xref: /openbmc/linux/drivers/nvme/target/tcp.c (revision 94b00cd6)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * NVMe over Fabrics TCP target.
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/inet.h>
15 #include <linux/llist.h>
16 #include <crypto/hash.h>
17 #include <trace/events/sock.h>
18 
19 #include "nvmet.h"
20 
21 #define NVMET_TCP_DEF_INLINE_DATA_SIZE	(4 * PAGE_SIZE)
22 #define NVMET_TCP_MAXH2CDATA		0x400000 /* 16M arbitrary limit */
23 
24 static int param_store_val(const char *str, int *val, int min, int max)
25 {
26 	int ret, new_val;
27 
28 	ret = kstrtoint(str, 10, &new_val);
29 	if (ret)
30 		return -EINVAL;
31 
32 	if (new_val < min || new_val > max)
33 		return -EINVAL;
34 
35 	*val = new_val;
36 	return 0;
37 }
38 
39 static int set_params(const char *str, const struct kernel_param *kp)
40 {
41 	return param_store_val(str, kp->arg, 0, INT_MAX);
42 }
43 
44 static const struct kernel_param_ops set_param_ops = {
45 	.set	= set_params,
46 	.get	= param_get_int,
47 };
48 
49 /* Define the socket priority to use for connections were it is desirable
50  * that the NIC consider performing optimized packet processing or filtering.
51  * A non-zero value being sufficient to indicate general consideration of any
52  * possible optimization.  Making it a module param allows for alternative
53  * values that may be unique for some NIC implementations.
54  */
55 static int so_priority;
56 device_param_cb(so_priority, &set_param_ops, &so_priority, 0644);
57 MODULE_PARM_DESC(so_priority, "nvmet tcp socket optimize priority: Default 0");
58 
59 /* Define a time period (in usecs) that io_work() shall sample an activated
60  * queue before determining it to be idle.  This optional module behavior
61  * can enable NIC solutions that support socket optimized packet processing
62  * using advanced interrupt moderation techniques.
63  */
64 static int idle_poll_period_usecs;
65 device_param_cb(idle_poll_period_usecs, &set_param_ops,
66 		&idle_poll_period_usecs, 0644);
67 MODULE_PARM_DESC(idle_poll_period_usecs,
68 		"nvmet tcp io_work poll till idle time period in usecs: Default 0");
69 
70 #define NVMET_TCP_RECV_BUDGET		8
71 #define NVMET_TCP_SEND_BUDGET		8
72 #define NVMET_TCP_IO_WORK_BUDGET	64
73 
74 enum nvmet_tcp_send_state {
75 	NVMET_TCP_SEND_DATA_PDU,
76 	NVMET_TCP_SEND_DATA,
77 	NVMET_TCP_SEND_R2T,
78 	NVMET_TCP_SEND_DDGST,
79 	NVMET_TCP_SEND_RESPONSE
80 };
81 
82 enum nvmet_tcp_recv_state {
83 	NVMET_TCP_RECV_PDU,
84 	NVMET_TCP_RECV_DATA,
85 	NVMET_TCP_RECV_DDGST,
86 	NVMET_TCP_RECV_ERR,
87 };
88 
89 enum {
90 	NVMET_TCP_F_INIT_FAILED = (1 << 0),
91 };
92 
93 struct nvmet_tcp_cmd {
94 	struct nvmet_tcp_queue		*queue;
95 	struct nvmet_req		req;
96 
97 	struct nvme_tcp_cmd_pdu		*cmd_pdu;
98 	struct nvme_tcp_rsp_pdu		*rsp_pdu;
99 	struct nvme_tcp_data_pdu	*data_pdu;
100 	struct nvme_tcp_r2t_pdu		*r2t_pdu;
101 
102 	u32				rbytes_done;
103 	u32				wbytes_done;
104 
105 	u32				pdu_len;
106 	u32				pdu_recv;
107 	int				sg_idx;
108 	struct msghdr			recv_msg;
109 	struct bio_vec			*iov;
110 	u32				flags;
111 
112 	struct list_head		entry;
113 	struct llist_node		lentry;
114 
115 	/* send state */
116 	u32				offset;
117 	struct scatterlist		*cur_sg;
118 	enum nvmet_tcp_send_state	state;
119 
120 	__le32				exp_ddgst;
121 	__le32				recv_ddgst;
122 };
123 
124 enum nvmet_tcp_queue_state {
125 	NVMET_TCP_Q_CONNECTING,
126 	NVMET_TCP_Q_LIVE,
127 	NVMET_TCP_Q_DISCONNECTING,
128 };
129 
130 struct nvmet_tcp_queue {
131 	struct socket		*sock;
132 	struct nvmet_tcp_port	*port;
133 	struct work_struct	io_work;
134 	struct nvmet_cq		nvme_cq;
135 	struct nvmet_sq		nvme_sq;
136 
137 	/* send state */
138 	struct nvmet_tcp_cmd	*cmds;
139 	unsigned int		nr_cmds;
140 	struct list_head	free_list;
141 	struct llist_head	resp_list;
142 	struct list_head	resp_send_list;
143 	int			send_list_len;
144 	struct nvmet_tcp_cmd	*snd_cmd;
145 
146 	/* recv state */
147 	int			offset;
148 	int			left;
149 	enum nvmet_tcp_recv_state rcv_state;
150 	struct nvmet_tcp_cmd	*cmd;
151 	union nvme_tcp_pdu	pdu;
152 
153 	/* digest state */
154 	bool			hdr_digest;
155 	bool			data_digest;
156 	struct ahash_request	*snd_hash;
157 	struct ahash_request	*rcv_hash;
158 
159 	unsigned long           poll_end;
160 
161 	spinlock_t		state_lock;
162 	enum nvmet_tcp_queue_state state;
163 
164 	struct sockaddr_storage	sockaddr;
165 	struct sockaddr_storage	sockaddr_peer;
166 	struct work_struct	release_work;
167 
168 	int			idx;
169 	struct list_head	queue_list;
170 
171 	struct nvmet_tcp_cmd	connect;
172 
173 	struct page_frag_cache	pf_cache;
174 
175 	void (*data_ready)(struct sock *);
176 	void (*state_change)(struct sock *);
177 	void (*write_space)(struct sock *);
178 };
179 
180 struct nvmet_tcp_port {
181 	struct socket		*sock;
182 	struct work_struct	accept_work;
183 	struct nvmet_port	*nport;
184 	struct sockaddr_storage addr;
185 	void (*data_ready)(struct sock *);
186 };
187 
188 static DEFINE_IDA(nvmet_tcp_queue_ida);
189 static LIST_HEAD(nvmet_tcp_queue_list);
190 static DEFINE_MUTEX(nvmet_tcp_queue_mutex);
191 
192 static struct workqueue_struct *nvmet_tcp_wq;
193 static const struct nvmet_fabrics_ops nvmet_tcp_ops;
194 static void nvmet_tcp_free_cmd(struct nvmet_tcp_cmd *c);
195 static void nvmet_tcp_free_cmd_buffers(struct nvmet_tcp_cmd *cmd);
196 
197 static inline u16 nvmet_tcp_cmd_tag(struct nvmet_tcp_queue *queue,
198 		struct nvmet_tcp_cmd *cmd)
199 {
200 	if (unlikely(!queue->nr_cmds)) {
201 		/* We didn't allocate cmds yet, send 0xffff */
202 		return USHRT_MAX;
203 	}
204 
205 	return cmd - queue->cmds;
206 }
207 
208 static inline bool nvmet_tcp_has_data_in(struct nvmet_tcp_cmd *cmd)
209 {
210 	return nvme_is_write(cmd->req.cmd) &&
211 		cmd->rbytes_done < cmd->req.transfer_len;
212 }
213 
214 static inline bool nvmet_tcp_need_data_in(struct nvmet_tcp_cmd *cmd)
215 {
216 	return nvmet_tcp_has_data_in(cmd) && !cmd->req.cqe->status;
217 }
218 
219 static inline bool nvmet_tcp_need_data_out(struct nvmet_tcp_cmd *cmd)
220 {
221 	return !nvme_is_write(cmd->req.cmd) &&
222 		cmd->req.transfer_len > 0 &&
223 		!cmd->req.cqe->status;
