projects / linux-2.6-block.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
nvme: implement In-Band authentication
[linux-2.6-block.git] / drivers / nvme / host / tcp.c
CommitLineData
3f2304f8
SG		 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>
1a9460ce 16#include <net/busy_poll.h>
3f2304f8
SG		 17
18#include "nvme.h"
19#include "fabrics.h"
20
21struct nvme_tcp_queue;
22
9912ade3
WM		 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 */
29static int so_priority;
30module_param(so_priority, int, 0644);
31MODULE_PARM_DESC(so_priority, "nvme tcp socket optimize priority");
32
841aee4d
CL		 33#ifdef CONFIG_DEBUG_LOCK_ALLOC
34/* lockdep can detect a circular dependency of the form
35 * sk_lock -> mmap_lock (page fault) -> fs locks -> sk_lock
36 * because dependencies are tracked for both nvme-tcp and user contexts. Using
37 * a separate class prevents lockdep from conflating nvme-tcp socket use with
38 * user-space socket API use.
39 */
40static struct lock_class_key nvme_tcp_sk_key[2];
41static struct lock_class_key nvme_tcp_slock_key[2];
42
43static void nvme_tcp_reclassify_socket(struct socket *sock)
44{
45 struct sock *sk = sock->sk;
46
47 if (WARN_ON_ONCE(!sock_allow_reclassification(sk)))
48 return;
49
50 switch (sk->sk_family) {
51 case AF_INET:
52 sock_lock_init_class_and_name(sk, "slock-AF_INET-NVME",
53 &nvme_tcp_slock_key[0],
54 "sk_lock-AF_INET-NVME",
55 &nvme_tcp_sk_key[0]);
56 break;
57 case AF_INET6:
58 sock_lock_init_class_and_name(sk, "slock-AF_INET6-NVME",
59 &nvme_tcp_slock_key[1],
60 "sk_lock-AF_INET6-NVME",
61 &nvme_tcp_sk_key[1]);
62 break;
63 default:
64 WARN_ON_ONCE(1);
65 }
66}
67#else
68static void nvme_tcp_reclassify_socket(struct socket *sock) { }
69#endif
70
3f2304f8
SG		 71enum nvme_tcp_send_state {
72 NVME_TCP_SEND_CMD_PDU = 0,
73 NVME_TCP_SEND_H2C_PDU,
74 NVME_TCP_SEND_DATA,
75 NVME_TCP_SEND_DDGST,
76};
77
78struct nvme_tcp_request {
79 struct nvme_request req;
80 void *pdu;
81 struct nvme_tcp_queue *queue;
82 u32 data_len;
83 u32 pdu_len;
84 u32 pdu_sent;
c2700d28
VP		 85 u32 h2cdata_left;
86 u32 h2cdata_offset;
3f2304f8 87 u16 ttag;
1ba2e507 88 __le16 status;
3f2304f8 89 struct list_head entry;
15ec928a 90 struct llist_node lentry;
a7273d40 91 __le32 ddgst;
3f2304f8
SG		 92
93 struct bio *curr_bio;
94 struct iov_iter iter;
95
96 /* send state */
97 size_t offset;
98 size_t data_sent;
99 enum nvme_tcp_send_state state;
100};
101
102enum nvme_tcp_queue_flags {
103 NVME_TCP_Q_ALLOCATED = 0,
104 NVME_TCP_Q_LIVE = 1,
72e5d757 105 NVME_TCP_Q_POLLING = 2,
3f2304f8
SG		 106};
107
108enum nvme_tcp_recv_state {
109 NVME_TCP_RECV_PDU = 0,
110 NVME_TCP_RECV_DATA,
111 NVME_TCP_RECV_DDGST,
112};
113
114struct nvme_tcp_ctrl;
115struct nvme_tcp_queue {
116 struct socket *sock;
117 struct work_struct io_work;
118 int io_cpu;
119
9ebbfe49 120 struct mutex queue_lock;
db5ad6b7 121 struct mutex send_mutex;
15ec928a 122 struct llist_head req_list;
3f2304f8 123 struct list_head send_list;
122e5b9f 124 bool more_requests;
3f2304f8
SG		 125
126 /* recv state */
127 void *pdu;
128 int pdu_remaining;
129 int pdu_offset;
130 size_t data_remaining;
131 size_t ddgst_remaining;
1a9460ce 132 unsigned int nr_cqe;
3f2304f8
SG		 133
134 /* send state */
135 struct nvme_tcp_request *request;
136
137 int queue_size;
c2700d28 138 u32 maxh2cdata;
3f2304f8
SG		 139 size_t cmnd_capsule_len;
140 struct nvme_tcp_ctrl *ctrl;
141 unsigned long flags;
142 bool rd_enabled;
143
144 bool hdr_digest;
145 bool data_digest;
146 struct ahash_request *rcv_hash;
147 struct ahash_request *snd_hash;
148 __le32 exp_ddgst;
149 __le32 recv_ddgst;
150
151 struct page_frag_cache pf_cache;
152
153 void (*state_change)(struct sock *);
154 void (*data_ready)(struct sock *);
155 void (*write_space)(struct sock *);
156};
157
158struct nvme_tcp_ctrl {
159 /* read only in the hot path */
160 struct nvme_tcp_queue *queues;
161 struct blk_mq_tag_set tag_set;
162
163 /* other member variables */
164 struct list_head list;
165 struct blk_mq_tag_set admin_tag_set;
166 struct sockaddr_storage addr;
167 struct sockaddr_storage src_addr;
168 struct nvme_ctrl ctrl;
169
170 struct work_struct err_work;
171 struct delayed_work connect_work;
172 struct nvme_tcp_request async_req;
64861993 173 u32 io_queues[HCTX_MAX_TYPES];
3f2304f8
SG		 174};
175
176static LIST_HEAD(nvme_tcp_ctrl_list);
177static DEFINE_MUTEX(nvme_tcp_ctrl_mutex);
178static struct workqueue_struct *nvme_tcp_wq;
6acbd961
RF		 179static const struct blk_mq_ops nvme_tcp_mq_ops;
180static const struct blk_mq_ops nvme_tcp_admin_mq_ops;
db5ad6b7 181static int nvme_tcp_try_send(struct nvme_tcp_queue *queue);
3f2304f8
SG		 182
183static inline struct nvme_tcp_ctrl *to_tcp_ctrl(struct nvme_ctrl *ctrl)
184{
185 return container_of(ctrl, struct nvme_tcp_ctrl, ctrl);
186}
187
188static inline int nvme_tcp_queue_id(struct nvme_tcp_queue *queue)
189{
190 return queue - queue->ctrl->queues;
191}
192
193static inline struct blk_mq_tags *nvme_tcp_tagset(struct nvme_tcp_queue *queue)
194{
195 u32 queue_idx = nvme_tcp_queue_id(queue);
196
197 if (queue_idx == 0)
198 return queue->ctrl->admin_tag_set.tags[queue_idx];
199 return queue->ctrl->tag_set.tags[queue_idx - 1];
200}
201
202static inline u8 nvme_tcp_hdgst_len(struct nvme_tcp_queue *queue)
203{
204 return queue->hdr_digest ? NVME_TCP_DIGEST_LENGTH : 0;
205}
206
207static inline u8 nvme_tcp_ddgst_len(struct nvme_tcp_queue *queue)
208{
209 return queue->data_digest ? NVME_TCP_DIGEST_LENGTH : 0;
210}
211
212static inline size_t nvme_tcp_inline_data_size(struct nvme_tcp_queue *queue)
213{
214 return queue->cmnd_capsule_len - sizeof(struct nvme_command);
215}
216
217static inline bool nvme_tcp_async_req(struct nvme_tcp_request *req)
218{
219 return req == &req->queue->ctrl->async_req;
220}
221
222static inline bool nvme_tcp_has_inline_data(struct nvme_tcp_request *req)
223{
224 struct request *rq;
3f2304f8
SG		 225
226 if (unlikely(nvme_tcp_async_req(req)))
227 return false; /* async events don't have a request */
228
229 rq = blk_mq_rq_from_pdu(req);
3f2304f8 230
25e5cb78
SG		 231 return rq_data_dir(rq) == WRITE && req->data_len &&
232 req->data_len <= nvme_tcp_inline_data_size(req->queue);
3f2304f8
SG		 233}
234
235static inline struct page *nvme_tcp_req_cur_page(struct nvme_tcp_request *req)
236{
237 return req->iter.bvec->bv_page;
238}
239
240static inline size_t nvme_tcp_req_cur_offset(struct nvme_tcp_request *req)
241{
242 return req->iter.bvec->bv_offset + req->iter.iov_offset;
243}
244
245static inline size_t nvme_tcp_req_cur_length(struct nvme_tcp_request *req)
246{
ca1ff67d 247 return min_t(size_t, iov_iter_single_seg_count(&req->iter),
3f2304f8
SG		 248 req->pdu_len - req->pdu_sent);
249}
250
3f2304f8
SG		 251static inline size_t nvme_tcp_pdu_data_left(struct nvme_tcp_request *req)
252{
253 return rq_data_dir(blk_mq_rq_from_pdu(req)) == WRITE ?
254 req->pdu_len - req->pdu_sent : 0;
255}
256
257static inline size_t nvme_tcp_pdu_last_send(struct nvme_tcp_request *req,
258 int len)
259{
260 return nvme_tcp_pdu_data_left(req) <= len;
261}
262
263static void nvme_tcp_init_iter(struct nvme_tcp_request *req,
264 unsigned int dir)
265{
266 struct request *rq = blk_mq_rq_from_pdu(req);
267 struct bio_vec *vec;
268 unsigned int size;
0dc9edaf 269 int nr_bvec;
3f2304f8
SG		 270 size_t offset;
271
272 if (rq->rq_flags & RQF_SPECIAL_PAYLOAD) {
273 vec = &rq->special_vec;
0dc9edaf 274 nr_bvec = 1;
3f2304f8
SG		 275 size = blk_rq_payload_bytes(rq);
276 offset = 0;
277 } else {
278 struct bio *bio = req->curr_bio;
0dc9edaf
SG		 279 struct bvec_iter bi;
280 struct bio_vec bv;
3f2304f8
SG		 281
282 vec = __bvec_iter_bvec(bio->bi_io_vec, bio->bi_iter);
0dc9edaf
SG		 283 nr_bvec = 0;
284 bio_for_each_bvec(bv, bio, bi) {
285 nr_bvec++;
286 }
3f2304f8
SG		 287 size = bio->bi_iter.bi_size;
288 offset = bio->bi_iter.bi_bvec_done;
289 }
290
0dc9edaf 291 iov_iter_bvec(&req->iter, dir, vec, nr_bvec, size);
3f2304f8
SG		 292 req->iter.iov_offset = offset;
293}
294
295static inline void nvme_tcp_advance_req(struct nvme_tcp_request *req,
296 int len)
297{
298 req->data_sent += len;
299 req->pdu_sent += len;
300 iov_iter_advance(&req->iter, len);
301 if (!iov_iter_count(&req->iter) &&
302 req->data_sent < req->data_len) {
303 req->curr_bio = req->curr_bio->bi_next;
304 nvme_tcp_init_iter(req, WRITE);
305 }
306}
307
5c11f7d9
SG		 308static inline void nvme_tcp_send_all(struct nvme_tcp_queue *queue)
309{
310 int ret;
311
312 /* drain the send queue as much as we can... */
313 do {
314 ret = nvme_tcp_try_send(queue);
315 } while (ret > 0);
316}
317
70f437fb
KB		 318static inline bool nvme_tcp_queue_more(struct nvme_tcp_queue *queue)
319{
320 return !list_empty(&queue->send_list) ||
321 !llist_empty(&queue->req_list) || queue->more_requests;
322}
323
db5ad6b7 324static inline void nvme_tcp_queue_request(struct nvme_tcp_request *req,
86f0348a 325 bool sync, bool last)
3f2304f8
SG		 326{
327 struct nvme_tcp_queue *queue = req->queue;
db5ad6b7 328 bool empty;
3f2304f8 329
15ec928a
SG		 330 empty = llist_add(&req->lentry, &queue->req_list) &&
331 list_empty(&queue->send_list) && !queue->request;
3f2304f8 332
db5ad6b7
SG		 333 /*
334 * if we're the first on the send_list and we can try to send
335 * directly, otherwise queue io_work. Also, only do that if we
336 * are on the same cpu, so we don't introduce contention.
