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;
58 struct list_head entry;
67 struct nvme_rdma_queue;
68 struct nvme_rdma_request {
70 struct nvme_rdma_qe sqe;
71 struct ib_sge sge[1 + NVME_RDMA_MAX_INLINE_SEGMENTS];
75 struct ib_reg_wr reg_wr;
76 struct ib_cqe reg_cqe;
77 struct nvme_rdma_queue *queue;
78 struct sg_table sg_table;
79 struct scatterlist first_sgl[];
82 enum nvme_rdma_queue_flags {
83 NVME_RDMA_Q_CONNECTED = (1 << 0),
84 NVME_RDMA_IB_QUEUE_ALLOCATED = (1 << 1),
85 NVME_RDMA_Q_DELETING = (1 << 2),
88 struct nvme_rdma_queue {
89 struct nvme_rdma_qe *rsp_ring;
92 size_t cmnd_capsule_len;
93 struct nvme_rdma_ctrl *ctrl;
94 struct nvme_rdma_device *device;
99 struct rdma_cm_id *cm_id;
101 struct completion cm_done;
104 struct nvme_rdma_ctrl {
105 /* read and written in the hot path */
108 /* read only in the hot path */
109 struct nvme_rdma_queue *queues;
112 /* other member variables */
113 struct blk_mq_tag_set tag_set;
114 struct work_struct delete_work;
115 struct work_struct reset_work;
116 struct work_struct err_work;
118 struct nvme_rdma_qe async_event_sqe;
121 struct delayed_work reconnect_work;
123 struct list_head list;
125 struct blk_mq_tag_set admin_tag_set;
126 struct nvme_rdma_device *device;
132 struct sockaddr addr;
133 struct sockaddr_in addr_in;
136 struct nvme_ctrl ctrl;
139 static inline struct nvme_rdma_ctrl *to_rdma_ctrl(struct nvme_ctrl *ctrl)
141 return container_of(ctrl, struct nvme_rdma_ctrl, ctrl);
144 static LIST_HEAD(device_list);
145 static DEFINE_MUTEX(device_list_mutex);
147 static LIST_HEAD(nvme_rdma_ctrl_list);
148 static DEFINE_MUTEX(nvme_rdma_ctrl_mutex);
150 static struct workqueue_struct *nvme_rdma_wq;
153 * Disabling this option makes small I/O goes faster, but is fundamentally
154 * unsafe. With it turned off we will have to register a global rkey that
155 * allows read and write access to all physical memory.
157 static bool register_always = true;
158 module_param(register_always, bool, 0444);
159 MODULE_PARM_DESC(register_always,
160 "Use memory registration even for contiguous memory regions");
162 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
163 struct rdma_cm_event *event);
164 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
166 /* XXX: really should move to a generic header sooner or later.. */
167 static inline void put_unaligned_le24(u32 val, u8 *p)
174 static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue *queue)
176 return queue - queue->ctrl->queues;
179 static inline size_t nvme_rdma_inline_data_size(struct nvme_rdma_queue *queue)
181 return queue->cmnd_capsule_len - sizeof(struct nvme_command);
184 static void nvme_rdma_free_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
185 size_t capsule_size, enum dma_data_direction dir)
187 ib_dma_unmap_single(ibdev, qe->dma, capsule_size, dir);
191 static int nvme_rdma_alloc_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
192 size_t capsule_size, enum dma_data_direction dir)
194 qe->data = kzalloc(capsule_size, GFP_KERNEL);
198 qe->dma = ib_dma_map_single(ibdev, qe->data, capsule_size, dir);
199 if (ib_dma_mapping_error(ibdev, qe->dma)) {
207 static void nvme_rdma_free_ring(struct ib_device *ibdev,
208 struct nvme_rdma_qe *ring, size_t ib_queue_size,
209 size_t capsule_size, enum dma_data_direction dir)
213 for (i = 0; i < ib_queue_size; i++)
214 nvme_rdma_free_qe(ibdev, &ring[i], capsule_size, dir);
218 static struct nvme_rdma_qe *nvme_rdma_alloc_ring(struct ib_device *ibdev,
219 size_t ib_queue_size, size_t capsule_size,
220 enum dma_data_direction dir)
222 struct nvme_rdma_qe *ring;
225 ring = kcalloc(ib_queue_size, sizeof(struct nvme_rdma_qe), GFP_KERNEL);
229 for (i = 0; i < ib_queue_size; i++) {
230 if (nvme_rdma_alloc_qe(ibdev, &ring[i], capsule_size, dir))
237 nvme_rdma_free_ring(ibdev, ring, i, capsule_size, dir);
241 static void nvme_rdma_qp_event(struct ib_event *event, void *context)
243 pr_debug("QP event %d\n", event->event);
246 static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue *queue)
248 wait_for_completion_interruptible_timeout(&queue->cm_done,
249 msecs_to_jiffies(NVME_RDMA_CONNECT_TIMEOUT_MS) + 1);
250 return queue->cm_error;
253 static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor)
255 struct nvme_rdma_device *dev = queue->device;
256 struct ib_qp_init_attr init_attr;
259 memset(&init_attr, 0, sizeof(init_attr));
260 init_attr.event_handler = nvme_rdma_qp_event;
262 init_attr.cap.max_send_wr = factor * queue->queue_size + 1;
264 init_attr.cap.max_recv_wr = queue->queue_size + 1;
265 init_attr.cap.max_recv_sge = 1;
266 init_attr.cap.max_send_sge = 1 + NVME_RDMA_MAX_INLINE_SEGMENTS;
267 init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
268 init_attr.qp_type = IB_QPT_RC;
269 init_attr.send_cq = queue->ib_cq;
270 init_attr.recv_cq = queue->ib_cq;
272 ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr);
274 queue->qp = queue->cm_id->qp;
278 static int nvme_rdma_reinit_request(void *data, struct request *rq)
280 struct nvme_rdma_ctrl *ctrl = data;
281 struct nvme_rdma_device *dev = ctrl->device;
282 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
285 if (!req->mr->need_inval)
288 ib_dereg_mr(req->mr);
290 req->mr = ib_alloc_mr(dev->pd, IB_MR_TYPE_MEM_REG,
292 if (IS_ERR(req->mr)) {
293 ret = PTR_ERR(req->mr);
298 req->mr->need_inval = false;
304 static void __nvme_rdma_exit_request(struct nvme_rdma_ctrl *ctrl,
305 struct request *rq, unsigned int queue_idx)
307 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
308 struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
309 struct nvme_rdma_device *dev = queue->device;
312 ib_dereg_mr(req->mr);
314 nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
318 static void nvme_rdma_exit_request(void *data, struct request *rq,
319 unsigned int hctx_idx, unsigned int rq_idx)
321 return __nvme_rdma_exit_request(data, rq, hctx_idx + 1);
324 static void nvme_rdma_exit_admin_request(void *data, struct request *rq,
325 unsigned int hctx_idx, unsigned int rq_idx)
327 return __nvme_rdma_exit_request(data, rq, 0);
330 static int __nvme_rdma_init_request(struct nvme_rdma_ctrl *ctrl,
331 struct request *rq, unsigned int queue_idx)
333 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
334 struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
335 struct nvme_rdma_device *dev = queue->device;
336 struct ib_device *ibdev = dev->dev;