224 }
225 
226 static inline bool nvmet_tcp_has_inline_data(struct nvmet_tcp_cmd *cmd)
227 {
228 	return nvme_is_write(cmd->req.cmd) && cmd->pdu_len &&
229 		!cmd->rbytes_done;
230 }
231 
232 static inline struct nvmet_tcp_cmd *
233 nvmet_tcp_get_cmd(struct nvmet_tcp_queue *queue)
234 {
235 	struct nvmet_tcp_cmd *cmd;
236 
237 	cmd = list_first_entry_or_null(&queue->free_list,
238 				struct nvmet_tcp_cmd, entry);
239 	if (!cmd)
240 		return NULL;
241 	list_del_init(&cmd->entry);
242 
243 	cmd->rbytes_done = cmd->wbytes_done = 0;
244 	cmd->pdu_len = 0;
245 	cmd->pdu_recv = 0;
246 	cmd->iov = NULL;
247 	cmd->flags = 0;
248 	return cmd;
249 }
250 
251 static inline void nvmet_tcp_put_cmd(struct nvmet_tcp_cmd *cmd)
252 {
253 	if (unlikely(cmd == &cmd->queue->connect))
254 		return;
255 
256 	list_add_tail(&cmd->entry, &cmd->queue->free_list);
257 }
258 
259 static inline int queue_cpu(struct nvmet_tcp_queue *queue)
260 {
261 	return queue->sock->sk->sk_incoming_cpu;
262 }
263 
264 static inline u8 nvmet_tcp_hdgst_len(struct nvmet_tcp_queue *queue)
265 {
266 	return queue->hdr_digest ? NVME_TCP_DIGEST_LENGTH : 0;
267 }
268 
269 static inline u8 nvmet_tcp_ddgst_len(struct nvmet_tcp_queue *queue)
270 {
271 	return queue->data_digest ? NVME_TCP_DIGEST_LENGTH : 0;
272 }
273 
274 static inline void nvmet_tcp_hdgst(struct ahash_request *hash,
275 		void *pdu, size_t len)
276 {
277 	struct scatterlist sg;
278 
279 	sg_init_one(&sg, pdu, len);
280 	ahash_request_set_crypt(hash, &sg, pdu + len, len);
281 	crypto_ahash_digest(hash);
282 }
283 
284 static int nvmet_tcp_verify_hdgst(struct nvmet_tcp_queue *queue,
285 	void *pdu, size_t len)
286 {
287 	struct nvme_tcp_hdr *hdr = pdu;
288 	__le32 recv_digest;
289 	__le32 exp_digest;
290 
291 	if (unlikely(!(hdr->flags & NVME_TCP_F_HDGST))) {
292 		pr_err("queue %d: header digest enabled but no header digest\n",
293 			queue->idx);
294 		return -EPROTO;
295 	}
296 
297 	recv_digest = *(__le32 *)(pdu + hdr->hlen);
298 	nvmet_tcp_hdgst(queue->rcv_hash, pdu, len);
299 	exp_digest = *(__le32 *)(pdu + hdr->hlen);
300 	if (recv_digest != exp_digest) {
301 		pr_err("queue %d: header digest error: recv %#x expected %#x\n",
302 			queue->idx, le32_to_cpu(recv_digest),
303 			le32_to_cpu(exp_digest));
304 		return -EPROTO;
305 	}
306 
307 	return 0;
308 }
309 
310 static int nvmet_tcp_check_ddgst(struct nvmet_tcp_queue *queue, void *pdu)
311 {
312 	struct nvme_tcp_hdr *hdr = pdu;
313 	u8 digest_len = nvmet_tcp_hdgst_len(queue);
314 	u32 len;
315 
316 	len = le32_to_cpu(hdr->plen) - hdr->hlen -
317 		(hdr->flags & NVME_TCP_F_HDGST ? digest_len : 0);
318 
319 	if (unlikely(len && !(hdr->flags & NVME_TCP_F_DDGST))) {
320 		pr_err("queue %d: data digest flag is cleared\n", queue->idx);
321 		return -EPROTO;
322 	}
323 
324 	return 0;
325 }
326 
327 static void nvmet_tcp_free_cmd_buffers(struct nvmet_tcp_cmd *cmd)
328 {
329 	kfree(cmd->iov);
330 	sgl_free(cmd->req.sg);
331 	cmd->iov = NULL;
332 	cmd->req.sg = NULL;
333 }
334 
335 static void nvmet_tcp_build_pdu_iovec(struct nvmet_tcp_cmd *cmd)
336 {
337 	struct bio_vec *iov = cmd->iov;
338 	struct scatterlist *sg;
339 	u32 length, offset, sg_offset;
340 	int nr_pages;
341 
342 	length = cmd->pdu_len;
343 	nr_pages = DIV_ROUND_UP(length, PAGE_SIZE);
344 	offset = cmd->rbytes_done;
345 	cmd->sg_idx = offset / PAGE_SIZE;
346 	sg_offset = offset % PAGE_SIZE;
347 	sg = &cmd->req.sg[cmd->sg_idx];
348 
349 	while (length) {
350 		u32 iov_len = min_t(u32, length, sg->length - sg_offset);
351 
352 		bvec_set_page(iov, sg_page(sg), iov_len,
353 				sg->offset + sg_offset);
354 
355 		length -= iov_len;
356 		sg = sg_next(sg);
357 		iov++;
358 		sg_offset = 0;
359 	}
360 
361 	iov_iter_bvec(&cmd->recv_msg.msg_iter, ITER_DEST, cmd->iov,
362 		      nr_pages, cmd->pdu_len);
363 }
364 
365 static void nvmet_tcp_fatal_error(struct nvmet_tcp_queue *queue)
366 {
367 	queue->rcv_state = NVMET_TCP_RECV_ERR;
368 	if (queue->nvme_sq.ctrl)
369 		nvmet_ctrl_fatal_error(queue->nvme_sq.ctrl);
370 	else
371 		kernel_sock_shutdown(queue->sock, SHUT_RDWR);
372 }
373 
374 static void nvmet_tcp_socket_error(struct nvmet_tcp_queue *queue, int status)
375 {
376 	queue->rcv_state = NVMET_TCP_RECV_ERR;
377 	if (status == -EPIPE || status == -ECONNRESET)
378 		kernel_sock_shutdown(queue->sock, SHUT_RDWR);
379 	else
380 		nvmet_tcp_fatal_error(queue);
381 }
382 
383 static int nvmet_tcp_map_data(struct nvmet_tcp_cmd *cmd)
384 {
385 	struct nvme_sgl_desc *sgl = &cmd->req.cmd->common.dptr.sgl;
386 	u32 len = le32_to_cpu(sgl->length);
387 
388 	if (!len)
389 		return 0;
390 
391 	if (sgl->type == ((NVME_SGL_FMT_DATA_DESC << 4) |
392 			  NVME_SGL_FMT_OFFSET)) {
393 		if (!nvme_is_write(cmd->req.cmd))
394 			return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
395 
396 		if (len > cmd->req.port->inline_data_size)
397 			return NVME_SC_SGL_INVALID_OFFSET | NVME_SC_DNR;
398 		cmd->pdu_len = len;
399 	}
400 	cmd->req.transfer_len += len;
401 
402 	cmd->req.sg = sgl_alloc(len, GFP_KERNEL, &cmd->req.sg_cnt);
403 	if (!cmd->req.sg)
404 		return NVME_SC_INTERNAL;
405 	cmd->cur_sg = cmd->req.sg;
406 
407 	if (nvmet_tcp_has_data_in(cmd)) {
408 		cmd->iov = kmalloc_array(cmd->req.sg_cnt,
409 				sizeof(*cmd->iov), GFP_KERNEL);
410 		if (!cmd->iov)
411 			goto err;
412 	}
413 
414 	return 0;
415 err:
416 	nvmet_tcp_free_cmd_buffers(cmd);
417 	return NVME_SC_INTERNAL;
418 }
419 
420 static void nvmet_tcp_calc_ddgst(struct ahash_request *hash,
421 		struct nvmet_tcp_cmd *cmd)
422 {
423 	ahash_request_set_crypt(hash, cmd->req.sg,
424 		(void *)&cmd->exp_ddgst, cmd->req.transfer_len);
425 	crypto_ahash_digest(hash);
426 }
427 
428 static void nvmet_setup_c2h_data_pdu(struct nvmet_tcp_cmd *cmd)
429 {
430 	struct nvme_tcp_data_pdu *pdu = cmd->data_pdu;
431 	struct nvmet_tcp_queue *queue = cmd->queue;
432 	u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
433 	u8 ddgst = nvmet_tcp_ddgst_len(cmd->queue);
434 
435 	cmd->offset = 0;
436 	cmd->state = NVMET_TCP_SEND_DATA_PDU;
437 
438 	pdu->hdr.type = nvme_tcp_c2h_data;
439 	pdu->hdr.flags = NVME_TCP_F_DATA_LAST | (queue->nvme_sq.sqhd_disabled ?
