xref: /linux/drivers/nvme/target/tcp.c (revision 63307d015b91e626c97bb82e88054af3d0b74643)
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 
18 #include "nvmet.h"
19 
20 #define NVMET_TCP_DEF_INLINE_DATA_SIZE	(4 * PAGE_SIZE)
21 
22 #define NVMET_TCP_RECV_BUDGET		8
23 #define NVMET_TCP_SEND_BUDGET		8
24 #define NVMET_TCP_IO_WORK_BUDGET	64
25 
26 enum nvmet_tcp_send_state {
27 	NVMET_TCP_SEND_DATA_PDU,
28 	NVMET_TCP_SEND_DATA,
29 	NVMET_TCP_SEND_R2T,
30 	NVMET_TCP_SEND_DDGST,
31 	NVMET_TCP_SEND_RESPONSE
32 };
33 
34 enum nvmet_tcp_recv_state {
35 	NVMET_TCP_RECV_PDU,
36 	NVMET_TCP_RECV_DATA,
37 	NVMET_TCP_RECV_DDGST,
38 	NVMET_TCP_RECV_ERR,
39 };
40 
41 enum {
42 	NVMET_TCP_F_INIT_FAILED = (1 << 0),
43 };
44 
45 struct nvmet_tcp_cmd {
46 	struct nvmet_tcp_queue		*queue;
47 	struct nvmet_req		req;
48 
49 	struct nvme_tcp_cmd_pdu		*cmd_pdu;
50 	struct nvme_tcp_rsp_pdu		*rsp_pdu;
51 	struct nvme_tcp_data_pdu	*data_pdu;
52 	struct nvme_tcp_r2t_pdu		*r2t_pdu;
53 
54 	u32				rbytes_done;
55 	u32				wbytes_done;
56 
57 	u32				pdu_len;
58 	u32				pdu_recv;
59 	int				sg_idx;
60 	int				nr_mapped;
61 	struct msghdr			recv_msg;
62 	struct kvec			*iov;
63 	u32				flags;
64 
65 	struct list_head		entry;
66 	struct llist_node		lentry;
67 
68 	/* send state */
69 	u32				offset;
70 	struct scatterlist		*cur_sg;
71 	enum nvmet_tcp_send_state	state;
72 
73 	__le32				exp_ddgst;
74 	__le32				recv_ddgst;
75 };
76 
77 enum nvmet_tcp_queue_state {
78 	NVMET_TCP_Q_CONNECTING,
79 	NVMET_TCP_Q_LIVE,
80 	NVMET_TCP_Q_DISCONNECTING,
81 };
82 
83 struct nvmet_tcp_queue {
84 	struct socket		*sock;
85 	struct nvmet_tcp_port	*port;
86 	struct work_struct	io_work;
87 	int			cpu;
88 	struct nvmet_cq		nvme_cq;
89 	struct nvmet_sq		nvme_sq;
90 
91 	/* send state */
92 	struct nvmet_tcp_cmd	*cmds;
93 	unsigned int		nr_cmds;
94 	struct list_head	free_list;
95 	struct llist_head	resp_list;
96 	struct list_head	resp_send_list;
97 	int			send_list_len;
98 	struct nvmet_tcp_cmd	*snd_cmd;
99 
100 	/* recv state */
101 	int			offset;
102 	int			left;
103 	enum nvmet_tcp_recv_state rcv_state;
104 	struct nvmet_tcp_cmd	*cmd;
105 	union nvme_tcp_pdu	pdu;
106 
107 	/* digest state */
108 	bool			hdr_digest;
109 	bool			data_digest;
110 	struct ahash_request	*snd_hash;
111 	struct ahash_request	*rcv_hash;
112 
113 	spinlock_t		state_lock;
114 	enum nvmet_tcp_queue_state state;
115 
116 	struct sockaddr_storage	sockaddr;
117 	struct sockaddr_storage	sockaddr_peer;
118 	struct work_struct	release_work;
119 
120 	int			idx;
121 	struct list_head	queue_list;
122 
123 	struct nvmet_tcp_cmd	connect;
124 
125 	struct page_frag_cache	pf_cache;
126 
127 	void (*data_ready)(struct sock *);
128 	void (*state_change)(struct sock *);
129 	void (*write_space)(struct sock *);
130 };
131 
132 struct nvmet_tcp_port {
133 	struct socket		*sock;
134 	struct work_struct	accept_work;
135 	struct nvmet_port	*nport;
136 	struct sockaddr_storage addr;
137 	int			last_cpu;
138 	void (*data_ready)(struct sock *);
139 };
140 
141 static DEFINE_IDA(nvmet_tcp_queue_ida);
142 static LIST_HEAD(nvmet_tcp_queue_list);
143 static DEFINE_MUTEX(nvmet_tcp_queue_mutex);
144 
145 static struct workqueue_struct *nvmet_tcp_wq;
146 static struct nvmet_fabrics_ops nvmet_tcp_ops;
147 static void nvmet_tcp_free_cmd(struct nvmet_tcp_cmd *c);
148 static void nvmet_tcp_finish_cmd(struct nvmet_tcp_cmd *cmd);
149 
150 static inline u16 nvmet_tcp_cmd_tag(struct nvmet_tcp_queue *queue,
151 		struct nvmet_tcp_cmd *cmd)
152 {
153 	return cmd - queue->cmds;
154 }
155 
156 static inline bool nvmet_tcp_has_data_in(struct nvmet_tcp_cmd *cmd)
157 {
158 	return nvme_is_write(cmd->req.cmd) &&
159 		cmd->rbytes_done < cmd->req.transfer_len;
160 }
161 
162 static inline bool nvmet_tcp_need_data_in(struct nvmet_tcp_cmd *cmd)
163 {
164 	return nvmet_tcp_has_data_in(cmd) && !cmd->req.cqe->status;
165 }
166 
167 static inline bool nvmet_tcp_need_data_out(struct nvmet_tcp_cmd *cmd)
168 {
169 	return !nvme_is_write(cmd->req.cmd) &&
170 		cmd->req.transfer_len > 0 &&
171 		!cmd->req.cqe->status;
172 }
173 
174 static inline bool nvmet_tcp_has_inline_data(struct nvmet_tcp_cmd *cmd)
175 {
176 	return nvme_is_write(cmd->req.cmd) && cmd->pdu_len &&
177 		!cmd->rbytes_done;
178 }
179 
180 static inline struct nvmet_tcp_cmd *
181 nvmet_tcp_get_cmd(struct nvmet_tcp_queue *queue)
182 {
183 	struct nvmet_tcp_cmd *cmd;
184 
185 	cmd = list_first_entry_or_null(&queue->free_list,
186 				struct nvmet_tcp_cmd, entry);
187 	if (!cmd)
188 		return NULL;
189 	list_del_init(&cmd->entry);
190 
191 	cmd->rbytes_done = cmd->wbytes_done = 0;
192 	cmd->pdu_len = 0;
193 	cmd->pdu_recv = 0;
194 	cmd->iov = NULL;
195 	cmd->flags = 0;
196 	return cmd;
197 }
198 
199 static inline void nvmet_tcp_put_cmd(struct nvmet_tcp_cmd *cmd)
200 {
201 	if (unlikely(cmd == &cmd->queue->connect))
202 		return;
203 
204 	list_add_tail(&cmd->entry, &cmd->queue->free_list);
205 }
206 
207 static inline u8 nvmet_tcp_hdgst_len(struct nvmet_tcp_queue *queue)
208 {
209 	return queue->hdr_digest ? NVME_TCP_DIGEST_LENGTH : 0;
210 }
211 
212 static inline u8 nvmet_tcp_ddgst_len(struct nvmet_tcp_queue *queue)
213 {
214 	return queue->data_digest ? NVME_TCP_DIGEST_LENGTH : 0;
215 }
216 
217 static inline void nvmet_tcp_hdgst(struct ahash_request *hash,
218 		void *pdu, size_t len)
219 {
220 	struct scatterlist sg;
221 
222 	sg_init_one(&sg, pdu, len);
223 	ahash_request_set_crypt(hash, &sg, pdu + len, len);
224 	crypto_ahash_digest(hash);
225 }
226 
227 static int nvmet_tcp_verify_hdgst(struct nvmet_tcp_queue *queue,
228 	void *pdu, size_t len)
229 {
230 	struct nvme_tcp_hdr *hdr = pdu;
231 	__le32 recv_digest;
232 	__le32 exp_digest;
233 
234 	if (unlikely(!(hdr->flags & NVME_TCP_F_HDGST))) {
235 		pr_err("queue %d: header digest enabled but no header digest\n",
236 			queue->idx);
237 		return -EPROTO;
238 	}
239 
240 	recv_digest = *(__le32 *)(pdu + hdr->hlen);
241 	nvmet_tcp_hdgst(queue->rcv_hash, pdu, len);
242 	exp_digest = *(__le32 *)(pdu + hdr->hlen);
243 	if (recv_digest != exp_digest) {
244 		pr_err("queue %d: header digest error: recv %#x expected %#x\n",
245 			queue->idx, le32_to_cpu(recv_digest),
