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;
74 struct list_head wait_list;
75 struct list_head free_list;
78 enum nvmet_rdma_queue_state {
79 NVMET_RDMA_Q_CONNECTING,
81 NVMET_RDMA_Q_DISCONNECTING,
84 struct nvmet_rdma_queue {
85 struct rdma_cm_id *cm_id;
86 struct nvmet_port *port;
89 struct nvmet_rdma_device *dev;
90 spinlock_t state_lock;
91 enum nvmet_rdma_queue_state state;
92 struct nvmet_cq nvme_cq;
93 struct nvmet_sq nvme_sq;
95 struct nvmet_rdma_rsp *rsps;
96 struct list_head free_rsps;
98 struct nvmet_rdma_cmd *cmds;
100 struct work_struct release_work;
101 struct list_head rsp_wait_list;
102 struct list_head rsp_wr_wait_list;
103 spinlock_t rsp_wr_wait_lock;
110 struct list_head queue_list;
113 struct nvmet_rdma_device {
114 struct ib_device *device;
117 struct nvmet_rdma_cmd *srq_cmds;
120 struct list_head entry;
121 int inline_data_size;
122 int inline_page_count;
125 static bool nvmet_rdma_use_srq;
126 module_param_named(use_srq, nvmet_rdma_use_srq, bool, 0444);
127 MODULE_PARM_DESC(use_srq, "Use shared receive queue.");
129 static DEFINE_IDA(nvmet_rdma_queue_ida);
130 static LIST_HEAD(nvmet_rdma_queue_list);
131 static DEFINE_MUTEX(nvmet_rdma_queue_mutex);
133 static LIST_HEAD(device_list);
134 static DEFINE_MUTEX(device_list_mutex);
136 static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp);
137 static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc);
138 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
139 static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc);
140 static void nvmet_rdma_qp_event(struct ib_event *event, void *priv);
141 static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue);
143 static const struct nvmet_fabrics_ops nvmet_rdma_ops;
145 static int num_pages(int len)
147 return 1 + (((len - 1) & PAGE_MASK) >> PAGE_SHIFT);
150 /* XXX: really should move to a generic header sooner or later.. */
151 static inline u32 get_unaligned_le24(const u8 *p)
153 return (u32)p[0] | (u32)p[1] << 8 | (u32)p[2] << 16;
156 static inline bool nvmet_rdma_need_data_in(struct nvmet_rdma_rsp *rsp)
158 return nvme_is_write(rsp->req.cmd) &&
159 rsp->req.transfer_len &&
160 !(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
163 static inline bool nvmet_rdma_need_data_out(struct nvmet_rdma_rsp *rsp)
165 return !nvme_is_write(rsp->req.cmd) &&
166 rsp->req.transfer_len &&
167 !rsp->req.rsp->status &&
168 !(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
171 static inline struct nvmet_rdma_rsp *
172 nvmet_rdma_get_rsp(struct nvmet_rdma_queue *queue)
174 struct nvmet_rdma_rsp *rsp;
177 spin_lock_irqsave(&queue->rsps_lock, flags);
178 rsp = list_first_entry_or_null(&queue->free_rsps,
179 struct nvmet_rdma_rsp, free_list);
181 list_del(&rsp->free_list);
182 spin_unlock_irqrestore(&queue->rsps_lock, flags);
184 if (unlikely(!rsp)) {
185 rsp = kmalloc(sizeof(*rsp), GFP_KERNEL);
188 rsp->allocated = true;
195 nvmet_rdma_put_rsp(struct nvmet_rdma_rsp *rsp)
199 if (rsp->allocated) {
204 spin_lock_irqsave(&rsp->queue->rsps_lock, flags);
205 list_add_tail(&rsp->free_list, &rsp->queue->free_rsps);
206 spin_unlock_irqrestore(&rsp->queue->rsps_lock, flags);
209 static void nvmet_rdma_free_inline_pages(struct nvmet_rdma_device *ndev,
210 struct nvmet_rdma_cmd *c)
212 struct scatterlist *sg;
216 if (!ndev->inline_data_size)
222 for (i = 0; i < ndev->inline_page_count; i++, sg++, sge++) {
224 ib_dma_unmap_page(ndev->device, sge->addr,
225 sge->length, DMA_FROM_DEVICE);
227 __free_page(sg_page(sg));
231 static int nvmet_rdma_alloc_inline_pages(struct nvmet_rdma_device *ndev,
232 struct nvmet_rdma_cmd *c)
234 struct scatterlist *sg;
240 if (!ndev->inline_data_size)
244 sg_init_table(sg, ndev->inline_page_count);
246 len = ndev->inline_data_size;
248 for (i = 0; i < ndev->inline_page_count; i++, sg++, sge++) {
249 pg = alloc_page(GFP_KERNEL);
252 sg_assign_page(sg, pg);
253 sge->addr = ib_dma_map_page(ndev->device,
254 pg, 0, PAGE_SIZE, DMA_FROM_DEVICE);
255 if (ib_dma_mapping_error(ndev->device, sge->addr))
257 sge->length = min_t(int, len, PAGE_SIZE);
258 sge->lkey = ndev->pd->local_dma_lkey;
264 for (; i >= 0; i--, sg--, sge--) {
266 ib_dma_unmap_page(ndev->device, sge->addr,
267 sge->length, DMA_FROM_DEVICE);
269 __free_page(sg_page(sg));
274 static int nvmet_rdma_alloc_cmd(struct nvmet_rdma_device *ndev,
275 struct nvmet_rdma_cmd *c, bool admin)
