1 | // SPDX-License-Identifier: GPL-2.0 |
2 | /* |
3 | * NVMe over Fabrics loopback device. |
4 | * Copyright (c) 2015-2016 HGST, a Western Digital Company. |
5 | */ |
6 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt |
7 | #include <linux/scatterlist.h> |
8 | #include <linux/blk-mq.h> |
9 | #include <linux/nvme.h> |
10 | #include <linux/module.h> |
11 | #include <linux/parser.h> |
12 | #include "nvmet.h" |
13 | #include "../host/nvme.h" |
14 | #include "../host/fabrics.h" |
15 | |
16 | #define NVME_LOOP_MAX_SEGMENTS 256 |
17 | |
18 | struct nvme_loop_iod { |
19 | struct nvme_request nvme_req; |
20 | struct nvme_command cmd; |
21 | struct nvme_completion cqe; |
22 | struct nvmet_req req; |
23 | struct nvme_loop_queue *queue; |
24 | struct work_struct work; |
25 | struct sg_table sg_table; |
26 | struct scatterlist first_sgl[]; |
27 | }; |
28 | |
29 | struct nvme_loop_ctrl { |
30 | struct nvme_loop_queue *queues; |
31 | |
32 | struct blk_mq_tag_set admin_tag_set; |
33 | |
34 | struct list_head list; |
35 | struct blk_mq_tag_set tag_set; |
36 | struct nvme_loop_iod async_event_iod; |
37 | struct nvme_ctrl ctrl; |
38 | |
39 | struct nvmet_port *port; |
40 | }; |
41 | |
42 | static inline struct nvme_loop_ctrl *to_loop_ctrl(struct nvme_ctrl *ctrl) |
43 | { |
44 | return container_of(ctrl, struct nvme_loop_ctrl, ctrl); |
45 | } |
46 | |
47 | enum nvme_loop_queue_flags { |
48 | NVME_LOOP_Q_LIVE = 0, |
49 | }; |
50 | |
51 | struct nvme_loop_queue { |
52 | struct nvmet_cq nvme_cq; |
53 | struct nvmet_sq nvme_sq; |
54 | struct nvme_loop_ctrl *ctrl; |
55 | unsigned long flags; |
56 | }; |
57 | |
58 | static LIST_HEAD(nvme_loop_ports); |
59 | static DEFINE_MUTEX(nvme_loop_ports_mutex); |
60 | |
61 | static LIST_HEAD(nvme_loop_ctrl_list); |
62 | static DEFINE_MUTEX(nvme_loop_ctrl_mutex); |
63 | |
64 | static void nvme_loop_queue_response(struct nvmet_req *nvme_req); |
65 | static void nvme_loop_delete_ctrl(struct nvmet_ctrl *ctrl); |
66 | |
67 | static const struct nvmet_fabrics_ops nvme_loop_ops; |
68 | |
69 | static inline int nvme_loop_queue_idx(struct nvme_loop_queue *queue) |
70 | { |
71 | return queue - queue->ctrl->queues; |
72 | } |
73 | |
74 | static void nvme_loop_complete_rq(struct request *req) |
75 | { |
76 | struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(rq: req); |
77 | |
78 | sg_free_table_chained(table: &iod->sg_table, NVME_INLINE_SG_CNT); |
79 | nvme_complete_rq(req); |
80 | } |
81 | |
82 | static struct blk_mq_tags *nvme_loop_tagset(struct nvme_loop_queue *queue) |
83 | { |
84 | u32 queue_idx = nvme_loop_queue_idx(queue); |
85 | |
86 | if (queue_idx == 0) |
87 | return queue->ctrl->admin_tag_set.tags[queue_idx]; |
88 | return queue->ctrl->tag_set.tags[queue_idx - 1]; |
89 | } |
90 | |
91 | static void nvme_loop_queue_response(struct nvmet_req *req) |
92 | { |
93 | struct nvme_loop_queue *queue = |
