xref: /linux/drivers/nvme/host/tcp.c (revision fbf46565c67c626849c7ce2a326972d3008d2a91)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * NVMe over Fabrics TCP host.
4  * Copyright (c) 2018 Lightbits Labs. All rights reserved.
5  */
6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
7 #include <linux/module.h>
8 #include <linux/init.h>
9 #include <linux/slab.h>
10 #include <linux/err.h>
11 #include <linux/nvme-tcp.h>
12 #include <net/sock.h>
13 #include <net/tcp.h>
14 #include <linux/blk-mq.h>
15 #include <crypto/hash.h>
16 #include <net/busy_poll.h>
17 
18 #include "nvme.h"
19 #include "fabrics.h"
20 
21 struct nvme_tcp_queue;
22 
23 /* Define the socket priority to use for connections were it is desirable
24  * that the NIC consider performing optimized packet processing or filtering.
25  * A non-zero value being sufficient to indicate general consideration of any
26  * possible optimization.  Making it a module param allows for alternative
27  * values that may be unique for some NIC implementations.
28  */
29 static int so_priority;
30 module_param(so_priority, int, 0644);
31 MODULE_PARM_DESC(so_priority, "nvme tcp socket optimize priority");
32 
33 #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  */
40 static struct lock_class_key nvme_tcp_sk_key[2];
41 static struct lock_class_key nvme_tcp_slock_key[2];
42 
43 static 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
68 static void nvme_tcp_reclassify_socket(struct socket *sock) { }
69 #endif
70 
71 enum 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 
78 struct 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;
85 	u32			h2cdata_left;
86 	u32			h2cdata_offset;
87 	u16			ttag;
88 	__le16			status;
89 	struct list_head	entry;
90 	struct llist_node	lentry;
91 	__le32			ddgst;
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 
102 enum nvme_tcp_queue_flags {
103 	NVME_TCP_Q_ALLOCATED	= 0,
104 	NVME_TCP_Q_LIVE		= 1,
105 	NVME_TCP_Q_POLLING	= 2,
106 };
107 
108 enum nvme_tcp_recv_state {
109 	NVME_TCP_RECV_PDU = 0,
110 	NVME_TCP_RECV_DATA,
111 	NVME_TCP_RECV_DDGST,
112 };
113 
114 struct nvme_tcp_ctrl;
115 struct nvme_tcp_queue {
116 	struct socket		*sock;
117 	struct work_struct	io_work;
118 	int			io_cpu;
119 
120 	struct mutex		queue_lock;
121 	struct mutex		send_mutex;
122 	struct llist_head	req_list;
123 	struct list_head	send_list;
124 
125 	/* recv state */
126 	void			*pdu;
127 	int			pdu_remaining;
128 	int			pdu_offset;
129 	size_t			data_remaining;
130 	size_t			ddgst_remaining;
131 	unsigned int		nr_cqe;
132 
133 	/* send state */
134 	struct nvme_tcp_request *request;
135 
136 	u32			maxh2cdata;
137 	size_t			cmnd_capsule_len;
138 	struct nvme_tcp_ctrl	*ctrl;
139 	unsigned long		flags;
140 	bool			rd_enabled;
141 
142 	bool			hdr_digest;
143 	bool			data_digest;
144 	struct ahash_request	*rcv_hash;
145 	struct ahash_request	*snd_hash;
146 	__le32			exp_ddgst;
147 	__le32			recv_ddgst;
148 
149 	struct page_frag_cache	pf_cache;
150 
151 	void (*state_change)(struct sock *);
152 	void (*data_ready)(struct sock *);
153 	void (*write_space)(struct sock *);
154 };
155 
156 struct nvme_tcp_ctrl {
157 	/* read only in the hot path */
158 	struct nvme_tcp_queue	*queues;
159 	struct blk_mq_tag_set	tag_set;
160 
161 	/* other member variables */
162 	struct list_head	list;
163 	struct blk_mq_tag_set	admin_tag_set;
164 	struct sockaddr_storage addr;
165 	struct sockaddr_storage src_addr;
166 	struct nvme_ctrl	ctrl;
167 
168 	struct work_struct	err_work;
169 	struct delayed_work	connect_work;
170 	struct nvme_tcp_request async_req;
171 	u32			io_queues[HCTX_MAX_TYPES];
172 };
173 
174 static LIST_HEAD(nvme_tcp_ctrl_list);
175 static DEFINE_MUTEX(nvme_tcp_ctrl_mutex);
176 static struct workqueue_struct *nvme_tcp_wq;
177 static const struct blk_mq_ops nvme_tcp_mq_ops;
178 static const struct blk_mq_ops nvme_tcp_admin_mq_ops;
179 static int nvme_tcp_try_send(struct nvme_tcp_queue *queue);
180 
181 static inline struct nvme_tcp_ctrl *to_tcp_ctrl(struct nvme_ctrl *ctrl)
182 {
183 	return container_of(ctrl, struct nvme_tcp_ctrl, ctrl);
184 }
185 
186 static inline int nvme_tcp_queue_id(struct nvme_tcp_queue *queue)
187 {
188 	return queue - queue->ctrl->queues;
189 }
190 
191 static inline struct blk_mq_tags *nvme_tcp_tagset(struct nvme_tcp_queue *queue)
192 {
193 	u32 queue_idx = nvme_tcp_queue_id(queue);
194 
195 	if (queue_idx == 0)
196 		return queue->ctrl->admin_tag_set.tags[queue_idx];
197 	return queue->ctrl->tag_set.tags[queue_idx - 1];
198 }
199 
200 static inline u8 nvme_tcp_hdgst_len(struct nvme_tcp_queue *queue)
201 {
202 	return queue->hdr_digest ? NVME_TCP_DIGEST_LENGTH : 0;
203 }
204 
205 static inline u8 nvme_tcp_ddgst_len(struct nvme_tcp_queue *queue)
206 {
207 	return queue->data_digest ? NVME_TCP_DIGEST_LENGTH : 0;
208 }
209 
210 static inline size_t nvme_tcp_inline_data_size(struct nvme_tcp_request *req)
211 {
212 	if (nvme_is_fabrics(req->req.cmd))
213 		return NVME_TCP_ADMIN_CCSZ;
214 	return req->queue->cmnd_capsule_len - sizeof(struct nvme_command);
215 }
216 
217 static inline bool nvme_tcp_async_req(struct nvme_tcp_request *req)
218 {
219 	return req == &req->queue->ctrl->async_req;
220 }
221 
222 static inline bool nvme_tcp_has_inline_data(struct nvme_tcp_request *req)
223 {
224 	struct request *rq;
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);
230 
231 	return rq_data_dir(rq) == WRITE && req->data_len &&
232 		req->data_len <= nvme_tcp_inline_data_size(req);
233 }
234 
235 static inline struct page *nvme_tcp_req_cur_page(struct nvme_tcp_request *req)
236 {
237 	return req->iter.bvec->bv_page;
238 }
239 
240 static 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 
245 static inline size_t nvme_tcp_req_cur_length(struct nvme_tcp_request *req)
246 {
247 	return min_t(size_t, iov_iter_single_seg_count(&req->iter),
248 			req->pdu_len - req->pdu_sent);
249 }
250 
251 static 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 
257 static 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 
263 static 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;
269 	int nr_bvec;
270 	size_t offset;
271 
272 	if (rq->rq_flags & RQF_SPECIAL_PAYLOAD) {
273 		vec = &rq->special_vec;
274 		nr_bvec = 1;
275 		size = blk_rq_payload_bytes(rq);
276 		offset = 0;
277 	} else {
278 		struct bio *bio = req->curr_bio;
279 		struct bvec_iter bi;
280 		struct bio_vec bv;
281 
282 		vec = __bvec_iter_bvec(bio->bi_io_vec, bio->bi_iter);
283 		nr_bvec = 0;
284 		bio_for_each_bvec(bv, bio, bi) {
285 			nr_bvec++;
286 		}
287 		size = bio->bi_iter.bi_size;
288 		offset = bio->bi_iter.bi_bvec_done;
289 	}
290 
291 	iov_iter_bvec(&req->iter, dir, vec, nr_bvec, size);
292 	req->iter.iov_offset = offset;
293 }
294 
295 static 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, ITER_SOURCE);
305 	}
306 }
307 
308 static 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 
318 static 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);
322 }
323 
324 static inline void nvme_tcp_queue_request(struct nvme_tcp_request *req,
325 		bool sync, bool last)
326 {
327 	struct nvme_tcp_queue *queue = req->queue;
328 	bool empty;
329 
330 	empty = llist_add(&req->lentry, &queue->req_list) &&
331 		list_empty(&queue->send_list) && !queue->request;
332 
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 	 */
338 	if (queue->io_cpu == raw_smp_processor_id() &&
339 	    sync && empty && mutex_trylock(&queue->send_mutex)) {
340 		nvme_tcp_send_all(queue);
341 		mutex_unlock(&queue->send_mutex);
342 	}
343 
344 	if (last && nvme_tcp_queue_more(queue))
345 		queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
346 }
347 
348 static void nvme_tcp_process_req_list(struct nvme_tcp_queue *queue)
349 {
350 	struct nvme_tcp_request *req;
351 	struct llist_node *node;
352 
353 	for (node = llist_del_all(&queue->req_list); node; node = node->next) {
354 		req = llist_entry(node, struct nvme_tcp_request, lentry);
355 		list_add(&req->entry, &queue->send_list);
356 	}
357 }
358 
359 static inline struct nvme_tcp_request *
360 nvme_tcp_fetch_request(struct nvme_tcp_queue *queue)
361 {
362 	struct nvme_tcp_request *req;
363 
364 	req = list_first_entry_or_null(&queue->send_list,
365 			struct nvme_tcp_request, entry);
366 	if (!req) {
367 		nvme_tcp_process_req_list(queue);
368 		req = list_first_entry_or_null(&queue->send_list,
369 				struct nvme_tcp_request, entry);
370 		if (unlikely(!req))
371 			return NULL;
372 	}
373 
374 	list_del(&req->entry);
375 	return req;
376 }
377 
378 static inline void nvme_tcp_ddgst_final(struct ahash_request *hash,
379 		__le32 *dgst)
380 {
381 	ahash_request_set_crypt(hash, NULL, (u8 *)dgst, 0);
382 	crypto_ahash_final(hash);
383 }
384 
385 static inline void nvme_tcp_ddgst_update(struct ahash_request *hash,
386 		struct page *page, off_t off, size_t len)
387 {
388 	struct scatterlist sg;
389 
390 	sg_init_table(&sg, 1);
391 	sg_set_page(&sg, page, len, off);
392 	ahash_request_set_crypt(hash, &sg, NULL, len);
393 	crypto_ahash_update(hash);
394 }
395 
396 static inline void nvme_tcp_hdgst(struct ahash_request *hash,
397 		void *pdu, size_t len)
398 {
399 	struct scatterlist sg;
400 
401 	sg_init_one(&sg, pdu, len);
402 	ahash_request_set_crypt(hash, &sg, pdu + len, len);
403 	crypto_ahash_digest(hash);
404 }
405 
406 static int nvme_tcp_verify_hdgst(struct nvme_tcp_queue *queue,
407 		void *pdu, size_t pdu_len)
408 {
409 	struct nvme_tcp_hdr *hdr = pdu;
410 	__le32 recv_digest;
411 	__le32 exp_digest;
412 
413 	if (unlikely(!(hdr->flags & NVME_TCP_F_HDGST))) {
414 		dev_err(queue->ctrl->ctrl.device,
415 			"queue %d: header digest flag is cleared\n",
416 			nvme_tcp_queue_id(queue));
417 		return -EPROTO;
418 	}
419 
420 	recv_digest = *(__le32 *)(pdu + hdr->hlen);
421 	nvme_tcp_hdgst(queue->rcv_hash, pdu, pdu_len);
422 	exp_digest = *(__le32 *)(pdu + hdr->hlen);
423 	if (recv_digest != exp_digest) {
424 		dev_err(queue->ctrl->ctrl.device,
425 			"header digest error: recv %#x expected %#x\n",
426 			le32_to_cpu(recv_digest), le32_to_cpu(exp_digest));
427 		return -EIO;
428 	}
429 
430 	return 0;
431 }
432 
433 static int nvme_tcp_check_ddgst(struct nvme_tcp_queue *queue, void *pdu)
434 {
435 	struct nvme_tcp_hdr *hdr = pdu;
436 	u8 digest_len = nvme_tcp_hdgst_len(queue);
437 	u32 len;
438 
439 	len = le32_to_cpu(hdr->plen) - hdr->hlen -
440 		((hdr->flags & NVME_TCP_F_HDGST) ? digest_len : 0);
441 
442 	if (unlikely(len && !(hdr->flags & NVME_TCP_F_DDGST))) {
443 		dev_err(queue->ctrl->ctrl.device,
444 			"queue %d: data digest flag is cleared\n",
445 		nvme_tcp_queue_id(queue));
446 		return -EPROTO;
447 	}
448 	crypto_ahash_init(queue->rcv_hash);
449 
450 	return 0;
451 }
452 
453 static void nvme_tcp_exit_request(struct blk_mq_tag_set *set,
454 		struct request *rq, unsigned int hctx_idx)
