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 <linux/nvme-rdma.h>
37 #define NVME_RDMA_CONNECT_TIMEOUT_MS 3000 /* 3 second */
39 #define NVME_RDMA_MAX_SEGMENT_SIZE 0xffffff /* 24-bit SGL field */
41 #define NVME_RDMA_MAX_SEGMENTS 256
43 #define NVME_RDMA_MAX_INLINE_SEGMENTS 1
46 * We handle AEN commands ourselves and don't even let the
47 * block layer know about them.
49 #define NVME_RDMA_NR_AEN_COMMANDS 1
50 #define NVME_RDMA_AQ_BLKMQ_DEPTH \
51 (NVMF_AQ_DEPTH - NVME_RDMA_NR_AEN_COMMANDS)
53 struct nvme_rdma_device {
54 struct ib_device *dev;
57 struct list_head entry;
66 struct nvme_rdma_queue;
67 struct nvme_rdma_request {
68 struct nvme_request req;
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 {
84 NVME_RDMA_Q_DELETING = 1,
87 struct nvme_rdma_queue {
88 struct nvme_rdma_qe *rsp_ring;
91 size_t cmnd_capsule_len;
92 struct nvme_rdma_ctrl *ctrl;
93 struct nvme_rdma_device *device;
98 struct rdma_cm_id *cm_id;
100 struct completion cm_done;
103 struct nvme_rdma_ctrl {
104 /* read only in the hot path */
105 struct nvme_rdma_queue *queues;
108 /* other member variables */
109 struct blk_mq_tag_set tag_set;
110 struct work_struct delete_work;
111 struct work_struct reset_work;
112 struct work_struct err_work;
114 struct nvme_rdma_qe async_event_sqe;
116 struct delayed_work reconnect_work;
118 struct list_head list;
120 struct blk_mq_tag_set admin_tag_set;
121 struct nvme_rdma_device *device;
126 struct sockaddr_storage addr;
127 struct sockaddr_storage src_addr;
129 struct nvme_ctrl ctrl;
132 static inline struct nvme_rdma_ctrl *to_rdma_ctrl(struct nvme_ctrl *ctrl)
134 return container_of(ctrl, struct nvme_rdma_ctrl, ctrl);
137 static LIST_HEAD(device_list);
138 static DEFINE_MUTEX(device_list_mutex);
140 static LIST_HEAD(nvme_rdma_ctrl_list);
141 static DEFINE_MUTEX(nvme_rdma_ctrl_mutex);
144 * Disabling this option makes small I/O goes faster, but is fundamentally
145 * unsafe. With it turned off we will have to register a global rkey that
146 * allows read and write access to all physical memory.
148 static bool register_always = true;
149 module_param(register_always, bool, 0444);
150 MODULE_PARM_DESC(register_always,
151 "Use memory registration even for contiguous memory regions");
153 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
154 struct rdma_cm_event *event);
155 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
157 /* XXX: really should move to a generic header sooner or later.. */
158 static inline void put_unaligned_le24(u32 val, u8 *p)
165 static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue *queue)
167 return queue - queue->ctrl->queues;
170 static inline size_t nvme_rdma_inline_data_size(struct nvme_rdma_queue *queue)
172 return queue->cmnd_capsule_len - sizeof(struct nvme_command);
175 static void nvme_rdma_free_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
176 size_t capsule_size, enum dma_data_direction dir)
178 ib_dma_unmap_single(ibdev, qe->dma, capsule_size, dir);
182 static int nvme_rdma_alloc_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
183 size_t capsule_size, enum dma_data_direction dir)
185 qe->data = kzalloc(capsule_size, GFP_KERNEL);
189 qe->dma = ib_dma_map_single(ibdev, qe->data, capsule_size, dir);
190 if (ib_dma_mapping_error(ibdev, qe->dma)) {
198 static void nvme_rdma_free_ring(struct ib_device *ibdev,
199 struct nvme_rdma_qe *ring, size_t ib_queue_size,
200 size_t capsule_size, enum dma_data_direction dir)
204 for (i = 0; i < ib_queue_size; i++)
205 nvme_rdma_free_qe(ibdev, &ring[i], capsule_size, dir);
209 static struct nvme_rdma_qe *nvme_rdma_alloc_ring(struct ib_device *ibdev,
210 size_t ib_queue_size, size_t capsule_size,
211 enum dma_data_direction dir)
213 struct nvme_rdma_qe *ring;
216 ring = kcalloc(ib_queue_size, sizeof(struct nvme_rdma_qe), GFP_KERNEL);
220 for (i = 0; i < ib_queue_size; i++) {
221 if (nvme_rdma_alloc_qe(ibdev, &ring[i], capsule_size, dir))
228 nvme_rdma_free_ring(ibdev, ring, i, capsule_size, dir);
232 static void nvme_rdma_qp_event(struct ib_event *event, void *context)
234 pr_debug("QP event %s (%d)\n",
235 ib_event_msg(event->event), event->event);
239 static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue *queue)
241 wait_for_completion_interruptible_timeout(&queue->cm_done,
242 msecs_to_jiffies(NVME_RDMA_CONNECT_TIMEOUT_MS) + 1);
243 return queue->cm_error;
246 static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor)
248 struct nvme_rdma_device *dev = queue->device;
249 struct ib_qp_init_attr init_attr;
252 memset(&init_attr, 0, sizeof(init_attr));
253 init_attr.event_handler = nvme_rdma_qp_event;
255 init_attr.cap.max_send_wr = factor * queue->queue_size + 1;
257 init_attr.cap.max_recv_wr = queue->queue_size + 1;
258 init_attr.cap.max_recv_sge = 1;
259 init_attr.cap.max_send_sge = 1 + NVME_RDMA_MAX_INLINE_SEGMENTS;
260 init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
261 init_attr.qp_type = IB_QPT_RC;
262 init_attr.send_cq = queue->ib_cq;
263 init_attr.recv_cq = queue->ib_cq;
265 ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr);
267 queue->qp = queue->cm_id->qp;
271 static int nvme_rdma_reinit_request(void *data, struct request *rq)
273 struct nvme_rdma_ctrl *ctrl = data;
274 struct nvme_rdma_device *dev = ctrl->device;
275 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
278 if (!req->mr->need_inval)
281 ib_dereg_mr(req->mr);
283 req->mr = ib_alloc_mr(dev->pd, IB_MR_TYPE_MEM_REG,
285 if (IS_ERR(req->mr)) {
286 ret = PTR_ERR(req->mr);
291 req->mr->need_inval = false;
297 static void nvme_rdma_exit_request(struct blk_mq_tag_set *set,
298 struct request *rq, unsigned int hctx_idx)
300 struct nvme_rdma_ctrl *ctrl = set->driver_data;
301 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
302 int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
303 struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