277 /* NVMe command / RDMA RECV */
278 c->nvme_cmd = kmalloc(sizeof(*c->nvme_cmd), GFP_KERNEL);
282 c->sge[0].addr = ib_dma_map_single(ndev->device, c->nvme_cmd,
283 sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
284 if (ib_dma_mapping_error(ndev->device, c->sge[0].addr))
287 c->sge[0].length = sizeof(*c->nvme_cmd);
288 c->sge[0].lkey = ndev->pd->local_dma_lkey;
290 if (!admin && nvmet_rdma_alloc_inline_pages(ndev, c))
293 c->cqe.done = nvmet_rdma_recv_done;
295 c->wr.wr_cqe = &c->cqe;
296 c->wr.sg_list = c->sge;
297 c->wr.num_sge = admin ? 1 : ndev->inline_page_count + 1;
302 ib_dma_unmap_single(ndev->device, c->sge[0].addr,
303 sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
311 static void nvmet_rdma_free_cmd(struct nvmet_rdma_device *ndev,
312 struct nvmet_rdma_cmd *c, bool admin)
315 nvmet_rdma_free_inline_pages(ndev, c);
316 ib_dma_unmap_single(ndev->device, c->sge[0].addr,
317 sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
321 static struct nvmet_rdma_cmd *
322 nvmet_rdma_alloc_cmds(struct nvmet_rdma_device *ndev,
323 int nr_cmds, bool admin)
325 struct nvmet_rdma_cmd *cmds;
326 int ret = -EINVAL, i;
328 cmds = kcalloc(nr_cmds, sizeof(struct nvmet_rdma_cmd), GFP_KERNEL);
332 for (i = 0; i < nr_cmds; i++) {
333 ret = nvmet_rdma_alloc_cmd(ndev, cmds + i, admin);
342 nvmet_rdma_free_cmd(ndev, cmds + i, admin);
348 static void nvmet_rdma_free_cmds(struct nvmet_rdma_device *ndev,
349 struct nvmet_rdma_cmd *cmds, int nr_cmds, bool admin)
353 for (i = 0; i < nr_cmds; i++)
354 nvmet_rdma_free_cmd(ndev, cmds + i, admin);
358 static int nvmet_rdma_alloc_rsp(struct nvmet_rdma_device *ndev,
359 struct nvmet_rdma_rsp *r)
361 /* NVMe CQE / RDMA SEND */
362 r->req.rsp = kmalloc(sizeof(*r->req.rsp), GFP_KERNEL);
366 r->send_sge.addr = ib_dma_map_single(ndev->device, r->req.rsp,
367 sizeof(*r->req.rsp), DMA_TO_DEVICE);
368 if (ib_dma_mapping_error(ndev->device, r->send_sge.addr))
371 r->send_sge.length = sizeof(*r->req.rsp);
372 r->send_sge.lkey = ndev->pd->local_dma_lkey;
374 r->send_cqe.done = nvmet_rdma_send_done;
376 r->send_wr.wr_cqe = &r->send_cqe;
377 r->send_wr.sg_list = &r->send_sge;
378 r->send_wr.num_sge = 1;
379 r->send_wr.send_flags = IB_SEND_SIGNALED;
381 /* Data In / RDMA READ */
382 r->read_cqe.done = nvmet_rdma_read_data_done;
391 static void nvmet_rdma_free_rsp(struct nvmet_rdma_device *ndev,
392 struct nvmet_rdma_rsp *r)
394 ib_dma_unmap_single(ndev->device, r->send_sge.addr,
395 sizeof(*r->req.rsp), DMA_TO_DEVICE);
400 nvmet_rdma_alloc_rsps(struct nvmet_rdma_queue *queue)
402 struct nvmet_rdma_device *ndev = queue->dev;
403 int nr_rsps = queue->recv_queue_size * 2;
404 int ret = -EINVAL, i;
406 queue->rsps = kcalloc(nr_rsps, sizeof(struct nvmet_rdma_rsp),
411 for (i = 0; i < nr_rsps; i++) {
412 struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
414 ret = nvmet_rdma_alloc_rsp(ndev, rsp);
418 list_add_tail(&rsp->free_list, &queue->free_rsps);
425 struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
427 list_del(&rsp->free_list);
428 nvmet_rdma_free_rsp(ndev, rsp);
435 static void nvmet_rdma_free_rsps(struct nvmet_rdma_queue *queue)
437 struct nvmet_rdma_device *ndev = queue->dev;
438 int i, nr_rsps = queue->recv_queue_size * 2;
440 for (i = 0; i < nr_rsps; i++) {
441 struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
443 list_del(&rsp->free_list);
444 nvmet_rdma_free_rsp(ndev, rsp);
449 static int nvmet_rdma_post_recv(struct nvmet_rdma_device *ndev,
450 struct nvmet_rdma_cmd *cmd)
454 ib_dma_sync_single_for_device(ndev->device,
455 cmd->sge[0].addr, cmd->sge[0].length,
459 ret = ib_post_srq_recv(ndev->srq, &cmd->wr, NULL);
461 ret = ib_post_recv(cmd->queue->cm_id->qp, &cmd->wr, NULL);
464 pr_err("post_recv cmd failed\n");
469 static void nvmet_rdma_process_wr_wait_list(struct nvmet_rdma_queue *queue)
471 spin_lock(&queue->rsp_wr_wait_lock);
472 while (!list_empty(&queue->rsp_wr_wait_list)) {
473 struct nvmet_rdma_rsp *rsp;
476 rsp = list_entry(queue->rsp_wr_wait_list.next,
477 struct nvmet_rdma_rsp, wait_list);
478 list_del(&rsp->wait_list);
480 spin_unlock(&queue->rsp_wr_wait_lock);
481 ret = nvmet_rdma_execute_command(rsp);
482 spin_lock(&queue->rsp_wr_wait_lock);
485 list_add(&rsp->wait_list, &queue->rsp_wr_wait_list);
489 spin_unlock(&queue->rsp_wr_wait_lock);
493 static void nvmet_rdma_release_rsp(struct nvmet_rdma_rsp *rsp)
