2 * NVMe over Fabrics RDMA host code.
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/module.h>
16 #include <linux/init.h>
17 #include <linux/slab.h>
18 #include <linux/err.h>
19 #include <linux/string.h>
20 #include <linux/atomic.h>
21 #include <linux/blk-mq.h>
22 #include <linux/types.h>
23 #include <linux/list.h>
24 #include <linux/mutex.h>
25 #include <linux/scatterlist.h>
26 #include <linux/nvme.h>
27 #include <asm/unaligned.h>
29 #include <rdma/ib_verbs.h>
30 #include <rdma/rdma_cm.h>
31 #include <rdma/ib_cm.h>
32 #include <linux/nvme-rdma.h>
38 #define NVME_RDMA_CONNECT_TIMEOUT_MS 1000 /* 1 second */
40 #define NVME_RDMA_MAX_SEGMENT_SIZE 0xffffff /* 24-bit SGL field */
42 #define NVME_RDMA_MAX_SEGMENTS 256
44 #define NVME_RDMA_MAX_INLINE_SEGMENTS 1
47 * We handle AEN commands ourselves and don't even let the
48 * block layer know about them.
50 #define NVME_RDMA_NR_AEN_COMMANDS 1
51 #define NVME_RDMA_AQ_BLKMQ_DEPTH \
52 (NVMF_AQ_DEPTH - NVME_RDMA_NR_AEN_COMMANDS)
54 struct nvme_rdma_device {
55 struct ib_device *dev;
59 struct list_head entry;
68 struct nvme_rdma_queue;
69 struct nvme_rdma_request {
71 struct nvme_rdma_qe sqe;
72 struct ib_sge sge[1 + NVME_RDMA_MAX_INLINE_SEGMENTS];
76 struct ib_reg_wr reg_wr;
77 struct ib_cqe reg_cqe;
78 struct nvme_rdma_queue *queue;
79 struct sg_table sg_table;
80 struct scatterlist first_sgl[];
83 enum nvme_rdma_queue_flags {
84 NVME_RDMA_Q_CONNECTED = (1 << 0),
87 struct nvme_rdma_queue {
88 struct nvme_rdma_qe *rsp_ring;
91 size_t cmnd_capsule_len;
92 struct nvme_rdma_ctrl *ctrl;
93 struct nvme_rdma_device *device;
98 struct rdma_cm_id *cm_id;
100 struct completion cm_done;
103 struct nvme_rdma_ctrl {
104 /* read and written in the hot path */
107 /* read only in the hot path */
108 struct nvme_rdma_queue *queues;
111 /* other member variables */
112 struct blk_mq_tag_set tag_set;
113 struct work_struct delete_work;
114 struct work_struct reset_work;
115 struct work_struct err_work;
117 struct nvme_rdma_qe async_event_sqe;
120 struct delayed_work reconnect_work;
122 struct list_head list;
124 struct blk_mq_tag_set admin_tag_set;
125 struct nvme_rdma_device *device;
131 struct sockaddr addr;
132 struct sockaddr_in addr_in;
135 struct nvme_ctrl ctrl;
138 static inline struct nvme_rdma_ctrl *to_rdma_ctrl(struct nvme_ctrl *ctrl)
140 return container_of(ctrl, struct nvme_rdma_ctrl, ctrl);
143 static LIST_HEAD(device_list);
144 static DEFINE_MUTEX(device_list_mutex);
146 static LIST_HEAD(nvme_rdma_ctrl_list);
147 static DEFINE_MUTEX(nvme_rdma_ctrl_mutex);
149 static struct workqueue_struct *nvme_rdma_wq;
152 * Disabling this option makes small I/O goes faster, but is fundamentally
153 * unsafe. With it turned off we will have to register a global rkey that
154 * allows read and write access to all physical memory.
156 static bool register_always = true;
157 module_param(register_always, bool, 0444);
158 MODULE_PARM_DESC(register_always,
159 "Use memory registration even for contiguous memory regions");
161 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
162 struct rdma_cm_event *event);
163 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
165 /* XXX: really should move to a generic header sooner or later.. */
166 static inline void put_unaligned_le24(u32 val, u8 *p)
173 static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue *queue)
175 return queue - queue->ctrl->queues;
178 static inline size_t nvme_rdma_inline_data_size(struct nvme_rdma_queue *queue)
180 return queue->cmnd_capsule_len - sizeof(struct nvme_command);
183 static void nvme_rdma_free_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
184 size_t capsule_size, enum dma_data_direction dir)
186 ib_dma_unmap_single(ibdev, qe->dma, capsule_size, dir);
190 static int nvme_rdma_alloc_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
191 size_t capsule_size, enum dma_data_direction dir)
193 qe->data = kzalloc(capsule_size, GFP_KERNEL);
197 qe->dma = ib_dma_map_single(ibdev, qe->data, capsule_size, dir);
198 if (ib_dma_mapping_error(ibdev, qe->dma)) {
206 static void nvme_rdma_free_ring(struct ib_device *ibdev,
207 struct nvme_rdma_qe *ring, size_t ib_queue_size,
208 size_t capsule_size, enum dma_data_direction dir)
212 for (i = 0; i < ib_queue_size; i++)
213 nvme_rdma_free_qe(ibdev, &ring[i], capsule_size, dir);
217 static struct nvme_rdma_qe *nvme_rdma_alloc_ring(struct ib_device *ibdev,
218 size_t ib_queue_size, size_t capsule_size,
219 enum dma_data_direction dir)
221 struct nvme_rdma_qe *ring;
224 ring = kcalloc(ib_queue_size, sizeof(struct nvme_rdma_qe), GFP_KERNEL);
228 for (i = 0; i < ib_queue_size; i++) {
229 if (nvme_rdma_alloc_qe(ibdev, &ring[i], capsule_size, dir))
236 nvme_rdma_free_ring(ibdev, ring, i, capsule_size, dir);
240 static void nvme_rdma_qp_event(struct ib_event *event, void *context)
242 pr_debug("QP event %d\n", event->event);
245 static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue *queue)
247 wait_for_completion_interruptible_timeout(&queue->cm_done,
248 msecs_to_jiffies(NVME_RDMA_CONNECT_TIMEOUT_MS) + 1);
249 return queue->cm_error;
252 static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor)
254 struct nvme_rdma_device *dev = queue->device;
255 struct ib_qp_init_attr init_attr;
258 memset(&init_attr, 0, sizeof(init_attr));
259 init_attr.event_handler = nvme_rdma_qp_event;
261 init_attr.cap.max_send_wr = factor * queue->queue_size + 1;
