]>
Commit | Line | Data |
---|---|---|
71102307 CH |
1 | /* |
2 | * NVMe over Fabrics RDMA host code. | |
3 | * Copyright (c) 2015-2016 HGST, a Western Digital Company. | |
4 | * | |
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. | |
8 | * | |
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 | |
12 | * more details. | |
13 | */ | |
14 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt | |
71102307 CH |
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> | |
71102307 CH |
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> | |
71102307 CH |
27 | #include <asm/unaligned.h> |
28 | ||
29 | #include <rdma/ib_verbs.h> | |
30 | #include <rdma/rdma_cm.h> | |
71102307 CH |
31 | #include <linux/nvme-rdma.h> |
32 | ||
33 | #include "nvme.h" | |
34 | #include "fabrics.h" | |
35 | ||
36 | ||
37 | #define NVME_RDMA_CONNECT_TIMEOUT_MS 1000 /* 1 second */ | |
38 | ||
39 | #define NVME_RDMA_MAX_SEGMENT_SIZE 0xffffff /* 24-bit SGL field */ | |
40 | ||
41 | #define NVME_RDMA_MAX_SEGMENTS 256 | |
42 | ||
43 | #define NVME_RDMA_MAX_INLINE_SEGMENTS 1 | |
44 | ||
71102307 CH |
45 | /* |
46 | * We handle AEN commands ourselves and don't even let the | |
47 | * block layer know about them. | |
48 | */ | |
49 | #define NVME_RDMA_NR_AEN_COMMANDS 1 | |
50 | #define NVME_RDMA_AQ_BLKMQ_DEPTH \ | |
51 | (NVMF_AQ_DEPTH - NVME_RDMA_NR_AEN_COMMANDS) | |
52 | ||
53 | struct nvme_rdma_device { | |
54 | struct ib_device *dev; | |
55 | struct ib_pd *pd; | |
71102307 CH |
56 | struct kref ref; |
57 | struct list_head entry; | |
58 | }; | |
59 | ||
60 | struct nvme_rdma_qe { | |
61 | struct ib_cqe cqe; | |
62 | void *data; | |
63 | u64 dma; | |
64 | }; | |
65 | ||
66 | struct nvme_rdma_queue; | |
67 | struct nvme_rdma_request { | |
d49187e9 | 68 | struct nvme_request req; |
71102307 CH |
69 | struct ib_mr *mr; |
70 | struct nvme_rdma_qe sqe; | |
71 | struct ib_sge sge[1 + NVME_RDMA_MAX_INLINE_SEGMENTS]; | |
72 | u32 num_sge; | |
73 | int nents; | |
74 | bool inline_data; | |
71102307 CH |
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[]; | |
80 | }; | |
81 | ||
82 | enum nvme_rdma_queue_flags { | |
83 | NVME_RDMA_Q_CONNECTED = (1 << 0), | |
f361e5a0 | 84 | NVME_RDMA_IB_QUEUE_ALLOCATED = (1 << 1), |
e89ca58f | 85 | NVME_RDMA_Q_DELETING = (1 << 2), |
71102307 CH |
86 | }; |
87 | ||
88 | struct nvme_rdma_queue { | |
89 | struct nvme_rdma_qe *rsp_ring; | |
90 | u8 sig_count; | |
91 | int queue_size; | |
92 | size_t cmnd_capsule_len; | |
93 | struct nvme_rdma_ctrl *ctrl; | |
94 | struct nvme_rdma_device *device; | |
95 | struct ib_cq *ib_cq; | |
96 | struct ib_qp *qp; | |
97 | ||
98 | unsigned long flags; | |
99 | struct rdma_cm_id *cm_id; | |
100 | int cm_error; | |
101 | struct completion cm_done; | |
102 | }; | |
103 | ||
104 | struct nvme_rdma_ctrl { | |
105 | /* read and written in the hot path */ | |
106 | spinlock_t lock; | |
107 | ||
108 | /* read only in the hot path */ | |
109 | struct nvme_rdma_queue *queues; | |
110 | u32 queue_count; | |
111 | ||
112 | /* other member variables */ | |
71102307 CH |
113 | struct blk_mq_tag_set tag_set; |
114 | struct work_struct delete_work; | |
115 | struct work_struct reset_work; | |
116 | struct work_struct err_work; | |
117 | ||
118 | struct nvme_rdma_qe async_event_sqe; | |
119 | ||
120 | int reconnect_delay; | |
121 | struct delayed_work reconnect_work; | |
122 | ||
123 | struct list_head list; | |
124 | ||
125 | struct blk_mq_tag_set admin_tag_set; | |
126 | struct nvme_rdma_device *device; | |
127 | ||
128 | u64 cap; | |
129 | u32 max_fr_pages; | |
130 | ||
131 | union { | |
132 | struct sockaddr addr; | |
133 | struct sockaddr_in addr_in; | |
134 | }; | |
135 | ||
136 | struct nvme_ctrl ctrl; | |
137 | }; | |
138 | ||
139 | static inline struct nvme_rdma_ctrl *to_rdma_ctrl(struct nvme_ctrl *ctrl) | |
140 | { | |
141 | return container_of(ctrl, struct nvme_rdma_ctrl, ctrl); | |
142 | } | |
143 | ||
144 | static LIST_HEAD(device_list); | |
145 | static DEFINE_MUTEX(device_list_mutex); | |
146 | ||
147 | static LIST_HEAD(nvme_rdma_ctrl_list); | |
148 | static DEFINE_MUTEX(nvme_rdma_ctrl_mutex); | |
149 | ||
150 | static struct workqueue_struct *nvme_rdma_wq; | |
151 | ||
152 | /* | |
153 | * Disabling this option makes small I/O goes faster, but is fundamentally | |
154 | * unsafe. With it turned off we will have to register a global rkey that | |
155 | * allows read and write access to all physical memory. | |
156 | */ | |
157 | static bool register_always = true; | |
158 | module_param(register_always, bool, 0444); | |
159 | MODULE_PARM_DESC(register_always, | |
160 | "Use memory registration even for contiguous memory regions"); | |
161 | ||
162 | static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id, | |
163 | struct rdma_cm_event *event); | |
164 | static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc); | |
71102307 CH |
165 | |
166 | /* XXX: really should move to a generic header sooner or later.. */ | |
167 | static inline void put_unaligned_le24(u32 val, u8 *p) | |
168 | { | |
169 | *p++ = val; | |
170 | *p++ = val >> 8; | |
171 | *p++ = val >> 16; | |
172 | } | |
173 | ||
174 | static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue *queue) | |
175 | { | |
176 | return queue - queue->ctrl->queues; | |
177 | } | |
178 | ||
179 | static inline size_t nvme_rdma_inline_data_size(struct nvme_rdma_queue *queue) | |
180 | { | |
181 | return queue->cmnd_capsule_len - sizeof(struct nvme_command); | |
182 | } | |
183 | ||
184 | static void nvme_rdma_free_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe, | |
185 | size_t capsule_size, enum dma_data_direction dir) | |
186 | { | |
187 | ib_dma_unmap_single(ibdev, qe->dma, capsule_size, dir); | |
188 | kfree(qe->data); | |
189 | } | |
190 | ||
191 | static int nvme_rdma_alloc_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe, | |
192 | size_t capsule_size, enum dma_data_direction dir) | |
193 | { | |
194 | qe->data = kzalloc(capsule_size, GFP_KERNEL); | |
195 | if (!qe->data) | |
196 | return -ENOMEM; | |
197 | ||
198 | qe->dma = ib_dma_map_single(ibdev, qe->data, capsule_size, dir); | |
199 | if (ib_dma_mapping_error(ibdev, qe->dma)) { | |
200 | kfree(qe->data); | |
201 | return -ENOMEM; | |
202 | } | |
203 | ||
204 | return 0; | |
205 | } | |
206 | ||
207 | static void nvme_rdma_free_ring(struct ib_device *ibdev, | |
208 | struct nvme_rdma_qe *ring, size_t ib_queue_size, | |
209 | size_t capsule_size, enum dma_data_direction dir) | |
210 | { | |
211 | int i; | |
212 | ||
213 | for (i = 0; i < ib_queue_size; i++) | |
214 | nvme_rdma_free_qe(ibdev, &ring[i], capsule_size, dir); | |
215 | kfree(ring); | |
216 | } | |
217 | ||
218 | static struct nvme_rdma_qe *nvme_rdma_alloc_ring(struct ib_device *ibdev, | |
219 | size_t ib_queue_size, size_t capsule_size, | |
220 | enum dma_data_direction dir) | |
221 | { | |
222 | struct nvme_rdma_qe *ring; | |
223 | int i; | |
224 | ||
225 | ring = kcalloc(ib_queue_size, sizeof(struct nvme_rdma_qe), GFP_KERNEL); | |
226 | if (!ring) | |
227 | return NULL; | |
228 | ||
229 | for (i = 0; i < ib_queue_size; i++) { | |
230 | if (nvme_rdma_alloc_qe(ibdev, &ring[i], capsule_size, dir)) | |
231 | goto out_free_ring; | |
232 | } | |
233 | ||
234 | return ring; | |
235 | ||
236 | out_free_ring: | |
237 | nvme_rdma_free_ring(ibdev, ring, i, capsule_size, dir); | |
238 | return NULL; | |
239 | } | |
240 | ||
241 | static void nvme_rdma_qp_event(struct ib_event *event, void *context) | |
242 | { | |
243 | pr_debug("QP event %d\n", event->event); | |
244 | } | |
245 | ||
246 | static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue *queue) | |
247 | { | |
248 | wait_for_completion_interruptible_timeout(&queue->cm_done, | |
249 | msecs_to_jiffies(NVME_RDMA_CONNECT_TIMEOUT_MS) + 1); | |
250 | return queue->cm_error; | |
251 | } | |
252 | ||
253 | static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor) | |
254 | { | |
255 | struct nvme_rdma_device *dev = queue->device; | |
256 | struct ib_qp_init_attr init_attr; | |
257 | int ret; | |
258 | ||
259 | memset(&init_attr, 0, sizeof(init_attr)); | |
260 | init_attr.event_handler = nvme_rdma_qp_event; | |
261 | /* +1 for drain */ | |
262 | init_attr.cap.max_send_wr = factor * queue->queue_size + 1; | |
263 | /* +1 for drain */ | |
264 | init_attr.cap.max_recv_wr = queue->queue_size + 1; | |
265 | init_attr.cap.max_recv_sge = 1; | |
266 | init_attr.cap.max_send_sge = 1 + NVME_RDMA_MAX_INLINE_SEGMENTS; | |
267 | init_attr.sq_sig_type = IB_SIGNAL_REQ_WR; | |
268 | init_attr.qp_type = IB_QPT_RC; | |
269 | init_attr.send_cq = queue->ib_cq; | |
270 | init_attr.recv_cq = queue->ib_cq; | |
271 | ||
272 | ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr); | |
273 | ||