495 struct nvmet_rdma_queue *queue = rsp->queue;
497 atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
500 rdma_rw_ctx_destroy(&rsp->rw, queue->cm_id->qp,
501 queue->cm_id->port_num, rsp->req.sg,
502 rsp->req.sg_cnt, nvmet_data_dir(&rsp->req));
505 if (rsp->req.sg != rsp->cmd->inline_sg)
506 nvmet_req_free_sgl(&rsp->req);
508 if (unlikely(!list_empty_careful(&queue->rsp_wr_wait_list)))
509 nvmet_rdma_process_wr_wait_list(queue);
511 nvmet_rdma_put_rsp(rsp);
514 static void nvmet_rdma_error_comp(struct nvmet_rdma_queue *queue)
516 if (queue->nvme_sq.ctrl) {
517 nvmet_ctrl_fatal_error(queue->nvme_sq.ctrl);
520 * we didn't setup the controller yet in case
521 * of admin connect error, just disconnect and
524 nvmet_rdma_queue_disconnect(queue);
528 static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
530 struct nvmet_rdma_rsp *rsp =
531 container_of(wc->wr_cqe, struct nvmet_rdma_rsp, send_cqe);
533 nvmet_rdma_release_rsp(rsp);
535 if (unlikely(wc->status != IB_WC_SUCCESS &&
536 wc->status != IB_WC_WR_FLUSH_ERR)) {
537 pr_err("SEND for CQE 0x%p failed with status %s (%d).\n",
538 wc->wr_cqe, ib_wc_status_msg(wc->status), wc->status);
539 nvmet_rdma_error_comp(rsp->queue);
543 static void nvmet_rdma_queue_response(struct nvmet_req *req)
545 struct nvmet_rdma_rsp *rsp =
546 container_of(req, struct nvmet_rdma_rsp, req);
547 struct rdma_cm_id *cm_id = rsp->queue->cm_id;
548 struct ib_send_wr *first_wr;
550 if (rsp->flags & NVMET_RDMA_REQ_INVALIDATE_RKEY) {
551 rsp->send_wr.opcode = IB_WR_SEND_WITH_INV;
552 rsp->send_wr.ex.invalidate_rkey = rsp->invalidate_rkey;
554 rsp->send_wr.opcode = IB_WR_SEND;
557 if (nvmet_rdma_need_data_out(rsp))
558 first_wr = rdma_rw_ctx_wrs(&rsp->rw, cm_id->qp,
559 cm_id->port_num, NULL, &rsp->send_wr);
561 first_wr = &rsp->send_wr;
563 nvmet_rdma_post_recv(rsp->queue->dev, rsp->cmd);
565 ib_dma_sync_single_for_device(rsp->queue->dev->device,
566 rsp->send_sge.addr, rsp->send_sge.length,
569 if (unlikely(ib_post_send(cm_id->qp, first_wr, NULL))) {
570 pr_err("sending cmd response failed\n");
571 nvmet_rdma_release_rsp(rsp);
575 static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc)
577 struct nvmet_rdma_rsp *rsp =
578 container_of(wc->wr_cqe, struct nvmet_rdma_rsp, read_cqe);
579 struct nvmet_rdma_queue *queue = cq->cq_context;
581 WARN_ON(rsp->n_rdma <= 0);
582 atomic_add(rsp->n_rdma, &queue->sq_wr_avail);
583 rdma_rw_ctx_destroy(&rsp->rw, queue->cm_id->qp,
584 queue->cm_id->port_num, rsp->req.sg,
585 rsp->req.sg_cnt, nvmet_data_dir(&rsp->req));
588 if (unlikely(wc->status != IB_WC_SUCCESS)) {
589 nvmet_req_uninit(&rsp->req);
590 nvmet_rdma_release_rsp(rsp);
591 if (wc->status != IB_WC_WR_FLUSH_ERR) {
592 pr_info("RDMA READ for CQE 0x%p failed with status %s (%d).\n",
593 wc->wr_cqe, ib_wc_status_msg(wc->status), wc->status);
594 nvmet_rdma_error_comp(queue);
599 nvmet_req_execute(&rsp->req);
602 static void nvmet_rdma_use_inline_sg(struct nvmet_rdma_rsp *rsp, u32 len,
605 int sg_count = num_pages(len);
606 struct scatterlist *sg;
609 sg = rsp->cmd->inline_sg;
610 for (i = 0; i < sg_count; i++, sg++) {
611 if (i < sg_count - 1)
616 sg->length = min_t(int, len, PAGE_SIZE - off);
622 rsp->req.sg = rsp->cmd->inline_sg;
623 rsp->req.sg_cnt = sg_count;
626 static u16 nvmet_rdma_map_sgl_inline(struct nvmet_rdma_rsp *rsp)
628 struct nvme_sgl_desc *sgl = &rsp->req.cmd->common.dptr.sgl;
629 u64 off = le64_to_cpu(sgl->addr);
630 u32 len = le32_to_cpu(sgl->length);
632 if (!nvme_is_write(rsp->req.cmd))
633 return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
635 if (off + len > rsp->queue->dev->inline_data_size) {
636 pr_err("invalid inline data offset!\n");
637 return NVME_SC_SGL_INVALID_OFFSET | NVME_SC_DNR;
640 /* no data command? */
644 nvmet_rdma_use_inline_sg(rsp, len, off);
645 rsp->flags |= NVMET_RDMA_REQ_INLINE_DATA;
646 rsp->req.transfer_len += len;
650 static u16 nvmet_rdma_map_sgl_keyed(struct nvmet_rdma_rsp *rsp,
651 struct nvme_keyed_sgl_desc *sgl, bool invalidate)
653 struct rdma_cm_id *cm_id = rsp->queue->cm_id;
654 u64 addr = le64_to_cpu(sgl->addr);
655 u32 key = get_unaligned_le32(sgl->key);
658 rsp->req.transfer_len = get_unaligned_le24(sgl->length);
660 /* no data command? */
661 if (!rsp->req.transfer_len)
664 ret = nvmet_req_alloc_sgl(&rsp->req);
668 ret = rdma_rw_ctx_init(&rsp->rw, cm_id->qp, cm_id->port_num,
