1 // SPDX-License-Identifier: GPL-2.0-or-later
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.
11 #define pr_fmt(fmt) KBUILD_MODNAME " L" __stringify(__LINE__) ": " fmt
13 #include <linux/module.h>
17 #include <rdma/ib_cm.h>
18 #include <rdma/ib_verbs.h>
19 #include "rtrs-srv-trace.h"
21 MODULE_DESCRIPTION("RDMA Transport Server");
22 MODULE_LICENSE("GPL");
24 /* Must be power of 2, see mask from mr->page_size in ib_sg_to_pages() */
25 #define DEFAULT_MAX_CHUNK_SIZE (128 << 10)
26 #define DEFAULT_SESS_QUEUE_DEPTH 512
27 #define MAX_HDR_SIZE PAGE_SIZE
29 static struct rtrs_rdma_dev_pd dev_pd;
30 const struct class rtrs_dev_class = {
31 .name = "rtrs-server",
33 static struct rtrs_srv_ib_ctx ib_ctx;
35 static int __read_mostly max_chunk_size = DEFAULT_MAX_CHUNK_SIZE;
36 static int __read_mostly sess_queue_depth = DEFAULT_SESS_QUEUE_DEPTH;
38 static bool always_invalidate = true;
39 module_param(always_invalidate, bool, 0444);
40 MODULE_PARM_DESC(always_invalidate,
41 "Invalidate memory registration for contiguous memory regions before accessing.");
43 module_param_named(max_chunk_size, max_chunk_size, int, 0444);
44 MODULE_PARM_DESC(max_chunk_size,
45 "Max size for each IO request, when change the unit is in byte (default: "
46 __stringify(DEFAULT_MAX_CHUNK_SIZE) "KB)");
48 module_param_named(sess_queue_depth, sess_queue_depth, int, 0444);
49 MODULE_PARM_DESC(sess_queue_depth,
50 "Number of buffers for pending I/O requests to allocate per session. Maximum: "
51 __stringify(MAX_SESS_QUEUE_DEPTH) " (default: "
52 __stringify(DEFAULT_SESS_QUEUE_DEPTH) ")");
54 static cpumask_t cq_affinity_mask = { CPU_BITS_ALL };
56 static struct workqueue_struct *rtrs_wq;
58 static inline struct rtrs_srv_con *to_srv_con(struct rtrs_con *c)
60 return container_of(c, struct rtrs_srv_con, c);
63 static bool rtrs_srv_change_state(struct rtrs_srv_path *srv_path,
64 enum rtrs_srv_state new_state)
66 enum rtrs_srv_state old_state;
69 spin_lock_irq(&srv_path->state_lock);
70 old_state = srv_path->state;
72 case RTRS_SRV_CONNECTED:
73 if (old_state == RTRS_SRV_CONNECTING)
76 case RTRS_SRV_CLOSING:
77 if (old_state == RTRS_SRV_CONNECTING ||
78 old_state == RTRS_SRV_CONNECTED)
82 if (old_state == RTRS_SRV_CLOSING)
89 srv_path->state = new_state;
90 spin_unlock_irq(&srv_path->state_lock);
95 static void free_id(struct rtrs_srv_op *id)
102 static void rtrs_srv_free_ops_ids(struct rtrs_srv_path *srv_path)
104 struct rtrs_srv_sess *srv = srv_path->srv;
107 if (srv_path->ops_ids) {
108 for (i = 0; i < srv->queue_depth; i++)
109 free_id(srv_path->ops_ids[i]);
110 kfree(srv_path->ops_ids);
111 srv_path->ops_ids = NULL;
115 static void rtrs_srv_rdma_done(struct ib_cq *cq, struct ib_wc *wc);
117 static struct ib_cqe io_comp_cqe = {
118 .done = rtrs_srv_rdma_done
121 static inline void rtrs_srv_inflight_ref_release(struct percpu_ref *ref)
123 struct rtrs_srv_path *srv_path = container_of(ref,
124 struct rtrs_srv_path,
127 percpu_ref_exit(&srv_path->ids_inflight_ref);
128 complete(&srv_path->complete_done);
131 static int rtrs_srv_alloc_ops_ids(struct rtrs_srv_path *srv_path)
133 struct rtrs_srv_sess *srv = srv_path->srv;
134 struct rtrs_srv_op *id;
137 srv_path->ops_ids = kcalloc(srv->queue_depth,
138 sizeof(*srv_path->ops_ids),
140 if (!srv_path->ops_ids)
143 for (i = 0; i < srv->queue_depth; ++i) {
144 id = kzalloc(sizeof(*id), GFP_KERNEL);
148 srv_path->ops_ids[i] = id;
151 ret = percpu_ref_init(&srv_path->ids_inflight_ref,
152 rtrs_srv_inflight_ref_release, 0, GFP_KERNEL);
154 pr_err("Percpu reference init failed\n");
157 init_completion(&srv_path->complete_done);
162 rtrs_srv_free_ops_ids(srv_path);
166 static inline void rtrs_srv_get_ops_ids(struct rtrs_srv_path *srv_path)
168 percpu_ref_get(&srv_path->ids_inflight_ref);
171 static inline void rtrs_srv_put_ops_ids(struct rtrs_srv_path *srv_path)
173 percpu_ref_put(&srv_path->ids_inflight_ref);
176 static void rtrs_srv_reg_mr_done(struct ib_cq *cq, struct ib_wc *wc)
178 struct rtrs_srv_con *con = to_srv_con(wc->qp->qp_context);
179 struct rtrs_path *s = con->c.path;
180 struct rtrs_srv_path *srv_path = to_srv_path(s);
182 if (wc->status != IB_WC_SUCCESS) {
183 rtrs_err(s, "REG MR failed: %s\n",
184 ib_wc_status_msg(wc->status));
185 close_path(srv_path);
190 static struct ib_cqe local_reg_cqe = {
191 .done = rtrs_srv_reg_mr_done
194 static int rdma_write_sg(struct rtrs_srv_op *id)
196 struct rtrs_path *s = id->con->c.path;
197 struct rtrs_srv_path *srv_path = to_srv_path(s);
198 dma_addr_t dma_addr = srv_path->dma_addr[id->msg_id];
199 struct rtrs_srv_mr *srv_mr;
200 struct ib_send_wr inv_wr;
201 struct ib_rdma_wr imm_wr;
202 struct ib_rdma_wr *wr = NULL;
203 enum ib_send_flags flags;
208 struct ib_reg_wr rwr;
209 struct ib_sge *plist;
212 sg_cnt = le16_to_cpu(id->rd_msg->sg_cnt);
213 need_inval = le16_to_cpu(id->rd_msg->flags) & RTRS_MSG_NEED_INVAL_F;
221 plist->addr = dma_addr + offset;
222 plist->length = le32_to_cpu(id->rd_msg->desc[0].len);
224 /* WR will fail with length error
227 if (plist->length == 0) {
228 rtrs_err(s, "Invalid RDMA-Write sg list length 0\n");
232 plist->lkey = srv_path->s.dev->ib_pd->local_dma_lkey;
233 offset += plist->length;
235 wr->wr.sg_list = plist;
237 wr->remote_addr = le64_to_cpu(id->rd_msg->desc[0].addr);
238 wr->rkey = le32_to_cpu(id->rd_msg->desc[0].key);
242 /* Only one key is actually used */
243 WARN_ON_ONCE(rkey != wr->rkey);
245 wr->wr.opcode = IB_WR_RDMA_WRITE;
246 wr->wr.wr_cqe = &io_comp_cqe;
247 wr->wr.ex.imm_data = 0;
248 wr->wr.send_flags = 0;
250 if (need_inval && always_invalidate) {
251 wr->wr.next = &rwr.wr;
252 rwr.wr.next = &inv_wr;
253 inv_wr.next = &imm_wr.wr;
254 } else if (always_invalidate) {
255 wr->wr.next = &rwr.wr;
256 rwr.wr.next = &imm_wr.wr;
257 } else if (need_inval) {
258 wr->wr.next = &inv_wr;
259 inv_wr.next = &imm_wr.wr;
261 wr->wr.next = &imm_wr.wr;
264 * From time to time we have to post signaled sends,
265 * or send queue will fill up and only QP reset can help.
267 flags = (atomic_inc_return(&id->con->c.wr_cnt) % s->signal_interval) ?
