1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
3 * Copyright (c) 2022 Hewlett Packard Enterprise, Inc. All rights reserved.
4 * Copyright (c) 2016 Mellanox Technologies Ltd. All rights reserved.
5 * Copyright (c) 2015 System Fabric Works, Inc. All rights reserved.
9 * rxe_mcast.c implements driver support for multicast transport.
10 * It is based on two data structures struct rxe_mcg ('mcg') and
11 * struct rxe_mca ('mca'). An mcg is allocated each time a qp is
12 * attached to a new mgid for the first time. These are indexed by
13 * a red-black tree using the mgid. This data structure is searched
14 * for the mcg when a multicast packet is received and when another
15 * qp is attached to the same mgid. It is cleaned up when the last qp
16 * is detached from the mcg. Each time a qp is attached to an mcg an
17 * mca is created. It holds a pointer to the qp and is added to a list
18 * of qp's that are attached to the mcg. The qp_list is used to replicate
19 * mcast packets in the rxe receive path.
25 * rxe_mcast_add - add multicast address to rxe device
26 * @rxe: rxe device object
27 * @mgid: multicast address as a gid
29 * Returns 0 on success else an error
31 static int rxe_mcast_add(struct rxe_dev *rxe, union ib_gid *mgid)
33 unsigned char ll_addr[ETH_ALEN];
35 ipv6_eth_mc_map((struct in6_addr *)mgid->raw, ll_addr);
37 return dev_mc_add(rxe->ndev, ll_addr);
41 * rxe_mcast_del - delete multicast address from rxe device
42 * @rxe: rxe device object
43 * @mgid: multicast address as a gid
45 * Returns 0 on success else an error
47 static int rxe_mcast_del(struct rxe_dev *rxe, union ib_gid *mgid)
49 unsigned char ll_addr[ETH_ALEN];
51 ipv6_eth_mc_map((struct in6_addr *)mgid->raw, ll_addr);
53 return dev_mc_del(rxe->ndev, ll_addr);
57 * __rxe_insert_mcg - insert an mcg into red-black tree (rxe->mcg_tree)
58 * @mcg: mcg object with an embedded red-black tree node
60 * Context: caller must hold a reference to mcg and rxe->mcg_lock and
61 * is responsible to avoid adding the same mcg twice to the tree.
63 static void __rxe_insert_mcg(struct rxe_mcg *mcg)
65 struct rb_root *tree = &mcg->rxe->mcg_tree;
66 struct rb_node **link = &tree->rb_node;
67 struct rb_node *node = NULL;
73 tmp = rb_entry(node, struct rxe_mcg, node);
75 cmp = memcmp(&tmp->mgid, &mcg->mgid, sizeof(mcg->mgid));
77 link = &(*link)->rb_left;
79 link = &(*link)->rb_right;
82 rb_link_node(&mcg->node, node, link);
83 rb_insert_color(&mcg->node, tree);
87 * __rxe_remove_mcg - remove an mcg from red-black tree holding lock
88 * @mcg: mcast group object with an embedded red-black tree node
90 * Context: caller must hold a reference to mcg and rxe->mcg_lock
92 static void __rxe_remove_mcg(struct rxe_mcg *mcg)
94 rb_erase(&mcg->node, &mcg->rxe->mcg_tree);
98 * __rxe_lookup_mcg - lookup mcg in rxe->mcg_tree while holding lock
99 * @rxe: rxe device object
100 * @mgid: multicast IP address
102 * Context: caller must hold rxe->mcg_lock
103 * Returns: mcg on success and takes a ref to mcg else NULL
105 static struct rxe_mcg *__rxe_lookup_mcg(struct rxe_dev *rxe,
108 struct rb_root *tree = &rxe->mcg_tree;
110 struct rb_node *node;
113 node = tree->rb_node;
116 mcg = rb_entry(node, struct rxe_mcg, node);
118 cmp = memcmp(&mcg->mgid, mgid, sizeof(*mgid));
121 node = node->rb_left;
123 node = node->rb_right;
129 kref_get(&mcg->ref_cnt);
137 * rxe_lookup_mcg - lookup up mcg in red-back tree
138 * @rxe: rxe device object
139 * @mgid: multicast IP address
141 * Returns: mcg if found else NULL
143 struct rxe_mcg *rxe_lookup_mcg(struct rxe_dev *rxe, union ib_gid *mgid)
147 spin_lock_bh(&rxe->mcg_lock);
148 mcg = __rxe_lookup_mcg(rxe, mgid);
149 spin_unlock_bh(&rxe->mcg_lock);
155 * __rxe_init_mcg - initialize a new mcg
157 * @mgid: multicast address as a gid
158 * @mcg: new mcg object
160 * Context: caller should hold rxe->mcg lock
162 static void __rxe_init_mcg(struct rxe_dev *rxe, union ib_gid *mgid,
165 kref_init(&mcg->ref_cnt);
166 memcpy(&mcg->mgid, mgid, sizeof(mcg->mgid));
167 INIT_LIST_HEAD(&mcg->qp_list);
170 /* caller holds a ref on mcg but that will be
171 * dropped when mcg goes out of scope. We need to take a ref
172 * on the pointer that will be saved in the red-black tree
173 * by __rxe_insert_mcg and used to lookup mcg from mgid later.
