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),
85 NVME_RDMA_IB_QUEUE_ALLOCATED = (1 << 1),
86 NVME_RDMA_Q_DELETING = (1 << 2),
89 struct nvme_rdma_queue {
90 struct nvme_rdma_qe *rsp_ring;
93 size_t cmnd_capsule_len;
94 struct nvme_rdma_ctrl *ctrl;
95 struct nvme_rdma_device *device;
100 struct rdma_cm_id *cm_id;
102 struct completion cm_done;
105 struct nvme_rdma_ctrl {
106 /* read and written in the hot path */
109 /* read only in the hot path */
110 struct nvme_rdma_queue *queues;
113 /* other member variables */
114 struct blk_mq_tag_set tag_set;
115 struct work_struct delete_work;
116 struct work_struct reset_work;
117 struct work_struct err_work;
119 struct nvme_rdma_qe async_event_sqe;
122 struct delayed_work reconnect_work;
124 struct list_head list;
126 struct blk_mq_tag_set admin_tag_set;
127 struct nvme_rdma_device *device;
133 struct sockaddr addr;
134 struct sockaddr_in addr_in;
137 struct nvme_ctrl ctrl;
140 static inline struct nvme_rdma_ctrl *to_rdma_ctrl(struct nvme_ctrl *ctrl)
142 return container_of(ctrl, struct nvme_rdma_ctrl, ctrl);
145 static LIST_HEAD(device_list);
146 static DEFINE_MUTEX(device_list_mutex);
148 static LIST_HEAD(nvme_rdma_ctrl_list);
149 static DEFINE_MUTEX(nvme_rdma_ctrl_mutex);
151 static struct workqueue_struct *nvme_rdma_wq;
154 * Disabling this option makes small I/O goes faster, but is fundamentally
155 * unsafe. With it turned off we will have to register a global rkey that
156 * allows read and write access to all physical memory.
158 static bool register_always = true;
159 module_param(register_always, bool, 0444);
160 MODULE_PARM_DESC(register_always,
161 "Use memory registration even for contiguous memory regions");
163 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
164 struct rdma_cm_event *event);
165 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
167 /* XXX: really should move to a generic header sooner or later.. */
168 static inline void put_unaligned_le24(u32 val, u8 *p)
175 static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue *queue)
177 return queue - queue->ctrl->queues;
180 static inline size_t nvme_rdma_inline_data_size(struct nvme_rdma_queue *queue)
182 return queue->cmnd_capsule_len - sizeof(struct nvme_command);
185 static void nvme_rdma_free_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
186 size_t capsule_size, enum dma_data_direction dir)
188 ib_dma_unmap_single(ibdev, qe->dma, capsule_size, dir);
192 static int nvme_rdma_alloc_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
193 size_t capsule_size, enum dma_data_direction dir)
195 qe->data = kzalloc(capsule_size, GFP_KERNEL);
199 qe->dma = ib_dma_map_single(ibdev, qe->data, capsule_size, dir);
200 if (ib_dma_mapping_error(ibdev, qe->dma)) {
208 static void nvme_rdma_free_ring(struct ib_device *ibdev,
209 struct nvme_rdma_qe *ring, size_t ib_queue_size,
210 size_t capsule_size, enum dma_data_direction dir)
214 for (i = 0; i < ib_queue_size; i++)
215 nvme_rdma_free_qe(ibdev, &ring[i], capsule_size, dir);
219 static struct nvme_rdma_qe *nvme_rdma_alloc_ring(struct ib_device *ibdev,
220 size_t ib_queue_size, size_t capsule_size,
221 enum dma_data_direction dir)
223 struct nvme_rdma_qe *ring;
226 ring = kcalloc(ib_queue_size, sizeof(struct nvme_rdma_qe), GFP_KERNEL);
230 for (i = 0; i < ib_queue_size; i++) {
231 if (nvme_rdma_alloc_qe(ibdev, &ring[i], capsule_size, dir))
238 nvme_rdma_free_ring(ibdev, ring, i, capsule_size, dir);
242 static void nvme_rdma_qp_event(struct ib_event *event, void *context)
244 pr_debug("QP event %d\n", event->event);
247 static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue *queue)
249 wait_for_completion_interruptible_timeout(&queue->cm_done,
250 msecs_to_jiffies(NVME_RDMA_CONNECT_TIMEOUT_MS) + 1);
251 return queue->cm_error;
254 static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor)
256 struct nvme_rdma_device *dev = queue->device;
257 struct ib_qp_init_attr init_attr;
260 memset(&init_attr, 0, sizeof(init_attr));
261 init_attr.event_handler = nvme_rdma_qp_event;
263 init_attr.cap.max_send_wr = factor * queue->queue_size + 1;
265 init_attr.cap.max_recv_wr = queue->queue_size + 1;
266 init_attr.cap.max_recv_sge = 1;
267 init_attr.cap.max_send_sge = 1 + NVME_RDMA_MAX_INLINE_SEGMENTS;
268 init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
269 init_attr.qp_type = IB_QPT_RC;
270 init_attr.send_cq = queue->ib_cq;
271 init_attr.recv_cq = queue->ib_cq;
273 ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr);
275 queue->qp = queue->cm_id->qp;
279 static int nvme_rdma_reinit_request(void *data, struct request *rq)
281 struct nvme_rdma_ctrl *ctrl = data;
282 struct nvme_rdma_device *dev = ctrl->device;
283 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
286 if (!req->mr->need_inval)
289 ib_dereg_mr(req->mr);
291 req->mr = ib_alloc_mr(dev->pd, IB_MR_TYPE_MEM_REG,
293 if (IS_ERR(req->mr)) {
294 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 if (!register_always)
411 ib_dereg_mr(ndev->mr);
412 ib_dealloc_pd(ndev->pd);
417 static void nvme_rdma_dev_put(struct nvme_rdma_device *dev)
419 kref_put(&dev->ref, nvme_rdma_free_dev);
422 static int nvme_rdma_dev_get(struct nvme_rdma_device *dev)
424 return kref_get_unless_zero(&dev->ref);
427 static struct nvme_rdma_device *
428 nvme_rdma_find_get_device(struct rdma_cm_id *cm_id)
430 struct nvme_rdma_device *ndev;
432 mutex_lock(&device_list_mutex);
433 list_for_each_entry(ndev, &device_list, entry) {
434 if (ndev->dev->node_guid == cm_id->device->node_guid &&
435 nvme_rdma_dev_get(ndev))
439 ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
443 ndev->dev = cm_id->device;
444 kref_init(&ndev->ref);
446 ndev->pd = ib_alloc_pd(ndev->dev);
447 if (IS_ERR(ndev->pd))
450 if (!register_always) {
451 ndev->mr = ib_get_dma_mr(ndev->pd,
452 IB_ACCESS_LOCAL_WRITE |
453 IB_ACCESS_REMOTE_READ |
454 IB_ACCESS_REMOTE_WRITE);
455 if (IS_ERR(ndev->mr))
459 if (!(ndev->dev->attrs.device_cap_flags &
460 IB_DEVICE_MEM_MGT_EXTENSIONS)) {
461 dev_err(&ndev->dev->dev,
462 "Memory registrations not supported.\n");
466 list_add(&ndev->entry, &device_list);
468 mutex_unlock(&device_list_mutex);
472 if (!register_always)
473 ib_dereg_mr(ndev->mr);
475 ib_dealloc_pd(ndev->pd);
479 mutex_unlock(&device_list_mutex);
483 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
485 struct nvme_rdma_device *dev;
486 struct ib_device *ibdev;
488 if (!test_and_clear_bit(NVME_RDMA_IB_QUEUE_ALLOCATED, &queue->flags))
493 rdma_destroy_qp(queue->cm_id);
494 ib_free_cq(queue->ib_cq);
496 nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
497 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
499 nvme_rdma_dev_put(dev);
502 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue,
503 struct nvme_rdma_device *dev)
505 struct ib_device *ibdev = dev->dev;
506 const int send_wr_factor = 3; /* MR, SEND, INV */
507 const int cq_factor = send_wr_factor + 1; /* + RECV */
508 int comp_vector, idx = nvme_rdma_queue_idx(queue);
515 * The admin queue is barely used once the controller is live, so don't
516 * bother to spread it out.