94 | container_of(req->sq, struct nvme_loop_queue, nvme_sq); |
95 | struct nvme_completion *cqe = req->cqe; |
96 | |
97 | /* |
98 | * AEN requests are special as they don't time out and can |
99 | * survive any kind of queue freeze and often don't respond to |
100 | * aborts. We don't even bother to allocate a struct request |
101 | * for them but rather special case them here. |
102 | */ |
103 | if (unlikely(nvme_is_aen_req(nvme_loop_queue_idx(queue), |
104 | cqe->command_id))) { |
105 | nvme_complete_async_event(ctrl: &queue->ctrl->ctrl, status: cqe->status, |
106 | res: &cqe->result); |
107 | } else { |
108 | struct request *rq; |
109 | |
110 | rq = nvme_find_rq(tags: nvme_loop_tagset(queue), command_id: cqe->command_id); |
111 | if (!rq) { |
112 | dev_err(queue->ctrl->ctrl.device, |
113 | "got bad command_id %#x on queue %d\n" , |
114 | cqe->command_id, nvme_loop_queue_idx(queue)); |
115 | return; |
116 | } |
117 | |
118 | if (!nvme_try_complete_req(req: rq, status: cqe->status, result: cqe->result)) |
119 | nvme_loop_complete_rq(req: rq); |
120 | } |
121 | } |
122 | |
123 | static void nvme_loop_execute_work(struct work_struct *work) |
124 | { |
125 | struct nvme_loop_iod *iod = |
126 | container_of(work, struct nvme_loop_iod, work); |
127 | |
128 | iod->req.execute(&iod->req); |
129 | } |
130 | |
131 | static blk_status_t nvme_loop_queue_rq(struct blk_mq_hw_ctx *hctx, |
132 | const struct blk_mq_queue_data *bd) |
133 | { |
134 | struct nvme_ns *ns = hctx->queue->queuedata; |
135 | struct nvme_loop_queue *queue = hctx->driver_data; |
136 | struct request *req = bd->rq; |
137 | struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(rq: req); |
138 | bool queue_ready = test_bit(NVME_LOOP_Q_LIVE, &queue->flags); |
139 | blk_status_t ret; |
140 | |
141 | if (!nvme_check_ready(ctrl: &queue->ctrl->ctrl, rq: req, queue_live: queue_ready)) |
142 | return nvme_fail_nonready_command(ctrl: &queue->ctrl->ctrl, req); |
143 | |
144 | ret = nvme_setup_cmd(ns, req); |
145 | if (ret) |
146 | return ret; |
147 | |
148 | nvme_start_request(rq: req); |
149 | iod->cmd.common.flags |= NVME_CMD_SGL_METABUF; |
150 | iod->req.port = queue->ctrl->port; |
151 | if (!nvmet_req_init(req: &iod->req, cq: &queue->nvme_cq, |
152 | sq: &queue->nvme_sq, ops: &nvme_loop_ops)) |
153 | return BLK_STS_OK; |
154 | |
155 | if (blk_rq_nr_phys_segments(rq: req)) { |
156 | iod->sg_table.sgl = iod->first_sgl; |
157 | if (sg_alloc_table_chained(table: &iod->sg_table, |
158 | nents: blk_rq_nr_phys_segments(rq: req), |
159 | first_chunk: iod->sg_table.sgl, NVME_INLINE_SG_CNT)) { |
160 | nvme_cleanup_cmd(req); |
161 | return BLK_STS_RESOURCE; |
162 | } |
163 | |
164 | iod->req.sg = iod->sg_table.sgl; |
165 | iod->req.sg_cnt = blk_rq_map_sg(q: req->q, rq: req, sglist: iod->sg_table.sgl); |
166 | iod->req.transfer_len = blk_rq_payload_bytes(rq: req); |
167 | } |