268 0 : IB_SEND_SIGNALED;
271 inv_wr.sg_list = NULL;
273 inv_wr.opcode = IB_WR_SEND_WITH_INV;
274 inv_wr.wr_cqe = &io_comp_cqe;
275 inv_wr.send_flags = 0;
276 inv_wr.ex.invalidate_rkey = rkey;
279 imm_wr.wr.next = NULL;
280 if (always_invalidate) {
281 struct rtrs_msg_rkey_rsp *msg;
283 srv_mr = &srv_path->mrs[id->msg_id];
284 rwr.wr.opcode = IB_WR_REG_MR;
285 rwr.wr.wr_cqe = &local_reg_cqe;
288 rwr.wr.send_flags = 0;
289 rwr.key = srv_mr->mr->rkey;
290 rwr.access = (IB_ACCESS_LOCAL_WRITE |
291 IB_ACCESS_REMOTE_WRITE);
292 msg = srv_mr->iu->buf;
293 msg->buf_id = cpu_to_le16(id->msg_id);
294 msg->type = cpu_to_le16(RTRS_MSG_RKEY_RSP);
295 msg->rkey = cpu_to_le32(srv_mr->mr->rkey);
297 list.addr = srv_mr->iu->dma_addr;
298 list.length = sizeof(*msg);
299 list.lkey = srv_path->s.dev->ib_pd->local_dma_lkey;
300 imm_wr.wr.sg_list = &list;
301 imm_wr.wr.num_sge = 1;
302 imm_wr.wr.opcode = IB_WR_SEND_WITH_IMM;
303 ib_dma_sync_single_for_device(srv_path->s.dev->ib_dev,
304 srv_mr->iu->dma_addr,
305 srv_mr->iu->size, DMA_TO_DEVICE);
307 imm_wr.wr.sg_list = NULL;
308 imm_wr.wr.num_sge = 0;
309 imm_wr.wr.opcode = IB_WR_RDMA_WRITE_WITH_IMM;
311 imm_wr.wr.send_flags = flags;
312 imm_wr.wr.ex.imm_data = cpu_to_be32(rtrs_to_io_rsp_imm(id->msg_id,
315 imm_wr.wr.wr_cqe = &io_comp_cqe;
316 ib_dma_sync_single_for_device(srv_path->s.dev->ib_dev, dma_addr,
317 offset, DMA_BIDIRECTIONAL);
319 err = ib_post_send(id->con->c.qp, &id->tx_wr.wr, NULL);
322 "Posting RDMA-Write-Request to QP failed, err: %d\n",
329 * send_io_resp_imm() - respond to client with empty IMM on failed READ/WRITE
330 * requests or on successful WRITE request.
331 * @con: the connection to send back result
332 * @id: the id associated with the IO
333 * @errno: the error number of the IO.
335 * Return 0 on success, errno otherwise.
337 static int send_io_resp_imm(struct rtrs_srv_con *con, struct rtrs_srv_op *id,
340 struct rtrs_path *s = con->c.path;
341 struct rtrs_srv_path *srv_path = to_srv_path(s);
342 struct ib_send_wr inv_wr, *wr = NULL;
343 struct ib_rdma_wr imm_wr;
344 struct ib_reg_wr rwr;
345 struct rtrs_srv_mr *srv_mr;
346 bool need_inval = false;
347 enum ib_send_flags flags;
351 if (id->dir == READ) {
352 struct rtrs_msg_rdma_read *rd_msg = id->rd_msg;
355 need_inval = le16_to_cpu(rd_msg->flags) &
356 RTRS_MSG_NEED_INVAL_F;
357 sg_cnt = le16_to_cpu(rd_msg->sg_cnt);
361 inv_wr.wr_cqe = &io_comp_cqe;
362 inv_wr.sg_list = NULL;
364 inv_wr.opcode = IB_WR_SEND_WITH_INV;
365 inv_wr.send_flags = 0;
366 /* Only one key is actually used */
367 inv_wr.ex.invalidate_rkey =
368 le32_to_cpu(rd_msg->desc[0].key);
376 trace_send_io_resp_imm(id, need_inval, always_invalidate, errno);
378 if (need_inval && always_invalidate) {
380 inv_wr.next = &rwr.wr;
381 rwr.wr.next = &imm_wr.wr;
382 } else if (always_invalidate) {
384 rwr.wr.next = &imm_wr.wr;
385 } else if (need_inval) {
387 inv_wr.next = &imm_wr.wr;
392 * From time to time we have to post signalled sends,
393 * or send queue will fill up and only QP reset can help.
395 flags = (atomic_inc_return(&con->c.wr_cnt) % s->signal_interval) ?
396 0 : IB_SEND_SIGNALED;
397 imm = rtrs_to_io_rsp_imm(id->msg_id, errno, need_inval);
398 imm_wr.wr.next = NULL;
399 if (always_invalidate) {
401 struct rtrs_msg_rkey_rsp *msg;
403 srv_mr = &srv_path->mrs[id->msg_id];
404 rwr.wr.next = &imm_wr.wr;
405 rwr.wr.opcode = IB_WR_REG_MR;
406 rwr.wr.wr_cqe = &local_reg_cqe;
408 rwr.wr.send_flags = 0;
410 rwr.key = srv_mr->mr->rkey;
411 rwr.access = (IB_ACCESS_LOCAL_WRITE |
412 IB_ACCESS_REMOTE_WRITE);
413 msg = srv_mr->iu->buf;
414 msg->buf_id = cpu_to_le16(id->msg_id);
415 msg->type = cpu_to_le16(RTRS_MSG_RKEY_RSP);
416 msg->rkey = cpu_to_le32(srv_mr->mr->rkey);
418 list.addr = srv_mr->iu->dma_addr;
419 list.length = sizeof(*msg);
420 list.lkey = srv_path->s.dev->ib_pd->local_dma_lkey;
421 imm_wr.wr.sg_list = &list;
422 imm_wr.wr.num_sge = 1;
423 imm_wr.wr.opcode = IB_WR_SEND_WITH_IMM;
424 ib_dma_sync_single_for_device(srv_path->s.dev->ib_dev,
425 srv_mr->iu->dma_addr,
426 srv_mr->iu->size, DMA_TO_DEVICE);
428 imm_wr.wr.sg_list = NULL;
429 imm_wr.wr.num_sge = 0;
430 imm_wr.wr.opcode = IB_WR_RDMA_WRITE_WITH_IMM;
432 imm_wr.wr.send_flags = flags;
433 imm_wr.wr.wr_cqe = &io_comp_cqe;
435 imm_wr.wr.ex.imm_data = cpu_to_be32(imm);
437 err = ib_post_send(id->con->c.qp, wr, NULL);
439 rtrs_err_rl(s, "Posting RDMA-Reply to QP failed, err: %d\n",
445 void close_path(struct rtrs_srv_path *srv_path)
447 if (rtrs_srv_change_state(srv_path, RTRS_SRV_CLOSING))
448 queue_work(rtrs_wq, &srv_path->close_work);
449 WARN_ON(srv_path->state != RTRS_SRV_CLOSING);
452 static inline const char *rtrs_srv_state_str(enum rtrs_srv_state state)
455 case RTRS_SRV_CONNECTING:
456 return "RTRS_SRV_CONNECTING";
457 case RTRS_SRV_CONNECTED:
458 return "RTRS_SRV_CONNECTED";
459 case RTRS_SRV_CLOSING:
460 return "RTRS_SRV_CLOSING";
461 case RTRS_SRV_CLOSED:
462 return "RTRS_SRV_CLOSED";
469 * rtrs_srv_resp_rdma() - Finish an RDMA request
471 * @id: Internal RTRS operation identifier
472 * @status: Response Code sent to the other side for this operation.
473 * 0 = success, <=0 error
476 * Finish a RDMA operation. A message is sent to the client and the
477 * corresponding memory areas will be released.
479 bool rtrs_srv_resp_rdma(struct rtrs_srv_op *id, int status)
481 struct rtrs_srv_path *srv_path;
482 struct rtrs_srv_con *con;
491 srv_path = to_srv_path(s);
495 if (srv_path->state != RTRS_SRV_CONNECTED) {
497 "Sending I/O response failed, server path %s is disconnected, path state %s\n",
498 kobject_name(&srv_path->kobj),
499 rtrs_srv_state_str(srv_path->state));
502 if (always_invalidate) {
503 struct rtrs_srv_mr *mr = &srv_path->mrs[id->msg_id];
505 ib_update_fast_reg_key(mr->mr, ib_inc_rkey(mr->mr->rkey));
507 if (atomic_sub_return(1, &con->c.sq_wr_avail) < 0) {
508 rtrs_err(s, "IB send queue full: srv_path=%s cid=%d\n",
509 kobject_name(&srv_path->kobj),
511 atomic_add(1, &con->c.sq_wr_avail);
512 spin_lock(&con->rsp_wr_wait_lock);
513 list_add_tail(&id->wait_list, &con->rsp_wr_wait_list);
514 spin_unlock(&con->rsp_wr_wait_lock);
518 if (status || id->dir == WRITE || !id->rd_msg->sg_cnt)
519 err = send_io_resp_imm(con, id, status);
521 err = rdma_write_sg(id);
524 rtrs_err_rl(s, "IO response failed: %d: srv_path=%s\n", err,
525 kobject_name(&srv_path->kobj));
526 close_path(srv_path);
529 rtrs_srv_put_ops_ids(srv_path);
532 EXPORT_SYMBOL(rtrs_srv_resp_rdma);
535 * rtrs_srv_set_sess_priv() - Set private pointer in rtrs_srv.
536 * @srv: Session pointer
537 * @priv: The private pointer that is associated with the session.
