xref: /openbmc/linux/drivers/nvme/target/tcp.c (revision b1a3e75e)
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 the socket priority to use for connections were it is desirable
23  * that the NIC consider performing optimized packet processing or filtering.
24  * A non-zero value being sufficient to indicate general consideration of any
25  * possible optimization.  Making it a module param allows for alternative
26  * values that may be unique for some NIC implementations.
27  */
28 static int so_priority;
29 module_param(so_priority, int, 0644);
30 MODULE_PARM_DESC(so_priority, "nvmet tcp socket optimize priority");
31 
32 #define NVMET_TCP_RECV_BUDGET		8
33 #define NVMET_TCP_SEND_BUDGET		8
34 #define NVMET_TCP_IO_WORK_BUDGET	64
35 
36 enum nvmet_tcp_send_state {
37 	NVMET_TCP_SEND_DATA_PDU,
38 	NVMET_TCP_SEND_DATA,
39 	NVMET_TCP_SEND_R2T,
40 	NVMET_TCP_SEND_DDGST,
41 	NVMET_TCP_SEND_RESPONSE
42 };
43 
44 enum nvmet_tcp_recv_state {
45 	NVMET_TCP_RECV_PDU,
46 	NVMET_TCP_RECV_DATA,
47 	NVMET_TCP_RECV_DDGST,
48 	NVMET_TCP_RECV_ERR,
49 };
50 
51 enum {
52 	NVMET_TCP_F_INIT_FAILED = (1 << 0),
53 };
54 
55 struct nvmet_tcp_cmd {
56 	struct nvmet_tcp_queue		*queue;
57 	struct nvmet_req		req;
58 
59 	struct nvme_tcp_cmd_pdu		*cmd_pdu;
60 	struct nvme_tcp_rsp_pdu		*rsp_pdu;
61 	struct nvme_tcp_data_pdu	*data_pdu;
62 	struct nvme_tcp_r2t_pdu		*r2t_pdu;
63 
64 	u32				rbytes_done;
65 	u32				wbytes_done;
66 
67 	u32				pdu_len;
68 	u32				pdu_recv;
69 	int				sg_idx;
70 	int				nr_mapped;
71 	struct msghdr			recv_msg;
72 	struct kvec			*iov;
73 	u32				flags;
74 
75 	struct list_head		entry;
76 	struct llist_node		lentry;
77 
78 	/* send state */
79 	u32				offset;
80 	struct scatterlist		*cur_sg;
81 	enum nvmet_tcp_send_state	state;
82 
83 	__le32				exp_ddgst;
84 	__le32				recv_ddgst;
85 };
86 
87 enum nvmet_tcp_queue_state {
88 	NVMET_TCP_Q_CONNECTING,
89 	NVMET_TCP_Q_LIVE,
90 	NVMET_TCP_Q_DISCONNECTING,
91 };
92 
93 struct nvmet_tcp_queue {
94 	struct socket		*sock;
95 	struct nvmet_tcp_port	*port;
96 	struct work_struct	io_work;
97 	int			cpu;
98 	struct nvmet_cq		nvme_cq;
99 	struct nvmet_sq		nvme_sq;
100 
101 	/* send state */
102 	struct nvmet_tcp_cmd	*cmds;
103 	unsigned int		nr_cmds;
104 	struct list_head	free_list;
105 	struct llist_head	resp_list;
106 	struct list_head	resp_send_list;
107 	int			send_list_len;
108 	struct nvmet_tcp_cmd	*snd_cmd;
109 
110 	/* recv state */
111 	int			offset;
112 	int			left;
113 	enum nvmet_tcp_recv_state rcv_state;
114 	struct nvmet_tcp_cmd	*cmd;
115 	union nvme_tcp_pdu	pdu;
116 
117 	/* digest state */
118 	bool			hdr_digest;
119 	bool			data_digest;
120 	struct ahash_request	*snd_hash;
121 	struct ahash_request	*rcv_hash;
122 
123 	spinlock_t		state_lock;
124 	enum nvmet_tcp_queue_state state;
125 
126 	struct sockaddr_storage	sockaddr;
127 	struct sockaddr_storage	sockaddr_peer;
128 	struct work_struct	release_work;
129 
130 	int			idx;
131 	struct list_head	queue_list;
132 
133 	struct nvmet_tcp_cmd	connect;
134 
135 	struct page_frag_cache	pf_cache;
136 
137 	void (*data_ready)(struct sock *);
138 	void (*state_change)(struct sock *);
139 	void (*write_space)(struct sock *);
140 };
141 
142 struct nvmet_tcp_port {
143 	struct socket		*sock;
144 	struct work_struct	accept_work;
145 	struct nvmet_port	*nport;
146 	struct sockaddr_storage addr;
147 	int			last_cpu;
148 	void (*data_ready)(struct sock *);
149 };
150 
151 static DEFINE_IDA(nvmet_tcp_queue_ida);
152 static LIST_HEAD(nvmet_tcp_queue_list);
153 static DEFINE_MUTEX(nvmet_tcp_queue_mutex);
154 
155 static struct workqueue_struct *nvmet_tcp_wq;
156 static const struct nvmet_fabrics_ops nvmet_tcp_ops;
157 static void nvmet_tcp_free_cmd(struct nvmet_tcp_cmd *c);
158 static void nvmet_tcp_finish_cmd(struct nvmet_tcp_cmd *cmd);
159 
160 static inline u16 nvmet_tcp_cmd_tag(struct nvmet_tcp_queue *queue,
161 		struct nvmet_tcp_cmd *cmd)
162 {
163 	return cmd - queue->cmds;
164 }
165 
166 static inline bool nvmet_tcp_has_data_in(struct nvmet_tcp_cmd *cmd)
167 {
168 	return nvme_is_write(cmd->req.cmd) &&
169 		cmd->rbytes_done < cmd->req.transfer_len;
170 }
171 
172 static inline bool nvmet_tcp_need_data_in(struct nvmet_tcp_cmd *cmd)
173 {
174 	return nvmet_tcp_has_data_in(cmd) && !cmd->req.cqe->status;
175 }
176 
177 static inline bool nvmet_tcp_need_data_out(struct nvmet_tcp_cmd *cmd)
178 {
179 	return !nvme_is_write(cmd->req.cmd) &&
180 		cmd->req.transfer_len > 0 &&
181 		!cmd->req.cqe->status;
182 }
183 
184 static inline bool nvmet_tcp_has_inline_data(struct nvmet_tcp_cmd *cmd)
185 {
186 	return nvme_is_write(cmd->req.cmd) && cmd->pdu_len &&
187 		!cmd->rbytes_done;
188 }
189 
190 static inline struct nvmet_tcp_cmd *
191 nvmet_tcp_get_cmd(struct nvmet_tcp_queue *queue)
192 {
193 	struct nvmet_tcp_cmd *cmd;
194 
195 	cmd = list_first_entry_or_null(&queue->free_list,
196 				struct nvmet_tcp_cmd, entry);
197 	if (!cmd)
198 		return NULL;
199 	list_del_init(&cmd->entry);
200 
201 	cmd->rbytes_done = cmd->wbytes_done = 0;
202 	cmd->pdu_len = 0;
203 	cmd->pdu_recv = 0;
204 	cmd->iov = NULL;
205 	cmd->flags = 0;
206 	return cmd;
207 }
208 
209 static inline void nvmet_tcp_put_cmd(struct nvmet_tcp_cmd *cmd)
210 {
211 	if (unlikely(cmd == &cmd->queue->connect))
212 		return;
213 
214 	list_add_tail(&cmd->entry, &cmd->queue->free_list);
215 }
216 
217 static inline u8 nvmet_tcp_hdgst_len(struct nvmet_tcp_queue *queue)
218 {
219 	return queue->hdr_digest ? NVME_TCP_DIGEST_LENGTH : 0;
220 }
221 
222 static inline u8 nvmet_tcp_ddgst_len(struct nvmet_tcp_queue *queue)
223 {
224 	return queue->data_digest ? NVME_TCP_DIGEST_LENGTH : 0;
225 }
226 
227 static inline void nvmet_tcp_hdgst(struct ahash_request *hash,
228 		void *pdu, size_t len)
229 {
230 	struct scatterlist sg;
231 
232 	sg_init_one(&sg, pdu, len);
233 	ahash_request_set_crypt(hash, &sg, pdu + len, len);
234 	crypto_ahash_digest(hash);
235 }
236 
237 static int nvmet_tcp_verify_hdgst(struct nvmet_tcp_queue *queue,
238 	void *pdu, size_t len)
239 {
240 	struct nvme_tcp_hdr *hdr = pdu;
241 	__le32 recv_digest;
242 	__le32 exp_digest;
243 
244 	if (unlikely(!(hdr->flags & NVME_TCP_F_HDGST))) {
245 		pr_err("queue %d: header digest enabled but no header digest\n",
246 			queue->idx);
247 		return -EPROTO;