337 */
bb833370 338 if (queue->io_cpu == raw_smp_processor_id() &&
db5ad6b7 339 sync && empty && mutex_trylock(&queue->send_mutex)) {
122e5b9f 340 queue->more_requests = !last;
5c11f7d9 341 nvme_tcp_send_all(queue);
122e5b9f 342 queue->more_requests = false;
db5ad6b7 343 mutex_unlock(&queue->send_mutex);
db5ad6b7 344 }
70f437fb
KB		 345
346 if (last && nvme_tcp_queue_more(queue))
347 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
3f2304f8
SG		 348}
349
15ec928a
SG		 350static void nvme_tcp_process_req_list(struct nvme_tcp_queue *queue)
351{
352 struct nvme_tcp_request *req;
353 struct llist_node *node;
354
355 for (node = llist_del_all(&queue->req_list); node; node = node->next) {
356 req = llist_entry(node, struct nvme_tcp_request, lentry);
357 list_add(&req->entry, &queue->send_list);
358 }
359}
360
3f2304f8
SG		 361static inline struct nvme_tcp_request *
362nvme_tcp_fetch_request(struct nvme_tcp_queue *queue)
363{
364 struct nvme_tcp_request *req;
365
3f2304f8
SG		 366 req = list_first_entry_or_null(&queue->send_list,
367 struct nvme_tcp_request, entry);
15ec928a
SG		 368 if (!req) {
369 nvme_tcp_process_req_list(queue);
370 req = list_first_entry_or_null(&queue->send_list,
371 struct nvme_tcp_request, entry);
372 if (unlikely(!req))
373 return NULL;
374 }
3f2304f8 375
15ec928a 376 list_del(&req->entry);
3f2304f8
SG		 377 return req;
378}
379
a7273d40
CH		 380static inline void nvme_tcp_ddgst_final(struct ahash_request *hash,
381 __le32 *dgst)
3f2304f8
SG		 382{
383 ahash_request_set_crypt(hash, NULL, (u8 *)dgst, 0);
384 crypto_ahash_final(hash);
385}
386
387static inline void nvme_tcp_ddgst_update(struct ahash_request *hash,
388 struct page *page, off_t off, size_t len)
389{
390 struct scatterlist sg;
391
392 sg_init_marker(&sg, 1);
393 sg_set_page(&sg, page, len, off);
394 ahash_request_set_crypt(hash, &sg, NULL, len);
395 crypto_ahash_update(hash);
396}
397
398static inline void nvme_tcp_hdgst(struct ahash_request *hash,
399 void *pdu, size_t len)
400{
401 struct scatterlist sg;
402
403 sg_init_one(&sg, pdu, len);
404 ahash_request_set_crypt(hash, &sg, pdu + len, len);
405 crypto_ahash_digest(hash);
406}
407
408static int nvme_tcp_verify_hdgst(struct nvme_tcp_queue *queue,
409 void *pdu, size_t pdu_len)
410{
411 struct nvme_tcp_hdr *hdr = pdu;
412 __le32 recv_digest;
413 __le32 exp_digest;
414
415 if (unlikely(!(hdr->flags & NVME_TCP_F_HDGST))) {
416 dev_err(queue->ctrl->ctrl.device,
417 "queue %d: header digest flag is cleared\n",
418 nvme_tcp_queue_id(queue));
419 return -EPROTO;
420 }
421
422 recv_digest = *(__le32 *)(pdu + hdr->hlen);
423 nvme_tcp_hdgst(queue->rcv_hash, pdu, pdu_len);
424 exp_digest = *(__le32 *)(pdu + hdr->hlen);
425 if (recv_digest != exp_digest) {
426 dev_err(queue->ctrl->ctrl.device,
427 "header digest error: recv %#x expected %#x\n",
428 le32_to_cpu(recv_digest), le32_to_cpu(exp_digest));
429 return -EIO;
430 }
431
432 return 0;
433}
434
435static int nvme_tcp_check_ddgst(struct nvme_tcp_queue *queue, void *pdu)
436{
437 struct nvme_tcp_hdr *hdr = pdu;
438 u8 digest_len = nvme_tcp_hdgst_len(queue);
439 u32 len;
440
441 len = le32_to_cpu(hdr->plen) - hdr->hlen -
442 ((hdr->flags & NVME_TCP_F_HDGST) ? digest_len : 0);
443
444 if (unlikely(len && !(hdr->flags & NVME_TCP_F_DDGST))) {
445 dev_err(queue->ctrl->ctrl.device,
446 "queue %d: data digest flag is cleared\n",
447 nvme_tcp_queue_id(queue));
448 return -EPROTO;
449 }
450 crypto_ahash_init(queue->rcv_hash);
451
452 return 0;
453}
454
455static void nvme_tcp_exit_request(struct blk_mq_tag_set *set,
456 struct request *rq, unsigned int hctx_idx)
457{
458 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
459
460 page_frag_free(req->pdu);
461}
462
463static int nvme_tcp_init_request(struct blk_mq_tag_set *set,
464 struct request *rq, unsigned int hctx_idx,
465 unsigned int numa_node)
466{
467 struct nvme_tcp_ctrl *ctrl = set->driver_data;
468 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
f4b9e6c9 469 struct nvme_tcp_cmd_pdu *pdu;
3f2304f8
SG		 470 int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
471 struct nvme_tcp_queue *queue = &ctrl->queues[queue_idx];
472 u8 hdgst = nvme_tcp_hdgst_len(queue);
473
474 req->pdu = page_frag_alloc(&queue->pf_cache,
475 sizeof(struct nvme_tcp_cmd_pdu) + hdgst,
476 GFP_KERNEL | __GFP_ZERO);
477 if (!req->pdu)
478 return -ENOMEM;
479
f4b9e6c9 480 pdu = req->pdu;
3f2304f8
SG		 481 req->queue = queue;
482 nvme_req(rq)->ctrl = &ctrl->ctrl;
f4b9e6c9 483 nvme_req(rq)->cmd = &pdu->cmd;
3f2304f8
SG		 484
485 return 0;
486}
487
488static int nvme_tcp_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
489 unsigned int hctx_idx)
490{
491 struct nvme_tcp_ctrl *ctrl = data;
492 struct nvme_tcp_queue *queue = &ctrl->queues[hctx_idx + 1];
493
494 hctx->driver_data = queue;
495 return 0;
496}
497
498static int nvme_tcp_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
499 unsigned int hctx_idx)
500{
501 struct nvme_tcp_ctrl *ctrl = data;
502 struct nvme_tcp_queue *queue = &ctrl->queues[0];
503
504 hctx->driver_data = queue;
505 return 0;
506}
507
508static enum nvme_tcp_recv_state
509nvme_tcp_recv_state(struct nvme_tcp_queue *queue)
510{
511 return (queue->pdu_remaining) ? NVME_TCP_RECV_PDU :
512 (queue->ddgst_remaining) ? NVME_TCP_RECV_DDGST :
513 NVME_TCP_RECV_DATA;
514}
515
516static void nvme_tcp_init_recv_ctx(struct nvme_tcp_queue *queue)
517{
518 queue->pdu_remaining = sizeof(struct nvme_tcp_rsp_pdu) +
519 nvme_tcp_hdgst_len(queue);
520 queue->pdu_offset = 0;
521 queue->data_remaining = -1;
522 queue->ddgst_remaining = 0;
523}
524
525static void nvme_tcp_error_recovery(struct nvme_ctrl *ctrl)
526{
527 if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_RESETTING))
528 return;
529
236187c4 530 dev_warn(ctrl->device, "starting error recovery\n");
97b2512a 531 queue_work(nvme_reset_wq, &to_tcp_ctrl(ctrl)->err_work);
3f2304f8
SG		 532}
533
534static int nvme_tcp_process_nvme_cqe(struct nvme_tcp_queue *queue,
535 struct nvme_completion *cqe)
536{
1ba2e507 537 struct nvme_tcp_request *req;
3f2304f8
SG		 538 struct request *rq;
539
e7006de6 540 rq = nvme_find_rq(nvme_tcp_tagset(queue), cqe->command_id);
3f2304f8
SG		 541 if (!rq) {
542 dev_err(queue->ctrl->ctrl.device,
e7006de6
SG		 543 "got bad cqe.command_id %#x on queue %d\n",
544 cqe->command_id, nvme_tcp_queue_id(queue));
3f2304f8
SG		 545 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
546 return -EINVAL;
547 }
548
1ba2e507
DW		 549 req = blk_mq_rq_to_pdu(rq);
550 if (req->status == cpu_to_le16(NVME_SC_SUCCESS))
551 req->status = cqe->status;
552
553 if (!nvme_try_complete_req(rq, req->status, cqe->result))
ff029451 554 nvme_complete_rq(rq);
1a9460ce 555 queue->nr_cqe++;
3f2304f8
SG		 556
557 return 0;
558}
559
560static int nvme_tcp_handle_c2h_data(struct nvme_tcp_queue *queue,
561 struct nvme_tcp_data_pdu *pdu)
562{
563 struct request *rq;
564
e7006de6 565 rq = nvme_find_rq(nvme_tcp_tagset(queue), pdu->command_id);
3f2304f8
SG		 566 if (!rq) {
567 dev_err(queue->ctrl->ctrl.device,
e7006de6
SG		 568 "got bad c2hdata.command_id %#x on queue %d\n",
569 pdu->command_id, nvme_tcp_queue_id(queue));
3f2304f8
SG		 570 return -ENOENT;
571 }
572
573 if (!blk_rq_payload_bytes(rq)) {
574 dev_err(queue->ctrl->ctrl.device,
575 "queue %d tag %#x unexpected data\n",
576 nvme_tcp_queue_id(queue), rq->tag);
577 return -EIO;
578 }
579
580 queue->data_remaining = le32_to_cpu(pdu->data_length);
581
602d674c
SG		 582 if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS &&
583 unlikely(!(pdu->hdr.flags & NVME_TCP_F_DATA_LAST))) {
584 dev_err(queue->ctrl->ctrl.device,
585 "queue %d tag %#x SUCCESS set but not last PDU\n",
586 nvme_tcp_queue_id(queue), rq->tag);
587 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
588 return -EPROTO;
589 }
590
3f2304f8 591 return 0;
3f2304f8
SG		 592}
593
594static int nvme_tcp_handle_comp(struct nvme_tcp_queue *queue,
595 struct nvme_tcp_rsp_pdu *pdu)
596{
597 struct nvme_completion *cqe = &pdu->cqe;
598 int ret = 0;
599
600 /*
601 * AEN requests are special as they don't time out and can
602 * survive any kind of queue freeze and often don't respond to
603 * aborts. We don't even bother to allocate a struct request
604 * for them but rather special case them here.
605 */
58a8df67
IR		 606 if (unlikely(nvme_is_aen_req(nvme_tcp_queue_id(queue),
607 cqe->command_id)))
3f2304f8
SG		 608 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
609 &cqe->result);
610 else
611 ret = nvme_tcp_process_nvme_cqe(queue, cqe);
612
613 return ret;
614}
615
c2700d28 616static void nvme_tcp_setup_h2c_data_pdu(struct nvme_tcp_request *req)
3f2304f8
SG		 617{
618 struct nvme_tcp_data_pdu *data = req->pdu;
619 struct nvme_tcp_queue *queue = req->queue;
620 struct request *rq = blk_mq_rq_from_pdu(req);
c2700d28 621 u32 h2cdata_sent = req->pdu_len;
3f2304f8
SG		 622 u8 hdgst = nvme_tcp_hdgst_len(queue);
623 u8 ddgst = nvme_tcp_ddgst_len(queue);
624
1d3ef9c3
VP		 625 req->state = NVME_TCP_SEND_H2C_PDU;
626 req->offset = 0;
c2700d28 627 req->pdu_len = min(req->h2cdata_left, queue->maxh2cdata);
3f2304f8 628 req->pdu_sent = 0;
c2700d28
VP		 629 req->h2cdata_left -= req->pdu_len;
630 req->h2cdata_offset += h2cdata_sent;
3f2304f8 631
3f2304f8
SG		 632 memset(data, 0, sizeof(*data));
633 data->hdr.type = nvme_tcp_h2c_data;
c2700d28
VP		 634 if (!req->h2cdata_left)
635 data->hdr.flags = NVME_TCP_F_DATA_LAST;
3f2304f8
SG		 636 if (queue->hdr_digest)
637 data->hdr.flags |= NVME_TCP_F_HDGST;
638 if (queue->data_digest)
639 data->hdr.flags |= NVME_TCP_F_DDGST;
640 data->hdr.hlen = sizeof(*data);
641 data->hdr.pdo = data->hdr.hlen + hdgst;
642 data->hdr.plen =
643 cpu_to_le32(data->hdr.hlen + hdgst + req->pdu_len + ddgst);
c2700d28 644 data->ttag = req->ttag;
e7006de6 645 data->command_id = nvme_cid(rq);
c2700d28 646 data->data_offset = cpu_to_le32(req->h2cdata_offset);
3f2304f8 647 data->data_length = cpu_to_le32(req->pdu_len);
3f2304f8
SG		 648}
649
650static int nvme_tcp_handle_r2t(struct nvme_tcp_queue *queue,
651 struct nvme_tcp_r2t_pdu *pdu)
652{
653 struct nvme_tcp_request *req;
654 struct request *rq;
1d3ef9c3 655 u32 r2t_length = le32_to_cpu(pdu->r2t_length);
c2700d28 656 u32 r2t_offset = le32_to_cpu(pdu->r2t_offset);
3f2304f8 657
e7006de6 658 rq = nvme_find_rq(nvme_tcp_tagset(queue), pdu->command_id);
3f2304f8
SG		 659 if (!rq) {
660 dev_err(queue->ctrl->ctrl.device,
e7006de6
SG		 661 "got bad r2t.command_id %#x on queue %d\n",
662 pdu->command_id, nvme_tcp_queue_id(queue));
3f2304f8
SG		 663 return -ENOENT;
664 }
665 req = blk_mq_rq_to_pdu(rq);
666
1d3ef9c3