339 BUG_ON(queue_idx >= ctrl->queue_count);
341 ret = nvme_rdma_alloc_qe(ibdev, &req->sqe, sizeof(struct nvme_command),
346 req->mr = ib_alloc_mr(dev->pd, IB_MR_TYPE_MEM_REG,
348 if (IS_ERR(req->mr)) {
349 ret = PTR_ERR(req->mr);
358 nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
363 static int nvme_rdma_init_request(void *data, struct request *rq,
364 unsigned int hctx_idx, unsigned int rq_idx,
365 unsigned int numa_node)
367 return __nvme_rdma_init_request(data, rq, hctx_idx + 1);
370 static int nvme_rdma_init_admin_request(void *data, struct request *rq,
371 unsigned int hctx_idx, unsigned int rq_idx,
372 unsigned int numa_node)
374 return __nvme_rdma_init_request(data, rq, 0);
377 static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
378 unsigned int hctx_idx)
380 struct nvme_rdma_ctrl *ctrl = data;
381 struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1];
383 BUG_ON(hctx_idx >= ctrl->queue_count);
385 hctx->driver_data = queue;
389 static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
390 unsigned int hctx_idx)
392 struct nvme_rdma_ctrl *ctrl = data;
393 struct nvme_rdma_queue *queue = &ctrl->queues[0];
395 BUG_ON(hctx_idx != 0);
397 hctx->driver_data = queue;
401 static void nvme_rdma_free_dev(struct kref *ref)
403 struct nvme_rdma_device *ndev =
404 container_of(ref, struct nvme_rdma_device, ref);
406 mutex_lock(&device_list_mutex);
407 list_del(&ndev->entry);
408 mutex_unlock(&device_list_mutex);
410 ib_dealloc_pd(ndev->pd);
414 static void nvme_rdma_dev_put(struct nvme_rdma_device *dev)
416 kref_put(&dev->ref, nvme_rdma_free_dev);
419 static int nvme_rdma_dev_get(struct nvme_rdma_device *dev)
421 return kref_get_unless_zero(&dev->ref);
424 static struct nvme_rdma_device *
425 nvme_rdma_find_get_device(struct rdma_cm_id *cm_id)
427 struct nvme_rdma_device *ndev;
429 mutex_lock(&device_list_mutex);
430 list_for_each_entry(ndev, &device_list, entry) {
431 if (ndev->dev->node_guid == cm_id->device->node_guid &&
432 nvme_rdma_dev_get(ndev))
436 ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
440 ndev->dev = cm_id->device;
441 kref_init(&ndev->ref);
443 ndev->pd = ib_alloc_pd(ndev->dev,
444 register_always ? 0 : IB_PD_UNSAFE_GLOBAL_RKEY);
445 if (IS_ERR(ndev->pd))
448 if (!(ndev->dev->attrs.device_cap_flags &
449 IB_DEVICE_MEM_MGT_EXTENSIONS)) {
450 dev_err(&ndev->dev->dev,
451 "Memory registrations not supported.\n");
455 list_add(&ndev->entry, &device_list);
457 mutex_unlock(&device_list_mutex);
461 ib_dealloc_pd(ndev->pd);
465 mutex_unlock(&device_list_mutex);
469 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
471 struct nvme_rdma_device *dev;
472 struct ib_device *ibdev;
474 if (!test_and_clear_bit(NVME_RDMA_IB_QUEUE_ALLOCATED, &queue->flags))
479 rdma_destroy_qp(queue->cm_id);
480 ib_free_cq(queue->ib_cq);
482 nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
483 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
485 nvme_rdma_dev_put(dev);
488 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue,
489 struct nvme_rdma_device *dev)
491 struct ib_device *ibdev = dev->dev;
492 const int send_wr_factor = 3; /* MR, SEND, INV */
493 const int cq_factor = send_wr_factor + 1; /* + RECV */
494 int comp_vector, idx = nvme_rdma_queue_idx(queue);
501 * The admin queue is barely used once the controller is live, so don't
502 * bother to spread it out.
507 comp_vector = idx % ibdev->num_comp_vectors;
510 /* +1 for ib_stop_cq */
511 queue->ib_cq = ib_alloc_cq(dev->dev, queue,
512 cq_factor * queue->queue_size + 1, comp_vector,
514 if (IS_ERR(queue->ib_cq)) {
515 ret = PTR_ERR(queue->ib_cq);
519 ret = nvme_rdma_create_qp(queue, send_wr_factor);
521 goto out_destroy_ib_cq;
523 queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size,
524 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
525 if (!queue->rsp_ring) {
529 set_bit(NVME_RDMA_IB_QUEUE_ALLOCATED, &queue->flags);
534 ib_destroy_qp(queue->qp);
536 ib_free_cq(queue->ib_cq);
541 static int nvme_rdma_init_queue(struct nvme_rdma_ctrl *ctrl,
542 int idx, size_t queue_size)
544 struct nvme_rdma_queue *queue;
547 queue = &ctrl->queues[idx];
549 init_completion(&queue->cm_done);
552 queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
554 queue->cmnd_capsule_len = sizeof(struct nvme_command);
556 queue->queue_size = queue_size;
558 queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue,
559 RDMA_PS_TCP, IB_QPT_RC);
560 if (IS_ERR(queue->cm_id)) {
561 dev_info(ctrl->ctrl.device,
562 "failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id));
563 return PTR_ERR(queue->cm_id);
566 queue->cm_error = -ETIMEDOUT;
567 ret = rdma_resolve_addr(queue->cm_id, NULL, &ctrl->addr,
568 NVME_RDMA_CONNECT_TIMEOUT_MS);
570 dev_info(ctrl->ctrl.device,
571 "rdma_resolve_addr failed (%d).\n", ret);
572 goto out_destroy_cm_id;
575 ret = nvme_rdma_wait_for_cm(queue);
577 dev_info(ctrl->ctrl.device,
578 "rdma_resolve_addr wait failed (%d).\n", ret);
579 goto out_destroy_cm_id;
582 clear_bit(NVME_RDMA_Q_DELETING, &queue->flags);
583 set_bit(NVME_RDMA_Q_CONNECTED, &queue->flags);
588 nvme_rdma_destroy_queue_ib(queue);
589 rdma_destroy_id(queue->cm_id);
593 static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
595 rdma_disconnect(queue->cm_id);
596 ib_drain_qp(queue->qp);
599 static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
601 nvme_rdma_destroy_queue_ib(queue);
602 rdma_destroy_id(queue->cm_id);
605 static void nvme_rdma_stop_and_free_queue(struct nvme_rdma_queue *queue)
607 if (test_and_set_bit(NVME_RDMA_Q_DELETING, &queue->flags))
609 nvme_rdma_stop_queue(queue);
610 nvme_rdma_free_queue(queue);
613 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl)
617 for (i = 1; i < ctrl->queue_count; i++)
618 nvme_rdma_stop_and_free_queue(&ctrl->queues[i]);
621 static int nvme_rdma_connect_io_queues(struct nvme_rdma_ctrl *ctrl)
625 for (i = 1; i < ctrl->queue_count; i++) {
626 ret = nvmf_connect_io_queue(&ctrl->ctrl, i);
634 static int nvme_rdma_init_io_queues(struct nvme_rdma_ctrl *ctrl)
638 for (i = 1; i < ctrl->queue_count; i++) {
639 ret = nvme_rdma_init_queue(ctrl, i,
640 ctrl->ctrl.opts->queue_size);
642 dev_info(ctrl->ctrl.device,
643 "failed to initialize i/o queue: %d\n", ret);
644 goto out_free_queues;
651 for (i--; i >= 1; i--)
652 nvme_rdma_stop_and_free_queue(&ctrl->queues[i]);
657 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl)
659 nvme_rdma_free_qe(ctrl->queues[0].device->dev, &ctrl->async_event_sqe,
