2 * NVMe over Fabrics loopback device.
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/scatterlist.h>
16 #include <linux/blk-mq.h>
17 #include <linux/nvme.h>
18 #include <linux/module.h>
19 #include <linux/parser.h>
21 #include "../host/nvme.h"
22 #include "../host/fabrics.h"
24 #define NVME_LOOP_MAX_SEGMENTS 256
26 struct nvme_loop_iod {
27 struct nvme_request nvme_req;
28 struct nvme_command cmd;
29 struct nvme_completion rsp;
31 struct nvme_loop_queue *queue;
32 struct work_struct work;
33 struct sg_table sg_table;
34 struct scatterlist first_sgl[];
37 struct nvme_loop_ctrl {
38 struct nvme_loop_queue *queues;
40 struct blk_mq_tag_set admin_tag_set;
42 struct list_head list;
43 struct blk_mq_tag_set tag_set;
44 struct nvme_loop_iod async_event_iod;
45 struct nvme_ctrl ctrl;
47 struct nvmet_ctrl *target_ctrl;
50 static inline struct nvme_loop_ctrl *to_loop_ctrl(struct nvme_ctrl *ctrl)
52 return container_of(ctrl, struct nvme_loop_ctrl, ctrl);
55 enum nvme_loop_queue_flags {
59 struct nvme_loop_queue {
60 struct nvmet_cq nvme_cq;
61 struct nvmet_sq nvme_sq;
62 struct nvme_loop_ctrl *ctrl;
66 static struct nvmet_port *nvmet_loop_port;
68 static LIST_HEAD(nvme_loop_ctrl_list);
69 static DEFINE_MUTEX(nvme_loop_ctrl_mutex);
71 static void nvme_loop_queue_response(struct nvmet_req *nvme_req);
72 static void nvme_loop_delete_ctrl(struct nvmet_ctrl *ctrl);
74 static const struct nvmet_fabrics_ops nvme_loop_ops;
76 static inline int nvme_loop_queue_idx(struct nvme_loop_queue *queue)
78 return queue - queue->ctrl->queues;
81 static void nvme_loop_complete_rq(struct request *req)
83 struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(req);
85 nvme_cleanup_cmd(req);
86 sg_free_table_chained(&iod->sg_table, true);
87 nvme_complete_rq(req);
90 static struct blk_mq_tags *nvme_loop_tagset(struct nvme_loop_queue *queue)
92 u32 queue_idx = nvme_loop_queue_idx(queue);
95 return queue->ctrl->admin_tag_set.tags[queue_idx];
96 return queue->ctrl->tag_set.tags[queue_idx - 1];
99 static void nvme_loop_queue_response(struct nvmet_req *req)
101 struct nvme_loop_queue *queue =
102 container_of(req->sq, struct nvme_loop_queue, nvme_sq);
103 struct nvme_completion *cqe = req->rsp;
106 * AEN requests are special as they don't time out and can
107 * survive any kind of queue freeze and often don't respond to
108 * aborts. We don't even bother to allocate a struct request
109 * for them but rather special case them here.
111 if (unlikely(nvme_loop_queue_idx(queue) == 0 &&
112 cqe->command_id >= NVME_AQ_BLK_MQ_DEPTH)) {
113 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
118 rq = blk_mq_tag_to_rq(nvme_loop_tagset(queue), cqe->command_id);
120 dev_err(queue->ctrl->ctrl.device,
121 "tag 0x%x on queue %d not found\n",
122 cqe->command_id, nvme_loop_queue_idx(queue));
126 nvme_end_request(rq, cqe->status, cqe->result);
130 static void nvme_loop_execute_work(struct work_struct *work)
132 struct nvme_loop_iod *iod =
133 container_of(work, struct nvme_loop_iod, work);
135 nvmet_req_execute(&iod->req);
138 static enum blk_eh_timer_return
139 nvme_loop_timeout(struct request *rq, bool reserved)
141 struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(rq);
143 /* queue error recovery */
144 nvme_reset_ctrl(&iod->queue->ctrl->ctrl);
146 /* fail with DNR on admin cmd timeout */
147 nvme_req(rq)->status = NVME_SC_ABORT_REQ | NVME_SC_DNR;
149 return BLK_EH_HANDLED;
152 static blk_status_t nvme_loop_queue_rq(struct blk_mq_hw_ctx *hctx,