304 struct nvme_rdma_device *dev = queue->device;
307 ib_dereg_mr(req->mr);
309 nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
313 static int nvme_rdma_init_request(struct blk_mq_tag_set *set,
314 struct request *rq, unsigned int hctx_idx,
315 unsigned int numa_node)
317 struct nvme_rdma_ctrl *ctrl = set->driver_data;
318 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
319 int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
320 struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
321 struct nvme_rdma_device *dev = queue->device;
322 struct ib_device *ibdev = dev->dev;
325 ret = nvme_rdma_alloc_qe(ibdev, &req->sqe, sizeof(struct nvme_command),
330 req->mr = ib_alloc_mr(dev->pd, IB_MR_TYPE_MEM_REG,
332 if (IS_ERR(req->mr)) {
333 ret = PTR_ERR(req->mr);
342 nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
347 static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
348 unsigned int hctx_idx)
350 struct nvme_rdma_ctrl *ctrl = data;
351 struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1];
353 BUG_ON(hctx_idx >= ctrl->queue_count);
355 hctx->driver_data = queue;
359 static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
360 unsigned int hctx_idx)
362 struct nvme_rdma_ctrl *ctrl = data;
363 struct nvme_rdma_queue *queue = &ctrl->queues[0];
365 BUG_ON(hctx_idx != 0);
367 hctx->driver_data = queue;
371 static void nvme_rdma_free_dev(struct kref *ref)
373 struct nvme_rdma_device *ndev =
374 container_of(ref, struct nvme_rdma_device, ref);
376 mutex_lock(&device_list_mutex);
377 list_del(&ndev->entry);
378 mutex_unlock(&device_list_mutex);
380 ib_dealloc_pd(ndev->pd);
384 static void nvme_rdma_dev_put(struct nvme_rdma_device *dev)
386 kref_put(&dev->ref, nvme_rdma_free_dev);
389 static int nvme_rdma_dev_get(struct nvme_rdma_device *dev)
391 return kref_get_unless_zero(&dev->ref);
394 static struct nvme_rdma_device *
395 nvme_rdma_find_get_device(struct rdma_cm_id *cm_id)
397 struct nvme_rdma_device *ndev;
399 mutex_lock(&device_list_mutex);
400 list_for_each_entry(ndev, &device_list, entry) {
401 if (ndev->dev->node_guid == cm_id->device->node_guid &&
402 nvme_rdma_dev_get(ndev))
406 ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
410 ndev->dev = cm_id->device;
411 kref_init(&ndev->ref);
413 ndev->pd = ib_alloc_pd(ndev->dev,
414 register_always ? 0 : IB_PD_UNSAFE_GLOBAL_RKEY);
415 if (IS_ERR(ndev->pd))
418 if (!(ndev->dev->attrs.device_cap_flags &
419 IB_DEVICE_MEM_MGT_EXTENSIONS)) {
420 dev_err(&ndev->dev->dev,
421 "Memory registrations not supported.\n");
425 list_add(&ndev->entry, &device_list);
427 mutex_unlock(&device_list_mutex);
431 ib_dealloc_pd(ndev->pd);
435 mutex_unlock(&device_list_mutex);
439 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
441 struct nvme_rdma_device *dev;
442 struct ib_device *ibdev;
446 rdma_destroy_qp(queue->cm_id);
447 ib_free_cq(queue->ib_cq);
449 nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
450 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
452 nvme_rdma_dev_put(dev);
455 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue)
457 struct ib_device *ibdev;
458 const int send_wr_factor = 3; /* MR, SEND, INV */
459 const int cq_factor = send_wr_factor + 1; /* + RECV */
460 int comp_vector, idx = nvme_rdma_queue_idx(queue);
463 queue->device = nvme_rdma_find_get_device(queue->cm_id);
464 if (!queue->device) {
465 dev_err(queue->cm_id->device->dev.parent,
466 "no client data found!\n");
467 return -ECONNREFUSED;
469 ibdev = queue->device->dev;
472 * The admin queue is barely used once the controller is live, so don't
473 * bother to spread it out.
478 comp_vector = idx % ibdev->num_comp_vectors;
481 /* +1 for ib_stop_cq */
482 queue->ib_cq = ib_alloc_cq(ibdev, queue,
483 cq_factor * queue->queue_size + 1,
484 comp_vector, IB_POLL_SOFTIRQ);
485 if (IS_ERR(queue->ib_cq)) {
486 ret = PTR_ERR(queue->ib_cq);
490 ret = nvme_rdma_create_qp(queue, send_wr_factor);
492 goto out_destroy_ib_cq;
494 queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size,
495 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
496 if (!queue->rsp_ring) {
504 ib_destroy_qp(queue->qp);
506 ib_free_cq(queue->ib_cq);
508 nvme_rdma_dev_put(queue->device);
512 static int nvme_rdma_init_queue(struct nvme_rdma_ctrl *ctrl,
513 int idx, size_t queue_size)
515 struct nvme_rdma_queue *queue;
516 struct sockaddr *src_addr = NULL;
519 queue = &ctrl->queues[idx];
521 init_completion(&queue->cm_done);
524 queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
526 queue->cmnd_capsule_len = sizeof(struct nvme_command);
528 queue->queue_size = queue_size;
530 queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue,
531 RDMA_PS_TCP, IB_QPT_RC);
532 if (IS_ERR(queue->cm_id)) {
533 dev_info(ctrl->ctrl.device,
534 "failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id));
535 return PTR_ERR(queue->cm_id);
538 if (ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR)
539 src_addr = (struct sockaddr *)&ctrl->src_addr;
541 queue->cm_error = -ETIMEDOUT;
542 ret = rdma_resolve_addr(queue->cm_id, src_addr,
543 (struct sockaddr *)&ctrl->addr,
544 NVME_RDMA_CONNECT_TIMEOUT_MS);
546 dev_info(ctrl->ctrl.device,
547 "rdma_resolve_addr failed (%d).\n", ret);
548 goto out_destroy_cm_id;
551 ret = nvme_rdma_wait_for_cm(queue);
553 dev_info(ctrl->ctrl.device,
554 "rdma_resolve_addr wait failed (%d).\n", ret);
555 goto out_destroy_cm_id;
558 clear_bit(NVME_RDMA_Q_DELETING, &queue->flags);
563 rdma_destroy_id(queue->cm_id);
567 static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
569 rdma_disconnect(queue->cm_id);
570 ib_drain_qp(queue->qp);
573 static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
575 nvme_rdma_destroy_queue_ib(queue);
576 rdma_destroy_id(queue->cm_id);
579 static void nvme_rdma_stop_and_free_queue(struct nvme_rdma_queue *queue)