521 comp_vector = idx % ibdev->num_comp_vectors;
524 /* +1 for ib_stop_cq */
525 queue->ib_cq = ib_alloc_cq(dev->dev, queue,
526 cq_factor * queue->queue_size + 1, comp_vector,
528 if (IS_ERR(queue->ib_cq)) {
529 ret = PTR_ERR(queue->ib_cq);
533 ret = nvme_rdma_create_qp(queue, send_wr_factor);
535 goto out_destroy_ib_cq;
537 queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size,
538 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
539 if (!queue->rsp_ring) {
543 set_bit(NVME_RDMA_IB_QUEUE_ALLOCATED, &queue->flags);
548 ib_destroy_qp(queue->qp);
550 ib_free_cq(queue->ib_cq);
555 static int nvme_rdma_init_queue(struct nvme_rdma_ctrl *ctrl,
556 int idx, size_t queue_size)
558 struct nvme_rdma_queue *queue;
561 queue = &ctrl->queues[idx];
564 init_completion(&queue->cm_done);
567 queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
569 queue->cmnd_capsule_len = sizeof(struct nvme_command);
571 queue->queue_size = queue_size;
573 queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue,
574 RDMA_PS_TCP, IB_QPT_RC);
575 if (IS_ERR(queue->cm_id)) {
576 dev_info(ctrl->ctrl.device,
577 "failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id));
578 return PTR_ERR(queue->cm_id);
581 queue->cm_error = -ETIMEDOUT;
582 ret = rdma_resolve_addr(queue->cm_id, NULL, &ctrl->addr,
583 NVME_RDMA_CONNECT_TIMEOUT_MS);
585 dev_info(ctrl->ctrl.device,
586 "rdma_resolve_addr failed (%d).\n", ret);
587 goto out_destroy_cm_id;
590 ret = nvme_rdma_wait_for_cm(queue);
592 dev_info(ctrl->ctrl.device,
593 "rdma_resolve_addr wait failed (%d).\n", ret);
594 goto out_destroy_cm_id;
597 set_bit(NVME_RDMA_Q_CONNECTED, &queue->flags);
602 nvme_rdma_destroy_queue_ib(queue);
603 rdma_destroy_id(queue->cm_id);
607 static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
609 rdma_disconnect(queue->cm_id);
610 ib_drain_qp(queue->qp);
613 static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
615 nvme_rdma_destroy_queue_ib(queue);
616 rdma_destroy_id(queue->cm_id);
619 static void nvme_rdma_stop_and_free_queue(struct nvme_rdma_queue *queue)
621 if (test_and_set_bit(NVME_RDMA_Q_DELETING, &queue->flags))
623 nvme_rdma_stop_queue(queue);
624 nvme_rdma_free_queue(queue);
627 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl)
631 for (i = 1; i < ctrl->queue_count; i++)
632 nvme_rdma_stop_and_free_queue(&ctrl->queues[i]);
635 static int nvme_rdma_connect_io_queues(struct nvme_rdma_ctrl *ctrl)
639 for (i = 1; i < ctrl->queue_count; i++) {
640 ret = nvmf_connect_io_queue(&ctrl->ctrl, i);
648 static int nvme_rdma_init_io_queues(struct nvme_rdma_ctrl *ctrl)
652 for (i = 1; i < ctrl->queue_count; i++) {
653 ret = nvme_rdma_init_queue(ctrl, i,
654 ctrl->ctrl.opts->queue_size);
656 dev_info(ctrl->ctrl.device,
657 "failed to initialize i/o queue: %d\n", ret);
658 goto out_free_queues;
665 for (i--; i >= 1; i--)
666 nvme_rdma_stop_and_free_queue(&ctrl->queues[i]);
671 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl)
673 nvme_rdma_free_qe(ctrl->queues[0].device->dev, &ctrl->async_event_sqe,
674 sizeof(struct nvme_command), DMA_TO_DEVICE);
675 nvme_rdma_stop_and_free_queue(&ctrl->queues[0]);
676 blk_cleanup_queue(ctrl->ctrl.admin_q);
677 blk_mq_free_tag_set(&ctrl->admin_tag_set);
678 nvme_rdma_dev_put(ctrl->device);
681 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
683 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
685 if (list_empty(&ctrl->list))
688 mutex_lock(&nvme_rdma_ctrl_mutex);
689 list_del(&ctrl->list);
690 mutex_unlock(&nvme_rdma_ctrl_mutex);
693 nvmf_free_options(nctrl->opts);
698 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
700 struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
701 struct nvme_rdma_ctrl, reconnect_work);
705 if (ctrl->queue_count > 1) {
706 nvme_rdma_free_io_queues(ctrl);
708 ret = blk_mq_reinit_tagset(&ctrl->tag_set);
713 nvme_rdma_stop_and_free_queue(&ctrl->queues[0]);
715 ret = blk_mq_reinit_tagset(&ctrl->admin_tag_set);
719 ret = nvme_rdma_init_queue(ctrl, 0, NVMF_AQ_DEPTH);
723 blk_mq_start_stopped_hw_queues(ctrl->ctrl.admin_q, true);
725 ret = nvmf_connect_admin_queue(&ctrl->ctrl);
729 ret = nvme_enable_ctrl(&ctrl->ctrl, ctrl->cap);
733 nvme_start_keep_alive(&ctrl->ctrl);
735 if (ctrl->queue_count > 1) {
736 ret = nvme_rdma_init_io_queues(ctrl);
740 ret = nvme_rdma_connect_io_queues(ctrl);
745 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
746 WARN_ON_ONCE(!changed);
748 if (ctrl->queue_count > 1) {
749 nvme_start_queues(&ctrl->ctrl);
750 nvme_queue_scan(&ctrl->ctrl);
751 nvme_queue_async_events(&ctrl->ctrl);
754 dev_info(ctrl->ctrl.device, "Successfully reconnected\n");
759 blk_mq_stop_hw_queues(ctrl->ctrl.admin_q);
761 /* Make sure we are not resetting/deleting */
762 if (ctrl->ctrl.state == NVME_CTRL_RECONNECTING) {
763 dev_info(ctrl->ctrl.device,
764 "Failed reconnect attempt, requeueing...\n");
765 queue_delayed_work(nvme_rdma_wq, &ctrl->reconnect_work,
766 ctrl->reconnect_delay * HZ);
770 static void nvme_rdma_error_recovery_work(struct work_struct *work)
772 struct nvme_rdma_ctrl *ctrl = container_of(work,
773 struct nvme_rdma_ctrl, err_work);
776 nvme_stop_keep_alive(&ctrl->ctrl);
778 for (i = 0; i < ctrl->queue_count; i++)
779 clear_bit(NVME_RDMA_Q_CONNECTED, &ctrl->queues[i].flags);
781 if (ctrl->queue_count > 1)