274 | queue->qp = queue->cm_id->qp; | |
275 | return ret; | |
276 | } | |
277 | ||
278 | static int nvme_rdma_reinit_request(void *data, struct request *rq) | |
279 | { | |
280 | struct nvme_rdma_ctrl *ctrl = data; | |
281 | struct nvme_rdma_device *dev = ctrl->device; | |
282 | struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq); | |
283 | int ret = 0; | |
284 | ||
f5b7b559 | 285 | if (!req->mr->need_inval) |
71102307 CH |
286 | goto out; |
287 | ||
288 | ib_dereg_mr(req->mr); | |
289 | ||
290 | req->mr = ib_alloc_mr(dev->pd, IB_MR_TYPE_MEM_REG, | |
291 | ctrl->max_fr_pages); | |
292 | if (IS_ERR(req->mr)) { | |
71102307 | 293 | ret = PTR_ERR(req->mr); |
458a9632 | 294 | req->mr = NULL; |
1bda18de | 295 | goto out; |
71102307 CH |
296 | } |
297 | ||
f5b7b559 | 298 | req->mr->need_inval = false; |
71102307 CH |
299 | |
300 | out: | |
301 | return ret; | |
302 | } | |
303 | ||
304 | static void __nvme_rdma_exit_request(struct nvme_rdma_ctrl *ctrl, | |
305 | struct request *rq, unsigned int queue_idx) | |
306 | { | |
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; | |
310 | ||
311 | if (req->mr) | |
312 | ib_dereg_mr(req->mr); | |
313 | ||
314 | nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command), | |
315 | DMA_TO_DEVICE); | |
316 | } | |
317 | ||
318 | static void nvme_rdma_exit_request(void *data, struct request *rq, | |
319 | unsigned int hctx_idx, unsigned int rq_idx) | |
320 | { | |
321 | return __nvme_rdma_exit_request(data, rq, hctx_idx + 1); | |
322 | } | |
323 | ||
324 | static void nvme_rdma_exit_admin_request(void *data, struct request *rq, | |
325 | unsigned int hctx_idx, unsigned int rq_idx) | |
326 | { | |
327 | return __nvme_rdma_exit_request(data, rq, 0); | |
328 | } | |
329 | ||
330 | static int __nvme_rdma_init_request(struct nvme_rdma_ctrl *ctrl, | |
331 | struct request *rq, unsigned int queue_idx) | |
332 | { | |
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; | |
337 | int ret; | |
338 | ||
339 | BUG_ON(queue_idx >= ctrl->queue_count); | |
340 | ||
341 | ret = nvme_rdma_alloc_qe(ibdev, &req->sqe, sizeof(struct nvme_command), | |
342 | DMA_TO_DEVICE); | |
343 | if (ret) | |
344 | return ret; | |
345 | ||
346 | req->mr = ib_alloc_mr(dev->pd, IB_MR_TYPE_MEM_REG, | |
347 | ctrl->max_fr_pages); | |
348 | if (IS_ERR(req->mr)) { | |
349 | ret = PTR_ERR(req->mr); | |
350 | goto out_free_qe; | |
351 | } | |
352 | ||
353 | req->queue = queue; | |
354 | ||
355 | return 0; | |
356 | ||
357 | out_free_qe: | |
358 | nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command), | |
359 | DMA_TO_DEVICE); | |
360 | return -ENOMEM; | |
361 | } | |
362 | ||
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) | |
366 | { | |
367 | return __nvme_rdma_init_request(data, rq, hctx_idx + 1); | |
368 | } | |
369 | ||
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) | |
373 | { | |
374 | return __nvme_rdma_init_request(data, rq, 0); | |
375 | } | |
376 | ||
377 | static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data, | |
378 | unsigned int hctx_idx) | |
379 | { | |
380 | struct nvme_rdma_ctrl *ctrl = data; | |
381 | struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1]; | |
382 | ||
383 | BUG_ON(hctx_idx >= ctrl->queue_count); | |
384 | ||
385 | hctx->driver_data = queue; | |
386 | return 0; | |
387 | } | |
388 | ||
389 | static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data, | |
390 | unsigned int hctx_idx) | |
391 | { | |
392 | struct nvme_rdma_ctrl *ctrl = data; | |
393 | struct nvme_rdma_queue *queue = &ctrl->queues[0]; | |
394 | ||
395 | BUG_ON(hctx_idx != 0); | |
396 | ||
397 | hctx->driver_data = queue; | |
398 | return 0; | |
399 | } | |
400 | ||
401 | static void nvme_rdma_free_dev(struct kref *ref) | |
402 | { | |
403 | struct nvme_rdma_device *ndev = | |
404 | container_of(ref, struct nvme_rdma_device, ref); | |
405 | ||
406 | mutex_lock(&device_list_mutex); | |
407 | list_del(&ndev->entry); | |
408 | mutex_unlock(&device_list_mutex); | |
409 | ||
71102307 | 410 | ib_dealloc_pd(ndev->pd); |
71102307 CH |
411 | kfree(ndev); |
412 | } | |
413 | ||
414 | static void nvme_rdma_dev_put(struct nvme_rdma_device *dev) | |
415 | { | |
416 | kref_put(&dev->ref, nvme_rdma_free_dev); | |
417 | } | |
418 | ||
419 | static int nvme_rdma_dev_get(struct nvme_rdma_device *dev) | |
420 | { | |
421 | return kref_get_unless_zero(&dev->ref); | |
422 | } | |
423 | ||
424 | static struct nvme_rdma_device * | |
425 | nvme_rdma_find_get_device(struct rdma_cm_id *cm_id) | |
426 | { | |
427 | struct nvme_rdma_device *ndev; | |
428 | ||
429 | mutex_lock(&device_list_mutex); | |
430 | list_for_each_entry(ndev, &device_list, entry) { | |
431 | if (ndev->dev->node_guid == cm_id->device->node_guid && | |
432 | nvme_rdma_dev_get(ndev)) | |
433 | goto out_unlock; | |
434 | } | |
435 | ||
436 | ndev = kzalloc(sizeof(*ndev), GFP_KERNEL); | |
437 | if (!ndev) | |
438 | goto out_err; | |
439 | ||
440 | ndev->dev = cm_id->device; | |
441 | kref_init(&ndev->ref); | |
442 | ||
11975e01 CH |
443 | ndev->pd = ib_alloc_pd(ndev->dev, |
444 | register_always ? 0 : IB_PD_UNSAFE_GLOBAL_RKEY); | |
71102307 CH |
445 | if (IS_ERR(ndev->pd)) |
446 | goto out_free_dev; | |
447 | ||
71102307 CH |
448 | if (!(ndev->dev->attrs.device_cap_flags & |
449 | IB_DEVICE_MEM_MGT_EXTENSIONS)) { | |
450 | dev_err(&ndev->dev->dev, | |
451 | "Memory registrations not supported.\n"); | |
11975e01 | 452 | goto out_free_pd; |
71102307 CH |
453 | } |
454 | ||
455 | list_add(&ndev->entry, &device_list); | |
456 | out_unlock: | |
457 | mutex_unlock(&device_list_mutex); | |
458 | return ndev; | |
459 | ||
71102307 CH |
460 | out_free_pd: |
461 | ib_dealloc_pd(ndev->pd); | |
462 | out_free_dev: | |
463 | kfree(ndev); | |
464 | out_err: | |
465 | mutex_unlock(&device_list_mutex); | |
466 | return NULL; | |
467 | } | |
468 | ||
469 | static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue) | |
470 | { | |
f361e5a0 SW |
471 | struct nvme_rdma_device *dev; |
472 | struct ib_device *ibdev; | |
71102307 | 473 | |
f361e5a0 SW |
474 | if (!test_and_clear_bit(NVME_RDMA_IB_QUEUE_ALLOCATED, &queue->flags)) |
475 | return; | |
476 | ||
477 | dev = queue->device; | |
478 | ibdev = dev->dev; | |
71102307 CH |
479 | rdma_destroy_qp(queue->cm_id); |
480 | ib_free_cq(queue->ib_cq); | |
481 | ||
482 | nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size, | |
483 | sizeof(struct nvme_completion), DMA_FROM_DEVICE); | |
484 | ||
485 | nvme_rdma_dev_put(dev); | |
486 | } | |
487 | ||
488 | static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue, | |
489 | struct nvme_rdma_device *dev) | |
490 | { | |
491 | struct ib_device *ibdev = dev->dev; | |
492 | const int send_wr_factor = 3; /* MR, SEND, INV */ | |
493 | const int cq_factor = send_wr_factor + 1; /* + RECV */ | |
494 | int comp_vector, idx = nvme_rdma_queue_idx(queue); | |
495 | ||
496 | int ret; | |
497 | ||
498 | queue->device = dev; | |
499 | ||
500 | /* | |
501 | * The admin queue is barely used once the controller is live, so don't | |
502 | * bother to spread it out. | |
503 | */ | |
504 | if (idx == 0) | |
505 | comp_vector = 0; | |
506 | else | |
507 | comp_vector = idx % ibdev->num_comp_vectors; | |
508 | ||
509 | ||
510 | /* +1 for ib_stop_cq */ | |
511 | queue->ib_cq = ib_alloc_cq(dev->dev, queue, | |
512 | cq_factor * queue->queue_size + 1, comp_vector, | |
513 | IB_POLL_SOFTIRQ); | |
514 | if (IS_ERR(queue->ib_cq)) { | |
515 | ret = PTR_ERR(queue->ib_cq); | |
516 | goto out; | |
517 | } | |
518 | ||
519 | ret = nvme_rdma_create_qp(queue, send_wr_factor); | |
520 | if (ret) | |
521 | goto out_destroy_ib_cq; | |
522 | ||
523 | queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size, | |
524 | sizeof(struct nvme_completion), DMA_FROM_DEVICE); | |
525 | if (!queue->rsp_ring) { | |
526 | ret = -ENOMEM; | |
527 | goto out_destroy_qp; | |
528 | } | |
f361e5a0 | 529 | set_bit(NVME_RDMA_IB_QUEUE_ALLOCATED, &queue->flags); |
71102307 CH |
530 | |
531 | return 0; | |
532 | ||
533 | out_destroy_qp: | |
534 | ib_destroy_qp(queue->qp); | |
535 | out_destroy_ib_cq: | |
536 | ib_free_cq(queue->ib_cq); | |
537 | out: | |
538 | return ret; | |
539 | } | |
540 | ||
541 | static int nvme_rdma_init_queue(struct nvme_rdma_ctrl *ctrl, | |
542 | int idx, size_t queue_size) | |
543 | { | |
544 | struct nvme_rdma_queue *queue; | |
545 | int ret; | |
546 | ||
547 | queue = &ctrl->queues[idx]; | |
548 | queue->ctrl = ctrl; | |
549 | init_completion(&queue->cm_done); | |
550 | ||
551 | if (idx > 0) | |
552 | queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16; | |
553 | else | |
554 | queue->cmnd_capsule_len = sizeof(struct nvme_command); | |
555 | ||
556 | queue->queue_size = queue_size; | |
557 | ||