669 rsp->req.sg, rsp->req.sg_cnt, 0, addr, key,
670 nvmet_data_dir(&rsp->req));
676 rsp->invalidate_rkey = key;
677 rsp->flags |= NVMET_RDMA_REQ_INVALIDATE_RKEY;
683 rsp->req.transfer_len = 0;
684 return NVME_SC_INTERNAL;
687 static u16 nvmet_rdma_map_sgl(struct nvmet_rdma_rsp *rsp)
689 struct nvme_keyed_sgl_desc *sgl = &rsp->req.cmd->common.dptr.ksgl;
691 switch (sgl->type >> 4) {
692 case NVME_SGL_FMT_DATA_DESC:
693 switch (sgl->type & 0xf) {
694 case NVME_SGL_FMT_OFFSET:
695 return nvmet_rdma_map_sgl_inline(rsp);
697 pr_err("invalid SGL subtype: %#x\n", sgl->type);
698 return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
700 case NVME_KEY_SGL_FMT_DATA_DESC:
701 switch (sgl->type & 0xf) {
702 case NVME_SGL_FMT_ADDRESS | NVME_SGL_FMT_INVALIDATE:
703 return nvmet_rdma_map_sgl_keyed(rsp, sgl, true);
704 case NVME_SGL_FMT_ADDRESS:
705 return nvmet_rdma_map_sgl_keyed(rsp, sgl, false);
707 pr_err("invalid SGL subtype: %#x\n", sgl->type);
708 return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
711 pr_err("invalid SGL type: %#x\n", sgl->type);
712 return NVME_SC_SGL_INVALID_TYPE | NVME_SC_DNR;
716 static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp)
718 struct nvmet_rdma_queue *queue = rsp->queue;
720 if (unlikely(atomic_sub_return(1 + rsp->n_rdma,
721 &queue->sq_wr_avail) < 0)) {
722 pr_debug("IB send queue full (needed %d): queue %u cntlid %u\n",
723 1 + rsp->n_rdma, queue->idx,
724 queue->nvme_sq.ctrl->cntlid);
725 atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
729 if (nvmet_rdma_need_data_in(rsp)) {
730 if (rdma_rw_ctx_post(&rsp->rw, queue->cm_id->qp,
731 queue->cm_id->port_num, &rsp->read_cqe, NULL))
732 nvmet_req_complete(&rsp->req, NVME_SC_DATA_XFER_ERROR);
734 nvmet_req_execute(&rsp->req);
740 static void nvmet_rdma_handle_command(struct nvmet_rdma_queue *queue,
741 struct nvmet_rdma_rsp *cmd)
745 ib_dma_sync_single_for_cpu(queue->dev->device,
746 cmd->cmd->sge[0].addr, cmd->cmd->sge[0].length,
748 ib_dma_sync_single_for_cpu(queue->dev->device,
749 cmd->send_sge.addr, cmd->send_sge.length,
752 cmd->req.p2p_client = &queue->dev->device->dev;
754 if (!nvmet_req_init(&cmd->req, &queue->nvme_cq,
755 &queue->nvme_sq, &nvmet_rdma_ops))
758 status = nvmet_rdma_map_sgl(cmd);
762 if (unlikely(!nvmet_rdma_execute_command(cmd))) {
763 spin_lock(&queue->rsp_wr_wait_lock);
764 list_add_tail(&cmd->wait_list, &queue->rsp_wr_wait_list);
765 spin_unlock(&queue->rsp_wr_wait_lock);
771 nvmet_req_complete(&cmd->req, status);
774 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
776 struct nvmet_rdma_cmd *cmd =
777 container_of(wc->wr_cqe, struct nvmet_rdma_cmd, cqe);
778 struct nvmet_rdma_queue *queue = cq->cq_context;
779 struct nvmet_rdma_rsp *rsp;
781 if (unlikely(wc->status != IB_WC_SUCCESS)) {
782 if (wc->status != IB_WC_WR_FLUSH_ERR) {
783 pr_err("RECV for CQE 0x%p failed with status %s (%d)\n",
784 wc->wr_cqe, ib_wc_status_msg(wc->status),
786 nvmet_rdma_error_comp(queue);
791 if (unlikely(wc->byte_len < sizeof(struct nvme_command))) {
792 pr_err("Ctrl Fatal Error: capsule size less than 64 bytes\n");
793 nvmet_rdma_error_comp(queue);
798 rsp = nvmet_rdma_get_rsp(queue);
799 if (unlikely(!rsp)) {
801 * we get here only under memory pressure,
802 * silently drop and have the host retry
803 * as we can't even fail it.
805 nvmet_rdma_post_recv(queue->dev, cmd);
811 rsp->req.cmd = cmd->nvme_cmd;
812 rsp->req.port = queue->port;
815 if (unlikely(queue->state != NVMET_RDMA_Q_LIVE)) {
818 spin_lock_irqsave(&queue->state_lock, flags);
819 if (queue->state == NVMET_RDMA_Q_CONNECTING)
820 list_add_tail(&rsp->wait_list, &queue->rsp_wait_list);
822 nvmet_rdma_put_rsp(rsp);
823 spin_unlock_irqrestore(&queue->state_lock, flags);
827 nvmet_rdma_handle_command(queue, rsp);
830 static void nvmet_rdma_destroy_srq(struct nvmet_rdma_device *ndev)
835 nvmet_rdma_free_cmds(ndev, ndev->srq_cmds, ndev->srq_size, false);
836 ib_destroy_srq(ndev->srq);
839 static int nvmet_rdma_init_srq(struct nvmet_rdma_device *ndev)
841 struct ib_srq_init_attr srq_attr = { NULL, };
846 srq_size = 4095; /* XXX: tune */
848 srq_attr.attr.max_wr = srq_size;
849 srq_attr.attr.max_sge = 1 + ndev->inline_page_count;
850 srq_attr.attr.srq_limit = 0;
851 srq_attr.srq_type = IB_SRQT_BASIC;
852 srq = ib_create_srq(ndev->pd, &srq_attr);
855 * If SRQs aren't supported we just go ahead and use normal
856 * non-shared receive queues.