782 nvme_stop_queues(&ctrl->ctrl);
783 blk_mq_stop_hw_queues(ctrl->ctrl.admin_q);
785 /* We must take care of fastfail/requeue all our inflight requests */
786 if (ctrl->queue_count > 1)
787 blk_mq_tagset_busy_iter(&ctrl->tag_set,
788 nvme_cancel_request, &ctrl->ctrl);
789 blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
790 nvme_cancel_request, &ctrl->ctrl);
792 dev_info(ctrl->ctrl.device, "reconnecting in %d seconds\n",
793 ctrl->reconnect_delay);
795 queue_delayed_work(nvme_rdma_wq, &ctrl->reconnect_work,
796 ctrl->reconnect_delay * HZ);
799 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
801 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RECONNECTING))
804 queue_work(nvme_rdma_wq, &ctrl->err_work);
807 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
810 struct nvme_rdma_queue *queue = cq->cq_context;
811 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
813 if (ctrl->ctrl.state == NVME_CTRL_LIVE)
814 dev_info(ctrl->ctrl.device,
815 "%s for CQE 0x%p failed with status %s (%d)\n",
817 ib_wc_status_msg(wc->status), wc->status);
818 nvme_rdma_error_recovery(ctrl);
821 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
823 if (unlikely(wc->status != IB_WC_SUCCESS))
824 nvme_rdma_wr_error(cq, wc, "MEMREG");
827 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
829 if (unlikely(wc->status != IB_WC_SUCCESS))
830 nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
833 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
834 struct nvme_rdma_request *req)
836 struct ib_send_wr *bad_wr;
837 struct ib_send_wr wr = {
838 .opcode = IB_WR_LOCAL_INV,
842 .ex.invalidate_rkey = req->mr->rkey,
845 req->reg_cqe.done = nvme_rdma_inv_rkey_done;
846 wr.wr_cqe = &req->reg_cqe;
848 return ib_post_send(queue->qp, &wr, &bad_wr);
851 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
854 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
855 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
856 struct nvme_rdma_device *dev = queue->device;
857 struct ib_device *ibdev = dev->dev;
860 if (!blk_rq_bytes(rq))
863 if (req->mr->need_inval) {
864 res = nvme_rdma_inv_rkey(queue, req);
866 dev_err(ctrl->ctrl.device,
867 "Queueing INV WR for rkey %#x failed (%d)\n",
869 nvme_rdma_error_recovery(queue->ctrl);
873 ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
874 req->nents, rq_data_dir(rq) ==
875 WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
877 nvme_cleanup_cmd(rq);
878 sg_free_table_chained(&req->sg_table, true);
881 static int nvme_rdma_set_sg_null(struct nvme_command *c)
883 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
886 put_unaligned_le24(0, sg->length);
887 put_unaligned_le32(0, sg->key);
888 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
892 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
893 struct nvme_rdma_request *req, struct nvme_command *c)
895 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
897 req->sge[1].addr = sg_dma_address(req->sg_table.sgl);
898 req->sge[1].length = sg_dma_len(req->sg_table.sgl);
899 req->sge[1].lkey = queue->device->pd->local_dma_lkey;
901 sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
902 sg->length = cpu_to_le32(sg_dma_len(req->sg_table.sgl));
903 sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
905 req->inline_data = true;
910 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
911 struct nvme_rdma_request *req, struct nvme_command *c)
913 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
915 sg->addr = cpu_to_le64(sg_dma_address(req->sg_table.sgl));
916 put_unaligned_le24(sg_dma_len(req->sg_table.sgl), sg->length);
917 put_unaligned_le32(queue->device->mr->rkey, sg->key);
918 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
922 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
923 struct nvme_rdma_request *req, struct nvme_command *c,
926 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
929 nr = ib_map_mr_sg(req->mr, req->sg_table.sgl, count, NULL, PAGE_SIZE);
936 ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
938 req->reg_cqe.done = nvme_rdma_memreg_done;
939 memset(&req->reg_wr, 0, sizeof(req->reg_wr));
940 req->reg_wr.wr.opcode = IB_WR_REG_MR;
941 req->reg_wr.wr.wr_cqe = &req->reg_cqe;
942 req->reg_wr.wr.num_sge = 0;
943 req->reg_wr.mr = req->mr;
944 req->reg_wr.key = req->mr->rkey;
945 req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
946 IB_ACCESS_REMOTE_READ |
947 IB_ACCESS_REMOTE_WRITE;
949 req->mr->need_inval = true;
951 sg->addr = cpu_to_le64(req->mr->iova);
952 put_unaligned_le24(req->mr->length, sg->length);
953 put_unaligned_le32(req->mr->rkey, sg->key);
954 sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
955 NVME_SGL_FMT_INVALIDATE;
960 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
961 struct request *rq, unsigned int map_len,
962 struct nvme_command *c)
964 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
965 struct nvme_rdma_device *dev = queue->device;
966 struct ib_device *ibdev = dev->dev;
971 req->inline_data = false;
972 req->mr->need_inval = false;
974 c->common.flags |= NVME_CMD_SGL_METABUF;
976 if (!blk_rq_bytes(rq))
977 return nvme_rdma_set_sg_null(c);
979 req->sg_table.sgl = req->first_sgl;
980 ret = sg_alloc_table_chained(&req->sg_table, rq->nr_phys_segments,
