xref: /linux/net/smc/smc_ib.c (revision fd639726bf15fca8ee1a00dce8e0096d0ad9bd18)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  *  Shared Memory Communications over RDMA (SMC-R) and RoCE
4  *
5  *  IB infrastructure:
6  *  Establish SMC-R as an Infiniband Client to be notified about added and
7  *  removed IB devices of type RDMA.
8  *  Determine device and port characteristics for these IB devices.
9  *
10  *  Copyright IBM Corp. 2016
11  *
12  *  Author(s):  Ursula Braun <ubraun@linux.vnet.ibm.com>
13  */
14 
15 #include <linux/random.h>
16 #include <linux/workqueue.h>
17 #include <linux/scatterlist.h>
18 #include <rdma/ib_verbs.h>
19 
20 #include "smc_pnet.h"
21 #include "smc_ib.h"
22 #include "smc_core.h"
23 #include "smc_wr.h"
24 #include "smc.h"
25 
26 #define SMC_QP_MIN_RNR_TIMER		5
27 #define SMC_QP_TIMEOUT			15 /* 4096 * 2 ** timeout usec */
28 #define SMC_QP_RETRY_CNT			7 /* 7: infinite */
29 #define SMC_QP_RNR_RETRY			7 /* 7: infinite */
30 
31 struct smc_ib_devices smc_ib_devices = {	/* smc-registered ib devices */
32 	.lock = __SPIN_LOCK_UNLOCKED(smc_ib_devices.lock),
33 	.list = LIST_HEAD_INIT(smc_ib_devices.list),
34 };
35 
36 #define SMC_LOCAL_SYSTEMID_RESET	"%%%%%%%"
37 
38 u8 local_systemid[SMC_SYSTEMID_LEN] = SMC_LOCAL_SYSTEMID_RESET;	/* unique system
39 								 * identifier
40 								 */
41 
42 static int smc_ib_modify_qp_init(struct smc_link *lnk)
43 {
44 	struct ib_qp_attr qp_attr;
45 
46 	memset(&qp_attr, 0, sizeof(qp_attr));
47 	qp_attr.qp_state = IB_QPS_INIT;
48 	qp_attr.pkey_index = 0;
49 	qp_attr.port_num = lnk->ibport;
50 	qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE
51 				| IB_ACCESS_REMOTE_WRITE;
52 	return ib_modify_qp(lnk->roce_qp, &qp_attr,
53 			    IB_QP_STATE | IB_QP_PKEY_INDEX |
54 			    IB_QP_ACCESS_FLAGS | IB_QP_PORT);
55 }
56 
57 static int smc_ib_modify_qp_rtr(struct smc_link *lnk)
58 {
59 	enum ib_qp_attr_mask qp_attr_mask =
60 		IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU | IB_QP_DEST_QPN |
61 		IB_QP_RQ_PSN | IB_QP_MAX_DEST_RD_ATOMIC | IB_QP_MIN_RNR_TIMER;
62 	struct ib_qp_attr qp_attr;
63 
64 	memset(&qp_attr, 0, sizeof(qp_attr));
65 	qp_attr.qp_state = IB_QPS_RTR;
66 	qp_attr.path_mtu = min(lnk->path_mtu, lnk->peer_mtu);
67 	qp_attr.ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
68 	rdma_ah_set_port_num(&qp_attr.ah_attr, lnk->ibport);
69 	rdma_ah_set_grh(&qp_attr.ah_attr, NULL, 0, 0, 1, 0);
70 	rdma_ah_set_dgid_raw(&qp_attr.ah_attr, lnk->peer_gid);
71 	memcpy(&qp_attr.ah_attr.roce.dmac, lnk->peer_mac,
72 	       sizeof(lnk->peer_mac));
73 	qp_attr.dest_qp_num = lnk->peer_qpn;
74 	qp_attr.rq_psn = lnk->peer_psn; /* starting receive packet seq # */
75 	qp_attr.max_dest_rd_atomic = 1; /* max # of resources for incoming
76 					 * requests
77 					 */
78 	qp_attr.min_rnr_timer = SMC_QP_MIN_RNR_TIMER;
79 
80 	return ib_modify_qp(lnk->roce_qp, &qp_attr, qp_attr_mask);
81 }
82 
83 int smc_ib_modify_qp_rts(struct smc_link *lnk)
84 {
85 	struct ib_qp_attr qp_attr;
86 
87 	memset(&qp_attr, 0, sizeof(qp_attr));