455 {
456 	struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
457 
458 	page_frag_free(req->pdu);
459 }
460 
461 static int nvme_tcp_init_request(struct blk_mq_tag_set *set,
462 		struct request *rq, unsigned int hctx_idx,
463 		unsigned int numa_node)
464 {
465 	struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(set->driver_data);
466 	struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
467 	struct nvme_tcp_cmd_pdu *pdu;
468 	int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
469 	struct nvme_tcp_queue *queue = &ctrl->queues[queue_idx];
470 	u8 hdgst = nvme_tcp_hdgst_len(queue);
471 
472 	req->pdu = page_frag_alloc(&queue->pf_cache,
473 		sizeof(struct nvme_tcp_cmd_pdu) + hdgst,
474 		GFP_KERNEL | __GFP_ZERO);
475 	if (!req->pdu)
476 		return -ENOMEM;
477 
478 	pdu = req->pdu;
479 	req->queue = queue;
480 	nvme_req(rq)->ctrl = &ctrl->ctrl;
481 	nvme_req(rq)->cmd = &pdu->cmd;
482 
483 	return 0;
484 }
485 
486 static int nvme_tcp_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
487 		unsigned int hctx_idx)
488 {
489 	struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(data);
490 	struct nvme_tcp_queue *queue = &ctrl->queues[hctx_idx + 1];
491 
492 	hctx->driver_data = queue;
493 	return 0;
494 }
495 
496 static int nvme_tcp_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
497 		unsigned int hctx_idx)
498 {
499 	struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(data);
500 	struct nvme_tcp_queue *queue = &ctrl->queues[0];
501 
502 	hctx->driver_data = queue;
503 	return 0;
504 }
505 
506 static enum nvme_tcp_recv_state
507 nvme_tcp_recv_state(struct nvme_tcp_queue *queue)
508 {
509 	return  (queue->pdu_remaining) ? NVME_TCP_RECV_PDU :
510 		(queue->ddgst_remaining) ? NVME_TCP_RECV_DDGST :
511 		NVME_TCP_RECV_DATA;
512 }
513 
514 static void nvme_tcp_init_recv_ctx(struct nvme_tcp_queue *queue)
515 {
516 	queue->pdu_remaining = sizeof(struct nvme_tcp_rsp_pdu) +
517 				nvme_tcp_hdgst_len(queue);
518 	queue->pdu_offset = 0;
519 	queue->data_remaining = -1;
520 	queue->ddgst_remaining = 0;
521 }
522 
523 static void nvme_tcp_error_recovery(struct nvme_ctrl *ctrl)
524 {
525 	if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_RESETTING))
526 		return;
527 
528 	dev_warn(ctrl->device, "starting error recovery\n");
529 	queue_work(nvme_reset_wq, &to_tcp_ctrl(ctrl)->err_work);
530 }
531 
532 static int nvme_tcp_process_nvme_cqe(struct nvme_tcp_queue *queue,
533 		struct nvme_completion *cqe)
534 {
535 	struct nvme_tcp_request *req;
536 	struct request *rq;
537 
538 	rq = nvme_find_rq(nvme_tcp_tagset(queue), cqe->command_id);
539 	if (!rq) {
540 		dev_err(queue->ctrl->ctrl.device,
541 			"got bad cqe.command_id %#x on queue %d\n",
542 			cqe->command_id, nvme_tcp_queue_id(queue));
543 		nvme_tcp_error_recovery(&queue->ctrl->ctrl);
544 		return -EINVAL;
545 	}
546 
547 	req = blk_mq_rq_to_pdu(rq);
548 	if (req->status == cpu_to_le16(NVME_SC_SUCCESS))
549 		req->status = cqe->status;
550 
551 	if (!nvme_try_complete_req(rq, req->status, cqe->result))
552 		nvme_complete_rq(rq);
553 	queue->nr_cqe++;
554 
555 	return 0;
556 }
557 
558 static int nvme_tcp_handle_c2h_data(struct nvme_tcp_queue *queue,
559 		struct nvme_tcp_data_pdu *pdu)
560 {
561 	struct request *rq;
562 
563 	rq = nvme_find_rq(nvme_tcp_tagset(queue), pdu->command_id);
564 	if (!rq) {
565 		dev_err(queue->ctrl->ctrl.device,
566 			"got bad c2hdata.command_id %#x on queue %d\n",
567 			pdu->command_id, nvme_tcp_queue_id(queue));
568 		return -ENOENT;
569 	}
570 
571 	if (!blk_rq_payload_bytes(rq)) {
572 		dev_err(queue->ctrl->ctrl.device,
573 			"queue %d tag %#x unexpected data\n",
574 			nvme_tcp_queue_id(queue), rq->tag);
575 		return -EIO;
576 	}
577 
578 	queue->data_remaining = le32_to_cpu(pdu->data_length);
579 
580 	if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS &&
581 	    unlikely(!(pdu->hdr.flags & NVME_TCP_F_DATA_LAST))) {
582 		dev_err(queue->ctrl->ctrl.device,
583 			"queue %d tag %#x SUCCESS set but not last PDU\n",
584 			nvme_tcp_queue_id(queue), rq->tag);
585 		nvme_tcp_error_recovery(&queue->ctrl->ctrl);
586 		return -EPROTO;
587 	}
588 
589 	return 0;
590 }
591 
592 static int nvme_tcp_handle_comp(struct nvme_tcp_queue *queue,
593 		struct nvme_tcp_rsp_pdu *pdu)
594 {
595 	struct nvme_completion *cqe = &pdu->cqe;
596 	int ret = 0;
597 
598 	/*
599 	 * AEN requests are special as they don't time out and can
600 	 * survive any kind of queue freeze and often don't respond to
601 	 * aborts.  We don't even bother to allocate a struct request
602 	 * for them but rather special case them here.
603 	 */
604 	if (unlikely(nvme_is_aen_req(nvme_tcp_queue_id(queue),
605 				     cqe->command_id)))
606 		nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
607 				&cqe->result);
608 	else
609 		ret = nvme_tcp_process_nvme_cqe(queue, cqe);
610 
611 	return ret;
612 }
613 
614 static void nvme_tcp_setup_h2c_data_pdu(struct nvme_tcp_request *req)
615 {
616 	struct nvme_tcp_data_pdu *data = req->pdu;
617 	struct nvme_tcp_queue *queue = req->queue;
618 	struct request *rq = blk_mq_rq_from_pdu(req);
619 	u32 h2cdata_sent = req->pdu_len;
620 	u8 hdgst = nvme_tcp_hdgst_len(queue);
621 	u8 ddgst = nvme_tcp_ddgst_len(queue);
622 
623 	req->state = NVME_TCP_SEND_H2C_PDU;
624 	req->offset = 0;
625 	req->pdu_len = min(req->h2cdata_left, queue->maxh2cdata);
626 	req->pdu_sent = 0;
627 	req->h2cdata_left -= req->pdu_len;
628 	req->h2cdata_offset += h2cdata_sent;
629 
630 	memset(data, 0, sizeof(*data));
631 	data->hdr.type = nvme_tcp_h2c_data;
632 	if (!req->h2cdata_left)
633 		data->hdr.flags = NVME_TCP_F_DATA_LAST;
634 	if (queue->hdr_digest)
635 		data->hdr.flags |= NVME_TCP_F_HDGST;
636 	if (queue->data_digest)
637 		data->hdr.flags |= NVME_TCP_F_DDGST;
638 	data->hdr.hlen = sizeof(*data);
639 	data->hdr.pdo = data->hdr.hlen + hdgst;
640 	data->hdr.plen =
641 		cpu_to_le32(data->hdr.hlen + hdgst + req->pdu_len + ddgst);
642 	data->ttag = req->ttag;
643 	data->command_id = nvme_cid(rq);
644 	data->data_offset = cpu_to_le32(req->h2cdata_offset);
645 	data->data_length = cpu_to_le32(req->pdu_len);
646 }
647 
648 static int nvme_tcp_handle_r2t(struct nvme_tcp_queue *queue,
649 		struct nvme_tcp_r2t_pdu *pdu)
650 {
651 	struct nvme_tcp_request *req;
652 	struct request *rq;
653 	u32 r2t_length = le32_to_cpu(pdu->r2t_length);
654 	u32 r2t_offset = le32_to_cpu(pdu->r2t_offset);
655 
656 	rq = nvme_find_rq(nvme_tcp_tagset(queue), pdu->command_id);
657 	if (!rq) {
658 		dev_err(queue->ctrl->ctrl.device,
659 			"got bad r2t.command_id %#x on queue %d\n",
660 			pdu->command_id, nvme_tcp_queue_id(queue));
661 		return -ENOENT;
662 	}
663 	req = blk_mq_rq_to_pdu(rq);
664 
665 	if (unlikely(!r2t_length)) {
666 		dev_err(queue->ctrl->ctrl.device,
667 			"req %d r2t len is %u, probably a bug...\n",
668 			rq->tag, r2t_length);
669 		return -EPROTO;
670 	}
671 
672 	if (unlikely(req->data_sent + r2t_length > req->data_len)) {
673 		dev_err(queue->ctrl->ctrl.device,
674 			"req %d r2t len %u exceeded data len %u (%zu sent)\n",
675 			rq->tag, r2t_length, req->data_len, req->data_sent);
676 		return -EPROTO;
677 	}
678 
679 	if (unlikely(r2t_offset < req->data_sent)) {
680 		dev_err(queue->ctrl->ctrl.device,
681 			"req %d unexpected r2t offset %u (expected %zu)\n",
682 			rq->tag, r2t_offset, req->data_sent);
683 		return -EPROTO;
684 	}
685 
686 	req->pdu_len = 0;
687 	req->h2cdata_left = r2t_length;
688 	req->h2cdata_offset = r2t_offset;
689 	req->ttag = pdu->ttag;
690 
691 	nvme_tcp_setup_h2c_data_pdu(req);
692 	nvme_tcp_queue_request(req, false, true);
693 
694 	return 0;
695 }
696 
697 static int nvme_tcp_recv_pdu(struct nvme_tcp_queue *queue, struct sk_buff *skb,
698 		unsigned int *offset, size_t *len)
699 {
700 	struct nvme_tcp_hdr *hdr;
701 	char *pdu = queue->pdu;
702 	size_t rcv_len = min_t(size_t, *len, queue->pdu_remaining);
703 	int ret;
704 
705 	ret = skb_copy_bits(skb, *offset,
706 		&pdu[queue->pdu_offset], rcv_len);
707 	if (unlikely(ret))
708 		return ret;
709 
710 	queue->pdu_remaining -= rcv_len;
711 	queue->pdu_offset += rcv_len;
712 	*offset += rcv_len;
713 	*len -= rcv_len;
714 	if (queue->pdu_remaining)
715 		return 0;
716 
717 	hdr = queue->pdu;
718 	if (queue->hdr_digest) {
719 		ret = nvme_tcp_verify_hdgst(queue, queue->pdu, hdr->hlen);
720 		if (unlikely(ret))
721 			return ret;
722 	}
723 
724 
725 	if (queue->data_digest) {
726 		ret = nvme_tcp_check_ddgst(queue, queue->pdu);
727 		if (unlikely(ret))
728 			return ret;
729 	}
730 
731 	switch (hdr->type) {
732 	case nvme_tcp_c2h_data:
733 		return nvme_tcp_handle_c2h_data(queue, (void *)queue->pdu);
734 	case nvme_tcp_rsp:
735 		nvme_tcp_init_recv_ctx(queue);
736 		return nvme_tcp_handle_comp(queue, (void *)queue->pdu);
737 	case nvme_tcp_r2t:
738 		nvme_tcp_init_recv_ctx(queue);
739 		return nvme_tcp_handle_r2t(queue, (void *)queue->pdu);
740 	default:
741 		dev_err(queue->ctrl->ctrl.device,
742 			"unsupported pdu type (%d)\n", hdr->type);
743 		return -EINVAL;
744 	}
745 }
746 
747 static inline void nvme_tcp_end_request(struct request *rq, u16 status)
748 {
749 	union nvme_result res = {};
750 
751 	if (!nvme_try_complete_req(rq, cpu_to_le16(status << 1), res))
752 		nvme_complete_rq(rq);
753 }
754 
755 static int nvme_tcp_recv_data(struct nvme_tcp_queue *queue, struct sk_buff *skb,
756 			      unsigned int *offset, size_t *len)
757 {
758 	struct nvme_tcp_data_pdu *pdu = (void *)queue->pdu;
759 	struct request *rq =
760 		nvme_cid_to_rq(nvme_tcp_tagset(queue), pdu->command_id);
761 	struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
762 
763 	while (true) {
764 		int recv_len, ret;
765 
766 		recv_len = min_t(size_t, *len, queue->data_remaining);
767 		if (!recv_len)
768 			break;
769 
770 		if (!iov_iter_count(&req->iter)) {
771 			req->curr_bio = req->curr_bio->bi_next;
772 
773 			/*
774 			 * If we don`t have any bios it means that controller
775 			 * sent more data than we requested, hence error
776 			 */
777 			if (!req->curr_bio) {
778 				dev_err(queue->ctrl->ctrl.device,
779 					"queue %d no space in request %#x",
780 					nvme_tcp_queue_id(queue), rq->tag);
781 				nvme_tcp_init_recv_ctx(queue);
782 				return -EIO;
783 			}
784 			nvme_tcp_init_iter(req, ITER_DEST);
785 		}
786 
787 		/* we can read only from what is left in this bio */
788 		recv_len = min_t(size_t, recv_len,
789 				iov_iter_count(&req->iter));
790 
791 		if (queue->data_digest)
792 			ret = skb_copy_and_hash_datagram_iter(skb, *offset,
793 				&req->iter, recv_len, queue->rcv_hash);
794 		else
795 			ret = skb_copy_datagram_iter(skb, *offset,
796 					&req->iter, recv_len);
797 		if (ret) {
798 			dev_err(queue->ctrl->ctrl.device,
799 				"queue %d failed to copy request %#x data",