581 if (test_and_set_bit(NVME_RDMA_Q_DELETING, &queue->flags))
583 nvme_rdma_stop_queue(queue);
584 nvme_rdma_free_queue(queue);
587 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl)
591 for (i = 1; i < ctrl->queue_count; i++)
592 nvme_rdma_stop_and_free_queue(&ctrl->queues[i]);
595 static int nvme_rdma_connect_io_queues(struct nvme_rdma_ctrl *ctrl)
599 for (i = 1; i < ctrl->queue_count; i++) {
600 ret = nvmf_connect_io_queue(&ctrl->ctrl, i);
602 dev_info(ctrl->ctrl.device,
603 "failed to connect i/o queue: %d\n", ret);
604 goto out_free_queues;
606 set_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[i].flags);
612 nvme_rdma_free_io_queues(ctrl);
616 static int nvme_rdma_init_io_queues(struct nvme_rdma_ctrl *ctrl)
618 struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
619 unsigned int nr_io_queues;
622 nr_io_queues = min(opts->nr_io_queues, num_online_cpus());
623 ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
627 ctrl->queue_count = nr_io_queues + 1;
628 if (ctrl->queue_count < 2)
631 dev_info(ctrl->ctrl.device,
632 "creating %d I/O queues.\n", nr_io_queues);
634 for (i = 1; i < ctrl->queue_count; i++) {
635 ret = nvme_rdma_init_queue(ctrl, i,
636 ctrl->ctrl.opts->queue_size);
638 dev_info(ctrl->ctrl.device,
639 "failed to initialize i/o queue: %d\n", ret);
640 goto out_free_queues;
647 for (i--; i >= 1; i--)
648 nvme_rdma_stop_and_free_queue(&ctrl->queues[i]);
653 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl)
655 nvme_rdma_free_qe(ctrl->queues[0].device->dev, &ctrl->async_event_sqe,
656 sizeof(struct nvme_command), DMA_TO_DEVICE);
657 nvme_rdma_stop_and_free_queue(&ctrl->queues[0]);
658 blk_cleanup_queue(ctrl->ctrl.admin_q);
659 blk_mq_free_tag_set(&ctrl->admin_tag_set);
660 nvme_rdma_dev_put(ctrl->device);
663 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
665 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
667 if (list_empty(&ctrl->list))
670 mutex_lock(&nvme_rdma_ctrl_mutex);
671 list_del(&ctrl->list);
672 mutex_unlock(&nvme_rdma_ctrl_mutex);
675 nvmf_free_options(nctrl->opts);
680 static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl *ctrl)
682 /* If we are resetting/deleting then do nothing */
683 if (ctrl->ctrl.state != NVME_CTRL_RECONNECTING) {
684 WARN_ON_ONCE(ctrl->ctrl.state == NVME_CTRL_NEW ||
685 ctrl->ctrl.state == NVME_CTRL_LIVE);
689 if (nvmf_should_reconnect(&ctrl->ctrl)) {
690 dev_info(ctrl->ctrl.device, "Reconnecting in %d seconds...\n",
691 ctrl->ctrl.opts->reconnect_delay);
692 queue_delayed_work(nvme_wq, &ctrl->reconnect_work,
693 ctrl->ctrl.opts->reconnect_delay * HZ);
695 dev_info(ctrl->ctrl.device, "Removing controller...\n");
696 queue_work(nvme_wq, &ctrl->delete_work);
700 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
702 struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
703 struct nvme_rdma_ctrl, reconnect_work);
707 ++ctrl->ctrl.nr_reconnects;
709 if (ctrl->queue_count > 1) {
710 nvme_rdma_free_io_queues(ctrl);
712 ret = blk_mq_reinit_tagset(&ctrl->tag_set);
717 nvme_rdma_stop_and_free_queue(&ctrl->queues[0]);
719 ret = blk_mq_reinit_tagset(&ctrl->admin_tag_set);
723 ret = nvme_rdma_init_queue(ctrl, 0, NVMF_AQ_DEPTH);
727 ret = nvmf_connect_admin_queue(&ctrl->ctrl);
731 set_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[0].flags);
733 ret = nvme_enable_ctrl(&ctrl->ctrl, ctrl->cap);
737 nvme_start_keep_alive(&ctrl->ctrl);
739 if (ctrl->queue_count > 1) {
740 ret = nvme_rdma_init_io_queues(ctrl);
744 ret = nvme_rdma_connect_io_queues(ctrl);
749 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
750 WARN_ON_ONCE(!changed);
751 ctrl->ctrl.nr_reconnects = 0;
753 if (ctrl->queue_count > 1) {
754 nvme_queue_scan(&ctrl->ctrl);
755 nvme_queue_async_events(&ctrl->ctrl);
758 dev_info(ctrl->ctrl.device, "Successfully reconnected\n");
763 dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d\n",
764 ctrl->ctrl.nr_reconnects);
765 nvme_rdma_reconnect_or_remove(ctrl);
768 static void nvme_rdma_error_recovery_work(struct work_struct *work)
770 struct nvme_rdma_ctrl *ctrl = container_of(work,
771 struct nvme_rdma_ctrl, err_work);
774 nvme_stop_keep_alive(&ctrl->ctrl);
776 for (i = 0; i < ctrl->queue_count; i++)
777 clear_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[i].flags);
779 if (ctrl->queue_count > 1)
780 nvme_stop_queues(&ctrl->ctrl);
781 blk_mq_stop_hw_queues(ctrl->ctrl.admin_q);
783 /* We must take care of fastfail/requeue all our inflight requests */
784 if (ctrl->queue_count > 1)
785 blk_mq_tagset_busy_iter(&ctrl->tag_set,
786 nvme_cancel_request, &ctrl->ctrl);
787 blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
788 nvme_cancel_request, &ctrl->ctrl);
791 * queues are not a live anymore, so restart the queues to fail fast
794 blk_mq_start_stopped_hw_queues(ctrl->ctrl.admin_q, true);
795 nvme_start_queues(&ctrl->ctrl);
797 nvme_rdma_reconnect_or_remove(ctrl);
800 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
802 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RECONNECTING))
805 queue_work(nvme_wq, &ctrl->err_work);
808 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
811 struct nvme_rdma_queue *queue = cq->cq_context;
812 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
814 if (ctrl->ctrl.state == NVME_CTRL_LIVE)
815 dev_info(ctrl->ctrl.device,
816 "%s for CQE 0x%p failed with status %s (%d)\n",
818 ib_wc_status_msg(wc->status), wc->status);
819 nvme_rdma_error_recovery(ctrl);
822 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
824 if (unlikely(wc->status != IB_WC_SUCCESS))
825 nvme_rdma_wr_error(cq, wc, "MEMREG");
828 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
830 if (unlikely(wc->status != IB_WC_SUCCESS))
831 nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
834 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
835 struct nvme_rdma_request *req)
837 struct ib_send_wr *bad_wr;
838 struct ib_send_wr wr = {
