2 * NVMe over Fabrics RDMA target.
3 * Copyright (c) 2015-2016 HGST, a Western Digital Company.
5 * This program is free software; you can redistribute it and/or modify it
6 * under the terms and conditions of the GNU General Public License,
7 * version 2, as published by the Free Software Foundation.
9 * This program is distributed in the hope it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
14 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
15 #include <linux/atomic.h>
16 #include <linux/ctype.h>
17 #include <linux/delay.h>
18 #include <linux/err.h>
19 #include <linux/init.h>
20 #include <linux/module.h>
21 #include <linux/nvme.h>
22 #include <linux/slab.h>
23 #include <linux/string.h>
24 #include <linux/wait.h>
25 #include <linux/inet.h>
26 #include <asm/unaligned.h>
28 #include <rdma/ib_verbs.h>
29 #include <rdma/rdma_cm.h>
32 #include <linux/nvme-rdma.h>
36 * We allow at least 1 page, up to 4 SGEs, and up to 16KB of inline data
38 #define NVMET_RDMA_DEFAULT_INLINE_DATA_SIZE PAGE_SIZE
39 #define NVMET_RDMA_MAX_INLINE_SGE 4
40 #define NVMET_RDMA_MAX_INLINE_DATA_SIZE max_t(int, SZ_16K, PAGE_SIZE)
42 struct nvmet_rdma_cmd {
43 struct ib_sge sge[NVMET_RDMA_MAX_INLINE_SGE + 1];
46 struct scatterlist inline_sg[NVMET_RDMA_MAX_INLINE_SGE];
47 struct nvme_command *nvme_cmd;
48 struct nvmet_rdma_queue *queue;
52 NVMET_RDMA_REQ_INLINE_DATA = (1 << 0),
53 NVMET_RDMA_REQ_INVALIDATE_RKEY = (1 << 1),
56 struct nvmet_rdma_rsp {
57 struct ib_sge send_sge;
58 struct ib_cqe send_cqe;
59 struct ib_send_wr send_wr;
61 struct nvmet_rdma_cmd *cmd;
62 struct nvmet_rdma_queue *queue;
64 struct ib_cqe read_cqe;
65 struct rdma_rw_ctx rw;
73 struct list_head wait_list;
74 struct list_head free_list;
77 enum nvmet_rdma_queue_state {
78 NVMET_RDMA_Q_CONNECTING,
80 NVMET_RDMA_Q_DISCONNECTING,
83 struct nvmet_rdma_queue {
84 struct rdma_cm_id *cm_id;
85 struct nvmet_port *port;
88 struct nvmet_rdma_device *dev;
89 spinlock_t state_lock;
90 enum nvmet_rdma_queue_state state;
91 struct nvmet_cq nvme_cq;
92 struct nvmet_sq nvme_sq;
94 struct nvmet_rdma_rsp *rsps;
95 struct list_head free_rsps;
97 struct nvmet_rdma_cmd *cmds;
99 struct work_struct release_work;
100 struct list_head rsp_wait_list;
101 struct list_head rsp_wr_wait_list;
102 spinlock_t rsp_wr_wait_lock;
109 struct list_head queue_list;
112 struct nvmet_rdma_device {
113 struct ib_device *device;
116 struct nvmet_rdma_cmd *srq_cmds;
119 struct list_head entry;
120 int inline_data_size;
121 int inline_page_count;
124 static bool nvmet_rdma_use_srq;
125 module_param_named(use_srq, nvmet_rdma_use_srq, bool, 0444);
126 MODULE_PARM_DESC(use_srq, "Use shared receive queue.");
128 static DEFINE_IDA(nvmet_rdma_queue_ida);
129 static LIST_HEAD(nvmet_rdma_queue_list);
130 static DEFINE_MUTEX(nvmet_rdma_queue_mutex);
132 static LIST_HEAD(device_list);
133 static DEFINE_MUTEX(device_list_mutex);
135 static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp);
136 static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc);
137 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
138 static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc);
139 static void nvmet_rdma_qp_event(struct ib_event *event, void *priv);
140 static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue);
142 static const struct nvmet_fabrics_ops nvmet_rdma_ops;
144 static int num_pages(int len)
146 return 1 + (((len - 1) & PAGE_MASK) >> PAGE_SHIFT);
149 /* XXX: really should move to a generic header sooner or later.. */
150 static inline u32 get_unaligned_le24(const u8 *p)
152 return (u32)p[0] | (u32)p[1] << 8 | (u32)p[2] << 16;
155 static inline bool nvmet_rdma_need_data_in(struct nvmet_rdma_rsp *rsp)
157 return nvme_is_write(rsp->req.cmd) &&
158 rsp->req.transfer_len &&
159 !(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
162 static inline bool nvmet_rdma_need_data_out(struct nvmet_rdma_rsp *rsp)
164 return !nvme_is_write(rsp->req.cmd) &&
165 rsp->req.transfer_len &&
166 !rsp->req.rsp->status &&
167 !(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
170 static inline struct nvmet_rdma_rsp *
171 nvmet_rdma_get_rsp(struct nvmet_rdma_queue *queue)
173 struct nvmet_rdma_rsp *rsp;
176 spin_lock_irqsave(&queue->rsps_lock, flags);
177 rsp = list_first_entry(&queue->free_rsps,
178 struct nvmet_rdma_rsp, free_list);
179 list_del(&rsp->free_list);
180 spin_unlock_irqrestore(&queue->rsps_lock, flags);
186 nvmet_rdma_put_rsp(struct nvmet_rdma_rsp *rsp)
190 spin_lock_irqsave(&rsp->queue->rsps_lock, flags);
191 list_add_tail(&rsp->free_list, &rsp->queue->free_rsps);
192 spin_unlock_irqrestore(&rsp->queue->rsps_lock, flags);
195 static void nvmet_rdma_free_inline_pages(struct nvmet_rdma_device *ndev,