VP		 667 if (unlikely(!r2t_length)) {
668 dev_err(queue->ctrl->ctrl.device,
669 "req %d r2t len is %u, probably a bug...\n",
670 rq->tag, r2t_length);
671 return -EPROTO;
672 }
3f2304f8 673
1d3ef9c3
VP		 674 if (unlikely(req->data_sent + r2t_length > req->data_len)) {
675 dev_err(queue->ctrl->ctrl.device,
676 "req %d r2t len %u exceeded data len %u (%zu sent)\n",
677 rq->tag, r2t_length, req->data_len, req->data_sent);
678 return -EPROTO;
679 }
680
c2700d28 681 if (unlikely(r2t_offset < req->data_sent)) {
1d3ef9c3
VP		 682 dev_err(queue->ctrl->ctrl.device,
683 "req %d unexpected r2t offset %u (expected %zu)\n",
c2700d28 684 rq->tag, r2t_offset, req->data_sent);
1d3ef9c3
VP		 685 return -EPROTO;
686 }
3f2304f8 687
c2700d28
VP		 688 req->pdu_len = 0;
689 req->h2cdata_left = r2t_length;
690 req->h2cdata_offset = r2t_offset;
691 req->ttag = pdu->ttag;
692
693 nvme_tcp_setup_h2c_data_pdu(req);
86f0348a 694 nvme_tcp_queue_request(req, false, true);
3f2304f8
SG		 695
696 return 0;
697}
698
699static int nvme_tcp_recv_pdu(struct nvme_tcp_queue *queue, struct sk_buff *skb,
700 unsigned int *offset, size_t *len)
701{
702 struct nvme_tcp_hdr *hdr;
703 char *pdu = queue->pdu;
704 size_t rcv_len = min_t(size_t, *len, queue->pdu_remaining);
705 int ret;
706
707 ret = skb_copy_bits(skb, *offset,
708 &pdu[queue->pdu_offset], rcv_len);
709 if (unlikely(ret))
710 return ret;
711
712 queue->pdu_remaining -= rcv_len;
713 queue->pdu_offset += rcv_len;
714 *offset += rcv_len;
715 *len -= rcv_len;
716 if (queue->pdu_remaining)
717 return 0;
718
719 hdr = queue->pdu;
720 if (queue->hdr_digest) {
721 ret = nvme_tcp_verify_hdgst(queue, queue->pdu, hdr->hlen);
722 if (unlikely(ret))
723 return ret;
724 }
725
726
727 if (queue->data_digest) {
728 ret = nvme_tcp_check_ddgst(queue, queue->pdu);
729 if (unlikely(ret))
730 return ret;
731 }
732
733 switch (hdr->type) {
734 case nvme_tcp_c2h_data:
6be18260 735 return nvme_tcp_handle_c2h_data(queue, (void *)queue->pdu);
3f2304f8
SG		 736 case nvme_tcp_rsp:
737 nvme_tcp_init_recv_ctx(queue);
6be18260 738 return nvme_tcp_handle_comp(queue, (void *)queue->pdu);
3f2304f8
SG		 739 case nvme_tcp_r2t:
740 nvme_tcp_init_recv_ctx(queue);
6be18260 741 return nvme_tcp_handle_r2t(queue, (void *)queue->pdu);
3f2304f8
SG		 742 default:
743 dev_err(queue->ctrl->ctrl.device,
744 "unsupported pdu type (%d)\n", hdr->type);
745 return -EINVAL;
746 }
3f2304f8
SG		 747}
748
988aef9e 749static inline void nvme_tcp_end_request(struct request *rq, u16 status)
602d674c
SG		 750{
751 union nvme_result res = {};
752
2eb81a33 753 if (!nvme_try_complete_req(rq, cpu_to_le16(status << 1), res))
ff029451 754 nvme_complete_rq(rq);
602d674c
SG		 755}
756
3f2304f8
SG		 757static int nvme_tcp_recv_data(struct nvme_tcp_queue *queue, struct sk_buff *skb,
758 unsigned int *offset, size_t *len)
759{
760 struct nvme_tcp_data_pdu *pdu = (void *)queue->pdu;
3b01a9d0 761 struct request *rq =
e7006de6 762 nvme_cid_to_rq(nvme_tcp_tagset(queue), pdu->command_id);
3b01a9d0 763 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
3f2304f8
SG		 764
765 while (true) {
766 int recv_len, ret;
767
768 recv_len = min_t(size_t, *len, queue->data_remaining);
769 if (!recv_len)
770 break;
771
772 if (!iov_iter_count(&req->iter)) {
773 req->curr_bio = req->curr_bio->bi_next;
774
775 /*
776 * If we don`t have any bios it means that controller
777 * sent more data than we requested, hence error
778 */
779 if (!req->curr_bio) {
780 dev_err(queue->ctrl->ctrl.device,
781 "queue %d no space in request %#x",
782 nvme_tcp_queue_id(queue), rq->tag);
783 nvme_tcp_init_recv_ctx(queue);
784 return -EIO;
785 }
786 nvme_tcp_init_iter(req, READ);
787 }
788
789 /* we can read only from what is left in this bio */
790 recv_len = min_t(size_t, recv_len,
791 iov_iter_count(&req->iter));
792
793 if (queue->data_digest)
794 ret = skb_copy_and_hash_datagram_iter(skb, *offset,
795 &req->iter, recv_len, queue->rcv_hash);
796 else
797 ret = skb_copy_datagram_iter(skb, *offset,
798 &req->iter, recv_len);
799 if (ret) {
800 dev_err(queue->ctrl->ctrl.device,
801 "queue %d failed to copy request %#x data",
802 nvme_tcp_queue_id(queue), rq->tag);
803 return ret;
804 }
805
806 *len -= recv_len;
807 *offset += recv_len;
808 queue->data_remaining -= recv_len;
809 }
810
811 if (!queue->data_remaining) {
812 if (queue->data_digest) {
813 nvme_tcp_ddgst_final(queue->rcv_hash, &queue->exp_ddgst);
814 queue->ddgst_remaining = NVME_TCP_DIGEST_LENGTH;
815 } else {
1a9460ce 816 if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
1ba2e507
DW		 817 nvme_tcp_end_request(rq,
818 le16_to_cpu(req->status));
1a9460ce
SG		 819 queue->nr_cqe++;
820 }
3f2304f8
SG		 821 nvme_tcp_init_recv_ctx(queue);
822 }
823 }
824
825 return 0;
826}
827
828static int nvme_tcp_recv_ddgst(struct nvme_tcp_queue *queue,
829 struct sk_buff *skb, unsigned int *offset, size_t *len)
830{
602d674c 831 struct nvme_tcp_data_pdu *pdu = (void *)queue->pdu;
3f2304f8
SG		 832 char *ddgst = (char *)&queue->recv_ddgst;
833 size_t recv_len = min_t(size_t, *len, queue->ddgst_remaining);
834 off_t off = NVME_TCP_DIGEST_LENGTH - queue->ddgst_remaining;
835 int ret;
836
837 ret = skb_copy_bits(skb, *offset, &ddgst[off], recv_len);
838 if (unlikely(ret))
839 return ret;
840
841 queue->ddgst_remaining -= recv_len;
842 *offset += recv_len;
843 *len -= recv_len;
844 if (queue->ddgst_remaining)
845 return 0;
846
847 if (queue->recv_ddgst != queue->exp_ddgst) {
1ba2e507
DW		 848 struct request *rq = nvme_cid_to_rq(nvme_tcp_tagset(queue),
849 pdu->command_id);
850 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
851
852 req->status = cpu_to_le16(NVME_SC_DATA_XFER_ERROR);
853
3f2304f8
SG		 854 dev_err(queue->ctrl->ctrl.device,
855 "data digest error: recv %#x expected %#x\n",
856 le32_to_cpu(queue->recv_ddgst),
857 le32_to_cpu(queue->exp_ddgst));
3f2304f8
SG		 858 }
859
602d674c 860 if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
e7006de6
SG		 861 struct request *rq = nvme_cid_to_rq(nvme_tcp_tagset(queue),
862 pdu->command_id);
1ba2e507 863 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
602d674c 864
1ba2e507 865 nvme_tcp_end_request(rq, le16_to_cpu(req->status));
1a9460ce 866 queue->nr_cqe++;
602d674c
SG		 867 }
868
3f2304f8
SG		 869 nvme_tcp_init_recv_ctx(queue);
870 return 0;
871}
872
873static int nvme_tcp_recv_skb(read_descriptor_t *desc, struct sk_buff *skb,
874 unsigned int offset, size_t len)
875{
876 struct nvme_tcp_queue *queue = desc->arg.data;
877 size_t consumed = len;
878 int result;
879
880 while (len) {
881 switch (nvme_tcp_recv_state(queue)) {
882 case NVME_TCP_RECV_PDU:
883 result = nvme_tcp_recv_pdu(queue, skb, &offset, &len);
884 break;
885 case NVME_TCP_RECV_DATA:
886 result = nvme_tcp_recv_data(queue, skb, &offset, &len);
887 break;
888 case NVME_TCP_RECV_DDGST:
889 result = nvme_tcp_recv_ddgst(queue, skb, &offset, &len);
890 break;
891 default:
892 result = -EFAULT;
893 }
894 if (result) {
895 dev_err(queue->ctrl->ctrl.device,
896 "receive failed: %d\n", result);
897 queue->rd_enabled = false;
898 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
899 return result;
900 }
901 }
902
903 return consumed;
904}
905
906static void nvme_tcp_data_ready(struct sock *sk)
907{
908 struct nvme_tcp_queue *queue;
909
386e5e6e 910 read_lock_bh(&sk->sk_callback_lock);
3f2304f8 911 queue = sk->sk_user_data;
72e5d757
SG		 912 if (likely(queue && queue->rd_enabled) &&
913 !test_bit(NVME_TCP_Q_POLLING, &queue->flags))
3f2304f8 914 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
386e5e6e 915 read_unlock_bh(&sk->sk_callback_lock);
3f2304f8
SG		 916}
917
918static void nvme_tcp_write_space(struct sock *sk)
919{
920 struct nvme_tcp_queue *queue;
921
922 read_lock_bh(&sk->sk_callback_lock);
923 queue = sk->sk_user_data;
924 if (likely(queue && sk_stream_is_writeable(sk))) {
925 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
926 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
927 }
928 read_unlock_bh(&sk->sk_callback_lock);
929}
930
931static void nvme_tcp_state_change(struct sock *sk)
932{
933 struct nvme_tcp_queue *queue;
934
8b73b45d 935 read_lock_bh(&sk->sk_callback_lock);
3f2304f8
SG		 936 queue = sk->sk_user_data;
937 if (!queue)
938 goto done;
939
940 switch (sk->sk_state) {
941 case TCP_CLOSE:
942 case TCP_CLOSE_WAIT:
943 case TCP_LAST_ACK:
944 case TCP_FIN_WAIT1:
945 case TCP_FIN_WAIT2:
3f2304f8
SG		 946 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
947 break;
948 default:
949 dev_info(queue->ctrl->ctrl.device,
950 "queue %d socket state %d\n",
951 nvme_tcp_queue_id(queue), sk->sk_state);
952 }
953
954 queue->state_change(sk);
955done:
8b73b45d 956 read_unlock_bh(&sk->sk_callback_lock);
3f2304f8
SG		 957}
958
959static inline void nvme_tcp_done_send_req(struct nvme_tcp_queue *queue)
960{
961 queue->request = NULL;
962}
963
964static void nvme_tcp_fail_request(struct nvme_tcp_request *req)
965{
63573807
SG		 966 if (nvme_tcp_async_req(req)) {
967 union nvme_result res = {};
968
969 nvme_complete_async_event(&req->queue->ctrl->ctrl,
970 cpu_to_le16(NVME_SC_HOST_PATH_ERROR), &res);
971 } else {
972 nvme_tcp_end_request(blk_mq_rq_from_pdu(req),
973 NVME_SC_HOST_PATH_ERROR);
974 }
3f2304f8
SG		 975}
976
977static int nvme_tcp_try_send_data(struct nvme_tcp_request *req)
978{
979 struct nvme_tcp_queue *queue = req->queue;
25e1f67e 980 int req_data_len = req->data_len;
c2700d28 981 u32 h2cdata_left = req->h2cdata_left;
3f2304f8
SG		 982
983 while (true) {
984 struct page *page = nvme_tcp_req_cur_page(req);
985 size_t offset = nvme_tcp_req_cur_offset(req);
986 size_t len = nvme_tcp_req_cur_length(req);
987 bool last = nvme_tcp_pdu_last_send(req, len);
25e1f67e 988 int req_data_sent = req->data_sent;
3f2304f8
SG		 989 int ret, flags = MSG_DONTWAIT;
990
122e5b9f 991 if (last && !queue->data_digest && !nvme_tcp_queue_more(queue))
3f2304f8
SG		 992 flags |= MSG_EOR;
993 else
5bb052d7 994 flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST;
3f2304f8 995
7d4194ab
CL		 996 if (sendpage_ok(page)) {
997 ret = kernel_sendpage(queue->sock, page, offset, len,
37c15219
MS		 998 flags);
999 } else {
7d4194ab 1000 ret = sock_no_sendpage(queue->sock, page, offset, len,
37c15219
MS		 1001 flags);
1002 }
3f2304f8
SG		 1003 if (ret <= 0)
1004 return ret;
1005
3f2304f8
SG		 1006 if (queue->data_digest)
1007 nvme_tcp_ddgst_update(queue->snd_hash, page,
1008 offset, ret);
1009
e371af03
SG		 1010 /*
1011 * update the request iterator except for the last payload send
1012 * in the request where we don't want to modify it as we may
1013 * compete with the RX path completing the request.
