xref: /openbmc/linux/drivers/nvme/target/tcp.c (revision 8ebc80a25f9d9bf7a8e368b266d5b740c485c362)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * NVMe over Fabrics TCP target.
4  * Copyright (c) 2018 Lightbits Labs. All rights reserved.
5  */
6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
7 #include <linux/module.h>
8 #include <linux/init.h>
9 #include <linux/slab.h>
10 #include <linux/err.h>
11 #include <linux/nvme-tcp.h>
12 #include <net/sock.h>
13 #include <net/tcp.h>
14 #include <linux/inet.h>
15 #include <linux/llist.h>
16 #include <crypto/hash.h>
17 #include <trace/events/sock.h>
18 
19 #include "nvmet.h"
20 
21 #define NVMET_TCP_DEF_INLINE_DATA_SIZE	(4 * PAGE_SIZE)
22 #define NVMET_TCP_MAXH2CDATA		0x400000 /* 16M arbitrary limit */
23 
param_store_val(const char * str,int * val,int min,int max)24 static int param_store_val(const char *str, int *val, int min, int max)
25 {
26 	int ret, new_val;
27 
28 	ret = kstrtoint(str, 10, &new_val);
29 	if (ret)
30 		return -EINVAL;
31 
32 	if (new_val < min || new_val > max)
33 		return -EINVAL;
34 
35 	*val = new_val;
36 	return 0;
37 }
38 
set_params(const char * str,const struct kernel_param * kp)39 static int set_params(const char *str, const struct kernel_param *kp)
40 {
41 	return param_store_val(str, kp->arg, 0, INT_MAX);
42 }
43 
44 static const struct kernel_param_ops set_param_ops = {
45 	.set	= set_params,
46 	.get	= param_get_int,
47 };
48 
49 /* Define the socket priority to use for connections were it is desirable
50  * that the NIC consider performing optimized packet processing or filtering.
51  * A non-zero value being sufficient to indicate general consideration of any
52  * possible optimization.  Making it a module param allows for alternative
53  * values that may be unique for some NIC implementations.
54  */
55 static int so_priority;
56 device_param_cb(so_priority, &set_param_ops, &so_priority, 0644);
57 MODULE_PARM_DESC(so_priority, "nvmet tcp socket optimize priority: Default 0");
58 
59 /* Define a time period (in usecs) that io_work() shall sample an activated
60  * queue before determining it to be idle.  This optional module behavior
61  * can enable NIC solutions that support socket optimized packet processing
62  * using advanced interrupt moderation techniques.
63  */
64 static int idle_poll_period_usecs;
65 device_param_cb(idle_poll_period_usecs, &set_param_ops,
66 		&idle_poll_period_usecs, 0644);
67 MODULE_PARM_DESC(idle_poll_period_usecs,
68 		"nvmet tcp io_work poll till idle time period in usecs: Default 0");
69 
70 #define NVMET_TCP_RECV_BUDGET		8
71 #define NVMET_TCP_SEND_BUDGET		8
72 #define NVMET_TCP_IO_WORK_BUDGET	64
73 
74 enum nvmet_tcp_send_state {
75 	NVMET_TCP_SEND_DATA_PDU,
76 	NVMET_TCP_SEND_DATA,
77 	NVMET_TCP_SEND_R2T,
