xref: /openbmc/linux/net/smc/smc_ib.c (revision f7d84fa7)
1 /*
2  *  Shared Memory Communications over RDMA (SMC-R) and RoCE
3  *
4  *  IB infrastructure:
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.
8  *
9  *  Copyright IBM Corp. 2016
10  *
11  *  Author(s):  Ursula Braun <ubraun@linux.vnet.ibm.com>
12  */
13 
14 #include <linux/random.h>
15 #include <linux/workqueue.h>
16 #include <rdma/ib_verbs.h>
17 
18 #include "smc_pnet.h"
19 #include "smc_ib.h"
20 #include "smc_core.h"
21 #include "smc_wr.h"
22 #include "smc.h"
23 
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 */
28 
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),
32 };
33 
34 #define SMC_LOCAL_SYSTEMID_RESET	"%%%%%%%"
35 
36 u8 local_systemid[SMC_SYSTEMID_LEN] = SMC_LOCAL_SYSTEMID_RESET;	/* unique system
37 								 * identifier
38 								 */
39 
40 static int smc_ib_modify_qp_init(struct smc_link *lnk)
41 {
42 	struct ib_qp_attr qp_attr;
43 
44 	memset(&qp_attr, 0, sizeof(qp_attr));
45 	qp_attr.qp_state = IB_QPS_INIT;
46 	qp_attr.pkey_index = 0;
47 	qp_attr.port_num = lnk->ibport;
48 	qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE
49 				| IB_ACCESS_REMOTE_WRITE;
50 	return ib_modify_qp(lnk->roce_qp, &qp_attr,
51 			    IB_QP_STATE | IB_QP_PKEY_INDEX |
52 			    IB_QP_ACCESS_FLAGS | IB_QP_PORT);
53 }
54 
55 static int smc_ib_modify_qp_rtr(struct smc_link *lnk)
56 {
57 	enum ib_qp_attr_mask qp_attr_mask =
58 		IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU | IB_QP_DEST_QPN |
59 		IB_QP_RQ_PSN | IB_QP_MAX_DEST_RD_ATOMIC | IB_QP_MIN_RNR_TIMER;
60 	struct ib_qp_attr qp_attr;
61 
62 	memset(&qp_attr, 0, sizeof(qp_attr));
63 	qp_attr.qp_state = IB_QPS_RTR;
64 	qp_attr.path_mtu = min(lnk->path_mtu, lnk->peer_mtu);
65 	qp_attr.ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
66 	rdma_ah_set_port_num(&qp_attr.ah_attr, lnk->ibport);
67 	rdma_ah_set_grh(&qp_attr.ah_attr, NULL, 0, 0, 1, 0);
68 	rdma_ah_set_dgid_raw(&qp_attr.ah_attr, lnk->peer_gid);
69 	memcpy(&qp_attr.ah_attr.roce.dmac, lnk->peer_mac,
70 	       sizeof(lnk->peer_mac));
71 	qp_attr.dest_qp_num = lnk->peer_qpn;
72 	qp_attr.rq_psn = lnk->peer_psn; /* starting receive packet seq # */
73 	qp_attr.max_dest_rd_atomic = 1; /* max # of resources for incoming
74 					 * requests
75 					 */
76 	qp_attr.min_rnr_timer = SMC_QP_MIN_RNR_TIMER;
77 
78 	return ib_modify_qp(lnk->roce_qp, &qp_attr, qp_attr_mask);
79 }
80 
81 int smc_ib_modify_qp_rts(struct smc_link *lnk)
82 {
83 	struct ib_qp_attr qp_attr;
84 
85 	memset(&qp_attr, 0, sizeof(qp_attr));
86 	qp_attr.qp_state = IB_QPS_RTS;
87 	qp_attr.timeout = SMC_QP_TIMEOUT;	/* local ack timeout */
88 	qp_attr.retry_cnt = SMC_QP_RETRY_CNT;	/* retry count */
89 	qp_attr.rnr_retry = SMC_QP_RNR_RETRY;	/* RNR retries, 7=infinite */