88 	qp_attr.qp_state = IB_QPS_RTS;
89 	qp_attr.timeout = SMC_QP_TIMEOUT;	/* local ack timeout */
90 	qp_attr.retry_cnt = SMC_QP_RETRY_CNT;	/* retry count */
91 	qp_attr.rnr_retry = SMC_QP_RNR_RETRY;	/* RNR retries, 7=infinite */
92 	qp_attr.sq_psn = lnk->psn_initial;	/* starting send packet seq # */
93 	qp_attr.max_rd_atomic = 1;	/* # of outstanding RDMA reads and
94 					 * atomic ops allowed
95 					 */
96 	return ib_modify_qp(lnk->roce_qp, &qp_attr,
97 			    IB_QP_STATE | IB_QP_TIMEOUT | IB_QP_RETRY_CNT |
98 			    IB_QP_SQ_PSN | IB_QP_RNR_RETRY |
99 			    IB_QP_MAX_QP_RD_ATOMIC);
100 }
101 
102 int smc_ib_modify_qp_reset(struct smc_link *lnk)
103 {
104 	struct ib_qp_attr qp_attr;
105 
106 	memset(&qp_attr, 0, sizeof(qp_attr));
107 	qp_attr.qp_state = IB_QPS_RESET;
108 	return ib_modify_qp(lnk->roce_qp, &qp_attr, IB_QP_STATE);
109 }
110 
111 int smc_ib_ready_link(struct smc_link *lnk)
112 {
113 	struct smc_link_group *lgr =
114 		container_of(lnk, struct smc_link_group, lnk[0]);
115 	int rc = 0;
116 
117 	rc = smc_ib_modify_qp_init(lnk);
118 	if (rc)
119 		goto out;
120 
121 	rc = smc_ib_modify_qp_rtr(lnk);
122 	if (rc)
123 		goto out;
124 	smc_wr_remember_qp_attr(lnk);
125 	rc = ib_req_notify_cq(lnk->smcibdev->roce_cq_recv,
126 			      IB_CQ_SOLICITED_MASK);
127 	if (rc)
128 		goto out;
129 	rc = smc_wr_rx_post_init(lnk);
130 	if (rc)
131 		goto out;
132 	smc_wr_remember_qp_attr(lnk);
133 
134 	if (lgr->role == SMC_SERV) {
135 		rc = smc_ib_modify_qp_rts(lnk);
136 		if (rc)
137 			goto out;
138 		smc_wr_remember_qp_attr(lnk);
139 	}
140 out:
141 	return rc;
142 }
143 
144 /* process context wrapper for might_sleep smc_ib_remember_port_attr */
145 static void smc_ib_port_event_work(struct work_struct *work)
146 {
147 	struct smc_ib_device *smcibdev = container_of(
148 		work, struct smc_ib_device, port_event_work);
149 	u8 port_idx;
150 
151 	for_each_set_bit(port_idx, &smcibdev->port_event_mask, SMC_MAX_PORTS) {
152 		smc_ib_remember_port_attr(smcibdev, port_idx + 1);
153 		clear_bit(port_idx, &smcibdev->port_event_mask);
154 	}
155 }
156 
157 /* can be called in IRQ context */
158 static void smc_ib_global_event_handler(struct ib_event_handler *handler,
159 					struct ib_event *ibevent)
160 {
161 	struct smc_ib_device *smcibdev;
162 	u8 port_idx;
163 
164 	smcibdev = container_of(handler, struct smc_ib_device, event_handler);
165 
166 	switch (ibevent->event) {
167 	case IB_EVENT_PORT_ERR:
168 		port_idx = ibevent->element.port_num - 1;
169 		set_bit(port_idx, &smcibdev->port_event_mask);
170 		schedule_work(&smcibdev->port_event_work);
171 		/* fall through */
172 	case IB_EVENT_DEVICE_FATAL:
173 		/* tbd in follow-on patch:
174 		 * abnormal close of corresponding connections
175 		 */
176 		break;
177 	case IB_EVENT_PORT_ACTIVE:
178 		port_idx = ibevent->element.port_num - 1;
179 		set_bit(port_idx, &smcibdev->port_event_mask);
180 		schedule_work(&smcibdev->port_event_work);
181 		break;
182 	default:
183 		break;
184 	}
185 }
186 