168 | |
169 | queue_work(wq: nvmet_wq, work: &iod->work); |
170 | return BLK_STS_OK; |
171 | } |
172 | |
173 | static void nvme_loop_submit_async_event(struct nvme_ctrl *arg) |
174 | { |
175 | struct nvme_loop_ctrl *ctrl = to_loop_ctrl(ctrl: arg); |
176 | struct nvme_loop_queue *queue = &ctrl->queues[0]; |
177 | struct nvme_loop_iod *iod = &ctrl->async_event_iod; |
178 | |
179 | memset(&iod->cmd, 0, sizeof(iod->cmd)); |
180 | iod->cmd.common.opcode = nvme_admin_async_event; |
181 | iod->cmd.common.command_id = NVME_AQ_BLK_MQ_DEPTH; |
182 | iod->cmd.common.flags |= NVME_CMD_SGL_METABUF; |
183 | |
184 | if (!nvmet_req_init(req: &iod->req, cq: &queue->nvme_cq, sq: &queue->nvme_sq, |
185 | ops: &nvme_loop_ops)) { |
186 | dev_err(ctrl->ctrl.device, "failed async event work\n" ); |
187 | return; |
188 | } |
189 | |
190 | queue_work(wq: nvmet_wq, work: &iod->work); |
191 | } |
192 | |
193 | static int nvme_loop_init_iod(struct nvme_loop_ctrl *ctrl, |
194 | struct nvme_loop_iod *iod, unsigned int queue_idx) |
195 | { |
196 | iod->req.cmd = &iod->cmd; |
197 | iod->req.cqe = &iod->cqe; |
198 | iod->queue = &ctrl->queues[queue_idx]; |
199 | INIT_WORK(&iod->work, nvme_loop_execute_work); |
200 | return 0; |
201 | } |
202 | |
203 | static int nvme_loop_init_request(struct blk_mq_tag_set *set, |
204 | struct request *req, unsigned int hctx_idx, |
205 | unsigned int numa_node) |
206 | { |
207 | struct nvme_loop_ctrl *ctrl = to_loop_ctrl(ctrl: set->driver_data); |
208 | struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(rq: req); |
209 | |
210 | nvme_req(req)->ctrl = &ctrl->ctrl; |
211 | nvme_req(req)->cmd = &iod->cmd; |
212 | return nvme_loop_init_iod(ctrl, iod: blk_mq_rq_to_pdu(rq: req), |
213 | queue_idx: (set == &ctrl->tag_set) ? hctx_idx + 1 : 0); |
214 | } |
215 | |
216 | static struct lock_class_key loop_hctx_fq_lock_key; |
217 | |
218 | static int nvme_loop_init_hctx(struct blk_mq_hw_ctx *hctx, void *data, |
219 | unsigned int hctx_idx) |
220 | { |
221 | struct nvme_loop_ctrl *ctrl = to_loop_ctrl(ctrl: data); |
222 | struct nvme_loop_queue *queue = &ctrl->queues[hctx_idx + 1]; |
223 | |
224 | BUG_ON(hctx_idx >= ctrl->ctrl.queue_count); |
225 | |
226 | /* |
227 | * flush_end_io() can be called recursively for us, so use our own |
228 | * lock class key for avoiding lockdep possible recursive locking, |
229 | * then we can remove the dynamically allocated lock class for each |
230 | * flush queue, that way may cause horrible boot delay. |
231 | */ |
232 | blk_mq_hctx_set_fq_lock_class(hctx, key: &loop_hctx_fq_lock_key); |
233 | |
234 | hctx->driver_data = queue; |
235 | return 0; |
236 | } |
237 | |
238 | static int nvme_loop_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data, |
239 | unsigned int hctx_idx) |
240 | { |
241 | struct nvme_loop_ctrl *ctrl = to_loop_ctrl(ctrl: data); |
242 | struct nvme_loop_queue *queue = &ctrl->queues[0]; |
