]>
Commit | Line | Data |
---|---|---|
5d8762d5 | 1 | // SPDX-License-Identifier: GPL-2.0 |
71102307 CH |
2 | /* |
3 | * NVMe over Fabrics RDMA host code. | |
4 | * Copyright (c) 2015-2016 HGST, a Western Digital Company. | |
71102307 CH |
5 | */ |
6 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt | |
71102307 CH |
7 | #include <linux/module.h> |
8 | #include <linux/init.h> | |
9 | #include <linux/slab.h> | |
f41725bb | 10 | #include <rdma/mr_pool.h> |
71102307 CH |
11 | #include <linux/err.h> |
12 | #include <linux/string.h> | |
71102307 CH |
13 | #include <linux/atomic.h> |
14 | #include <linux/blk-mq.h> | |
0b36658c | 15 | #include <linux/blk-mq-rdma.h> |
71102307 CH |
16 | #include <linux/types.h> |
17 | #include <linux/list.h> | |
18 | #include <linux/mutex.h> | |
19 | #include <linux/scatterlist.h> | |
20 | #include <linux/nvme.h> | |
71102307 CH |
21 | #include <asm/unaligned.h> |
22 | ||
23 | #include <rdma/ib_verbs.h> | |
24 | #include <rdma/rdma_cm.h> | |
71102307 CH |
25 | #include <linux/nvme-rdma.h> |
26 | ||
27 | #include "nvme.h" | |
28 | #include "fabrics.h" | |
29 | ||
30 | ||
782d820c | 31 | #define NVME_RDMA_CONNECT_TIMEOUT_MS 3000 /* 3 second */ |
71102307 | 32 | |
71102307 CH |
33 | #define NVME_RDMA_MAX_SEGMENTS 256 |
34 | ||
64a741c1 | 35 | #define NVME_RDMA_MAX_INLINE_SEGMENTS 4 |
71102307 | 36 | |
71102307 | 37 | struct nvme_rdma_device { |
f87c89ad MG |
38 | struct ib_device *dev; |
39 | struct ib_pd *pd; | |
71102307 CH |
40 | struct kref ref; |
41 | struct list_head entry; | |
64a741c1 | 42 | unsigned int num_inline_segments; |
71102307 CH |
43 | }; |
44 | ||
45 | struct nvme_rdma_qe { | |
46 | struct ib_cqe cqe; | |
47 | void *data; | |
48 | u64 dma; | |
49 | }; | |
50 | ||
51 | struct nvme_rdma_queue; | |
52 | struct nvme_rdma_request { | |
d49187e9 | 53 | struct nvme_request req; |
71102307 CH |
54 | struct ib_mr *mr; |
55 | struct nvme_rdma_qe sqe; | |
4af7f7ff SG |
56 | union nvme_result result; |
57 | __le16 status; | |
58 | refcount_t ref; | |
71102307 CH |
59 | struct ib_sge sge[1 + NVME_RDMA_MAX_INLINE_SEGMENTS]; |
60 | u32 num_sge; | |
61 | int nents; | |
71102307 CH |
62 | struct ib_reg_wr reg_wr; |
63 | struct ib_cqe reg_cqe; | |
64 | struct nvme_rdma_queue *queue; | |
65 | struct sg_table sg_table; | |
66 | struct scatterlist first_sgl[]; | |
67 | }; | |
68 | ||
69 | enum nvme_rdma_queue_flags { | |
5013e98b SG |
70 | NVME_RDMA_Q_ALLOCATED = 0, |
71 | NVME_RDMA_Q_LIVE = 1, | |
eb1bd249 | 72 | NVME_RDMA_Q_TR_READY = 2, |
71102307 CH |
73 | }; |
74 | ||
75 | struct nvme_rdma_queue { | |
76 | struct nvme_rdma_qe *rsp_ring; | |
71102307 CH |
77 | int queue_size; |
78 | size_t cmnd_capsule_len; | |
79 | struct nvme_rdma_ctrl *ctrl; | |
80 | struct nvme_rdma_device *device; | |
81 | struct ib_cq *ib_cq; | |
82 | struct ib_qp *qp; | |
83 | ||
84 | unsigned long flags; | |
85 | struct rdma_cm_id *cm_id; | |
86 | int cm_error; | |
87 | struct completion cm_done; | |
88 | }; | |
89 | ||
90 | struct nvme_rdma_ctrl { | |
71102307 CH |
91 | /* read only in the hot path */ |
92 | struct nvme_rdma_queue *queues; | |
71102307 CH |
93 | |
94 | /* other member variables */ | |
71102307 | 95 | struct blk_mq_tag_set tag_set; |
71102307 CH |
96 | struct work_struct err_work; |
97 | ||
98 | struct nvme_rdma_qe async_event_sqe; | |
99 | ||
71102307 CH |
100 | struct delayed_work reconnect_work; |
101 | ||
102 | struct list_head list; | |
103 | ||
104 | struct blk_mq_tag_set admin_tag_set; | |
105 | struct nvme_rdma_device *device; | |
106 | ||
71102307 CH |
107 | u32 max_fr_pages; |
108 | ||
0928f9b4 SG |
109 | struct sockaddr_storage addr; |
110 | struct sockaddr_storage src_addr; | |
71102307 CH |
111 | |
112 | struct nvme_ctrl ctrl; | |
64a741c1 | 113 | bool use_inline_data; |
b1064d3e | 114 | u32 io_queues[HCTX_MAX_TYPES]; |
71102307 CH |
115 | }; |
116 | ||
117 | static inline struct nvme_rdma_ctrl *to_rdma_ctrl(struct nvme_ctrl *ctrl) | |
118 | { | |
119 | return container_of(ctrl, struct nvme_rdma_ctrl, ctrl); | |
120 | } | |
121 | ||
122 | static LIST_HEAD(device_list); | |
123 | static DEFINE_MUTEX(device_list_mutex); | |
124 | ||
125 | static LIST_HEAD(nvme_rdma_ctrl_list); | |
126 | static DEFINE_MUTEX(nvme_rdma_ctrl_mutex); | |
127 | ||
71102307 CH |
128 | /* |
129 | * Disabling this option makes small I/O goes faster, but is fundamentally | |
130 | * unsafe. With it turned off we will have to register a global rkey that | |
131 | * allows read and write access to all physical memory. | |
132 | */ | |
133 | static bool register_always = true; | |
134 | module_param(register_always, bool, 0444); | |
135 | MODULE_PARM_DESC(register_always, | |
136 | "Use memory registration even for contiguous memory regions"); | |
137 | ||
138 | static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id, | |
139 | struct rdma_cm_event *event); | |
140 | static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc); | |
71102307 | 141 | |
90af3512 SG |
142 | static const struct blk_mq_ops nvme_rdma_mq_ops; |
143 | static const struct blk_mq_ops nvme_rdma_admin_mq_ops; | |
144 | ||
71102307 CH |
145 | /* XXX: really should move to a generic header sooner or later.. */ |
146 | static inline void put_unaligned_le24(u32 val, u8 *p) | |
147 | { | |
148 | *p++ = val; | |
149 | *p++ = val >> 8; | |
150 | *p++ = val >> 16; | |
151 | } | |
152 | ||
153 | static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue *queue) | |
154 | { | |
155 | return queue - queue->ctrl->queues; | |
156 | } | |
157 | ||
ff8519f9 SG |
158 | static bool nvme_rdma_poll_queue(struct nvme_rdma_queue *queue) |
159 | { | |
160 | return nvme_rdma_queue_idx(queue) > | |
b1064d3e SG |
161 | queue->ctrl->io_queues[HCTX_TYPE_DEFAULT] + |
162 | queue->ctrl->io_queues[HCTX_TYPE_READ]; | |
ff8519f9 SG |
163 | } |
164 | ||
71102307 CH |
165 | static inline size_t nvme_rdma_inline_data_size(struct nvme_rdma_queue *queue) |
166 | { | |
167 | return queue->cmnd_capsule_len - sizeof(struct nvme_command); | |
168 | } | |
169 | ||
170 | static void nvme_rdma_free_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe, | |
171 | size_t capsule_size, enum dma_data_direction dir) | |
172 | { | |
173 | ib_dma_unmap_single(ibdev, qe->dma, capsule_size, dir); | |
174 | kfree(qe->data); | |
175 | } | |
176 | ||
177 | static int nvme_rdma_alloc_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe, | |
178 | size_t capsule_size, enum dma_data_direction dir) | |
179 | { | |
180 | qe->data = kzalloc(capsule_size, GFP_KERNEL); | |
181 | if (!qe->data) | |
182 | return -ENOMEM; | |
183 | ||
184 | qe->dma = ib_dma_map_single(ibdev, qe->data, capsule_size, dir); | |
185 | if (ib_dma_mapping_error(ibdev, qe->dma)) { | |
186 | kfree(qe->data); | |
6344d02d | 187 | qe->data = NULL; |
71102307 CH |
188 | return -ENOMEM; |
189 | } | |
190 | ||
191 | return 0; | |
192 | } | |
193 | ||
194 | static void nvme_rdma_free_ring(struct ib_device *ibdev, | |
195 | struct nvme_rdma_qe *ring, size_t ib_queue_size, | |
196 | size_t capsule_size, enum dma_data_direction dir) | |
197 | { | |
198 | int i; | |
199 | ||
200 | for (i = 0; i < ib_queue_size; i++) | |
201 | nvme_rdma_free_qe(ibdev, &ring[i], capsule_size, dir); | |
202 | kfree(ring); | |
203 | } | |
204 | ||
205 | static struct nvme_rdma_qe *nvme_rdma_alloc_ring(struct ib_device *ibdev, | |
206 | size_t ib_queue_size, size_t capsule_size, | |
207 | enum dma_data_direction dir) | |
208 | { | |
209 | struct nvme_rdma_qe *ring; | |
210 | int i; | |
211 | ||
212 | ring = kcalloc(ib_queue_size, sizeof(struct nvme_rdma_qe), GFP_KERNEL); | |
213 | if (!ring) | |
214 | return NULL; | |
215 | ||
62f99b62 MG |
216 | /* |
217 | * Bind the CQEs (post recv buffers) DMA mapping to the RDMA queue | |
218 | * lifetime. It's safe, since any chage in the underlying RDMA device | |
219 | * will issue error recovery and queue re-creation. | |
220 | */ | |
71102307 CH |
221 | for (i = 0; i < ib_queue_size; i++) { |
222 | if (nvme_rdma_alloc_qe(ibdev, &ring[i], capsule_size, dir)) | |
223 | goto out_free_ring; | |
224 | } | |
225 | ||
226 | return ring; | |
227 | ||
228 | out_free_ring: | |
229 | nvme_rdma_free_ring(ibdev, ring, i, capsule_size, dir); | |
230 | return NULL; | |
231 | } | |
232 | ||
233 | static void nvme_rdma_qp_event(struct ib_event *event, void *context) | |
234 | { | |
27a4beef MG |
235 | pr_debug("QP event %s (%d)\n", |
236 | ib_event_msg(event->event), event->event); | |
237 | ||
71102307 CH |
238 | } |
239 | ||
240 | static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue *queue) | |
241 | { | |
35da77d5 BVA |
242 | int ret; |
243 | ||
244 | ret = wait_for_completion_interruptible_timeout(&queue->cm_done, | |
71102307 | 245 | msecs_to_jiffies(NVME_RDMA_CONNECT_TIMEOUT_MS) + 1); |
35da77d5 BVA |
246 | if (ret < 0) |
247 | return ret; | |
248 | if (ret == 0) | |
249 | return -ETIMEDOUT; | |
250 | WARN_ON_ONCE(queue->cm_error > 0); | |
71102307 CH |
251 | return queue->cm_error; |
252 | } | |
253 | ||
254 | static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor) | |
255 | { | |
256 | struct nvme_rdma_device *dev = queue->device; | |
257 | struct ib_qp_init_attr init_attr; | |
258 | int ret; | |
259 | ||
260 | memset(&init_attr, 0, sizeof(init_attr)); | |
261 | init_attr.event_handler = nvme_rdma_qp_event; | |
262 | /* +1 for drain */ | |
263 | init_attr.cap.max_send_wr = factor * queue->queue_size + 1; | |
264 | /* +1 for drain */ | |
265 | init_attr.cap.max_recv_wr = queue->queue_size + 1; | |
266 | init_attr.cap.max_recv_sge = 1; | |
64a741c1 | 267 | init_attr.cap.max_send_sge = 1 + dev->num_inline_segments; |
71102307 CH |
268 | init_attr.sq_sig_type = IB_SIGNAL_REQ_WR; |
269 | init_attr.qp_type = IB_QPT_RC; | |
270 | init_attr.send_cq = queue->ib_cq; | |
271 | init_attr.recv_cq = queue->ib_cq; | |
272 | ||
273 | ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr); | |
274 | ||
275 | queue->qp = queue->cm_id->qp; | |
276 | return ret; | |
277 | } | |
278 | ||
385475ee CH |
279 | static void nvme_rdma_exit_request(struct blk_mq_tag_set *set, |
280 | struct request *rq, unsigned int hctx_idx) | |
71102307 CH |
281 | { |
282 | struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq); | |
71102307 | 283 | |
62f99b62 | 284 | kfree(req->sqe.data); |
71102307 CH |
285 | } |
286 | ||
385475ee CH |