187 void smc_ib_dealloc_protection_domain(struct smc_link *lnk)
188 {
189 	ib_dealloc_pd(lnk->roce_pd);
190 	lnk->roce_pd = NULL;
191 }
192 
193 int smc_ib_create_protection_domain(struct smc_link *lnk)
194 {
195 	int rc;
196 
197 	lnk->roce_pd = ib_alloc_pd(lnk->smcibdev->ibdev, 0);
198 	rc = PTR_ERR_OR_ZERO(lnk->roce_pd);
199 	if (IS_ERR(lnk->roce_pd))
200 		lnk->roce_pd = NULL;
201 	return rc;
202 }
203 
204 static void smc_ib_qp_event_handler(struct ib_event *ibevent, void *priv)
205 {
206 	switch (ibevent->event) {
207 	case IB_EVENT_DEVICE_FATAL:
208 	case IB_EVENT_GID_CHANGE:
209 	case IB_EVENT_PORT_ERR:
210 	case IB_EVENT_QP_ACCESS_ERR:
211 		/* tbd in follow-on patch:
212 		 * abnormal close of corresponding connections
213 		 */
214 		break;
215 	default:
216 		break;
217 	}
218 }
219 
220 void smc_ib_destroy_queue_pair(struct smc_link *lnk)
221 {
222 	ib_destroy_qp(lnk->roce_qp);
223 	lnk->roce_qp = NULL;
224 }
225 
226 /* create a queue pair within the protection domain for a link */
227 int smc_ib_create_queue_pair(struct smc_link *lnk)
228 {
229 	struct ib_qp_init_attr qp_attr = {
230 		.event_handler = smc_ib_qp_event_handler,
231 		.qp_context = lnk,
232 		.send_cq = lnk->smcibdev->roce_cq_send,
233 		.recv_cq = lnk->smcibdev->roce_cq_recv,
234 		.srq = NULL,
235 		.cap = {
236 				/* include unsolicited rdma_writes as well,
237 				 * there are max. 2 RDMA_WRITE per 1 WR_SEND
238 				 */
239 			.max_send_wr = SMC_WR_BUF_CNT * 3,
240 			.max_recv_wr = SMC_WR_BUF_CNT * 3,
241 			.max_send_sge = SMC_IB_MAX_SEND_SGE,
242 			.max_recv_sge = 1,
243 		},
244 		.sq_sig_type = IB_SIGNAL_REQ_WR,
245 		.qp_type = IB_QPT_RC,
246 	};
247 	int rc;
248 
249 	lnk->roce_qp = ib_create_qp(lnk->roce_pd, &qp_attr);
250 	rc = PTR_ERR_OR_ZERO(lnk->roce_qp);
251 	if (IS_ERR(lnk->roce_qp))
252 		lnk->roce_qp = NULL;
253 	else
254 		smc_wr_remember_qp_attr(lnk);
255 	return rc;
256 }
257 
258 void smc_ib_put_memory_region(struct ib_mr *mr)
259 {
260 	ib_dereg_mr(mr);
261 }
262 
263 static int smc_ib_map_mr_sg(struct smc_buf_desc *buf_slot)
264 {
265 	unsigned int offset = 0;
266 	int sg_num;
267 
268 	/* map the largest prefix of a dma mapped SG list */
269 	sg_num = ib_map_mr_sg(buf_slot->mr_rx[SMC_SINGLE_LINK],
270 			      buf_slot->sgt[SMC_SINGLE_LINK].sgl,
271 			      buf_slot->sgt[SMC_SINGLE_LINK].orig_nents,
272 			      &offset, PAGE_SIZE);
273 
274 	return sg_num;
275 }
276 
277 /* Allocate a memory region and map the dma mapped SG list of buf_slot */
278 int smc_ib_get_memory_region(struct ib_pd *pd, int access_flags,
279 			     struct smc_buf_desc *buf_slot)
280 {
281 	if (buf_slot->mr_rx[SMC_SINGLE_LINK])
282 		return 0; /* already done */
283 
284 	buf_slot->mr_rx[SMC_SINGLE_LINK] =
285 		ib_alloc_mr(pd, IB_MR_TYPE_MEM_REG, 1 << buf_slot->order);
286 	if (IS_ERR(buf_slot->mr_rx[SMC_SINGLE_LINK])) {
287 		int rc;
288 
289 		rc = PTR_ERR(buf_slot->mr_rx[SMC_SINGLE_LINK]);
290 		buf_slot->mr_rx[SMC_SINGLE_LINK] = NULL;
291 		return rc;
292 	}
293 
294 	if (smc_ib_map_mr_sg(buf_slot) != 1)