248 	}
249 
250 	recv_digest = *(__le32 *)(pdu + hdr->hlen);
251 	nvmet_tcp_hdgst(queue->rcv_hash, pdu, len);
252 	exp_digest = *(__le32 *)(pdu + hdr->hlen);
253 	if (recv_digest != exp_digest) {
254 		pr_err("queue %d: header digest error: recv %#x expected %#x\n",
255 			queue->idx, le32_to_cpu(recv_digest),
256 			le32_to_cpu(exp_digest));
257 		return -EPROTO;
258 	}
259 
260 	return 0;
261 }
262 
263 static int nvmet_tcp_check_ddgst(struct nvmet_tcp_queue *queue, void *pdu)
264 {
265 	struct nvme_tcp_hdr *hdr = pdu;
266 	u8 digest_len = nvmet_tcp_hdgst_len(queue);
267 	u32 len;
268 
269 	len = le32_to_cpu(hdr->plen) - hdr->hlen -
270 		(hdr->flags & NVME_TCP_F_HDGST ? digest_len : 0);
271 
272 	if (unlikely(len && !(hdr->flags & NVME_TCP_F_DDGST))) {
273 		pr_err("queue %d: data digest flag is cleared\n", queue->idx);
274 		return -EPROTO;
275 	}
276 
277 	return 0;
278 }
279 
280 static void nvmet_tcp_unmap_pdu_iovec(struct nvmet_tcp_cmd *cmd)
281 {
282 	struct scatterlist *sg;
283 	int i;
284 
285 	sg = &cmd->req.sg[cmd->sg_idx];
286 
287 	for (i = 0; i < cmd->nr_mapped; i++)
288 		kunmap(sg_page(&sg[i]));
289 }
290 
291 static void nvmet_tcp_map_pdu_iovec(struct nvmet_tcp_cmd *cmd)
292 {
293 	struct kvec *iov = cmd->iov;
294 	struct scatterlist *sg;
295 	u32 length, offset, sg_offset;
296 
297 	length = cmd->pdu_len;
298 	cmd->nr_mapped = DIV_ROUND_UP(length, PAGE_SIZE);
299 	offset = cmd->rbytes_done;
300 	cmd->sg_idx = DIV_ROUND_UP(offset, PAGE_SIZE);
301 	sg_offset = offset % PAGE_SIZE;
302 	sg = &cmd->req.sg[cmd->sg_idx];
303 
304 	while (length) {
305 		u32 iov_len = min_t(u32, length, sg->length - sg_offset);
306 
307 		iov->iov_base = kmap(sg_page(sg)) + sg->offset + sg_offset;
308 		iov->iov_len = iov_len;
309 
310 		length -= iov_len;
311 		sg = sg_next(sg);
312 		iov++;
313 	}
314 
315 	iov_iter_kvec(&cmd->recv_msg.msg_iter, READ, cmd->iov,
316 		cmd->nr_mapped, cmd->pdu_len);
317 }
318 
319 static void nvmet_tcp_fatal_error(struct nvmet_tcp_queue *queue)
320 {
321 	queue->rcv_state = NVMET_TCP_RECV_ERR;
322 	if (queue->nvme_sq.ctrl)
323 		nvmet_ctrl_fatal_error(queue->nvme_sq.ctrl);
324 	else
325 		kernel_sock_shutdown(queue->sock, SHUT_RDWR);
326 }
327 
328 static void nvmet_tcp_socket_error(struct nvmet_tcp_queue *queue, int status)
329 {
330 	if (status == -EPIPE || status == -ECONNRESET)
331 		kernel_sock_shutdown(queue->sock, SHUT_RDWR);
332 	else
333 		nvmet_tcp_fatal_error(queue);
334 }
335 
336 static int nvmet_tcp_map_data(struct nvmet_tcp_cmd *cmd)
337 {
338 	struct nvme_sgl_desc *sgl = &cmd->req.cmd->common.dptr.sgl;
339 	u32 len = le32_to_cpu(sgl->length);
340 
341 	if (!len)
342 		return 0;
343 
344 	if (sgl->type == ((NVME_SGL_FMT_DATA_DESC << 4) |
345 			  NVME_SGL_FMT_OFFSET)) {
346 		if (!nvme_is_write(cmd->req.cmd))
347 			return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
348 
349 		if (len > cmd->req.port->inline_data_size)
350 			return NVME_SC_SGL_INVALID_OFFSET | NVME_SC_DNR;
351 		cmd->pdu_len = len;
352 	}
353 	cmd->req.transfer_len += len;
354 
355 	cmd->req.sg = sgl_alloc(len, GFP_KERNEL, &cmd->req.sg_cnt);
356 	if (!cmd->req.sg)
357 		return NVME_SC_INTERNAL;
358 	cmd->cur_sg = cmd->req.sg;
359 
360 	if (nvmet_tcp_has_data_in(cmd)) {
361 		cmd->iov = kmalloc_array(cmd->req.sg_cnt,
362 				sizeof(*cmd->iov), GFP_KERNEL);
363 		if (!cmd->iov)
364 			goto err;
365 	}
366 
367 	return 0;
368 err:
369 	sgl_free(cmd->req.sg);
370 	return NVME_SC_INTERNAL;
371 }
372 
373 static void nvmet_tcp_ddgst(struct ahash_request *hash,
374 		struct nvmet_tcp_cmd *cmd)
375 {
376 	ahash_request_set_crypt(hash, cmd->req.sg,
377 		(void *)&cmd->exp_ddgst, cmd->req.transfer_len);
378 	crypto_ahash_digest(hash);
379 }
380 
381 static void nvmet_setup_c2h_data_pdu(struct nvmet_tcp_cmd *cmd)
382 {
383 	struct nvme_tcp_data_pdu *pdu = cmd->data_pdu;
384 	struct nvmet_tcp_queue *queue = cmd->queue;
385 	u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
386 	u8 ddgst = nvmet_tcp_ddgst_len(cmd->queue);
387 
388 	cmd->offset = 0;
389 	cmd->state = NVMET_TCP_SEND_DATA_PDU;
390 
391 	pdu->hdr.type = nvme_tcp_c2h_data;
392 	pdu->hdr.flags = NVME_TCP_F_DATA_LAST | (queue->nvme_sq.sqhd_disabled ?
393 						NVME_TCP_F_DATA_SUCCESS : 0);
394 	pdu->hdr.hlen = sizeof(*pdu);
395 	pdu->hdr.pdo = pdu->hdr.hlen + hdgst;
396 	pdu->hdr.plen =
397 		cpu_to_le32(pdu->hdr.hlen + hdgst +
398 				cmd->req.transfer_len + ddgst);
399 	pdu->command_id = cmd->req.cqe->command_id;
400 	pdu->data_length = cpu_to_le32(cmd->req.transfer_len);
401 	pdu->data_offset = cpu_to_le32(cmd->wbytes_done);
402 
403 	if (queue->data_digest) {
404 		pdu->hdr.flags |= NVME_TCP_F_DDGST;
405 		nvmet_tcp_ddgst(queue->snd_hash, cmd);
406 	}
407 
408 	if (cmd->queue->hdr_digest) {
409 		pdu->hdr.flags |= NVME_TCP_F_HDGST;
410 		nvmet_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
411 	}
412 }
413 
414 static void nvmet_setup_r2t_pdu(struct nvmet_tcp_cmd *cmd)
415 {
416 	struct nvme_tcp_r2t_pdu *pdu = cmd->r2t_pdu;
417 	struct nvmet_tcp_queue *queue = cmd->queue;
418 	u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
419 
420 	cmd->offset = 0;
421 	cmd->state = NVMET_TCP_SEND_R2T;
422 
423 	pdu->hdr.type = nvme_tcp_r2t;
424 	pdu->hdr.flags = 0;
425 	pdu->hdr.hlen = sizeof(*pdu);
426 	pdu->hdr.pdo = 0;
427 	pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst);
428 
429 	pdu->command_id = cmd->req.cmd->common.command_id;
430 	pdu->ttag = nvmet_tcp_cmd_tag(cmd->queue, cmd);
431 	pdu->r2t_length = cpu_to_le32(cmd->req.transfer_len - cmd->rbytes_done);
432 	pdu->r2t_offset = cpu_to_le32(cmd->rbytes_done);
433 	if (cmd->queue->hdr_digest) {
434 		pdu->hdr.flags |= NVME_TCP_F_HDGST;
435 		nvmet_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
436 	}
437 }
438 
439 static void nvmet_setup_response_pdu(struct nvmet_tcp_cmd *cmd)
440 {
441 	struct nvme_tcp_rsp_pdu *pdu = cmd->rsp_pdu;
442 	struct nvmet_tcp_queue *queue = cmd->queue;
443 	u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
444 
445 	cmd->offset = 0;
446 	cmd->state = NVMET_TCP_SEND_RESPONSE;
447 
448 	pdu->hdr.type = nvme_tcp_rsp;
449 	pdu->hdr.flags = 0;
450 	pdu->hdr.hlen = sizeof(*pdu);
451 	pdu->hdr.pdo = 0;
452 	pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst);