440 						NVME_TCP_F_DATA_SUCCESS : 0);
441 	pdu->hdr.hlen = sizeof(*pdu);
442 	pdu->hdr.pdo = pdu->hdr.hlen + hdgst;
443 	pdu->hdr.plen =
444 		cpu_to_le32(pdu->hdr.hlen + hdgst +
445 				cmd->req.transfer_len + ddgst);
446 	pdu->command_id = cmd->req.cqe->command_id;
447 	pdu->data_length = cpu_to_le32(cmd->req.transfer_len);
448 	pdu->data_offset = cpu_to_le32(cmd->wbytes_done);
449 
450 	if (queue->data_digest) {
451 		pdu->hdr.flags |= NVME_TCP_F_DDGST;
452 		nvmet_tcp_calc_ddgst(queue->snd_hash, cmd);
453 	}
454 
455 	if (cmd->queue->hdr_digest) {
456 		pdu->hdr.flags |= NVME_TCP_F_HDGST;
457 		nvmet_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
458 	}
459 }
460 
461 static void nvmet_setup_r2t_pdu(struct nvmet_tcp_cmd *cmd)
462 {
463 	struct nvme_tcp_r2t_pdu *pdu = cmd->r2t_pdu;
464 	struct nvmet_tcp_queue *queue = cmd->queue;
465 	u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
466 
467 	cmd->offset = 0;
468 	cmd->state = NVMET_TCP_SEND_R2T;
469 
470 	pdu->hdr.type = nvme_tcp_r2t;
471 	pdu->hdr.flags = 0;
472 	pdu->hdr.hlen = sizeof(*pdu);
473 	pdu->hdr.pdo = 0;
474 	pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst);
475 
476 	pdu->command_id = cmd->req.cmd->common.command_id;
477 	pdu->ttag = nvmet_tcp_cmd_tag(cmd->queue, cmd);
478 	pdu->r2t_length = cpu_to_le32(cmd->req.transfer_len - cmd->rbytes_done);
479 	pdu->r2t_offset = cpu_to_le32(cmd->rbytes_done);
480 	if (cmd->queue->hdr_digest) {
481 		pdu->hdr.flags |= NVME_TCP_F_HDGST;
482 		nvmet_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
483 	}
484 }
485 
486 static void nvmet_setup_response_pdu(struct nvmet_tcp_cmd *cmd)
487 {
488 	struct nvme_tcp_rsp_pdu *pdu = cmd->rsp_pdu;
489 	struct nvmet_tcp_queue *queue = cmd->queue;
490 	u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
491 
492 	cmd->offset = 0;
493 	cmd->state = NVMET_TCP_SEND_RESPONSE;
494 
495 	pdu->hdr.type = nvme_tcp_rsp;
496 	pdu->hdr.flags = 0;
497 	pdu->hdr.hlen = sizeof(*pdu);
498 	pdu->hdr.pdo = 0;
499 	pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst);
500 	if (cmd->queue->hdr_digest) {
501 		pdu->hdr.flags |= NVME_TCP_F_HDGST;
502 		nvmet_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
503 	}
504 }
505 
506 static void nvmet_tcp_process_resp_list(struct nvmet_tcp_queue *queue)
507 {
508 	struct llist_node *node;
509 	struct nvmet_tcp_cmd *cmd;
510 
511 	for (node = llist_del_all(&queue->resp_list); node; node = node->next) {
512 		cmd = llist_entry(node, struct nvmet_tcp_cmd, lentry);
513 		list_add(&cmd->entry, &queue->resp_send_list);
514 		queue->send_list_len++;
515 	}
516 }
517 
518 static struct nvmet_tcp_cmd *nvmet_tcp_fetch_cmd(struct nvmet_tcp_queue *queue)
519 {
520 	queue->snd_cmd = list_first_entry_or_null(&queue->resp_send_list,
521 				struct nvmet_tcp_cmd, entry);
522 	if (!queue->snd_cmd) {
523 		nvmet_tcp_process_resp_list(queue);
524 		queue->snd_cmd =
525 			list_first_entry_or_null(&queue->resp_send_list,
526 					struct nvmet_tcp_cmd, entry);
527 		if (unlikely(!queue->snd_cmd))
528 			return NULL;
529 	}
530 
531 	list_del_init(&queue->snd_cmd->entry);
532 	queue->send_list_len--;
533 
534 	if (nvmet_tcp_need_data_out(queue->snd_cmd))
535 		nvmet_setup_c2h_data_pdu(queue->snd_cmd);
536 	else if (nvmet_tcp_need_data_in(queue->snd_cmd))
537 		nvmet_setup_r2t_pdu(queue->snd_cmd);
538 	else
539 		nvmet_setup_response_pdu(queue->snd_cmd);
540 
541 	return queue->snd_cmd;
542 }
543 
544 static void nvmet_tcp_queue_response(struct nvmet_req *req)
545 {
546 	struct nvmet_tcp_cmd *cmd =
547 		container_of(req, struct nvmet_tcp_cmd, req);
548 	struct nvmet_tcp_queue	*queue = cmd->queue;
549 	struct nvme_sgl_desc *sgl;
550 	u32 len;
551 
552 	if (unlikely(cmd == queue->cmd)) {
553 		sgl = &cmd->req.cmd->common.dptr.sgl;
554 		len = le32_to_cpu(sgl->length);
555 
556 		/*
557 		 * Wait for inline data before processing the response.
558 		 * Avoid using helpers, this might happen before
559 		 * nvmet_req_init is completed.
560 		 */
561 		if (queue->rcv_state == NVMET_TCP_RECV_PDU &&
562 		    len && len <= cmd->req.port->inline_data_size &&
563 		    nvme_is_write(cmd->req.cmd))
564 			return;
565 	}
566 
567 	llist_add(&cmd->lentry, &queue->resp_list);
568 	queue_work_on(queue_cpu(queue), nvmet_tcp_wq, &cmd->queue->io_work);
569 }
570 
571 static void nvmet_tcp_execute_request(struct nvmet_tcp_cmd *cmd)
572 {
573 	if (unlikely(cmd->flags & NVMET_TCP_F_INIT_FAILED))
574 		nvmet_tcp_queue_response(&cmd->req);
575 	else
576 		cmd->req.execute(&cmd->req);
577 }
578 
579 static int nvmet_try_send_data_pdu(struct nvmet_tcp_cmd *cmd)
580 {
581 	struct msghdr msg = {
582 		.msg_flags = MSG_DONTWAIT | MSG_MORE | MSG_SPLICE_PAGES,
583 	};
584 	struct bio_vec bvec;
585 	u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
586 	int left = sizeof(*cmd->data_pdu) - cmd->offset + hdgst;
587 	int ret;
588 
589 	bvec_set_virt(&bvec, (void *)cmd->data_pdu + cmd->offset, left);
590 	iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, &bvec, 1, left);
591 	ret = sock_sendmsg(cmd->queue->sock, &msg);
592 	if (ret <= 0)
593 		return ret;
594 
595 	cmd->offset += ret;
596 	left -= ret;
597 
598 	if (left)
599 		return -EAGAIN;
600 
601 	cmd->state = NVMET_TCP_SEND_DATA;
602 	cmd->offset  = 0;
603 	return 1;
604 }
605 
606 static int nvmet_try_send_data(struct nvmet_tcp_cmd *cmd, bool last_in_batch)
607 {
608 	struct nvmet_tcp_queue *queue = cmd->queue;
609 	int ret;
610 
611 	while (cmd->cur_sg) {
612 		struct msghdr msg = {
613 			.msg_flags = MSG_DONTWAIT | MSG_SPLICE_PAGES,
614 		};
615 		struct page *page = sg_page(cmd->cur_sg);
616 		struct bio_vec bvec;
617 		u32 left = cmd->cur_sg->length - cmd->offset;
618 
619 		if ((!last_in_batch && cmd->queue->send_list_len) ||
620 		    cmd->wbytes_done + left < cmd->req.transfer_len ||
621 		    queue->data_digest || !queue->nvme_sq.sqhd_disabled)
622 			msg.msg_flags |= MSG_MORE;
623 
624 		bvec_set_page(&bvec, page, left, cmd->offset);
625 		iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, &bvec, 1, left);
626 		ret = sock_sendmsg(cmd->queue->sock, &msg);
627 		if (ret <= 0)
628 			return ret;
629 
630 		cmd->offset += ret;
631 		cmd->wbytes_done += ret;
632 
633 		/* Done with sg?*/
634 		if (cmd->offset == cmd->cur_sg->length) {
635 			cmd->cur_sg = sg_next(cmd->cur_sg);
636 			cmd->offset = 0;
637 		}
638 	}
639 
640 	if (queue->data_digest) {
641 		cmd->state = NVMET_TCP_SEND_DDGST;
642 		cmd->offset = 0;
643 	} else {
644 		if (queue->nvme_sq.sqhd_disabled) {
645 			cmd->queue->snd_cmd = NULL;
646 			nvmet_tcp_put_cmd(cmd);
647 		} else {
648 			nvmet_setup_response_pdu(cmd);
649 		}
650 	}
651 
652 	if (queue->nvme_sq.sqhd_disabled)
653 		nvmet_tcp_free_cmd_buffers(cmd);
654 
655 	return 1;
656 
657 }
658 
659 static int nvmet_try_send_response(struct nvmet_tcp_cmd *cmd,
660 		bool last_in_batch)
661 {
662 	struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_SPLICE_PAGES, };
663 	struct bio_vec bvec;
664 	u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
665 	int left = sizeof(*cmd->rsp_pdu) - cmd->offset + hdgst;
666 	int ret;
667 
668 	if (!last_in_batch && cmd->queue->send_list_len)
669 		msg.msg_flags |= MSG_MORE;
670 	else
671 		msg.msg_flags |= MSG_EOR;
672 
673 	bvec_set_virt(&bvec, (void *)cmd->rsp_pdu + cmd->offset, left);
674 	iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, &bvec, 1, left);