90 	qp_attr.sq_psn = lnk->psn_initial;	/* starting send packet seq # */
91 	qp_attr.max_rd_atomic = 1;	/* # of outstanding RDMA reads and
92 					 * atomic ops allowed
93 					 */
94 	return ib_modify_qp(lnk->roce_qp, &qp_attr,
95 			    IB_QP_STATE | IB_QP_TIMEOUT | IB_QP_RETRY_CNT |
96 			    IB_QP_SQ_PSN | IB_QP_RNR_RETRY |
97 			    IB_QP_MAX_QP_RD_ATOMIC);
98 }
99 
100 int smc_ib_modify_qp_reset(struct smc_link *lnk)
101 {
102 	struct ib_qp_attr qp_attr;
103 
104 	memset(&qp_attr, 0, sizeof(qp_attr));
105 	qp_attr.qp_state = IB_QPS_RESET;
106 	return ib_modify_qp(lnk->roce_qp, &qp_attr, IB_QP_STATE);
107 }
108 
109 int smc_ib_ready_link(struct smc_link *lnk)
110 {
111 	struct smc_link_group *lgr =
112 		container_of(lnk, struct smc_link_group, lnk[0]);
113 	int rc = 0;
114 
115 	rc = smc_ib_modify_qp_init(lnk);
116 	if (rc)
117 		goto out;
118 
119 	rc = smc_ib_modify_qp_rtr(lnk);
120 	if (rc)
121 		goto out;
122 	smc_wr_remember_qp_attr(lnk);
123 	rc = ib_req_notify_cq(lnk->smcibdev->roce_cq_recv,
124 			      IB_CQ_SOLICITED_MASK);
125 	if (rc)
126 		goto out;
127 	rc = smc_wr_rx_post_init(lnk);
128 	if (rc)
129 		goto out;
130 	smc_wr_remember_qp_attr(lnk);
131 
132 	if (lgr->role == SMC_SERV) {
133 		rc = smc_ib_modify_qp_rts(lnk);
134 		if (rc)
135 			goto out;
136 		smc_wr_remember_qp_attr(lnk);
137 	}
138 out:
139 	return rc;
140 }
141 
142 /* process context wrapper for might_sleep smc_ib_remember_port_attr */
143 static void smc_ib_port_event_work(struct work_struct *work)
144 {
145 	struct smc_ib_device *smcibdev = container_of(
146 		work, struct smc_ib_device, port_event_work);
147 	u8 port_idx;
148 
149 	for_each_set_bit(port_idx, &smcibdev->port_event_mask, SMC_MAX_PORTS) {
150 		smc_ib_remember_port_attr(smcibdev, port_idx + 1);
151 		clear_bit(port_idx, &smcibdev->port_event_mask);
152 	}
153 }
154 
155 /* can be called in IRQ context */
156 static void smc_ib_global_event_handler(struct ib_event_handler *handler,
157 					struct ib_event *ibevent)
158 {
159 	struct smc_ib_device *smcibdev;
160 	u8 port_idx;
161 
162 	smcibdev = container_of(handler, struct smc_ib_device, event_handler);
163 
164 	switch (ibevent->event) {
165 	case IB_EVENT_PORT_ERR:
166 		port_idx = ibevent->element.port_num - 1;
167 		set_bit(port_idx, &smcibdev->port_event_mask);
168 		schedule_work(&smcibdev->port_event_work);
169 		/* fall through */
170 	case IB_EVENT_DEVICE_FATAL:
171 		/* tbd in follow-on patch:
172 		 * abnormal close of corresponding connections
173 		 */
174 		break;
175 	case IB_EVENT_PORT_ACTIVE:
176 		port_idx = ibevent->element.port_num - 1;
177 		set_bit(port_idx, &smcibdev->port_event_mask);
178 		schedule_work(&smcibdev->port_event_work);
179 		break;
