1
2
3
4
5
6
7
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
18MODULE_DESCRIPTION("RDMA Transport Core");
19MODULE_LICENSE("GPL");
20
21struct rtrs_iu *rtrs_iu_alloc(u32 queue_size, 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(queue_size, sizeof(*ius), gfp_mask);
30 if (!ius)
31 return NULL;
32 for (i = 0; i < queue_size; 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 goto err;
42
43 iu->cqe.done = done;
44 iu->size = size;
45 }
46 return ius;
47err:
48 rtrs_iu_free(ius, dma_dev, i);
49 return NULL;
50}
51EXPORT_SYMBOL_GPL(rtrs_iu_alloc);
52
53void rtrs_iu_free(struct rtrs_iu *ius, struct ib_device *ibdev, u32 queue_size)
54{
55 struct rtrs_iu *iu;
56 int i;
57
58 if (!ius)
59 return;
60
61 for (i = 0; i < queue_size; i++) {
62 iu = &ius[i];
63 ib_dma_unmap_single(ibdev, iu->dma_addr, iu->size, iu->direction);
64 kfree(iu->buf);
65 }
66 kfree(ius);
67}
68EXPORT_SYMBOL_GPL(rtrs_iu_free);
69
70int rtrs_iu_post_recv(struct rtrs_con *con, struct rtrs_iu *iu)
71{
72 struct rtrs_sess *sess = con->sess;
73 struct ib_recv_wr wr;
74 struct ib_sge list;
75
76 list.addr = iu->dma_addr;
77 list.length = iu->size;
78 list.lkey = sess->dev->ib_pd->local_dma_lkey;
79
80 if (list.length == 0) {
81 rtrs_wrn(con->sess,
82 "Posting receive work request failed, sg list is empty\n");
83 return -EINVAL;
84 }
85 wr = (struct ib_recv_wr) {
86 .wr_cqe = &iu->cqe,
87 .sg_list = &list,
88 .num_sge = 1,
89 };
90
91 return ib_post_recv(con->qp, &wr, NULL);
92}
93EXPORT_SYMBOL_GPL(rtrs_iu_post_recv);
94
95int rtrs_post_recv_empty(struct rtrs_con *con, struct ib_cqe *cqe)
96{
97 struct ib_recv_wr wr;
98
99 wr = (struct ib_recv_wr) {
100 .wr_cqe = cqe,
101 };
102
103 return ib_post_recv(con->qp, &wr, NULL);
104}
105EXPORT_SYMBOL_GPL(rtrs_post_recv_empty);
106
107static int rtrs_post_send(struct ib_qp *qp, struct ib_send_wr *head,
108 struct ib_send_wr *wr)
109{
110 if (head) {
111 struct ib_send_wr *tail = head;
112
113 while (tail->next)
114 tail = tail->next;
115 tail->next = wr;
116 } else {
117 head = wr;
118 }
119
120 return ib_post_send(qp, head, NULL);
121}
122
123int rtrs_iu_post_send(struct rtrs_con *con, struct rtrs_iu *iu, size_t size,
124 struct ib_send_wr *head)
125{
126 struct rtrs_sess *sess = con->sess;
127 struct ib_send_wr wr;
128 struct ib_sge list;
129
130 if (WARN_ON(size == 0))
131 return -EINVAL;
132
133 list.addr = iu->dma_addr;
134 list.length = size;
135 list.lkey = sess->dev->ib_pd->local_dma_lkey;
136
137 wr = (struct ib_send_wr) {
138 .wr_cqe = &iu->cqe,
139 .sg_list = &list,
140 .num_sge = 1,
141 .opcode = IB_WR_SEND,
142 .send_flags = IB_SEND_SIGNALED,
143 };
144
145 return rtrs_post_send(con->qp, head, &wr);
146}
147EXPORT_SYMBOL_GPL(rtrs_iu_post_send);
148
149int rtrs_iu_post_rdma_write_imm(struct rtrs_con *con, struct rtrs_iu *iu,
150 struct ib_sge *sge, unsigned int num_sge,
151 u32 rkey, u64 rdma_addr, u32 imm_data,
152 enum ib_send_flags flags,
153 struct ib_send_wr *head)
154{
155 struct ib_rdma_wr wr;
156 int i;
157
158 wr = (struct ib_rdma_wr) {
159 .wr.wr_cqe = &iu->cqe,
160 .wr.sg_list = sge,
161 .wr.num_sge = num_sge,
162 .rkey = rkey,
163 .remote_addr = rdma_addr,
164 .wr.opcode = IB_WR_RDMA_WRITE_WITH_IMM,
165 .wr.ex.imm_data = cpu_to_be32(imm_data),
166 .wr.send_flags = flags,
167 };
168
169
170
171
172
173 for (i = 0; i < num_sge; i++)