196 struct nvmet_rdma_cmd *c)
198 struct scatterlist *sg;
202 if (!ndev->inline_data_size)
208 for (i = 0; i < ndev->inline_page_count; i++, sg++, sge++) {
210 ib_dma_unmap_page(ndev->device, sge->addr,
211 sge->length, DMA_FROM_DEVICE);
213 __free_page(sg_page(sg));
217 static int nvmet_rdma_alloc_inline_pages(struct nvmet_rdma_device *ndev,
218 struct nvmet_rdma_cmd *c)
220 struct scatterlist *sg;
226 if (!ndev->inline_data_size)
230 sg_init_table(sg, ndev->inline_page_count);
232 len = ndev->inline_data_size;
234 for (i = 0; i < ndev->inline_page_count; i++, sg++, sge++) {
235 pg = alloc_page(GFP_KERNEL);
238 sg_assign_page(sg, pg);
239 sge->addr = ib_dma_map_page(ndev->device,
240 pg, 0, PAGE_SIZE, DMA_FROM_DEVICE);
241 if (ib_dma_mapping_error(ndev->device, sge->addr))
243 sge->length = min_t(int, len, PAGE_SIZE);
244 sge->lkey = ndev->pd->local_dma_lkey;
250 for (; i >= 0; i--, sg--, sge--) {
252 ib_dma_unmap_page(ndev->device, sge->addr,
253 sge->length, DMA_FROM_DEVICE);
255 __free_page(sg_page(sg));
260 static int nvmet_rdma_alloc_cmd(struct nvmet_rdma_device *ndev,
261 struct nvmet_rdma_cmd *c, bool admin)
263 /* NVMe command / RDMA RECV */
264 c->nvme_cmd = kmalloc(sizeof(*c->nvme_cmd), GFP_KERNEL);
268 c->sge[0].addr = ib_dma_map_single(ndev->device, c->nvme_cmd,
269 sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
270 if (ib_dma_mapping_error(ndev->device, c->sge[0].addr))
273 c->sge[0].length = sizeof(*c->nvme_cmd);
274 c->sge[0].lkey = ndev->pd->local_dma_lkey;
276 if (!admin && nvmet_rdma_alloc_inline_pages(ndev, c))
279 c->cqe.done = nvmet_rdma_recv_done;
281 c->wr.wr_cqe = &c->cqe;
282 c->wr.sg_list = c->sge;
283 c->wr.num_sge = admin ? 1 : ndev->inline_page_count + 1;
288 ib_dma_unmap_single(ndev->device, c->sge[0].addr,
289 sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
297 static void nvmet_rdma_free_cmd(struct nvmet_rdma_device *ndev,
298 struct nvmet_rdma_cmd *c, bool admin)
301 nvmet_rdma_free_inline_pages(ndev, c);
302 ib_dma_unmap_single(ndev->device, c->sge[0].addr,
303 sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
307 static struct nvmet_rdma_cmd *
308 nvmet_rdma_alloc_cmds(struct nvmet_rdma_device *ndev,
309 int nr_cmds, bool admin)
311 struct nvmet_rdma_cmd *cmds;
312 int ret = -EINVAL, i;
314 cmds = kcalloc(nr_cmds, sizeof(struct nvmet_rdma_cmd), GFP_KERNEL);
318 for (i = 0; i < nr_cmds; i++) {
319 ret = nvmet_rdma_alloc_cmd(ndev, cmds + i, admin);
328 nvmet_rdma_free_cmd(ndev, cmds + i, admin);
334 static void nvmet_rdma_free_cmds(struct nvmet_rdma_device *ndev,
335 struct nvmet_rdma_cmd *cmds, int nr_cmds, bool admin)
339 for (i = 0; i < nr_cmds; i++)
340 nvmet_rdma_free_cmd(ndev, cmds + i, admin);
344 static int nvmet_rdma_alloc_rsp(struct nvmet_rdma_device *ndev,
345 struct nvmet_rdma_rsp *r)
347 /* NVMe CQE / RDMA SEND */
348 r->req.rsp = kmalloc(sizeof(*r->req.rsp), GFP_KERNEL);
352 r->send_sge.addr = ib_dma_map_single(ndev->device, r->req.rsp,
353 sizeof(*r->req.rsp), DMA_TO_DEVICE);
354 if (ib_dma_mapping_error(ndev->device, r->send_sge.addr))
357 r->send_sge.length = sizeof(*r->req.rsp);
358 r->send_sge.lkey = ndev->pd->local_dma_lkey;
360 r->send_cqe.done = nvmet_rdma_send_done;
362 r->send_wr.wr_cqe = &r->send_cqe;
363 r->send_wr.sg_list = &r->send_sge;
364 r->send_wr.num_sge = 1;
365 r->send_wr.send_flags = IB_SEND_SIGNALED;
367 /* Data In / RDMA READ */
368 r->read_cqe.done = nvmet_rdma_read_data_done;
377 static void nvmet_rdma_free_rsp(struct nvmet_rdma_device *ndev,
378 struct nvmet_rdma_rsp *r)
380 ib_dma_unmap_single(ndev->device, r->send_sge.addr,
381 sizeof(*r->req.rsp), DMA_TO_DEVICE);
386 nvmet_rdma_alloc_rsps(struct nvmet_rdma_queue *queue)
388 struct nvmet_rdma_device *ndev = queue->dev;
389 int nr_rsps = queue->recv_queue_size * 2;
390 int ret = -EINVAL, i;
392 queue->rsps = kcalloc(nr_rsps, sizeof(struct nvmet_rdma_rsp),
397 for (i = 0; i < nr_rsps; i++) {
398 struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
400 ret = nvmet_rdma_alloc_rsp(ndev, rsp);
404 list_add_tail(&rsp->free_list, &queue->free_rsps);
411 struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
413 list_del(&rsp->free_list);
414 nvmet_rdma_free_rsp(ndev, rsp);
421 static void nvmet_rdma_free_rsps(struct nvmet_rdma_queue *queue)
423 struct nvmet_rdma_device *ndev = queue->dev;
424 int i, nr_rsps = queue->recv_queue_size * 2;
426 for (i = 0; i < nr_rsps; i++) {
427 struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
429 list_del(&rsp->free_list);
430 nvmet_rdma_free_rsp(ndev, rsp);
435 static int nvmet_rdma_post_recv(struct nvmet_rdma_device *ndev,
436 struct nvmet_rdma_cmd *cmd)
438 struct ib_recv_wr *bad_wr;