180 	default:
181 		break;
182 	}
183 }
184 
185 void smc_ib_dealloc_protection_domain(struct smc_link *lnk)
186 {
187 	ib_dealloc_pd(lnk->roce_pd);
188 	lnk->roce_pd = NULL;
189 }
190 
191 int smc_ib_create_protection_domain(struct smc_link *lnk)
192 {
193 	int rc;
194 
195 	lnk->roce_pd = ib_alloc_pd(lnk->smcibdev->ibdev,
196 				   IB_PD_UNSAFE_GLOBAL_RKEY);
197 	rc = PTR_ERR_OR_ZERO(lnk->roce_pd);
198 	if (IS_ERR(lnk->roce_pd))
199 		lnk->roce_pd = NULL;
200 	return rc;
201 }
202 
203 static void smc_ib_qp_event_handler(struct ib_event *ibevent, void *priv)
204 {
205 	switch (ibevent->event) {
206 	case IB_EVENT_DEVICE_FATAL:
207 	case IB_EVENT_GID_CHANGE:
208 	case IB_EVENT_PORT_ERR:
209 	case IB_EVENT_QP_ACCESS_ERR:
210 		/* tbd in follow-on patch:
211 		 * abnormal close of corresponding connections
212 		 */
213 		break;
214 	default:
215 		break;
216 	}
217 }
218 
219 void smc_ib_destroy_queue_pair(struct smc_link *lnk)
220 {
221 	ib_destroy_qp(lnk->roce_qp);
222 	lnk->roce_qp = NULL;
223 }
224 
225 /* create a queue pair within the protection domain for a link */
226 int smc_ib_create_queue_pair(struct smc_link *lnk)
227 {
228 	struct ib_qp_init_attr qp_attr = {
229 		.event_handler = smc_ib_qp_event_handler,
230 		.qp_context = lnk,
231 		.send_cq = lnk->smcibdev->roce_cq_send,
232 		.recv_cq = lnk->smcibdev->roce_cq_recv,
233 		.srq = NULL,
234 		.cap = {
235 			.max_send_wr = SMC_WR_BUF_CNT,
236 				/* include unsolicited rdma_writes as well,
237 				 * there are max. 2 RDMA_WRITE per 1 WR_SEND
238 				 */
239 			.max_recv_wr = SMC_WR_BUF_CNT * 3,
240 			.max_send_sge = SMC_IB_MAX_SEND_SGE,
241 			.max_recv_sge = 1,
242 		},
243 		.sq_sig_type = IB_SIGNAL_REQ_WR,
244 		.qp_type = IB_QPT_RC,
245 	};
246 	int rc;
247 
248 	lnk->roce_qp = ib_create_qp(lnk->roce_pd, &qp_attr);
249 	rc = PTR_ERR_OR_ZERO(lnk->roce_qp);
250 	if (IS_ERR(lnk->roce_qp))
251 		lnk->roce_qp = NULL;
252 	else
253 		smc_wr_remember_qp_attr(lnk);
254 	return rc;
255 }
256 
257 /* map a new TX or RX buffer to DMA */
258 int smc_ib_buf_map(struct smc_ib_device *smcibdev, int buf_size,
259 		   struct smc_buf_desc *buf_slot,
260 		   enum dma_data_direction data_direction)
261 {
262 	int rc = 0;
263 
264 	if (buf_slot->dma_addr[SMC_SINGLE_LINK])
265 		return rc; /* already mapped */
266 	buf_slot->dma_addr[SMC_SINGLE_LINK] =
267 		ib_dma_map_single(smcibdev->ibdev, buf_slot->cpu_addr,
268 				  buf_size, data_direction);
269 	if (ib_dma_mapping_error(smcibdev->ibdev,
270 				 buf_slot->dma_addr[SMC_SINGLE_LINK]))
271 		rc = -EIO;
272 	return rc;
273 }
274 
275 void smc_ib_buf_unmap(struct smc_ib_device *smcibdev, int buf_size,
276 		      struct smc_buf_desc *buf_slot,
277 		      enum dma_data_direction data_direction)
