xref: /linux/drivers/infiniband/ulp/rtrs/rtrs.c (revision a06c3fad49a50d5d5eb078f93e70f4d3eca5d5a5)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * RDMA Transport Layer
4  *
5  * Copyright (c) 2014 - 2018 ProfitBricks GmbH. All rights reserved.
6  * Copyright (c) 2018 - 2019 1&1 IONOS Cloud GmbH. All rights reserved.
7  * Copyright (c) 2019 - 2020 1&1 IONOS SE. All rights reserved.
8  */
9 #undef pr_fmt
10 #define pr_fmt(fmt) KBUILD_MODNAME " L" __stringify(__LINE__) ": " fmt
11 
12 #include <linux/module.h>
13 #include <linux/inet.h>
14 
15 #include "rtrs-pri.h"
16 #include "rtrs-log.h"
17 
18 MODULE_DESCRIPTION("RDMA Transport Core");
19 MODULE_LICENSE("GPL");
20 
21 struct rtrs_iu *rtrs_iu_alloc(u32 iu_num, size_t size, gfp_t gfp_mask,
22 			      struct ib_device *dma_dev,
23 			      enum dma_data_direction dir,
24 			      void (*done)(struct ib_cq *cq, struct ib_wc *wc))
25 {
26 	struct rtrs_iu *ius, *iu;
27 	int i;
28 
29 	ius = kcalloc(iu_num, sizeof(*ius), gfp_mask);
30 	if (!ius)
31 		return NULL;
32 	for (i = 0; i < iu_num; i++) {
33 		iu = &ius[i];
34 		iu->direction = dir;
35 		iu->buf = kzalloc(size, gfp_mask);
36 		if (!iu->buf)
37 			goto err;
38 
39 		iu->dma_addr = ib_dma_map_single(dma_dev, iu->buf, size, dir);
40 		if (ib_dma_mapping_error(dma_dev, iu->dma_addr)) {
41 			kfree(iu->buf);
42 			goto err;
43 		}
44 
45 		iu->cqe.done  = done;
46 		iu->size      = size;
47 	}
48 	return ius;
49 err:
50 	rtrs_iu_free(ius, dma_dev, i);
51 	return NULL;
52 }
53 EXPORT_SYMBOL_GPL(rtrs_iu_alloc);
54 
55 void rtrs_iu_free(struct rtrs_iu *ius, struct ib_device *ibdev, u32 queue_num)
56 {
57 	struct rtrs_iu *iu;
58 	int i;
59 
60 	if (!ius)
61 		return;
62 
63 	for (i = 0; i < queue_num; i++) {
64 		iu = &ius[i];
65 		ib_dma_unmap_single(ibdev, iu->dma_addr, iu->size, iu->direction);
66 		kfree(iu->buf);
67 	}
68 	kfree(ius);
69 }
70 EXPORT_SYMBOL_GPL(rtrs_iu_free);
71 
72 int rtrs_iu_post_recv(struct rtrs_con *con, struct rtrs_iu *iu)
73 {
74 	struct rtrs_path *path = con->path;
75 	struct ib_recv_wr wr;
76 	struct ib_sge list;
77 
78 	list.addr   = iu->dma_addr;
79 	list.length = iu->size;
80 	list.lkey   = path->dev->ib_pd->local_dma_lkey;
81 
82 	if (list.length == 0) {
83 		rtrs_wrn(con->path,
84 			  "Posting receive work request failed, sg list is empty\n");
85 		return -EINVAL;
86 	}
87 	wr = (struct ib_recv_wr) {
88 		.wr_cqe  = &iu->cqe,
89 		.sg_list = &list,
90 		.num_sge = 1,
91 	};
92 
93 	return ib_post_recv(con->qp, &wr, NULL);
94 }
95 EXPORT_SYMBOL_GPL(rtrs_iu_post_recv);
96 
97 int rtrs_post_recv_empty(struct rtrs_con *con, struct ib_cqe *cqe)
98 {
99 	struct ib_recv_wr wr;
100 