1014 */
25e1f67e 1015 if (req_data_sent + ret < req_data_len)
e371af03
SG		 1016 nvme_tcp_advance_req(req, ret);
1017
1018 /* fully successful last send in current PDU */
3f2304f8
SG		 1019 if (last && ret == len) {
1020 if (queue->data_digest) {
1021 nvme_tcp_ddgst_final(queue->snd_hash,
1022 &req->ddgst);
1023 req->state = NVME_TCP_SEND_DDGST;
1024 req->offset = 0;
1025 } else {
c2700d28
VP		 1026 if (h2cdata_left)
1027 nvme_tcp_setup_h2c_data_pdu(req);
1028 else
1029 nvme_tcp_done_send_req(queue);
3f2304f8
SG		 1030 }
1031 return 1;
1032 }
1033 }
1034 return -EAGAIN;
1035}
1036
1037static int nvme_tcp_try_send_cmd_pdu(struct nvme_tcp_request *req)
1038{
1039 struct nvme_tcp_queue *queue = req->queue;
1040 struct nvme_tcp_cmd_pdu *pdu = req->pdu;
1041 bool inline_data = nvme_tcp_has_inline_data(req);
3f2304f8
SG		 1042 u8 hdgst = nvme_tcp_hdgst_len(queue);
1043 int len = sizeof(*pdu) + hdgst - req->offset;
5bb052d7 1044 int flags = MSG_DONTWAIT;
3f2304f8
SG		 1045 int ret;
1046
122e5b9f 1047 if (inline_data || nvme_tcp_queue_more(queue))
5bb052d7
SG		 1048 flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST;
1049 else
1050 flags |= MSG_EOR;
1051
3f2304f8
SG		 1052 if (queue->hdr_digest && !req->offset)
1053 nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
1054
1055 ret = kernel_sendpage(queue->sock, virt_to_page(pdu),
1056 offset_in_page(pdu) + req->offset, len, flags);
1057 if (unlikely(ret <= 0))
1058 return ret;
1059
1060 len -= ret;
1061 if (!len) {
1062 if (inline_data) {
1063 req->state = NVME_TCP_SEND_DATA;
1064 if (queue->data_digest)
1065 crypto_ahash_init(queue->snd_hash);
3f2304f8
SG		 1066 } else {
1067 nvme_tcp_done_send_req(queue);
1068 }
1069 return 1;
1070 }
1071 req->offset += ret;
1072
1073 return -EAGAIN;
1074}
1075
1076static int nvme_tcp_try_send_data_pdu(struct nvme_tcp_request *req)
1077{
1078 struct nvme_tcp_queue *queue = req->queue;
1079 struct nvme_tcp_data_pdu *pdu = req->pdu;
1080 u8 hdgst = nvme_tcp_hdgst_len(queue);
1081 int len = sizeof(*pdu) - req->offset + hdgst;
1082 int ret;
1083
1084 if (queue->hdr_digest && !req->offset)
1085 nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
1086
c2700d28
VP		 1087 if (!req->h2cdata_left)
1088 ret = kernel_sendpage(queue->sock, virt_to_page(pdu),
1089 offset_in_page(pdu) + req->offset, len,
1090 MSG_DONTWAIT | MSG_MORE | MSG_SENDPAGE_NOTLAST);
1091 else
1092 ret = sock_no_sendpage(queue->sock, virt_to_page(pdu),
1093 offset_in_page(pdu) + req->offset, len,
1094 MSG_DONTWAIT | MSG_MORE);
3f2304f8
SG		 1095 if (unlikely(ret <= 0))
1096 return ret;
1097
1098 len -= ret;
1099 if (!len) {
1100 req->state = NVME_TCP_SEND_DATA;
1101 if (queue->data_digest)
1102 crypto_ahash_init(queue->snd_hash);
3f2304f8
SG		 1103 return 1;
1104 }
1105 req->offset += ret;
1106
1107 return -EAGAIN;
1108}
1109
1110static int nvme_tcp_try_send_ddgst(struct nvme_tcp_request *req)
1111{
1112 struct nvme_tcp_queue *queue = req->queue;
ce7723e9 1113 size_t offset = req->offset;
c2700d28 1114 u32 h2cdata_left = req->h2cdata_left;
3f2304f8 1115 int ret;
122e5b9f 1116 struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
3f2304f8 1117 struct kvec iov = {
d89b9f3b 1118 .iov_base = (u8 *)&req->ddgst + req->offset,
3f2304f8
SG		 1119 .iov_len = NVME_TCP_DIGEST_LENGTH - req->offset
1120 };
1121
122e5b9f
SG		 1122 if (nvme_tcp_queue_more(queue))
1123 msg.msg_flags |= MSG_MORE;
1124 else
1125 msg.msg_flags |= MSG_EOR;
1126
3f2304f8
SG		 1127 ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
1128 if (unlikely(ret <= 0))
1129 return ret;
1130
ce7723e9 1131 if (offset + ret == NVME_TCP_DIGEST_LENGTH) {
c2700d28
VP		 1132 if (h2cdata_left)
1133 nvme_tcp_setup_h2c_data_pdu(req);
1134 else
1135 nvme_tcp_done_send_req(queue);
3f2304f8
SG		 1136 return 1;
1137 }
1138
1139 req->offset += ret;
1140 return -EAGAIN;
1141}
1142
1143static int nvme_tcp_try_send(struct nvme_tcp_queue *queue)
1144{
1145 struct nvme_tcp_request *req;
1146 int ret = 1;
1147
1148 if (!queue->request) {
1149 queue->request = nvme_tcp_fetch_request(queue);
1150 if (!queue->request)
1151 return 0;
1152 }
1153 req = queue->request;
1154
1155 if (req->state == NVME_TCP_SEND_CMD_PDU) {
1156 ret = nvme_tcp_try_send_cmd_pdu(req);
1157 if (ret <= 0)
1158 goto done;
1159 if (!nvme_tcp_has_inline_data(req))
1160 return ret;
1161 }
1162
1163 if (req->state == NVME_TCP_SEND_H2C_PDU) {
1164 ret = nvme_tcp_try_send_data_pdu(req);
1165 if (ret <= 0)
1166 goto done;
1167 }
1168
1169 if (req->state == NVME_TCP_SEND_DATA) {
1170 ret = nvme_tcp_try_send_data(req);
1171 if (ret <= 0)
1172 goto done;
1173 }
1174
1175 if (req->state == NVME_TCP_SEND_DDGST)
1176 ret = nvme_tcp_try_send_ddgst(req);
1177done:
5ff4e112 1178 if (ret == -EAGAIN) {
3f2304f8 1179 ret = 0;
5ff4e112
SG		 1180 } else if (ret < 0) {
1181 dev_err(queue->ctrl->ctrl.device,
1182 "failed to send request %d\n", ret);
41d07df7 1183 nvme_tcp_fail_request(queue->request);
5ff4e112
SG		 1184 nvme_tcp_done_send_req(queue);
1185 }
3f2304f8
SG		 1186 return ret;
1187}
1188
1189static int nvme_tcp_try_recv(struct nvme_tcp_queue *queue)
1190{
10407ec9
PBT		 1191 struct socket *sock = queue->sock;
1192 struct sock *sk = sock->sk;
3f2304f8
SG		 1193 read_descriptor_t rd_desc;
1194 int consumed;
1195
1196 rd_desc.arg.data = queue;
1197 rd_desc.count = 1;
1198 lock_sock(sk);
1a9460ce 1199 queue->nr_cqe = 0;
10407ec9 1200 consumed = sock->ops->read_sock(sk, &rd_desc, nvme_tcp_recv_skb);
3f2304f8
SG		 1201 release_sock(sk);
1202 return consumed;
1203}
1204
1205static void nvme_tcp_io_work(struct work_struct *w)
1206{
1207 struct nvme_tcp_queue *queue =
1208 container_of(w, struct nvme_tcp_queue, io_work);
ddef2957 1209 unsigned long deadline = jiffies + msecs_to_jiffies(1);
3f2304f8
SG		 1210
1211 do {
1212 bool pending = false;
1213 int result;
1214
db5ad6b7
SG		 1215 if (mutex_trylock(&queue->send_mutex)) {
1216 result = nvme_tcp_try_send(queue);
1217 mutex_unlock(&queue->send_mutex);
1218 if (result > 0)
1219 pending = true;
1220 else if (unlikely(result < 0))
1221 break;
70f437fb 1222 }
3f2304f8
SG		 1223
1224 result = nvme_tcp_try_recv(queue);
1225 if (result > 0)
1226 pending = true;
761ad26c 1227 else if (unlikely(result < 0))
39d06079 1228 return;
3f2304f8
SG		 1229
1230 if (!pending)
1231 return;
1232
ddef2957 1233 } while (!time_after(jiffies, deadline)); /* quota is exhausted */
3f2304f8
SG		 1234
1235 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
1236}
1237
1238static void nvme_tcp_free_crypto(struct nvme_tcp_queue *queue)
1239{
1240 struct crypto_ahash *tfm = crypto_ahash_reqtfm(queue->rcv_hash);
1241
1242 ahash_request_free(queue->rcv_hash);
1243 ahash_request_free(queue->snd_hash);
1244 crypto_free_ahash(tfm);
1245}
1246
1247static int nvme_tcp_alloc_crypto(struct nvme_tcp_queue *queue)
1248{
1249 struct crypto_ahash *tfm;
1250
1251 tfm = crypto_alloc_ahash("crc32c", 0, CRYPTO_ALG_ASYNC);
1252 if (IS_ERR(tfm))
1253 return PTR_ERR(tfm);
1254
1255 queue->snd_hash = ahash_request_alloc(tfm, GFP_KERNEL);
1256 if (!queue->snd_hash)
1257 goto free_tfm;
1258 ahash_request_set_callback(queue->snd_hash, 0, NULL, NULL);
1259
1260 queue->rcv_hash = ahash_request_alloc(tfm, GFP_KERNEL);
1261 if (!queue->rcv_hash)
1262 goto free_snd_hash;
1263 ahash_request_set_callback(queue->rcv_hash, 0, NULL, NULL);
1264
1265 return 0;
1266free_snd_hash:
1267 ahash_request_free(queue->snd_hash);
1268free_tfm:
1269 crypto_free_ahash(tfm);
1270 return -ENOMEM;
1271}
1272
1273static void nvme_tcp_free_async_req(struct nvme_tcp_ctrl *ctrl)
1274{
1275 struct nvme_tcp_request *async = &ctrl->async_req;
1276
1277 page_frag_free(async->pdu);
1278}
1279
1280static int nvme_tcp_alloc_async_req(struct nvme_tcp_ctrl *ctrl)
1281{
1282 struct nvme_tcp_queue *queue = &ctrl->queues[0];
1283 struct nvme_tcp_request *async = &ctrl->async_req;
1284 u8 hdgst = nvme_tcp_hdgst_len(queue);
1285
1286 async->pdu = page_frag_alloc(&queue->pf_cache,
1287 sizeof(struct nvme_tcp_cmd_pdu) + hdgst,
1288 GFP_KERNEL | __GFP_ZERO);
1289 if (!async->pdu)
1290 return -ENOMEM;
1291
1292 async->queue = &ctrl->queues[0];
1293 return 0;
1294}
1295
1296static void nvme_tcp_free_queue(struct nvme_ctrl *nctrl, int qid)
1297{
a5053c92 1298 struct page *page;
3f2304f8
SG		 1299 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1300 struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1301
1302 if (!test_and_clear_bit(NVME_TCP_Q_ALLOCATED, &queue->flags))
1303 return;
1304
1305 if (queue->hdr_digest || queue->data_digest)
1306 nvme_tcp_free_crypto(queue);
1307
a5053c92
ML		 1308 if (queue->pf_cache.va) {
1309 page = virt_to_head_page(queue->pf_cache.va);
1310 __page_frag_cache_drain(page, queue->pf_cache.pagecnt_bias);
1311 queue->pf_cache.va = NULL;
1312 }
3f2304f8
SG		 1313 sock_release(queue->sock);
1314 kfree(queue->pdu);
d48f92cd 1315 mutex_destroy(&queue->send_mutex);
9ebbfe49 1316 mutex_destroy(&queue->queue_lock);
3f2304f8
SG		 1317}
1318
1319static int nvme_tcp_init_connection(struct nvme_tcp_queue *queue)
1320{
1321 struct nvme_tcp_icreq_pdu *icreq;
1322 struct nvme_tcp_icresp_pdu *icresp;
1323 struct msghdr msg = {};
1324 struct kvec iov;
1325 bool ctrl_hdgst, ctrl_ddgst;
c2700d28 1326 u32 maxh2cdata;
3f2304f8
SG		 1327 int ret;
1328
1329 icreq = kzalloc(sizeof(*icreq), GFP_KERNEL);
1330 if (!icreq)
1331 return -ENOMEM;
1332
1333 icresp = kzalloc(sizeof(*icresp), GFP_KERNEL);
1334 if (!icresp) {
1335 ret = -ENOMEM;
1336 goto free_icreq;
1337 }
1338
1339 icreq->hdr.type = nvme_tcp_icreq;
1340 icreq->hdr.hlen = sizeof(*icreq);
1341 icreq->hdr.pdo = 0;
1342 icreq->hdr.plen = cpu_to_le32(icreq->hdr.hlen);
1343 icreq->pfv = cpu_to_le16(NVME_TCP_PFV_1_0);
1344 icreq->maxr2t = 0; /* single inflight r2t supported */
1345 icreq->hpda = 0; /* no alignment constraint */
1346 if (queue->hdr_digest)
1347 icreq->digest |= NVME_TCP_HDR_DIGEST_ENABLE;
1348 if (queue->data_digest)
1349 icreq->digest |= NVME_TCP_DATA_DIGEST_ENABLE;
1350
1351 iov.iov_base = icreq;
1352 iov.iov_len = sizeof(*icreq);
1353 ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
1354 if (ret < 0)
1355 goto free_icresp;
1356
1357 memset(&msg, 0, sizeof(msg));
1358 iov.iov_base = icresp;
1359 iov.iov_len = sizeof(*icresp);
1360 ret = kernel_recvmsg(queue->sock, &msg, &iov, 1,
1361 iov.iov_len, msg.msg_flags);
1362 if (ret < 0)
1363 goto free_icresp;
1364
1365 ret = -EINVAL;
1366 if (icresp->hdr.type != nvme_tcp_icresp) {
1367 pr_err("queue %d: bad type returned %d\n",
1368 nvme_tcp_queue_id(queue), icresp->hdr.type);
1369 goto free_icresp;
1370 }
1371
1372 if (le32_to_cpu(icresp->hdr.plen) != sizeof(*icresp)) {
1373 pr_err("queue %d: bad pdu length returned %d\n",
1374 nvme_tcp_queue_id(queue), icresp->hdr.plen);
1375 goto free_icresp;
1376 }
1377
1378 if (icresp->pfv != NVME_TCP_PFV_1_0) {
1379 pr_err("queue %d: bad pfv returned %d\n",
1380 nvme_tcp_queue_id(queue), icresp->pfv);
1381 goto free_icresp;
1382 }
1383
1384 ctrl_ddgst = !!(icresp->digest & NVME_TCP_DATA_DIGEST_ENABLE);
1385 if ((queue->data_digest && !ctrl_ddgst) ||
1386 (!queue->data_digest && ctrl_ddgst)) {
1387 pr_err("queue %d: data digest mismatch host: %s ctrl: %s\n",
1388 nvme_tcp_queue_id(queue),
1389 queue->data_digest ? "enabled" : "disabled",
1390 ctrl_ddgst ? "enabled" : "disabled");
1391 goto free_icresp;
1392 }
1393
1394 ctrl_hdgst = !!(icresp->digest & NVME_TCP_HDR_DIGEST_ENABLE);
1395 if ((queue->hdr_digest && !ctrl_hdgst) ||
1396 (!queue->hdr_digest && ctrl_hdgst)) {
1397 pr_err("queue %d: header digest mismatch host: %s ctrl: %s\n",
1398 nvme_tcp_queue_id(queue),
1399 queue->hdr_digest ? "enabled" : "disabled",
1400 ctrl_hdgst ? "enabled" : "disabled");
1401 goto free_icresp;
1402 }
1403
1404 if (icresp->cpda != 0) {
1405 pr_err("queue %d: unsupported cpda returned %d\n",
1406 nvme_tcp_queue_id(queue), icresp->cpda);
1407 goto free_icresp;
1408 }
1409
c2700d28
VP		 1410 maxh2cdata = le32_to_cpu(icresp->maxdata);
1411 if ((maxh2cdata % 4) || (maxh2cdata < NVME_TCP_MIN_MAXH2CDATA)) {
1412 pr_err("queue %d: invalid maxh2cdata returned %u\n",
1413 nvme_tcp_queue_id(queue), maxh2cdata);
1414 goto free_icresp;
1415 }
1416 queue->maxh2cdata = maxh2cdata;
1417
3f2304f8
SG		 1418 ret = 0;
1419free_icresp:
1420 kfree(icresp);
1421free_icreq:
1422 kfree(icreq);
1423 return ret;
1424}
1425
40510a63
SG		 1426static bool nvme_tcp_admin_queue(struct nvme_tcp_queue *queue)
1427{
1428 return nvme_tcp_queue_id(queue) == 0;
1429}
1430
1431static bool nvme_tcp_default_queue(struct nvme_tcp_queue *queue)
1432{
1433 struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1434 int qid = nvme_tcp_queue_id(queue);
1435
1436 return !nvme_tcp_admin_queue(queue) &&
1437 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT];
1438}
1439