174 * Inserting mcg makes it visible to outside so this should
175 * be done last after the object is ready.
177 kref_get(&mcg->ref_cnt);
178 __rxe_insert_mcg(mcg);
182 * rxe_get_mcg - lookup or allocate a mcg
183 * @rxe: rxe device object
184 * @mgid: multicast IP address as a gid
186 * Returns: mcg on success else ERR_PTR(error)
188 static struct rxe_mcg *rxe_get_mcg(struct rxe_dev *rxe, union ib_gid *mgid)
190 struct rxe_mcg *mcg, *tmp;
193 if (rxe->attr.max_mcast_grp == 0)
194 return ERR_PTR(-EINVAL);
196 /* check to see if mcg already exists */
197 mcg = rxe_lookup_mcg(rxe, mgid);
201 /* check to see if we have reached limit */
202 if (atomic_inc_return(&rxe->mcg_num) > rxe->attr.max_mcast_grp) {
207 /* speculative alloc of new mcg */
208 mcg = kzalloc(sizeof(*mcg), GFP_KERNEL);
214 spin_lock_bh(&rxe->mcg_lock);
215 /* re-check to see if someone else just added it */
216 tmp = __rxe_lookup_mcg(rxe, mgid);
218 spin_unlock_bh(&rxe->mcg_lock);
219 atomic_dec(&rxe->mcg_num);
224 __rxe_init_mcg(rxe, mgid, mcg);
225 spin_unlock_bh(&rxe->mcg_lock);
227 /* add mcast address outside of lock */
228 err = rxe_mcast_add(rxe, mgid);
234 atomic_dec(&rxe->mcg_num);
239 * rxe_cleanup_mcg - cleanup mcg for kref_put
240 * @kref: struct kref embnedded in mcg
242 void rxe_cleanup_mcg(struct kref *kref)
244 struct rxe_mcg *mcg = container_of(kref, typeof(*mcg), ref_cnt);
250 * __rxe_destroy_mcg - destroy mcg object holding rxe->mcg_lock
251 * @mcg: the mcg object
253 * Context: caller is holding rxe->mcg_lock
254 * no qp's are attached to mcg
256 static void __rxe_destroy_mcg(struct rxe_mcg *mcg)
258 struct rxe_dev *rxe = mcg->rxe;
260 /* remove mcg from red-black tree then drop ref */
261 __rxe_remove_mcg(mcg);
262 kref_put(&mcg->ref_cnt, rxe_cleanup_mcg);
264 atomic_dec(&rxe->mcg_num);
268 * rxe_destroy_mcg - destroy mcg object
269 * @mcg: the mcg object
271 * Context: no qp's are attached to mcg
273 static void rxe_destroy_mcg(struct rxe_mcg *mcg)
275 /* delete mcast address outside of lock */
276 rxe_mcast_del(mcg->rxe, &mcg->mgid);
278 spin_lock_bh(&mcg->rxe->mcg_lock);
279 __rxe_destroy_mcg(mcg);
280 spin_unlock_bh(&mcg->rxe->mcg_lock);
284 * __rxe_init_mca - initialize a new mca holding lock
287 * @mca: empty space for new mca
289 * Context: caller must hold references on qp and mcg, rxe->mcg_lock
290 * and pass memory for new mca
292 * Returns: 0 on success else an error
294 static int __rxe_init_mca(struct rxe_qp *qp, struct rxe_mcg *mcg,
297 struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
300 n = atomic_inc_return(&rxe->mcg_attach);
301 if (n > rxe->attr.max_total_mcast_qp_attach) {
302 atomic_dec(&rxe->mcg_attach);
306 n = atomic_inc_return(&mcg->qp_num);
307 if (n > rxe->attr.max_mcast_qp_attach) {
308 atomic_dec(&mcg->qp_num);
309 atomic_dec(&rxe->mcg_attach);
313 atomic_inc(&qp->mcg_num);
318 list_add_tail(&mca->qp_list, &mcg->qp_list);
324 * rxe_attach_mcg - attach qp to mcg if not already attached
328 * Context: caller must hold reference on qp and mcg.
