?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
treewide: remove redundant IS_ERR() before error code check
[linux/fpc-iii.git] / drivers / nvme / host / tcp.c
blob6d43b23a0fc8bc15d2870da5fa850802ac9b8356
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 enum nvme_tcp_send_state {
24 NVME_TCP_SEND_CMD_PDU = 0,
25 NVME_TCP_SEND_H2C_PDU,
26 NVME_TCP_SEND_DATA,
27 NVME_TCP_SEND_DDGST,
28 };
29
30 struct nvme_tcp_request {
31 struct nvme_request req;
32 void *pdu;
33 struct nvme_tcp_queue *queue;
34 u32 data_len;
35 u32 pdu_len;
36 u32 pdu_sent;
37 u16 ttag;
38 struct list_head entry;
39 __le32 ddgst;
40
41 struct bio *curr_bio;
42 struct iov_iter iter;
43
44 /* send state */
45 size_t offset;
46 size_t data_sent;
47 enum nvme_tcp_send_state state;
48 };
49
50 enum nvme_tcp_queue_flags {
51 NVME_TCP_Q_ALLOCATED = 0,
52 NVME_TCP_Q_LIVE = 1,
53 };
54
55 enum nvme_tcp_recv_state {
56 NVME_TCP_RECV_PDU = 0,
57 NVME_TCP_RECV_DATA,
58 NVME_TCP_RECV_DDGST,
59 };
60
61 struct nvme_tcp_ctrl;
62 struct nvme_tcp_queue {
63 struct socket *sock;
64 struct work_struct io_work;
65 int io_cpu;
66
67 spinlock_t lock;
68 struct list_head send_list;
69
70 /* recv state */
71 void *pdu;
72 int pdu_remaining;
73 int pdu_offset;
74 size_t data_remaining;
75 size_t ddgst_remaining;
76 unsigned int nr_cqe;
77
78 /* send state */
79 struct nvme_tcp_request *request;
80
81 int queue_size;
82 size_t cmnd_capsule_len;
83 struct nvme_tcp_ctrl *ctrl;
84 unsigned long flags;
85 bool rd_enabled;
86
87 bool hdr_digest;
88 bool data_digest;
89 struct ahash_request *rcv_hash;
90 struct ahash_request *snd_hash;
91 __le32 exp_ddgst;
92 __le32 recv_ddgst;
93
94 struct page_frag_cache pf_cache;
95
96 void (*state_change)(struct sock *);
97 void (*data_ready)(struct sock *);
98 void (*write_space)(struct sock *);
99 };
100
101 struct nvme_tcp_ctrl {
102 /* read only in the hot path */
103 struct nvme_tcp_queue *queues;
104 struct blk_mq_tag_set tag_set;
105
106 /* other member variables */
107 struct list_head list;
108 struct blk_mq_tag_set admin_tag_set;
109 struct sockaddr_storage addr;
110 struct sockaddr_storage src_addr;
111 struct nvme_ctrl ctrl;
112
113 struct work_struct err_work;
114 struct delayed_work connect_work;
115 struct nvme_tcp_request async_req;
116 u32 io_queues[HCTX_MAX_TYPES];
117 };
118
119 static LIST_HEAD(nvme_tcp_ctrl_list);
120 static DEFINE_MUTEX(nvme_tcp_ctrl_mutex);
121 static struct workqueue_struct *nvme_tcp_wq;
122 static struct blk_mq_ops nvme_tcp_mq_ops;
123 static struct blk_mq_ops nvme_tcp_admin_mq_ops;
124
125 static inline struct nvme_tcp_ctrl *to_tcp_ctrl(struct nvme_ctrl *ctrl)
126 {
127 return container_of(ctrl, struct nvme_tcp_ctrl, ctrl);
128 }
129
130 static inline int nvme_tcp_queue_id(struct nvme_tcp_queue *queue)
131 {
132 return queue - queue->ctrl->queues;
133 }
134
135 static inline struct blk_mq_tags *nvme_tcp_tagset(struct nvme_tcp_queue *queue)
136 {
137 u32 queue_idx = nvme_tcp_queue_id(queue);
138
139 if (queue_idx == 0)
140 return queue->ctrl->admin_tag_set.tags[queue_idx];
141 return queue->ctrl->tag_set.tags[queue_idx - 1];
142 }
143
144 static inline u8 nvme_tcp_hdgst_len(struct nvme_tcp_queue *queue)
145 {
146 return queue->hdr_digest ? NVME_TCP_DIGEST_LENGTH : 0;
147 }
148
149 static inline u8 nvme_tcp_ddgst_len(struct nvme_tcp_queue *queue)
150 {
151 return queue->data_digest ? NVME_TCP_DIGEST_LENGTH : 0;
152 }
153
154 static inline size_t nvme_tcp_inline_data_size(struct nvme_tcp_queue *queue)
155 {
156 return queue->cmnd_capsule_len - sizeof(struct nvme_command);
157 }
158
159 static inline bool nvme_tcp_async_req(struct nvme_tcp_request *req)
160 {
161 return req == &req->queue->ctrl->async_req;
162 }
163
164 static inline bool nvme_tcp_has_inline_data(struct nvme_tcp_request *req)
165 {
166 struct request *rq;
167 unsigned int bytes;
168
169 if (unlikely(nvme_tcp_async_req(req)))
170 return false; /* async events don't have a request */
171
172 rq = blk_mq_rq_from_pdu(req);
173 bytes = blk_rq_payload_bytes(rq);
174
175 return rq_data_dir(rq) == WRITE && bytes &&
176 bytes <= nvme_tcp_inline_data_size(req->queue);
177 }
178
179 static inline struct page *nvme_tcp_req_cur_page(struct nvme_tcp_request *req)
180 {
181 return req->iter.bvec->bv_page;
182 }
183
184 static inline size_t nvme_tcp_req_cur_offset(struct nvme_tcp_request *req)
185 {
186 return req->iter.bvec->bv_offset + req->iter.iov_offset;
187 }
188
189 static inline size_t nvme_tcp_req_cur_length(struct nvme_tcp_request *req)
190 {
191 return min_t(size_t, req->iter.bvec->bv_len - req->iter.iov_offset,
192 req->pdu_len - req->pdu_sent);
193 }
194
195 static inline size_t nvme_tcp_req_offset(struct nvme_tcp_request *req)
196 {
197 return req->iter.iov_offset;
198 }
199
200 static inline size_t nvme_tcp_pdu_data_left(struct nvme_tcp_request *req)
201 {
202 return rq_data_dir(blk_mq_rq_from_pdu(req)) == WRITE ?
203 req->pdu_len - req->pdu_sent : 0;
204 }
205
206 static inline size_t nvme_tcp_pdu_last_send(struct nvme_tcp_request *req,
207 int len)
208 {
209 return nvme_tcp_pdu_data_left(req) <= len;
210 }
211
212 static void nvme_tcp_init_iter(struct nvme_tcp_request *req,
213 unsigned int dir)
214 {
215 struct request *rq = blk_mq_rq_from_pdu(req);
216 struct bio_vec *vec;
217 unsigned int size;
218 int nsegs;
219 size_t offset;
220
221 if (rq->rq_flags & RQF_SPECIAL_PAYLOAD) {
222 vec = &rq->special_vec;
223 nsegs = 1;
224 size = blk_rq_payload_bytes(rq);
225 offset = 0;
226 } else {
227 struct bio *bio = req->curr_bio;
228
229 vec = __bvec_iter_bvec(bio->bi_io_vec, bio->bi_iter);
230 nsegs = bio_segments(bio);
231 size = bio->bi_iter.bi_size;
232 offset = bio->bi_iter.bi_bvec_done;
233 }
234
235 iov_iter_bvec(&req->iter, dir, vec, nsegs, size);
236 req->iter.iov_offset = offset;
237 }
238
239 static inline void nvme_tcp_advance_req(struct nvme_tcp_request *req,
240 int len)
241 {
242 req->data_sent += len;
243 req->pdu_sent += len;
244 iov_iter_advance(&req->iter, len);
245 if (!iov_iter_count(&req->iter) &&
246 req->data_sent < req->data_len) {
247 req->curr_bio = req->curr_bio->bi_next;
248 nvme_tcp_init_iter(req, WRITE);
249 }
250 }
251
252 static inline void nvme_tcp_queue_request(struct nvme_tcp_request *req)
253 {
254 struct nvme_tcp_queue *queue = req->queue;
255
256 spin_lock(&queue->lock);
257 list_add_tail(&req->entry, &queue->send_list);
258 spin_unlock(&queue->lock);
259
260 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
261 }
262
263 static inline struct nvme_tcp_request *
264 nvme_tcp_fetch_request(struct nvme_tcp_queue *queue)
265 {
266 struct nvme_tcp_request *req;
267
268 spin_lock(&queue->lock);
269 req = list_first_entry_or_null(&queue->send_list,
270 struct nvme_tcp_request, entry);
271 if (req)
272 list_del(&req->entry);
273 spin_unlock(&queue->lock);
274
275 return req;
276 }
277
278 static inline void nvme_tcp_ddgst_final(struct ahash_request *hash,
279 __le32 *dgst)
280 {
281 ahash_request_set_crypt(hash, NULL, (u8 *)dgst, 0);
282 crypto_ahash_final(hash);
283 }
284
285 static inline void nvme_tcp_ddgst_update(struct ahash_request *hash,
286 struct page *page, off_t off, size_t len)
287 {
288 struct scatterlist sg;
289
290 sg_init_marker(&sg, 1);
291 sg_set_page(&sg, page, len, off);
292 ahash_request_set_crypt(hash, &sg, NULL, len);
293 crypto_ahash_update(hash);
294 }
295
296 static inline void nvme_tcp_hdgst(struct ahash_request *hash,
297 void *pdu, size_t len)
298 {
299 struct scatterlist sg;
300
301 sg_init_one(&sg, pdu, len);
302 ahash_request_set_crypt(hash, &sg, pdu + len, len);
303 crypto_ahash_digest(hash);
304 }
305
306 static int nvme_tcp_verify_hdgst(struct nvme_tcp_queue *queue,
307 void *pdu, size_t pdu_len)
308 {
309 struct nvme_tcp_hdr *hdr = pdu;
310 __le32 recv_digest;
311 __le32 exp_digest;
312
313 if (unlikely(!(hdr->flags & NVME_TCP_F_HDGST))) {
314 dev_err(queue->ctrl->ctrl.device,
315 "queue %d: header digest flag is cleared\n",
316 nvme_tcp_queue_id(queue));
317 return -EPROTO;
318 }
319
320 recv_digest = *(__le32 *)(pdu + hdr->hlen);
321 nvme_tcp_hdgst(queue->rcv_hash, pdu, pdu_len);
322 exp_digest = *(__le32 *)(pdu + hdr->hlen);
323 if (recv_digest != exp_digest) {
324 dev_err(queue->ctrl->ctrl.device,
325 "header digest error: recv %#x expected %#x\n",
326 le32_to_cpu(recv_digest), le32_to_cpu(exp_digest));
327 return -EIO;
328 }
329
330 return 0;
331 }
332
333 static int nvme_tcp_check_ddgst(struct nvme_tcp_queue *queue, void *pdu)
334 {
335 struct nvme_tcp_hdr *hdr = pdu;
336 u8 digest_len = nvme_tcp_hdgst_len(queue);
337 u32 len;
338
339 len = le32_to_cpu(hdr->plen) - hdr->hlen -
340 ((hdr->flags & NVME_TCP_F_HDGST) ? digest_len : 0);
341
342 if (unlikely(len && !(hdr->flags & NVME_TCP_F_DDGST))) {
343 dev_err(queue->ctrl->ctrl.device,
344 "queue %d: data digest flag is cleared\n",
345 nvme_tcp_queue_id(queue));
346 return -EPROTO;
347 }