441 ib_dma_sync_single_for_device(ndev->device,
442 cmd->sge[0].addr, cmd->sge[0].length,
446 ret = ib_post_srq_recv(ndev->srq, &cmd->wr, &bad_wr);
448 ret = ib_post_recv(cmd->queue->cm_id->qp, &cmd->wr, &bad_wr);
451 pr_err("post_recv cmd failed\n");
456 static void nvmet_rdma_process_wr_wait_list(struct nvmet_rdma_queue *queue)
458 spin_lock(&queue->rsp_wr_wait_lock);
459 while (!list_empty(&queue->rsp_wr_wait_list)) {
460 struct nvmet_rdma_rsp *rsp;
463 rsp = list_entry(queue->rsp_wr_wait_list.next,
464 struct nvmet_rdma_rsp, wait_list);
465 list_del(&rsp->wait_list);
467 spin_unlock(&queue->rsp_wr_wait_lock);
468 ret = nvmet_rdma_execute_command(rsp);
469 spin_lock(&queue->rsp_wr_wait_lock);
472 list_add(&rsp->wait_list, &queue->rsp_wr_wait_list);
476 spin_unlock(&queue->rsp_wr_wait_lock);
480 static void nvmet_rdma_release_rsp(struct nvmet_rdma_rsp *rsp)
482 struct nvmet_rdma_queue *queue = rsp->queue;
484 atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
487 rdma_rw_ctx_destroy(&rsp->rw, queue->cm_id->qp,
488 queue->cm_id->port_num, rsp->req.sg,
489 rsp->req.sg_cnt, nvmet_data_dir(&rsp->req));
492 if (rsp->req.sg != rsp->cmd->inline_sg)
493 sgl_free(rsp->req.sg);
495 if (unlikely(!list_empty_careful(&queue->rsp_wr_wait_list)))
496 nvmet_rdma_process_wr_wait_list(queue);
498 nvmet_rdma_put_rsp(rsp);
501 static void nvmet_rdma_error_comp(struct nvmet_rdma_queue *queue)
503 if (queue->nvme_sq.ctrl) {
504 nvmet_ctrl_fatal_error(queue->nvme_sq.ctrl);
507 * we didn't setup the controller yet in case
508 * of admin connect error, just disconnect and
511 nvmet_rdma_queue_disconnect(queue);
515 static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
517 struct nvmet_rdma_rsp *rsp =
518 container_of(wc->wr_cqe, struct nvmet_rdma_rsp, send_cqe);
520 nvmet_rdma_release_rsp(rsp);
522 if (unlikely(wc->status != IB_WC_SUCCESS &&
523 wc->status != IB_WC_WR_FLUSH_ERR)) {
524 pr_err("SEND for CQE 0x%p failed with status %s (%d).\n",
525 wc->wr_cqe, ib_wc_status_msg(wc->status), wc->status);
526 nvmet_rdma_error_comp(rsp->queue);
530 static void nvmet_rdma_queue_response(struct nvmet_req *req)
532 struct nvmet_rdma_rsp *rsp =
533 container_of(req, struct nvmet_rdma_rsp, req);
534 struct rdma_cm_id *cm_id = rsp->queue->cm_id;
535 struct ib_send_wr *first_wr, *bad_wr;
537 if (rsp->flags & NVMET_RDMA_REQ_INVALIDATE_RKEY) {
538 rsp->send_wr.opcode = IB_WR_SEND_WITH_INV;
539 rsp->send_wr.ex.invalidate_rkey = rsp->invalidate_rkey;
541 rsp->send_wr.opcode = IB_WR_SEND;
544 if (nvmet_rdma_need_data_out(rsp))
545 first_wr = rdma_rw_ctx_wrs(&rsp->rw, cm_id->qp,
546 cm_id->port_num, NULL, &rsp->send_wr);
548 first_wr = &rsp->send_wr;
550 nvmet_rdma_post_recv(rsp->queue->dev, rsp->cmd);
552 ib_dma_sync_single_for_device(rsp->queue->dev->device,
553 rsp->send_sge.addr, rsp->send_sge.length,
556 if (unlikely(ib_post_send(cm_id->qp, first_wr, &bad_wr))) {
557 pr_err("sending cmd response failed\n");
558 nvmet_rdma_release_rsp(rsp);
562 static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc)
564 struct nvmet_rdma_rsp *rsp =
565 container_of(wc->wr_cqe, struct nvmet_rdma_rsp, read_cqe);
566 struct nvmet_rdma_queue *queue = cq->cq_context;
568 WARN_ON(rsp->n_rdma <= 0);
569 atomic_add(rsp->n_rdma, &queue->sq_wr_avail);
570 rdma_rw_ctx_destroy(&rsp->rw, queue->cm_id->qp,
571 queue->cm_id->port_num, rsp->req.sg,
572 rsp->req.sg_cnt, nvmet_data_dir(&rsp->req));
575 if (unlikely(wc->status != IB_WC_SUCCESS)) {
576 nvmet_req_uninit(&rsp->req);
577 nvmet_rdma_release_rsp(rsp);
578 if (wc->status != IB_WC_WR_FLUSH_ERR) {
579 pr_info("RDMA READ for CQE 0x%p failed with status %s (%d).\n",
580 wc->wr_cqe, ib_wc_status_msg(wc->status), wc->status);
581 nvmet_rdma_error_comp(queue);
586 nvmet_req_execute(&rsp->req);
589 static void nvmet_rdma_use_inline_sg(struct nvmet_rdma_rsp *rsp, u32 len,
592 int sg_count = num_pages(len);
593 struct scatterlist *sg;
596 sg = rsp->cmd->inline_sg;
597 for (i = 0; i < sg_count; i++, sg++) {
598 if (i < sg_count - 1)
603 sg->length = min_t(int, len, PAGE_SIZE - off);
609 rsp->req.sg = rsp->cmd->inline_sg;
610 rsp->req.sg_cnt = sg_count;
613 static u16 nvmet_rdma_map_sgl_inline(struct nvmet_rdma_rsp *rsp)
615 struct nvme_sgl_desc *sgl = &rsp->req.cmd->common.dptr.sgl;
616 u64 off = le64_to_cpu(sgl->addr);
617 u32 len = le32_to_cpu(sgl->length);
619 if (!nvme_is_write(rsp->req.cmd))
620 return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
622 if (off + len > rsp->queue->dev->inline_data_size) {
623 pr_err("invalid inline data offset!\n");
624 return NVME_SC_SGL_INVALID_OFFSET | NVME_SC_DNR;
627 /* no data command? */
631 nvmet_rdma_use_inline_sg(rsp, len, off);
632 rsp->flags |= NVMET_RDMA_REQ_INLINE_DATA;
