?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
dm writecache: add cond_resched to loop in persistent_memory_claim()
[linux/fpc-iii.git] / drivers / nvme / host / tcp.c
blob4862fa962011dc631fe25f6a913ce44f13f00172
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * NVMe over Fabrics TCP host.
4 * Copyright (c) 2018 Lightbits Labs. All rights reserved.
5 */
6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
7 #include <linux/module.h>
8 #include <linux/init.h>
9 #include <linux/slab.h>
10 #include <linux/err.h>
11 #include <linux/nvme-tcp.h>
12 #include <net/sock.h>
13 #include <net/tcp.h>
14 #include <linux/blk-mq.h>
15 #include <crypto/hash.h>
16 #include <net/busy_poll.h>
17
18 #include "nvme.h"
19 #include "fabrics.h"
20
21 struct nvme_tcp_queue;
22
23 /* Define the socket priority to use for connections were it is desirable
24 * that the NIC consider performing optimized packet processing or filtering.
25 * A non-zero value being sufficient to indicate general consideration of any
26 * possible optimization. Making it a module param allows for alternative
27 * values that may be unique for some NIC implementations.
28 */
29 static int so_priority;
30 module_param(so_priority, int, 0644);
31 MODULE_PARM_DESC(so_priority, "nvme tcp socket optimize priority");
32
33 enum nvme_tcp_send_state {
34 NVME_TCP_SEND_CMD_PDU = 0,
35 NVME_TCP_SEND_H2C_PDU,
36 NVME_TCP_SEND_DATA,
37 NVME_TCP_SEND_DDGST,
38 };
39
40 struct nvme_tcp_request {
41 struct nvme_request req;
42 void *pdu;
43 struct nvme_tcp_queue *queue;
44 u32 data_len;
45 u32 pdu_len;
46 u32 pdu_sent;
47 u16 ttag;
48 struct list_head entry;
49 __le32 ddgst;
50
51 struct bio *curr_bio;
52 struct iov_iter iter;
53
54 /* send state */
55 size_t offset;
56 size_t data_sent;
57 enum nvme_tcp_send_state state;
58 };
59
60 enum nvme_tcp_queue_flags {
61 NVME_TCP_Q_ALLOCATED = 0,
62 NVME_TCP_Q_LIVE = 1,
63 };
64
65 enum nvme_tcp_recv_state {
66 NVME_TCP_RECV_PDU = 0,
67 NVME_TCP_RECV_DATA,
68 NVME_TCP_RECV_DDGST,
69 };
70
71 struct nvme_tcp_ctrl;
72 struct nvme_tcp_queue {
73 struct socket *sock;
74 struct work_struct io_work;
75 int io_cpu;
76
77 spinlock_t lock;
78 struct list_head send_list;
79
80 /* recv state */
81 void *pdu;
82 int pdu_remaining;
83 int pdu_offset;
84 size_t data_remaining;
85 size_t ddgst_remaining;
86 unsigned int nr_cqe;
87
88 /* send state */
89 struct nvme_tcp_request *request;
90
91 int queue_size;
92 size_t cmnd_capsule_len;
93 struct nvme_tcp_ctrl *ctrl;
94 unsigned long flags;
95 bool rd_enabled;
96
97 bool hdr_digest;
98 bool data_digest;
99 struct ahash_request *rcv_hash;
100 struct ahash_request *snd_hash;
101 __le32 exp_ddgst;
102 __le32 recv_ddgst;
103
104 struct page_frag_cache pf_cache;
105
106 void (*state_change)(struct sock *);
107 void (*data_ready)(struct sock *);
108 void (*write_space)(struct sock *);
109 };
110
111 struct nvme_tcp_ctrl {
112 /* read only in the hot path */
113 struct nvme_tcp_queue *queues;
114 struct blk_mq_tag_set tag_set;
115
116 /* other member variables */
117 struct list_head list;
118 struct blk_mq_tag_set admin_tag_set;
119 struct sockaddr_storage addr;
120 struct sockaddr_storage src_addr;
121 struct nvme_ctrl ctrl;
122
123 struct work_struct err_work;
124 struct delayed_work connect_work;
125 struct nvme_tcp_request async_req;
126 u32 io_queues[HCTX_MAX_TYPES];
127 };
128
129 static LIST_HEAD(nvme_tcp_ctrl_list);
130 static DEFINE_MUTEX(nvme_tcp_ctrl_mutex);
131 static struct workqueue_struct *nvme_tcp_wq;
132 static struct blk_mq_ops nvme_tcp_mq_ops;
133 static struct blk_mq_ops nvme_tcp_admin_mq_ops;
134
135 static inline struct nvme_tcp_ctrl *to_tcp_ctrl(struct nvme_ctrl *ctrl)
136 {
137 return container_of(ctrl, struct nvme_tcp_ctrl, ctrl);
138 }
139
140 static inline int nvme_tcp_queue_id(struct nvme_tcp_queue *queue)
141 {
142 return queue - queue->ctrl->queues;
143 }
144
145 static inline struct blk_mq_tags *nvme_tcp_tagset(struct nvme_tcp_queue *queue)
146 {
147 u32 queue_idx = nvme_tcp_queue_id(queue);
148
149 if (queue_idx == 0)
150 return queue->ctrl->admin_tag_set.tags[queue_idx];
151 return queue->ctrl->tag_set.tags[queue_idx - 1];
152 }
153
154 static inline u8 nvme_tcp_hdgst_len(struct nvme_tcp_queue *queue)
155 {
156 return queue->hdr_digest ? NVME_TCP_DIGEST_LENGTH : 0;
157 }
158
159 static inline u8 nvme_tcp_ddgst_len(struct nvme_tcp_queue *queue)
160 {
161 return queue->data_digest ? NVME_TCP_DIGEST_LENGTH : 0;
162 }
163
164 static inline size_t nvme_tcp_inline_data_size(struct nvme_tcp_queue *queue)
165 {
166 return queue->cmnd_capsule_len - sizeof(struct nvme_command);
167 }
168
169 static inline bool nvme_tcp_async_req(struct nvme_tcp_request *req)
170 {
171 return req == &req->queue->ctrl->async_req;
172 }
173
174 static inline bool nvme_tcp_has_inline_data(struct nvme_tcp_request *req)
175 {
176 struct request *rq;
177
178 if (unlikely(nvme_tcp_async_req(req)))
179 return false; /* async events don't have a request */
180
181 rq = blk_mq_rq_from_pdu(req);
182
183 return rq_data_dir(rq) == WRITE && req->data_len &&
184 req->data_len <= nvme_tcp_inline_data_size(req->queue);
185 }
186
187 static inline struct page *nvme_tcp_req_cur_page(struct nvme_tcp_request *req)
188 {
189 return req->iter.bvec->bv_page;
190 }
191
192 static inline size_t nvme_tcp_req_cur_offset(struct nvme_tcp_request *req)
193 {
194 return req->iter.bvec->bv_offset + req->iter.iov_offset;
195 }
196
197 static inline size_t nvme_tcp_req_cur_length(struct nvme_tcp_request *req)
198 {
199 return min_t(size_t, req->iter.bvec->bv_len - req->iter.iov_offset,
200 req->pdu_len - req->pdu_sent);
201 }
202
203 static inline size_t nvme_tcp_req_offset(struct nvme_tcp_request *req)
204 {
205 return req->iter.iov_offset;
206 }
207
208 static inline size_t nvme_tcp_pdu_data_left(struct nvme_tcp_request *req)
209 {
210 return rq_data_dir(blk_mq_rq_from_pdu(req)) == WRITE ?
211 req->pdu_len - req->pdu_sent : 0;
212 }
213
214 static inline size_t nvme_tcp_pdu_last_send(struct nvme_tcp_request *req,
215 int len)
216 {
217 return nvme_tcp_pdu_data_left(req) <= len;
218 }
219
220 static void nvme_tcp_init_iter(struct nvme_tcp_request *req,
221 unsigned int dir)
222 {
223 struct request *rq = blk_mq_rq_from_pdu(req);
224 struct bio_vec *vec;
225 unsigned int size;
226 int nsegs;
227 size_t offset;
228
229 if (rq->rq_flags & RQF_SPECIAL_PAYLOAD) {
230 vec = &rq->special_vec;
231 nsegs = 1;
232 size = blk_rq_payload_bytes(rq);
233 offset = 0;
234 } else {
235 struct bio *bio = req->curr_bio;
236
237 vec = __bvec_iter_bvec(bio->bi_io_vec, bio->bi_iter);
238 nsegs = bio_segments(bio);
239 size = bio->bi_iter.bi_size;
240 offset = bio->bi_iter.bi_bvec_done;
241 }
242
243 iov_iter_bvec(&req->iter, dir, vec, nsegs, size);
244 req->iter.iov_offset = offset;
245 }
246
247 static inline void nvme_tcp_advance_req(struct nvme_tcp_request *req,
248 int len)
249 {
250 req->data_sent += len;
251 req->pdu_sent += len;
252 iov_iter_advance(&req->iter, len);
253 if (!iov_iter_count(&req->iter) &&
254 req->data_sent < req->data_len) {
255 req->curr_bio = req->curr_bio->bi_next;
256 nvme_tcp_init_iter(req, WRITE);
257 }
258 }
259
260 static inline void nvme_tcp_queue_request(struct nvme_tcp_request *req)
261 {
262 struct nvme_tcp_queue *queue = req->queue;
263
264 spin_lock(&queue->lock);
265 list_add_tail(&req->entry, &queue->send_list);
266 spin_unlock(&queue->lock);
267
268 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
269 }
270
271 static inline struct nvme_tcp_request *
272 nvme_tcp_fetch_request(struct nvme_tcp_queue *queue)
273 {
274 struct nvme_tcp_request *req;
275
276 spin_lock(&queue->lock);
277 req = list_first_entry_or_null(&queue->send_list,
278 struct nvme_tcp_request, entry);
279 if (req)
280 list_del(&req->entry);
281 spin_unlock(&queue->lock);
282
283 return req;
284 }
285
286 static inline void nvme_tcp_ddgst_final(struct ahash_request *hash,
287 __le32 *dgst)
288 {
289 ahash_request_set_crypt(hash, NULL, (u8 *)dgst, 0);
290 crypto_ahash_final(hash);
291 }
292
293 static inline void nvme_tcp_ddgst_update(struct ahash_request *hash,
294 struct page *page, off_t off, size_t len)
295 {
296 struct scatterlist sg;
297
298 sg_init_marker(&sg, 1);
299 sg_set_page(&sg, page, len, off);
300 ahash_request_set_crypt(hash, &sg, NULL, len);
301 crypto_ahash_update(hash);
302 }
303
304 static inline void nvme_tcp_hdgst(struct ahash_request *hash,
305 void *pdu, size_t len)
306 {
307 struct scatterlist sg;
308
309 sg_init_one(&sg, pdu, len);
310 ahash_request_set_crypt(hash, &sg, pdu + len, len);
311 crypto_ahash_digest(hash);
312 }
313
314 static int nvme_tcp_verify_hdgst(struct nvme_tcp_queue *queue,
315 void *pdu, size_t pdu_len)
316 {
317 struct nvme_tcp_hdr *hdr = pdu;
318 __le32 recv_digest;
319 __le32 exp_digest;
320
321 if (unlikely(!(hdr->flags & NVME_TCP_F_HDGST))) {
322 dev_err(queue->ctrl->ctrl.device,
323 "queue %d: header digest flag is cleared\n",
324 nvme_tcp_queue_id(queue));
325 return -EPROTO;
326 }
327
328 recv_digest = *(__le32 *)(pdu + hdr->hlen);
329 nvme_tcp_hdgst(queue->rcv_hash, pdu, pdu_len);
330 exp_digest = *(__le32 *)(pdu + hdr->hlen);
331 if (recv_digest != exp_digest) {
332 dev_err(queue->ctrl->ctrl.device,
333 "header digest error: recv %#x expected %#x\n",
334 le32_to_cpu(recv_digest), le32_to_cpu(exp_digest));
335 return -EIO;
336 }
337
338 return 0;
339 }
340
341 static int nvme_tcp_check_ddgst(struct nvme_tcp_queue *queue, void *pdu)
342 {
343 struct nvme_tcp_hdr *hdr = pdu;
344 u8 digest_len = nvme_tcp_hdgst_len(queue);
345 u32 len;
346
347 len = le32_to_cpu(hdr->plen) - hdr->hlen -