660 sizeof(struct nvme_command), DMA_TO_DEVICE);
661 nvme_rdma_stop_and_free_queue(&ctrl->queues[0]);
662 blk_cleanup_queue(ctrl->ctrl.admin_q);
663 blk_mq_free_tag_set(&ctrl->admin_tag_set);
664 nvme_rdma_dev_put(ctrl->device);
667 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
669 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
671 if (list_empty(&ctrl->list))
674 mutex_lock(&nvme_rdma_ctrl_mutex);
675 list_del(&ctrl->list);
676 mutex_unlock(&nvme_rdma_ctrl_mutex);
679 nvmf_free_options(nctrl->opts);
684 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
686 struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
687 struct nvme_rdma_ctrl, reconnect_work);
691 if (ctrl->queue_count > 1) {
692 nvme_rdma_free_io_queues(ctrl);
694 ret = blk_mq_reinit_tagset(&ctrl->tag_set);
699 nvme_rdma_stop_and_free_queue(&ctrl->queues[0]);
701 ret = blk_mq_reinit_tagset(&ctrl->admin_tag_set);
705 ret = nvme_rdma_init_queue(ctrl, 0, NVMF_AQ_DEPTH);
709 blk_mq_start_stopped_hw_queues(ctrl->ctrl.admin_q, true);
711 ret = nvmf_connect_admin_queue(&ctrl->ctrl);
715 ret = nvme_enable_ctrl(&ctrl->ctrl, ctrl->cap);
719 nvme_start_keep_alive(&ctrl->ctrl);
721 if (ctrl->queue_count > 1) {
722 ret = nvme_rdma_init_io_queues(ctrl);
726 ret = nvme_rdma_connect_io_queues(ctrl);
731 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
732 WARN_ON_ONCE(!changed);
734 if (ctrl->queue_count > 1) {
735 nvme_start_queues(&ctrl->ctrl);
736 nvme_queue_scan(&ctrl->ctrl);
737 nvme_queue_async_events(&ctrl->ctrl);
740 dev_info(ctrl->ctrl.device, "Successfully reconnected\n");
745 blk_mq_stop_hw_queues(ctrl->ctrl.admin_q);
747 /* Make sure we are not resetting/deleting */
748 if (ctrl->ctrl.state == NVME_CTRL_RECONNECTING) {
749 dev_info(ctrl->ctrl.device,
750 "Failed reconnect attempt, requeueing...\n");
751 queue_delayed_work(nvme_rdma_wq, &ctrl->reconnect_work,
752 ctrl->reconnect_delay * HZ);
756 static void nvme_rdma_error_recovery_work(struct work_struct *work)
758 struct nvme_rdma_ctrl *ctrl = container_of(work,
759 struct nvme_rdma_ctrl, err_work);
762 nvme_stop_keep_alive(&ctrl->ctrl);
764 for (i = 0; i < ctrl->queue_count; i++)
765 clear_bit(NVME_RDMA_Q_CONNECTED, &ctrl->queues[i].flags);
767 if (ctrl->queue_count > 1)
768 nvme_stop_queues(&ctrl->ctrl);
769 blk_mq_stop_hw_queues(ctrl->ctrl.admin_q);
771 /* We must take care of fastfail/requeue all our inflight requests */
772 if (ctrl->queue_count > 1)
773 blk_mq_tagset_busy_iter(&ctrl->tag_set,
774 nvme_cancel_request, &ctrl->ctrl);
775 blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
776 nvme_cancel_request, &ctrl->ctrl);
778 dev_info(ctrl->ctrl.device, "reconnecting in %d seconds\n",
779 ctrl->reconnect_delay);
781 queue_delayed_work(nvme_rdma_wq, &ctrl->reconnect_work,
782 ctrl->reconnect_delay * HZ);
785 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
787 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RECONNECTING))
790 queue_work(nvme_rdma_wq, &ctrl->err_work);
793 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
796 struct nvme_rdma_queue *queue = cq->cq_context;
797 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
799 if (ctrl->ctrl.state == NVME_CTRL_LIVE)
800 dev_info(ctrl->ctrl.device,
801 "%s for CQE 0x%p failed with status %s (%d)\n",
803 ib_wc_status_msg(wc->status), wc->status);
804 nvme_rdma_error_recovery(ctrl);
807 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
809 if (unlikely(wc->status != IB_WC_SUCCESS))
810 nvme_rdma_wr_error(cq, wc, "MEMREG");
813 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
815 if (unlikely(wc->status != IB_WC_SUCCESS))
816 nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
819 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
820 struct nvme_rdma_request *req)
822 struct ib_send_wr *bad_wr;
823 struct ib_send_wr wr = {
824 .opcode = IB_WR_LOCAL_INV,
828 .ex.invalidate_rkey = req->mr->rkey,
831 req->reg_cqe.done = nvme_rdma_inv_rkey_done;
832 wr.wr_cqe = &req->reg_cqe;
834 return ib_post_send(queue->qp, &wr, &bad_wr);
837 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
840 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
841 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
842 struct nvme_rdma_device *dev = queue->device;
843 struct ib_device *ibdev = dev->dev;
846 if (!blk_rq_bytes(rq))
849 if (req->mr->need_inval) {
850 res = nvme_rdma_inv_rkey(queue, req);
852 dev_err(ctrl->ctrl.device,
853 "Queueing INV WR for rkey %#x failed (%d)\n",
855 nvme_rdma_error_recovery(queue->ctrl);
859 ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
860 req->nents, rq_data_dir(rq) ==
861 WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
863 nvme_cleanup_cmd(rq);
864 sg_free_table_chained(&req->sg_table, true);
867 static int nvme_rdma_set_sg_null(struct nvme_command *c)
869 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
872 put_unaligned_le24(0, sg->length);
873 put_unaligned_le32(0, sg->key);
874 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
878 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
879 struct nvme_rdma_request *req, struct nvme_command *c)
881 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
883 req->sge[1].addr = sg_dma_address(req->sg_table.sgl);
884 req->sge[1].length = sg_dma_len(req->sg_table.sgl);
885 req->sge[1].lkey = queue->device->pd->local_dma_lkey;
887 sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
888 sg->length = cpu_to_le32(sg_dma_len(req->sg_table.sgl));
889 sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
891 req->inline_data = true;
896 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
897 struct nvme_rdma_request *req, struct nvme_command *c)
899 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
901 sg->addr = cpu_to_le64(sg_dma_address(req->sg_table.sgl));
902 put_unaligned_le24(sg_dma_len(req->sg_table.sgl), sg->length);
903 put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
904 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
908 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
909 struct nvme_rdma_request *req, struct nvme_command *c,
912 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
915 nr = ib_map_mr_sg(req->mr, req->sg_table.sgl, count, NULL, PAGE_SIZE);
922 ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
924 req->reg_cqe.done = nvme_rdma_memreg_done;
925 memset(&req->reg_wr, 0, sizeof(req->reg_wr));