839 .opcode = IB_WR_LOCAL_INV,
843 .ex.invalidate_rkey = req->mr->rkey,
846 req->reg_cqe.done = nvme_rdma_inv_rkey_done;
847 wr.wr_cqe = &req->reg_cqe;
849 return ib_post_send(queue->qp, &wr, &bad_wr);
852 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
855 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
856 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
857 struct nvme_rdma_device *dev = queue->device;
858 struct ib_device *ibdev = dev->dev;
861 if (!blk_rq_bytes(rq))
864 if (req->mr->need_inval) {
865 res = nvme_rdma_inv_rkey(queue, req);
867 dev_err(ctrl->ctrl.device,
868 "Queueing INV WR for rkey %#x failed (%d)\n",
870 nvme_rdma_error_recovery(queue->ctrl);
874 ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
875 req->nents, rq_data_dir(rq) ==
876 WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
878 nvme_cleanup_cmd(rq);
879 sg_free_table_chained(&req->sg_table, true);
882 static int nvme_rdma_set_sg_null(struct nvme_command *c)
884 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
887 put_unaligned_le24(0, sg->length);
888 put_unaligned_le32(0, sg->key);
889 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
893 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
894 struct nvme_rdma_request *req, struct nvme_command *c)
896 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
898 req->sge[1].addr = sg_dma_address(req->sg_table.sgl);
899 req->sge[1].length = sg_dma_len(req->sg_table.sgl);
900 req->sge[1].lkey = queue->device->pd->local_dma_lkey;
902 sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
903 sg->length = cpu_to_le32(sg_dma_len(req->sg_table.sgl));
904 sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
906 req->inline_data = true;
911 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
912 struct nvme_rdma_request *req, struct nvme_command *c)
914 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
916 sg->addr = cpu_to_le64(sg_dma_address(req->sg_table.sgl));
917 put_unaligned_le24(sg_dma_len(req->sg_table.sgl), sg->length);
918 put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
919 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
923 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
924 struct nvme_rdma_request *req, struct nvme_command *c,
927 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
930 nr = ib_map_mr_sg(req->mr, req->sg_table.sgl, count, NULL, PAGE_SIZE);
937 ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
939 req->reg_cqe.done = nvme_rdma_memreg_done;
940 memset(&req->reg_wr, 0, sizeof(req->reg_wr));
941 req->reg_wr.wr.opcode = IB_WR_REG_MR;
942 req->reg_wr.wr.wr_cqe = &req->reg_cqe;
943 req->reg_wr.wr.num_sge = 0;
944 req->reg_wr.mr = req->mr;
945 req->reg_wr.key = req->mr->rkey;
946 req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
947 IB_ACCESS_REMOTE_READ |
948 IB_ACCESS_REMOTE_WRITE;
950 req->mr->need_inval = true;
952 sg->addr = cpu_to_le64(req->mr->iova);
953 put_unaligned_le24(req->mr->length, sg->length);
954 put_unaligned_le32(req->mr->rkey, sg->key);
955 sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
956 NVME_SGL_FMT_INVALIDATE;
961 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
962 struct request *rq, 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;
970 req->inline_data = false;
971 req->mr->need_inval = false;
973 c->common.flags |= NVME_CMD_SGL_METABUF;
975 if (!blk_rq_bytes(rq))
976 return nvme_rdma_set_sg_null(c);
978 req->sg_table.sgl = req->first_sgl;
979 ret = sg_alloc_table_chained(&req->sg_table,
980 blk_rq_nr_phys_segments(rq), req->sg_table.sgl);
984 req->nents = blk_rq_map_sg(rq->q, rq, req->sg_table.sgl);
986 count = ib_dma_map_sg(ibdev, req->sg_table.sgl, req->nents,
987 rq_data_dir(rq) == WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
988 if (unlikely(count <= 0)) {
989 sg_free_table_chained(&req->sg_table, true);
994 if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) &&
995 blk_rq_payload_bytes(rq) <=
996 nvme_rdma_inline_data_size(queue))
997 return nvme_rdma_map_sg_inline(queue, req, c);
999 if (dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY)
1000 return nvme_rdma_map_sg_single(queue, req, c);
1003 return nvme_rdma_map_sg_fr(queue, req, c, count);
1006 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1008 if (unlikely(wc->status != IB_WC_SUCCESS))
1009 nvme_rdma_wr_error(cq, wc, "SEND");
1012 static inline int nvme_rdma_queue_sig_limit(struct nvme_rdma_queue *queue)
1017 * We signal completion every queue depth/2 and also handle the
1018 * degenerated case of a device with queue_depth=1, where we
1019 * would need to signal every message.
1021 sig_limit = max(queue->queue_size / 2, 1);
1022 return (++queue->sig_count % sig_limit) == 0;
1025 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1026 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1027 struct ib_send_wr *first, bool flush)
1029 struct ib_send_wr wr, *bad_wr;
1032 sge->addr = qe->dma;
1033 sge->length = sizeof(struct nvme_command),
1034 sge->lkey = queue->device->pd->local_dma_lkey;
1036 qe->cqe.done = nvme_rdma_send_done;
1039 wr.wr_cqe = &qe->cqe;
1041 wr.num_sge = num_sge;
1042 wr.opcode = IB_WR_SEND;
1046 * Unsignalled send completions are another giant desaster in the
1047 * IB Verbs spec: If we don't regularly post signalled sends
1048 * the send queue will fill up and only a QP reset will rescue us.
1049 * Would have been way to obvious to handle this in hardware or
1050 * at least the RDMA stack..
1052 * Always signal the flushes. The magic request used for the flush
1053 * sequencer is not allocated in our driver's tagset and it's
1054 * triggered to be freed by blk_cleanup_queue(). So we need to
1055 * always mark it as signaled to ensure that the "wr_cqe", which is
1056 * embedded in request's payload, is not freed when __ib_process_cq()
1057 * calls wr_cqe->done().