858 pr_info("SRQ requested but not supported.\n");
862 ndev->srq_cmds = nvmet_rdma_alloc_cmds(ndev, srq_size, false);
863 if (IS_ERR(ndev->srq_cmds)) {
864 ret = PTR_ERR(ndev->srq_cmds);
865 goto out_destroy_srq;
869 ndev->srq_size = srq_size;
871 for (i = 0; i < srq_size; i++) {
872 ret = nvmet_rdma_post_recv(ndev, &ndev->srq_cmds[i]);
880 nvmet_rdma_free_cmds(ndev, ndev->srq_cmds, ndev->srq_size, false);
886 static void nvmet_rdma_free_dev(struct kref *ref)
888 struct nvmet_rdma_device *ndev =
889 container_of(ref, struct nvmet_rdma_device, ref);
891 mutex_lock(&device_list_mutex);
892 list_del(&ndev->entry);
893 mutex_unlock(&device_list_mutex);
895 nvmet_rdma_destroy_srq(ndev);
896 ib_dealloc_pd(ndev->pd);
901 static struct nvmet_rdma_device *
902 nvmet_rdma_find_get_device(struct rdma_cm_id *cm_id)
904 struct nvmet_port *port = cm_id->context;
905 struct nvmet_rdma_device *ndev;
906 int inline_page_count;
907 int inline_sge_count;
910 mutex_lock(&device_list_mutex);
911 list_for_each_entry(ndev, &device_list, entry) {
912 if (ndev->device->node_guid == cm_id->device->node_guid &&
913 kref_get_unless_zero(&ndev->ref))
917 ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
921 inline_page_count = num_pages(port->inline_data_size);
922 inline_sge_count = max(cm_id->device->attrs.max_sge_rd,
923 cm_id->device->attrs.max_recv_sge) - 1;
924 if (inline_page_count > inline_sge_count) {
925 pr_warn("inline_data_size %d cannot be supported by device %s. Reducing to %lu.\n",
926 port->inline_data_size, cm_id->device->name,
927 inline_sge_count * PAGE_SIZE);
928 port->inline_data_size = inline_sge_count * PAGE_SIZE;
929 inline_page_count = inline_sge_count;
931 ndev->inline_data_size = port->inline_data_size;
932 ndev->inline_page_count = inline_page_count;
933 ndev->device = cm_id->device;
934 kref_init(&ndev->ref);
936 ndev->pd = ib_alloc_pd(ndev->device, 0);
937 if (IS_ERR(ndev->pd))
940 if (nvmet_rdma_use_srq) {
941 ret = nvmet_rdma_init_srq(ndev);
946 list_add(&ndev->entry, &device_list);
948 mutex_unlock(&device_list_mutex);
949 pr_debug("added %s.\n", ndev->device->name);
953 ib_dealloc_pd(ndev->pd);
957 mutex_unlock(&device_list_mutex);
961 static int nvmet_rdma_create_queue_ib(struct nvmet_rdma_queue *queue)
963 struct ib_qp_init_attr qp_attr;
964 struct nvmet_rdma_device *ndev = queue->dev;
965 int comp_vector, nr_cqe, ret, i;
968 * Spread the io queues across completion vectors,
969 * but still keep all admin queues on vector 0.
971 comp_vector = !queue->host_qid ? 0 :
972 queue->idx % ndev->device->num_comp_vectors;
975 * Reserve CQ slots for RECV + RDMA_READ/RDMA_WRITE + RDMA_SEND.
977 nr_cqe = queue->recv_queue_size + 2 * queue->send_queue_size;
979 queue->cq = ib_alloc_cq(ndev->device, queue,
980 nr_cqe + 1, comp_vector,
982 if (IS_ERR(queue->cq)) {
983 ret = PTR_ERR(queue->cq);
984 pr_err("failed to create CQ cqe= %d ret= %d\n",
989 memset(&qp_attr, 0, sizeof(qp_attr));
990 qp_attr.qp_context = queue;
991 qp_attr.event_handler = nvmet_rdma_qp_event;
992 qp_attr.send_cq = queue->cq;
993 qp_attr.recv_cq = queue->cq;
994 qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
995 qp_attr.qp_type = IB_QPT_RC;
997 qp_attr.cap.max_send_wr = queue->send_queue_size + 1;
998 qp_attr.cap.max_rdma_ctxs = queue->send_queue_size;
999 qp_attr.cap.max_send_sge = max(ndev->device->attrs.max_sge_rd,
1000 ndev->device->attrs.max_send_sge);
1003 qp_attr.srq = ndev->srq;
1006 qp_attr.cap.max_recv_wr = 1 + queue->recv_queue_size;
1007 qp_attr.cap.max_recv_sge = 1 + ndev->inline_page_count;
1010 ret = rdma_create_qp(queue->cm_id, ndev->pd, &qp_attr);
1012 pr_err("failed to create_qp ret= %d\n", ret);
1013 goto err_destroy_cq;
1016 atomic_set(&queue->sq_wr_avail, qp_attr.cap.max_send_wr);
1018 pr_debug("%s: max_cqe= %d max_sge= %d sq_size = %d cm_id= %p\n",
1019 __func__, queue->cq->cqe, qp_attr.cap.max_send_sge,
1020 qp_attr.cap.max_send_wr, queue->cm_id);
1023 for (i = 0; i < queue->recv_queue_size; i++) {
1024 queue->cmds[i].queue = queue;
1025 ret = nvmet_rdma_post_recv(ndev, &queue->cmds[i]);
1027 goto err_destroy_qp;
1035 rdma_destroy_qp(queue->cm_id);
1037 ib_free_cq(queue->cq);
1041 static void nvmet_rdma_destroy_queue_ib(struct nvmet_rdma_queue *queue)
1043 struct ib_qp *qp = queue->cm_id->qp;
1046 rdma_destroy_id(queue->cm_id);
1048 ib_free_cq(queue->cq);
1051 static void nvmet_rdma_free_queue(struct nvmet_rdma_queue *queue)
1053 pr_debug("freeing queue %d\n", queue->idx);
1055 nvmet_sq_destroy(&queue->nvme_sq);
1057 nvmet_rdma_destroy_queue_ib(queue);
1058 if (!queue->dev->srq) {
1059 nvmet_rdma_free_cmds(queue->dev, queue->cmds,
1060 queue->recv_queue_size,
1063 nvmet_rdma_free_rsps(queue);
1064 ida_simple_remove(&nvmet_rdma_queue_ida, queue->idx);
1068 static void nvmet_rdma_release_queue_work(struct work_struct *w)