153 const struct blk_mq_queue_data *bd)
155 struct nvme_ns *ns = hctx->queue->queuedata;
156 struct nvme_loop_queue *queue = hctx->driver_data;
157 struct request *req = bd->rq;
158 struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(req);
161 ret = nvmf_check_if_ready(&queue->ctrl->ctrl, req,
162 test_bit(NVME_LOOP_Q_LIVE, &queue->flags), true);
166 ret = nvme_setup_cmd(ns, req, &iod->cmd);
170 blk_mq_start_request(req);
171 iod->cmd.common.flags |= NVME_CMD_SGL_METABUF;
172 iod->req.port = nvmet_loop_port;
173 if (!nvmet_req_init(&iod->req, &queue->nvme_cq,
174 &queue->nvme_sq, &nvme_loop_ops))
177 if (blk_rq_payload_bytes(req)) {
178 iod->sg_table.sgl = iod->first_sgl;
179 if (sg_alloc_table_chained(&iod->sg_table,
180 blk_rq_nr_phys_segments(req),
182 return BLK_STS_RESOURCE;
184 iod->req.sg = iod->sg_table.sgl;
185 iod->req.sg_cnt = blk_rq_map_sg(req->q, req, iod->sg_table.sgl);
186 iod->req.transfer_len = blk_rq_payload_bytes(req);
189 schedule_work(&iod->work);
193 static void nvme_loop_submit_async_event(struct nvme_ctrl *arg)
195 struct nvme_loop_ctrl *ctrl = to_loop_ctrl(arg);
196 struct nvme_loop_queue *queue = &ctrl->queues[0];
197 struct nvme_loop_iod *iod = &ctrl->async_event_iod;
199 memset(&iod->cmd, 0, sizeof(iod->cmd));
200 iod->cmd.common.opcode = nvme_admin_async_event;
201 iod->cmd.common.command_id = NVME_AQ_BLK_MQ_DEPTH;
202 iod->cmd.common.flags |= NVME_CMD_SGL_METABUF;
204 if (!nvmet_req_init(&iod->req, &queue->nvme_cq, &queue->nvme_sq,
206 dev_err(ctrl->ctrl.device, "failed async event work\n");
210 schedule_work(&iod->work);
213 static int nvme_loop_init_iod(struct nvme_loop_ctrl *ctrl,
214 struct nvme_loop_iod *iod, unsigned int queue_idx)
216 iod->req.cmd = &iod->cmd;
217 iod->req.rsp = &iod->rsp;
218 iod->queue = &ctrl->queues[queue_idx];
219 INIT_WORK(&iod->work, nvme_loop_execute_work);
223 static int nvme_loop_init_request(struct blk_mq_tag_set *set,
224 struct request *req, unsigned int hctx_idx,
225 unsigned int numa_node)
227 struct nvme_loop_ctrl *ctrl = set->driver_data;
229 return nvme_loop_init_iod(ctrl, blk_mq_rq_to_pdu(req),
230 (set == &ctrl->tag_set) ? hctx_idx + 1 : 0);
233 static int nvme_loop_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
234 unsigned int hctx_idx)
236 struct nvme_loop_ctrl *ctrl = data;
237 struct nvme_loop_queue *queue = &ctrl->queues[hctx_idx + 1];
239 BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
241 hctx->driver_data = queue;
245 static int nvme_loop_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
246 unsigned int hctx_idx)
248 struct nvme_loop_ctrl *ctrl = data;
249 struct nvme_loop_queue *queue = &ctrl->queues[0];
251 BUG_ON(hctx_idx != 0);
253 hctx->driver_data = queue;
257 static const struct blk_mq_ops nvme_loop_mq_ops = {
258 .queue_rq = nvme_loop_queue_rq,
259 .complete = nvme_loop_complete_rq,
260 .init_request = nvme_loop_init_request,
261 .init_hctx = nvme_loop_init_hctx,
262 .timeout = nvme_loop_timeout,
265 static const struct blk_mq_ops nvme_loop_admin_mq_ops = {
266 .queue_rq = nvme_loop_queue_rq,
267 .complete = nvme_loop_complete_rq,
268 .init_request = nvme_loop_init_request,
269 .init_hctx = nvme_loop_init_admin_hctx,
270 .timeout = nvme_loop_timeout,
273 static void nvme_loop_destroy_admin_queue(struct nvme_loop_ctrl *ctrl)
275 clear_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[0].flags);
276 nvmet_sq_destroy(&ctrl->queues[0].nvme_sq);