1059 if (nvme_rdma_queue_sig_limit(queue) || flush)
1060 wr.send_flags |= IB_SEND_SIGNALED;
1067 ret = ib_post_send(queue->qp, first, &bad_wr);
1069 dev_err(queue->ctrl->ctrl.device,
1070 "%s failed with error code %d\n", __func__, ret);
1075 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1076 struct nvme_rdma_qe *qe)
1078 struct ib_recv_wr wr, *bad_wr;
1082 list.addr = qe->dma;
1083 list.length = sizeof(struct nvme_completion);
1084 list.lkey = queue->device->pd->local_dma_lkey;
1086 qe->cqe.done = nvme_rdma_recv_done;
1089 wr.wr_cqe = &qe->cqe;
1093 ret = ib_post_recv(queue->qp, &wr, &bad_wr);
1095 dev_err(queue->ctrl->ctrl.device,
1096 "%s failed with error code %d\n", __func__, ret);
1101 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1103 u32 queue_idx = nvme_rdma_queue_idx(queue);
1106 return queue->ctrl->admin_tag_set.tags[queue_idx];
1107 return queue->ctrl->tag_set.tags[queue_idx - 1];
1110 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg, int aer_idx)
1112 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1113 struct nvme_rdma_queue *queue = &ctrl->queues[0];
1114 struct ib_device *dev = queue->device->dev;
1115 struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1116 struct nvme_command *cmd = sqe->data;
1120 if (WARN_ON_ONCE(aer_idx != 0))
1123 ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1125 memset(cmd, 0, sizeof(*cmd));
1126 cmd->common.opcode = nvme_admin_async_event;
1127 cmd->common.command_id = NVME_RDMA_AQ_BLKMQ_DEPTH;
1128 cmd->common.flags |= NVME_CMD_SGL_METABUF;
1129 nvme_rdma_set_sg_null(cmd);
1131 ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1134 ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL, false);
1138 static int nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1139 struct nvme_completion *cqe, struct ib_wc *wc, int tag)
1142 struct nvme_rdma_request *req;
1145 rq = blk_mq_tag_to_rq(nvme_rdma_tagset(queue), cqe->command_id);
1147 dev_err(queue->ctrl->ctrl.device,
1148 "tag 0x%x on QP %#x not found\n",
1149 cqe->command_id, queue->qp->qp_num);
1150 nvme_rdma_error_recovery(queue->ctrl);
1153 req = blk_mq_rq_to_pdu(rq);
1158 if ((wc->wc_flags & IB_WC_WITH_INVALIDATE) &&
1159 wc->ex.invalidate_rkey == req->mr->rkey)
1160 req->mr->need_inval = false;
1162 nvme_end_request(rq, cqe->status, cqe->result);
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->status,
1193 ret = nvme_rdma_process_nvme_rsp(queue, cqe, wc, tag);
1194 ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1196 nvme_rdma_post_recv(queue, qe);
1200 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1202 __nvme_rdma_recv_done(cq, wc, -1);
1205 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1209 for (i = 0; i < queue->queue_size; i++) {
1210 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1212 goto out_destroy_queue_ib;
1217 out_destroy_queue_ib:
1218 nvme_rdma_destroy_queue_ib(queue);
1222 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1223 struct rdma_cm_event *ev)
1225 struct rdma_cm_id *cm_id = queue->cm_id;
1226 int status = ev->status;
1227 const char *rej_msg;
1228 const struct nvme_rdma_cm_rej *rej_data;
1231 rej_msg = rdma_reject_msg(cm_id, status);
1232 rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len);
1234 if (rej_data && rej_data_len >= sizeof(u16)) {
1235 u16 sts = le16_to_cpu(rej_data->sts);
1237 dev_err(queue->ctrl->ctrl.device,
1238 "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1239 status, rej_msg, sts, nvme_rdma_cm_msg(sts));
1241 dev_err(queue->ctrl->ctrl.device,
1242 "Connect rejected: status %d (%s).\n", status, rej_msg);
1248 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1252 ret = nvme_rdma_create_queue_ib(queue);
1256 ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1258 dev_err(queue->ctrl->ctrl.device,
1259 "rdma_resolve_route failed (%d).\n",
1261 goto out_destroy_queue;
1267 nvme_rdma_destroy_queue_ib(queue);
1271 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1273 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1274 struct rdma_conn_param param = { };
1275 struct nvme_rdma_cm_req priv = { };
1278 param.qp_num = queue->qp->qp_num;
1279 param.flow_control = 1;
1281 param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1282 /* maximum retry count */
1283 param.retry_count = 7;
1284 param.rnr_retry_count = 7;
1285 param.private_data = &priv;
1286 param.private_data_len = sizeof(priv);
1288 priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1289 priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1291 * set the admin queue depth to the minimum size
1292 * specified by the Fabrics standard.
1294 if (priv.qid == 0) {
1295 priv.hrqsize = cpu_to_le16(NVMF_AQ_DEPTH);
1296 priv.hsqsize = cpu_to_le16(NVMF_AQ_DEPTH - 1);
1299 * current interpretation of the fabrics spec
1300 * is at minimum you make hrqsize sqsize+1, or a
1301 * 1's based representation of sqsize.
1303 priv.hrqsize = cpu_to_le16(queue->queue_size);
1304 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1307 ret = rdma_connect(queue->cm_id, ¶m);
1309 dev_err(ctrl->ctrl.device,
1310 "rdma_connect failed (%d).\n", ret);
1311 goto out_destroy_queue_ib;
1316 out_destroy_queue_ib:
1317 nvme_rdma_destroy_queue_ib(queue);
1321 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1322 struct rdma_cm_event *ev)
1324 struct nvme_rdma_queue *queue = cm_id->context;
1327 dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1328 rdma_event_msg(ev->event), ev->event,
1331 switch (ev->event) {
1332 case RDMA_CM_EVENT_ADDR_RESOLVED:
1333 cm_error = nvme_rdma_addr_resolved(queue);
1335 case RDMA_CM_EVENT_ROUTE_RESOLVED:
1336 cm_error = nvme_rdma_route_resolved(queue);
1338 case RDMA_CM_EVENT_ESTABLISHED:
1339 queue->cm_error = nvme_rdma_conn_established(queue);
1340 /* complete cm_done regardless of success/failure */
1341 complete(&queue->cm_done);
1343 case RDMA_CM_EVENT_REJECTED:
1344 nvme_rdma_destroy_queue_ib(queue);
1345 cm_error = nvme_rdma_conn_rejected(queue, ev);
1347 case RDMA_CM_EVENT_ROUTE_ERROR:
1348 case RDMA_CM_EVENT_CONNECT_ERROR:
1349 case RDMA_CM_EVENT_UNREACHABLE:
1350 nvme_rdma_destroy_queue_ib(queue);
1351 case RDMA_CM_EVENT_ADDR_ERROR:
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 nvme_req(rq)->status = NVME_SC_ABORT_REQ | NVME_SC_DNR;
1392 return BLK_EH_HANDLED;
1396 * We cannot accept any other command until the Connect command has completed.