348 crypto_ahash_init(queue->rcv_hash);
349
350 return 0;
351 }
352
353 static void nvme_tcp_exit_request(struct blk_mq_tag_set *set,
354 struct request *rq, unsigned int hctx_idx)
355 {
356 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
357
358 page_frag_free(req->pdu);
359 }
360
361 static int nvme_tcp_init_request(struct blk_mq_tag_set *set,
362 struct request *rq, unsigned int hctx_idx,
363 unsigned int numa_node)
364 {
365 struct nvme_tcp_ctrl *ctrl = set->driver_data;
366 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
367 int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
368 struct nvme_tcp_queue *queue = &ctrl->queues[queue_idx];
369 u8 hdgst = nvme_tcp_hdgst_len(queue);
370
371 req->pdu = page_frag_alloc(&queue->pf_cache,
372 sizeof(struct nvme_tcp_cmd_pdu) + hdgst,
373 GFP_KERNEL | __GFP_ZERO);
374 if (!req->pdu)
375 return -ENOMEM;
376
377 req->queue = queue;
378 nvme_req(rq)->ctrl = &ctrl->ctrl;
379
380 return 0;
381 }
382
383 static int nvme_tcp_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
384 unsigned int hctx_idx)
385 {
386 struct nvme_tcp_ctrl *ctrl = data;
387 struct nvme_tcp_queue *queue = &ctrl->queues[hctx_idx + 1];
388
389 hctx->driver_data = queue;
390 return 0;
391 }
392
393 static int nvme_tcp_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
394 unsigned int hctx_idx)
395 {
396 struct nvme_tcp_ctrl *ctrl = data;
397 struct nvme_tcp_queue *queue = &ctrl->queues[0];
398
399 hctx->driver_data = queue;
400 return 0;
401 }
402
403 static enum nvme_tcp_recv_state
404 nvme_tcp_recv_state(struct nvme_tcp_queue *queue)
405 {
406 return (queue->pdu_remaining) ? NVME_TCP_RECV_PDU :
407 (queue->ddgst_remaining) ? NVME_TCP_RECV_DDGST :
408 NVME_TCP_RECV_DATA;
409 }
410
411 static void nvme_tcp_init_recv_ctx(struct nvme_tcp_queue *queue)
412 {
413 queue->pdu_remaining = sizeof(struct nvme_tcp_rsp_pdu) +
414 nvme_tcp_hdgst_len(queue);
415 queue->pdu_offset = 0;
416 queue->data_remaining = -1;
417 queue->ddgst_remaining = 0;
418 }
419
420 static void nvme_tcp_error_recovery(struct nvme_ctrl *ctrl)
421 {
422 if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_RESETTING))
423 return;
424
425 queue_work(nvme_wq, &to_tcp_ctrl(ctrl)->err_work);
426 }
427
428 static int nvme_tcp_process_nvme_cqe(struct nvme_tcp_queue *queue,
429 struct nvme_completion *cqe)
430 {
431 struct request *rq;
432
433 rq = blk_mq_tag_to_rq(nvme_tcp_tagset(queue), cqe->command_id);
434 if (!rq) {
435 dev_err(queue->ctrl->ctrl.device,
436 "queue %d tag 0x%x not found\n",
437 nvme_tcp_queue_id(queue), cqe->command_id);
438 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
439 return -EINVAL;
440 }
441
442 nvme_end_request(rq, cqe->status, cqe->result);
443 queue->nr_cqe++;
444
445 return 0;
446 }
447
448 static int nvme_tcp_handle_c2h_data(struct nvme_tcp_queue *queue,
449 struct nvme_tcp_data_pdu *pdu)
450 {
451 struct request *rq;
452
453 rq = blk_mq_tag_to_rq(nvme_tcp_tagset(queue), pdu->command_id);
454 if (!rq) {
455 dev_err(queue->ctrl->ctrl.device,
456 "queue %d tag %#x not found\n",
457 nvme_tcp_queue_id(queue), pdu->command_id);
458 return -ENOENT;
459 }
460
461 if (!blk_rq_payload_bytes(rq)) {
462 dev_err(queue->ctrl->ctrl.device,
463 "queue %d tag %#x unexpected data\n",
464 nvme_tcp_queue_id(queue), rq->tag);
465 return -EIO;
466 }
467
468 queue->data_remaining = le32_to_cpu(pdu->data_length);
469
470 if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS &&
471 unlikely(!(pdu->hdr.flags & NVME_TCP_F_DATA_LAST))) {
472 dev_err(queue->ctrl->ctrl.device,
473 "queue %d tag %#x SUCCESS set but not last PDU\n",
474 nvme_tcp_queue_id(queue), rq->tag);
475 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
476 return -EPROTO;
477 }
478
479 return 0;
480 }
481
482 static int nvme_tcp_handle_comp(struct nvme_tcp_queue *queue,
483 struct nvme_tcp_rsp_pdu *pdu)
484 {
485 struct nvme_completion *cqe = &pdu->cqe;
486 int ret = 0;
487
488 /*
489 * AEN requests are special as they don't time out and can
490 * survive any kind of queue freeze and often don't respond to
491 * aborts. We don't even bother to allocate a struct request
492 * for them but rather special case them here.
493 */
494 if (unlikely(nvme_is_aen_req(nvme_tcp_queue_id(queue),
495 cqe->command_id)))
496 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
497 &cqe->result);
498 else
499 ret = nvme_tcp_process_nvme_cqe(queue, cqe);
500
501 return ret;
502 }
503
504 static int nvme_tcp_setup_h2c_data_pdu(struct nvme_tcp_request *req,
505 struct nvme_tcp_r2t_pdu *pdu)
506 {
507 struct nvme_tcp_data_pdu *data = req->pdu;
508 struct nvme_tcp_queue *queue = req->queue;
509 struct request *rq = blk_mq_rq_from_pdu(req);
510 u8 hdgst = nvme_tcp_hdgst_len(queue);
511 u8 ddgst = nvme_tcp_ddgst_len(queue);
512
513 req->pdu_len = le32_to_cpu(pdu->r2t_length);
514 req->pdu_sent = 0;
515
516 if (unlikely(req->data_sent + req->pdu_len > req->data_len)) {
517 dev_err(queue->ctrl->ctrl.device,
518 "req %d r2t len %u exceeded data len %u (%zu sent)\n",
519 rq->tag, req->pdu_len, req->data_len,
520 req->data_sent);
521 return -EPROTO;
522 }
523
524 if (unlikely(le32_to_cpu(pdu->r2t_offset) < req->data_sent)) {
525 dev_err(queue->ctrl->ctrl.device,
526 "req %d unexpected r2t offset %u (expected %zu)\n",
527 rq->tag, le32_to_cpu(pdu->r2t_offset),
528 req->data_sent);
529 return -EPROTO;
530 }
531
532 memset(data, 0, sizeof(*data));
533 data->hdr.type = nvme_tcp_h2c_data;
534 data->hdr.flags = NVME_TCP_F_DATA_LAST;
535 if (queue->hdr_digest)
536 data->hdr.flags |= NVME_TCP_F_HDGST;
537 if (queue->data_digest)
538 data->hdr.flags |= NVME_TCP_F_DDGST;
539 data->hdr.hlen = sizeof(*data);
540 data->hdr.pdo = data->hdr.hlen + hdgst;
541 data->hdr.plen =
542 cpu_to_le32(data->hdr.hlen + hdgst + req->pdu_len + ddgst);
543 data->ttag = pdu->ttag;
544 data->command_id = rq->tag;
545 data->data_offset = cpu_to_le32(req->data_sent);
546 data->data_length = cpu_to_le32(req->pdu_len);
547 return 0;
548 }
549
550 static int nvme_tcp_handle_r2t(struct nvme_tcp_queue *queue,
551 struct nvme_tcp_r2t_pdu *pdu)
552 {
553 struct nvme_tcp_request *req;
554 struct request *rq;
555 int ret;
556
557 rq = blk_mq_tag_to_rq(nvme_tcp_tagset(queue), pdu->command_id);
558 if (!rq) {
559 dev_err(queue->ctrl->ctrl.device,
560 "queue %d tag %#x not found\n",
561 nvme_tcp_queue_id(queue), pdu->command_id);
562 return -ENOENT;
563 }
564 req = blk_mq_rq_to_pdu(rq);
565
566 ret = nvme_tcp_setup_h2c_data_pdu(req, pdu);
567 if (unlikely(ret))
568 return ret;
569
570 req->state = NVME_TCP_SEND_H2C_PDU;
571 req->offset = 0;
572
573 nvme_tcp_queue_request(req);
574
575 return 0;
576 }
577
578 static int nvme_tcp_recv_pdu(struct nvme_tcp_queue *queue, struct sk_buff *skb,
579 unsigned int *offset, size_t *len)
580 {
581 struct nvme_tcp_hdr *hdr;
582 char *pdu = queue->pdu;
583 size_t rcv_len = min_t(size_t, *len, queue->pdu_remaining);
584 int ret;
585
586 ret = skb_copy_bits(skb, *offset,
587 &pdu[queue->pdu_offset], rcv_len);
588 if (unlikely(ret))
589 return ret;
590
591 queue->pdu_remaining -= rcv_len;
592 queue->pdu_offset += rcv_len;
593 *offset += rcv_len;
594 *len -= rcv_len;
595 if (queue->pdu_remaining)
596 return 0;
597
598 hdr = queue->pdu;
599 if (queue->hdr_digest) {
600 ret = nvme_tcp_verify_hdgst(queue, queue->pdu, hdr->hlen);
601 if (unlikely(ret))
602 return ret;
603 }
604
605
606 if (queue->data_digest) {
607 ret = nvme_tcp_check_ddgst(queue, queue->pdu);
608 if (unlikely(ret))
609 return ret;
610 }
611
612 switch (hdr->type) {
613 case nvme_tcp_c2h_data:
614 return nvme_tcp_handle_c2h_data(queue, (void *)queue->pdu);
615 case nvme_tcp_rsp:
616 nvme_tcp_init_recv_ctx(queue);
617 return nvme_tcp_handle_comp(queue, (void *)queue->pdu);
618 case nvme_tcp_r2t:
619 nvme_tcp_init_recv_ctx(queue);
620 return nvme_tcp_handle_r2t(queue, (void *)queue->pdu);
621 default:
622 dev_err(queue->ctrl->ctrl.device,
623 "unsupported pdu type (%d)\n", hdr->type);
624 return -EINVAL;
625 }
626 }
627
628 static inline void nvme_tcp_end_request(struct request *rq, u16 status)
629 {
630 union nvme_result res = {};
631
632 nvme_end_request(rq, cpu_to_le16(status << 1), res);
633 }
634
635 static int nvme_tcp_recv_data(struct nvme_tcp_queue *queue, struct sk_buff *skb,
636 unsigned int *offset, size_t *len)
637 {
638 struct nvme_tcp_data_pdu *pdu = (void *)queue->pdu;
639 struct nvme_tcp_request *req;
640 struct request *rq;
641
642 rq = blk_mq_tag_to_rq(nvme_tcp_tagset(queue), pdu->command_id);
643 if (!rq) {
644 dev_err(queue->ctrl->ctrl.device,
645 "queue %d tag %#x not found\n",
646 nvme_tcp_queue_id(queue), pdu->command_id);
647 return -ENOENT;
648 }
649 req = blk_mq_rq_to_pdu(rq);
650
651 while (true) {
652 int recv_len, ret;
653
654 recv_len = min_t(size_t, *len, queue->data_remaining);
655 if (!recv_len)
656 break;
657
658 if (!iov_iter_count(&req->iter)) {