348 ((hdr->flags & NVME_TCP_F_HDGST) ? digest_len : 0);
349
350 if (unlikely(len && !(hdr->flags & NVME_TCP_F_DDGST))) {
351 dev_err(queue->ctrl->ctrl.device,
352 "queue %d: data digest flag is cleared\n",
353 nvme_tcp_queue_id(queue));
354 return -EPROTO;
355 }
356 crypto_ahash_init(queue->rcv_hash);
357
358 return 0;
359 }
360
361 static void nvme_tcp_exit_request(struct blk_mq_tag_set *set,
362 struct request *rq, unsigned int hctx_idx)
363 {
364 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
365
366 page_frag_free(req->pdu);
367 }
368
369 static int nvme_tcp_init_request(struct blk_mq_tag_set *set,
370 struct request *rq, unsigned int hctx_idx,
371 unsigned int numa_node)
372 {
373 struct nvme_tcp_ctrl *ctrl = set->driver_data;
374 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
375 int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
376 struct nvme_tcp_queue *queue = &ctrl->queues[queue_idx];
377 u8 hdgst = nvme_tcp_hdgst_len(queue);
378
379 req->pdu = page_frag_alloc(&queue->pf_cache,
380 sizeof(struct nvme_tcp_cmd_pdu) + hdgst,
381 GFP_KERNEL | __GFP_ZERO);
382 if (!req->pdu)
383 return -ENOMEM;
384
385 req->queue = queue;
386 nvme_req(rq)->ctrl = &ctrl->ctrl;
387
388 return 0;
389 }
390
391 static int nvme_tcp_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
392 unsigned int hctx_idx)
393 {
394 struct nvme_tcp_ctrl *ctrl = data;
395 struct nvme_tcp_queue *queue = &ctrl->queues[hctx_idx + 1];
396
397 hctx->driver_data = queue;
398 return 0;
399 }
400
401 static int nvme_tcp_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
402 unsigned int hctx_idx)
403 {
404 struct nvme_tcp_ctrl *ctrl = data;
405 struct nvme_tcp_queue *queue = &ctrl->queues[0];
406
407 hctx->driver_data = queue;
408 return 0;
409 }
410
411 static enum nvme_tcp_recv_state
412 nvme_tcp_recv_state(struct nvme_tcp_queue *queue)
413 {
414 return (queue->pdu_remaining) ? NVME_TCP_RECV_PDU :
415 (queue->ddgst_remaining) ? NVME_TCP_RECV_DDGST :
416 NVME_TCP_RECV_DATA;
417 }
418
419 static void nvme_tcp_init_recv_ctx(struct nvme_tcp_queue *queue)
420 {
421 queue->pdu_remaining = sizeof(struct nvme_tcp_rsp_pdu) +
422 nvme_tcp_hdgst_len(queue);
423 queue->pdu_offset = 0;
424 queue->data_remaining = -1;
425 queue->ddgst_remaining = 0;
426 }
427
428 static void nvme_tcp_error_recovery(struct nvme_ctrl *ctrl)
429 {
430 if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_RESETTING))
431 return;
432
433 queue_work(nvme_reset_wq, &to_tcp_ctrl(ctrl)->err_work);
434 }
435
436 static int nvme_tcp_process_nvme_cqe(struct nvme_tcp_queue *queue,
437 struct nvme_completion *cqe)
438 {
439 struct request *rq;
440
441 rq = blk_mq_tag_to_rq(nvme_tcp_tagset(queue), cqe->command_id);
442 if (!rq) {
443 dev_err(queue->ctrl->ctrl.device,
444 "queue %d tag 0x%x not found\n",
445 nvme_tcp_queue_id(queue), cqe->command_id);
446 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
447 return -EINVAL;
448 }
449
450 nvme_end_request(rq, cqe->status, cqe->result);
451 queue->nr_cqe++;
452
453 return 0;
454 }
455
456 static int nvme_tcp_handle_c2h_data(struct nvme_tcp_queue *queue,
457 struct nvme_tcp_data_pdu *pdu)
458 {
459 struct request *rq;
460
461 rq = blk_mq_tag_to_rq(nvme_tcp_tagset(queue), pdu->command_id);
462 if (!rq) {
463 dev_err(queue->ctrl->ctrl.device,
464 "queue %d tag %#x not found\n",
465 nvme_tcp_queue_id(queue), pdu->command_id);
466 return -ENOENT;
467 }
468
469 if (!blk_rq_payload_bytes(rq)) {
470 dev_err(queue->ctrl->ctrl.device,
471 "queue %d tag %#x unexpected data\n",
472 nvme_tcp_queue_id(queue), rq->tag);
473 return -EIO;
474 }
475
476 queue->data_remaining = le32_to_cpu(pdu->data_length);
477
478 if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS &&
479 unlikely(!(pdu->hdr.flags & NVME_TCP_F_DATA_LAST))) {
480 dev_err(queue->ctrl->ctrl.device,
481 "queue %d tag %#x SUCCESS set but not last PDU\n",
482 nvme_tcp_queue_id(queue), rq->tag);
483 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
484 return -EPROTO;
485 }
486
487 return 0;
488 }
489
490 static int nvme_tcp_handle_comp(struct nvme_tcp_queue *queue,
491 struct nvme_tcp_rsp_pdu *pdu)
492 {
493 struct nvme_completion *cqe = &pdu->cqe;
494 int ret = 0;
495
496 /*
497 * AEN requests are special as they don't time out and can
498 * survive any kind of queue freeze and often don't respond to
499 * aborts. We don't even bother to allocate a struct request
500 * for them but rather special case them here.
501 */
502 if (unlikely(nvme_is_aen_req(nvme_tcp_queue_id(queue),
503 cqe->command_id)))
504 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
505 &cqe->result);
506 else
507 ret = nvme_tcp_process_nvme_cqe(queue, cqe);
508
509 return ret;
510 }
511
512 static int nvme_tcp_setup_h2c_data_pdu(struct nvme_tcp_request *req,
513 struct nvme_tcp_r2t_pdu *pdu)
514 {
515 struct nvme_tcp_data_pdu *data = req->pdu;
516 struct nvme_tcp_queue *queue = req->queue;
517 struct request *rq = blk_mq_rq_from_pdu(req);
518 u8 hdgst = nvme_tcp_hdgst_len(queue);
519 u8 ddgst = nvme_tcp_ddgst_len(queue);
520
521 req->pdu_len = le32_to_cpu(pdu->r2t_length);
522 req->pdu_sent = 0;
523
524 if (unlikely(req->data_sent + req->pdu_len > req->data_len)) {
525 dev_err(queue->ctrl->ctrl.device,
526 "req %d r2t len %u exceeded data len %u (%zu sent)\n",
527 rq->tag, req->pdu_len, req->data_len,
528 req->data_sent);
529 return -EPROTO;
530 }
531
532 if (unlikely(le32_to_cpu(pdu->r2t_offset) < req->data_sent)) {
533 dev_err(queue->ctrl->ctrl.device,
534 "req %d unexpected r2t offset %u (expected %zu)\n",
535 rq->tag, le32_to_cpu(pdu->r2t_offset),
536 req->data_sent);
537 return -EPROTO;
538 }
539
540 memset(data, 0, sizeof(*data));
541 data->hdr.type = nvme_tcp_h2c_data;
542 data->hdr.flags = NVME_TCP_F_DATA_LAST;
543 if (queue->hdr_digest)
544 data->hdr.flags |= NVME_TCP_F_HDGST;
545 if (queue->data_digest)
546 data->hdr.flags |= NVME_TCP_F_DDGST;
547 data->hdr.hlen = sizeof(*data);
548 data->hdr.pdo = data->hdr.hlen + hdgst;
549 data->hdr.plen =
550 cpu_to_le32(data->hdr.hlen + hdgst + req->pdu_len + ddgst);
551 data->ttag = pdu->ttag;
552 data->command_id = rq->tag;
553 data->data_offset = cpu_to_le32(req->data_sent);
554 data->data_length = cpu_to_le32(req->pdu_len);
555 return 0;
556 }
557
558 static int nvme_tcp_handle_r2t(struct nvme_tcp_queue *queue,
559 struct nvme_tcp_r2t_pdu *pdu)
560 {
561 struct nvme_tcp_request *req;
562 struct request *rq;
563 int ret;
564
565 rq = blk_mq_tag_to_rq(nvme_tcp_tagset(queue), pdu->command_id);
566 if (!rq) {
567 dev_err(queue->ctrl->ctrl.device,
568 "queue %d tag %#x not found\n",
569 nvme_tcp_queue_id(queue), pdu->command_id);
570 return -ENOENT;
571 }
572 req = blk_mq_rq_to_pdu(rq);
573
574 ret = nvme_tcp_setup_h2c_data_pdu(req, pdu);
575 if (unlikely(ret))
576 return ret;
577
578 req->state = NVME_TCP_SEND_H2C_PDU;
579 req->offset = 0;
580
581 nvme_tcp_queue_request(req);
582
583 return 0;
584 }
585
586 static int nvme_tcp_recv_pdu(struct nvme_tcp_queue *queue, struct sk_buff *skb,
587 unsigned int *offset, size_t *len)
588 {
589 struct nvme_tcp_hdr *hdr;
590 char *pdu = queue->pdu;
591 size_t rcv_len = min_t(size_t, *len, queue->pdu_remaining);
592 int ret;
593
594 ret = skb_copy_bits(skb, *offset,
595 &pdu[queue->pdu_offset], rcv_len);
596 if (unlikely(ret))
597 return ret;
598
599 queue->pdu_remaining -= rcv_len;
600 queue->pdu_offset += rcv_len;
601 *offset += rcv_len;
602 *len -= rcv_len;
603 if (queue->pdu_remaining)
604 return 0;
605
606 hdr = queue->pdu;
607 if (queue->hdr_digest) {
608 ret = nvme_tcp_verify_hdgst(queue, queue->pdu, hdr->hlen);
609 if (unlikely(ret))
610 return ret;
611 }
612
613
614 if (queue->data_digest) {
615 ret = nvme_tcp_check_ddgst(queue, queue->pdu);
616 if (unlikely(ret))
617 return ret;
618 }
619
620 switch (hdr->type) {
621 case nvme_tcp_c2h_data:
622 return nvme_tcp_handle_c2h_data(queue, (void *)queue->pdu);
623 case nvme_tcp_rsp:
624 nvme_tcp_init_recv_ctx(queue);
625 return nvme_tcp_handle_comp(queue, (void *)queue->pdu);
626 case nvme_tcp_r2t:
627 nvme_tcp_init_recv_ctx(queue);
628 return nvme_tcp_handle_r2t(queue, (void *)queue->pdu);
629 default:
630 dev_err(queue->ctrl->ctrl.device,
631 "unsupported pdu type (%d)\n", hdr->type);
632 return -EINVAL;
633 }
634 }
635
636 static inline void nvme_tcp_end_request(struct request *rq, u16 status)
637 {
638 union nvme_result res = {};
639
640 nvme_end_request(rq, cpu_to_le16(status << 1), res);
641 }
642
643 static int nvme_tcp_recv_data(struct nvme_tcp_queue *queue, struct sk_buff *skb,
644 unsigned int *offset, size_t *len)
645 {
646 struct nvme_tcp_data_pdu *pdu = (void *)queue->pdu;
647 struct nvme_tcp_request *req;
648 struct request *rq;
649
650 rq = blk_mq_tag_to_rq(nvme_tcp_tagset(queue), pdu->command_id);
651 if (!rq) {
652 dev_err(queue->ctrl->ctrl.device,
653 "queue %d tag %#x not found\n",
654 nvme_tcp_queue_id(queue), pdu->command_id);
655 return -ENOENT;
656 }
657 req = blk_mq_rq_to_pdu(rq);
658
659 while (true) {
660 int recv_len, ret;
661
662 recv_len = min_t(size_t, *len, queue->data_remaining);
663 if (!recv_len)
664 break;
665
666 if (!iov_iter_count(&req->iter)) {
667 req->curr_bio = req->curr_bio->bi_next;
668
669 /*
670 * If we don`t have any bios it means that controller
671 * sent more data than we requested, hence error
672 */
673 if (!req->curr_bio) {
674 dev_err(queue->ctrl->ctrl.device,
675 "queue %d no space in request %#x",
676 nvme_tcp_queue_id(queue), rq->tag);
677 nvme_tcp_init_recv_ctx(queue);
678 return -EIO;
679 }
680 nvme_tcp_init_iter(req, READ);
681 }
682
683 /* we can read only from what is left in this bio */