263 init_attr.cap.max_recv_wr = queue->queue_size + 1;
264 init_attr.cap.max_recv_sge = 1;
265 init_attr.cap.max_send_sge = 1 + NVME_RDMA_MAX_INLINE_SEGMENTS;
266 init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
267 init_attr.qp_type = IB_QPT_RC;
268 init_attr.send_cq = queue->ib_cq;
269 init_attr.recv_cq = queue->ib_cq;
271 ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr);
273 queue->qp = queue->cm_id->qp;
277 static int nvme_rdma_reinit_request(void *data, struct request *rq)
279 struct nvme_rdma_ctrl *ctrl = data;
280 struct nvme_rdma_device *dev = ctrl->device;
281 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
284 if (!req->mr->need_inval)
287 ib_dereg_mr(req->mr);
289 req->mr = ib_alloc_mr(dev->pd, IB_MR_TYPE_MEM_REG,
291 if (IS_ERR(req->mr)) {
292 ret = PTR_ERR(req->mr);
296 req->mr->need_inval = false;
302 static void __nvme_rdma_exit_request(struct nvme_rdma_ctrl *ctrl,
303 struct request *rq, unsigned int queue_idx)
305 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
306 struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
307 struct nvme_rdma_device *dev = queue->device;
310 ib_dereg_mr(req->mr);
312 nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
316 static void nvme_rdma_exit_request(void *data, struct request *rq,
317 unsigned int hctx_idx, unsigned int rq_idx)
319 return __nvme_rdma_exit_request(data, rq, hctx_idx + 1);
322 static void nvme_rdma_exit_admin_request(void *data, struct request *rq,
323 unsigned int hctx_idx, unsigned int rq_idx)
325 return __nvme_rdma_exit_request(data, rq, 0);
328 static int __nvme_rdma_init_request(struct nvme_rdma_ctrl *ctrl,
329 struct request *rq, unsigned int queue_idx)
331 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
332 struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
333 struct nvme_rdma_device *dev = queue->device;
334 struct ib_device *ibdev = dev->dev;
337 BUG_ON(queue_idx >= ctrl->queue_count);
339 ret = nvme_rdma_alloc_qe(ibdev, &req->sqe, sizeof(struct nvme_command),
344 req->mr = ib_alloc_mr(dev->pd, IB_MR_TYPE_MEM_REG,
346 if (IS_ERR(req->mr)) {
347 ret = PTR_ERR(req->mr);
356 nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
361 static int nvme_rdma_init_request(void *data, struct request *rq,
362 unsigned int hctx_idx, unsigned int rq_idx,
363 unsigned int numa_node)
365 return __nvme_rdma_init_request(data, rq, hctx_idx + 1);
368 static int nvme_rdma_init_admin_request(void *data, struct request *rq,
369 unsigned int hctx_idx, unsigned int rq_idx,
370 unsigned int numa_node)
372 return __nvme_rdma_init_request(data, rq, 0);
375 static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
376 unsigned int hctx_idx)
378 struct nvme_rdma_ctrl *ctrl = data;
379 struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1];
381 BUG_ON(hctx_idx >= ctrl->queue_count);
383 hctx->driver_data = queue;
387 static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
388 unsigned int hctx_idx)
390 struct nvme_rdma_ctrl *ctrl = data;
391 struct nvme_rdma_queue *queue = &ctrl->queues[0];
393 BUG_ON(hctx_idx != 0);
395 hctx->driver_data = queue;
399 static void nvme_rdma_free_dev(struct kref *ref)
401 struct nvme_rdma_device *ndev =
402 container_of(ref, struct nvme_rdma_device, ref);
404 mutex_lock(&device_list_mutex);
405 list_del(&ndev->entry);
406 mutex_unlock(&device_list_mutex);
408 if (!register_always)
409 ib_dereg_mr(ndev->mr);
410 ib_dealloc_pd(ndev->pd);
415 static void nvme_rdma_dev_put(struct nvme_rdma_device *dev)
417 kref_put(&dev->ref, nvme_rdma_free_dev);
420 static int nvme_rdma_dev_get(struct nvme_rdma_device *dev)
422 return kref_get_unless_zero(&dev->ref);
425 static struct nvme_rdma_device *
426 nvme_rdma_find_get_device(struct rdma_cm_id *cm_id)
428 struct nvme_rdma_device *ndev;
430 mutex_lock(&device_list_mutex);
431 list_for_each_entry(ndev, &device_list, entry) {
432 if (ndev->dev->node_guid == cm_id->device->node_guid &&
433 nvme_rdma_dev_get(ndev))
437 ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
441 ndev->dev = cm_id->device;
442 kref_init(&ndev->ref);
444 ndev->pd = ib_alloc_pd(ndev->dev);
445 if (IS_ERR(ndev->pd))
448 if (!register_always) {
449 ndev->mr = ib_get_dma_mr(ndev->pd,
450 IB_ACCESS_LOCAL_WRITE |
451 IB_ACCESS_REMOTE_READ |
452 IB_ACCESS_REMOTE_WRITE);
453 if (IS_ERR(ndev->mr))
457 if (!(ndev->dev->attrs.device_cap_flags &
458 IB_DEVICE_MEM_MGT_EXTENSIONS)) {
459 dev_err(&ndev->dev->dev,
460 "Memory registrations not supported.\n");
464 list_add(&ndev->entry, &device_list);
466 mutex_unlock(&device_list_mutex);
470 if (!register_always)
471 ib_dereg_mr(ndev->mr);
473 ib_dealloc_pd(ndev->pd);
477 mutex_unlock(&device_list_mutex);
481 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
483 struct nvme_rdma_device *dev = queue->device;
484 struct ib_device *ibdev = dev->dev;
486 rdma_destroy_qp(queue->cm_id);
487 ib_free_cq(queue->ib_cq);
489 nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
490 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
492 nvme_rdma_dev_put(dev);
495 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue,
496 struct nvme_rdma_device *dev)
498 struct ib_device *ibdev = dev->dev;
499 const int send_wr_factor = 3; /* MR, SEND, INV */
500 const int cq_factor = send_wr_factor + 1; /* + RECV */
501 int comp_vector, idx = nvme_rdma_queue_idx(queue);
508 * The admin queue is barely used once the controller is live, so don't
509 * bother to spread it out.