539 void rtrs_srv_set_sess_priv(struct rtrs_srv_sess *srv, void *priv)
543 EXPORT_SYMBOL(rtrs_srv_set_sess_priv);
545 static void unmap_cont_bufs(struct rtrs_srv_path *srv_path)
549 for (i = 0; i < srv_path->mrs_num; i++) {
550 struct rtrs_srv_mr *srv_mr;
552 srv_mr = &srv_path->mrs[i];
553 rtrs_iu_free(srv_mr->iu, srv_path->s.dev->ib_dev, 1);
554 ib_dereg_mr(srv_mr->mr);
555 ib_dma_unmap_sg(srv_path->s.dev->ib_dev, srv_mr->sgt.sgl,
556 srv_mr->sgt.nents, DMA_BIDIRECTIONAL);
557 sg_free_table(&srv_mr->sgt);
559 kfree(srv_path->mrs);
562 static int map_cont_bufs(struct rtrs_srv_path *srv_path)
564 struct rtrs_srv_sess *srv = srv_path->srv;
565 struct rtrs_path *ss = &srv_path->s;
567 unsigned int chunk_bits;
568 int chunks_per_mr = 1;
570 struct sg_table *sgt;
573 * Here we map queue_depth chunks to MR. Firstly we have to
574 * figure out how many chunks can we map per MR.
576 if (always_invalidate) {
578 * in order to do invalidate for each chunks of memory, we needs
579 * more memory regions.
581 mrs_num = srv->queue_depth;
584 srv_path->s.dev->ib_dev->attrs.max_fast_reg_page_list_len;
585 mrs_num = DIV_ROUND_UP(srv->queue_depth, chunks_per_mr);
586 chunks_per_mr = DIV_ROUND_UP(srv->queue_depth, mrs_num);
589 srv_path->mrs = kcalloc(mrs_num, sizeof(*srv_path->mrs), GFP_KERNEL);
593 for (srv_path->mrs_num = 0; srv_path->mrs_num < mrs_num;
594 srv_path->mrs_num++) {
595 struct rtrs_srv_mr *srv_mr = &srv_path->mrs[srv_path->mrs_num];
596 struct scatterlist *s;
597 int nr, nr_sgt, chunks;
600 chunks = chunks_per_mr * srv_path->mrs_num;
601 if (!always_invalidate)
602 chunks_per_mr = min_t(int, chunks_per_mr,
603 srv->queue_depth - chunks);
605 err = sg_alloc_table(sgt, chunks_per_mr, GFP_KERNEL);
609 for_each_sg(sgt->sgl, s, chunks_per_mr, i)
610 sg_set_page(s, srv->chunks[chunks + i],
613 nr_sgt = ib_dma_map_sg(srv_path->s.dev->ib_dev, sgt->sgl,
614 sgt->nents, DMA_BIDIRECTIONAL);
619 mr = ib_alloc_mr(srv_path->s.dev->ib_pd, IB_MR_TYPE_MEM_REG,
625 nr = ib_map_mr_sg(mr, sgt->sgl, nr_sgt,
626 NULL, max_chunk_size);
628 err = nr < 0 ? nr : -EINVAL;
632 if (always_invalidate) {
633 srv_mr->iu = rtrs_iu_alloc(1,
634 sizeof(struct rtrs_msg_rkey_rsp),
635 GFP_KERNEL, srv_path->s.dev->ib_dev,
636 DMA_TO_DEVICE, rtrs_srv_rdma_done);
639 rtrs_err(ss, "rtrs_iu_alloc(), err: %d\n", err);
643 /* Eventually dma addr for each chunk can be cached */
644 for_each_sg(sgt->sgl, s, nr_sgt, i)
645 srv_path->dma_addr[chunks + i] = sg_dma_address(s);
647 ib_update_fast_reg_key(mr, ib_inc_rkey(mr->rkey));
651 chunk_bits = ilog2(srv->queue_depth - 1) + 1;
652 srv_path->mem_bits = (MAX_IMM_PAYL_BITS - chunk_bits);
659 ib_dma_unmap_sg(srv_path->s.dev->ib_dev, sgt->sgl,
660 sgt->nents, DMA_BIDIRECTIONAL);
664 unmap_cont_bufs(srv_path);
669 static void rtrs_srv_hb_err_handler(struct rtrs_con *c)
671 close_path(to_srv_path(c->path));
674 static void rtrs_srv_init_hb(struct rtrs_srv_path *srv_path)
676 rtrs_init_hb(&srv_path->s, &io_comp_cqe,
679 rtrs_srv_hb_err_handler,
683 static void rtrs_srv_start_hb(struct rtrs_srv_path *srv_path)
685 rtrs_start_hb(&srv_path->s);
688 static void rtrs_srv_stop_hb(struct rtrs_srv_path *srv_path)
690 rtrs_stop_hb(&srv_path->s);
693 static void rtrs_srv_info_rsp_done(struct ib_cq *cq, struct ib_wc *wc)
695 struct rtrs_srv_con *con = to_srv_con(wc->qp->qp_context);
696 struct rtrs_path *s = con->c.path;
697 struct rtrs_srv_path *srv_path = to_srv_path(s);
700 iu = container_of(wc->wr_cqe, struct rtrs_iu, cqe);
701 rtrs_iu_free(iu, srv_path->s.dev->ib_dev, 1);
703 if (wc->status != IB_WC_SUCCESS) {
704 rtrs_err(s, "Sess info response send failed: %s\n",
705 ib_wc_status_msg(wc->status));
706 close_path(srv_path);
709 WARN_ON(wc->opcode != IB_WC_SEND);
712 static void rtrs_srv_path_up(struct rtrs_srv_path *srv_path)
714 struct rtrs_srv_sess *srv = srv_path->srv;
715 struct rtrs_srv_ctx *ctx = srv->ctx;
718 mutex_lock(&srv->paths_ev_mutex);
719 up = ++srv->paths_up;
721 ctx->ops.link_ev(srv, RTRS_SRV_LINK_EV_CONNECTED, NULL);
722 mutex_unlock(&srv->paths_ev_mutex);
724 /* Mark session as established */
725 srv_path->established = true;
728 static void rtrs_srv_path_down(struct rtrs_srv_path *srv_path)
730 struct rtrs_srv_sess *srv = srv_path->srv;
731 struct rtrs_srv_ctx *ctx = srv->ctx;
733 if (!srv_path->established)
736 srv_path->established = false;
737 mutex_lock(&srv->paths_ev_mutex);
738 WARN_ON(!srv->paths_up);
739 if (--srv->paths_up == 0)
740 ctx->ops.link_ev(srv, RTRS_SRV_LINK_EV_DISCONNECTED, srv->priv);
741 mutex_unlock(&srv->paths_ev_mutex);
744 static bool exist_pathname(struct rtrs_srv_ctx *ctx,
745 const char *pathname, const uuid_t *path_uuid)
747 struct rtrs_srv_sess *srv;
748 struct rtrs_srv_path *srv_path;
751 mutex_lock(&ctx->srv_mutex);
752 list_for_each_entry(srv, &ctx->srv_list, ctx_list) {
753 mutex_lock(&srv->paths_mutex);
755 /* when a client with same uuid and same sessname tried to add a path */
756 if (uuid_equal(&srv->paths_uuid, path_uuid)) {
757 mutex_unlock(&srv->paths_mutex);
761 list_for_each_entry(srv_path, &srv->paths_list, s.entry) {
762 if (strlen(srv_path->s.sessname) == strlen(pathname) &&
763 !strcmp(srv_path->s.sessname, pathname)) {
768 mutex_unlock(&srv->paths_mutex);
772 mutex_unlock(&ctx->srv_mutex);
776 static int post_recv_path(struct rtrs_srv_path *srv_path);
777 static int rtrs_rdma_do_reject(struct rdma_cm_id *cm_id, int errno);
779 static int process_info_req(struct rtrs_srv_con *con,
780 struct rtrs_msg_info_req *msg)
782 struct rtrs_path *s = con->c.path;
783 struct rtrs_srv_path *srv_path = to_srv_path(s);
784 struct ib_send_wr *reg_wr = NULL;
785 struct rtrs_msg_info_rsp *rsp;
786 struct rtrs_iu *tx_iu;
787 struct ib_reg_wr *rwr;
791 err = post_recv_path(srv_path);
793 rtrs_err(s, "post_recv_path(), err: %d\n", err);
797 if (strchr(msg->pathname, '/') || strchr(msg->pathname, '.')) {
798 rtrs_err(s, "pathname cannot contain / and .\n");
802 if (exist_pathname(srv_path->srv->ctx,
803 msg->pathname, &srv_path->srv->paths_uuid)) {
804 rtrs_err(s, "pathname is duplicated: %s\n", msg->pathname);
807 strscpy(srv_path->s.sessname, msg->pathname,
808 sizeof(srv_path->s.sessname));
810 rwr = kcalloc(srv_path->mrs_num, sizeof(*rwr), GFP_KERNEL);
814 tx_sz = sizeof(*rsp);
815 tx_sz += sizeof(rsp->desc[0]) * srv_path->mrs_num;
816 tx_iu = rtrs_iu_alloc(1, tx_sz, GFP_KERNEL, srv_path->s.dev->ib_dev,
817 DMA_TO_DEVICE, rtrs_srv_info_rsp_done);
824 rsp->type = cpu_to_le16(RTRS_MSG_INFO_RSP);
825 rsp->sg_cnt = cpu_to_le16(srv_path->mrs_num);
827 for (mri = 0; mri < srv_path->mrs_num; mri++) {
828 struct ib_mr *mr = srv_path->mrs[mri].mr;
830 rsp->desc[mri].addr = cpu_to_le64(mr->iova);
831 rsp->desc[mri].key = cpu_to_le32(mr->rkey);
832 rsp->desc[mri].len = cpu_to_le32(mr->length);
835 * Fill in reg MR request and chain them *backwards*
837 rwr[mri].wr.next = mri ? &rwr[mri - 1].wr : NULL;
838 rwr[mri].wr.opcode = IB_WR_REG_MR;
839 rwr[mri].wr.wr_cqe = &local_reg_cqe;
840 rwr[mri].wr.num_sge = 0;
841 rwr[mri].wr.send_flags = 0;
843 rwr[mri].key = mr->rkey;
844 rwr[mri].access = (IB_ACCESS_LOCAL_WRITE |
845 IB_ACCESS_REMOTE_WRITE);
846 reg_wr = &rwr[mri].wr;
849 err = rtrs_srv_create_path_files(srv_path);
852 kobject_get(&srv_path->kobj);
853 get_device(&srv_path->srv->dev);
854 rtrs_srv_change_state(srv_path, RTRS_SRV_CONNECTED);
855 rtrs_srv_start_hb(srv_path);
858 * We do not account number of established connections at the current
859 * moment, we rely on the client, which should send info request when
860 * all connections are successfully established. Thus, simply notify
861 * listener with a proper event if we are the first path.