985 nents = blk_rq_map_sg(rq->q, rq, req->sg_table.sgl);
986 BUG_ON(nents > rq->nr_phys_segments);
989 count = ib_dma_map_sg(ibdev, req->sg_table.sgl, nents,
990 rq_data_dir(rq) == WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
991 if (unlikely(count <= 0)) {
992 sg_free_table_chained(&req->sg_table, true);
997 if (rq_data_dir(rq) == WRITE &&
998 map_len <= nvme_rdma_inline_data_size(queue) &&
999 nvme_rdma_queue_idx(queue))
1000 return nvme_rdma_map_sg_inline(queue, req, c);
1002 if (!register_always)
1003 return nvme_rdma_map_sg_single(queue, req, c);
1006 return nvme_rdma_map_sg_fr(queue, req, c, count);
1009 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1011 if (unlikely(wc->status != IB_WC_SUCCESS))
1012 nvme_rdma_wr_error(cq, wc, "SEND");
1015 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1016 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1017 struct ib_send_wr *first, bool flush)
1019 struct ib_send_wr wr, *bad_wr;
1022 sge->addr = qe->dma;
1023 sge->length = sizeof(struct nvme_command),
1024 sge->lkey = queue->device->pd->local_dma_lkey;
1026 qe->cqe.done = nvme_rdma_send_done;
1029 wr.wr_cqe = &qe->cqe;
1031 wr.num_sge = num_sge;
1032 wr.opcode = IB_WR_SEND;
1036 * Unsignalled send completions are another giant desaster in the
1037 * IB Verbs spec: If we don't regularly post signalled sends
1038 * the send queue will fill up and only a QP reset will rescue us.
1039 * Would have been way to obvious to handle this in hardware or
1040 * at least the RDMA stack..
1042 * This messy and racy code sniplet is copy and pasted from the iSER
1043 * initiator, and the magic '32' comes from there as well.
1045 * Always signal the flushes. The magic request used for the flush
1046 * sequencer is not allocated in our driver's tagset and it's
1047 * triggered to be freed by blk_cleanup_queue(). So we need to
1048 * always mark it as signaled to ensure that the "wr_cqe", which is
1049 * embeded in request's payload, is not freed when __ib_process_cq()
1050 * calls wr_cqe->done().
1052 if ((++queue->sig_count % 32) == 0 || flush)
1053 wr.send_flags |= IB_SEND_SIGNALED;
1060 ret = ib_post_send(queue->qp, first, &bad_wr);
1062 dev_err(queue->ctrl->ctrl.device,
1063 "%s failed with error code %d\n", __func__, ret);
1068 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1069 struct nvme_rdma_qe *qe)
1071 struct ib_recv_wr wr, *bad_wr;
1075 list.addr = qe->dma;
1076 list.length = sizeof(struct nvme_completion);
1077 list.lkey = queue->device->pd->local_dma_lkey;
1079 qe->cqe.done = nvme_rdma_recv_done;
1082 wr.wr_cqe = &qe->cqe;
1086 ret = ib_post_recv(queue->qp, &wr, &bad_wr);
1088 dev_err(queue->ctrl->ctrl.device,
1089 "%s failed with error code %d\n", __func__, ret);
1094 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1096 u32 queue_idx = nvme_rdma_queue_idx(queue);
1099 return queue->ctrl->admin_tag_set.tags[queue_idx];
1100 return queue->ctrl->tag_set.tags[queue_idx - 1];
1103 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg, int aer_idx)
1105 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1106 struct nvme_rdma_queue *queue = &ctrl->queues[0];
1107 struct ib_device *dev = queue->device->dev;
1108 struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1109 struct nvme_command *cmd = sqe->data;
1113 if (WARN_ON_ONCE(aer_idx != 0))
1116 ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1118 memset(cmd, 0, sizeof(*cmd));
1119 cmd->common.opcode = nvme_admin_async_event;
1120 cmd->common.command_id = NVME_RDMA_AQ_BLKMQ_DEPTH;
1121 cmd->common.flags |= NVME_CMD_SGL_METABUF;
1122 nvme_rdma_set_sg_null(cmd);
1124 ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1127 ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL, false);
1131 static int nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1132 struct nvme_completion *cqe, struct ib_wc *wc, int tag)
1134 u16 status = le16_to_cpu(cqe->status);
1136 struct nvme_rdma_request *req;
1141 rq = blk_mq_tag_to_rq(nvme_rdma_tagset(queue), cqe->command_id);
1143 dev_err(queue->ctrl->ctrl.device,
1144 "tag 0x%x on QP %#x not found\n",
1145 cqe->command_id, queue->qp->qp_num);
1146 nvme_rdma_error_recovery(queue->ctrl);
1149 req = blk_mq_rq_to_pdu(rq);
1151 if (rq->cmd_type == REQ_TYPE_DRV_PRIV && rq->special)
1152 memcpy(rq->special, cqe, sizeof(*cqe));
1157 if ((wc->wc_flags & IB_WC_WITH_INVALIDATE) &&
1158 wc->ex.invalidate_rkey == req->mr->rkey)
1159 req->mr->need_inval = false;
1161 blk_mq_complete_request(rq, status);
1166 static int __nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc, int tag)
1168 struct nvme_rdma_qe *qe =
1169 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1170 struct nvme_rdma_queue *queue = cq->cq_context;
1171 struct ib_device *ibdev = queue->device->dev;
1172 struct nvme_completion *cqe = qe->data;
1173 const size_t len = sizeof(struct nvme_completion);
1176 if (unlikely(wc->status != IB_WC_SUCCESS)) {
1177 nvme_rdma_wr_error(cq, wc, "RECV");
1181 ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1183 * AEN requests are special as they don't time out and can
1184 * survive any kind of queue freeze and often don't respond to
1185 * aborts. We don't even bother to allocate a struct request
1186 * for them but rather special case them here.