659 req->curr_bio = req->curr_bio->bi_next;
660
661 /*
662 * If we don`t have any bios it means that controller
663 * sent more data than we requested, hence error
664 */
665 if (!req->curr_bio) {
666 dev_err(queue->ctrl->ctrl.device,
667 "queue %d no space in request %#x",
668 nvme_tcp_queue_id(queue), rq->tag);
669 nvme_tcp_init_recv_ctx(queue);
670 return -EIO;
671 }
672 nvme_tcp_init_iter(req, READ);
673 }
674
675 /* we can read only from what is left in this bio */
676 recv_len = min_t(size_t, recv_len,
677 iov_iter_count(&req->iter));
678
679 if (queue->data_digest)
680 ret = skb_copy_and_hash_datagram_iter(skb, *offset,
681 &req->iter, recv_len, queue->rcv_hash);
682 else
683 ret = skb_copy_datagram_iter(skb, *offset,
684 &req->iter, recv_len);
685 if (ret) {
686 dev_err(queue->ctrl->ctrl.device,
687 "queue %d failed to copy request %#x data",
688 nvme_tcp_queue_id(queue), rq->tag);
689 return ret;
690 }
691
692 *len -= recv_len;
693 *offset += recv_len;
694 queue->data_remaining -= recv_len;
695 }
696
697 if (!queue->data_remaining) {
698 if (queue->data_digest) {
699 nvme_tcp_ddgst_final(queue->rcv_hash, &queue->exp_ddgst);
700 queue->ddgst_remaining = NVME_TCP_DIGEST_LENGTH;
701 } else {
702 if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
703 nvme_tcp_end_request(rq, NVME_SC_SUCCESS);
704 queue->nr_cqe++;
705 }
706 nvme_tcp_init_recv_ctx(queue);
707 }
708 }
709
710 return 0;
711 }
712
713 static int nvme_tcp_recv_ddgst(struct nvme_tcp_queue *queue,
714 struct sk_buff *skb, unsigned int *offset, size_t *len)
715 {
716 struct nvme_tcp_data_pdu *pdu = (void *)queue->pdu;
717 char *ddgst = (char *)&queue->recv_ddgst;
718 size_t recv_len = min_t(size_t, *len, queue->ddgst_remaining);
719 off_t off = NVME_TCP_DIGEST_LENGTH - queue->ddgst_remaining;
720 int ret;
721
722 ret = skb_copy_bits(skb, *offset, &ddgst[off], recv_len);
723 if (unlikely(ret))
724 return ret;
725
726 queue->ddgst_remaining -= recv_len;
727 *offset += recv_len;
728 *len -= recv_len;
729 if (queue->ddgst_remaining)
730 return 0;
731
732 if (queue->recv_ddgst != queue->exp_ddgst) {
733 dev_err(queue->ctrl->ctrl.device,
734 "data digest error: recv %#x expected %#x\n",
735 le32_to_cpu(queue->recv_ddgst),
736 le32_to_cpu(queue->exp_ddgst));
737 return -EIO;
738 }
739
740 if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
741 struct request *rq = blk_mq_tag_to_rq(nvme_tcp_tagset(queue),
742 pdu->command_id);
743
744 nvme_tcp_end_request(rq, NVME_SC_SUCCESS);
745 queue->nr_cqe++;
746 }
747
748 nvme_tcp_init_recv_ctx(queue);
749 return 0;
750 }
751
752 static int nvme_tcp_recv_skb(read_descriptor_t *desc, struct sk_buff *skb,
753 unsigned int offset, size_t len)
754 {
755 struct nvme_tcp_queue *queue = desc->arg.data;
756 size_t consumed = len;
757 int result;
758
759 while (len) {
760 switch (nvme_tcp_recv_state(queue)) {
761 case NVME_TCP_RECV_PDU:
762 result = nvme_tcp_recv_pdu(queue, skb, &offset, &len);
763 break;
764 case NVME_TCP_RECV_DATA:
765 result = nvme_tcp_recv_data(queue, skb, &offset, &len);
766 break;
767 case NVME_TCP_RECV_DDGST:
768 result = nvme_tcp_recv_ddgst(queue, skb, &offset, &len);
769 break;
770 default:
771 result = -EFAULT;
772 }
773 if (result) {
774 dev_err(queue->ctrl->ctrl.device,
775 "receive failed: %d\n", result);
776 queue->rd_enabled = false;
777 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
778 return result;
779 }
780 }
781
782 return consumed;
783 }
784
785 static void nvme_tcp_data_ready(struct sock *sk)
786 {
787 struct nvme_tcp_queue *queue;
788
789 read_lock(&sk->sk_callback_lock);
790 queue = sk->sk_user_data;
791 if (likely(queue && queue->rd_enabled))
792 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
793 read_unlock(&sk->sk_callback_lock);
794 }
795
796 static void nvme_tcp_write_space(struct sock *sk)
797 {
798 struct nvme_tcp_queue *queue;
799
800 read_lock_bh(&sk->sk_callback_lock);
801 queue = sk->sk_user_data;
802 if (likely(queue && sk_stream_is_writeable(sk))) {
803 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
804 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
805 }
806 read_unlock_bh(&sk->sk_callback_lock);
807 }
808
809 static void nvme_tcp_state_change(struct sock *sk)
810 {
811 struct nvme_tcp_queue *queue;
812
813 read_lock(&sk->sk_callback_lock);
814 queue = sk->sk_user_data;
815 if (!queue)
816 goto done;
817
818 switch (sk->sk_state) {
819 case TCP_CLOSE:
820 case TCP_CLOSE_WAIT:
821 case TCP_LAST_ACK:
822 case TCP_FIN_WAIT1:
823 case TCP_FIN_WAIT2:
824 /* fallthrough */
825 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
826 break;
827 default:
828 dev_info(queue->ctrl->ctrl.device,
829 "queue %d socket state %d\n",
830 nvme_tcp_queue_id(queue), sk->sk_state);
831 }
832
833 queue->state_change(sk);
834 done:
835 read_unlock(&sk->sk_callback_lock);
836 }
837
838 static inline void nvme_tcp_done_send_req(struct nvme_tcp_queue *queue)
839 {
840 queue->request = NULL;
841 }
842
843 static void nvme_tcp_fail_request(struct nvme_tcp_request *req)
844 {
845 nvme_tcp_end_request(blk_mq_rq_from_pdu(req), NVME_SC_HOST_PATH_ERROR);
846 }
847
848 static int nvme_tcp_try_send_data(struct nvme_tcp_request *req)
849 {
850 struct nvme_tcp_queue *queue = req->queue;
851
852 while (true) {
853 struct page *page = nvme_tcp_req_cur_page(req);
854 size_t offset = nvme_tcp_req_cur_offset(req);
855 size_t len = nvme_tcp_req_cur_length(req);
856 bool last = nvme_tcp_pdu_last_send(req, len);
857 int ret, flags = MSG_DONTWAIT;
858
859 if (last && !queue->data_digest)
860 flags |= MSG_EOR;
861 else
862 flags |= MSG_MORE;
863
864 /* can't zcopy slab pages */
865 if (unlikely(PageSlab(page))) {
866 ret = sock_no_sendpage(queue->sock, page, offset, len,
867 flags);
868 } else {
869 ret = kernel_sendpage(queue->sock, page, offset, len,
870 flags);
871 }
872 if (ret <= 0)
873 return ret;
874
875 nvme_tcp_advance_req(req, ret);
876 if (queue->data_digest)
877 nvme_tcp_ddgst_update(queue->snd_hash, page,
878 offset, ret);
879
880 /* fully successful last write*/
881 if (last && ret == len) {
882 if (queue->data_digest) {
883 nvme_tcp_ddgst_final(queue->snd_hash,
884 &req->ddgst);
885 req->state = NVME_TCP_SEND_DDGST;
886 req->offset = 0;
887 } else {
888 nvme_tcp_done_send_req(queue);
889 }
890 return 1;
891 }
892 }
893 return -EAGAIN;
894 }
895
896 static int nvme_tcp_try_send_cmd_pdu(struct nvme_tcp_request *req)
897 {
898 struct nvme_tcp_queue *queue = req->queue;
899 struct nvme_tcp_cmd_pdu *pdu = req->pdu;
900 bool inline_data = nvme_tcp_has_inline_data(req);
901 int flags = MSG_DONTWAIT | (inline_data ? MSG_MORE : MSG_EOR);
902 u8 hdgst = nvme_tcp_hdgst_len(queue);
903 int len = sizeof(*pdu) + hdgst - req->offset;
904 int ret;
905
906 if (queue->hdr_digest && !req->offset)
907 nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
908
909 ret = kernel_sendpage(queue->sock, virt_to_page(pdu),
910 offset_in_page(pdu) + req->offset, len, flags);
911 if (unlikely(ret <= 0))
912 return ret;
913
914 len -= ret;
915 if (!len) {
916 if (inline_data) {
917 req->state = NVME_TCP_SEND_DATA;
918 if (queue->data_digest)
919 crypto_ahash_init(queue->snd_hash);
920 nvme_tcp_init_iter(req, WRITE);
921 } else {
922 nvme_tcp_done_send_req(queue);
923 }
924 return 1;
925 }
926 req->offset += ret;
927
928 return -EAGAIN;
929 }
930
931 static int nvme_tcp_try_send_data_pdu(struct nvme_tcp_request *req)
932 {
933 struct nvme_tcp_queue *queue = req->queue;
934 struct nvme_tcp_data_pdu *pdu = req->pdu;
935 u8 hdgst = nvme_tcp_hdgst_len(queue);
936 int len = sizeof(*pdu) - req->offset + hdgst;
937 int ret;
938
939 if (queue->hdr_digest && !req->offset)
940 nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
941
942 ret = kernel_sendpage(queue->sock, virt_to_page(pdu),
943 offset_in_page(pdu) + req->offset, len,
944 MSG_DONTWAIT | MSG_MORE);
945 if (unlikely(ret <= 0))
946 return ret;
947
948 len -= ret;
949 if (!len) {
950 req->state = NVME_TCP_SEND_DATA;
951 if (queue->data_digest)
952 crypto_ahash_init(queue->snd_hash);
953 if (!req->data_sent)
954 nvme_tcp_init_iter(req, WRITE);
955 return 1;
956 }
957 req->offset += ret;
958
959 return -EAGAIN;
960 }
961
962 static int nvme_tcp_try_send_ddgst(struct nvme_tcp_request *req)
963 {
964 struct nvme_tcp_queue *queue = req->queue;
965 int ret;
966 struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_EOR };
967 struct kvec iov = {
968 .iov_base = &req->ddgst + req->offset,
969 .iov_len = NVME_TCP_DIGEST_LENGTH - req->offset
970 };
971
972 ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
973 if (unlikely(ret <= 0))
974 return ret;
975
976 if (req->offset + ret == NVME_TCP_DIGEST_LENGTH) {
977 nvme_tcp_done_send_req(queue);
978 return 1;
979 }
980
981 req->offset += ret;
982 return -EAGAIN;
983 }
984
985 static int nvme_tcp_try_send(struct nvme_tcp_queue *queue)
986 {
987 struct nvme_tcp_request *req;
988 int ret = 1;