78 	NVMET_TCP_SEND_DDGST,
79 	NVMET_TCP_SEND_RESPONSE
80 };
81 
82 enum nvmet_tcp_recv_state {
83 	NVMET_TCP_RECV_PDU,
84 	NVMET_TCP_RECV_DATA,
85 	NVMET_TCP_RECV_DDGST,
86 	NVMET_TCP_RECV_ERR,
87 };
88 
89 enum {
90 	NVMET_TCP_F_INIT_FAILED = (1 << 0),
91 };
92 
93 struct nvmet_tcp_cmd {
94 	struct nvmet_tcp_queue		*queue;
95 	struct nvmet_req		req;
96 
97 	struct nvme_tcp_cmd_pdu		*cmd_pdu;
98 	struct nvme_tcp_rsp_pdu		*rsp_pdu;
99 	struct nvme_tcp_data_pdu	*data_pdu;
100 	struct nvme_tcp_r2t_pdu		*r2t_pdu;
101 
102 	u32				rbytes_done;
103 	u32				wbytes_done;
104 
105 	u32				pdu_len;
106 	u32				pdu_recv;
107 	int				sg_idx;
108 	struct msghdr			recv_msg;
109 	struct bio_vec			*iov;
110 	u32				flags;
111 
112 	struct list_head		entry;
113 	struct llist_node		lentry;
114 
115 	/* send state */
116 	u32				offset;
117 	struct scatterlist		*cur_sg;
118 	enum nvmet_tcp_send_state	state;
119 
120 	__le32				exp_ddgst;
121 	__le32				recv_ddgst;
122 };
123 
124 enum nvmet_tcp_queue_state {
125 	NVMET_TCP_Q_CONNECTING,
126 	NVMET_TCP_Q_LIVE,
127 	NVMET_TCP_Q_DISCONNECTING,
128 };
129 
130 struct nvmet_tcp_queue {
131 	struct socket		*sock;
132 	struct nvmet_tcp_port	*port;
133 	struct work_struct	io_work;
134 	struct nvmet_cq		nvme_cq;
135 	struct nvmet_sq		nvme_sq;
136 
137 	/* send state */
138 	struct nvmet_tcp_cmd	*cmds;
139 	unsigned int		nr_cmds;
140 	struct list_head	free_list;
141 	struct llist_head	resp_list;
142 	struct list_head	resp_send_list;
143 	int			send_list_len;
144 	struct nvmet_tcp_cmd	*snd_cmd;
145 
146 	/* recv state */
147 	int			offset;
148 	int			left;
149 	enum nvmet_tcp_recv_state rcv_state;
150 	struct nvmet_tcp_cmd	*cmd;
151 	union nvme_tcp_pdu	pdu;
152 
153 	/* digest state */
154 	bool			hdr_digest;
155 	bool			data_digest;
156 	struct ahash_request	*snd_hash;
157 	struct ahash_request	*rcv_hash;
158 
159 	unsigned long           poll_end;
160 
161 	spinlock_t		state_lock;
162 	enum nvmet_tcp_queue_state state;
163 
164 	struct sockaddr_storage	sockaddr;
165 	struct sockaddr_storage	sockaddr_peer;
166 	struct work_struct	release_work;
167 
168 	int			idx;
169 	struct list_head	queue_list;
170 
171 	struct nvmet_tcp_cmd	connect;
172 
173 	struct page_frag_cache	pf_cache;
174 
175 	void (*data_ready)(struct sock *);
176 	void (*state_change)(struct sock *);
177 	void (*write_space)(struct sock *);
178 };
179 
180 struct nvmet_tcp_port {
181 	struct socket		*sock;
182 	struct work_struct	accept_work;
183 	struct nvmet_port	*nport;
184 	struct sockaddr_storage addr;
185 	void (*data_ready)(struct sock *);
186 };
187 
188 static DEFINE_IDA(nvmet_tcp_queue_ida);
189 static LIST_HEAD(nvmet_tcp_queue_list);
190 static DEFINE_MUTEX(nvmet_tcp_queue_mutex);
191 
192 static struct workqueue_struct *nvmet_tcp_wq;