101 	wr = (struct ib_recv_wr) {
102 		.wr_cqe  = cqe,
103 	};
104 
105 	return ib_post_recv(con->qp, &wr, NULL);
106 }
107 EXPORT_SYMBOL_GPL(rtrs_post_recv_empty);
108 
109 static int rtrs_post_send(struct ib_qp *qp, struct ib_send_wr *head,
110 			  struct ib_send_wr *wr, struct ib_send_wr *tail)
111 {
112 	if (head) {
113 		struct ib_send_wr *next = head;
114 
115 		while (next->next)
116 			next = next->next;
117 		next->next = wr;
118 	} else {
119 		head = wr;
120 	}
121 
122 	if (tail)
123 		wr->next = tail;
124 
125 	return ib_post_send(qp, head, NULL);
126 }
127 
128 int rtrs_iu_post_send(struct rtrs_con *con, struct rtrs_iu *iu, size_t size,
129 		       struct ib_send_wr *head)
130 {
131 	struct rtrs_path *path = con->path;
132 	struct ib_send_wr wr;
133 	struct ib_sge list;
134 
135 	if (WARN_ON(size == 0))
136 		return -EINVAL;
137 
138 	list.addr   = iu->dma_addr;
139 	list.length = size;
140 	list.lkey   = path->dev->ib_pd->local_dma_lkey;
141 
142 	wr = (struct ib_send_wr) {
143 		.wr_cqe     = &iu->cqe,
144 		.sg_list    = &list,
145 		.num_sge    = 1,
146 		.opcode     = IB_WR_SEND,
147 		.send_flags = IB_SEND_SIGNALED,
148 	};
149 
150 	return rtrs_post_send(con->qp, head, &wr, NULL);
151 }
152 EXPORT_SYMBOL_GPL(rtrs_iu_post_send);
153 
154 int rtrs_iu_post_rdma_write_imm(struct rtrs_con *con, struct rtrs_iu *iu,
155 				struct ib_sge *sge, unsigned int num_sge,
156 				u32 rkey, u64 rdma_addr, u32 imm_data,
157 				enum ib_send_flags flags,
158 				struct ib_send_wr *head,
159 				struct ib_send_wr *tail)
160 {
161 	struct ib_rdma_wr wr;
162 	int i;
163 
164 	wr = (struct ib_rdma_wr) {
165 		.wr.wr_cqe	  = &iu->cqe,
166 		.wr.sg_list	  = sge,
167 		.wr.num_sge	  = num_sge,
168 		.rkey		  = rkey,
169 		.remote_addr	  = rdma_addr,
170 		.wr.opcode	  = IB_WR_RDMA_WRITE_WITH_IMM,
171 		.wr.ex.imm_data = cpu_to_be32(imm_data),
172 		.wr.send_flags  = flags,
173 	};
174 
175 	/*
176 	 * If one of the sges has 0 size, the operation will fail with a
177 	 * length error
178 	 */
179 	for (i = 0; i < num_sge; i++)
180 		if (WARN_ONCE(sge[i].length == 0, "sg %d is zero length\n", i))
181 			return -EINVAL;
182 
183 	return rtrs_post_send(con->qp, head, &wr.wr, tail);
184 }
185 EXPORT_SYMBOL_GPL(rtrs_iu_post_rdma_write_imm);
186 
187 static int rtrs_post_rdma_write_imm_empty(struct rtrs_con *con,
188 					  struct ib_cqe *cqe,
189 					  u32 imm_data,
190 					  struct ib_send_wr *head)
191 {
192 	struct ib_rdma_wr wr;
193 	struct rtrs_path *path = con->path;
194 	enum ib_send_flags sflags;
195 
196 	atomic_dec_if_positive(&con->sq_wr_avail);
197 	sflags = (atomic_inc_return(&con->wr_cnt) % path->signal_interval) ?