453 	if (cmd->queue->hdr_digest) {
454 		pdu->hdr.flags |= NVME_TCP_F_HDGST;
455 		nvmet_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
456 	}
457 }
458 
459 static void nvmet_tcp_process_resp_list(struct nvmet_tcp_queue *queue)
460 {
461 	struct llist_node *node;
462 	struct nvmet_tcp_cmd *cmd;
463 
464 	for (node = llist_del_all(&queue->resp_list); node; node = node->next) {
465 		cmd = llist_entry(node, struct nvmet_tcp_cmd, lentry);
466 		list_add(&cmd->entry, &queue->resp_send_list);
467 		queue->send_list_len++;
468 	}
469 }
470 
471 static struct nvmet_tcp_cmd *nvmet_tcp_fetch_cmd(struct nvmet_tcp_queue *queue)
472 {
473 	queue->snd_cmd = list_first_entry_or_null(&queue->resp_send_list,
474 				struct nvmet_tcp_cmd, entry);
475 	if (!queue->snd_cmd) {
476 		nvmet_tcp_process_resp_list(queue);
477 		queue->snd_cmd =
478 			list_first_entry_or_null(&queue->resp_send_list,
479 					struct nvmet_tcp_cmd, entry);
480 		if (unlikely(!queue->snd_cmd))
481 			return NULL;
482 	}
483 
484 	list_del_init(&queue->snd_cmd->entry);
485 	queue->send_list_len--;
486 
487 	if (nvmet_tcp_need_data_out(queue->snd_cmd))
488 		nvmet_setup_c2h_data_pdu(queue->snd_cmd);
489 	else if (nvmet_tcp_need_data_in(queue->snd_cmd))
490 		nvmet_setup_r2t_pdu(queue->snd_cmd);
491 	else
492 		nvmet_setup_response_pdu(queue->snd_cmd);
493 
494 	return queue->snd_cmd;
495 }
496 
497 static void nvmet_tcp_queue_response(struct nvmet_req *req)
498 {
499 	struct nvmet_tcp_cmd *cmd =
500 		container_of(req, struct nvmet_tcp_cmd, req);
501 	struct nvmet_tcp_queue	*queue = cmd->queue;
502 
503 	llist_add(&cmd->lentry, &queue->resp_list);
504 	queue_work_on(cmd->queue->cpu, nvmet_tcp_wq, &cmd->queue->io_work);
505 }
506 
507 static int nvmet_try_send_data_pdu(struct nvmet_tcp_cmd *cmd)
508 {
509 	u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
510 	int left = sizeof(*cmd->data_pdu) - cmd->offset + hdgst;
511 	int ret;
512 
513 	ret = kernel_sendpage(cmd->queue->sock, virt_to_page(cmd->data_pdu),
514 			offset_in_page(cmd->data_pdu) + cmd->offset,
515 			left, MSG_DONTWAIT | MSG_MORE | MSG_SENDPAGE_NOTLAST);
516 	if (ret <= 0)
517 		return ret;
518 
519 	cmd->offset += ret;
520 	left -= ret;
521 
522 	if (left)
523 		return -EAGAIN;
524 
525 	cmd->state = NVMET_TCP_SEND_DATA;
526 	cmd->offset  = 0;
527 	return 1;
528 }
529 
530 static int nvmet_try_send_data(struct nvmet_tcp_cmd *cmd, bool last_in_batch)
531 {
532 	struct nvmet_tcp_queue *queue = cmd->queue;
533 	int ret;
534 
535 	while (cmd->cur_sg) {
536 		struct page *page = sg_page(cmd->cur_sg);
537 		u32 left = cmd->cur_sg->length - cmd->offset;
538 		int flags = MSG_DONTWAIT;
539 
540 		if ((!last_in_batch && cmd->queue->send_list_len) ||
541 		    cmd->wbytes_done + left < cmd->req.transfer_len ||
542 		    queue->data_digest || !queue->nvme_sq.sqhd_disabled)
543 			flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST;
544 
545 		ret = kernel_sendpage(cmd->queue->sock, page, cmd->offset,
546 					left, flags);
547 		if (ret <= 0)
548 			return ret;
549 
550 		cmd->offset += ret;
551 		cmd->wbytes_done += ret;
552 
553 		/* Done with sg?*/
554 		if (cmd->offset == cmd->cur_sg->length) {
555 			cmd->cur_sg = sg_next(cmd->cur_sg);
556 			cmd->offset = 0;
557 		}
558 	}
559 
560 	if (queue->data_digest) {
561 		cmd->state = NVMET_TCP_SEND_DDGST;
562 		cmd->offset = 0;
563 	} else {
564 		if (queue->nvme_sq.sqhd_disabled) {
565 			cmd->queue->snd_cmd = NULL;
566 			nvmet_tcp_put_cmd(cmd);
567 		} else {
568 			nvmet_setup_response_pdu(cmd);
569 		}
570 	}
571 
572 	if (queue->nvme_sq.sqhd_disabled) {
573 		kfree(cmd->iov);
574 		sgl_free(cmd->req.sg);
575 	}
576 
577 	return 1;
578 
579 }
580 
581 static int nvmet_try_send_response(struct nvmet_tcp_cmd *cmd,
582 		bool last_in_batch)
583 {
584 	u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
585 	int left = sizeof(*cmd->rsp_pdu) - cmd->offset + hdgst;
586 	int flags = MSG_DONTWAIT;
587 	int ret;
588 
589 	if (!last_in_batch && cmd->queue->send_list_len)
590 		flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST;
591 	else
592 		flags |= MSG_EOR;
593 
594 	ret = kernel_sendpage(cmd->queue->sock, virt_to_page(cmd->rsp_pdu),
595 		offset_in_page(cmd->rsp_pdu) + cmd->offset, left, flags);
596 	if (ret <= 0)
597 		return ret;
598 	cmd->offset += ret;
599 	left -= ret;
600 
601 	if (left)
602 		return -EAGAIN;
603 
604 	kfree(cmd->iov);
605 	sgl_free(cmd->req.sg);
606 	cmd->queue->snd_cmd = NULL;
607 	nvmet_tcp_put_cmd(cmd);
608 	return 1;
609 }
610 
611 static int nvmet_try_send_r2t(struct nvmet_tcp_cmd *cmd, bool last_in_batch)
612 {
613 	u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
614 	int left = sizeof(*cmd->r2t_pdu) - cmd->offset + hdgst;
615 	int flags = MSG_DONTWAIT;
616 	int ret;
617 
618 	if (!last_in_batch && cmd->queue->send_list_len)
619 		flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST;
620 	else
621 		flags |= MSG_EOR;
622 
623 	ret = kernel_sendpage(cmd->queue->sock, virt_to_page(cmd->r2t_pdu),
624 		offset_in_page(cmd->r2t_pdu) + cmd->offset, left, flags);
625 	if (ret <= 0)
626 		return ret;
627 	cmd->offset += ret;
628 	left -= ret;
629 
630 	if (left)
631 		return -EAGAIN;
632 
633 	cmd->queue->snd_cmd = NULL;
634 	return 1;
635 }
636 
637 static int nvmet_try_send_ddgst(struct nvmet_tcp_cmd *cmd, bool last_in_batch)
638 {
639 	struct nvmet_tcp_queue *queue = cmd->queue;
640 	struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
641 	struct kvec iov = {
642 		.iov_base = &cmd->exp_ddgst + cmd->offset,
643 		.iov_len = NVME_TCP_DIGEST_LENGTH - cmd->offset
644 	};
645 	int ret;
646 
647 	if (!last_in_batch && cmd->queue->send_list_len)
648 		msg.msg_flags |= MSG_MORE;
649 	else
650 		msg.msg_flags |= MSG_EOR;
651 
652 	ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
653 	if (unlikely(ret <= 0))
654 		return ret;
655 
656 	cmd->offset += ret;
657 
658 	if (queue->nvme_sq.sqhd_disabled) {
659 		cmd->queue->snd_cmd = NULL;
660 		nvmet_tcp_put_cmd(cmd);
661 	} else {
662 		nvmet_setup_response_pdu(cmd);
663 	}
664 	return 1;
665 }
666 
667 static int nvmet_tcp_try_send_one(struct nvmet_tcp_queue *queue,
668 		bool last_in_batch)
669 {
670 	struct nvmet_tcp_cmd *cmd = queue->snd_cmd;
671 	int ret = 0;
672 