193 static const struct nvmet_fabrics_ops nvmet_tcp_ops;
194 static void nvmet_tcp_free_cmd(struct nvmet_tcp_cmd *c);
195 static void nvmet_tcp_free_cmd_buffers(struct nvmet_tcp_cmd *cmd);
196 
nvmet_tcp_cmd_tag(struct nvmet_tcp_queue * queue,struct nvmet_tcp_cmd * cmd)197 static inline u16 nvmet_tcp_cmd_tag(struct nvmet_tcp_queue *queue,
198 		struct nvmet_tcp_cmd *cmd)
199 {
200 	if (unlikely(!queue->nr_cmds)) {
201 		/* We didn't allocate cmds yet, send 0xffff */
202 		return USHRT_MAX;
203 	}
204 
205 	return cmd - queue->cmds;
206 }
207 
nvmet_tcp_has_data_in(struct nvmet_tcp_cmd * cmd)208 static inline bool nvmet_tcp_has_data_in(struct nvmet_tcp_cmd *cmd)
209 {
210 	return nvme_is_write(cmd->req.cmd) &&
211 		cmd->rbytes_done < cmd->req.transfer_len;
212 }
213 
nvmet_tcp_need_data_in(struct nvmet_tcp_cmd * cmd)214 static inline bool nvmet_tcp_need_data_in(struct nvmet_tcp_cmd *cmd)
215 {
216 	return nvmet_tcp_has_data_in(cmd) && !cmd->req.cqe->status;
217 }
218 
nvmet_tcp_need_data_out(struct nvmet_tcp_cmd * cmd)219 static inline bool nvmet_tcp_need_data_out(struct nvmet_tcp_cmd *cmd)
220 {
221 	return !nvme_is_write(cmd->req.cmd) &&
222 		cmd->req.transfer_len > 0 &&
223 		!cmd->req.cqe->status;
224 }
225 
nvmet_tcp_has_inline_data(struct nvmet_tcp_cmd * cmd)226 static inline bool nvmet_tcp_has_inline_data(struct nvmet_tcp_cmd *cmd)
227 {
228 	return nvme_is_write(cmd->req.cmd) && cmd->pdu_len &&
229 		!cmd->rbytes_done;
230 }
231 
232 static inline struct nvmet_tcp_cmd *
nvmet_tcp_get_cmd(struct nvmet_tcp_queue * queue)233 nvmet_tcp_get_cmd(struct nvmet_tcp_queue *queue)
234 {
235 	struct nvmet_tcp_cmd *cmd;
236 
237 	cmd = list_first_entry_or_null(&queue->free_list,
238 				struct nvmet_tcp_cmd, entry);
239 	if (!cmd)
240 		return NULL;
241 	list_del_init(&cmd->entry);
242 
243 	cmd->rbytes_done = cmd->wbytes_done = 0;
244 	cmd->pdu_len = 0;
245 	cmd->pdu_recv = 0;
246 	cmd->iov = NULL;
247 	cmd->flags = 0;
248 	return cmd;
249 }
250 
nvmet_tcp_put_cmd(struct nvmet_tcp_cmd * cmd)251 static inline void nvmet_tcp_put_cmd(struct nvmet_tcp_cmd *cmd)
252 {
253 	if (unlikely(cmd == &cmd->queue->connect))
254 		return;
255 
256 	list_add_tail(&cmd->entry, &cmd->queue->free_list);
257 }
258 
queue_cpu(struct nvmet_tcp_queue * queue)259 static inline int queue_cpu(struct nvmet_tcp_queue *queue)
260 {
261 	return queue->sock->sk->sk_incoming_cpu;
262 }
263 
nvmet_tcp_hdgst_len(struct nvmet_tcp_queue * queue)264 static inline u8 nvmet_tcp_hdgst_len(struct nvmet_tcp_queue *queue)
265 {
266 	return queue->hdr_digest ? NVME_TCP_DIGEST_LENGTH : 0;
267 }
268 
nvmet_tcp_ddgst_len(struct nvmet_tcp_queue * queue)269 static inline u8 nvmet_tcp_ddgst_len(struct nvmet_tcp_queue *queue)
270 {
271 	return queue->data_digest ? NVME_TCP_DIGEST_LENGTH : 0;