558 | queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue, | |
559 | RDMA_PS_TCP, IB_QPT_RC); | |
560 | if (IS_ERR(queue->cm_id)) { | |
561 | dev_info(ctrl->ctrl.device, | |
562 | "failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id)); | |
563 | return PTR_ERR(queue->cm_id); | |
564 | } | |
565 | ||
566 | queue->cm_error = -ETIMEDOUT; | |
567 | ret = rdma_resolve_addr(queue->cm_id, NULL, &ctrl->addr, | |
568 | NVME_RDMA_CONNECT_TIMEOUT_MS); | |
569 | if (ret) { | |
570 | dev_info(ctrl->ctrl.device, | |
571 | "rdma_resolve_addr failed (%d).\n", ret); | |
572 | goto out_destroy_cm_id; | |
573 | } | |
574 | ||
575 | ret = nvme_rdma_wait_for_cm(queue); | |
576 | if (ret) { | |
577 | dev_info(ctrl->ctrl.device, | |
578 | "rdma_resolve_addr wait failed (%d).\n", ret); | |
579 | goto out_destroy_cm_id; | |
580 | } | |
581 | ||
3b4ac786 | 582 | clear_bit(NVME_RDMA_Q_DELETING, &queue->flags); |
71102307 CH |
583 | set_bit(NVME_RDMA_Q_CONNECTED, &queue->flags); |
584 | ||
585 | return 0; | |
586 | ||
587 | out_destroy_cm_id: | |
f361e5a0 | 588 | nvme_rdma_destroy_queue_ib(queue); |
71102307 CH |
589 | rdma_destroy_id(queue->cm_id); |
590 | return ret; | |
591 | } | |
592 | ||
593 | static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue) | |
594 | { | |
595 | rdma_disconnect(queue->cm_id); | |
596 | ib_drain_qp(queue->qp); | |
597 | } | |
598 | ||
599 | static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue) | |
600 | { | |
601 | nvme_rdma_destroy_queue_ib(queue); | |
602 | rdma_destroy_id(queue->cm_id); | |
603 | } | |
604 | ||
605 | static void nvme_rdma_stop_and_free_queue(struct nvme_rdma_queue *queue) | |
606 | { | |
e89ca58f | 607 | if (test_and_set_bit(NVME_RDMA_Q_DELETING, &queue->flags)) |
71102307 CH |
608 | return; |
609 | nvme_rdma_stop_queue(queue); | |
610 | nvme_rdma_free_queue(queue); | |
611 | } | |
612 | ||
613 | static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl) | |
614 | { | |
615 | int i; | |
616 | ||
617 | for (i = 1; i < ctrl->queue_count; i++) | |
618 | nvme_rdma_stop_and_free_queue(&ctrl->queues[i]); | |
619 | } | |
620 | ||
621 | static int nvme_rdma_connect_io_queues(struct nvme_rdma_ctrl *ctrl) | |
622 | { | |
623 | int i, ret = 0; | |
624 | ||
625 | for (i = 1; i < ctrl->queue_count; i++) { | |
626 | ret = nvmf_connect_io_queue(&ctrl->ctrl, i); | |
627 | if (ret) | |
628 | break; | |
629 | } | |
630 | ||
631 | return ret; | |
632 | } | |
633 | ||
634 | static int nvme_rdma_init_io_queues(struct nvme_rdma_ctrl *ctrl) | |
635 | { | |
636 | int i, ret; | |
637 | ||
638 | for (i = 1; i < ctrl->queue_count; i++) { | |
c5af8654 JF |
639 | ret = nvme_rdma_init_queue(ctrl, i, |
640 | ctrl->ctrl.opts->queue_size); | |
71102307 CH |
641 | if (ret) { |
642 | dev_info(ctrl->ctrl.device, | |
643 | "failed to initialize i/o queue: %d\n", ret); | |
644 | goto out_free_queues; | |
645 | } | |
646 | } | |
647 | ||
648 | return 0; | |
649 | ||
650 | out_free_queues: | |
f361e5a0 | 651 | for (i--; i >= 1; i--) |
71102307 CH |
652 | nvme_rdma_stop_and_free_queue(&ctrl->queues[i]); |
653 | ||
654 | return ret; | |
655 | } | |
656 | ||
657 | static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl) | |
658 | { | |
659 | nvme_rdma_free_qe(ctrl->queues[0].device->dev, &ctrl->async_event_sqe, | |
660 | sizeof(struct nvme_command), DMA_TO_DEVICE); | |
661 | nvme_rdma_stop_and_free_queue(&ctrl->queues[0]); | |
662 | blk_cleanup_queue(ctrl->ctrl.admin_q); | |
663 | blk_mq_free_tag_set(&ctrl->admin_tag_set); | |
664 | nvme_rdma_dev_put(ctrl->device); | |
665 | } | |
666 | ||
667 | static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl) | |
668 | { | |
669 | struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl); | |
670 | ||
671 | if (list_empty(&ctrl->list)) | |
672 | goto free_ctrl; | |
673 | ||
674 | mutex_lock(&nvme_rdma_ctrl_mutex); | |
675 | list_del(&ctrl->list); | |
676 | mutex_unlock(&nvme_rdma_ctrl_mutex); | |
677 | ||
71102307 CH |
678 | kfree(ctrl->queues); |
679 | nvmf_free_options(nctrl->opts); | |
680 | free_ctrl: | |
681 | kfree(ctrl); | |
682 | } | |
683 | ||
684 | static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work) | |
685 | { | |
686 | struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work), | |
687 | struct nvme_rdma_ctrl, reconnect_work); | |
688 | bool changed; | |
689 | int ret; | |
690 | ||
691 | if (ctrl->queue_count > 1) { | |
692 | nvme_rdma_free_io_queues(ctrl); | |
693 | ||
694 | ret = blk_mq_reinit_tagset(&ctrl->tag_set); | |
695 | if (ret) | |
696 | goto requeue; | |
697 | } | |
698 | ||
699 | nvme_rdma_stop_and_free_queue(&ctrl->queues[0]); | |
700 | ||
701 | ret = blk_mq_reinit_tagset(&ctrl->admin_tag_set); | |
702 | if (ret) | |
703 | goto requeue; | |
704 | ||
705 | ret = nvme_rdma_init_queue(ctrl, 0, NVMF_AQ_DEPTH); | |
706 | if (ret) | |
707 | goto requeue; | |
708 | ||
709 | blk_mq_start_stopped_hw_queues(ctrl->ctrl.admin_q, true); | |
710 | ||
711 | ret = nvmf_connect_admin_queue(&ctrl->ctrl); | |
712 | if (ret) | |
713 | goto stop_admin_q; | |
714 | ||
715 | ret = nvme_enable_ctrl(&ctrl->ctrl, ctrl->cap); | |
716 | if (ret) | |
717 | goto stop_admin_q; | |
718 | ||
719 | nvme_start_keep_alive(&ctrl->ctrl); | |
720 | ||
721 | if (ctrl->queue_count > 1) { | |
722 | ret = nvme_rdma_init_io_queues(ctrl); | |
723 | if (ret) | |
724 | goto stop_admin_q; | |
725 | ||
726 | ret = nvme_rdma_connect_io_queues(ctrl); | |
727 | if (ret) | |
728 | goto stop_admin_q; | |
729 | } | |
730 | ||
731 | changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE); | |
732 | WARN_ON_ONCE(!changed); | |
733 | ||
5f372eb3 | 734 | if (ctrl->queue_count > 1) { |
71102307 | 735 | nvme_start_queues(&ctrl->ctrl); |
5f372eb3 | 736 | nvme_queue_scan(&ctrl->ctrl); |
3ef1b4b2 | 737 | nvme_queue_async_events(&ctrl->ctrl); |
5f372eb3 | 738 | } |
71102307 CH |
739 | |
740 | dev_info(ctrl->ctrl.device, "Successfully reconnected\n"); | |
741 | ||
742 | return; | |
743 | ||
744 | stop_admin_q: | |
745 | blk_mq_stop_hw_queues(ctrl->ctrl.admin_q); | |
746 | requeue: | |
747 | /* Make sure we are not resetting/deleting */ | |
748 | if (ctrl->ctrl.state == NVME_CTRL_RECONNECTING) { | |
749 | dev_info(ctrl->ctrl.device, | |
750 | "Failed reconnect attempt, requeueing...\n"); | |
751 | queue_delayed_work(nvme_rdma_wq, &ctrl->reconnect_work, | |
752 | ctrl->reconnect_delay * HZ); | |
753 | } | |
754 | } | |
755 | ||
756 | static void nvme_rdma_error_recovery_work(struct work_struct *work) | |
757 | { | |
758 | struct nvme_rdma_ctrl *ctrl = container_of(work, | |
759 | struct nvme_rdma_ctrl, err_work); | |
e89ca58f | 760 | int i; |
71102307 CH |
761 | |
762 | nvme_stop_keep_alive(&ctrl->ctrl); | |
e89ca58f SG |
763 | |
764 | for (i = 0; i < ctrl->queue_count; i++) | |
765 | clear_bit(NVME_RDMA_Q_CONNECTED, &ctrl->queues[i].flags); | |
766 | ||
71102307 CH |
767 | if (ctrl->queue_count > 1) |
768 | nvme_stop_queues(&ctrl->ctrl); | |
769 | blk_mq_stop_hw_queues(ctrl->ctrl.admin_q); | |
770 | ||
771 | /* We must take care of fastfail/requeue all our inflight requests */ | |
772 | if (ctrl->queue_count > 1) | |
773 | blk_mq_tagset_busy_iter(&ctrl->tag_set, | |
774 | nvme_cancel_request, &ctrl->ctrl); | |
775 | blk_mq_tagset_busy_iter(&ctrl->admin_tag_set, | |
776 | nvme_cancel_request, &ctrl->ctrl); | |
777 | ||
778 | dev_info(ctrl->ctrl.device, "reconnecting in %d seconds\n", | |
779 | ctrl->reconnect_delay); | |
780 | ||
781 | queue_delayed_work(nvme_rdma_wq, &ctrl->reconnect_work, | |
782 | ctrl->reconnect_delay * HZ); | |
783 | } | |
784 | ||
785 | static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl) | |
786 | { | |
787 | if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RECONNECTING)) | |
788 | return; | |
789 | ||
790 | queue_work(nvme_rdma_wq, &ctrl->err_work); | |
791 | } | |
792 | ||
793 | static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc, | |
794 | const char *op) | |
795 | { | |
796 | struct nvme_rdma_queue *queue = cq->cq_context; | |
797 | struct nvme_rdma_ctrl *ctrl = queue->ctrl; | |
798 | ||
799 | if (ctrl->ctrl.state == NVME_CTRL_LIVE) | |
800 | dev_info(ctrl->ctrl.device, | |
801 | "%s for CQE 0x%p failed with status %s (%d)\n", | |
802 | op, wc->wr_cqe, | |
803 | ib_wc_status_msg(wc->status), wc->status); | |
804 | nvme_rdma_error_recovery(ctrl); | |
805 | } | |
806 | ||
807 | static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc) | |
808 | { | |
809 | if (unlikely(wc->status != IB_WC_SUCCESS)) | |
810 | nvme_rdma_wr_error(cq, wc, "MEMREG"); | |
811 | } | |
812 | ||
813 | static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc) | |
814 | { | |
815 | if (unlikely(wc->status != IB_WC_SUCCESS)) | |
816 | nvme_rdma_wr_error(cq, wc, "LOCAL_INV"); | |
817 | } | |
818 | ||