673 	if (!cmd || queue->state == NVMET_TCP_Q_DISCONNECTING) {
674 		cmd = nvmet_tcp_fetch_cmd(queue);
675 		if (unlikely(!cmd))
676 			return 0;
677 	}
678 
679 	if (cmd->state == NVMET_TCP_SEND_DATA_PDU) {
680 		ret = nvmet_try_send_data_pdu(cmd);
681 		if (ret <= 0)
682 			goto done_send;
683 	}
684 
685 	if (cmd->state == NVMET_TCP_SEND_DATA) {
686 		ret = nvmet_try_send_data(cmd, last_in_batch);
687 		if (ret <= 0)
688 			goto done_send;
689 	}
690 
691 	if (cmd->state == NVMET_TCP_SEND_DDGST) {
692 		ret = nvmet_try_send_ddgst(cmd, last_in_batch);
693 		if (ret <= 0)
694 			goto done_send;
695 	}
696 
697 	if (cmd->state == NVMET_TCP_SEND_R2T) {
698 		ret = nvmet_try_send_r2t(cmd, last_in_batch);
699 		if (ret <= 0)
700 			goto done_send;
701 	}
702 
703 	if (cmd->state == NVMET_TCP_SEND_RESPONSE)
704 		ret = nvmet_try_send_response(cmd, last_in_batch);
705 
706 done_send:
707 	if (ret < 0) {
708 		if (ret == -EAGAIN)
709 			return 0;
710 		return ret;
711 	}
712 
713 	return 1;
714 }
715 
716 static int nvmet_tcp_try_send(struct nvmet_tcp_queue *queue,
717 		int budget, int *sends)
718 {
719 	int i, ret = 0;
720 
721 	for (i = 0; i < budget; i++) {
722 		ret = nvmet_tcp_try_send_one(queue, i == budget - 1);
723 		if (unlikely(ret < 0)) {
724 			nvmet_tcp_socket_error(queue, ret);
725 			goto done;
726 		} else if (ret == 0) {
727 			break;
728 		}
729 		(*sends)++;
730 	}
731 done:
732 	return ret;
733 }
734 
735 static void nvmet_prepare_receive_pdu(struct nvmet_tcp_queue *queue)
736 {
737 	queue->offset = 0;
738 	queue->left = sizeof(struct nvme_tcp_hdr);
739 	queue->cmd = NULL;
740 	queue->rcv_state = NVMET_TCP_RECV_PDU;
741 }
742 
743 static void nvmet_tcp_free_crypto(struct nvmet_tcp_queue *queue)
744 {
745 	struct crypto_ahash *tfm = crypto_ahash_reqtfm(queue->rcv_hash);
746 
747 	ahash_request_free(queue->rcv_hash);
748 	ahash_request_free(queue->snd_hash);
749 	crypto_free_ahash(tfm);
750 }
751 
752 static int nvmet_tcp_alloc_crypto(struct nvmet_tcp_queue *queue)
753 {
754 	struct crypto_ahash *tfm;
755 
756 	tfm = crypto_alloc_ahash("crc32c", 0, CRYPTO_ALG_ASYNC);
757 	if (IS_ERR(tfm))
758 		return PTR_ERR(tfm);
759 
760 	queue->snd_hash = ahash_request_alloc(tfm, GFP_KERNEL);
761 	if (!queue->snd_hash)
762 		goto free_tfm;
763 	ahash_request_set_callback(queue->snd_hash, 0, NULL, NULL);
764 
765 	queue->rcv_hash = ahash_request_alloc(tfm, GFP_KERNEL);
766 	if (!queue->rcv_hash)
767 		goto free_snd_hash;
768 	ahash_request_set_callback(queue->rcv_hash, 0, NULL, NULL);
769 
770 	return 0;
771 free_snd_hash:
772 	ahash_request_free(queue->snd_hash);
773 free_tfm:
774 	crypto_free_ahash(tfm);
775 	return -ENOMEM;
776 }
777 
778 
779 static int nvmet_tcp_handle_icreq(struct nvmet_tcp_queue *queue)
780 {
781 	struct nvme_tcp_icreq_pdu *icreq = &queue->pdu.icreq;
782 	struct nvme_tcp_icresp_pdu *icresp = &queue->pdu.icresp;
783 	struct msghdr msg = {};
784 	struct kvec iov;
785 	int ret;
786 
787 	if (le32_to_cpu(icreq->hdr.plen) != sizeof(struct nvme_tcp_icreq_pdu)) {
788 		pr_err("bad nvme-tcp pdu length (%d)\n",
789 			le32_to_cpu(icreq->hdr.plen));
790 		nvmet_tcp_fatal_error(queue);
791 	}
792 
793 	if (icreq->pfv != NVME_TCP_PFV_1_0) {
794 		pr_err("queue %d: bad pfv %d\n", queue->idx, icreq->pfv);
795 		return -EPROTO;
796 	}
797 
798 	if (icreq->hpda != 0) {
799 		pr_err("queue %d: unsupported hpda %d\n", queue->idx,
800 			icreq->hpda);
801 		return -EPROTO;
802 	}
803 
804 	queue->hdr_digest = !!(icreq->digest & NVME_TCP_HDR_DIGEST_ENABLE);
805 	queue->data_digest = !!(icreq->digest & NVME_TCP_DATA_DIGEST_ENABLE);
806 	if (queue->hdr_digest || queue->data_digest) {
807 		ret = nvmet_tcp_alloc_crypto(queue);
808 		if (ret)
809 			return ret;
810 	}
811 
812 	memset(icresp, 0, sizeof(*icresp));
813 	icresp->hdr.type = nvme_tcp_icresp;
814 	icresp->hdr.hlen = sizeof(*icresp);
815 	icresp->hdr.pdo = 0;
816 	icresp->hdr.plen = cpu_to_le32(icresp->hdr.hlen);
817 	icresp->pfv = cpu_to_le16(NVME_TCP_PFV_1_0);
818 	icresp->maxdata = cpu_to_le32(0x400000); /* 16M arbitrary limit */
819 	icresp->cpda = 0;
820 	if (queue->hdr_digest)
821 		icresp->digest |= NVME_TCP_HDR_DIGEST_ENABLE;
822 	if (queue->data_digest)
823 		icresp->digest |= NVME_TCP_DATA_DIGEST_ENABLE;
824 
825 	iov.iov_base = icresp;
826 	iov.iov_len = sizeof(*icresp);
827 	ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
828 	if (ret < 0)
829 		goto free_crypto;
830 
831 	queue->state = NVMET_TCP_Q_LIVE;
832 	nvmet_prepare_receive_pdu(queue);
833 	return 0;
834 free_crypto:
835 	if (queue->hdr_digest || queue->data_digest)
836 		nvmet_tcp_free_crypto(queue);
837 	return ret;
838 }
839 
840 static void nvmet_tcp_handle_req_failure(struct nvmet_tcp_queue *queue,
841 		struct nvmet_tcp_cmd *cmd, struct nvmet_req *req)
842 {
843 	size_t data_len = le32_to_cpu(req->cmd->common.dptr.sgl.length);
844 	int ret;
845 
846 	if (!nvme_is_write(cmd->req.cmd) ||
847 	    data_len > cmd->req.port->inline_data_size) {
848 		nvmet_prepare_receive_pdu(queue);
849 		return;
850 	}
851 
852 	ret = nvmet_tcp_map_data(cmd);
853 	if (unlikely(ret)) {
854 		pr_err("queue %d: failed to map data\n", queue->idx);
855 		nvmet_tcp_fatal_error(queue);
856 		return;
857 	}
858 
859 	queue->rcv_state = NVMET_TCP_RECV_DATA;
860 	nvmet_tcp_map_pdu_iovec(cmd);
861 	cmd->flags |= NVMET_TCP_F_INIT_FAILED;
862 }
863 
864 static int nvmet_tcp_handle_h2c_data_pdu(struct nvmet_tcp_queue *queue)
865 {
866 	struct nvme_tcp_data_pdu *data = &queue->pdu.data;
867 	struct nvmet_tcp_cmd *cmd;
868 
869 	cmd = &queue->cmds[data->ttag];
870 
871 	if (le32_to_cpu(data->data_offset) != cmd->rbytes_done) {
872 		pr_err("ttag %u unexpected data offset %u (expected %u)\n",
873 			data->ttag, le32_to_cpu(data->data_offset),
874 			cmd->rbytes_done);
875 		/* FIXME: use path and transport errors */
876 		nvmet_req_complete(&cmd->req,
877 			NVME_SC_INVALID_FIELD | NVME_SC_DNR);
878 		return -EPROTO;
879 	}
880 
881 	cmd->pdu_len = le32_to_cpu(data->data_length);
882 	cmd->pdu_recv = 0;
883 	nvmet_tcp_map_pdu_iovec(cmd);
884 	queue->cmd = cmd;
885 	queue->rcv_state = NVMET_TCP_RECV_DATA;
886 
887 	return 0;
888 }
889 
890 static int nvmet_tcp_done_recv_pdu(struct nvmet_tcp_queue *queue)
891 {
892 	struct nvme_tcp_hdr *hdr = &queue->pdu.cmd.hdr;