926 req->reg_wr.wr.opcode = IB_WR_REG_MR;
927 req->reg_wr.wr.wr_cqe = &req->reg_cqe;
928 req->reg_wr.wr.num_sge = 0;
929 req->reg_wr.mr = req->mr;
930 req->reg_wr.key = req->mr->rkey;
931 req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
932 IB_ACCESS_REMOTE_READ |
933 IB_ACCESS_REMOTE_WRITE;
935 req->mr->need_inval = true;
937 sg->addr = cpu_to_le64(req->mr->iova);
938 put_unaligned_le24(req->mr->length, sg->length);
939 put_unaligned_le32(req->mr->rkey, sg->key);
940 sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
941 NVME_SGL_FMT_INVALIDATE;
946 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
947 struct request *rq, unsigned int map_len,
948 struct nvme_command *c)
950 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
951 struct nvme_rdma_device *dev = queue->device;
952 struct ib_device *ibdev = dev->dev;
957 req->inline_data = false;
958 req->mr->need_inval = false;
960 c->common.flags |= NVME_CMD_SGL_METABUF;
962 if (!blk_rq_bytes(rq))
963 return nvme_rdma_set_sg_null(c);
965 req->sg_table.sgl = req->first_sgl;
966 ret = sg_alloc_table_chained(&req->sg_table, rq->nr_phys_segments,
971 nents = blk_rq_map_sg(rq->q, rq, req->sg_table.sgl);
972 BUG_ON(nents > rq->nr_phys_segments);
975 count = ib_dma_map_sg(ibdev, req->sg_table.sgl, nents,
976 rq_data_dir(rq) == WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
977 if (unlikely(count <= 0)) {
978 sg_free_table_chained(&req->sg_table, true);
983 if (rq_data_dir(rq) == WRITE &&
984 map_len <= nvme_rdma_inline_data_size(queue) &&
985 nvme_rdma_queue_idx(queue))
986 return nvme_rdma_map_sg_inline(queue, req, c);
988 if (dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY)
989 return nvme_rdma_map_sg_single(queue, req, c);
992 return nvme_rdma_map_sg_fr(queue, req, c, count);
995 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
997 if (unlikely(wc->status != IB_WC_SUCCESS))
998 nvme_rdma_wr_error(cq, wc, "SEND");
1001 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1002 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1003 struct ib_send_wr *first, bool flush)
1005 struct ib_send_wr wr, *bad_wr;
1008 sge->addr = qe->dma;
1009 sge->length = sizeof(struct nvme_command),
1010 sge->lkey = queue->device->pd->local_dma_lkey;
1012 qe->cqe.done = nvme_rdma_send_done;
1015 wr.wr_cqe = &qe->cqe;
1017 wr.num_sge = num_sge;
1018 wr.opcode = IB_WR_SEND;
1022 * Unsignalled send completions are another giant desaster in the
1023 * IB Verbs spec: If we don't regularly post signalled sends
1024 * the send queue will fill up and only a QP reset will rescue us.
1025 * Would have been way to obvious to handle this in hardware or
1026 * at least the RDMA stack..
1028 * This messy and racy code sniplet is copy and pasted from the iSER
1029 * initiator, and the magic '32' comes from there as well.
1031 * Always signal the flushes. The magic request used for the flush
1032 * sequencer is not allocated in our driver's tagset and it's
1033 * triggered to be freed by blk_cleanup_queue(). So we need to
1034 * always mark it as signaled to ensure that the "wr_cqe", which is
1035 * embeded in request's payload, is not freed when __ib_process_cq()
1036 * calls wr_cqe->done().
1038 if ((++queue->sig_count % 32) == 0 || flush)
1039 wr.send_flags |= IB_SEND_SIGNALED;
1046 ret = ib_post_send(queue->qp, first, &bad_wr);
1048 dev_err(queue->ctrl->ctrl.device,
1049 "%s failed with error code %d\n", __func__, ret);
1054 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1055 struct nvme_rdma_qe *qe)
1057 struct ib_recv_wr wr, *bad_wr;
1061 list.addr = qe->dma;
1062 list.length = sizeof(struct nvme_completion);
1063 list.lkey = queue->device->pd->local_dma_lkey;
1065 qe->cqe.done = nvme_rdma_recv_done;
1068 wr.wr_cqe = &qe->cqe;
1072 ret = ib_post_recv(queue->qp, &wr, &bad_wr);
1074 dev_err(queue->ctrl->ctrl.device,
1075 "%s failed with error code %d\n", __func__, ret);
1080 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1082 u32 queue_idx = nvme_rdma_queue_idx(queue);
1085 return queue->ctrl->admin_tag_set.tags[queue_idx];
1086 return queue->ctrl->tag_set.tags[queue_idx - 1];
1089 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg, int aer_idx)
1091 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1092 struct nvme_rdma_queue *queue = &ctrl->queues[0];
1093 struct ib_device *dev = queue->device->dev;
1094 struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1095 struct nvme_command *cmd = sqe->data;
1099 if (WARN_ON_ONCE(aer_idx != 0))
1102 ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1104 memset(cmd, 0, sizeof(*cmd));
1105 cmd->common.opcode = nvme_admin_async_event;
1106 cmd->common.command_id = NVME_RDMA_AQ_BLKMQ_DEPTH;
1107 cmd->common.flags |= NVME_CMD_SGL_METABUF;
1108 nvme_rdma_set_sg_null(cmd);
1110 ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1113 ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL, false);
1117 static int nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1118 struct nvme_completion *cqe, struct ib_wc *wc, int tag)
1120 u16 status = le16_to_cpu(cqe->status);
1122 struct nvme_rdma_request *req;
1127 rq = blk_mq_tag_to_rq(nvme_rdma_tagset(queue), cqe->command_id);
1129 dev_err(queue->ctrl->ctrl.device,
1130 "tag 0x%x on QP %#x not found\n",
1131 cqe->command_id, queue->qp->qp_num);
1132 nvme_rdma_error_recovery(queue->ctrl);
1135 req = blk_mq_rq_to_pdu(rq);
1137 if (rq->cmd_type == REQ_TYPE_DRV_PRIV && rq->special)
1138 memcpy(rq->special, cqe, sizeof(*cqe));
1143 if ((wc->wc_flags & IB_WC_WITH_INVALIDATE) &&
1144 wc->ex.invalidate_rkey == req->mr->rkey)
1145 req->mr->need_inval = false;
1147 blk_mq_complete_request(rq, status);
1152 static int __nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc, int tag)
1154 struct nvme_rdma_qe *qe =
1155 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1156 struct nvme_rdma_queue *queue = cq->cq_context;
1157 struct ib_device *ibdev = queue->device->dev;
1158 struct nvme_completion *cqe = qe->data;
1159 const size_t len = sizeof(struct nvme_completion);
1162 if (unlikely(wc->status != IB_WC_SUCCESS)) {
1163 nvme_rdma_wr_error(cq, wc, "RECV");
1167 ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1169 * AEN requests are special as they don't time out and can
1170 * survive any kind of queue freeze and often don't respond to
1171 * aborts. We don't even bother to allocate a struct request
1172 * for them but rather special case them here.