819 | static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue, | |
820 | struct nvme_rdma_request *req) | |
821 | { | |
822 | struct ib_send_wr *bad_wr; | |
823 | struct ib_send_wr wr = { | |
824 | .opcode = IB_WR_LOCAL_INV, | |
825 | .next = NULL, | |
826 | .num_sge = 0, | |
827 | .send_flags = 0, | |
828 | .ex.invalidate_rkey = req->mr->rkey, | |
829 | }; | |
830 | ||
831 | req->reg_cqe.done = nvme_rdma_inv_rkey_done; | |
832 | wr.wr_cqe = &req->reg_cqe; | |
833 | ||
834 | return ib_post_send(queue->qp, &wr, &bad_wr); | |
835 | } | |
836 | ||
837 | static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue, | |
838 | struct request *rq) | |
839 | { | |
840 | struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq); | |
841 | struct nvme_rdma_ctrl *ctrl = queue->ctrl; | |
842 | struct nvme_rdma_device *dev = queue->device; | |
843 | struct ib_device *ibdev = dev->dev; | |
844 | int res; | |
845 | ||
846 | if (!blk_rq_bytes(rq)) | |
847 | return; | |
848 | ||
f5b7b559 | 849 | if (req->mr->need_inval) { |
71102307 CH |
850 | res = nvme_rdma_inv_rkey(queue, req); |
851 | if (res < 0) { | |
852 | dev_err(ctrl->ctrl.device, | |
853 | "Queueing INV WR for rkey %#x failed (%d)\n", | |
854 | req->mr->rkey, res); | |
855 | nvme_rdma_error_recovery(queue->ctrl); | |
856 | } | |
857 | } | |
858 | ||
859 | ib_dma_unmap_sg(ibdev, req->sg_table.sgl, | |
860 | req->nents, rq_data_dir(rq) == | |
861 | WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE); | |
862 | ||
863 | nvme_cleanup_cmd(rq); | |
864 | sg_free_table_chained(&req->sg_table, true); | |
865 | } | |
866 | ||
867 | static int nvme_rdma_set_sg_null(struct nvme_command *c) | |
868 | { | |
869 | struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl; | |
870 | ||
871 | sg->addr = 0; | |
872 | put_unaligned_le24(0, sg->length); | |
873 | put_unaligned_le32(0, sg->key); | |
874 | sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4; | |
875 | return 0; | |
876 | } | |
877 | ||
878 | static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue, | |
879 | struct nvme_rdma_request *req, struct nvme_command *c) | |
880 | { | |
881 | struct nvme_sgl_desc *sg = &c->common.dptr.sgl; | |
882 | ||
883 | req->sge[1].addr = sg_dma_address(req->sg_table.sgl); | |
884 | req->sge[1].length = sg_dma_len(req->sg_table.sgl); | |
885 | req->sge[1].lkey = queue->device->pd->local_dma_lkey; | |
886 | ||
887 | sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff); | |
888 | sg->length = cpu_to_le32(sg_dma_len(req->sg_table.sgl)); | |
889 | sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET; | |
890 | ||
891 | req->inline_data = true; | |
892 | req->num_sge++; | |
893 | return 0; | |
894 | } | |
895 | ||
896 | static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue, | |
897 | struct nvme_rdma_request *req, struct nvme_command *c) | |
898 | { | |
899 | struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl; | |
900 | ||
901 | sg->addr = cpu_to_le64(sg_dma_address(req->sg_table.sgl)); | |
902 | put_unaligned_le24(sg_dma_len(req->sg_table.sgl), sg->length); | |
11975e01 | 903 | put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key); |
71102307 CH |
904 | sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4; |
905 | return 0; | |
906 | } | |
907 | ||
908 | static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue, | |
909 | struct nvme_rdma_request *req, struct nvme_command *c, | |
910 | int count) | |
911 | { | |
912 | struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl; | |
913 | int nr; | |
914 | ||
915 | nr = ib_map_mr_sg(req->mr, req->sg_table.sgl, count, NULL, PAGE_SIZE); | |
916 | if (nr < count) { | |
917 | if (nr < 0) | |
918 | return nr; | |
919 | return -EINVAL; | |
920 | } | |
921 | ||
922 | ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey)); | |
923 | ||
924 | req->reg_cqe.done = nvme_rdma_memreg_done; | |
925 | memset(&req->reg_wr, 0, sizeof(req->reg_wr)); | |
926 | req->reg_wr.wr.opcode = IB_WR_REG_MR; | |
927 | req->reg_wr.wr.wr_cqe = &req->reg_cqe; | |
928 | req->reg_wr.wr.num_sge = 0; | |
929 | req->reg_wr.mr = req->mr; | |
930 | req->reg_wr.key = req->mr->rkey; | |
931 | req->reg_wr.access = IB_ACCESS_LOCAL_WRITE | | |
932 | IB_ACCESS_REMOTE_READ | | |
933 | IB_ACCESS_REMOTE_WRITE; | |
934 | ||
f5b7b559 | 935 | req->mr->need_inval = true; |
71102307 CH |
936 | |
937 | sg->addr = cpu_to_le64(req->mr->iova); | |
938 | put_unaligned_le24(req->mr->length, sg->length); | |
939 | put_unaligned_le32(req->mr->rkey, sg->key); | |
940 | sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) | | |
941 | NVME_SGL_FMT_INVALIDATE; | |
942 | ||
943 | return 0; | |
944 | } | |
945 | ||
946 | static int nvme_rdma_map_data(struct nvme_rdma_queue *queue, | |
947 | struct request *rq, unsigned int map_len, | |
948 | struct nvme_command *c) | |
949 | { | |
950 | struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq); | |
951 | struct nvme_rdma_device *dev = queue->device; | |
952 | struct ib_device *ibdev = dev->dev; | |
953 | int nents, count; | |
954 | int ret; | |
955 | ||
956 | req->num_sge = 1; | |
957 | req->inline_data = false; | |
f5b7b559 | 958 | req->mr->need_inval = false; |
71102307 CH |
959 | |
960 | c->common.flags |= NVME_CMD_SGL_METABUF; | |
961 | ||
962 | if (!blk_rq_bytes(rq)) | |
963 | return nvme_rdma_set_sg_null(c); | |
964 | ||
965 | req->sg_table.sgl = req->first_sgl; | |
966 | ret = sg_alloc_table_chained(&req->sg_table, rq->nr_phys_segments, | |
967 | req->sg_table.sgl); | |
968 | if (ret) | |
969 | return -ENOMEM; | |
970 | ||
971 | nents = blk_rq_map_sg(rq->q, rq, req->sg_table.sgl); | |
972 | BUG_ON(nents > rq->nr_phys_segments); | |
973 | req->nents = nents; | |
974 | ||
975 | count = ib_dma_map_sg(ibdev, req->sg_table.sgl, nents, | |
976 | rq_data_dir(rq) == WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE); | |
977 | if (unlikely(count <= 0)) { | |
978 | sg_free_table_chained(&req->sg_table, true); | |
979 | return -EIO; | |
980 | } | |
981 | ||
982 | if (count == 1) { | |
983 | if (rq_data_dir(rq) == WRITE && | |
984 | map_len <= nvme_rdma_inline_data_size(queue) && | |
985 | nvme_rdma_queue_idx(queue)) | |
986 | return nvme_rdma_map_sg_inline(queue, req, c); | |
987 | ||
11975e01 | 988 | if (dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY) |
71102307 CH |
989 | return nvme_rdma_map_sg_single(queue, req, c); |
990 | } | |
991 | ||
992 | return nvme_rdma_map_sg_fr(queue, req, c, count); | |
993 | } | |
994 | ||
995 | static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc) | |
996 | { | |
997 | if (unlikely(wc->status != IB_WC_SUCCESS)) | |
998 | nvme_rdma_wr_error(cq, wc, "SEND"); | |
999 | } | |
1000 | ||
1001 | static int nvme_rdma_post_send(struct nvme_rdma_queue *queue, | |
1002 | struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge, | |
1003 | struct ib_send_wr *first, bool flush) | |
1004 | { | |
1005 | struct ib_send_wr wr, *bad_wr; | |
1006 | int ret; | |
1007 | ||
1008 | sge->addr = qe->dma; | |
1009 | sge->length = sizeof(struct nvme_command), | |
1010 | sge->lkey = queue->device->pd->local_dma_lkey; | |
1011 | ||
1012 | qe->cqe.done = nvme_rdma_send_done; | |
1013 | ||
1014 | wr.next = NULL; | |
1015 | wr.wr_cqe = &qe->cqe; | |
1016 | wr.sg_list = sge; | |
1017 | wr.num_sge = num_sge; | |
1018 | wr.opcode = IB_WR_SEND; | |
1019 | wr.send_flags = 0; | |
1020 | ||
1021 | /* | |
1022 | * Unsignalled send completions are another giant desaster in the | |
1023 | * IB Verbs spec: If we don't regularly post signalled sends | |
1024 | * the send queue will fill up and only a QP reset will rescue us. | |
1025 | * Would have been way to obvious to handle this in hardware or | |
1026 | * at least the RDMA stack.. | |
1027 | * | |
1028 | * This messy and racy code sniplet is copy and pasted from the iSER | |
1029 | * initiator, and the magic '32' comes from there as well. | |
1030 | * | |
1031 | * Always signal the flushes. The magic request used for the flush | |
1032 | * sequencer is not allocated in our driver's tagset and it's | |
1033 | * triggered to be freed by blk_cleanup_queue(). So we need to | |
1034 | * always mark it as signaled to ensure that the "wr_cqe", which is | |
1035 | * embeded in request's payload, is not freed when __ib_process_cq() | |
1036 | * calls wr_cqe->done(). | |
1037 | */ | |
1038 | if ((++queue->sig_count % 32) == 0 || flush) | |
1039 | wr.send_flags |= IB_SEND_SIGNALED; | |
1040 | ||
1041 | if (first) | |
1042 | first->next = ≀ | |
1043 | else | |
1044 | first = ≀ | |
1045 | ||
1046 | ret = ib_post_send(queue->qp, first, &bad_wr); | |
1047 | if (ret) { | |
1048 | dev_err(queue->ctrl->ctrl.device, | |
1049 | "%s failed with error code %d\n", __func__, ret); | |
1050 | } | |
1051 | return ret; | |