243 | |
244 | BUG_ON(hctx_idx != 0); |
245 | |
246 | hctx->driver_data = queue; |
247 | return 0; |
248 | } |
249 | |
250 | static const struct blk_mq_ops nvme_loop_mq_ops = { |
251 | .queue_rq = nvme_loop_queue_rq, |
252 | .complete = nvme_loop_complete_rq, |
253 | .init_request = nvme_loop_init_request, |
254 | .init_hctx = nvme_loop_init_hctx, |
255 | }; |
256 | |
257 | static const struct blk_mq_ops nvme_loop_admin_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_admin_hctx, |
262 | }; |
263 | |
264 | static void nvme_loop_destroy_admin_queue(struct nvme_loop_ctrl *ctrl) |
265 | { |
266 | if (!test_and_clear_bit(nr: NVME_LOOP_Q_LIVE, addr: &ctrl->queues[0].flags)) |
267 | return; |
268 | nvmet_sq_destroy(sq: &ctrl->queues[0].nvme_sq); |
269 | nvme_remove_admin_tag_set(ctrl: &ctrl->ctrl); |
270 | } |
271 | |
272 | static void nvme_loop_free_ctrl(struct nvme_ctrl *nctrl) |
273 | { |
274 | struct nvme_loop_ctrl *ctrl = to_loop_ctrl(ctrl: nctrl); |
275 | |
276 | if (list_empty(head: &ctrl->list)) |
277 | goto free_ctrl; |
278 | |
279 | mutex_lock(&nvme_loop_ctrl_mutex); |
280 | list_del(entry: &ctrl->list); |
281 | mutex_unlock(lock: &nvme_loop_ctrl_mutex); |
282 | |
283 | if (nctrl->tagset) |
284 | nvme_remove_io_tag_set(ctrl: nctrl); |
285 | kfree(objp: ctrl->queues); |
286 | nvmf_free_options(opts: nctrl->opts); |
287 | free_ctrl: |
288 | kfree(objp: ctrl); |
289 | } |
290 | |
291 | static void nvme_loop_destroy_io_queues(struct nvme_loop_ctrl *ctrl) |
292 | { |
293 | int i; |
294 | |
295 | for (i = 1; i < ctrl->ctrl.queue_count; i++) { |
296 | clear_bit(nr: NVME_LOOP_Q_LIVE, addr: &ctrl->queues[i].flags); |
297 | nvmet_sq_destroy(sq: &ctrl->queues[i].nvme_sq); |
298 | } |
299 | ctrl->ctrl.queue_count = 1; |
300 | } |
301 | |
302 | static int nvme_loop_init_io_queues(struct nvme_loop_ctrl *ctrl) |
303 | { |
304 | struct nvmf_ctrl_options *opts = ctrl->ctrl.opts; |
305 | unsigned int nr_io_queues; |
306 | int ret, i; |
307 | |
308 | nr_io_queues = min(opts->nr_io_queues, num_online_cpus()); |
309 | ret = nvme_set_queue_count(ctrl: &ctrl->ctrl, count: &nr_io_queues); |
310 | if (ret || !nr_io_queues) |
311 | return ret; |
312 | |
313 | dev_info(ctrl->ctrl.device, "creating %d I/O queues.\n" , nr_io_queues); |
314 | |
315 | for (i = 1; i <= nr_io_queues; i++) { |
316 | ctrl->queues[i].ctrl = ctrl; |
317 | ret = nvmet_sq_init(sq: &ctrl->queues[i].nvme_sq); |
318 | if (ret) |
319 | goto out_destroy_queues; |
320 | |
321 | ctrl->ctrl.queue_count++; |
322 | } |
323 | |
324 | return 0; |
325 | |
326 | out_destroy_queues: |
327 | nvme_loop_destroy_io_queues(ctrl); |
328 | return ret; |
329 | } |
330 | |
331 | static int nvme_loop_connect_io_queues(struct nvme_loop_ctrl *ctrl) |
332 | { |
333 | int i, ret; |
334 | |
335 | for (i = 1; i < ctrl->ctrl.queue_count; i++) { |
336 | ret = nvmf_connect_io_queue(ctrl: &ctrl->ctrl, qid: i); |