863 rtrs_srv_path_up(srv_path);
865 ib_dma_sync_single_for_device(srv_path->s.dev->ib_dev,
867 tx_iu->size, DMA_TO_DEVICE);
869 /* Send info response */
870 err = rtrs_iu_post_send(&con->c, tx_iu, tx_sz, reg_wr);
872 rtrs_err(s, "rtrs_iu_post_send(), err: %d\n", err);
874 rtrs_iu_free(tx_iu, srv_path->s.dev->ib_dev, 1);
882 static void rtrs_srv_info_req_done(struct ib_cq *cq, struct ib_wc *wc)
884 struct rtrs_srv_con *con = to_srv_con(wc->qp->qp_context);
885 struct rtrs_path *s = con->c.path;
886 struct rtrs_srv_path *srv_path = to_srv_path(s);
887 struct rtrs_msg_info_req *msg;
893 iu = container_of(wc->wr_cqe, struct rtrs_iu, cqe);
894 if (wc->status != IB_WC_SUCCESS) {
895 rtrs_err(s, "Sess info request receive failed: %s\n",
896 ib_wc_status_msg(wc->status));
899 WARN_ON(wc->opcode != IB_WC_RECV);
901 if (wc->byte_len < sizeof(*msg)) {
902 rtrs_err(s, "Sess info request is malformed: size %d\n",
906 ib_dma_sync_single_for_cpu(srv_path->s.dev->ib_dev, iu->dma_addr,
907 iu->size, DMA_FROM_DEVICE);
909 if (le16_to_cpu(msg->type) != RTRS_MSG_INFO_REQ) {
910 rtrs_err(s, "Sess info request is malformed: type %d\n",
911 le16_to_cpu(msg->type));
914 err = process_info_req(con, msg);
919 rtrs_iu_free(iu, srv_path->s.dev->ib_dev, 1);
922 close_path(srv_path);
926 static int post_recv_info_req(struct rtrs_srv_con *con)
928 struct rtrs_path *s = con->c.path;
929 struct rtrs_srv_path *srv_path = to_srv_path(s);
930 struct rtrs_iu *rx_iu;
933 rx_iu = rtrs_iu_alloc(1, sizeof(struct rtrs_msg_info_req),
934 GFP_KERNEL, srv_path->s.dev->ib_dev,
935 DMA_FROM_DEVICE, rtrs_srv_info_req_done);
938 /* Prepare for getting info response */
939 err = rtrs_iu_post_recv(&con->c, rx_iu);
941 rtrs_err(s, "rtrs_iu_post_recv(), err: %d\n", err);
942 rtrs_iu_free(rx_iu, srv_path->s.dev->ib_dev, 1);
949 static int post_recv_io(struct rtrs_srv_con *con, size_t q_size)
953 for (i = 0; i < q_size; i++) {
954 err = rtrs_post_recv_empty(&con->c, &io_comp_cqe);
962 static int post_recv_path(struct rtrs_srv_path *srv_path)
964 struct rtrs_srv_sess *srv = srv_path->srv;
965 struct rtrs_path *s = &srv_path->s;
969 for (cid = 0; cid < srv_path->s.con_num; cid++) {
971 q_size = SERVICE_CON_QUEUE_DEPTH;
973 q_size = srv->queue_depth;
975 err = post_recv_io(to_srv_con(srv_path->s.con[cid]), q_size);
977 rtrs_err(s, "post_recv_io(), err: %d\n", err);
985 static void process_read(struct rtrs_srv_con *con,
986 struct rtrs_msg_rdma_read *msg,
989 struct rtrs_path *s = con->c.path;
990 struct rtrs_srv_path *srv_path = to_srv_path(s);
991 struct rtrs_srv_sess *srv = srv_path->srv;
992 struct rtrs_srv_ctx *ctx = srv->ctx;
993 struct rtrs_srv_op *id;
995 size_t usr_len, data_len;
999 if (srv_path->state != RTRS_SRV_CONNECTED) {
1001 "Processing read request failed, session is disconnected, sess state %s\n",
1002 rtrs_srv_state_str(srv_path->state));
1005 if (msg->sg_cnt != 1 && msg->sg_cnt != 0) {
1007 "Processing read request failed, invalid message\n");
1010 rtrs_srv_get_ops_ids(srv_path);
1011 rtrs_srv_update_rdma_stats(srv_path->stats, off, READ);
1012 id = srv_path->ops_ids[buf_id];
1015 id->msg_id = buf_id;
1017 usr_len = le16_to_cpu(msg->usr_len);
1018 data_len = off - usr_len;
1019 data = page_address(srv->chunks[buf_id]);
1020 ret = ctx->ops.rdma_ev(srv->priv, id, data, data_len,
1021 data + data_len, usr_len);
1025 "Processing read request failed, user module cb reported for msg_id %d, err: %d\n",
1033 ret = send_io_resp_imm(con, id, ret);
1036 "Sending err msg for failed RDMA-Write-Req failed, msg_id %d, err: %d\n",
1038 close_path(srv_path);
1040 rtrs_srv_put_ops_ids(srv_path);
1043 static void process_write(struct rtrs_srv_con *con,
1044 struct rtrs_msg_rdma_write *req,
1045 u32 buf_id, u32 off)
1047 struct rtrs_path *s = con->c.path;
1048 struct rtrs_srv_path *srv_path = to_srv_path(s);
1049 struct rtrs_srv_sess *srv = srv_path->srv;
1050 struct rtrs_srv_ctx *ctx = srv->ctx;
1051 struct rtrs_srv_op *id;
1053 size_t data_len, usr_len;
1057 if (srv_path->state != RTRS_SRV_CONNECTED) {
1059 "Processing write request failed, session is disconnected, sess state %s\n",
1060 rtrs_srv_state_str(srv_path->state));
1063 rtrs_srv_get_ops_ids(srv_path);
1064 rtrs_srv_update_rdma_stats(srv_path->stats, off, WRITE);
1065 id = srv_path->ops_ids[buf_id];
1068 id->msg_id = buf_id;
1070 usr_len = le16_to_cpu(req->usr_len);
1071 data_len = off - usr_len;
1072 data = page_address(srv->chunks[buf_id]);
1073 ret = ctx->ops.rdma_ev(srv->priv, id, data, data_len,
1074 data + data_len, usr_len);
1077 "Processing write request failed, user module callback reports err: %d\n",
1085 ret = send_io_resp_imm(con, id, ret);
1088 "Processing write request failed, sending I/O response failed, msg_id %d, err: %d\n",
1090 close_path(srv_path);
1092 rtrs_srv_put_ops_ids(srv_path);
1095 static void process_io_req(struct rtrs_srv_con *con, void *msg,
1098 struct rtrs_path *s = con->c.path;
1099 struct rtrs_srv_path *srv_path = to_srv_path(s);
1100 struct rtrs_msg_rdma_hdr *hdr;
1103 ib_dma_sync_single_for_cpu(srv_path->s.dev->ib_dev,
1104 srv_path->dma_addr[id],
1105 max_chunk_size, DMA_BIDIRECTIONAL);
1107 type = le16_to_cpu(hdr->type);
1110 case RTRS_MSG_WRITE:
1111 process_write(con, msg, id, off);
1114 process_read(con, msg, id, off);
1118 "Processing I/O request failed, unknown message type received: 0x%02x\n",
1126 close_path(srv_path);
1129 static void rtrs_srv_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
1131 struct rtrs_srv_mr *mr =
1132 container_of(wc->wr_cqe, typeof(*mr), inv_cqe);
1133 struct rtrs_srv_con *con = to_srv_con(wc->qp->qp_context);
1134 struct rtrs_path *s = con->c.path;
1135 struct rtrs_srv_path *srv_path = to_srv_path(s);