246 			le32_to_cpu(exp_digest));
247 		return -EPROTO;
248 	}
249 
250 	return 0;
251 }
252 
253 static int nvmet_tcp_check_ddgst(struct nvmet_tcp_queue *queue, void *pdu)
254 {
255 	struct nvme_tcp_hdr *hdr = pdu;
256 	u8 digest_len = nvmet_tcp_hdgst_len(queue);
257 	u32 len;
258 
259 	len = le32_to_cpu(hdr->plen) - hdr->hlen -
260 		(hdr->flags & NVME_TCP_F_HDGST ? digest_len : 0);
261 
262 	if (unlikely(len && !(hdr->flags & NVME_TCP_F_DDGST))) {
263 		pr_err("queue %d: data digest flag is cleared\n", queue->idx);
264 		return -EPROTO;
265 	}
266 
267 	return 0;
268 }
269 
270 static void nvmet_tcp_unmap_pdu_iovec(struct nvmet_tcp_cmd *cmd)
271 {
272 	struct scatterlist *sg;
273 	int i;
274 
275 	sg = &cmd->req.sg[cmd->sg_idx];
276 
277 	for (i = 0; i < cmd->nr_mapped; i++)
278 		kunmap(sg_page(&sg[i]));
279 }
280 
281 static void nvmet_tcp_map_pdu_iovec(struct nvmet_tcp_cmd *cmd)
282 {
283 	struct kvec *iov = cmd->iov;
284 	struct scatterlist *sg;
285 	u32 length, offset, sg_offset;
286 
287 	length = cmd->pdu_len;
288 	cmd->nr_mapped = DIV_ROUND_UP(length, PAGE_SIZE);
289 	offset = cmd->rbytes_done;
290 	cmd->sg_idx = DIV_ROUND_UP(offset, PAGE_SIZE);
291 	sg_offset = offset % PAGE_SIZE;
292 	sg = &cmd->req.sg[cmd->sg_idx];
293 
294 	while (length) {
295 		u32 iov_len = min_t(u32, length, sg->length - sg_offset);
296 
297 		iov->iov_base = kmap(sg_page(sg)) + sg->offset + sg_offset;
298 		iov->iov_len = iov_len;
299 
300 		length -= iov_len;
301 		sg = sg_next(sg);
302 		iov++;
303 	}
304 
305 	iov_iter_kvec(&cmd->recv_msg.msg_iter, READ, cmd->iov,
306 		cmd->nr_mapped, cmd->pdu_len);
307 }
308 
309 static void nvmet_tcp_fatal_error(struct nvmet_tcp_queue *queue)
310 {
311 	queue->rcv_state = NVMET_TCP_RECV_ERR;
312 	if (queue->nvme_sq.ctrl)
313 		nvmet_ctrl_fatal_error(queue->nvme_sq.ctrl);
314 	else
315 		kernel_sock_shutdown(queue->sock, SHUT_RDWR);
316 }
317 
318 static int nvmet_tcp_map_data(struct nvmet_tcp_cmd *cmd)
319 {
320 	struct nvme_sgl_desc *sgl = &cmd->req.cmd->common.dptr.sgl;
321 	u32 len = le32_to_cpu(sgl->length);
322 
323 	if (!cmd->req.data_len)
324 		return 0;
325 
326 	if (sgl->type == ((NVME_SGL_FMT_DATA_DESC << 4) |
327 			  NVME_SGL_FMT_OFFSET)) {
328 		if (!nvme_is_write(cmd->req.cmd))
329 			return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
330 
331 		if (len > cmd->req.port->inline_data_size)
332 			return NVME_SC_SGL_INVALID_OFFSET | NVME_SC_DNR;
333 		cmd->pdu_len = len;
334 	}
335 	cmd->req.transfer_len += len;
336 
337 	cmd->req.sg = sgl_alloc(len, GFP_KERNEL, &cmd->req.sg_cnt);
338 	if (!cmd->req.sg)
339 		return NVME_SC_INTERNAL;
340 	cmd->cur_sg = cmd->req.sg;
341 
342 	if (nvmet_tcp_has_data_in(cmd)) {
343 		cmd->iov = kmalloc_array(cmd->req.sg_cnt,
344 				sizeof(*cmd->iov), GFP_KERNEL);
345 		if (!cmd->iov)
346 			goto err;
347 	}
348 
349 	return 0;
350 err:
351 	sgl_free(cmd->req.sg);
352 	return NVME_SC_INTERNAL;
353 }
354 
355 static void nvmet_tcp_ddgst(struct ahash_request *hash,
356 		struct nvmet_tcp_cmd *cmd)
357 {
358 	ahash_request_set_crypt(hash, cmd->req.sg,
359 		(void *)&cmd->exp_ddgst, cmd->req.transfer_len);
360 	crypto_ahash_digest(hash);
361 }
362 
363 static void nvmet_setup_c2h_data_pdu(struct nvmet_tcp_cmd *cmd)
364 {
365 	struct nvme_tcp_data_pdu *pdu = cmd->data_pdu;
366 	struct nvmet_tcp_queue *queue = cmd->queue;
367 	u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
368 	u8 ddgst = nvmet_tcp_ddgst_len(cmd->queue);
369 
370 	cmd->offset = 0;
371 	cmd->state = NVMET_TCP_SEND_DATA_PDU;
372 
373 	pdu->hdr.type = nvme_tcp_c2h_data;
374 	pdu->hdr.flags = NVME_TCP_F_DATA_LAST | (queue->nvme_sq.sqhd_disabled ?
375 						NVME_TCP_F_DATA_SUCCESS : 0);
376 	pdu->hdr.hlen = sizeof(*pdu);
377 	pdu->hdr.pdo = pdu->hdr.hlen + hdgst;
378 	pdu->hdr.plen =
379 		cpu_to_le32(pdu->hdr.hlen + hdgst +
380 				cmd->req.transfer_len + ddgst);
381 	pdu->command_id = cmd->req.cqe->command_id;
382 	pdu->data_length = cpu_to_le32(cmd->req.transfer_len);
383 	pdu->data_offset = cpu_to_le32(cmd->wbytes_done);
384 
385 	if (queue->data_digest) {
386 		pdu->hdr.flags |= NVME_TCP_F_DDGST;
387 		nvmet_tcp_ddgst(queue->snd_hash, cmd);
388 	}
389 
390 	if (cmd->queue->hdr_digest) {
391 		pdu->hdr.flags |= NVME_TCP_F_HDGST;
392 		nvmet_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
393 	}
394 }
395 
396 static void nvmet_setup_r2t_pdu(struct nvmet_tcp_cmd *cmd)
397 {
398 	struct nvme_tcp_r2t_pdu *pdu = cmd->r2t_pdu;
399 	struct nvmet_tcp_queue *queue = cmd->queue;
400 	u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
401 
402 	cmd->offset = 0;
403 	cmd->state = NVMET_TCP_SEND_R2T;
404 
405 	pdu->hdr.type = nvme_tcp_r2t;
406 	pdu->hdr.flags = 0;
407 	pdu->hdr.hlen = sizeof(*pdu);
408 	pdu->hdr.pdo = 0;
409 	pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst);
410 
411 	pdu->command_id = cmd->req.cmd->common.command_id;
412 	pdu->ttag = nvmet_tcp_cmd_tag(cmd->queue, cmd);
413 	pdu->r2t_length = cpu_to_le32(cmd->req.transfer_len - cmd->rbytes_done);
414 	pdu->r2t_offset = cpu_to_le32(cmd->rbytes_done);
415 	if (cmd->queue->hdr_digest) {
416 		pdu->hdr.flags |= NVME_TCP_F_HDGST;
417 		nvmet_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
418 	}
419 }
420 
421 static void nvmet_setup_response_pdu(struct nvmet_tcp_cmd *cmd)
422 {
423 	struct nvme_tcp_rsp_pdu *pdu = cmd->rsp_pdu;
424 	struct nvmet_tcp_queue *queue = cmd->queue;
425 	u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
426 
427 	cmd->offset = 0;
428 	cmd->state = NVMET_TCP_SEND_RESPONSE;
429 
430 	pdu->hdr.type = nvme_tcp_rsp;
431 	pdu->hdr.flags = 0;
432 	pdu->hdr.hlen = sizeof(*pdu);
433 	pdu->hdr.pdo = 0;
434 	pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst);
435 	if (cmd->queue->hdr_digest) {
436 		pdu->hdr.flags |= NVME_TCP_F_HDGST;
437 		nvmet_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
438 	}
439 }
440 
441 static void nvmet_tcp_process_resp_list(struct nvmet_tcp_queue *queue)
442 {
443 	struct llist_node *node;
444 
445 	node = llist_del_all(&queue->resp_list);
446 	if (!node)
447 		return;
448 
449 	while (node) {
450 		struct nvmet_tcp_cmd *cmd = llist_entry(node,
451 					struct nvmet_tcp_cmd, lentry);