337 | if (ret) |
338 | return ret; |
339 | set_bit(nr: NVME_LOOP_Q_LIVE, addr: &ctrl->queues[i].flags); |
340 | } |
341 | |
342 | return 0; |
343 | } |
344 | |
345 | static int nvme_loop_configure_admin_queue(struct nvme_loop_ctrl *ctrl) |
346 | { |
347 | int error; |
348 | |
349 | ctrl->queues[0].ctrl = ctrl; |
350 | error = nvmet_sq_init(sq: &ctrl->queues[0].nvme_sq); |
351 | if (error) |
352 | return error; |
353 | ctrl->ctrl.queue_count = 1; |
354 | |
355 | error = nvme_alloc_admin_tag_set(ctrl: &ctrl->ctrl, set: &ctrl->admin_tag_set, |
356 | ops: &nvme_loop_admin_mq_ops, |
357 | cmd_size: sizeof(struct nvme_loop_iod) + |
358 | NVME_INLINE_SG_CNT * sizeof(struct scatterlist)); |
359 | if (error) |
360 | goto out_free_sq; |
361 | |
362 | /* reset stopped state for the fresh admin queue */ |
363 | clear_bit(nr: NVME_CTRL_ADMIN_Q_STOPPED, addr: &ctrl->ctrl.flags); |
364 | |
365 | error = nvmf_connect_admin_queue(ctrl: &ctrl->ctrl); |
366 | if (error) |
367 | goto out_cleanup_tagset; |
368 | |
369 | set_bit(nr: NVME_LOOP_Q_LIVE, addr: &ctrl->queues[0].flags); |
370 | |
371 | error = nvme_enable_ctrl(ctrl: &ctrl->ctrl); |
372 | if (error) |
373 | goto out_cleanup_tagset; |
374 | |
375 | ctrl->ctrl.max_hw_sectors = |
376 | (NVME_LOOP_MAX_SEGMENTS - 1) << PAGE_SECTORS_SHIFT; |
377 | |
378 | nvme_unquiesce_admin_queue(ctrl: &ctrl->ctrl); |
379 | |
380 | error = nvme_init_ctrl_finish(ctrl: &ctrl->ctrl, was_suspended: false); |
381 | if (error) |
382 | goto out_cleanup_tagset; |
383 | |
384 | return 0; |
385 | |
386 | out_cleanup_tagset: |
387 | clear_bit(nr: NVME_LOOP_Q_LIVE, addr: &ctrl->queues[0].flags); |
388 | nvme_remove_admin_tag_set(ctrl: &ctrl->ctrl); |
389 | out_free_sq: |
390 | nvmet_sq_destroy(sq: &ctrl->queues[0].nvme_sq); |
391 | return error; |
392 | } |
393 | |
394 | static void nvme_loop_shutdown_ctrl(struct nvme_loop_ctrl *ctrl) |
395 | { |
396 | if (ctrl->ctrl.queue_count > 1) { |
397 | nvme_quiesce_io_queues(ctrl: &ctrl->ctrl); |
398 | nvme_cancel_tagset(ctrl: &ctrl->ctrl); |
399 | nvme_loop_destroy_io_queues(ctrl); |
400 | } |
401 | |
402 | nvme_quiesce_admin_queue(ctrl: &ctrl->ctrl); |
403 | if (ctrl->ctrl.state == NVME_CTRL_LIVE) |
404 | nvme_disable_ctrl(ctrl: &ctrl->ctrl, shutdown: true); |
405 | |
406 | nvme_cancel_admin_tagset(ctrl: &ctrl->ctrl); |
407 | nvme_loop_destroy_admin_queue(ctrl); |
408 | } |
409 | |
410 | static void nvme_loop_delete_ctrl_host(struct nvme_ctrl *ctrl) |
411 | { |
412 | nvme_loop_shutdown_ctrl(ctrl: to_loop_ctrl(ctrl)); |
413 | } |
414 | |
415 | static void nvme_loop_delete_ctrl(struct nvmet_ctrl *nctrl) |
416 | { |
417 | struct nvme_loop_ctrl *ctrl; |
418 | |
419 | mutex_lock(&nvme_loop_ctrl_mutex); |
420 | list_for_each_entry(ctrl, &nvme_loop_ctrl_list, list) { |
421 | if (ctrl->ctrl.cntlid == nctrl->cntlid) |
422 | nvme_delete_ctrl(ctrl: &ctrl->ctrl); |