1136 struct rtrs_srv_sess *srv = srv_path->srv;
1140 if (wc->status != IB_WC_SUCCESS) {
1141 rtrs_err(s, "Failed IB_WR_LOCAL_INV: %s\n",
1142 ib_wc_status_msg(wc->status));
1143 close_path(srv_path);
1145 msg_id = mr->msg_id;
1147 data = page_address(srv->chunks[msg_id]) + off;
1148 process_io_req(con, data, msg_id, off);
1151 static int rtrs_srv_inv_rkey(struct rtrs_srv_con *con,
1152 struct rtrs_srv_mr *mr)
1154 struct ib_send_wr wr = {
1155 .opcode = IB_WR_LOCAL_INV,
1156 .wr_cqe = &mr->inv_cqe,
1157 .send_flags = IB_SEND_SIGNALED,
1158 .ex.invalidate_rkey = mr->mr->rkey,
1160 mr->inv_cqe.done = rtrs_srv_inv_rkey_done;
1162 return ib_post_send(con->c.qp, &wr, NULL);
1165 static void rtrs_rdma_process_wr_wait_list(struct rtrs_srv_con *con)
1167 spin_lock(&con->rsp_wr_wait_lock);
1168 while (!list_empty(&con->rsp_wr_wait_list)) {
1169 struct rtrs_srv_op *id;
1172 id = list_entry(con->rsp_wr_wait_list.next,
1173 struct rtrs_srv_op, wait_list);
1174 list_del(&id->wait_list);
1176 spin_unlock(&con->rsp_wr_wait_lock);
1177 ret = rtrs_srv_resp_rdma(id, id->status);
1178 spin_lock(&con->rsp_wr_wait_lock);
1181 list_add(&id->wait_list, &con->rsp_wr_wait_list);
1185 spin_unlock(&con->rsp_wr_wait_lock);
1188 static void rtrs_srv_rdma_done(struct ib_cq *cq, struct ib_wc *wc)
1190 struct rtrs_srv_con *con = to_srv_con(wc->qp->qp_context);
1191 struct rtrs_path *s = con->c.path;
1192 struct rtrs_srv_path *srv_path = to_srv_path(s);
1193 struct rtrs_srv_sess *srv = srv_path->srv;
1194 u32 imm_type, imm_payload;
1197 if (wc->status != IB_WC_SUCCESS) {
1198 if (wc->status != IB_WC_WR_FLUSH_ERR) {
1200 "%s (wr_cqe: %p, type: %d, vendor_err: 0x%x, len: %u)\n",
1201 ib_wc_status_msg(wc->status), wc->wr_cqe,
1202 wc->opcode, wc->vendor_err, wc->byte_len);
1203 close_path(srv_path);
1208 switch (wc->opcode) {
1209 case IB_WC_RECV_RDMA_WITH_IMM:
1211 * post_recv() RDMA write completions of IO reqs (read/write)
1214 if (WARN_ON(wc->wr_cqe != &io_comp_cqe))
1216 err = rtrs_post_recv_empty(&con->c, &io_comp_cqe);
1218 rtrs_err(s, "rtrs_post_recv(), err: %d\n", err);
1219 close_path(srv_path);
1222 rtrs_from_imm(be32_to_cpu(wc->ex.imm_data),
1223 &imm_type, &imm_payload);
1224 if (imm_type == RTRS_IO_REQ_IMM) {
1228 msg_id = imm_payload >> srv_path->mem_bits;
1229 off = imm_payload & ((1 << srv_path->mem_bits) - 1);
1230 if (msg_id >= srv->queue_depth || off >= max_chunk_size) {
1231 rtrs_err(s, "Wrong msg_id %u, off %u\n",
1233 close_path(srv_path);
1236 if (always_invalidate) {
1237 struct rtrs_srv_mr *mr = &srv_path->mrs[msg_id];
1240 mr->msg_id = msg_id;
1241 err = rtrs_srv_inv_rkey(con, mr);
1243 rtrs_err(s, "rtrs_post_recv(), err: %d\n",
1245 close_path(srv_path);
1249 data = page_address(srv->chunks[msg_id]) + off;
1250 process_io_req(con, data, msg_id, off);
1252 } else if (imm_type == RTRS_HB_MSG_IMM) {
1253 WARN_ON(con->c.cid);
1254 rtrs_send_hb_ack(&srv_path->s);
1255 } else if (imm_type == RTRS_HB_ACK_IMM) {
1256 WARN_ON(con->c.cid);
1257 srv_path->s.hb_missed_cnt = 0;
1259 rtrs_wrn(s, "Unknown IMM type %u\n", imm_type);
1262 case IB_WC_RDMA_WRITE:
1265 * post_send() RDMA write completions of IO reqs (read/write)
1268 atomic_add(s->signal_interval, &con->c.sq_wr_avail);
1270 if (!list_empty_careful(&con->rsp_wr_wait_list))
1271 rtrs_rdma_process_wr_wait_list(con);
1275 rtrs_wrn(s, "Unexpected WC type: %d\n", wc->opcode);
1281 * rtrs_srv_get_path_name() - Get rtrs_srv peer hostname.
1283 * @pathname: Pathname buffer
1284 * @len: Length of sessname buffer
1286 int rtrs_srv_get_path_name(struct rtrs_srv_sess *srv, char *pathname,
1289 struct rtrs_srv_path *srv_path;
1290 int err = -ENOTCONN;
1292 mutex_lock(&srv->paths_mutex);
1293 list_for_each_entry(srv_path, &srv->paths_list, s.entry) {
1294 if (srv_path->state != RTRS_SRV_CONNECTED)
1296 strscpy(pathname, srv_path->s.sessname,
1297 min_t(size_t, sizeof(srv_path->s.sessname), len));
1301 mutex_unlock(&srv->paths_mutex);
1305 EXPORT_SYMBOL(rtrs_srv_get_path_name);
1308 * rtrs_srv_get_queue_depth() - Get rtrs_srv qdepth.
1311 int rtrs_srv_get_queue_depth(struct rtrs_srv_sess *srv)
1313 return srv->queue_depth;
1315 EXPORT_SYMBOL(rtrs_srv_get_queue_depth);
1317 static int find_next_bit_ring(struct rtrs_srv_path *srv_path)
1319 struct ib_device *ib_dev = srv_path->s.dev->ib_dev;
1322 v = cpumask_next(srv_path->cur_cq_vector, &cq_affinity_mask);
1323 if (v >= nr_cpu_ids || v >= ib_dev->num_comp_vectors)
1324 v = cpumask_first(&cq_affinity_mask);
1328 static int rtrs_srv_get_next_cq_vector(struct rtrs_srv_path *srv_path)
1330 srv_path->cur_cq_vector = find_next_bit_ring(srv_path);
1332 return srv_path->cur_cq_vector;
1335 static void rtrs_srv_dev_release(struct device *dev)
1337 struct rtrs_srv_sess *srv = container_of(dev, struct rtrs_srv_sess,
1343 static void free_srv(struct rtrs_srv_sess *srv)
1347 WARN_ON(refcount_read(&srv->refcount));
1348 for (i = 0; i < srv->queue_depth; i++)
1349 __free_pages(srv->chunks[i], get_order(max_chunk_size));
1351 mutex_destroy(&srv->paths_mutex);
1352 mutex_destroy(&srv->paths_ev_mutex);
1353 /* last put to release the srv structure */
1354 put_device(&srv->dev);
1357 static struct rtrs_srv_sess *get_or_create_srv(struct rtrs_srv_ctx *ctx,
1358 const uuid_t *paths_uuid,
1361 struct rtrs_srv_sess *srv;
1364 mutex_lock(&ctx->srv_mutex);
1365 list_for_each_entry(srv, &ctx->srv_list, ctx_list) {
1366 if (uuid_equal(&srv->paths_uuid, paths_uuid) &&
1367 refcount_inc_not_zero(&srv->refcount)) {
1368 mutex_unlock(&ctx->srv_mutex);
1372 mutex_unlock(&ctx->srv_mutex);
1374 * If this request is not the first connection request from the
1375 * client for this session then fail and return error.