684 recv_len = min_t(size_t, recv_len,
685 iov_iter_count(&req->iter));
686
687 if (queue->data_digest)
688 ret = skb_copy_and_hash_datagram_iter(skb, *offset,
689 &req->iter, recv_len, queue->rcv_hash);
690 else
691 ret = skb_copy_datagram_iter(skb, *offset,
692 &req->iter, recv_len);
693 if (ret) {
694 dev_err(queue->ctrl->ctrl.device,
695 "queue %d failed to copy request %#x data",
696 nvme_tcp_queue_id(queue), rq->tag);
697 return ret;
698 }
699
700 *len -= recv_len;
701 *offset += recv_len;
702 queue->data_remaining -= recv_len;
703 }
704
705 if (!queue->data_remaining) {
706 if (queue->data_digest) {
707 nvme_tcp_ddgst_final(queue->rcv_hash, &queue->exp_ddgst);
708 queue->ddgst_remaining = NVME_TCP_DIGEST_LENGTH;
709 } else {
710 if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
711 nvme_tcp_end_request(rq, NVME_SC_SUCCESS);
712 queue->nr_cqe++;
713 }
714 nvme_tcp_init_recv_ctx(queue);
715 }
716 }
717
718 return 0;
719 }
720
721 static int nvme_tcp_recv_ddgst(struct nvme_tcp_queue *queue,
722 struct sk_buff *skb, unsigned int *offset, size_t *len)
723 {
724 struct nvme_tcp_data_pdu *pdu = (void *)queue->pdu;
725 char *ddgst = (char *)&queue->recv_ddgst;
726 size_t recv_len = min_t(size_t, *len, queue->ddgst_remaining);
727 off_t off = NVME_TCP_DIGEST_LENGTH - queue->ddgst_remaining;
728 int ret;
729
730 ret = skb_copy_bits(skb, *offset, &ddgst[off], recv_len);
731 if (unlikely(ret))
732 return ret;
733
734 queue->ddgst_remaining -= recv_len;
735 *offset += recv_len;
736 *len -= recv_len;
737 if (queue->ddgst_remaining)
738 return 0;
739
740 if (queue->recv_ddgst != queue->exp_ddgst) {
741 dev_err(queue->ctrl->ctrl.device,
742 "data digest error: recv %#x expected %#x\n",
743 le32_to_cpu(queue->recv_ddgst),
744 le32_to_cpu(queue->exp_ddgst));
745 return -EIO;
746 }
747
748 if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
749 struct request *rq = blk_mq_tag_to_rq(nvme_tcp_tagset(queue),
750 pdu->command_id);
751
752 nvme_tcp_end_request(rq, NVME_SC_SUCCESS);
753 queue->nr_cqe++;
754 }
755
756 nvme_tcp_init_recv_ctx(queue);
757 return 0;
758 }
759
760 static int nvme_tcp_recv_skb(read_descriptor_t *desc, struct sk_buff *skb,
761 unsigned int offset, size_t len)
762 {
763 struct nvme_tcp_queue *queue = desc->arg.data;
764 size_t consumed = len;
765 int result;
766
767 while (len) {
768 switch (nvme_tcp_recv_state(queue)) {
769 case NVME_TCP_RECV_PDU:
770 result = nvme_tcp_recv_pdu(queue, skb, &offset, &len);
771 break;
772 case NVME_TCP_RECV_DATA:
773 result = nvme_tcp_recv_data(queue, skb, &offset, &len);
774 break;
775 case NVME_TCP_RECV_DDGST:
776 result = nvme_tcp_recv_ddgst(queue, skb, &offset, &len);
777 break;
778 default:
779 result = -EFAULT;
780 }
781 if (result) {
782 dev_err(queue->ctrl->ctrl.device,
783 "receive failed: %d\n", result);
784 queue->rd_enabled = false;
785 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
786 return result;
787 }
788 }
789
790 return consumed;
791 }
792
793 static void nvme_tcp_data_ready(struct sock *sk)
794 {
795 struct nvme_tcp_queue *queue;
796
797 read_lock_bh(&sk->sk_callback_lock);
798 queue = sk->sk_user_data;
799 if (likely(queue && queue->rd_enabled))
800 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
801 read_unlock_bh(&sk->sk_callback_lock);
802 }
803
804 static void nvme_tcp_write_space(struct sock *sk)
805 {
806 struct nvme_tcp_queue *queue;
807
808 read_lock_bh(&sk->sk_callback_lock);
809 queue = sk->sk_user_data;
810 if (likely(queue && sk_stream_is_writeable(sk))) {
811 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
812 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
813 }
814 read_unlock_bh(&sk->sk_callback_lock);
815 }
816
817 static void nvme_tcp_state_change(struct sock *sk)
818 {
819 struct nvme_tcp_queue *queue;
820
821 read_lock(&sk->sk_callback_lock);
822 queue = sk->sk_user_data;
823 if (!queue)
824 goto done;
825
826 switch (sk->sk_state) {
827 case TCP_CLOSE:
828 case TCP_CLOSE_WAIT:
829 case TCP_LAST_ACK:
830 case TCP_FIN_WAIT1:
831 case TCP_FIN_WAIT2:
832 /* fallthrough */
833 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
834 break;
835 default:
836 dev_info(queue->ctrl->ctrl.device,
837 "queue %d socket state %d\n",
838 nvme_tcp_queue_id(queue), sk->sk_state);
839 }
840
841 queue->state_change(sk);
842 done:
843 read_unlock(&sk->sk_callback_lock);
844 }
845
846 static inline void nvme_tcp_done_send_req(struct nvme_tcp_queue *queue)
847 {
848 queue->request = NULL;
849 }
850
851 static void nvme_tcp_fail_request(struct nvme_tcp_request *req)
852 {
853 nvme_tcp_end_request(blk_mq_rq_from_pdu(req), NVME_SC_HOST_PATH_ERROR);
854 }
855
856 static int nvme_tcp_try_send_data(struct nvme_tcp_request *req)
857 {
858 struct nvme_tcp_queue *queue = req->queue;
859
860 while (true) {
861 struct page *page = nvme_tcp_req_cur_page(req);
862 size_t offset = nvme_tcp_req_cur_offset(req);
863 size_t len = nvme_tcp_req_cur_length(req);
864 bool last = nvme_tcp_pdu_last_send(req, len);
865 int ret, flags = MSG_DONTWAIT;
866
867 if (last && !queue->data_digest)
868 flags |= MSG_EOR;
869 else
870 flags |= MSG_MORE;
871
872 /* can't zcopy slab pages */
873 if (unlikely(PageSlab(page))) {
874 ret = sock_no_sendpage(queue->sock, page, offset, len,
875 flags);
876 } else {
877 ret = kernel_sendpage(queue->sock, page, offset, len,
878 flags);
879 }
880 if (ret <= 0)
881 return ret;
882
883 nvme_tcp_advance_req(req, ret);
884 if (queue->data_digest)
885 nvme_tcp_ddgst_update(queue->snd_hash, page,
886 offset, ret);
887
888 /* fully successful last write*/
889 if (last && ret == len) {
890 if (queue->data_digest) {
891 nvme_tcp_ddgst_final(queue->snd_hash,
892 &req->ddgst);
893 req->state = NVME_TCP_SEND_DDGST;
894 req->offset = 0;
895 } else {
896 nvme_tcp_done_send_req(queue);
897 }
898 return 1;
899 }
900 }
901 return -EAGAIN;
902 }
903
904 static int nvme_tcp_try_send_cmd_pdu(struct nvme_tcp_request *req)
905 {
906 struct nvme_tcp_queue *queue = req->queue;
907 struct nvme_tcp_cmd_pdu *pdu = req->pdu;
908 bool inline_data = nvme_tcp_has_inline_data(req);
909 int flags = MSG_DONTWAIT | (inline_data ? MSG_MORE : MSG_EOR);
910 u8 hdgst = nvme_tcp_hdgst_len(queue);
911 int len = sizeof(*pdu) + hdgst - req->offset;
912 int ret;
913
914 if (queue->hdr_digest && !req->offset)
915 nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
916
917 ret = kernel_sendpage(queue->sock, virt_to_page(pdu),
918 offset_in_page(pdu) + req->offset, len, flags);
919 if (unlikely(ret <= 0))
920 return ret;
921
922 len -= ret;
923 if (!len) {
924 if (inline_data) {
925 req->state = NVME_TCP_SEND_DATA;
926 if (queue->data_digest)
927 crypto_ahash_init(queue->snd_hash);
928 nvme_tcp_init_iter(req, WRITE);
929 } else {
930 nvme_tcp_done_send_req(queue);
931 }
932 return 1;
933 }
934 req->offset += ret;
935
936 return -EAGAIN;
937 }
938
939 static int nvme_tcp_try_send_data_pdu(struct nvme_tcp_request *req)
940 {
941 struct nvme_tcp_queue *queue = req->queue;
942 struct nvme_tcp_data_pdu *pdu = req->pdu;
943 u8 hdgst = nvme_tcp_hdgst_len(queue);
944 int len = sizeof(*pdu) - req->offset + hdgst;
945 int ret;
946
947 if (queue->hdr_digest && !req->offset)
948 nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
949
950 ret = kernel_sendpage(queue->sock, virt_to_page(pdu),
951 offset_in_page(pdu) + req->offset, len,
952 MSG_DONTWAIT | MSG_MORE);
953 if (unlikely(ret <= 0))
954 return ret;
955
956 len -= ret;
957 if (!len) {
958 req->state = NVME_TCP_SEND_DATA;
959 if (queue->data_digest)
960 crypto_ahash_init(queue->snd_hash);
961 if (!req->data_sent)
962 nvme_tcp_init_iter(req, WRITE);
963 return 1;
964 }
965 req->offset += ret;
966
967 return -EAGAIN;
968 }
969
970 static int nvme_tcp_try_send_ddgst(struct nvme_tcp_request *req)
971 {
972 struct nvme_tcp_queue *queue = req->queue;
973 int ret;
974 struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_EOR };
975 struct kvec iov = {
976 .iov_base = &req->ddgst + req->offset,
977 .iov_len = NVME_TCP_DIGEST_LENGTH - req->offset
978 };
979
980 ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
981 if (unlikely(ret <= 0))
982 return ret;
983
984 if (req->offset + ret == NVME_TCP_DIGEST_LENGTH) {
985 nvme_tcp_done_send_req(queue);
986 return 1;
987 }
988
989 req->offset += ret;
990 return -EAGAIN;
991 }
992
993 static int nvme_tcp_try_send(struct nvme_tcp_queue *queue)
994 {
995 struct nvme_tcp_request *req;
996 int ret = 1;
997
998 if (!queue->request) {
999 queue->request = nvme_tcp_fetch_request(queue);
1000 if (!queue->request)
1001 return 0;
1002 }
1003 req = queue->request;
1004
1005 if (req->state == NVME_TCP_SEND_CMD_PDU) {
1006 ret = nvme_tcp_try_send_cmd_pdu(req);
1007 if (ret <= 0)
1008 goto done;
1009 if (!nvme_tcp_has_inline_data(req))
1010 return ret;
1011 }
1012
1013 if (req->state == NVME_TCP_SEND_H2C_PDU) {
1014 ret = nvme_tcp_try_send_data_pdu(req);
1015 if (ret <= 0)
1016 goto done;
1017 }
1018
1019 if (req->state == NVME_TCP_SEND_DATA) {
1020 ret = nvme_tcp_try_send_data(req);
1021 if (ret <= 0)
1022 goto done;
1023 }
1024
1025 if (req->state == NVME_TCP_SEND_DDGST)
1026 ret = nvme_tcp_try_send_ddgst(req);
1027 done:
1028 if (ret == -EAGAIN) {
1029 ret = 0;
1030 } else if (ret < 0) {
1031 dev_err(queue->ctrl->ctrl.device,
1032 "failed to send request %d\n", ret);
1033 if (ret != -EPIPE && ret != -ECONNRESET)
1034 nvme_tcp_fail_request(queue->request);
1035 nvme_tcp_done_send_req(queue);
1036 }
1037 return ret;
1038 }
1039
1040 static int nvme_tcp_try_recv(struct nvme_tcp_queue *queue)
1041 {
1042 struct socket *sock = queue->sock;
1043 struct sock *sk = sock->sk;
1044 read_descriptor_t rd_desc;