989
990 if (!queue->request) {
991 queue->request = nvme_tcp_fetch_request(queue);
992 if (!queue->request)
993 return 0;
994 }
995 req = queue->request;
996
997 if (req->state == NVME_TCP_SEND_CMD_PDU) {
998 ret = nvme_tcp_try_send_cmd_pdu(req);
999 if (ret <= 0)
1000 goto done;
1001 if (!nvme_tcp_has_inline_data(req))
1002 return ret;
1003 }
1004
1005 if (req->state == NVME_TCP_SEND_H2C_PDU) {
1006 ret = nvme_tcp_try_send_data_pdu(req);
1007 if (ret <= 0)
1008 goto done;
1009 }
1010
1011 if (req->state == NVME_TCP_SEND_DATA) {
1012 ret = nvme_tcp_try_send_data(req);
1013 if (ret <= 0)
1014 goto done;
1015 }
1016
1017 if (req->state == NVME_TCP_SEND_DDGST)
1018 ret = nvme_tcp_try_send_ddgst(req);
1019 done:
1020 if (ret == -EAGAIN)
1021 ret = 0;
1022 return ret;
1023 }
1024
1025 static int nvme_tcp_try_recv(struct nvme_tcp_queue *queue)
1026 {
1027 struct socket *sock = queue->sock;
1028 struct sock *sk = sock->sk;
1029 read_descriptor_t rd_desc;
1030 int consumed;
1031
1032 rd_desc.arg.data = queue;
1033 rd_desc.count = 1;
1034 lock_sock(sk);
1035 queue->nr_cqe = 0;
1036 consumed = sock->ops->read_sock(sk, &rd_desc, nvme_tcp_recv_skb);
1037 release_sock(sk);
1038 return consumed;
1039 }
1040
1041 static void nvme_tcp_io_work(struct work_struct *w)
1042 {
1043 struct nvme_tcp_queue *queue =
1044 container_of(w, struct nvme_tcp_queue, io_work);
1045 unsigned long deadline = jiffies + msecs_to_jiffies(1);
1046
1047 do {
1048 bool pending = false;
1049 int result;
1050
1051 result = nvme_tcp_try_send(queue);
1052 if (result > 0) {
1053 pending = true;
1054 } else if (unlikely(result < 0)) {
1055 dev_err(queue->ctrl->ctrl.device,
1056 "failed to send request %d\n", result);
1057 if (result != -EPIPE)
1058 nvme_tcp_fail_request(queue->request);
1059 nvme_tcp_done_send_req(queue);
1060 return;
1061 }
1062
1063 result = nvme_tcp_try_recv(queue);
1064 if (result > 0)
1065 pending = true;
1066
1067 if (!pending)
1068 return;
1069
1070 } while (!time_after(jiffies, deadline)); /* quota is exhausted */
1071
1072 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
1073 }
1074
1075 static void nvme_tcp_free_crypto(struct nvme_tcp_queue *queue)
1076 {
1077 struct crypto_ahash *tfm = crypto_ahash_reqtfm(queue->rcv_hash);
1078
1079 ahash_request_free(queue->rcv_hash);
1080 ahash_request_free(queue->snd_hash);
1081 crypto_free_ahash(tfm);
1082 }
1083
1084 static int nvme_tcp_alloc_crypto(struct nvme_tcp_queue *queue)
1085 {
1086 struct crypto_ahash *tfm;
1087
1088 tfm = crypto_alloc_ahash("crc32c", 0, CRYPTO_ALG_ASYNC);
1089 if (IS_ERR(tfm))
1090 return PTR_ERR(tfm);
1091
1092 queue->snd_hash = ahash_request_alloc(tfm, GFP_KERNEL);
1093 if (!queue->snd_hash)
1094 goto free_tfm;
1095 ahash_request_set_callback(queue->snd_hash, 0, NULL, NULL);
1096
1097 queue->rcv_hash = ahash_request_alloc(tfm, GFP_KERNEL);
1098 if (!queue->rcv_hash)
1099 goto free_snd_hash;
1100 ahash_request_set_callback(queue->rcv_hash, 0, NULL, NULL);
1101
1102 return 0;
1103 free_snd_hash:
1104 ahash_request_free(queue->snd_hash);
1105 free_tfm:
1106 crypto_free_ahash(tfm);
1107 return -ENOMEM;
1108 }
1109
1110 static void nvme_tcp_free_async_req(struct nvme_tcp_ctrl *ctrl)
1111 {
1112 struct nvme_tcp_request *async = &ctrl->async_req;
1113
1114 page_frag_free(async->pdu);
1115 }
1116
1117 static int nvme_tcp_alloc_async_req(struct nvme_tcp_ctrl *ctrl)
1118 {
1119 struct nvme_tcp_queue *queue = &ctrl->queues[0];
1120 struct nvme_tcp_request *async = &ctrl->async_req;
1121 u8 hdgst = nvme_tcp_hdgst_len(queue);
1122
1123 async->pdu = page_frag_alloc(&queue->pf_cache,
1124 sizeof(struct nvme_tcp_cmd_pdu) + hdgst,
1125 GFP_KERNEL | __GFP_ZERO);
1126 if (!async->pdu)
1127 return -ENOMEM;
1128
1129 async->queue = &ctrl->queues[0];
1130 return 0;
1131 }
1132
1133 static void nvme_tcp_free_queue(struct nvme_ctrl *nctrl, int qid)
1134 {
1135 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1136 struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1137
1138 if (!test_and_clear_bit(NVME_TCP_Q_ALLOCATED, &queue->flags))
1139 return;
1140
1141 if (queue->hdr_digest || queue->data_digest)
1142 nvme_tcp_free_crypto(queue);
1143
1144 sock_release(queue->sock);
1145 kfree(queue->pdu);
1146 }
1147
1148 static int nvme_tcp_init_connection(struct nvme_tcp_queue *queue)
1149 {
1150 struct nvme_tcp_icreq_pdu *icreq;
1151 struct nvme_tcp_icresp_pdu *icresp;
1152 struct msghdr msg = {};
1153 struct kvec iov;
1154 bool ctrl_hdgst, ctrl_ddgst;
1155 int ret;
1156
1157 icreq = kzalloc(sizeof(*icreq), GFP_KERNEL);
1158 if (!icreq)
1159 return -ENOMEM;
1160
1161 icresp = kzalloc(sizeof(*icresp), GFP_KERNEL);
1162 if (!icresp) {
1163 ret = -ENOMEM;
1164 goto free_icreq;
1165 }
1166
1167 icreq->hdr.type = nvme_tcp_icreq;
1168 icreq->hdr.hlen = sizeof(*icreq);
1169 icreq->hdr.pdo = 0;
1170 icreq->hdr.plen = cpu_to_le32(icreq->hdr.hlen);
1171 icreq->pfv = cpu_to_le16(NVME_TCP_PFV_1_0);
1172 icreq->maxr2t = 0; /* single inflight r2t supported */
1173 icreq->hpda = 0; /* no alignment constraint */
1174 if (queue->hdr_digest)
1175 icreq->digest |= NVME_TCP_HDR_DIGEST_ENABLE;
1176 if (queue->data_digest)
1177 icreq->digest |= NVME_TCP_DATA_DIGEST_ENABLE;
1178
1179 iov.iov_base = icreq;
1180 iov.iov_len = sizeof(*icreq);
1181 ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
1182 if (ret < 0)
1183 goto free_icresp;
1184
1185 memset(&msg, 0, sizeof(msg));
1186 iov.iov_base = icresp;
1187 iov.iov_len = sizeof(*icresp);
1188 ret = kernel_recvmsg(queue->sock, &msg, &iov, 1,
1189 iov.iov_len, msg.msg_flags);
1190 if (ret < 0)
1191 goto free_icresp;
1192
1193 ret = -EINVAL;
1194 if (icresp->hdr.type != nvme_tcp_icresp) {
1195 pr_err("queue %d: bad type returned %d\n",
1196 nvme_tcp_queue_id(queue), icresp->hdr.type);
1197 goto free_icresp;
1198 }
1199
1200 if (le32_to_cpu(icresp->hdr.plen) != sizeof(*icresp)) {
1201 pr_err("queue %d: bad pdu length returned %d\n",
1202 nvme_tcp_queue_id(queue), icresp->hdr.plen);
1203 goto free_icresp;
1204 }
1205
1206 if (icresp->pfv != NVME_TCP_PFV_1_0) {
1207 pr_err("queue %d: bad pfv returned %d\n",
1208 nvme_tcp_queue_id(queue), icresp->pfv);
1209 goto free_icresp;
1210 }
1211
1212 ctrl_ddgst = !!(icresp->digest & NVME_TCP_DATA_DIGEST_ENABLE);
1213 if ((queue->data_digest && !ctrl_ddgst) ||
1214 (!queue->data_digest && ctrl_ddgst)) {
1215 pr_err("queue %d: data digest mismatch host: %s ctrl: %s\n",
1216 nvme_tcp_queue_id(queue),
1217 queue->data_digest ? "enabled" : "disabled",
1218 ctrl_ddgst ? "enabled" : "disabled");
1219 goto free_icresp;
1220 }
1221
1222 ctrl_hdgst = !!(icresp->digest & NVME_TCP_HDR_DIGEST_ENABLE);
1223 if ((queue->hdr_digest && !ctrl_hdgst) ||
1224 (!queue->hdr_digest && ctrl_hdgst)) {
1225 pr_err("queue %d: header digest mismatch host: %s ctrl: %s\n",
1226 nvme_tcp_queue_id(queue),
1227 queue->hdr_digest ? "enabled" : "disabled",
1228 ctrl_hdgst ? "enabled" : "disabled");
1229 goto free_icresp;
1230 }
1231
1232 if (icresp->cpda != 0) {
1233 pr_err("queue %d: unsupported cpda returned %d\n",
1234 nvme_tcp_queue_id(queue), icresp->cpda);
1235 goto free_icresp;
1236 }
1237
1238 ret = 0;
1239 free_icresp:
1240 kfree(icresp);
1241 free_icreq:
1242 kfree(icreq);
1243 return ret;
1244 }
1245
1246 static int nvme_tcp_alloc_queue(struct nvme_ctrl *nctrl,
1247 int qid, size_t queue_size)
1248 {
1249 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1250 struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1251 struct linger sol = { .l_onoff = 1, .l_linger = 0 };
1252 int ret, opt, rcv_pdu_size, n;
1253
1254 queue->ctrl = ctrl;
1255 INIT_LIST_HEAD(&queue->send_list);
1256 spin_lock_init(&queue->lock);
1257 INIT_WORK(&queue->io_work, nvme_tcp_io_work);
1258 queue->queue_size = queue_size;
1259
1260 if (qid > 0)
1261 queue->cmnd_capsule_len = nctrl->ioccsz * 16;
1262 else
1263 queue->cmnd_capsule_len = sizeof(struct nvme_command) +
1264 NVME_TCP_ADMIN_CCSZ;
1265
1266 ret = sock_create(ctrl->addr.ss_family, SOCK_STREAM,
1267 IPPROTO_TCP, &queue->sock);
1268 if (ret) {
1269 dev_err(nctrl->device,
1270 "failed to create socket: %d\n", ret);
1271 return ret;
1272 }
1273
1274 /* Single syn retry */
1275 opt = 1;
1276 ret = kernel_setsockopt(queue->sock, IPPROTO_TCP, TCP_SYNCNT,
1277 (char *)&opt, sizeof(opt));
1278 if (ret) {
1279 dev_err(nctrl->device,
1280 "failed to set TCP_SYNCNT sock opt %d\n", ret);
1281 goto err_sock;
1282 }
1283
1284 /* Set TCP no delay */
1285 opt = 1;
1286 ret = kernel_setsockopt(queue->sock, IPPROTO_TCP,
1287 TCP_NODELAY, (char *)&opt, sizeof(opt));
1288 if (ret) {
1289 dev_err(nctrl->device,
1290 "failed to set TCP_NODELAY sock opt %d\n", ret);
1291 goto err_sock;
1292 }
1293
1294 /*
1295 * Cleanup whatever is sitting in the TCP transmit queue on socket
1296 * close. This is done to prevent stale data from being sent should
1297 * the network connection be restored before TCP times out.