277 blk_cleanup_queue(ctrl->ctrl.admin_q);
278 blk_mq_free_tag_set(&ctrl->admin_tag_set);
281 static void nvme_loop_free_ctrl(struct nvme_ctrl *nctrl)
283 struct nvme_loop_ctrl *ctrl = to_loop_ctrl(nctrl);
285 if (list_empty(&ctrl->list))
288 mutex_lock(&nvme_loop_ctrl_mutex);
289 list_del(&ctrl->list);
290 mutex_unlock(&nvme_loop_ctrl_mutex);
293 blk_cleanup_queue(ctrl->ctrl.connect_q);
294 blk_mq_free_tag_set(&ctrl->tag_set);
297 nvmf_free_options(nctrl->opts);
302 static void nvme_loop_destroy_io_queues(struct nvme_loop_ctrl *ctrl)
306 for (i = 1; i < ctrl->ctrl.queue_count; i++) {
307 clear_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[i].flags);
308 nvmet_sq_destroy(&ctrl->queues[i].nvme_sq);
312 static int nvme_loop_init_io_queues(struct nvme_loop_ctrl *ctrl)
314 struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
315 unsigned int nr_io_queues;
318 nr_io_queues = min(opts->nr_io_queues, num_online_cpus());
319 ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
320 if (ret || !nr_io_queues)
323 dev_info(ctrl->ctrl.device, "creating %d I/O queues.\n", nr_io_queues);
325 for (i = 1; i <= nr_io_queues; i++) {
326 ctrl->queues[i].ctrl = ctrl;
327 ret = nvmet_sq_init(&ctrl->queues[i].nvme_sq);
329 goto out_destroy_queues;
331 ctrl->ctrl.queue_count++;
337 nvme_loop_destroy_io_queues(ctrl);
341 static int nvme_loop_connect_io_queues(struct nvme_loop_ctrl *ctrl)
345 for (i = 1; i < ctrl->ctrl.queue_count; i++) {
346 ret = nvmf_connect_io_queue(&ctrl->ctrl, i);
349 set_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[i].flags);
355 static int nvme_loop_configure_admin_queue(struct nvme_loop_ctrl *ctrl)
359 memset(&ctrl->admin_tag_set, 0, sizeof(ctrl->admin_tag_set));
360 ctrl->admin_tag_set.ops = &nvme_loop_admin_mq_ops;
361 ctrl->admin_tag_set.queue_depth = NVME_AQ_MQ_TAG_DEPTH;
362 ctrl->admin_tag_set.reserved_tags = 2; /* connect + keep-alive */
363 ctrl->admin_tag_set.numa_node = NUMA_NO_NODE;
364 ctrl->admin_tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
365 SG_CHUNK_SIZE * sizeof(struct scatterlist);
366 ctrl->admin_tag_set.driver_data = ctrl;
367 ctrl->admin_tag_set.nr_hw_queues = 1;
368 ctrl->admin_tag_set.timeout = ADMIN_TIMEOUT;
369 ctrl->admin_tag_set.flags = BLK_MQ_F_NO_SCHED;
371 ctrl->queues[0].ctrl = ctrl;
372 error = nvmet_sq_init(&ctrl->queues[0].nvme_sq);
375 ctrl->ctrl.queue_count = 1;
377 error = blk_mq_alloc_tag_set(&ctrl->admin_tag_set);
380 ctrl->ctrl.admin_tagset = &ctrl->admin_tag_set;
382 ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
383 if (IS_ERR(ctrl->ctrl.admin_q)) {
384 error = PTR_ERR(ctrl->ctrl.admin_q);
385 goto out_free_tagset;
388 error = nvmf_connect_admin_queue(&ctrl->ctrl);
390 goto out_cleanup_queue;
392 set_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[0].flags);
394 error = nvmf_reg_read64(&ctrl->ctrl, NVME_REG_CAP, &ctrl->ctrl.cap);
396 dev_err(ctrl->ctrl.device,
397 "prop_get NVME_REG_CAP failed\n");
398 goto out_cleanup_queue;
402 min_t(int, NVME_CAP_MQES(ctrl->ctrl.cap), ctrl->ctrl.sqsize);
404 error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
406 goto out_cleanup_queue;
408 ctrl->ctrl.max_hw_sectors =
409 (NVME_LOOP_MAX_SEGMENTS - 1) << (PAGE_SHIFT - 9);
411 error = nvme_init_identify(&ctrl->ctrl);
413 goto out_cleanup_queue;
418 blk_cleanup_queue(ctrl->ctrl.admin_q);
420 blk_mq_free_tag_set(&ctrl->admin_tag_set);
422 nvmet_sq_destroy(&ctrl->queues[0].nvme_sq);
426 static void nvme_loop_shutdown_ctrl(struct nvme_loop_ctrl *ctrl)