1052 | } | |
1053 | ||
1054 | static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue, | |
1055 | struct nvme_rdma_qe *qe) | |
1056 | { | |
1057 | struct ib_recv_wr wr, *bad_wr; | |
1058 | struct ib_sge list; | |
1059 | int ret; | |
1060 | ||
1061 | list.addr = qe->dma; | |
1062 | list.length = sizeof(struct nvme_completion); | |
1063 | list.lkey = queue->device->pd->local_dma_lkey; | |
1064 | ||
1065 | qe->cqe.done = nvme_rdma_recv_done; | |
1066 | ||
1067 | wr.next = NULL; | |
1068 | wr.wr_cqe = &qe->cqe; | |
1069 | wr.sg_list = &list; | |
1070 | wr.num_sge = 1; | |
1071 | ||
1072 | ret = ib_post_recv(queue->qp, &wr, &bad_wr); | |
1073 | if (ret) { | |
1074 | dev_err(queue->ctrl->ctrl.device, | |
1075 | "%s failed with error code %d\n", __func__, ret); | |
1076 | } | |
1077 | return ret; | |
1078 | } | |
1079 | ||
1080 | static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue) | |
1081 | { | |
1082 | u32 queue_idx = nvme_rdma_queue_idx(queue); | |
1083 | ||
1084 | if (queue_idx == 0) | |
1085 | return queue->ctrl->admin_tag_set.tags[queue_idx]; | |
1086 | return queue->ctrl->tag_set.tags[queue_idx - 1]; | |
1087 | } | |
1088 | ||
1089 | static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg, int aer_idx) | |
1090 | { | |
1091 | struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg); | |
1092 | struct nvme_rdma_queue *queue = &ctrl->queues[0]; | |
1093 | struct ib_device *dev = queue->device->dev; | |
1094 | struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe; | |
1095 | struct nvme_command *cmd = sqe->data; | |
1096 | struct ib_sge sge; | |
1097 | int ret; | |
1098 | ||
1099 | if (WARN_ON_ONCE(aer_idx != 0)) | |
1100 | return; | |
1101 | ||
1102 | ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE); | |
1103 | ||
1104 | memset(cmd, 0, sizeof(*cmd)); | |
1105 | cmd->common.opcode = nvme_admin_async_event; | |
1106 | cmd->common.command_id = NVME_RDMA_AQ_BLKMQ_DEPTH; | |
1107 | cmd->common.flags |= NVME_CMD_SGL_METABUF; | |
1108 | nvme_rdma_set_sg_null(cmd); | |
1109 | ||
1110 | ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd), | |
1111 | DMA_TO_DEVICE); | |
1112 | ||
1113 | ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL, false); | |
1114 | WARN_ON_ONCE(ret); | |
1115 | } | |
1116 | ||
1117 | static int nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue, | |
1118 | struct nvme_completion *cqe, struct ib_wc *wc, int tag) | |
1119 | { | |
71102307 CH |
1120 | struct request *rq; |
1121 | struct nvme_rdma_request *req; | |
1122 | int ret = 0; | |
1123 | ||
71102307 CH |
1124 | rq = blk_mq_tag_to_rq(nvme_rdma_tagset(queue), cqe->command_id); |
1125 | if (!rq) { | |
1126 | dev_err(queue->ctrl->ctrl.device, | |
1127 | "tag 0x%x on QP %#x not found\n", | |
1128 | cqe->command_id, queue->qp->qp_num); | |
1129 | nvme_rdma_error_recovery(queue->ctrl); | |
1130 | return ret; | |
1131 | } | |
1132 | req = blk_mq_rq_to_pdu(rq); | |
1133 | ||
71102307 CH |
1134 | if (rq->tag == tag) |
1135 | ret = 1; | |
1136 | ||
1137 | if ((wc->wc_flags & IB_WC_WITH_INVALIDATE) && | |
1138 | wc->ex.invalidate_rkey == req->mr->rkey) | |
f5b7b559 | 1139 | req->mr->need_inval = false; |
71102307 | 1140 | |
d49187e9 CH |
1141 | req->req.result = cqe->result; |
1142 | blk_mq_complete_request(rq, le16_to_cpu(cqe->status) >> 1); | |
71102307 CH |
1143 | return ret; |
1144 | } | |
1145 | ||
1146 | static int __nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc, int tag) | |
1147 | { | |
1148 | struct nvme_rdma_qe *qe = | |
1149 | container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe); | |
1150 | struct nvme_rdma_queue *queue = cq->cq_context; | |
1151 | struct ib_device *ibdev = queue->device->dev; | |
1152 | struct nvme_completion *cqe = qe->data; | |
1153 | const size_t len = sizeof(struct nvme_completion); | |
1154 | int ret = 0; | |
1155 | ||
1156 | if (unlikely(wc->status != IB_WC_SUCCESS)) { | |
1157 | nvme_rdma_wr_error(cq, wc, "RECV"); | |
1158 | return 0; | |
1159 | } | |
1160 | ||
1161 | ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE); | |
1162 | /* | |
1163 | * AEN requests are special as they don't time out and can | |
1164 | * survive any kind of queue freeze and often don't respond to | |
1165 | * aborts. We don't even bother to allocate a struct request | |
1166 | * for them but rather special case them here. | |
1167 | */ | |
1168 | if (unlikely(nvme_rdma_queue_idx(queue) == 0 && | |
1169 | cqe->command_id >= NVME_RDMA_AQ_BLKMQ_DEPTH)) | |
7bf58533 CH |
1170 | nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status, |
1171 | &cqe->result); | |
71102307 CH |
1172 | else |
1173 | ret = nvme_rdma_process_nvme_rsp(queue, cqe, wc, tag); | |
1174 | ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE); | |
1175 | ||
1176 | nvme_rdma_post_recv(queue, qe); | |
1177 | return ret; | |
1178 | } | |
1179 | ||
1180 | static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc) | |
1181 | { | |
1182 | __nvme_rdma_recv_done(cq, wc, -1); | |
1183 | } | |
1184 | ||
1185 | static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue) | |
1186 | { | |
1187 | int ret, i; | |
1188 | ||
1189 | for (i = 0; i < queue->queue_size; i++) { | |
1190 | ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]); | |
1191 | if (ret) | |
1192 | goto out_destroy_queue_ib; | |
1193 | } | |
1194 | ||
1195 | return 0; | |
1196 | ||
1197 | out_destroy_queue_ib: | |
1198 | nvme_rdma_destroy_queue_ib(queue); | |
1199 | return ret; | |
1200 | } | |
1201 | ||
1202 | static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue, | |
1203 | struct rdma_cm_event *ev) | |
1204 | { | |
1205 | if (ev->param.conn.private_data_len) { | |
1206 | struct nvme_rdma_cm_rej *rej = | |
1207 | (struct nvme_rdma_cm_rej *)ev->param.conn.private_data; | |
1208 | ||
1209 | dev_err(queue->ctrl->ctrl.device, | |
1210 | "Connect rejected, status %d.", le16_to_cpu(rej->sts)); | |
1211 | /* XXX: Think of something clever to do here... */ | |
1212 | } else { | |
1213 | dev_err(queue->ctrl->ctrl.device, | |
1214 | "Connect rejected, no private data.\n"); | |
1215 | } | |
1216 | ||
1217 | return -ECONNRESET; | |
1218 | } | |
1219 | ||
1220 | static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue) | |
1221 | { | |
1222 | struct nvme_rdma_device *dev; | |
1223 | int ret; | |
1224 | ||
1225 | dev = nvme_rdma_find_get_device(queue->cm_id); | |
1226 | if (!dev) { | |
1227 | dev_err(queue->cm_id->device->dma_device, | |
1228 | "no client data found!\n"); | |
1229 | return -ECONNREFUSED; | |
1230 | } | |
1231 | ||
1232 | ret = nvme_rdma_create_queue_ib(queue, dev); | |
1233 | if (ret) { | |
1234 | nvme_rdma_dev_put(dev); | |
1235 | goto out; | |
1236 | } | |
1237 | ||
1238 | ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS); | |
1239 | if (ret) { | |
1240 | dev_err(queue->ctrl->ctrl.device, | |
1241 | "rdma_resolve_route failed (%d).\n", | |
1242 | queue->cm_error); | |
1243 | goto out_destroy_queue; | |
1244 | } | |
1245 | ||
1246 | return 0; | |
1247 | ||
1248 | out_destroy_queue: | |
1249 | nvme_rdma_destroy_queue_ib(queue); | |
1250 | out: | |
1251 | return ret; | |
1252 | } | |
1253 | ||
1254 | static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue) | |
1255 | { | |
1256 | struct nvme_rdma_ctrl *ctrl = queue->ctrl; | |
1257 | struct rdma_conn_param param = { }; | |
0b857b44 | 1258 | struct nvme_rdma_cm_req priv = { }; |
71102307 CH |
1259 | int ret; |
1260 | ||
1261 | param.qp_num = queue->qp->qp_num; | |
1262 | param.flow_control = 1; | |
1263 | ||
1264 | param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom; | |
2ac17c28 SG |
1265 | /* maximum retry count */ |
1266 | param.retry_count = 7; | |
71102307 CH |
1267 | param.rnr_retry_count = 7; |
1268 | param.private_data = &priv; | |
1269 | param.private_data_len = sizeof(priv); | |
1270 | ||
1271 | priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0); | |
1272 | priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue)); | |
f994d9dc JF |
1273 | /* |
1274 | * set the admin queue depth to the minimum size | |
1275 | * specified by the Fabrics standard. | |
1276 | */ | |
1277 | if (priv.qid == 0) { | |
1278 | priv.hrqsize = cpu_to_le16(NVMF_AQ_DEPTH); | |
1279 | priv.hsqsize = cpu_to_le16(NVMF_AQ_DEPTH - 1); | |
1280 | } else { | |
c5af8654 JF |
1281 | /* |
1282 | * current interpretation of the fabrics spec | |
1283 | * is at minimum you make hrqsize sqsize+1, or a | |
1284 | * 1's based representation of sqsize. | |
1285 | */ | |
f994d9dc | 1286 | priv.hrqsize = cpu_to_le16(queue->queue_size); |
c5af8654 | 1287 | priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize); |
f994d9dc | 1288 | } |
71102307 CH |
1289 | |
1290 | ret = rdma_connect(queue->cm_id, ¶m); | |