893 	struct nvme_command *nvme_cmd = &queue->pdu.cmd.cmd;
894 	struct nvmet_req *req;
895 	int ret;
896 
897 	if (unlikely(queue->state == NVMET_TCP_Q_CONNECTING)) {
898 		if (hdr->type != nvme_tcp_icreq) {
899 			pr_err("unexpected pdu type (%d) before icreq\n",
900 				hdr->type);
901 			nvmet_tcp_fatal_error(queue);
902 			return -EPROTO;
903 		}
904 		return nvmet_tcp_handle_icreq(queue);
905 	}
906 
907 	if (hdr->type == nvme_tcp_h2c_data) {
908 		ret = nvmet_tcp_handle_h2c_data_pdu(queue);
909 		if (unlikely(ret))
910 			return ret;
911 		return 0;
912 	}
913 
914 	queue->cmd = nvmet_tcp_get_cmd(queue);
915 	if (unlikely(!queue->cmd)) {
916 		/* This should never happen */
917 		pr_err("queue %d: out of commands (%d) send_list_len: %d, opcode: %d",
918 			queue->idx, queue->nr_cmds, queue->send_list_len,
919 			nvme_cmd->common.opcode);
920 		nvmet_tcp_fatal_error(queue);
921 		return -ENOMEM;
922 	}
923 
924 	req = &queue->cmd->req;
925 	memcpy(req->cmd, nvme_cmd, sizeof(*nvme_cmd));
926 
927 	if (unlikely(!nvmet_req_init(req, &queue->nvme_cq,
928 			&queue->nvme_sq, &nvmet_tcp_ops))) {
929 		pr_err("failed cmd %p id %d opcode %d, data_len: %d\n",
930 			req->cmd, req->cmd->common.command_id,
931 			req->cmd->common.opcode,
932 			le32_to_cpu(req->cmd->common.dptr.sgl.length));
933 
934 		nvmet_tcp_handle_req_failure(queue, queue->cmd, req);
935 		return -EAGAIN;
936 	}
937 
938 	ret = nvmet_tcp_map_data(queue->cmd);
939 	if (unlikely(ret)) {
940 		pr_err("queue %d: failed to map data\n", queue->idx);
941 		if (nvmet_tcp_has_inline_data(queue->cmd))
942 			nvmet_tcp_fatal_error(queue);
943 		else
944 			nvmet_req_complete(req, ret);
945 		ret = -EAGAIN;
946 		goto out;
947 	}
948 
949 	if (nvmet_tcp_need_data_in(queue->cmd)) {
950 		if (nvmet_tcp_has_inline_data(queue->cmd)) {
951 			queue->rcv_state = NVMET_TCP_RECV_DATA;
952 			nvmet_tcp_map_pdu_iovec(queue->cmd);
953 			return 0;
954 		}
955 		/* send back R2T */
956 		nvmet_tcp_queue_response(&queue->cmd->req);
957 		goto out;
958 	}
959 
960 	queue->cmd->req.execute(&queue->cmd->req);
961 out:
962 	nvmet_prepare_receive_pdu(queue);
963 	return ret;
964 }
965 
966 static const u8 nvme_tcp_pdu_sizes[] = {
967 	[nvme_tcp_icreq]	= sizeof(struct nvme_tcp_icreq_pdu),
968 	[nvme_tcp_cmd]		= sizeof(struct nvme_tcp_cmd_pdu),
969 	[nvme_tcp_h2c_data]	= sizeof(struct nvme_tcp_data_pdu),
970 };
971 
972 static inline u8 nvmet_tcp_pdu_size(u8 type)
973 {
974 	size_t idx = type;
975 
976 	return (idx < ARRAY_SIZE(nvme_tcp_pdu_sizes) &&
977 		nvme_tcp_pdu_sizes[idx]) ?
978 			nvme_tcp_pdu_sizes[idx] : 0;
979 }
980 
981 static inline bool nvmet_tcp_pdu_valid(u8 type)
982 {
983 	switch (type) {
984 	case nvme_tcp_icreq:
985 	case nvme_tcp_cmd:
986 	case nvme_tcp_h2c_data:
987 		/* fallthru */
988 		return true;
989 	}
990 
991 	return false;
992 }
993 
994 static int nvmet_tcp_try_recv_pdu(struct nvmet_tcp_queue *queue)
995 {
996 	struct nvme_tcp_hdr *hdr = &queue->pdu.cmd.hdr;
997 	int len;
998 	struct kvec iov;
999 	struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
1000 
1001 recv:
1002 	iov.iov_base = (void *)&queue->pdu + queue->offset;
1003 	iov.iov_len = queue->left;
1004 	len = kernel_recvmsg(queue->sock, &msg, &iov, 1,
1005 			iov.iov_len, msg.msg_flags);
1006 	if (unlikely(len < 0))
1007 		return len;
1008 
1009 	queue->offset += len;
1010 	queue->left -= len;
1011 	if (queue->left)
1012 		return -EAGAIN;
1013 
1014 	if (queue->offset == sizeof(struct nvme_tcp_hdr)) {
1015 		u8 hdgst = nvmet_tcp_hdgst_len(queue);
1016 
1017 		if (unlikely(!nvmet_tcp_pdu_valid(hdr->type))) {
1018 			pr_err("unexpected pdu type %d\n", hdr->type);
1019 			nvmet_tcp_fatal_error(queue);
1020 			return -EIO;
1021 		}
1022 
1023 		if (unlikely(hdr->hlen != nvmet_tcp_pdu_size(hdr->type))) {
1024 			pr_err("pdu %d bad hlen %d\n", hdr->type, hdr->hlen);
1025 			return -EIO;
1026 		}
1027 
1028 		queue->left = hdr->hlen - queue->offset + hdgst;
1029 		goto recv;
1030 	}
1031 
1032 	if (queue->hdr_digest &&
1033 	    nvmet_tcp_verify_hdgst(queue, &queue->pdu, queue->offset)) {
1034 		nvmet_tcp_fatal_error(queue); /* fatal */
1035 		return -EPROTO;
1036 	}
1037 
1038 	if (queue->data_digest &&
1039 	    nvmet_tcp_check_ddgst(queue, &queue->pdu)) {
1040 		nvmet_tcp_fatal_error(queue); /* fatal */
1041 		return -EPROTO;
1042 	}
1043 
1044 	return nvmet_tcp_done_recv_pdu(queue);
1045 }
1046 
1047 static void nvmet_tcp_prep_recv_ddgst(struct nvmet_tcp_cmd *cmd)
1048 {
1049 	struct nvmet_tcp_queue *queue = cmd->queue;
1050 
1051 	nvmet_tcp_ddgst(queue->rcv_hash, cmd);
1052 	queue->offset = 0;
1053 	queue->left = NVME_TCP_DIGEST_LENGTH;
1054 	queue->rcv_state = NVMET_TCP_RECV_DDGST;
1055 }
1056 
1057 static int nvmet_tcp_try_recv_data(struct nvmet_tcp_queue *queue)
1058 {
1059 	struct nvmet_tcp_cmd  *cmd = queue->cmd;
1060 	int ret;
1061 
1062 	while (msg_data_left(&cmd->recv_msg)) {
1063 		ret = sock_recvmsg(cmd->queue->sock, &cmd->recv_msg,
1064 			cmd->recv_msg.msg_flags);
1065 		if (ret <= 0)
1066 			return ret;
1067 
1068 		cmd->pdu_recv += ret;
1069 		cmd->rbytes_done += ret;
1070 	}
1071 
1072 	nvmet_tcp_unmap_pdu_iovec(cmd);
1073 
1074 	if (!(cmd->flags & NVMET_TCP_F_INIT_FAILED) &&
1075 	    cmd->rbytes_done == cmd->req.transfer_len) {
1076 		if (queue->data_digest) {
1077 			nvmet_tcp_prep_recv_ddgst(cmd);
1078 			return 0;
1079 		}
1080 		cmd->req.execute(&cmd->req);
1081 	}
1082 
1083 	nvmet_prepare_receive_pdu(queue);
1084 	return 0;
1085 }
1086 
1087 static int nvmet_tcp_try_recv_ddgst(struct nvmet_tcp_queue *queue)
1088 {
1089 	struct nvmet_tcp_cmd *cmd = queue->cmd;
1090 	int ret;
1091 	struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
1092 	struct kvec iov = {
1093 		.iov_base = (void *)&cmd->recv_ddgst + queue->offset,
1094 		.iov_len = queue->left
1095 	};
1096 
1097 	ret = kernel_recvmsg(queue->sock, &msg, &iov, 1,
1098 			iov.iov_len, msg.msg_flags);
1099 	if (unlikely(ret < 0))
1100 		return ret;
1101 
1102 	queue->offset += ret;
1103 	queue->left -= ret;
1104 	if (queue->left)
1105 		return -EAGAIN;
1106 
1107 	if (queue->data_digest && cmd->exp_ddgst != cmd->recv_ddgst) {
1108 		pr_err("queue %d: cmd %d pdu (%d) data digest error: recv %#x expected %#x\n",
1109 			queue->idx, cmd->req.cmd->common.command_id,