800 				nvme_tcp_queue_id(queue), rq->tag);
801 			return ret;
802 		}
803 
804 		*len -= recv_len;
805 		*offset += recv_len;
806 		queue->data_remaining -= recv_len;
807 	}
808 
809 	if (!queue->data_remaining) {
810 		if (queue->data_digest) {
811 			nvme_tcp_ddgst_final(queue->rcv_hash, &queue->exp_ddgst);
812 			queue->ddgst_remaining = NVME_TCP_DIGEST_LENGTH;
813 		} else {
814 			if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
815 				nvme_tcp_end_request(rq,
816 						le16_to_cpu(req->status));
817 				queue->nr_cqe++;
818 			}
819 			nvme_tcp_init_recv_ctx(queue);
820 		}
821 	}
822 
823 	return 0;
824 }
825 
826 static int nvme_tcp_recv_ddgst(struct nvme_tcp_queue *queue,
827 		struct sk_buff *skb, unsigned int *offset, size_t *len)
828 {
829 	struct nvme_tcp_data_pdu *pdu = (void *)queue->pdu;
830 	char *ddgst = (char *)&queue->recv_ddgst;
831 	size_t recv_len = min_t(size_t, *len, queue->ddgst_remaining);
832 	off_t off = NVME_TCP_DIGEST_LENGTH - queue->ddgst_remaining;
833 	int ret;
834 
835 	ret = skb_copy_bits(skb, *offset, &ddgst[off], recv_len);
836 	if (unlikely(ret))
837 		return ret;
838 
839 	queue->ddgst_remaining -= recv_len;
840 	*offset += recv_len;
841 	*len -= recv_len;
842 	if (queue->ddgst_remaining)
843 		return 0;
844 
845 	if (queue->recv_ddgst != queue->exp_ddgst) {
846 		struct request *rq = nvme_cid_to_rq(nvme_tcp_tagset(queue),
847 					pdu->command_id);
848 		struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
849 
850 		req->status = cpu_to_le16(NVME_SC_DATA_XFER_ERROR);
851 
852 		dev_err(queue->ctrl->ctrl.device,
853 			"data digest error: recv %#x expected %#x\n",
854 			le32_to_cpu(queue->recv_ddgst),
855 			le32_to_cpu(queue->exp_ddgst));
856 	}
857 
858 	if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
859 		struct request *rq = nvme_cid_to_rq(nvme_tcp_tagset(queue),
860 					pdu->command_id);
861 		struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
862 
863 		nvme_tcp_end_request(rq, le16_to_cpu(req->status));
864 		queue->nr_cqe++;
865 	}
866 
867 	nvme_tcp_init_recv_ctx(queue);
868 	return 0;
869 }
870 
871 static int nvme_tcp_recv_skb(read_descriptor_t *desc, struct sk_buff *skb,
872 			     unsigned int offset, size_t len)
873 {
874 	struct nvme_tcp_queue *queue = desc->arg.data;
875 	size_t consumed = len;
876 	int result;
877 
878 	while (len) {
879 		switch (nvme_tcp_recv_state(queue)) {
880 		case NVME_TCP_RECV_PDU:
881 			result = nvme_tcp_recv_pdu(queue, skb, &offset, &len);
882 			break;
883 		case NVME_TCP_RECV_DATA:
884 			result = nvme_tcp_recv_data(queue, skb, &offset, &len);
885 			break;
886 		case NVME_TCP_RECV_DDGST:
887 			result = nvme_tcp_recv_ddgst(queue, skb, &offset, &len);
888 			break;
889 		default:
890 			result = -EFAULT;
891 		}
892 		if (result) {
893 			dev_err(queue->ctrl->ctrl.device,
894 				"receive failed:  %d\n", result);
895 			queue->rd_enabled = false;
896 			nvme_tcp_error_recovery(&queue->ctrl->ctrl);
897 			return result;
898 		}
899 	}
900 
901 	return consumed;
902 }
903 
904 static void nvme_tcp_data_ready(struct sock *sk)
905 {
906 	struct nvme_tcp_queue *queue;
907 
908 	read_lock_bh(&sk->sk_callback_lock);
909 	queue = sk->sk_user_data;
910 	if (likely(queue && queue->rd_enabled) &&
911 	    !test_bit(NVME_TCP_Q_POLLING, &queue->flags))
912 		queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
913 	read_unlock_bh(&sk->sk_callback_lock);
914 }
915 
916 static void nvme_tcp_write_space(struct sock *sk)
917 {
918 	struct nvme_tcp_queue *queue;
919 
920 	read_lock_bh(&sk->sk_callback_lock);
921 	queue = sk->sk_user_data;
922 	if (likely(queue && sk_stream_is_writeable(sk))) {
923 		clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
924 		queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
925 	}
926 	read_unlock_bh(&sk->sk_callback_lock);
927 }
928 
929 static void nvme_tcp_state_change(struct sock *sk)
930 {
931 	struct nvme_tcp_queue *queue;
932 
933 	read_lock_bh(&sk->sk_callback_lock);
934 	queue = sk->sk_user_data;
935 	if (!queue)
936 		goto done;
937 
938 	switch (sk->sk_state) {
939 	case TCP_CLOSE:
940 	case TCP_CLOSE_WAIT:
941 	case TCP_LAST_ACK:
942 	case TCP_FIN_WAIT1:
943 	case TCP_FIN_WAIT2:
944 		nvme_tcp_error_recovery(&queue->ctrl->ctrl);
945 		break;
946 	default:
947 		dev_info(queue->ctrl->ctrl.device,
948 			"queue %d socket state %d\n",
949 			nvme_tcp_queue_id(queue), sk->sk_state);
950 	}
951 
952 	queue->state_change(sk);
953 done:
954 	read_unlock_bh(&sk->sk_callback_lock);
955 }
956 
957 static inline void nvme_tcp_done_send_req(struct nvme_tcp_queue *queue)
958 {
959 	queue->request = NULL;
960 }
961 
962 static void nvme_tcp_fail_request(struct nvme_tcp_request *req)
963 {
964 	if (nvme_tcp_async_req(req)) {
965 		union nvme_result res = {};
966 
967 		nvme_complete_async_event(&req->queue->ctrl->ctrl,
968 				cpu_to_le16(NVME_SC_HOST_PATH_ERROR), &res);
969 	} else {
970 		nvme_tcp_end_request(blk_mq_rq_from_pdu(req),
971 				NVME_SC_HOST_PATH_ERROR);
972 	}
973 }
974 
975 static int nvme_tcp_try_send_data(struct nvme_tcp_request *req)
976 {
977 	struct nvme_tcp_queue *queue = req->queue;
978 	int req_data_len = req->data_len;
979 	u32 h2cdata_left = req->h2cdata_left;
980 
981 	while (true) {
982 		struct page *page = nvme_tcp_req_cur_page(req);
983 		size_t offset = nvme_tcp_req_cur_offset(req);
984 		size_t len = nvme_tcp_req_cur_length(req);
985 		bool last = nvme_tcp_pdu_last_send(req, len);
986 		int req_data_sent = req->data_sent;
987 		int ret, flags = MSG_DONTWAIT;
988 
989 		if (last && !queue->data_digest && !nvme_tcp_queue_more(queue))
990 			flags |= MSG_EOR;
991 		else
992 			flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST;
993 
994 		if (sendpage_ok(page)) {
995 			ret = kernel_sendpage(queue->sock, page, offset, len,
996 					flags);
997 		} else {
998 			ret = sock_no_sendpage(queue->sock, page, offset, len,
999 					flags);
1000 		}
1001 		if (ret <= 0)
1002 			return ret;
1003 
1004 		if (queue->data_digest)
1005 			nvme_tcp_ddgst_update(queue->snd_hash, page,
1006 					offset, ret);
1007 
1008 		/*
1009 		 * update the request iterator except for the last payload send
1010 		 * in the request where we don't want to modify it as we may
1011 		 * compete with the RX path completing the request.
1012 		 */
1013 		if (req_data_sent + ret < req_data_len)
1014 			nvme_tcp_advance_req(req, ret);
1015 
1016 		/* fully successful last send in current PDU */
1017 		if (last && ret == len) {
1018 			if (queue->data_digest) {
1019 				nvme_tcp_ddgst_final(queue->snd_hash,
1020 					&req->ddgst);
1021 				req->state = NVME_TCP_SEND_DDGST;
1022 				req->offset = 0;
1023 			} else {
1024 				if (h2cdata_left)
1025 					nvme_tcp_setup_h2c_data_pdu(req);
1026 				else
1027 					nvme_tcp_done_send_req(queue);
1028 			}
1029 			return 1;
1030 		}
1031 	}
1032 	return -EAGAIN;
1033 }
1034 
1035 static int nvme_tcp_try_send_cmd_pdu(struct nvme_tcp_request *req)
1036 {
1037 	struct nvme_tcp_queue *queue = req->queue;
1038 	struct nvme_tcp_cmd_pdu *pdu = req->pdu;
1039 	bool inline_data = nvme_tcp_has_inline_data(req);
1040 	u8 hdgst = nvme_tcp_hdgst_len(queue);
1041 	int len = sizeof(*pdu) + hdgst - req->offset;
1042 	int flags = MSG_DONTWAIT;
1043 	int ret;
1044 
1045 	if (inline_data || nvme_tcp_queue_more(queue))
1046 		flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST;
1047 	else
1048 		flags |= MSG_EOR;
1049 
1050 	if (queue->hdr_digest && !req->offset)
1051 		nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
1052 
1053 	ret = kernel_sendpage(queue->sock, virt_to_page(pdu),
1054 			offset_in_page(pdu) + req->offset, len,  flags);
1055 	if (unlikely(ret <= 0))
1056 		return ret;
1057 
1058 	len -= ret;
1059 	if (!len) {
1060 		if (inline_data) {
1061 			req->state = NVME_TCP_SEND_DATA;
1062 			if (queue->data_digest)
1063 				crypto_ahash_init(queue->snd_hash);
1064 		} else {
1065 			nvme_tcp_done_send_req(queue);
1066 		}
1067 		return 1;
1068 	}
1069 	req->offset += ret;
1070 
1071 	return -EAGAIN;
1072 }
1073 
1074 static int nvme_tcp_try_send_data_pdu(struct nvme_tcp_request *req)
1075 {
1076 	struct nvme_tcp_queue *queue = req->queue;
1077 	struct nvme_tcp_data_pdu *pdu = req->pdu;
1078 	u8 hdgst = nvme_tcp_hdgst_len(queue);
1079 	int len = sizeof(*pdu) - req->offset + hdgst;
1080 	int ret;
1081 
1082 	if (queue->hdr_digest && !req->offset)
1083 		nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
1084 
1085 	if (!req->h2cdata_left)
1086 		ret = kernel_sendpage(queue->sock, virt_to_page(pdu),
1087 				offset_in_page(pdu) + req->offset, len,
1088 				MSG_DONTWAIT | MSG_MORE | MSG_SENDPAGE_NOTLAST);
1089 	else
1090 		ret = sock_no_sendpage(queue->sock, virt_to_page(pdu),
1091 				offset_in_page(pdu) + req->offset, len,
1092 				MSG_DONTWAIT | MSG_MORE);
1093 	if (unlikely(ret <= 0))
1094 		return ret;
1095 
1096 	len -= ret;
1097 	if (!len) {
1098 		req->state = NVME_TCP_SEND_DATA;
1099 		if (queue->data_digest)
1100 			crypto_ahash_init(queue->snd_hash);
1101 		return 1;
1102 	}
1103 	req->offset += ret;
1104 
1105 	return -EAGAIN;
1106 }
1107 
1108 static int nvme_tcp_try_send_ddgst(struct nvme_tcp_request *req)
1109 {
1110 	struct nvme_tcp_queue *queue = req->queue;
1111 	size_t offset = req->offset;
1112 	u32 h2cdata_left = req->h2cdata_left;
1113 	int ret;
1114 	struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
1115 	struct kvec iov = {
1116 		.iov_base = (u8 *)&req->ddgst + req->offset,
1117 		.iov_len = NVME_TCP_DIGEST_LENGTH - req->offset
1118 	};
1119 
1120 	if (nvme_tcp_queue_more(queue))
1121 		msg.msg_flags |= MSG_MORE;
1122 	else
1123 		msg.msg_flags |= MSG_EOR;
1124 
1125 	ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
1126 	if (unlikely(ret <= 0))
1127 		return ret;
1128 
1129 	if (offset + ret == NVME_TCP_DIGEST_LENGTH) {
1130 		if (h2cdata_left)
1131 			nvme_tcp_setup_h2c_data_pdu(req);
1132 		else
1133 			nvme_tcp_done_send_req(queue);
1134 		return 1;
1135 	}
1136 
1137 	req->offset += ret;
1138 	return -EAGAIN;
1139 }
1140 
1141 static int nvme_tcp_try_send(struct nvme_tcp_queue *queue)
1142 {
1143 	struct nvme_tcp_request *req;
1144 	unsigned int noreclaim_flag;
1145 	int ret = 1;
1146 
1147 	if (!queue->request) {
1148 		queue->request = nvme_tcp_fetch_request(queue);
1149 		if (!queue->request)
1150 			return 0;
1151 	}
1152 	req = queue->request;
1153 
1154 	noreclaim_flag = memalloc_noreclaim_save();
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 			goto out;
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);
1177 done:
1178 	if (ret == -EAGAIN) {
1179 		ret = 0;
1180 	} else if (ret < 0) {
1181 		dev_err(queue->ctrl->ctrl.device,
1182 			"failed to send request %d\n", ret);
1183 		nvme_tcp_fail_request(queue->request);
1184 		nvme_tcp_done_send_req(queue);