1045 int consumed;
1046
1047 rd_desc.arg.data = queue;
1048 rd_desc.count = 1;
1049 lock_sock(sk);
1050 queue->nr_cqe = 0;
1051 consumed = sock->ops->read_sock(sk, &rd_desc, nvme_tcp_recv_skb);
1052 release_sock(sk);
1053 return consumed;
1054 }
1055
1056 static void nvme_tcp_io_work(struct work_struct *w)
1057 {
1058 struct nvme_tcp_queue *queue =
1059 container_of(w, struct nvme_tcp_queue, io_work);
1060 unsigned long deadline = jiffies + msecs_to_jiffies(1);
1061
1062 do {
1063 bool pending = false;
1064 int result;
1065
1066 result = nvme_tcp_try_send(queue);
1067 if (result > 0)
1068 pending = true;
1069 else if (unlikely(result < 0))
1070 break;
1071
1072 result = nvme_tcp_try_recv(queue);
1073 if (result > 0)
1074 pending = true;
1075 else if (unlikely(result < 0))
1076 return;
1077
1078 if (!pending)
1079 return;
1080
1081 } while (!time_after(jiffies, deadline)); /* quota is exhausted */
1082
1083 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
1084 }
1085
1086 static void nvme_tcp_free_crypto(struct nvme_tcp_queue *queue)
1087 {
1088 struct crypto_ahash *tfm = crypto_ahash_reqtfm(queue->rcv_hash);
1089
1090 ahash_request_free(queue->rcv_hash);
1091 ahash_request_free(queue->snd_hash);
1092 crypto_free_ahash(tfm);
1093 }
1094
1095 static int nvme_tcp_alloc_crypto(struct nvme_tcp_queue *queue)
1096 {
1097 struct crypto_ahash *tfm;
1098
1099 tfm = crypto_alloc_ahash("crc32c", 0, CRYPTO_ALG_ASYNC);
1100 if (IS_ERR(tfm))
1101 return PTR_ERR(tfm);
1102
1103 queue->snd_hash = ahash_request_alloc(tfm, GFP_KERNEL);
1104 if (!queue->snd_hash)
1105 goto free_tfm;
1106 ahash_request_set_callback(queue->snd_hash, 0, NULL, NULL);
1107
1108 queue->rcv_hash = ahash_request_alloc(tfm, GFP_KERNEL);
1109 if (!queue->rcv_hash)
1110 goto free_snd_hash;
1111 ahash_request_set_callback(queue->rcv_hash, 0, NULL, NULL);
1112
1113 return 0;
1114 free_snd_hash:
1115 ahash_request_free(queue->snd_hash);
1116 free_tfm:
1117 crypto_free_ahash(tfm);
1118 return -ENOMEM;
1119 }
1120
1121 static void nvme_tcp_free_async_req(struct nvme_tcp_ctrl *ctrl)
1122 {
1123 struct nvme_tcp_request *async = &ctrl->async_req;
1124
1125 page_frag_free(async->pdu);
1126 }
1127
1128 static int nvme_tcp_alloc_async_req(struct nvme_tcp_ctrl *ctrl)
1129 {
1130 struct nvme_tcp_queue *queue = &ctrl->queues[0];
1131 struct nvme_tcp_request *async = &ctrl->async_req;
1132 u8 hdgst = nvme_tcp_hdgst_len(queue);
1133
1134 async->pdu = page_frag_alloc(&queue->pf_cache,
1135 sizeof(struct nvme_tcp_cmd_pdu) + hdgst,
1136 GFP_KERNEL | __GFP_ZERO);
1137 if (!async->pdu)
1138 return -ENOMEM;
1139
1140 async->queue = &ctrl->queues[0];
1141 return 0;
1142 }
1143
1144 static void nvme_tcp_free_queue(struct nvme_ctrl *nctrl, int qid)
1145 {
1146 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1147 struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1148
1149 if (!test_and_clear_bit(NVME_TCP_Q_ALLOCATED, &queue->flags))
1150 return;
1151
1152 if (queue->hdr_digest || queue->data_digest)
1153 nvme_tcp_free_crypto(queue);
1154
1155 sock_release(queue->sock);
1156 kfree(queue->pdu);
1157 }
1158
1159 static int nvme_tcp_init_connection(struct nvme_tcp_queue *queue)
1160 {
1161 struct nvme_tcp_icreq_pdu *icreq;
1162 struct nvme_tcp_icresp_pdu *icresp;
1163 struct msghdr msg = {};
1164 struct kvec iov;
1165 bool ctrl_hdgst, ctrl_ddgst;
1166 int ret;
1167
1168 icreq = kzalloc(sizeof(*icreq), GFP_KERNEL);
1169 if (!icreq)
1170 return -ENOMEM;
1171
1172 icresp = kzalloc(sizeof(*icresp), GFP_KERNEL);
1173 if (!icresp) {
1174 ret = -ENOMEM;
1175 goto free_icreq;
1176 }
1177
1178 icreq->hdr.type = nvme_tcp_icreq;
1179 icreq->hdr.hlen = sizeof(*icreq);
1180 icreq->hdr.pdo = 0;
1181 icreq->hdr.plen = cpu_to_le32(icreq->hdr.hlen);
1182 icreq->pfv = cpu_to_le16(NVME_TCP_PFV_1_0);
1183 icreq->maxr2t = 0; /* single inflight r2t supported */
1184 icreq->hpda = 0; /* no alignment constraint */
1185 if (queue->hdr_digest)
1186 icreq->digest |= NVME_TCP_HDR_DIGEST_ENABLE;
1187 if (queue->data_digest)
1188 icreq->digest |= NVME_TCP_DATA_DIGEST_ENABLE;
1189
1190 iov.iov_base = icreq;
1191 iov.iov_len = sizeof(*icreq);
1192 ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
1193 if (ret < 0)
1194 goto free_icresp;
1195
1196 memset(&msg, 0, sizeof(msg));
1197 iov.iov_base = icresp;
1198 iov.iov_len = sizeof(*icresp);
1199 ret = kernel_recvmsg(queue->sock, &msg, &iov, 1,
1200 iov.iov_len, msg.msg_flags);
1201 if (ret < 0)
1202 goto free_icresp;
1203
1204 ret = -EINVAL;
1205 if (icresp->hdr.type != nvme_tcp_icresp) {
1206 pr_err("queue %d: bad type returned %d\n",
1207 nvme_tcp_queue_id(queue), icresp->hdr.type);
1208 goto free_icresp;
1209 }
1210
1211 if (le32_to_cpu(icresp->hdr.plen) != sizeof(*icresp)) {
1212 pr_err("queue %d: bad pdu length returned %d\n",
1213 nvme_tcp_queue_id(queue), icresp->hdr.plen);
1214 goto free_icresp;
1215 }
1216
1217 if (icresp->pfv != NVME_TCP_PFV_1_0) {
1218 pr_err("queue %d: bad pfv returned %d\n",
1219 nvme_tcp_queue_id(queue), icresp->pfv);
1220 goto free_icresp;
1221 }
1222
1223 ctrl_ddgst = !!(icresp->digest & NVME_TCP_DATA_DIGEST_ENABLE);
1224 if ((queue->data_digest && !ctrl_ddgst) ||
1225 (!queue->data_digest && ctrl_ddgst)) {
1226 pr_err("queue %d: data digest mismatch host: %s ctrl: %s\n",
1227 nvme_tcp_queue_id(queue),
1228 queue->data_digest ? "enabled" : "disabled",
1229 ctrl_ddgst ? "enabled" : "disabled");
1230 goto free_icresp;
1231 }
1232
1233 ctrl_hdgst = !!(icresp->digest & NVME_TCP_HDR_DIGEST_ENABLE);
1234 if ((queue->hdr_digest && !ctrl_hdgst) ||
1235 (!queue->hdr_digest && ctrl_hdgst)) {
1236 pr_err("queue %d: header digest mismatch host: %s ctrl: %s\n",
1237 nvme_tcp_queue_id(queue),
1238 queue->hdr_digest ? "enabled" : "disabled",
1239 ctrl_hdgst ? "enabled" : "disabled");
1240 goto free_icresp;
1241 }
1242
1243 if (icresp->cpda != 0) {
1244 pr_err("queue %d: unsupported cpda returned %d\n",
1245 nvme_tcp_queue_id(queue), icresp->cpda);
1246 goto free_icresp;
1247 }
1248
1249 ret = 0;
1250 free_icresp:
1251 kfree(icresp);
1252 free_icreq:
1253 kfree(icreq);
1254 return ret;
1255 }
1256
1257 static bool nvme_tcp_admin_queue(struct nvme_tcp_queue *queue)
1258 {
1259 return nvme_tcp_queue_id(queue) == 0;
1260 }
1261
1262 static bool nvme_tcp_default_queue(struct nvme_tcp_queue *queue)
1263 {
1264 struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1265 int qid = nvme_tcp_queue_id(queue);
1266
1267 return !nvme_tcp_admin_queue(queue) &&
1268 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT];
1269 }
1270
1271 static bool nvme_tcp_read_queue(struct nvme_tcp_queue *queue)
1272 {
1273 struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1274 int qid = nvme_tcp_queue_id(queue);
1275
1276 return !nvme_tcp_admin_queue(queue) &&
1277 !nvme_tcp_default_queue(queue) &&
1278 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT] +
1279 ctrl->io_queues[HCTX_TYPE_READ];
1280 }
1281
1282 static bool nvme_tcp_poll_queue(struct nvme_tcp_queue *queue)
1283 {
1284 struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1285 int qid = nvme_tcp_queue_id(queue);
1286
1287 return !nvme_tcp_admin_queue(queue) &&
1288 !nvme_tcp_default_queue(queue) &&
1289 !nvme_tcp_read_queue(queue) &&
1290 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT] +
1291 ctrl->io_queues[HCTX_TYPE_READ] +
1292 ctrl->io_queues[HCTX_TYPE_POLL];
1293 }
1294
1295 static void nvme_tcp_set_queue_io_cpu(struct nvme_tcp_queue *queue)
1296 {
1297 struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1298 int qid = nvme_tcp_queue_id(queue);
1299 int n = 0;
1300
1301 if (nvme_tcp_default_queue(queue))
1302 n = qid - 1;
1303 else if (nvme_tcp_read_queue(queue))
1304 n = qid - ctrl->io_queues[HCTX_TYPE_DEFAULT] - 1;
1305 else if (nvme_tcp_poll_queue(queue))
1306 n = qid - ctrl->io_queues[HCTX_TYPE_DEFAULT] -
1307 ctrl->io_queues[HCTX_TYPE_READ] - 1;
1308 queue->io_cpu = cpumask_next_wrap(n - 1, cpu_online_mask, -1, false);
1309 }
1310
1311 static int nvme_tcp_alloc_queue(struct nvme_ctrl *nctrl,
1312 int qid, size_t queue_size)
1313 {
1314 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1315 struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1316 struct linger sol = { .l_onoff = 1, .l_linger = 0 };
1317 int ret, opt, rcv_pdu_size;
1318
1319 queue->ctrl = ctrl;
1320 INIT_LIST_HEAD(&queue->send_list);
1321 spin_lock_init(&queue->lock);
1322 INIT_WORK(&queue->io_work, nvme_tcp_io_work);
1323 queue->queue_size = queue_size;
1324
1325 if (qid > 0)
1326 queue->cmnd_capsule_len = nctrl->ioccsz * 16;
1327 else
1328 queue->cmnd_capsule_len = sizeof(struct nvme_command) +
1329 NVME_TCP_ADMIN_CCSZ;
1330
1331 ret = sock_create(ctrl->addr.ss_family, SOCK_STREAM,
1332 IPPROTO_TCP, &queue->sock);
1333 if (ret) {
1334 dev_err(nctrl->device,
1335 "failed to create socket: %d\n", ret);
1336 return ret;
1337 }
1338
1339 /* Single syn retry */
1340 opt = 1;
1341 ret = kernel_setsockopt(queue->sock, IPPROTO_TCP, TCP_SYNCNT,
1342 (char *)&opt, sizeof(opt));
1343 if (ret) {
1344 dev_err(nctrl->device,
1345 "failed to set TCP_SYNCNT sock opt %d\n", ret);
1346 goto err_sock;
1347 }
1348
1349 /* Set TCP no delay */
1350 opt = 1;
1351 ret = kernel_setsockopt(queue->sock, IPPROTO_TCP,
1352 TCP_NODELAY, (char *)&opt, sizeof(opt));
1353 if (ret) {
1354 dev_err(nctrl->device,
1355 "failed to set TCP_NODELAY sock opt %d\n", ret);
1356 goto err_sock;
1357 }
1358
1359 /*
1360 * Cleanup whatever is sitting in the TCP transmit queue on socket
1361 * close. This is done to prevent stale data from being sent should
1362 * the network connection be restored before TCP times out.