287 | static int nvme_rdma_init_request(struct blk_mq_tag_set *set, |
288 | struct request *rq, unsigned int hctx_idx, | |
289 | unsigned int numa_node) | |
71102307 | 290 | { |
385475ee | 291 | struct nvme_rdma_ctrl *ctrl = set->driver_data; |
71102307 | 292 | struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq); |
385475ee | 293 | int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0; |
71102307 | 294 | struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx]; |
71102307 | 295 | |
59e29ce6 | 296 | nvme_req(rq)->ctrl = &ctrl->ctrl; |
62f99b62 MG |
297 | req->sqe.data = kzalloc(sizeof(struct nvme_command), GFP_KERNEL); |
298 | if (!req->sqe.data) | |
299 | return -ENOMEM; | |
71102307 | 300 | |
71102307 CH |
301 | req->queue = queue; |
302 | ||
303 | return 0; | |
71102307 CH |
304 | } |
305 | ||
71102307 CH |
306 | static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data, |
307 | unsigned int hctx_idx) | |
308 | { | |
309 | struct nvme_rdma_ctrl *ctrl = data; | |
310 | struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1]; | |
311 | ||
d858e5f0 | 312 | BUG_ON(hctx_idx >= ctrl->ctrl.queue_count); |
71102307 CH |
313 | |
314 | hctx->driver_data = queue; | |
315 | return 0; | |
316 | } | |
317 | ||
318 | static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data, | |
319 | unsigned int hctx_idx) | |
320 | { | |
321 | struct nvme_rdma_ctrl *ctrl = data; | |
322 | struct nvme_rdma_queue *queue = &ctrl->queues[0]; | |
323 | ||
324 | BUG_ON(hctx_idx != 0); | |
325 | ||
326 | hctx->driver_data = queue; | |
327 | return 0; | |
328 | } | |
329 | ||
330 | static void nvme_rdma_free_dev(struct kref *ref) | |
331 | { | |
332 | struct nvme_rdma_device *ndev = | |
333 | container_of(ref, struct nvme_rdma_device, ref); | |
334 | ||
335 | mutex_lock(&device_list_mutex); | |
336 | list_del(&ndev->entry); | |
337 | mutex_unlock(&device_list_mutex); | |
338 | ||
71102307 | 339 | ib_dealloc_pd(ndev->pd); |
71102307 CH |
340 | kfree(ndev); |
341 | } | |
342 | ||
343 | static void nvme_rdma_dev_put(struct nvme_rdma_device *dev) | |
344 | { | |
345 | kref_put(&dev->ref, nvme_rdma_free_dev); | |
346 | } | |
347 | ||
348 | static int nvme_rdma_dev_get(struct nvme_rdma_device *dev) | |
349 | { | |
350 | return kref_get_unless_zero(&dev->ref); | |
351 | } | |
352 | ||
353 | static struct nvme_rdma_device * | |
354 | nvme_rdma_find_get_device(struct rdma_cm_id *cm_id) | |
355 | { | |
356 | struct nvme_rdma_device *ndev; | |
357 | ||
358 | mutex_lock(&device_list_mutex); | |
359 | list_for_each_entry(ndev, &device_list, entry) { | |
360 | if (ndev->dev->node_guid == cm_id->device->node_guid && | |
361 | nvme_rdma_dev_get(ndev)) | |
362 | goto out_unlock; | |
363 | } | |
364 | ||
365 | ndev = kzalloc(sizeof(*ndev), GFP_KERNEL); | |
366 | if (!ndev) | |
367 | goto out_err; | |
368 | ||
369 | ndev->dev = cm_id->device; | |
370 | kref_init(&ndev->ref); | |
371 | ||
11975e01 CH |
372 | ndev->pd = ib_alloc_pd(ndev->dev, |
373 | register_always ? 0 : IB_PD_UNSAFE_GLOBAL_RKEY); | |
71102307 CH |
374 | if (IS_ERR(ndev->pd)) |
375 | goto out_free_dev; | |
376 | ||
71102307 CH |
377 | if (!(ndev->dev->attrs.device_cap_flags & |
378 | IB_DEVICE_MEM_MGT_EXTENSIONS)) { | |
379 | dev_err(&ndev->dev->dev, | |
380 | "Memory registrations not supported.\n"); | |
11975e01 | 381 | goto out_free_pd; |
71102307 CH |
382 | } |
383 | ||
64a741c1 | 384 | ndev->num_inline_segments = min(NVME_RDMA_MAX_INLINE_SEGMENTS, |
0a3173a5 | 385 | ndev->dev->attrs.max_send_sge - 1); |
71102307 CH |
386 | list_add(&ndev->entry, &device_list); |
387 | out_unlock: | |
388 | mutex_unlock(&device_list_mutex); | |
389 | return ndev; | |
390 | ||
71102307 CH |
391 | out_free_pd: |
392 | ib_dealloc_pd(ndev->pd); | |
393 | out_free_dev: | |
394 | kfree(ndev); | |
395 | out_err: | |
396 | mutex_unlock(&device_list_mutex); | |
397 | return NULL; | |
398 | } | |
399 | ||
400 | static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue) | |
401 | { | |
eb1bd249 MG |
402 | struct nvme_rdma_device *dev; |
403 | struct ib_device *ibdev; | |
404 | ||
405 | if (!test_and_clear_bit(NVME_RDMA_Q_TR_READY, &queue->flags)) | |
406 | return; | |
407 | ||
408 | dev = queue->device; | |
409 | ibdev = dev->dev; | |
71102307 | 410 | |
f41725bb IR |
411 | ib_mr_pool_destroy(queue->qp, &queue->qp->rdma_mrs); |
412 | ||
eb1bd249 MG |
413 | /* |
414 | * The cm_id object might have been destroyed during RDMA connection | |
415 | * establishment error flow to avoid getting other cma events, thus | |
416 | * the destruction of the QP shouldn't use rdma_cm API. | |
417 | */ | |
418 | ib_destroy_qp(queue->qp); | |
71102307 CH |
419 | ib_free_cq(queue->ib_cq); |
420 | ||
421 | nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size, | |
422 | sizeof(struct nvme_completion), DMA_FROM_DEVICE); | |
423 | ||
424 | nvme_rdma_dev_put(dev); | |
425 | } | |
426 | ||
f41725bb IR |
427 | static int nvme_rdma_get_max_fr_pages(struct ib_device *ibdev) |
428 | { | |
429 | return min_t(u32, NVME_RDMA_MAX_SEGMENTS, | |
ff13c1b8 | 430 | ibdev->attrs.max_fast_reg_page_list_len - 1); |
f41725bb IR |
431 | } |
432 | ||
ca6e95bb | 433 | static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue) |
71102307 | 434 | { |
ca6e95bb | 435 | struct ib_device *ibdev; |
71102307 CH |
436 | const int send_wr_factor = 3; /* MR, SEND, INV */ |
437 | const int cq_factor = send_wr_factor + 1; /* + RECV */ | |
438 | int comp_vector, idx = nvme_rdma_queue_idx(queue); | |
ff8519f9 | 439 | enum ib_poll_context poll_ctx; |
ff13c1b8 | 440 | int ret, pages_per_mr; |
71102307 | 441 | |
ca6e95bb SG |
442 | queue->device = nvme_rdma_find_get_device(queue->cm_id); |
443 | if (!queue->device) { | |
444 | dev_err(queue->cm_id->device->dev.parent, | |
445 | "no client data found!\n"); | |
446 | return -ECONNREFUSED; | |
447 | } | |
448 | ibdev = queue->device->dev; | |
71102307 CH |
449 | |
450 | /* | |
0b36658c SG |
451 | * Spread I/O queues completion vectors according their queue index. |
452 | * Admin queues can always go on completion vector 0. | |
71102307 | 453 | */ |
0b36658c | 454 | comp_vector = idx == 0 ? idx : idx - 1; |
71102307 | 455 | |
ff8519f9 SG |
456 | /* Polling queues need direct cq polling context */ |
457 | if (nvme_rdma_poll_queue(queue)) | |
458 | poll_ctx = IB_POLL_DIRECT; | |
459 | else | |
460 | poll_ctx = IB_POLL_SOFTIRQ; | |
461 | ||
71102307 | 462 | /* +1 for ib_stop_cq */ |
ca6e95bb SG |
463 | queue->ib_cq = ib_alloc_cq(ibdev, queue, |
464 | cq_factor * queue->queue_size + 1, | |
ff8519f9 | 465 | comp_vector, poll_ctx); |
71102307 CH |
466 | if (IS_ERR(queue->ib_cq)) { |
467 | ret = PTR_ERR(queue->ib_cq); | |
ca6e95bb | 468 | goto out_put_dev; |
71102307 CH |
469 | } |
470 | ||
471 | ret = nvme_rdma_create_qp(queue, send_wr_factor); | |
472 | if (ret) | |
473 | goto out_destroy_ib_cq; | |
474 | ||
475 | queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size, | |
476 | sizeof(struct nvme_completion), DMA_FROM_DEVICE); | |
477 | if (!queue->rsp_ring) { | |
478 | ret = -ENOMEM; | |
479 | goto out_destroy_qp; | |
480 | } | |
481 | ||
ff13c1b8 MG |
482 | /* |
483 | * Currently we don't use SG_GAPS MR's so if the first entry is | |
484 | * misaligned we'll end up using two entries for a single data page, | |
485 | * so one additional entry is required. | |
486 | */ | |
487 | pages_per_mr = nvme_rdma_get_max_fr_pages(ibdev) + 1; | |
f41725bb IR |
488 | ret = ib_mr_pool_init(queue->qp, &queue->qp->rdma_mrs, |
489 | queue->queue_size, | |
490 | IB_MR_TYPE_MEM_REG, | |
ff13c1b8 | 491 | pages_per_mr, 0); |
f41725bb IR |
492 | if (ret) { |
493 | dev_err(queue->ctrl->ctrl.device, | |
494 | "failed to initialize MR pool sized %d for QID %d\n", | |
495 | queue->queue_size, idx); | |
496 | goto out_destroy_ring; | |
497 | } | |
498 | ||
eb1bd249 MG |
499 | set_bit(NVME_RDMA_Q_TR_READY, &queue->flags); |
500 | ||
71102307 CH |
501 | return 0; |
502 | ||
f41725bb IR |
503 | out_destroy_ring: |
504 | nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size, | |
505 | sizeof(struct nvme_completion), DMA_FROM_DEVICE); | |
71102307 | 506 | out_destroy_qp: |
1f61def9 | 507 | rdma_destroy_qp(queue->cm_id); |
71102307 CH |
508 | out_destroy_ib_cq: |
509 | ib_free_cq(queue->ib_cq); | |
ca6e95bb SG |
510 | out_put_dev: |
511 | nvme_rdma_dev_put(queue->device); | |
71102307 CH |
512 | return ret; |
513 | } | |
514 | ||
41e8cfa1 | 515 | static int nvme_rdma_alloc_queue(struct nvme_rdma_ctrl *ctrl, |
71102307 CH |
516 | int idx, size_t queue_size) |
517 | { | |
518 | struct nvme_rdma_queue *queue; | |
8f4e8dac | 519 | struct sockaddr *src_addr = NULL; |
71102307 CH |
520 | int ret; |
521 | ||
522 | queue = &ctrl->queues[idx]; | |
523 | queue->ctrl = ctrl; | |
524 | init_completion(&queue->cm_done); | |
525 | ||
526 | if (idx > 0) | |
527 | queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16; | |
528 | else | |
529 | queue->cmnd_capsule_len = sizeof(struct nvme_command); | |
530 | ||
531 | queue->queue_size = queue_size; | |
532 | ||
533 | queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue, | |
534 | RDMA_PS_TCP, IB_QPT_RC); | |
535 | if (IS_ERR(queue->cm_id)) { | |
536 | dev_info(ctrl->ctrl.device, | |
537 | "failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id)); | |
538 | return PTR_ERR(queue->cm_id); | |
539 | } | |
540 | ||
8f4e8dac | 541 | if (ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR) |
0928f9b4 | 542 | src_addr = (struct sockaddr *)&ctrl->src_addr; |
8f4e8dac | 543 | |
0928f9b4 SG |
544 | queue->cm_error = -ETIMEDOUT; |
545 | ret = rdma_resolve_addr(queue->cm_id, src_addr, | |
546 | (struct sockaddr *)&ctrl->addr, | |
71102307 CH |
547 | NVME_RDMA_CONNECT_TIMEOUT_MS); |
548 | if (ret) { | |
549 | dev_info(ctrl->ctrl.device, | |
550 | "rdma_resolve_addr failed (%d).\n", ret); | |
551 | goto out_destroy_cm_id; | |
552 | } | |
553 | ||
554 | ret = nvme_rdma_wait_for_cm(queue); | |
555 | if (ret) { | |
556 | dev_info(ctrl->ctrl.device, | |
d8bfceeb | 557 | "rdma connection establishment failed (%d)\n", ret); |
71102307 CH |
558 | goto out_destroy_cm_id; |
559 | } | |
560 | ||
5013e98b | 561 | set_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags); |
71102307 CH |
562 | |
563 | return 0; | |
564 | ||
565 | out_destroy_cm_id: | |
566 | rdma_destroy_id(queue->cm_id); | |