1070 struct nvmet_rdma_queue *queue =
1071 container_of(w, struct nvmet_rdma_queue, release_work);
1072 struct nvmet_rdma_device *dev = queue->dev;
1074 nvmet_rdma_free_queue(queue);
1076 kref_put(&dev->ref, nvmet_rdma_free_dev);
1080 nvmet_rdma_parse_cm_connect_req(struct rdma_conn_param *conn,
1081 struct nvmet_rdma_queue *queue)
1083 struct nvme_rdma_cm_req *req;
1085 req = (struct nvme_rdma_cm_req *)conn->private_data;
1086 if (!req || conn->private_data_len == 0)
1087 return NVME_RDMA_CM_INVALID_LEN;
1089 if (le16_to_cpu(req->recfmt) != NVME_RDMA_CM_FMT_1_0)
1090 return NVME_RDMA_CM_INVALID_RECFMT;
1092 queue->host_qid = le16_to_cpu(req->qid);
1095 * req->hsqsize corresponds to our recv queue size plus 1
1096 * req->hrqsize corresponds to our send queue size
1098 queue->recv_queue_size = le16_to_cpu(req->hsqsize) + 1;
1099 queue->send_queue_size = le16_to_cpu(req->hrqsize);
1101 if (!queue->host_qid && queue->recv_queue_size > NVME_AQ_DEPTH)
1102 return NVME_RDMA_CM_INVALID_HSQSIZE;
1104 /* XXX: Should we enforce some kind of max for IO queues? */
1109 static int nvmet_rdma_cm_reject(struct rdma_cm_id *cm_id,
1110 enum nvme_rdma_cm_status status)
1112 struct nvme_rdma_cm_rej rej;
1114 pr_debug("rejecting connect request: status %d (%s)\n",
1115 status, nvme_rdma_cm_msg(status));
1117 rej.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1118 rej.sts = cpu_to_le16(status);
1120 return rdma_reject(cm_id, (void *)&rej, sizeof(rej));
1123 static struct nvmet_rdma_queue *
1124 nvmet_rdma_alloc_queue(struct nvmet_rdma_device *ndev,
1125 struct rdma_cm_id *cm_id,
1126 struct rdma_cm_event *event)
1128 struct nvmet_rdma_queue *queue;
1131 queue = kzalloc(sizeof(*queue), GFP_KERNEL);
1133 ret = NVME_RDMA_CM_NO_RSC;
1137 ret = nvmet_sq_init(&queue->nvme_sq);
1139 ret = NVME_RDMA_CM_NO_RSC;
1140 goto out_free_queue;
1143 ret = nvmet_rdma_parse_cm_connect_req(&event->param.conn, queue);
1145 goto out_destroy_sq;
1148 * Schedules the actual release because calling rdma_destroy_id from
1149 * inside a CM callback would trigger a deadlock. (great API design..)
1151 INIT_WORK(&queue->release_work, nvmet_rdma_release_queue_work);
1153 queue->cm_id = cm_id;
1155 spin_lock_init(&queue->state_lock);
1156 queue->state = NVMET_RDMA_Q_CONNECTING;
1157 INIT_LIST_HEAD(&queue->rsp_wait_list);
1158 INIT_LIST_HEAD(&queue->rsp_wr_wait_list);
1159 spin_lock_init(&queue->rsp_wr_wait_lock);
1160 INIT_LIST_HEAD(&queue->free_rsps);
1161 spin_lock_init(&queue->rsps_lock);
1162 INIT_LIST_HEAD(&queue->queue_list);
1164 queue->idx = ida_simple_get(&nvmet_rdma_queue_ida, 0, 0, GFP_KERNEL);
1165 if (queue->idx < 0) {
1166 ret = NVME_RDMA_CM_NO_RSC;
1167 goto out_destroy_sq;
1170 ret = nvmet_rdma_alloc_rsps(queue);
1172 ret = NVME_RDMA_CM_NO_RSC;
1173 goto out_ida_remove;
1177 queue->cmds = nvmet_rdma_alloc_cmds(ndev,
1178 queue->recv_queue_size,
1180 if (IS_ERR(queue->cmds)) {
1181 ret = NVME_RDMA_CM_NO_RSC;
1182 goto out_free_responses;
1186 ret = nvmet_rdma_create_queue_ib(queue);
1188 pr_err("%s: creating RDMA queue failed (%d).\n",
1190 ret = NVME_RDMA_CM_NO_RSC;
1198 nvmet_rdma_free_cmds(queue->dev, queue->cmds,
1199 queue->recv_queue_size,
1203 nvmet_rdma_free_rsps(queue);
1205 ida_simple_remove(&nvmet_rdma_queue_ida, queue->idx);
1207 nvmet_sq_destroy(&queue->nvme_sq);
1211 nvmet_rdma_cm_reject(cm_id, ret);
1215 static void nvmet_rdma_qp_event(struct ib_event *event, void *priv)
1217 struct nvmet_rdma_queue *queue = priv;
1219 switch (event->event) {
1220 case IB_EVENT_COMM_EST:
1221 rdma_notify(queue->cm_id, event->event);
1224 pr_err("received IB QP event: %s (%d)\n",
1225 ib_event_msg(event->event), event->event);
1230 static int nvmet_rdma_cm_accept(struct rdma_cm_id *cm_id,
1231 struct nvmet_rdma_queue *queue,
1232 struct rdma_conn_param *p)
1234 struct rdma_conn_param param = { };
1235 struct nvme_rdma_cm_rep priv = { };
1238 param.rnr_retry_count = 7;
1239 param.flow_control = 1;
1240 param.initiator_depth = min_t(u8, p->initiator_depth,
1241 queue->dev->device->attrs.max_qp_init_rd_atom);
1242 param.private_data = &priv;
1243 param.private_data_len = sizeof(priv);
1244 priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1245 priv.crqsize = cpu_to_le16(queue->recv_queue_size);
1247 ret = rdma_accept(cm_id, ¶m);
1249 pr_err("rdma_accept failed (error code = %d)\n", ret);
1254 static int nvmet_rdma_queue_connect(struct rdma_cm_id *cm_id,
1255 struct rdma_cm_event *event)
1257 struct nvmet_rdma_device *ndev;
1258 struct nvmet_rdma_queue *queue;
1261 ndev = nvmet_rdma_find_get_device(cm_id);
1263 nvmet_rdma_cm_reject(cm_id, NVME_RDMA_CM_NO_RSC);
1264 return -ECONNREFUSED;
1267 queue = nvmet_rdma_alloc_queue(ndev, cm_id, event);
1272 queue->port = cm_id->context;
1274 if (queue->host_qid == 0) {
1275 /* Let inflight controller teardown complete */