278 {
279 	if (!buf_slot->dma_addr[SMC_SINGLE_LINK])
280 		return; /* already unmapped */
281 	ib_dma_unmap_single(smcibdev->ibdev, *buf_slot->dma_addr, buf_size,
282 			    data_direction);
283 	buf_slot->dma_addr[SMC_SINGLE_LINK] = 0;
284 }
285 
286 static int smc_ib_fill_gid_and_mac(struct smc_ib_device *smcibdev, u8 ibport)
287 {
288 	struct net_device *ndev;
289 	int rc;
290 
291 	rc = ib_query_gid(smcibdev->ibdev, ibport, 0,
292 			  &smcibdev->gid[ibport - 1], NULL);
293 	/* the SMC protocol requires specification of the roce MAC address;
294 	 * if net_device cannot be determined, it can be derived from gid 0
295 	 */
296 	ndev = smcibdev->ibdev->get_netdev(smcibdev->ibdev, ibport);
297 	if (ndev) {
298 		memcpy(&smcibdev->mac, ndev->dev_addr, ETH_ALEN);
299 	} else if (!rc) {
300 		memcpy(&smcibdev->mac[ibport - 1][0],
301 		       &smcibdev->gid[ibport - 1].raw[8], 3);
302 		memcpy(&smcibdev->mac[ibport - 1][3],
303 		       &smcibdev->gid[ibport - 1].raw[13], 3);
304 		smcibdev->mac[ibport - 1][0] &= ~0x02;
305 	}
306 	return rc;
307 }
308 
309 /* Create an identifier unique for this instance of SMC-R.
310  * The MAC-address of the first active registered IB device
311  * plus a random 2-byte number is used to create this identifier.
312  * This name is delivered to the peer during connection initialization.
313  */
314 static inline void smc_ib_define_local_systemid(struct smc_ib_device *smcibdev,
315 						u8 ibport)
316 {
317 	memcpy(&local_systemid[2], &smcibdev->mac[ibport - 1],
318 	       sizeof(smcibdev->mac[ibport - 1]));
319 	get_random_bytes(&local_systemid[0], 2);
320 }
321 
322 bool smc_ib_port_active(struct smc_ib_device *smcibdev, u8 ibport)
323 {
324 	return smcibdev->pattr[ibport - 1].state == IB_PORT_ACTIVE;
325 }
326 
327 int smc_ib_remember_port_attr(struct smc_ib_device *smcibdev, u8 ibport)
328 {
329 	int rc;
330 
331 	memset(&smcibdev->pattr[ibport - 1], 0,
332 	       sizeof(smcibdev->pattr[ibport - 1]));
333 	rc = ib_query_port(smcibdev->ibdev, ibport,
334 			   &smcibdev->pattr[ibport - 1]);
335 	if (rc)
336 		goto out;
337 	rc = smc_ib_fill_gid_and_mac(smcibdev, ibport);
338 	if (rc)
339 		goto out;
340 	if (!strncmp(local_systemid, SMC_LOCAL_SYSTEMID_RESET,
341 		     sizeof(local_systemid)) &&
342 	    smc_ib_port_active(smcibdev, ibport))
343 		/* create unique system identifier */
344 		smc_ib_define_local_systemid(smcibdev, ibport);
345 out:
346 	return rc;
347 }
348 
349 long smc_ib_setup_per_ibdev(struct smc_ib_device *smcibdev)
350 {
351 	struct ib_cq_init_attr cqattr =	{
352 		.cqe = SMC_WR_MAX_CQE, .comp_vector = 0 };
353 	long rc;
354 
355 	smcibdev->roce_cq_send = ib_create_cq(smcibdev->ibdev,
356 					      smc_wr_tx_cq_handler, NULL,