1398 static inline blk_status_t
1399 nvme_rdma_queue_is_ready(struct nvme_rdma_queue *queue, struct request *rq)
1401 if (unlikely(!test_bit(NVME_RDMA_Q_LIVE, &queue->flags))) {
1402 struct nvme_command *cmd = nvme_req(rq)->cmd;
1404 if (!blk_rq_is_passthrough(rq) ||
1405 cmd->common.opcode != nvme_fabrics_command ||
1406 cmd->fabrics.fctype != nvme_fabrics_type_connect) {
1408 * reconnecting state means transport disruption, which
1409 * can take a long time and even might fail permanently,
1410 * so we can't let incoming I/O be requeued forever.
1411 * fail it fast to allow upper layers a chance to
1414 if (queue->ctrl->ctrl.state == NVME_CTRL_RECONNECTING)
1415 return BLK_STS_IOERR;
1416 return BLK_STS_RESOURCE; /* try again later */
1423 static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
1424 const struct blk_mq_queue_data *bd)
1426 struct nvme_ns *ns = hctx->queue->queuedata;
1427 struct nvme_rdma_queue *queue = hctx->driver_data;
1428 struct request *rq = bd->rq;
1429 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1430 struct nvme_rdma_qe *sqe = &req->sqe;
1431 struct nvme_command *c = sqe->data;
1433 struct ib_device *dev;
1437 WARN_ON_ONCE(rq->tag < 0);
1439 ret = nvme_rdma_queue_is_ready(queue, rq);
1443 dev = queue->device->dev;
1444 ib_dma_sync_single_for_cpu(dev, sqe->dma,
1445 sizeof(struct nvme_command), DMA_TO_DEVICE);
1447 ret = nvme_setup_cmd(ns, rq, c);
1451 blk_mq_start_request(rq);
1453 err = nvme_rdma_map_data(queue, rq, c);
1455 dev_err(queue->ctrl->ctrl.device,
1456 "Failed to map data (%d)\n", err);
1457 nvme_cleanup_cmd(rq);
1461 ib_dma_sync_single_for_device(dev, sqe->dma,
1462 sizeof(struct nvme_command), DMA_TO_DEVICE);
1464 if (req_op(rq) == REQ_OP_FLUSH)
1466 err = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
1467 req->mr->need_inval ? &req->reg_wr.wr : NULL, flush);
1469 nvme_rdma_unmap_data(queue, rq);
1475 if (err == -ENOMEM || err == -EAGAIN)
1476 return BLK_STS_RESOURCE;
1477 return BLK_STS_IOERR;
1480 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx, unsigned int tag)
1482 struct nvme_rdma_queue *queue = hctx->driver_data;
1483 struct ib_cq *cq = queue->ib_cq;
1487 while (ib_poll_cq(cq, 1, &wc) > 0) {
1488 struct ib_cqe *cqe = wc.wr_cqe;
1491 if (cqe->done == nvme_rdma_recv_done)
1492 found |= __nvme_rdma_recv_done(cq, &wc, tag);
1501 static void nvme_rdma_complete_rq(struct request *rq)
1503 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1505 nvme_rdma_unmap_data(req->queue, rq);
1506 nvme_complete_rq(rq);
1509 static const struct blk_mq_ops nvme_rdma_mq_ops = {
1510 .queue_rq = nvme_rdma_queue_rq,
1511 .complete = nvme_rdma_complete_rq,
1512 .init_request = nvme_rdma_init_request,
1513 .exit_request = nvme_rdma_exit_request,
1514 .reinit_request = nvme_rdma_reinit_request,
1515 .init_hctx = nvme_rdma_init_hctx,
1516 .poll = nvme_rdma_poll,
1517 .timeout = nvme_rdma_timeout,
1520 static const struct blk_mq_ops nvme_rdma_admin_mq_ops = {
1521 .queue_rq = nvme_rdma_queue_rq,
1522 .complete = nvme_rdma_complete_rq,
1523 .init_request = nvme_rdma_init_request,
1524 .exit_request = nvme_rdma_exit_request,
1525 .reinit_request = nvme_rdma_reinit_request,
1526 .init_hctx = nvme_rdma_init_admin_hctx,
1527 .timeout = nvme_rdma_timeout,
1530 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl)
1534 error = nvme_rdma_init_queue(ctrl, 0, NVMF_AQ_DEPTH);
1538 ctrl->device = ctrl->queues[0].device;
1541 * We need a reference on the device as long as the tag_set is alive,
1542 * as the MRs in the request structures need a valid ib_device.
1545 if (!nvme_rdma_dev_get(ctrl->device))
1546 goto out_free_queue;
1548 ctrl->max_fr_pages = min_t(u32, NVME_RDMA_MAX_SEGMENTS,
1549 ctrl->device->dev->attrs.max_fast_reg_page_list_len);
1551 memset(&ctrl->admin_tag_set, 0, sizeof(ctrl->admin_tag_set));
1552 ctrl->admin_tag_set.ops = &nvme_rdma_admin_mq_ops;
1553 ctrl->admin_tag_set.queue_depth = NVME_RDMA_AQ_BLKMQ_DEPTH;
1554 ctrl->admin_tag_set.reserved_tags = 2; /* connect + keep-alive */
1555 ctrl->admin_tag_set.numa_node = NUMA_NO_NODE;
1556 ctrl->admin_tag_set.cmd_size = sizeof(struct nvme_rdma_request) +
1557 SG_CHUNK_SIZE * sizeof(struct scatterlist);
1558 ctrl->admin_tag_set.driver_data = ctrl;
1559 ctrl->admin_tag_set.nr_hw_queues = 1;
1560 ctrl->admin_tag_set.timeout = ADMIN_TIMEOUT;
1562 error = blk_mq_alloc_tag_set(&ctrl->admin_tag_set);
1566 ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
1567 if (IS_ERR(ctrl->ctrl.admin_q)) {
1568 error = PTR_ERR(ctrl->ctrl.admin_q);
1569 goto out_free_tagset;
1572 error = nvmf_connect_admin_queue(&ctrl->ctrl);
1574 goto out_cleanup_queue;
1576 set_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[0].flags);
1578 error = nvmf_reg_read64(&ctrl->ctrl, NVME_REG_CAP, &ctrl->cap);
1580 dev_err(ctrl->ctrl.device,
1581 "prop_get NVME_REG_CAP failed\n");