452 
453 		list_add(&cmd->entry, &queue->resp_send_list);
454 		node = node->next;
455 		queue->send_list_len++;
456 	}
457 }
458 
459 static struct nvmet_tcp_cmd *nvmet_tcp_fetch_cmd(struct nvmet_tcp_queue *queue)
460 {
461 	queue->snd_cmd = list_first_entry_or_null(&queue->resp_send_list,
462 				struct nvmet_tcp_cmd, entry);
463 	if (!queue->snd_cmd) {
464 		nvmet_tcp_process_resp_list(queue);
465 		queue->snd_cmd =
466 			list_first_entry_or_null(&queue->resp_send_list,
467 					struct nvmet_tcp_cmd, entry);
468 		if (unlikely(!queue->snd_cmd))
469 			return NULL;
470 	}
471 
472 	list_del_init(&queue->snd_cmd->entry);
473 	queue->send_list_len--;
474 
475 	if (nvmet_tcp_need_data_out(queue->snd_cmd))
476 		nvmet_setup_c2h_data_pdu(queue->snd_cmd);
477 	else if (nvmet_tcp_need_data_in(queue->snd_cmd))
478 		nvmet_setup_r2t_pdu(queue->snd_cmd);
479 	else
480 		nvmet_setup_response_pdu(queue->snd_cmd);
481 
482 	return queue->snd_cmd;
483 }
484 
485 static void nvmet_tcp_queue_response(struct nvmet_req *req)
486 {
487 	struct nvmet_tcp_cmd *cmd =
488 		container_of(req, struct nvmet_tcp_cmd, req);
489 	struct nvmet_tcp_queue	*queue = cmd->queue;
490 
491 	llist_add(&cmd->lentry, &queue->resp_list);
492 	queue_work_on(cmd->queue->cpu, nvmet_tcp_wq, &cmd->queue->io_work);
493 }
494 
495 static int nvmet_try_send_data_pdu(struct nvmet_tcp_cmd *cmd)
496 {
497 	u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
498 	int left = sizeof(*cmd->data_pdu) - cmd->offset + hdgst;
499 	int ret;
500 
501 	ret = kernel_sendpage(cmd->queue->sock, virt_to_page(cmd->data_pdu),
502 			offset_in_page(cmd->data_pdu) + cmd->offset,
503 			left, MSG_DONTWAIT | MSG_MORE);
504 	if (ret <= 0)
505 		return ret;
506 
507 	cmd->offset += ret;
508 	left -= ret;
509 
510 	if (left)
511 		return -EAGAIN;
512 
513 	cmd->state = NVMET_TCP_SEND_DATA;
514 	cmd->offset  = 0;
515 	return 1;
516 }
517 
518 static int nvmet_try_send_data(struct nvmet_tcp_cmd *cmd)
519 {
520 	struct nvmet_tcp_queue *queue = cmd->queue;
521 	int ret;
522 
523 	while (cmd->cur_sg) {
524 		struct page *page = sg_page(cmd->cur_sg);
525 		u32 left = cmd->cur_sg->length - cmd->offset;
526 
527 		ret = kernel_sendpage(cmd->queue->sock, page, cmd->offset,
528 					left, MSG_DONTWAIT | MSG_MORE);
529 		if (ret <= 0)
530 			return ret;
531 
532 		cmd->offset += ret;
533 		cmd->wbytes_done += ret;
534 
535 		/* Done with sg?*/
536 		if (cmd->offset == cmd->cur_sg->length) {
537 			cmd->cur_sg = sg_next(cmd->cur_sg);
538 			cmd->offset = 0;
539 		}
540 	}
541 
542 	if (queue->data_digest) {
543 		cmd->state = NVMET_TCP_SEND_DDGST;
544 		cmd->offset = 0;
545 	} else {
546 		if (queue->nvme_sq.sqhd_disabled) {
547 			cmd->queue->snd_cmd = NULL;
548 			nvmet_tcp_put_cmd(cmd);
549 		} else {
550 			nvmet_setup_response_pdu(cmd);
551 		}
552 	}
553 
554 	if (queue->nvme_sq.sqhd_disabled) {
555 		kfree(cmd->iov);
556 		sgl_free(cmd->req.sg);
557 	}
558 
559 	return 1;
560 
561 }
562 
563 static int nvmet_try_send_response(struct nvmet_tcp_cmd *cmd,
564 		bool last_in_batch)
565 {
566 	u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
567 	int left = sizeof(*cmd->rsp_pdu) - cmd->offset + hdgst;
568 	int flags = MSG_DONTWAIT;
569 	int ret;
570 
571 	if (!last_in_batch && cmd->queue->send_list_len)
572 		flags |= MSG_MORE;
573 	else
574 		flags |= MSG_EOR;
575 
576 	ret = kernel_sendpage(cmd->queue->sock, virt_to_page(cmd->rsp_pdu),
577 		offset_in_page(cmd->rsp_pdu) + cmd->offset, left, flags);
578 	if (ret <= 0)
579 		return ret;
580 	cmd->offset += ret;
581 	left -= ret;
582 
583 	if (left)
584 		return -EAGAIN;
585 
586 	kfree(cmd->iov);
587 	sgl_free(cmd->req.sg);
588 	cmd->queue->snd_cmd = NULL;
589 	nvmet_tcp_put_cmd(cmd);
590 	return 1;
591 }
592 
593 static int nvmet_try_send_r2t(struct nvmet_tcp_cmd *cmd, bool last_in_batch)
594 {
595 	u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
596 	int left = sizeof(*cmd->r2t_pdu) - cmd->offset + hdgst;
597 	int flags = MSG_DONTWAIT;
598 	int ret;
599 
600 	if (!last_in_batch && cmd->queue->send_list_len)
601 		flags |= MSG_MORE;
602 	else
603 		flags |= MSG_EOR;
604 
605 	ret = kernel_sendpage(cmd->queue->sock, virt_to_page(cmd->r2t_pdu),
606 		offset_in_page(cmd->r2t_pdu) + cmd->offset, left, flags);
607 	if (ret <= 0)
608 		return ret;
609 	cmd->offset += ret;
610 	left -= ret;
611 
612 	if (left)
613 		return -EAGAIN;
614 
615 	cmd->queue->snd_cmd = NULL;
616 	return 1;
617 }
618 
619 static int nvmet_try_send_ddgst(struct nvmet_tcp_cmd *cmd)
620 {
621 	struct nvmet_tcp_queue *queue = cmd->queue;
622 	struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
623 	struct kvec iov = {
624 		.iov_base = &cmd->exp_ddgst + cmd->offset,
625 		.iov_len = NVME_TCP_DIGEST_LENGTH - cmd->offset
626 	};
627 	int ret;
628 
629 	ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
630 	if (unlikely(ret <= 0))
631 		return ret;
632 
633 	cmd->offset += ret;
634 
635 	if (queue->nvme_sq.sqhd_disabled) {
636 		cmd->queue->snd_cmd = NULL;
637 		nvmet_tcp_put_cmd(cmd);
638 	} else {
639 		nvmet_setup_response_pdu(cmd);
640 	}
641 	return 1;
642 }
643 
644 static int nvmet_tcp_try_send_one(struct nvmet_tcp_queue *queue,
645 		bool last_in_batch)
646 {
647 	struct nvmet_tcp_cmd *cmd = queue->snd_cmd;
648 	int ret = 0;
649 
650 	if (!cmd || queue->state == NVMET_TCP_Q_DISCONNECTING) {
651 		cmd = nvmet_tcp_fetch_cmd(queue);
652 		if (unlikely(!cmd))
653 			return 0;
654 	}
655 
656 	if (cmd->state == NVMET_TCP_SEND_DATA_PDU) {
657 		ret = nvmet_try_send_data_pdu(cmd);
658 		if (ret <= 0)
659 			goto done_send;
660 	}
661 
662 	if (cmd->state == NVMET_TCP_SEND_DATA) {
663 		ret = nvmet_try_send_data(cmd);
664 		if (ret <= 0)
665 			goto done_send;
666 	}
667 
668 	if (cmd->state == NVMET_TCP_SEND_DDGST) {
669 		ret = nvmet_try_send_ddgst(cmd);
670 		if (ret <= 0)
671 			goto done_send;
672 	}
673 
674 	if (cmd->state == NVMET_TCP_SEND_R2T) {
675 		ret = nvmet_try_send_r2t(cmd, last_in_batch);
676 		if (ret <= 0)
677 			goto done_send;
678 	}
679 
680 	if (cmd->state == NVMET_TCP_SEND_RESPONSE)
681 		ret = nvmet_try_send_response(cmd, last_in_batch);
682 
683 done_send:
684 	if (ret < 0) {
685 		if (ret == -EAGAIN)
686 			return 0;
687 		return ret;
688 	}
689 
690 	return 1;
691 }
692 
693 static int nvmet_tcp_try_send(struct nvmet_tcp_queue *queue,
694 		int budget, int *sends)
695 {
696 	int i, ret = 0;
697 
698 	for (i = 0; i < budget; i++) {
699 		ret = nvmet_tcp_try_send_one(queue, i == budget - 1);
700 		if (ret <= 0)
701 			break;
702 		(*sends)++;
703 	}
704 
705 	return ret;
706 }
707 
708 static void nvmet_prepare_receive_pdu(struct nvmet_tcp_queue *queue)
709 {
710 	queue->offset = 0;
711 	queue->left = sizeof(struct nvme_tcp_hdr);
712 	queue->cmd = NULL;
713 	queue->rcv_state = NVMET_TCP_RECV_PDU;
714 }
715 
716 static void nvmet_tcp_free_crypto(struct nvmet_tcp_queue *queue)
717 {
718 	struct crypto_ahash *tfm = crypto_ahash_reqtfm(queue->rcv_hash);
719 
720 	ahash_request_free(queue->rcv_hash);
721 	ahash_request_free(queue->snd_hash);
722 	crypto_free_ahash(tfm);
723 }
724 
725 static int nvmet_tcp_alloc_crypto(struct nvmet_tcp_queue *queue)
726 {
727 	struct crypto_ahash *tfm;
728 
729 	tfm = crypto_alloc_ahash("crc32c", 0, CRYPTO_ALG_ASYNC);
730 	if (IS_ERR(tfm))
731 		return PTR_ERR(tfm);
732 
733 	queue->snd_hash = ahash_request_alloc(tfm, GFP_KERNEL);
734 	if (!queue->snd_hash)
735 		goto free_tfm;
736 	ahash_request_set_callback(queue->snd_hash, 0, NULL, NULL);
737 
738 	queue->rcv_hash = ahash_request_alloc(tfm, GFP_KERNEL);
739 	if (!queue->rcv_hash)
740 		goto free_snd_hash;
741 	ahash_request_set_callback(queue->rcv_hash, 0, NULL, NULL);
742 
743 	return 0;
744 free_snd_hash:
745 	ahash_request_free(queue->snd_hash);
746 free_tfm:
747 	crypto_free_ahash(tfm);
748 	return -ENOMEM;
749 }
750 
751 
752 static int nvmet_tcp_handle_icreq(struct nvmet_tcp_queue *queue)
753 {
754 	struct nvme_tcp_icreq_pdu *icreq = &queue->pdu.icreq;
755 	struct nvme_tcp_icresp_pdu *icresp = &queue->pdu.icresp;
756 	struct msghdr msg = {};
757 	struct kvec iov;
758 	int ret;
759 
760 	if (le32_to_cpu(icreq->hdr.plen) != sizeof(struct nvme_tcp_icreq_pdu)) {
761 		pr_err("bad nvme-tcp pdu length (%d)\n",
762 			le32_to_cpu(icreq->hdr.plen));
763 		nvmet_tcp_fatal_error(queue);
764 	}
765 
766 	if (icreq->pfv != NVME_TCP_PFV_1_0) {
767 		pr_err("queue %d: bad pfv %d\n", queue->idx, icreq->pfv);
768 		return -EPROTO;
769 	}
770 
771 	if (icreq->hpda != 0) {
772 		pr_err("queue %d: unsupported hpda %d\n", queue->idx,
773 			icreq->hpda);
774 		return -EPROTO;
775 	}
776 
777 	queue->hdr_digest = !!(icreq->digest & NVME_TCP_HDR_DIGEST_ENABLE);
778 	queue->data_digest = !!(icreq->digest & NVME_TCP_DATA_DIGEST_ENABLE);
779 	if (queue->hdr_digest || queue->data_digest) {
780 		ret = nvmet_tcp_alloc_crypto(queue);
781 		if (ret)
782 			return ret;
783 	}
784 
785 	memset(icresp, 0, sizeof(*icresp));
786 	icresp->hdr.type = nvme_tcp_icresp;
787 	icresp->hdr.hlen = sizeof(*icresp);
788 	icresp->hdr.pdo = 0;
789 	icresp->hdr.plen = cpu_to_le32(icresp->hdr.hlen);
790 	icresp->pfv = cpu_to_le16(NVME_TCP_PFV_1_0);
791 	icresp->maxdata = cpu_to_le32(0xffff); /* FIXME: support r2t */
792 	icresp->cpda = 0;
793 	if (queue->hdr_digest)
794 		icresp->digest |= NVME_TCP_HDR_DIGEST_ENABLE;
795 	if (queue->data_digest)
796 		icresp->digest |= NVME_TCP_DATA_DIGEST_ENABLE;
797 
798 	iov.iov_base = icresp;
799 	iov.iov_len = sizeof(*icresp);
800 	ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
801 	if (ret < 0)
802 		goto free_crypto;
803 
804 	queue->state = NVMET_TCP_Q_LIVE;
805 	nvmet_prepare_receive_pdu(queue);
806 	return 0;
807 free_crypto:
808 	if (queue->hdr_digest || queue->data_digest)
809 		nvmet_tcp_free_crypto(queue);
810 	return ret;
811 }
812 
813 static void nvmet_tcp_handle_req_failure(struct nvmet_tcp_queue *queue,
814 		struct nvmet_tcp_cmd *cmd, struct nvmet_req *req)
815 {
816 	int ret;
817 
818 	/* recover the expected data transfer length */
819 	req->data_len = le32_to_cpu(req->cmd->common.dptr.sgl.length);
820 
821 	if (!nvme_is_write(cmd->req.cmd) ||
822 	    req->data_len > cmd->req.port->inline_data_size) {
823 		nvmet_prepare_receive_pdu(queue);
824 		return;
825 	}
826 
827 	ret = nvmet_tcp_map_data(cmd);
828 	if (unlikely(ret)) {
829 		pr_err("queue %d: failed to map data\n", queue->idx);
830 		nvmet_tcp_fatal_error(queue);
831 		return;
832 	}
833 
834 	queue->rcv_state = NVMET_TCP_RECV_DATA;
835 	nvmet_tcp_map_pdu_iovec(cmd);
836 	cmd->flags |= NVMET_TCP_F_INIT_FAILED;
837 }
838 
839 static int nvmet_tcp_handle_h2c_data_pdu(struct nvmet_tcp_queue *queue)
840 {
841 	struct nvme_tcp_data_pdu *data = &queue->pdu.data;
842 	struct nvmet_tcp_cmd *cmd;
843 
844 	cmd = &queue->cmds[data->ttag];
845 
846 	if (le32_to_cpu(data->data_offset) != cmd->rbytes_done) {
847 		pr_err("ttag %u unexpected data offset %u (expected %u)\n",
848 			data->ttag, le32_to_cpu(data->data_offset),
849 			cmd->rbytes_done);
850 		/* FIXME: use path and transport errors */
851 		nvmet_req_complete(&cmd->req,
852 			NVME_SC_INVALID_FIELD | NVME_SC_DNR);
853 		return -EPROTO;
854 	}
855 
856 	cmd->pdu_len = le32_to_cpu(data->data_length);
857 	cmd->pdu_recv = 0;
858 	nvmet_tcp_map_pdu_iovec(cmd);
859 	queue->cmd = cmd;
860 	queue->rcv_state = NVMET_TCP_RECV_DATA;
861 
862 	return 0;
863 }
864 
865 static int nvmet_tcp_done_recv_pdu(struct nvmet_tcp_queue *queue)
866 {
867 	struct nvme_tcp_hdr *hdr = &queue->pdu.cmd.hdr;
868 	struct nvme_command *nvme_cmd = &queue->pdu.cmd.cmd;
869 	struct nvmet_req *req;
870 	int ret;
871 
872 	if (unlikely(queue->state == NVMET_TCP_Q_CONNECTING)) {
873 		if (hdr->type != nvme_tcp_icreq) {
874 			pr_err("unexpected pdu type (%d) before icreq\n",
875 				hdr->type);
876 			nvmet_tcp_fatal_error(queue);
877 			return -EPROTO;
878 		}
879 		return nvmet_tcp_handle_icreq(queue);
880 	}
881 
882 	if (hdr->type == nvme_tcp_h2c_data) {
883 		ret = nvmet_tcp_handle_h2c_data_pdu(queue);
884 		if (unlikely(ret))
885 			return ret;
886 		return 0;
887 	}
888 
889 	queue->cmd = nvmet_tcp_get_cmd(queue);
890 	if (unlikely(!queue->cmd)) {
891 		/* This should never happen */
892 		pr_err("queue %d: out of commands (%d) send_list_len: %d, opcode: %d",
893 			queue->idx, queue->nr_cmds, queue->send_list_len,
894 			nvme_cmd->common.opcode);