423 | } |
424 | mutex_unlock(lock: &nvme_loop_ctrl_mutex); |
425 | } |
426 | |
427 | static void nvme_loop_reset_ctrl_work(struct work_struct *work) |
428 | { |
429 | struct nvme_loop_ctrl *ctrl = |
430 | container_of(work, struct nvme_loop_ctrl, ctrl.reset_work); |
431 | int ret; |
432 | |
433 | nvme_stop_ctrl(ctrl: &ctrl->ctrl); |
434 | nvme_loop_shutdown_ctrl(ctrl); |
435 | |
436 | if (!nvme_change_ctrl_state(ctrl: &ctrl->ctrl, new_state: NVME_CTRL_CONNECTING)) { |
437 | if (ctrl->ctrl.state != NVME_CTRL_DELETING && |
438 | ctrl->ctrl.state != NVME_CTRL_DELETING_NOIO) |
439 | /* state change failure for non-deleted ctrl? */ |
440 | WARN_ON_ONCE(1); |
441 | return; |
442 | } |
443 | |
444 | ret = nvme_loop_configure_admin_queue(ctrl); |
445 | if (ret) |
446 | goto out_disable; |
447 | |
448 | ret = nvme_loop_init_io_queues(ctrl); |
449 | if (ret) |
450 | goto out_destroy_admin; |
451 | |
452 | ret = nvme_loop_connect_io_queues(ctrl); |
453 | if (ret) |
454 | goto out_destroy_io; |
455 | |
456 | blk_mq_update_nr_hw_queues(set: &ctrl->tag_set, |
457 | nr_hw_queues: ctrl->ctrl.queue_count - 1); |
458 | |
459 | if (!nvme_change_ctrl_state(ctrl: &ctrl->ctrl, new_state: NVME_CTRL_LIVE)) |
460 | WARN_ON_ONCE(1); |
461 | |
462 | nvme_start_ctrl(ctrl: &ctrl->ctrl); |
463 | |
464 | return; |
465 | |
466 | out_destroy_io: |
467 | nvme_loop_destroy_io_queues(ctrl); |
468 | out_destroy_admin: |
469 | nvme_loop_destroy_admin_queue(ctrl); |
470 | out_disable: |
471 | dev_warn(ctrl->ctrl.device, "Removing after reset failure\n" ); |
472 | nvme_uninit_ctrl(ctrl: &ctrl->ctrl); |
473 | } |
474 | |
475 | static const struct nvme_ctrl_ops nvme_loop_ctrl_ops = { |
476 | .name = "loop" , |
477 | .module = THIS_MODULE, |
478 | .flags = NVME_F_FABRICS, |
479 | .reg_read32 = nvmf_reg_read32, |
480 | .reg_read64 = nvmf_reg_read64, |
481 | .reg_write32 = nvmf_reg_write32, |
482 | .free_ctrl = nvme_loop_free_ctrl, |
483 | .submit_async_event = nvme_loop_submit_async_event, |
484 | .delete_ctrl = nvme_loop_delete_ctrl_host, |
485 | .get_address = nvmf_get_address, |
486 | }; |
487 | |
488 | static int nvme_loop_create_io_queues(struct nvme_loop_ctrl *ctrl) |
489 | { |
490 | int ret; |
491 | |
492 | ret = nvme_loop_init_io_queues(ctrl); |
493 | if (ret) |
494 | return ret; |
495 | |
496 | ret = nvme_alloc_io_tag_set(ctrl: &ctrl->ctrl, set: &ctrl->tag_set, |
497 | ops: &nvme_loop_mq_ops, nr_maps: 1, |
498 | cmd_size: sizeof(struct nvme_loop_iod) + |
499 | NVME_INLINE_SG_CNT * sizeof(struct scatterlist)); |
500 | if (ret) |
501 | goto out_destroy_queues; |
502 | |
503 | ret = nvme_loop_connect_io_queues(ctrl); |
504 | if (ret) |
505 | goto out_cleanup_tagset; |
506 | |
507 | return 0; |
508 | |
509 | out_cleanup_tagset: |
510 | nvme_remove_io_tag_set(ctrl: &ctrl->ctrl); |
511 | out_destroy_queues: |
512 | nvme_loop_destroy_io_queues(ctrl); |