272 }
273 
nvmet_tcp_hdgst(struct ahash_request * hash,void * pdu,size_t len)274 static inline void nvmet_tcp_hdgst(struct ahash_request *hash,
275 		void *pdu, size_t len)
276 {
277 	struct scatterlist sg;
278 
279 	sg_init_one(&sg, pdu, len);
280 	ahash_request_set_crypt(hash, &sg, pdu + len, len);
281 	crypto_ahash_digest(hash);
282 }
283 
nvmet_tcp_verify_hdgst(struct nvmet_tcp_queue * queue,void * pdu,size_t len)284 static int nvmet_tcp_verify_hdgst(struct nvmet_tcp_queue *queue,
285 	void *pdu, size_t len)
286 {
287 	struct nvme_tcp_hdr *hdr = pdu;
288 	__le32 recv_digest;
289 	__le32 exp_digest;
290 
291 	if (unlikely(!(hdr->flags & NVME_TCP_F_HDGST))) {
292 		pr_err("queue %d: header digest enabled but no header digest\n",
293 			queue->idx);
294 		return -EPROTO;
295 	}
296 
297 	recv_digest = *(__le32 *)(pdu + hdr->hlen);
298 	nvmet_tcp_hdgst(queue->rcv_hash, pdu, len);
299 	exp_digest = *(__le32 *)(pdu + hdr->hlen);
300 	if (recv_digest != exp_digest) {
301 		pr_err("queue %d: header digest error: recv %#x expected %#x\n",
302 			queue->idx, le32_to_cpu(recv_digest),
303 			le32_to_cpu(exp_digest));
304 		return -EPROTO;
305 	}
306 
307 	return 0;
308 }
309 
nvmet_tcp_check_ddgst(struct nvmet_tcp_queue * queue,void * pdu)310 static int nvmet_tcp_check_ddgst(struct nvmet_tcp_queue *queue, void *pdu)
311 {
312 	struct nvme_tcp_hdr *hdr = pdu;
313 	u8 digest_len = nvmet_tcp_hdgst_len(queue);
314 	u32 len;
315 
316 	len = le32_to_cpu(hdr->plen) - hdr->hlen -
317 		(hdr->flags & NVME_TCP_F_HDGST ? digest_len : 0);
318 
319 	if (unlikely(len && !(hdr->flags & NVME_TCP_F_DDGST))) {
320 		pr_err("queue %d: data digest flag is cleared\n", queue->idx);
321 		return -EPROTO;
322 	}
323 
324 	return 0;
325 }
326 
327 /* If cmd buffers are NULL, no operation is performed */
nvmet_tcp_free_cmd_buffers(struct nvmet_tcp_cmd * cmd)328 static void nvmet_tcp_free_cmd_buffers(struct nvmet_tcp_cmd *cmd)
329 {
330 	kfree(cmd->iov);
331 	sgl_free(cmd->req.sg);
332 	cmd->iov = NULL;
333 	cmd->req.sg = NULL;
334 }
335 
nvmet_tcp_build_pdu_iovec(struct nvmet_tcp_cmd * cmd)336 static void nvmet_tcp_build_pdu_iovec(struct nvmet_tcp_cmd *cmd)
337 {
338 	struct bio_vec *iov = cmd->iov;
339 	struct scatterlist *sg;
340 	u32 length, offset, sg_offset;
341 	int nr_pages;
342 
343 	length = cmd->pdu_len;
344 	nr_pages = DIV_ROUND_UP(length, PAGE_SIZE);
345 	offset = cmd->rbytes_done;
346 	cmd->sg_idx = offset / PAGE_SIZE;
347 	sg_offset = offset % PAGE_SIZE;
348 	sg = &cmd->req.sg[cmd->sg_idx];
349 
350 	while (length) {
351 		u32 iov_len = min_t(u32, length, sg->length - sg_offset);
352 
353 		bvec_set_page(iov, sg_page(sg), iov_len,
354 				sg->offset + sg_offset);
355 
356 		length -= iov_len;
357 		sg = sg_next(sg);