1378 pr_err_ratelimited("Error: Not the first connection request for this session\n");
1379 return ERR_PTR(-ENXIO);
1382 /* need to allocate a new srv */
1383 srv = kzalloc(sizeof(*srv), GFP_KERNEL);
1385 return ERR_PTR(-ENOMEM);
1387 INIT_LIST_HEAD(&srv->paths_list);
1388 mutex_init(&srv->paths_mutex);
1389 mutex_init(&srv->paths_ev_mutex);
1390 uuid_copy(&srv->paths_uuid, paths_uuid);
1391 srv->queue_depth = sess_queue_depth;
1393 device_initialize(&srv->dev);
1394 srv->dev.release = rtrs_srv_dev_release;
1396 srv->chunks = kcalloc(srv->queue_depth, sizeof(*srv->chunks),
1401 for (i = 0; i < srv->queue_depth; i++) {
1402 srv->chunks[i] = alloc_pages(GFP_KERNEL,
1403 get_order(max_chunk_size));
1404 if (!srv->chunks[i])
1405 goto err_free_chunks;
1407 refcount_set(&srv->refcount, 1);
1408 mutex_lock(&ctx->srv_mutex);
1409 list_add(&srv->ctx_list, &ctx->srv_list);
1410 mutex_unlock(&ctx->srv_mutex);
1416 __free_pages(srv->chunks[i], get_order(max_chunk_size));
1421 return ERR_PTR(-ENOMEM);
1424 static void put_srv(struct rtrs_srv_sess *srv)
1426 if (refcount_dec_and_test(&srv->refcount)) {
1427 struct rtrs_srv_ctx *ctx = srv->ctx;
1429 WARN_ON(srv->dev.kobj.state_in_sysfs);
1431 mutex_lock(&ctx->srv_mutex);
1432 list_del(&srv->ctx_list);
1433 mutex_unlock(&ctx->srv_mutex);
1438 static void __add_path_to_srv(struct rtrs_srv_sess *srv,
1439 struct rtrs_srv_path *srv_path)
1441 list_add_tail(&srv_path->s.entry, &srv->paths_list);
1443 WARN_ON(srv->paths_num >= MAX_PATHS_NUM);
1446 static void del_path_from_srv(struct rtrs_srv_path *srv_path)
1448 struct rtrs_srv_sess *srv = srv_path->srv;
1453 mutex_lock(&srv->paths_mutex);
1454 list_del(&srv_path->s.entry);
1455 WARN_ON(!srv->paths_num);
1457 mutex_unlock(&srv->paths_mutex);
1460 /* return true if addresses are the same, error other wise */
1461 static int sockaddr_cmp(const struct sockaddr *a, const struct sockaddr *b)
1463 switch (a->sa_family) {
1465 return memcmp(&((struct sockaddr_ib *)a)->sib_addr,
1466 &((struct sockaddr_ib *)b)->sib_addr,
1467 sizeof(struct ib_addr)) &&
1468 (b->sa_family == AF_IB);
1470 return memcmp(&((struct sockaddr_in *)a)->sin_addr,
1471 &((struct sockaddr_in *)b)->sin_addr,
1472 sizeof(struct in_addr)) &&
1473 (b->sa_family == AF_INET);
1475 return memcmp(&((struct sockaddr_in6 *)a)->sin6_addr,
1476 &((struct sockaddr_in6 *)b)->sin6_addr,
1477 sizeof(struct in6_addr)) &&
1478 (b->sa_family == AF_INET6);
1484 static bool __is_path_w_addr_exists(struct rtrs_srv_sess *srv,
1485 struct rdma_addr *addr)
1487 struct rtrs_srv_path *srv_path;
1489 list_for_each_entry(srv_path, &srv->paths_list, s.entry)
1490 if (!sockaddr_cmp((struct sockaddr *)&srv_path->s.dst_addr,
1491 (struct sockaddr *)&addr->dst_addr) &&
1492 !sockaddr_cmp((struct sockaddr *)&srv_path->s.src_addr,
1493 (struct sockaddr *)&addr->src_addr))
1499 static void free_path(struct rtrs_srv_path *srv_path)
1501 if (srv_path->kobj.state_in_sysfs) {
1502 kobject_del(&srv_path->kobj);
1503 kobject_put(&srv_path->kobj);
1505 free_percpu(srv_path->stats->rdma_stats);
1506 kfree(srv_path->stats);
1511 static void rtrs_srv_close_work(struct work_struct *work)
1513 struct rtrs_srv_path *srv_path;
1514 struct rtrs_srv_con *con;
1517 srv_path = container_of(work, typeof(*srv_path), close_work);
1519 rtrs_srv_destroy_path_files(srv_path);
1520 rtrs_srv_stop_hb(srv_path);
1522 for (i = 0; i < srv_path->s.con_num; i++) {
1523 if (!srv_path->s.con[i])
1525 con = to_srv_con(srv_path->s.con[i]);
1526 rdma_disconnect(con->c.cm_id);
1527 ib_drain_qp(con->c.qp);
1531 * Degrade ref count to the usual model with a single shared
1534 percpu_ref_kill(&srv_path->ids_inflight_ref);
1536 /* Wait for all completion */
1537 wait_for_completion(&srv_path->complete_done);
1539 /* Notify upper layer if we are the last path */
1540 rtrs_srv_path_down(srv_path);
1542 unmap_cont_bufs(srv_path);
1543 rtrs_srv_free_ops_ids(srv_path);
1545 for (i = 0; i < srv_path->s.con_num; i++) {
1546 if (!srv_path->s.con[i])
1548 con = to_srv_con(srv_path->s.con[i]);
1549 rtrs_cq_qp_destroy(&con->c);
1550 rdma_destroy_id(con->c.cm_id);
1553 rtrs_ib_dev_put(srv_path->s.dev);
1555 del_path_from_srv(srv_path);
1556 put_srv(srv_path->srv);
1557 srv_path->srv = NULL;
1558 rtrs_srv_change_state(srv_path, RTRS_SRV_CLOSED);
1560 kfree(srv_path->dma_addr);
1561 kfree(srv_path->s.con);
1562 free_path(srv_path);
1565 static int rtrs_rdma_do_accept(struct rtrs_srv_path *srv_path,
1566 struct rdma_cm_id *cm_id)
1568 struct rtrs_srv_sess *srv = srv_path->srv;
1569 struct rtrs_msg_conn_rsp msg;
1570 struct rdma_conn_param param;
1573 param = (struct rdma_conn_param) {
1574 .rnr_retry_count = 7,
1575 .private_data = &msg,
1576 .private_data_len = sizeof(msg),
1579 msg = (struct rtrs_msg_conn_rsp) {
1580 .magic = cpu_to_le16(RTRS_MAGIC),
1581 .version = cpu_to_le16(RTRS_PROTO_VER),
1582 .queue_depth = cpu_to_le16(srv->queue_depth),
1583 .max_io_size = cpu_to_le32(max_chunk_size - MAX_HDR_SIZE),
1584 .max_hdr_size = cpu_to_le32(MAX_HDR_SIZE),
1587 if (always_invalidate)
1588 msg.flags = cpu_to_le32(RTRS_MSG_NEW_RKEY_F);
1590 err = rdma_accept(cm_id, ¶m);
1592 pr_err("rdma_accept(), err: %d\n", err);
1597 static int rtrs_rdma_do_reject(struct rdma_cm_id *cm_id, int errno)
1599 struct rtrs_msg_conn_rsp msg;
1602 msg = (struct rtrs_msg_conn_rsp) {
1603 .magic = cpu_to_le16(RTRS_MAGIC),
1604 .version = cpu_to_le16(RTRS_PROTO_VER),
1605 .errno = cpu_to_le16(errno),
1608 err = rdma_reject(cm_id, &msg, sizeof(msg), IB_CM_REJ_CONSUMER_DEFINED);
1610 pr_err("rdma_reject(), err: %d\n", err);
1612 /* Bounce errno back */
1616 static struct rtrs_srv_path *
1617 __find_path(struct rtrs_srv_sess *srv, const uuid_t *sess_uuid)
1619 struct rtrs_srv_path *srv_path;
1621 list_for_each_entry(srv_path, &srv->paths_list, s.entry) {
1622 if (uuid_equal(&srv_path->s.uuid, sess_uuid))
1629 static int create_con(struct rtrs_srv_path *srv_path,
1630 struct rdma_cm_id *cm_id,
1633 struct rtrs_srv_sess *srv = srv_path->srv;
1634 struct rtrs_path *s = &srv_path->s;
1635 struct rtrs_srv_con *con;
1637 u32 cq_num, max_send_wr, max_recv_wr, wr_limit;
1640 con = kzalloc(sizeof(*con), GFP_KERNEL);
1646 spin_lock_init(&con->rsp_wr_wait_lock);
1647 INIT_LIST_HEAD(&con->rsp_wr_wait_list);
1648 con->c.cm_id = cm_id;
1649 con->c.path = &srv_path->s;
1651 atomic_set(&con->c.wr_cnt, 1);
1652 wr_limit = srv_path->s.dev->ib_dev->attrs.max_qp_wr;
1654 if (con->c.cid == 0) {
1656 * All receive and all send (each requiring invalidate)
1657 * + 2 for drain and heartbeat
1659 max_send_wr = min_t(int, wr_limit,
1660 SERVICE_CON_QUEUE_DEPTH * 2 + 2);
1661 max_recv_wr = max_send_wr;
1662 s->signal_interval = min_not_zero(srv->queue_depth,
1663 (size_t)SERVICE_CON_QUEUE_DEPTH);
1665 /* when always_invlaidate enalbed, we need linv+rinv+mr+imm */
1666 if (always_invalidate)
1668 min_t(int, wr_limit,
1669 srv->queue_depth * (1 + 4) + 1);
1672 min_t(int, wr_limit,
1673 srv->queue_depth * (1 + 2) + 1);
1675 max_recv_wr = srv->queue_depth + 1;
1677 cq_num = max_send_wr + max_recv_wr;
1678 atomic_set(&con->c.sq_wr_avail, max_send_wr);
1679 cq_vector = rtrs_srv_get_next_cq_vector(srv_path);
1681 /* TODO: SOFTIRQ can be faster, but be careful with softirq context */
1682 err = rtrs_cq_qp_create(&srv_path->s, &con->c, 1, cq_vector, cq_num,
1683 max_send_wr, max_recv_wr,
1686 rtrs_err(s, "rtrs_cq_qp_create(), err: %d\n", err);
1689 if (con->c.cid == 0) {
1690 err = post_recv_info_req(con);
1694 WARN_ON(srv_path->s.con[cid]);
1695 srv_path->s.con[cid] = &con->c;
1698 * Change context from server to current connection. The other
1699 * way is to use cm_id->qp->qp_context, which does not work on OFED.