514 comp_vector = idx % ibdev->num_comp_vectors;
517 /* +1 for ib_stop_cq */
518 queue->ib_cq = ib_alloc_cq(dev->dev, queue,
519 cq_factor * queue->queue_size + 1, comp_vector,
521 if (IS_ERR(queue->ib_cq)) {
522 ret = PTR_ERR(queue->ib_cq);
526 ret = nvme_rdma_create_qp(queue, send_wr_factor);
528 goto out_destroy_ib_cq;
530 queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size,
531 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
532 if (!queue->rsp_ring) {
540 ib_destroy_qp(queue->qp);
542 ib_free_cq(queue->ib_cq);
547 static int nvme_rdma_init_queue(struct nvme_rdma_ctrl *ctrl,
548 int idx, size_t queue_size)
550 struct nvme_rdma_queue *queue;
553 queue = &ctrl->queues[idx];
555 init_completion(&queue->cm_done);
558 queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
560 queue->cmnd_capsule_len = sizeof(struct nvme_command);
562 queue->queue_size = queue_size;
564 queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue,
565 RDMA_PS_TCP, IB_QPT_RC);
566 if (IS_ERR(queue->cm_id)) {
567 dev_info(ctrl->ctrl.device,
568 "failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id));
569 return PTR_ERR(queue->cm_id);
572 queue->cm_error = -ETIMEDOUT;
573 ret = rdma_resolve_addr(queue->cm_id, NULL, &ctrl->addr,
574 NVME_RDMA_CONNECT_TIMEOUT_MS);
576 dev_info(ctrl->ctrl.device,
577 "rdma_resolve_addr failed (%d).\n", ret);
578 goto out_destroy_cm_id;
581 ret = nvme_rdma_wait_for_cm(queue);
583 dev_info(ctrl->ctrl.device,
584 "rdma_resolve_addr wait failed (%d).\n", ret);
585 goto out_destroy_cm_id;
588 set_bit(NVME_RDMA_Q_CONNECTED, &queue->flags);
593 rdma_destroy_id(queue->cm_id);
597 static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
599 rdma_disconnect(queue->cm_id);
600 ib_drain_qp(queue->qp);
603 static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
605 nvme_rdma_destroy_queue_ib(queue);
606 rdma_destroy_id(queue->cm_id);
609 static void nvme_rdma_stop_and_free_queue(struct nvme_rdma_queue *queue)
611 if (!test_and_clear_bit(NVME_RDMA_Q_CONNECTED, &queue->flags))
613 nvme_rdma_stop_queue(queue);
614 nvme_rdma_free_queue(queue);
617 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl)
621 for (i = 1; i < ctrl->queue_count; i++)
622 nvme_rdma_stop_and_free_queue(&ctrl->queues[i]);
625 static int nvme_rdma_connect_io_queues(struct nvme_rdma_ctrl *ctrl)
629 for (i = 1; i < ctrl->queue_count; i++) {
630 ret = nvmf_connect_io_queue(&ctrl->ctrl, i);
638 static int nvme_rdma_init_io_queues(struct nvme_rdma_ctrl *ctrl)
642 for (i = 1; i < ctrl->queue_count; i++) {
643 ret = nvme_rdma_init_queue(ctrl, i,
644 ctrl->ctrl.opts->queue_size);
646 dev_info(ctrl->ctrl.device,
647 "failed to initialize i/o queue: %d\n", ret);
648 goto out_free_queues;
656 nvme_rdma_stop_and_free_queue(&ctrl->queues[i]);
661 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl)
663 nvme_rdma_free_qe(ctrl->queues[0].device->dev, &ctrl->async_event_sqe,
664 sizeof(struct nvme_command), DMA_TO_DEVICE);
665 nvme_rdma_stop_and_free_queue(&ctrl->queues[0]);
666 blk_cleanup_queue(ctrl->ctrl.admin_q);
667 blk_mq_free_tag_set(&ctrl->admin_tag_set);
668 nvme_rdma_dev_put(ctrl->device);
671 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
673 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
675 if (list_empty(&ctrl->list))
678 mutex_lock(&nvme_rdma_ctrl_mutex);
679 list_del(&ctrl->list);
680 mutex_unlock(&nvme_rdma_ctrl_mutex);
683 nvmf_free_options(nctrl->opts);
688 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
690 struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
691 struct nvme_rdma_ctrl, reconnect_work);
695 if (ctrl->queue_count > 1) {
696 nvme_rdma_free_io_queues(ctrl);
698 ret = blk_mq_reinit_tagset(&ctrl->tag_set);
703 nvme_rdma_stop_and_free_queue(&ctrl->queues[0]);
705 ret = blk_mq_reinit_tagset(&ctrl->admin_tag_set);
709 ret = nvme_rdma_init_queue(ctrl, 0, NVMF_AQ_DEPTH);
713 blk_mq_start_stopped_hw_queues(ctrl->ctrl.admin_q, true);
715 ret = nvmf_connect_admin_queue(&ctrl->ctrl);
719 ret = nvme_enable_ctrl(&ctrl->ctrl, ctrl->cap);
723 nvme_start_keep_alive(&ctrl->ctrl);
725 if (ctrl->queue_count > 1) {
726 ret = nvme_rdma_init_io_queues(ctrl);
730 ret = nvme_rdma_connect_io_queues(ctrl);
735 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
736 WARN_ON_ONCE(!changed);
738 if (ctrl->queue_count > 1) {
739 nvme_start_queues(&ctrl->ctrl);
740 nvme_queue_scan(&ctrl->ctrl);
741 nvme_queue_async_events(&ctrl->ctrl);
744 dev_info(ctrl->ctrl.device, "Successfully reconnected\n");
749 blk_mq_stop_hw_queues(ctrl->ctrl.admin_q);
751 /* Make sure we are not resetting/deleting */
752 if (ctrl->ctrl.state == NVME_CTRL_RECONNECTING) {
753 dev_info(ctrl->ctrl.device,
754 "Failed reconnect attempt, requeueing...\n");
755 queue_delayed_work(nvme_rdma_wq, &ctrl->reconnect_work,
756 ctrl->reconnect_delay * HZ);
760 static void nvme_rdma_error_recovery_work(struct work_struct *work)
762 struct nvme_rdma_ctrl *ctrl = container_of(work,
763 struct nvme_rdma_ctrl, err_work);
765 nvme_stop_keep_alive(&ctrl->ctrl);
766 if (ctrl->queue_count > 1)
767 nvme_stop_queues(&ctrl->ctrl);
768 blk_mq_stop_hw_queues(ctrl->ctrl.admin_q);
770 /* We must take care of fastfail/requeue all our inflight requests */
771 if (ctrl->queue_count > 1)
772 blk_mq_tagset_busy_iter(&ctrl->tag_set,
773 nvme_cancel_request, &ctrl->ctrl);
774 blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
775 nvme_cancel_request, &ctrl->ctrl);
777 dev_info(ctrl->ctrl.device, "reconnecting in %d seconds\n",
778 ctrl->reconnect_delay);
780 queue_delayed_work(nvme_rdma_wq, &ctrl->reconnect_work,
781 ctrl->reconnect_delay * HZ);
784 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
786 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RECONNECTING))
789 queue_work(nvme_rdma_wq, &ctrl->err_work);
792 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
795 struct nvme_rdma_queue *queue = cq->cq_context;
796 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
798 if (ctrl->ctrl.state == NVME_CTRL_LIVE)
799 dev_info(ctrl->ctrl.device,
800 "%s for CQE 0x%p failed with status %s (%d)\n",
802 ib_wc_status_msg(wc->status), wc->status);
803 nvme_rdma_error_recovery(ctrl);
806 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
808 if (unlikely(wc->status != IB_WC_SUCCESS))
809 nvme_rdma_wr_error(cq, wc, "MEMREG");
812 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
814 if (unlikely(wc->status != IB_WC_SUCCESS))
815 nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
818 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
819 struct nvme_rdma_request *req)
821 struct ib_send_wr *bad_wr;
822 struct ib_send_wr wr = {
823 .opcode = IB_WR_LOCAL_INV,
827 .ex.invalidate_rkey = req->mr->rkey,
830 req->reg_cqe.done = nvme_rdma_inv_rkey_done;
831 wr.wr_cqe = &req->reg_cqe;
833 return ib_post_send(queue->qp, &wr, &bad_wr);
836 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
839 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
840 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
841 struct nvme_rdma_device *dev = queue->device;
842 struct ib_device *ibdev = dev->dev;
845 if (!blk_rq_bytes(rq))
848 if (req->mr->need_inval) {
849 res = nvme_rdma_inv_rkey(queue, req);
851 dev_err(ctrl->ctrl.device,
852 "Queueing INV WR for rkey %#x failed (%d)\n",
854 nvme_rdma_error_recovery(queue->ctrl);
858 ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
859 req->nents, rq_data_dir(rq) ==
860 WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
862 nvme_cleanup_cmd(rq);
863 sg_free_table_chained(&req->sg_table, true);
866 static int nvme_rdma_set_sg_null(struct nvme_command *c)
868 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
871 put_unaligned_le24(0, sg->length);
872 put_unaligned_le32(0, sg->key);
873 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