1440static bool nvme_tcp_read_queue(struct nvme_tcp_queue *queue)
1441{
1442 struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1443 int qid = nvme_tcp_queue_id(queue);
1444
1445 return !nvme_tcp_admin_queue(queue) &&
1446 !nvme_tcp_default_queue(queue) &&
1447 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT] +
1448 ctrl->io_queues[HCTX_TYPE_READ];
1449}
1450
1451static bool nvme_tcp_poll_queue(struct nvme_tcp_queue *queue)
1452{
1453 struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1454 int qid = nvme_tcp_queue_id(queue);
1455
1456 return !nvme_tcp_admin_queue(queue) &&
1457 !nvme_tcp_default_queue(queue) &&
1458 !nvme_tcp_read_queue(queue) &&
1459 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT] +
1460 ctrl->io_queues[HCTX_TYPE_READ] +
1461 ctrl->io_queues[HCTX_TYPE_POLL];
1462}
1463
1464static void nvme_tcp_set_queue_io_cpu(struct nvme_tcp_queue *queue)
1465{
1466 struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1467 int qid = nvme_tcp_queue_id(queue);
1468 int n = 0;
1469
1470 if (nvme_tcp_default_queue(queue))
1471 n = qid - 1;
1472 else if (nvme_tcp_read_queue(queue))
1473 n = qid - ctrl->io_queues[HCTX_TYPE_DEFAULT] - 1;
1474 else if (nvme_tcp_poll_queue(queue))
1475 n = qid - ctrl->io_queues[HCTX_TYPE_DEFAULT] -
1476 ctrl->io_queues[HCTX_TYPE_READ] - 1;
1477 queue->io_cpu = cpumask_next_wrap(n - 1, cpu_online_mask, -1, false);
1478}
1479
3f2304f8
SG		 1480static int nvme_tcp_alloc_queue(struct nvme_ctrl *nctrl,
1481 int qid, size_t queue_size)
1482{
1483 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1484 struct nvme_tcp_queue *queue = &ctrl->queues[qid];
6ebf71ba 1485 int ret, rcv_pdu_size;
3f2304f8 1486
9ebbfe49 1487 mutex_init(&queue->queue_lock);
3f2304f8 1488 queue->ctrl = ctrl;
15ec928a 1489 init_llist_head(&queue->req_list);
3f2304f8 1490 INIT_LIST_HEAD(&queue->send_list);
db5ad6b7 1491 mutex_init(&queue->send_mutex);
3f2304f8
SG		 1492 INIT_WORK(&queue->io_work, nvme_tcp_io_work);
1493 queue->queue_size = queue_size;
1494
1495 if (qid > 0)
9924b030 1496 queue->cmnd_capsule_len = nctrl->ioccsz * 16;
3f2304f8
SG		 1497 else
1498 queue->cmnd_capsule_len = sizeof(struct nvme_command) +
1499 NVME_TCP_ADMIN_CCSZ;
1500
1501 ret = sock_create(ctrl->addr.ss_family, SOCK_STREAM,
1502 IPPROTO_TCP, &queue->sock);
1503 if (ret) {
9924b030 1504 dev_err(nctrl->device,
3f2304f8 1505 "failed to create socket: %d\n", ret);
9ebbfe49 1506 goto err_destroy_mutex;
3f2304f8
SG		 1507 }
1508
841aee4d
CL		 1509 nvme_tcp_reclassify_socket(queue->sock);
1510
3f2304f8 1511 /* Single syn retry */
557eadfc 1512 tcp_sock_set_syncnt(queue->sock->sk, 1);
3f2304f8
SG		 1513
1514 /* Set TCP no delay */
12abc5ee 1515 tcp_sock_set_nodelay(queue->sock->sk);
3f2304f8
SG		 1516
1517 /*
1518 * Cleanup whatever is sitting in the TCP transmit queue on socket
1519 * close. This is done to prevent stale data from being sent should
1520 * the network connection be restored before TCP times out.
1521 */
c433594c 1522 sock_no_linger(queue->sock->sk);
3f2304f8 1523
6e434967
CH		 1524 if (so_priority > 0)
1525 sock_set_priority(queue->sock->sk, so_priority);
9912ade3 1526
bb13985d 1527 /* Set socket type of service */
6ebf71ba
CH		 1528 if (nctrl->opts->tos >= 0)
1529 ip_sock_set_tos(queue->sock->sk, nctrl->opts->tos);
bb13985d 1530
adc99fd3
SG		 1531 /* Set 10 seconds timeout for icresp recvmsg */
1532 queue->sock->sk->sk_rcvtimeo = 10 * HZ;
1533
3f2304f8 1534 queue->sock->sk->sk_allocation = GFP_ATOMIC;
40510a63 1535 nvme_tcp_set_queue_io_cpu(queue);
3f2304f8
SG		 1536 queue->request = NULL;
1537 queue->data_remaining = 0;
1538 queue->ddgst_remaining = 0;
1539 queue->pdu_remaining = 0;
1540 queue->pdu_offset = 0;
1541 sk_set_memalloc(queue->sock->sk);
1542
9924b030 1543 if (nctrl->opts->mask & NVMF_OPT_HOST_TRADDR) {
3f2304f8
SG		 1544 ret = kernel_bind(queue->sock, (struct sockaddr *)&ctrl->src_addr,
1545 sizeof(ctrl->src_addr));
1546 if (ret) {
9924b030 1547 dev_err(nctrl->device,
3f2304f8
SG		 1548 "failed to bind queue %d socket %d\n",
1549 qid, ret);
1550 goto err_sock;
1551 }
1552 }
1553
3ede8f72
MB		 1554 if (nctrl->opts->mask & NVMF_OPT_HOST_IFACE) {
1555 char *iface = nctrl->opts->host_iface;
1556 sockptr_t optval = KERNEL_SOCKPTR(iface);
1557
1558 ret = sock_setsockopt(queue->sock, SOL_SOCKET, SO_BINDTODEVICE,
1559 optval, strlen(iface));
1560 if (ret) {
1561 dev_err(nctrl->device,
1562 "failed to bind to interface %s queue %d err %d\n",
1563 iface, qid, ret);
1564 goto err_sock;
1565 }
1566 }
1567
3f2304f8
SG		 1568 queue->hdr_digest = nctrl->opts->hdr_digest;
1569 queue->data_digest = nctrl->opts->data_digest;
1570 if (queue->hdr_digest || queue->data_digest) {
1571 ret = nvme_tcp_alloc_crypto(queue);
1572 if (ret) {
9924b030 1573 dev_err(nctrl->device,
3f2304f8
SG		 1574 "failed to allocate queue %d crypto\n", qid);
1575 goto err_sock;
1576 }
1577 }
1578
1579 rcv_pdu_size = sizeof(struct nvme_tcp_rsp_pdu) +
1580 nvme_tcp_hdgst_len(queue);
1581 queue->pdu = kmalloc(rcv_pdu_size, GFP_KERNEL);
1582 if (!queue->pdu) {
1583 ret = -ENOMEM;
1584 goto err_crypto;
1585 }
1586
9924b030 1587 dev_dbg(nctrl->device, "connecting queue %d\n",
3f2304f8
SG		 1588 nvme_tcp_queue_id(queue));
1589
1590 ret = kernel_connect(queue->sock, (struct sockaddr *)&ctrl->addr,
1591 sizeof(ctrl->addr), 0);
1592 if (ret) {
9924b030 1593 dev_err(nctrl->device,
3f2304f8
SG		 1594 "failed to connect socket: %d\n", ret);
1595 goto err_rcv_pdu;
1596 }
1597
1598 ret = nvme_tcp_init_connection(queue);
1599 if (ret)
1600 goto err_init_connect;
1601
1602 queue->rd_enabled = true;
1603 set_bit(NVME_TCP_Q_ALLOCATED, &queue->flags);
1604 nvme_tcp_init_recv_ctx(queue);
1605
1606 write_lock_bh(&queue->sock->sk->sk_callback_lock);
1607 queue->sock->sk->sk_user_data = queue;
1608 queue->state_change = queue->sock->sk->sk_state_change;
1609 queue->data_ready = queue->sock->sk->sk_data_ready;
1610 queue->write_space = queue->sock->sk->sk_write_space;
1611 queue->sock->sk->sk_data_ready = nvme_tcp_data_ready;
1612 queue->sock->sk->sk_state_change = nvme_tcp_state_change;
1613 queue->sock->sk->sk_write_space = nvme_tcp_write_space;
ac1c4e18 1614#ifdef CONFIG_NET_RX_BUSY_POLL
1a9460ce 1615 queue->sock->sk->sk_ll_usec = 1;
ac1c4e18 1616#endif
3f2304f8
SG		 1617 write_unlock_bh(&queue->sock->sk->sk_callback_lock);
1618
1619 return 0;
1620
1621err_init_connect:
1622 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1623err_rcv_pdu:
1624 kfree(queue->pdu);
1625err_crypto:
1626 if (queue->hdr_digest || queue->data_digest)
1627 nvme_tcp_free_crypto(queue);
1628err_sock:
1629 sock_release(queue->sock);
1630 queue->sock = NULL;
9ebbfe49 1631err_destroy_mutex:
d48f92cd 1632 mutex_destroy(&queue->send_mutex);
9ebbfe49 1633 mutex_destroy(&queue->queue_lock);
3f2304f8
SG		 1634 return ret;
1635}
1636
1637static void nvme_tcp_restore_sock_calls(struct nvme_tcp_queue *queue)
1638{
1639 struct socket *sock = queue->sock;
1640
1641 write_lock_bh(&sock->sk->sk_callback_lock);
1642 sock->sk->sk_user_data = NULL;
1643 sock->sk->sk_data_ready = queue->data_ready;
1644 sock->sk->sk_state_change = queue->state_change;
1645 sock->sk->sk_write_space = queue->write_space;
1646 write_unlock_bh(&sock->sk->sk_callback_lock);
1647}
1648
1649static void __nvme_tcp_stop_queue(struct nvme_tcp_queue *queue)
1650{
1651 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1652 nvme_tcp_restore_sock_calls(queue);
1653 cancel_work_sync(&queue->io_work);
1654}
1655
1656static void nvme_tcp_stop_queue(struct nvme_ctrl *nctrl, int qid)
1657{
1658 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1659 struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1660
9ebbfe49
CL		 1661 mutex_lock(&queue->queue_lock);
1662 if (test_and_clear_bit(NVME_TCP_Q_LIVE, &queue->flags))
1663 __nvme_tcp_stop_queue(queue);
1664 mutex_unlock(&queue->queue_lock);
3f2304f8
SG		 1665}
1666
1667static int nvme_tcp_start_queue(struct nvme_ctrl *nctrl, int idx)
1668{
1669 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1670 int ret;
1671
1672 if (idx)
be42a33b 1673 ret = nvmf_connect_io_queue(nctrl, idx);
3f2304f8
SG		 1674 else
1675 ret = nvmf_connect_admin_queue(nctrl);
1676
1677 if (!ret) {
1678 set_bit(NVME_TCP_Q_LIVE, &ctrl->queues[idx].flags);
1679 } else {
f34e2589
SG		 1680 if (test_bit(NVME_TCP_Q_ALLOCATED, &ctrl->queues[idx].flags))
1681 __nvme_tcp_stop_queue(&ctrl->queues[idx]);
3f2304f8
SG		 1682 dev_err(nctrl->device,
1683 "failed to connect queue: %d ret=%d\n", idx, ret);
1684 }
1685 return ret;
1686}
1687
1688static struct blk_mq_tag_set *nvme_tcp_alloc_tagset(struct nvme_ctrl *nctrl,
1689 bool admin)
1690{
1691 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1692 struct blk_mq_tag_set *set;
1693 int ret;
1694
1695 if (admin) {
1696 set = &ctrl->admin_tag_set;
1697 memset(set, 0, sizeof(*set));
1698 set->ops = &nvme_tcp_admin_mq_ops;
1699 set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
ed01fee2 1700 set->reserved_tags = NVMF_RESERVED_TAGS;
610c8235 1701 set->numa_node = nctrl->numa_node;
db5ad6b7 1702 set->flags = BLK_MQ_F_BLOCKING;
3f2304f8
SG		 1703 set->cmd_size = sizeof(struct nvme_tcp_request);
1704 set->driver_data = ctrl;
1705 set->nr_hw_queues = 1;
dc96f938 1706 set->timeout = NVME_ADMIN_TIMEOUT;
3f2304f8
SG		 1707 } else {
1708 set = &ctrl->tag_set;
1709 memset(set, 0, sizeof(*set));
1710 set->ops = &nvme_tcp_mq_ops;
1711 set->queue_depth = nctrl->sqsize + 1;
ed01fee2 1712 set->reserved_tags = NVMF_RESERVED_TAGS;
610c8235 1713 set->numa_node = nctrl->numa_node;
db5ad6b7 1714 set->flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_BLOCKING;
3f2304f8
SG		 1715 set->cmd_size = sizeof(struct nvme_tcp_request);
1716 set->driver_data = ctrl;
1717 set->nr_hw_queues = nctrl->queue_count - 1;
1718 set->timeout = NVME_IO_TIMEOUT;
1a9460ce 1719 set->nr_maps = nctrl->opts->nr_poll_queues ? HCTX_MAX_TYPES : 2;
3f2304f8
SG		 1720 }
1721
1722 ret = blk_mq_alloc_tag_set(set);
1723 if (ret)
1724 return ERR_PTR(ret);
1725
1726 return set;
1727}
1728
1729static void nvme_tcp_free_admin_queue(struct nvme_ctrl *ctrl)
1730{
1731 if (to_tcp_ctrl(ctrl)->async_req.pdu) {
ceb1e087 1732 cancel_work_sync(&ctrl->async_event_work);
3f2304f8
SG		 1733 nvme_tcp_free_async_req(to_tcp_ctrl(ctrl));
1734 to_tcp_ctrl(ctrl)->async_req.pdu = NULL;
1735 }
1736
1737 nvme_tcp_free_queue(ctrl, 0);
1738}
1739
1740static void nvme_tcp_free_io_queues(struct nvme_ctrl *ctrl)
1741{
1742 int i;
1743
1744 for (i = 1; i < ctrl->queue_count; i++)
1745 nvme_tcp_free_queue(ctrl, i);
1746}
1747
1748static void nvme_tcp_stop_io_queues(struct nvme_ctrl *ctrl)
1749{
1750 int i;
1751
1752 for (i = 1; i < ctrl->queue_count; i++)
1753 nvme_tcp_stop_queue(ctrl, i);
1754}
1755
1756static int nvme_tcp_start_io_queues(struct nvme_ctrl *ctrl)
1757{
462b8b2d 1758 int i, ret;
3f2304f8
SG		 1759
1760 for (i = 1; i < ctrl->queue_count; i++) {
1761 ret = nvme_tcp_start_queue(ctrl, i);
1762 if (ret)
1763 goto out_stop_queues;
1764 }
1765
1766 return 0;
1767
1768out_stop_queues:
1769 for (i--; i >= 1; i--)
1770 nvme_tcp_stop_queue(ctrl, i);
1771 return ret;
1772}
1773
1774static int nvme_tcp_alloc_admin_queue(struct nvme_ctrl *ctrl)
1775{
1776 int ret;
1777
1778 ret = nvme_tcp_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
1779 if (ret)
1780 return ret;
1781
1782 ret = nvme_tcp_alloc_async_req(to_tcp_ctrl(ctrl));
1783 if (ret)
1784 goto out_free_queue;
1785
1786 return 0;
1787
1788out_free_queue:
1789 nvme_tcp_free_queue(ctrl, 0);
1790 return ret;
1791}
1792
efb973b1 1793static int __nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
3f2304f8
SG		 1794{
1795 int i, ret;
1796
1797 for (i = 1; i < ctrl->queue_count; i++) {
a387935c 1798 ret = nvme_tcp_alloc_queue(ctrl, i, ctrl->sqsize + 1);
3f2304f8
SG		 1799 if (ret)
1800 goto out_free_queues;
1801 }
1802
1803 return 0;
1804
1805out_free_queues:
1806 for (i--; i >= 1; i--)
1807 nvme_tcp_free_queue(ctrl, i);
1808
1809 return ret;
1810}
1811
1812static unsigned int nvme_tcp_nr_io_queues(struct nvme_ctrl *ctrl)
1813{
873946f4
SG		 1814 unsigned int nr_io_queues;
1815
1816 nr_io_queues = min(ctrl->opts->nr_io_queues, num_online_cpus());
1817 nr_io_queues += min(ctrl->opts->nr_write_queues, num_online_cpus());
1a9460ce 1818 nr_io_queues += min(ctrl->opts->nr_poll_queues, num_online_cpus());
873946f4
SG		 1819
1820 return nr_io_queues;
3f2304f8
SG		 1821}
1822
64861993
SG		 1823static void nvme_tcp_set_io_queues(struct nvme_ctrl *nctrl,
1824 unsigned int nr_io_queues)
1825{
1826 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1827 struct nvmf_ctrl_options *opts = nctrl->opts;
1828
1829 if (opts->nr_write_queues && opts->nr_io_queues < nr_io_queues) {
1830 /*
1831 * separate read/write queues
1832 * hand out dedicated default queues only after we have
1833 * sufficient read queues.