675 	ret = sock_sendmsg(cmd->queue->sock, &msg);
676 	if (ret <= 0)
677 		return ret;
678 	cmd->offset += ret;
679 	left -= ret;
680 
681 	if (left)
682 		return -EAGAIN;
683 
684 	nvmet_tcp_free_cmd_buffers(cmd);
685 	cmd->queue->snd_cmd = NULL;
686 	nvmet_tcp_put_cmd(cmd);
687 	return 1;
688 }
689 
690 static int nvmet_try_send_r2t(struct nvmet_tcp_cmd *cmd, bool last_in_batch)
691 {
692 	struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_SPLICE_PAGES, };
693 	struct bio_vec bvec;
694 	u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
695 	int left = sizeof(*cmd->r2t_pdu) - cmd->offset + hdgst;
696 	int ret;
697 
698 	if (!last_in_batch && cmd->queue->send_list_len)
699 		msg.msg_flags |= MSG_MORE;
700 	else
701 		msg.msg_flags |= MSG_EOR;
702 
703 	bvec_set_virt(&bvec, (void *)cmd->r2t_pdu + cmd->offset, left);
704 	iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, &bvec, 1, left);
705 	ret = sock_sendmsg(cmd->queue->sock, &msg);
706 	if (ret <= 0)
707 		return ret;
708 	cmd->offset += ret;
709 	left -= ret;
710 
711 	if (left)
712 		return -EAGAIN;
713 
714 	cmd->queue->snd_cmd = NULL;
715 	return 1;
716 }
717 
718 static int nvmet_try_send_ddgst(struct nvmet_tcp_cmd *cmd, bool last_in_batch)
719 {
720 	struct nvmet_tcp_queue *queue = cmd->queue;
721 	int left = NVME_TCP_DIGEST_LENGTH - cmd->offset;
722 	struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
723 	struct kvec iov = {
724 		.iov_base = (u8 *)&cmd->exp_ddgst + cmd->offset,
725 		.iov_len = left
726 	};
727 	int ret;
728 
729 	if (!last_in_batch && cmd->queue->send_list_len)
730 		msg.msg_flags |= MSG_MORE;
731 	else
732 		msg.msg_flags |= MSG_EOR;
733 
734 	ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
735 	if (unlikely(ret <= 0))
736 		return ret;
737 
738 	cmd->offset += ret;
739 	left -= ret;
740 
741 	if (left)
742 		return -EAGAIN;
743 
744 	if (queue->nvme_sq.sqhd_disabled) {
745 		cmd->queue->snd_cmd = NULL;
746 		nvmet_tcp_put_cmd(cmd);
747 	} else {
748 		nvmet_setup_response_pdu(cmd);
749 	}
750 	return 1;
751 }
752 
753 static int nvmet_tcp_try_send_one(struct nvmet_tcp_queue *queue,
754 		bool last_in_batch)
755 {
756 	struct nvmet_tcp_cmd *cmd = queue->snd_cmd;
757 	int ret = 0;
758 
759 	if (!cmd || queue->state == NVMET_TCP_Q_DISCONNECTING) {
760 		cmd = nvmet_tcp_fetch_cmd(queue);
761 		if (unlikely(!cmd))
762 			return 0;
763 	}
764 
765 	if (cmd->state == NVMET_TCP_SEND_DATA_PDU) {
766 		ret = nvmet_try_send_data_pdu(cmd);
767 		if (ret <= 0)
768 			goto done_send;
769 	}
770 
771 	if (cmd->state == NVMET_TCP_SEND_DATA) {
772 		ret = nvmet_try_send_data(cmd, last_in_batch);
773 		if (ret <= 0)
774 			goto done_send;
775 	}
776 
777 	if (cmd->state == NVMET_TCP_SEND_DDGST) {
778 		ret = nvmet_try_send_ddgst(cmd, last_in_batch);
779 		if (ret <= 0)
780 			goto done_send;
781 	}
782 
783 	if (cmd->state == NVMET_TCP_SEND_R2T) {
784 		ret = nvmet_try_send_r2t(cmd, last_in_batch);
785 		if (ret <= 0)
786 			goto done_send;
787 	}
788 
789 	if (cmd->state == NVMET_TCP_SEND_RESPONSE)
790 		ret = nvmet_try_send_response(cmd, last_in_batch);
791 
792 done_send:
793 	if (ret < 0) {
794 		if (ret == -EAGAIN)
795 			return 0;
796 		return ret;
797 	}
798 
799 	return 1;
800 }
801 
802 static int nvmet_tcp_try_send(struct nvmet_tcp_queue *queue,
803 		int budget, int *sends)
804 {
805 	int i, ret = 0;
806 
807 	for (i = 0; i < budget; i++) {
808 		ret = nvmet_tcp_try_send_one(queue, i == budget - 1);
809 		if (unlikely(ret < 0)) {
810 			nvmet_tcp_socket_error(queue, ret);
811 			goto done;
812 		} else if (ret == 0) {
813 			break;
814 		}
815 		(*sends)++;
816 	}
817 done:
818 	return ret;
819 }
820 
821 static void nvmet_prepare_receive_pdu(struct nvmet_tcp_queue *queue)
822 {
823 	queue->offset = 0;
824 	queue->left = sizeof(struct nvme_tcp_hdr);
825 	queue->cmd = NULL;
826 	queue->rcv_state = NVMET_TCP_RECV_PDU;
827 }
828 
829 static void nvmet_tcp_free_crypto(struct nvmet_tcp_queue *queue)
830 {
831 	struct crypto_ahash *tfm = crypto_ahash_reqtfm(queue->rcv_hash);
832 
833 	ahash_request_free(queue->rcv_hash);
834 	ahash_request_free(queue->snd_hash);
835 	crypto_free_ahash(tfm);
836 }
837 
838 static int nvmet_tcp_alloc_crypto(struct nvmet_tcp_queue *queue)
839 {
840 	struct crypto_ahash *tfm;
841 
842 	tfm = crypto_alloc_ahash("crc32c", 0, CRYPTO_ALG_ASYNC);
843 	if (IS_ERR(tfm))
844 		return PTR_ERR(tfm);
845 
846 	queue->snd_hash = ahash_request_alloc(tfm, GFP_KERNEL);
847 	if (!queue->snd_hash)
848 		goto free_tfm;
849 	ahash_request_set_callback(queue->snd_hash, 0, NULL, NULL);
850 
851 	queue->rcv_hash = ahash_request_alloc(tfm, GFP_KERNEL);
852 	if (!queue->rcv_hash)
853 		goto free_snd_hash;
854 	ahash_request_set_callback(queue->rcv_hash, 0, NULL, NULL);
855 
856 	return 0;
857 free_snd_hash:
858 	ahash_request_free(queue->snd_hash);
859 free_tfm:
860 	crypto_free_ahash(tfm);
861 	return -ENOMEM;
862 }
863 
864 
865 static int nvmet_tcp_handle_icreq(struct nvmet_tcp_queue *queue)
866 {
867 	struct nvme_tcp_icreq_pdu *icreq = &queue->pdu.icreq;
868 	struct nvme_tcp_icresp_pdu *icresp = &queue->pdu.icresp;
869 	struct msghdr msg = {};
870 	struct kvec iov;
871 	int ret;
872 
873 	if (le32_to_cpu(icreq->hdr.plen) != sizeof(struct nvme_tcp_icreq_pdu)) {
874 		pr_err("bad nvme-tcp pdu length (%d)\n",
875 			le32_to_cpu(icreq->hdr.plen));
876 		nvmet_tcp_fatal_error(queue);
877 	}
878 
879 	if (icreq->pfv != NVME_TCP_PFV_1_0) {
880 		pr_err("queue %d: bad pfv %d\n", queue->idx, icreq->pfv);
881 		return -EPROTO;
882 	}
883 
884 	if (icreq->hpda != 0) {
885 		pr_err("queue %d: unsupported hpda %d\n", queue->idx,
886 			icreq->hpda);
887 		return -EPROTO;
888 	}
889 
890 	queue->hdr_digest = !!(icreq->digest & NVME_TCP_HDR_DIGEST_ENABLE);
891 	queue->data_digest = !!(icreq->digest & NVME_TCP_DATA_DIGEST_ENABLE);
892 	if (queue->hdr_digest || queue->data_digest) {
893 		ret = nvmet_tcp_alloc_crypto(queue);
894 		if (ret)
895 			return ret;
896 	}
897 
898 	memset(icresp, 0, sizeof(*icresp));
899 	icresp->hdr.type = nvme_tcp_icresp;
900 	icresp->hdr.hlen = sizeof(*icresp);
901 	icresp->hdr.pdo = 0;
902 	icresp->hdr.plen = cpu_to_le32(icresp->hdr.hlen);
903 	icresp->pfv = cpu_to_le16(NVME_TCP_PFV_1_0);
904 	icresp->maxdata = cpu_to_le32(NVMET_TCP_MAXH2CDATA);
905 	icresp->cpda = 0;
906 	if (queue->hdr_digest)
907 		icresp->digest |= NVME_TCP_HDR_DIGEST_ENABLE;
908 	if (queue->data_digest)
909 		icresp->digest |= NVME_TCP_DATA_DIGEST_ENABLE;
910 
911 	iov.iov_base = icresp;
912 	iov.iov_len = sizeof(*icresp);
913 	ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
914 	if (ret < 0)
915 		return ret; /* queue removal will cleanup */
916 
917 	queue->state = NVMET_TCP_Q_LIVE;
918 	nvmet_prepare_receive_pdu(queue);
919 	return 0;
920 }
921 
922 static void nvmet_tcp_handle_req_failure(struct nvmet_tcp_queue *queue,
923 		struct nvmet_tcp_cmd *cmd, struct nvmet_req *req)
924 {
925 	size_t data_len = le32_to_cpu(req->cmd->common.dptr.sgl.length);
926 	int ret;
927 
928 	/*
929 	 * This command has not been processed yet, hence we are trying to
930 	 * figure out if there is still pending data left to receive. If
931 	 * we don't, we can simply prepare for the next pdu and bail out,
932 	 * otherwise we will need to prepare a buffer and receive the
933 	 * stale data before continuing forward.