877 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
878 struct nvme_rdma_request *req, struct nvme_command *c)
880 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
882 req->sge[1].addr = sg_dma_address(req->sg_table.sgl);
883 req->sge[1].length = sg_dma_len(req->sg_table.sgl);
884 req->sge[1].lkey = queue->device->pd->local_dma_lkey;
886 sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
887 sg->length = cpu_to_le32(sg_dma_len(req->sg_table.sgl));
888 sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
890 req->inline_data = true;
895 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
896 struct nvme_rdma_request *req, struct nvme_command *c)
898 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
900 sg->addr = cpu_to_le64(sg_dma_address(req->sg_table.sgl));
901 put_unaligned_le24(sg_dma_len(req->sg_table.sgl), sg->length);
902 put_unaligned_le32(queue->device->mr->rkey, sg->key);
903 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
907 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
908 struct nvme_rdma_request *req, struct nvme_command *c,
911 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
914 nr = ib_map_mr_sg(req->mr, req->sg_table.sgl, count, NULL, PAGE_SIZE);
921 ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
923 req->reg_cqe.done = nvme_rdma_memreg_done;
924 memset(&req->reg_wr, 0, sizeof(req->reg_wr));
925 req->reg_wr.wr.opcode = IB_WR_REG_MR;
926 req->reg_wr.wr.wr_cqe = &req->reg_cqe;
927 req->reg_wr.wr.num_sge = 0;
928 req->reg_wr.mr = req->mr;
929 req->reg_wr.key = req->mr->rkey;
930 req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
931 IB_ACCESS_REMOTE_READ |
932 IB_ACCESS_REMOTE_WRITE;
934 req->mr->need_inval = true;
936 sg->addr = cpu_to_le64(req->mr->iova);
937 put_unaligned_le24(req->mr->length, sg->length);
938 put_unaligned_le32(req->mr->rkey, sg->key);
939 sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
940 NVME_SGL_FMT_INVALIDATE;
945 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
946 struct request *rq, unsigned int map_len,
947 struct nvme_command *c)
949 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
950 struct nvme_rdma_device *dev = queue->device;
951 struct ib_device *ibdev = dev->dev;
956 req->inline_data = false;
957 req->mr->need_inval = false;
959 c->common.flags |= NVME_CMD_SGL_METABUF;
961 if (!blk_rq_bytes(rq))
962 return nvme_rdma_set_sg_null(c);
964 req->sg_table.sgl = req->first_sgl;
965 ret = sg_alloc_table_chained(&req->sg_table, rq->nr_phys_segments,
970 nents = blk_rq_map_sg(rq->q, rq, req->sg_table.sgl);
971 BUG_ON(nents > rq->nr_phys_segments);
974 count = ib_dma_map_sg(ibdev, req->sg_table.sgl, nents,
975 rq_data_dir(rq) == WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
976 if (unlikely(count <= 0)) {
977 sg_free_table_chained(&req->sg_table, true);
982 if (rq_data_dir(rq) == WRITE &&
983 map_len <= nvme_rdma_inline_data_size(queue) &&
984 nvme_rdma_queue_idx(queue))
985 return nvme_rdma_map_sg_inline(queue, req, c);
987 if (!register_always)
988 return nvme_rdma_map_sg_single(queue, req, c);
991 return nvme_rdma_map_sg_fr(queue, req, c, count);
994 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
996 if (unlikely(wc->status != IB_WC_SUCCESS))
997 nvme_rdma_wr_error(cq, wc, "SEND");
1000 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1001 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1002 struct ib_send_wr *first, bool flush)
1004 struct ib_send_wr wr, *bad_wr;
1007 sge->addr = qe->dma;
1008 sge->length = sizeof(struct nvme_command),
1009 sge->lkey = queue->device->pd->local_dma_lkey;
1011 qe->cqe.done = nvme_rdma_send_done;
1014 wr.wr_cqe = &qe->cqe;
1016 wr.num_sge = num_sge;
1017 wr.opcode = IB_WR_SEND;
1021 * Unsignalled send completions are another giant desaster in the
1022 * IB Verbs spec: If we don't regularly post signalled sends
1023 * the send queue will fill up and only a QP reset will rescue us.
1024 * Would have been way to obvious to handle this in hardware or
1025 * at least the RDMA stack..
1027 * This messy and racy code sniplet is copy and pasted from the iSER
1028 * initiator, and the magic '32' comes from there as well.
1030 * Always signal the flushes. The magic request used for the flush
1031 * sequencer is not allocated in our driver's tagset and it's
1032 * triggered to be freed by blk_cleanup_queue(). So we need to
1033 * always mark it as signaled to ensure that the "wr_cqe", which is
1034 * embeded in request's payload, is not freed when __ib_process_cq()
1035 * calls wr_cqe->done().
1037 if ((++queue->sig_count % 32) == 0 || flush)
1038 wr.send_flags |= IB_SEND_SIGNALED;
1045 ret = ib_post_send(queue->qp, first, &bad_wr);
1047 dev_err(queue->ctrl->ctrl.device,
1048 "%s failed with error code %d\n", __func__, ret);
1053 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1054 struct nvme_rdma_qe *qe)
1056 struct ib_recv_wr wr, *bad_wr;
1060 list.addr = qe->dma;
1061 list.length = sizeof(struct nvme_completion);
1062 list.lkey = queue->device->pd->local_dma_lkey;
1064 qe->cqe.done = nvme_rdma_recv_done;
1067 wr.wr_cqe = &qe->cqe;
1071 ret = ib_post_recv(queue->qp, &wr, &bad_wr);
1073 dev_err(queue->ctrl->ctrl.device,
1074 "%s failed with error code %d\n", __func__, ret);
1079 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1081 u32 queue_idx = nvme_rdma_queue_idx(queue);
1084 return queue->ctrl->admin_tag_set.tags[queue_idx];
1085 return queue->ctrl->tag_set.tags[queue_idx - 1];
1088 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg, int aer_idx)
1090 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1091 struct nvme_rdma_queue *queue = &ctrl->queues[0];
1092 struct ib_device *dev = queue->device->dev;
1093 struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1094 struct nvme_command *cmd = sqe->data;
1098 if (WARN_ON_ONCE(aer_idx != 0))
1101 ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1103 memset(cmd, 0, sizeof(*cmd));
1104 cmd->common.opcode = nvme_admin_async_event;
1105 cmd->common.command_id = NVME_RDMA_AQ_BLKMQ_DEPTH;
1106 cmd->common.flags |= NVME_CMD_SGL_METABUF;
1107 nvme_rdma_set_sg_null(cmd);
1109 ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1112 ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL, false);
1116 static int nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1117 struct nvme_completion *cqe, struct ib_wc *wc, int tag)
1119 u16 status = le16_to_cpu(cqe->status);
1121 struct nvme_rdma_request *req;
1126 rq = blk_mq_tag_to_rq(nvme_rdma_tagset(queue), cqe->command_id);
1128 dev_err(queue->ctrl->ctrl.device,
1129 "tag 0x%x on QP %#x not found\n",
1130 cqe->command_id, queue->qp->qp_num);
1131 nvme_rdma_error_recovery(queue->ctrl);
1134 req = blk_mq_rq_to_pdu(rq);
1136 if (rq->cmd_type == REQ_TYPE_DRV_PRIV && rq->special)
1137 memcpy(rq->special, cqe, sizeof(*cqe));
1142 if ((wc->wc_flags & IB_WC_WITH_INVALIDATE) &&
1143 wc->ex.invalidate_rkey == req->mr->rkey)
1144 req->mr->need_inval = false;
1146 blk_mq_complete_request(rq, status);
1151 static int __nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc, int tag)
1153 struct nvme_rdma_qe *qe =
1154 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1155 struct nvme_rdma_queue *queue = cq->cq_context;
1156 struct ib_device *ibdev = queue->device->dev;
1157 struct nvme_completion *cqe = qe->data;
1158 const size_t len = sizeof(struct nvme_completion);
1161 if (unlikely(wc->status != IB_WC_SUCCESS)) {
1162 nvme_rdma_wr_error(cq, wc, "RECV");
1166 ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1168 * AEN requests are special as they don't time out and can
1169 * survive any kind of queue freeze and often don't respond to
1170 * aborts. We don't even bother to allocate a struct request
1171 * for them but rather special case them here.