1110 			queue->pdu.cmd.hdr.type, le32_to_cpu(cmd->recv_ddgst),
1111 			le32_to_cpu(cmd->exp_ddgst));
1112 		nvmet_tcp_finish_cmd(cmd);
1113 		nvmet_tcp_fatal_error(queue);
1114 		ret = -EPROTO;
1115 		goto out;
1116 	}
1117 
1118 	if (!(cmd->flags & NVMET_TCP_F_INIT_FAILED) &&
1119 	    cmd->rbytes_done == cmd->req.transfer_len)
1120 		cmd->req.execute(&cmd->req);
1121 	ret = 0;
1122 out:
1123 	nvmet_prepare_receive_pdu(queue);
1124 	return ret;
1125 }
1126 
1127 static int nvmet_tcp_try_recv_one(struct nvmet_tcp_queue *queue)
1128 {
1129 	int result = 0;
1130 
1131 	if (unlikely(queue->rcv_state == NVMET_TCP_RECV_ERR))
1132 		return 0;
1133 
1134 	if (queue->rcv_state == NVMET_TCP_RECV_PDU) {
1135 		result = nvmet_tcp_try_recv_pdu(queue);
1136 		if (result != 0)
1137 			goto done_recv;
1138 	}
1139 
1140 	if (queue->rcv_state == NVMET_TCP_RECV_DATA) {
1141 		result = nvmet_tcp_try_recv_data(queue);
1142 		if (result != 0)
1143 			goto done_recv;
1144 	}
1145 
1146 	if (queue->rcv_state == NVMET_TCP_RECV_DDGST) {
1147 		result = nvmet_tcp_try_recv_ddgst(queue);
1148 		if (result != 0)
1149 			goto done_recv;
1150 	}
1151 
1152 done_recv:
1153 	if (result < 0) {
1154 		if (result == -EAGAIN)
1155 			return 0;
1156 		return result;
1157 	}
1158 	return 1;
1159 }
1160 
1161 static int nvmet_tcp_try_recv(struct nvmet_tcp_queue *queue,
1162 		int budget, int *recvs)
1163 {
1164 	int i, ret = 0;
1165 
1166 	for (i = 0; i < budget; i++) {
1167 		ret = nvmet_tcp_try_recv_one(queue);
1168 		if (unlikely(ret < 0)) {
1169 			nvmet_tcp_socket_error(queue, ret);
1170 			goto done;
1171 		} else if (ret == 0) {
1172 			break;
1173 		}
1174 		(*recvs)++;
1175 	}
1176 done:
1177 	return ret;
1178 }
1179 
1180 static void nvmet_tcp_schedule_release_queue(struct nvmet_tcp_queue *queue)
1181 {
1182 	spin_lock(&queue->state_lock);
1183 	if (queue->state != NVMET_TCP_Q_DISCONNECTING) {
1184 		queue->state = NVMET_TCP_Q_DISCONNECTING;
1185 		schedule_work(&queue->release_work);
1186 	}
1187 	spin_unlock(&queue->state_lock);
1188 }
1189 
1190 static void nvmet_tcp_io_work(struct work_struct *w)
1191 {
1192 	struct nvmet_tcp_queue *queue =
1193 		container_of(w, struct nvmet_tcp_queue, io_work);
1194 	bool pending;
1195 	int ret, ops = 0;
1196 
1197 	do {
1198 		pending = false;
1199 
1200 		ret = nvmet_tcp_try_recv(queue, NVMET_TCP_RECV_BUDGET, &ops);
1201 		if (ret > 0)
1202 			pending = true;
1203 		else if (ret < 0)
1204 			return;
1205 
1206 		ret = nvmet_tcp_try_send(queue, NVMET_TCP_SEND_BUDGET, &ops);
1207 		if (ret > 0)
1208 			pending = true;
1209 		else if (ret < 0)
1210 			return;
1211 
1212 	} while (pending && ops < NVMET_TCP_IO_WORK_BUDGET);
1213 
1214 	/*
1215 	 * We exahusted our budget, requeue our selves
1216 	 */
1217 	if (pending)
1218 		queue_work_on(queue->cpu, nvmet_tcp_wq, &queue->io_work);
1219 }
1220 
1221 static int nvmet_tcp_alloc_cmd(struct nvmet_tcp_queue *queue,
1222 		struct nvmet_tcp_cmd *c)
1223 {
1224 	u8 hdgst = nvmet_tcp_hdgst_len(queue);
1225 
1226 	c->queue = queue;
1227 	c->req.port = queue->port->nport;
1228 
1229 	c->cmd_pdu = page_frag_alloc(&queue->pf_cache,
1230 			sizeof(*c->cmd_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO);
1231 	if (!c->cmd_pdu)
1232 		return -ENOMEM;
1233 	c->req.cmd = &c->cmd_pdu->cmd;
1234 
1235 	c->rsp_pdu = page_frag_alloc(&queue->pf_cache,
1236 			sizeof(*c->rsp_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO);
1237 	if (!c->rsp_pdu)
1238 		goto out_free_cmd;
1239 	c->req.cqe = &c->rsp_pdu->cqe;
1240 
1241 	c->data_pdu = page_frag_alloc(&queue->pf_cache,
1242 			sizeof(*c->data_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO);
1243 	if (!c->data_pdu)
1244 		goto out_free_rsp;
1245 
1246 	c->r2t_pdu = page_frag_alloc(&queue->pf_cache,
1247 			sizeof(*c->r2t_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO);
1248 	if (!c->r2t_pdu)
1249 		goto out_free_data;
1250 
1251 	c->recv_msg.msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
1252 
1253 	list_add_tail(&c->entry, &queue->free_list);
1254 
1255 	return 0;
1256 out_free_data:
1257 	page_frag_free(c->data_pdu);
1258 out_free_rsp:
1259 	page_frag_free(c->rsp_pdu);
1260 out_free_cmd:
1261 	page_frag_free(c->cmd_pdu);
1262 	return -ENOMEM;
1263 }
1264 
1265 static void nvmet_tcp_free_cmd(struct nvmet_tcp_cmd *c)
1266 {
1267 	page_frag_free(c->r2t_pdu);
1268 	page_frag_free(c->data_pdu);
1269 	page_frag_free(c->rsp_pdu);
1270 	page_frag_free(c->cmd_pdu);
1271 }
1272 
1273 static int nvmet_tcp_alloc_cmds(struct nvmet_tcp_queue *queue)
1274 {
1275 	struct nvmet_tcp_cmd *cmds;
1276 	int i, ret = -EINVAL, nr_cmds = queue->nr_cmds;
1277 
1278 	cmds = kcalloc(nr_cmds, sizeof(struct nvmet_tcp_cmd), GFP_KERNEL);
1279 	if (!cmds)
1280 		goto out;
1281 
1282 	for (i = 0; i < nr_cmds; i++) {
1283 		ret = nvmet_tcp_alloc_cmd(queue, cmds + i);
1284 		if (ret)
1285 			goto out_free;
1286 	}
1287 
1288 	queue->cmds = cmds;
1289 
1290 	return 0;
1291 out_free:
1292 	while (--i >= 0)
1293 		nvmet_tcp_free_cmd(cmds + i);
1294 	kfree(cmds);
1295 out:
1296 	return ret;
1297 }
1298 
1299 static void nvmet_tcp_free_cmds(struct nvmet_tcp_queue *queue)
1300 {
1301 	struct nvmet_tcp_cmd *cmds = queue->cmds;
1302 	int i;
1303 
1304 	for (i = 0; i < queue->nr_cmds; i++)
1305 		nvmet_tcp_free_cmd(cmds + i);
1306 
1307 	nvmet_tcp_free_cmd(&queue->connect);
1308 	kfree(cmds);
1309 }
1310 
1311 static void nvmet_tcp_restore_socket_callbacks(struct nvmet_tcp_queue *queue)
1312 {
1313 	struct socket *sock = queue->sock;
1314 
1315 	write_lock_bh(&sock->sk->sk_callback_lock);
1316 	sock->sk->sk_data_ready =  queue->data_ready;
1317 	sock->sk->sk_state_change = queue->state_change;
1318 	sock->sk->sk_write_space = queue->write_space;
1319 	sock->sk->sk_user_data = NULL;
1320 	write_unlock_bh(&sock->sk->sk_callback_lock);
1321 }
1322 
1323 static void nvmet_tcp_finish_cmd(struct nvmet_tcp_cmd *cmd)
1324 {
1325 	nvmet_req_uninit(&cmd->req);
1326 	nvmet_tcp_unmap_pdu_iovec(cmd);
1327 	kfree(cmd->iov);
1328 	sgl_free(cmd->req.sg);
1329 }
1330 
1331 static void nvmet_tcp_uninit_data_in_cmds(struct nvmet_tcp_queue *queue)
1332 {
1333 	struct nvmet_tcp_cmd *cmd = queue->cmds;
1334 	int i;
1335 
1336 	for (i = 0; i < queue->nr_cmds; i++, cmd++) {