eb1bd249 | 567 | nvme_rdma_destroy_queue_ib(queue); |
71102307 CH |
568 | return ret; |
569 | } | |
570 | ||
d94211b8 SG |
571 | static void __nvme_rdma_stop_queue(struct nvme_rdma_queue *queue) |
572 | { | |
573 | rdma_disconnect(queue->cm_id); | |
574 | ib_drain_qp(queue->qp); | |
575 | } | |
576 | ||
71102307 CH |
577 | static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue) |
578 | { | |
a57bd541 SG |
579 | if (!test_and_clear_bit(NVME_RDMA_Q_LIVE, &queue->flags)) |
580 | return; | |
d94211b8 | 581 | __nvme_rdma_stop_queue(queue); |
71102307 CH |
582 | } |
583 | ||
584 | static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue) | |
585 | { | |
5013e98b | 586 | if (!test_and_clear_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags)) |
a57bd541 SG |
587 | return; |
588 | ||
71102307 CH |
589 | nvme_rdma_destroy_queue_ib(queue); |
590 | rdma_destroy_id(queue->cm_id); | |
591 | } | |
592 | ||
a57bd541 | 593 | static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl) |
71102307 | 594 | { |
a57bd541 SG |
595 | int i; |
596 | ||
597 | for (i = 1; i < ctrl->ctrl.queue_count; i++) | |
598 | nvme_rdma_free_queue(&ctrl->queues[i]); | |
71102307 CH |
599 | } |
600 | ||
a57bd541 | 601 | static void nvme_rdma_stop_io_queues(struct nvme_rdma_ctrl *ctrl) |
71102307 CH |
602 | { |
603 | int i; | |
604 | ||
d858e5f0 | 605 | for (i = 1; i < ctrl->ctrl.queue_count; i++) |
a57bd541 | 606 | nvme_rdma_stop_queue(&ctrl->queues[i]); |
71102307 CH |
607 | } |
608 | ||
68e16fcf SG |
609 | static int nvme_rdma_start_queue(struct nvme_rdma_ctrl *ctrl, int idx) |
610 | { | |
ff8519f9 SG |
611 | struct nvme_rdma_queue *queue = &ctrl->queues[idx]; |
612 | bool poll = nvme_rdma_poll_queue(queue); | |
68e16fcf SG |
613 | int ret; |
614 | ||
615 | if (idx) | |
ff8519f9 | 616 | ret = nvmf_connect_io_queue(&ctrl->ctrl, idx, poll); |
68e16fcf SG |
617 | else |
618 | ret = nvmf_connect_admin_queue(&ctrl->ctrl); | |
619 | ||
d94211b8 | 620 | if (!ret) { |
ff8519f9 | 621 | set_bit(NVME_RDMA_Q_LIVE, &queue->flags); |
d94211b8 | 622 | } else { |
67b483dd SG |
623 | if (test_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags)) |
624 | __nvme_rdma_stop_queue(queue); | |
68e16fcf SG |
625 | dev_info(ctrl->ctrl.device, |
626 | "failed to connect queue: %d ret=%d\n", idx, ret); | |
d94211b8 | 627 | } |
68e16fcf SG |
628 | return ret; |
629 | } | |
630 | ||
631 | static int nvme_rdma_start_io_queues(struct nvme_rdma_ctrl *ctrl) | |
71102307 CH |
632 | { |
633 | int i, ret = 0; | |
634 | ||
d858e5f0 | 635 | for (i = 1; i < ctrl->ctrl.queue_count; i++) { |
68e16fcf SG |
636 | ret = nvme_rdma_start_queue(ctrl, i); |
637 | if (ret) | |
a57bd541 | 638 | goto out_stop_queues; |
71102307 CH |
639 | } |
640 | ||
c8dbc37c SW |
641 | return 0; |
642 | ||
a57bd541 | 643 | out_stop_queues: |
68e16fcf SG |
644 | for (i--; i >= 1; i--) |
645 | nvme_rdma_stop_queue(&ctrl->queues[i]); | |
71102307 CH |
646 | return ret; |
647 | } | |
648 | ||
41e8cfa1 | 649 | static int nvme_rdma_alloc_io_queues(struct nvme_rdma_ctrl *ctrl) |
71102307 | 650 | { |
c248c643 | 651 | struct nvmf_ctrl_options *opts = ctrl->ctrl.opts; |
0b36658c | 652 | struct ib_device *ibdev = ctrl->device->dev; |
5651cd3c SG |
653 | unsigned int nr_io_queues, nr_default_queues; |
654 | unsigned int nr_read_queues, nr_poll_queues; | |
71102307 CH |
655 | int i, ret; |
656 | ||
5651cd3c SG |
657 | nr_read_queues = min_t(unsigned int, ibdev->num_comp_vectors, |
658 | min(opts->nr_io_queues, num_online_cpus())); | |
659 | nr_default_queues = min_t(unsigned int, ibdev->num_comp_vectors, | |
660 | min(opts->nr_write_queues, num_online_cpus())); | |
661 | nr_poll_queues = min(opts->nr_poll_queues, num_online_cpus()); | |
662 | nr_io_queues = nr_read_queues + nr_default_queues + nr_poll_queues; | |
b65bb777 | 663 | |
c248c643 SG |
664 | ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues); |
665 | if (ret) | |
666 | return ret; | |
667 | ||
d858e5f0 SG |
668 | ctrl->ctrl.queue_count = nr_io_queues + 1; |
669 | if (ctrl->ctrl.queue_count < 2) | |
c248c643 SG |
670 | return 0; |
671 | ||
672 | dev_info(ctrl->ctrl.device, | |
673 | "creating %d I/O queues.\n", nr_io_queues); | |
674 | ||
5651cd3c SG |
675 | if (opts->nr_write_queues && nr_read_queues < nr_io_queues) { |
676 | /* | |
677 | * separate read/write queues | |
678 | * hand out dedicated default queues only after we have | |
679 | * sufficient read queues. | |
680 | */ | |
681 | ctrl->io_queues[HCTX_TYPE_READ] = nr_read_queues; | |
682 | nr_io_queues -= ctrl->io_queues[HCTX_TYPE_READ]; | |
683 | ctrl->io_queues[HCTX_TYPE_DEFAULT] = | |
684 | min(nr_default_queues, nr_io_queues); | |
685 | nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT]; | |
686 | } else { | |
687 | /* | |
688 | * shared read/write queues | |
689 | * either no write queues were requested, or we don't have | |
690 | * sufficient queue count to have dedicated default queues. | |
691 | */ | |
692 | ctrl->io_queues[HCTX_TYPE_DEFAULT] = | |
693 | min(nr_read_queues, nr_io_queues); | |
694 | nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT]; | |
695 | } | |
696 | ||
697 | if (opts->nr_poll_queues && nr_io_queues) { | |
698 | /* map dedicated poll queues only if we have queues left */ | |
699 | ctrl->io_queues[HCTX_TYPE_POLL] = | |
700 | min(nr_poll_queues, nr_io_queues); | |
701 | } | |
702 | ||
d858e5f0 | 703 | for (i = 1; i < ctrl->ctrl.queue_count; i++) { |
41e8cfa1 SG |
704 | ret = nvme_rdma_alloc_queue(ctrl, i, |
705 | ctrl->ctrl.sqsize + 1); | |
706 | if (ret) | |
71102307 | 707 | goto out_free_queues; |
71102307 CH |
708 | } |
709 | ||
710 | return 0; | |
711 | ||
712 | out_free_queues: | |
f361e5a0 | 713 | for (i--; i >= 1; i--) |
a57bd541 | 714 | nvme_rdma_free_queue(&ctrl->queues[i]); |
71102307 CH |
715 | |
716 | return ret; | |
717 | } | |
718 | ||
b28a308e SG |
719 | static struct blk_mq_tag_set *nvme_rdma_alloc_tagset(struct nvme_ctrl *nctrl, |
720 | bool admin) | |
721 | { | |
722 | struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl); | |
723 | struct blk_mq_tag_set *set; | |
724 | int ret; | |
725 | ||
726 | if (admin) { | |
727 | set = &ctrl->admin_tag_set; | |
728 | memset(set, 0, sizeof(*set)); | |
729 | set->ops = &nvme_rdma_admin_mq_ops; | |
38dabe21 | 730 | set->queue_depth = NVME_AQ_MQ_TAG_DEPTH; |
b28a308e | 731 | set->reserved_tags = 2; /* connect + keep-alive */ |
103e515e | 732 | set->numa_node = nctrl->numa_node; |
b28a308e SG |
733 | set->cmd_size = sizeof(struct nvme_rdma_request) + |
734 | SG_CHUNK_SIZE * sizeof(struct scatterlist); | |
735 | set->driver_data = ctrl; | |
736 | set->nr_hw_queues = 1; | |
737 | set->timeout = ADMIN_TIMEOUT; | |
94f29d4f | 738 | set->flags = BLK_MQ_F_NO_SCHED; |
b28a308e SG |
739 | } else { |
740 | set = &ctrl->tag_set; | |
741 | memset(set, 0, sizeof(*set)); | |
742 | set->ops = &nvme_rdma_mq_ops; | |
5e77d61c | 743 | set->queue_depth = nctrl->sqsize + 1; |
b28a308e | 744 | set->reserved_tags = 1; /* fabric connect */ |
103e515e | 745 | set->numa_node = nctrl->numa_node; |
b28a308e SG |
746 | set->flags = BLK_MQ_F_SHOULD_MERGE; |
747 | set->cmd_size = sizeof(struct nvme_rdma_request) + | |
748 | SG_CHUNK_SIZE * sizeof(struct scatterlist); | |
749 | set->driver_data = ctrl; | |
750 | set->nr_hw_queues = nctrl->queue_count - 1; | |
751 | set->timeout = NVME_IO_TIMEOUT; | |
ff8519f9 | 752 | set->nr_maps = nctrl->opts->nr_poll_queues ? HCTX_MAX_TYPES : 2; |
b28a308e SG |
753 | } |
754 | ||
755 | ret = blk_mq_alloc_tag_set(set); | |
756 | if (ret) | |
87fd1253 | 757 | return ERR_PTR(ret); |
b28a308e SG |
758 | |
759 | return set; | |
b28a308e SG |
760 | } |
761 | ||
3f02fffb SG |
762 | static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl, |
763 | bool remove) | |
71102307 | 764 | { |
3f02fffb SG |
765 | if (remove) { |
766 | blk_cleanup_queue(ctrl->ctrl.admin_q); | |
e7832cb4 | 767 | blk_cleanup_queue(ctrl->ctrl.fabrics_q); |
87fd1253 | 768 | blk_mq_free_tag_set(ctrl->ctrl.admin_tagset); |
3f02fffb | 769 | } |
682630f0 SG |
770 | if (ctrl->async_event_sqe.data) { |
771 | nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe, | |
772 | sizeof(struct nvme_command), DMA_TO_DEVICE); | |
773 | ctrl->async_event_sqe.data = NULL; | |
774 | } | |
a57bd541 | 775 | nvme_rdma_free_queue(&ctrl->queues[0]); |
71102307 CH |
776 | } |
777 | ||
3f02fffb SG |
778 | static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl, |
779 | bool new) | |
90af3512 SG |
780 | { |
781 | int error; | |
782 | ||
41e8cfa1 | 783 | error = nvme_rdma_alloc_queue(ctrl, 0, NVME_AQ_DEPTH); |
90af3512 SG |
784 | if (error) |
785 | return error; | |
786 | ||
787 | ctrl->device = ctrl->queues[0].device; | |
103e515e | 788 | ctrl->ctrl.numa_node = dev_to_node(ctrl->device->dev->dma_device); |
90af3512 | 789 | |
f41725bb | 790 | ctrl->max_fr_pages = nvme_rdma_get_max_fr_pages(ctrl->device->dev); |
90af3512 | 791 | |
62f99b62 MG |
792 | /* |
793 | * Bind the async event SQE DMA mapping to the admin queue lifetime. | |
794 | * It's safe, since any chage in the underlying RDMA device will issue | |
795 | * error recovery and queue re-creation. | |
796 | */ | |
94e42213 SG |
797 | error = nvme_rdma_alloc_qe(ctrl->device->dev, &ctrl->async_event_sqe, |
798 | sizeof(struct nvme_command), DMA_TO_DEVICE); | |
799 | if (error) | |
800 | goto out_free_queue; | |
801 | ||
3f02fffb SG |
802 | if (new) { |
803 | ctrl->ctrl.admin_tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, true); | |
f04b9cc8 SG |
804 | if (IS_ERR(ctrl->ctrl.admin_tagset)) { |
805 | error = PTR_ERR(ctrl->ctrl.admin_tagset); | |
94e42213 | 806 | goto out_free_async_qe; |
f04b9cc8 | 807 | } |
90af3512 | 808 | |
e7832cb4 SG |
809 | ctrl->ctrl.fabrics_q = blk_mq_init_queue(&ctrl->admin_tag_set); |
810 | if (IS_ERR(ctrl->ctrl.fabrics_q)) { | |
811 | error = PTR_ERR(ctrl->ctrl.fabrics_q); | |
812 | goto out_free_tagset; | |
813 | } | |
814 | ||
3f02fffb SG |
815 | ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set); |
816 | if (IS_ERR(ctrl->ctrl.admin_q)) { | |
817 | error = PTR_ERR(ctrl->ctrl.admin_q); | |
e7832cb4 | 818 | goto out_cleanup_fabrics_q; |
3f02fffb | 819 | } |
90af3512 SG |
820 | } |
821 | ||
68e16fcf | 822 | error = nvme_rdma_start_queue(ctrl, 0); |
90af3512 SG |
823 | if (error) |
824 | goto out_cleanup_queue; | |
825 | ||
c0f2f45b | 826 | error = nvme_enable_ctrl(&ctrl->ctrl); |