295 		return -EINVAL;
296 
297 	return 0;
298 }
299 
300 /* synchronize buffer usage for cpu access */
301 void smc_ib_sync_sg_for_cpu(struct smc_ib_device *smcibdev,
302 			    struct smc_buf_desc *buf_slot,
303 			    enum dma_data_direction data_direction)
304 {
305 	struct scatterlist *sg;
306 	unsigned int i;
307 
308 	/* for now there is just one DMA address */
309 	for_each_sg(buf_slot->sgt[SMC_SINGLE_LINK].sgl, sg,
310 		    buf_slot->sgt[SMC_SINGLE_LINK].nents, i) {
311 		if (!sg_dma_len(sg))
312 			break;
313 		ib_dma_sync_single_for_cpu(smcibdev->ibdev,
314 					   sg_dma_address(sg),
315 					   sg_dma_len(sg),
316 					   data_direction);
317 	}
318 }
319 
320 /* synchronize buffer usage for device access */
321 void smc_ib_sync_sg_for_device(struct smc_ib_device *smcibdev,
322 			       struct smc_buf_desc *buf_slot,
323 			       enum dma_data_direction data_direction)
324 {
325 	struct scatterlist *sg;
326 	unsigned int i;
327 
328 	/* for now there is just one DMA address */
329 	for_each_sg(buf_slot->sgt[SMC_SINGLE_LINK].sgl, sg,
330 		    buf_slot->sgt[SMC_SINGLE_LINK].nents, i) {
331 		if (!sg_dma_len(sg))
332 			break;
333 		ib_dma_sync_single_for_device(smcibdev->ibdev,
334 					      sg_dma_address(sg),
335 					      sg_dma_len(sg),
336 					      data_direction);
337 	}
338 }
339 
340 /* Map a new TX or RX buffer SG-table to DMA */
341 int smc_ib_buf_map_sg(struct smc_ib_device *smcibdev,
342 		      struct smc_buf_desc *buf_slot,
343 		      enum dma_data_direction data_direction)
344 {
345 	int mapped_nents;
346 
347 	mapped_nents = ib_dma_map_sg(smcibdev->ibdev,
348 				     buf_slot->sgt[SMC_SINGLE_LINK].sgl,
349 				     buf_slot->sgt[SMC_SINGLE_LINK].orig_nents,
350 				     data_direction);
351 	if (!mapped_nents)
352 		return -ENOMEM;
353 
354 	return mapped_nents;
355 }
356 
357 void smc_ib_buf_unmap_sg(struct smc_ib_device *smcibdev,
358 			 struct smc_buf_desc *buf_slot,
359 			 enum dma_data_direction data_direction)
360 {
361 	if (!buf_slot->sgt[SMC_SINGLE_LINK].sgl->dma_address)
362 		return; /* already unmapped */
363 
364 	ib_dma_unmap_sg(smcibdev->ibdev,
365 			buf_slot->sgt[SMC_SINGLE_LINK].sgl,
366 			buf_slot->sgt[SMC_SINGLE_LINK].orig_nents,
367 			data_direction);
368 	buf_slot->sgt[SMC_SINGLE_LINK].sgl->dma_address = 0;
369 }
370 
371 static int smc_ib_fill_gid_and_mac(struct smc_ib_device *smcibdev, u8 ibport)
372 {
373 	struct ib_gid_attr gattr;
374 	int rc;
375 
376 	rc = ib_query_gid(smcibdev->ibdev, ibport, 0,
377 			  &smcibdev->gid[ibport - 1], &gattr);
378 	if (rc || !gattr.ndev)
379 		return -ENODEV;
380 
381 	memcpy(smcibdev->mac[ibport - 1], gattr.ndev->dev_addr, ETH_ALEN);
382 	dev_put(gattr.ndev);
383 	return 0;
384 }
385 
386 /* Create an identifier unique for this instance of SMC-R.
387  * The MAC-address of the first active registered IB device
388  * plus a random 2-byte number is used to create this identifier.
389  * This name is delivered to the peer during connection initialization.