1363 */
1364 ret = kernel_setsockopt(queue->sock, SOL_SOCKET, SO_LINGER,
1365 (char *)&sol, sizeof(sol));
1366 if (ret) {
1367 dev_err(nctrl->device,
1368 "failed to set SO_LINGER sock opt %d\n", ret);
1369 goto err_sock;
1370 }
1371
1372 if (so_priority > 0) {
1373 ret = kernel_setsockopt(queue->sock, SOL_SOCKET, SO_PRIORITY,
1374 (char *)&so_priority, sizeof(so_priority));
1375 if (ret) {
1376 dev_err(ctrl->ctrl.device,
1377 "failed to set SO_PRIORITY sock opt, ret %d\n",
1378 ret);
1379 goto err_sock;
1380 }
1381 }
1382
1383 /* Set socket type of service */
1384 if (nctrl->opts->tos >= 0) {
1385 opt = nctrl->opts->tos;
1386 ret = kernel_setsockopt(queue->sock, SOL_IP, IP_TOS,
1387 (char *)&opt, sizeof(opt));
1388 if (ret) {
1389 dev_err(nctrl->device,
1390 "failed to set IP_TOS sock opt %d\n", ret);
1391 goto err_sock;
1392 }
1393 }
1394
1395 queue->sock->sk->sk_allocation = GFP_ATOMIC;
1396 nvme_tcp_set_queue_io_cpu(queue);
1397 queue->request = NULL;
1398 queue->data_remaining = 0;
1399 queue->ddgst_remaining = 0;
1400 queue->pdu_remaining = 0;
1401 queue->pdu_offset = 0;
1402 sk_set_memalloc(queue->sock->sk);
1403
1404 if (nctrl->opts->mask & NVMF_OPT_HOST_TRADDR) {
1405 ret = kernel_bind(queue->sock, (struct sockaddr *)&ctrl->src_addr,
1406 sizeof(ctrl->src_addr));
1407 if (ret) {
1408 dev_err(nctrl->device,
1409 "failed to bind queue %d socket %d\n",
1410 qid, ret);
1411 goto err_sock;
1412 }
1413 }
1414
1415 queue->hdr_digest = nctrl->opts->hdr_digest;
1416 queue->data_digest = nctrl->opts->data_digest;
1417 if (queue->hdr_digest || queue->data_digest) {
1418 ret = nvme_tcp_alloc_crypto(queue);
1419 if (ret) {
1420 dev_err(nctrl->device,
1421 "failed to allocate queue %d crypto\n", qid);
1422 goto err_sock;
1423 }
1424 }
1425
1426 rcv_pdu_size = sizeof(struct nvme_tcp_rsp_pdu) +
1427 nvme_tcp_hdgst_len(queue);
1428 queue->pdu = kmalloc(rcv_pdu_size, GFP_KERNEL);
1429 if (!queue->pdu) {
1430 ret = -ENOMEM;
1431 goto err_crypto;
1432 }
1433
1434 dev_dbg(nctrl->device, "connecting queue %d\n",
1435 nvme_tcp_queue_id(queue));
1436
1437 ret = kernel_connect(queue->sock, (struct sockaddr *)&ctrl->addr,
1438 sizeof(ctrl->addr), 0);
1439 if (ret) {
1440 dev_err(nctrl->device,
1441 "failed to connect socket: %d\n", ret);
1442 goto err_rcv_pdu;
1443 }
1444
1445 ret = nvme_tcp_init_connection(queue);
1446 if (ret)
1447 goto err_init_connect;
1448
1449 queue->rd_enabled = true;
1450 set_bit(NVME_TCP_Q_ALLOCATED, &queue->flags);
1451 nvme_tcp_init_recv_ctx(queue);
1452
1453 write_lock_bh(&queue->sock->sk->sk_callback_lock);
1454 queue->sock->sk->sk_user_data = queue;
1455 queue->state_change = queue->sock->sk->sk_state_change;
1456 queue->data_ready = queue->sock->sk->sk_data_ready;
1457 queue->write_space = queue->sock->sk->sk_write_space;
1458 queue->sock->sk->sk_data_ready = nvme_tcp_data_ready;
1459 queue->sock->sk->sk_state_change = nvme_tcp_state_change;
1460 queue->sock->sk->sk_write_space = nvme_tcp_write_space;
1461 #ifdef CONFIG_NET_RX_BUSY_POLL
1462 queue->sock->sk->sk_ll_usec = 1;
1463 #endif
1464 write_unlock_bh(&queue->sock->sk->sk_callback_lock);
1465
1466 return 0;
1467
1468 err_init_connect:
1469 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1470 err_rcv_pdu:
1471 kfree(queue->pdu);
1472 err_crypto:
1473 if (queue->hdr_digest || queue->data_digest)
1474 nvme_tcp_free_crypto(queue);
1475 err_sock:
1476 sock_release(queue->sock);
1477 queue->sock = NULL;
1478 return ret;
1479 }
1480
1481 static void nvme_tcp_restore_sock_calls(struct nvme_tcp_queue *queue)
1482 {
1483 struct socket *sock = queue->sock;
1484
1485 write_lock_bh(&sock->sk->sk_callback_lock);
1486 sock->sk->sk_user_data = NULL;
1487 sock->sk->sk_data_ready = queue->data_ready;
1488 sock->sk->sk_state_change = queue->state_change;
1489 sock->sk->sk_write_space = queue->write_space;
1490 write_unlock_bh(&sock->sk->sk_callback_lock);
1491 }
1492
1493 static void __nvme_tcp_stop_queue(struct nvme_tcp_queue *queue)
1494 {
1495 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1496 nvme_tcp_restore_sock_calls(queue);
1497 cancel_work_sync(&queue->io_work);
1498 }
1499
1500 static void nvme_tcp_stop_queue(struct nvme_ctrl *nctrl, int qid)
1501 {
1502 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1503 struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1504
1505 if (!test_and_clear_bit(NVME_TCP_Q_LIVE, &queue->flags))
1506 return;
1507
1508 __nvme_tcp_stop_queue(queue);
1509 }
1510
1511 static int nvme_tcp_start_queue(struct nvme_ctrl *nctrl, int idx)
1512 {
1513 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1514 int ret;
1515
1516 if (idx)
1517 ret = nvmf_connect_io_queue(nctrl, idx, false);
1518 else
1519 ret = nvmf_connect_admin_queue(nctrl);
1520
1521 if (!ret) {
1522 set_bit(NVME_TCP_Q_LIVE, &ctrl->queues[idx].flags);
1523 } else {
1524 if (test_bit(NVME_TCP_Q_ALLOCATED, &ctrl->queues[idx].flags))
1525 __nvme_tcp_stop_queue(&ctrl->queues[idx]);
1526 dev_err(nctrl->device,
1527 "failed to connect queue: %d ret=%d\n", idx, ret);
1528 }
1529 return ret;
1530 }
1531
1532 static struct blk_mq_tag_set *nvme_tcp_alloc_tagset(struct nvme_ctrl *nctrl,
1533 bool admin)
1534 {
1535 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1536 struct blk_mq_tag_set *set;
1537 int ret;
1538
1539 if (admin) {
1540 set = &ctrl->admin_tag_set;
1541 memset(set, 0, sizeof(*set));
1542 set->ops = &nvme_tcp_admin_mq_ops;
1543 set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
1544 set->reserved_tags = 2; /* connect + keep-alive */
1545 set->numa_node = NUMA_NO_NODE;
1546 set->cmd_size = sizeof(struct nvme_tcp_request);
1547 set->driver_data = ctrl;
1548 set->nr_hw_queues = 1;
1549 set->timeout = ADMIN_TIMEOUT;
1550 } else {
1551 set = &ctrl->tag_set;
1552 memset(set, 0, sizeof(*set));
1553 set->ops = &nvme_tcp_mq_ops;
1554 set->queue_depth = nctrl->sqsize + 1;
1555 set->reserved_tags = 1; /* fabric connect */
1556 set->numa_node = NUMA_NO_NODE;
1557 set->flags = BLK_MQ_F_SHOULD_MERGE;
1558 set->cmd_size = sizeof(struct nvme_tcp_request);
1559 set->driver_data = ctrl;
1560 set->nr_hw_queues = nctrl->queue_count - 1;
1561 set->timeout = NVME_IO_TIMEOUT;
1562 set->nr_maps = nctrl->opts->nr_poll_queues ? HCTX_MAX_TYPES : 2;
1563 }
1564
1565 ret = blk_mq_alloc_tag_set(set);
1566 if (ret)
1567 return ERR_PTR(ret);
1568
1569 return set;
1570 }
1571
1572 static void nvme_tcp_free_admin_queue(struct nvme_ctrl *ctrl)
1573 {
1574 if (to_tcp_ctrl(ctrl)->async_req.pdu) {
1575 nvme_tcp_free_async_req(to_tcp_ctrl(ctrl));
1576 to_tcp_ctrl(ctrl)->async_req.pdu = NULL;
1577 }
1578
1579 nvme_tcp_free_queue(ctrl, 0);
1580 }
1581
1582 static void nvme_tcp_free_io_queues(struct nvme_ctrl *ctrl)
1583 {
1584 int i;
1585
1586 for (i = 1; i < ctrl->queue_count; i++)
1587 nvme_tcp_free_queue(ctrl, i);
1588 }
1589
1590 static void nvme_tcp_stop_io_queues(struct nvme_ctrl *ctrl)
1591 {
1592 int i;
1593
1594 for (i = 1; i < ctrl->queue_count; i++)
1595 nvme_tcp_stop_queue(ctrl, i);
1596 }
1597
1598 static int nvme_tcp_start_io_queues(struct nvme_ctrl *ctrl)
1599 {
1600 int i, ret = 0;
1601
1602 for (i = 1; i < ctrl->queue_count; i++) {
1603 ret = nvme_tcp_start_queue(ctrl, i);
1604 if (ret)
1605 goto out_stop_queues;
1606 }
1607
1608 return 0;
1609
1610 out_stop_queues:
1611 for (i--; i >= 1; i--)
1612 nvme_tcp_stop_queue(ctrl, i);
1613 return ret;
1614 }
1615
1616 static int nvme_tcp_alloc_admin_queue(struct nvme_ctrl *ctrl)
1617 {
1618 int ret;
1619
1620 ret = nvme_tcp_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
1621 if (ret)
1622 return ret;
1623
1624 ret = nvme_tcp_alloc_async_req(to_tcp_ctrl(ctrl));
1625 if (ret)
1626 goto out_free_queue;
1627
1628 return 0;
1629
1630 out_free_queue:
1631 nvme_tcp_free_queue(ctrl, 0);
1632 return ret;
1633 }
1634
1635 static int __nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
1636 {
1637 int i, ret;
1638
1639 for (i = 1; i < ctrl->queue_count; i++) {
1640 ret = nvme_tcp_alloc_queue(ctrl, i,
1641 ctrl->sqsize + 1);
1642 if (ret)
1643 goto out_free_queues;
1644 }
1645
1646 return 0;
1647
1648 out_free_queues:
1649 for (i--; i >= 1; i--)
1650 nvme_tcp_free_queue(ctrl, i);
1651
1652 return ret;
1653 }
1654
1655 static unsigned int nvme_tcp_nr_io_queues(struct nvme_ctrl *ctrl)
1656 {
1657 unsigned int nr_io_queues;
1658
1659 nr_io_queues = min(ctrl->opts->nr_io_queues, num_online_cpus());
1660 nr_io_queues += min(ctrl->opts->nr_write_queues, num_online_cpus());
1661 nr_io_queues += min(ctrl->opts->nr_poll_queues, num_online_cpus());
1662
1663 return nr_io_queues;
1664 }
1665
1666 static void nvme_tcp_set_io_queues(struct nvme_ctrl *nctrl,
1667 unsigned int nr_io_queues)
1668 {
1669 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1670 struct nvmf_ctrl_options *opts = nctrl->opts;
1671
1672 if (opts->nr_write_queues && opts->nr_io_queues < nr_io_queues) {
1673 /*
1674 * separate read/write queues
1675 * hand out dedicated default queues only after we have
1676 * sufficient read queues.