1185 	}
1186 out:
1187 	memalloc_noreclaim_restore(noreclaim_flag);
1188 	return ret;
1189 }
1190 
1191 static int nvme_tcp_try_recv(struct nvme_tcp_queue *queue)
1192 {
1193 	struct socket *sock = queue->sock;
1194 	struct sock *sk = sock->sk;
1195 	read_descriptor_t rd_desc;
1196 	int consumed;
1197 
1198 	rd_desc.arg.data = queue;
1199 	rd_desc.count = 1;
1200 	lock_sock(sk);
1201 	queue->nr_cqe = 0;
1202 	consumed = sock->ops->read_sock(sk, &rd_desc, nvme_tcp_recv_skb);
1203 	release_sock(sk);
1204 	return consumed;
1205 }
1206 
1207 static void nvme_tcp_io_work(struct work_struct *w)
1208 {
1209 	struct nvme_tcp_queue *queue =
1210 		container_of(w, struct nvme_tcp_queue, io_work);
1211 	unsigned long deadline = jiffies + msecs_to_jiffies(1);
1212 
1213 	do {
1214 		bool pending = false;
1215 		int result;
1216 
1217 		if (mutex_trylock(&queue->send_mutex)) {
1218 			result = nvme_tcp_try_send(queue);
1219 			mutex_unlock(&queue->send_mutex);
1220 			if (result > 0)
1221 				pending = true;
1222 			else if (unlikely(result < 0))
1223 				break;
1224 		}
1225 
1226 		result = nvme_tcp_try_recv(queue);
1227 		if (result > 0)
1228 			pending = true;
1229 		else if (unlikely(result < 0))
1230 			return;
1231 
1232 		if (!pending || !queue->rd_enabled)
1233 			return;
1234 
1235 	} while (!time_after(jiffies, deadline)); /* quota is exhausted */
1236 
1237 	queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
1238 }
1239 
1240 static void nvme_tcp_free_crypto(struct nvme_tcp_queue *queue)
1241 {
1242 	struct crypto_ahash *tfm = crypto_ahash_reqtfm(queue->rcv_hash);
1243 
1244 	ahash_request_free(queue->rcv_hash);
1245 	ahash_request_free(queue->snd_hash);
1246 	crypto_free_ahash(tfm);
1247 }
1248 
1249 static int nvme_tcp_alloc_crypto(struct nvme_tcp_queue *queue)
1250 {
1251 	struct crypto_ahash *tfm;
1252 
1253 	tfm = crypto_alloc_ahash("crc32c", 0, CRYPTO_ALG_ASYNC);
1254 	if (IS_ERR(tfm))
1255 		return PTR_ERR(tfm);
1256 
1257 	queue->snd_hash = ahash_request_alloc(tfm, GFP_KERNEL);
1258 	if (!queue->snd_hash)
1259 		goto free_tfm;
1260 	ahash_request_set_callback(queue->snd_hash, 0, NULL, NULL);
1261 
1262 	queue->rcv_hash = ahash_request_alloc(tfm, GFP_KERNEL);
1263 	if (!queue->rcv_hash)
1264 		goto free_snd_hash;
1265 	ahash_request_set_callback(queue->rcv_hash, 0, NULL, NULL);
1266 
1267 	return 0;
1268 free_snd_hash:
1269 	ahash_request_free(queue->snd_hash);
1270 free_tfm:
1271 	crypto_free_ahash(tfm);
1272 	return -ENOMEM;
1273 }
1274 
1275 static void nvme_tcp_free_async_req(struct nvme_tcp_ctrl *ctrl)
1276 {
1277 	struct nvme_tcp_request *async = &ctrl->async_req;
1278 
1279 	page_frag_free(async->pdu);
1280 }
1281 
1282 static int nvme_tcp_alloc_async_req(struct nvme_tcp_ctrl *ctrl)
1283 {
1284 	struct nvme_tcp_queue *queue = &ctrl->queues[0];
1285 	struct nvme_tcp_request *async = &ctrl->async_req;
1286 	u8 hdgst = nvme_tcp_hdgst_len(queue);
1287 
1288 	async->pdu = page_frag_alloc(&queue->pf_cache,
1289 		sizeof(struct nvme_tcp_cmd_pdu) + hdgst,
1290 		GFP_KERNEL | __GFP_ZERO);
1291 	if (!async->pdu)
1292 		return -ENOMEM;
1293 
1294 	async->queue = &ctrl->queues[0];
1295 	return 0;
1296 }
1297 
1298 static void nvme_tcp_free_queue(struct nvme_ctrl *nctrl, int qid)
1299 {
1300 	struct page *page;
1301 	struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1302 	struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1303 	unsigned int noreclaim_flag;
1304 
1305 	if (!test_and_clear_bit(NVME_TCP_Q_ALLOCATED, &queue->flags))
1306 		return;
1307 
1308 	if (queue->hdr_digest || queue->data_digest)
1309 		nvme_tcp_free_crypto(queue);
1310 
1311 	if (queue->pf_cache.va) {
1312 		page = virt_to_head_page(queue->pf_cache.va);
1313 		__page_frag_cache_drain(page, queue->pf_cache.pagecnt_bias);
1314 		queue->pf_cache.va = NULL;
1315 	}
1316 
1317 	noreclaim_flag = memalloc_noreclaim_save();
1318 	sock_release(queue->sock);
1319 	memalloc_noreclaim_restore(noreclaim_flag);
1320 
1321 	kfree(queue->pdu);
1322 	mutex_destroy(&queue->send_mutex);
1323 	mutex_destroy(&queue->queue_lock);
1324 }
1325 
1326 static int nvme_tcp_init_connection(struct nvme_tcp_queue *queue)
1327 {
1328 	struct nvme_tcp_icreq_pdu *icreq;
1329 	struct nvme_tcp_icresp_pdu *icresp;
1330 	struct msghdr msg = {};
1331 	struct kvec iov;
1332 	bool ctrl_hdgst, ctrl_ddgst;
1333 	u32 maxh2cdata;
1334 	int ret;
1335 
1336 	icreq = kzalloc(sizeof(*icreq), GFP_KERNEL);
1337 	if (!icreq)
1338 		return -ENOMEM;
1339 
1340 	icresp = kzalloc(sizeof(*icresp), GFP_KERNEL);
1341 	if (!icresp) {
1342 		ret = -ENOMEM;
1343 		goto free_icreq;
1344 	}
1345 
1346 	icreq->hdr.type = nvme_tcp_icreq;
1347 	icreq->hdr.hlen = sizeof(*icreq);
1348 	icreq->hdr.pdo = 0;
1349 	icreq->hdr.plen = cpu_to_le32(icreq->hdr.hlen);
1350 	icreq->pfv = cpu_to_le16(NVME_TCP_PFV_1_0);
1351 	icreq->maxr2t = 0; /* single inflight r2t supported */
1352 	icreq->hpda = 0; /* no alignment constraint */
1353 	if (queue->hdr_digest)
1354 		icreq->digest |= NVME_TCP_HDR_DIGEST_ENABLE;
1355 	if (queue->data_digest)
1356 		icreq->digest |= NVME_TCP_DATA_DIGEST_ENABLE;
1357 
1358 	iov.iov_base = icreq;
1359 	iov.iov_len = sizeof(*icreq);
1360 	ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
1361 	if (ret < 0)
1362 		goto free_icresp;
1363 
1364 	memset(&msg, 0, sizeof(msg));
1365 	iov.iov_base = icresp;
1366 	iov.iov_len = sizeof(*icresp);
1367 	ret = kernel_recvmsg(queue->sock, &msg, &iov, 1,
1368 			iov.iov_len, msg.msg_flags);
1369 	if (ret < 0)
1370 		goto free_icresp;
1371 
1372 	ret = -EINVAL;
1373 	if (icresp->hdr.type != nvme_tcp_icresp) {
1374 		pr_err("queue %d: bad type returned %d\n",
1375 			nvme_tcp_queue_id(queue), icresp->hdr.type);
1376 		goto free_icresp;
1377 	}
1378 
1379 	if (le32_to_cpu(icresp->hdr.plen) != sizeof(*icresp)) {
1380 		pr_err("queue %d: bad pdu length returned %d\n",
1381 			nvme_tcp_queue_id(queue), icresp->hdr.plen);
1382 		goto free_icresp;
1383 	}
1384 
1385 	if (icresp->pfv != NVME_TCP_PFV_1_0) {
1386 		pr_err("queue %d: bad pfv returned %d\n",
1387 			nvme_tcp_queue_id(queue), icresp->pfv);
1388 		goto free_icresp;
1389 	}
1390 
1391 	ctrl_ddgst = !!(icresp->digest & NVME_TCP_DATA_DIGEST_ENABLE);
1392 	if ((queue->data_digest && !ctrl_ddgst) ||
1393 	    (!queue->data_digest && ctrl_ddgst)) {
1394 		pr_err("queue %d: data digest mismatch host: %s ctrl: %s\n",
1395 			nvme_tcp_queue_id(queue),
1396 			queue->data_digest ? "enabled" : "disabled",
1397 			ctrl_ddgst ? "enabled" : "disabled");
1398 		goto free_icresp;
1399 	}
1400 
1401 	ctrl_hdgst = !!(icresp->digest & NVME_TCP_HDR_DIGEST_ENABLE);
1402 	if ((queue->hdr_digest && !ctrl_hdgst) ||
1403 	    (!queue->hdr_digest && ctrl_hdgst)) {
1404 		pr_err("queue %d: header digest mismatch host: %s ctrl: %s\n",
1405 			nvme_tcp_queue_id(queue),
1406 			queue->hdr_digest ? "enabled" : "disabled",
1407 			ctrl_hdgst ? "enabled" : "disabled");
1408 		goto free_icresp;
1409 	}
1410 
1411 	if (icresp->cpda != 0) {
1412 		pr_err("queue %d: unsupported cpda returned %d\n",
1413 			nvme_tcp_queue_id(queue), icresp->cpda);
1414 		goto free_icresp;
1415 	}
1416 
1417 	maxh2cdata = le32_to_cpu(icresp->maxdata);
1418 	if ((maxh2cdata % 4) || (maxh2cdata < NVME_TCP_MIN_MAXH2CDATA)) {
1419 		pr_err("queue %d: invalid maxh2cdata returned %u\n",
1420 		       nvme_tcp_queue_id(queue), maxh2cdata);
1421 		goto free_icresp;
1422 	}
1423 	queue->maxh2cdata = maxh2cdata;
1424 
1425 	ret = 0;
1426 free_icresp:
1427 	kfree(icresp);
1428 free_icreq:
1429 	kfree(icreq);
1430 	return ret;
1431 }
1432 
1433 static bool nvme_tcp_admin_queue(struct nvme_tcp_queue *queue)
1434 {
1435 	return nvme_tcp_queue_id(queue) == 0;
1436 }
1437 
1438 static bool nvme_tcp_default_queue(struct nvme_tcp_queue *queue)
1439 {
1440 	struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1441 	int qid = nvme_tcp_queue_id(queue);
1442 
1443 	return !nvme_tcp_admin_queue(queue) &&
1444 		qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT];
1445 }
1446 
1447 static bool nvme_tcp_read_queue(struct nvme_tcp_queue *queue)
1448 {
1449 	struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1450 	int qid = nvme_tcp_queue_id(queue);
1451 
1452 	return !nvme_tcp_admin_queue(queue) &&
1453 		!nvme_tcp_default_queue(queue) &&
1454 		qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT] +
1455 			  ctrl->io_queues[HCTX_TYPE_READ];
1456 }
1457 
1458 static bool nvme_tcp_poll_queue(struct nvme_tcp_queue *queue)
1459 {
1460 	struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1461 	int qid = nvme_tcp_queue_id(queue);
1462 
1463 	return !nvme_tcp_admin_queue(queue) &&
1464 		!nvme_tcp_default_queue(queue) &&
1465 		!nvme_tcp_read_queue(queue) &&
1466 		qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT] +
1467 			  ctrl->io_queues[HCTX_TYPE_READ] +
1468 			  ctrl->io_queues[HCTX_TYPE_POLL];
1469 }
1470 
1471 static void nvme_tcp_set_queue_io_cpu(struct nvme_tcp_queue *queue)
1472 {
1473 	struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1474 	int qid = nvme_tcp_queue_id(queue);
1475 	int n = 0;
1476 
1477 	if (nvme_tcp_default_queue(queue))
1478 		n = qid - 1;
1479 	else if (nvme_tcp_read_queue(queue))
1480 		n = qid - ctrl->io_queues[HCTX_TYPE_DEFAULT] - 1;
1481 	else if (nvme_tcp_poll_queue(queue))
1482 		n = qid - ctrl->io_queues[HCTX_TYPE_DEFAULT] -
1483 				ctrl->io_queues[HCTX_TYPE_READ] - 1;
1484 	queue->io_cpu = cpumask_next_wrap(n - 1, cpu_online_mask, -1, false);
1485 }
1486 
1487 static int nvme_tcp_alloc_queue(struct nvme_ctrl *nctrl, int qid)
1488 {
1489 	struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1490 	struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1491 	int ret, rcv_pdu_size;
1492 
1493 	mutex_init(&queue->queue_lock);
1494 	queue->ctrl = ctrl;
1495 	init_llist_head(&queue->req_list);
1496 	INIT_LIST_HEAD(&queue->send_list);
1497 	mutex_init(&queue->send_mutex);
1498 	INIT_WORK(&queue->io_work, nvme_tcp_io_work);
1499 
1500 	if (qid > 0)
1501 		queue->cmnd_capsule_len = nctrl->ioccsz * 16;
1502 	else
1503 		queue->cmnd_capsule_len = sizeof(struct nvme_command) +
1504 						NVME_TCP_ADMIN_CCSZ;
1505 
1506 	ret = sock_create(ctrl->addr.ss_family, SOCK_STREAM,
1507 			IPPROTO_TCP, &queue->sock);
1508 	if (ret) {
1509 		dev_err(nctrl->device,
1510 			"failed to create socket: %d\n", ret);
1511 		goto err_destroy_mutex;
1512 	}
1513 
1514 	nvme_tcp_reclassify_socket(queue->sock);
1515 
1516 	/* Single syn retry */
1517 	tcp_sock_set_syncnt(queue->sock->sk, 1);
1518 
1519 	/* Set TCP no delay */
1520 	tcp_sock_set_nodelay(queue->sock->sk);
1521 
1522 	/*
1523 	 * Cleanup whatever is sitting in the TCP transmit queue on socket
1524 	 * close. This is done to prevent stale data from being sent should
1525 	 * the network connection be restored before TCP times out.