1701 cm_id->context = &con->c;
1706 rtrs_cq_qp_destroy(&con->c);
1714 static struct rtrs_srv_path *__alloc_path(struct rtrs_srv_sess *srv,
1715 struct rdma_cm_id *cm_id,
1716 unsigned int con_num,
1717 unsigned int recon_cnt,
1720 struct rtrs_srv_path *srv_path;
1723 struct rtrs_addr path;
1725 if (srv->paths_num >= MAX_PATHS_NUM) {
1729 if (__is_path_w_addr_exists(srv, &cm_id->route.addr)) {
1731 pr_err("Path with same addr exists\n");
1734 srv_path = kzalloc(sizeof(*srv_path), GFP_KERNEL);
1738 srv_path->stats = kzalloc(sizeof(*srv_path->stats), GFP_KERNEL);
1739 if (!srv_path->stats)
1742 srv_path->stats->rdma_stats = alloc_percpu(struct rtrs_srv_stats_rdma_stats);
1743 if (!srv_path->stats->rdma_stats)
1744 goto err_free_stats;
1746 srv_path->stats->srv_path = srv_path;
1748 srv_path->dma_addr = kcalloc(srv->queue_depth,
1749 sizeof(*srv_path->dma_addr),
1751 if (!srv_path->dma_addr)
1752 goto err_free_percpu;
1754 srv_path->s.con = kcalloc(con_num, sizeof(*srv_path->s.con),
1756 if (!srv_path->s.con)
1757 goto err_free_dma_addr;
1759 srv_path->state = RTRS_SRV_CONNECTING;
1760 srv_path->srv = srv;
1761 srv_path->cur_cq_vector = -1;
1762 srv_path->s.dst_addr = cm_id->route.addr.dst_addr;
1763 srv_path->s.src_addr = cm_id->route.addr.src_addr;
1765 /* temporary until receiving session-name from client */
1766 path.src = &srv_path->s.src_addr;
1767 path.dst = &srv_path->s.dst_addr;
1768 rtrs_addr_to_str(&path, str, sizeof(str));
1769 strscpy(srv_path->s.sessname, str, sizeof(srv_path->s.sessname));
1771 srv_path->s.con_num = con_num;
1772 srv_path->s.irq_con_num = con_num;
1773 srv_path->s.recon_cnt = recon_cnt;
1774 uuid_copy(&srv_path->s.uuid, uuid);
1775 spin_lock_init(&srv_path->state_lock);
1776 INIT_WORK(&srv_path->close_work, rtrs_srv_close_work);
1777 rtrs_srv_init_hb(srv_path);
1779 srv_path->s.dev = rtrs_ib_dev_find_or_add(cm_id->device, &dev_pd);
1780 if (!srv_path->s.dev) {
1784 err = map_cont_bufs(srv_path);
1788 err = rtrs_srv_alloc_ops_ids(srv_path);
1790 goto err_unmap_bufs;
1792 __add_path_to_srv(srv, srv_path);
1797 unmap_cont_bufs(srv_path);
1799 rtrs_ib_dev_put(srv_path->s.dev);
1801 kfree(srv_path->s.con);
1803 kfree(srv_path->dma_addr);
1805 free_percpu(srv_path->stats->rdma_stats);
1807 kfree(srv_path->stats);
1811 return ERR_PTR(err);
1814 static int rtrs_rdma_connect(struct rdma_cm_id *cm_id,
1815 const struct rtrs_msg_conn_req *msg,
1818 struct rtrs_srv_ctx *ctx = cm_id->context;
1819 struct rtrs_srv_path *srv_path;
1820 struct rtrs_srv_sess *srv;
1822 u16 version, con_num, cid;
1824 int err = -ECONNRESET;
1826 if (len < sizeof(*msg)) {
1827 pr_err("Invalid RTRS connection request\n");
1830 if (le16_to_cpu(msg->magic) != RTRS_MAGIC) {
1831 pr_err("Invalid RTRS magic\n");
1834 version = le16_to_cpu(msg->version);
1835 if (version >> 8 != RTRS_PROTO_VER_MAJOR) {
1836 pr_err("Unsupported major RTRS version: %d, expected %d\n",
1837 version >> 8, RTRS_PROTO_VER_MAJOR);
1840 con_num = le16_to_cpu(msg->cid_num);
1841 if (con_num > 4096) {
1843 pr_err("Too many connections requested: %d\n", con_num);
1846 cid = le16_to_cpu(msg->cid);
1847 if (cid >= con_num) {
1849 pr_err("Incorrect cid: %d >= %d\n", cid, con_num);
1852 recon_cnt = le16_to_cpu(msg->recon_cnt);
1853 srv = get_or_create_srv(ctx, &msg->paths_uuid, msg->first_conn);
1856 pr_err("get_or_create_srv(), error %d\n", err);
1859 mutex_lock(&srv->paths_mutex);
1860 srv_path = __find_path(srv, &msg->sess_uuid);
1862 struct rtrs_path *s = &srv_path->s;
1864 /* Session already holds a reference */
1867 if (srv_path->state != RTRS_SRV_CONNECTING) {
1868 rtrs_err(s, "Session in wrong state: %s\n",
1869 rtrs_srv_state_str(srv_path->state));
1870 mutex_unlock(&srv->paths_mutex);
1876 if (con_num != s->con_num || cid >= s->con_num) {
1877 rtrs_err(s, "Incorrect request: %d, %d\n",
1879 mutex_unlock(&srv->paths_mutex);
1883 rtrs_err(s, "Connection already exists: %d\n",
1885 mutex_unlock(&srv->paths_mutex);
1889 srv_path = __alloc_path(srv, cm_id, con_num, recon_cnt,
1891 if (IS_ERR(srv_path)) {
1892 mutex_unlock(&srv->paths_mutex);
1894 err = PTR_ERR(srv_path);
1895 pr_err("RTRS server session allocation failed: %d\n", err);
1899 err = create_con(srv_path, cm_id, cid);
1901 rtrs_err((&srv_path->s), "create_con(), error %d\n", err);
1902 rtrs_rdma_do_reject(cm_id, err);
1904 * Since session has other connections we follow normal way
1905 * through workqueue, but still return an error to tell cma.c
1906 * to call rdma_destroy_id() for current connection.
1908 goto close_and_return_err;
1910 err = rtrs_rdma_do_accept(srv_path, cm_id);
1912 rtrs_err((&srv_path->s), "rtrs_rdma_do_accept(), error %d\n", err);
1913 rtrs_rdma_do_reject(cm_id, err);
1915 * Since current connection was successfully added to the
1916 * session we follow normal way through workqueue to close the
1917 * session, thus return 0 to tell cma.c we call
1918 * rdma_destroy_id() ourselves.