357 					      smcibdev, &cqattr);
358 	rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_send);
359 	if (IS_ERR(smcibdev->roce_cq_send)) {
360 		smcibdev->roce_cq_send = NULL;
361 		return rc;
362 	}
363 	smcibdev->roce_cq_recv = ib_create_cq(smcibdev->ibdev,
364 					      smc_wr_rx_cq_handler, NULL,
365 					      smcibdev, &cqattr);
366 	rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_recv);
367 	if (IS_ERR(smcibdev->roce_cq_recv)) {
368 		smcibdev->roce_cq_recv = NULL;
369 		goto err;
370 	}
371 	INIT_IB_EVENT_HANDLER(&smcibdev->event_handler, smcibdev->ibdev,
372 			      smc_ib_global_event_handler);
373 	ib_register_event_handler(&smcibdev->event_handler);
374 	smc_wr_add_dev(smcibdev);
375 	smcibdev->initialized = 1;
376 	return rc;
377 
378 err:
379 	ib_destroy_cq(smcibdev->roce_cq_send);
380 	return rc;
381 }
382 
383 static void smc_ib_cleanup_per_ibdev(struct smc_ib_device *smcibdev)
384 {
385 	if (!smcibdev->initialized)
386 		return;
387 	smc_wr_remove_dev(smcibdev);
388 	ib_unregister_event_handler(&smcibdev->event_handler);
389 	ib_destroy_cq(smcibdev->roce_cq_recv);
390 	ib_destroy_cq(smcibdev->roce_cq_send);
391 }
392 
393 static struct ib_client smc_ib_client;
394 
395 /* callback function for ib_register_client() */
396 static void smc_ib_add_dev(struct ib_device *ibdev)
397 {
398 	struct smc_ib_device *smcibdev;
399 
400 	if (ibdev->node_type != RDMA_NODE_IB_CA)
401 		return;
402 
403 	smcibdev = kzalloc(sizeof(*smcibdev), GFP_KERNEL);
404 	if (!smcibdev)
405 		return;
406 
407 	smcibdev->ibdev = ibdev;
408 	INIT_WORK(&smcibdev->port_event_work, smc_ib_port_event_work);
409 
410 	spin_lock(&smc_ib_devices.lock);
411 	list_add_tail(&smcibdev->list, &smc_ib_devices.list);
412 	spin_unlock(&smc_ib_devices.lock);
413 	ib_set_client_data(ibdev, &smc_ib_client, smcibdev);
414 }
415 
416 /* callback function for ib_register_client() */
417 static void smc_ib_remove_dev(struct ib_device *ibdev, void *client_data)
418 {
419 	struct smc_ib_device *smcibdev;
420 
421 	smcibdev = ib_get_client_data(ibdev, &smc_ib_client);
422 	ib_set_client_data(ibdev, &smc_ib_client, NULL);
423 	spin_lock(&smc_ib_devices.lock);
424 	list_del_init(&smcibdev->list); /* remove from smc_ib_devices */
425 	spin_unlock(&smc_ib_devices.lock);
426 	smc_pnet_remove_by_ibdev(smcibdev);
427 	smc_ib_cleanup_per_ibdev(smcibdev);
428 	kfree(smcibdev);
429 }
430 
431 static struct ib_client smc_ib_client = {
432 	.name	= "smc_ib",
433 	.add	= smc_ib_add_dev,
434 	.remove = smc_ib_remove_dev,
435 };
436 
437 int __init smc_ib_register_client(void)
438 {
439 	return ib_register_client(&smc_ib_client);
440 }
441 
442 void smc_ib_unregister_client(void)
443 {
444 	ib_unregister_client(&smc_ib_client);
445 }
446