513 | return ret; |
514 | } |
515 | |
516 | static struct nvmet_port *nvme_loop_find_port(struct nvme_ctrl *ctrl) |
517 | { |
518 | struct nvmet_port *p, *found = NULL; |
519 | |
520 | mutex_lock(&nvme_loop_ports_mutex); |
521 | list_for_each_entry(p, &nvme_loop_ports, entry) { |
522 | /* if no transport address is specified use the first port */ |
523 | if ((ctrl->opts->mask & NVMF_OPT_TRADDR) && |
524 | strcmp(ctrl->opts->traddr, p->disc_addr.traddr)) |
525 | continue; |
526 | found = p; |
527 | break; |
528 | } |
529 | mutex_unlock(lock: &nvme_loop_ports_mutex); |
530 | return found; |
531 | } |
532 | |
533 | static struct nvme_ctrl *nvme_loop_create_ctrl(struct device *dev, |
534 | struct nvmf_ctrl_options *opts) |
535 | { |
536 | struct nvme_loop_ctrl *ctrl; |
537 | int ret; |
538 | |
539 | ctrl = kzalloc(size: sizeof(*ctrl), GFP_KERNEL); |
540 | if (!ctrl) |
541 | return ERR_PTR(error: -ENOMEM); |
542 | ctrl->ctrl.opts = opts; |
543 | INIT_LIST_HEAD(list: &ctrl->list); |
544 | |
545 | INIT_WORK(&ctrl->ctrl.reset_work, nvme_loop_reset_ctrl_work); |
546 | |
547 | ret = nvme_init_ctrl(ctrl: &ctrl->ctrl, dev, ops: &nvme_loop_ctrl_ops, |
548 | quirks: 0 /* no quirks, we're perfect! */); |
549 | if (ret) { |
550 | kfree(objp: ctrl); |
551 | goto out; |
552 | } |
553 | |
554 | if (!nvme_change_ctrl_state(ctrl: &ctrl->ctrl, new_state: NVME_CTRL_CONNECTING)) |
555 | WARN_ON_ONCE(1); |
556 | |
557 | ret = -ENOMEM; |
558 | |
559 | ctrl->ctrl.kato = opts->kato; |
560 | ctrl->port = nvme_loop_find_port(ctrl: &ctrl->ctrl); |
561 | |
562 | ctrl->queues = kcalloc(n: opts->nr_io_queues + 1, size: sizeof(*ctrl->queues), |
563 | GFP_KERNEL); |
564 | if (!ctrl->queues) |
565 | goto out_uninit_ctrl; |
566 | |
567 | ret = nvme_loop_configure_admin_queue(ctrl); |
568 | if (ret) |
569 | goto out_free_queues; |
570 | |
571 | if (opts->queue_size > ctrl->ctrl.maxcmd) { |
572 | /* warn if maxcmd is lower than queue_size */ |
573 | dev_warn(ctrl->ctrl.device, |
574 | "queue_size %zu > ctrl maxcmd %u, clamping down\n" , |
575 | opts->queue_size, ctrl->ctrl.maxcmd); |
576 | opts->queue_size = ctrl->ctrl.maxcmd; |
577 | } |
578 | ctrl->ctrl.sqsize = opts->queue_size - 1; |
579 | |
580 | if (opts->nr_io_queues) { |
581 | ret = nvme_loop_create_io_queues(ctrl); |
582 | if (ret) |
583 | goto out_remove_admin_queue; |
584 | } |
585 | |
586 | nvme_loop_init_iod(ctrl, iod: &ctrl->async_event_iod, queue_idx: 0); |
587 | |
588 | dev_info(ctrl->ctrl.device, |
589 | "new ctrl: \"%s\"\n" , ctrl->ctrl.opts->subsysnqn); |
590 | |
591 | if (!nvme_change_ctrl_state(ctrl: &ctrl->ctrl, new_state: NVME_CTRL_LIVE)) |
592 | WARN_ON_ONCE(1); |
593 | |
594 | mutex_lock(&nvme_loop_ctrl_mutex); |
595 | list_add_tail(new: &ctrl->list, head: &nvme_loop_ctrl_list); |
596 | mutex_unlock(lock: &nvme_loop_ctrl_mutex); |
597 | |
598 | nvme_start_ctrl(ctrl: &ctrl->ctrl); |
599 | |