1921 goto close_and_return_err;
1923 mutex_unlock(&srv->paths_mutex);
1928 return rtrs_rdma_do_reject(cm_id, err);
1930 close_and_return_err:
1931 mutex_unlock(&srv->paths_mutex);
1932 close_path(srv_path);
1937 static int rtrs_srv_rdma_cm_handler(struct rdma_cm_id *cm_id,
1938 struct rdma_cm_event *ev)
1940 struct rtrs_srv_path *srv_path = NULL;
1941 struct rtrs_path *s = NULL;
1942 struct rtrs_con *c = NULL;
1944 if (ev->event == RDMA_CM_EVENT_CONNECT_REQUEST)
1946 * In case of error cma.c will destroy cm_id,
1947 * see cma_process_remove()
1949 return rtrs_rdma_connect(cm_id, ev->param.conn.private_data,
1950 ev->param.conn.private_data_len);
1954 srv_path = to_srv_path(s);
1956 switch (ev->event) {
1957 case RDMA_CM_EVENT_ESTABLISHED:
1960 case RDMA_CM_EVENT_REJECTED:
1961 case RDMA_CM_EVENT_CONNECT_ERROR:
1962 case RDMA_CM_EVENT_UNREACHABLE:
1963 rtrs_err(s, "CM error (CM event: %s, err: %d)\n",
1964 rdma_event_msg(ev->event), ev->status);
1966 case RDMA_CM_EVENT_DISCONNECTED:
1967 case RDMA_CM_EVENT_ADDR_CHANGE:
1968 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1969 case RDMA_CM_EVENT_DEVICE_REMOVAL:
1970 close_path(srv_path);
1973 pr_err("Ignoring unexpected CM event %s, err %d\n",
1974 rdma_event_msg(ev->event), ev->status);
1981 static struct rdma_cm_id *rtrs_srv_cm_init(struct rtrs_srv_ctx *ctx,
1982 struct sockaddr *addr,
1983 enum rdma_ucm_port_space ps)
1985 struct rdma_cm_id *cm_id;
1988 cm_id = rdma_create_id(&init_net, rtrs_srv_rdma_cm_handler,
1989 ctx, ps, IB_QPT_RC);
1990 if (IS_ERR(cm_id)) {
1991 ret = PTR_ERR(cm_id);
1992 pr_err("Creating id for RDMA connection failed, err: %d\n",
1996 ret = rdma_bind_addr(cm_id, addr);
1998 pr_err("Binding RDMA address failed, err: %d\n", ret);
2001 ret = rdma_listen(cm_id, 64);
2003 pr_err("Listening on RDMA connection failed, err: %d\n",
2011 rdma_destroy_id(cm_id);
2014 return ERR_PTR(ret);
2017 static int rtrs_srv_rdma_init(struct rtrs_srv_ctx *ctx, u16 port)
2019 struct sockaddr_in6 sin = {
2020 .sin6_family = AF_INET6,
2021 .sin6_addr = IN6ADDR_ANY_INIT,
2022 .sin6_port = htons(port),
2024 struct sockaddr_ib sib = {
2025 .sib_family = AF_IB,
2026 .sib_sid = cpu_to_be64(RDMA_IB_IP_PS_IB | port),
2027 .sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL),
2028 .sib_pkey = cpu_to_be16(0xffff),
2030 struct rdma_cm_id *cm_ip, *cm_ib;
2034 * We accept both IPoIB and IB connections, so we need to keep
2035 * two cm id's, one for each socket type and port space.
2036 * If the cm initialization of one of the id's fails, we abort
2039 cm_ip = rtrs_srv_cm_init(ctx, (struct sockaddr *)&sin, RDMA_PS_TCP);
2041 return PTR_ERR(cm_ip);
2043 cm_ib = rtrs_srv_cm_init(ctx, (struct sockaddr *)&sib, RDMA_PS_IB);
2044 if (IS_ERR(cm_ib)) {
2045 ret = PTR_ERR(cm_ib);
2049 ctx->cm_id_ip = cm_ip;
2050 ctx->cm_id_ib = cm_ib;
2055 rdma_destroy_id(cm_ip);
2060 static struct rtrs_srv_ctx *alloc_srv_ctx(struct rtrs_srv_ops *ops)
2062 struct rtrs_srv_ctx *ctx;
2064 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
2069 mutex_init(&ctx->srv_mutex);
2070 INIT_LIST_HEAD(&ctx->srv_list);
2075 static void free_srv_ctx(struct rtrs_srv_ctx *ctx)
2077 WARN_ON(!list_empty(&ctx->srv_list));
2078 mutex_destroy(&ctx->srv_mutex);
2082 static int rtrs_srv_add_one(struct ib_device *device)
2084 struct rtrs_srv_ctx *ctx;
2087 mutex_lock(&ib_ctx.ib_dev_mutex);
2088 if (ib_ctx.ib_dev_count)
2092 * Since our CM IDs are NOT bound to any ib device we will create them
2095 ctx = ib_ctx.srv_ctx;
2096 ret = rtrs_srv_rdma_init(ctx, ib_ctx.port);
2099 * We errored out here.
2100 * According to the ib code, if we encounter an error here then the
2101 * error code is ignored, and no more calls to our ops are made.
2103 pr_err("Failed to initialize RDMA connection");
2109 * Keep a track on the number of ib devices added
2111 ib_ctx.ib_dev_count++;
2114 mutex_unlock(&ib_ctx.ib_dev_mutex);
2118 static void rtrs_srv_remove_one(struct ib_device *device, void *client_data)
2120 struct rtrs_srv_ctx *ctx;
2122 mutex_lock(&ib_ctx.ib_dev_mutex);
2123 ib_ctx.ib_dev_count--;
2125 if (ib_ctx.ib_dev_count)
2129 * Since our CM IDs are NOT bound to any ib device we will remove them
2130 * only once, when the last device is removed
2132 ctx = ib_ctx.srv_ctx;
2133 rdma_destroy_id(ctx->cm_id_ip);
2134 rdma_destroy_id(ctx->cm_id_ib);
2137 mutex_unlock(&ib_ctx.ib_dev_mutex);
2140 static struct ib_client rtrs_srv_client = {
2141 .name = "rtrs_server",
2142 .add = rtrs_srv_add_one,
2143 .remove = rtrs_srv_remove_one
2147 * rtrs_srv_open() - open RTRS server context
2148 * @ops: callback functions
2149 * @port: port to listen on
2151 * Creates server context with specified callbacks.
2153 * Return a valid pointer on success otherwise PTR_ERR.
2155 struct rtrs_srv_ctx *rtrs_srv_open(struct rtrs_srv_ops *ops, u16 port)
2157 struct rtrs_srv_ctx *ctx;
2160 ctx = alloc_srv_ctx(ops);
2162 return ERR_PTR(-ENOMEM);
2164 mutex_init(&ib_ctx.ib_dev_mutex);
2165 ib_ctx.srv_ctx = ctx;
2168 err = ib_register_client(&rtrs_srv_client);
2171 return ERR_PTR(err);
2176 EXPORT_SYMBOL(rtrs_srv_open);
2178 static void close_paths(struct rtrs_srv_sess *srv)
2180 struct rtrs_srv_path *srv_path;
2182 mutex_lock(&srv->paths_mutex);
2183 list_for_each_entry(srv_path, &srv->paths_list, s.entry)
2184 close_path(srv_path);
2185 mutex_unlock(&srv->paths_mutex);
2188 static void close_ctx(struct rtrs_srv_ctx *ctx)
2190 struct rtrs_srv_sess *srv;
2192 mutex_lock(&ctx->srv_mutex);
2193 list_for_each_entry(srv, &ctx->srv_list, ctx_list)
2195 mutex_unlock(&ctx->srv_mutex);
2196 flush_workqueue(rtrs_wq);
2200 * rtrs_srv_close() - close RTRS server context
2201 * @ctx: pointer to server context
2203 * Closes RTRS server context with all client sessions.
2205 void rtrs_srv_close(struct rtrs_srv_ctx *ctx)
2207 ib_unregister_client(&rtrs_srv_client);
2208 mutex_destroy(&ib_ctx.ib_dev_mutex);
2212 EXPORT_SYMBOL(rtrs_srv_close);
2214 static int check_module_params(void)
2216 if (sess_queue_depth < 1 || sess_queue_depth > MAX_SESS_QUEUE_DEPTH) {
2217 pr_err("Invalid sess_queue_depth value %d, has to be >= %d, <= %d.\n",
2218 sess_queue_depth, 1, MAX_SESS_QUEUE_DEPTH);
2221 if (max_chunk_size < MIN_CHUNK_SIZE || !is_power_of_2(max_chunk_size)) {
2222 pr_err("Invalid max_chunk_size value %d, has to be >= %d and should be power of two.\n",
2223 max_chunk_size, MIN_CHUNK_SIZE);
2228 * Check if IB immediate data size is enough to hold the mem_id and the
2229 * offset inside the memory chunk
2231 if ((ilog2(sess_queue_depth - 1) + 1) +
2232 (ilog2(max_chunk_size - 1) + 1) > MAX_IMM_PAYL_BITS) {
2233 pr_err("RDMA immediate size (%db) not enough to encode %d buffers of size %dB. Reduce 'sess_queue_depth' or 'max_chunk_size' parameters.\n",
2234 MAX_IMM_PAYL_BITS, sess_queue_depth, max_chunk_size);
2241 static int __init rtrs_server_init(void)
2245 pr_info("Loading module %s, proto %s: (max_chunk_size: %d (pure IO %ld, headers %ld) , sess_queue_depth: %d, always_invalidate: %d)\n",
2246 KBUILD_MODNAME, RTRS_PROTO_VER_STRING,
2247 max_chunk_size, max_chunk_size - MAX_HDR_SIZE, MAX_HDR_SIZE,
2248 sess_queue_depth, always_invalidate);
2250 rtrs_rdma_dev_pd_init(0, &dev_pd);
2252 err = check_module_params();
2254 pr_err("Failed to load module, invalid module parameters, err: %d\n",
2258 err = class_register(&rtrs_dev_class);
2262 rtrs_wq = alloc_workqueue("rtrs_server_wq", 0, 0);
2271 class_unregister(&rtrs_dev_class);
2276 static void __exit rtrs_server_exit(void)
2278 destroy_workqueue(rtrs_wq);
2279 class_unregister(&rtrs_dev_class);
2280 rtrs_rdma_dev_pd_deinit(&dev_pd);
2283 module_init(rtrs_server_init);
2284 module_exit(rtrs_server_exit);