1173 if (unlikely(nvme_rdma_queue_idx(queue) == 0 &&
1174 cqe->command_id >= NVME_RDMA_AQ_BLKMQ_DEPTH))
1175 nvme_complete_async_event(&queue->ctrl->ctrl, cqe);
1177 ret = nvme_rdma_process_nvme_rsp(queue, cqe, wc, tag);
1178 ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1180 nvme_rdma_post_recv(queue, qe);
1184 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1186 __nvme_rdma_recv_done(cq, wc, -1);
1189 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1193 for (i = 0; i < queue->queue_size; i++) {
1194 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1196 goto out_destroy_queue_ib;
1201 out_destroy_queue_ib:
1202 nvme_rdma_destroy_queue_ib(queue);
1206 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1207 struct rdma_cm_event *ev)
1209 if (ev->param.conn.private_data_len) {
1210 struct nvme_rdma_cm_rej *rej =
1211 (struct nvme_rdma_cm_rej *)ev->param.conn.private_data;
1213 dev_err(queue->ctrl->ctrl.device,
1214 "Connect rejected, status %d.", le16_to_cpu(rej->sts));
1215 /* XXX: Think of something clever to do here... */
1217 dev_err(queue->ctrl->ctrl.device,
1218 "Connect rejected, no private data.\n");
1224 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1226 struct nvme_rdma_device *dev;
1229 dev = nvme_rdma_find_get_device(queue->cm_id);
1231 dev_err(queue->cm_id->device->dma_device,
1232 "no client data found!\n");
1233 return -ECONNREFUSED;
1236 ret = nvme_rdma_create_queue_ib(queue, dev);
1238 nvme_rdma_dev_put(dev);
1242 ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1244 dev_err(queue->ctrl->ctrl.device,
1245 "rdma_resolve_route failed (%d).\n",
1247 goto out_destroy_queue;
1253 nvme_rdma_destroy_queue_ib(queue);
1258 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1260 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1261 struct rdma_conn_param param = { };
1262 struct nvme_rdma_cm_req priv = { };
1265 param.qp_num = queue->qp->qp_num;
1266 param.flow_control = 1;
1268 param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1269 /* maximum retry count */
1270 param.retry_count = 7;
1271 param.rnr_retry_count = 7;
1272 param.private_data = &priv;
1273 param.private_data_len = sizeof(priv);
1275 priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1276 priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1278 * set the admin queue depth to the minimum size
1279 * specified by the Fabrics standard.
1281 if (priv.qid == 0) {
1282 priv.hrqsize = cpu_to_le16(NVMF_AQ_DEPTH);
1283 priv.hsqsize = cpu_to_le16(NVMF_AQ_DEPTH - 1);
1286 * current interpretation of the fabrics spec
1287 * is at minimum you make hrqsize sqsize+1, or a
1288 * 1's based representation of sqsize.
1290 priv.hrqsize = cpu_to_le16(queue->queue_size);
1291 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1294 ret = rdma_connect(queue->cm_id, ¶m);
1296 dev_err(ctrl->ctrl.device,
1297 "rdma_connect failed (%d).\n", ret);
1298 goto out_destroy_queue_ib;
1303 out_destroy_queue_ib:
1304 nvme_rdma_destroy_queue_ib(queue);
1309 * nvme_rdma_device_unplug() - Handle RDMA device unplug
1310 * @queue: Queue that owns the cm_id that caught the event
1312 * DEVICE_REMOVAL event notifies us that the RDMA device is about
1313 * to unplug so we should take care of destroying our RDMA resources.
1314 * This event will be generated for each allocated cm_id.
1316 * In our case, the RDMA resources are managed per controller and not
1317 * only per queue. So the way we handle this is we trigger an implicit
1318 * controller deletion upon the first DEVICE_REMOVAL event we see, and
1319 * hold the event inflight until the controller deletion is completed.
1321 * One exception that we need to handle is the destruction of the cm_id
1322 * that caught the event. Since we hold the callout until the controller
1323 * deletion is completed, we'll deadlock if the controller deletion will
1324 * call rdma_destroy_id on this queue's cm_id. Thus, we claim ownership
1325 * of destroying this queue before-hand, destroy the queue resources,
1326 * then queue the controller deletion which won't destroy this queue and
1327 * we destroy the cm_id implicitely by returning a non-zero rc to the callout.