90af3512 | 827 | if (error) |
2e050f00 | 828 | goto out_stop_queue; |
90af3512 | 829 | |
ff13c1b8 MG |
830 | ctrl->ctrl.max_segments = ctrl->max_fr_pages; |
831 | ctrl->ctrl.max_hw_sectors = ctrl->max_fr_pages << (ilog2(SZ_4K) - 9); | |
90af3512 | 832 | |
e7832cb4 SG |
833 | blk_mq_unquiesce_queue(ctrl->ctrl.admin_q); |
834 | ||
90af3512 SG |
835 | error = nvme_init_identify(&ctrl->ctrl); |
836 | if (error) | |
2e050f00 | 837 | goto out_stop_queue; |
90af3512 | 838 | |
90af3512 SG |
839 | return 0; |
840 | ||
2e050f00 JW |
841 | out_stop_queue: |
842 | nvme_rdma_stop_queue(&ctrl->queues[0]); | |
90af3512 | 843 | out_cleanup_queue: |
3f02fffb SG |
844 | if (new) |
845 | blk_cleanup_queue(ctrl->ctrl.admin_q); | |
e7832cb4 SG |
846 | out_cleanup_fabrics_q: |
847 | if (new) | |
848 | blk_cleanup_queue(ctrl->ctrl.fabrics_q); | |
90af3512 | 849 | out_free_tagset: |
3f02fffb | 850 | if (new) |
87fd1253 | 851 | blk_mq_free_tag_set(ctrl->ctrl.admin_tagset); |
94e42213 SG |
852 | out_free_async_qe: |
853 | nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe, | |
854 | sizeof(struct nvme_command), DMA_TO_DEVICE); | |
6344d02d | 855 | ctrl->async_event_sqe.data = NULL; |
90af3512 SG |
856 | out_free_queue: |
857 | nvme_rdma_free_queue(&ctrl->queues[0]); | |
858 | return error; | |
859 | } | |
860 | ||
a57bd541 SG |
861 | static void nvme_rdma_destroy_io_queues(struct nvme_rdma_ctrl *ctrl, |
862 | bool remove) | |
863 | { | |
a57bd541 SG |
864 | if (remove) { |
865 | blk_cleanup_queue(ctrl->ctrl.connect_q); | |
87fd1253 | 866 | blk_mq_free_tag_set(ctrl->ctrl.tagset); |
a57bd541 SG |
867 | } |
868 | nvme_rdma_free_io_queues(ctrl); | |
869 | } | |
870 | ||
871 | static int nvme_rdma_configure_io_queues(struct nvme_rdma_ctrl *ctrl, bool new) | |
872 | { | |
873 | int ret; | |
874 | ||
41e8cfa1 | 875 | ret = nvme_rdma_alloc_io_queues(ctrl); |
a57bd541 SG |
876 | if (ret) |
877 | return ret; | |
878 | ||
879 | if (new) { | |
880 | ctrl->ctrl.tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, false); | |
f04b9cc8 SG |
881 | if (IS_ERR(ctrl->ctrl.tagset)) { |
882 | ret = PTR_ERR(ctrl->ctrl.tagset); | |
a57bd541 | 883 | goto out_free_io_queues; |
f04b9cc8 | 884 | } |
a57bd541 SG |
885 | |
886 | ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set); | |
887 | if (IS_ERR(ctrl->ctrl.connect_q)) { | |
888 | ret = PTR_ERR(ctrl->ctrl.connect_q); | |
889 | goto out_free_tag_set; | |
890 | } | |
891 | } else { | |
a57bd541 SG |
892 | blk_mq_update_nr_hw_queues(&ctrl->tag_set, |
893 | ctrl->ctrl.queue_count - 1); | |
894 | } | |
895 | ||
68e16fcf | 896 | ret = nvme_rdma_start_io_queues(ctrl); |
a57bd541 SG |
897 | if (ret) |
898 | goto out_cleanup_connect_q; | |
899 | ||
900 | return 0; | |
901 | ||
902 | out_cleanup_connect_q: | |
903 | if (new) | |
904 | blk_cleanup_queue(ctrl->ctrl.connect_q); | |
905 | out_free_tag_set: | |
906 | if (new) | |
87fd1253 | 907 | blk_mq_free_tag_set(ctrl->ctrl.tagset); |
a57bd541 SG |
908 | out_free_io_queues: |
909 | nvme_rdma_free_io_queues(ctrl); | |
910 | return ret; | |
71102307 CH |
911 | } |
912 | ||
75862c72 SG |
913 | static void nvme_rdma_teardown_admin_queue(struct nvme_rdma_ctrl *ctrl, |
914 | bool remove) | |
915 | { | |
916 | blk_mq_quiesce_queue(ctrl->ctrl.admin_q); | |
917 | nvme_rdma_stop_queue(&ctrl->queues[0]); | |
622b8b68 | 918 | if (ctrl->ctrl.admin_tagset) { |
1007709d SG |
919 | blk_mq_tagset_busy_iter(ctrl->ctrl.admin_tagset, |
920 | nvme_cancel_request, &ctrl->ctrl); | |
622b8b68 ML |
921 | blk_mq_tagset_wait_completed_request(ctrl->ctrl.admin_tagset); |
922 | } | |
e7832cb4 SG |
923 | if (remove) |
924 | blk_mq_unquiesce_queue(ctrl->ctrl.admin_q); | |
75862c72 SG |
925 | nvme_rdma_destroy_admin_queue(ctrl, remove); |
926 | } | |
927 | ||
928 | static void nvme_rdma_teardown_io_queues(struct nvme_rdma_ctrl *ctrl, | |
929 | bool remove) | |
930 | { | |
931 | if (ctrl->ctrl.queue_count > 1) { | |
932 | nvme_stop_queues(&ctrl->ctrl); | |
933 | nvme_rdma_stop_io_queues(ctrl); | |
622b8b68 | 934 | if (ctrl->ctrl.tagset) { |
1007709d SG |
935 | blk_mq_tagset_busy_iter(ctrl->ctrl.tagset, |
936 | nvme_cancel_request, &ctrl->ctrl); | |
622b8b68 ML |
937 | blk_mq_tagset_wait_completed_request(ctrl->ctrl.tagset); |
938 | } | |
75862c72 SG |
939 | if (remove) |
940 | nvme_start_queues(&ctrl->ctrl); | |
941 | nvme_rdma_destroy_io_queues(ctrl, remove); | |
942 | } | |
943 | } | |
944 | ||
71102307 CH |
945 | static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl) |
946 | { | |
947 | struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl); | |
948 | ||
949 | if (list_empty(&ctrl->list)) | |
950 | goto free_ctrl; | |
951 | ||
952 | mutex_lock(&nvme_rdma_ctrl_mutex); | |
953 | list_del(&ctrl->list); | |
954 | mutex_unlock(&nvme_rdma_ctrl_mutex); | |
955 | ||
71102307 CH |
956 | nvmf_free_options(nctrl->opts); |
957 | free_ctrl: | |
3d064101 | 958 | kfree(ctrl->queues); |
71102307 CH |
959 | kfree(ctrl); |
960 | } | |
961 | ||
fd8563ce SG |
962 | static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl *ctrl) |
963 | { | |
964 | /* If we are resetting/deleting then do nothing */ | |
ad6a0a52 | 965 | if (ctrl->ctrl.state != NVME_CTRL_CONNECTING) { |
fd8563ce SG |
966 | WARN_ON_ONCE(ctrl->ctrl.state == NVME_CTRL_NEW || |
967 | ctrl->ctrl.state == NVME_CTRL_LIVE); | |
968 | return; | |
969 | } | |
970 | ||
971 | if (nvmf_should_reconnect(&ctrl->ctrl)) { | |
972 | dev_info(ctrl->ctrl.device, "Reconnecting in %d seconds...\n", | |
973 | ctrl->ctrl.opts->reconnect_delay); | |
9a6327d2 | 974 | queue_delayed_work(nvme_wq, &ctrl->reconnect_work, |
fd8563ce SG |
975 | ctrl->ctrl.opts->reconnect_delay * HZ); |
976 | } else { | |
12fa1304 | 977 | nvme_delete_ctrl(&ctrl->ctrl); |
fd8563ce SG |
978 | } |
979 | } | |
980 | ||
c66e2998 | 981 | static int nvme_rdma_setup_ctrl(struct nvme_rdma_ctrl *ctrl, bool new) |
71102307 | 982 | { |
c66e2998 | 983 | int ret = -EINVAL; |
71102307 | 984 | bool changed; |
71102307 | 985 | |
c66e2998 | 986 | ret = nvme_rdma_configure_admin_queue(ctrl, new); |
71102307 | 987 | if (ret) |
c66e2998 SG |
988 | return ret; |
989 | ||
990 | if (ctrl->ctrl.icdoff) { | |
991 | dev_err(ctrl->ctrl.device, "icdoff is not supported!\n"); | |
992 | goto destroy_admin; | |
993 | } | |
994 | ||
995 | if (!(ctrl->ctrl.sgls & (1 << 2))) { | |
996 | dev_err(ctrl->ctrl.device, | |
997 | "Mandatory keyed sgls are not supported!\n"); | |
998 | goto destroy_admin; | |
999 | } | |
1000 | ||
1001 | if (ctrl->ctrl.opts->queue_size > ctrl->ctrl.sqsize + 1) { | |
1002 | dev_warn(ctrl->ctrl.device, | |
1003 | "queue_size %zu > ctrl sqsize %u, clamping down\n", | |
1004 | ctrl->ctrl.opts->queue_size, ctrl->ctrl.sqsize + 1); | |
1005 | } | |
1006 | ||
1007 | if (ctrl->ctrl.sqsize + 1 > ctrl->ctrl.maxcmd) { | |
1008 | dev_warn(ctrl->ctrl.device, | |
1009 | "sqsize %u > ctrl maxcmd %u, clamping down\n", | |
1010 | ctrl->ctrl.sqsize + 1, ctrl->ctrl.maxcmd); | |
1011 | ctrl->ctrl.sqsize = ctrl->ctrl.maxcmd - 1; | |
1012 | } | |
71102307 | 1013 | |
64a741c1 SW |
1014 | if (ctrl->ctrl.sgls & (1 << 20)) |
1015 | ctrl->use_inline_data = true; | |
71102307 | 1016 | |
d858e5f0 | 1017 | if (ctrl->ctrl.queue_count > 1) { |
c66e2998 | 1018 | ret = nvme_rdma_configure_io_queues(ctrl, new); |
71102307 | 1019 | if (ret) |
5e1fe61d | 1020 | goto destroy_admin; |
71102307 CH |
1021 | } |
1022 | ||
1023 | changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE); | |
0a960afd SG |
1024 | if (!changed) { |
1025 | /* state change failure is ok if we're in DELETING state */ | |
1026 | WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING); | |
c66e2998 SG |
1027 | ret = -EINVAL; |
1028 | goto destroy_io; | |
0a960afd SG |
1029 | } |
1030 | ||
d09f2b45 | 1031 | nvme_start_ctrl(&ctrl->ctrl); |
c66e2998 SG |
1032 | return 0; |
1033 | ||
1034 | destroy_io: | |
1035 | if (ctrl->ctrl.queue_count > 1) | |
1036 | nvme_rdma_destroy_io_queues(ctrl, new); | |
1037 | destroy_admin: | |
1038 | nvme_rdma_stop_queue(&ctrl->queues[0]); | |
1039 | nvme_rdma_destroy_admin_queue(ctrl, new); | |
1040 | return ret; | |
1041 | } | |
1042 | ||
1043 | static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work) | |
1044 | { | |
1045 | struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work), | |
1046 | struct nvme_rdma_ctrl, reconnect_work); | |
1047 | ||
1048 | ++ctrl->ctrl.nr_reconnects; | |
1049 | ||
1050 | if (nvme_rdma_setup_ctrl(ctrl, false)) | |
1051 | goto requeue; | |
71102307 | 1052 | |
5e1fe61d SG |
1053 | dev_info(ctrl->ctrl.device, "Successfully reconnected (%d attempts)\n", |
1054 | ctrl->ctrl.nr_reconnects); | |
1055 | ||
1056 | ctrl->ctrl.nr_reconnects = 0; | |
71102307 CH |
1057 | |
1058 | return; | |
1059 | ||
71102307 | 1060 | requeue: |
fd8563ce | 1061 | dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d\n", |
fdf9dfa8 | 1062 | ctrl->ctrl.nr_reconnects); |
fd8563ce | 1063 | nvme_rdma_reconnect_or_remove(ctrl); |
71102307 CH |
1064 | } |
1065 | ||
1066 | static void nvme_rdma_error_recovery_work(struct work_struct *work) | |
1067 | { | |
1068 | struct nvme_rdma_ctrl *ctrl = container_of(work, | |
1069 | struct nvme_rdma_ctrl, err_work); | |
1070 | ||
e4d753d7 | 1071 | nvme_stop_keep_alive(&ctrl->ctrl); |
75862c72 | 1072 | nvme_rdma_teardown_io_queues(ctrl, false); |
e818a5b4 | 1073 | nvme_start_queues(&ctrl->ctrl); |
75862c72 | 1074 | nvme_rdma_teardown_admin_queue(ctrl, false); |
e7832cb4 | 1075 | blk_mq_unquiesce_queue(ctrl->ctrl.admin_q); |
e818a5b4 | 1076 | |
ad6a0a52 | 1077 | if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) { |
187c0832 NC |
1078 | /* state change failure is ok if we're in DELETING state */ |
1079 | WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING); | |
d5bf4b7f SG |
1080 | return; |
1081 | } | |
1082 | ||
fd8563ce | 1083 | nvme_rdma_reconnect_or_remove(ctrl); |
71102307 CH |
1084 | } |
1085 | ||
1086 | static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl) | |
1087 | { | |
d5bf4b7f | 1088 | if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING)) |
71102307 CH |
1089 | return; |
1090 | ||
9a6327d2 | 1091 | queue_work(nvme_wq, &ctrl->err_work); |
71102307 CH |
1092 | } |
1093 | ||
1094 | static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc, | |
1095 | const char *op) | |
1096 | { | |
1097 | struct nvme_rdma_queue *queue = cq->cq_context; | |