390  */
391 static inline void smc_ib_define_local_systemid(struct smc_ib_device *smcibdev,
392 						u8 ibport)
393 {
394 	memcpy(&local_systemid[2], &smcibdev->mac[ibport - 1],
395 	       sizeof(smcibdev->mac[ibport - 1]));
396 	get_random_bytes(&local_systemid[0], 2);
397 }
398 
399 bool smc_ib_port_active(struct smc_ib_device *smcibdev, u8 ibport)
400 {
401 	return smcibdev->pattr[ibport - 1].state == IB_PORT_ACTIVE;
402 }
403 
404 int smc_ib_remember_port_attr(struct smc_ib_device *smcibdev, u8 ibport)
405 {
406 	int rc;
407 
408 	memset(&smcibdev->pattr[ibport - 1], 0,
409 	       sizeof(smcibdev->pattr[ibport - 1]));
410 	rc = ib_query_port(smcibdev->ibdev, ibport,
411 			   &smcibdev->pattr[ibport - 1]);
412 	if (rc)
413 		goto out;
414 	/* the SMC protocol requires specification of the RoCE MAC address */
415 	rc = smc_ib_fill_gid_and_mac(smcibdev, ibport);
416 	if (rc)
417 		goto out;
418 	if (!strncmp(local_systemid, SMC_LOCAL_SYSTEMID_RESET,
419 		     sizeof(local_systemid)) &&
420 	    smc_ib_port_active(smcibdev, ibport))
421 		/* create unique system identifier */
422 		smc_ib_define_local_systemid(smcibdev, ibport);
423 out:
424 	return rc;
425 }
426 
427 long smc_ib_setup_per_ibdev(struct smc_ib_device *smcibdev)
428 {
429 	struct ib_cq_init_attr cqattr =	{
430 		.cqe = SMC_WR_MAX_CQE, .comp_vector = 0 };
431 	long rc;
432 
433 	smcibdev->roce_cq_send = ib_create_cq(smcibdev->ibdev,
434 					      smc_wr_tx_cq_handler, NULL,
435 					      smcibdev, &cqattr);
436 	rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_send);
437 	if (IS_ERR(smcibdev->roce_cq_send)) {
438 		smcibdev->roce_cq_send = NULL;
439 		return rc;
440 	}
441 	smcibdev->roce_cq_recv = ib_create_cq(smcibdev->ibdev,
442 					      smc_wr_rx_cq_handler, NULL,
443 					      smcibdev, &cqattr);
444 	rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_recv);
445 	if (IS_ERR(smcibdev->roce_cq_recv)) {
446 		smcibdev->roce_cq_recv = NULL;
447 		goto err;
448 	}
449 	INIT_IB_EVENT_HANDLER(&smcibdev->event_handler, smcibdev->ibdev,
450 			      smc_ib_global_event_handler);
451 	ib_register_event_handler(&smcibdev->event_handler);
452 	smc_wr_add_dev(smcibdev);
453 	smcibdev->initialized = 1;
454 	return rc;
455 
456 err:
457 	ib_destroy_cq(smcibdev->roce_cq_send);
458 	return rc;
459 }
460 
461 static void smc_ib_cleanup_per_ibdev(struct smc_ib_device *smcibdev)
462 {
463 	if (!smcibdev->initialized)
464 		return;
465 	smc_wr_remove_dev(smcibdev);
466 	ib_unregister_event_handler(&smcibdev->event_handler);
467 	ib_destroy_cq(smcibdev->roce_cq_recv);
468 	ib_destroy_cq(smcibdev->roce_cq_send);
469 }
470 
471 static struct ib_client smc_ib_client;
472 
473 /* callback function for ib_register_client() */
474 static void smc_ib_add_dev(struct ib_device *ibdev)
475 {
476 	struct smc_ib_device *smcibdev;
477 
478 	if (ibdev->node_type != RDMA_NODE_IB_CA)
479 		return;
480 
481 	smcibdev = kzalloc(sizeof(*smcibdev), GFP_KERNEL);
482 	if (!smcibdev)
483 		return;
484 
485 	smcibdev->ibdev = ibdev;
486 	INIT_WORK(&smcibdev->port_event_work, smc_ib_port_event_work);
487 
488 	spin_lock(&smc_ib_devices.lock);
489 	list_add_tail(&smcibdev->list, &smc_ib_devices.list);
490 	spin_unlock(&smc_ib_devices.lock);
491 	ib_set_client_data(ibdev, &smc_ib_client, smcibdev);
492 }
493 
494 /* callback function for ib_register_client() */
495 static void smc_ib_remove_dev(struct ib_device *ibdev, void *client_data)
496 {
497 	struct smc_ib_device *smcibdev;
498 
499 	smcibdev = ib_get_client_data(ibdev, &smc_ib_client);
500 	ib_set_client_data(ibdev, &smc_ib_client, NULL);
501 	spin_lock(&smc_ib_devices.lock);
502 	list_del_init(&smcibdev->list); /* remove from smc_ib_devices */
503 	spin_unlock(&smc_ib_devices.lock);
504 	smc_pnet_remove_by_ibdev(smcibdev);
505 	smc_ib_cleanup_per_ibdev(smcibdev);
506 	kfree(smcibdev);
507 }
508 
509 static struct ib_client smc_ib_client = {
510 	.name	= "smc_ib",
511 	.add	= smc_ib_add_dev,
512 	.remove = smc_ib_remove_dev,
513 };
514 
515 int __init smc_ib_register_client(void)
516 {
517 	return ib_register_client(&smc_ib_client);
518 }
519 
520 void smc_ib_unregister_client(void)
521 {
522 	ib_unregister_client(&smc_ib_client);
523 }
524