174 if (WARN_ON(sge[i].length == 0))
175 return -EINVAL;
176
177 return rtrs_post_send(con->qp, head, &wr.wr);
178}
179EXPORT_SYMBOL_GPL(rtrs_iu_post_rdma_write_imm);
180
181int rtrs_post_rdma_write_imm_empty(struct rtrs_con *con, struct ib_cqe *cqe,
182 u32 imm_data, enum ib_send_flags flags,
183 struct ib_send_wr *head)
184{
185 struct ib_rdma_wr wr;
186
187 wr = (struct ib_rdma_wr) {
188 .wr.wr_cqe = cqe,
189 .wr.send_flags = flags,
190 .wr.opcode = IB_WR_RDMA_WRITE_WITH_IMM,
191 .wr.ex.imm_data = cpu_to_be32(imm_data),
192 };
193
194 return rtrs_post_send(con->qp, head, &wr.wr);
195}
196EXPORT_SYMBOL_GPL(rtrs_post_rdma_write_imm_empty);
197
198static void qp_event_handler(struct ib_event *ev, void *ctx)
199{
200 struct rtrs_con *con = ctx;
201
202 switch (ev->event) {
203 case IB_EVENT_COMM_EST:
204 rtrs_info(con->sess, "QP event %s (%d) received\n",
205 ib_event_msg(ev->event), ev->event);
206 rdma_notify(con->cm_id, IB_EVENT_COMM_EST);
207 break;
208 default:
209 rtrs_info(con->sess, "Unhandled QP event %s (%d) received\n",
210 ib_event_msg(ev->event), ev->event);
211 break;
212 }
213}
214
215static int create_cq(struct rtrs_con *con, int cq_vector, u16 cq_size,
216 enum ib_poll_context poll_ctx)
217{
218 struct rdma_cm_id *cm_id = con->cm_id;
219 struct ib_cq *cq;
220
221 cq = ib_alloc_cq(cm_id->device, con, cq_size,
222 cq_vector, poll_ctx);
223 if (IS_ERR(cq)) {
224 rtrs_err(con->sess, "Creating completion queue failed, errno: %ld\n",
225 PTR_ERR(cq));
226 return PTR_ERR(cq);
227 }
228 con->cq = cq;
229
230 return 0;
231}
232
233static int create_qp(struct rtrs_con *con, struct ib_pd *pd,
234 u32 max_send_wr, u32 max_recv_wr, u32 max_sge)
235{
236 struct ib_qp_init_attr init_attr = {NULL};
237 struct rdma_cm_id *cm_id = con->cm_id;
238 int ret;
239
240 init_attr.cap.max_send_wr = max_send_wr;
241 init_attr.cap.max_recv_wr = max_recv_wr;
242 init_attr.cap.max_recv_sge = 1;
243 init_attr.event_handler = qp_event_handler;
244 init_attr.qp_context = con;
245 init_attr.cap.max_send_sge = max_sge;
246
247 init_attr.qp_type = IB_QPT_RC;
248 init_attr.send_cq = con->cq;
249 init_attr.recv_cq = con->cq;
250 init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
251
252 ret = rdma_create_qp(cm_id, pd, &init_attr);
253 if (ret) {
254 rtrs_err(con->sess, "Creating QP failed, err: %d\n", ret);
255 return ret;
256 }
257 con->qp = cm_id->qp;
258
259 return ret;
260}
261
262int rtrs_cq_qp_create(struct rtrs_sess *sess, struct rtrs_con *con,
263 u32 max_send_sge, int cq_vector, int cq_size,
264 u32 max_send_wr, u32 max_recv_wr,
265 enum ib_poll_context poll_ctx)
266{
267 int err;
268
269 err = create_cq(con, cq_vector, cq_size, poll_ctx);
270 if (err)
271 return err;
272
273 err = create_qp(con, sess->dev->ib_pd, max_send_wr, max_recv_wr,
274 max_send_sge);
275 if (err) {
276 ib_free_cq(con->cq);
277 con->cq = NULL;
278 return err;
279 }
280 con->sess = sess;
281
282 return 0;
283}
284EXPORT_SYMBOL_GPL(rtrs_cq_qp_create);
285
286void rtrs_cq_qp_destroy(struct rtrs_con *con)
287{
288 if (con->qp) {
289 rdma_destroy_qp(con->cm_id);
290 con->qp = NULL;
291 }
292 if (con->cq) {
293 ib_free_cq(con->cq);
294 con->cq = NULL;
295 }
296}
297EXPORT_SYMBOL_GPL(rtrs_cq_qp_destroy);
298
299static void schedule_hb(struct rtrs_sess *sess)
300{
301 queue_delayed_work(sess->hb_wq, &sess->hb_dwork,