1188 if (unlikely(nvme_rdma_queue_idx(queue) == 0 &&
1189 cqe->command_id >= NVME_RDMA_AQ_BLKMQ_DEPTH))
1190 nvme_complete_async_event(&queue->ctrl->ctrl, cqe);
1192 ret = nvme_rdma_process_nvme_rsp(queue, cqe, wc, tag);
1193 ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1195 nvme_rdma_post_recv(queue, qe);
1199 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1201 __nvme_rdma_recv_done(cq, wc, -1);
1204 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1208 for (i = 0; i < queue->queue_size; i++) {
1209 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1211 goto out_destroy_queue_ib;
1216 out_destroy_queue_ib:
1217 nvme_rdma_destroy_queue_ib(queue);
1221 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1222 struct rdma_cm_event *ev)
1224 if (ev->param.conn.private_data_len) {
1225 struct nvme_rdma_cm_rej *rej =
1226 (struct nvme_rdma_cm_rej *)ev->param.conn.private_data;
1228 dev_err(queue->ctrl->ctrl.device,
1229 "Connect rejected, status %d.", le16_to_cpu(rej->sts));
1230 /* XXX: Think of something clever to do here... */
1232 dev_err(queue->ctrl->ctrl.device,
1233 "Connect rejected, no private data.\n");
1239 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1241 struct nvme_rdma_device *dev;
1244 dev = nvme_rdma_find_get_device(queue->cm_id);
1246 dev_err(queue->cm_id->device->dma_device,
1247 "no client data found!\n");
1248 return -ECONNREFUSED;
1251 ret = nvme_rdma_create_queue_ib(queue, dev);
1253 nvme_rdma_dev_put(dev);
1257 ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1259 dev_err(queue->ctrl->ctrl.device,
1260 "rdma_resolve_route failed (%d).\n",
1262 goto out_destroy_queue;
1268 nvme_rdma_destroy_queue_ib(queue);
1273 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1275 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1276 struct rdma_conn_param param = { };
1277 struct nvme_rdma_cm_req priv = { };
1280 param.qp_num = queue->qp->qp_num;
1281 param.flow_control = 1;
1283 param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1284 /* maximum retry count */
1285 param.retry_count = 7;
1286 param.rnr_retry_count = 7;
1287 param.private_data = &priv;
1288 param.private_data_len = sizeof(priv);
1290 priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1291 priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1293 * set the admin queue depth to the minimum size
1294 * specified by the Fabrics standard.
1296 if (priv.qid == 0) {
1297 priv.hrqsize = cpu_to_le16(NVMF_AQ_DEPTH);
1298 priv.hsqsize = cpu_to_le16(NVMF_AQ_DEPTH - 1);
1301 * current interpretation of the fabrics spec
1302 * is at minimum you make hrqsize sqsize+1, or a
1303 * 1's based representation of sqsize.
1305 priv.hrqsize = cpu_to_le16(queue->queue_size);
1306 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1309 ret = rdma_connect(queue->cm_id, ¶m);
1311 dev_err(ctrl->ctrl.device,
1312 "rdma_connect failed (%d).\n", ret);
1313 goto out_destroy_queue_ib;
1318 out_destroy_queue_ib:
1319 nvme_rdma_destroy_queue_ib(queue);
1323 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1324 struct rdma_cm_event *ev)
1326 struct nvme_rdma_queue *queue = cm_id->context;
1329 dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1330 rdma_event_msg(ev->event), ev->event,
1333 switch (ev->event) {
1334 case RDMA_CM_EVENT_ADDR_RESOLVED:
1335 cm_error = nvme_rdma_addr_resolved(queue);
1337 case RDMA_CM_EVENT_ROUTE_RESOLVED:
1338 cm_error = nvme_rdma_route_resolved(queue);
1340 case RDMA_CM_EVENT_ESTABLISHED:
1341 queue->cm_error = nvme_rdma_conn_established(queue);
1342 /* complete cm_done regardless of success/failure */
1343 complete(&queue->cm_done);
1345 case RDMA_CM_EVENT_REJECTED:
1346 cm_error = nvme_rdma_conn_rejected(queue, ev);
1348 case RDMA_CM_EVENT_ADDR_ERROR:
1349 case RDMA_CM_EVENT_ROUTE_ERROR:
1350 case RDMA_CM_EVENT_CONNECT_ERROR:
1351 case RDMA_CM_EVENT_UNREACHABLE:
1352 dev_dbg(queue->ctrl->ctrl.device,
1353 "CM error event %d\n", ev->event);
1354 cm_error = -ECONNRESET;
1356 case RDMA_CM_EVENT_DISCONNECTED:
1357 case RDMA_CM_EVENT_ADDR_CHANGE:
1358 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1359 dev_dbg(queue->ctrl->ctrl.device,
1360 "disconnect received - connection closed\n");
1361 nvme_rdma_error_recovery(queue->ctrl);
1363 case RDMA_CM_EVENT_DEVICE_REMOVAL:
1364 /* device removal is handled via the ib_client API */
1367 dev_err(queue->ctrl->ctrl.device,
1368 "Unexpected RDMA CM event (%d)\n", ev->event);
1369 nvme_rdma_error_recovery(queue->ctrl);
1374 queue->cm_error = cm_error;
1375 complete(&queue->cm_done);
1381 static enum blk_eh_timer_return
1382 nvme_rdma_timeout(struct request *rq, bool reserved)
1384 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1386 /* queue error recovery */
1387 nvme_rdma_error_recovery(req->queue->ctrl);
1389 /* fail with DNR on cmd timeout */
1390 rq->errors = NVME_SC_ABORT_REQ | NVME_SC_DNR;
1392 return BLK_EH_HANDLED;
1395 static int nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
1396 const struct blk_mq_queue_data *bd)
1398 struct nvme_ns *ns = hctx->queue->queuedata;
1399 struct nvme_rdma_queue *queue = hctx->driver_data;
1400 struct request *rq = bd->rq;
1401 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1402 struct nvme_rdma_qe *sqe = &req->sqe;
1403 struct nvme_command *c = sqe->data;
1405 struct ib_device *dev;
1406 unsigned int map_len;
1409 WARN_ON_ONCE(rq->tag < 0);
1411 dev = queue->device->dev;
1412 ib_dma_sync_single_for_cpu(dev, sqe->dma,
1413 sizeof(struct nvme_command), DMA_TO_DEVICE);
1415 ret = nvme_setup_cmd(ns, rq, c);
1419 c->common.command_id = rq->tag;
1420 blk_mq_start_request(rq);
1422 map_len = nvme_map_len(rq);
1423 ret = nvme_rdma_map_data(queue, rq, map_len, c);
1425 dev_err(queue->ctrl->ctrl.device,
1426 "Failed to map data (%d)\n", ret);
1427 nvme_cleanup_cmd(rq);
1431 ib_dma_sync_single_for_device(dev, sqe->dma,
1432 sizeof(struct nvme_command), DMA_TO_DEVICE);
1434 if (rq->cmd_type == REQ_TYPE_FS && req_op(rq) == REQ_OP_FLUSH)
1436 ret = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
1437 req->mr->need_inval ? &req->reg_wr.wr : NULL, flush);
1439 nvme_rdma_unmap_data(queue, rq);
1443 return BLK_MQ_RQ_QUEUE_OK;
1445 return (ret == -ENOMEM || ret == -EAGAIN) ?
