_PEND_REQS_PER_FUNC);
966 release_group(group, 0);
970 ++group->func[slave].num_pend_reqs;
973 mutex_unlock(&group->lock);
974 release_group(group, 0);
975 return 1; /* consumed */
976 case IB_SA_METHOD_GET_TABLE:
977 case IB_MGMT_METHOD_GET_RESP:
978 case IB_SA_METHOD_GET_TABLE_RESP:
979 case IB_SA_METHOD_DELETE_RESP:
980 return 0; /* not consumed, pass-through */
982 mcg_warn("In multiplex, port %d, func %d: unexpected MCMember method: 0x%x, dropping\n",
983 port, slave, sa_mad->mad_hdr.method);
984 return 1; /* consumed */
988 static ssize_t sysfs_show_group(struct device *dev,
989 struct device_attribute *attr, char *buf)
991 struct mcast_group *group =
992 container_of(attr, struct mcast_group, dentry);
993 struct mcast_req *req = NULL;
994 char pending_str[40];
999 if (group->state == MCAST_IDLE)
1000 sprintf(state_str, "%s", get_state_string(group->state));
1002 sprintf(state_str, "%s(TID=0x%llx)",
1003 get_state_string(group->state),
1004 be64_to_cpu(group->last_req_tid));
1005 if (list_empty(&group->pending_list)) {
1006 sprintf(pending_str, "No");
1008 req = list_first_entry(&group->pending_list, struct mcast_req, group_list);
1009 sprintf(pending_str, "Yes(TID=0x%llx)",
1010 be64_to_cpu(req->sa_mad.mad_hdr.tid));
1012 len += sprintf(buf + len, "%1d [%02d,%02d,%02d] %4d %4s %5s ",
1013 group->rec.scope_join_state & 0xf,
1014 group->members[2], group->members[1], group->members[0],
1015 atomic_read(&group->refcount),
1018 for (f = 0; f < MAX_VFS; ++f)
1019 if (group->func[f].state == MCAST_MEMBER)
1020 len += sprintf(buf + len, "%d[%1x] ",
1021 f, group->func[f].join_state);
1023 len += sprintf(buf + len, "\t\t(%4hx %4x %2x %2x %2x %2x %2x "
1024 "%4x %4x %2x %2x)\n",
1025 be16_to_cpu(group->rec.pkey),
1026 be32_to_cpu(group->rec.qkey),
1027 (group->rec.mtusel_mtu & 0xc0) >> 6,
1028 group->rec.mtusel_mtu & 0x3f,
1030 (group->rec.ratesel_rate & 0xc0) >> 6,
1031 group->rec.ratesel_rate & 0x3f,
1032 (be32_to_cpu(group->rec.sl_flowlabel_hoplimit) & 0xf0000000) >> 28,
1033 (be32_to_cpu(group->rec.sl_flowlabel_hoplimit) & 0x0fffff00) >> 8,
1034 be32_to_cpu(group->rec.sl_flowlabel_hoplimit) & 0x000000ff,
1035 group->rec.proxy_join);
1040 int mlx4_ib_mcg_port_init(struct mlx4_ib_demux_ctx *ctx)
1044 atomic_set(&ctx->tid, 0);
1045 sprintf(name, "mlx4_ib_mcg%d", ctx->port);
1046 ctx->mcg_wq = create_singlethread_workqueue(name);
1050 mutex_init(&ctx->mcg_table_lock);
1051 ctx->mcg_table = RB_ROOT;
1052 INIT_LIST_HEAD(&ctx->mcg_mgid0_list);
1058 static void force_clean_group(struct mcast_group *group)
1060 struct mcast_req *req, *tmp
1062 list_for_each_entry_safe(req, tmp, &group->pending_list, group_list) {
1063 list_del(&req->group_list);
1066 del_sysfs_port_mcg_attr(group->demux->dev, group->demux->port, &group->dentry.attr);
1067 rb_erase(&group->node, &group->demux->mcg_table);
1071 static void _mlx4_ib_mcg_port_cleanup(struct mlx4_ib_demux_ctx *ctx, int destroy_wq)
1075 struct mcast_group *group;
1079 for (i = 0; i < MAX_VFS; ++i)
1080 clean_vf_mcast(ctx, i);
1082 end = jiffies + msecs_to_jiffies(MAD_TIMEOUT_MS + 3000);
1085 mutex_lock(&ctx->mcg_table_lock);
1086 for (p = rb_first(&ctx->mcg_table); p; p = rb_next(p))
1088 mutex_unlock(&ctx->mcg_table_lock);
1093 } while (time_after(end, jiffies));
1095 flush_workqueue(ctx->mcg_wq);
1097 destroy_workqueue(ctx->mcg_wq);