934 	 */
935 	if (!nvme_is_write(cmd->req.cmd) || !data_len ||
936 	    data_len > cmd->req.port->inline_data_size) {
937 		nvmet_prepare_receive_pdu(queue);
938 		return;
939 	}
940 
941 	ret = nvmet_tcp_map_data(cmd);
942 	if (unlikely(ret)) {
943 		pr_err("queue %d: failed to map data\n", queue->idx);
944 		nvmet_tcp_fatal_error(queue);
945 		return;
946 	}
947 
948 	queue->rcv_state = NVMET_TCP_RECV_DATA;
949 	nvmet_tcp_build_pdu_iovec(cmd);
950 	cmd->flags |= NVMET_TCP_F_INIT_FAILED;
951 }
952 
953 static int nvmet_tcp_handle_h2c_data_pdu(struct nvmet_tcp_queue *queue)
954 {
955 	struct nvme_tcp_data_pdu *data = &queue->pdu.data;
956 	struct nvmet_tcp_cmd *cmd;
957 	unsigned int exp_data_len;
958 
959 	if (likely(queue->nr_cmds)) {
960 		if (unlikely(data->ttag >= queue->nr_cmds)) {
961 			pr_err("queue %d: received out of bound ttag %u, nr_cmds %u\n",
962 				queue->idx, data->ttag, queue->nr_cmds);
963 			nvmet_tcp_fatal_error(queue);
964 			return -EPROTO;
965 		}
966 		cmd = &queue->cmds[data->ttag];
967 	} else {
968 		cmd = &queue->connect;
969 	}
970 
971 	if (le32_to_cpu(data->data_offset) != cmd->rbytes_done) {
972 		pr_err("ttag %u unexpected data offset %u (expected %u)\n",
973 			data->ttag, le32_to_cpu(data->data_offset),
974 			cmd->rbytes_done);
975 		/* FIXME: use path and transport errors */
976 		nvmet_tcp_fatal_error(queue);
977 		return -EPROTO;
978 	}
979 
980 	exp_data_len = le32_to_cpu(data->hdr.plen) -
981 			nvmet_tcp_hdgst_len(queue) -
982 			nvmet_tcp_ddgst_len(queue) -
983 			sizeof(*data);
984 
985 	cmd->pdu_len = le32_to_cpu(data->data_length);
986 	if (unlikely(cmd->pdu_len != exp_data_len ||
987 		     cmd->pdu_len == 0 ||
988 		     cmd->pdu_len > NVMET_TCP_MAXH2CDATA)) {
989 		pr_err("H2CData PDU len %u is invalid\n", cmd->pdu_len);
990 		/* FIXME: use proper transport errors */
991 		nvmet_tcp_fatal_error(queue);
992 		return -EPROTO;
993 	}
994 	cmd->pdu_recv = 0;
995 	nvmet_tcp_build_pdu_iovec(cmd);
996 	queue->cmd = cmd;
997 	queue->rcv_state = NVMET_TCP_RECV_DATA;
998 
999 	return 0;
1000 }
1001 
1002 static int nvmet_tcp_done_recv_pdu(struct nvmet_tcp_queue *queue)
1003 {
1004 	struct nvme_tcp_hdr *hdr = &queue->pdu.cmd.hdr;
1005 	struct nvme_command *nvme_cmd = &queue->pdu.cmd.cmd;
1006 	struct nvmet_req *req;
1007 	int ret;
1008 
1009 	if (unlikely(queue->state == NVMET_TCP_Q_CONNECTING)) {
1010 		if (hdr->type != nvme_tcp_icreq) {
1011 			pr_err("unexpected pdu type (%d) before icreq\n",
1012 				hdr->type);
1013 			nvmet_tcp_fatal_error(queue);
1014 			return -EPROTO;
1015 		}
1016 		return nvmet_tcp_handle_icreq(queue);
1017 	}
1018 
1019 	if (unlikely(hdr->type == nvme_tcp_icreq)) {
1020 		pr_err("queue %d: received icreq pdu in state %d\n",
1021 			queue->idx, queue->state);
1022 		nvmet_tcp_fatal_error(queue);
1023 		return -EPROTO;
1024 	}
1025 
1026 	if (hdr->type == nvme_tcp_h2c_data) {
1027 		ret = nvmet_tcp_handle_h2c_data_pdu(queue);
1028 		if (unlikely(ret))
1029 			return ret;
1030 		return 0;
1031 	}
1032 
1033 	queue->cmd = nvmet_tcp_get_cmd(queue);
1034 	if (unlikely(!queue->cmd)) {
1035 		/* This should never happen */
1036 		pr_err("queue %d: out of commands (%d) send_list_len: %d, opcode: %d",
1037 			queue->idx, queue->nr_cmds, queue->send_list_len,
1038 			nvme_cmd->common.opcode);
1039 		nvmet_tcp_fatal_error(queue);
1040 		return -ENOMEM;
1041 	}
1042 
1043 	req = &queue->cmd->req;
1044 	memcpy(req->cmd, nvme_cmd, sizeof(*nvme_cmd));
1045 
1046 	if (unlikely(!nvmet_req_init(req, &queue->nvme_cq,
1047 			&queue->nvme_sq, &nvmet_tcp_ops))) {
1048 		pr_err("failed cmd %p id %d opcode %d, data_len: %d\n",
1049 			req->cmd, req->cmd->common.command_id,
1050 			req->cmd->common.opcode,
1051 			le32_to_cpu(req->cmd->common.dptr.sgl.length));
1052 
1053 		nvmet_tcp_handle_req_failure(queue, queue->cmd, req);
1054 		return 0;
1055 	}
1056 
1057 	ret = nvmet_tcp_map_data(queue->cmd);
1058 	if (unlikely(ret)) {
1059 		pr_err("queue %d: failed to map data\n", queue->idx);
1060 		if (nvmet_tcp_has_inline_data(queue->cmd))
1061 			nvmet_tcp_fatal_error(queue);
1062 		else
1063 			nvmet_req_complete(req, ret);
1064 		ret = -EAGAIN;
1065 		goto out;
1066 	}
1067 
1068 	if (nvmet_tcp_need_data_in(queue->cmd)) {
1069 		if (nvmet_tcp_has_inline_data(queue->cmd)) {
1070 			queue->rcv_state = NVMET_TCP_RECV_DATA;
1071 			nvmet_tcp_build_pdu_iovec(queue->cmd);
1072 			return 0;
1073 		}
1074 		/* send back R2T */
1075 		nvmet_tcp_queue_response(&queue->cmd->req);
1076 		goto out;
1077 	}
1078 
1079 	queue->cmd->req.execute(&queue->cmd->req);
1080 out:
1081 	nvmet_prepare_receive_pdu(queue);
1082 	return ret;
1083 }
1084 
1085 static const u8 nvme_tcp_pdu_sizes[] = {
1086 	[nvme_tcp_icreq]	= sizeof(struct nvme_tcp_icreq_pdu),
1087 	[nvme_tcp_cmd]		= sizeof(struct nvme_tcp_cmd_pdu),
1088 	[nvme_tcp_h2c_data]	= sizeof(struct nvme_tcp_data_pdu),
1089 };
1090 
1091 static inline u8 nvmet_tcp_pdu_size(u8 type)
1092 {
1093 	size_t idx = type;
1094 
1095 	return (idx < ARRAY_SIZE(nvme_tcp_pdu_sizes) &&
1096 		nvme_tcp_pdu_sizes[idx]) ?
1097 			nvme_tcp_pdu_sizes[idx] : 0;
1098 }
1099 
1100 static inline bool nvmet_tcp_pdu_valid(u8 type)
1101 {
1102 	switch (type) {
1103 	case nvme_tcp_icreq:
1104 	case nvme_tcp_cmd:
1105 	case nvme_tcp_h2c_data:
1106 		/* fallthru */
1107 		return true;
1108 	}
1109 
1110 	return false;
1111 }
1112 
1113 static int nvmet_tcp_try_recv_pdu(struct nvmet_tcp_queue *queue)
1114 {
1115 	struct nvme_tcp_hdr *hdr = &queue->pdu.cmd.hdr;
1116 	int len;
1117 	struct kvec iov;
1118 	struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
1119 
1120 recv:
1121 	iov.iov_base = (void *)&queue->pdu + queue->offset;
1122 	iov.iov_len = queue->left;
1123 	len = kernel_recvmsg(queue->sock, &msg, &iov, 1,
1124 			iov.iov_len, msg.msg_flags);
1125 	if (unlikely(len < 0))
1126 		return len;
1127 
1128 	queue->offset += len;
1129 	queue->left -= len;
1130 	if (queue->left)
1131 		return -EAGAIN;
1132 
1133 	if (queue->offset == sizeof(struct nvme_tcp_hdr)) {
1134 		u8 hdgst = nvmet_tcp_hdgst_len(queue);
1135 
1136 		if (unlikely(!nvmet_tcp_pdu_valid(hdr->type))) {
1137 			pr_err("unexpected pdu type %d\n", hdr->type);
1138 			nvmet_tcp_fatal_error(queue);
1139 			return -EIO;
1140 		}
1141 
1142 		if (unlikely(hdr->hlen != nvmet_tcp_pdu_size(hdr->type))) {
1143 			pr_err("pdu %d bad hlen %d\n", hdr->type, hdr->hlen);
1144 			return -EIO;
1145 		}
1146 
1147 		queue->left = hdr->hlen - queue->offset + hdgst;
1148 		goto recv;
1149 	}
1150 
1151 	if (queue->hdr_digest &&
1152 	    nvmet_tcp_verify_hdgst(queue, &queue->pdu, hdr->hlen)) {
1153 		nvmet_tcp_fatal_error(queue); /* fatal */
1154 		return -EPROTO;
1155 	}
1156 
1157 	if (queue->data_digest &&
1158 	    nvmet_tcp_check_ddgst(queue, &queue->pdu)) {
1159 		nvmet_tcp_fatal_error(queue); /* fatal */
1160 		return -EPROTO;
1161 	}
1162 
1163 	return nvmet_tcp_done_recv_pdu(queue);
1164 }
1165 
1166 static void nvmet_tcp_prep_recv_ddgst(struct nvmet_tcp_cmd *cmd)
1167 {
1168 	struct nvmet_tcp_queue *queue = cmd->queue;
1169 
1170 	nvmet_tcp_calc_ddgst(queue->rcv_hash, cmd);
1171 	queue->offset = 0;
1172 	queue->left = NVME_TCP_DIGEST_LENGTH;
1173 	queue->rcv_state = NVMET_TCP_RECV_DDGST;
1174 }
1175 
1176 static int nvmet_tcp_try_recv_data(struct nvmet_tcp_queue *queue)
1177 {
1178 	struct nvmet_tcp_cmd  *cmd = queue->cmd;
1179 	int ret;