1298 */
1299 ret = kernel_setsockopt(queue->sock, SOL_SOCKET, SO_LINGER,
1300 (char *)&sol, sizeof(sol));
1301 if (ret) {
1302 dev_err(nctrl->device,
1303 "failed to set SO_LINGER sock opt %d\n", ret);
1304 goto err_sock;
1305 }
1306
1307 /* Set socket type of service */
1308 if (nctrl->opts->tos >= 0) {
1309 opt = nctrl->opts->tos;
1310 ret = kernel_setsockopt(queue->sock, SOL_IP, IP_TOS,
1311 (char *)&opt, sizeof(opt));
1312 if (ret) {
1313 dev_err(nctrl->device,
1314 "failed to set IP_TOS sock opt %d\n", ret);
1315 goto err_sock;
1316 }
1317 }
1318
1319 queue->sock->sk->sk_allocation = GFP_ATOMIC;
1320 if (!qid)
1321 n = 0;
1322 else
1323 n = (qid - 1) % num_online_cpus();
1324 queue->io_cpu = cpumask_next_wrap(n - 1, cpu_online_mask, -1, false);
1325 queue->request = NULL;
1326 queue->data_remaining = 0;
1327 queue->ddgst_remaining = 0;
1328 queue->pdu_remaining = 0;
1329 queue->pdu_offset = 0;
1330 sk_set_memalloc(queue->sock->sk);
1331
1332 if (nctrl->opts->mask & NVMF_OPT_HOST_TRADDR) {
1333 ret = kernel_bind(queue->sock, (struct sockaddr *)&ctrl->src_addr,
1334 sizeof(ctrl->src_addr));
1335 if (ret) {
1336 dev_err(nctrl->device,
1337 "failed to bind queue %d socket %d\n",
1338 qid, ret);
1339 goto err_sock;
1340 }
1341 }
1342
1343 queue->hdr_digest = nctrl->opts->hdr_digest;
1344 queue->data_digest = nctrl->opts->data_digest;
1345 if (queue->hdr_digest || queue->data_digest) {
1346 ret = nvme_tcp_alloc_crypto(queue);
1347 if (ret) {
1348 dev_err(nctrl->device,
1349 "failed to allocate queue %d crypto\n", qid);
1350 goto err_sock;
1351 }
1352 }
1353
1354 rcv_pdu_size = sizeof(struct nvme_tcp_rsp_pdu) +
1355 nvme_tcp_hdgst_len(queue);
1356 queue->pdu = kmalloc(rcv_pdu_size, GFP_KERNEL);
1357 if (!queue->pdu) {
1358 ret = -ENOMEM;
1359 goto err_crypto;
1360 }
1361
1362 dev_dbg(nctrl->device, "connecting queue %d\n",
1363 nvme_tcp_queue_id(queue));
1364
1365 ret = kernel_connect(queue->sock, (struct sockaddr *)&ctrl->addr,
1366 sizeof(ctrl->addr), 0);
1367 if (ret) {
1368 dev_err(nctrl->device,
1369 "failed to connect socket: %d\n", ret);
1370 goto err_rcv_pdu;
1371 }
1372
1373 ret = nvme_tcp_init_connection(queue);
1374 if (ret)
1375 goto err_init_connect;
1376
1377 queue->rd_enabled = true;
1378 set_bit(NVME_TCP_Q_ALLOCATED, &queue->flags);
1379 nvme_tcp_init_recv_ctx(queue);
1380
1381 write_lock_bh(&queue->sock->sk->sk_callback_lock);
1382 queue->sock->sk->sk_user_data = queue;
1383 queue->state_change = queue->sock->sk->sk_state_change;
1384 queue->data_ready = queue->sock->sk->sk_data_ready;
1385 queue->write_space = queue->sock->sk->sk_write_space;
1386 queue->sock->sk->sk_data_ready = nvme_tcp_data_ready;
1387 queue->sock->sk->sk_state_change = nvme_tcp_state_change;
1388 queue->sock->sk->sk_write_space = nvme_tcp_write_space;
1389 #ifdef CONFIG_NET_RX_BUSY_POLL
1390 queue->sock->sk->sk_ll_usec = 1;
1391 #endif
1392 write_unlock_bh(&queue->sock->sk->sk_callback_lock);
1393
1394 return 0;
1395
1396 err_init_connect:
1397 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1398 err_rcv_pdu:
1399 kfree(queue->pdu);
1400 err_crypto:
1401 if (queue->hdr_digest || queue->data_digest)
1402 nvme_tcp_free_crypto(queue);
1403 err_sock:
1404 sock_release(queue->sock);
1405 queue->sock = NULL;
1406 return ret;
1407 }
1408
1409 static void nvme_tcp_restore_sock_calls(struct nvme_tcp_queue *queue)
1410 {
1411 struct socket *sock = queue->sock;
1412
1413 write_lock_bh(&sock->sk->sk_callback_lock);
1414 sock->sk->sk_user_data = NULL;
1415 sock->sk->sk_data_ready = queue->data_ready;
1416 sock->sk->sk_state_change = queue->state_change;
1417 sock->sk->sk_write_space = queue->write_space;
1418 write_unlock_bh(&sock->sk->sk_callback_lock);
1419 }
1420
1421 static void __nvme_tcp_stop_queue(struct nvme_tcp_queue *queue)
1422 {
1423 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1424 nvme_tcp_restore_sock_calls(queue);
1425 cancel_work_sync(&queue->io_work);
1426 }
1427
1428 static void nvme_tcp_stop_queue(struct nvme_ctrl *nctrl, int qid)
1429 {
1430 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1431 struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1432
1433 if (!test_and_clear_bit(NVME_TCP_Q_LIVE, &queue->flags))
1434 return;
1435
1436 __nvme_tcp_stop_queue(queue);
1437 }
1438
1439 static int nvme_tcp_start_queue(struct nvme_ctrl *nctrl, int idx)
1440 {
1441 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1442 int ret;
1443
1444 if (idx)
1445 ret = nvmf_connect_io_queue(nctrl, idx, false);
1446 else
1447 ret = nvmf_connect_admin_queue(nctrl);
1448
1449 if (!ret) {
1450 set_bit(NVME_TCP_Q_LIVE, &ctrl->queues[idx].flags);
1451 } else {
1452 if (test_bit(NVME_TCP_Q_ALLOCATED, &ctrl->queues[idx].flags))
1453 __nvme_tcp_stop_queue(&ctrl->queues[idx]);
1454 dev_err(nctrl->device,
1455 "failed to connect queue: %d ret=%d\n", idx, ret);
1456 }
1457 return ret;
1458 }
1459
1460 static struct blk_mq_tag_set *nvme_tcp_alloc_tagset(struct nvme_ctrl *nctrl,
1461 bool admin)
1462 {
1463 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1464 struct blk_mq_tag_set *set;
1465 int ret;
1466
1467 if (admin) {
1468 set = &ctrl->admin_tag_set;
1469 memset(set, 0, sizeof(*set));
1470 set->ops = &nvme_tcp_admin_mq_ops;
1471 set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
1472 set->reserved_tags = 2; /* connect + keep-alive */
1473 set->numa_node = NUMA_NO_NODE;
1474 set->cmd_size = sizeof(struct nvme_tcp_request);
1475 set->driver_data = ctrl;
1476 set->nr_hw_queues = 1;
1477 set->timeout = ADMIN_TIMEOUT;
1478 } else {
1479 set = &ctrl->tag_set;
1480 memset(set, 0, sizeof(*set));
1481 set->ops = &nvme_tcp_mq_ops;
1482 set->queue_depth = nctrl->sqsize + 1;
1483 set->reserved_tags = 1; /* fabric connect */
1484 set->numa_node = NUMA_NO_NODE;
1485 set->flags = BLK_MQ_F_SHOULD_MERGE;
1486 set->cmd_size = sizeof(struct nvme_tcp_request);
1487 set->driver_data = ctrl;
1488 set->nr_hw_queues = nctrl->queue_count - 1;
1489 set->timeout = NVME_IO_TIMEOUT;
1490 set->nr_maps = nctrl->opts->nr_poll_queues ? HCTX_MAX_TYPES : 2;
1491 }
1492
1493 ret = blk_mq_alloc_tag_set(set);
1494 if (ret)
1495 return ERR_PTR(ret);
1496
1497 return set;
1498 }
1499
1500 static void nvme_tcp_free_admin_queue(struct nvme_ctrl *ctrl)
1501 {
1502 if (to_tcp_ctrl(ctrl)->async_req.pdu) {
1503 nvme_tcp_free_async_req(to_tcp_ctrl(ctrl));
1504 to_tcp_ctrl(ctrl)->async_req.pdu = NULL;
1505 }
1506
1507 nvme_tcp_free_queue(ctrl, 0);
1508 }
1509
1510 static void nvme_tcp_free_io_queues(struct nvme_ctrl *ctrl)
1511 {
1512 int i;
1513
1514 for (i = 1; i < ctrl->queue_count; i++)
1515 nvme_tcp_free_queue(ctrl, i);
1516 }
1517
1518 static void nvme_tcp_stop_io_queues(struct nvme_ctrl *ctrl)
1519 {
1520 int i;
1521
1522 for (i = 1; i < ctrl->queue_count; i++)
1523 nvme_tcp_stop_queue(ctrl, i);
1524 }
1525
1526 static int nvme_tcp_start_io_queues(struct nvme_ctrl *ctrl)
1527 {
1528 int i, ret = 0;
1529
1530 for (i = 1; i < ctrl->queue_count; i++) {
1531 ret = nvme_tcp_start_queue(ctrl, i);
1532 if (ret)
1533 goto out_stop_queues;
1534 }
1535
1536 return 0;
1537
1538 out_stop_queues:
1539 for (i--; i >= 1; i--)
1540 nvme_tcp_stop_queue(ctrl, i);
1541 return ret;
1542 }
1543
1544 static int nvme_tcp_alloc_admin_queue(struct nvme_ctrl *ctrl)
1545 {
1546 int ret;
1547
1548 ret = nvme_tcp_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
1549 if (ret)
1550 return ret;
1551
1552 ret = nvme_tcp_alloc_async_req(to_tcp_ctrl(ctrl));
1553 if (ret)
1554 goto out_free_queue;
1555
1556 return 0;
1557
1558 out_free_queue:
1559 nvme_tcp_free_queue(ctrl, 0);
1560 return ret;
1561 }
1562
1563 static int __nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
1564 {
1565 int i, ret;
1566
1567 for (i = 1; i < ctrl->queue_count; i++) {
1568 ret = nvme_tcp_alloc_queue(ctrl, i,
1569 ctrl->sqsize + 1);
1570 if (ret)
1571 goto out_free_queues;
1572 }
1573
1574 return 0;
1575
1576 out_free_queues:
1577 for (i--; i >= 1; i--)
1578 nvme_tcp_free_queue(ctrl, i);
1579
1580 return ret;
1581 }
1582
1583 static unsigned int nvme_tcp_nr_io_queues(struct nvme_ctrl *ctrl)
1584 {
1585 unsigned int nr_io_queues;
1586
1587 nr_io_queues = min(ctrl->opts->nr_io_queues, num_online_cpus());
1588 nr_io_queues += min(ctrl->opts->nr_write_queues, num_online_cpus());
1589 nr_io_queues += min(ctrl->opts->nr_poll_queues, num_online_cpus());
1590
1591 return nr_io_queues;
1592 }
1593
1594 static void nvme_tcp_set_io_queues(struct nvme_ctrl *nctrl,
1595 unsigned int nr_io_queues)
1596 {
1597 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1598 struct nvmf_ctrl_options *opts = nctrl->opts;
1599
1600 if (opts->nr_write_queues && opts->nr_io_queues < nr_io_queues) {
1601 /*
1602 * separate read/write queues
1603 * hand out dedicated default queues only after we have
1604 * sufficient read queues.