633 rsp->req.transfer_len += len;
637 static u16 nvmet_rdma_map_sgl_keyed(struct nvmet_rdma_rsp *rsp,
638 struct nvme_keyed_sgl_desc *sgl, bool invalidate)
640 struct rdma_cm_id *cm_id = rsp->queue->cm_id;
641 u64 addr = le64_to_cpu(sgl->addr);
642 u32 len = get_unaligned_le24(sgl->length);
643 u32 key = get_unaligned_le32(sgl->key);
646 /* no data command? */
650 rsp->req.sg = sgl_alloc(len, GFP_KERNEL, &rsp->req.sg_cnt);
652 return NVME_SC_INTERNAL;
654 ret = rdma_rw_ctx_init(&rsp->rw, cm_id->qp, cm_id->port_num,
655 rsp->req.sg, rsp->req.sg_cnt, 0, addr, key,
656 nvmet_data_dir(&rsp->req));
658 return NVME_SC_INTERNAL;
659 rsp->req.transfer_len += len;
663 rsp->invalidate_rkey = key;
664 rsp->flags |= NVMET_RDMA_REQ_INVALIDATE_RKEY;
670 static u16 nvmet_rdma_map_sgl(struct nvmet_rdma_rsp *rsp)
672 struct nvme_keyed_sgl_desc *sgl = &rsp->req.cmd->common.dptr.ksgl;
674 switch (sgl->type >> 4) {
675 case NVME_SGL_FMT_DATA_DESC:
676 switch (sgl->type & 0xf) {
677 case NVME_SGL_FMT_OFFSET:
678 return nvmet_rdma_map_sgl_inline(rsp);
680 pr_err("invalid SGL subtype: %#x\n", sgl->type);
681 return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
683 case NVME_KEY_SGL_FMT_DATA_DESC:
684 switch (sgl->type & 0xf) {
685 case NVME_SGL_FMT_ADDRESS | NVME_SGL_FMT_INVALIDATE:
686 return nvmet_rdma_map_sgl_keyed(rsp, sgl, true);
687 case NVME_SGL_FMT_ADDRESS:
688 return nvmet_rdma_map_sgl_keyed(rsp, sgl, false);
690 pr_err("invalid SGL subtype: %#x\n", sgl->type);
691 return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
694 pr_err("invalid SGL type: %#x\n", sgl->type);
695 return NVME_SC_SGL_INVALID_TYPE | NVME_SC_DNR;
699 static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp)
701 struct nvmet_rdma_queue *queue = rsp->queue;
703 if (unlikely(atomic_sub_return(1 + rsp->n_rdma,
704 &queue->sq_wr_avail) < 0)) {
705 pr_debug("IB send queue full (needed %d): queue %u cntlid %u\n",
706 1 + rsp->n_rdma, queue->idx,
707 queue->nvme_sq.ctrl->cntlid);
708 atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
712 if (nvmet_rdma_need_data_in(rsp)) {
713 if (rdma_rw_ctx_post(&rsp->rw, queue->cm_id->qp,
714 queue->cm_id->port_num, &rsp->read_cqe, NULL))
715 nvmet_req_complete(&rsp->req, NVME_SC_DATA_XFER_ERROR);
717 nvmet_req_execute(&rsp->req);
723 static void nvmet_rdma_handle_command(struct nvmet_rdma_queue *queue,
724 struct nvmet_rdma_rsp *cmd)
728 ib_dma_sync_single_for_cpu(queue->dev->device,
729 cmd->cmd->sge[0].addr, cmd->cmd->sge[0].length,
731 ib_dma_sync_single_for_cpu(queue->dev->device,
732 cmd->send_sge.addr, cmd->send_sge.length,
735 if (!nvmet_req_init(&cmd->req, &queue->nvme_cq,
736 &queue->nvme_sq, &nvmet_rdma_ops))
739 status = nvmet_rdma_map_sgl(cmd);
743 if (unlikely(!nvmet_rdma_execute_command(cmd))) {
744 spin_lock(&queue->rsp_wr_wait_lock);
745 list_add_tail(&cmd->wait_list, &queue->rsp_wr_wait_list);
746 spin_unlock(&queue->rsp_wr_wait_lock);
752 nvmet_req_complete(&cmd->req, status);
755 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
757 struct nvmet_rdma_cmd *cmd =
758 container_of(wc->wr_cqe, struct nvmet_rdma_cmd, cqe);
759 struct nvmet_rdma_queue *queue = cq->cq_context;
760 struct nvmet_rdma_rsp *rsp;
762 if (unlikely(wc->status != IB_WC_SUCCESS)) {
763 if (wc->status != IB_WC_WR_FLUSH_ERR) {
764 pr_err("RECV for CQE 0x%p failed with status %s (%d)\n",
765 wc->wr_cqe, ib_wc_status_msg(wc->status),
767 nvmet_rdma_error_comp(queue);
772 if (unlikely(wc->byte_len < sizeof(struct nvme_command))) {
773 pr_err("Ctrl Fatal Error: capsule size less than 64 bytes\n");
774 nvmet_rdma_error_comp(queue);
779 rsp = nvmet_rdma_get_rsp(queue);
783 rsp->req.cmd = cmd->nvme_cmd;
784 rsp->req.port = queue->port;
787 if (unlikely(queue->state != NVMET_RDMA_Q_LIVE)) {
790 spin_lock_irqsave(&queue->state_lock, flags);
791 if (queue->state == NVMET_RDMA_Q_CONNECTING)
792 list_add_tail(&rsp->wait_list, &queue->rsp_wait_list);
794 nvmet_rdma_put_rsp(rsp);
795 spin_unlock_irqrestore(&queue->state_lock, flags);
799 nvmet_rdma_handle_command(queue, rsp);
802 static void nvmet_rdma_destroy_srq(struct nvmet_rdma_device *ndev)
807 nvmet_rdma_free_cmds(ndev, ndev->srq_cmds, ndev->srq_size, false);
808 ib_destroy_srq(ndev->srq);
811 static int nvmet_rdma_init_srq(struct nvmet_rdma_device *ndev)
813 struct ib_srq_init_attr srq_attr = { NULL, };
818 srq_size = 4095; /* XXX: tune */
820 srq_attr.attr.max_wr = srq_size;
821 srq_attr.attr.max_sge = 1 + ndev->inline_page_count;
822 srq_attr.attr.srq_limit = 0;
823 srq_attr.srq_type = IB_SRQT_BASIC;
824 srq = ib_create_srq(ndev->pd, &srq_attr);
827 * If SRQs aren't supported we just go ahead and use normal
828 * non-shared receive queues.