1180 
1181 	while (msg_data_left(&cmd->recv_msg)) {
1182 		ret = sock_recvmsg(cmd->queue->sock, &cmd->recv_msg,
1183 			cmd->recv_msg.msg_flags);
1184 		if (ret <= 0)
1185 			return ret;
1186 
1187 		cmd->pdu_recv += ret;
1188 		cmd->rbytes_done += ret;
1189 	}
1190 
1191 	if (queue->data_digest) {
1192 		nvmet_tcp_prep_recv_ddgst(cmd);
1193 		return 0;
1194 	}
1195 
1196 	if (cmd->rbytes_done == cmd->req.transfer_len)
1197 		nvmet_tcp_execute_request(cmd);
1198 
1199 	nvmet_prepare_receive_pdu(queue);
1200 	return 0;
1201 }
1202 
1203 static int nvmet_tcp_try_recv_ddgst(struct nvmet_tcp_queue *queue)
1204 {
1205 	struct nvmet_tcp_cmd *cmd = queue->cmd;
1206 	int ret;
1207 	struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
1208 	struct kvec iov = {
1209 		.iov_base = (void *)&cmd->recv_ddgst + queue->offset,
1210 		.iov_len = queue->left
1211 	};
1212 
1213 	ret = kernel_recvmsg(queue->sock, &msg, &iov, 1,
1214 			iov.iov_len, msg.msg_flags);
1215 	if (unlikely(ret < 0))
1216 		return ret;
1217 
1218 	queue->offset += ret;
1219 	queue->left -= ret;
1220 	if (queue->left)
1221 		return -EAGAIN;
1222 
1223 	if (queue->data_digest && cmd->exp_ddgst != cmd->recv_ddgst) {
1224 		pr_err("queue %d: cmd %d pdu (%d) data digest error: recv %#x expected %#x\n",
1225 			queue->idx, cmd->req.cmd->common.command_id,
1226 			queue->pdu.cmd.hdr.type, le32_to_cpu(cmd->recv_ddgst),
1227 			le32_to_cpu(cmd->exp_ddgst));
1228 		nvmet_req_uninit(&cmd->req);
1229 		nvmet_tcp_free_cmd_buffers(cmd);
1230 		nvmet_tcp_fatal_error(queue);
1231 		ret = -EPROTO;
1232 		goto out;
1233 	}
1234 
1235 	if (cmd->rbytes_done == cmd->req.transfer_len)
1236 		nvmet_tcp_execute_request(cmd);
1237 
1238 	ret = 0;
1239 out:
1240 	nvmet_prepare_receive_pdu(queue);
1241 	return ret;
1242 }
1243 
1244 static int nvmet_tcp_try_recv_one(struct nvmet_tcp_queue *queue)
1245 {
1246 	int result = 0;
1247 
1248 	if (unlikely(queue->rcv_state == NVMET_TCP_RECV_ERR))
1249 		return 0;
1250 
1251 	if (queue->rcv_state == NVMET_TCP_RECV_PDU) {
1252 		result = nvmet_tcp_try_recv_pdu(queue);
1253 		if (result != 0)
1254 			goto done_recv;
1255 	}
1256 
1257 	if (queue->rcv_state == NVMET_TCP_RECV_DATA) {
1258 		result = nvmet_tcp_try_recv_data(queue);
1259 		if (result != 0)
1260 			goto done_recv;
1261 	}
1262 
1263 	if (queue->rcv_state == NVMET_TCP_RECV_DDGST) {
1264 		result = nvmet_tcp_try_recv_ddgst(queue);
1265 		if (result != 0)
1266 			goto done_recv;
1267 	}
1268 
1269 done_recv:
1270 	if (result < 0) {
1271 		if (result == -EAGAIN)
1272 			return 0;
1273 		return result;
1274 	}
1275 	return 1;
1276 }
1277 
1278 static int nvmet_tcp_try_recv(struct nvmet_tcp_queue *queue,
1279 		int budget, int *recvs)
1280 {
1281 	int i, ret = 0;
1282 
1283 	for (i = 0; i < budget; i++) {
1284 		ret = nvmet_tcp_try_recv_one(queue);
1285 		if (unlikely(ret < 0)) {
1286 			nvmet_tcp_socket_error(queue, ret);
1287 			goto done;
1288 		} else if (ret == 0) {
1289 			break;
1290 		}
1291 		(*recvs)++;
1292 	}
1293 done:
1294 	return ret;
1295 }
1296 
1297 static void nvmet_tcp_schedule_release_queue(struct nvmet_tcp_queue *queue)
1298 {
1299 	spin_lock(&queue->state_lock);
1300 	if (queue->state != NVMET_TCP_Q_DISCONNECTING) {
1301 		queue->state = NVMET_TCP_Q_DISCONNECTING;
1302 		queue_work(nvmet_wq, &queue->release_work);
1303 	}
1304 	spin_unlock(&queue->state_lock);
1305 }
1306 
1307 static inline void nvmet_tcp_arm_queue_deadline(struct nvmet_tcp_queue *queue)
1308 {
1309 	queue->poll_end = jiffies + usecs_to_jiffies(idle_poll_period_usecs);
1310 }
1311 
1312 static bool nvmet_tcp_check_queue_deadline(struct nvmet_tcp_queue *queue,
1313 		int ops)
1314 {
1315 	if (!idle_poll_period_usecs)
1316 		return false;
1317 
1318 	if (ops)
1319 		nvmet_tcp_arm_queue_deadline(queue);
1320 
1321 	return !time_after(jiffies, queue->poll_end);
1322 }
1323 
1324 static void nvmet_tcp_io_work(struct work_struct *w)
1325 {
1326 	struct nvmet_tcp_queue *queue =
1327 		container_of(w, struct nvmet_tcp_queue, io_work);
1328 	bool pending;
1329 	int ret, ops = 0;
1330 
1331 	do {
1332 		pending = false;
1333 
1334 		ret = nvmet_tcp_try_recv(queue, NVMET_TCP_RECV_BUDGET, &ops);
1335 		if (ret > 0)
1336 			pending = true;
1337 		else if (ret < 0)
1338 			return;
1339 
1340 		ret = nvmet_tcp_try_send(queue, NVMET_TCP_SEND_BUDGET, &ops);
1341 		if (ret > 0)
1342 			pending = true;
1343 		else if (ret < 0)
1344 			return;
1345 
1346 	} while (pending && ops < NVMET_TCP_IO_WORK_BUDGET);
1347 
1348 	/*
1349 	 * Requeue the worker if idle deadline period is in progress or any
1350 	 * ops activity was recorded during the do-while loop above.
1351 	 */
1352 	if (nvmet_tcp_check_queue_deadline(queue, ops) || pending)
1353 		queue_work_on(queue_cpu(queue), nvmet_tcp_wq, &queue->io_work);
1354 }
1355 
1356 static int nvmet_tcp_alloc_cmd(struct nvmet_tcp_queue *queue,
1357 		struct nvmet_tcp_cmd *c)
1358 {
1359 	u8 hdgst = nvmet_tcp_hdgst_len(queue);
1360 
1361 	c->queue = queue;
1362 	c->req.port = queue->port->nport;
1363 
1364 	c->cmd_pdu = page_frag_alloc(&queue->pf_cache,
1365 			sizeof(*c->cmd_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO);
1366 	if (!c->cmd_pdu)
1367 		return -ENOMEM;
1368 	c->req.cmd = &c->cmd_pdu->cmd;
1369 
1370 	c->rsp_pdu = page_frag_alloc(&queue->pf_cache,
1371 			sizeof(*c->rsp_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO);
1372 	if (!c->rsp_pdu)
1373 		goto out_free_cmd;
1374 	c->req.cqe = &c->rsp_pdu->cqe;
1375 
1376 	c->data_pdu = page_frag_alloc(&queue->pf_cache,
1377 			sizeof(*c->data_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO);
1378 	if (!c->data_pdu)
1379 		goto out_free_rsp;
1380 
1381 	c->r2t_pdu = page_frag_alloc(&queue->pf_cache,
1382 			sizeof(*c->r2t_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO);
1383 	if (!c->r2t_pdu)
1384 		goto out_free_data;
1385 
1386 	c->recv_msg.msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
1387 
1388 	list_add_tail(&c->entry, &queue->free_list);
1389 
1390 	return 0;
1391 out_free_data:
1392 	page_frag_free(c->data_pdu);
1393 out_free_rsp:
1394 	page_frag_free(c->rsp_pdu);
1395 out_free_cmd:
1396 	page_frag_free(c->cmd_pdu);
1397 	return -ENOMEM;
1398 }
1399 
1400 static void nvmet_tcp_free_cmd(struct nvmet_tcp_cmd *c)
1401 {
1402 	page_frag_free(c->r2t_pdu);
1403 	page_frag_free(c->data_pdu);
1404 	page_frag_free(c->rsp_pdu);
1405 	page_frag_free(c->cmd_pdu);
1406 }
1407 
1408 static int nvmet_tcp_alloc_cmds(struct nvmet_tcp_queue *queue)
1409 {
1410 	struct nvmet_tcp_cmd *cmds;
1411 	int i, ret = -EINVAL, nr_cmds = queue->nr_cmds;
1412 
1413 	cmds = kcalloc(nr_cmds, sizeof(struct nvmet_tcp_cmd), GFP_KERNEL);
1414 	if (!cmds)
1415 		goto out;
1416 
1417 	for (i = 0; i < nr_cmds; i++) {
1418 		ret = nvmet_tcp_alloc_cmd(queue, cmds + i);
1419 		if (ret)
1420 			goto out_free;
1421 	}
1422 
1423 	queue->cmds = cmds;
1424 
1425 	return 0;
1426 out_free:
1427 	while (--i >= 0)
1428 		nvmet_tcp_free_cmd(cmds + i);
1429 	kfree(cmds);
1430 out:
1431 	return ret;
1432 }
1433 
1434 static void nvmet_tcp_free_cmds(struct nvmet_tcp_queue *queue)
1435 {
1436 	struct nvmet_tcp_cmd *cmds = queue->cmds;
1437 	int i;
1438 
1439 	for (i = 0; i < queue->nr_cmds; i++)
1440 		nvmet_tcp_free_cmd(cmds + i);
1441 
1442 	nvmet_tcp_free_cmd(&queue->connect);
1443 	kfree(cmds);
1444 }
1445 