1099 mutex_lock(&ctx->mcg_table_lock);
1100 while ((p = rb_first(&ctx->mcg_table)) != NULL) {
1101 group = rb_entry(p, struct mcast_group, node);
1102 if (atomic_read(&group->refcount))
1103 mcg_warn_group(group, "group refcount %d!!! (pointer %p)\n", atomic_read(&group->refcount), group);
1105 force_clean_group(group);
1107 mutex_unlock(&ctx->mcg_table_lock);
1111 struct work_struct work;
1112 struct mlx4_ib_demux_ctx *ctx;
1116 static void mcg_clean_task(struct work_struct *work)
1118 struct clean_work *cw = container_of(work, struct clean_work, work);
1120 _mlx4_ib_mcg_port_cleanup(cw->ctx, cw->destroy_wq);
1121 cw->ctx->flushing = 0;
1125 void mlx4_ib_mcg_port_cleanup(struct mlx4_ib_demux_ctx *ctx, int destroy_wq)
1127 struct clean_work *work;
1135 _mlx4_ib_mcg_port_cleanup(ctx, destroy_wq);
1140 work = kmalloc(sizeof *work, GFP_KERNEL);
1143 mcg_warn("failed allocating work for cleanup\n");
1148 work->destroy_wq = destroy_wq;
1149 INIT_WORK(&work->work, mcg_clean_task);
1150 queue_work(clean_wq, &work->work);
1153 static void build_leave_mad(struct mcast_req *req)
1155 struct ib_sa_mad *mad = &req->sa_mad;
1157 mad->mad_hdr.method = IB_SA_METHOD_DELETE;
1161 static void clear_pending_reqs(struct mcast_group *group, int vf)
1163 struct mcast_req *req, *tmp, *group_first = NULL;
1167 if (!list_empty(&group->pending_list))
1168 group_first = list_first_entry(&group->pending_list, struct mcast_req, group_list);
1170 list_for_each_entry_safe(req, tmp, &group->func[vf].pending, func_list) {
1172 if (group_first == req &&
1173 (group->state == MCAST_JOIN_SENT ||
1174 group->state == MCAST_LEAVE_SENT)) {
1175 clear = cancel_delayed_work(&group->timeout_work);
1177 group->state = MCAST_IDLE;
1180 --group->func[vf].num_pend_reqs;
1181 list_del(&req->group_list);
1182 list_del(&req->func_list);
1184 atomic_dec(&group->refcount);
1188 if (!pend && (!list_empty(&group->func[vf].pending) || group->func[vf].num_pend_reqs)) {
1189 mcg_warn_group(group, "DRIVER BUG: list_empty %d, num_pend_reqs %d\n",
1190 list_empty(&group->func[vf].pending), group->func[vf].num_pend_reqs);
1194 static int push_deleteing_req(struct mcast_group *group, int slave)
1196 struct mcast_req *req;
1197 struct mcast_req *pend_req;
1199 if (!group->func[slave].join_state)
1202 req = kzalloc(sizeof *req, GFP_KERNEL);
1204 mcg_warn_group(group, "failed allocation - may leave stall groups\n");
1208 if (!list_empty(&group->func[slave].pending)) {
1209 pend_req = list_entry(group->func[slave].pending.prev, struct mcast_req, group_list);
1210 if (pend_req->clean) {
1219 ++group->func[slave].num_pend_reqs;
1220 build_leave_mad(req);
1225 void clean_vf_mcast(struct mlx4_ib_demux_ctx *ctx, int slave)
1227 struct mcast_group *group;
1230 mutex_lock(&ctx->mcg_table_lock);
1231 for (p = rb_first(&ctx->mcg_table); p; p = rb_next(p)) {
1232 group = rb_entry(p, struct mcast_group, node);
1233 mutex_lock(&group->lock);
1234 if (atomic_read(&group->refcount)) {
1235 /* clear pending requests of this VF */
1236 clear_pending_reqs(group, slave);
1237 push_deleteing_req(group, slave);
1239 mutex_unlock(&group->lock);
1241 mutex_unlock(&ctx->mcg_table_lock);
1245 int mlx4_ib_mcg_init(void)
1247 clean_wq = create_singlethread_workqueue("mlx4_ib_mcg");
1254 void mlx4_ib_mcg_destroy(void)
1256 destroy_workqueue(clean_wq);