895 		nvmet_tcp_fatal_error(queue);
896 		return -ENOMEM;
897 	}
898 
899 	req = &queue->cmd->req;
900 	memcpy(req->cmd, nvme_cmd, sizeof(*nvme_cmd));
901 
902 	if (unlikely(!nvmet_req_init(req, &queue->nvme_cq,
903 			&queue->nvme_sq, &nvmet_tcp_ops))) {
904 		pr_err("failed cmd %p id %d opcode %d, data_len: %d\n",
905 			req->cmd, req->cmd->common.command_id,
906 			req->cmd->common.opcode,
907 			le32_to_cpu(req->cmd->common.dptr.sgl.length));
908 
909 		nvmet_tcp_handle_req_failure(queue, queue->cmd, req);
910 		return -EAGAIN;
911 	}
912 
913 	ret = nvmet_tcp_map_data(queue->cmd);
914 	if (unlikely(ret)) {
915 		pr_err("queue %d: failed to map data\n", queue->idx);
916 		if (nvmet_tcp_has_inline_data(queue->cmd))
917 			nvmet_tcp_fatal_error(queue);
918 		else
919 			nvmet_req_complete(req, ret);
920 		ret = -EAGAIN;
921 		goto out;
922 	}
923 
924 	if (nvmet_tcp_need_data_in(queue->cmd)) {
925 		if (nvmet_tcp_has_inline_data(queue->cmd)) {
926 			queue->rcv_state = NVMET_TCP_RECV_DATA;
927 			nvmet_tcp_map_pdu_iovec(queue->cmd);
928 			return 0;
929 		}
930 		/* send back R2T */
931 		nvmet_tcp_queue_response(&queue->cmd->req);
932 		goto out;
933 	}
934 
935 	nvmet_req_execute(&queue->cmd->req);
936 out:
937 	nvmet_prepare_receive_pdu(queue);
938 	return ret;
939 }
940 
941 static const u8 nvme_tcp_pdu_sizes[] = {
942 	[nvme_tcp_icreq]	= sizeof(struct nvme_tcp_icreq_pdu),
943 	[nvme_tcp_cmd]		= sizeof(struct nvme_tcp_cmd_pdu),
944 	[nvme_tcp_h2c_data]	= sizeof(struct nvme_tcp_data_pdu),
945 };
946 
947 static inline u8 nvmet_tcp_pdu_size(u8 type)
948 {
949 	size_t idx = type;
950 
951 	return (idx < ARRAY_SIZE(nvme_tcp_pdu_sizes) &&
952 		nvme_tcp_pdu_sizes[idx]) ?
953 			nvme_tcp_pdu_sizes[idx] : 0;
954 }
955 
956 static inline bool nvmet_tcp_pdu_valid(u8 type)
957 {
958 	switch (type) {
959 	case nvme_tcp_icreq:
960 	case nvme_tcp_cmd:
961 	case nvme_tcp_h2c_data:
962 		/* fallthru */
963 		return true;
964 	}
965 
966 	return false;
967 }
968 
969 static int nvmet_tcp_try_recv_pdu(struct nvmet_tcp_queue *queue)
970 {
971 	struct nvme_tcp_hdr *hdr = &queue->pdu.cmd.hdr;
972 	int len;
973 	struct kvec iov;
974 	struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
975 
976 recv:
977 	iov.iov_base = (void *)&queue->pdu + queue->offset;
978 	iov.iov_len = queue->left;
979 	len = kernel_recvmsg(queue->sock, &msg, &iov, 1,
980 			iov.iov_len, msg.msg_flags);
981 	if (unlikely(len < 0))
982 		return len;
983 
984 	queue->offset += len;
985 	queue->left -= len;
986 	if (queue->left)
987 		return -EAGAIN;
988 
989 	if (queue->offset == sizeof(struct nvme_tcp_hdr)) {
990 		u8 hdgst = nvmet_tcp_hdgst_len(queue);
991 
992 		if (unlikely(!nvmet_tcp_pdu_valid(hdr->type))) {
993 			pr_err("unexpected pdu type %d\n", hdr->type);
994 			nvmet_tcp_fatal_error(queue);
995 			return -EIO;
996 		}
997 
998 		if (unlikely(hdr->hlen != nvmet_tcp_pdu_size(hdr->type))) {
999 			pr_err("pdu %d bad hlen %d\n", hdr->type, hdr->hlen);
1000 			return -EIO;
1001 		}
1002 
1003 		queue->left = hdr->hlen - queue->offset + hdgst;
1004 		goto recv;
1005 	}
1006 
1007 	if (queue->hdr_digest &&
1008 	    nvmet_tcp_verify_hdgst(queue, &queue->pdu, queue->offset)) {
1009 		nvmet_tcp_fatal_error(queue); /* fatal */
1010 		return -EPROTO;
1011 	}
1012 
1013 	if (queue->data_digest &&
1014 	    nvmet_tcp_check_ddgst(queue, &queue->pdu)) {
1015 		nvmet_tcp_fatal_error(queue); /* fatal */
1016 		return -EPROTO;
1017 	}
1018 
1019 	return nvmet_tcp_done_recv_pdu(queue);
1020 }
1021 
1022 static void nvmet_tcp_prep_recv_ddgst(struct nvmet_tcp_cmd *cmd)
1023 {
1024 	struct nvmet_tcp_queue *queue = cmd->queue;
1025 
1026 	nvmet_tcp_ddgst(queue->rcv_hash, cmd);
1027 	queue->offset = 0;
1028 	queue->left = NVME_TCP_DIGEST_LENGTH;
1029 	queue->rcv_state = NVMET_TCP_RECV_DDGST;
1030 }
1031 
1032 static int nvmet_tcp_try_recv_data(struct nvmet_tcp_queue *queue)
1033 {
1034 	struct nvmet_tcp_cmd  *cmd = queue->cmd;
1035 	int ret;
1036 
1037 	while (msg_data_left(&cmd->recv_msg)) {
1038 		ret = sock_recvmsg(cmd->queue->sock, &cmd->recv_msg,
1039 			cmd->recv_msg.msg_flags);
1040 		if (ret <= 0)
1041 			return ret;
1042 
1043 		cmd->pdu_recv += ret;
1044 		cmd->rbytes_done += ret;
1045 	}
1046 
1047 	nvmet_tcp_unmap_pdu_iovec(cmd);
1048 
1049 	if (!(cmd->flags & NVMET_TCP_F_INIT_FAILED) &&
1050 	    cmd->rbytes_done == cmd->req.transfer_len) {
1051 		if (queue->data_digest) {
1052 			nvmet_tcp_prep_recv_ddgst(cmd);
1053 			return 0;
1054 		}
1055 		nvmet_req_execute(&cmd->req);
1056 	}
1057 
1058 	nvmet_prepare_receive_pdu(queue);
1059 	return 0;
1060 }
1061 
1062 static int nvmet_tcp_try_recv_ddgst(struct nvmet_tcp_queue *queue)
1063 {
1064 	struct nvmet_tcp_cmd *cmd = queue->cmd;
1065 	int ret;
1066 	struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
1067 	struct kvec iov = {
1068 		.iov_base = (void *)&cmd->recv_ddgst + queue->offset,
1069 		.iov_len = queue->left
1070 	};
1071 
1072 	ret = kernel_recvmsg(queue->sock, &msg, &iov, 1,
1073 			iov.iov_len, msg.msg_flags);
1074 	if (unlikely(ret < 0))
1075 		return ret;
1076 
1077 	queue->offset += ret;
1078 	queue->left -= ret;
1079 	if (queue->left)
1080 		return -EAGAIN;
1081 
1082 	if (queue->data_digest && cmd->exp_ddgst != cmd->recv_ddgst) {
1083 		pr_err("queue %d: cmd %d pdu (%d) data digest error: recv %#x expected %#x\n",
1084 			queue->idx, cmd->req.cmd->common.command_id,
1085 			queue->pdu.cmd.hdr.type, le32_to_cpu(cmd->recv_ddgst),
1086 			le32_to_cpu(cmd->exp_ddgst));
1087 		nvmet_tcp_finish_cmd(cmd);
1088 		nvmet_tcp_fatal_error(queue);
1089 		ret = -EPROTO;
1090 		goto out;
1091 	}
1092 
1093 	if (!(cmd->flags & NVMET_TCP_F_INIT_FAILED) &&
1094 	    cmd->rbytes_done == cmd->req.transfer_len)
1095 		nvmet_req_execute(&cmd->req);
1096 	ret = 0;
1097 out:
1098 	nvmet_prepare_receive_pdu(queue);
1099 	return ret;
1100 }
1101 
1102 static int nvmet_tcp_try_recv_one(struct nvmet_tcp_queue *queue)
1103 {
1104 	int result = 0;
1105 
1106 	if (unlikely(queue->rcv_state == NVMET_TCP_RECV_ERR))
1107 		return 0;
1108 
1109 	if (queue->rcv_state == NVMET_TCP_RECV_PDU) {
1110 		result = nvmet_tcp_try_recv_pdu(queue);