1291 | if (ret) { | |
1292 | dev_err(ctrl->ctrl.device, | |
1293 | "rdma_connect failed (%d).\n", ret); | |
1294 | goto out_destroy_queue_ib; | |
1295 | } | |
1296 | ||
1297 | return 0; | |
1298 | ||
1299 | out_destroy_queue_ib: | |
1300 | nvme_rdma_destroy_queue_ib(queue); | |
1301 | return ret; | |
1302 | } | |
1303 | ||
71102307 CH |
1304 | static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id, |
1305 | struct rdma_cm_event *ev) | |
1306 | { | |
1307 | struct nvme_rdma_queue *queue = cm_id->context; | |
1308 | int cm_error = 0; | |
1309 | ||
1310 | dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n", | |
1311 | rdma_event_msg(ev->event), ev->event, | |
1312 | ev->status, cm_id); | |
1313 | ||
1314 | switch (ev->event) { | |
1315 | case RDMA_CM_EVENT_ADDR_RESOLVED: | |
1316 | cm_error = nvme_rdma_addr_resolved(queue); | |
1317 | break; | |
1318 | case RDMA_CM_EVENT_ROUTE_RESOLVED: | |
1319 | cm_error = nvme_rdma_route_resolved(queue); | |
1320 | break; | |
1321 | case RDMA_CM_EVENT_ESTABLISHED: | |
1322 | queue->cm_error = nvme_rdma_conn_established(queue); | |
1323 | /* complete cm_done regardless of success/failure */ | |
1324 | complete(&queue->cm_done); | |
1325 | return 0; | |
1326 | case RDMA_CM_EVENT_REJECTED: | |
1327 | cm_error = nvme_rdma_conn_rejected(queue, ev); | |
1328 | break; | |
1329 | case RDMA_CM_EVENT_ADDR_ERROR: | |
1330 | case RDMA_CM_EVENT_ROUTE_ERROR: | |
1331 | case RDMA_CM_EVENT_CONNECT_ERROR: | |
1332 | case RDMA_CM_EVENT_UNREACHABLE: | |
1333 | dev_dbg(queue->ctrl->ctrl.device, | |
1334 | "CM error event %d\n", ev->event); | |
1335 | cm_error = -ECONNRESET; | |
1336 | break; | |
1337 | case RDMA_CM_EVENT_DISCONNECTED: | |
1338 | case RDMA_CM_EVENT_ADDR_CHANGE: | |
1339 | case RDMA_CM_EVENT_TIMEWAIT_EXIT: | |
1340 | dev_dbg(queue->ctrl->ctrl.device, | |
1341 | "disconnect received - connection closed\n"); | |
1342 | nvme_rdma_error_recovery(queue->ctrl); | |
1343 | break; | |
1344 | case RDMA_CM_EVENT_DEVICE_REMOVAL: | |
e87a911f SW |
1345 | /* device removal is handled via the ib_client API */ |
1346 | break; | |
71102307 CH |
1347 | default: |
1348 | dev_err(queue->ctrl->ctrl.device, | |
1349 | "Unexpected RDMA CM event (%d)\n", ev->event); | |
1350 | nvme_rdma_error_recovery(queue->ctrl); | |
1351 | break; | |
1352 | } | |
1353 | ||
1354 | if (cm_error) { | |
1355 | queue->cm_error = cm_error; | |
1356 | complete(&queue->cm_done); | |
1357 | } | |
1358 | ||
1359 | return 0; | |
1360 | } | |
1361 | ||
1362 | static enum blk_eh_timer_return | |
1363 | nvme_rdma_timeout(struct request *rq, bool reserved) | |
1364 | { | |
1365 | struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq); | |
1366 | ||
1367 | /* queue error recovery */ | |
1368 | nvme_rdma_error_recovery(req->queue->ctrl); | |
1369 | ||
1370 | /* fail with DNR on cmd timeout */ | |
1371 | rq->errors = NVME_SC_ABORT_REQ | NVME_SC_DNR; | |
1372 | ||
1373 | return BLK_EH_HANDLED; | |
1374 | } | |
1375 | ||
1376 | static int nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx, | |
1377 | const struct blk_mq_queue_data *bd) | |
1378 | { | |
1379 | struct nvme_ns *ns = hctx->queue->queuedata; | |
1380 | struct nvme_rdma_queue *queue = hctx->driver_data; | |
1381 | struct request *rq = bd->rq; | |
1382 | struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq); | |
1383 | struct nvme_rdma_qe *sqe = &req->sqe; | |
1384 | struct nvme_command *c = sqe->data; | |
1385 | bool flush = false; | |
1386 | struct ib_device *dev; | |
1387 | unsigned int map_len; | |
1388 | int ret; | |
1389 | ||
1390 | WARN_ON_ONCE(rq->tag < 0); | |
1391 | ||
1392 | dev = queue->device->dev; | |
1393 | ib_dma_sync_single_for_cpu(dev, sqe->dma, | |
1394 | sizeof(struct nvme_command), DMA_TO_DEVICE); | |
1395 | ||
1396 | ret = nvme_setup_cmd(ns, rq, c); | |
bac0000a | 1397 | if (ret != BLK_MQ_RQ_QUEUE_OK) |
71102307 CH |
1398 | return ret; |
1399 | ||
1400 | c->common.command_id = rq->tag; | |
1401 | blk_mq_start_request(rq); | |
1402 | ||
1403 | map_len = nvme_map_len(rq); | |
1404 | ret = nvme_rdma_map_data(queue, rq, map_len, c); | |
1405 | if (ret < 0) { | |
1406 | dev_err(queue->ctrl->ctrl.device, | |
1407 | "Failed to map data (%d)\n", ret); | |
1408 | nvme_cleanup_cmd(rq); | |
1409 | goto err; | |
1410 | } | |
1411 | ||
1412 | ib_dma_sync_single_for_device(dev, sqe->dma, | |
1413 | sizeof(struct nvme_command), DMA_TO_DEVICE); | |
1414 | ||
1415 | if (rq->cmd_type == REQ_TYPE_FS && req_op(rq) == REQ_OP_FLUSH) | |
1416 | flush = true; | |
1417 | ret = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge, | |
f5b7b559 | 1418 | req->mr->need_inval ? &req->reg_wr.wr : NULL, flush); |
71102307 CH |
1419 | if (ret) { |
1420 | nvme_rdma_unmap_data(queue, rq); | |
1421 | goto err; | |
1422 | } | |
1423 | ||
1424 | return BLK_MQ_RQ_QUEUE_OK; | |
1425 | err: | |
1426 | return (ret == -ENOMEM || ret == -EAGAIN) ? | |
1427 | BLK_MQ_RQ_QUEUE_BUSY : BLK_MQ_RQ_QUEUE_ERROR; | |
1428 | } | |
1429 | ||
1430 | static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx, unsigned int tag) | |
1431 | { | |
1432 | struct nvme_rdma_queue *queue = hctx->driver_data; | |
1433 | struct ib_cq *cq = queue->ib_cq; | |
1434 | struct ib_wc wc; | |
1435 | int found = 0; | |
1436 | ||
1437 | ib_req_notify_cq(cq, IB_CQ_NEXT_COMP); | |
1438 | while (ib_poll_cq(cq, 1, &wc) > 0) { | |
1439 | struct ib_cqe *cqe = wc.wr_cqe; | |
1440 | ||
1441 | if (cqe) { | |
1442 | if (cqe->done == nvme_rdma_recv_done) | |
1443 | found |= __nvme_rdma_recv_done(cq, &wc, tag); | |
1444 | else | |
1445 | cqe->done(cq, &wc); | |
1446 | } | |
1447 | } | |
1448 | ||
1449 | return found; | |
1450 | } | |
1451 | ||
1452 | static void nvme_rdma_complete_rq(struct request *rq) | |
1453 | { | |
1454 | struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq); | |
1455 | struct nvme_rdma_queue *queue = req->queue; | |
1456 | int error = 0; | |
1457 | ||
1458 | nvme_rdma_unmap_data(queue, rq); | |
1459 | ||
1460 | if (unlikely(rq->errors)) { | |
1461 | if (nvme_req_needs_retry(rq, rq->errors)) { | |
1462 | nvme_requeue_req(rq); | |
1463 | return; | |
1464 | } | |
1465 | ||
1466 | if (rq->cmd_type == REQ_TYPE_DRV_PRIV) | |
1467 | error = rq->errors; | |
1468 | else | |
1469 | error = nvme_error_status(rq->errors); | |
1470 | } | |
1471 | ||
1472 | blk_mq_end_request(rq, error); | |
1473 | } | |
1474 | ||
1475 | static struct blk_mq_ops nvme_rdma_mq_ops = { | |
1476 | .queue_rq = nvme_rdma_queue_rq, | |
1477 | .complete = nvme_rdma_complete_rq, | |
71102307 CH |
1478 | .init_request = nvme_rdma_init_request, |
1479 | .exit_request = nvme_rdma_exit_request, | |
1480 | .reinit_request = nvme_rdma_reinit_request, | |
1481 | .init_hctx = nvme_rdma_init_hctx, | |
1482 | .poll = nvme_rdma_poll, | |
1483 | .timeout = nvme_rdma_timeout, | |
1484 | }; | |
1485 | ||
1486 | static struct blk_mq_ops nvme_rdma_admin_mq_ops = { | |
1487 | .queue_rq = nvme_rdma_queue_rq, | |
1488 | .complete = nvme_rdma_complete_rq, | |
71102307 CH |
1489 | .init_request = nvme_rdma_init_admin_request, |
1490 | .exit_request = nvme_rdma_exit_admin_request, | |
1491 | .reinit_request = nvme_rdma_reinit_request, | |
1492 | .init_hctx = nvme_rdma_init_admin_hctx, | |
1493 | .timeout = nvme_rdma_timeout, | |
1494 | }; | |
1495 | ||
1496 | static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl) | |
1497 | { | |
1498 | int error; | |
1499 | ||
1500 | error = nvme_rdma_init_queue(ctrl, 0, NVMF_AQ_DEPTH); | |
1501 | if (error) | |
1502 | return error; | |
1503 | ||
1504 | ctrl->device = ctrl->queues[0].device; | |
1505 | ||
1506 | /* | |
1507 | * We need a reference on the device as long as the tag_set is alive, | |
1508 | * as the MRs in the request structures need a valid ib_device. | |
1509 | */ | |
1510 | error = -EINVAL; | |
1511 | if (!nvme_rdma_dev_get(ctrl->device)) | |
1512 | goto out_free_queue; | |
1513 | ||
1514 | ctrl->max_fr_pages = min_t(u32, NVME_RDMA_MAX_SEGMENTS, | |
1515 | ctrl->device->dev->attrs.max_fast_reg_page_list_len); | |
1516 | ||
1517 | memset(&ctrl->admin_tag_set, 0, sizeof(ctrl->admin_tag_set)); | |
1518 | ctrl->admin_tag_set.ops = &nvme_rdma_admin_mq_ops; | |
1519 | ctrl->admin_tag_set.queue_depth = NVME_RDMA_AQ_BLKMQ_DEPTH; | |
1520 | ctrl->admin_tag_set.reserved_tags = 2; /* connect + keep-alive */ | |
1521 | ctrl->admin_tag_set.numa_node = NUMA_NO_NODE; | |
1522 | ctrl->admin_tag_set.cmd_size = sizeof(struct nvme_rdma_request) + | |
1523 | SG_CHUNK_SIZE * sizeof(struct scatterlist); | |
1524 | ctrl->admin_tag_set.driver_data = ctrl; | |
1525 | ctrl->admin_tag_set.nr_hw_queues = 1; | |
1526 | ctrl->admin_tag_set.timeout = ADMIN_TIMEOUT; | |
1527 | ||
1528 | error = blk_mq_alloc_tag_set(&ctrl->admin_tag_set); | |
1529 | if (error) | |