1526 	 */
1527 	sock_no_linger(queue->sock->sk);
1528 
1529 	if (so_priority > 0)
1530 		sock_set_priority(queue->sock->sk, so_priority);
1531 
1532 	/* Set socket type of service */
1533 	if (nctrl->opts->tos >= 0)
1534 		ip_sock_set_tos(queue->sock->sk, nctrl->opts->tos);
1535 
1536 	/* Set 10 seconds timeout for icresp recvmsg */
1537 	queue->sock->sk->sk_rcvtimeo = 10 * HZ;
1538 
1539 	queue->sock->sk->sk_allocation = GFP_ATOMIC;
1540 	queue->sock->sk->sk_use_task_frag = false;
1541 	nvme_tcp_set_queue_io_cpu(queue);
1542 	queue->request = NULL;
1543 	queue->data_remaining = 0;
1544 	queue->ddgst_remaining = 0;
1545 	queue->pdu_remaining = 0;
1546 	queue->pdu_offset = 0;
1547 	sk_set_memalloc(queue->sock->sk);
1548 
1549 	if (nctrl->opts->mask & NVMF_OPT_HOST_TRADDR) {
1550 		ret = kernel_bind(queue->sock, (struct sockaddr *)&ctrl->src_addr,
1551 			sizeof(ctrl->src_addr));
1552 		if (ret) {
1553 			dev_err(nctrl->device,
1554 				"failed to bind queue %d socket %d\n",
1555 				qid, ret);
1556 			goto err_sock;
1557 		}
1558 	}
1559 
1560 	if (nctrl->opts->mask & NVMF_OPT_HOST_IFACE) {
1561 		char *iface = nctrl->opts->host_iface;
1562 		sockptr_t optval = KERNEL_SOCKPTR(iface);
1563 
1564 		ret = sock_setsockopt(queue->sock, SOL_SOCKET, SO_BINDTODEVICE,
1565 				      optval, strlen(iface));
1566 		if (ret) {
1567 			dev_err(nctrl->device,
1568 			  "failed to bind to interface %s queue %d err %d\n",
1569 			  iface, qid, ret);
1570 			goto err_sock;
1571 		}
1572 	}
1573 
1574 	queue->hdr_digest = nctrl->opts->hdr_digest;
1575 	queue->data_digest = nctrl->opts->data_digest;
1576 	if (queue->hdr_digest || queue->data_digest) {
1577 		ret = nvme_tcp_alloc_crypto(queue);
1578 		if (ret) {
1579 			dev_err(nctrl->device,
1580 				"failed to allocate queue %d crypto\n", qid);
1581 			goto err_sock;
1582 		}
1583 	}
1584 
1585 	rcv_pdu_size = sizeof(struct nvme_tcp_rsp_pdu) +
1586 			nvme_tcp_hdgst_len(queue);
1587 	queue->pdu = kmalloc(rcv_pdu_size, GFP_KERNEL);
1588 	if (!queue->pdu) {
1589 		ret = -ENOMEM;
1590 		goto err_crypto;
1591 	}
1592 
1593 	dev_dbg(nctrl->device, "connecting queue %d\n",
1594 			nvme_tcp_queue_id(queue));
1595 
1596 	ret = kernel_connect(queue->sock, (struct sockaddr *)&ctrl->addr,
1597 		sizeof(ctrl->addr), 0);
1598 	if (ret) {
1599 		dev_err(nctrl->device,
1600 			"failed to connect socket: %d\n", ret);
1601 		goto err_rcv_pdu;
1602 	}
1603 
1604 	ret = nvme_tcp_init_connection(queue);
1605 	if (ret)
1606 		goto err_init_connect;
1607 
1608 	queue->rd_enabled = true;
1609 	set_bit(NVME_TCP_Q_ALLOCATED, &queue->flags);
1610 	nvme_tcp_init_recv_ctx(queue);
1611 
1612 	write_lock_bh(&queue->sock->sk->sk_callback_lock);
1613 	queue->sock->sk->sk_user_data = queue;
1614 	queue->state_change = queue->sock->sk->sk_state_change;
1615 	queue->data_ready = queue->sock->sk->sk_data_ready;
1616 	queue->write_space = queue->sock->sk->sk_write_space;
1617 	queue->sock->sk->sk_data_ready = nvme_tcp_data_ready;
1618 	queue->sock->sk->sk_state_change = nvme_tcp_state_change;
1619 	queue->sock->sk->sk_write_space = nvme_tcp_write_space;
1620 #ifdef CONFIG_NET_RX_BUSY_POLL
1621 	queue->sock->sk->sk_ll_usec = 1;
1622 #endif
1623 	write_unlock_bh(&queue->sock->sk->sk_callback_lock);
1624 
1625 	return 0;
1626 
1627 err_init_connect:
1628 	kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1629 err_rcv_pdu:
1630 	kfree(queue->pdu);
1631 err_crypto:
1632 	if (queue->hdr_digest || queue->data_digest)
1633 		nvme_tcp_free_crypto(queue);
1634 err_sock:
1635 	sock_release(queue->sock);
1636 	queue->sock = NULL;
1637 err_destroy_mutex:
1638 	mutex_destroy(&queue->send_mutex);
1639 	mutex_destroy(&queue->queue_lock);
1640 	return ret;
1641 }
1642 
1643 static void nvme_tcp_restore_sock_calls(struct nvme_tcp_queue *queue)
1644 {
1645 	struct socket *sock = queue->sock;
1646 
1647 	write_lock_bh(&sock->sk->sk_callback_lock);
1648 	sock->sk->sk_user_data  = NULL;
1649 	sock->sk->sk_data_ready = queue->data_ready;
1650 	sock->sk->sk_state_change = queue->state_change;
1651 	sock->sk->sk_write_space  = queue->write_space;
1652 	write_unlock_bh(&sock->sk->sk_callback_lock);
1653 }
1654 
1655 static void __nvme_tcp_stop_queue(struct nvme_tcp_queue *queue)
1656 {
1657 	kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1658 	nvme_tcp_restore_sock_calls(queue);
1659 	cancel_work_sync(&queue->io_work);
1660 }
1661 
1662 static void nvme_tcp_stop_queue(struct nvme_ctrl *nctrl, int qid)
1663 {
1664 	struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1665 	struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1666 
1667 	if (!test_bit(NVME_TCP_Q_ALLOCATED, &queue->flags))
1668 		return;
1669 
1670 	mutex_lock(&queue->queue_lock);
1671 	if (test_and_clear_bit(NVME_TCP_Q_LIVE, &queue->flags))
1672 		__nvme_tcp_stop_queue(queue);
1673 	mutex_unlock(&queue->queue_lock);
1674 }
1675 
1676 static int nvme_tcp_start_queue(struct nvme_ctrl *nctrl, int idx)
1677 {
1678 	struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1679 	int ret;
1680 
1681 	if (idx)
1682 		ret = nvmf_connect_io_queue(nctrl, idx);
1683 	else
1684 		ret = nvmf_connect_admin_queue(nctrl);
1685 
1686 	if (!ret) {
1687 		set_bit(NVME_TCP_Q_LIVE, &ctrl->queues[idx].flags);
1688 	} else {
1689 		if (test_bit(NVME_TCP_Q_ALLOCATED, &ctrl->queues[idx].flags))
1690 			__nvme_tcp_stop_queue(&ctrl->queues[idx]);
1691 		dev_err(nctrl->device,
1692 			"failed to connect queue: %d ret=%d\n", idx, ret);
1693 	}
1694 	return ret;
1695 }
1696 
1697 static void nvme_tcp_free_admin_queue(struct nvme_ctrl *ctrl)
1698 {
1699 	if (to_tcp_ctrl(ctrl)->async_req.pdu) {
1700 		cancel_work_sync(&ctrl->async_event_work);
1701 		nvme_tcp_free_async_req(to_tcp_ctrl(ctrl));
1702 		to_tcp_ctrl(ctrl)->async_req.pdu = NULL;
1703 	}
1704 
1705 	nvme_tcp_free_queue(ctrl, 0);
1706 }
1707 
1708 static void nvme_tcp_free_io_queues(struct nvme_ctrl *ctrl)
1709 {
1710 	int i;
1711 
1712 	for (i = 1; i < ctrl->queue_count; i++)
1713 		nvme_tcp_free_queue(ctrl, i);
1714 }
1715 
1716 static void nvme_tcp_stop_io_queues(struct nvme_ctrl *ctrl)
1717 {
1718 	int i;
1719 
1720 	for (i = 1; i < ctrl->queue_count; i++)
1721 		nvme_tcp_stop_queue(ctrl, i);
1722 }
1723 
1724 static int nvme_tcp_start_io_queues(struct nvme_ctrl *ctrl,
1725 				    int first, int last)
1726 {
1727 	int i, ret;
1728 
1729 	for (i = first; i < last; i++) {
1730 		ret = nvme_tcp_start_queue(ctrl, i);
1731 		if (ret)
1732 			goto out_stop_queues;
1733 	}
1734 
1735 	return 0;
1736 
1737 out_stop_queues:
1738 	for (i--; i >= first; i--)
1739 		nvme_tcp_stop_queue(ctrl, i);
1740 	return ret;
1741 }
1742 
1743 static int nvme_tcp_alloc_admin_queue(struct nvme_ctrl *ctrl)
1744 {
1745 	int ret;
1746 
1747 	ret = nvme_tcp_alloc_queue(ctrl, 0);
1748 	if (ret)
1749 		return ret;
1750 
1751 	ret = nvme_tcp_alloc_async_req(to_tcp_ctrl(ctrl));
1752 	if (ret)
1753 		goto out_free_queue;
1754 
1755 	return 0;
1756 
1757 out_free_queue:
1758 	nvme_tcp_free_queue(ctrl, 0);
1759 	return ret;
1760 }
1761 
1762 static int __nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
1763 {
1764 	int i, ret;
1765 
1766 	for (i = 1; i < ctrl->queue_count; i++) {
1767 		ret = nvme_tcp_alloc_queue(ctrl, i);
1768 		if (ret)
1769 			goto out_free_queues;
1770 	}
1771 
1772 	return 0;
1773 
1774 out_free_queues:
1775 	for (i--; i >= 1; i--)
1776 		nvme_tcp_free_queue(ctrl, i);
1777 
1778 	return ret;
1779 }
1780 
1781 static unsigned int nvme_tcp_nr_io_queues(struct nvme_ctrl *ctrl)
1782 {
1783 	unsigned int nr_io_queues;
1784 
1785 	nr_io_queues = min(ctrl->opts->nr_io_queues, num_online_cpus());
1786 	nr_io_queues += min(ctrl->opts->nr_write_queues, num_online_cpus());
1787 	nr_io_queues += min(ctrl->opts->nr_poll_queues, num_online_cpus());
1788 
1789 	return nr_io_queues;
1790 }
1791 
1792 static void nvme_tcp_set_io_queues(struct nvme_ctrl *nctrl,
1793 		unsigned int nr_io_queues)
1794 {
1795 	struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1796 	struct nvmf_ctrl_options *opts = nctrl->opts;
1797 
1798 	if (opts->nr_write_queues && opts->nr_io_queues < nr_io_queues) {
1799 		/*
1800 		 * separate read/write queues
1801 		 * hand out dedicated default queues only after we have
1802 		 * sufficient read queues.
