2 * Shared Memory Communications over RDMA (SMC-R) and RoCE
5 * Establish SMC-R as an Infiniband Client to be notified about added and
6 * removed IB devices of type RDMA.
7 * Determine device and port characteristics for these IB devices.
9 * Copyright IBM Corp. 2016
11 * Author(s): Ursula Braun <ubraun@linux.vnet.ibm.com>
14 #include <linux/random.h>
15 #include <linux/workqueue.h>
16 #include <rdma/ib_verbs.h>
24 #define SMC_QP_MIN_RNR_TIMER 5
25 #define SMC_QP_TIMEOUT 15 /* 4096 * 2 ** timeout usec */
26 #define SMC_QP_RETRY_CNT 7 /* 7: infinite */
27 #define SMC_QP_RNR_RETRY 7 /* 7: infinite */
29 struct smc_ib_devices smc_ib_devices = { /* smc-registered ib devices */
30 .lock = __SPIN_LOCK_UNLOCKED(smc_ib_devices.lock),
31 .list = LIST_HEAD_INIT(smc_ib_devices.list),
34 #define SMC_LOCAL_SYSTEMID_RESET "%%%%%%%"
36 u8 local_systemid[SMC_SYSTEMID_LEN] = SMC_LOCAL_SYSTEMID_RESET; /* unique system
40 int smc_ib_get_memory_region(struct ib_pd *pd, int access_flags,
46 return 0; /* already done */
48 /* obtain unique key -
49 * next invocation of get_dma_mr returns a different key!
51 *mr = pd->device->get_dma_mr(pd, access_flags);
52 rc = PTR_ERR_OR_ZERO(*mr);
58 static int smc_ib_modify_qp_init(struct smc_link *lnk)
60 struct ib_qp_attr qp_attr;
62 memset(&qp_attr, 0, sizeof(qp_attr));
63 qp_attr.qp_state = IB_QPS_INIT;
64 qp_attr.pkey_index = 0;
65 qp_attr.port_num = lnk->ibport;
66 qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE
67 | IB_ACCESS_REMOTE_WRITE;
68 return ib_modify_qp(lnk->roce_qp, &qp_attr,
69 IB_QP_STATE | IB_QP_PKEY_INDEX |
70 IB_QP_ACCESS_FLAGS | IB_QP_PORT);
73 static int smc_ib_modify_qp_rtr(struct smc_link *lnk)
75 enum ib_qp_attr_mask qp_attr_mask =
76 IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU | IB_QP_DEST_QPN |
77 IB_QP_RQ_PSN | IB_QP_MAX_DEST_RD_ATOMIC | IB_QP_MIN_RNR_TIMER;
78 struct ib_qp_attr qp_attr;
80 memset(&qp_attr, 0, sizeof(qp_attr));
81 qp_attr.qp_state = IB_QPS_RTR;
82 qp_attr.path_mtu = min(lnk->path_mtu, lnk->peer_mtu);
83 qp_attr.ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
84 rdma_ah_set_port_num(&qp_attr.ah_attr, lnk->ibport);
85 rdma_ah_set_grh(&qp_attr.ah_attr, NULL, 0, 0, 1, 0);
86 rdma_ah_set_dgid_raw(&qp_attr.ah_attr, lnk->peer_gid);
87 memcpy(&qp_attr.ah_attr.roce.dmac, lnk->peer_mac,
88 sizeof(lnk->peer_mac));
89 qp_attr.dest_qp_num = lnk->peer_qpn;
90 qp_attr.rq_psn = lnk->peer_psn; /* starting receive packet seq # */
91 qp_attr.max_dest_rd_atomic = 1; /* max # of resources for incoming
94 qp_attr.min_rnr_timer = SMC_QP_MIN_RNR_TIMER;
96 return ib_modify_qp(lnk->roce_qp, &qp_attr, qp_attr_mask);
99 int smc_ib_modify_qp_rts(struct smc_link *lnk)
101 struct ib_qp_attr qp_attr;
103 memset(&qp_attr, 0, sizeof(qp_attr));
104 qp_attr.qp_state = IB_QPS_RTS;