1329 static int nvme_rdma_device_unplug(struct nvme_rdma_queue *queue)
1331 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1334 /* Own the controller deletion */
1335 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_DELETING))
1338 dev_warn(ctrl->ctrl.device,
1339 "Got rdma device removal event, deleting ctrl\n");
1341 /* Get rid of reconnect work if its running */
1342 cancel_delayed_work_sync(&ctrl->reconnect_work);
1344 /* Disable the queue so ctrl delete won't free it */
1345 if (test_and_clear_bit(NVME_RDMA_Q_CONNECTED, &queue->flags)) {
1346 /* Free this queue ourselves */
1347 nvme_rdma_stop_queue(queue);
1348 nvme_rdma_destroy_queue_ib(queue);
1350 /* Return non-zero so the cm_id will destroy implicitly */
1354 /* Queue controller deletion */
1355 queue_work(nvme_rdma_wq, &ctrl->delete_work);
1356 flush_work(&ctrl->delete_work);
1360 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1361 struct rdma_cm_event *ev)
1363 struct nvme_rdma_queue *queue = cm_id->context;
1366 dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1367 rdma_event_msg(ev->event), ev->event,
1370 switch (ev->event) {
1371 case RDMA_CM_EVENT_ADDR_RESOLVED:
1372 cm_error = nvme_rdma_addr_resolved(queue);
1374 case RDMA_CM_EVENT_ROUTE_RESOLVED:
1375 cm_error = nvme_rdma_route_resolved(queue);
1377 case RDMA_CM_EVENT_ESTABLISHED:
1378 queue->cm_error = nvme_rdma_conn_established(queue);
1379 /* complete cm_done regardless of success/failure */
1380 complete(&queue->cm_done);
1382 case RDMA_CM_EVENT_REJECTED:
1383 cm_error = nvme_rdma_conn_rejected(queue, ev);
1385 case RDMA_CM_EVENT_ADDR_ERROR:
1386 case RDMA_CM_EVENT_ROUTE_ERROR:
1387 case RDMA_CM_EVENT_CONNECT_ERROR:
1388 case RDMA_CM_EVENT_UNREACHABLE:
1389 dev_dbg(queue->ctrl->ctrl.device,
1390 "CM error event %d\n", ev->event);
1391 cm_error = -ECONNRESET;
1393 case RDMA_CM_EVENT_DISCONNECTED:
1394 case RDMA_CM_EVENT_ADDR_CHANGE:
1395 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1396 dev_dbg(queue->ctrl->ctrl.device,
1397 "disconnect received - connection closed\n");
1398 nvme_rdma_error_recovery(queue->ctrl);
1400 case RDMA_CM_EVENT_DEVICE_REMOVAL:
1401 /* return 1 means impliciy CM ID destroy */
1402 return nvme_rdma_device_unplug(queue);
1404 dev_err(queue->ctrl->ctrl.device,
1405 "Unexpected RDMA CM event (%d)\n", ev->event);
1406 nvme_rdma_error_recovery(queue->ctrl);
1411 queue->cm_error = cm_error;
1412 complete(&queue->cm_done);
1418 static enum blk_eh_timer_return
1419 nvme_rdma_timeout(struct request *rq, bool reserved)
1421 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1423 /* queue error recovery */
1424 nvme_rdma_error_recovery(req->queue->ctrl);
1426 /* fail with DNR on cmd timeout */
1427 rq->errors = NVME_SC_ABORT_REQ | NVME_SC_DNR;
1429 return BLK_EH_HANDLED;
1432 static int nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
1433 const struct blk_mq_queue_data *bd)
1435 struct nvme_ns *ns = hctx->queue->queuedata;
1436 struct nvme_rdma_queue *queue = hctx->driver_data;
1437 struct request *rq = bd->rq;
1438 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1439 struct nvme_rdma_qe *sqe = &req->sqe;
1440 struct nvme_command *c = sqe->data;
1442 struct ib_device *dev;
1443 unsigned int map_len;
1446 WARN_ON_ONCE(rq->tag < 0);
1448 dev = queue->device->dev;
1449 ib_dma_sync_single_for_cpu(dev, sqe->dma,
1450 sizeof(struct nvme_command), DMA_TO_DEVICE);
1452 ret = nvme_setup_cmd(ns, rq, c);
1456 c->common.command_id = rq->tag;
1457 blk_mq_start_request(rq);
1459 map_len = nvme_map_len(rq);
1460 ret = nvme_rdma_map_data(queue, rq, map_len, c);
1462 dev_err(queue->ctrl->ctrl.device,
1463 "Failed to map data (%d)\n", ret);
1464 nvme_cleanup_cmd(rq);
1468 ib_dma_sync_single_for_device(dev, sqe->dma,
1469 sizeof(struct nvme_command), DMA_TO_DEVICE);
1471 if (rq->cmd_type == REQ_TYPE_FS && req_op(rq) == REQ_OP_FLUSH)
1473 ret = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
1474 req->mr->need_inval ? &req->reg_wr.wr : NULL, flush);
1476 nvme_rdma_unmap_data(queue, rq);
1480 return BLK_MQ_RQ_QUEUE_OK;
1482 return (ret == -ENOMEM || ret == -EAGAIN) ?
1483 BLK_MQ_RQ_QUEUE_BUSY : BLK_MQ_RQ_QUEUE_ERROR;
1486 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx, unsigned int tag)
1488 struct nvme_rdma_queue *queue = hctx->driver_data;
1489 struct ib_cq *cq = queue->ib_cq;
1493 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1494 while (ib_poll_cq(cq, 1, &wc) > 0) {
1495 struct ib_cqe *cqe = wc.wr_cqe;
1498 if (cqe->done == nvme_rdma_recv_done)
1499 found |= __nvme_rdma_recv_done(cq, &wc, tag);
1508 static void nvme_rdma_complete_rq(struct request *rq)
1510 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1511 struct nvme_rdma_queue *queue = req->queue;
1514 nvme_rdma_unmap_data(queue, rq);
1516 if (unlikely(rq->errors)) {
1517 if (nvme_req_needs_retry(rq, rq->errors)) {
1518 nvme_requeue_req(rq);
1522 if (rq->cmd_type == REQ_TYPE_DRV_PRIV)
1525 error = nvme_error_status(rq->errors);
1528 blk_mq_end_request(rq, error);
1531 static struct blk_mq_ops nvme_rdma_mq_ops = {
1532 .queue_rq = nvme_rdma_queue_rq,
1533 .complete = nvme_rdma_complete_rq,
1534 .init_request = nvme_rdma_init_request,
1535 .exit_request = nvme_rdma_exit_request,
1536 .reinit_request = nvme_rdma_reinit_request,
1537 .init_hctx = nvme_rdma_init_hctx,
1538 .poll = nvme_rdma_poll,
1539 .timeout = nvme_rdma_timeout,
1542 static struct blk_mq_ops nvme_rdma_admin_mq_ops = {
1543 .queue_rq = nvme_rdma_queue_rq,
1544 .complete = nvme_rdma_complete_rq,
1545 .init_request = nvme_rdma_init_admin_request,
1546 .exit_request = nvme_rdma_exit_admin_request,
1547 .reinit_request = nvme_rdma_reinit_request,
1548 .init_hctx = nvme_rdma_init_admin_hctx,
1549 .timeout = nvme_rdma_timeout,
1552 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl)
1556 error = nvme_rdma_init_queue(ctrl, 0, NVMF_AQ_DEPTH);
1560 ctrl->device = ctrl->queues[0].device;
1563 * We need a reference on the device as long as the tag_set is alive,
1564 * as the MRs in the request structures need a valid ib_device.