830 pr_info("SRQ requested but not supported.\n");
834 ndev->srq_cmds = nvmet_rdma_alloc_cmds(ndev, srq_size, false);
835 if (IS_ERR(ndev->srq_cmds)) {
836 ret = PTR_ERR(ndev->srq_cmds);
837 goto out_destroy_srq;
841 ndev->srq_size = srq_size;
843 for (i = 0; i < srq_size; i++) {
844 ret = nvmet_rdma_post_recv(ndev, &ndev->srq_cmds[i]);
852 nvmet_rdma_free_cmds(ndev, ndev->srq_cmds, ndev->srq_size, false);
858 static void nvmet_rdma_free_dev(struct kref *ref)
860 struct nvmet_rdma_device *ndev =
861 container_of(ref, struct nvmet_rdma_device, ref);
863 mutex_lock(&device_list_mutex);
864 list_del(&ndev->entry);
865 mutex_unlock(&device_list_mutex);
867 nvmet_rdma_destroy_srq(ndev);
868 ib_dealloc_pd(ndev->pd);
873 static struct nvmet_rdma_device *
874 nvmet_rdma_find_get_device(struct rdma_cm_id *cm_id)
876 struct nvmet_port *port = cm_id->context;
877 struct nvmet_rdma_device *ndev;
878 int inline_page_count;
879 int inline_sge_count;
882 mutex_lock(&device_list_mutex);
883 list_for_each_entry(ndev, &device_list, entry) {
884 if (ndev->device->node_guid == cm_id->device->node_guid &&
885 kref_get_unless_zero(&ndev->ref))
889 ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
893 inline_page_count = num_pages(port->inline_data_size);
894 inline_sge_count = max(cm_id->device->attrs.max_sge_rd,
895 cm_id->device->attrs.max_sge) - 1;
896 if (inline_page_count > inline_sge_count) {
897 pr_warn("inline_data_size %d cannot be supported by device %s. Reducing to %lu.\n",
898 port->inline_data_size, cm_id->device->name,
899 inline_sge_count * PAGE_SIZE);
900 port->inline_data_size = inline_sge_count * PAGE_SIZE;
901 inline_page_count = inline_sge_count;
903 ndev->inline_data_size = port->inline_data_size;
904 ndev->inline_page_count = inline_page_count;
905 ndev->device = cm_id->device;
906 kref_init(&ndev->ref);
908 ndev->pd = ib_alloc_pd(ndev->device, 0);
909 if (IS_ERR(ndev->pd))
912 if (nvmet_rdma_use_srq) {
913 ret = nvmet_rdma_init_srq(ndev);
918 list_add(&ndev->entry, &device_list);
920 mutex_unlock(&device_list_mutex);
921 pr_debug("added %s.\n", ndev->device->name);
925 ib_dealloc_pd(ndev->pd);
929 mutex_unlock(&device_list_mutex);
933 static int nvmet_rdma_create_queue_ib(struct nvmet_rdma_queue *queue)
935 struct ib_qp_init_attr qp_attr;
936 struct nvmet_rdma_device *ndev = queue->dev;
937 int comp_vector, nr_cqe, ret, i;
940 * Spread the io queues across completion vectors,
941 * but still keep all admin queues on vector 0.
943 comp_vector = !queue->host_qid ? 0 :
944 queue->idx % ndev->device->num_comp_vectors;
947 * Reserve CQ slots for RECV + RDMA_READ/RDMA_WRITE + RDMA_SEND.
949 nr_cqe = queue->recv_queue_size + 2 * queue->send_queue_size;
951 queue->cq = ib_alloc_cq(ndev->device, queue,
952 nr_cqe + 1, comp_vector,
954 if (IS_ERR(queue->cq)) {
955 ret = PTR_ERR(queue->cq);
956 pr_err("failed to create CQ cqe= %d ret= %d\n",
961 memset(&qp_attr, 0, sizeof(qp_attr));
962 qp_attr.qp_context = queue;
963 qp_attr.event_handler = nvmet_rdma_qp_event;
964 qp_attr.send_cq = queue->cq;
965 qp_attr.recv_cq = queue->cq;
966 qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
967 qp_attr.qp_type = IB_QPT_RC;
969 qp_attr.cap.max_send_wr = queue->send_queue_size + 1;
970 qp_attr.cap.max_rdma_ctxs = queue->send_queue_size;
971 qp_attr.cap.max_send_sge = max(ndev->device->attrs.max_sge_rd,
972 ndev->device->attrs.max_sge);
975 qp_attr.srq = ndev->srq;
978 qp_attr.cap.max_recv_wr = 1 + queue->recv_queue_size;
979 qp_attr.cap.max_recv_sge = 1 + ndev->inline_page_count;
982 ret = rdma_create_qp(queue->cm_id, ndev->pd, &qp_attr);
984 pr_err("failed to create_qp ret= %d\n", ret);
988 atomic_set(&queue->sq_wr_avail, qp_attr.cap.max_send_wr);
990 pr_debug("%s: max_cqe= %d max_sge= %d sq_size = %d cm_id= %p\n",
991 __func__, queue->cq->cqe, qp_attr.cap.max_send_sge,
992 qp_attr.cap.max_send_wr, queue->cm_id);
995 for (i = 0; i < queue->recv_queue_size; i++) {
996 queue->cmds[i].queue = queue;
997 ret = nvmet_rdma_post_recv(ndev, &queue->cmds[i]);
1007 rdma_destroy_qp(queue->cm_id);
1009 ib_free_cq(queue->cq);
1013 static void nvmet_rdma_destroy_queue_ib(struct nvmet_rdma_queue *queue)
1015 struct ib_qp *qp = queue->cm_id->qp;
1018 rdma_destroy_id(queue->cm_id);
1020 ib_free_cq(queue->cq);
1023 static void nvmet_rdma_free_queue(struct nvmet_rdma_queue *queue)
1025 pr_debug("freeing queue %d\n", queue->idx);
1027 nvmet_sq_destroy(&queue->nvme_sq);