105 qp_attr.timeout = SMC_QP_TIMEOUT; /* local ack timeout */
106 qp_attr.retry_cnt = SMC_QP_RETRY_CNT; /* retry count */
107 qp_attr.rnr_retry = SMC_QP_RNR_RETRY; /* RNR retries, 7=infinite */
108 qp_attr.sq_psn = lnk->psn_initial; /* starting send packet seq # */
109 qp_attr.max_rd_atomic = 1; /* # of outstanding RDMA reads and
112 return ib_modify_qp(lnk->roce_qp, &qp_attr,
113 IB_QP_STATE | IB_QP_TIMEOUT | IB_QP_RETRY_CNT |
114 IB_QP_SQ_PSN | IB_QP_RNR_RETRY |
115 IB_QP_MAX_QP_RD_ATOMIC);
118 int smc_ib_modify_qp_reset(struct smc_link *lnk)
120 struct ib_qp_attr qp_attr;
122 memset(&qp_attr, 0, sizeof(qp_attr));
123 qp_attr.qp_state = IB_QPS_RESET;
124 return ib_modify_qp(lnk->roce_qp, &qp_attr, IB_QP_STATE);
127 int smc_ib_ready_link(struct smc_link *lnk)
129 struct smc_link_group *lgr =
130 container_of(lnk, struct smc_link_group, lnk[0]);
133 rc = smc_ib_modify_qp_init(lnk);
137 rc = smc_ib_modify_qp_rtr(lnk);
140 smc_wr_remember_qp_attr(lnk);
141 rc = ib_req_notify_cq(lnk->smcibdev->roce_cq_recv,
142 IB_CQ_SOLICITED_MASK);
145 rc = smc_wr_rx_post_init(lnk);
148 smc_wr_remember_qp_attr(lnk);
150 if (lgr->role == SMC_SERV) {
151 rc = smc_ib_modify_qp_rts(lnk);
154 smc_wr_remember_qp_attr(lnk);
160 /* process context wrapper for might_sleep smc_ib_remember_port_attr */
161 static void smc_ib_port_event_work(struct work_struct *work)
163 struct smc_ib_device *smcibdev = container_of(
164 work, struct smc_ib_device, port_event_work);
167 for_each_set_bit(port_idx, &smcibdev->port_event_mask, SMC_MAX_PORTS) {
168 smc_ib_remember_port_attr(smcibdev, port_idx + 1);
169 clear_bit(port_idx, &smcibdev->port_event_mask);
173 /* can be called in IRQ context */
174 static void smc_ib_global_event_handler(struct ib_event_handler *handler,
175 struct ib_event *ibevent)
177 struct smc_ib_device *smcibdev;
180 smcibdev = container_of(handler, struct smc_ib_device, event_handler);
182 switch (ibevent->event) {
183 case IB_EVENT_PORT_ERR:
184 port_idx = ibevent->element.port_num - 1;
185 set_bit(port_idx, &smcibdev->port_event_mask);
186 schedule_work(&smcibdev->port_event_work);
188 case IB_EVENT_DEVICE_FATAL:
189 /* tbd in follow-on patch:
190 * abnormal close of corresponding connections
193 case IB_EVENT_PORT_ACTIVE:
194 port_idx = ibevent->element.port_num - 1;
195 set_bit(port_idx, &smcibdev->port_event_mask);
196 schedule_work(&smcibdev->port_event_work);
203 void smc_ib_dealloc_protection_domain(struct smc_link *lnk)
205 ib_dealloc_pd(lnk->roce_pd);
209 int smc_ib_create_protection_domain(struct smc_link *lnk)
213 lnk->roce_pd = ib_alloc_pd(lnk->smcibdev->ibdev, 0);
214 rc = PTR_ERR_OR_ZERO(lnk->roce_pd);
215 if (IS_ERR(lnk->roce_pd))
220 static void smc_ib_qp_event_handler(struct ib_event *ibevent, void *priv)
222 switch (ibevent->event) {
223 case IB_EVENT_DEVICE_FATAL:
224 case IB_EVENT_GID_CHANGE:
225 case IB_EVENT_PORT_ERR:
226 case IB_EVENT_QP_ACCESS_ERR:
227 /* tbd in follow-on patch:
228 * abnormal close of corresponding connections
236 void smc_ib_destroy_queue_pair(struct smc_link *lnk)
238 ib_destroy_qp(lnk->roce_qp);
242 /* create a queue pair within the protection domain for a link */
243 int smc_ib_create_queue_pair(struct smc_link *lnk)
245 struct ib_qp_init_attr qp_attr = {
246 .event_handler = smc_ib_qp_event_handler,
248 .send_cq = lnk->smcibdev->roce_cq_send,
249 .recv_cq = lnk->smcibdev->roce_cq_recv,
252 .max_send_wr = SMC_WR_BUF_CNT,
253 /* include unsolicited rdma_writes as well,
254 * there are max. 2 RDMA_WRITE per 1 WR_SEND
256 .max_recv_wr = SMC_WR_BUF_CNT * 3,
257 .max_send_sge = SMC_IB_MAX_SEND_SGE,
260 .sq_sig_type = IB_SIGNAL_REQ_WR,
261 .qp_type = IB_QPT_RC,
265 lnk->roce_qp = ib_create_qp(lnk->roce_pd, &qp_attr);
266 rc = PTR_ERR_OR_ZERO(lnk->roce_qp);
267 if (IS_ERR(lnk->roce_qp))
270 smc_wr_remember_qp_attr(lnk);
274 /* map a new TX or RX buffer to DMA */
275 int smc_ib_buf_map(struct smc_ib_device *smcibdev, int buf_size,
276 struct smc_buf_desc *buf_slot,
277 enum dma_data_direction data_direction)
281 if (buf_slot->dma_addr[SMC_SINGLE_LINK])
282 return rc; /* already mapped */
283 buf_slot->dma_addr[SMC_SINGLE_LINK] =
284 ib_dma_map_single(smcibdev->ibdev, buf_slot->cpu_addr,
285 buf_size, data_direction);
286 if (ib_dma_mapping_error(smcibdev->ibdev,
287 buf_slot->dma_addr[SMC_SINGLE_LINK]))
292 void smc_ib_buf_unmap(struct smc_ib_device *smcibdev, int buf_size,
293 struct smc_buf_desc *buf_slot,
294 enum dma_data_direction data_direction)
296 if (!buf_slot->dma_addr[SMC_SINGLE_LINK])
297 return; /* already unmapped */
298 ib_dma_unmap_single(smcibdev->ibdev, *buf_slot->dma_addr, buf_size,
300 buf_slot->dma_addr[SMC_SINGLE_LINK] = 0;
303 static int smc_ib_fill_gid_and_mac(struct smc_ib_device *smcibdev, u8 ibport)
305 struct net_device *ndev;
308 rc = ib_query_gid(smcibdev->ibdev, ibport, 0,
309 &smcibdev->gid[ibport - 1], NULL);
310 /* the SMC protocol requires specification of the roce MAC address;
311 * if net_device cannot be determined, it can be derived from gid 0
313 ndev = smcibdev->ibdev->get_netdev(smcibdev->ibdev, ibport);
315 memcpy(&smcibdev->mac, ndev->dev_addr, ETH_ALEN);
317 memcpy(&smcibdev->mac[ibport - 1][0],
318 &smcibdev->gid[ibport - 1].raw[8], 3);
319 memcpy(&smcibdev->mac[ibport - 1][3],
320 &smcibdev->gid[ibport - 1].raw[13], 3);
321 smcibdev->mac[ibport - 1][0] &= ~0x02;
326 /* Create an identifier unique for this instance of SMC-R.
327 * The MAC-address of the first active registered IB device
328 * plus a random 2-byte number is used to create this identifier.
329 * This name is delivered to the peer during connection initialization.