302 msecs_to_jiffies(sess->hb_interval_ms));
303}
304
305void rtrs_send_hb_ack(struct rtrs_sess *sess)
306{
307 struct rtrs_con *usr_con = sess->con[0];
308 u32 imm;
309 int err;
310
311 imm = rtrs_to_imm(RTRS_HB_ACK_IMM, 0);
312 err = rtrs_post_rdma_write_imm_empty(usr_con, sess->hb_cqe, imm,
313 0, NULL);
314 if (err) {
315 sess->hb_err_handler(usr_con);
316 return;
317 }
318}
319EXPORT_SYMBOL_GPL(rtrs_send_hb_ack);
320
321static void hb_work(struct work_struct *work)
322{
323 struct rtrs_con *usr_con;
324 struct rtrs_sess *sess;
325 u32 imm;
326 int err;
327
328 sess = container_of(to_delayed_work(work), typeof(*sess), hb_dwork);
329 usr_con = sess->con[0];
330
331 if (sess->hb_missed_cnt > sess->hb_missed_max) {
332 sess->hb_err_handler(usr_con);
333 return;
334 }
335 if (sess->hb_missed_cnt++) {
336
337 schedule_hb(sess);
338 return;
339 }
340 imm = rtrs_to_imm(RTRS_HB_MSG_IMM, 0);
341 err = rtrs_post_rdma_write_imm_empty(usr_con, sess->hb_cqe, imm,
342 0, NULL);
343 if (err) {
344 sess->hb_err_handler(usr_con);
345 return;
346 }
347
348 schedule_hb(sess);
349}
350
351void rtrs_init_hb(struct rtrs_sess *sess, struct ib_cqe *cqe,
352 unsigned int interval_ms, unsigned int missed_max,
353 void (*err_handler)(struct rtrs_con *con),
354 struct workqueue_struct *wq)
355{
356 sess->hb_cqe = cqe;
357 sess->hb_interval_ms = interval_ms;
358 sess->hb_err_handler = err_handler;
359 sess->hb_wq = wq;
360 sess->hb_missed_max = missed_max;
361 sess->hb_missed_cnt = 0;
362 INIT_DELAYED_WORK(&sess->hb_dwork, hb_work);
363}
364EXPORT_SYMBOL_GPL(rtrs_init_hb);
365
366void rtrs_start_hb(struct rtrs_sess *sess)
367{
368 schedule_hb(sess);
369}
370EXPORT_SYMBOL_GPL(rtrs_start_hb);
371
372void rtrs_stop_hb(struct rtrs_sess *sess)
373{
374 cancel_delayed_work_sync(&sess->hb_dwork);
375 sess->hb_missed_cnt = 0;
376 sess->hb_missed_max = 0;
377}
378EXPORT_SYMBOL_GPL(rtrs_stop_hb);
379
380static int rtrs_str_gid_to_sockaddr(const char *addr, size_t len,
381 short port, struct sockaddr_storage *dst)
382{
383 struct sockaddr_ib *dst_ib = (struct sockaddr_ib *)dst;
384 int ret;
385
386
387
388
389
390 ret = in6_pton(addr, len, dst_ib->sib_addr.sib_raw, '\0', NULL);
391 if (ret == 0)
392 return -EINVAL;
393
394 dst_ib->sib_family = AF_IB;
395
396
397
398
399 dst_ib->sib_sid = cpu_to_be64(RDMA_IB_IP_PS_IB | port);
400 dst_ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL);
401 dst_ib->sib_pkey = cpu_to_be16(0xffff);
402
403 return 0;
404}
405
406
407
408
409
410
411
412
413
414
415
416
417
418static int rtrs_str_to_sockaddr(const char *addr, size_t len,
419 u16 port, struct sockaddr_storage *dst)
420{
421 if (strncmp(addr, "gid:", 4) == 0) {
422 return rtrs_str_gid_to_sockaddr(addr + 4, len - 4, port, dst);
423 } else if (strncmp(addr, "ip:", 3) == 0) {
424 char port_str[8];
425 char *cpy;
426 int err;
427
428 snprintf(port_str, sizeof(port_str), "%u", port);
429 cpy = kstrndup(addr + 3, len - 3, GFP_KERNEL);
430 err = cpy ? inet_pton_with_scope(&init_net, AF_UNSPEC,
431 cpy, port_str, dst) : -ENOMEM;
432 kfree(cpy);
433
434 return err;
435 }
436 return -EPROTONOSUPPORT;
437}
438
439
440
441
442
443
444
445
446
447
448int sockaddr_to_str(const struct sockaddr *addr, char *buf, size_t len)
449{
450
451 switch (addr->sa_family) {
452 case AF_IB:
453 return scnprintf(buf, len, "gid:%pI6",