1029 nvmet_rdma_destroy_queue_ib(queue);
1030 if (!queue->dev->srq) {
1031 nvmet_rdma_free_cmds(queue->dev, queue->cmds,
1032 queue->recv_queue_size,
1035 nvmet_rdma_free_rsps(queue);
1036 ida_simple_remove(&nvmet_rdma_queue_ida, queue->idx);
1040 static void nvmet_rdma_release_queue_work(struct work_struct *w)
1042 struct nvmet_rdma_queue *queue =
1043 container_of(w, struct nvmet_rdma_queue, release_work);
1044 struct nvmet_rdma_device *dev = queue->dev;
1046 nvmet_rdma_free_queue(queue);
1048 kref_put(&dev->ref, nvmet_rdma_free_dev);
1052 nvmet_rdma_parse_cm_connect_req(struct rdma_conn_param *conn,
1053 struct nvmet_rdma_queue *queue)
1055 struct nvme_rdma_cm_req *req;
1057 req = (struct nvme_rdma_cm_req *)conn->private_data;
1058 if (!req || conn->private_data_len == 0)
1059 return NVME_RDMA_CM_INVALID_LEN;
1061 if (le16_to_cpu(req->recfmt) != NVME_RDMA_CM_FMT_1_0)
1062 return NVME_RDMA_CM_INVALID_RECFMT;
1064 queue->host_qid = le16_to_cpu(req->qid);
1067 * req->hsqsize corresponds to our recv queue size plus 1
1068 * req->hrqsize corresponds to our send queue size
1070 queue->recv_queue_size = le16_to_cpu(req->hsqsize) + 1;
1071 queue->send_queue_size = le16_to_cpu(req->hrqsize);
1073 if (!queue->host_qid && queue->recv_queue_size > NVME_AQ_DEPTH)
1074 return NVME_RDMA_CM_INVALID_HSQSIZE;
1076 /* XXX: Should we enforce some kind of max for IO queues? */
1081 static int nvmet_rdma_cm_reject(struct rdma_cm_id *cm_id,
1082 enum nvme_rdma_cm_status status)
1084 struct nvme_rdma_cm_rej rej;
1086 pr_debug("rejecting connect request: status %d (%s)\n",
1087 status, nvme_rdma_cm_msg(status));
1089 rej.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1090 rej.sts = cpu_to_le16(status);
1092 return rdma_reject(cm_id, (void *)&rej, sizeof(rej));
1095 static struct nvmet_rdma_queue *
1096 nvmet_rdma_alloc_queue(struct nvmet_rdma_device *ndev,
1097 struct rdma_cm_id *cm_id,
1098 struct rdma_cm_event *event)
1100 struct nvmet_rdma_queue *queue;
1103 queue = kzalloc(sizeof(*queue), GFP_KERNEL);
1105 ret = NVME_RDMA_CM_NO_RSC;
1109 ret = nvmet_sq_init(&queue->nvme_sq);
1111 ret = NVME_RDMA_CM_NO_RSC;
1112 goto out_free_queue;
1115 ret = nvmet_rdma_parse_cm_connect_req(&event->param.conn, queue);
1117 goto out_destroy_sq;
1120 * Schedules the actual release because calling rdma_destroy_id from
1121 * inside a CM callback would trigger a deadlock. (great API design..)
1123 INIT_WORK(&queue->release_work, nvmet_rdma_release_queue_work);
1125 queue->cm_id = cm_id;
1127 spin_lock_init(&queue->state_lock);
1128 queue->state = NVMET_RDMA_Q_CONNECTING;
1129 INIT_LIST_HEAD(&queue->rsp_wait_list);
1130 INIT_LIST_HEAD(&queue->rsp_wr_wait_list);
1131 spin_lock_init(&queue->rsp_wr_wait_lock);
1132 INIT_LIST_HEAD(&queue->free_rsps);
1133 spin_lock_init(&queue->rsps_lock);
1134 INIT_LIST_HEAD(&queue->queue_list);
1136 queue->idx = ida_simple_get(&nvmet_rdma_queue_ida, 0, 0, GFP_KERNEL);
1137 if (queue->idx < 0) {
1138 ret = NVME_RDMA_CM_NO_RSC;
1139 goto out_destroy_sq;
1142 ret = nvmet_rdma_alloc_rsps(queue);
1144 ret = NVME_RDMA_CM_NO_RSC;
1145 goto out_ida_remove;
1149 queue->cmds = nvmet_rdma_alloc_cmds(ndev,
1150 queue->recv_queue_size,
1152 if (IS_ERR(queue->cmds)) {
1153 ret = NVME_RDMA_CM_NO_RSC;
1154 goto out_free_responses;
1158 ret = nvmet_rdma_create_queue_ib(queue);
1160 pr_err("%s: creating RDMA queue failed (%d).\n",
1162 ret = NVME_RDMA_CM_NO_RSC;
1170 nvmet_rdma_free_cmds(queue->dev, queue->cmds,
1171 queue->recv_queue_size,
1175 nvmet_rdma_free_rsps(queue);
1177 ida_simple_remove(&nvmet_rdma_queue_ida, queue->idx);
1179 nvmet_sq_destroy(&queue->nvme_sq);
1183 nvmet_rdma_cm_reject(cm_id, ret);
1187 static void nvmet_rdma_qp_event(struct ib_event *event, void *priv)
1189 struct nvmet_rdma_queue *queue = priv;
1191 switch (event->event) {
1192 case IB_EVENT_COMM_EST:
1193 rdma_notify(queue->cm_id, event->event);
1196 pr_err("received IB QP event: %s (%d)\n",
1197 ib_event_msg(event->event), event->event);
1202 static int nvmet_rdma_cm_accept(struct rdma_cm_id *cm_id,
1203 struct nvmet_rdma_queue *queue,
1204 struct rdma_conn_param *p)
1206 struct rdma_conn_param param = { };
1207 struct nvme_rdma_cm_rep priv = { };
1210 param.rnr_retry_count = 7;
1211 param.flow_control = 1;
1212 param.initiator_depth = min_t(u8, p->initiator_depth,
1213 queue->dev->device->attrs.max_qp_init_rd_atom);
1214 param.private_data = &priv;
1215 param.private_data_len = sizeof(priv);