331 static inline void smc_ib_define_local_systemid(struct smc_ib_device *smcibdev,
334 memcpy(&local_systemid[2], &smcibdev->mac[ibport - 1],
335 sizeof(smcibdev->mac[ibport - 1]));
336 get_random_bytes(&local_systemid[0], 2);
339 bool smc_ib_port_active(struct smc_ib_device *smcibdev, u8 ibport)
341 return smcibdev->pattr[ibport - 1].state == IB_PORT_ACTIVE;
344 int smc_ib_remember_port_attr(struct smc_ib_device *smcibdev, u8 ibport)
348 memset(&smcibdev->pattr[ibport - 1], 0,
349 sizeof(smcibdev->pattr[ibport - 1]));
350 rc = ib_query_port(smcibdev->ibdev, ibport,
351 &smcibdev->pattr[ibport - 1]);
354 rc = smc_ib_fill_gid_and_mac(smcibdev, ibport);
357 if (!strncmp(local_systemid, SMC_LOCAL_SYSTEMID_RESET,
358 sizeof(local_systemid)) &&
359 smc_ib_port_active(smcibdev, ibport))
360 /* create unique system identifier */
361 smc_ib_define_local_systemid(smcibdev, ibport);
366 long smc_ib_setup_per_ibdev(struct smc_ib_device *smcibdev)
368 struct ib_cq_init_attr cqattr = {
369 .cqe = SMC_WR_MAX_CQE, .comp_vector = 0 };
372 smcibdev->roce_cq_send = ib_create_cq(smcibdev->ibdev,
373 smc_wr_tx_cq_handler, NULL,
375 rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_send);
376 if (IS_ERR(smcibdev->roce_cq_send)) {
377 smcibdev->roce_cq_send = NULL;
380 smcibdev->roce_cq_recv = ib_create_cq(smcibdev->ibdev,
381 smc_wr_rx_cq_handler, NULL,
383 rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_recv);
384 if (IS_ERR(smcibdev->roce_cq_recv)) {
385 smcibdev->roce_cq_recv = NULL;
388 INIT_IB_EVENT_HANDLER(&smcibdev->event_handler, smcibdev->ibdev,
389 smc_ib_global_event_handler);
390 ib_register_event_handler(&smcibdev->event_handler);
391 smc_wr_add_dev(smcibdev);
392 smcibdev->initialized = 1;
396 ib_destroy_cq(smcibdev->roce_cq_send);
400 static void smc_ib_cleanup_per_ibdev(struct smc_ib_device *smcibdev)
402 if (!smcibdev->initialized)
404 smc_wr_remove_dev(smcibdev);
405 ib_unregister_event_handler(&smcibdev->event_handler);
406 ib_destroy_cq(smcibdev->roce_cq_recv);
407 ib_destroy_cq(smcibdev->roce_cq_send);
410 static struct ib_client smc_ib_client;
412 /* callback function for ib_register_client() */
413 static void smc_ib_add_dev(struct ib_device *ibdev)
415 struct smc_ib_device *smcibdev;
417 if (ibdev->node_type != RDMA_NODE_IB_CA)
420 smcibdev = kzalloc(sizeof(*smcibdev), GFP_KERNEL);
424 smcibdev->ibdev = ibdev;
425 INIT_WORK(&smcibdev->port_event_work, smc_ib_port_event_work);
427 spin_lock(&smc_ib_devices.lock);
428 list_add_tail(&smcibdev->list, &smc_ib_devices.list);
429 spin_unlock(&smc_ib_devices.lock);
430 ib_set_client_data(ibdev, &smc_ib_client, smcibdev);
433 /* callback function for ib_register_client() */
434 static void smc_ib_remove_dev(struct ib_device *ibdev, void *client_data)
436 struct smc_ib_device *smcibdev;
438 smcibdev = ib_get_client_data(ibdev, &smc_ib_client);
439 ib_set_client_data(ibdev, &smc_ib_client, NULL);
440 spin_lock(&smc_ib_devices.lock);
441 list_del_init(&smcibdev->list); /* remove from smc_ib_devices */
442 spin_unlock(&smc_ib_devices.lock);
443 smc_pnet_remove_by_ibdev(smcibdev);
444 smc_ib_cleanup_per_ibdev(smcibdev);
448 static struct ib_client smc_ib_client = {
450 .add = smc_ib_add_dev,
451 .remove = smc_ib_remove_dev,
454 int __init smc_ib_register_client(void)
456 return ib_register_client(&smc_ib_client);
459 void smc_ib_unregister_client(void)
461 ib_unregister_client(&smc_ib_client);