1605 */
1606 ctrl->io_queues[HCTX_TYPE_READ] = opts->nr_io_queues;
1607 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_READ];
1608 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
1609 min(opts->nr_write_queues, nr_io_queues);
1610 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
1611 } else {
1612 /*
1613 * shared read/write queues
1614 * either no write queues were requested, or we don't have
1615 * sufficient queue count to have dedicated default queues.
1616 */
1617 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
1618 min(opts->nr_io_queues, nr_io_queues);
1619 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
1620 }
1621
1622 if (opts->nr_poll_queues && nr_io_queues) {
1623 /* map dedicated poll queues only if we have queues left */
1624 ctrl->io_queues[HCTX_TYPE_POLL] =
1625 min(opts->nr_poll_queues, nr_io_queues);
1626 }
1627 }
1628
1629 static int nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
1630 {
1631 unsigned int nr_io_queues;
1632 int ret;
1633
1634 nr_io_queues = nvme_tcp_nr_io_queues(ctrl);
1635 ret = nvme_set_queue_count(ctrl, &nr_io_queues);
1636 if (ret)
1637 return ret;
1638
1639 ctrl->queue_count = nr_io_queues + 1;
1640 if (ctrl->queue_count < 2)
1641 return 0;
1642
1643 dev_info(ctrl->device,
1644 "creating %d I/O queues.\n", nr_io_queues);
1645
1646 nvme_tcp_set_io_queues(ctrl, nr_io_queues);
1647
1648 return __nvme_tcp_alloc_io_queues(ctrl);
1649 }
1650
1651 static void nvme_tcp_destroy_io_queues(struct nvme_ctrl *ctrl, bool remove)
1652 {
1653 nvme_tcp_stop_io_queues(ctrl);
1654 if (remove) {
1655 blk_cleanup_queue(ctrl->connect_q);
1656 blk_mq_free_tag_set(ctrl->tagset);
1657 }
1658 nvme_tcp_free_io_queues(ctrl);
1659 }
1660
1661 static int nvme_tcp_configure_io_queues(struct nvme_ctrl *ctrl, bool new)
1662 {
1663 int ret;
1664
1665 ret = nvme_tcp_alloc_io_queues(ctrl);
1666 if (ret)
1667 return ret;
1668
1669 if (new) {
1670 ctrl->tagset = nvme_tcp_alloc_tagset(ctrl, false);
1671 if (IS_ERR(ctrl->tagset)) {
1672 ret = PTR_ERR(ctrl->tagset);
1673 goto out_free_io_queues;
1674 }
1675
1676 ctrl->connect_q = blk_mq_init_queue(ctrl->tagset);
1677 if (IS_ERR(ctrl->connect_q)) {
1678 ret = PTR_ERR(ctrl->connect_q);
1679 goto out_free_tag_set;
1680 }
1681 } else {
1682 blk_mq_update_nr_hw_queues(ctrl->tagset,
1683 ctrl->queue_count - 1);
1684 }
1685
1686 ret = nvme_tcp_start_io_queues(ctrl);
1687 if (ret)
1688 goto out_cleanup_connect_q;
1689
1690 return 0;
1691
1692 out_cleanup_connect_q:
1693 if (new)
1694 blk_cleanup_queue(ctrl->connect_q);
1695 out_free_tag_set:
1696 if (new)
1697 blk_mq_free_tag_set(ctrl->tagset);
1698 out_free_io_queues:
1699 nvme_tcp_free_io_queues(ctrl);
1700 return ret;
1701 }
1702
1703 static void nvme_tcp_destroy_admin_queue(struct nvme_ctrl *ctrl, bool remove)
1704 {
1705 nvme_tcp_stop_queue(ctrl, 0);
1706 if (remove) {
1707 blk_cleanup_queue(ctrl->admin_q);
1708 blk_cleanup_queue(ctrl->fabrics_q);
1709 blk_mq_free_tag_set(ctrl->admin_tagset);
1710 }
1711 nvme_tcp_free_admin_queue(ctrl);
1712 }
1713
1714 static int nvme_tcp_configure_admin_queue(struct nvme_ctrl *ctrl, bool new)
1715 {
1716 int error;
1717
1718 error = nvme_tcp_alloc_admin_queue(ctrl);
1719 if (error)
1720 return error;
1721
1722 if (new) {
1723 ctrl->admin_tagset = nvme_tcp_alloc_tagset(ctrl, true);
1724 if (IS_ERR(ctrl->admin_tagset)) {
1725 error = PTR_ERR(ctrl->admin_tagset);
1726 goto out_free_queue;
1727 }
1728
1729 ctrl->fabrics_q = blk_mq_init_queue(ctrl->admin_tagset);
1730 if (IS_ERR(ctrl->fabrics_q)) {
1731 error = PTR_ERR(ctrl->fabrics_q);
1732 goto out_free_tagset;
1733 }
1734
1735 ctrl->admin_q = blk_mq_init_queue(ctrl->admin_tagset);
1736 if (IS_ERR(ctrl->admin_q)) {
1737 error = PTR_ERR(ctrl->admin_q);
1738 goto out_cleanup_fabrics_q;
1739 }
1740 }
1741
1742 error = nvme_tcp_start_queue(ctrl, 0);
1743 if (error)
1744 goto out_cleanup_queue;
1745
1746 error = nvme_enable_ctrl(ctrl);
1747 if (error)
1748 goto out_stop_queue;
1749
1750 blk_mq_unquiesce_queue(ctrl->admin_q);
1751
1752 error = nvme_init_identify(ctrl);
1753 if (error)
1754 goto out_stop_queue;
1755
1756 return 0;
1757
1758 out_stop_queue:
1759 nvme_tcp_stop_queue(ctrl, 0);
1760 out_cleanup_queue:
1761 if (new)
1762 blk_cleanup_queue(ctrl->admin_q);
1763 out_cleanup_fabrics_q:
1764 if (new)
1765 blk_cleanup_queue(ctrl->fabrics_q);
1766 out_free_tagset:
1767 if (new)
1768 blk_mq_free_tag_set(ctrl->admin_tagset);
1769 out_free_queue:
1770 nvme_tcp_free_admin_queue(ctrl);
1771 return error;
1772 }
1773
1774 static void nvme_tcp_teardown_admin_queue(struct nvme_ctrl *ctrl,
1775 bool remove)
1776 {
1777 blk_mq_quiesce_queue(ctrl->admin_q);
1778 nvme_tcp_stop_queue(ctrl, 0);
1779 if (ctrl->admin_tagset) {
1780 blk_mq_tagset_busy_iter(ctrl->admin_tagset,
1781 nvme_cancel_request, ctrl);
1782 blk_mq_tagset_wait_completed_request(ctrl->admin_tagset);
1783 }
1784 if (remove)
1785 blk_mq_unquiesce_queue(ctrl->admin_q);
1786 nvme_tcp_destroy_admin_queue(ctrl, remove);
1787 }
1788
1789 static void nvme_tcp_teardown_io_queues(struct nvme_ctrl *ctrl,
1790 bool remove)
1791 {
1792 if (ctrl->queue_count <= 1)
1793 return;
1794 nvme_stop_queues(ctrl);
1795 nvme_tcp_stop_io_queues(ctrl);
1796 if (ctrl->tagset) {
1797 blk_mq_tagset_busy_iter(ctrl->tagset,
1798 nvme_cancel_request, ctrl);
1799 blk_mq_tagset_wait_completed_request(ctrl->tagset);
1800 }
1801 if (remove)
1802 nvme_start_queues(ctrl);
1803 nvme_tcp_destroy_io_queues(ctrl, remove);
1804 }
1805
1806 static void nvme_tcp_reconnect_or_remove(struct nvme_ctrl *ctrl)
1807 {
1808 /* If we are resetting/deleting then do nothing */
1809 if (ctrl->state != NVME_CTRL_CONNECTING) {
1810 WARN_ON_ONCE(ctrl->state == NVME_CTRL_NEW ||
1811 ctrl->state == NVME_CTRL_LIVE);
1812 return;
1813 }
1814
1815 if (nvmf_should_reconnect(ctrl)) {
1816 dev_info(ctrl->device, "Reconnecting in %d seconds...\n",
1817 ctrl->opts->reconnect_delay);
1818 queue_delayed_work(nvme_wq, &to_tcp_ctrl(ctrl)->connect_work,
1819 ctrl->opts->reconnect_delay * HZ);
1820 } else {
1821 dev_info(ctrl->device, "Removing controller...\n");
1822 nvme_delete_ctrl(ctrl);
1823 }
1824 }
1825
1826 static int nvme_tcp_setup_ctrl(struct nvme_ctrl *ctrl, bool new)
1827 {
1828 struct nvmf_ctrl_options *opts = ctrl->opts;
1829 int ret;
1830
1831 ret = nvme_tcp_configure_admin_queue(ctrl, new);
1832 if (ret)
1833 return ret;
1834
1835 if (ctrl->icdoff) {
1836 dev_err(ctrl->device, "icdoff is not supported!\n");
1837 goto destroy_admin;
1838 }
1839
1840 if (opts->queue_size > ctrl->sqsize + 1)
1841 dev_warn(ctrl->device,
1842 "queue_size %zu > ctrl sqsize %u, clamping down\n",
1843 opts->queue_size, ctrl->sqsize + 1);
1844
1845 if (ctrl->sqsize + 1 > ctrl->maxcmd) {
1846 dev_warn(ctrl->device,