1276 flush_scheduled_work();
1279 ret = nvmet_rdma_cm_accept(cm_id, queue, &event->param.conn);
1281 schedule_work(&queue->release_work);
1282 /* Destroying rdma_cm id is not needed here */
1286 mutex_lock(&nvmet_rdma_queue_mutex);
1287 list_add_tail(&queue->queue_list, &nvmet_rdma_queue_list);
1288 mutex_unlock(&nvmet_rdma_queue_mutex);
1293 kref_put(&ndev->ref, nvmet_rdma_free_dev);
1298 static void nvmet_rdma_queue_established(struct nvmet_rdma_queue *queue)
1300 unsigned long flags;
1302 spin_lock_irqsave(&queue->state_lock, flags);
1303 if (queue->state != NVMET_RDMA_Q_CONNECTING) {
1304 pr_warn("trying to establish a connected queue\n");
1307 queue->state = NVMET_RDMA_Q_LIVE;
1309 while (!list_empty(&queue->rsp_wait_list)) {
1310 struct nvmet_rdma_rsp *cmd;
1312 cmd = list_first_entry(&queue->rsp_wait_list,
1313 struct nvmet_rdma_rsp, wait_list);
1314 list_del(&cmd->wait_list);
1316 spin_unlock_irqrestore(&queue->state_lock, flags);
1317 nvmet_rdma_handle_command(queue, cmd);
1318 spin_lock_irqsave(&queue->state_lock, flags);
1322 spin_unlock_irqrestore(&queue->state_lock, flags);
1325 static void __nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
1327 bool disconnect = false;
1328 unsigned long flags;
1330 pr_debug("cm_id= %p queue->state= %d\n", queue->cm_id, queue->state);
1332 spin_lock_irqsave(&queue->state_lock, flags);
1333 switch (queue->state) {
1334 case NVMET_RDMA_Q_CONNECTING:
1335 case NVMET_RDMA_Q_LIVE:
1336 queue->state = NVMET_RDMA_Q_DISCONNECTING;
1339 case NVMET_RDMA_Q_DISCONNECTING:
1342 spin_unlock_irqrestore(&queue->state_lock, flags);
1345 rdma_disconnect(queue->cm_id);
1346 schedule_work(&queue->release_work);
1350 static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
1352 bool disconnect = false;
1354 mutex_lock(&nvmet_rdma_queue_mutex);
1355 if (!list_empty(&queue->queue_list)) {
1356 list_del_init(&queue->queue_list);
1359 mutex_unlock(&nvmet_rdma_queue_mutex);
1362 __nvmet_rdma_queue_disconnect(queue);
1365 static void nvmet_rdma_queue_connect_fail(struct rdma_cm_id *cm_id,
1366 struct nvmet_rdma_queue *queue)
1368 WARN_ON_ONCE(queue->state != NVMET_RDMA_Q_CONNECTING);
1370 mutex_lock(&nvmet_rdma_queue_mutex);
1371 if (!list_empty(&queue->queue_list))
1372 list_del_init(&queue->queue_list);
1373 mutex_unlock(&nvmet_rdma_queue_mutex);
1375 pr_err("failed to connect queue %d\n", queue->idx);
1376 schedule_work(&queue->release_work);
1380 * nvme_rdma_device_removal() - Handle RDMA device removal
1381 * @cm_id: rdma_cm id, used for nvmet port
1382 * @queue: nvmet rdma queue (cm id qp_context)
1384 * DEVICE_REMOVAL event notifies us that the RDMA device is about
1385 * to unplug. Note that this event can be generated on a normal
1386 * queue cm_id and/or a device bound listener cm_id (where in this
1387 * case queue will be null).
1389 * We registered an ib_client to handle device removal for queues,
1390 * so we only need to handle the listening port cm_ids. In this case
1391 * we nullify the priv to prevent double cm_id destruction and destroying
1392 * the cm_id implicitely by returning a non-zero rc to the callout.
1394 static int nvmet_rdma_device_removal(struct rdma_cm_id *cm_id,
1395 struct nvmet_rdma_queue *queue)
1397 struct nvmet_port *port;
1401 * This is a queue cm_id. we have registered
1402 * an ib_client to handle queues removal
1403 * so don't interfear and just return.
1408 port = cm_id->context;
1411 * This is a listener cm_id. Make sure that
1412 * future remove_port won't invoke a double
1413 * cm_id destroy. use atomic xchg to make sure
1414 * we don't compete with remove_port.
1416 if (xchg(&port->priv, NULL) != cm_id)
1420 * We need to return 1 so that the core will destroy
1421 * it's own ID. What a great API design..
1426 static int nvmet_rdma_cm_handler(struct rdma_cm_id *cm_id,
1427 struct rdma_cm_event *event)
1429 struct nvmet_rdma_queue *queue = NULL;
1433 queue = cm_id->qp->qp_context;
1435 pr_debug("%s (%d): status %d id %p\n",
1436 rdma_event_msg(event->event), event->event,
1437 event->status, cm_id);
1439 switch (event->event) {
1440 case RDMA_CM_EVENT_CONNECT_REQUEST:
1441 ret = nvmet_rdma_queue_connect(cm_id, event);
1443 case RDMA_CM_EVENT_ESTABLISHED:
1444 nvmet_rdma_queue_established(queue);
1446 case RDMA_CM_EVENT_ADDR_CHANGE:
1447 case RDMA_CM_EVENT_DISCONNECTED:
1448 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1449 nvmet_rdma_queue_disconnect(queue);
1451 case RDMA_CM_EVENT_DEVICE_REMOVAL:
1452 ret = nvmet_rdma_device_removal(cm_id, queue);
1454 case RDMA_CM_EVENT_REJECTED:
1455 pr_debug("Connection rejected: %s\n",
1456 rdma_reject_msg(cm_id, event->status));
1458 case RDMA_CM_EVENT_UNREACHABLE:
1459 case RDMA_CM_EVENT_CONNECT_ERROR:
1460 nvmet_rdma_queue_connect_fail(cm_id, queue);
1463 pr_err("received unrecognized RDMA CM event %d\n",
