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];
34 struct net_device *ndev;
37 ndev = rxe_ib_device_get_netdev(&rxe->ib_dev);
41 ipv6_eth_mc_map((struct in6_addr *)mgid->raw, ll_addr);
43 ret = dev_mc_add(ndev, ll_addr);
50 * rxe_mcast_del - delete multicast address from rxe device
51 * @rxe: rxe device object
52 * @mgid: multicast address as a gid
54 * Returns 0 on success else an error
56 static int rxe_mcast_del(struct rxe_dev *rxe, union ib_gid *mgid)
58 unsigned char ll_addr[ETH_ALEN];
59 struct net_device *ndev;
62 ndev = rxe_ib_device_get_netdev(&rxe->ib_dev);
66 ipv6_eth_mc_map((struct in6_addr *)mgid->raw, ll_addr);
68 ret = dev_mc_del(ndev, ll_addr);
75 * __rxe_insert_mcg - insert an mcg into red-black tree (rxe->mcg_tree)
76 * @mcg: mcg object with an embedded red-black tree node
78 * Context: caller must hold a reference to mcg and rxe->mcg_lock and
79 * is responsible to avoid adding the same mcg twice to the tree.
81 static void __rxe_insert_mcg(struct rxe_mcg *mcg)
83 struct rb_root *tree = &mcg->rxe->mcg_tree;
84 struct rb_node **link = &tree->rb_node;
85 struct rb_node *node = NULL;
91 tmp = rb_entry(node, struct rxe_mcg, node);
93 cmp = memcmp(&tmp->mgid, &mcg->mgid, sizeof(mcg->mgid));
95 link = &(*link)->rb_left;
97 link = &(*link)->rb_right;
100 rb_link_node(&mcg->node, node, link);
101 rb_insert_color(&mcg->node, tree);
105 * __rxe_remove_mcg - remove an mcg from red-black tree holding lock
106 * @mcg: mcast group object with an embedded red-black tree node
108 * Context: caller must hold a reference to mcg and rxe->mcg_lock
110 static void __rxe_remove_mcg(struct rxe_mcg *mcg)
112 rb_erase(&mcg->node, &mcg->rxe->mcg_tree);
116 * __rxe_lookup_mcg - lookup mcg in rxe->mcg_tree while holding lock
117 * @rxe: rxe device object
118 * @mgid: multicast IP address
120 * Context: caller must hold rxe->mcg_lock
121 * Returns: mcg on success and takes a ref to mcg else NULL
123 static struct rxe_mcg *__rxe_lookup_mcg(struct rxe_dev *rxe,
126 struct rb_root *tree = &rxe->mcg_tree;
128 struct rb_node *node;
131 node = tree->rb_node;
134 mcg = rb_entry(node, struct rxe_mcg, node);
136 cmp = memcmp(&mcg->mgid, mgid, sizeof(*mgid));
139 node = node->rb_left;
141 node = node->rb_right;
147 kref_get(&mcg->ref_cnt);
155 * rxe_lookup_mcg - lookup up mcg in red-back tree
156 * @rxe: rxe device object
157 * @mgid: multicast IP address
159 * Returns: mcg if found else NULL
161 struct rxe_mcg *rxe_lookup_mcg(struct rxe_dev *rxe, union ib_gid *mgid)
165 spin_lock_bh(&rxe->mcg_lock);
166 mcg = __rxe_lookup_mcg(rxe, mgid);
167 spin_unlock_bh(&rxe->mcg_lock);
173 * __rxe_init_mcg - initialize a new mcg
175 * @mgid: multicast address as a gid
176 * @mcg: new mcg object
178 * Context: caller should hold rxe->mcg lock
180 static void __rxe_init_mcg(struct rxe_dev *rxe, union ib_gid *mgid,
183 kref_init(&mcg->ref_cnt);
184 memcpy(&mcg->mgid, mgid, sizeof(mcg->mgid));
185 INIT_LIST_HEAD(&mcg->qp_list);
188 /* caller holds a ref on mcg but that will be
189 * dropped when mcg goes out of scope. We need to take a ref
190 * on the pointer that will be saved in the red-black tree
191 * by __rxe_insert_mcg and used to lookup mcg from mgid later.