198 		0 : IB_SEND_SIGNALED;
199 
200 	wr = (struct ib_rdma_wr) {
201 		.wr.wr_cqe	= cqe,
202 		.wr.send_flags	= sflags,
203 		.wr.opcode	= IB_WR_RDMA_WRITE_WITH_IMM,
204 		.wr.ex.imm_data	= cpu_to_be32(imm_data),
205 	};
206 
207 	return rtrs_post_send(con->qp, head, &wr.wr, NULL);
208 }
209 
210 static void qp_event_handler(struct ib_event *ev, void *ctx)
211 {
212 	struct rtrs_con *con = ctx;
213 
214 	switch (ev->event) {
215 	case IB_EVENT_COMM_EST:
216 		rtrs_info(con->path, "QP event %s (%d) received\n",
217 			   ib_event_msg(ev->event), ev->event);
218 		rdma_notify(con->cm_id, IB_EVENT_COMM_EST);
219 		break;
220 	default:
221 		rtrs_info(con->path, "Unhandled QP event %s (%d) received\n",
222 			   ib_event_msg(ev->event), ev->event);
223 		break;
224 	}
225 }
226 
227 static bool is_pollqueue(struct rtrs_con *con)
228 {
229 	return con->cid >= con->path->irq_con_num;
230 }
231 
232 static int create_cq(struct rtrs_con *con, int cq_vector, int nr_cqe,
233 		     enum ib_poll_context poll_ctx)
234 {
235 	struct rdma_cm_id *cm_id = con->cm_id;
236 	struct ib_cq *cq;
237 
238 	if (is_pollqueue(con))
239 		cq = ib_alloc_cq(cm_id->device, con, nr_cqe, cq_vector,
240 				 poll_ctx);
241 	else
242 		cq = ib_cq_pool_get(cm_id->device, nr_cqe, cq_vector, poll_ctx);
243 
244 	if (IS_ERR(cq)) {
245 		rtrs_err(con->path, "Creating completion queue failed, errno: %pe\n",
246 			  cq);
247 		return PTR_ERR(cq);
248 	}
249 	con->cq = cq;
250 	con->nr_cqe = nr_cqe;
251 
252 	return 0;
253 }
254 
255 static int create_qp(struct rtrs_con *con, struct ib_pd *pd,
256 		     u32 max_send_wr, u32 max_recv_wr, u32 max_sge)
257 {
258 	struct ib_qp_init_attr init_attr = {};
259 	struct rdma_cm_id *cm_id = con->cm_id;
260 	int ret;
261 
262 	init_attr.cap.max_send_wr = max_send_wr;
263 	init_attr.cap.max_recv_wr = max_recv_wr;
264 	init_attr.cap.max_recv_sge = 1;
265 	init_attr.event_handler = qp_event_handler;
266 	init_attr.qp_context = con;
267 	init_attr.cap.max_send_sge = max_sge;
268 
269 	init_attr.qp_type = IB_QPT_RC;
270 	init_attr.send_cq = con->cq;
271 	init_attr.recv_cq = con->cq;
272 	init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
273 
274 	ret = rdma_create_qp(cm_id, pd, &init_attr);
275 	if (ret) {
276 		rtrs_err(con->path, "Creating QP failed, err: %d\n", ret);
277 		return ret;
278 	}
279 	con->qp = cm_id->qp;
280 
281 	return ret;
282 }
283 
284 static void destroy_cq(struct rtrs_con *con)
285 {
286 	if (con->cq) {
287 		if (is_pollqueue(con))
288 			ib_free_cq(con->cq);
289 		else
290 			ib_cq_pool_put(con->cq, con->nr_cqe);
291 	}
292 	con->cq = NULL;
293 }
294 
295 int rtrs_cq_qp_create(struct rtrs_path *path, struct rtrs_con *con,
296 		       u32 max_send_sge, int cq_vector, int nr_cqe,
297 		       u32 max_send_wr, u32 max_recv_wr,
298 		       enum ib_poll_context poll_ctx)
299 {
300 	int err;
301 
302 	err = create_cq(con, cq_vector, nr_cqe, poll_ctx);
303 	if (err)
304 		return err;
305 
306 	err = create_qp(con, path->dev->ib_pd, max_send_wr, max_recv_wr,
307 			max_send_sge);
308 	if (err) {
309 		destroy_cq(con);
310 		return err;
311 	}
312 	con->path = path;
313 
314 	return 0;
315 }
316 EXPORT_SYMBOL_GPL(rtrs_cq_qp_create);
317 
318 void rtrs_cq_qp_destroy(struct rtrs_con *con)
319 {
320 	if (con->qp) {
321 		rdma_destroy_qp(con->cm_id);
322 		con->qp = NULL;
323 	}
324 	destroy_cq(con);
325 }
326 EXPORT_SYMBOL_GPL(rtrs_cq_qp_destroy);
327 
328 static void schedule_hb(struct rtrs_path *path)
329 {
330 	queue_delayed_work(path->hb_wq, &path->hb_dwork,