428 if (ctrl->ctrl.queue_count > 1) {
429 nvme_stop_queues(&ctrl->ctrl);
430 blk_mq_tagset_busy_iter(&ctrl->tag_set,
431 nvme_cancel_request, &ctrl->ctrl);
432 nvme_loop_destroy_io_queues(ctrl);
435 if (ctrl->ctrl.state == NVME_CTRL_LIVE)
436 nvme_shutdown_ctrl(&ctrl->ctrl);
438 blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
439 blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
440 nvme_cancel_request, &ctrl->ctrl);
441 blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
442 nvme_loop_destroy_admin_queue(ctrl);
445 static void nvme_loop_delete_ctrl_host(struct nvme_ctrl *ctrl)
447 nvme_loop_shutdown_ctrl(to_loop_ctrl(ctrl));
450 static void nvme_loop_delete_ctrl(struct nvmet_ctrl *nctrl)
452 struct nvme_loop_ctrl *ctrl;
454 mutex_lock(&nvme_loop_ctrl_mutex);
455 list_for_each_entry(ctrl, &nvme_loop_ctrl_list, list) {
456 if (ctrl->ctrl.cntlid == nctrl->cntlid)
457 nvme_delete_ctrl(&ctrl->ctrl);
459 mutex_unlock(&nvme_loop_ctrl_mutex);
462 static void nvme_loop_reset_ctrl_work(struct work_struct *work)
464 struct nvme_loop_ctrl *ctrl =
465 container_of(work, struct nvme_loop_ctrl, ctrl.reset_work);
469 nvme_stop_ctrl(&ctrl->ctrl);
470 nvme_loop_shutdown_ctrl(ctrl);
472 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
473 /* state change failure should never happen */
478 ret = nvme_loop_configure_admin_queue(ctrl);
482 ret = nvme_loop_init_io_queues(ctrl);
484 goto out_destroy_admin;
486 ret = nvme_loop_connect_io_queues(ctrl);
490 blk_mq_update_nr_hw_queues(&ctrl->tag_set,
491 ctrl->ctrl.queue_count - 1);
493 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
494 WARN_ON_ONCE(!changed);
496 nvme_start_ctrl(&ctrl->ctrl);
501 nvme_loop_destroy_io_queues(ctrl);
503 nvme_loop_destroy_admin_queue(ctrl);
505 dev_warn(ctrl->ctrl.device, "Removing after reset failure\n");
506 nvme_uninit_ctrl(&ctrl->ctrl);
507 nvme_put_ctrl(&ctrl->ctrl);
510 static const struct nvme_ctrl_ops nvme_loop_ctrl_ops = {
512 .module = THIS_MODULE,
513 .flags = NVME_F_FABRICS,
514 .reg_read32 = nvmf_reg_read32,
515 .reg_read64 = nvmf_reg_read64,
516 .reg_write32 = nvmf_reg_write32,
517 .free_ctrl = nvme_loop_free_ctrl,
518 .submit_async_event = nvme_loop_submit_async_event,
519 .delete_ctrl = nvme_loop_delete_ctrl_host,
522 static int nvme_loop_create_io_queues(struct nvme_loop_ctrl *ctrl)
526 ret = nvme_loop_init_io_queues(ctrl);
530 memset(&ctrl->tag_set, 0, sizeof(ctrl->tag_set));
531 ctrl->tag_set.ops = &nvme_loop_mq_ops;
532 ctrl->tag_set.queue_depth = ctrl->ctrl.opts->queue_size;
533 ctrl->tag_set.reserved_tags = 1; /* fabric connect */
534 ctrl->tag_set.numa_node = NUMA_NO_NODE;
535 ctrl->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
536 ctrl->tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
537 SG_CHUNK_SIZE * sizeof(struct scatterlist);
538 ctrl->tag_set.driver_data = ctrl;
539 ctrl->tag_set.nr_hw_queues = ctrl->ctrl.queue_count - 1;
540 ctrl->tag_set.timeout = NVME_IO_TIMEOUT;
541 ctrl->ctrl.tagset = &ctrl->tag_set;
543 ret = blk_mq_alloc_tag_set(&ctrl->tag_set);
545 goto out_destroy_queues;
547 ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
548 if (IS_ERR(ctrl->ctrl.connect_q)) {
549 ret = PTR_ERR(ctrl->ctrl.connect_q);
550 goto out_free_tagset;
553 ret = nvme_loop_connect_io_queues(ctrl);
555 goto out_cleanup_connect_q;
559 out_cleanup_connect_q:
560 blk_cleanup_queue(ctrl->ctrl.connect_q);
562 blk_mq_free_tag_set(&ctrl->tag_set);
564 nvme_loop_destroy_io_queues(ctrl);