1677 */
1678 ctrl->io_queues[HCTX_TYPE_READ] = opts->nr_io_queues;
1679 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_READ];
1680 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
1681 min(opts->nr_write_queues, nr_io_queues);
1682 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
1683 } else {
1684 /*
1685 * shared read/write queues
1686 * either no write queues were requested, or we don't have
1687 * sufficient queue count to have dedicated default queues.
1688 */
1689 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
1690 min(opts->nr_io_queues, nr_io_queues);
1691 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
1692 }
1693
1694 if (opts->nr_poll_queues && nr_io_queues) {
1695 /* map dedicated poll queues only if we have queues left */
1696 ctrl->io_queues[HCTX_TYPE_POLL] =
1697 min(opts->nr_poll_queues, nr_io_queues);
1698 }
1699 }
1700
1701 static int nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
1702 {
1703 unsigned int nr_io_queues;
1704 int ret;
1705
1706 nr_io_queues = nvme_tcp_nr_io_queues(ctrl);
1707 ret = nvme_set_queue_count(ctrl, &nr_io_queues);
1708 if (ret)
1709 return ret;
1710
1711 ctrl->queue_count = nr_io_queues + 1;
1712 if (ctrl->queue_count < 2)
1713 return 0;
1714
1715 dev_info(ctrl->device,
1716 "creating %d I/O queues.\n", nr_io_queues);
1717
1718 nvme_tcp_set_io_queues(ctrl, nr_io_queues);
1719
1720 return __nvme_tcp_alloc_io_queues(ctrl);
1721 }
1722
1723 static void nvme_tcp_destroy_io_queues(struct nvme_ctrl *ctrl, bool remove)
1724 {
1725 nvme_tcp_stop_io_queues(ctrl);
1726 if (remove) {
1727 blk_cleanup_queue(ctrl->connect_q);
1728 blk_mq_free_tag_set(ctrl->tagset);
1729 }
1730 nvme_tcp_free_io_queues(ctrl);
1731 }
1732
1733 static int nvme_tcp_configure_io_queues(struct nvme_ctrl *ctrl, bool new)
1734 {
1735 int ret;
1736
1737 ret = nvme_tcp_alloc_io_queues(ctrl);
1738 if (ret)
1739 return ret;
1740
1741 if (new) {
1742 ctrl->tagset = nvme_tcp_alloc_tagset(ctrl, false);
1743 if (IS_ERR(ctrl->tagset)) {
1744 ret = PTR_ERR(ctrl->tagset);
1745 goto out_free_io_queues;
1746 }
1747
1748 ctrl->connect_q = blk_mq_init_queue(ctrl->tagset);
1749 if (IS_ERR(ctrl->connect_q)) {
1750 ret = PTR_ERR(ctrl->connect_q);
1751 goto out_free_tag_set;
1752 }
1753 } else {
1754 blk_mq_update_nr_hw_queues(ctrl->tagset,
1755 ctrl->queue_count - 1);
1756 }
1757
1758 ret = nvme_tcp_start_io_queues(ctrl);
1759 if (ret)
1760 goto out_cleanup_connect_q;
1761
1762 return 0;
1763
1764 out_cleanup_connect_q:
1765 if (new)
1766 blk_cleanup_queue(ctrl->connect_q);
1767 out_free_tag_set:
1768 if (new)
1769 blk_mq_free_tag_set(ctrl->tagset);
1770 out_free_io_queues:
1771 nvme_tcp_free_io_queues(ctrl);
1772 return ret;
1773 }
1774
1775 static void nvme_tcp_destroy_admin_queue(struct nvme_ctrl *ctrl, bool remove)
1776 {
1777 nvme_tcp_stop_queue(ctrl, 0);
1778 if (remove) {
1779 blk_cleanup_queue(ctrl->admin_q);
1780 blk_cleanup_queue(ctrl->fabrics_q);
1781 blk_mq_free_tag_set(ctrl->admin_tagset);
1782 }
1783 nvme_tcp_free_admin_queue(ctrl);
1784 }
1785
1786 static int nvme_tcp_configure_admin_queue(struct nvme_ctrl *ctrl, bool new)
1787 {
1788 int error;
1789
1790 error = nvme_tcp_alloc_admin_queue(ctrl);
1791 if (error)
1792 return error;
1793
1794 if (new) {
1795 ctrl->admin_tagset = nvme_tcp_alloc_tagset(ctrl, true);
1796 if (IS_ERR(ctrl->admin_tagset)) {
1797 error = PTR_ERR(ctrl->admin_tagset);
1798 goto out_free_queue;
1799 }
1800
1801 ctrl->fabrics_q = blk_mq_init_queue(ctrl->admin_tagset);
1802 if (IS_ERR(ctrl->fabrics_q)) {
1803 error = PTR_ERR(ctrl->fabrics_q);
1804 goto out_free_tagset;
1805 }
1806
1807 ctrl->admin_q = blk_mq_init_queue(ctrl->admin_tagset);
1808 if (IS_ERR(ctrl->admin_q)) {
1809 error = PTR_ERR(ctrl->admin_q);
1810 goto out_cleanup_fabrics_q;
1811 }
1812 }
1813
1814 error = nvme_tcp_start_queue(ctrl, 0);
1815 if (error)
1816 goto out_cleanup_queue;
1817
1818 error = nvme_enable_ctrl(ctrl);
1819 if (error)
1820 goto out_stop_queue;
1821
1822 blk_mq_unquiesce_queue(ctrl->admin_q);
1823
1824 error = nvme_init_identify(ctrl);
1825 if (error)
1826 goto out_stop_queue;
1827
1828 return 0;
1829
1830 out_stop_queue:
1831 nvme_tcp_stop_queue(ctrl, 0);
1832 out_cleanup_queue:
1833 if (new)
1834 blk_cleanup_queue(ctrl->admin_q);
1835 out_cleanup_fabrics_q:
1836 if (new)
1837 blk_cleanup_queue(ctrl->fabrics_q);
1838 out_free_tagset:
1839 if (new)
1840 blk_mq_free_tag_set(ctrl->admin_tagset);
1841 out_free_queue:
1842 nvme_tcp_free_admin_queue(ctrl);
1843 return error;
1844 }
1845
1846 static void nvme_tcp_teardown_admin_queue(struct nvme_ctrl *ctrl,
1847 bool remove)
1848 {
1849 blk_mq_quiesce_queue(ctrl->admin_q);
1850 nvme_tcp_stop_queue(ctrl, 0);
1851 if (ctrl->admin_tagset) {
1852 blk_mq_tagset_busy_iter(ctrl->admin_tagset,
1853 nvme_cancel_request, ctrl);
1854 blk_mq_tagset_wait_completed_request(ctrl->admin_tagset);
1855 }
1856 if (remove)
1857 blk_mq_unquiesce_queue(ctrl->admin_q);
1858 nvme_tcp_destroy_admin_queue(ctrl, remove);
1859 }
1860
1861 static void nvme_tcp_teardown_io_queues(struct nvme_ctrl *ctrl,
1862 bool remove)
1863 {
1864 if (ctrl->queue_count <= 1)
1865 return;
1866 nvme_stop_queues(ctrl);
1867 nvme_tcp_stop_io_queues(ctrl);
1868 if (ctrl->tagset) {
1869 blk_mq_tagset_busy_iter(ctrl->tagset,
1870 nvme_cancel_request, ctrl);
1871 blk_mq_tagset_wait_completed_request(ctrl->tagset);
1872 }
1873 if (remove)
1874 nvme_start_queues(ctrl);
1875 nvme_tcp_destroy_io_queues(ctrl, remove);
1876 }
1877
1878 static void nvme_tcp_reconnect_or_remove(struct nvme_ctrl *ctrl)
1879 {
1880 /* If we are resetting/deleting then do nothing */
1881 if (ctrl->state != NVME_CTRL_CONNECTING) {
1882 WARN_ON_ONCE(ctrl->state == NVME_CTRL_NEW ||
1883 ctrl->state == NVME_CTRL_LIVE);
1884 return;
1885 }
1886
1887 if (nvmf_should_reconnect(ctrl)) {
1888 dev_info(ctrl->device, "Reconnecting in %d seconds...\n",
1889 ctrl->opts->reconnect_delay);
1890 queue_delayed_work(nvme_wq, &to_tcp_ctrl(ctrl)->connect_work,
1891 ctrl->opts->reconnect_delay * HZ);
1892 } else {
1893 dev_info(ctrl->device, "Removing controller...\n");
1894 nvme_delete_ctrl(ctrl);
1895 }
1896 }
1897
1898 static int nvme_tcp_setup_ctrl(struct nvme_ctrl *ctrl, bool new)
1899 {
1900 struct nvmf_ctrl_options *opts = ctrl->opts;
1901 int ret;
1902
1903 ret = nvme_tcp_configure_admin_queue(ctrl, new);
1904 if (ret)
1905 return ret;
1906
1907 if (ctrl->icdoff) {
1908 dev_err(ctrl->device, "icdoff is not supported!\n");
1909 goto destroy_admin;
1910 }
1911
1912 if (opts->queue_size > ctrl->sqsize + 1)
1913 dev_warn(ctrl->device,
1914 "queue_size %zu > ctrl sqsize %u, clamping down\n",
1915 opts->queue_size, ctrl->sqsize + 1);
1916
1917 if (ctrl->sqsize + 1 > ctrl->maxcmd) {
1918 dev_warn(ctrl->device,
1919 "sqsize %u > ctrl maxcmd %u, clamping down\n",
1920 ctrl->sqsize + 1, ctrl->maxcmd);
1921 ctrl->sqsize = ctrl->maxcmd - 1;
1922 }
1923
1924 if (ctrl->queue_count > 1) {
1925 ret = nvme_tcp_configure_io_queues(ctrl, new);
1926 if (ret)
1927 goto destroy_admin;
1928 }
1929
1930 if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_LIVE)) {
1931 /*
1932 * state change failure is ok if we're in DELETING state,
1933 * unless we're during creation of a new controller to
1934 * avoid races with teardown flow.