1174 if (unlikely(nvme_rdma_queue_idx(queue) == 0 &&
1175 cqe->command_id >= NVME_RDMA_AQ_BLKMQ_DEPTH))
1176 nvme_complete_async_event(&queue->ctrl->ctrl, cqe);
1178 ret = nvme_rdma_process_nvme_rsp(queue, cqe, wc, tag);
1179 ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1181 nvme_rdma_post_recv(queue, qe);
1185 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1187 __nvme_rdma_recv_done(cq, wc, -1);
1190 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1194 for (i = 0; i < queue->queue_size; i++) {
1195 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1197 goto out_destroy_queue_ib;
1202 out_destroy_queue_ib:
1203 nvme_rdma_destroy_queue_ib(queue);
1207 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1208 struct rdma_cm_event *ev)
1210 if (ev->param.conn.private_data_len) {
1211 struct nvme_rdma_cm_rej *rej =
1212 (struct nvme_rdma_cm_rej *)ev->param.conn.private_data;
1214 dev_err(queue->ctrl->ctrl.device,
1215 "Connect rejected, status %d.", le16_to_cpu(rej->sts));
1216 /* XXX: Think of something clever to do here... */
1218 dev_err(queue->ctrl->ctrl.device,
1219 "Connect rejected, no private data.\n");
1225 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1227 struct nvme_rdma_device *dev;
1230 dev = nvme_rdma_find_get_device(queue->cm_id);
1232 dev_err(queue->cm_id->device->dma_device,
1233 "no client data found!\n");
1234 return -ECONNREFUSED;
1237 ret = nvme_rdma_create_queue_ib(queue, dev);
1239 nvme_rdma_dev_put(dev);
1243 ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1245 dev_err(queue->ctrl->ctrl.device,
1246 "rdma_resolve_route failed (%d).\n",
1248 goto out_destroy_queue;
1254 nvme_rdma_destroy_queue_ib(queue);
1259 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1261 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1262 struct rdma_conn_param param = { };
1263 struct nvme_rdma_cm_req priv = { };
1266 param.qp_num = queue->qp->qp_num;
1267 param.flow_control = 1;
1269 param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1270 /* maximum retry count */
1271 param.retry_count = 7;
1272 param.rnr_retry_count = 7;
1273 param.private_data = &priv;
1274 param.private_data_len = sizeof(priv);
1276 priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1277 priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1279 * set the admin queue depth to the minimum size
1280 * specified by the Fabrics standard.
1282 if (priv.qid == 0) {
1283 priv.hrqsize = cpu_to_le16(NVMF_AQ_DEPTH);
1284 priv.hsqsize = cpu_to_le16(NVMF_AQ_DEPTH - 1);
1287 * current interpretation of the fabrics spec
1288 * is at minimum you make hrqsize sqsize+1, or a
1289 * 1's based representation of sqsize.
1291 priv.hrqsize = cpu_to_le16(queue->queue_size);
1292 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1295 ret = rdma_connect(queue->cm_id, ¶m);
1297 dev_err(ctrl->ctrl.device,
1298 "rdma_connect failed (%d).\n", ret);
1299 goto out_destroy_queue_ib;
1304 out_destroy_queue_ib:
1305 nvme_rdma_destroy_queue_ib(queue);
1309 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1310 struct rdma_cm_event *ev)
1312 struct nvme_rdma_queue *queue = cm_id->context;
1315 dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1316 rdma_event_msg(ev->event), ev->event,
1319 switch (ev->event) {
1320 case RDMA_CM_EVENT_ADDR_RESOLVED:
1321 cm_error = nvme_rdma_addr_resolved(queue);
1323 case RDMA_CM_EVENT_ROUTE_RESOLVED:
1324 cm_error = nvme_rdma_route_resolved(queue);
1326 case RDMA_CM_EVENT_ESTABLISHED:
1327 queue->cm_error = nvme_rdma_conn_established(queue);
1328 /* complete cm_done regardless of success/failure */
1329 complete(&queue->cm_done);
1331 case RDMA_CM_EVENT_REJECTED:
1332 cm_error = nvme_rdma_conn_rejected(queue, ev);
1334 case RDMA_CM_EVENT_ADDR_ERROR:
1335 case RDMA_CM_EVENT_ROUTE_ERROR:
1336 case RDMA_CM_EVENT_CONNECT_ERROR:
1337 case RDMA_CM_EVENT_UNREACHABLE:
1338 dev_dbg(queue->ctrl->ctrl.device,
1339 "CM error event %d\n", ev->event);
1340 cm_error = -ECONNRESET;
1342 case RDMA_CM_EVENT_DISCONNECTED:
1343 case RDMA_CM_EVENT_ADDR_CHANGE:
1344 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1345 dev_dbg(queue->ctrl->ctrl.device,
1346 "disconnect received - connection closed\n");
1347 nvme_rdma_error_recovery(queue->ctrl);
1349 case RDMA_CM_EVENT_DEVICE_REMOVAL:
1350 /* device removal is handled via the ib_client API */
1353 dev_err(queue->ctrl->ctrl.device,
1354 "Unexpected RDMA CM event (%d)\n", ev->event);
1355 nvme_rdma_error_recovery(queue->ctrl);
1360 queue->cm_error = cm_error;
1361 complete(&queue->cm_done);
1367 static enum blk_eh_timer_return
1368 nvme_rdma_timeout(struct request *rq, bool reserved)
1370 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1372 /* queue error recovery */
1373 nvme_rdma_error_recovery(req->queue->ctrl);
1375 /* fail with DNR on cmd timeout */
1376 rq->errors = NVME_SC_ABORT_REQ | NVME_SC_DNR;
1378 return BLK_EH_HANDLED;
1381 static int nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
1382 const struct blk_mq_queue_data *bd)
1384 struct nvme_ns *ns = hctx->queue->queuedata;
1385 struct nvme_rdma_queue *queue = hctx->driver_data;
1386 struct request *rq = bd->rq;
1387 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1388 struct nvme_rdma_qe *sqe = &req->sqe;
1389 struct nvme_command *c = sqe->data;
1391 struct ib_device *dev;
1392 unsigned int map_len;
1395 WARN_ON_ONCE(rq->tag < 0);
1397 dev = queue->device->dev;
1398 ib_dma_sync_single_for_cpu(dev, sqe->dma,
1399 sizeof(struct nvme_command), DMA_TO_DEVICE);
1401 ret = nvme_setup_cmd(ns, rq, c);
1405 c->common.command_id = rq->tag;
1406 blk_mq_start_request(rq);
1408 map_len = nvme_map_len(rq);
1409 ret = nvme_rdma_map_data(queue, rq, map_len, c);
1411 dev_err(queue->ctrl->ctrl.device,
1412 "Failed to map data (%d)\n", ret);
1413 nvme_cleanup_cmd(rq);
1417 ib_dma_sync_single_for_device(dev, sqe->dma,
1418 sizeof(struct nvme_command), DMA_TO_DEVICE);
1420 if (rq->cmd_type == REQ_TYPE_FS && req_op(rq) == REQ_OP_FLUSH)
1422 ret = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
1423 req->mr->need_inval ? &req->reg_wr.wr : NULL, flush);
1425 nvme_rdma_unmap_data(queue, rq);
1429 return BLK_MQ_RQ_QUEUE_OK;
1431 return (ret == -ENOMEM || ret == -EAGAIN) ?