1582 goto out_cleanup_queue;
1586 min_t(int, NVME_CAP_MQES(ctrl->cap), ctrl->ctrl.sqsize);
1588 error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->cap);
1590 goto out_cleanup_queue;
1592 ctrl->ctrl.max_hw_sectors =
1593 (ctrl->max_fr_pages - 1) << (PAGE_SHIFT - 9);
1595 error = nvme_init_identify(&ctrl->ctrl);
1597 goto out_cleanup_queue;
1599 error = nvme_rdma_alloc_qe(ctrl->queues[0].device->dev,
1600 &ctrl->async_event_sqe, sizeof(struct nvme_command),
1603 goto out_cleanup_queue;
1605 nvme_start_keep_alive(&ctrl->ctrl);
1610 blk_cleanup_queue(ctrl->ctrl.admin_q);
1612 /* disconnect and drain the queue before freeing the tagset */
1613 nvme_rdma_stop_queue(&ctrl->queues[0]);
1614 blk_mq_free_tag_set(&ctrl->admin_tag_set);
1616 nvme_rdma_dev_put(ctrl->device);
1618 nvme_rdma_free_queue(&ctrl->queues[0]);
1622 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl)
1624 nvme_stop_keep_alive(&ctrl->ctrl);
1625 cancel_work_sync(&ctrl->err_work);
1626 cancel_delayed_work_sync(&ctrl->reconnect_work);
1628 if (ctrl->queue_count > 1) {
1629 nvme_stop_queues(&ctrl->ctrl);
1630 blk_mq_tagset_busy_iter(&ctrl->tag_set,
1631 nvme_cancel_request, &ctrl->ctrl);
1632 nvme_rdma_free_io_queues(ctrl);
1635 if (test_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[0].flags))
1636 nvme_shutdown_ctrl(&ctrl->ctrl);
1638 blk_mq_stop_hw_queues(ctrl->ctrl.admin_q);
1639 blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
1640 nvme_cancel_request, &ctrl->ctrl);
1641 nvme_rdma_destroy_admin_queue(ctrl);
1644 static void __nvme_rdma_remove_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
1646 nvme_uninit_ctrl(&ctrl->ctrl);
1648 nvme_rdma_shutdown_ctrl(ctrl);
1650 if (ctrl->ctrl.tagset) {
1651 blk_cleanup_queue(ctrl->ctrl.connect_q);
1652 blk_mq_free_tag_set(&ctrl->tag_set);
1653 nvme_rdma_dev_put(ctrl->device);
1656 nvme_put_ctrl(&ctrl->ctrl);
1659 static void nvme_rdma_del_ctrl_work(struct work_struct *work)
1661 struct nvme_rdma_ctrl *ctrl = container_of(work,
1662 struct nvme_rdma_ctrl, delete_work);
1664 __nvme_rdma_remove_ctrl(ctrl, true);
1667 static int __nvme_rdma_del_ctrl(struct nvme_rdma_ctrl *ctrl)
1669 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_DELETING))
1672 if (!queue_work(nvme_wq, &ctrl->delete_work))
1678 static int nvme_rdma_del_ctrl(struct nvme_ctrl *nctrl)
1680 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
1684 * Keep a reference until all work is flushed since
1685 * __nvme_rdma_del_ctrl can free the ctrl mem
1687 if (!kref_get_unless_zero(&ctrl->ctrl.kref))
1689 ret = __nvme_rdma_del_ctrl(ctrl);
1691 flush_work(&ctrl->delete_work);
1692 nvme_put_ctrl(&ctrl->ctrl);
1696 static void nvme_rdma_remove_ctrl_work(struct work_struct *work)
1698 struct nvme_rdma_ctrl *ctrl = container_of(work,
1699 struct nvme_rdma_ctrl, delete_work);
1701 __nvme_rdma_remove_ctrl(ctrl, false);
1704 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
1706 struct nvme_rdma_ctrl *ctrl = container_of(work,
1707 struct nvme_rdma_ctrl, reset_work);
1711 nvme_rdma_shutdown_ctrl(ctrl);
1713 ret = nvme_rdma_configure_admin_queue(ctrl);
1715 /* ctrl is already shutdown, just remove the ctrl */
1716 INIT_WORK(&ctrl->delete_work, nvme_rdma_remove_ctrl_work);
1720 if (ctrl->queue_count > 1) {
1721 ret = blk_mq_reinit_tagset(&ctrl->tag_set);
1725 ret = nvme_rdma_init_io_queues(ctrl);
1729 ret = nvme_rdma_connect_io_queues(ctrl);
1734 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1735 WARN_ON_ONCE(!changed);
1737 if (ctrl->queue_count > 1) {
1738 nvme_start_queues(&ctrl->ctrl);
1739 nvme_queue_scan(&ctrl->ctrl);
1740 nvme_queue_async_events(&ctrl->ctrl);
1746 /* Deleting this dead controller... */
1747 dev_warn(ctrl->ctrl.device, "Removing after reset failure\n");
1748 WARN_ON(!queue_work(nvme_wq, &ctrl->delete_work));
1751 static int nvme_rdma_reset_ctrl(struct nvme_ctrl *nctrl)
1753 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
1755 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
1758 if (!queue_work(nvme_wq, &ctrl->reset_work))
1761 flush_work(&ctrl->reset_work);
1766 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
1768 .module = THIS_MODULE,
1769 .flags = NVME_F_FABRICS,
1770 .reg_read32 = nvmf_reg_read32,
1771 .reg_read64 = nvmf_reg_read64,
1772 .reg_write32 = nvmf_reg_write32,
1773 .reset_ctrl = nvme_rdma_reset_ctrl,
1774 .free_ctrl = nvme_rdma_free_ctrl,
1775 .submit_async_event = nvme_rdma_submit_async_event,
1776 .delete_ctrl = nvme_rdma_del_ctrl,
1777 .get_subsysnqn = nvmf_get_subsysnqn,
1778 .get_address = nvmf_get_address,
1781 static int nvme_rdma_create_io_queues(struct nvme_rdma_ctrl *ctrl)
1785 ret = nvme_rdma_init_io_queues(ctrl);
1790 * We need a reference on the device as long as the tag_set is alive,
1791 * as the MRs in the request structures need a valid ib_device.