1446 static void nvmet_tcp_restore_socket_callbacks(struct nvmet_tcp_queue *queue)
1447 {
1448 	struct socket *sock = queue->sock;
1449 
1450 	write_lock_bh(&sock->sk->sk_callback_lock);
1451 	sock->sk->sk_data_ready =  queue->data_ready;
1452 	sock->sk->sk_state_change = queue->state_change;
1453 	sock->sk->sk_write_space = queue->write_space;
1454 	sock->sk->sk_user_data = NULL;
1455 	write_unlock_bh(&sock->sk->sk_callback_lock);
1456 }
1457 
1458 static void nvmet_tcp_uninit_data_in_cmds(struct nvmet_tcp_queue *queue)
1459 {
1460 	struct nvmet_tcp_cmd *cmd = queue->cmds;
1461 	int i;
1462 
1463 	for (i = 0; i < queue->nr_cmds; i++, cmd++) {
1464 		if (nvmet_tcp_need_data_in(cmd))
1465 			nvmet_req_uninit(&cmd->req);
1466 	}
1467 
1468 	if (!queue->nr_cmds && nvmet_tcp_need_data_in(&queue->connect)) {
1469 		/* failed in connect */
1470 		nvmet_req_uninit(&queue->connect.req);
1471 	}
1472 }
1473 
1474 static void nvmet_tcp_free_cmd_data_in_buffers(struct nvmet_tcp_queue *queue)
1475 {
1476 	struct nvmet_tcp_cmd *cmd = queue->cmds;
1477 	int i;
1478 
1479 	for (i = 0; i < queue->nr_cmds; i++, cmd++) {
1480 		if (nvmet_tcp_need_data_in(cmd))
1481 			nvmet_tcp_free_cmd_buffers(cmd);
1482 	}
1483 
1484 	if (!queue->nr_cmds && nvmet_tcp_need_data_in(&queue->connect))
1485 		nvmet_tcp_free_cmd_buffers(&queue->connect);
1486 }
1487 
1488 static void nvmet_tcp_release_queue_work(struct work_struct *w)
1489 {
1490 	struct page *page;
1491 	struct nvmet_tcp_queue *queue =
1492 		container_of(w, struct nvmet_tcp_queue, release_work);
1493 
1494 	mutex_lock(&nvmet_tcp_queue_mutex);
1495 	list_del_init(&queue->queue_list);
1496 	mutex_unlock(&nvmet_tcp_queue_mutex);
1497 
1498 	nvmet_tcp_restore_socket_callbacks(queue);
1499 	cancel_work_sync(&queue->io_work);
1500 	/* stop accepting incoming data */
1501 	queue->rcv_state = NVMET_TCP_RECV_ERR;
1502 
1503 	nvmet_tcp_uninit_data_in_cmds(queue);
1504 	nvmet_sq_destroy(&queue->nvme_sq);
1505 	cancel_work_sync(&queue->io_work);
1506 	nvmet_tcp_free_cmd_data_in_buffers(queue);
1507 	sock_release(queue->sock);
1508 	nvmet_tcp_free_cmds(queue);
1509 	if (queue->hdr_digest || queue->data_digest)
1510 		nvmet_tcp_free_crypto(queue);
1511 	ida_free(&nvmet_tcp_queue_ida, queue->idx);
1512 
1513 	page = virt_to_head_page(queue->pf_cache.va);
1514 	__page_frag_cache_drain(page, queue->pf_cache.pagecnt_bias);
1515 	kfree(queue);
1516 }
1517 
1518 static void nvmet_tcp_data_ready(struct sock *sk)
1519 {
1520 	struct nvmet_tcp_queue *queue;
1521 
1522 	trace_sk_data_ready(sk);
1523 
1524 	read_lock_bh(&sk->sk_callback_lock);
1525 	queue = sk->sk_user_data;
1526 	if (likely(queue))
1527 		queue_work_on(queue_cpu(queue), nvmet_tcp_wq, &queue->io_work);
1528 	read_unlock_bh(&sk->sk_callback_lock);
1529 }
1530 
1531 static void nvmet_tcp_write_space(struct sock *sk)
1532 {
1533 	struct nvmet_tcp_queue *queue;
1534 
1535 	read_lock_bh(&sk->sk_callback_lock);
1536 	queue = sk->sk_user_data;
1537 	if (unlikely(!queue))
1538 		goto out;
1539 
1540 	if (unlikely(queue->state == NVMET_TCP_Q_CONNECTING)) {
1541 		queue->write_space(sk);
1542 		goto out;
1543 	}
1544 
1545 	if (sk_stream_is_writeable(sk)) {
1546 		clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1547 		queue_work_on(queue_cpu(queue), nvmet_tcp_wq, &queue->io_work);
1548 	}
1549 out:
1550 	read_unlock_bh(&sk->sk_callback_lock);
1551 }
1552 
1553 static void nvmet_tcp_state_change(struct sock *sk)
1554 {
1555 	struct nvmet_tcp_queue *queue;
1556 
1557 	read_lock_bh(&sk->sk_callback_lock);
1558 	queue = sk->sk_user_data;
1559 	if (!queue)
1560 		goto done;
1561 
1562 	switch (sk->sk_state) {
1563 	case TCP_FIN_WAIT2:
1564 	case TCP_LAST_ACK:
1565 		break;
1566 	case TCP_FIN_WAIT1:
1567 	case TCP_CLOSE_WAIT:
1568 	case TCP_CLOSE:
1569 		/* FALLTHRU */
1570 		nvmet_tcp_schedule_release_queue(queue);
1571 		break;
1572 	default:
1573 		pr_warn("queue %d unhandled state %d\n",
1574 			queue->idx, sk->sk_state);
1575 	}
1576 done:
1577 	read_unlock_bh(&sk->sk_callback_lock);
1578 }
1579 
1580 static int nvmet_tcp_set_queue_sock(struct nvmet_tcp_queue *queue)
1581 {
1582 	struct socket *sock = queue->sock;
1583 	struct inet_sock *inet = inet_sk(sock->sk);
1584 	int ret;
1585 
1586 	ret = kernel_getsockname(sock,
1587 		(struct sockaddr *)&queue->sockaddr);
1588 	if (ret < 0)
1589 		return ret;
1590 
1591 	ret = kernel_getpeername(sock,
1592 		(struct sockaddr *)&queue->sockaddr_peer);
1593 	if (ret < 0)
1594 		return ret;
1595 
1596 	/*
1597 	 * Cleanup whatever is sitting in the TCP transmit queue on socket
1598 	 * close. This is done to prevent stale data from being sent should
1599 	 * the network connection be restored before TCP times out.
1600 	 */
1601 	sock_no_linger(sock->sk);
1602 
1603 	if (so_priority > 0)
1604 		sock_set_priority(sock->sk, so_priority);
1605 
1606 	/* Set socket type of service */
1607 	if (inet->rcv_tos > 0)
1608 		ip_sock_set_tos(sock->sk, inet->rcv_tos);
1609 
1610 	ret = 0;
1611 	write_lock_bh(&sock->sk->sk_callback_lock);
1612 	if (sock->sk->sk_state != TCP_ESTABLISHED) {
1613 		/*
1614 		 * If the socket is already closing, don't even start
1615 		 * consuming it
1616 		 */
1617 		ret = -ENOTCONN;
1618 	} else {
1619 		sock->sk->sk_user_data = queue;
1620 		queue->data_ready = sock->sk->sk_data_ready;
1621 		sock->sk->sk_data_ready = nvmet_tcp_data_ready;
1622 		queue->state_change = sock->sk->sk_state_change;
1623 		sock->sk->sk_state_change = nvmet_tcp_state_change;
1624 		queue->write_space = sock->sk->sk_write_space;
1625 		sock->sk->sk_write_space = nvmet_tcp_write_space;
1626 		if (idle_poll_period_usecs)
1627 			nvmet_tcp_arm_queue_deadline(queue);
1628 		queue_work_on(queue_cpu(queue), nvmet_tcp_wq, &queue->io_work);
1629 	}
1630 	write_unlock_bh(&sock->sk->sk_callback_lock);
1631 
1632 	return ret;
1633 }
1634 
1635 static int nvmet_tcp_alloc_queue(struct nvmet_tcp_port *port,
1636 		struct socket *newsock)
1637 {
1638 	struct nvmet_tcp_queue *queue;
1639 	int ret;
1640 
1641 	queue = kzalloc(sizeof(*queue), GFP_KERNEL);
1642 	if (!queue)
1643 		return -ENOMEM;
1644 
1645 	INIT_WORK(&queue->release_work, nvmet_tcp_release_queue_work);
1646 	INIT_WORK(&queue->io_work, nvmet_tcp_io_work);
1647 	queue->sock = newsock;
1648 	queue->port = port;
1649 	queue->nr_cmds = 0;
1650 	spin_lock_init(&queue->state_lock);
1651 	queue->state = NVMET_TCP_Q_CONNECTING;
1652 	INIT_LIST_HEAD(&queue->free_list);
1653 	init_llist_head(&queue->resp_list);
1654 	INIT_LIST_HEAD(&queue->resp_send_list);
1655 
1656 	queue->idx = ida_alloc(&nvmet_tcp_queue_ida, GFP_KERNEL);
1657 	if (queue->idx < 0) {
1658 		ret = queue->idx;
1659 		goto out_free_queue;
1660 	}
1661 
1662 	ret = nvmet_tcp_alloc_cmd(queue, &queue->connect);
1663 	if (ret)
1664 		goto out_ida_remove;
1665 
1666 	ret = nvmet_sq_init(&queue->nvme_sq);
1667 	if (ret)
1668 		goto out_free_connect;
1669 
1670 	nvmet_prepare_receive_pdu(queue);
1671 
1672 	mutex_lock(&nvmet_tcp_queue_mutex);
1673 	list_add_tail(&queue->queue_list, &nvmet_tcp_queue_list);
1674 	mutex_unlock(&nvmet_tcp_queue_mutex);
1675 
1676 	ret = nvmet_tcp_set_queue_sock(queue);
1677 	if (ret)
1678 		goto out_destroy_sq;
1679 
1680 	return 0;
1681 out_destroy_sq:
1682 	mutex_lock(&nvmet_tcp_queue_mutex);
1683 	list_del_init(&queue->queue_list);
1684 	mutex_unlock(&nvmet_tcp_queue_mutex);
1685 	nvmet_sq_destroy(&queue->nvme_sq);
1686 out_free_connect:
1687 	nvmet_tcp_free_cmd(&queue->connect);
1688 out_ida_remove:
1689 	ida_free(&nvmet_tcp_queue_ida, queue->idx);
1690 out_free_queue:
1691 	kfree(queue);
1692 	return ret;
1693 }
1694 
1695 static void nvmet_tcp_accept_work(struct work_struct *w)
1696 {
1697 	struct nvmet_tcp_port *port =
1698 		container_of(w, struct nvmet_tcp_port, accept_work);
1699 	struct socket *newsock;
1700 	int ret;
1701 
1702 	while (true) {
1703 		ret = kernel_accept(port->sock, &newsock, O_NONBLOCK);
1704 		if (ret < 0) {
1705 			if (ret != -EAGAIN)
1706 				pr_warn("failed to accept err=%d\n", ret);
1707 			return;
1708 		}
1709 		ret = nvmet_tcp_alloc_queue(port, newsock);
1710 		if (ret) {
1711 			pr_err("failed to allocate queue\n");
1712 			sock_release(newsock);
1713 		}
1714 	}
1715 }
1716 
1717 static void nvmet_tcp_listen_data_ready(struct sock *sk)
1718 {
1719 	struct nvmet_tcp_port *port;
1720 
1721 	trace_sk_data_ready(sk);
1722 
1723 	read_lock_bh(&sk->sk_callback_lock);
1724 	port = sk->sk_user_data;
1725 	if (!port)
1726 		goto out;
1727 
1728 	if (sk->sk_state == TCP_LISTEN)
1729 		queue_work(nvmet_wq, &port->accept_work);
1730 out:
1731 	read_unlock_bh(&sk->sk_callback_lock);
1732 }
1733 
1734 static int nvmet_tcp_add_port(struct nvmet_port *nport)
1735 {
1736 	struct nvmet_tcp_port *port;
1737 	__kernel_sa_family_t af;
1738 	int ret;
1739 
1740 	port = kzalloc(sizeof(*port), GFP_KERNEL);
1741 	if (!port)
1742 		return -ENOMEM;
1743 
1744 	switch (nport->disc_addr.adrfam) {
1745 	case NVMF_ADDR_FAMILY_IP4:
1746 		af = AF_INET;
1747 		break;
1748 	case NVMF_ADDR_FAMILY_IP6:
1749 		af = AF_INET6;
1750 		break;
1751 	default:
1752 		pr_err("address family %d not supported\n",
1753 				nport->disc_addr.adrfam);
1754 		ret = -EINVAL;
1755 		goto err_port;
1756 	}
1757 
1758 	ret = inet_pton_with_scope(&init_net, af, nport->disc_addr.traddr,
1759 			nport->disc_addr.trsvcid, &port->addr);
1760 	if (ret) {
1761 		pr_err("malformed ip/port passed: %s:%s\n",
1762 			nport->disc_addr.traddr, nport->disc_addr.trsvcid);
1763 		goto err_port;
1764 	}
1765 
1766 	port->nport = nport;
1767 	INIT_WORK(&port->accept_work, nvmet_tcp_accept_work);
1768 	if (port->nport->inline_data_size < 0)
1769 		port->nport->inline_data_size = NVMET_TCP_DEF_INLINE_DATA_SIZE;
1770 
1771 	ret = sock_create(port->addr.ss_family, SOCK_STREAM,
1772 				IPPROTO_TCP, &port->sock);
1773 	if (ret) {
1774 		pr_err("failed to create a socket\n");
1775 		goto err_port;
1776 	}
1777 
1778 	port->sock->sk->sk_user_data = port;
1779 	port->data_ready = port->sock->sk->sk_data_ready;
1780 	port->sock->sk->sk_data_ready = nvmet_tcp_listen_data_ready;
1781 	sock_set_reuseaddr(port->sock->sk);
1782 	tcp_sock_set_nodelay(port->sock->sk);
1783 	if (so_priority > 0)
1784 		sock_set_priority(port->sock->sk, so_priority);
1785 
1786 	ret = kernel_bind(port->sock, (struct sockaddr *)&port->addr,
1787 			sizeof(port->addr));
1788 	if (ret) {
1789 		pr_err("failed to bind port socket %d\n", ret);
1790 		goto err_sock;
1791 	}
1792 
1793 	ret = kernel_listen(port->sock, 128);
1794 	if (ret) {
1795 		pr_err("failed to listen %d on port sock\n", ret);
1796 		goto err_sock;
1797 	}
1798 
1799 	nport->priv = port;
1800 	pr_info("enabling port %d (%pISpc)\n",
1801 		le16_to_cpu(nport->disc_addr.portid), &port->addr);
1802 
1803 	return 0;
1804 
1805 err_sock:
1806 	sock_release(port->sock);
1807 err_port:
1808 	kfree(port);
1809 	return ret;
1810 }
1811 
1812 static void nvmet_tcp_destroy_port_queues(struct nvmet_tcp_port *port)
1813 {
1814 	struct nvmet_tcp_queue *queue;
1815 
1816 	mutex_lock(&nvmet_tcp_queue_mutex);
1817 	list_for_each_entry(queue, &nvmet_tcp_queue_list, queue_list)
1818 		if (queue->port == port)
1819 			kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1820 	mutex_unlock(&nvmet_tcp_queue_mutex);
1821 }
1822 
1823 static void nvmet_tcp_remove_port(struct nvmet_port *nport)
1824 {
1825 	struct nvmet_tcp_port *port = nport->priv;
1826 
1827 	write_lock_bh(&port->sock->sk->sk_callback_lock);
1828 	port->sock->sk->sk_data_ready = port->data_ready;
1829 	port->sock->sk->sk_user_data = NULL;
1830 	write_unlock_bh(&port->sock->sk->sk_callback_lock);
1831 	cancel_work_sync(&port->accept_work);
1832 	/*
1833 	 * Destroy the remaining queues, which are not belong to any
1834 	 * controller yet.
1835 	 */
1836 	nvmet_tcp_destroy_port_queues(port);
1837 
1838 	sock_release(port->sock);
1839 	kfree(port);
1840 }
1841 
1842 static void nvmet_tcp_delete_ctrl(struct nvmet_ctrl *ctrl)
1843 {
1844 	struct nvmet_tcp_queue *queue;
1845 
1846 	mutex_lock(&nvmet_tcp_queue_mutex);
1847 	list_for_each_entry(queue, &nvmet_tcp_queue_list, queue_list)
1848 		if (queue->nvme_sq.ctrl == ctrl)
1849 			kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1850 	mutex_unlock(&nvmet_tcp_queue_mutex);
1851 }
1852 
1853 static u16 nvmet_tcp_install_queue(struct nvmet_sq *sq)
1854 {
1855 	struct nvmet_tcp_queue *queue =
1856 		container_of(sq, struct nvmet_tcp_queue, nvme_sq);
1857 
1858 	if (sq->qid == 0) {
1859 		/* Let inflight controller teardown complete */
1860 		flush_workqueue(nvmet_wq);
1861 	}
1862 
1863 	queue->nr_cmds = sq->size * 2;
1864 	if (nvmet_tcp_alloc_cmds(queue))
1865 		return NVME_SC_INTERNAL;
1866 	return 0;
1867 }
1868 
1869 static void nvmet_tcp_disc_port_addr(struct nvmet_req *req,
1870 		struct nvmet_port *nport, char *traddr)
1871 {
1872 	struct nvmet_tcp_port *port = nport->priv;
1873 
1874 	if (inet_addr_is_any((struct sockaddr *)&port->addr)) {
1875 		struct nvmet_tcp_cmd *cmd =
1876 			container_of(req, struct nvmet_tcp_cmd, req);
1877 		struct nvmet_tcp_queue *queue = cmd->queue;
1878 
1879 		sprintf(traddr, "%pISc", (struct sockaddr *)&queue->sockaddr);
1880 	} else {
1881 		memcpy(traddr, nport->disc_addr.traddr, NVMF_TRADDR_SIZE);
1882 	}
1883 }
1884 
1885 static const struct nvmet_fabrics_ops nvmet_tcp_ops = {
1886 	.owner			= THIS_MODULE,
1887 	.type			= NVMF_TRTYPE_TCP,
1888 	.msdbd			= 1,
1889 	.add_port		= nvmet_tcp_add_port,
1890 	.remove_port		= nvmet_tcp_remove_port,
1891 	.queue_response		= nvmet_tcp_queue_response,
1892 	.delete_ctrl		= nvmet_tcp_delete_ctrl,
1893 	.install_queue		= nvmet_tcp_install_queue,
1894 	.disc_traddr		= nvmet_tcp_disc_port_addr,
1895 };
1896 
1897 static int __init nvmet_tcp_init(void)
1898 {
1899 	int ret;
1900 
1901 	nvmet_tcp_wq = alloc_workqueue("nvmet_tcp_wq",
1902 				WQ_MEM_RECLAIM | WQ_HIGHPRI, 0);
1903 	if (!nvmet_tcp_wq)
1904 		return -ENOMEM;
1905 
1906 	ret = nvmet_register_transport(&nvmet_tcp_ops);
1907 	if (ret)
1908 		goto err;
1909 
1910 	return 0;
1911 err:
1912 	destroy_workqueue(nvmet_tcp_wq);
1913 	return ret;
1914 }
1915 
1916 static void __exit nvmet_tcp_exit(void)
1917 {
1918 	struct nvmet_tcp_queue *queue;
1919 
1920 	nvmet_unregister_transport(&nvmet_tcp_ops);
1921 
1922 	flush_workqueue(nvmet_wq);
1923 	mutex_lock(&nvmet_tcp_queue_mutex);
1924 	list_for_each_entry(queue, &nvmet_tcp_queue_list, queue_list)
1925 		kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1926 	mutex_unlock(&nvmet_tcp_queue_mutex);
1927 	flush_workqueue(nvmet_wq);
1928 
1929 	destroy_workqueue(nvmet_tcp_wq);
1930 	ida_destroy(&nvmet_tcp_queue_ida);
1931 }
1932 
1933 module_init(nvmet_tcp_init);
1934 module_exit(nvmet_tcp_exit);
1935 
1936 MODULE_LICENSE("GPL v2");
1937 MODULE_ALIAS("nvmet-transport-3"); /* 3 == NVMF_TRTYPE_TCP */
1938