1530 | goto out_put_dev; | |
1531 | ||
1532 | ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set); | |
1533 | if (IS_ERR(ctrl->ctrl.admin_q)) { | |
1534 | error = PTR_ERR(ctrl->ctrl.admin_q); | |
1535 | goto out_free_tagset; | |
1536 | } | |
1537 | ||
1538 | error = nvmf_connect_admin_queue(&ctrl->ctrl); | |
1539 | if (error) | |
1540 | goto out_cleanup_queue; | |
1541 | ||
1542 | error = nvmf_reg_read64(&ctrl->ctrl, NVME_REG_CAP, &ctrl->cap); | |
1543 | if (error) { | |
1544 | dev_err(ctrl->ctrl.device, | |
1545 | "prop_get NVME_REG_CAP failed\n"); | |
1546 | goto out_cleanup_queue; | |
1547 | } | |
1548 | ||
1549 | ctrl->ctrl.sqsize = | |
1550 | min_t(int, NVME_CAP_MQES(ctrl->cap) + 1, ctrl->ctrl.sqsize); | |
1551 | ||
1552 | error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->cap); | |
1553 | if (error) | |
1554 | goto out_cleanup_queue; | |
1555 | ||
1556 | ctrl->ctrl.max_hw_sectors = | |
1557 | (ctrl->max_fr_pages - 1) << (PAGE_SHIFT - 9); | |
1558 | ||
1559 | error = nvme_init_identify(&ctrl->ctrl); | |
1560 | if (error) | |
1561 | goto out_cleanup_queue; | |
1562 | ||
1563 | error = nvme_rdma_alloc_qe(ctrl->queues[0].device->dev, | |
1564 | &ctrl->async_event_sqe, sizeof(struct nvme_command), | |
1565 | DMA_TO_DEVICE); | |
1566 | if (error) | |
1567 | goto out_cleanup_queue; | |
1568 | ||
1569 | nvme_start_keep_alive(&ctrl->ctrl); | |
1570 | ||
1571 | return 0; | |
1572 | ||
1573 | out_cleanup_queue: | |
1574 | blk_cleanup_queue(ctrl->ctrl.admin_q); | |
1575 | out_free_tagset: | |
1576 | /* disconnect and drain the queue before freeing the tagset */ | |
1577 | nvme_rdma_stop_queue(&ctrl->queues[0]); | |
1578 | blk_mq_free_tag_set(&ctrl->admin_tag_set); | |
1579 | out_put_dev: | |
1580 | nvme_rdma_dev_put(ctrl->device); | |
1581 | out_free_queue: | |
1582 | nvme_rdma_free_queue(&ctrl->queues[0]); | |
1583 | return error; | |
1584 | } | |
1585 | ||
1586 | static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl) | |
1587 | { | |
1588 | nvme_stop_keep_alive(&ctrl->ctrl); | |
1589 | cancel_work_sync(&ctrl->err_work); | |
1590 | cancel_delayed_work_sync(&ctrl->reconnect_work); | |
1591 | ||
1592 | if (ctrl->queue_count > 1) { | |
1593 | nvme_stop_queues(&ctrl->ctrl); | |
1594 | blk_mq_tagset_busy_iter(&ctrl->tag_set, | |
1595 | nvme_cancel_request, &ctrl->ctrl); | |
1596 | nvme_rdma_free_io_queues(ctrl); | |
1597 | } | |
1598 | ||
45862ebc | 1599 | if (test_bit(NVME_RDMA_Q_CONNECTED, &ctrl->queues[0].flags)) |
71102307 CH |
1600 | nvme_shutdown_ctrl(&ctrl->ctrl); |
1601 | ||
1602 | blk_mq_stop_hw_queues(ctrl->ctrl.admin_q); | |
1603 | blk_mq_tagset_busy_iter(&ctrl->admin_tag_set, | |
1604 | nvme_cancel_request, &ctrl->ctrl); | |
1605 | nvme_rdma_destroy_admin_queue(ctrl); | |
1606 | } | |
1607 | ||
2461a8dd SG |
1608 | static void __nvme_rdma_remove_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown) |
1609 | { | |
1610 | nvme_uninit_ctrl(&ctrl->ctrl); | |
1611 | if (shutdown) | |
1612 | nvme_rdma_shutdown_ctrl(ctrl); | |
a34ca17a SG |
1613 | |
1614 | if (ctrl->ctrl.tagset) { | |
1615 | blk_cleanup_queue(ctrl->ctrl.connect_q); | |
1616 | blk_mq_free_tag_set(&ctrl->tag_set); | |
1617 | nvme_rdma_dev_put(ctrl->device); | |
1618 | } | |
1619 | ||
2461a8dd SG |
1620 | nvme_put_ctrl(&ctrl->ctrl); |
1621 | } | |
1622 | ||
71102307 CH |
1623 | static void nvme_rdma_del_ctrl_work(struct work_struct *work) |
1624 | { | |
1625 | struct nvme_rdma_ctrl *ctrl = container_of(work, | |
1626 | struct nvme_rdma_ctrl, delete_work); | |
1627 | ||
2461a8dd | 1628 | __nvme_rdma_remove_ctrl(ctrl, true); |
71102307 CH |
1629 | } |
1630 | ||
1631 | static int __nvme_rdma_del_ctrl(struct nvme_rdma_ctrl *ctrl) | |
1632 | { | |
1633 | if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_DELETING)) | |
1634 | return -EBUSY; | |
1635 | ||
1636 | if (!queue_work(nvme_rdma_wq, &ctrl->delete_work)) | |
1637 | return -EBUSY; | |
1638 | ||
1639 | return 0; | |
1640 | } | |
1641 | ||
1642 | static int nvme_rdma_del_ctrl(struct nvme_ctrl *nctrl) | |
1643 | { | |
1644 | struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl); | |
cdbecc8d | 1645 | int ret = 0; |
71102307 | 1646 | |
cdbecc8d SW |
1647 | /* |
1648 | * Keep a reference until all work is flushed since | |
1649 | * __nvme_rdma_del_ctrl can free the ctrl mem | |
1650 | */ | |
1651 | if (!kref_get_unless_zero(&ctrl->ctrl.kref)) | |
1652 | return -EBUSY; | |
71102307 | 1653 | ret = __nvme_rdma_del_ctrl(ctrl); |
cdbecc8d SW |
1654 | if (!ret) |
1655 | flush_work(&ctrl->delete_work); | |
1656 | nvme_put_ctrl(&ctrl->ctrl); | |
1657 | return ret; | |
71102307 CH |
1658 | } |
1659 | ||
1660 | static void nvme_rdma_remove_ctrl_work(struct work_struct *work) | |
1661 | { | |
1662 | struct nvme_rdma_ctrl *ctrl = container_of(work, | |
1663 | struct nvme_rdma_ctrl, delete_work); | |
1664 | ||
2461a8dd | 1665 | __nvme_rdma_remove_ctrl(ctrl, false); |
71102307 CH |
1666 | } |
1667 | ||
1668 | static void nvme_rdma_reset_ctrl_work(struct work_struct *work) | |
1669 | { | |
1670 | struct nvme_rdma_ctrl *ctrl = container_of(work, | |
1671 | struct nvme_rdma_ctrl, reset_work); | |
1672 | int ret; | |
1673 | bool changed; | |
1674 | ||
1675 | nvme_rdma_shutdown_ctrl(ctrl); | |
1676 | ||
1677 | ret = nvme_rdma_configure_admin_queue(ctrl); | |
1678 | if (ret) { | |
1679 | /* ctrl is already shutdown, just remove the ctrl */ | |
1680 | INIT_WORK(&ctrl->delete_work, nvme_rdma_remove_ctrl_work); | |
1681 | goto del_dead_ctrl; | |
1682 | } | |
1683 | ||
1684 | if (ctrl->queue_count > 1) { | |
1685 | ret = blk_mq_reinit_tagset(&ctrl->tag_set); | |
1686 | if (ret) | |
1687 | goto del_dead_ctrl; | |
1688 | ||
1689 | ret = nvme_rdma_init_io_queues(ctrl); | |
1690 | if (ret) | |
1691 | goto del_dead_ctrl; | |
1692 | ||
1693 | ret = nvme_rdma_connect_io_queues(ctrl); | |
1694 | if (ret) | |
1695 | goto del_dead_ctrl; | |
1696 | } | |
1697 | ||
1698 | changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE); | |
1699 | WARN_ON_ONCE(!changed); | |
1700 | ||
1701 | if (ctrl->queue_count > 1) { | |
1702 | nvme_start_queues(&ctrl->ctrl); | |
1703 | nvme_queue_scan(&ctrl->ctrl); | |
3ef1b4b2 | 1704 | nvme_queue_async_events(&ctrl->ctrl); |
71102307 CH |
1705 | } |
1706 | ||
1707 | return; | |
1708 | ||
1709 | del_dead_ctrl: | |
1710 | /* Deleting this dead controller... */ | |
1711 | dev_warn(ctrl->ctrl.device, "Removing after reset failure\n"); | |
1712 | WARN_ON(!queue_work(nvme_rdma_wq, &ctrl->delete_work)); | |
1713 | } | |
1714 | ||
1715 | static int nvme_rdma_reset_ctrl(struct nvme_ctrl *nctrl) | |
1716 | { | |
1717 | struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl); | |
1718 | ||
1719 | if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING)) | |
1720 | return -EBUSY; | |
1721 | ||
1722 | if (!queue_work(nvme_rdma_wq, &ctrl->reset_work)) | |
1723 | return -EBUSY; | |
1724 | ||
1725 | flush_work(&ctrl->reset_work); | |
1726 | ||
1727 | return 0; | |
1728 | } | |
1729 | ||
1730 | static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = { | |
1731 | .name = "rdma", | |
1732 | .module = THIS_MODULE, | |
1733 | .is_fabrics = true, | |
1734 | .reg_read32 = nvmf_reg_read32, | |
1735 | .reg_read64 = nvmf_reg_read64, | |
1736 | .reg_write32 = nvmf_reg_write32, | |
1737 | .reset_ctrl = nvme_rdma_reset_ctrl, | |
1738 | .free_ctrl = nvme_rdma_free_ctrl, | |
1739 | .submit_async_event = nvme_rdma_submit_async_event, | |
1740 | .delete_ctrl = nvme_rdma_del_ctrl, | |
1741 | .get_subsysnqn = nvmf_get_subsysnqn, | |
1742 | .get_address = nvmf_get_address, | |
1743 | }; | |
1744 | ||
1745 | static int nvme_rdma_create_io_queues(struct nvme_rdma_ctrl *ctrl) | |
1746 | { | |
1747 | struct nvmf_ctrl_options *opts = ctrl->ctrl.opts; | |
1748 | int ret; | |
1749 | ||
1750 | ret = nvme_set_queue_count(&ctrl->ctrl, &opts->nr_io_queues); | |
1751 | if (ret) | |
1752 | return ret; | |
1753 | ||
1754 | ctrl->queue_count = opts->nr_io_queues + 1; | |
1755 | if (ctrl->queue_count < 2) | |
1756 | return 0; | |
1757 | ||
1758 | dev_info(ctrl->ctrl.device, | |
1759 | "creating %d I/O queues.\n", opts->nr_io_queues); | |
1760 | ||
1761 | ret = nvme_rdma_init_io_queues(ctrl); | |
1762 | if (ret) | |
1763 | return ret; | |
1764 | ||
1765 | /* | |
1766 | * We need a reference on the device as long as the tag_set is alive, | |
1767 | * as the MRs in the request structures need a valid ib_device. | |
1768 | */ | |
1769 | ret = -EINVAL; | |
1770 | if (!nvme_rdma_dev_get(ctrl->device)) | |
1771 | goto out_free_io_queues; | |
1772 | ||
1773 | memset(&ctrl->tag_set, 0, sizeof(ctrl->tag_set)); | |
1774 | ctrl->tag_set.ops = &nvme_rdma_mq_ops; | |
c5af8654 | 1775 | ctrl->tag_set.queue_depth = ctrl->ctrl.opts->queue_size; |
71102307 CH |
1776 | ctrl->tag_set.reserved_tags = 1; /* fabric connect */ |