1098 | struct nvme_rdma_ctrl *ctrl = queue->ctrl; | |
1099 | ||
1100 | if (ctrl->ctrl.state == NVME_CTRL_LIVE) | |
1101 | dev_info(ctrl->ctrl.device, | |
1102 | "%s for CQE 0x%p failed with status %s (%d)\n", | |
1103 | op, wc->wr_cqe, | |
1104 | ib_wc_status_msg(wc->status), wc->status); | |
1105 | nvme_rdma_error_recovery(ctrl); | |
1106 | } | |
1107 | ||
1108 | static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc) | |
1109 | { | |
1110 | if (unlikely(wc->status != IB_WC_SUCCESS)) | |
1111 | nvme_rdma_wr_error(cq, wc, "MEMREG"); | |
1112 | } | |
1113 | ||
1114 | static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc) | |
1115 | { | |
2f122e4f SG |
1116 | struct nvme_rdma_request *req = |
1117 | container_of(wc->wr_cqe, struct nvme_rdma_request, reg_cqe); | |
1118 | struct request *rq = blk_mq_rq_from_pdu(req); | |
1119 | ||
1120 | if (unlikely(wc->status != IB_WC_SUCCESS)) { | |
71102307 | 1121 | nvme_rdma_wr_error(cq, wc, "LOCAL_INV"); |
2f122e4f SG |
1122 | return; |
1123 | } | |
1124 | ||
1125 | if (refcount_dec_and_test(&req->ref)) | |
1126 | nvme_end_request(rq, req->status, req->result); | |
1127 | ||
71102307 CH |
1128 | } |
1129 | ||
1130 | static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue, | |
1131 | struct nvme_rdma_request *req) | |
1132 | { | |
71102307 CH |
1133 | struct ib_send_wr wr = { |
1134 | .opcode = IB_WR_LOCAL_INV, | |
1135 | .next = NULL, | |
1136 | .num_sge = 0, | |
2f122e4f | 1137 | .send_flags = IB_SEND_SIGNALED, |
71102307 CH |
1138 | .ex.invalidate_rkey = req->mr->rkey, |
1139 | }; | |
1140 | ||
1141 | req->reg_cqe.done = nvme_rdma_inv_rkey_done; | |
1142 | wr.wr_cqe = &req->reg_cqe; | |
1143 | ||
45e3cc1a | 1144 | return ib_post_send(queue->qp, &wr, NULL); |
71102307 CH |
1145 | } |
1146 | ||
1147 | static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue, | |
1148 | struct request *rq) | |
1149 | { | |
1150 | struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq); | |
71102307 CH |
1151 | struct nvme_rdma_device *dev = queue->device; |
1152 | struct ib_device *ibdev = dev->dev; | |
71102307 | 1153 | |
34e08191 | 1154 | if (!blk_rq_nr_phys_segments(rq)) |
71102307 CH |
1155 | return; |
1156 | ||
f41725bb IR |
1157 | if (req->mr) { |
1158 | ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr); | |
1159 | req->mr = NULL; | |
1160 | } | |
1161 | ||
bc31c1ee | 1162 | ib_dma_unmap_sg(ibdev, req->sg_table.sgl, req->nents, rq_dma_dir(rq)); |
71102307 CH |
1163 | |
1164 | nvme_cleanup_cmd(rq); | |
4635873c | 1165 | sg_free_table_chained(&req->sg_table, SG_CHUNK_SIZE); |
71102307 CH |
1166 | } |
1167 | ||
1168 | static int nvme_rdma_set_sg_null(struct nvme_command *c) | |
1169 | { | |
1170 | struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl; | |
1171 | ||
1172 | sg->addr = 0; | |
1173 | put_unaligned_le24(0, sg->length); | |
1174 | put_unaligned_le32(0, sg->key); | |
1175 | sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4; | |
1176 | return 0; | |
1177 | } | |
1178 | ||
1179 | static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue, | |
64a741c1 SW |
1180 | struct nvme_rdma_request *req, struct nvme_command *c, |
1181 | int count) | |
71102307 CH |
1182 | { |
1183 | struct nvme_sgl_desc *sg = &c->common.dptr.sgl; | |
64a741c1 SW |
1184 | struct scatterlist *sgl = req->sg_table.sgl; |
1185 | struct ib_sge *sge = &req->sge[1]; | |
1186 | u32 len = 0; | |
1187 | int i; | |
71102307 | 1188 | |
64a741c1 SW |
1189 | for (i = 0; i < count; i++, sgl++, sge++) { |
1190 | sge->addr = sg_dma_address(sgl); | |
1191 | sge->length = sg_dma_len(sgl); | |
1192 | sge->lkey = queue->device->pd->local_dma_lkey; | |
1193 | len += sge->length; | |
1194 | } | |
71102307 CH |
1195 | |
1196 | sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff); | |
64a741c1 | 1197 | sg->length = cpu_to_le32(len); |
71102307 CH |
1198 | sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET; |
1199 | ||
64a741c1 | 1200 | req->num_sge += count; |
71102307 CH |
1201 | return 0; |
1202 | } | |
1203 | ||
1204 | static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue, | |
1205 | struct nvme_rdma_request *req, struct nvme_command *c) | |
1206 | { | |
1207 | struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl; | |
1208 | ||
1209 | sg->addr = cpu_to_le64(sg_dma_address(req->sg_table.sgl)); | |
1210 | put_unaligned_le24(sg_dma_len(req->sg_table.sgl), sg->length); | |
11975e01 | 1211 | put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key); |
71102307 CH |
1212 | sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4; |
1213 | return 0; | |
1214 | } | |
1215 | ||
1216 | static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue, | |
1217 | struct nvme_rdma_request *req, struct nvme_command *c, | |
1218 | int count) | |
1219 | { | |
1220 | struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl; | |
1221 | int nr; | |
1222 | ||
f41725bb IR |
1223 | req->mr = ib_mr_pool_get(queue->qp, &queue->qp->rdma_mrs); |
1224 | if (WARN_ON_ONCE(!req->mr)) | |
1225 | return -EAGAIN; | |
1226 | ||
b925a2dc MG |
1227 | /* |
1228 | * Align the MR to a 4K page size to match the ctrl page size and | |
1229 | * the block virtual boundary. | |
1230 | */ | |
1231 | nr = ib_map_mr_sg(req->mr, req->sg_table.sgl, count, NULL, SZ_4K); | |
a7b7c7a1 | 1232 | if (unlikely(nr < count)) { |
f41725bb IR |
1233 | ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr); |
1234 | req->mr = NULL; | |
71102307 CH |
1235 | if (nr < 0) |
1236 | return nr; | |
1237 | return -EINVAL; | |
1238 | } | |
1239 | ||
1240 | ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey)); | |
1241 | ||
1242 | req->reg_cqe.done = nvme_rdma_memreg_done; | |
1243 | memset(&req->reg_wr, 0, sizeof(req->reg_wr)); | |
1244 | req->reg_wr.wr.opcode = IB_WR_REG_MR; | |
1245 | req->reg_wr.wr.wr_cqe = &req->reg_cqe; | |
1246 | req->reg_wr.wr.num_sge = 0; | |
1247 | req->reg_wr.mr = req->mr; | |
1248 | req->reg_wr.key = req->mr->rkey; | |
1249 | req->reg_wr.access = IB_ACCESS_LOCAL_WRITE | | |
1250 | IB_ACCESS_REMOTE_READ | | |
1251 | IB_ACCESS_REMOTE_WRITE; | |
1252 | ||
71102307 CH |
1253 | sg->addr = cpu_to_le64(req->mr->iova); |
1254 | put_unaligned_le24(req->mr->length, sg->length); | |
1255 | put_unaligned_le32(req->mr->rkey, sg->key); | |
1256 | sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) | | |
1257 | NVME_SGL_FMT_INVALIDATE; | |
1258 | ||
1259 | return 0; | |
1260 | } | |
1261 | ||
1262 | static int nvme_rdma_map_data(struct nvme_rdma_queue *queue, | |
b131c61d | 1263 | struct request *rq, struct nvme_command *c) |
71102307 CH |
1264 | { |
1265 | struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq); | |
1266 | struct nvme_rdma_device *dev = queue->device; | |
1267 | struct ib_device *ibdev = dev->dev; | |
f9d03f96 | 1268 | int count, ret; |
71102307 CH |
1269 | |
1270 | req->num_sge = 1; | |
4af7f7ff | 1271 | refcount_set(&req->ref, 2); /* send and recv completions */ |
71102307 CH |
1272 | |
1273 | c->common.flags |= NVME_CMD_SGL_METABUF; | |
1274 | ||
34e08191 | 1275 | if (!blk_rq_nr_phys_segments(rq)) |
71102307 CH |
1276 | return nvme_rdma_set_sg_null(c); |
1277 | ||
1278 | req->sg_table.sgl = req->first_sgl; | |
f9d03f96 | 1279 | ret = sg_alloc_table_chained(&req->sg_table, |
4635873c ML |
1280 | blk_rq_nr_phys_segments(rq), req->sg_table.sgl, |
1281 | SG_CHUNK_SIZE); | |
71102307 CH |
1282 | if (ret) |
1283 | return -ENOMEM; | |
1284 | ||
f9d03f96 | 1285 | req->nents = blk_rq_map_sg(rq->q, rq, req->sg_table.sgl); |
71102307 | 1286 | |
f9d03f96 | 1287 | count = ib_dma_map_sg(ibdev, req->sg_table.sgl, req->nents, |
bc31c1ee | 1288 | rq_dma_dir(rq)); |
71102307 | 1289 | if (unlikely(count <= 0)) { |
94423a8f MG |
1290 | ret = -EIO; |
1291 | goto out_free_table; | |
71102307 CH |
1292 | } |
1293 | ||
64a741c1 | 1294 | if (count <= dev->num_inline_segments) { |
b131c61d | 1295 | if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) && |
64a741c1 | 1296 | queue->ctrl->use_inline_data && |
b131c61d | 1297 | blk_rq_payload_bytes(rq) <= |
94423a8f | 1298 | nvme_rdma_inline_data_size(queue)) { |
64a741c1 | 1299 | ret = nvme_rdma_map_sg_inline(queue, req, c, count); |
94423a8f MG |
1300 | goto out; |
1301 | } | |
71102307 | 1302 | |
64a741c1 | 1303 | if (count == 1 && dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY) { |
94423a8f MG |
1304 | ret = nvme_rdma_map_sg_single(queue, req, c); |
1305 | goto out; | |
1306 | } | |
71102307 CH |
1307 | } |
1308 | ||
94423a8f MG |
1309 | ret = nvme_rdma_map_sg_fr(queue, req, c, count); |
1310 | out: | |
1311 | if (unlikely(ret)) | |
1312 | goto out_unmap_sg; | |
1313 | ||
1314 | return 0; | |
1315 | ||
1316 | out_unmap_sg: | |
bc31c1ee | 1317 | ib_dma_unmap_sg(ibdev, req->sg_table.sgl, req->nents, rq_dma_dir(rq)); |
94423a8f | 1318 | out_free_table: |
4635873c | 1319 | sg_free_table_chained(&req->sg_table, SG_CHUNK_SIZE); |
94423a8f | 1320 | return ret; |
71102307 CH |
1321 | } |
1322 | ||
1323 | static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc) | |
1324 | { | |
4af7f7ff SG |
1325 | struct nvme_rdma_qe *qe = |
1326 | container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe); | |
1327 | struct nvme_rdma_request *req = | |
1328 | container_of(qe, struct nvme_rdma_request, sqe); | |
1329 | struct request *rq = blk_mq_rq_from_pdu(req); | |
1330 | ||
1331 | if (unlikely(wc->status != IB_WC_SUCCESS)) { | |
71102307 | 1332 | nvme_rdma_wr_error(cq, wc, "SEND"); |
4af7f7ff SG |
1333 | return; |
1334 | } | |
1335 | ||
1336 | if (refcount_dec_and_test(&req->ref)) | |
1337 | nvme_end_request(rq, req->status, req->result); | |
71102307 CH |
1338 | } |
1339 | ||
1340 | static int nvme_rdma_post_send(struct nvme_rdma_queue *queue, | |
1341 | struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge, | |
b4b591c8 | 1342 | struct ib_send_wr *first) |
71102307 | 1343 | { |
45e3cc1a | 1344 | struct ib_send_wr wr; |
71102307 CH |
1345 | int ret; |
1346 | ||
1347 | sge->addr = qe->dma; | |
1348 | sge->length = sizeof(struct nvme_command), | |
1349 | sge->lkey = queue->device->pd->local_dma_lkey; | |
1350 | ||
71102307 CH |
1351 | wr.next = NULL; |
1352 | wr.wr_cqe = &qe->cqe; | |
1353 | wr.sg_list = sge; | |
1354 | wr.num_sge = num_sge; | |
1355 | wr.opcode = IB_WR_SEND; | |
b4b591c8 | 1356 | wr.send_flags = IB_SEND_SIGNALED; |
71102307 CH |