568 static struct nvme_ctrl *nvme_loop_create_ctrl(struct device *dev,
569 struct nvmf_ctrl_options *opts)
571 struct nvme_loop_ctrl *ctrl;
575 ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
577 return ERR_PTR(-ENOMEM);
578 ctrl->ctrl.opts = opts;
579 INIT_LIST_HEAD(&ctrl->list);
581 INIT_WORK(&ctrl->ctrl.reset_work, nvme_loop_reset_ctrl_work);
583 ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_loop_ctrl_ops,
584 0 /* no quirks, we're perfect! */);
590 ctrl->ctrl.sqsize = opts->queue_size - 1;
591 ctrl->ctrl.kato = opts->kato;
593 ctrl->queues = kcalloc(opts->nr_io_queues + 1, sizeof(*ctrl->queues),
596 goto out_uninit_ctrl;
598 ret = nvme_loop_configure_admin_queue(ctrl);
600 goto out_free_queues;
602 if (opts->queue_size > ctrl->ctrl.maxcmd) {
603 /* warn if maxcmd is lower than queue_size */
604 dev_warn(ctrl->ctrl.device,
605 "queue_size %zu > ctrl maxcmd %u, clamping down\n",
606 opts->queue_size, ctrl->ctrl.maxcmd);
607 opts->queue_size = ctrl->ctrl.maxcmd;
610 if (opts->nr_io_queues) {
611 ret = nvme_loop_create_io_queues(ctrl);
613 goto out_remove_admin_queue;
616 nvme_loop_init_iod(ctrl, &ctrl->async_event_iod, 0);
618 dev_info(ctrl->ctrl.device,
619 "new ctrl: \"%s\"\n", ctrl->ctrl.opts->subsysnqn);
621 nvme_get_ctrl(&ctrl->ctrl);
623 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
624 WARN_ON_ONCE(!changed);
626 mutex_lock(&nvme_loop_ctrl_mutex);
627 list_add_tail(&ctrl->list, &nvme_loop_ctrl_list);
628 mutex_unlock(&nvme_loop_ctrl_mutex);
630 nvme_start_ctrl(&ctrl->ctrl);
634 out_remove_admin_queue:
635 nvme_loop_destroy_admin_queue(ctrl);
639 nvme_uninit_ctrl(&ctrl->ctrl);
641 nvme_put_ctrl(&ctrl->ctrl);
647 static int nvme_loop_add_port(struct nvmet_port *port)
650 * XXX: disalow adding more than one port so
651 * there is no connection rejections when a
652 * a subsystem is assigned to a port for which
653 * loop doesn't have a pointer.
654 * This scenario would be possible if we allowed
655 * more than one port to be added and a subsystem
656 * was assigned to a port other than nvmet_loop_port.
662 nvmet_loop_port = port;
666 static void nvme_loop_remove_port(struct nvmet_port *port)
668 if (port == nvmet_loop_port)
669 nvmet_loop_port = NULL;
672 static const struct nvmet_fabrics_ops nvme_loop_ops = {
673 .owner = THIS_MODULE,
674 .type = NVMF_TRTYPE_LOOP,
675 .add_port = nvme_loop_add_port,
676 .remove_port = nvme_loop_remove_port,
677 .queue_response = nvme_loop_queue_response,
678 .delete_ctrl = nvme_loop_delete_ctrl,
681 static struct nvmf_transport_ops nvme_loop_transport = {
683 .module = THIS_MODULE,
684 .create_ctrl = nvme_loop_create_ctrl,
687 static int __init nvme_loop_init_module(void)
691 ret = nvmet_register_transport(&nvme_loop_ops);
695 ret = nvmf_register_transport(&nvme_loop_transport);
697 nvmet_unregister_transport(&nvme_loop_ops);
702 static void __exit nvme_loop_cleanup_module(void)
704 struct nvme_loop_ctrl *ctrl, *next;
706 nvmf_unregister_transport(&nvme_loop_transport);
707 nvmet_unregister_transport(&nvme_loop_ops);
709 mutex_lock(&nvme_loop_ctrl_mutex);
710 list_for_each_entry_safe(ctrl, next, &nvme_loop_ctrl_list, list)
711 nvme_delete_ctrl(&ctrl->ctrl);
712 mutex_unlock(&nvme_loop_ctrl_mutex);
714 flush_workqueue(nvme_delete_wq);
717 module_init(nvme_loop_init_module);
718 module_exit(nvme_loop_cleanup_module);
720 MODULE_LICENSE("GPL v2");
721 MODULE_ALIAS("nvmet-transport-254"); /* 254 == NVMF_TRTYPE_LOOP */