1847 "sqsize %u > ctrl maxcmd %u, clamping down\n",
1848 ctrl->sqsize + 1, ctrl->maxcmd);
1849 ctrl->sqsize = ctrl->maxcmd - 1;
1850 }
1851
1852 if (ctrl->queue_count > 1) {
1853 ret = nvme_tcp_configure_io_queues(ctrl, new);
1854 if (ret)
1855 goto destroy_admin;
1856 }
1857
1858 if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_LIVE)) {
1859 /* state change failure is ok if we're in DELETING state */
1860 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING);
1861 ret = -EINVAL;
1862 goto destroy_io;
1863 }
1864
1865 nvme_start_ctrl(ctrl);
1866 return 0;
1867
1868 destroy_io:
1869 if (ctrl->queue_count > 1)
1870 nvme_tcp_destroy_io_queues(ctrl, new);
1871 destroy_admin:
1872 nvme_tcp_stop_queue(ctrl, 0);
1873 nvme_tcp_destroy_admin_queue(ctrl, new);
1874 return ret;
1875 }
1876
1877 static void nvme_tcp_reconnect_ctrl_work(struct work_struct *work)
1878 {
1879 struct nvme_tcp_ctrl *tcp_ctrl = container_of(to_delayed_work(work),
1880 struct nvme_tcp_ctrl, connect_work);
1881 struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
1882
1883 ++ctrl->nr_reconnects;
1884
1885 if (nvme_tcp_setup_ctrl(ctrl, false))
1886 goto requeue;
1887
1888 dev_info(ctrl->device, "Successfully reconnected (%d attempt)\n",
1889 ctrl->nr_reconnects);
1890
1891 ctrl->nr_reconnects = 0;
1892
1893 return;
1894
1895 requeue:
1896 dev_info(ctrl->device, "Failed reconnect attempt %d\n",
1897 ctrl->nr_reconnects);
1898 nvme_tcp_reconnect_or_remove(ctrl);
1899 }
1900
1901 static void nvme_tcp_error_recovery_work(struct work_struct *work)
1902 {
1903 struct nvme_tcp_ctrl *tcp_ctrl = container_of(work,
1904 struct nvme_tcp_ctrl, err_work);
1905 struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
1906
1907 nvme_stop_keep_alive(ctrl);
1908 nvme_tcp_teardown_io_queues(ctrl, false);
1909 /* unquiesce to fail fast pending requests */
1910 nvme_start_queues(ctrl);
1911 nvme_tcp_teardown_admin_queue(ctrl, false);
1912 blk_mq_unquiesce_queue(ctrl->admin_q);
1913
1914 if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
1915 /* state change failure is ok if we're in DELETING state */
1916 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING);
1917 return;
1918 }
1919
1920 nvme_tcp_reconnect_or_remove(ctrl);
1921 }
1922
1923 static void nvme_tcp_teardown_ctrl(struct nvme_ctrl *ctrl, bool shutdown)
1924 {
1925 cancel_work_sync(&to_tcp_ctrl(ctrl)->err_work);
1926 cancel_delayed_work_sync(&to_tcp_ctrl(ctrl)->connect_work);
1927
1928 nvme_tcp_teardown_io_queues(ctrl, shutdown);
1929 blk_mq_quiesce_queue(ctrl->admin_q);
1930 if (shutdown)
1931 nvme_shutdown_ctrl(ctrl);
1932 else
1933 nvme_disable_ctrl(ctrl);
1934 nvme_tcp_teardown_admin_queue(ctrl, shutdown);
1935 }
1936
1937 static void nvme_tcp_delete_ctrl(struct nvme_ctrl *ctrl)
1938 {
1939 nvme_tcp_teardown_ctrl(ctrl, true);
1940 }
1941
1942 static void nvme_reset_ctrl_work(struct work_struct *work)
1943 {
1944 struct nvme_ctrl *ctrl =
1945 container_of(work, struct nvme_ctrl, reset_work);
1946
1947 nvme_stop_ctrl(ctrl);
1948 nvme_tcp_teardown_ctrl(ctrl, false);
1949
1950 if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
1951 /* state change failure is ok if we're in DELETING state */
1952 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING);
1953 return;
1954 }
1955
1956 if (nvme_tcp_setup_ctrl(ctrl, false))
1957 goto out_fail;
1958
1959 return;
1960
1961 out_fail:
1962 ++ctrl->nr_reconnects;
1963 nvme_tcp_reconnect_or_remove(ctrl);
1964 }
1965
1966 static void nvme_tcp_free_ctrl(struct nvme_ctrl *nctrl)
1967 {
1968 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1969
1970 if (list_empty(&ctrl->list))
1971 goto free_ctrl;
1972
1973 mutex_lock(&nvme_tcp_ctrl_mutex);
1974 list_del(&ctrl->list);
1975 mutex_unlock(&nvme_tcp_ctrl_mutex);
1976
1977 nvmf_free_options(nctrl->opts);
1978 free_ctrl:
1979 kfree(ctrl->queues);
1980 kfree(ctrl);
1981 }
1982
1983 static void nvme_tcp_set_sg_null(struct nvme_command *c)
1984 {
1985 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1986
1987 sg->addr = 0;
1988 sg->length = 0;
1989 sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
1990 NVME_SGL_FMT_TRANSPORT_A;
1991 }
1992
1993 static void nvme_tcp_set_sg_inline(struct nvme_tcp_queue *queue,
1994 struct nvme_command *c, u32 data_len)
1995 {
1996 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1997
1998 sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
1999 sg->length = cpu_to_le32(data_len);
2000 sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
2001 }
2002
2003 static void nvme_tcp_set_sg_host_data(struct nvme_command *c,
2004 u32 data_len)
2005 {
2006 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2007
2008 sg->addr = 0;
2009 sg->length = cpu_to_le32(data_len);
2010 sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
2011 NVME_SGL_FMT_TRANSPORT_A;
2012 }
2013
2014 static void nvme_tcp_submit_async_event(struct nvme_ctrl *arg)
2015 {
2016 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(arg);
2017 struct nvme_tcp_queue *queue = &ctrl->queues[0];
2018 struct nvme_tcp_cmd_pdu *pdu = ctrl->async_req.pdu;
2019 struct nvme_command *cmd = &pdu->cmd;
2020 u8 hdgst = nvme_tcp_hdgst_len(queue);
2021
2022 memset(pdu, 0, sizeof(*pdu));
2023 pdu->hdr.type = nvme_tcp_cmd;
2024 if (queue->hdr_digest)
2025 pdu->hdr.flags |= NVME_TCP_F_HDGST;
2026 pdu->hdr.hlen = sizeof(*pdu);
2027 pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst);
2028
2029 cmd->common.opcode = nvme_admin_async_event;
2030 cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
2031 cmd->common.flags |= NVME_CMD_SGL_METABUF;
2032 nvme_tcp_set_sg_null(cmd);
2033
2034 ctrl->async_req.state = NVME_TCP_SEND_CMD_PDU;
2035 ctrl->async_req.offset = 0;
2036 ctrl->async_req.curr_bio = NULL;
2037 ctrl->async_req.data_len = 0;
2038
2039 nvme_tcp_queue_request(&ctrl->async_req);
2040 }
2041
2042 static enum blk_eh_timer_return
2043 nvme_tcp_timeout(struct request *rq, bool reserved)
2044 {
2045 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2046 struct nvme_tcp_ctrl *ctrl = req->queue->ctrl;
2047 struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2048
2049 /*
2050 * Restart the timer if a controller reset is already scheduled. Any
2051 * timed out commands would be handled before entering the connecting
2052 * state.
2053 */
2054 if (ctrl->ctrl.state == NVME_CTRL_RESETTING)
2055 return BLK_EH_RESET_TIMER;
2056
2057 dev_warn(ctrl->ctrl.device,
2058 "queue %d: timeout request %#x type %d\n",
2059 nvme_tcp_queue_id(req->queue), rq->tag, pdu->hdr.type);
2060
2061 if (ctrl->ctrl.state != NVME_CTRL_LIVE) {
2062 /*
2063 * Teardown immediately if controller times out while starting
2064 * or we are already started error recovery. all outstanding
2065 * requests are completed on shutdown, so we return BLK_EH_DONE.