331 			   msecs_to_jiffies(path->hb_interval_ms));
332 }
333 
334 void rtrs_send_hb_ack(struct rtrs_path *path)
335 {
336 	struct rtrs_con *usr_con = path->con[0];
337 	u32 imm;
338 	int err;
339 
340 	imm = rtrs_to_imm(RTRS_HB_ACK_IMM, 0);
341 	err = rtrs_post_rdma_write_imm_empty(usr_con, path->hb_cqe, imm,
342 					     NULL);
343 	if (err) {
344 		rtrs_err(path, "send HB ACK failed, errno: %d\n", err);
345 		path->hb_err_handler(usr_con);
346 		return;
347 	}
348 }
349 EXPORT_SYMBOL_GPL(rtrs_send_hb_ack);
350 
351 static void hb_work(struct work_struct *work)
352 {
353 	struct rtrs_con *usr_con;
354 	struct rtrs_path *path;
355 	u32 imm;
356 	int err;
357 
358 	path = container_of(to_delayed_work(work), typeof(*path), hb_dwork);
359 	usr_con = path->con[0];
360 
361 	if (path->hb_missed_cnt > path->hb_missed_max) {
362 		rtrs_err(path, "HB missed max reached.\n");
363 		path->hb_err_handler(usr_con);
364 		return;
365 	}
366 	if (path->hb_missed_cnt++) {
367 		/* Reschedule work without sending hb */
368 		schedule_hb(path);
369 		return;
370 	}
371 
372 	path->hb_last_sent = ktime_get();
373 
374 	imm = rtrs_to_imm(RTRS_HB_MSG_IMM, 0);
375 	err = rtrs_post_rdma_write_imm_empty(usr_con, path->hb_cqe, imm,
376 					     NULL);
377 	if (err) {
378 		rtrs_err(path, "HB send failed, errno: %d\n", err);
379 		path->hb_err_handler(usr_con);
380 		return;
381 	}
382 
383 	schedule_hb(path);
384 }
385 
386 void rtrs_init_hb(struct rtrs_path *path, struct ib_cqe *cqe,
387 		  unsigned int interval_ms, unsigned int missed_max,
388 		  void (*err_handler)(struct rtrs_con *con),
389 		  struct workqueue_struct *wq)
390 {
391 	path->hb_cqe = cqe;
392 	path->hb_interval_ms = interval_ms;
393 	path->hb_err_handler = err_handler;
394 	path->hb_wq = wq;
395 	path->hb_missed_max = missed_max;
396 	path->hb_missed_cnt = 0;
397 	INIT_DELAYED_WORK(&path->hb_dwork, hb_work);
398 }
399 EXPORT_SYMBOL_GPL(rtrs_init_hb);
400 
401 void rtrs_start_hb(struct rtrs_path *path)
402 {
403 	schedule_hb(path);
404 }
405 EXPORT_SYMBOL_GPL(rtrs_start_hb);
406 
407 void rtrs_stop_hb(struct rtrs_path *path)
408 {
409 	cancel_delayed_work_sync(&path->hb_dwork);
410 	path->hb_missed_cnt = 0;
411 }
412 EXPORT_SYMBOL_GPL(rtrs_stop_hb);
413 
414 static int rtrs_str_gid_to_sockaddr(const char *addr, size_t len,
415 				     short port, struct sockaddr_storage *dst)
416 {
417 	struct sockaddr_ib *dst_ib = (struct sockaddr_ib *)dst;
418 	int ret;
419 
420 	/*
421 	 * We can use some of the IPv6 functions since GID is a valid
422 	 * IPv6 address format
423 	 */
424 	ret = in6_pton(addr, len, dst_ib->sib_addr.sib_raw, '\0', NULL);
425 	if (ret == 0)
426 		return -EINVAL;
427 
428 	dst_ib->sib_family = AF_IB;
429 	/*
430 	 * Use the same TCP server port number as the IB service ID
431 	 * on the IB port space range
432 	 */
433 	dst_ib->sib_sid = cpu_to_be64(RDMA_IB_IP_PS_IB | port);
434 	dst_ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL);
435 	dst_ib->sib_pkey = cpu_to_be16(0xffff);
436 
437 	return 0;
438 }
439 
440 /**
441  * rtrs_str_to_sockaddr() - Convert rtrs address string to sockaddr
442  * @addr:	String representation of an addr (IPv4, IPv6 or IB GID):
443  *              - "ip:192.168.1.1"
444  *              - "ip:fe80::200:5aee:feaa:20a2"
445  *              - "gid:fe80::200:5aee:feaa:20a2"
446  * @len:        String address length
447  * @port:	Destination port
448  * @dst:	Destination sockaddr structure
449  *
450  * Returns 0 if conversion successful. Non-zero on error.