600 | return &ctrl->ctrl; |
601 | |
602 | out_remove_admin_queue: |
603 | nvme_loop_destroy_admin_queue(ctrl); |
604 | out_free_queues: |
605 | kfree(objp: ctrl->queues); |
606 | out_uninit_ctrl: |
607 | nvme_uninit_ctrl(ctrl: &ctrl->ctrl); |
608 | nvme_put_ctrl(ctrl: &ctrl->ctrl); |
609 | out: |
610 | if (ret > 0) |
611 | ret = -EIO; |
612 | return ERR_PTR(error: ret); |
613 | } |
614 | |
615 | static int nvme_loop_add_port(struct nvmet_port *port) |
616 | { |
617 | mutex_lock(&nvme_loop_ports_mutex); |
618 | list_add_tail(new: &port->entry, head: &nvme_loop_ports); |
619 | mutex_unlock(lock: &nvme_loop_ports_mutex); |
620 | return 0; |
621 | } |
622 | |
623 | static void nvme_loop_remove_port(struct nvmet_port *port) |
624 | { |
625 | mutex_lock(&nvme_loop_ports_mutex); |
626 | list_del_init(entry: &port->entry); |
627 | mutex_unlock(lock: &nvme_loop_ports_mutex); |
628 | |
629 | /* |
630 | * Ensure any ctrls that are in the process of being |
631 | * deleted are in fact deleted before we return |
632 | * and free the port. This is to prevent active |
633 | * ctrls from using a port after it's freed. |
634 | */ |
635 | flush_workqueue(nvme_delete_wq); |
636 | } |
637 | |
638 | static const struct nvmet_fabrics_ops nvme_loop_ops = { |
639 | .owner = THIS_MODULE, |
640 | .type = NVMF_TRTYPE_LOOP, |
641 | .add_port = nvme_loop_add_port, |
642 | .remove_port = nvme_loop_remove_port, |
643 | .queue_response = nvme_loop_queue_response, |
644 | .delete_ctrl = nvme_loop_delete_ctrl, |
645 | }; |
646 | |
647 | static struct nvmf_transport_ops nvme_loop_transport = { |
648 | .name = "loop" , |
649 | .module = THIS_MODULE, |
650 | .create_ctrl = nvme_loop_create_ctrl, |
651 | .allowed_opts = NVMF_OPT_TRADDR, |
652 | }; |
653 | |
654 | static int __init nvme_loop_init_module(void) |
655 | { |
656 | int ret; |
657 | |
658 | ret = nvmet_register_transport(ops: &nvme_loop_ops); |
659 | if (ret) |
660 | return ret; |
661 | |
662 | ret = nvmf_register_transport(ops: &nvme_loop_transport); |
663 | if (ret) |
664 | nvmet_unregister_transport(ops: &nvme_loop_ops); |
665 | |
666 | return ret; |
667 | } |
668 | |
669 | static void __exit nvme_loop_cleanup_module(void) |
670 | { |
671 | struct nvme_loop_ctrl *ctrl, *next; |
672 | |
673 | nvmf_unregister_transport(ops: &nvme_loop_transport); |
674 | nvmet_unregister_transport(ops: &nvme_loop_ops); |
675 | |
676 | mutex_lock(&nvme_loop_ctrl_mutex); |
677 | list_for_each_entry_safe(ctrl, next, &nvme_loop_ctrl_list, list) |
678 | nvme_delete_ctrl(ctrl: &ctrl->ctrl); |
679 | mutex_unlock(lock: &nvme_loop_ctrl_mutex); |
680 | |
681 | flush_workqueue(nvme_delete_wq); |
682 | } |
683 | |
684 | module_init(nvme_loop_init_module); |
685 | module_exit(nvme_loop_cleanup_module); |
686 | |
687 | MODULE_LICENSE("GPL v2" ); |
688 | MODULE_ALIAS("nvmet-transport-254" ); /* 254 == NVMF_TRTYPE_LOOP */ |
689 | |