1432 BLK_MQ_RQ_QUEUE_BUSY : BLK_MQ_RQ_QUEUE_ERROR;
1435 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx, unsigned int tag)
1437 struct nvme_rdma_queue *queue = hctx->driver_data;
1438 struct ib_cq *cq = queue->ib_cq;
1442 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1443 while (ib_poll_cq(cq, 1, &wc) > 0) {
1444 struct ib_cqe *cqe = wc.wr_cqe;
1447 if (cqe->done == nvme_rdma_recv_done)
1448 found |= __nvme_rdma_recv_done(cq, &wc, tag);
1457 static void nvme_rdma_complete_rq(struct request *rq)
1459 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1460 struct nvme_rdma_queue *queue = req->queue;
1463 nvme_rdma_unmap_data(queue, rq);
1465 if (unlikely(rq->errors)) {
1466 if (nvme_req_needs_retry(rq, rq->errors)) {
1467 nvme_requeue_req(rq);
1471 if (rq->cmd_type == REQ_TYPE_DRV_PRIV)
1474 error = nvme_error_status(rq->errors);
1477 blk_mq_end_request(rq, error);
1480 static struct blk_mq_ops nvme_rdma_mq_ops = {
1481 .queue_rq = nvme_rdma_queue_rq,
1482 .complete = nvme_rdma_complete_rq,
1483 .init_request = nvme_rdma_init_request,
1484 .exit_request = nvme_rdma_exit_request,
1485 .reinit_request = nvme_rdma_reinit_request,
1486 .init_hctx = nvme_rdma_init_hctx,
1487 .poll = nvme_rdma_poll,
1488 .timeout = nvme_rdma_timeout,
1491 static struct blk_mq_ops nvme_rdma_admin_mq_ops = {
1492 .queue_rq = nvme_rdma_queue_rq,
1493 .complete = nvme_rdma_complete_rq,
1494 .init_request = nvme_rdma_init_admin_request,
1495 .exit_request = nvme_rdma_exit_admin_request,
1496 .reinit_request = nvme_rdma_reinit_request,
1497 .init_hctx = nvme_rdma_init_admin_hctx,
1498 .timeout = nvme_rdma_timeout,
1501 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl)
1505 error = nvme_rdma_init_queue(ctrl, 0, NVMF_AQ_DEPTH);
1509 ctrl->device = ctrl->queues[0].device;
1512 * We need a reference on the device as long as the tag_set is alive,
1513 * as the MRs in the request structures need a valid ib_device.
1516 if (!nvme_rdma_dev_get(ctrl->device))
1517 goto out_free_queue;
1519 ctrl->max_fr_pages = min_t(u32, NVME_RDMA_MAX_SEGMENTS,
1520 ctrl->device->dev->attrs.max_fast_reg_page_list_len);
1522 memset(&ctrl->admin_tag_set, 0, sizeof(ctrl->admin_tag_set));
1523 ctrl->admin_tag_set.ops = &nvme_rdma_admin_mq_ops;
1524 ctrl->admin_tag_set.queue_depth = NVME_RDMA_AQ_BLKMQ_DEPTH;
1525 ctrl->admin_tag_set.reserved_tags = 2; /* connect + keep-alive */
1526 ctrl->admin_tag_set.numa_node = NUMA_NO_NODE;
1527 ctrl->admin_tag_set.cmd_size = sizeof(struct nvme_rdma_request) +
1528 SG_CHUNK_SIZE * sizeof(struct scatterlist);
1529 ctrl->admin_tag_set.driver_data = ctrl;
1530 ctrl->admin_tag_set.nr_hw_queues = 1;
1531 ctrl->admin_tag_set.timeout = ADMIN_TIMEOUT;
1533 error = blk_mq_alloc_tag_set(&ctrl->admin_tag_set);
1537 ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
1538 if (IS_ERR(ctrl->ctrl.admin_q)) {
1539 error = PTR_ERR(ctrl->ctrl.admin_q);
1540 goto out_free_tagset;
1543 error = nvmf_connect_admin_queue(&ctrl->ctrl);
1545 goto out_cleanup_queue;
1547 error = nvmf_reg_read64(&ctrl->ctrl, NVME_REG_CAP, &ctrl->cap);
1549 dev_err(ctrl->ctrl.device,
1550 "prop_get NVME_REG_CAP failed\n");
1551 goto out_cleanup_queue;
1555 min_t(int, NVME_CAP_MQES(ctrl->cap) + 1, ctrl->ctrl.sqsize);
1557 error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->cap);
1559 goto out_cleanup_queue;
1561 ctrl->ctrl.max_hw_sectors =
1562 (ctrl->max_fr_pages - 1) << (PAGE_SHIFT - 9);
1564 error = nvme_init_identify(&ctrl->ctrl);
1566 goto out_cleanup_queue;
1568 error = nvme_rdma_alloc_qe(ctrl->queues[0].device->dev,
1569 &ctrl->async_event_sqe, sizeof(struct nvme_command),
1572 goto out_cleanup_queue;
1574 nvme_start_keep_alive(&ctrl->ctrl);
1579 blk_cleanup_queue(ctrl->ctrl.admin_q);
1581 /* disconnect and drain the queue before freeing the tagset */
1582 nvme_rdma_stop_queue(&ctrl->queues[0]);
1583 blk_mq_free_tag_set(&ctrl->admin_tag_set);
1585 nvme_rdma_dev_put(ctrl->device);
1587 nvme_rdma_free_queue(&ctrl->queues[0]);
1591 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl)
1593 nvme_stop_keep_alive(&ctrl->ctrl);
1594 cancel_work_sync(&ctrl->err_work);