1471 static void nvmet_rdma_delete_ctrl(struct nvmet_ctrl *ctrl)
1473 struct nvmet_rdma_queue *queue;
1476 mutex_lock(&nvmet_rdma_queue_mutex);
1477 list_for_each_entry(queue, &nvmet_rdma_queue_list, queue_list) {
1478 if (queue->nvme_sq.ctrl == ctrl) {
1479 list_del_init(&queue->queue_list);
1480 mutex_unlock(&nvmet_rdma_queue_mutex);
1482 __nvmet_rdma_queue_disconnect(queue);
1486 mutex_unlock(&nvmet_rdma_queue_mutex);
1489 static int nvmet_rdma_add_port(struct nvmet_port *port)
1491 struct rdma_cm_id *cm_id;
1492 struct sockaddr_storage addr = { };
1493 __kernel_sa_family_t af;
1496 switch (port->disc_addr.adrfam) {
1497 case NVMF_ADDR_FAMILY_IP4:
1500 case NVMF_ADDR_FAMILY_IP6:
1504 pr_err("address family %d not supported\n",
1505 port->disc_addr.adrfam);
1509 if (port->inline_data_size < 0) {
1510 port->inline_data_size = NVMET_RDMA_DEFAULT_INLINE_DATA_SIZE;
1511 } else if (port->inline_data_size > NVMET_RDMA_MAX_INLINE_DATA_SIZE) {
1512 pr_warn("inline_data_size %u is too large, reducing to %u\n",
1513 port->inline_data_size,
1514 NVMET_RDMA_MAX_INLINE_DATA_SIZE);
1515 port->inline_data_size = NVMET_RDMA_MAX_INLINE_DATA_SIZE;
1518 ret = inet_pton_with_scope(&init_net, af, port->disc_addr.traddr,
1519 port->disc_addr.trsvcid, &addr);
1521 pr_err("malformed ip/port passed: %s:%s\n",
1522 port->disc_addr.traddr, port->disc_addr.trsvcid);
1526 cm_id = rdma_create_id(&init_net, nvmet_rdma_cm_handler, port,
1527 RDMA_PS_TCP, IB_QPT_RC);
1528 if (IS_ERR(cm_id)) {
1529 pr_err("CM ID creation failed\n");
1530 return PTR_ERR(cm_id);
1534 * Allow both IPv4 and IPv6 sockets to bind a single port
1537 ret = rdma_set_afonly(cm_id, 1);
1539 pr_err("rdma_set_afonly failed (%d)\n", ret);
1540 goto out_destroy_id;
1543 ret = rdma_bind_addr(cm_id, (struct sockaddr *)&addr);
1545 pr_err("binding CM ID to %pISpcs failed (%d)\n",
1546 (struct sockaddr *)&addr, ret);
1547 goto out_destroy_id;
1550 ret = rdma_listen(cm_id, 128);
1552 pr_err("listening to %pISpcs failed (%d)\n",
1553 (struct sockaddr *)&addr, ret);
1554 goto out_destroy_id;
1557 pr_info("enabling port %d (%pISpcs)\n",
1558 le16_to_cpu(port->disc_addr.portid), (struct sockaddr *)&addr);
1563 rdma_destroy_id(cm_id);
1567 static void nvmet_rdma_remove_port(struct nvmet_port *port)
1569 struct rdma_cm_id *cm_id = xchg(&port->priv, NULL);
1572 rdma_destroy_id(cm_id);
1575 static void nvmet_rdma_disc_port_addr(struct nvmet_req *req,
1576 struct nvmet_port *port, char *traddr)
1578 struct rdma_cm_id *cm_id = port->priv;
1580 if (inet_addr_is_any((struct sockaddr *)&cm_id->route.addr.src_addr)) {
1581 struct nvmet_rdma_rsp *rsp =
1582 container_of(req, struct nvmet_rdma_rsp, req);
1583 struct rdma_cm_id *req_cm_id = rsp->queue->cm_id;
1584 struct sockaddr *addr = (void *)&req_cm_id->route.addr.src_addr;
1586 sprintf(traddr, "%pISc", addr);
1588 memcpy(traddr, port->disc_addr.traddr, NVMF_TRADDR_SIZE);
1592 static const struct nvmet_fabrics_ops nvmet_rdma_ops = {
1593 .owner = THIS_MODULE,
1594 .type = NVMF_TRTYPE_RDMA,
1596 .has_keyed_sgls = 1,
1597 .add_port = nvmet_rdma_add_port,
1598 .remove_port = nvmet_rdma_remove_port,
1599 .queue_response = nvmet_rdma_queue_response,
1600 .delete_ctrl = nvmet_rdma_delete_ctrl,
1601 .disc_traddr = nvmet_rdma_disc_port_addr,
1604 static void nvmet_rdma_remove_one(struct ib_device *ib_device, void *client_data)
1606 struct nvmet_rdma_queue *queue, *tmp;
1607 struct nvmet_rdma_device *ndev;
1610 mutex_lock(&device_list_mutex);
1611 list_for_each_entry(ndev, &device_list, entry) {
1612 if (ndev->device == ib_device) {
1617 mutex_unlock(&device_list_mutex);
1623 * IB Device that is used by nvmet controllers is being removed,
1624 * delete all queues using this device.
1626 mutex_lock(&nvmet_rdma_queue_mutex);
1627 list_for_each_entry_safe(queue, tmp, &nvmet_rdma_queue_list,
1629 if (queue->dev->device != ib_device)
1632 pr_info("Removing queue %d\n", queue->idx);
1633 list_del_init(&queue->queue_list);
1634 __nvmet_rdma_queue_disconnect(queue);
1636 mutex_unlock(&nvmet_rdma_queue_mutex);
1638 flush_scheduled_work();
1641 static struct ib_client nvmet_rdma_ib_client = {
1642 .name = "nvmet_rdma",
1643 .remove = nvmet_rdma_remove_one
1646 static int __init nvmet_rdma_init(void)
1650 ret = ib_register_client(&nvmet_rdma_ib_client);
1654 ret = nvmet_register_transport(&nvmet_rdma_ops);
1661 ib_unregister_client(&nvmet_rdma_ib_client);
1665 static void __exit nvmet_rdma_exit(void)
1667 nvmet_unregister_transport(&nvmet_rdma_ops);
1668 ib_unregister_client(&nvmet_rdma_ib_client);
1669 WARN_ON_ONCE(!list_empty(&nvmet_rdma_queue_list));
1670 ida_destroy(&nvmet_rdma_queue_ida);
1673 module_init(nvmet_rdma_init);
1674 module_exit(nvmet_rdma_exit);
1676 MODULE_LICENSE("GPL v2");
1677 MODULE_ALIAS("nvmet-transport-1"); /* 1 == NVMF_TRTYPE_RDMA */