1216 priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1217 priv.crqsize = cpu_to_le16(queue->recv_queue_size);
1219 ret = rdma_accept(cm_id, ¶m);
1221 pr_err("rdma_accept failed (error code = %d)\n", ret);
1226 static int nvmet_rdma_queue_connect(struct rdma_cm_id *cm_id,
1227 struct rdma_cm_event *event)
1229 struct nvmet_rdma_device *ndev;
1230 struct nvmet_rdma_queue *queue;
1233 ndev = nvmet_rdma_find_get_device(cm_id);
1235 nvmet_rdma_cm_reject(cm_id, NVME_RDMA_CM_NO_RSC);
1236 return -ECONNREFUSED;
1239 queue = nvmet_rdma_alloc_queue(ndev, cm_id, event);
1244 queue->port = cm_id->context;
1246 if (queue->host_qid == 0) {
1247 /* Let inflight controller teardown complete */
1248 flush_scheduled_work();
1251 ret = nvmet_rdma_cm_accept(cm_id, queue, &event->param.conn);
1253 schedule_work(&queue->release_work);
1254 /* Destroying rdma_cm id is not needed here */
1258 mutex_lock(&nvmet_rdma_queue_mutex);
1259 list_add_tail(&queue->queue_list, &nvmet_rdma_queue_list);
1260 mutex_unlock(&nvmet_rdma_queue_mutex);
1265 kref_put(&ndev->ref, nvmet_rdma_free_dev);
1270 static void nvmet_rdma_queue_established(struct nvmet_rdma_queue *queue)
1272 unsigned long flags;
1274 spin_lock_irqsave(&queue->state_lock, flags);
1275 if (queue->state != NVMET_RDMA_Q_CONNECTING) {
1276 pr_warn("trying to establish a connected queue\n");
1279 queue->state = NVMET_RDMA_Q_LIVE;
1281 while (!list_empty(&queue->rsp_wait_list)) {
1282 struct nvmet_rdma_rsp *cmd;
1284 cmd = list_first_entry(&queue->rsp_wait_list,
1285 struct nvmet_rdma_rsp, wait_list);
1286 list_del(&cmd->wait_list);
1288 spin_unlock_irqrestore(&queue->state_lock, flags);
1289 nvmet_rdma_handle_command(queue, cmd);
1290 spin_lock_irqsave(&queue->state_lock, flags);
1294 spin_unlock_irqrestore(&queue->state_lock, flags);
1297 static void __nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
1299 bool disconnect = false;
1300 unsigned long flags;
1302 pr_debug("cm_id= %p queue->state= %d\n", queue->cm_id, queue->state);
1304 spin_lock_irqsave(&queue->state_lock, flags);
1305 switch (queue->state) {
1306 case NVMET_RDMA_Q_CONNECTING:
1307 case NVMET_RDMA_Q_LIVE:
1308 queue->state = NVMET_RDMA_Q_DISCONNECTING;
1311 case NVMET_RDMA_Q_DISCONNECTING:
1314 spin_unlock_irqrestore(&queue->state_lock, flags);
1317 rdma_disconnect(queue->cm_id);
1318 schedule_work(&queue->release_work);
1322 static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
1324 bool disconnect = false;
1326 mutex_lock(&nvmet_rdma_queue_mutex);
1327 if (!list_empty(&queue->queue_list)) {
1328 list_del_init(&queue->queue_list);
1331 mutex_unlock(&nvmet_rdma_queue_mutex);
1334 __nvmet_rdma_queue_disconnect(queue);
1337 static void nvmet_rdma_queue_connect_fail(struct rdma_cm_id *cm_id,
1338 struct nvmet_rdma_queue *queue)
1340 WARN_ON_ONCE(queue->state != NVMET_RDMA_Q_CONNECTING);
1342 mutex_lock(&nvmet_rdma_queue_mutex);
1343 if (!list_empty(&queue->queue_list))
1344 list_del_init(&queue->queue_list);
1345 mutex_unlock(&nvmet_rdma_queue_mutex);
1347 pr_err("failed to connect queue %d\n", queue->idx);
1348 schedule_work(&queue->release_work);
1352 * nvme_rdma_device_removal() - Handle RDMA device removal
1353 * @cm_id: rdma_cm id, used for nvmet port
1354 * @queue: nvmet rdma queue (cm id qp_context)
1356 * DEVICE_REMOVAL event notifies us that the RDMA device is about
1357 * to unplug. Note that this event can be generated on a normal
1358 * queue cm_id and/or a device bound listener cm_id (where in this
1359 * case queue will be null).
1361 * We registered an ib_client to handle device removal for queues,
1362 * so we only need to handle the listening port cm_ids. In this case
1363 * we nullify the priv to prevent double cm_id destruction and destroying
1364 * the cm_id implicitely by returning a non-zero rc to the callout.
1366 static int nvmet_rdma_device_removal(struct rdma_cm_id *cm_id,
1367 struct nvmet_rdma_queue *queue)
1369 struct nvmet_port *port;
1373 * This is a queue cm_id. we have registered
1374 * an ib_client to handle queues removal
1375 * so don't interfear and just return.
1380 port = cm_id->context;
1383 * This is a listener cm_id. Make sure that
1384 * future remove_port won't invoke a double
1385 * cm_id destroy. use atomic xchg to make sure
1386 * we don't compete with remove_port.
1388 if (xchg(&port->priv, NULL) != cm_id)
1392 * We need to return 1 so that the core will destroy
1393 * it's own ID. What a great API design..