329 * Returns: 0 on success else an error
331 static int rxe_attach_mcg(struct rxe_mcg *mcg, struct rxe_qp *qp)
333 struct rxe_dev *rxe = mcg->rxe;
334 struct rxe_mca *mca, *tmp;
337 /* check to see if the qp is already a member of the group */
338 spin_lock_bh(&rxe->mcg_lock);
339 list_for_each_entry(mca, &mcg->qp_list, qp_list) {
341 spin_unlock_bh(&rxe->mcg_lock);
345 spin_unlock_bh(&rxe->mcg_lock);
347 /* speculative alloc new mca without using GFP_ATOMIC */
348 mca = kzalloc(sizeof(*mca), GFP_KERNEL);
352 spin_lock_bh(&rxe->mcg_lock);
353 /* re-check to see if someone else just attached qp */
354 list_for_each_entry(tmp, &mcg->qp_list, qp_list) {
362 err = __rxe_init_mca(qp, mcg, mca);
366 spin_unlock_bh(&rxe->mcg_lock);
371 * __rxe_cleanup_mca - cleanup mca object holding lock
375 * Context: caller must hold a reference to mcg and rxe->mcg_lock
377 static void __rxe_cleanup_mca(struct rxe_mca *mca, struct rxe_mcg *mcg)
379 list_del(&mca->qp_list);
381 atomic_dec(&mcg->qp_num);
382 atomic_dec(&mcg->rxe->mcg_attach);
383 atomic_dec(&mca->qp->mcg_num);
390 * rxe_detach_mcg - detach qp from mcg
394 * Returns: 0 on success else an error if qp is not attached.
396 static int rxe_detach_mcg(struct rxe_mcg *mcg, struct rxe_qp *qp)
398 struct rxe_dev *rxe = mcg->rxe;
399 struct rxe_mca *mca, *tmp;
401 spin_lock_bh(&rxe->mcg_lock);
402 list_for_each_entry_safe(mca, tmp, &mcg->qp_list, qp_list) {
404 __rxe_cleanup_mca(mca, mcg);
406 /* if the number of qp's attached to the
407 * mcast group falls to zero go ahead and
408 * tear it down. This will not free the
409 * object since we are still holding a ref
412 if (atomic_read(&mcg->qp_num) <= 0)
413 __rxe_destroy_mcg(mcg);
415 spin_unlock_bh(&rxe->mcg_lock);
420 /* we didn't find the qp on the list */
421 spin_unlock_bh(&rxe->mcg_lock);
426 * rxe_attach_mcast - attach qp to multicast group (see IBA-11.3.1)
427 * @ibqp: (IB) qp object
428 * @mgid: multicast IP address
429 * @mlid: multicast LID, ignored for RoCEv2 (see IBA-A17.5.6)
431 * Returns: 0 on success else an errno
433 int rxe_attach_mcast(struct ib_qp *ibqp, union ib_gid *mgid, u16 mlid)
436 struct rxe_dev *rxe = to_rdev(ibqp->device);
437 struct rxe_qp *qp = to_rqp(ibqp);
440 /* takes a ref on mcg if successful */
441 mcg = rxe_get_mcg(rxe, mgid);
445 err = rxe_attach_mcg(mcg, qp);
447 /* if we failed to attach the first qp to mcg tear it down */
448 if (atomic_read(&mcg->qp_num) == 0)
449 rxe_destroy_mcg(mcg);
451 kref_put(&mcg->ref_cnt, rxe_cleanup_mcg);
457 * rxe_detach_mcast - detach qp from multicast group (see IBA-11.3.2)
458 * @ibqp: address of (IB) qp object
459 * @mgid: multicast IP address
460 * @mlid: multicast LID, ignored for RoCEv2 (see IBA-A17.5.6)
462 * Returns: 0 on success else an errno
464 int rxe_detach_mcast(struct ib_qp *ibqp, union ib_gid *mgid, u16 mlid)
466 struct rxe_dev *rxe = to_rdev(ibqp->device);
467 struct rxe_qp *qp = to_rqp(ibqp);
471 mcg = rxe_lookup_mcg(rxe, mgid);
475 err = rxe_detach_mcg(mcg, qp);
476 kref_put(&mcg->ref_cnt, rxe_cleanup_mcg);