1935 */
1936 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING);
1937 WARN_ON_ONCE(new);
1938 ret = -EINVAL;
1939 goto destroy_io;
1940 }
1941
1942 nvme_start_ctrl(ctrl);
1943 return 0;
1944
1945 destroy_io:
1946 if (ctrl->queue_count > 1)
1947 nvme_tcp_destroy_io_queues(ctrl, new);
1948 destroy_admin:
1949 nvme_tcp_stop_queue(ctrl, 0);
1950 nvme_tcp_destroy_admin_queue(ctrl, new);
1951 return ret;
1952 }
1953
1954 static void nvme_tcp_reconnect_ctrl_work(struct work_struct *work)
1955 {
1956 struct nvme_tcp_ctrl *tcp_ctrl = container_of(to_delayed_work(work),
1957 struct nvme_tcp_ctrl, connect_work);
1958 struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
1959
1960 ++ctrl->nr_reconnects;
1961
1962 if (nvme_tcp_setup_ctrl(ctrl, false))
1963 goto requeue;
1964
1965 dev_info(ctrl->device, "Successfully reconnected (%d attempt)\n",
1966 ctrl->nr_reconnects);
1967
1968 ctrl->nr_reconnects = 0;
1969
1970 return;
1971
1972 requeue:
1973 dev_info(ctrl->device, "Failed reconnect attempt %d\n",
1974 ctrl->nr_reconnects);
1975 nvme_tcp_reconnect_or_remove(ctrl);
1976 }
1977
1978 static void nvme_tcp_error_recovery_work(struct work_struct *work)
1979 {
1980 struct nvme_tcp_ctrl *tcp_ctrl = container_of(work,
1981 struct nvme_tcp_ctrl, err_work);
1982 struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
1983
1984 nvme_stop_keep_alive(ctrl);
1985 nvme_tcp_teardown_io_queues(ctrl, false);
1986 /* unquiesce to fail fast pending requests */
1987 nvme_start_queues(ctrl);
1988 nvme_tcp_teardown_admin_queue(ctrl, false);
1989 blk_mq_unquiesce_queue(ctrl->admin_q);
1990
1991 if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
1992 /* state change failure is ok if we're in DELETING state */
1993 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING);
1994 return;
1995 }
1996
1997 nvme_tcp_reconnect_or_remove(ctrl);
1998 }
1999
2000 static void nvme_tcp_teardown_ctrl(struct nvme_ctrl *ctrl, bool shutdown)
2001 {
2002 cancel_work_sync(&to_tcp_ctrl(ctrl)->err_work);
2003 cancel_delayed_work_sync(&to_tcp_ctrl(ctrl)->connect_work);
2004
2005 nvme_tcp_teardown_io_queues(ctrl, shutdown);
2006 blk_mq_quiesce_queue(ctrl->admin_q);
2007 if (shutdown)
2008 nvme_shutdown_ctrl(ctrl);
2009 else
2010 nvme_disable_ctrl(ctrl);
2011 nvme_tcp_teardown_admin_queue(ctrl, shutdown);
2012 }
2013
2014 static void nvme_tcp_delete_ctrl(struct nvme_ctrl *ctrl)
2015 {
2016 nvme_tcp_teardown_ctrl(ctrl, true);
2017 }
2018
2019 static void nvme_reset_ctrl_work(struct work_struct *work)
2020 {
2021 struct nvme_ctrl *ctrl =
2022 container_of(work, struct nvme_ctrl, reset_work);
2023
2024 nvme_stop_ctrl(ctrl);
2025 nvme_tcp_teardown_ctrl(ctrl, false);
2026
2027 if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
2028 /* state change failure is ok if we're in DELETING state */
2029 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING);
2030 return;
2031 }
2032
2033 if (nvme_tcp_setup_ctrl(ctrl, false))
2034 goto out_fail;
2035
2036 return;
2037
2038 out_fail:
2039 ++ctrl->nr_reconnects;
2040 nvme_tcp_reconnect_or_remove(ctrl);
2041 }
2042
2043 static void nvme_tcp_free_ctrl(struct nvme_ctrl *nctrl)
2044 {
2045 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
2046
2047 if (list_empty(&ctrl->list))
2048 goto free_ctrl;
2049
2050 mutex_lock(&nvme_tcp_ctrl_mutex);
2051 list_del(&ctrl->list);
2052 mutex_unlock(&nvme_tcp_ctrl_mutex);
2053
2054 nvmf_free_options(nctrl->opts);
2055 free_ctrl:
2056 kfree(ctrl->queues);
2057 kfree(ctrl);
2058 }
2059
2060 static void nvme_tcp_set_sg_null(struct nvme_command *c)
2061 {
2062 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2063
2064 sg->addr = 0;
2065 sg->length = 0;
2066 sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
2067 NVME_SGL_FMT_TRANSPORT_A;
2068 }
2069
2070 static void nvme_tcp_set_sg_inline(struct nvme_tcp_queue *queue,
2071 struct nvme_command *c, u32 data_len)
2072 {
2073 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2074
2075 sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
2076 sg->length = cpu_to_le32(data_len);
2077 sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
2078 }
2079
2080 static void nvme_tcp_set_sg_host_data(struct nvme_command *c,
2081 u32 data_len)
2082 {
2083 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2084
2085 sg->addr = 0;
2086 sg->length = cpu_to_le32(data_len);
2087 sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
2088 NVME_SGL_FMT_TRANSPORT_A;
2089 }
2090
2091 static void nvme_tcp_submit_async_event(struct nvme_ctrl *arg)
2092 {
2093 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(arg);
2094 struct nvme_tcp_queue *queue = &ctrl->queues[0];
2095 struct nvme_tcp_cmd_pdu *pdu = ctrl->async_req.pdu;
2096 struct nvme_command *cmd = &pdu->cmd;
2097 u8 hdgst = nvme_tcp_hdgst_len(queue);
2098
2099 memset(pdu, 0, sizeof(*pdu));
2100 pdu->hdr.type = nvme_tcp_cmd;
2101 if (queue->hdr_digest)
2102 pdu->hdr.flags |= NVME_TCP_F_HDGST;
2103 pdu->hdr.hlen = sizeof(*pdu);
2104 pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst);
2105
2106 cmd->common.opcode = nvme_admin_async_event;
2107 cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
2108 cmd->common.flags |= NVME_CMD_SGL_METABUF;
2109 nvme_tcp_set_sg_null(cmd);
2110
2111 ctrl->async_req.state = NVME_TCP_SEND_CMD_PDU;
2112 ctrl->async_req.offset = 0;
2113 ctrl->async_req.curr_bio = NULL;
2114 ctrl->async_req.data_len = 0;
2115
2116 nvme_tcp_queue_request(&ctrl->async_req);
2117 }
2118
2119 static enum blk_eh_timer_return
2120 nvme_tcp_timeout(struct request *rq, bool reserved)
2121 {
2122 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2123 struct nvme_tcp_ctrl *ctrl = req->queue->ctrl;
2124 struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2125
2126 /*
2127 * Restart the timer if a controller reset is already scheduled. Any
2128 * timed out commands would be handled before entering the connecting
2129 * state.
2130 */
2131 if (ctrl->ctrl.state == NVME_CTRL_RESETTING)
2132 return BLK_EH_RESET_TIMER;
2133
2134 dev_warn(ctrl->ctrl.device,
2135 "queue %d: timeout request %#x type %d\n",
2136 nvme_tcp_queue_id(req->queue), rq->tag, pdu->hdr.type);
2137
2138 if (ctrl->ctrl.state != NVME_CTRL_LIVE) {
2139 /*
2140 * Teardown immediately if controller times out while starting
2141 * or we are already started error recovery. all outstanding
2142 * requests are completed on shutdown, so we return BLK_EH_DONE.
2143 */
2144 flush_work(&ctrl->err_work);
2145 nvme_tcp_teardown_io_queues(&ctrl->ctrl, false);
2146 nvme_tcp_teardown_admin_queue(&ctrl->ctrl, false);
2147 return BLK_EH_DONE;
2148 }
2149
2150 dev_warn(ctrl->ctrl.device, "starting error recovery\n");
2151 nvme_tcp_error_recovery(&ctrl->ctrl);
2152
2153 return BLK_EH_RESET_TIMER;
2154 }
2155
2156 static blk_status_t nvme_tcp_map_data(struct nvme_tcp_queue *queue,
2157 struct request *rq)
2158 {
2159 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2160 struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2161 struct nvme_command *c = &pdu->cmd;
2162
2163 c->common.flags |= NVME_CMD_SGL_METABUF;
2164
2165 if (!blk_rq_nr_phys_segments(rq))
2166 nvme_tcp_set_sg_null(c);
2167 else if (rq_data_dir(rq) == WRITE &&
2168 req->data_len <= nvme_tcp_inline_data_size(queue))
2169 nvme_tcp_set_sg_inline(queue, c, req->data_len);
2170 else
2171 nvme_tcp_set_sg_host_data(c, req->data_len);
2172
2173 return 0;
2174 }
2175
2176 static blk_status_t nvme_tcp_setup_cmd_pdu(struct nvme_ns *ns,
2177 struct request *rq)
2178 {
2179 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2180 struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2181 struct nvme_tcp_queue *queue = req->queue;
2182 u8 hdgst = nvme_tcp_hdgst_len(queue), ddgst = 0;
2183 blk_status_t ret;
2184
2185 ret = nvme_setup_cmd(ns, rq, &pdu->cmd);
2186 if (ret)
2187 return ret;
2188
2189 req->state = NVME_TCP_SEND_CMD_PDU;
2190 req->offset = 0;
2191 req->data_sent = 0;
2192 req->pdu_len = 0;
2193 req->pdu_sent = 0;
2194 req->data_len = blk_rq_nr_phys_segments(rq) ?