192 * Inserting mcg makes it visible to outside so this should
193 * be done last after the object is ready.
195 kref_get(&mcg->ref_cnt);
196 __rxe_insert_mcg(mcg);
200 * rxe_get_mcg - lookup or allocate a mcg
201 * @rxe: rxe device object
202 * @mgid: multicast IP address as a gid
204 * Returns: mcg on success else ERR_PTR(error)
206 static struct rxe_mcg *rxe_get_mcg(struct rxe_dev *rxe, union ib_gid *mgid)
208 struct rxe_mcg *mcg, *tmp;
211 if (rxe->attr.max_mcast_grp == 0)
212 return ERR_PTR(-EINVAL);
214 /* check to see if mcg already exists */
215 mcg = rxe_lookup_mcg(rxe, mgid);
219 /* check to see if we have reached limit */
220 if (atomic_inc_return(&rxe->mcg_num) > rxe->attr.max_mcast_grp) {
225 /* speculative alloc of new mcg */
226 mcg = kzalloc(sizeof(*mcg), GFP_KERNEL);
232 spin_lock_bh(&rxe->mcg_lock);
233 /* re-check to see if someone else just added it */
234 tmp = __rxe_lookup_mcg(rxe, mgid);
236 spin_unlock_bh(&rxe->mcg_lock);
237 atomic_dec(&rxe->mcg_num);
242 __rxe_init_mcg(rxe, mgid, mcg);
243 spin_unlock_bh(&rxe->mcg_lock);
245 /* add mcast address outside of lock */
246 err = rxe_mcast_add(rxe, mgid);
252 atomic_dec(&rxe->mcg_num);
257 * rxe_cleanup_mcg - cleanup mcg for kref_put
258 * @kref: struct kref embnedded in mcg
260 void rxe_cleanup_mcg(struct kref *kref)
262 struct rxe_mcg *mcg = container_of(kref, typeof(*mcg), ref_cnt);
268 * __rxe_destroy_mcg - destroy mcg object holding rxe->mcg_lock
269 * @mcg: the mcg object
271 * Context: caller is holding rxe->mcg_lock
272 * no qp's are attached to mcg
274 static void __rxe_destroy_mcg(struct rxe_mcg *mcg)
276 struct rxe_dev *rxe = mcg->rxe;
278 /* remove mcg from red-black tree then drop ref */
279 __rxe_remove_mcg(mcg);
280 kref_put(&mcg->ref_cnt, rxe_cleanup_mcg);
282 atomic_dec(&rxe->mcg_num);
286 * rxe_destroy_mcg - destroy mcg object
287 * @mcg: the mcg object
289 * Context: no qp's are attached to mcg
291 static void rxe_destroy_mcg(struct rxe_mcg *mcg)
293 /* delete mcast address outside of lock */
294 rxe_mcast_del(mcg->rxe, &mcg->mgid);
296 spin_lock_bh(&mcg->rxe->mcg_lock);
297 __rxe_destroy_mcg(mcg);
298 spin_unlock_bh(&mcg->rxe->mcg_lock);
302 * __rxe_init_mca - initialize a new mca holding lock
305 * @mca: empty space for new mca
307 * Context: caller must hold references on qp and mcg, rxe->mcg_lock
308 * and pass memory for new mca
310 * Returns: 0 on success else an error
312 static int __rxe_init_mca(struct rxe_qp *qp, struct rxe_mcg *mcg,
315 struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
318 n = atomic_inc_return(&rxe->mcg_attach);
319 if (n > rxe->attr.max_total_mcast_qp_attach) {
320 atomic_dec(&rxe->mcg_attach);
324 n = atomic_inc_return(&mcg->qp_num);
325 if (n > rxe->attr.max_mcast_qp_attach) {
326 atomic_dec(&mcg->qp_num);
327 atomic_dec(&rxe->mcg_attach);
331 atomic_inc(&qp->mcg_num);
336 list_add_tail(&mca->qp_list, &mcg->qp_list);
342 * rxe_attach_mcg - attach qp to mcg if not already attached
346 * Context: caller must hold reference on qp and mcg.