454 &((struct sockaddr_ib *)addr)->sib_addr.sib_raw);
455 case AF_INET:
456 return scnprintf(buf, len, "ip:%pI4",
457 &((struct sockaddr_in *)addr)->sin_addr);
458 case AF_INET6:
459 return scnprintf(buf, len, "ip:%pI6c",
460 &((struct sockaddr_in6 *)addr)->sin6_addr);
461 }
462 return scnprintf(buf, len, "<invalid address family>");
463}
464EXPORT_SYMBOL(sockaddr_to_str);
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480int rtrs_addr_to_sockaddr(const char *str, size_t len, u16 port,
481 struct rtrs_addr *addr)
482{
483 const char *d;
484
485 d = strchr(str, ',');
486 if (!d)
487 d = strchr(str, '@');
488 if (d) {
489 if (rtrs_str_to_sockaddr(str, d - str, 0, addr->src))
490 return -EINVAL;
491 d += 1;
492 len -= d - str;
493 str = d;
494
495 } else {
496 addr->src = NULL;
497 }
498 return rtrs_str_to_sockaddr(str, len, port, addr->dst);
499}
500EXPORT_SYMBOL(rtrs_addr_to_sockaddr);
501
502void rtrs_rdma_dev_pd_init(enum ib_pd_flags pd_flags,
503 struct rtrs_rdma_dev_pd *pool)
504{
505 WARN_ON(pool->ops && (!pool->ops->alloc ^ !pool->ops->free));
506 INIT_LIST_HEAD(&pool->list);
507 mutex_init(&pool->mutex);
508 pool->pd_flags = pd_flags;
509}
510EXPORT_SYMBOL(rtrs_rdma_dev_pd_init);
511
512void rtrs_rdma_dev_pd_deinit(struct rtrs_rdma_dev_pd *pool)
513{
514 mutex_destroy(&pool->mutex);
515 WARN_ON(!list_empty(&pool->list));
516}
517EXPORT_SYMBOL(rtrs_rdma_dev_pd_deinit);
518
519static void dev_free(struct kref *ref)
520{
521 struct rtrs_rdma_dev_pd *pool;
522 struct rtrs_ib_dev *dev;
523
524 dev = container_of(ref, typeof(*dev), ref);
525 pool = dev->pool;
526
527 mutex_lock(&pool->mutex);
528 list_del(&dev->entry);
529 mutex_unlock(&pool->mutex);
530
531 if (pool->ops && pool->ops->deinit)
532 pool->ops->deinit(dev);
533
534 ib_dealloc_pd(dev->ib_pd);
535
536 if (pool->ops && pool->ops->free)
537 pool->ops->free(dev);
538 else
539 kfree(dev);
540}
541
542int rtrs_ib_dev_put(struct rtrs_ib_dev *dev)
543{
544 return kref_put(&dev->ref, dev_free);
545}
546EXPORT_SYMBOL(rtrs_ib_dev_put);
547
548static int rtrs_ib_dev_get(struct rtrs_ib_dev *dev)
549{
550 return kref_get_unless_zero(&dev->ref);
551}
552
553struct rtrs_ib_dev *
554rtrs_ib_dev_find_or_add(struct ib_device *ib_dev,
555 struct rtrs_rdma_dev_pd *pool)
556{
557 struct rtrs_ib_dev *dev;
558
559 mutex_lock(&pool->mutex);
560 list_for_each_entry(dev, &pool->list, entry) {
561 if (dev->ib_dev->node_guid == ib_dev->node_guid &&
562 rtrs_ib_dev_get(dev))
563 goto out_unlock;
564 }
565 mutex_unlock(&pool->mutex);
566 if (pool->ops && pool->ops->alloc)
567 dev = pool->ops->alloc();
568 else
569 dev = kzalloc(sizeof(*dev), GFP_KERNEL);
570 if (IS_ERR_OR_NULL(dev))
571 goto out_err;
572
573 kref_init(&dev->ref);
574 dev->pool = pool;
575 dev->ib_dev = ib_dev;
576 dev->ib_pd = ib_alloc_pd(ib_dev, pool->pd_flags);
577 if (IS_ERR(dev->ib_pd))
578 goto out_free_dev;
579
580 if (pool->ops && pool->ops->init && pool->ops->init(dev))
581 goto out_free_pd;
582
583 mutex_lock(&pool->mutex);
584 list_add(&dev->entry, &pool->list);
585out_unlock:
586 mutex_unlock(&pool->mutex);
587 return dev;
588
589out_free_pd:
590 ib_dealloc_pd(dev->ib_pd);
591out_free_dev:
592 if (pool->ops && pool->ops->free)
593 pool->ops->free(dev);
594 else
595 kfree(dev);
596out_err:
597 return NULL;
598}
599EXPORT_SYMBOL(rtrs_ib_dev_find_or_add);
600