1357 | |
1358 | if (first) | |
1359 | first->next = ≀ | |
1360 | else | |
1361 | first = ≀ | |
1362 | ||
45e3cc1a | 1363 | ret = ib_post_send(queue->qp, first, NULL); |
a7b7c7a1 | 1364 | if (unlikely(ret)) { |
71102307 CH |
1365 | dev_err(queue->ctrl->ctrl.device, |
1366 | "%s failed with error code %d\n", __func__, ret); | |
1367 | } | |
1368 | return ret; | |
1369 | } | |
1370 | ||
1371 | static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue, | |
1372 | struct nvme_rdma_qe *qe) | |
1373 | { | |
45e3cc1a | 1374 | struct ib_recv_wr wr; |
71102307 CH |
1375 | struct ib_sge list; |
1376 | int ret; | |
1377 | ||
1378 | list.addr = qe->dma; | |
1379 | list.length = sizeof(struct nvme_completion); | |
1380 | list.lkey = queue->device->pd->local_dma_lkey; | |
1381 | ||
1382 | qe->cqe.done = nvme_rdma_recv_done; | |
1383 | ||
1384 | wr.next = NULL; | |
1385 | wr.wr_cqe = &qe->cqe; | |
1386 | wr.sg_list = &list; | |
1387 | wr.num_sge = 1; | |
1388 | ||
45e3cc1a | 1389 | ret = ib_post_recv(queue->qp, &wr, NULL); |
a7b7c7a1 | 1390 | if (unlikely(ret)) { |
71102307 CH |
1391 | dev_err(queue->ctrl->ctrl.device, |
1392 | "%s failed with error code %d\n", __func__, ret); | |
1393 | } | |
1394 | return ret; | |
1395 | } | |
1396 | ||
1397 | static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue) | |
1398 | { | |
1399 | u32 queue_idx = nvme_rdma_queue_idx(queue); | |
1400 | ||
1401 | if (queue_idx == 0) | |
1402 | return queue->ctrl->admin_tag_set.tags[queue_idx]; | |
1403 | return queue->ctrl->tag_set.tags[queue_idx - 1]; | |
1404 | } | |
1405 | ||
b4b591c8 SG |
1406 | static void nvme_rdma_async_done(struct ib_cq *cq, struct ib_wc *wc) |
1407 | { | |
1408 | if (unlikely(wc->status != IB_WC_SUCCESS)) | |
1409 | nvme_rdma_wr_error(cq, wc, "ASYNC"); | |
1410 | } | |
1411 | ||
ad22c355 | 1412 | static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg) |
71102307 CH |
1413 | { |
1414 | struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg); | |
1415 | struct nvme_rdma_queue *queue = &ctrl->queues[0]; | |
1416 | struct ib_device *dev = queue->device->dev; | |
1417 | struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe; | |
1418 | struct nvme_command *cmd = sqe->data; | |
1419 | struct ib_sge sge; | |
1420 | int ret; | |
1421 | ||
71102307 CH |
1422 | ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE); |
1423 | ||
1424 | memset(cmd, 0, sizeof(*cmd)); | |
1425 | cmd->common.opcode = nvme_admin_async_event; | |
38dabe21 | 1426 | cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH; |
71102307 CH |
1427 | cmd->common.flags |= NVME_CMD_SGL_METABUF; |
1428 | nvme_rdma_set_sg_null(cmd); | |
1429 | ||
b4b591c8 SG |
1430 | sqe->cqe.done = nvme_rdma_async_done; |
1431 | ||
71102307 CH |
1432 | ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd), |
1433 | DMA_TO_DEVICE); | |
1434 | ||
b4b591c8 | 1435 | ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL); |
71102307 CH |
1436 | WARN_ON_ONCE(ret); |
1437 | } | |
1438 | ||
1052b8ac JA |
1439 | static void nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue, |
1440 | struct nvme_completion *cqe, struct ib_wc *wc) | |
71102307 | 1441 | { |
71102307 CH |
1442 | struct request *rq; |
1443 | struct nvme_rdma_request *req; | |
71102307 | 1444 | |
71102307 CH |
1445 | rq = blk_mq_tag_to_rq(nvme_rdma_tagset(queue), cqe->command_id); |
1446 | if (!rq) { | |
1447 | dev_err(queue->ctrl->ctrl.device, | |
1448 | "tag 0x%x on QP %#x not found\n", | |
1449 | cqe->command_id, queue->qp->qp_num); | |
1450 | nvme_rdma_error_recovery(queue->ctrl); | |
1052b8ac | 1451 | return; |
71102307 CH |
1452 | } |
1453 | req = blk_mq_rq_to_pdu(rq); | |
1454 | ||
4af7f7ff SG |
1455 | req->status = cqe->status; |
1456 | req->result = cqe->result; | |
71102307 | 1457 | |
3ef0279b SG |
1458 | if (wc->wc_flags & IB_WC_WITH_INVALIDATE) { |
1459 | if (unlikely(wc->ex.invalidate_rkey != req->mr->rkey)) { | |
1460 | dev_err(queue->ctrl->ctrl.device, | |
1461 | "Bogus remote invalidation for rkey %#x\n", | |
1462 | req->mr->rkey); | |
1463 | nvme_rdma_error_recovery(queue->ctrl); | |
1464 | } | |
f41725bb | 1465 | } else if (req->mr) { |
1052b8ac JA |
1466 | int ret; |
1467 | ||
2f122e4f SG |
1468 | ret = nvme_rdma_inv_rkey(queue, req); |
1469 | if (unlikely(ret < 0)) { | |
1470 | dev_err(queue->ctrl->ctrl.device, | |
1471 | "Queueing INV WR for rkey %#x failed (%d)\n", | |
1472 | req->mr->rkey, ret); | |
1473 | nvme_rdma_error_recovery(queue->ctrl); | |
1474 | } | |
1475 | /* the local invalidation completion will end the request */ | |
1052b8ac | 1476 | return; |
2f122e4f | 1477 | } |
71102307 | 1478 | |
1052b8ac | 1479 | if (refcount_dec_and_test(&req->ref)) |
4af7f7ff | 1480 | nvme_end_request(rq, req->status, req->result); |
71102307 CH |
1481 | } |
1482 | ||
1052b8ac | 1483 | static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc) |
71102307 CH |
1484 | { |
1485 | struct nvme_rdma_qe *qe = | |
1486 | container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe); | |
1487 | struct nvme_rdma_queue *queue = cq->cq_context; | |
1488 | struct ib_device *ibdev = queue->device->dev; | |
1489 | struct nvme_completion *cqe = qe->data; | |
1490 | const size_t len = sizeof(struct nvme_completion); | |
71102307 CH |
1491 | |
1492 | if (unlikely(wc->status != IB_WC_SUCCESS)) { | |
1493 | nvme_rdma_wr_error(cq, wc, "RECV"); | |
1052b8ac | 1494 | return; |
71102307 CH |
1495 | } |
1496 | ||
1497 | ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE); | |
1498 | /* | |
1499 | * AEN requests are special as they don't time out and can | |
1500 | * survive any kind of queue freeze and often don't respond to | |
1501 | * aborts. We don't even bother to allocate a struct request | |
1502 | * for them but rather special case them here. | |
1503 | */ | |
1504 | if (unlikely(nvme_rdma_queue_idx(queue) == 0 && | |
38dabe21 | 1505 | cqe->command_id >= NVME_AQ_BLK_MQ_DEPTH)) |
7bf58533 CH |
1506 | nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status, |
1507 | &cqe->result); | |
71102307 | 1508 | else |
1052b8ac | 1509 | nvme_rdma_process_nvme_rsp(queue, cqe, wc); |
71102307 CH |
1510 | ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE); |
1511 | ||
1512 | nvme_rdma_post_recv(queue, qe); | |
71102307 CH |
1513 | } |
1514 | ||
1515 | static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue) | |
1516 | { | |
1517 | int ret, i; | |
1518 | ||
1519 | for (i = 0; i < queue->queue_size; i++) { | |
1520 | ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]); | |
1521 | if (ret) | |
1522 | goto out_destroy_queue_ib; | |
1523 | } | |
1524 | ||
1525 | return 0; | |
1526 | ||
1527 | out_destroy_queue_ib: | |
1528 | nvme_rdma_destroy_queue_ib(queue); | |
1529 | return ret; | |
1530 | } | |
1531 | ||
1532 | static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue, | |
1533 | struct rdma_cm_event *ev) | |
1534 | { | |
7f03953c SW |
1535 | struct rdma_cm_id *cm_id = queue->cm_id; |
1536 | int status = ev->status; | |
1537 | const char *rej_msg; | |
1538 | const struct nvme_rdma_cm_rej *rej_data; | |
1539 | u8 rej_data_len; | |
1540 | ||
1541 | rej_msg = rdma_reject_msg(cm_id, status); | |
1542 | rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len); | |
1543 | ||
1544 | if (rej_data && rej_data_len >= sizeof(u16)) { | |
1545 | u16 sts = le16_to_cpu(rej_data->sts); | |
71102307 CH |
1546 | |
1547 | dev_err(queue->ctrl->ctrl.device, | |
7f03953c SW |
1548 | "Connect rejected: status %d (%s) nvme status %d (%s).\n", |
1549 | status, rej_msg, sts, nvme_rdma_cm_msg(sts)); | |
71102307 CH |
1550 | } else { |
1551 | dev_err(queue->ctrl->ctrl.device, | |
7f03953c | 1552 | "Connect rejected: status %d (%s).\n", status, rej_msg); |
71102307 CH |
1553 | } |
1554 | ||
1555 | return -ECONNRESET; | |
1556 | } | |
1557 | ||
1558 | static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue) | |
1559 | { | |
e63440d6 | 1560 | struct nvme_ctrl *ctrl = &queue->ctrl->ctrl; |
71102307 CH |
1561 | int ret; |
1562 | ||
ca6e95bb SG |
1563 | ret = nvme_rdma_create_queue_ib(queue); |
1564 | if (ret) | |
1565 | return ret; | |
71102307 | 1566 | |
e63440d6 IR |
1567 | if (ctrl->opts->tos >= 0) |
1568 | rdma_set_service_type(queue->cm_id, ctrl->opts->tos); | |
71102307 CH |
1569 | ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS); |
1570 | if (ret) { | |
e63440d6 | 1571 | dev_err(ctrl->device, "rdma_resolve_route failed (%d).\n", |
71102307 CH |
1572 | queue->cm_error); |
1573 | goto out_destroy_queue; | |
1574 | } | |
1575 | ||
1576 | return 0; | |
1577 | ||
1578 | out_destroy_queue: | |
1579 | nvme_rdma_destroy_queue_ib(queue); | |
71102307 CH |
1580 | return ret; |
1581 | } | |
1582 | ||
1583 | static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue) | |
1584 | { | |
1585 | struct nvme_rdma_ctrl *ctrl = queue->ctrl; | |
1586 | struct rdma_conn_param param = { }; | |
0b857b44 | 1587 | struct nvme_rdma_cm_req priv = { }; |
71102307 CH |
1588 | int ret; |
1589 | ||
1590 | param.qp_num = queue->qp->qp_num; | |
1591 | param.flow_control = 1; | |
1592 | ||
1593 | param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom; | |
2ac17c28 SG |
1594 | /* maximum retry count */ |
1595 | param.retry_count = 7; | |
71102307 CH |
1596 | param.rnr_retry_count = 7; |
1597 | param.private_data = &priv; | |
1598 | param.private_data_len = sizeof(priv); | |
1599 | ||
1600 | priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0); | |
1601 | priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue)); | |
f994d9dc JF |
1602 | /* |
1603 | * set the admin queue depth to the minimum size | |
1604 | * specified by the Fabrics standard. | |
1605 | */ | |
1606 | if (priv.qid == 0) { | |
7aa1f427 SG |
1607 | priv.hrqsize = cpu_to_le16(NVME_AQ_DEPTH); |
1608 | priv.hsqsize = cpu_to_le16(NVME_AQ_DEPTH - 1); | |
f994d9dc | 1609 | } else { |
c5af8654 JF |
1610 | /* |
1611 | * current interpretation of the fabrics spec | |
1612 | * is at minimum you make hrqsize sqsize+1, or a | |
1613 | * 1's based representation of sqsize. | |
1614 | */ | |
f994d9dc | 1615 | priv.hrqsize = cpu_to_le16(queue->queue_size); |
c5af8654 | 1616 | priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize); |
f994d9dc | 1617 | } |
71102307 CH |
1618 | |
1619 | ret = rdma_connect(queue->cm_id, ¶m); | |
1620 | if (ret) { | |
1621 | dev_err(ctrl->ctrl.device, | |