1111 		if (result != 0)
1112 			goto done_recv;
1113 	}
1114 
1115 	if (queue->rcv_state == NVMET_TCP_RECV_DATA) {
1116 		result = nvmet_tcp_try_recv_data(queue);
1117 		if (result != 0)
1118 			goto done_recv;
1119 	}
1120 
1121 	if (queue->rcv_state == NVMET_TCP_RECV_DDGST) {
1122 		result = nvmet_tcp_try_recv_ddgst(queue);
1123 		if (result != 0)
1124 			goto done_recv;
1125 	}
1126 
1127 done_recv:
1128 	if (result < 0) {
1129 		if (result == -EAGAIN)
1130 			return 0;
1131 		return result;
1132 	}
1133 	return 1;
1134 }
1135 
1136 static int nvmet_tcp_try_recv(struct nvmet_tcp_queue *queue,
1137 		int budget, int *recvs)
1138 {
1139 	int i, ret = 0;
1140 
1141 	for (i = 0; i < budget; i++) {
1142 		ret = nvmet_tcp_try_recv_one(queue);
1143 		if (ret <= 0)
1144 			break;
1145 		(*recvs)++;
1146 	}
1147 
1148 	return ret;
1149 }
1150 
1151 static void nvmet_tcp_schedule_release_queue(struct nvmet_tcp_queue *queue)
1152 {
1153 	spin_lock(&queue->state_lock);
1154 	if (queue->state != NVMET_TCP_Q_DISCONNECTING) {
1155 		queue->state = NVMET_TCP_Q_DISCONNECTING;
1156 		schedule_work(&queue->release_work);
1157 	}
1158 	spin_unlock(&queue->state_lock);
1159 }
1160 
1161 static void nvmet_tcp_io_work(struct work_struct *w)
1162 {
1163 	struct nvmet_tcp_queue *queue =
1164 		container_of(w, struct nvmet_tcp_queue, io_work);
1165 	bool pending;
1166 	int ret, ops = 0;
1167 
1168 	do {
1169 		pending = false;
1170 
1171 		ret = nvmet_tcp_try_recv(queue, NVMET_TCP_RECV_BUDGET, &ops);
1172 		if (ret > 0) {
1173 			pending = true;
1174 		} else if (ret < 0) {
1175 			if (ret == -EPIPE || ret == -ECONNRESET)
1176 				kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1177 			else
1178 				nvmet_tcp_fatal_error(queue);
1179 			return;
1180 		}
1181 
1182 		ret = nvmet_tcp_try_send(queue, NVMET_TCP_SEND_BUDGET, &ops);
1183 		if (ret > 0) {
1184 			/* transmitted message/data */
1185 			pending = true;
1186 		} else if (ret < 0) {
1187 			if (ret == -EPIPE || ret == -ECONNRESET)
1188 				kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1189 			else
1190 				nvmet_tcp_fatal_error(queue);
1191 			return;
1192 		}
1193 
1194 	} while (pending && ops < NVMET_TCP_IO_WORK_BUDGET);
1195 
1196 	/*
1197 	 * We exahusted our budget, requeue our selves
1198 	 */
1199 	if (pending)
1200 		queue_work_on(queue->cpu, nvmet_tcp_wq, &queue->io_work);
1201 }
1202 
1203 static int nvmet_tcp_alloc_cmd(struct nvmet_tcp_queue *queue,
1204 		struct nvmet_tcp_cmd *c)
1205 {
1206 	u8 hdgst = nvmet_tcp_hdgst_len(queue);
1207 
1208 	c->queue = queue;
1209 	c->req.port = queue->port->nport;
1210 
1211 	c->cmd_pdu = page_frag_alloc(&queue->pf_cache,
1212 			sizeof(*c->cmd_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO);
1213 	if (!c->cmd_pdu)
1214 		return -ENOMEM;
1215 	c->req.cmd = &c->cmd_pdu->cmd;
1216 
1217 	c->rsp_pdu = page_frag_alloc(&queue->pf_cache,
1218 			sizeof(*c->rsp_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO);
1219 	if (!c->rsp_pdu)
1220 		goto out_free_cmd;
1221 	c->req.cqe = &c->rsp_pdu->cqe;
1222 
1223 	c->data_pdu = page_frag_alloc(&queue->pf_cache,
1224 			sizeof(*c->data_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO);
1225 	if (!c->data_pdu)
1226 		goto out_free_rsp;
1227 
1228 	c->r2t_pdu = page_frag_alloc(&queue->pf_cache,
1229 			sizeof(*c->r2t_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO);
1230 	if (!c->r2t_pdu)
1231 		goto out_free_data;
1232 
1233 	c->recv_msg.msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
1234 
1235 	list_add_tail(&c->entry, &queue->free_list);
1236 
1237 	return 0;
1238 out_free_data:
1239 	page_frag_free(c->data_pdu);
1240 out_free_rsp:
1241 	page_frag_free(c->rsp_pdu);
1242 out_free_cmd:
1243 	page_frag_free(c->cmd_pdu);
1244 	return -ENOMEM;
1245 }
1246 
1247 static void nvmet_tcp_free_cmd(struct nvmet_tcp_cmd *c)
1248 {
1249 	page_frag_free(c->r2t_pdu);
1250 	page_frag_free(c->data_pdu);
1251 	page_frag_free(c->rsp_pdu);
1252 	page_frag_free(c->cmd_pdu);
1253 }
1254 
1255 static int nvmet_tcp_alloc_cmds(struct nvmet_tcp_queue *queue)
1256 {
1257 	struct nvmet_tcp_cmd *cmds;
1258 	int i, ret = -EINVAL, nr_cmds = queue->nr_cmds;
1259 
1260 	cmds = kcalloc(nr_cmds, sizeof(struct nvmet_tcp_cmd), GFP_KERNEL);
1261 	if (!cmds)
1262 		goto out;
1263 
1264 	for (i = 0; i < nr_cmds; i++) {
1265 		ret = nvmet_tcp_alloc_cmd(queue, cmds + i);
1266 		if (ret)
1267 			goto out_free;
1268 	}
1269 
1270 	queue->cmds = cmds;
1271 
1272 	return 0;
1273 out_free:
1274 	while (--i >= 0)
1275 		nvmet_tcp_free_cmd(cmds + i);
1276 	kfree(cmds);
1277 out:
1278 	return ret;
1279 }
1280 
1281 static void nvmet_tcp_free_cmds(struct nvmet_tcp_queue *queue)
1282 {
1283 	struct nvmet_tcp_cmd *cmds = queue->cmds;
1284 	int i;
1285 
1286 	for (i = 0; i < queue->nr_cmds; i++)
1287 		nvmet_tcp_free_cmd(cmds + i);
1288 
1289 	nvmet_tcp_free_cmd(&queue->connect);
1290 	kfree(cmds);
1291 }
1292 
1293 static void nvmet_tcp_restore_socket_callbacks(struct nvmet_tcp_queue *queue)
1294 {
1295 	struct socket *sock = queue->sock;
1296 
1297 	write_lock_bh(&sock->sk->sk_callback_lock);
1298 	sock->sk->sk_data_ready =  queue->data_ready;
1299 	sock->sk->sk_state_change = queue->state_change;
1300 	sock->sk->sk_write_space = queue->write_space;
1301 	sock->sk->sk_user_data = NULL;
1302 	write_unlock_bh(&sock->sk->sk_callback_lock);
1303 }
1304 
1305 static void nvmet_tcp_finish_cmd(struct nvmet_tcp_cmd *cmd)
1306 {
1307 	nvmet_req_uninit(&cmd->req);
1308 	nvmet_tcp_unmap_pdu_iovec(cmd);
1309 	sgl_free(cmd->req.sg);
1310 }
1311 
1312 static void nvmet_tcp_uninit_data_in_cmds(struct nvmet_tcp_queue *queue)
1313 {
1314 	struct nvmet_tcp_cmd *cmd = queue->cmds;
1315 	int i;
1316 
1317 	for (i = 0; i < queue->nr_cmds; i++, cmd++) {
1318 		if (nvmet_tcp_need_data_in(cmd))
1319 			nvmet_tcp_finish_cmd(cmd);
1320 	}
1321 
1322 	if (!queue->nr_cmds && nvmet_tcp_need_data_in(&queue->connect)) {
1323 		/* failed in connect */
1324 		nvmet_tcp_finish_cmd(&queue->connect);
1325 	}
1326 }
1327 
1328 static void nvmet_tcp_release_queue_work(struct work_struct *w)
1329 {
1330 	struct nvmet_tcp_queue *queue =
1331 		container_of(w, struct nvmet_tcp_queue, release_work);