451  */
452 static int rtrs_str_to_sockaddr(const char *addr, size_t len,
453 				u16 port, struct sockaddr_storage *dst)
454 {
455 	if (strncmp(addr, "gid:", 4) == 0) {
456 		return rtrs_str_gid_to_sockaddr(addr + 4, len - 4, port, dst);
457 	} else if (strncmp(addr, "ip:", 3) == 0) {
458 		char port_str[8];
459 		char *cpy;
460 		int err;
461 
462 		snprintf(port_str, sizeof(port_str), "%u", port);
463 		cpy = kstrndup(addr + 3, len - 3, GFP_KERNEL);
464 		err = cpy ? inet_pton_with_scope(&init_net, AF_UNSPEC,
465 						 cpy, port_str, dst) : -ENOMEM;
466 		kfree(cpy);
467 
468 		return err;
469 	}
470 	return -EPROTONOSUPPORT;
471 }
472 
473 /**
474  * sockaddr_to_str() - convert sockaddr to a string.
475  * @addr:	the sockadddr structure to be converted.
476  * @buf:	string containing socket addr.
477  * @len:	string length.
478  *
479  * The return value is the number of characters written into buf not
480  * including the trailing '\0'. If len is == 0 the function returns 0..
481  */
482 int sockaddr_to_str(const struct sockaddr *addr, char *buf, size_t len)
483 {
484 	switch (addr->sa_family) {
485 	case AF_IB:
486 		return scnprintf(buf, len, "gid:%pI6",
487 			&((struct sockaddr_ib *)addr)->sib_addr.sib_raw);
488 	case AF_INET:
489 		return scnprintf(buf, len, "ip:%pI4",
490 			&((struct sockaddr_in *)addr)->sin_addr);
491 	case AF_INET6:
492 		return scnprintf(buf, len, "ip:%pI6c",
493 			  &((struct sockaddr_in6 *)addr)->sin6_addr);
494 	}
495 	return scnprintf(buf, len, "<invalid address family>");
496 }
497 EXPORT_SYMBOL(sockaddr_to_str);
498 
499 /**
500  * rtrs_addr_to_str() - convert rtrs_addr to a string "src@dst"
501  * @addr:	the rtrs_addr structure to be converted
502  * @buf:	string containing source and destination addr of a path
503  *		separated by '@' I.e. "ip:1.1.1.1@ip:1.1.1.2"
504  *		"ip:1.1.1.1@ip:1.1.1.2".
505  * @len:	string length
506  *
507  * The return value is the number of characters written into buf not
508  * including the trailing '\0'.
509  */
510 int rtrs_addr_to_str(const struct rtrs_addr *addr, char *buf, size_t len)
511 {
512 	int cnt;
513 
514 	cnt = sockaddr_to_str((struct sockaddr *)addr->src,
515 			      buf, len);
516 	cnt += scnprintf(buf + cnt, len - cnt, "@");
517 	sockaddr_to_str((struct sockaddr *)addr->dst,
518 			buf + cnt, len - cnt);
519 	return cnt;
520 }
521 EXPORT_SYMBOL(rtrs_addr_to_str);
522 
523 /**
524  * rtrs_addr_to_sockaddr() - convert path string "src,dst" or "src@dst"
525  * to sockaddreses
526  * @str:	string containing source and destination addr of a path
527  *		separated by ',' or '@' I.e. "ip:1.1.1.1,ip:1.1.1.2" or
528  *		"ip:1.1.1.1@ip:1.1.1.2". If str contains only one address it's
529  *		considered to be destination.
530  * @len:	string length
531  * @port:	Destination port number.