1446 BLK_MQ_RQ_QUEUE_BUSY : BLK_MQ_RQ_QUEUE_ERROR;
1449 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx, unsigned int tag)
1451 struct nvme_rdma_queue *queue = hctx->driver_data;
1452 struct ib_cq *cq = queue->ib_cq;
1456 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1457 while (ib_poll_cq(cq, 1, &wc) > 0) {
1458 struct ib_cqe *cqe = wc.wr_cqe;
1461 if (cqe->done == nvme_rdma_recv_done)
1462 found |= __nvme_rdma_recv_done(cq, &wc, tag);
1471 static void nvme_rdma_complete_rq(struct request *rq)
1473 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1474 struct nvme_rdma_queue *queue = req->queue;
1477 nvme_rdma_unmap_data(queue, rq);
1479 if (unlikely(rq->errors)) {
1480 if (nvme_req_needs_retry(rq, rq->errors)) {
1481 nvme_requeue_req(rq);
1485 if (rq->cmd_type == REQ_TYPE_DRV_PRIV)
1488 error = nvme_error_status(rq->errors);
1491 blk_mq_end_request(rq, error);
1494 static struct blk_mq_ops nvme_rdma_mq_ops = {
1495 .queue_rq = nvme_rdma_queue_rq,
1496 .complete = nvme_rdma_complete_rq,
1497 .map_queue = blk_mq_map_queue,
1498 .init_request = nvme_rdma_init_request,
1499 .exit_request = nvme_rdma_exit_request,
1500 .reinit_request = nvme_rdma_reinit_request,
1501 .init_hctx = nvme_rdma_init_hctx,
1502 .poll = nvme_rdma_poll,
1503 .timeout = nvme_rdma_timeout,
1506 static struct blk_mq_ops nvme_rdma_admin_mq_ops = {
1507 .queue_rq = nvme_rdma_queue_rq,
1508 .complete = nvme_rdma_complete_rq,
1509 .map_queue = blk_mq_map_queue,
1510 .init_request = nvme_rdma_init_admin_request,
1511 .exit_request = nvme_rdma_exit_admin_request,
1512 .reinit_request = nvme_rdma_reinit_request,
1513 .init_hctx = nvme_rdma_init_admin_hctx,
1514 .timeout = nvme_rdma_timeout,
1517 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl)
1521 error = nvme_rdma_init_queue(ctrl, 0, NVMF_AQ_DEPTH);
1525 ctrl->device = ctrl->queues[0].device;
1528 * We need a reference on the device as long as the tag_set is alive,
1529 * as the MRs in the request structures need a valid ib_device.
1532 if (!nvme_rdma_dev_get(ctrl->device))
1533 goto out_free_queue;
1535 ctrl->max_fr_pages = min_t(u32, NVME_RDMA_MAX_SEGMENTS,
1536 ctrl->device->dev->attrs.max_fast_reg_page_list_len);
1538 memset(&ctrl->admin_tag_set, 0, sizeof(ctrl->admin_tag_set));
1539 ctrl->admin_tag_set.ops = &nvme_rdma_admin_mq_ops;
1540 ctrl->admin_tag_set.queue_depth = NVME_RDMA_AQ_BLKMQ_DEPTH;
1541 ctrl->admin_tag_set.reserved_tags = 2; /* connect + keep-alive */
1542 ctrl->admin_tag_set.numa_node = NUMA_NO_NODE;
1543 ctrl->admin_tag_set.cmd_size = sizeof(struct nvme_rdma_request) +
1544 SG_CHUNK_SIZE * sizeof(struct scatterlist);
1545 ctrl->admin_tag_set.driver_data = ctrl;
1546 ctrl->admin_tag_set.nr_hw_queues = 1;
1547 ctrl->admin_tag_set.timeout = ADMIN_TIMEOUT;
1549 error = blk_mq_alloc_tag_set(&ctrl->admin_tag_set);
1553 ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
1554 if (IS_ERR(ctrl->ctrl.admin_q)) {
1555 error = PTR_ERR(ctrl->ctrl.admin_q);
1556 goto out_free_tagset;
1559 error = nvmf_connect_admin_queue(&ctrl->ctrl);
1561 goto out_cleanup_queue;
1563 error = nvmf_reg_read64(&ctrl->ctrl, NVME_REG_CAP, &ctrl->cap);
1565 dev_err(ctrl->ctrl.device,
1566 "prop_get NVME_REG_CAP failed\n");
1567 goto out_cleanup_queue;
1571 min_t(int, NVME_CAP_MQES(ctrl->cap) + 1, ctrl->ctrl.sqsize);
1573 error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->cap);
1575 goto out_cleanup_queue;
1577 ctrl->ctrl.max_hw_sectors =
1578 (ctrl->max_fr_pages - 1) << (PAGE_SHIFT - 9);
1580 error = nvme_init_identify(&ctrl->ctrl);
1582 goto out_cleanup_queue;
1584 error = nvme_rdma_alloc_qe(ctrl->queues[0].device->dev,
1585 &ctrl->async_event_sqe, sizeof(struct nvme_command),
1588 goto out_cleanup_queue;
1590 nvme_start_keep_alive(&ctrl->ctrl);
1595 blk_cleanup_queue(ctrl->ctrl.admin_q);
1597 /* disconnect and drain the queue before freeing the tagset */
1598 nvme_rdma_stop_queue(&ctrl->queues[0]);
1599 blk_mq_free_tag_set(&ctrl->admin_tag_set);
1601 nvme_rdma_dev_put(ctrl->device);
1603 nvme_rdma_free_queue(&ctrl->queues[0]);
1607 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl)
1609 nvme_stop_keep_alive(&ctrl->ctrl);