1777 | ctrl->tag_set.numa_node = NUMA_NO_NODE; | |
1778 | ctrl->tag_set.flags = BLK_MQ_F_SHOULD_MERGE; | |
1779 | ctrl->tag_set.cmd_size = sizeof(struct nvme_rdma_request) + | |
1780 | SG_CHUNK_SIZE * sizeof(struct scatterlist); | |
1781 | ctrl->tag_set.driver_data = ctrl; | |
1782 | ctrl->tag_set.nr_hw_queues = ctrl->queue_count - 1; | |
1783 | ctrl->tag_set.timeout = NVME_IO_TIMEOUT; | |
1784 | ||
1785 | ret = blk_mq_alloc_tag_set(&ctrl->tag_set); | |
1786 | if (ret) | |
1787 | goto out_put_dev; | |
1788 | ctrl->ctrl.tagset = &ctrl->tag_set; | |
1789 | ||
1790 | ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set); | |
1791 | if (IS_ERR(ctrl->ctrl.connect_q)) { | |
1792 | ret = PTR_ERR(ctrl->ctrl.connect_q); | |
1793 | goto out_free_tag_set; | |
1794 | } | |
1795 | ||
1796 | ret = nvme_rdma_connect_io_queues(ctrl); | |
1797 | if (ret) | |
1798 | goto out_cleanup_connect_q; | |
1799 | ||
1800 | return 0; | |
1801 | ||
1802 | out_cleanup_connect_q: | |
1803 | blk_cleanup_queue(ctrl->ctrl.connect_q); | |
1804 | out_free_tag_set: | |
1805 | blk_mq_free_tag_set(&ctrl->tag_set); | |
1806 | out_put_dev: | |
1807 | nvme_rdma_dev_put(ctrl->device); | |
1808 | out_free_io_queues: | |
1809 | nvme_rdma_free_io_queues(ctrl); | |
1810 | return ret; | |
1811 | } | |
1812 | ||
1813 | static int nvme_rdma_parse_ipaddr(struct sockaddr_in *in_addr, char *p) | |
1814 | { | |
1815 | u8 *addr = (u8 *)&in_addr->sin_addr.s_addr; | |
1816 | size_t buflen = strlen(p); | |
1817 | ||
1818 | /* XXX: handle IPv6 addresses */ | |
1819 | ||
1820 | if (buflen > INET_ADDRSTRLEN) | |
1821 | return -EINVAL; | |
1822 | if (in4_pton(p, buflen, addr, '\0', NULL) == 0) | |
1823 | return -EINVAL; | |
1824 | in_addr->sin_family = AF_INET; | |
1825 | return 0; | |
1826 | } | |
1827 | ||
1828 | static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev, | |
1829 | struct nvmf_ctrl_options *opts) | |
1830 | { | |
1831 | struct nvme_rdma_ctrl *ctrl; | |
1832 | int ret; | |
1833 | bool changed; | |
1834 | ||
1835 | ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL); | |
1836 | if (!ctrl) | |
1837 | return ERR_PTR(-ENOMEM); | |
1838 | ctrl->ctrl.opts = opts; | |
1839 | INIT_LIST_HEAD(&ctrl->list); | |
1840 | ||
1841 | ret = nvme_rdma_parse_ipaddr(&ctrl->addr_in, opts->traddr); | |
1842 | if (ret) { | |
1843 | pr_err("malformed IP address passed: %s\n", opts->traddr); | |
1844 | goto out_free_ctrl; | |
1845 | } | |
1846 | ||
1847 | if (opts->mask & NVMF_OPT_TRSVCID) { | |
1848 | u16 port; | |
1849 | ||
1850 | ret = kstrtou16(opts->trsvcid, 0, &port); | |
1851 | if (ret) | |
1852 | goto out_free_ctrl; | |
1853 | ||
1854 | ctrl->addr_in.sin_port = cpu_to_be16(port); | |
1855 | } else { | |
1856 | ctrl->addr_in.sin_port = cpu_to_be16(NVME_RDMA_IP_PORT); | |
1857 | } | |
1858 | ||
1859 | ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops, | |
1860 | 0 /* no quirks, we're perfect! */); | |
1861 | if (ret) | |
1862 | goto out_free_ctrl; | |
1863 | ||
1864 | ctrl->reconnect_delay = opts->reconnect_delay; | |
1865 | INIT_DELAYED_WORK(&ctrl->reconnect_work, | |
1866 | nvme_rdma_reconnect_ctrl_work); | |
1867 | INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work); | |
1868 | INIT_WORK(&ctrl->delete_work, nvme_rdma_del_ctrl_work); | |
1869 | INIT_WORK(&ctrl->reset_work, nvme_rdma_reset_ctrl_work); | |
1870 | spin_lock_init(&ctrl->lock); | |
1871 | ||
1872 | ctrl->queue_count = opts->nr_io_queues + 1; /* +1 for admin queue */ | |
c5af8654 | 1873 | ctrl->ctrl.sqsize = opts->queue_size - 1; |
71102307 CH |
1874 | ctrl->ctrl.kato = opts->kato; |
1875 | ||
1876 | ret = -ENOMEM; | |
1877 | ctrl->queues = kcalloc(ctrl->queue_count, sizeof(*ctrl->queues), | |
1878 | GFP_KERNEL); | |
1879 | if (!ctrl->queues) | |
1880 | goto out_uninit_ctrl; | |
1881 | ||
1882 | ret = nvme_rdma_configure_admin_queue(ctrl); | |
1883 | if (ret) | |
1884 | goto out_kfree_queues; | |
1885 | ||
1886 | /* sanity check icdoff */ | |
1887 | if (ctrl->ctrl.icdoff) { | |
1888 | dev_err(ctrl->ctrl.device, "icdoff is not supported!\n"); | |
1889 | goto out_remove_admin_queue; | |
1890 | } | |
1891 | ||
1892 | /* sanity check keyed sgls */ | |
1893 | if (!(ctrl->ctrl.sgls & (1 << 20))) { | |
1894 | dev_err(ctrl->ctrl.device, "Mandatory keyed sgls are not support\n"); | |
1895 | goto out_remove_admin_queue; | |
1896 | } | |
1897 | ||
1898 | if (opts->queue_size > ctrl->ctrl.maxcmd) { | |
1899 | /* warn if maxcmd is lower than queue_size */ | |
1900 | dev_warn(ctrl->ctrl.device, | |
1901 | "queue_size %zu > ctrl maxcmd %u, clamping down\n", | |
1902 | opts->queue_size, ctrl->ctrl.maxcmd); | |
1903 | opts->queue_size = ctrl->ctrl.maxcmd; | |
1904 | } | |
1905 | ||
76c08bf4 SJ |
1906 | if (opts->queue_size > ctrl->ctrl.sqsize + 1) { |
1907 | /* warn if sqsize is lower than queue_size */ | |
1908 | dev_warn(ctrl->ctrl.device, | |
1909 | "queue_size %zu > ctrl sqsize %u, clamping down\n", | |
1910 | opts->queue_size, ctrl->ctrl.sqsize + 1); | |
1911 | opts->queue_size = ctrl->ctrl.sqsize + 1; | |
1912 | } | |
1913 | ||
71102307 CH |
1914 | if (opts->nr_io_queues) { |
1915 | ret = nvme_rdma_create_io_queues(ctrl); | |
1916 | if (ret) | |
1917 | goto out_remove_admin_queue; | |
1918 | } | |
1919 | ||
1920 | changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE); | |
1921 | WARN_ON_ONCE(!changed); | |
1922 | ||
1923 | dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISp\n", | |
1924 | ctrl->ctrl.opts->subsysnqn, &ctrl->addr); | |
1925 | ||
1926 | kref_get(&ctrl->ctrl.kref); | |
1927 | ||
1928 | mutex_lock(&nvme_rdma_ctrl_mutex); | |
1929 | list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list); | |
1930 | mutex_unlock(&nvme_rdma_ctrl_mutex); | |
1931 | ||
1932 | if (opts->nr_io_queues) { | |
1933 | nvme_queue_scan(&ctrl->ctrl); | |
1934 | nvme_queue_async_events(&ctrl->ctrl); | |
1935 | } | |
1936 | ||
1937 | return &ctrl->ctrl; | |
1938 | ||
1939 | out_remove_admin_queue: | |
1940 | nvme_stop_keep_alive(&ctrl->ctrl); | |
1941 | nvme_rdma_destroy_admin_queue(ctrl); | |
1942 | out_kfree_queues: | |
1943 | kfree(ctrl->queues); | |
1944 | out_uninit_ctrl: | |
1945 | nvme_uninit_ctrl(&ctrl->ctrl); | |
1946 | nvme_put_ctrl(&ctrl->ctrl); | |
1947 | if (ret > 0) | |
1948 | ret = -EIO; | |
1949 | return ERR_PTR(ret); | |
1950 | out_free_ctrl: | |
1951 | kfree(ctrl); | |
1952 | return ERR_PTR(ret); | |
1953 | } | |
1954 | ||
1955 | static struct nvmf_transport_ops nvme_rdma_transport = { | |
1956 | .name = "rdma", | |
1957 | .required_opts = NVMF_OPT_TRADDR, | |
2ac17c28 | 1958 | .allowed_opts = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY, |
71102307 CH |
1959 | .create_ctrl = nvme_rdma_create_ctrl, |
1960 | }; | |
1961 | ||
e87a911f SW |
1962 | static void nvme_rdma_add_one(struct ib_device *ib_device) |
1963 | { | |
1964 | } | |
1965 | ||
1966 | static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data) | |
1967 | { | |
1968 | struct nvme_rdma_ctrl *ctrl; | |
1969 | ||
1970 | /* Delete all controllers using this device */ | |
1971 | mutex_lock(&nvme_rdma_ctrl_mutex); | |
1972 | list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) { | |
1973 | if (ctrl->device->dev != ib_device) | |
1974 | continue; | |
1975 | dev_info(ctrl->ctrl.device, | |
1976 | "Removing ctrl: NQN \"%s\", addr %pISp\n", | |
1977 | ctrl->ctrl.opts->subsysnqn, &ctrl->addr); | |
1978 | __nvme_rdma_del_ctrl(ctrl); | |
1979 | } | |
1980 | mutex_unlock(&nvme_rdma_ctrl_mutex); | |
1981 | ||
1982 | flush_workqueue(nvme_rdma_wq); | |
1983 | } | |
1984 | ||
1985 | static struct ib_client nvme_rdma_ib_client = { | |
1986 | .name = "nvme_rdma", | |
1987 | .add = nvme_rdma_add_one, | |
1988 | .remove = nvme_rdma_remove_one | |
1989 | }; | |
1990 | ||
71102307 CH |
1991 | static int __init nvme_rdma_init_module(void) |
1992 | { | |
e87a911f SW |
1993 | int ret; |
1994 | ||
71102307 CH |
1995 | nvme_rdma_wq = create_workqueue("nvme_rdma_wq"); |
1996 | if (!nvme_rdma_wq) | |
1997 | return -ENOMEM; | |
1998 | ||
e87a911f SW |
1999 | ret = ib_register_client(&nvme_rdma_ib_client); |
2000 | if (ret) { | |
2001 | destroy_workqueue(nvme_rdma_wq); | |
2002 | return ret; | |
2003 | } | |
2004 | ||
71102307 CH |
2005 | nvmf_register_transport(&nvme_rdma_transport); |
2006 | return 0; | |
2007 | } | |
2008 | ||
2009 | static void __exit nvme_rdma_cleanup_module(void) | |
2010 | { | |
71102307 | 2011 | nvmf_unregister_transport(&nvme_rdma_transport); |
e87a911f | 2012 | ib_unregister_client(&nvme_rdma_ib_client); |
71102307 CH |
2013 | destroy_workqueue(nvme_rdma_wq); |
2014 | } | |
2015 | ||
2016 | module_init(nvme_rdma_init_module); | |
2017 | module_exit(nvme_rdma_cleanup_module); | |
2018 | ||
2019 | MODULE_LICENSE("GPL v2"); |