1794 if (!nvme_rdma_dev_get(ctrl->device))
1795 goto out_free_io_queues;
1797 memset(&ctrl->tag_set, 0, sizeof(ctrl->tag_set));
1798 ctrl->tag_set.ops = &nvme_rdma_mq_ops;
1799 ctrl->tag_set.queue_depth = ctrl->ctrl.opts->queue_size;
1800 ctrl->tag_set.reserved_tags = 1; /* fabric connect */
1801 ctrl->tag_set.numa_node = NUMA_NO_NODE;
1802 ctrl->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
1803 ctrl->tag_set.cmd_size = sizeof(struct nvme_rdma_request) +
1804 SG_CHUNK_SIZE * sizeof(struct scatterlist);
1805 ctrl->tag_set.driver_data = ctrl;
1806 ctrl->tag_set.nr_hw_queues = ctrl->queue_count - 1;
1807 ctrl->tag_set.timeout = NVME_IO_TIMEOUT;
1809 ret = blk_mq_alloc_tag_set(&ctrl->tag_set);
1812 ctrl->ctrl.tagset = &ctrl->tag_set;
1814 ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
1815 if (IS_ERR(ctrl->ctrl.connect_q)) {
1816 ret = PTR_ERR(ctrl->ctrl.connect_q);
1817 goto out_free_tag_set;
1820 ret = nvme_rdma_connect_io_queues(ctrl);
1822 goto out_cleanup_connect_q;
1826 out_cleanup_connect_q:
1827 blk_cleanup_queue(ctrl->ctrl.connect_q);
1829 blk_mq_free_tag_set(&ctrl->tag_set);
1831 nvme_rdma_dev_put(ctrl->device);
1833 nvme_rdma_free_io_queues(ctrl);
1837 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
1838 struct nvmf_ctrl_options *opts)
1840 struct nvme_rdma_ctrl *ctrl;
1845 ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1847 return ERR_PTR(-ENOMEM);
1848 ctrl->ctrl.opts = opts;
1849 INIT_LIST_HEAD(&ctrl->list);
1851 if (opts->mask & NVMF_OPT_TRSVCID)
1852 port = opts->trsvcid;
1854 port = __stringify(NVME_RDMA_IP_PORT);
1856 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1857 opts->traddr, port, &ctrl->addr);
1859 pr_err("malformed address passed: %s:%s\n", opts->traddr, port);
1863 if (opts->mask & NVMF_OPT_HOST_TRADDR) {
1864 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1865 opts->host_traddr, NULL, &ctrl->src_addr);
1867 pr_err("malformed src address passed: %s\n",
1873 ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
1874 0 /* no quirks, we're perfect! */);
1878 INIT_DELAYED_WORK(&ctrl->reconnect_work,
1879 nvme_rdma_reconnect_ctrl_work);
1880 INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
1881 INIT_WORK(&ctrl->delete_work, nvme_rdma_del_ctrl_work);
1882 INIT_WORK(&ctrl->reset_work, nvme_rdma_reset_ctrl_work);
1884 ctrl->queue_count = opts->nr_io_queues + 1; /* +1 for admin queue */
1885 ctrl->ctrl.sqsize = opts->queue_size - 1;
1886 ctrl->ctrl.kato = opts->kato;
1889 ctrl->queues = kcalloc(ctrl->queue_count, sizeof(*ctrl->queues),
1892 goto out_uninit_ctrl;
1894 ret = nvme_rdma_configure_admin_queue(ctrl);
1896 goto out_kfree_queues;
1898 /* sanity check icdoff */
1899 if (ctrl->ctrl.icdoff) {
1900 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
1902 goto out_remove_admin_queue;
1905 /* sanity check keyed sgls */
1906 if (!(ctrl->ctrl.sgls & (1 << 20))) {
1907 dev_err(ctrl->ctrl.device, "Mandatory keyed sgls are not support\n");
1909 goto out_remove_admin_queue;
1912 if (opts->queue_size > ctrl->ctrl.maxcmd) {
1913 /* warn if maxcmd is lower than queue_size */
1914 dev_warn(ctrl->ctrl.device,
1915 "queue_size %zu > ctrl maxcmd %u, clamping down\n",
1916 opts->queue_size, ctrl->ctrl.maxcmd);
1917 opts->queue_size = ctrl->ctrl.maxcmd;
1920 if (opts->queue_size > ctrl->ctrl.sqsize + 1) {
1921 /* warn if sqsize is lower than queue_size */
1922 dev_warn(ctrl->ctrl.device,
1923 "queue_size %zu > ctrl sqsize %u, clamping down\n",
1924 opts->queue_size, ctrl->ctrl.sqsize + 1);
1925 opts->queue_size = ctrl->ctrl.sqsize + 1;
1928 if (opts->nr_io_queues) {
1929 ret = nvme_rdma_create_io_queues(ctrl);
1931 goto out_remove_admin_queue;
1934 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1935 WARN_ON_ONCE(!changed);
1937 dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs\n",
1938 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
1940 kref_get(&ctrl->ctrl.kref);
1942 mutex_lock(&nvme_rdma_ctrl_mutex);
1943 list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
1944 mutex_unlock(&nvme_rdma_ctrl_mutex);
1946 if (opts->nr_io_queues) {
1947 nvme_queue_scan(&ctrl->ctrl);
1948 nvme_queue_async_events(&ctrl->ctrl);
1953 out_remove_admin_queue:
1954 nvme_stop_keep_alive(&ctrl->ctrl);
1955 nvme_rdma_destroy_admin_queue(ctrl);
1957 kfree(ctrl->queues);
1959 nvme_uninit_ctrl(&ctrl->ctrl);
1960 nvme_put_ctrl(&ctrl->ctrl);
1963 return ERR_PTR(ret);
1966 return ERR_PTR(ret);
1969 static struct nvmf_transport_ops nvme_rdma_transport = {
1971 .required_opts = NVMF_OPT_TRADDR,
1972 .allowed_opts = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
1973 NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO,
1974 .create_ctrl = nvme_rdma_create_ctrl,
1977 static void nvme_rdma_add_one(struct ib_device *ib_device)
1981 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
1983 struct nvme_rdma_ctrl *ctrl;
1985 /* Delete all controllers using this device */
1986 mutex_lock(&nvme_rdma_ctrl_mutex);
1987 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
1988 if (ctrl->device->dev != ib_device)
1990 dev_info(ctrl->ctrl.device,
1991 "Removing ctrl: NQN \"%s\", addr %pISp\n",
1992 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
1993 __nvme_rdma_del_ctrl(ctrl);
1995 mutex_unlock(&nvme_rdma_ctrl_mutex);
1997 flush_workqueue(nvme_wq);
2000 static struct ib_client nvme_rdma_ib_client = {
2001 .name = "nvme_rdma",
2002 .add = nvme_rdma_add_one,
2003 .remove = nvme_rdma_remove_one
2006 static int __init nvme_rdma_init_module(void)
2010 ret = ib_register_client(&nvme_rdma_ib_client);
2014 ret = nvmf_register_transport(&nvme_rdma_transport);
2016 goto err_unreg_client;
2021 ib_unregister_client(&nvme_rdma_ib_client);
2025 static void __exit nvme_rdma_cleanup_module(void)
2027 nvmf_unregister_transport(&nvme_rdma_transport);
2028 ib_unregister_client(&nvme_rdma_ib_client);
2031 module_init(nvme_rdma_init_module);
2032 module_exit(nvme_rdma_cleanup_module);
2034 MODULE_LICENSE("GPL v2");