1834 */
1835 ctrl->io_queues[HCTX_TYPE_READ] = opts->nr_io_queues;
1836 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_READ];
1837 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
1838 min(opts->nr_write_queues, nr_io_queues);
1839 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
1840 } else {
1841 /*
1842 * shared read/write queues
1843 * either no write queues were requested, or we don't have
1844 * sufficient queue count to have dedicated default queues.
1845 */
1846 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
1847 min(opts->nr_io_queues, nr_io_queues);
1848 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
1849 }
1a9460ce
SG		 1850
1851 if (opts->nr_poll_queues && nr_io_queues) {
1852 /* map dedicated poll queues only if we have queues left */
1853 ctrl->io_queues[HCTX_TYPE_POLL] =
1854 min(opts->nr_poll_queues, nr_io_queues);
1855 }
64861993
SG		 1856}
1857
efb973b1 1858static int nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
3f2304f8
SG		 1859{
1860 unsigned int nr_io_queues;
1861 int ret;
1862
1863 nr_io_queues = nvme_tcp_nr_io_queues(ctrl);
1864 ret = nvme_set_queue_count(ctrl, &nr_io_queues);
1865 if (ret)
1866 return ret;
1867
664227fd 1868 if (nr_io_queues == 0) {
72f57242
SG		 1869 dev_err(ctrl->device,
1870 "unable to set any I/O queues\n");
1871 return -ENOMEM;
1872 }
3f2304f8 1873
664227fd 1874 ctrl->queue_count = nr_io_queues + 1;
3f2304f8
SG		 1875 dev_info(ctrl->device,
1876 "creating %d I/O queues.\n", nr_io_queues);
1877
64861993
SG		 1878 nvme_tcp_set_io_queues(ctrl, nr_io_queues);
1879
efb973b1 1880 return __nvme_tcp_alloc_io_queues(ctrl);
3f2304f8
SG		 1881}
1882
1883static void nvme_tcp_destroy_io_queues(struct nvme_ctrl *ctrl, bool remove)
1884{
1885 nvme_tcp_stop_io_queues(ctrl);
1886 if (remove) {
6f8191fd 1887 blk_mq_destroy_queue(ctrl->connect_q);
3f2304f8
SG		 1888 blk_mq_free_tag_set(ctrl->tagset);
1889 }
1890 nvme_tcp_free_io_queues(ctrl);
1891}
1892
1893static int nvme_tcp_configure_io_queues(struct nvme_ctrl *ctrl, bool new)
1894{
1895 int ret;
1896
efb973b1 1897 ret = nvme_tcp_alloc_io_queues(ctrl);
3f2304f8
SG		 1898 if (ret)
1899 return ret;
1900
1901 if (new) {
1902 ctrl->tagset = nvme_tcp_alloc_tagset(ctrl, false);
1903 if (IS_ERR(ctrl->tagset)) {
1904 ret = PTR_ERR(ctrl->tagset);
1905 goto out_free_io_queues;
1906 }
1907
72e8b5cd
CK		 1908 ret = nvme_ctrl_init_connect_q(ctrl);
1909 if (ret)
e85037a2 1910 goto out_free_tag_set;
3f2304f8
SG		 1911 }
1912
1913 ret = nvme_tcp_start_io_queues(ctrl);
1914 if (ret)
1915 goto out_cleanup_connect_q;
1916
2875b0ae
SG		 1917 if (!new) {
1918 nvme_start_queues(ctrl);
e5c01f4f
SG		 1919 if (!nvme_wait_freeze_timeout(ctrl, NVME_IO_TIMEOUT)) {
1920 /*
1921 * If we timed out waiting for freeze we are likely to
1922 * be stuck. Fail the controller initialization just
1923 * to be safe.
1924 */
1925 ret = -ENODEV;
1926 goto out_wait_freeze_timed_out;
1927 }
2875b0ae
SG		 1928 blk_mq_update_nr_hw_queues(ctrl->tagset,
1929 ctrl->queue_count - 1);
1930 nvme_unfreeze(ctrl);
1931 }
1932
3f2304f8
SG		 1933 return 0;
1934
e5c01f4f
SG		 1935out_wait_freeze_timed_out:
1936 nvme_stop_queues(ctrl);
70a99574 1937 nvme_sync_io_queues(ctrl);
e5c01f4f 1938 nvme_tcp_stop_io_queues(ctrl);
3f2304f8 1939out_cleanup_connect_q:
70a99574 1940 nvme_cancel_tagset(ctrl);
e85037a2 1941 if (new)
6f8191fd 1942 blk_mq_destroy_queue(ctrl->connect_q);
3f2304f8
SG		 1943out_free_tag_set:
1944 if (new)
1945 blk_mq_free_tag_set(ctrl->tagset);
1946out_free_io_queues:
1947 nvme_tcp_free_io_queues(ctrl);
1948 return ret;
1949}
1950
1951static void nvme_tcp_destroy_admin_queue(struct nvme_ctrl *ctrl, bool remove)
1952{
1953 nvme_tcp_stop_queue(ctrl, 0);
1954 if (remove) {
6f8191fd
CH		 1955 blk_mq_destroy_queue(ctrl->admin_q);
1956 blk_mq_destroy_queue(ctrl->fabrics_q);
3f2304f8
SG		 1957 blk_mq_free_tag_set(ctrl->admin_tagset);
1958 }
1959 nvme_tcp_free_admin_queue(ctrl);
1960}
1961
1962static int nvme_tcp_configure_admin_queue(struct nvme_ctrl *ctrl, bool new)
1963{
1964 int error;
1965
1966 error = nvme_tcp_alloc_admin_queue(ctrl);
1967 if (error)
1968 return error;
1969
1970 if (new) {
1971 ctrl->admin_tagset = nvme_tcp_alloc_tagset(ctrl, true);
1972 if (IS_ERR(ctrl->admin_tagset)) {
1973 error = PTR_ERR(ctrl->admin_tagset);
1974 goto out_free_queue;
1975 }
1976
e7832cb4
SG		 1977 ctrl->fabrics_q = blk_mq_init_queue(ctrl->admin_tagset);
1978 if (IS_ERR(ctrl->fabrics_q)) {
1979 error = PTR_ERR(ctrl->fabrics_q);
1980 goto out_free_tagset;
1981 }
1982
3f2304f8
SG		 1983 ctrl->admin_q = blk_mq_init_queue(ctrl->admin_tagset);
1984 if (IS_ERR(ctrl->admin_q)) {
1985 error = PTR_ERR(ctrl->admin_q);
e7832cb4 1986 goto out_cleanup_fabrics_q;
3f2304f8
SG		 1987 }
1988 }
1989
1990 error = nvme_tcp_start_queue(ctrl, 0);
1991 if (error)
1992 goto out_cleanup_queue;
1993
c0f2f45b 1994 error = nvme_enable_ctrl(ctrl);
3f2304f8
SG		 1995 if (error)
1996 goto out_stop_queue;
1997
6ca1d902 1998 nvme_start_admin_queue(ctrl);
e7832cb4 1999
f21c4769 2000 error = nvme_init_ctrl_finish(ctrl);
3f2304f8 2001 if (error)
70a99574 2002 goto out_quiesce_queue;
3f2304f8
SG		 2003
2004 return 0;
2005
70a99574 2006out_quiesce_queue:
6ca1d902 2007 nvme_stop_admin_queue(ctrl);
70a99574 2008 blk_sync_queue(ctrl->admin_q);
3f2304f8
SG		 2009out_stop_queue:
2010 nvme_tcp_stop_queue(ctrl, 0);
70a99574 2011 nvme_cancel_admin_tagset(ctrl);
3f2304f8
SG		 2012out_cleanup_queue:
2013 if (new)
6f8191fd 2014 blk_mq_destroy_queue(ctrl->admin_q);
e7832cb4
SG		 2015out_cleanup_fabrics_q:
2016 if (new)
6f8191fd 2017 blk_mq_destroy_queue(ctrl->fabrics_q);
3f2304f8
SG		 2018out_free_tagset:
2019 if (new)
2020 blk_mq_free_tag_set(ctrl->admin_tagset);
2021out_free_queue:
2022 nvme_tcp_free_admin_queue(ctrl);
2023 return error;
2024}
2025
2026static void nvme_tcp_teardown_admin_queue(struct nvme_ctrl *ctrl,
2027 bool remove)
2028{
6ca1d902 2029 nvme_stop_admin_queue(ctrl);
d6f66210 2030 blk_sync_queue(ctrl->admin_q);
3f2304f8 2031 nvme_tcp_stop_queue(ctrl, 0);
563c8158 2032 nvme_cancel_admin_tagset(ctrl);
e7832cb4 2033 if (remove)
6ca1d902 2034 nvme_start_admin_queue(ctrl);
3f2304f8
SG		 2035 nvme_tcp_destroy_admin_queue(ctrl, remove);
2036}
2037
2038static void nvme_tcp_teardown_io_queues(struct nvme_ctrl *ctrl,
2039 bool remove)
2040{
2041 if (ctrl->queue_count <= 1)
d6f66210 2042 return;
6ca1d902 2043 nvme_stop_admin_queue(ctrl);
2875b0ae 2044 nvme_start_freeze(ctrl);
3f2304f8 2045 nvme_stop_queues(ctrl);
d6f66210 2046 nvme_sync_io_queues(ctrl);
3f2304f8 2047 nvme_tcp_stop_io_queues(ctrl);
563c8158 2048 nvme_cancel_tagset(ctrl);
3f2304f8
SG		 2049 if (remove)
2050 nvme_start_queues(ctrl);
2051 nvme_tcp_destroy_io_queues(ctrl, remove);
2052}
2053
2054static void nvme_tcp_reconnect_or_remove(struct nvme_ctrl *ctrl)
2055{
2056 /* If we are resetting/deleting then do nothing */
2057 if (ctrl->state != NVME_CTRL_CONNECTING) {
2058 WARN_ON_ONCE(ctrl->state == NVME_CTRL_NEW ||
2059 ctrl->state == NVME_CTRL_LIVE);
2060 return;
2061 }
2062
2063 if (nvmf_should_reconnect(ctrl)) {
2064 dev_info(ctrl->device, "Reconnecting in %d seconds...\n",
2065 ctrl->opts->reconnect_delay);
2066 queue_delayed_work(nvme_wq, &to_tcp_ctrl(ctrl)->connect_work,
2067 ctrl->opts->reconnect_delay * HZ);
2068 } else {
2069 dev_info(ctrl->device, "Removing controller...\n");
2070 nvme_delete_ctrl(ctrl);
2071 }
2072}
2073
2074static int nvme_tcp_setup_ctrl(struct nvme_ctrl *ctrl, bool new)
2075{
2076 struct nvmf_ctrl_options *opts = ctrl->opts;
312910f4 2077 int ret;
3f2304f8
SG		 2078
2079 ret = nvme_tcp_configure_admin_queue(ctrl, new);
2080 if (ret)
2081 return ret;
2082
2083 if (ctrl->icdoff) {
522af60c 2084 ret = -EOPNOTSUPP;
3f2304f8
SG		 2085 dev_err(ctrl->device, "icdoff is not supported!\n");
2086 goto destroy_admin;
2087 }
2088
3b54064f 2089 if (!nvme_ctrl_sgl_supported(ctrl)) {
522af60c 2090 ret = -EOPNOTSUPP;
73ffcefc
MG		 2091 dev_err(ctrl->device, "Mandatory sgls are not supported!\n");
2092 goto destroy_admin;
2093 }
2094
3f2304f8
SG		 2095 if (opts->queue_size > ctrl->sqsize + 1)
2096 dev_warn(ctrl->device,
2097 "queue_size %zu > ctrl sqsize %u, clamping down\n",
2098 opts->queue_size, ctrl->sqsize + 1);
2099
2100 if (ctrl->sqsize + 1 > ctrl->maxcmd) {
2101 dev_warn(ctrl->device,
2102 "sqsize %u > ctrl maxcmd %u, clamping down\n",
2103 ctrl->sqsize + 1, ctrl->maxcmd);
2104 ctrl->sqsize = ctrl->maxcmd - 1;
2105 }
2106
2107 if (ctrl->queue_count > 1) {
2108 ret = nvme_tcp_configure_io_queues(ctrl, new);
2109 if (ret)
2110 goto destroy_admin;
2111 }
2112
2113 if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_LIVE)) {
bea54ef5 2114 /*
ecca390e 2115 * state change failure is ok if we started ctrl delete,
bea54ef5
IR		 2116 * unless we're during creation of a new controller to
2117 * avoid races with teardown flow.