1567 if (!nvme_rdma_dev_get(ctrl->device))
1568 goto out_free_queue;
1570 ctrl->max_fr_pages = min_t(u32, NVME_RDMA_MAX_SEGMENTS,
1571 ctrl->device->dev->attrs.max_fast_reg_page_list_len);
1573 memset(&ctrl->admin_tag_set, 0, sizeof(ctrl->admin_tag_set));
1574 ctrl->admin_tag_set.ops = &nvme_rdma_admin_mq_ops;
1575 ctrl->admin_tag_set.queue_depth = NVME_RDMA_AQ_BLKMQ_DEPTH;
1576 ctrl->admin_tag_set.reserved_tags = 2; /* connect + keep-alive */
1577 ctrl->admin_tag_set.numa_node = NUMA_NO_NODE;
1578 ctrl->admin_tag_set.cmd_size = sizeof(struct nvme_rdma_request) +
1579 SG_CHUNK_SIZE * sizeof(struct scatterlist);
1580 ctrl->admin_tag_set.driver_data = ctrl;
1581 ctrl->admin_tag_set.nr_hw_queues = 1;
1582 ctrl->admin_tag_set.timeout = ADMIN_TIMEOUT;
1584 error = blk_mq_alloc_tag_set(&ctrl->admin_tag_set);
1588 ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
1589 if (IS_ERR(ctrl->ctrl.admin_q)) {
1590 error = PTR_ERR(ctrl->ctrl.admin_q);
1591 goto out_free_tagset;
1594 error = nvmf_connect_admin_queue(&ctrl->ctrl);
1596 goto out_cleanup_queue;
1598 error = nvmf_reg_read64(&ctrl->ctrl, NVME_REG_CAP, &ctrl->cap);
1600 dev_err(ctrl->ctrl.device,
1601 "prop_get NVME_REG_CAP failed\n");
1602 goto out_cleanup_queue;
1606 min_t(int, NVME_CAP_MQES(ctrl->cap) + 1, ctrl->ctrl.sqsize);
1608 error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->cap);
1610 goto out_cleanup_queue;
1612 ctrl->ctrl.max_hw_sectors =
1613 (ctrl->max_fr_pages - 1) << (PAGE_SHIFT - 9);
1615 error = nvme_init_identify(&ctrl->ctrl);
1617 goto out_cleanup_queue;
1619 error = nvme_rdma_alloc_qe(ctrl->queues[0].device->dev,
1620 &ctrl->async_event_sqe, sizeof(struct nvme_command),
1623 goto out_cleanup_queue;
1625 nvme_start_keep_alive(&ctrl->ctrl);
1630 blk_cleanup_queue(ctrl->ctrl.admin_q);
1632 /* disconnect and drain the queue before freeing the tagset */
1633 nvme_rdma_stop_queue(&ctrl->queues[0]);
1634 blk_mq_free_tag_set(&ctrl->admin_tag_set);
1636 nvme_rdma_dev_put(ctrl->device);
1638 nvme_rdma_free_queue(&ctrl->queues[0]);
1642 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl)
1644 nvme_stop_keep_alive(&ctrl->ctrl);
1645 cancel_work_sync(&ctrl->err_work);
1646 cancel_delayed_work_sync(&ctrl->reconnect_work);
1648 if (ctrl->queue_count > 1) {
1649 nvme_stop_queues(&ctrl->ctrl);
1650 blk_mq_tagset_busy_iter(&ctrl->tag_set,
1651 nvme_cancel_request, &ctrl->ctrl);
1652 nvme_rdma_free_io_queues(ctrl);
1655 if (test_bit(NVME_RDMA_Q_CONNECTED, &ctrl->queues[0].flags))
1656 nvme_shutdown_ctrl(&ctrl->ctrl);
1658 blk_mq_stop_hw_queues(ctrl->ctrl.admin_q);
1659 blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
1660 nvme_cancel_request, &ctrl->ctrl);
1661 nvme_rdma_destroy_admin_queue(ctrl);
1664 static void __nvme_rdma_remove_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
1666 nvme_uninit_ctrl(&ctrl->ctrl);
1668 nvme_rdma_shutdown_ctrl(ctrl);
1670 if (ctrl->ctrl.tagset) {
1671 blk_cleanup_queue(ctrl->ctrl.connect_q);
1672 blk_mq_free_tag_set(&ctrl->tag_set);
1673 nvme_rdma_dev_put(ctrl->device);
1676 nvme_put_ctrl(&ctrl->ctrl);
1679 static void nvme_rdma_del_ctrl_work(struct work_struct *work)
1681 struct nvme_rdma_ctrl *ctrl = container_of(work,
1682 struct nvme_rdma_ctrl, delete_work);
1684 __nvme_rdma_remove_ctrl(ctrl, true);
1687 static int __nvme_rdma_del_ctrl(struct nvme_rdma_ctrl *ctrl)
1689 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_DELETING))
1692 if (!queue_work(nvme_rdma_wq, &ctrl->delete_work))
1698 static int nvme_rdma_del_ctrl(struct nvme_ctrl *nctrl)
1700 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
1703 ret = __nvme_rdma_del_ctrl(ctrl);
1707 flush_work(&ctrl->delete_work);
1712 static void nvme_rdma_remove_ctrl_work(struct work_struct *work)
1714 struct nvme_rdma_ctrl *ctrl = container_of(work,
1715 struct nvme_rdma_ctrl, delete_work);
1717 __nvme_rdma_remove_ctrl(ctrl, false);
1720 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
1722 struct nvme_rdma_ctrl *ctrl = container_of(work,
1723 struct nvme_rdma_ctrl, reset_work);
1727 nvme_rdma_shutdown_ctrl(ctrl);
1729 ret = nvme_rdma_configure_admin_queue(ctrl);
1731 /* ctrl is already shutdown, just remove the ctrl */
1732 INIT_WORK(&ctrl->delete_work, nvme_rdma_remove_ctrl_work);
1736 if (ctrl->queue_count > 1) {
1737 ret = blk_mq_reinit_tagset(&ctrl->tag_set);
1741 ret = nvme_rdma_init_io_queues(ctrl);
1745 ret = nvme_rdma_connect_io_queues(ctrl);
1750 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1751 WARN_ON_ONCE(!changed);
1753 if (ctrl->queue_count > 1) {
1754 nvme_start_queues(&ctrl->ctrl);
1755 nvme_queue_scan(&ctrl->ctrl);
1756 nvme_queue_async_events(&ctrl->ctrl);
1762 /* Deleting this dead controller... */
1763 dev_warn(ctrl->ctrl.device, "Removing after reset failure\n");
1764 WARN_ON(!queue_work(nvme_rdma_wq, &ctrl->delete_work));
1767 static int nvme_rdma_reset_ctrl(struct nvme_ctrl *nctrl)
1769 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
1771 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
1774 if (!queue_work(nvme_rdma_wq, &ctrl->reset_work))
1777 flush_work(&ctrl->reset_work);
1782 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
1784 .module = THIS_MODULE,
1786 .reg_read32 = nvmf_reg_read32,
1787 .reg_read64 = nvmf_reg_read64,
1788 .reg_write32 = nvmf_reg_write32,
1789 .reset_ctrl = nvme_rdma_reset_ctrl,
1790 .free_ctrl = nvme_rdma_free_ctrl,
1791 .submit_async_event = nvme_rdma_submit_async_event,
1792 .delete_ctrl = nvme_rdma_del_ctrl,
1793 .get_subsysnqn = nvmf_get_subsysnqn,
1794 .get_address = nvmf_get_address,
1797 static int nvme_rdma_create_io_queues(struct nvme_rdma_ctrl *ctrl)
1799 struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
1802 ret = nvme_set_queue_count(&ctrl->ctrl, &opts->nr_io_queues);
1806 ctrl->queue_count = opts->nr_io_queues + 1;
1807 if (ctrl->queue_count < 2)
1810 dev_info(ctrl->ctrl.device,
1811 "creating %d I/O queues.\n", opts->nr_io_queues);
1813 ret = nvme_rdma_init_io_queues(ctrl);
1818 * We need a reference on the device as long as the tag_set is alive,
1819 * as the MRs in the request structures need a valid ib_device.