1337 		if (nvmet_tcp_need_data_in(cmd))
1338 			nvmet_tcp_finish_cmd(cmd);
1339 	}
1340 
1341 	if (!queue->nr_cmds && nvmet_tcp_need_data_in(&queue->connect)) {
1342 		/* failed in connect */
1343 		nvmet_tcp_finish_cmd(&queue->connect);
1344 	}
1345 }
1346 
1347 static void nvmet_tcp_release_queue_work(struct work_struct *w)
1348 {
1349 	struct nvmet_tcp_queue *queue =
1350 		container_of(w, struct nvmet_tcp_queue, release_work);
1351 
1352 	mutex_lock(&nvmet_tcp_queue_mutex);
1353 	list_del_init(&queue->queue_list);
1354 	mutex_unlock(&nvmet_tcp_queue_mutex);
1355 
1356 	nvmet_tcp_restore_socket_callbacks(queue);
1357 	flush_work(&queue->io_work);
1358 
1359 	nvmet_tcp_uninit_data_in_cmds(queue);
1360 	nvmet_sq_destroy(&queue->nvme_sq);
1361 	cancel_work_sync(&queue->io_work);
1362 	sock_release(queue->sock);
1363 	nvmet_tcp_free_cmds(queue);
1364 	if (queue->hdr_digest || queue->data_digest)
1365 		nvmet_tcp_free_crypto(queue);
1366 	ida_simple_remove(&nvmet_tcp_queue_ida, queue->idx);
1367 
1368 	kfree(queue);
1369 }
1370 
1371 static void nvmet_tcp_data_ready(struct sock *sk)
1372 {
1373 	struct nvmet_tcp_queue *queue;
1374 
1375 	read_lock_bh(&sk->sk_callback_lock);
1376 	queue = sk->sk_user_data;
1377 	if (likely(queue))
1378 		queue_work_on(queue->cpu, nvmet_tcp_wq, &queue->io_work);
1379 	read_unlock_bh(&sk->sk_callback_lock);
1380 }
1381 
1382 static void nvmet_tcp_write_space(struct sock *sk)
1383 {
1384 	struct nvmet_tcp_queue *queue;
1385 
1386 	read_lock_bh(&sk->sk_callback_lock);
1387 	queue = sk->sk_user_data;
1388 	if (unlikely(!queue))
1389 		goto out;
1390 
1391 	if (unlikely(queue->state == NVMET_TCP_Q_CONNECTING)) {
1392 		queue->write_space(sk);
1393 		goto out;
1394 	}
1395 
1396 	if (sk_stream_is_writeable(sk)) {
1397 		clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1398 		queue_work_on(queue->cpu, nvmet_tcp_wq, &queue->io_work);
1399 	}
1400 out:
1401 	read_unlock_bh(&sk->sk_callback_lock);
1402 }
1403 
1404 static void nvmet_tcp_state_change(struct sock *sk)
1405 {
1406 	struct nvmet_tcp_queue *queue;
1407 
1408 	write_lock_bh(&sk->sk_callback_lock);
1409 	queue = sk->sk_user_data;
1410 	if (!queue)
1411 		goto done;
1412 
1413 	switch (sk->sk_state) {
1414 	case TCP_FIN_WAIT1:
1415 	case TCP_CLOSE_WAIT:
1416 	case TCP_CLOSE:
1417 		/* FALLTHRU */
1418 		sk->sk_user_data = NULL;
1419 		nvmet_tcp_schedule_release_queue(queue);
1420 		break;
1421 	default:
1422 		pr_warn("queue %d unhandled state %d\n",
1423 			queue->idx, sk->sk_state);
1424 	}
1425 done:
1426 	write_unlock_bh(&sk->sk_callback_lock);
1427 }
1428 
1429 static int nvmet_tcp_set_queue_sock(struct nvmet_tcp_queue *queue)
1430 {
1431 	struct socket *sock = queue->sock;
1432 	struct inet_sock *inet = inet_sk(sock->sk);
1433 	int ret;
1434 
1435 	ret = kernel_getsockname(sock,
1436 		(struct sockaddr *)&queue->sockaddr);
1437 	if (ret < 0)
1438 		return ret;
1439 
1440 	ret = kernel_getpeername(sock,
1441 		(struct sockaddr *)&queue->sockaddr_peer);
1442 	if (ret < 0)
1443 		return ret;
1444 
1445 	/*
1446 	 * Cleanup whatever is sitting in the TCP transmit queue on socket
1447 	 * close. This is done to prevent stale data from being sent should
1448 	 * the network connection be restored before TCP times out.
1449 	 */
1450 	sock_no_linger(sock->sk);
1451 
1452 	if (so_priority > 0)
1453 		sock_set_priority(sock->sk, so_priority);
1454 
1455 	/* Set socket type of service */
1456 	if (inet->rcv_tos > 0)
1457 		ip_sock_set_tos(sock->sk, inet->rcv_tos);
1458 
1459 	write_lock_bh(&sock->sk->sk_callback_lock);
1460 	sock->sk->sk_user_data = queue;
1461 	queue->data_ready = sock->sk->sk_data_ready;
1462 	sock->sk->sk_data_ready = nvmet_tcp_data_ready;
1463 	queue->state_change = sock->sk->sk_state_change;
1464 	sock->sk->sk_state_change = nvmet_tcp_state_change;
1465 	queue->write_space = sock->sk->sk_write_space;
1466 	sock->sk->sk_write_space = nvmet_tcp_write_space;
1467 	write_unlock_bh(&sock->sk->sk_callback_lock);
1468 
1469 	return 0;
1470 }
1471 
1472 static int nvmet_tcp_alloc_queue(struct nvmet_tcp_port *port,
1473 		struct socket *newsock)
1474 {
1475 	struct nvmet_tcp_queue *queue;
1476 	int ret;
1477 
1478 	queue = kzalloc(sizeof(*queue), GFP_KERNEL);
1479 	if (!queue)
1480 		return -ENOMEM;
1481 
1482 	INIT_WORK(&queue->release_work, nvmet_tcp_release_queue_work);
1483 	INIT_WORK(&queue->io_work, nvmet_tcp_io_work);
1484 	queue->sock = newsock;
1485 	queue->port = port;
1486 	queue->nr_cmds = 0;
1487 	spin_lock_init(&queue->state_lock);
1488 	queue->state = NVMET_TCP_Q_CONNECTING;
1489 	INIT_LIST_HEAD(&queue->free_list);
1490 	init_llist_head(&queue->resp_list);
1491 	INIT_LIST_HEAD(&queue->resp_send_list);
1492 
1493 	queue->idx = ida_simple_get(&nvmet_tcp_queue_ida, 0, 0, GFP_KERNEL);
1494 	if (queue->idx < 0) {
1495 		ret = queue->idx;
1496 		goto out_free_queue;
1497 	}
1498 
1499 	ret = nvmet_tcp_alloc_cmd(queue, &queue->connect);
1500 	if (ret)
1501 		goto out_ida_remove;
1502 
1503 	ret = nvmet_sq_init(&queue->nvme_sq);
1504 	if (ret)
1505 		goto out_free_connect;
1506 
1507 	port->last_cpu = cpumask_next_wrap(port->last_cpu,
1508 				cpu_online_mask, -1, false);
1509 	queue->cpu = port->last_cpu;
1510 	nvmet_prepare_receive_pdu(queue);
1511 
1512 	mutex_lock(&nvmet_tcp_queue_mutex);
1513 	list_add_tail(&queue->queue_list, &nvmet_tcp_queue_list);
1514 	mutex_unlock(&nvmet_tcp_queue_mutex);
1515 
1516 	ret = nvmet_tcp_set_queue_sock(queue);
1517 	if (ret)
1518 		goto out_destroy_sq;
1519 
1520 	queue_work_on(queue->cpu, nvmet_tcp_wq, &queue->io_work);
1521 
1522 	return 0;
1523 out_destroy_sq:
1524 	mutex_lock(&nvmet_tcp_queue_mutex);
1525 	list_del_init(&queue->queue_list);
1526 	mutex_unlock(&nvmet_tcp_queue_mutex);
1527 	nvmet_sq_destroy(&queue->nvme_sq);
1528 out_free_connect:
1529 	nvmet_tcp_free_cmd(&queue->connect);
1530 out_ida_remove:
1531 	ida_simple_remove(&nvmet_tcp_queue_ida, queue->idx);
1532 out_free_queue:
1533 	kfree(queue);
1534 	return ret;
1535 }
1536 
1537 static void nvmet_tcp_accept_work(struct work_struct *w)
1538 {
1539 	struct nvmet_tcp_port *port =
1540 		container_of(w, struct nvmet_tcp_port, accept_work);
1541 	struct socket *newsock;
1542 	int ret;
1543 
1544 	while (true) {
1545 		ret = kernel_accept(port->sock, &newsock, O_NONBLOCK);