358 		iov++;
359 		sg_offset = 0;
360 	}
361 
362 	iov_iter_bvec(&cmd->recv_msg.msg_iter, ITER_DEST, cmd->iov,
363 		      nr_pages, cmd->pdu_len);
364 }
365 
nvmet_tcp_fatal_error(struct nvmet_tcp_queue * queue)366 static void nvmet_tcp_fatal_error(struct nvmet_tcp_queue *queue)
367 {
368 	queue->rcv_state = NVMET_TCP_RECV_ERR;
369 	if (queue->nvme_sq.ctrl)
370 		nvmet_ctrl_fatal_error(queue->nvme_sq.ctrl);
371 	else
372 		kernel_sock_shutdown(queue->sock, SHUT_RDWR);
373 }
374 
nvmet_tcp_socket_error(struct nvmet_tcp_queue * queue,int status)375 static void nvmet_tcp_socket_error(struct nvmet_tcp_queue *queue, int status)
376 {
377 	queue->rcv_state = NVMET_TCP_RECV_ERR;
378 	if (status == -EPIPE || status == -ECONNRESET)
379 		kernel_sock_shutdown(queue->sock, SHUT_RDWR);
380 	else
381 		nvmet_tcp_fatal_error(queue);
382 }
383 
nvmet_tcp_map_data(struct nvmet_tcp_cmd * cmd)384 static int nvmet_tcp_map_data(struct nvmet_tcp_cmd *cmd)
385 {
386 	struct nvme_sgl_desc *sgl = &cmd->req.cmd->common.dptr.sgl;
387 	u32 len = le32_to_cpu(sgl->length);
388 
389 	if (!len)
390 		return 0;
391 
392 	if (sgl->type == ((NVME_SGL_FMT_DATA_DESC << 4) |
393 			  NVME_SGL_FMT_OFFSET)) {
394 		if (!nvme_is_write(cmd->req.cmd))
395 			return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
396 
397 		if (len > cmd->req.port->inline_data_size)
398 			return NVME_SC_SGL_INVALID_OFFSET | NVME_SC_DNR;
399 		cmd->pdu_len = len;
400 	}
401 	cmd->req.transfer_len += len;
402 
403 	cmd->req.sg = sgl_alloc(len, GFP_KERNEL, &cmd->req.sg_cnt);
404 	if (!cmd->req.sg)
405 		return NVME_SC_INTERNAL;
406 	cmd->cur_sg = cmd->req.sg;
407 
408 	if (nvmet_tcp_has_data_in(cmd)) {
409 		cmd->iov = kmalloc_array(cmd->req.sg_cnt,
410 				sizeof(*cmd->iov), GFP_KERNEL);
411 		if (!cmd->iov)
412 			goto err;
413 	}
414 
415 	return 0;
416 err:
417 	nvmet_tcp_free_cmd_buffers(cmd);
418 	return NVME_SC_INTERNAL;
419 }
420 
nvmet_tcp_calc_ddgst(struct ahash_request * hash,struct nvmet_tcp_cmd * cmd)421 static void nvmet_tcp_calc_ddgst(struct ahash_request *hash,
422 		struct nvmet_tcp_cmd *cmd)
423 {
424 	ahash_request_set_crypt(hash, cmd->req.sg,
425 		(void *)&cmd->exp_ddgst, cmd->req.transfer_len);
426 	crypto_ahash_digest(hash);
427 }
428 
nvmet_setup_c2h_data_pdu(struct nvmet_tcp_cmd * cmd)429 static void nvmet_setup_c2h_data_pdu(struct nvmet_tcp_cmd *cmd)
430 {
431 	struct nvme_tcp_data_pdu *pdu = cmd->data_pdu;
432 	struct nvmet_tcp_queue *queue = cmd->queue;
433 	u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
434 	u8 ddgst = nvmet_tcp_ddgst_len(cmd->queue);
435 
436 	cmd->offset = 0;
437 	cmd->state = NVMET_TCP_SEND_DATA_PDU;
438 
439 	pdu->hdr.type = nvme_tcp_c2h_data;
440 	pdu->hdr.flags = NVME_TCP_F_DATA_LAST | (queue->nvme_sq.sqhd_disabled ?