1822 if (!nvme_rdma_dev_get(ctrl->device))
1823 goto out_free_io_queues;
1825 memset(&ctrl->tag_set, 0, sizeof(ctrl->tag_set));
1826 ctrl->tag_set.ops = &nvme_rdma_mq_ops;
1827 ctrl->tag_set.queue_depth = ctrl->ctrl.opts->queue_size;
1828 ctrl->tag_set.reserved_tags = 1; /* fabric connect */
1829 ctrl->tag_set.numa_node = NUMA_NO_NODE;
1830 ctrl->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
1831 ctrl->tag_set.cmd_size = sizeof(struct nvme_rdma_request) +
1832 SG_CHUNK_SIZE * sizeof(struct scatterlist);
1833 ctrl->tag_set.driver_data = ctrl;
1834 ctrl->tag_set.nr_hw_queues = ctrl->queue_count - 1;
1835 ctrl->tag_set.timeout = NVME_IO_TIMEOUT;
1837 ret = blk_mq_alloc_tag_set(&ctrl->tag_set);
1840 ctrl->ctrl.tagset = &ctrl->tag_set;
1842 ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
1843 if (IS_ERR(ctrl->ctrl.connect_q)) {
1844 ret = PTR_ERR(ctrl->ctrl.connect_q);
1845 goto out_free_tag_set;
1848 ret = nvme_rdma_connect_io_queues(ctrl);
1850 goto out_cleanup_connect_q;
1854 out_cleanup_connect_q:
1855 blk_cleanup_queue(ctrl->ctrl.connect_q);
1857 blk_mq_free_tag_set(&ctrl->tag_set);
1859 nvme_rdma_dev_put(ctrl->device);
1861 nvme_rdma_free_io_queues(ctrl);
1865 static int nvme_rdma_parse_ipaddr(struct sockaddr_in *in_addr, char *p)
1867 u8 *addr = (u8 *)&in_addr->sin_addr.s_addr;
1868 size_t buflen = strlen(p);
1870 /* XXX: handle IPv6 addresses */
1872 if (buflen > INET_ADDRSTRLEN)
1874 if (in4_pton(p, buflen, addr, '\0', NULL) == 0)
1876 in_addr->sin_family = AF_INET;
1880 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
1881 struct nvmf_ctrl_options *opts)
1883 struct nvme_rdma_ctrl *ctrl;
1887 ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1889 return ERR_PTR(-ENOMEM);
1890 ctrl->ctrl.opts = opts;
1891 INIT_LIST_HEAD(&ctrl->list);
1893 ret = nvme_rdma_parse_ipaddr(&ctrl->addr_in, opts->traddr);
1895 pr_err("malformed IP address passed: %s\n", opts->traddr);
1899 if (opts->mask & NVMF_OPT_TRSVCID) {
1902 ret = kstrtou16(opts->trsvcid, 0, &port);
1906 ctrl->addr_in.sin_port = cpu_to_be16(port);
1908 ctrl->addr_in.sin_port = cpu_to_be16(NVME_RDMA_IP_PORT);
1911 ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
1912 0 /* no quirks, we're perfect! */);
1916 ctrl->reconnect_delay = opts->reconnect_delay;
1917 INIT_DELAYED_WORK(&ctrl->reconnect_work,
1918 nvme_rdma_reconnect_ctrl_work);
1919 INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
1920 INIT_WORK(&ctrl->delete_work, nvme_rdma_del_ctrl_work);
1921 INIT_WORK(&ctrl->reset_work, nvme_rdma_reset_ctrl_work);
1922 spin_lock_init(&ctrl->lock);
1924 ctrl->queue_count = opts->nr_io_queues + 1; /* +1 for admin queue */
1925 ctrl->ctrl.sqsize = opts->queue_size - 1;
1926 ctrl->ctrl.kato = opts->kato;
1929 ctrl->queues = kcalloc(ctrl->queue_count, sizeof(*ctrl->queues),
1932 goto out_uninit_ctrl;
1934 ret = nvme_rdma_configure_admin_queue(ctrl);
1936 goto out_kfree_queues;
1938 /* sanity check icdoff */
1939 if (ctrl->ctrl.icdoff) {
1940 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
1941 goto out_remove_admin_queue;
1944 /* sanity check keyed sgls */
1945 if (!(ctrl->ctrl.sgls & (1 << 20))) {
1946 dev_err(ctrl->ctrl.device, "Mandatory keyed sgls are not support\n");
1947 goto out_remove_admin_queue;
1950 if (opts->queue_size > ctrl->ctrl.maxcmd) {
1951 /* warn if maxcmd is lower than queue_size */
1952 dev_warn(ctrl->ctrl.device,
1953 "queue_size %zu > ctrl maxcmd %u, clamping down\n",
1954 opts->queue_size, ctrl->ctrl.maxcmd);
1955 opts->queue_size = ctrl->ctrl.maxcmd;
1958 if (opts->nr_io_queues) {
1959 ret = nvme_rdma_create_io_queues(ctrl);
1961 goto out_remove_admin_queue;
1964 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1965 WARN_ON_ONCE(!changed);
1967 dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISp\n",
1968 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
1970 kref_get(&ctrl->ctrl.kref);
1972 mutex_lock(&nvme_rdma_ctrl_mutex);
1973 list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
1974 mutex_unlock(&nvme_rdma_ctrl_mutex);
1976 if (opts->nr_io_queues) {
1977 nvme_queue_scan(&ctrl->ctrl);
1978 nvme_queue_async_events(&ctrl->ctrl);
1983 out_remove_admin_queue:
1984 nvme_stop_keep_alive(&ctrl->ctrl);
1985 nvme_rdma_destroy_admin_queue(ctrl);
1987 kfree(ctrl->queues);
1989 nvme_uninit_ctrl(&ctrl->ctrl);
1990 nvme_put_ctrl(&ctrl->ctrl);
1993 return ERR_PTR(ret);
1996 return ERR_PTR(ret);
1999 static struct nvmf_transport_ops nvme_rdma_transport = {
2001 .required_opts = NVMF_OPT_TRADDR,
2002 .allowed_opts = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY,
2003 .create_ctrl = nvme_rdma_create_ctrl,
2006 static int __init nvme_rdma_init_module(void)
2008 nvme_rdma_wq = create_workqueue("nvme_rdma_wq");
2012 nvmf_register_transport(&nvme_rdma_transport);
2016 static void __exit nvme_rdma_cleanup_module(void)
2018 struct nvme_rdma_ctrl *ctrl;
2020 nvmf_unregister_transport(&nvme_rdma_transport);
2022 mutex_lock(&nvme_rdma_ctrl_mutex);
2023 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list)
2024 __nvme_rdma_del_ctrl(ctrl);
2025 mutex_unlock(&nvme_rdma_ctrl_mutex);
2027 destroy_workqueue(nvme_rdma_wq);
2030 module_init(nvme_rdma_init_module);
2031 module_exit(nvme_rdma_cleanup_module);
2033 MODULE_LICENSE("GPL v2");