2118 */
ecca390e
SG		 2119 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING &&
2120 ctrl->state != NVME_CTRL_DELETING_NOIO);
bea54ef5 2121 WARN_ON_ONCE(new);
3f2304f8
SG		 2122 ret = -EINVAL;
2123 goto destroy_io;
2124 }
2125
2126 nvme_start_ctrl(ctrl);
2127 return 0;
2128
2129destroy_io:
70a99574
CL		 2130 if (ctrl->queue_count > 1) {
2131 nvme_stop_queues(ctrl);
2132 nvme_sync_io_queues(ctrl);
2133 nvme_tcp_stop_io_queues(ctrl);
2134 nvme_cancel_tagset(ctrl);
3f2304f8 2135 nvme_tcp_destroy_io_queues(ctrl, new);
70a99574 2136 }
3f2304f8 2137destroy_admin:
6ca1d902 2138 nvme_stop_admin_queue(ctrl);
70a99574 2139 blk_sync_queue(ctrl->admin_q);
3f2304f8 2140 nvme_tcp_stop_queue(ctrl, 0);
70a99574 2141 nvme_cancel_admin_tagset(ctrl);
3f2304f8
SG		 2142 nvme_tcp_destroy_admin_queue(ctrl, new);
2143 return ret;
2144}
2145
2146static void nvme_tcp_reconnect_ctrl_work(struct work_struct *work)
2147{
2148 struct nvme_tcp_ctrl *tcp_ctrl = container_of(to_delayed_work(work),
2149 struct nvme_tcp_ctrl, connect_work);
2150 struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
2151
2152 ++ctrl->nr_reconnects;
2153
2154 if (nvme_tcp_setup_ctrl(ctrl, false))
2155 goto requeue;
2156
56a77d26 2157 dev_info(ctrl->device, "Successfully reconnected (%d attempt)\n",
3f2304f8
SG		 2158 ctrl->nr_reconnects);
2159
2160 ctrl->nr_reconnects = 0;
2161
2162 return;
2163
2164requeue:
2165 dev_info(ctrl->device, "Failed reconnect attempt %d\n",
2166 ctrl->nr_reconnects);
2167 nvme_tcp_reconnect_or_remove(ctrl);
2168}
2169
2170static void nvme_tcp_error_recovery_work(struct work_struct *work)
2171{
2172 struct nvme_tcp_ctrl *tcp_ctrl = container_of(work,
2173 struct nvme_tcp_ctrl, err_work);
2174 struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
2175
f50fff73 2176 nvme_auth_stop(ctrl);
3f2304f8 2177 nvme_stop_keep_alive(ctrl);
ff9fc7eb 2178 flush_work(&ctrl->async_event_work);
3f2304f8
SG		 2179 nvme_tcp_teardown_io_queues(ctrl, false);
2180 /* unquiesce to fail fast pending requests */
2181 nvme_start_queues(ctrl);
2182 nvme_tcp_teardown_admin_queue(ctrl, false);
6ca1d902 2183 nvme_start_admin_queue(ctrl);
3f2304f8
SG		 2184
2185 if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
ecca390e
SG		 2186 /* state change failure is ok if we started ctrl delete */
2187 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING &&
2188 ctrl->state != NVME_CTRL_DELETING_NOIO);
3f2304f8
SG		 2189 return;
2190 }
2191
2192 nvme_tcp_reconnect_or_remove(ctrl);
2193}
2194
2195static void nvme_tcp_teardown_ctrl(struct nvme_ctrl *ctrl, bool shutdown)
2196{
2197 nvme_tcp_teardown_io_queues(ctrl, shutdown);
6ca1d902 2198 nvme_stop_admin_queue(ctrl);
3f2304f8
SG		 2199 if (shutdown)
2200 nvme_shutdown_ctrl(ctrl);
2201 else
b5b05048 2202 nvme_disable_ctrl(ctrl);
3f2304f8
SG		 2203 nvme_tcp_teardown_admin_queue(ctrl, shutdown);
2204}
2205
2206static void nvme_tcp_delete_ctrl(struct nvme_ctrl *ctrl)
2207{
2208 nvme_tcp_teardown_ctrl(ctrl, true);
2209}
2210
2211static void nvme_reset_ctrl_work(struct work_struct *work)
2212{
2213 struct nvme_ctrl *ctrl =
2214 container_of(work, struct nvme_ctrl, reset_work);
2215
2216 nvme_stop_ctrl(ctrl);
2217 nvme_tcp_teardown_ctrl(ctrl, false);
2218
2219 if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
ecca390e
SG		 2220 /* state change failure is ok if we started ctrl delete */
2221 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING &&
2222 ctrl->state != NVME_CTRL_DELETING_NOIO);
3f2304f8
SG		 2223 return;
2224 }
2225
2226 if (nvme_tcp_setup_ctrl(ctrl, false))
2227 goto out_fail;
2228
2229 return;
2230
2231out_fail:
2232 ++ctrl->nr_reconnects;
2233 nvme_tcp_reconnect_or_remove(ctrl);
2234}
2235
f7f70f4a
RL		 2236static void nvme_tcp_stop_ctrl(struct nvme_ctrl *ctrl)
2237{
2238 cancel_work_sync(&to_tcp_ctrl(ctrl)->err_work);
2239 cancel_delayed_work_sync(&to_tcp_ctrl(ctrl)->connect_work);
2240}
2241
3f2304f8
SG		 2242static void nvme_tcp_free_ctrl(struct nvme_ctrl *nctrl)
2243{
2244 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
2245
2246 if (list_empty(&ctrl->list))
2247 goto free_ctrl;
2248
2249 mutex_lock(&nvme_tcp_ctrl_mutex);
2250 list_del(&ctrl->list);
2251 mutex_unlock(&nvme_tcp_ctrl_mutex);
2252
2253 nvmf_free_options(nctrl->opts);
2254free_ctrl:
2255 kfree(ctrl->queues);
2256 kfree(ctrl);
2257}
2258
2259static void nvme_tcp_set_sg_null(struct nvme_command *c)
2260{
2261 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2262
2263 sg->addr = 0;
2264 sg->length = 0;
2265 sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
2266 NVME_SGL_FMT_TRANSPORT_A;
2267}
2268
2269static void nvme_tcp_set_sg_inline(struct nvme_tcp_queue *queue,
2270 struct nvme_command *c, u32 data_len)
2271{
2272 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2273
2274 sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
2275 sg->length = cpu_to_le32(data_len);
2276 sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
2277}
2278
2279static void nvme_tcp_set_sg_host_data(struct nvme_command *c,
2280 u32 data_len)
2281{
2282 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2283
2284 sg->addr = 0;
2285 sg->length = cpu_to_le32(data_len);
2286 sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
2287 NVME_SGL_FMT_TRANSPORT_A;
2288}
2289
2290static void nvme_tcp_submit_async_event(struct nvme_ctrl *arg)
2291{
2292 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(arg);
2293 struct nvme_tcp_queue *queue = &ctrl->queues[0];
2294 struct nvme_tcp_cmd_pdu *pdu = ctrl->async_req.pdu;
2295 struct nvme_command *cmd = &pdu->cmd;
2296 u8 hdgst = nvme_tcp_hdgst_len(queue);
2297
2298 memset(pdu, 0, sizeof(*pdu));
2299 pdu->hdr.type = nvme_tcp_cmd;
2300 if (queue->hdr_digest)
2301 pdu->hdr.flags |= NVME_TCP_F_HDGST;
2302 pdu->hdr.hlen = sizeof(*pdu);
2303 pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst);
2304
2305 cmd->common.opcode = nvme_admin_async_event;
2306 cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
2307 cmd->common.flags |= NVME_CMD_SGL_METABUF;
2308 nvme_tcp_set_sg_null(cmd);
2309
2310 ctrl->async_req.state = NVME_TCP_SEND_CMD_PDU;
2311 ctrl->async_req.offset = 0;
2312 ctrl->async_req.curr_bio = NULL;
2313 ctrl->async_req.data_len = 0;
2314
86f0348a 2315 nvme_tcp_queue_request(&ctrl->async_req, true, true);
3f2304f8
SG		 2316}
2317
236187c4
SG		 2318static void nvme_tcp_complete_timed_out(struct request *rq)
2319{
2320 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2321 struct nvme_ctrl *ctrl = &req->queue->ctrl->ctrl;
2322
236187c4 2323 nvme_tcp_stop_queue(ctrl, nvme_tcp_queue_id(req->queue));
93ba75c9 2324 nvmf_complete_timed_out_request(rq);
236187c4
SG		 2325}
2326
9bdb4833 2327static enum blk_eh_timer_return nvme_tcp_timeout(struct request *rq)
3f2304f8
SG		 2328{
2329 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
236187c4 2330 struct nvme_ctrl *ctrl = &req->queue->ctrl->ctrl;
3f2304f8
SG		 2331 struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2332
236187c4 2333 dev_warn(ctrl->device,
3f2304f8 2334 "queue %d: timeout request %#x type %d\n",
39d57757 2335 nvme_tcp_queue_id(req->queue), rq->tag, pdu->hdr.type);
3f2304f8 2336
236187c4 2337 if (ctrl->state != NVME_CTRL_LIVE) {
39d57757 2338 /*
236187c4
SG		 2339 * If we are resetting, connecting or deleting we should
2340 * complete immediately because we may block controller
2341 * teardown or setup sequence
2342 * - ctrl disable/shutdown fabrics requests
2343 * - connect requests
2344 * - initialization admin requests
2345 * - I/O requests that entered after unquiescing and
2346 * the controller stopped responding
2347 *
2348 * All other requests should be cancelled by the error
2349 * recovery work, so it's fine that we fail it here.
39d57757 2350 */
236187c4 2351 nvme_tcp_complete_timed_out(rq);
3f2304f8
SG		 2352 return BLK_EH_DONE;
2353 }
2354
236187c4
SG		 2355 /*
2356 * LIVE state should trigger the normal error recovery which will
2357 * handle completing this request.
2358 */
2359 nvme_tcp_error_recovery(ctrl);
3f2304f8
SG		 2360 return BLK_EH_RESET_TIMER;
2361}
2362
2363static blk_status_t nvme_tcp_map_data(struct nvme_tcp_queue *queue,
2364 struct request *rq)
2365{
2366 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2367 struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2368 struct nvme_command *c = &pdu->cmd;
2369
2370 c->common.flags |= NVME_CMD_SGL_METABUF;
2371
25e5cb78
SG		 2372 if (!blk_rq_nr_phys_segments(rq))
2373 nvme_tcp_set_sg_null(c);
2374 else if (rq_data_dir(rq) == WRITE &&
3f2304f8
SG		 2375 req->data_len <= nvme_tcp_inline_data_size(queue))
2376 nvme_tcp_set_sg_inline(queue, c, req->data_len);
2377 else
2378 nvme_tcp_set_sg_host_data(c, req->data_len);
2379
2380 return 0;
2381}
2382
2383static blk_status_t nvme_tcp_setup_cmd_pdu(struct nvme_ns *ns,
2384 struct request *rq)
2385{
2386 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2387 struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2388 struct nvme_tcp_queue *queue = req->queue;
2389 u8 hdgst = nvme_tcp_hdgst_len(queue), ddgst = 0;
2390 blk_status_t ret;
2391
f4b9e6c9 2392 ret = nvme_setup_cmd(ns, rq);
3f2304f8
SG		 2393 if (ret)
2394 return ret;
2395
2396 req->state = NVME_TCP_SEND_CMD_PDU;
1ba2e507 2397 req->status = cpu_to_le16(NVME_SC_SUCCESS);
3f2304f8
SG		 2398 req->offset = 0;
2399 req->data_sent = 0;
2400 req->pdu_len = 0;
2401 req->pdu_sent = 0;
c2700d28 2402 req->h2cdata_left = 0;
25e5cb78
SG		 2403 req->data_len = blk_rq_nr_phys_segments(rq) ?