1546 		if (ret < 0) {
1547 			if (ret != -EAGAIN)
1548 				pr_warn("failed to accept err=%d\n", ret);
1549 			return;
1550 		}
1551 		ret = nvmet_tcp_alloc_queue(port, newsock);
1552 		if (ret) {
1553 			pr_err("failed to allocate queue\n");
1554 			sock_release(newsock);
1555 		}
1556 	}
1557 }
1558 
1559 static void nvmet_tcp_listen_data_ready(struct sock *sk)
1560 {
1561 	struct nvmet_tcp_port *port;
1562 
1563 	read_lock_bh(&sk->sk_callback_lock);
1564 	port = sk->sk_user_data;
1565 	if (!port)
1566 		goto out;
1567 
1568 	if (sk->sk_state == TCP_LISTEN)
1569 		schedule_work(&port->accept_work);
1570 out:
1571 	read_unlock_bh(&sk->sk_callback_lock);
1572 }
1573 
1574 static int nvmet_tcp_add_port(struct nvmet_port *nport)
1575 {
1576 	struct nvmet_tcp_port *port;
1577 	__kernel_sa_family_t af;
1578 	int ret;
1579 
1580 	port = kzalloc(sizeof(*port), GFP_KERNEL);
1581 	if (!port)
1582 		return -ENOMEM;
1583 
1584 	switch (nport->disc_addr.adrfam) {
1585 	case NVMF_ADDR_FAMILY_IP4:
1586 		af = AF_INET;
1587 		break;
1588 	case NVMF_ADDR_FAMILY_IP6:
1589 		af = AF_INET6;
1590 		break;
1591 	default:
1592 		pr_err("address family %d not supported\n",
1593 				nport->disc_addr.adrfam);
1594 		ret = -EINVAL;
1595 		goto err_port;
1596 	}
1597 
1598 	ret = inet_pton_with_scope(&init_net, af, nport->disc_addr.traddr,
1599 			nport->disc_addr.trsvcid, &port->addr);
1600 	if (ret) {
1601 		pr_err("malformed ip/port passed: %s:%s\n",
1602 			nport->disc_addr.traddr, nport->disc_addr.trsvcid);
1603 		goto err_port;
1604 	}
1605 
1606 	port->nport = nport;
1607 	port->last_cpu = -1;
1608 	INIT_WORK(&port->accept_work, nvmet_tcp_accept_work);
1609 	if (port->nport->inline_data_size < 0)
1610 		port->nport->inline_data_size = NVMET_TCP_DEF_INLINE_DATA_SIZE;
1611 
1612 	ret = sock_create(port->addr.ss_family, SOCK_STREAM,
1613 				IPPROTO_TCP, &port->sock);
1614 	if (ret) {
1615 		pr_err("failed to create a socket\n");
1616 		goto err_port;
1617 	}
1618 
1619 	port->sock->sk->sk_user_data = port;
1620 	port->data_ready = port->sock->sk->sk_data_ready;
1621 	port->sock->sk->sk_data_ready = nvmet_tcp_listen_data_ready;
1622 	sock_set_reuseaddr(port->sock->sk);
1623 	tcp_sock_set_nodelay(port->sock->sk);
1624 	if (so_priority > 0)
1625 		sock_set_priority(port->sock->sk, so_priority);
1626 
1627 	ret = kernel_bind(port->sock, (struct sockaddr *)&port->addr,
1628 			sizeof(port->addr));
1629 	if (ret) {
1630 		pr_err("failed to bind port socket %d\n", ret);
1631 		goto err_sock;
1632 	}
1633 
1634 	ret = kernel_listen(port->sock, 128);
1635 	if (ret) {
1636 		pr_err("failed to listen %d on port sock\n", ret);
1637 		goto err_sock;
1638 	}
1639 
1640 	nport->priv = port;
1641 	pr_info("enabling port %d (%pISpc)\n",
1642 		le16_to_cpu(nport->disc_addr.portid), &port->addr);
1643 
1644 	return 0;
1645 
1646 err_sock:
1647 	sock_release(port->sock);
1648 err_port:
1649 	kfree(port);
1650 	return ret;
1651 }
1652 
1653 static void nvmet_tcp_remove_port(struct nvmet_port *nport)
1654 {
1655 	struct nvmet_tcp_port *port = nport->priv;
1656 
1657 	write_lock_bh(&port->sock->sk->sk_callback_lock);
1658 	port->sock->sk->sk_data_ready = port->data_ready;
1659 	port->sock->sk->sk_user_data = NULL;
1660 	write_unlock_bh(&port->sock->sk->sk_callback_lock);
1661 	cancel_work_sync(&port->accept_work);
1662 
1663 	sock_release(port->sock);
1664 	kfree(port);
1665 }
1666 
1667 static void nvmet_tcp_delete_ctrl(struct nvmet_ctrl *ctrl)
1668 {
1669 	struct nvmet_tcp_queue *queue;
1670 
1671 	mutex_lock(&nvmet_tcp_queue_mutex);
1672 	list_for_each_entry(queue, &nvmet_tcp_queue_list, queue_list)
1673 		if (queue->nvme_sq.ctrl == ctrl)
1674 			kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1675 	mutex_unlock(&nvmet_tcp_queue_mutex);
1676 }
1677 
1678 static u16 nvmet_tcp_install_queue(struct nvmet_sq *sq)
1679 {
1680 	struct nvmet_tcp_queue *queue =
1681 		container_of(sq, struct nvmet_tcp_queue, nvme_sq);
1682 
1683 	if (sq->qid == 0) {
1684 		/* Let inflight controller teardown complete */
1685 		flush_scheduled_work();
1686 	}
1687 
1688 	queue->nr_cmds = sq->size * 2;
1689 	if (nvmet_tcp_alloc_cmds(queue))
1690 		return NVME_SC_INTERNAL;
1691 	return 0;
1692 }
1693 
1694 static void nvmet_tcp_disc_port_addr(struct nvmet_req *req,
1695 		struct nvmet_port *nport, char *traddr)
1696 {
1697 	struct nvmet_tcp_port *port = nport->priv;
1698 
1699 	if (inet_addr_is_any((struct sockaddr *)&port->addr)) {
1700 		struct nvmet_tcp_cmd *cmd =
1701 			container_of(req, struct nvmet_tcp_cmd, req);
1702 		struct nvmet_tcp_queue *queue = cmd->queue;
1703 
1704 		sprintf(traddr, "%pISc", (struct sockaddr *)&queue->sockaddr);
1705 	} else {
1706 		memcpy(traddr, nport->disc_addr.traddr, NVMF_TRADDR_SIZE);
1707 	}
1708 }
1709 
1710 static const struct nvmet_fabrics_ops nvmet_tcp_ops = {
1711 	.owner			= THIS_MODULE,
1712 	.type			= NVMF_TRTYPE_TCP,
1713 	.msdbd			= 1,
1714 	.add_port		= nvmet_tcp_add_port,
1715 	.remove_port		= nvmet_tcp_remove_port,
1716 	.queue_response		= nvmet_tcp_queue_response,
1717 	.delete_ctrl		= nvmet_tcp_delete_ctrl,
1718 	.install_queue		= nvmet_tcp_install_queue,
1719 	.disc_traddr		= nvmet_tcp_disc_port_addr,
1720 };
1721 
1722 static int __init nvmet_tcp_init(void)
1723 {
1724 	int ret;
1725 
1726 	nvmet_tcp_wq = alloc_workqueue("nvmet_tcp_wq", WQ_HIGHPRI, 0);
1727 	if (!nvmet_tcp_wq)
1728 		return -ENOMEM;
1729 
1730 	ret = nvmet_register_transport(&nvmet_tcp_ops);
1731 	if (ret)
1732 		goto err;
1733 
1734 	return 0;
1735 err:
1736 	destroy_workqueue(nvmet_tcp_wq);
1737 	return ret;
1738 }
1739 
1740 static void __exit nvmet_tcp_exit(void)
1741 {
1742 	struct nvmet_tcp_queue *queue;
1743 
1744 	nvmet_unregister_transport(&nvmet_tcp_ops);
1745 
1746 	flush_scheduled_work();
1747 	mutex_lock(&nvmet_tcp_queue_mutex);
1748 	list_for_each_entry(queue, &nvmet_tcp_queue_list, queue_list)
1749 		kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1750 	mutex_unlock(&nvmet_tcp_queue_mutex);
1751 	flush_scheduled_work();
1752 
1753 	destroy_workqueue(nvmet_tcp_wq);
1754 }
1755 
1756 module_init(nvmet_tcp_init);
1757 module_exit(nvmet_tcp_exit);
1758 
1759 MODULE_LICENSE("GPL v2");
1760 MODULE_ALIAS("nvmet-transport-3"); /* 3 == NVMF_TRTYPE_TCP */
1761