1622 | "rdma_connect failed (%d).\n", ret); | |
1623 | goto out_destroy_queue_ib; | |
1624 | } | |
1625 | ||
1626 | return 0; | |
1627 | ||
1628 | out_destroy_queue_ib: | |
1629 | nvme_rdma_destroy_queue_ib(queue); | |
1630 | return ret; | |
1631 | } | |
1632 | ||
71102307 CH |
1633 | static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id, |
1634 | struct rdma_cm_event *ev) | |
1635 | { | |
1636 | struct nvme_rdma_queue *queue = cm_id->context; | |
1637 | int cm_error = 0; | |
1638 | ||
1639 | dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n", | |
1640 | rdma_event_msg(ev->event), ev->event, | |
1641 | ev->status, cm_id); | |
1642 | ||
1643 | switch (ev->event) { | |
1644 | case RDMA_CM_EVENT_ADDR_RESOLVED: | |
1645 | cm_error = nvme_rdma_addr_resolved(queue); | |
1646 | break; | |
1647 | case RDMA_CM_EVENT_ROUTE_RESOLVED: | |
1648 | cm_error = nvme_rdma_route_resolved(queue); | |
1649 | break; | |
1650 | case RDMA_CM_EVENT_ESTABLISHED: | |
1651 | queue->cm_error = nvme_rdma_conn_established(queue); | |
1652 | /* complete cm_done regardless of success/failure */ | |
1653 | complete(&queue->cm_done); | |
1654 | return 0; | |
1655 | case RDMA_CM_EVENT_REJECTED: | |
abf87d5e | 1656 | nvme_rdma_destroy_queue_ib(queue); |
71102307 CH |
1657 | cm_error = nvme_rdma_conn_rejected(queue, ev); |
1658 | break; | |
71102307 CH |
1659 | case RDMA_CM_EVENT_ROUTE_ERROR: |
1660 | case RDMA_CM_EVENT_CONNECT_ERROR: | |
1661 | case RDMA_CM_EVENT_UNREACHABLE: | |
abf87d5e | 1662 | nvme_rdma_destroy_queue_ib(queue); |
249090f9 | 1663 | /* fall through */ |
abf87d5e | 1664 | case RDMA_CM_EVENT_ADDR_ERROR: |
71102307 CH |
1665 | dev_dbg(queue->ctrl->ctrl.device, |
1666 | "CM error event %d\n", ev->event); | |
1667 | cm_error = -ECONNRESET; | |
1668 | break; | |
1669 | case RDMA_CM_EVENT_DISCONNECTED: | |
1670 | case RDMA_CM_EVENT_ADDR_CHANGE: | |
1671 | case RDMA_CM_EVENT_TIMEWAIT_EXIT: | |
1672 | dev_dbg(queue->ctrl->ctrl.device, | |
1673 | "disconnect received - connection closed\n"); | |
1674 | nvme_rdma_error_recovery(queue->ctrl); | |
1675 | break; | |
1676 | case RDMA_CM_EVENT_DEVICE_REMOVAL: | |
e87a911f SW |
1677 | /* device removal is handled via the ib_client API */ |
1678 | break; | |
71102307 CH |
1679 | default: |
1680 | dev_err(queue->ctrl->ctrl.device, | |
1681 | "Unexpected RDMA CM event (%d)\n", ev->event); | |
1682 | nvme_rdma_error_recovery(queue->ctrl); | |
1683 | break; | |
1684 | } | |
1685 | ||
1686 | if (cm_error) { | |
1687 | queue->cm_error = cm_error; | |
1688 | complete(&queue->cm_done); | |
1689 | } | |
1690 | ||
1691 | return 0; | |
1692 | } | |
1693 | ||
1694 | static enum blk_eh_timer_return | |
1695 | nvme_rdma_timeout(struct request *rq, bool reserved) | |
1696 | { | |
1697 | struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq); | |
4c174e63 SG |
1698 | struct nvme_rdma_queue *queue = req->queue; |
1699 | struct nvme_rdma_ctrl *ctrl = queue->ctrl; | |
71102307 | 1700 | |
4c174e63 SG |
1701 | dev_warn(ctrl->ctrl.device, "I/O %d QID %d timeout\n", |
1702 | rq->tag, nvme_rdma_queue_idx(queue)); | |
e62a538d | 1703 | |
4c174e63 SG |
1704 | if (ctrl->ctrl.state != NVME_CTRL_LIVE) { |
1705 | /* | |
1706 | * Teardown immediately if controller times out while starting | |
1707 | * or we are already started error recovery. all outstanding | |
1708 | * requests are completed on shutdown, so we return BLK_EH_DONE. | |
1709 | */ | |
1710 | flush_work(&ctrl->err_work); | |
1711 | nvme_rdma_teardown_io_queues(ctrl, false); | |
1712 | nvme_rdma_teardown_admin_queue(ctrl, false); | |
1713 | return BLK_EH_DONE; | |
1714 | } | |
71102307 | 1715 | |
4c174e63 SG |
1716 | dev_warn(ctrl->ctrl.device, "starting error recovery\n"); |
1717 | nvme_rdma_error_recovery(ctrl); | |
71102307 | 1718 | |
4c174e63 | 1719 | return BLK_EH_RESET_TIMER; |
71102307 CH |
1720 | } |
1721 | ||
fc17b653 | 1722 | static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx, |
71102307 CH |
1723 | const struct blk_mq_queue_data *bd) |
1724 | { | |
1725 | struct nvme_ns *ns = hctx->queue->queuedata; | |
1726 | struct nvme_rdma_queue *queue = hctx->driver_data; | |
1727 | struct request *rq = bd->rq; | |
1728 | struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq); | |
1729 | struct nvme_rdma_qe *sqe = &req->sqe; | |
1730 | struct nvme_command *c = sqe->data; | |
71102307 | 1731 | struct ib_device *dev; |
3bc32bb1 | 1732 | bool queue_ready = test_bit(NVME_RDMA_Q_LIVE, &queue->flags); |
fc17b653 CH |
1733 | blk_status_t ret; |
1734 | int err; | |
71102307 CH |
1735 | |
1736 | WARN_ON_ONCE(rq->tag < 0); | |
1737 | ||
3bc32bb1 | 1738 | if (!nvmf_check_ready(&queue->ctrl->ctrl, rq, queue_ready)) |
6cdefc6e | 1739 | return nvmf_fail_nonready_command(&queue->ctrl->ctrl, rq); |
553cd9ef | 1740 | |
71102307 | 1741 | dev = queue->device->dev; |
62f99b62 MG |
1742 | |
1743 | req->sqe.dma = ib_dma_map_single(dev, req->sqe.data, | |
1744 | sizeof(struct nvme_command), | |
1745 | DMA_TO_DEVICE); | |
1746 | err = ib_dma_mapping_error(dev, req->sqe.dma); | |
1747 | if (unlikely(err)) | |
1748 | return BLK_STS_RESOURCE; | |
1749 | ||
71102307 CH |
1750 | ib_dma_sync_single_for_cpu(dev, sqe->dma, |
1751 | sizeof(struct nvme_command), DMA_TO_DEVICE); | |
1752 | ||
1753 | ret = nvme_setup_cmd(ns, rq, c); | |
fc17b653 | 1754 | if (ret) |
62f99b62 | 1755 | goto unmap_qe; |
71102307 | 1756 | |
71102307 CH |
1757 | blk_mq_start_request(rq); |
1758 | ||
fc17b653 | 1759 | err = nvme_rdma_map_data(queue, rq, c); |
a7b7c7a1 | 1760 | if (unlikely(err < 0)) { |
71102307 | 1761 | dev_err(queue->ctrl->ctrl.device, |
fc17b653 | 1762 | "Failed to map data (%d)\n", err); |
71102307 CH |
1763 | nvme_cleanup_cmd(rq); |
1764 | goto err; | |
1765 | } | |
1766 | ||
b4b591c8 SG |
1767 | sqe->cqe.done = nvme_rdma_send_done; |
1768 | ||
71102307 CH |
1769 | ib_dma_sync_single_for_device(dev, sqe->dma, |
1770 | sizeof(struct nvme_command), DMA_TO_DEVICE); | |
1771 | ||
fc17b653 | 1772 | err = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge, |
f41725bb | 1773 | req->mr ? &req->reg_wr.wr : NULL); |
a7b7c7a1 | 1774 | if (unlikely(err)) { |
71102307 CH |
1775 | nvme_rdma_unmap_data(queue, rq); |
1776 | goto err; | |
1777 | } | |
1778 | ||
fc17b653 | 1779 | return BLK_STS_OK; |
62f99b62 | 1780 | |
71102307 | 1781 | err: |
fc17b653 | 1782 | if (err == -ENOMEM || err == -EAGAIN) |
62f99b62 MG |
1783 | ret = BLK_STS_RESOURCE; |
1784 | else | |
1785 | ret = BLK_STS_IOERR; | |
1786 | unmap_qe: | |
1787 | ib_dma_unmap_single(dev, req->sqe.dma, sizeof(struct nvme_command), | |
1788 | DMA_TO_DEVICE); | |
1789 | return ret; | |
71102307 CH |
1790 | } |
1791 | ||
ff8519f9 SG |
1792 | static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx) |
1793 | { | |
1794 | struct nvme_rdma_queue *queue = hctx->driver_data; | |
1795 | ||
1796 | return ib_process_cq_direct(queue->ib_cq, -1); | |
1797 | } | |
1798 | ||
71102307 CH |
1799 | static void nvme_rdma_complete_rq(struct request *rq) |
1800 | { | |
1801 | struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq); | |
62f99b62 MG |
1802 | struct nvme_rdma_queue *queue = req->queue; |
1803 | struct ib_device *ibdev = queue->device->dev; | |
71102307 | 1804 | |
62f99b62 MG |
1805 | nvme_rdma_unmap_data(queue, rq); |
1806 | ib_dma_unmap_single(ibdev, req->sqe.dma, sizeof(struct nvme_command), | |
1807 | DMA_TO_DEVICE); | |
77f02a7a | 1808 | nvme_complete_rq(rq); |
71102307 CH |
1809 | } |
1810 | ||
0b36658c SG |
1811 | static int nvme_rdma_map_queues(struct blk_mq_tag_set *set) |
1812 | { | |
1813 | struct nvme_rdma_ctrl *ctrl = set->driver_data; | |
5651cd3c | 1814 | struct nvmf_ctrl_options *opts = ctrl->ctrl.opts; |
0b36658c | 1815 | |
5651cd3c | 1816 | if (opts->nr_write_queues && ctrl->io_queues[HCTX_TYPE_READ]) { |
b65bb777 | 1817 | /* separate read/write queues */ |
5651cd3c SG |
1818 | set->map[HCTX_TYPE_DEFAULT].nr_queues = |
1819 | ctrl->io_queues[HCTX_TYPE_DEFAULT]; | |
1820 | set->map[HCTX_TYPE_DEFAULT].queue_offset = 0; | |
1821 | set->map[HCTX_TYPE_READ].nr_queues = | |
1822 | ctrl->io_queues[HCTX_TYPE_READ]; | |
b65bb777 | 1823 | set->map[HCTX_TYPE_READ].queue_offset = |
5651cd3c | 1824 | ctrl->io_queues[HCTX_TYPE_DEFAULT]; |
b65bb777 | 1825 | } else { |
5651cd3c SG |
1826 | /* shared read/write queues */ |
1827 | set->map[HCTX_TYPE_DEFAULT].nr_queues = | |
1828 | ctrl->io_queues[HCTX_TYPE_DEFAULT]; | |
1829 | set->map[HCTX_TYPE_DEFAULT].queue_offset = 0; | |
1830 | set->map[HCTX_TYPE_READ].nr_queues = | |
1831 | ctrl->io_queues[HCTX_TYPE_DEFAULT]; | |
b65bb777 SG |
1832 | set->map[HCTX_TYPE_READ].queue_offset = 0; |
1833 | } | |
1834 | blk_mq_rdma_map_queues(&set->map[HCTX_TYPE_DEFAULT], | |
1835 | ctrl->device->dev, 0); | |
1836 | blk_mq_rdma_map_queues(&set->map[HCTX_TYPE_READ], | |
1837 | ctrl->device->dev, 0); | |
ff8519f9 | 1838 | |
5651cd3c SG |
1839 | if (opts->nr_poll_queues && ctrl->io_queues[HCTX_TYPE_POLL]) { |
1840 | /* map dedicated poll queues only if we have queues left */ | |
ff8519f9 | 1841 | set->map[HCTX_TYPE_POLL].nr_queues = |
b1064d3e | 1842 | ctrl->io_queues[HCTX_TYPE_POLL]; |
ff8519f9 | 1843 | set->map[HCTX_TYPE_POLL].queue_offset = |
5651cd3c SG |
1844 | ctrl->io_queues[HCTX_TYPE_DEFAULT] + |
1845 | ctrl->io_queues[HCTX_TYPE_READ]; | |
ff8519f9 SG |
1846 | blk_mq_map_queues(&set->map[HCTX_TYPE_POLL]); |
1847 | } | |
5651cd3c SG |
1848 | |
1849 | dev_info(ctrl->ctrl.device, | |
1850 | "mapped %d/%d/%d default/read/poll queues.\n", | |
1851 | ctrl->io_queues[HCTX_TYPE_DEFAULT], | |
1852 | ctrl->io_queues[HCTX_TYPE_READ], | |
1853 | ctrl->io_queues[HCTX_TYPE_POLL]); | |
1854 | ||
b65bb777 | 1855 | return 0; |
0b36658c SG |
1856 | } |
1857 | ||
f363b089 | 1858 | static const struct blk_mq_ops nvme_rdma_mq_ops = { |
71102307 CH |
1859 | .queue_rq = nvme_rdma_queue_rq, |
1860 | .complete = nvme_rdma_complete_rq, | |
71102307 CH |
1861 | .init_request = nvme_rdma_init_request, |
1862 | .exit_request = nvme_rdma_exit_request, | |
71102307 | 1863 | .init_hctx = nvme_rdma_init_hctx, |
71102307 | 1864 | .timeout = nvme_rdma_timeout, |
0b36658c | 1865 | .map_queues = nvme_rdma_map_queues, |
ff8519f9 | 1866 | .poll = nvme_rdma_poll, |
71102307 CH |
1867 | }; |
1868 | ||
f363b089 | 1869 | static const struct blk_mq_ops nvme_rdma_admin_mq_ops = { |
71102307 CH |
1870 | .queue_rq = nvme_rdma_queue_rq, |