1595 cancel_delayed_work_sync(&ctrl->reconnect_work);
1597 if (ctrl->queue_count > 1) {
1598 nvme_stop_queues(&ctrl->ctrl);
1599 blk_mq_tagset_busy_iter(&ctrl->tag_set,
1600 nvme_cancel_request, &ctrl->ctrl);
1601 nvme_rdma_free_io_queues(ctrl);
1604 if (test_bit(NVME_RDMA_Q_CONNECTED, &ctrl->queues[0].flags))
1605 nvme_shutdown_ctrl(&ctrl->ctrl);
1607 blk_mq_stop_hw_queues(ctrl->ctrl.admin_q);
1608 blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
1609 nvme_cancel_request, &ctrl->ctrl);
1610 nvme_rdma_destroy_admin_queue(ctrl);
1613 static void __nvme_rdma_remove_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
1615 nvme_uninit_ctrl(&ctrl->ctrl);
1617 nvme_rdma_shutdown_ctrl(ctrl);
1619 if (ctrl->ctrl.tagset) {
1620 blk_cleanup_queue(ctrl->ctrl.connect_q);
1621 blk_mq_free_tag_set(&ctrl->tag_set);
1622 nvme_rdma_dev_put(ctrl->device);
1625 nvme_put_ctrl(&ctrl->ctrl);
1628 static void nvme_rdma_del_ctrl_work(struct work_struct *work)
1630 struct nvme_rdma_ctrl *ctrl = container_of(work,
1631 struct nvme_rdma_ctrl, delete_work);
1633 __nvme_rdma_remove_ctrl(ctrl, true);
1636 static int __nvme_rdma_del_ctrl(struct nvme_rdma_ctrl *ctrl)
1638 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_DELETING))
1641 if (!queue_work(nvme_rdma_wq, &ctrl->delete_work))
1647 static int nvme_rdma_del_ctrl(struct nvme_ctrl *nctrl)
1649 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
1653 * Keep a reference until all work is flushed since
1654 * __nvme_rdma_del_ctrl can free the ctrl mem
1656 if (!kref_get_unless_zero(&ctrl->ctrl.kref))
1658 ret = __nvme_rdma_del_ctrl(ctrl);
1660 flush_work(&ctrl->delete_work);
1661 nvme_put_ctrl(&ctrl->ctrl);
1665 static void nvme_rdma_remove_ctrl_work(struct work_struct *work)
1667 struct nvme_rdma_ctrl *ctrl = container_of(work,
1668 struct nvme_rdma_ctrl, delete_work);
1670 __nvme_rdma_remove_ctrl(ctrl, false);
1673 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
1675 struct nvme_rdma_ctrl *ctrl = container_of(work,
1676 struct nvme_rdma_ctrl, reset_work);
1680 nvme_rdma_shutdown_ctrl(ctrl);
1682 ret = nvme_rdma_configure_admin_queue(ctrl);
1684 /* ctrl is already shutdown, just remove the ctrl */
1685 INIT_WORK(&ctrl->delete_work, nvme_rdma_remove_ctrl_work);
1689 if (ctrl->queue_count > 1) {
1690 ret = blk_mq_reinit_tagset(&ctrl->tag_set);
1694 ret = nvme_rdma_init_io_queues(ctrl);
1698 ret = nvme_rdma_connect_io_queues(ctrl);
1703 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1704 WARN_ON_ONCE(!changed);
1706 if (ctrl->queue_count > 1) {
1707 nvme_start_queues(&ctrl->ctrl);
1708 nvme_queue_scan(&ctrl->ctrl);
1709 nvme_queue_async_events(&ctrl->ctrl);
1715 /* Deleting this dead controller... */
1716 dev_warn(ctrl->ctrl.device, "Removing after reset failure\n");
1717 WARN_ON(!queue_work(nvme_rdma_wq, &ctrl->delete_work));
1720 static int nvme_rdma_reset_ctrl(struct nvme_ctrl *nctrl)
1722 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
1724 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
1727 if (!queue_work(nvme_rdma_wq, &ctrl->reset_work))
1730 flush_work(&ctrl->reset_work);
1735 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
1737 .module = THIS_MODULE,
1739 .reg_read32 = nvmf_reg_read32,
1740 .reg_read64 = nvmf_reg_read64,
1741 .reg_write32 = nvmf_reg_write32,
1742 .reset_ctrl = nvme_rdma_reset_ctrl,
1743 .free_ctrl = nvme_rdma_free_ctrl,
1744 .submit_async_event = nvme_rdma_submit_async_event,
1745 .delete_ctrl = nvme_rdma_del_ctrl,
1746 .get_subsysnqn = nvmf_get_subsysnqn,
1747 .get_address = nvmf_get_address,
1750 static int nvme_rdma_create_io_queues(struct nvme_rdma_ctrl *ctrl)
1752 struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
1755 ret = nvme_set_queue_count(&ctrl->ctrl, &opts->nr_io_queues);
1759 ctrl->queue_count = opts->nr_io_queues + 1;
1760 if (ctrl->queue_count < 2)
1763 dev_info(ctrl->ctrl.device,
1764 "creating %d I/O queues.\n", opts->nr_io_queues);
1766 ret = nvme_rdma_init_io_queues(ctrl);
1771 * We need a reference on the device as long as the tag_set is alive,
1772 * as the MRs in the request structures need a valid ib_device.