532  * @addr:	will be set to the source/destination address or to NULL
533  *		if str doesn't contain any source address.
534  *
535  * Returns zero if conversion successful. Non-zero otherwise.
536  */
537 int rtrs_addr_to_sockaddr(const char *str, size_t len, u16 port,
538 			  struct rtrs_addr *addr)
539 {
540 	const char *d;
541 
542 	d = strchr(str, ',');
543 	if (!d)
544 		d = strchr(str, '@');
545 	if (d) {
546 		if (rtrs_str_to_sockaddr(str, d - str, 0, addr->src))
547 			return -EINVAL;
548 		d += 1;
549 		len -= d - str;
550 		str  = d;
551 
552 	} else {
553 		addr->src = NULL;
554 	}
555 	return rtrs_str_to_sockaddr(str, len, port, addr->dst);
556 }
557 EXPORT_SYMBOL(rtrs_addr_to_sockaddr);
558 
559 void rtrs_rdma_dev_pd_init(enum ib_pd_flags pd_flags,
560 			    struct rtrs_rdma_dev_pd *pool)
561 {
562 	INIT_LIST_HEAD(&pool->list);
563 	mutex_init(&pool->mutex);
564 	pool->pd_flags = pd_flags;
565 }
566 EXPORT_SYMBOL(rtrs_rdma_dev_pd_init);
567 
568 void rtrs_rdma_dev_pd_deinit(struct rtrs_rdma_dev_pd *pool)
569 {
570 	mutex_destroy(&pool->mutex);
571 	WARN_ON(!list_empty(&pool->list));
572 }
573 EXPORT_SYMBOL(rtrs_rdma_dev_pd_deinit);
574 
575 static void dev_free(struct kref *ref)
576 {
577 	struct rtrs_rdma_dev_pd *pool;
578 	struct rtrs_ib_dev *dev;
579 
580 	dev = container_of(ref, typeof(*dev), ref);
581 	pool = dev->pool;
582 
583 	mutex_lock(&pool->mutex);
584 	list_del(&dev->entry);
585 	mutex_unlock(&pool->mutex);
586 
587 	ib_dealloc_pd(dev->ib_pd);
588 	kfree(dev);
589 }
590 
591 int rtrs_ib_dev_put(struct rtrs_ib_dev *dev)
592 {
593 	return kref_put(&dev->ref, dev_free);
594 }
595 EXPORT_SYMBOL(rtrs_ib_dev_put);
596 
597 static int rtrs_ib_dev_get(struct rtrs_ib_dev *dev)
598 {
599 	return kref_get_unless_zero(&dev->ref);
600 }
601 
602 struct rtrs_ib_dev *
603 rtrs_ib_dev_find_or_add(struct ib_device *ib_dev,
604 			 struct rtrs_rdma_dev_pd *pool)
605 {
606 	struct rtrs_ib_dev *dev;
607 
608 	mutex_lock(&pool->mutex);
609 	list_for_each_entry(dev, &pool->list, entry) {
610 		if (dev->ib_dev->node_guid == ib_dev->node_guid &&
611 		    rtrs_ib_dev_get(dev))
612 			goto out_unlock;
613 	}
614 	mutex_unlock(&pool->mutex);
615 	dev = kzalloc(sizeof(*dev), GFP_KERNEL);
616 	if (!dev)
617 		goto out_err;
618 
619 	kref_init(&dev->ref);
620 	dev->pool = pool;
621 	dev->ib_dev = ib_dev;
622 	dev->ib_pd = ib_alloc_pd(ib_dev, pool->pd_flags);
623 	if (IS_ERR(dev->ib_pd))
624 		goto out_free_dev;
625 
626 	if (pool->ops && pool->ops->init && pool->ops->init(dev))
627 		goto out_free_pd;
628 
629 	mutex_lock(&pool->mutex);
630 	list_add(&dev->entry, &pool->list);
631 out_unlock:
632 	mutex_unlock(&pool->mutex);
633 	return dev;
634 
635 out_free_pd:
636 	ib_dealloc_pd(dev->ib_pd);
637 out_free_dev:
638 	kfree(dev);
639 out_err:
640 	return NULL;
641 }
642 EXPORT_SYMBOL(rtrs_ib_dev_find_or_add);
643