2195 blk_rq_payload_bytes(rq) : 0;
2196 req->curr_bio = rq->bio;
2197
2198 if (rq_data_dir(rq) == WRITE &&
2199 req->data_len <= nvme_tcp_inline_data_size(queue))
2200 req->pdu_len = req->data_len;
2201 else if (req->curr_bio)
2202 nvme_tcp_init_iter(req, READ);
2203
2204 pdu->hdr.type = nvme_tcp_cmd;
2205 pdu->hdr.flags = 0;
2206 if (queue->hdr_digest)
2207 pdu->hdr.flags |= NVME_TCP_F_HDGST;
2208 if (queue->data_digest && req->pdu_len) {
2209 pdu->hdr.flags |= NVME_TCP_F_DDGST;
2210 ddgst = nvme_tcp_ddgst_len(queue);
2211 }
2212 pdu->hdr.hlen = sizeof(*pdu);
2213 pdu->hdr.pdo = req->pdu_len ? pdu->hdr.hlen + hdgst : 0;
2214 pdu->hdr.plen =
2215 cpu_to_le32(pdu->hdr.hlen + hdgst + req->pdu_len + ddgst);
2216
2217 ret = nvme_tcp_map_data(queue, rq);
2218 if (unlikely(ret)) {
2219 nvme_cleanup_cmd(rq);
2220 dev_err(queue->ctrl->ctrl.device,
2221 "Failed to map data (%d)\n", ret);
2222 return ret;
2223 }
2224
2225 return 0;
2226 }
2227
2228 static blk_status_t nvme_tcp_queue_rq(struct blk_mq_hw_ctx *hctx,
2229 const struct blk_mq_queue_data *bd)
2230 {
2231 struct nvme_ns *ns = hctx->queue->queuedata;
2232 struct nvme_tcp_queue *queue = hctx->driver_data;
2233 struct request *rq = bd->rq;
2234 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2235 bool queue_ready = test_bit(NVME_TCP_Q_LIVE, &queue->flags);
2236 blk_status_t ret;
2237
2238 if (!nvmf_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
2239 return nvmf_fail_nonready_command(&queue->ctrl->ctrl, rq);
2240
2241 ret = nvme_tcp_setup_cmd_pdu(ns, rq);
2242 if (unlikely(ret))
2243 return ret;
2244
2245 blk_mq_start_request(rq);
2246
2247 nvme_tcp_queue_request(req);
2248
2249 return BLK_STS_OK;
2250 }
2251
2252 static int nvme_tcp_map_queues(struct blk_mq_tag_set *set)
2253 {
2254 struct nvme_tcp_ctrl *ctrl = set->driver_data;
2255 struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
2256
2257 if (opts->nr_write_queues && ctrl->io_queues[HCTX_TYPE_READ]) {
2258 /* separate read/write queues */
2259 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2260 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2261 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2262 set->map[HCTX_TYPE_READ].nr_queues =
2263 ctrl->io_queues[HCTX_TYPE_READ];
2264 set->map[HCTX_TYPE_READ].queue_offset =
2265 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2266 } else {
2267 /* shared read/write queues */
2268 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2269 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2270 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2271 set->map[HCTX_TYPE_READ].nr_queues =
2272 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2273 set->map[HCTX_TYPE_READ].queue_offset = 0;
2274 }
2275 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
2276 blk_mq_map_queues(&set->map[HCTX_TYPE_READ]);
2277
2278 if (opts->nr_poll_queues && ctrl->io_queues[HCTX_TYPE_POLL]) {
2279 /* map dedicated poll queues only if we have queues left */
2280 set->map[HCTX_TYPE_POLL].nr_queues =
2281 ctrl->io_queues[HCTX_TYPE_POLL];
2282 set->map[HCTX_TYPE_POLL].queue_offset =
2283 ctrl->io_queues[HCTX_TYPE_DEFAULT] +
2284 ctrl->io_queues[HCTX_TYPE_READ];
2285 blk_mq_map_queues(&set->map[HCTX_TYPE_POLL]);
2286 }
2287
2288 dev_info(ctrl->ctrl.device,
2289 "mapped %d/%d/%d default/read/poll queues.\n",
2290 ctrl->io_queues[HCTX_TYPE_DEFAULT],
2291 ctrl->io_queues[HCTX_TYPE_READ],
2292 ctrl->io_queues[HCTX_TYPE_POLL]);
2293
2294 return 0;
2295 }
2296
2297 static int nvme_tcp_poll(struct blk_mq_hw_ctx *hctx)
2298 {
2299 struct nvme_tcp_queue *queue = hctx->driver_data;
2300 struct sock *sk = queue->sock->sk;
2301
2302 if (!test_bit(NVME_TCP_Q_LIVE, &queue->flags))
2303 return 0;
2304
2305 if (sk_can_busy_loop(sk) && skb_queue_empty_lockless(&sk->sk_receive_queue))
2306 sk_busy_loop(sk, true);
2307 nvme_tcp_try_recv(queue);
2308 return queue->nr_cqe;
2309 }
2310
2311 static struct blk_mq_ops nvme_tcp_mq_ops = {
2312 .queue_rq = nvme_tcp_queue_rq,
2313 .complete = nvme_complete_rq,
2314 .init_request = nvme_tcp_init_request,
2315 .exit_request = nvme_tcp_exit_request,
2316 .init_hctx = nvme_tcp_init_hctx,
2317 .timeout = nvme_tcp_timeout,
2318 .map_queues = nvme_tcp_map_queues,
2319 .poll = nvme_tcp_poll,
2320 };
2321
2322 static struct blk_mq_ops nvme_tcp_admin_mq_ops = {
2323 .queue_rq = nvme_tcp_queue_rq,
2324 .complete = nvme_complete_rq,
2325 .init_request = nvme_tcp_init_request,
2326 .exit_request = nvme_tcp_exit_request,
2327 .init_hctx = nvme_tcp_init_admin_hctx,
2328 .timeout = nvme_tcp_timeout,
2329 };
2330
2331 static const struct nvme_ctrl_ops nvme_tcp_ctrl_ops = {
2332 .name = "tcp",
2333 .module = THIS_MODULE,
2334 .flags = NVME_F_FABRICS,
2335 .reg_read32 = nvmf_reg_read32,
2336 .reg_read64 = nvmf_reg_read64,
2337 .reg_write32 = nvmf_reg_write32,
2338 .free_ctrl = nvme_tcp_free_ctrl,
2339 .submit_async_event = nvme_tcp_submit_async_event,
2340 .delete_ctrl = nvme_tcp_delete_ctrl,
2341 .get_address = nvmf_get_address,
2342 };
2343
2344 static bool
2345 nvme_tcp_existing_controller(struct nvmf_ctrl_options *opts)
2346 {
2347 struct nvme_tcp_ctrl *ctrl;
2348 bool found = false;
2349
2350 mutex_lock(&nvme_tcp_ctrl_mutex);
2351 list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list) {
2352 found = nvmf_ip_options_match(&ctrl->ctrl, opts);
2353 if (found)
2354 break;
2355 }
2356 mutex_unlock(&nvme_tcp_ctrl_mutex);
2357
2358 return found;
2359 }
2360
2361 static struct nvme_ctrl *nvme_tcp_create_ctrl(struct device *dev,
2362 struct nvmf_ctrl_options *opts)
2363 {
2364 struct nvme_tcp_ctrl *ctrl;
2365 int ret;
2366
2367 ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
2368 if (!ctrl)
2369 return ERR_PTR(-ENOMEM);
2370
2371 INIT_LIST_HEAD(&ctrl->list);
2372 ctrl->ctrl.opts = opts;
2373 ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues +
2374 opts->nr_poll_queues + 1;
2375 ctrl->ctrl.sqsize = opts->queue_size - 1;
2376 ctrl->ctrl.kato = opts->kato;
2377
2378 INIT_DELAYED_WORK(&ctrl->connect_work,
2379 nvme_tcp_reconnect_ctrl_work);
2380 INIT_WORK(&ctrl->err_work, nvme_tcp_error_recovery_work);
2381 INIT_WORK(&ctrl->ctrl.reset_work, nvme_reset_ctrl_work);
2382
2383 if (!(opts->mask & NVMF_OPT_TRSVCID)) {
2384 opts->trsvcid =
2385 kstrdup(__stringify(NVME_TCP_DISC_PORT), GFP_KERNEL);
2386 if (!opts->trsvcid) {
2387 ret = -ENOMEM;
2388 goto out_free_ctrl;
2389 }
2390 opts->mask |= NVMF_OPT_TRSVCID;
2391 }
2392
2393 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2394 opts->traddr, opts->trsvcid, &ctrl->addr);
2395 if (ret) {
2396 pr_err("malformed address passed: %s:%s\n",
2397 opts->traddr, opts->trsvcid);
2398 goto out_free_ctrl;
2399 }
2400
2401 if (opts->mask & NVMF_OPT_HOST_TRADDR) {
2402 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2403 opts->host_traddr, NULL, &ctrl->src_addr);
2404 if (ret) {
2405 pr_err("malformed src address passed: %s\n",
2406 opts->host_traddr);
2407 goto out_free_ctrl;
2408 }
2409 }
2410
2411 if (!opts->duplicate_connect && nvme_tcp_existing_controller(opts)) {
2412 ret = -EALREADY;
2413 goto out_free_ctrl;
2414 }
2415
2416 ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
2417 GFP_KERNEL);
2418 if (!ctrl->queues) {
2419 ret = -ENOMEM;
2420 goto out_free_ctrl;
2421 }
2422
2423 ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_tcp_ctrl_ops, 0);
2424 if (ret)
2425 goto out_kfree_queues;
2426
2427 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
2428 WARN_ON_ONCE(1);
2429 ret = -EINTR;
2430 goto out_uninit_ctrl;
2431 }
2432
2433 ret = nvme_tcp_setup_ctrl(&ctrl->ctrl, true);
2434 if (ret)
2435 goto out_uninit_ctrl;
2436
2437 dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISp\n",
2438 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
2439
2440 mutex_lock(&nvme_tcp_ctrl_mutex);
2441 list_add_tail(&ctrl->list, &nvme_tcp_ctrl_list);
2442 mutex_unlock(&nvme_tcp_ctrl_mutex);
2443
2444 return &ctrl->ctrl;
2445
2446 out_uninit_ctrl:
2447 nvme_uninit_ctrl(&ctrl->ctrl);
2448 nvme_put_ctrl(&ctrl->ctrl);
2449 if (ret > 0)
2450 ret = -EIO;
2451 return ERR_PTR(ret);
2452 out_kfree_queues:
2453 kfree(ctrl->queues);
2454 out_free_ctrl:
2455 kfree(ctrl);
2456 return ERR_PTR(ret);
2457 }
2458
2459 static struct nvmf_transport_ops nvme_tcp_transport = {
2460 .name = "tcp",
2461 .module = THIS_MODULE,
2462 .required_opts = NVMF_OPT_TRADDR,
2463 .allowed_opts = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2464 NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
2465 NVMF_OPT_HDR_DIGEST | NVMF_OPT_DATA_DIGEST |
2466 NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES |
2467 NVMF_OPT_TOS,
2468 .create_ctrl = nvme_tcp_create_ctrl,
2469 };
2470
2471 static int __init nvme_tcp_init_module(void)
2472 {
2473 nvme_tcp_wq = alloc_workqueue("nvme_tcp_wq",
2474 WQ_MEM_RECLAIM | WQ_HIGHPRI, 0);
2475 if (!nvme_tcp_wq)
2476 return -ENOMEM;
2477
2478 nvmf_register_transport(&nvme_tcp_transport);
2479 return 0;
2480 }
2481
2482 static void __exit nvme_tcp_cleanup_module(void)
2483 {
2484 struct nvme_tcp_ctrl *ctrl;
2485
2486 nvmf_unregister_transport(&nvme_tcp_transport);
2487
2488 mutex_lock(&nvme_tcp_ctrl_mutex);
2489 list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list)
2490 nvme_delete_ctrl(&ctrl->ctrl);
2491 mutex_unlock(&nvme_tcp_ctrl_mutex);
2492 flush_workqueue(nvme_delete_wq);
2493
2494 destroy_workqueue(nvme_tcp_wq);
2495 }
2496
2497 module_init(nvme_tcp_init_module);
2498 module_exit(nvme_tcp_cleanup_module);
2499
2500 MODULE_LICENSE("GPL v2");
Linux with added FPC-III support