2404 blk_rq_payload_bytes(rq) : 0;
3f2304f8 2405 req->curr_bio = rq->bio;
e11e5116 2406 if (req->curr_bio && req->data_len)
cb9b870f 2407 nvme_tcp_init_iter(req, rq_data_dir(rq));
3f2304f8
SG		 2408
2409 if (rq_data_dir(rq) == WRITE &&
2410 req->data_len <= nvme_tcp_inline_data_size(queue))
2411 req->pdu_len = req->data_len;
3f2304f8
SG		 2412
2413 pdu->hdr.type = nvme_tcp_cmd;
2414 pdu->hdr.flags = 0;
2415 if (queue->hdr_digest)
2416 pdu->hdr.flags |= NVME_TCP_F_HDGST;
2417 if (queue->data_digest && req->pdu_len) {
2418 pdu->hdr.flags |= NVME_TCP_F_DDGST;
2419 ddgst = nvme_tcp_ddgst_len(queue);
2420 }
2421 pdu->hdr.hlen = sizeof(*pdu);
2422 pdu->hdr.pdo = req->pdu_len ? pdu->hdr.hlen + hdgst : 0;
2423 pdu->hdr.plen =
2424 cpu_to_le32(pdu->hdr.hlen + hdgst + req->pdu_len + ddgst);
2425
2426 ret = nvme_tcp_map_data(queue, rq);
2427 if (unlikely(ret)) {
28a4cac4 2428 nvme_cleanup_cmd(rq);
3f2304f8
SG		 2429 dev_err(queue->ctrl->ctrl.device,
2430 "Failed to map data (%d)\n", ret);
2431 return ret;
2432 }
2433
2434 return 0;
2435}
2436
86f0348a
SG		 2437static void nvme_tcp_commit_rqs(struct blk_mq_hw_ctx *hctx)
2438{
2439 struct nvme_tcp_queue *queue = hctx->driver_data;
2440
2441 if (!llist_empty(&queue->req_list))
2442 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
2443}
2444
3f2304f8
SG		 2445static blk_status_t nvme_tcp_queue_rq(struct blk_mq_hw_ctx *hctx,
2446 const struct blk_mq_queue_data *bd)
2447{
2448 struct nvme_ns *ns = hctx->queue->queuedata;
2449 struct nvme_tcp_queue *queue = hctx->driver_data;
2450 struct request *rq = bd->rq;
2451 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2452 bool queue_ready = test_bit(NVME_TCP_Q_LIVE, &queue->flags);
2453 blk_status_t ret;
2454
a9715744
TC		 2455 if (!nvme_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
2456 return nvme_fail_nonready_command(&queue->ctrl->ctrl, rq);
3f2304f8
SG		 2457
2458 ret = nvme_tcp_setup_cmd_pdu(ns, rq);
2459 if (unlikely(ret))
2460 return ret;
2461
2462 blk_mq_start_request(rq);
2463
86f0348a 2464 nvme_tcp_queue_request(req, true, bd->last);
3f2304f8
SG		 2465
2466 return BLK_STS_OK;
2467}
2468
873946f4
SG		 2469static int nvme_tcp_map_queues(struct blk_mq_tag_set *set)
2470{
2471 struct nvme_tcp_ctrl *ctrl = set->driver_data;
64861993 2472 struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
873946f4 2473
64861993 2474 if (opts->nr_write_queues && ctrl->io_queues[HCTX_TYPE_READ]) {
873946f4
SG		 2475 /* separate read/write queues */
2476 set->map[HCTX_TYPE_DEFAULT].nr_queues =
64861993
SG		 2477 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2478 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2479 set->map[HCTX_TYPE_READ].nr_queues =
2480 ctrl->io_queues[HCTX_TYPE_READ];
873946f4 2481 set->map[HCTX_TYPE_READ].queue_offset =
64861993 2482 ctrl->io_queues[HCTX_TYPE_DEFAULT];
873946f4 2483 } else {
64861993 2484 /* shared read/write queues */
873946f4 2485 set->map[HCTX_TYPE_DEFAULT].nr_queues =
64861993
SG		 2486 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2487 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2488 set->map[HCTX_TYPE_READ].nr_queues =
2489 ctrl->io_queues[HCTX_TYPE_DEFAULT];
873946f4
SG		 2490 set->map[HCTX_TYPE_READ].queue_offset = 0;
2491 }
2492 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
2493 blk_mq_map_queues(&set->map[HCTX_TYPE_READ]);
64861993 2494
1a9460ce
SG		 2495 if (opts->nr_poll_queues && ctrl->io_queues[HCTX_TYPE_POLL]) {
2496 /* map dedicated poll queues only if we have queues left */
2497 set->map[HCTX_TYPE_POLL].nr_queues =
2498 ctrl->io_queues[HCTX_TYPE_POLL];
2499 set->map[HCTX_TYPE_POLL].queue_offset =
2500 ctrl->io_queues[HCTX_TYPE_DEFAULT] +
2501 ctrl->io_queues[HCTX_TYPE_READ];
2502 blk_mq_map_queues(&set->map[HCTX_TYPE_POLL]);
2503 }
2504
64861993 2505 dev_info(ctrl->ctrl.device,
1a9460ce 2506 "mapped %d/%d/%d default/read/poll queues.\n",
64861993 2507 ctrl->io_queues[HCTX_TYPE_DEFAULT],
1a9460ce
SG		 2508 ctrl->io_queues[HCTX_TYPE_READ],
2509 ctrl->io_queues[HCTX_TYPE_POLL]);
64861993 2510
873946f4
SG		 2511 return 0;
2512}
2513
5a72e899 2514static int nvme_tcp_poll(struct blk_mq_hw_ctx *hctx, struct io_comp_batch *iob)
1a9460ce
SG		 2515{
2516 struct nvme_tcp_queue *queue = hctx->driver_data;
2517 struct sock *sk = queue->sock->sk;
2518
f86e5bf8
SG		 2519 if (!test_bit(NVME_TCP_Q_LIVE, &queue->flags))
2520 return 0;
2521
72e5d757 2522 set_bit(NVME_TCP_Q_POLLING, &queue->flags);
3f926af3 2523 if (sk_can_busy_loop(sk) && skb_queue_empty_lockless(&sk->sk_receive_queue))
1a9460ce
SG		 2524 sk_busy_loop(sk, true);
2525 nvme_tcp_try_recv(queue);
72e5d757 2526 clear_bit(NVME_TCP_Q_POLLING, &queue->flags);
1a9460ce
SG		 2527 return queue->nr_cqe;
2528}
2529
6acbd961 2530static const struct blk_mq_ops nvme_tcp_mq_ops = {
3f2304f8 2531 .queue_rq = nvme_tcp_queue_rq,
86f0348a 2532 .commit_rqs = nvme_tcp_commit_rqs,
3f2304f8
SG		 2533 .complete = nvme_complete_rq,
2534 .init_request = nvme_tcp_init_request,
2535 .exit_request = nvme_tcp_exit_request,
2536 .init_hctx = nvme_tcp_init_hctx,
2537 .timeout = nvme_tcp_timeout,
873946f4 2538 .map_queues = nvme_tcp_map_queues,
1a9460ce 2539 .poll = nvme_tcp_poll,
3f2304f8
SG		 2540};
2541
6acbd961 2542static const struct blk_mq_ops nvme_tcp_admin_mq_ops = {
3f2304f8
SG		 2543 .queue_rq = nvme_tcp_queue_rq,
2544 .complete = nvme_complete_rq,
2545 .init_request = nvme_tcp_init_request,
2546 .exit_request = nvme_tcp_exit_request,
2547 .init_hctx = nvme_tcp_init_admin_hctx,
2548 .timeout = nvme_tcp_timeout,
2549};
2550
2551static const struct nvme_ctrl_ops nvme_tcp_ctrl_ops = {
2552 .name = "tcp",
2553 .module = THIS_MODULE,
2554 .flags = NVME_F_FABRICS,
2555 .reg_read32 = nvmf_reg_read32,
2556 .reg_read64 = nvmf_reg_read64,
2557 .reg_write32 = nvmf_reg_write32,
2558 .free_ctrl = nvme_tcp_free_ctrl,
2559 .submit_async_event = nvme_tcp_submit_async_event,
2560 .delete_ctrl = nvme_tcp_delete_ctrl,
2561 .get_address = nvmf_get_address,
f7f70f4a 2562 .stop_ctrl = nvme_tcp_stop_ctrl,
3f2304f8
SG		 2563};
2564
2565static bool
2566nvme_tcp_existing_controller(struct nvmf_ctrl_options *opts)
2567{
2568 struct nvme_tcp_ctrl *ctrl;
2569 bool found = false;
2570
2571 mutex_lock(&nvme_tcp_ctrl_mutex);
2572 list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list) {
2573 found = nvmf_ip_options_match(&ctrl->ctrl, opts);
2574 if (found)
2575 break;
2576 }
2577 mutex_unlock(&nvme_tcp_ctrl_mutex);
2578
2579 return found;
2580}
2581
2582static struct nvme_ctrl *nvme_tcp_create_ctrl(struct device *dev,
2583 struct nvmf_ctrl_options *opts)
2584{
2585 struct nvme_tcp_ctrl *ctrl;
2586 int ret;
2587
2588 ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
2589 if (!ctrl)
2590 return ERR_PTR(-ENOMEM);
2591
2592 INIT_LIST_HEAD(&ctrl->list);
2593 ctrl->ctrl.opts = opts;
1a9460ce
SG		 2594 ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues +
2595 opts->nr_poll_queues + 1;
3f2304f8
SG		 2596 ctrl->ctrl.sqsize = opts->queue_size - 1;
2597 ctrl->ctrl.kato = opts->kato;
2598
2599 INIT_DELAYED_WORK(&ctrl->connect_work,
2600 nvme_tcp_reconnect_ctrl_work);
2601 INIT_WORK(&ctrl->err_work, nvme_tcp_error_recovery_work);
2602 INIT_WORK(&ctrl->ctrl.reset_work, nvme_reset_ctrl_work);
2603
2604 if (!(opts->mask & NVMF_OPT_TRSVCID)) {
2605 opts->trsvcid =
2606 kstrdup(__stringify(NVME_TCP_DISC_PORT), GFP_KERNEL);
2607 if (!opts->trsvcid) {
2608 ret = -ENOMEM;
2609 goto out_free_ctrl;
2610 }
2611 opts->mask |= NVMF_OPT_TRSVCID;
2612 }
2613
2614 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2615 opts->traddr, opts->trsvcid, &ctrl->addr);
2616 if (ret) {
2617 pr_err("malformed address passed: %s:%s\n",
2618 opts->traddr, opts->trsvcid);
2619 goto out_free_ctrl;
2620 }
2621
2622 if (opts->mask & NVMF_OPT_HOST_TRADDR) {
2623 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2624 opts->host_traddr, NULL, &ctrl->src_addr);
2625 if (ret) {
2626 pr_err("malformed src address passed: %s\n",
2627 opts->host_traddr);
2628 goto out_free_ctrl;
2629 }
2630 }
2631
3ede8f72 2632 if (opts->mask & NVMF_OPT_HOST_IFACE) {
8b43ced6 2633 if (!__dev_get_by_name(&init_net, opts->host_iface)) {
3ede8f72
MB		 2634 pr_err("invalid interface passed: %s\n",
2635 opts->host_iface);
2636 ret = -ENODEV;
2637 goto out_free_ctrl;
2638 }
2639 }
2640
3f2304f8
SG		 2641 if (!opts->duplicate_connect && nvme_tcp_existing_controller(opts)) {
2642 ret = -EALREADY;
2643 goto out_free_ctrl;
2644 }
2645
873946f4 2646 ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
3f2304f8
SG		 2647 GFP_KERNEL);
2648 if (!ctrl->queues) {
2649 ret = -ENOMEM;
2650 goto out_free_ctrl;
2651 }
2652
2653 ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_tcp_ctrl_ops, 0);
2654 if (ret)
2655 goto out_kfree_queues;
2656
2657 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
2658 WARN_ON_ONCE(1);
2659 ret = -EINTR;
2660 goto out_uninit_ctrl;
2661 }
2662
2663 ret = nvme_tcp_setup_ctrl(&ctrl->ctrl, true);
2664 if (ret)
2665 goto out_uninit_ctrl;
2666
2667 dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISp\n",
e5ea42fa 2668 nvmf_ctrl_subsysnqn(&ctrl->ctrl), &ctrl->addr);
3f2304f8 2669
3f2304f8
SG		 2670 mutex_lock(&nvme_tcp_ctrl_mutex);
2671 list_add_tail(&ctrl->list, &nvme_tcp_ctrl_list);
2672 mutex_unlock(&nvme_tcp_ctrl_mutex);
2673
2674 return &ctrl->ctrl;
2675
2676out_uninit_ctrl:
2677 nvme_uninit_ctrl(&ctrl->ctrl);
2678 nvme_put_ctrl(&ctrl->ctrl);
2679 if (ret > 0)
2680 ret = -EIO;
2681 return ERR_PTR(ret);
2682out_kfree_queues:
2683 kfree(ctrl->queues);
2684out_free_ctrl:
2685 kfree(ctrl);
2686 return ERR_PTR(ret);
2687}
2688
2689static struct nvmf_transport_ops nvme_tcp_transport = {
2690 .name = "tcp",
2691 .module = THIS_MODULE,
2692 .required_opts = NVMF_OPT_TRADDR,
2693 .allowed_opts = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2694 NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
873946f4 2695 NVMF_OPT_HDR_DIGEST | NVMF_OPT_DATA_DIGEST |
bb13985d 2696 NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES |
3ede8f72 2697 NVMF_OPT_TOS | NVMF_OPT_HOST_IFACE,
3f2304f8
SG		 2698 .create_ctrl = nvme_tcp_create_ctrl,
2699};
2700
2701static int __init nvme_tcp_init_module(void)
2702{
2703 nvme_tcp_wq = alloc_workqueue("nvme_tcp_wq",
2704 WQ_MEM_RECLAIM | WQ_HIGHPRI, 0);
2705 if (!nvme_tcp_wq)
2706 return -ENOMEM;
2707
2708 nvmf_register_transport(&nvme_tcp_transport);
2709 return 0;
2710}
2711
2712static void __exit nvme_tcp_cleanup_module(void)
2713{
2714 struct nvme_tcp_ctrl *ctrl;
2715
2716 nvmf_unregister_transport(&nvme_tcp_transport);
2717
2718 mutex_lock(&nvme_tcp_ctrl_mutex);
2719 list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list)
2720 nvme_delete_ctrl(&ctrl->ctrl);
2721 mutex_unlock(&nvme_tcp_ctrl_mutex);
2722 flush_workqueue(nvme_delete_wq);
2723
2724 destroy_workqueue(nvme_tcp_wq);
2725}
2726
2727module_init(nvme_tcp_init_module);
2728module_exit(nvme_tcp_cleanup_module);
2729
2730MODULE_LICENSE("GPL v2");
Linux 6.x block layer and io_uring tree(s)