1332 
1333 	mutex_lock(&nvmet_tcp_queue_mutex);
1334 	list_del_init(&queue->queue_list);
1335 	mutex_unlock(&nvmet_tcp_queue_mutex);
1336 
1337 	nvmet_tcp_restore_socket_callbacks(queue);
1338 	flush_work(&queue->io_work);
1339 
1340 	nvmet_tcp_uninit_data_in_cmds(queue);
1341 	nvmet_sq_destroy(&queue->nvme_sq);
1342 	cancel_work_sync(&queue->io_work);
1343 	sock_release(queue->sock);
1344 	nvmet_tcp_free_cmds(queue);
1345 	if (queue->hdr_digest || queue->data_digest)
1346 		nvmet_tcp_free_crypto(queue);
1347 	ida_simple_remove(&nvmet_tcp_queue_ida, queue->idx);
1348 
1349 	kfree(queue);
1350 }
1351 
1352 static void nvmet_tcp_data_ready(struct sock *sk)
1353 {
1354 	struct nvmet_tcp_queue *queue;
1355 
1356 	read_lock_bh(&sk->sk_callback_lock);
1357 	queue = sk->sk_user_data;
1358 	if (likely(queue))
1359 		queue_work_on(queue->cpu, nvmet_tcp_wq, &queue->io_work);
1360 	read_unlock_bh(&sk->sk_callback_lock);
1361 }
1362 
1363 static void nvmet_tcp_write_space(struct sock *sk)
1364 {
1365 	struct nvmet_tcp_queue *queue;
1366 
1367 	read_lock_bh(&sk->sk_callback_lock);
1368 	queue = sk->sk_user_data;
1369 	if (unlikely(!queue))
1370 		goto out;
1371 
1372 	if (unlikely(queue->state == NVMET_TCP_Q_CONNECTING)) {
1373 		queue->write_space(sk);
1374 		goto out;
1375 	}
1376 
1377 	if (sk_stream_is_writeable(sk)) {
1378 		clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1379 		queue_work_on(queue->cpu, nvmet_tcp_wq, &queue->io_work);
1380 	}
1381 out:
1382 	read_unlock_bh(&sk->sk_callback_lock);
1383 }
1384 
1385 static void nvmet_tcp_state_change(struct sock *sk)
1386 {
1387 	struct nvmet_tcp_queue *queue;
1388 
1389 	write_lock_bh(&sk->sk_callback_lock);
1390 	queue = sk->sk_user_data;
1391 	if (!queue)
1392 		goto done;
1393 
1394 	switch (sk->sk_state) {
1395 	case TCP_FIN_WAIT1:
1396 	case TCP_CLOSE_WAIT:
1397 	case TCP_CLOSE:
1398 		/* FALLTHRU */
1399 		sk->sk_user_data = NULL;
1400 		nvmet_tcp_schedule_release_queue(queue);
1401 		break;
1402 	default:
1403 		pr_warn("queue %d unhandled state %d\n",
1404 			queue->idx, sk->sk_state);
1405 	}
1406 done:
1407 	write_unlock_bh(&sk->sk_callback_lock);
1408 }
1409 
1410 static int nvmet_tcp_set_queue_sock(struct nvmet_tcp_queue *queue)
1411 {
1412 	struct socket *sock = queue->sock;
1413 	struct linger sol = { .l_onoff = 1, .l_linger = 0 };
1414 	int ret;
1415 
1416 	ret = kernel_getsockname(sock,
1417 		(struct sockaddr *)&queue->sockaddr);
1418 	if (ret < 0)
1419 		return ret;
1420 
1421 	ret = kernel_getpeername(sock,
1422 		(struct sockaddr *)&queue->sockaddr_peer);
1423 	if (ret < 0)
1424 		return ret;
1425 
1426 	/*
1427 	 * Cleanup whatever is sitting in the TCP transmit queue on socket
1428 	 * close. This is done to prevent stale data from being sent should
1429 	 * the network connection be restored before TCP times out.
1430 	 */
1431 	ret = kernel_setsockopt(sock, SOL_SOCKET, SO_LINGER,
1432 			(char *)&sol, sizeof(sol));
1433 	if (ret)
1434 		return ret;
1435 
1436 	write_lock_bh(&sock->sk->sk_callback_lock);
1437 	sock->sk->sk_user_data = queue;
1438 	queue->data_ready = sock->sk->sk_data_ready;
1439 	sock->sk->sk_data_ready = nvmet_tcp_data_ready;
1440 	queue->state_change = sock->sk->sk_state_change;
1441 	sock->sk->sk_state_change = nvmet_tcp_state_change;
1442 	queue->write_space = sock->sk->sk_write_space;
1443 	sock->sk->sk_write_space = nvmet_tcp_write_space;
1444 	write_unlock_bh(&sock->sk->sk_callback_lock);
1445 
1446 	return 0;
1447 }
1448 
1449 static int nvmet_tcp_alloc_queue(struct nvmet_tcp_port *port,
1450 		struct socket *newsock)
1451 {
1452 	struct nvmet_tcp_queue *queue;
1453 	int ret;
1454 
1455 	queue = kzalloc(sizeof(*queue), GFP_KERNEL);
1456 	if (!queue)
1457 		return -ENOMEM;
1458 
1459 	INIT_WORK(&queue->release_work, nvmet_tcp_release_queue_work);
1460 	INIT_WORK(&queue->io_work, nvmet_tcp_io_work);
1461 	queue->sock = newsock;
1462 	queue->port = port;
1463 	queue->nr_cmds = 0;
1464 	spin_lock_init(&queue->state_lock);
1465 	queue->state = NVMET_TCP_Q_CONNECTING;
1466 	INIT_LIST_HEAD(&queue->free_list);
1467 	init_llist_head(&queue->resp_list);
1468 	INIT_LIST_HEAD(&queue->resp_send_list);
1469 
1470 	queue->idx = ida_simple_get(&nvmet_tcp_queue_ida, 0, 0, GFP_KERNEL);
1471 	if (queue->idx < 0) {
1472 		ret = queue->idx;
1473 		goto out_free_queue;
1474 	}
1475 
1476 	ret = nvmet_tcp_alloc_cmd(queue, &queue->connect);
1477 	if (ret)
1478 		goto out_ida_remove;
1479 
1480 	ret = nvmet_sq_init(&queue->nvme_sq);
1481 	if (ret)
1482 		goto out_free_connect;
1483 
1484 	port->last_cpu = cpumask_next_wrap(port->last_cpu,
1485 				cpu_online_mask, -1, false);
1486 	queue->cpu = port->last_cpu;
1487 	nvmet_prepare_receive_pdu(queue);
1488 
1489 	mutex_lock(&nvmet_tcp_queue_mutex);
1490 	list_add_tail(&queue->queue_list, &nvmet_tcp_queue_list);
1491 	mutex_unlock(&nvmet_tcp_queue_mutex);
1492 
1493 	ret = nvmet_tcp_set_queue_sock(queue);
1494 	if (ret)
1495 		goto out_destroy_sq;
1496 
1497 	queue_work_on(queue->cpu, nvmet_tcp_wq, &queue->io_work);
1498 
1499 	return 0;
1500 out_destroy_sq:
1501 	mutex_lock(&nvmet_tcp_queue_mutex);
1502 	list_del_init(&queue->queue_list);
1503 	mutex_unlock(&nvmet_tcp_queue_mutex);
1504 	nvmet_sq_destroy(&queue->nvme_sq);
1505 out_free_connect:
1506 	nvmet_tcp_free_cmd(&queue->connect);
1507 out_ida_remove:
1508 	ida_simple_remove(&nvmet_tcp_queue_ida, queue->idx);
1509 out_free_queue:
1510 	kfree(queue);
1511 	return ret;
1512 }
1513 
1514 static void nvmet_tcp_accept_work(struct work_struct *w)
1515 {
1516 	struct nvmet_tcp_port *port =
1517 		container_of(w, struct nvmet_tcp_port, accept_work);
1518 	struct socket *newsock;
1519 	int ret;
1520 
1521 	while (true) {
1522 		ret = kernel_accept(port->sock, &newsock, O_NONBLOCK);
1523 		if (ret < 0) {
1524 			if (ret != -EAGAIN)
1525 				pr_warn("failed to accept err=%d\n", ret);
1526 			return;
1527 		}
1528 		ret = nvmet_tcp_alloc_queue(port, newsock);
1529 		if (ret) {
1530 			pr_err("failed to allocate queue\n");
1531 			sock_release(newsock);
1532 		}
1533 	}
1534 }
1535 
1536 static void nvmet_tcp_listen_data_ready(struct sock *sk)
1537 {
1538 	struct nvmet_tcp_port *port;