1610 cancel_work_sync(&ctrl->err_work);
1611 cancel_delayed_work_sync(&ctrl->reconnect_work);
1613 if (ctrl->queue_count > 1) {
1614 nvme_stop_queues(&ctrl->ctrl);
1615 blk_mq_tagset_busy_iter(&ctrl->tag_set,
1616 nvme_cancel_request, &ctrl->ctrl);
1617 nvme_rdma_free_io_queues(ctrl);
1620 if (test_bit(NVME_RDMA_Q_CONNECTED, &ctrl->queues[0].flags))
1621 nvme_shutdown_ctrl(&ctrl->ctrl);
1623 blk_mq_stop_hw_queues(ctrl->ctrl.admin_q);
1624 blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
1625 nvme_cancel_request, &ctrl->ctrl);
1626 nvme_rdma_destroy_admin_queue(ctrl);
1629 static void __nvme_rdma_remove_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
1631 nvme_uninit_ctrl(&ctrl->ctrl);
1633 nvme_rdma_shutdown_ctrl(ctrl);
1635 if (ctrl->ctrl.tagset) {
1636 blk_cleanup_queue(ctrl->ctrl.connect_q);
1637 blk_mq_free_tag_set(&ctrl->tag_set);
1638 nvme_rdma_dev_put(ctrl->device);
1641 nvme_put_ctrl(&ctrl->ctrl);
1644 static void nvme_rdma_del_ctrl_work(struct work_struct *work)
1646 struct nvme_rdma_ctrl *ctrl = container_of(work,
1647 struct nvme_rdma_ctrl, delete_work);
1649 __nvme_rdma_remove_ctrl(ctrl, true);
1652 static int __nvme_rdma_del_ctrl(struct nvme_rdma_ctrl *ctrl)
1654 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_DELETING))
1657 if (!queue_work(nvme_rdma_wq, &ctrl->delete_work))
1663 static int nvme_rdma_del_ctrl(struct nvme_ctrl *nctrl)
1665 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
1669 * Keep a reference until all work is flushed since
1670 * __nvme_rdma_del_ctrl can free the ctrl mem
1672 if (!kref_get_unless_zero(&ctrl->ctrl.kref))
1674 ret = __nvme_rdma_del_ctrl(ctrl);
1676 flush_work(&ctrl->delete_work);
1677 nvme_put_ctrl(&ctrl->ctrl);
1681 static void nvme_rdma_remove_ctrl_work(struct work_struct *work)
1683 struct nvme_rdma_ctrl *ctrl = container_of(work,
1684 struct nvme_rdma_ctrl, delete_work);
1686 __nvme_rdma_remove_ctrl(ctrl, false);
1689 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
1691 struct nvme_rdma_ctrl *ctrl = container_of(work,
1692 struct nvme_rdma_ctrl, reset_work);
1696 nvme_rdma_shutdown_ctrl(ctrl);
1698 ret = nvme_rdma_configure_admin_queue(ctrl);
1700 /* ctrl is already shutdown, just remove the ctrl */
1701 INIT_WORK(&ctrl->delete_work, nvme_rdma_remove_ctrl_work);
1705 if (ctrl->queue_count > 1) {
1706 ret = blk_mq_reinit_tagset(&ctrl->tag_set);
1710 ret = nvme_rdma_init_io_queues(ctrl);
1714 ret = nvme_rdma_connect_io_queues(ctrl);
1719 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1720 WARN_ON_ONCE(!changed);
1722 if (ctrl->queue_count > 1) {
1723 nvme_start_queues(&ctrl->ctrl);
1724 nvme_queue_scan(&ctrl->ctrl);
1725 nvme_queue_async_events(&ctrl->ctrl);
1731 /* Deleting this dead controller... */
1732 dev_warn(ctrl->ctrl.device, "Removing after reset failure\n");
1733 WARN_ON(!queue_work(nvme_rdma_wq, &ctrl->delete_work));
1736 static int nvme_rdma_reset_ctrl(struct nvme_ctrl *nctrl)
1738 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
1740 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
1743 if (!queue_work(nvme_rdma_wq, &ctrl->reset_work))
1746 flush_work(&ctrl->reset_work);
1751 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
1753 .module = THIS_MODULE,
1755 .reg_read32 = nvmf_reg_read32,
1756 .reg_read64 = nvmf_reg_read64,
1757 .reg_write32 = nvmf_reg_write32,
1758 .reset_ctrl = nvme_rdma_reset_ctrl,
1759 .free_ctrl = nvme_rdma_free_ctrl,
1760 .submit_async_event = nvme_rdma_submit_async_event,
1761 .delete_ctrl = nvme_rdma_del_ctrl,
1762 .get_subsysnqn = nvmf_get_subsysnqn,
1763 .get_address = nvmf_get_address,
1766 static int nvme_rdma_create_io_queues(struct nvme_rdma_ctrl *ctrl)
1768 struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
1771 ret = nvme_set_queue_count(&ctrl->ctrl, &opts->nr_io_queues);
1775 ctrl->queue_count = opts->nr_io_queues + 1;
1776 if (ctrl->queue_count < 2)
1779 dev_info(ctrl->ctrl.device,
1780 "creating %d I/O queues.\n", opts->nr_io_queues);
1782 ret = nvme_rdma_init_io_queues(ctrl);
1787 * We need a reference on the device as long as the tag_set is alive,
1788 * as the MRs in the request structures need a valid ib_device.