2066 */
2067 flush_work(&ctrl->err_work);
2068 nvme_tcp_teardown_io_queues(&ctrl->ctrl, false);
2069 nvme_tcp_teardown_admin_queue(&ctrl->ctrl, false);
2070 return BLK_EH_DONE;
2071 }
2072
2073 dev_warn(ctrl->ctrl.device, "starting error recovery\n");
2074 nvme_tcp_error_recovery(&ctrl->ctrl);
2075
2076 return BLK_EH_RESET_TIMER;
2077 }
2078
2079 static blk_status_t nvme_tcp_map_data(struct nvme_tcp_queue *queue,
2080 struct request *rq)
2081 {
2082 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2083 struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2084 struct nvme_command *c = &pdu->cmd;
2085
2086 c->common.flags |= NVME_CMD_SGL_METABUF;
2087
2088 if (rq_data_dir(rq) == WRITE && req->data_len &&
2089 req->data_len <= nvme_tcp_inline_data_size(queue))
2090 nvme_tcp_set_sg_inline(queue, c, req->data_len);
2091 else
2092 nvme_tcp_set_sg_host_data(c, req->data_len);
2093
2094 return 0;
2095 }
2096
2097 static blk_status_t nvme_tcp_setup_cmd_pdu(struct nvme_ns *ns,
2098 struct request *rq)
2099 {
2100 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2101 struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2102 struct nvme_tcp_queue *queue = req->queue;
2103 u8 hdgst = nvme_tcp_hdgst_len(queue), ddgst = 0;
2104 blk_status_t ret;
2105
2106 ret = nvme_setup_cmd(ns, rq, &pdu->cmd);
2107 if (ret)
2108 return ret;
2109
2110 req->state = NVME_TCP_SEND_CMD_PDU;
2111 req->offset = 0;
2112 req->data_sent = 0;
2113 req->pdu_len = 0;
2114 req->pdu_sent = 0;
2115 req->data_len = blk_rq_payload_bytes(rq);
2116 req->curr_bio = rq->bio;
2117
2118 if (rq_data_dir(rq) == WRITE &&
2119 req->data_len <= nvme_tcp_inline_data_size(queue))
2120 req->pdu_len = req->data_len;
2121 else if (req->curr_bio)
2122 nvme_tcp_init_iter(req, READ);
2123
2124 pdu->hdr.type = nvme_tcp_cmd;
2125 pdu->hdr.flags = 0;
2126 if (queue->hdr_digest)
2127 pdu->hdr.flags |= NVME_TCP_F_HDGST;
2128 if (queue->data_digest && req->pdu_len) {
2129 pdu->hdr.flags |= NVME_TCP_F_DDGST;
2130 ddgst = nvme_tcp_ddgst_len(queue);
2131 }
2132 pdu->hdr.hlen = sizeof(*pdu);
2133 pdu->hdr.pdo = req->pdu_len ? pdu->hdr.hlen + hdgst : 0;
2134 pdu->hdr.plen =
2135 cpu_to_le32(pdu->hdr.hlen + hdgst + req->pdu_len + ddgst);
2136
2137 ret = nvme_tcp_map_data(queue, rq);
2138 if (unlikely(ret)) {
2139 nvme_cleanup_cmd(rq);
2140 dev_err(queue->ctrl->ctrl.device,
2141 "Failed to map data (%d)\n", ret);
2142 return ret;
2143 }
2144
2145 return 0;
2146 }
2147
2148 static blk_status_t nvme_tcp_queue_rq(struct blk_mq_hw_ctx *hctx,
2149 const struct blk_mq_queue_data *bd)
2150 {
2151 struct nvme_ns *ns = hctx->queue->queuedata;
2152 struct nvme_tcp_queue *queue = hctx->driver_data;
2153 struct request *rq = bd->rq;
2154 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2155 bool queue_ready = test_bit(NVME_TCP_Q_LIVE, &queue->flags);
2156 blk_status_t ret;
2157
2158 if (!nvmf_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
2159 return nvmf_fail_nonready_command(&queue->ctrl->ctrl, rq);
2160
2161 ret = nvme_tcp_setup_cmd_pdu(ns, rq);
2162 if (unlikely(ret))
2163 return ret;
2164
2165 blk_mq_start_request(rq);
2166
2167 nvme_tcp_queue_request(req);
2168
2169 return BLK_STS_OK;
2170 }
2171
2172 static int nvme_tcp_map_queues(struct blk_mq_tag_set *set)
2173 {
2174 struct nvme_tcp_ctrl *ctrl = set->driver_data;
2175 struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
2176
2177 if (opts->nr_write_queues && ctrl->io_queues[HCTX_TYPE_READ]) {
2178 /* separate read/write queues */
2179 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2180 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2181 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2182 set->map[HCTX_TYPE_READ].nr_queues =
2183 ctrl->io_queues[HCTX_TYPE_READ];
2184 set->map[HCTX_TYPE_READ].queue_offset =
2185 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2186 } else {
2187 /* shared read/write queues */
2188 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2189 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2190 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2191 set->map[HCTX_TYPE_READ].nr_queues =
2192 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2193 set->map[HCTX_TYPE_READ].queue_offset = 0;
2194 }
2195 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
2196 blk_mq_map_queues(&set->map[HCTX_TYPE_READ]);
2197
2198 if (opts->nr_poll_queues && ctrl->io_queues[HCTX_TYPE_POLL]) {
2199 /* map dedicated poll queues only if we have queues left */
2200 set->map[HCTX_TYPE_POLL].nr_queues =
2201 ctrl->io_queues[HCTX_TYPE_POLL];
2202 set->map[HCTX_TYPE_POLL].queue_offset =
2203 ctrl->io_queues[HCTX_TYPE_DEFAULT] +
2204 ctrl->io_queues[HCTX_TYPE_READ];
2205 blk_mq_map_queues(&set->map[HCTX_TYPE_POLL]);
2206 }
2207
2208 dev_info(ctrl->ctrl.device,
2209 "mapped %d/%d/%d default/read/poll queues.\n",
2210 ctrl->io_queues[HCTX_TYPE_DEFAULT],
2211 ctrl->io_queues[HCTX_TYPE_READ],
2212 ctrl->io_queues[HCTX_TYPE_POLL]);
2213
2214 return 0;
2215 }
2216
2217 static int nvme_tcp_poll(struct blk_mq_hw_ctx *hctx)
2218 {
2219 struct nvme_tcp_queue *queue = hctx->driver_data;
2220 struct sock *sk = queue->sock->sk;
2221
2222 if (sk_can_busy_loop(sk) && skb_queue_empty_lockless(&sk->sk_receive_queue))
2223 sk_busy_loop(sk, true);
2224 nvme_tcp_try_recv(queue);
2225 return queue->nr_cqe;
2226 }
2227
2228 static struct blk_mq_ops nvme_tcp_mq_ops = {
2229 .queue_rq = nvme_tcp_queue_rq,
2230 .complete = nvme_complete_rq,
2231 .init_request = nvme_tcp_init_request,
2232 .exit_request = nvme_tcp_exit_request,
2233 .init_hctx = nvme_tcp_init_hctx,
2234 .timeout = nvme_tcp_timeout,
2235 .map_queues = nvme_tcp_map_queues,
2236 .poll = nvme_tcp_poll,
2237 };
2238
2239 static struct blk_mq_ops nvme_tcp_admin_mq_ops = {
2240 .queue_rq = nvme_tcp_queue_rq,
2241 .complete = nvme_complete_rq,
2242 .init_request = nvme_tcp_init_request,
2243 .exit_request = nvme_tcp_exit_request,
2244 .init_hctx = nvme_tcp_init_admin_hctx,
2245 .timeout = nvme_tcp_timeout,
2246 };
2247
2248 static const struct nvme_ctrl_ops nvme_tcp_ctrl_ops = {
2249 .name = "tcp",
2250 .module = THIS_MODULE,
2251 .flags = NVME_F_FABRICS,
2252 .reg_read32 = nvmf_reg_read32,
2253 .reg_read64 = nvmf_reg_read64,
2254 .reg_write32 = nvmf_reg_write32,
2255 .free_ctrl = nvme_tcp_free_ctrl,
2256 .submit_async_event = nvme_tcp_submit_async_event,
2257 .delete_ctrl = nvme_tcp_delete_ctrl,
2258 .get_address = nvmf_get_address,
2259 };
2260
2261 static bool
2262 nvme_tcp_existing_controller(struct nvmf_ctrl_options *opts)
2263 {
2264 struct nvme_tcp_ctrl *ctrl;
2265 bool found = false;
2266
2267 mutex_lock(&nvme_tcp_ctrl_mutex);
2268 list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list) {
2269 found = nvmf_ip_options_match(&ctrl->ctrl, opts);
2270 if (found)
2271 break;
2272 }
2273 mutex_unlock(&nvme_tcp_ctrl_mutex);
2274
2275 return found;
2276 }
2277
2278 static struct nvme_ctrl *nvme_tcp_create_ctrl(struct device *dev,
2279 struct nvmf_ctrl_options *opts)
2280 {
2281 struct nvme_tcp_ctrl *ctrl;
2282 int ret;
2283
2284 ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
2285 if (!ctrl)
2286 return ERR_PTR(-ENOMEM);
2287
2288 INIT_LIST_HEAD(&ctrl->list);
2289 ctrl->ctrl.opts = opts;
2290 ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues +
2291 opts->nr_poll_queues + 1;
2292 ctrl->ctrl.sqsize = opts->queue_size - 1;
2293 ctrl->ctrl.kato = opts->kato;
2294
2295 INIT_DELAYED_WORK(&ctrl->connect_work,
2296 nvme_tcp_reconnect_ctrl_work);
2297 INIT_WORK(&ctrl->err_work, nvme_tcp_error_recovery_work);
2298 INIT_WORK(&ctrl->ctrl.reset_work, nvme_reset_ctrl_work);
2299
2300 if (!(opts->mask & NVMF_OPT_TRSVCID)) {
2301 opts->trsvcid =
2302 kstrdup(__stringify(NVME_TCP_DISC_PORT), GFP_KERNEL);
2303 if (!opts->trsvcid) {
2304 ret = -ENOMEM;
2305 goto out_free_ctrl;
2306 }
2307 opts->mask |= NVMF_OPT_TRSVCID;
2308 }
2309
2310 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2311 opts->traddr, opts->trsvcid, &ctrl->addr);
2312 if (ret) {
2313 pr_err("malformed address passed: %s:%s\n",
2314 opts->traddr, opts->trsvcid);
2315 goto out_free_ctrl;
2316 }
2317
2318 if (opts->mask & NVMF_OPT_HOST_TRADDR) {
2319 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2320 opts->host_traddr, NULL, &ctrl->src_addr);
2321 if (ret) {
2322 pr_err("malformed src address passed: %s\n",
2323 opts->host_traddr);
2324 goto out_free_ctrl;
2325 }
2326 }
2327
2328 if (!opts->duplicate_connect && nvme_tcp_existing_controller(opts)) {
2329 ret = -EALREADY;
2330 goto out_free_ctrl;
2331 }
2332
2333 ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
2334 GFP_KERNEL);
2335 if (!ctrl->queues) {
2336 ret = -ENOMEM;
2337 goto out_free_ctrl;
2338 }
2339
2340 ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_tcp_ctrl_ops, 0);
2341 if (ret)
2342 goto out_kfree_queues;
2343
2344 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
2345 WARN_ON_ONCE(1);
2346 ret = -EINTR;
2347 goto out_uninit_ctrl;
2348 }
2349
2350 ret = nvme_tcp_setup_ctrl(&ctrl->ctrl, true);
2351 if (ret)
2352 goto out_uninit_ctrl;
2353
2354 dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISp\n",
2355 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
2356
2357 nvme_get_ctrl(&ctrl->ctrl);
2358
2359 mutex_lock(&nvme_tcp_ctrl_mutex);
2360 list_add_tail(&ctrl->list, &nvme_tcp_ctrl_list);
2361 mutex_unlock(&nvme_tcp_ctrl_mutex);
2362
2363 return &ctrl->ctrl;
2364
2365 out_uninit_ctrl:
2366 nvme_uninit_ctrl(&ctrl->ctrl);
2367 nvme_put_ctrl(&ctrl->ctrl);
2368 if (ret > 0)
2369 ret = -EIO;
2370 return ERR_PTR(ret);
2371 out_kfree_queues:
2372 kfree(ctrl->queues);
2373 out_free_ctrl:
2374 kfree(ctrl);
2375 return ERR_PTR(ret);
2376 }
2377
2378 static struct nvmf_transport_ops nvme_tcp_transport = {
2379 .name = "tcp",
2380 .module = THIS_MODULE,
2381 .required_opts = NVMF_OPT_TRADDR,
2382 .allowed_opts = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2383 NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
2384 NVMF_OPT_HDR_DIGEST | NVMF_OPT_DATA_DIGEST |
2385 NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES |
2386 NVMF_OPT_TOS,
2387 .create_ctrl = nvme_tcp_create_ctrl,
2388 };
2389
2390 static int __init nvme_tcp_init_module(void)
2391 {
2392 nvme_tcp_wq = alloc_workqueue("nvme_tcp_wq",
2393 WQ_MEM_RECLAIM | WQ_HIGHPRI, 0);
2394 if (!nvme_tcp_wq)
2395 return -ENOMEM;
2396
2397 nvmf_register_transport(&nvme_tcp_transport);
2398 return 0;
2399 }
2400
2401 static void __exit nvme_tcp_cleanup_module(void)
2402 {
2403 struct nvme_tcp_ctrl *ctrl;
2404
2405 nvmf_unregister_transport(&nvme_tcp_transport);
2406
2407 mutex_lock(&nvme_tcp_ctrl_mutex);
2408 list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list)
2409 nvme_delete_ctrl(&ctrl->ctrl);
2410 mutex_unlock(&nvme_tcp_ctrl_mutex);
2411 flush_workqueue(nvme_delete_wq);
2412
2413 destroy_workqueue(nvme_tcp_wq);
2414 }
2415
2416 module_init(nvme_tcp_init_module);
2417 module_exit(nvme_tcp_cleanup_module);
2418
2419 MODULE_LICENSE("GPL v2");
Linux with added FPC-III support