1775 if (!nvme_rdma_dev_get(ctrl->device))
1776 goto out_free_io_queues;
1778 memset(&ctrl->tag_set, 0, sizeof(ctrl->tag_set));
1779 ctrl->tag_set.ops = &nvme_rdma_mq_ops;
1780 ctrl->tag_set.queue_depth = ctrl->ctrl.opts->queue_size;
1781 ctrl->tag_set.reserved_tags = 1; /* fabric connect */
1782 ctrl->tag_set.numa_node = NUMA_NO_NODE;
1783 ctrl->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
1784 ctrl->tag_set.cmd_size = sizeof(struct nvme_rdma_request) +
1785 SG_CHUNK_SIZE * sizeof(struct scatterlist);
1786 ctrl->tag_set.driver_data = ctrl;
1787 ctrl->tag_set.nr_hw_queues = ctrl->queue_count - 1;
1788 ctrl->tag_set.timeout = NVME_IO_TIMEOUT;
1790 ret = blk_mq_alloc_tag_set(&ctrl->tag_set);
1793 ctrl->ctrl.tagset = &ctrl->tag_set;
1795 ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
1796 if (IS_ERR(ctrl->ctrl.connect_q)) {
1797 ret = PTR_ERR(ctrl->ctrl.connect_q);
1798 goto out_free_tag_set;
1801 ret = nvme_rdma_connect_io_queues(ctrl);
1803 goto out_cleanup_connect_q;
1807 out_cleanup_connect_q:
1808 blk_cleanup_queue(ctrl->ctrl.connect_q);
1810 blk_mq_free_tag_set(&ctrl->tag_set);
1812 nvme_rdma_dev_put(ctrl->device);
1814 nvme_rdma_free_io_queues(ctrl);
1818 static int nvme_rdma_parse_ipaddr(struct sockaddr_in *in_addr, char *p)
1820 u8 *addr = (u8 *)&in_addr->sin_addr.s_addr;
1821 size_t buflen = strlen(p);
1823 /* XXX: handle IPv6 addresses */
1825 if (buflen > INET_ADDRSTRLEN)
1827 if (in4_pton(p, buflen, addr, '\0', NULL) == 0)
1829 in_addr->sin_family = AF_INET;
1833 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
1834 struct nvmf_ctrl_options *opts)
1836 struct nvme_rdma_ctrl *ctrl;
1840 ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1842 return ERR_PTR(-ENOMEM);
1843 ctrl->ctrl.opts = opts;
1844 INIT_LIST_HEAD(&ctrl->list);
1846 ret = nvme_rdma_parse_ipaddr(&ctrl->addr_in, opts->traddr);
1848 pr_err("malformed IP address passed: %s\n", opts->traddr);
1852 if (opts->mask & NVMF_OPT_TRSVCID) {
1855 ret = kstrtou16(opts->trsvcid, 0, &port);
1859 ctrl->addr_in.sin_port = cpu_to_be16(port);
1861 ctrl->addr_in.sin_port = cpu_to_be16(NVME_RDMA_IP_PORT);
1864 ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
1865 0 /* no quirks, we're perfect! */);
1869 ctrl->reconnect_delay = opts->reconnect_delay;
1870 INIT_DELAYED_WORK(&ctrl->reconnect_work,
1871 nvme_rdma_reconnect_ctrl_work);
1872 INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
1873 INIT_WORK(&ctrl->delete_work, nvme_rdma_del_ctrl_work);
1874 INIT_WORK(&ctrl->reset_work, nvme_rdma_reset_ctrl_work);
1875 spin_lock_init(&ctrl->lock);
1877 ctrl->queue_count = opts->nr_io_queues + 1; /* +1 for admin queue */
1878 ctrl->ctrl.sqsize = opts->queue_size - 1;
1879 ctrl->ctrl.kato = opts->kato;
1882 ctrl->queues = kcalloc(ctrl->queue_count, sizeof(*ctrl->queues),
1885 goto out_uninit_ctrl;
1887 ret = nvme_rdma_configure_admin_queue(ctrl);
1889 goto out_kfree_queues;
1891 /* sanity check icdoff */
1892 if (ctrl->ctrl.icdoff) {
1893 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
1894 goto out_remove_admin_queue;
1897 /* sanity check keyed sgls */
1898 if (!(ctrl->ctrl.sgls & (1 << 20))) {
1899 dev_err(ctrl->ctrl.device, "Mandatory keyed sgls are not support\n");
1900 goto out_remove_admin_queue;
1903 if (opts->queue_size > ctrl->ctrl.maxcmd) {
1904 /* warn if maxcmd is lower than queue_size */
1905 dev_warn(ctrl->ctrl.device,
1906 "queue_size %zu > ctrl maxcmd %u, clamping down\n",
1907 opts->queue_size, ctrl->ctrl.maxcmd);
1908 opts->queue_size = ctrl->ctrl.maxcmd;
1911 if (opts->nr_io_queues) {
1912 ret = nvme_rdma_create_io_queues(ctrl);
1914 goto out_remove_admin_queue;
1917 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1918 WARN_ON_ONCE(!changed);
1920 dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISp\n",
1921 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
1923 kref_get(&ctrl->ctrl.kref);
1925 mutex_lock(&nvme_rdma_ctrl_mutex);
1926 list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
1927 mutex_unlock(&nvme_rdma_ctrl_mutex);
1929 if (opts->nr_io_queues) {
1930 nvme_queue_scan(&ctrl->ctrl);
1931 nvme_queue_async_events(&ctrl->ctrl);
1936 out_remove_admin_queue:
1937 nvme_stop_keep_alive(&ctrl->ctrl);
1938 nvme_rdma_destroy_admin_queue(ctrl);
1940 kfree(ctrl->queues);
1942 nvme_uninit_ctrl(&ctrl->ctrl);
1943 nvme_put_ctrl(&ctrl->ctrl);
1946 return ERR_PTR(ret);
1949 return ERR_PTR(ret);
1952 static struct nvmf_transport_ops nvme_rdma_transport = {
1954 .required_opts = NVMF_OPT_TRADDR,
1955 .allowed_opts = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY,
1956 .create_ctrl = nvme_rdma_create_ctrl,
1959 static void nvme_rdma_add_one(struct ib_device *ib_device)
1963 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
1965 struct nvme_rdma_ctrl *ctrl;
1967 /* Delete all controllers using this device */
1968 mutex_lock(&nvme_rdma_ctrl_mutex);
1969 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
1970 if (ctrl->device->dev != ib_device)
1972 dev_info(ctrl->ctrl.device,
1973 "Removing ctrl: NQN \"%s\", addr %pISp\n",
1974 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
1975 __nvme_rdma_del_ctrl(ctrl);
1977 mutex_unlock(&nvme_rdma_ctrl_mutex);
1979 flush_workqueue(nvme_rdma_wq);
1982 static struct ib_client nvme_rdma_ib_client = {
1983 .name = "nvme_rdma",
1984 .add = nvme_rdma_add_one,
1985 .remove = nvme_rdma_remove_one
1988 static int __init nvme_rdma_init_module(void)
1992 nvme_rdma_wq = create_workqueue("nvme_rdma_wq");
1996 ret = ib_register_client(&nvme_rdma_ib_client);
1998 destroy_workqueue(nvme_rdma_wq);
2002 nvmf_register_transport(&nvme_rdma_transport);
2006 static void __exit nvme_rdma_cleanup_module(void)
2008 nvmf_unregister_transport(&nvme_rdma_transport);
2009 ib_unregister_client(&nvme_rdma_ib_client);
2010 destroy_workqueue(nvme_rdma_wq);
2013 module_init(nvme_rdma_init_module);
2014 module_exit(nvme_rdma_cleanup_module);
2016 MODULE_LICENSE("GPL v2");