1791 if (!nvme_rdma_dev_get(ctrl->device))
1792 goto out_free_io_queues;
1794 memset(&ctrl->tag_set, 0, sizeof(ctrl->tag_set));
1795 ctrl->tag_set.ops = &nvme_rdma_mq_ops;
1796 ctrl->tag_set.queue_depth = ctrl->ctrl.opts->queue_size;
1797 ctrl->tag_set.reserved_tags = 1; /* fabric connect */
1798 ctrl->tag_set.numa_node = NUMA_NO_NODE;
1799 ctrl->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
1800 ctrl->tag_set.cmd_size = sizeof(struct nvme_rdma_request) +
1801 SG_CHUNK_SIZE * sizeof(struct scatterlist);
1802 ctrl->tag_set.driver_data = ctrl;
1803 ctrl->tag_set.nr_hw_queues = ctrl->queue_count - 1;
1804 ctrl->tag_set.timeout = NVME_IO_TIMEOUT;
1806 ret = blk_mq_alloc_tag_set(&ctrl->tag_set);
1809 ctrl->ctrl.tagset = &ctrl->tag_set;
1811 ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
1812 if (IS_ERR(ctrl->ctrl.connect_q)) {
1813 ret = PTR_ERR(ctrl->ctrl.connect_q);
1814 goto out_free_tag_set;
1817 ret = nvme_rdma_connect_io_queues(ctrl);
1819 goto out_cleanup_connect_q;
1823 out_cleanup_connect_q:
1824 blk_cleanup_queue(ctrl->ctrl.connect_q);
1826 blk_mq_free_tag_set(&ctrl->tag_set);
1828 nvme_rdma_dev_put(ctrl->device);
1830 nvme_rdma_free_io_queues(ctrl);
1834 static int nvme_rdma_parse_ipaddr(struct sockaddr_in *in_addr, char *p)
1836 u8 *addr = (u8 *)&in_addr->sin_addr.s_addr;
1837 size_t buflen = strlen(p);
1839 /* XXX: handle IPv6 addresses */
1841 if (buflen > INET_ADDRSTRLEN)
1843 if (in4_pton(p, buflen, addr, '\0', NULL) == 0)
1845 in_addr->sin_family = AF_INET;
1849 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
1850 struct nvmf_ctrl_options *opts)
1852 struct nvme_rdma_ctrl *ctrl;
1856 ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1858 return ERR_PTR(-ENOMEM);
1859 ctrl->ctrl.opts = opts;
1860 INIT_LIST_HEAD(&ctrl->list);
1862 ret = nvme_rdma_parse_ipaddr(&ctrl->addr_in, opts->traddr);
1864 pr_err("malformed IP address passed: %s\n", opts->traddr);
1868 if (opts->mask & NVMF_OPT_TRSVCID) {
1871 ret = kstrtou16(opts->trsvcid, 0, &port);
1875 ctrl->addr_in.sin_port = cpu_to_be16(port);
1877 ctrl->addr_in.sin_port = cpu_to_be16(NVME_RDMA_IP_PORT);
1880 ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
1881 0 /* no quirks, we're perfect! */);
1885 ctrl->reconnect_delay = opts->reconnect_delay;
1886 INIT_DELAYED_WORK(&ctrl->reconnect_work,
1887 nvme_rdma_reconnect_ctrl_work);
1888 INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
1889 INIT_WORK(&ctrl->delete_work, nvme_rdma_del_ctrl_work);
1890 INIT_WORK(&ctrl->reset_work, nvme_rdma_reset_ctrl_work);
1891 spin_lock_init(&ctrl->lock);
1893 ctrl->queue_count = opts->nr_io_queues + 1; /* +1 for admin queue */
1894 ctrl->ctrl.sqsize = opts->queue_size - 1;
1895 ctrl->ctrl.kato = opts->kato;
1898 ctrl->queues = kcalloc(ctrl->queue_count, sizeof(*ctrl->queues),
1901 goto out_uninit_ctrl;
1903 ret = nvme_rdma_configure_admin_queue(ctrl);
1905 goto out_kfree_queues;
1907 /* sanity check icdoff */
1908 if (ctrl->ctrl.icdoff) {
1909 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
1910 goto out_remove_admin_queue;
1913 /* sanity check keyed sgls */
1914 if (!(ctrl->ctrl.sgls & (1 << 20))) {
1915 dev_err(ctrl->ctrl.device, "Mandatory keyed sgls are not support\n");
1916 goto out_remove_admin_queue;
1919 if (opts->queue_size > ctrl->ctrl.maxcmd) {
1920 /* warn if maxcmd is lower than queue_size */
1921 dev_warn(ctrl->ctrl.device,
1922 "queue_size %zu > ctrl maxcmd %u, clamping down\n",
1923 opts->queue_size, ctrl->ctrl.maxcmd);
1924 opts->queue_size = ctrl->ctrl.maxcmd;
1927 if (opts->nr_io_queues) {
1928 ret = nvme_rdma_create_io_queues(ctrl);
1930 goto out_remove_admin_queue;
1933 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1934 WARN_ON_ONCE(!changed);
1936 dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISp\n",
1937 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
1939 kref_get(&ctrl->ctrl.kref);
1941 mutex_lock(&nvme_rdma_ctrl_mutex);
1942 list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
1943 mutex_unlock(&nvme_rdma_ctrl_mutex);
1945 if (opts->nr_io_queues) {
1946 nvme_queue_scan(&ctrl->ctrl);
1947 nvme_queue_async_events(&ctrl->ctrl);
1952 out_remove_admin_queue:
1953 nvme_stop_keep_alive(&ctrl->ctrl);
1954 nvme_rdma_destroy_admin_queue(ctrl);
1956 kfree(ctrl->queues);
1958 nvme_uninit_ctrl(&ctrl->ctrl);
1959 nvme_put_ctrl(&ctrl->ctrl);
1962 return ERR_PTR(ret);
1965 return ERR_PTR(ret);
1968 static struct nvmf_transport_ops nvme_rdma_transport = {
1970 .required_opts = NVMF_OPT_TRADDR,
1971 .allowed_opts = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY,
1972 .create_ctrl = nvme_rdma_create_ctrl,
1975 static void nvme_rdma_add_one(struct ib_device *ib_device)
1979 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
1981 struct nvme_rdma_ctrl *ctrl;
1983 /* Delete all controllers using this device */
1984 mutex_lock(&nvme_rdma_ctrl_mutex);
1985 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
1986 if (ctrl->device->dev != ib_device)
1988 dev_info(ctrl->ctrl.device,
1989 "Removing ctrl: NQN \"%s\", addr %pISp\n",
1990 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
1991 __nvme_rdma_del_ctrl(ctrl);
1993 mutex_unlock(&nvme_rdma_ctrl_mutex);
1995 flush_workqueue(nvme_rdma_wq);
1998 static struct ib_client nvme_rdma_ib_client = {
1999 .name = "nvme_rdma",
2000 .add = nvme_rdma_add_one,
2001 .remove = nvme_rdma_remove_one
2004 static int __init nvme_rdma_init_module(void)
2008 nvme_rdma_wq = create_workqueue("nvme_rdma_wq");
2012 ret = ib_register_client(&nvme_rdma_ib_client);
2014 destroy_workqueue(nvme_rdma_wq);
2018 nvmf_register_transport(&nvme_rdma_transport);
2022 static void __exit nvme_rdma_cleanup_module(void)
2024 nvmf_unregister_transport(&nvme_rdma_transport);
2025 ib_unregister_client(&nvme_rdma_ib_client);
2026 destroy_workqueue(nvme_rdma_wq);
2029 module_init(nvme_rdma_init_module);
2030 module_exit(nvme_rdma_cleanup_module);
2032 MODULE_LICENSE("GPL v2");