1398 static int nvmet_rdma_cm_handler(struct rdma_cm_id *cm_id,
1399 struct rdma_cm_event *event)
1401 struct nvmet_rdma_queue *queue = NULL;
1405 queue = cm_id->qp->qp_context;
1407 pr_debug("%s (%d): status %d id %p\n",
1408 rdma_event_msg(event->event), event->event,
1409 event->status, cm_id);
1411 switch (event->event) {
1412 case RDMA_CM_EVENT_CONNECT_REQUEST:
1413 ret = nvmet_rdma_queue_connect(cm_id, event);
1415 case RDMA_CM_EVENT_ESTABLISHED:
1416 nvmet_rdma_queue_established(queue);
1418 case RDMA_CM_EVENT_ADDR_CHANGE:
1419 case RDMA_CM_EVENT_DISCONNECTED:
1420 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1421 nvmet_rdma_queue_disconnect(queue);
1423 case RDMA_CM_EVENT_DEVICE_REMOVAL:
1424 ret = nvmet_rdma_device_removal(cm_id, queue);
1426 case RDMA_CM_EVENT_REJECTED:
1427 pr_debug("Connection rejected: %s\n",
1428 rdma_reject_msg(cm_id, event->status));
1430 case RDMA_CM_EVENT_UNREACHABLE:
1431 case RDMA_CM_EVENT_CONNECT_ERROR:
1432 nvmet_rdma_queue_connect_fail(cm_id, queue);
1435 pr_err("received unrecognized RDMA CM event %d\n",
1443 static void nvmet_rdma_delete_ctrl(struct nvmet_ctrl *ctrl)
1445 struct nvmet_rdma_queue *queue;
1448 mutex_lock(&nvmet_rdma_queue_mutex);
1449 list_for_each_entry(queue, &nvmet_rdma_queue_list, queue_list) {
1450 if (queue->nvme_sq.ctrl == ctrl) {
1451 list_del_init(&queue->queue_list);
1452 mutex_unlock(&nvmet_rdma_queue_mutex);
1454 __nvmet_rdma_queue_disconnect(queue);
1458 mutex_unlock(&nvmet_rdma_queue_mutex);
1461 static int nvmet_rdma_add_port(struct nvmet_port *port)
1463 struct rdma_cm_id *cm_id;
1464 struct sockaddr_storage addr = { };
1465 __kernel_sa_family_t af;
1468 switch (port->disc_addr.adrfam) {
1469 case NVMF_ADDR_FAMILY_IP4:
1472 case NVMF_ADDR_FAMILY_IP6:
1476 pr_err("address family %d not supported\n",
1477 port->disc_addr.adrfam);
1481 if (port->inline_data_size < 0) {
1482 port->inline_data_size = NVMET_RDMA_DEFAULT_INLINE_DATA_SIZE;
1483 } else if (port->inline_data_size > NVMET_RDMA_MAX_INLINE_DATA_SIZE) {
1484 pr_warn("inline_data_size %u is too large, reducing to %u\n",
1485 port->inline_data_size,
1486 NVMET_RDMA_MAX_INLINE_DATA_SIZE);
1487 port->inline_data_size = NVMET_RDMA_MAX_INLINE_DATA_SIZE;
1490 ret = inet_pton_with_scope(&init_net, af, port->disc_addr.traddr,
1491 port->disc_addr.trsvcid, &addr);
1493 pr_err("malformed ip/port passed: %s:%s\n",
1494 port->disc_addr.traddr, port->disc_addr.trsvcid);
1498 cm_id = rdma_create_id(&init_net, nvmet_rdma_cm_handler, port,
1499 RDMA_PS_TCP, IB_QPT_RC);
1500 if (IS_ERR(cm_id)) {
1501 pr_err("CM ID creation failed\n");
1502 return PTR_ERR(cm_id);
1506 * Allow both IPv4 and IPv6 sockets to bind a single port
1509 ret = rdma_set_afonly(cm_id, 1);
1511 pr_err("rdma_set_afonly failed (%d)\n", ret);
1512 goto out_destroy_id;
1515 ret = rdma_bind_addr(cm_id, (struct sockaddr *)&addr);
1517 pr_err("binding CM ID to %pISpcs failed (%d)\n",
1518 (struct sockaddr *)&addr, ret);
1519 goto out_destroy_id;
1522 ret = rdma_listen(cm_id, 128);
1524 pr_err("listening to %pISpcs failed (%d)\n",
1525 (struct sockaddr *)&addr, ret);
1526 goto out_destroy_id;
1529 pr_info("enabling port %d (%pISpcs)\n",
1530 le16_to_cpu(port->disc_addr.portid), (struct sockaddr *)&addr);
1535 rdma_destroy_id(cm_id);
1539 static void nvmet_rdma_remove_port(struct nvmet_port *port)
1541 struct rdma_cm_id *cm_id = xchg(&port->priv, NULL);
1544 rdma_destroy_id(cm_id);
1547 static void nvmet_rdma_disc_port_addr(struct nvmet_req *req,
1548 struct nvmet_port *port, char *traddr)
1550 struct rdma_cm_id *cm_id = port->priv;
1552 if (inet_addr_is_any((struct sockaddr *)&cm_id->route.addr.src_addr)) {
1553 struct nvmet_rdma_rsp *rsp =
1554 container_of(req, struct nvmet_rdma_rsp, req);
1555 struct rdma_cm_id *req_cm_id = rsp->queue->cm_id;
1556 struct sockaddr *addr = (void *)&req_cm_id->route.addr.src_addr;
1558 sprintf(traddr, "%pISc", addr);
1560 memcpy(traddr, port->disc_addr.traddr, NVMF_TRADDR_SIZE);
1564 static const struct nvmet_fabrics_ops nvmet_rdma_ops = {
1565 .owner = THIS_MODULE,
1566 .type = NVMF_TRTYPE_RDMA,
1568 .has_keyed_sgls = 1,
1569 .add_port = nvmet_rdma_add_port,
1570 .remove_port = nvmet_rdma_remove_port,
1571 .queue_response = nvmet_rdma_queue_response,
1572 .delete_ctrl = nvmet_rdma_delete_ctrl,
1573 .disc_traddr = nvmet_rdma_disc_port_addr,
1576 static void nvmet_rdma_remove_one(struct ib_device *ib_device, void *client_data)
1578 struct nvmet_rdma_queue *queue, *tmp;
1579 struct nvmet_rdma_device *ndev;
1582 mutex_lock(&device_list_mutex);
1583 list_for_each_entry(ndev, &device_list, entry) {
1584 if (ndev->device == ib_device) {
1589 mutex_unlock(&device_list_mutex);
1595 * IB Device that is used by nvmet controllers is being removed,
1596 * delete all queues using this device.
1598 mutex_lock(&nvmet_rdma_queue_mutex);
1599 list_for_each_entry_safe(queue, tmp, &nvmet_rdma_queue_list,
1601 if (queue->dev->device != ib_device)
1604 pr_info("Removing queue %d\n", queue->idx);
1605 list_del_init(&queue->queue_list);
1606 __nvmet_rdma_queue_disconnect(queue);
1608 mutex_unlock(&nvmet_rdma_queue_mutex);
1610 flush_scheduled_work();
1613 static struct ib_client nvmet_rdma_ib_client = {
1614 .name = "nvmet_rdma",
1615 .remove = nvmet_rdma_remove_one
1618 static int __init nvmet_rdma_init(void)
1622 ret = ib_register_client(&nvmet_rdma_ib_client);
1626 ret = nvmet_register_transport(&nvmet_rdma_ops);
1633 ib_unregister_client(&nvmet_rdma_ib_client);
1637 static void __exit nvmet_rdma_exit(void)
1639 nvmet_unregister_transport(&nvmet_rdma_ops);
1640 ib_unregister_client(&nvmet_rdma_ib_client);
1641 WARN_ON_ONCE(!list_empty(&nvmet_rdma_queue_list));
1642 ida_destroy(&nvmet_rdma_queue_ida);
1645 module_init(nvmet_rdma_init);
1646 module_exit(nvmet_rdma_exit);
1648 MODULE_LICENSE("GPL v2");
1649 MODULE_ALIAS("nvmet-transport-1"); /* 1 == NVMF_TRTYPE_RDMA */