347 * Returns: 0 on success else an error
349 static int rxe_attach_mcg(struct rxe_mcg *mcg, struct rxe_qp *qp)
351 struct rxe_dev *rxe = mcg->rxe;
352 struct rxe_mca *mca, *tmp;
355 /* check to see if the qp is already a member of the group */
356 spin_lock_bh(&rxe->mcg_lock);
357 list_for_each_entry(mca, &mcg->qp_list, qp_list) {
359 spin_unlock_bh(&rxe->mcg_lock);
363 spin_unlock_bh(&rxe->mcg_lock);
365 /* speculative alloc new mca without using GFP_ATOMIC */
366 mca = kzalloc(sizeof(*mca), GFP_KERNEL);
370 spin_lock_bh(&rxe->mcg_lock);
371 /* re-check to see if someone else just attached qp */
372 list_for_each_entry(tmp, &mcg->qp_list, qp_list) {
380 err = __rxe_init_mca(qp, mcg, mca);
384 spin_unlock_bh(&rxe->mcg_lock);
389 * __rxe_cleanup_mca - cleanup mca object holding lock
393 * Context: caller must hold a reference to mcg and rxe->mcg_lock
395 static void __rxe_cleanup_mca(struct rxe_mca *mca, struct rxe_mcg *mcg)
397 list_del(&mca->qp_list);
399 atomic_dec(&mcg->qp_num);
400 atomic_dec(&mcg->rxe->mcg_attach);
401 atomic_dec(&mca->qp->mcg_num);
408 * rxe_detach_mcg - detach qp from mcg
412 * Returns: 0 on success else an error if qp is not attached.
414 static int rxe_detach_mcg(struct rxe_mcg *mcg, struct rxe_qp *qp)
416 struct rxe_dev *rxe = mcg->rxe;
417 struct rxe_mca *mca, *tmp;
419 spin_lock_bh(&rxe->mcg_lock);
420 list_for_each_entry_safe(mca, tmp, &mcg->qp_list, qp_list) {
422 __rxe_cleanup_mca(mca, mcg);
424 /* if the number of qp's attached to the
425 * mcast group falls to zero go ahead and
426 * tear it down. This will not free the
427 * object since we are still holding a ref
430 if (atomic_read(&mcg->qp_num) <= 0)
431 __rxe_destroy_mcg(mcg);
433 spin_unlock_bh(&rxe->mcg_lock);
438 /* we didn't find the qp on the list */
439 spin_unlock_bh(&rxe->mcg_lock);
444 * rxe_attach_mcast - attach qp to multicast group (see IBA-11.3.1)
445 * @ibqp: (IB) qp object
446 * @mgid: multicast IP address
447 * @mlid: multicast LID, ignored for RoCEv2 (see IBA-A17.5.6)
449 * Returns: 0 on success else an errno
451 int rxe_attach_mcast(struct ib_qp *ibqp, union ib_gid *mgid, u16 mlid)
454 struct rxe_dev *rxe = to_rdev(ibqp->device);
455 struct rxe_qp *qp = to_rqp(ibqp);
458 /* takes a ref on mcg if successful */
459 mcg = rxe_get_mcg(rxe, mgid);
463 err = rxe_attach_mcg(mcg, qp);
465 /* if we failed to attach the first qp to mcg tear it down */
466 if (atomic_read(&mcg->qp_num) == 0)
467 rxe_destroy_mcg(mcg);
469 kref_put(&mcg->ref_cnt, rxe_cleanup_mcg);
475 * rxe_detach_mcast - detach qp from multicast group (see IBA-11.3.2)
476 * @ibqp: address of (IB) qp object
477 * @mgid: multicast IP address
478 * @mlid: multicast LID, ignored for RoCEv2 (see IBA-A17.5.6)
480 * Returns: 0 on success else an errno
482 int rxe_detach_mcast(struct ib_qp *ibqp, union ib_gid *mgid, u16 mlid)
484 struct rxe_dev *rxe = to_rdev(ibqp->device);
485 struct rxe_qp *qp = to_rqp(ibqp);
489 mcg = rxe_lookup_mcg(rxe, mgid);
493 err = rxe_detach_mcg(mcg, qp);
494 kref_put(&mcg->ref_cnt, rxe_cleanup_mcg);