1871 | .complete = nvme_rdma_complete_rq, | |
385475ee CH |
1872 | .init_request = nvme_rdma_init_request, |
1873 | .exit_request = nvme_rdma_exit_request, | |
71102307 CH |
1874 | .init_hctx = nvme_rdma_init_admin_hctx, |
1875 | .timeout = nvme_rdma_timeout, | |
1876 | }; | |
1877 | ||
18398af2 | 1878 | static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown) |
71102307 | 1879 | { |
794a4cb3 SG |
1880 | cancel_work_sync(&ctrl->err_work); |
1881 | cancel_delayed_work_sync(&ctrl->reconnect_work); | |
1882 | ||
75862c72 | 1883 | nvme_rdma_teardown_io_queues(ctrl, shutdown); |
e7832cb4 | 1884 | blk_mq_quiesce_queue(ctrl->ctrl.admin_q); |
18398af2 | 1885 | if (shutdown) |
71102307 | 1886 | nvme_shutdown_ctrl(&ctrl->ctrl); |
18398af2 | 1887 | else |
b5b05048 | 1888 | nvme_disable_ctrl(&ctrl->ctrl); |
75862c72 | 1889 | nvme_rdma_teardown_admin_queue(ctrl, shutdown); |
71102307 CH |
1890 | } |
1891 | ||
c5017e85 | 1892 | static void nvme_rdma_delete_ctrl(struct nvme_ctrl *ctrl) |
2461a8dd | 1893 | { |
e9bc2587 | 1894 | nvme_rdma_shutdown_ctrl(to_rdma_ctrl(ctrl), true); |
71102307 CH |
1895 | } |
1896 | ||
71102307 CH |
1897 | static void nvme_rdma_reset_ctrl_work(struct work_struct *work) |
1898 | { | |
d86c4d8e CH |
1899 | struct nvme_rdma_ctrl *ctrl = |
1900 | container_of(work, struct nvme_rdma_ctrl, ctrl.reset_work); | |
71102307 | 1901 | |
d09f2b45 | 1902 | nvme_stop_ctrl(&ctrl->ctrl); |
18398af2 | 1903 | nvme_rdma_shutdown_ctrl(ctrl, false); |
71102307 | 1904 | |
ad6a0a52 | 1905 | if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) { |
d5bf4b7f SG |
1906 | /* state change failure should never happen */ |
1907 | WARN_ON_ONCE(1); | |
1908 | return; | |
1909 | } | |
1910 | ||
c66e2998 | 1911 | if (nvme_rdma_setup_ctrl(ctrl, false)) |
370ae6e4 | 1912 | goto out_fail; |
71102307 | 1913 | |
71102307 CH |
1914 | return; |
1915 | ||
370ae6e4 | 1916 | out_fail: |
8000d1fd NC |
1917 | ++ctrl->ctrl.nr_reconnects; |
1918 | nvme_rdma_reconnect_or_remove(ctrl); | |
71102307 CH |
1919 | } |
1920 | ||
71102307 CH |
1921 | static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = { |
1922 | .name = "rdma", | |
1923 | .module = THIS_MODULE, | |
d3d5b87d | 1924 | .flags = NVME_F_FABRICS, |
71102307 CH |
1925 | .reg_read32 = nvmf_reg_read32, |
1926 | .reg_read64 = nvmf_reg_read64, | |
1927 | .reg_write32 = nvmf_reg_write32, | |
71102307 CH |
1928 | .free_ctrl = nvme_rdma_free_ctrl, |
1929 | .submit_async_event = nvme_rdma_submit_async_event, | |
c5017e85 | 1930 | .delete_ctrl = nvme_rdma_delete_ctrl, |
71102307 CH |
1931 | .get_address = nvmf_get_address, |
1932 | }; | |
1933 | ||
36e835f2 JS |
1934 | /* |
1935 | * Fails a connection request if it matches an existing controller | |
1936 | * (association) with the same tuple: | |
1937 | * <Host NQN, Host ID, local address, remote address, remote port, SUBSYS NQN> | |
1938 | * | |
1939 | * if local address is not specified in the request, it will match an | |
1940 | * existing controller with all the other parameters the same and no | |
1941 | * local port address specified as well. | |
1942 | * | |
1943 | * The ports don't need to be compared as they are intrinsically | |
1944 | * already matched by the port pointers supplied. | |
1945 | */ | |
1946 | static bool | |
1947 | nvme_rdma_existing_controller(struct nvmf_ctrl_options *opts) | |
1948 | { | |
1949 | struct nvme_rdma_ctrl *ctrl; | |
1950 | bool found = false; | |
1951 | ||
1952 | mutex_lock(&nvme_rdma_ctrl_mutex); | |
1953 | list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) { | |
b7c7be6f | 1954 | found = nvmf_ip_options_match(&ctrl->ctrl, opts); |
36e835f2 JS |
1955 | if (found) |
1956 | break; | |
1957 | } | |
1958 | mutex_unlock(&nvme_rdma_ctrl_mutex); | |
1959 | ||
1960 | return found; | |
1961 | } | |
1962 | ||
71102307 CH |
1963 | static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev, |
1964 | struct nvmf_ctrl_options *opts) | |
1965 | { | |
1966 | struct nvme_rdma_ctrl *ctrl; | |
1967 | int ret; | |
1968 | bool changed; | |
1969 | ||
1970 | ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL); | |
1971 | if (!ctrl) | |
1972 | return ERR_PTR(-ENOMEM); | |
1973 | ctrl->ctrl.opts = opts; | |
1974 | INIT_LIST_HEAD(&ctrl->list); | |
1975 | ||
bb59b8e5 SG |
1976 | if (!(opts->mask & NVMF_OPT_TRSVCID)) { |
1977 | opts->trsvcid = | |
1978 | kstrdup(__stringify(NVME_RDMA_IP_PORT), GFP_KERNEL); | |
1979 | if (!opts->trsvcid) { | |
1980 | ret = -ENOMEM; | |
1981 | goto out_free_ctrl; | |
1982 | } | |
1983 | opts->mask |= NVMF_OPT_TRSVCID; | |
1984 | } | |
0928f9b4 SG |
1985 | |
1986 | ret = inet_pton_with_scope(&init_net, AF_UNSPEC, | |
bb59b8e5 | 1987 | opts->traddr, opts->trsvcid, &ctrl->addr); |
71102307 | 1988 | if (ret) { |
bb59b8e5 SG |
1989 | pr_err("malformed address passed: %s:%s\n", |
1990 | opts->traddr, opts->trsvcid); | |
71102307 CH |
1991 | goto out_free_ctrl; |
1992 | } | |
1993 | ||
8f4e8dac | 1994 | if (opts->mask & NVMF_OPT_HOST_TRADDR) { |
0928f9b4 SG |
1995 | ret = inet_pton_with_scope(&init_net, AF_UNSPEC, |
1996 | opts->host_traddr, NULL, &ctrl->src_addr); | |
8f4e8dac | 1997 | if (ret) { |
0928f9b4 | 1998 | pr_err("malformed src address passed: %s\n", |
8f4e8dac MG |
1999 | opts->host_traddr); |
2000 | goto out_free_ctrl; | |
2001 | } | |
2002 | } | |
2003 | ||
36e835f2 JS |
2004 | if (!opts->duplicate_connect && nvme_rdma_existing_controller(opts)) { |
2005 | ret = -EALREADY; | |
2006 | goto out_free_ctrl; | |
2007 | } | |
2008 | ||
71102307 CH |
2009 | INIT_DELAYED_WORK(&ctrl->reconnect_work, |
2010 | nvme_rdma_reconnect_ctrl_work); | |
2011 | INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work); | |
d86c4d8e | 2012 | INIT_WORK(&ctrl->ctrl.reset_work, nvme_rdma_reset_ctrl_work); |
71102307 | 2013 | |
ff8519f9 SG |
2014 | ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues + |
2015 | opts->nr_poll_queues + 1; | |
c5af8654 | 2016 | ctrl->ctrl.sqsize = opts->queue_size - 1; |
71102307 CH |
2017 | ctrl->ctrl.kato = opts->kato; |
2018 | ||
2019 | ret = -ENOMEM; | |
d858e5f0 | 2020 | ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues), |
71102307 CH |
2021 | GFP_KERNEL); |
2022 | if (!ctrl->queues) | |
3d064101 SG |
2023 | goto out_free_ctrl; |
2024 | ||
2025 | ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops, | |
2026 | 0 /* no quirks, we're perfect! */); | |
2027 | if (ret) | |
2028 | goto out_kfree_queues; | |
71102307 | 2029 | |
b754a32c MG |
2030 | changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING); |
2031 | WARN_ON_ONCE(!changed); | |
2032 | ||
c66e2998 | 2033 | ret = nvme_rdma_setup_ctrl(ctrl, true); |
71102307 | 2034 | if (ret) |
3d064101 | 2035 | goto out_uninit_ctrl; |
71102307 | 2036 | |
0928f9b4 | 2037 | dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs\n", |
71102307 CH |
2038 | ctrl->ctrl.opts->subsysnqn, &ctrl->addr); |
2039 | ||
d22524a4 | 2040 | nvme_get_ctrl(&ctrl->ctrl); |
71102307 CH |
2041 | |
2042 | mutex_lock(&nvme_rdma_ctrl_mutex); | |
2043 | list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list); | |
2044 | mutex_unlock(&nvme_rdma_ctrl_mutex); | |
2045 | ||
71102307 CH |
2046 | return &ctrl->ctrl; |
2047 | ||
71102307 CH |
2048 | out_uninit_ctrl: |
2049 | nvme_uninit_ctrl(&ctrl->ctrl); | |
2050 | nvme_put_ctrl(&ctrl->ctrl); | |
2051 | if (ret > 0) | |
2052 | ret = -EIO; | |
2053 | return ERR_PTR(ret); | |
3d064101 SG |
2054 | out_kfree_queues: |
2055 | kfree(ctrl->queues); | |
71102307 CH |
2056 | out_free_ctrl: |
2057 | kfree(ctrl); | |
2058 | return ERR_PTR(ret); | |
2059 | } | |
2060 | ||
2061 | static struct nvmf_transport_ops nvme_rdma_transport = { | |
2062 | .name = "rdma", | |
0de5cd36 | 2063 | .module = THIS_MODULE, |
71102307 | 2064 | .required_opts = NVMF_OPT_TRADDR, |
8f4e8dac | 2065 | .allowed_opts = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY | |
b65bb777 | 2066 | NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO | |
e63440d6 IR |
2067 | NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES | |
2068 | NVMF_OPT_TOS, | |
71102307 CH |
2069 | .create_ctrl = nvme_rdma_create_ctrl, |
2070 | }; | |
2071 | ||
e87a911f SW |
2072 | static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data) |
2073 | { | |
2074 | struct nvme_rdma_ctrl *ctrl; | |
9bad0404 MG |
2075 | struct nvme_rdma_device *ndev; |
2076 | bool found = false; | |
2077 | ||
2078 | mutex_lock(&device_list_mutex); | |
2079 | list_for_each_entry(ndev, &device_list, entry) { | |
2080 | if (ndev->dev == ib_device) { | |
2081 | found = true; | |
2082 | break; | |
2083 | } | |
2084 | } | |
2085 | mutex_unlock(&device_list_mutex); | |
2086 | ||
2087 | if (!found) | |
2088 | return; | |
e87a911f SW |
2089 | |
2090 | /* Delete all controllers using this device */ | |
2091 | mutex_lock(&nvme_rdma_ctrl_mutex); | |
2092 | list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) { | |
2093 | if (ctrl->device->dev != ib_device) | |
2094 | continue; | |
c5017e85 | 2095 | nvme_delete_ctrl(&ctrl->ctrl); |
e87a911f SW |
2096 | } |
2097 | mutex_unlock(&nvme_rdma_ctrl_mutex); | |
2098 | ||
b227c59b | 2099 | flush_workqueue(nvme_delete_wq); |
e87a911f SW |
2100 | } |
2101 | ||
2102 | static struct ib_client nvme_rdma_ib_client = { | |
2103 | .name = "nvme_rdma", | |
e87a911f SW |
2104 | .remove = nvme_rdma_remove_one |
2105 | }; | |
2106 | ||
71102307 CH |
2107 | static int __init nvme_rdma_init_module(void) |
2108 | { | |
e87a911f SW |
2109 | int ret; |
2110 | ||
e87a911f | 2111 | ret = ib_register_client(&nvme_rdma_ib_client); |
a56c79cf | 2112 | if (ret) |
9a6327d2 | 2113 | return ret; |
a56c79cf SG |
2114 | |
2115 | ret = nvmf_register_transport(&nvme_rdma_transport); | |
2116 | if (ret) | |
2117 | goto err_unreg_client; | |
e87a911f | 2118 | |
a56c79cf | 2119 | return 0; |
e87a911f | 2120 | |
a56c79cf SG |
2121 | err_unreg_client: |
2122 | ib_unregister_client(&nvme_rdma_ib_client); | |
a56c79cf | 2123 | return ret; |
71102307 CH |
2124 | } |
2125 | ||
2126 | static void __exit nvme_rdma_cleanup_module(void) | |
2127 | { | |
71102307 | 2128 | nvmf_unregister_transport(&nvme_rdma_transport); |
e87a911f | 2129 | ib_unregister_client(&nvme_rdma_ib_client); |
71102307 CH |
2130 | } |
2131 | ||
2132 | module_init(nvme_rdma_init_module); | |
2133 | module_exit(nvme_rdma_cleanup_module); | |
2134 | ||
2135 | MODULE_LICENSE("GPL v2"); |