1803 		 */
1804 		ctrl->io_queues[HCTX_TYPE_READ] = opts->nr_io_queues;
1805 		nr_io_queues -= ctrl->io_queues[HCTX_TYPE_READ];
1806 		ctrl->io_queues[HCTX_TYPE_DEFAULT] =
1807 			min(opts->nr_write_queues, nr_io_queues);
1808 		nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
1809 	} else {
1810 		/*
1811 		 * shared read/write queues
1812 		 * either no write queues were requested, or we don't have
1813 		 * sufficient queue count to have dedicated default queues.
1814 		 */
1815 		ctrl->io_queues[HCTX_TYPE_DEFAULT] =
1816 			min(opts->nr_io_queues, nr_io_queues);
1817 		nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
1818 	}
1819 
1820 	if (opts->nr_poll_queues && nr_io_queues) {
1821 		/* map dedicated poll queues only if we have queues left */
1822 		ctrl->io_queues[HCTX_TYPE_POLL] =
1823 			min(opts->nr_poll_queues, nr_io_queues);
1824 	}
1825 }
1826 
1827 static int nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
1828 {
1829 	unsigned int nr_io_queues;
1830 	int ret;
1831 
1832 	nr_io_queues = nvme_tcp_nr_io_queues(ctrl);
1833 	ret = nvme_set_queue_count(ctrl, &nr_io_queues);
1834 	if (ret)
1835 		return ret;
1836 
1837 	if (nr_io_queues == 0) {
1838 		dev_err(ctrl->device,
1839 			"unable to set any I/O queues\n");
1840 		return -ENOMEM;
1841 	}
1842 
1843 	ctrl->queue_count = nr_io_queues + 1;
1844 	dev_info(ctrl->device,
1845 		"creating %d I/O queues.\n", nr_io_queues);
1846 
1847 	nvme_tcp_set_io_queues(ctrl, nr_io_queues);
1848 
1849 	return __nvme_tcp_alloc_io_queues(ctrl);
1850 }
1851 
1852 static void nvme_tcp_destroy_io_queues(struct nvme_ctrl *ctrl, bool remove)
1853 {
1854 	nvme_tcp_stop_io_queues(ctrl);
1855 	if (remove)
1856 		nvme_remove_io_tag_set(ctrl);
1857 	nvme_tcp_free_io_queues(ctrl);
1858 }
1859 
1860 static int nvme_tcp_configure_io_queues(struct nvme_ctrl *ctrl, bool new)
1861 {
1862 	int ret, nr_queues;
1863 
1864 	ret = nvme_tcp_alloc_io_queues(ctrl);
1865 	if (ret)
1866 		return ret;
1867 
1868 	if (new) {
1869 		ret = nvme_alloc_io_tag_set(ctrl, &to_tcp_ctrl(ctrl)->tag_set,
1870 				&nvme_tcp_mq_ops,
1871 				ctrl->opts->nr_poll_queues ? HCTX_MAX_TYPES : 2,
1872 				sizeof(struct nvme_tcp_request));
1873 		if (ret)
1874 			goto out_free_io_queues;
1875 	}
1876 
1877 	/*
1878 	 * Only start IO queues for which we have allocated the tagset
1879 	 * and limitted it to the available queues. On reconnects, the
1880 	 * queue number might have changed.
1881 	 */
1882 	nr_queues = min(ctrl->tagset->nr_hw_queues + 1, ctrl->queue_count);
1883 	ret = nvme_tcp_start_io_queues(ctrl, 1, nr_queues);
1884 	if (ret)
1885 		goto out_cleanup_connect_q;
1886 
1887 	if (!new) {
1888 		nvme_unquiesce_io_queues(ctrl);
1889 		if (!nvme_wait_freeze_timeout(ctrl, NVME_IO_TIMEOUT)) {
1890 			/*
1891 			 * If we timed out waiting for freeze we are likely to
1892 			 * be stuck.  Fail the controller initialization just
1893 			 * to be safe.
1894 			 */
1895 			ret = -ENODEV;
1896 			goto out_wait_freeze_timed_out;
1897 		}
1898 		blk_mq_update_nr_hw_queues(ctrl->tagset,
1899 			ctrl->queue_count - 1);
1900 		nvme_unfreeze(ctrl);
1901 	}
1902 
1903 	/*
1904 	 * If the number of queues has increased (reconnect case)
1905 	 * start all new queues now.
1906 	 */
1907 	ret = nvme_tcp_start_io_queues(ctrl, nr_queues,
1908 				       ctrl->tagset->nr_hw_queues + 1);
1909 	if (ret)
1910 		goto out_wait_freeze_timed_out;
1911 
1912 	return 0;
1913 
1914 out_wait_freeze_timed_out:
1915 	nvme_quiesce_io_queues(ctrl);
1916 	nvme_sync_io_queues(ctrl);
1917 	nvme_tcp_stop_io_queues(ctrl);
1918 out_cleanup_connect_q:
1919 	nvme_cancel_tagset(ctrl);
1920 	if (new)
1921 		nvme_remove_io_tag_set(ctrl);
1922 out_free_io_queues:
1923 	nvme_tcp_free_io_queues(ctrl);
1924 	return ret;
1925 }
1926 
1927 static void nvme_tcp_destroy_admin_queue(struct nvme_ctrl *ctrl, bool remove)
1928 {
1929 	nvme_tcp_stop_queue(ctrl, 0);
1930 	if (remove)
1931 		nvme_remove_admin_tag_set(ctrl);
1932 	nvme_tcp_free_admin_queue(ctrl);
1933 }
1934 
1935 static int nvme_tcp_configure_admin_queue(struct nvme_ctrl *ctrl, bool new)
1936 {
1937 	int error;
1938 
1939 	error = nvme_tcp_alloc_admin_queue(ctrl);
1940 	if (error)
1941 		return error;
1942 
1943 	if (new) {
1944 		error = nvme_alloc_admin_tag_set(ctrl,
1945 				&to_tcp_ctrl(ctrl)->admin_tag_set,
1946 				&nvme_tcp_admin_mq_ops,
1947 				sizeof(struct nvme_tcp_request));
1948 		if (error)
1949 			goto out_free_queue;
1950 	}
1951 
1952 	error = nvme_tcp_start_queue(ctrl, 0);
1953 	if (error)
1954 		goto out_cleanup_tagset;
1955 
1956 	error = nvme_enable_ctrl(ctrl);
1957 	if (error)
1958 		goto out_stop_queue;
1959 
1960 	nvme_unquiesce_admin_queue(ctrl);
1961 
1962 	error = nvme_init_ctrl_finish(ctrl, false);
1963 	if (error)
1964 		goto out_quiesce_queue;
1965 
1966 	return 0;
1967 
1968 out_quiesce_queue:
1969 	nvme_quiesce_admin_queue(ctrl);
1970 	blk_sync_queue(ctrl->admin_q);
1971 out_stop_queue:
1972 	nvme_tcp_stop_queue(ctrl, 0);
1973 	nvme_cancel_admin_tagset(ctrl);
1974 out_cleanup_tagset:
1975 	if (new)
1976 		nvme_remove_admin_tag_set(ctrl);
1977 out_free_queue:
1978 	nvme_tcp_free_admin_queue(ctrl);
1979 	return error;
1980 }
1981 
1982 static void nvme_tcp_teardown_admin_queue(struct nvme_ctrl *ctrl,
1983 		bool remove)
1984 {
1985 	nvme_quiesce_admin_queue(ctrl);
1986 	blk_sync_queue(ctrl->admin_q);
1987 	nvme_tcp_stop_queue(ctrl, 0);
1988 	nvme_cancel_admin_tagset(ctrl);
1989 	if (remove)
1990 		nvme_unquiesce_admin_queue(ctrl);
1991 	nvme_tcp_destroy_admin_queue(ctrl, remove);
1992 }
1993 
1994 static void nvme_tcp_teardown_io_queues(struct nvme_ctrl *ctrl,
1995 		bool remove)
1996 {
1997 	if (ctrl->queue_count <= 1)
1998 		return;
1999 	nvme_quiesce_admin_queue(ctrl);
2000 	nvme_start_freeze(ctrl);
2001 	nvme_quiesce_io_queues(ctrl);
2002 	nvme_sync_io_queues(ctrl);
2003 	nvme_tcp_stop_io_queues(ctrl);
2004 	nvme_cancel_tagset(ctrl);
2005 	if (remove)
2006 		nvme_unquiesce_io_queues(ctrl);
2007 	nvme_tcp_destroy_io_queues(ctrl, remove);
2008 }
2009 
2010 static void nvme_tcp_reconnect_or_remove(struct nvme_ctrl *ctrl)
2011 {
2012 	/* If we are resetting/deleting then do nothing */
2013 	if (ctrl->state != NVME_CTRL_CONNECTING) {
2014 		WARN_ON_ONCE(ctrl->state == NVME_CTRL_NEW ||
2015 			ctrl->state == NVME_CTRL_LIVE);
2016 		return;
2017 	}
2018 
2019 	if (nvmf_should_reconnect(ctrl)) {
2020 		dev_info(ctrl->device, "Reconnecting in %d seconds...\n",
2021 			ctrl->opts->reconnect_delay);
2022 		queue_delayed_work(nvme_wq, &to_tcp_ctrl(ctrl)->connect_work,
2023 				ctrl->opts->reconnect_delay * HZ);
2024 	} else {
2025 		dev_info(ctrl->device, "Removing controller...\n");
2026 		nvme_delete_ctrl(ctrl);
2027 	}
2028 }
2029 
2030 static int nvme_tcp_setup_ctrl(struct nvme_ctrl *ctrl, bool new)
2031 {
2032 	struct nvmf_ctrl_options *opts = ctrl->opts;
2033 	int ret;
2034 
2035 	ret = nvme_tcp_configure_admin_queue(ctrl, new);
2036 	if (ret)
2037 		return ret;
2038 
2039 	if (ctrl->icdoff) {
2040 		ret = -EOPNOTSUPP;
2041 		dev_err(ctrl->device, "icdoff is not supported!\n");
2042 		goto destroy_admin;
2043 	}
2044 
2045 	if (!nvme_ctrl_sgl_supported(ctrl)) {
2046 		ret = -EOPNOTSUPP;
2047 		dev_err(ctrl->device, "Mandatory sgls are not supported!\n");
2048 		goto destroy_admin;
2049 	}
2050 
2051 	if (opts->queue_size > ctrl->sqsize + 1)
2052 		dev_warn(ctrl->device,
2053 			"queue_size %zu > ctrl sqsize %u, clamping down\n",
2054 			opts->queue_size, ctrl->sqsize + 1);
2055 
2056 	if (ctrl->sqsize + 1 > ctrl->maxcmd) {
2057 		dev_warn(ctrl->device,
2058 			"sqsize %u > ctrl maxcmd %u, clamping down\n",
2059 			ctrl->sqsize + 1, ctrl->maxcmd);
2060 		ctrl->sqsize = ctrl->maxcmd - 1;
2061 	}
2062 
2063 	if (ctrl->queue_count > 1) {
2064 		ret = nvme_tcp_configure_io_queues(ctrl, new);
2065 		if (ret)
2066 			goto destroy_admin;
2067 	}
2068 
2069 	if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_LIVE)) {
2070 		/*
2071 		 * state change failure is ok if we started ctrl delete,
2072 		 * unless we're during creation of a new controller to
2073 		 * avoid races with teardown flow.