1539 
1540 	read_lock_bh(&sk->sk_callback_lock);
1541 	port = sk->sk_user_data;
1542 	if (!port)
1543 		goto out;
1544 
1545 	if (sk->sk_state == TCP_LISTEN)
1546 		schedule_work(&port->accept_work);
1547 out:
1548 	read_unlock_bh(&sk->sk_callback_lock);
1549 }
1550 
1551 static int nvmet_tcp_add_port(struct nvmet_port *nport)
1552 {
1553 	struct nvmet_tcp_port *port;
1554 	__kernel_sa_family_t af;
1555 	int opt, ret;
1556 
1557 	port = kzalloc(sizeof(*port), GFP_KERNEL);
1558 	if (!port)
1559 		return -ENOMEM;
1560 
1561 	switch (nport->disc_addr.adrfam) {
1562 	case NVMF_ADDR_FAMILY_IP4:
1563 		af = AF_INET;
1564 		break;
1565 	case NVMF_ADDR_FAMILY_IP6:
1566 		af = AF_INET6;
1567 		break;
1568 	default:
1569 		pr_err("address family %d not supported\n",
1570 				nport->disc_addr.adrfam);
1571 		ret = -EINVAL;
1572 		goto err_port;
1573 	}
1574 
1575 	ret = inet_pton_with_scope(&init_net, af, nport->disc_addr.traddr,
1576 			nport->disc_addr.trsvcid, &port->addr);
1577 	if (ret) {
1578 		pr_err("malformed ip/port passed: %s:%s\n",
1579 			nport->disc_addr.traddr, nport->disc_addr.trsvcid);
1580 		goto err_port;
1581 	}
1582 
1583 	port->nport = nport;
1584 	port->last_cpu = -1;
1585 	INIT_WORK(&port->accept_work, nvmet_tcp_accept_work);
1586 	if (port->nport->inline_data_size < 0)
1587 		port->nport->inline_data_size = NVMET_TCP_DEF_INLINE_DATA_SIZE;
1588 
1589 	ret = sock_create(port->addr.ss_family, SOCK_STREAM,
1590 				IPPROTO_TCP, &port->sock);
1591 	if (ret) {
1592 		pr_err("failed to create a socket\n");
1593 		goto err_port;
1594 	}
1595 
1596 	port->sock->sk->sk_user_data = port;
1597 	port->data_ready = port->sock->sk->sk_data_ready;
1598 	port->sock->sk->sk_data_ready = nvmet_tcp_listen_data_ready;
1599 
1600 	opt = 1;
1601 	ret = kernel_setsockopt(port->sock, IPPROTO_TCP,
1602 			TCP_NODELAY, (char *)&opt, sizeof(opt));
1603 	if (ret) {
1604 		pr_err("failed to set TCP_NODELAY sock opt %d\n", ret);
1605 		goto err_sock;
1606 	}
1607 
1608 	ret = kernel_setsockopt(port->sock, SOL_SOCKET, SO_REUSEADDR,
1609 			(char *)&opt, sizeof(opt));
1610 	if (ret) {
1611 		pr_err("failed to set SO_REUSEADDR sock opt %d\n", ret);
1612 		goto err_sock;
1613 	}
1614 
1615 	ret = kernel_bind(port->sock, (struct sockaddr *)&port->addr,
1616 			sizeof(port->addr));
1617 	if (ret) {
1618 		pr_err("failed to bind port socket %d\n", ret);
1619 		goto err_sock;
1620 	}
1621 
1622 	ret = kernel_listen(port->sock, 128);
1623 	if (ret) {
1624 		pr_err("failed to listen %d on port sock\n", ret);
1625 		goto err_sock;
1626 	}
1627 
1628 	nport->priv = port;
1629 	pr_info("enabling port %d (%pISpc)\n",
1630 		le16_to_cpu(nport->disc_addr.portid), &port->addr);
1631 
1632 	return 0;
1633 
1634 err_sock:
1635 	sock_release(port->sock);
1636 err_port:
1637 	kfree(port);
1638 	return ret;
1639 }
1640 
1641 static void nvmet_tcp_remove_port(struct nvmet_port *nport)
1642 {
1643 	struct nvmet_tcp_port *port = nport->priv;
1644 
1645 	write_lock_bh(&port->sock->sk->sk_callback_lock);
1646 	port->sock->sk->sk_data_ready = port->data_ready;
1647 	port->sock->sk->sk_user_data = NULL;
1648 	write_unlock_bh(&port->sock->sk->sk_callback_lock);
1649 	cancel_work_sync(&port->accept_work);
1650 
1651 	sock_release(port->sock);
1652 	kfree(port);
1653 }
1654 
1655 static void nvmet_tcp_delete_ctrl(struct nvmet_ctrl *ctrl)
1656 {
1657 	struct nvmet_tcp_queue *queue;
1658 
1659 	mutex_lock(&nvmet_tcp_queue_mutex);
1660 	list_for_each_entry(queue, &nvmet_tcp_queue_list, queue_list)
1661 		if (queue->nvme_sq.ctrl == ctrl)
1662 			kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1663 	mutex_unlock(&nvmet_tcp_queue_mutex);
1664 }
1665 
1666 static u16 nvmet_tcp_install_queue(struct nvmet_sq *sq)
1667 {
1668 	struct nvmet_tcp_queue *queue =
1669 		container_of(sq, struct nvmet_tcp_queue, nvme_sq);
1670 
1671 	if (sq->qid == 0) {
1672 		/* Let inflight controller teardown complete */
1673 		flush_scheduled_work();
1674 	}
1675 
1676 	queue->nr_cmds = sq->size * 2;
1677 	if (nvmet_tcp_alloc_cmds(queue))
1678 		return NVME_SC_INTERNAL;
1679 	return 0;
1680 }
1681 
1682 static void nvmet_tcp_disc_port_addr(struct nvmet_req *req,
1683 		struct nvmet_port *nport, char *traddr)
1684 {
1685 	struct nvmet_tcp_port *port = nport->priv;
1686 
1687 	if (inet_addr_is_any((struct sockaddr *)&port->addr)) {
1688 		struct nvmet_tcp_cmd *cmd =
1689 			container_of(req, struct nvmet_tcp_cmd, req);
1690 		struct nvmet_tcp_queue *queue = cmd->queue;
1691 
1692 		sprintf(traddr, "%pISc", (struct sockaddr *)&queue->sockaddr);
1693 	} else {
1694 		memcpy(traddr, nport->disc_addr.traddr, NVMF_TRADDR_SIZE);
1695 	}
1696 }
1697 
1698 static struct nvmet_fabrics_ops nvmet_tcp_ops = {
1699 	.owner			= THIS_MODULE,
1700 	.type			= NVMF_TRTYPE_TCP,
1701 	.msdbd			= 1,
1702 	.has_keyed_sgls		= 0,
1703 	.add_port		= nvmet_tcp_add_port,
1704 	.remove_port		= nvmet_tcp_remove_port,
1705 	.queue_response		= nvmet_tcp_queue_response,
1706 	.delete_ctrl		= nvmet_tcp_delete_ctrl,
1707 	.install_queue		= nvmet_tcp_install_queue,
1708 	.disc_traddr		= nvmet_tcp_disc_port_addr,
1709 };
1710 
1711 static int __init nvmet_tcp_init(void)
1712 {
1713 	int ret;
1714 
1715 	nvmet_tcp_wq = alloc_workqueue("nvmet_tcp_wq", WQ_HIGHPRI, 0);
1716 	if (!nvmet_tcp_wq)
1717 		return -ENOMEM;
1718 
1719 	ret = nvmet_register_transport(&nvmet_tcp_ops);
1720 	if (ret)
1721 		goto err;
1722 
1723 	return 0;
1724 err:
1725 	destroy_workqueue(nvmet_tcp_wq);
1726 	return ret;
1727 }
1728 
1729 static void __exit nvmet_tcp_exit(void)
1730 {
1731 	struct nvmet_tcp_queue *queue;
1732 
1733 	nvmet_unregister_transport(&nvmet_tcp_ops);
1734 
1735 	flush_scheduled_work();
1736 	mutex_lock(&nvmet_tcp_queue_mutex);
1737 	list_for_each_entry(queue, &nvmet_tcp_queue_list, queue_list)
1738 		kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1739 	mutex_unlock(&nvmet_tcp_queue_mutex);
1740 	flush_scheduled_work();
1741 
1742 	destroy_workqueue(nvmet_tcp_wq);
1743 }
1744 
1745 module_init(nvmet_tcp_init);
1746 module_exit(nvmet_tcp_exit);
1747 
1748 MODULE_LICENSE("GPL v2");
1749 MODULE_ALIAS("nvmet-transport-3"); /* 3 == NVMF_TRTYPE_TCP */
1750