2074 		 */
2075 		WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING &&
2076 			     ctrl->state != NVME_CTRL_DELETING_NOIO);
2077 		WARN_ON_ONCE(new);
2078 		ret = -EINVAL;
2079 		goto destroy_io;
2080 	}
2081 
2082 	nvme_start_ctrl(ctrl);
2083 	return 0;
2084 
2085 destroy_io:
2086 	if (ctrl->queue_count > 1) {
2087 		nvme_quiesce_io_queues(ctrl);
2088 		nvme_sync_io_queues(ctrl);
2089 		nvme_tcp_stop_io_queues(ctrl);
2090 		nvme_cancel_tagset(ctrl);
2091 		nvme_tcp_destroy_io_queues(ctrl, new);
2092 	}
2093 destroy_admin:
2094 	nvme_quiesce_admin_queue(ctrl);
2095 	blk_sync_queue(ctrl->admin_q);
2096 	nvme_tcp_stop_queue(ctrl, 0);
2097 	nvme_cancel_admin_tagset(ctrl);
2098 	nvme_tcp_destroy_admin_queue(ctrl, new);
2099 	return ret;
2100 }
2101 
2102 static void nvme_tcp_reconnect_ctrl_work(struct work_struct *work)
2103 {
2104 	struct nvme_tcp_ctrl *tcp_ctrl = container_of(to_delayed_work(work),
2105 			struct nvme_tcp_ctrl, connect_work);
2106 	struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
2107 
2108 	++ctrl->nr_reconnects;
2109 
2110 	if (nvme_tcp_setup_ctrl(ctrl, false))
2111 		goto requeue;
2112 
2113 	dev_info(ctrl->device, "Successfully reconnected (%d attempt)\n",
2114 			ctrl->nr_reconnects);
2115 
2116 	ctrl->nr_reconnects = 0;
2117 
2118 	return;
2119 
2120 requeue:
2121 	dev_info(ctrl->device, "Failed reconnect attempt %d\n",
2122 			ctrl->nr_reconnects);
2123 	nvme_tcp_reconnect_or_remove(ctrl);
2124 }
2125 
2126 static void nvme_tcp_error_recovery_work(struct work_struct *work)
2127 {
2128 	struct nvme_tcp_ctrl *tcp_ctrl = container_of(work,
2129 				struct nvme_tcp_ctrl, err_work);
2130 	struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
2131 
2132 	nvme_stop_keep_alive(ctrl);
2133 	flush_work(&ctrl->async_event_work);
2134 	nvme_tcp_teardown_io_queues(ctrl, false);
2135 	/* unquiesce to fail fast pending requests */
2136 	nvme_unquiesce_io_queues(ctrl);
2137 	nvme_tcp_teardown_admin_queue(ctrl, false);
2138 	nvme_unquiesce_admin_queue(ctrl);
2139 	nvme_auth_stop(ctrl);
2140 
2141 	if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
2142 		/* state change failure is ok if we started ctrl delete */
2143 		WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING &&
2144 			     ctrl->state != NVME_CTRL_DELETING_NOIO);
2145 		return;
2146 	}
2147 
2148 	nvme_tcp_reconnect_or_remove(ctrl);
2149 }
2150 
2151 static void nvme_tcp_teardown_ctrl(struct nvme_ctrl *ctrl, bool shutdown)
2152 {
2153 	nvme_tcp_teardown_io_queues(ctrl, shutdown);
2154 	nvme_quiesce_admin_queue(ctrl);
2155 	nvme_disable_ctrl(ctrl, shutdown);
2156 	nvme_tcp_teardown_admin_queue(ctrl, shutdown);
2157 }
2158 
2159 static void nvme_tcp_delete_ctrl(struct nvme_ctrl *ctrl)
2160 {
2161 	nvme_tcp_teardown_ctrl(ctrl, true);
2162 }
2163 
2164 static void nvme_reset_ctrl_work(struct work_struct *work)
2165 {
2166 	struct nvme_ctrl *ctrl =
2167 		container_of(work, struct nvme_ctrl, reset_work);
2168 
2169 	nvme_stop_ctrl(ctrl);
2170 	nvme_tcp_teardown_ctrl(ctrl, false);
2171 
2172 	if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
2173 		/* state change failure is ok if we started ctrl delete */
2174 		WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING &&
2175 			     ctrl->state != NVME_CTRL_DELETING_NOIO);
2176 		return;
2177 	}
2178 
2179 	if (nvme_tcp_setup_ctrl(ctrl, false))
2180 		goto out_fail;
2181 
2182 	return;
2183 
2184 out_fail:
2185 	++ctrl->nr_reconnects;
2186 	nvme_tcp_reconnect_or_remove(ctrl);
2187 }
2188 
2189 static void nvme_tcp_stop_ctrl(struct nvme_ctrl *ctrl)
2190 {
2191 	flush_work(&to_tcp_ctrl(ctrl)->err_work);
2192 	cancel_delayed_work_sync(&to_tcp_ctrl(ctrl)->connect_work);
2193 }
2194 
2195 static void nvme_tcp_free_ctrl(struct nvme_ctrl *nctrl)
2196 {
2197 	struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
2198 
2199 	if (list_empty(&ctrl->list))
2200 		goto free_ctrl;
2201 
2202 	mutex_lock(&nvme_tcp_ctrl_mutex);
2203 	list_del(&ctrl->list);
2204 	mutex_unlock(&nvme_tcp_ctrl_mutex);
2205 
2206 	nvmf_free_options(nctrl->opts);
2207 free_ctrl:
2208 	kfree(ctrl->queues);
2209 	kfree(ctrl);
2210 }
2211 
2212 static void nvme_tcp_set_sg_null(struct nvme_command *c)
2213 {
2214 	struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2215 
2216 	sg->addr = 0;
2217 	sg->length = 0;
2218 	sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
2219 			NVME_SGL_FMT_TRANSPORT_A;
2220 }
2221 
2222 static void nvme_tcp_set_sg_inline(struct nvme_tcp_queue *queue,
2223 		struct nvme_command *c, u32 data_len)
2224 {
2225 	struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2226 
2227 	sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
2228 	sg->length = cpu_to_le32(data_len);
2229 	sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
2230 }
2231 
2232 static void nvme_tcp_set_sg_host_data(struct nvme_command *c,
2233 		u32 data_len)
2234 {
2235 	struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2236 
2237 	sg->addr = 0;
2238 	sg->length = cpu_to_le32(data_len);
2239 	sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
2240 			NVME_SGL_FMT_TRANSPORT_A;
2241 }
2242 
2243 static void nvme_tcp_submit_async_event(struct nvme_ctrl *arg)
2244 {
2245 	struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(arg);
2246 	struct nvme_tcp_queue *queue = &ctrl->queues[0];
2247 	struct nvme_tcp_cmd_pdu *pdu = ctrl->async_req.pdu;
2248 	struct nvme_command *cmd = &pdu->cmd;
2249 	u8 hdgst = nvme_tcp_hdgst_len(queue);
2250 
2251 	memset(pdu, 0, sizeof(*pdu));
2252 	pdu->hdr.type = nvme_tcp_cmd;
2253 	if (queue->hdr_digest)
2254 		pdu->hdr.flags |= NVME_TCP_F_HDGST;
2255 	pdu->hdr.hlen = sizeof(*pdu);
2256 	pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst);
2257 
2258 	cmd->common.opcode = nvme_admin_async_event;
2259 	cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
2260 	cmd->common.flags |= NVME_CMD_SGL_METABUF;
2261 	nvme_tcp_set_sg_null(cmd);
2262 
2263 	ctrl->async_req.state = NVME_TCP_SEND_CMD_PDU;
2264 	ctrl->async_req.offset = 0;
2265 	ctrl->async_req.curr_bio = NULL;
2266 	ctrl->async_req.data_len = 0;
2267 
2268 	nvme_tcp_queue_request(&ctrl->async_req, true, true);
2269 }
2270 
2271 static void nvme_tcp_complete_timed_out(struct request *rq)
2272 {
2273 	struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2274 	struct nvme_ctrl *ctrl = &req->queue->ctrl->ctrl;
2275 
2276 	nvme_tcp_stop_queue(ctrl, nvme_tcp_queue_id(req->queue));
2277 	nvmf_complete_timed_out_request(rq);
2278 }
2279 
2280 static enum blk_eh_timer_return nvme_tcp_timeout(struct request *rq)
2281 {
2282 	struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2283 	struct nvme_ctrl *ctrl = &req->queue->ctrl->ctrl;
2284 	struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2285 
2286 	dev_warn(ctrl->device,
2287 		"queue %d: timeout request %#x type %d\n",
2288 		nvme_tcp_queue_id(req->queue), rq->tag, pdu->hdr.type);
2289 
2290 	if (ctrl->state != NVME_CTRL_LIVE) {
2291 		/*
2292 		 * If we are resetting, connecting or deleting we should
2293 		 * complete immediately because we may block controller
2294 		 * teardown or setup sequence
2295 		 * - ctrl disable/shutdown fabrics requests
2296 		 * - connect requests
2297 		 * - initialization admin requests
2298 		 * - I/O requests that entered after unquiescing and
2299 		 *   the controller stopped responding
2300 		 *
2301 		 * All other requests should be cancelled by the error
2302 		 * recovery work, so it's fine that we fail it here.
2303 		 */
2304 		nvme_tcp_complete_timed_out(rq);
2305 		return BLK_EH_DONE;
2306 	}
2307 
2308 	/*
2309 	 * LIVE state should trigger the normal error recovery which will
2310 	 * handle completing this request.
2311 	 */
2312 	nvme_tcp_error_recovery(ctrl);
2313 	return BLK_EH_RESET_TIMER;
2314 }
2315 
2316 static blk_status_t nvme_tcp_map_data(struct nvme_tcp_queue *queue,
2317 			struct request *rq)
2318 {
2319 	struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2320 	struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2321 	struct nvme_command *c = &pdu->cmd;
2322 
2323 	c->common.flags |= NVME_CMD_SGL_METABUF;
2324 
2325 	if (!blk_rq_nr_phys_segments(rq))
2326 		nvme_tcp_set_sg_null(c);
2327 	else if (rq_data_dir(rq) == WRITE &&
2328 	    req->data_len <= nvme_tcp_inline_data_size(req))
2329 		nvme_tcp_set_sg_inline(queue, c, req->data_len);
2330 	else
2331 		nvme_tcp_set_sg_host_data(c, req->data_len);
2332 
2333 	return 0;
2334 }
2335 
2336 static blk_status_t nvme_tcp_setup_cmd_pdu(struct nvme_ns *ns,
2337 		struct request *rq)
2338 {
2339 	struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2340 	struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2341 	struct nvme_tcp_queue *queue = req->queue;
2342 	u8 hdgst = nvme_tcp_hdgst_len(queue), ddgst = 0;
2343 	blk_status_t ret;
2344 
2345 	ret = nvme_setup_cmd(ns, rq);
2346 	if (ret)
2347 		return ret;
2348 
2349 	req->state = NVME_TCP_SEND_CMD_PDU;
2350 	req->status = cpu_to_le16(NVME_SC_SUCCESS);
2351 	req->offset = 0;
2352 	req->data_sent = 0;
2353 	req->pdu_len = 0;
2354 	req->pdu_sent = 0;
2355 	req->h2cdata_left = 0;
2356 	req->data_len = blk_rq_nr_phys_segments(rq) ?
2357 				blk_rq_payload_bytes(rq) : 0;
2358 	req->curr_bio = rq->bio;
2359 	if (req->curr_bio && req->data_len)
2360 		nvme_tcp_init_iter(req, rq_data_dir(rq));
2361 
2362 	if (rq_data_dir(rq) == WRITE &&
2363 	    req->data_len <= nvme_tcp_inline_data_size(req))
2364 		req->pdu_len = req->data_len;
2365 
2366 	pdu->hdr.type = nvme_tcp_cmd;
2367 	pdu->hdr.flags = 0;
2368 	if (queue->hdr_digest)
2369 		pdu->hdr.flags |= NVME_TCP_F_HDGST;
2370 	if (queue->data_digest && req->pdu_len) {
2371 		pdu->hdr.flags |= NVME_TCP_F_DDGST;
2372 		ddgst = nvme_tcp_ddgst_len(queue);
2373 	}
2374 	pdu->hdr.hlen = sizeof(*pdu);
2375 	pdu->hdr.pdo = req->pdu_len ? pdu->hdr.hlen + hdgst : 0;
2376 	pdu->hdr.plen =
2377 		cpu_to_le32(pdu->hdr.hlen + hdgst + req->pdu_len + ddgst);
2378 
2379 	ret = nvme_tcp_map_data(queue, rq);
2380 	if (unlikely(ret)) {
2381 		nvme_cleanup_cmd(rq);
2382 		dev_err(queue->ctrl->ctrl.device,
2383 			"Failed to map data (%d)\n", ret);
2384 		return ret;
2385 	}
2386 
2387 	return 0;
2388 }
2389 
2390 static void nvme_tcp_commit_rqs(struct blk_mq_hw_ctx *hctx)
2391 {
2392 	struct nvme_tcp_queue *queue = hctx->driver_data;
2393 
2394 	if (!llist_empty(&queue->req_list))
2395 		queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
2396 }
2397 
2398 static blk_status_t nvme_tcp_queue_rq(struct blk_mq_hw_ctx *hctx,
2399 		const struct blk_mq_queue_data *bd)
2400 {
2401 	struct nvme_ns *ns = hctx->queue->queuedata;
2402 	struct nvme_tcp_queue *queue = hctx->driver_data;
2403 	struct request *rq = bd->rq;
2404 	struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2405 	bool queue_ready = test_bit(NVME_TCP_Q_LIVE, &queue->flags);
2406 	blk_status_t ret;
2407 
2408 	if (!nvme_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
2409 		return nvme_fail_nonready_command(&queue->ctrl->ctrl, rq);
2410 
2411 	ret = nvme_tcp_setup_cmd_pdu(ns, rq);
2412 	if (unlikely(ret))
2413 		return ret;
2414 
2415 	nvme_start_request(rq);
2416 
2417 	nvme_tcp_queue_request(req, true, bd->last);
2418 
2419 	return BLK_STS_OK;
2420 }
2421 
2422 static void nvme_tcp_map_queues(struct blk_mq_tag_set *set)
2423 {
2424 	struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(set->driver_data);
2425 	struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
2426 
2427 	if (opts->nr_write_queues && ctrl->io_queues[HCTX_TYPE_READ]) {
2428 		/* separate read/write queues */
2429 		set->map[HCTX_TYPE_DEFAULT].nr_queues =
2430 			ctrl->io_queues[HCTX_TYPE_DEFAULT];
2431 		set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2432 		set->map[HCTX_TYPE_READ].nr_queues =
2433 			ctrl->io_queues[HCTX_TYPE_READ];
2434 		set->map[HCTX_TYPE_READ].queue_offset =
2435 			ctrl->io_queues[HCTX_TYPE_DEFAULT];
2436 	} else {
2437 		/* shared read/write queues */
2438 		set->map[HCTX_TYPE_DEFAULT].nr_queues =
2439 			ctrl->io_queues[HCTX_TYPE_DEFAULT];
2440 		set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2441 		set->map[HCTX_TYPE_READ].nr_queues =
2442 			ctrl->io_queues[HCTX_TYPE_DEFAULT];
2443 		set->map[HCTX_TYPE_READ].queue_offset = 0;
2444 	}
2445 	blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
2446 	blk_mq_map_queues(&set->map[HCTX_TYPE_READ]);
2447 
2448 	if (opts->nr_poll_queues && ctrl->io_queues[HCTX_TYPE_POLL]) {
2449 		/* map dedicated poll queues only if we have queues left */
2450 		set->map[HCTX_TYPE_POLL].nr_queues =
2451 				ctrl->io_queues[HCTX_TYPE_POLL];
2452 		set->map[HCTX_TYPE_POLL].queue_offset =
2453 			ctrl->io_queues[HCTX_TYPE_DEFAULT] +
2454 			ctrl->io_queues[HCTX_TYPE_READ];
2455 		blk_mq_map_queues(&set->map[HCTX_TYPE_POLL]);
2456 	}
2457 
2458 	dev_info(ctrl->ctrl.device,
2459 		"mapped %d/%d/%d default/read/poll queues.\n",
2460 		ctrl->io_queues[HCTX_TYPE_DEFAULT],
2461 		ctrl->io_queues[HCTX_TYPE_READ],
2462 		ctrl->io_queues[HCTX_TYPE_POLL]);
2463 }
2464 
2465 static int nvme_tcp_poll(struct blk_mq_hw_ctx *hctx, struct io_comp_batch *iob)
2466 {
2467 	struct nvme_tcp_queue *queue = hctx->driver_data;
2468 	struct sock *sk = queue->sock->sk;
2469 
2470 	if (!test_bit(NVME_TCP_Q_LIVE, &queue->flags))
2471 		return 0;
2472 
2473 	set_bit(NVME_TCP_Q_POLLING, &queue->flags);
2474 	if (sk_can_busy_loop(sk) && skb_queue_empty_lockless(&sk->sk_receive_queue))
2475 		sk_busy_loop(sk, true);
2476 	nvme_tcp_try_recv(queue);
2477 	clear_bit(NVME_TCP_Q_POLLING, &queue->flags);
2478 	return queue->nr_cqe;
2479 }
2480 
2481 static int nvme_tcp_get_address(struct nvme_ctrl *ctrl, char *buf, int size)
2482 {
2483 	struct nvme_tcp_queue *queue = &to_tcp_ctrl(ctrl)->queues[0];
2484 	struct sockaddr_storage src_addr;
2485 	int ret, len;
2486 
2487 	len = nvmf_get_address(ctrl, buf, size);
2488 
2489 	ret = kernel_getsockname(queue->sock, (struct sockaddr *)&src_addr);
2490 	if (ret > 0) {
2491 		if (len > 0)
2492 			len--; /* strip trailing newline */
2493 		len += scnprintf(buf + len, size - len, "%ssrc_addr=%pISc\n",
2494 				(len) ? "," : "", &src_addr);
2495 	}
2496 
2497 	return len;
2498 }
2499 
2500 static const struct blk_mq_ops nvme_tcp_mq_ops = {
2501 	.queue_rq	= nvme_tcp_queue_rq,
2502 	.commit_rqs	= nvme_tcp_commit_rqs,
2503 	.complete	= nvme_complete_rq,
2504 	.init_request	= nvme_tcp_init_request,
2505 	.exit_request	= nvme_tcp_exit_request,
2506 	.init_hctx	= nvme_tcp_init_hctx,
2507 	.timeout	= nvme_tcp_timeout,
2508 	.map_queues	= nvme_tcp_map_queues,
2509 	.poll		= nvme_tcp_poll,
2510 };
2511 
2512 static const struct blk_mq_ops nvme_tcp_admin_mq_ops = {
2513 	.queue_rq	= nvme_tcp_queue_rq,
2514 	.complete	= nvme_complete_rq,
2515 	.init_request	= nvme_tcp_init_request,
2516 	.exit_request	= nvme_tcp_exit_request,
2517 	.init_hctx	= nvme_tcp_init_admin_hctx,
2518 	.timeout	= nvme_tcp_timeout,
2519 };
2520 
2521 static const struct nvme_ctrl_ops nvme_tcp_ctrl_ops = {
2522 	.name			= "tcp",
2523 	.module			= THIS_MODULE,
2524 	.flags			= NVME_F_FABRICS | NVME_F_BLOCKING,
2525 	.reg_read32		= nvmf_reg_read32,
2526 	.reg_read64		= nvmf_reg_read64,
2527 	.reg_write32		= nvmf_reg_write32,
2528 	.free_ctrl		= nvme_tcp_free_ctrl,
2529 	.submit_async_event	= nvme_tcp_submit_async_event,
2530 	.delete_ctrl		= nvme_tcp_delete_ctrl,
2531 	.get_address		= nvme_tcp_get_address,
2532 	.stop_ctrl		= nvme_tcp_stop_ctrl,
2533 };
2534 
2535 static bool
2536 nvme_tcp_existing_controller(struct nvmf_ctrl_options *opts)
2537 {
2538 	struct nvme_tcp_ctrl *ctrl;
2539 	bool found = false;
2540 
2541 	mutex_lock(&nvme_tcp_ctrl_mutex);
2542 	list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list) {
2543 		found = nvmf_ip_options_match(&ctrl->ctrl, opts);
2544 		if (found)
2545 			break;
2546 	}
2547 	mutex_unlock(&nvme_tcp_ctrl_mutex);
2548 
2549 	return found;
2550 }
2551 
2552 static struct nvme_ctrl *nvme_tcp_create_ctrl(struct device *dev,
2553 		struct nvmf_ctrl_options *opts)
2554 {
2555 	struct nvme_tcp_ctrl *ctrl;
2556 	int ret;
2557 
2558 	ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
2559 	if (!ctrl)
2560 		return ERR_PTR(-ENOMEM);
2561 
2562 	INIT_LIST_HEAD(&ctrl->list);
2563 	ctrl->ctrl.opts = opts;
2564 	ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues +
2565 				opts->nr_poll_queues + 1;
2566 	ctrl->ctrl.sqsize = opts->queue_size - 1;
2567 	ctrl->ctrl.kato = opts->kato;
2568 
2569 	INIT_DELAYED_WORK(&ctrl->connect_work,
2570 			nvme_tcp_reconnect_ctrl_work);
2571 	INIT_WORK(&ctrl->err_work, nvme_tcp_error_recovery_work);
2572 	INIT_WORK(&ctrl->ctrl.reset_work, nvme_reset_ctrl_work);
2573 
2574 	if (!(opts->mask & NVMF_OPT_TRSVCID)) {
2575 		opts->trsvcid =
2576 			kstrdup(__stringify(NVME_TCP_DISC_PORT), GFP_KERNEL);
2577 		if (!opts->trsvcid) {
2578 			ret = -ENOMEM;
2579 			goto out_free_ctrl;
2580 		}
2581 		opts->mask |= NVMF_OPT_TRSVCID;
2582 	}
2583 
2584 	ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2585 			opts->traddr, opts->trsvcid, &ctrl->addr);
2586 	if (ret) {
2587 		pr_err("malformed address passed: %s:%s\n",
2588 			opts->traddr, opts->trsvcid);
2589 		goto out_free_ctrl;
2590 	}
2591 
2592 	if (opts->mask & NVMF_OPT_HOST_TRADDR) {
2593 		ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2594 			opts->host_traddr, NULL, &ctrl->src_addr);
2595 		if (ret) {
2596 			pr_err("malformed src address passed: %s\n",
2597 			       opts->host_traddr);
2598 			goto out_free_ctrl;
2599 		}
2600 	}
2601 
2602 	if (opts->mask & NVMF_OPT_HOST_IFACE) {
2603 		if (!__dev_get_by_name(&init_net, opts->host_iface)) {
2604 			pr_err("invalid interface passed: %s\n",
2605 			       opts->host_iface);
2606 			ret = -ENODEV;
2607 			goto out_free_ctrl;
2608 		}
2609 	}
2610 
2611 	if (!opts->duplicate_connect && nvme_tcp_existing_controller(opts)) {
2612 		ret = -EALREADY;
2613 		goto out_free_ctrl;
2614 	}
2615 
2616 	ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
2617 				GFP_KERNEL);
2618 	if (!ctrl->queues) {
2619 		ret = -ENOMEM;
2620 		goto out_free_ctrl;
2621 	}
2622 
2623 	ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_tcp_ctrl_ops, 0);
2624 	if (ret)
2625 		goto out_kfree_queues;
2626 
2627 	if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
2628 		WARN_ON_ONCE(1);
2629 		ret = -EINTR;
2630 		goto out_uninit_ctrl;
2631 	}
2632 
2633 	ret = nvme_tcp_setup_ctrl(&ctrl->ctrl, true);
2634 	if (ret)
2635 		goto out_uninit_ctrl;
2636 
2637 	dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISp\n",
2638 		nvmf_ctrl_subsysnqn(&ctrl->ctrl), &ctrl->addr);
2639 
2640 	mutex_lock(&nvme_tcp_ctrl_mutex);
2641 	list_add_tail(&ctrl->list, &nvme_tcp_ctrl_list);
2642 	mutex_unlock(&nvme_tcp_ctrl_mutex);
2643 
2644 	return &ctrl->ctrl;
2645 
2646 out_uninit_ctrl:
2647 	nvme_uninit_ctrl(&ctrl->ctrl);
2648 	nvme_put_ctrl(&ctrl->ctrl);
2649 	if (ret > 0)
2650 		ret = -EIO;
2651 	return ERR_PTR(ret);
2652 out_kfree_queues:
2653 	kfree(ctrl->queues);
2654 out_free_ctrl:
2655 	kfree(ctrl);
2656 	return ERR_PTR(ret);
2657 }
2658 
2659 static struct nvmf_transport_ops nvme_tcp_transport = {
2660 	.name		= "tcp",
2661 	.module		= THIS_MODULE,
2662 	.required_opts	= NVMF_OPT_TRADDR,
2663 	.allowed_opts	= NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2664 			  NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
2665 			  NVMF_OPT_HDR_DIGEST | NVMF_OPT_DATA_DIGEST |
2666 			  NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES |
2667 			  NVMF_OPT_TOS | NVMF_OPT_HOST_IFACE,
2668 	.create_ctrl	= nvme_tcp_create_ctrl,
2669 };
2670 
2671 static int __init nvme_tcp_init_module(void)
2672 {
2673 	nvme_tcp_wq = alloc_workqueue("nvme_tcp_wq",
2674 			WQ_MEM_RECLAIM | WQ_HIGHPRI, 0);
2675 	if (!nvme_tcp_wq)
2676 		return -ENOMEM;
2677 
2678 	nvmf_register_transport(&nvme_tcp_transport);
2679 	return 0;
2680 }
2681 
2682 static void __exit nvme_tcp_cleanup_module(void)
2683 {
2684 	struct nvme_tcp_ctrl *ctrl;
2685 
2686 	nvmf_unregister_transport(&nvme_tcp_transport);
2687 
2688 	mutex_lock(&nvme_tcp_ctrl_mutex);
2689 	list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list)
2690 		nvme_delete_ctrl(&ctrl->ctrl);
2691 	mutex_unlock(&nvme_tcp_ctrl_mutex);
2692 	flush_workqueue(nvme_delete_wq);
2693 
2694 	destroy_workqueue(nvme_tcp_wq);
2695 }
2696 
2697 module_init(nvme_tcp_init_module);
2698 module_exit(nvme_tcp_cleanup_module);
2699 
2700 MODULE_LICENSE("GPL v2");
2701