441 						NVME_TCP_F_DATA_SUCCESS : 0);
442 	pdu->hdr.hlen = sizeof(*pdu);
443 	pdu->hdr.pdo = pdu->hdr.hlen + hdgst;
444 	pdu->hdr.plen =
445 		cpu_to_le32(pdu->hdr.hlen + hdgst +
446 				cmd->req.transfer_len + ddgst);
447 	pdu->command_id = cmd->req.cqe->command_id;
448 	pdu->data_length = cpu_to_le32(cmd->req.transfer_len);
449 	pdu->data_offset = cpu_to_le32(cmd->wbytes_done);
450 
451 	if (queue->data_digest) {
452 		pdu->hdr.flags |= NVME_TCP_F_DDGST;
453 		nvmet_tcp_calc_ddgst(queue->snd_hash, cmd);
454 	}
455 
456 	if (cmd->queue->hdr_digest) {
457 		pdu->hdr.flags |= NVME_TCP_F_HDGST;
458 		nvmet_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
459 	}
460 }
461 
nvmet_setup_r2t_pdu(struct nvmet_tcp_cmd * cmd)462 static void nvmet_setup_r2t_pdu(struct nvmet_tcp_cmd *cmd)
463 {
464 	struct nvme_tcp_r2t_pdu *pdu = cmd->r2t_pdu;
465 	struct nvmet_tcp_queue *queue = cmd->queue;
466 	u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
467 
468 	cmd->offset = 0;
469 	cmd->state = NVMET_TCP_SEND_R2T;
470 
471 	pdu->hdr.type = nvme_tcp_r2t;
472 	pdu->hdr.flags = 0;
473 	pdu->hdr.hlen = sizeof(*pdu);
474 	pdu->hdr.pdo = 0;
475 	pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst);
476 
477 	pdu->command_id = cmd->req.cmd->common.command_id;
478 	pdu->ttag = nvmet_tcp_cmd_tag(cmd->queue, cmd);
479 	pdu->r2t_length = cpu_to_le32(cmd->req.transfer_len - cmd->rbytes_done);
480 	pdu->r2t_offset = cpu_to_le32(cmd->rbytes_done);
481 	if (cmd->queue->hdr_digest) {
482 		pdu->hdr.flags |= NVME_TCP_F_HDGST;
483 		nvmet_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
484 	}
485 }
486 
nvmet_setup_response_pdu(struct nvmet_tcp_cmd * cmd)487 static void nvmet_setup_response_pdu(struct nvmet_tcp_cmd *cmd)
488 {
489 	struct nvme_tcp_rsp_pdu *pdu = cmd->rsp_pdu;
490 	struct nvmet_tcp_queue *queue = cmd->queue;
491 	u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
492 
493 	cmd->offset = 0;
494 	cmd->state = NVMET_TCP_SEND_RESPONSE;
495 
496 	pdu->hdr.type = nvme_tcp_rsp;
497 	pdu->hdr.flags = 0;
498 	pdu->hdr.hlen = sizeof(*pdu);
499 	pdu->hdr.pdo = 0;
500 	pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst);
501 	if (cmd->queue->hdr_digest) {
502 		pdu->hdr.flags |= NVME_TCP_F_HDGST;
503 		nvmet_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
504 	}
505 }
506 
nvmet_tcp_process_resp_list(struct nvmet_tcp_queue * queue)507 static void nvmet_tcp_process_resp_list(struct nvmet_tcp_queue *queue)
508 {
509 	struct llist_node *node;
510 	struct nvmet_tcp_cmd *cmd;
511 
512 	for (node = llist_del_all(&queue->resp_list); node; node = node->next) {
513 		cmd = llist_entry(node, struct nvmet_tcp_cmd, lentry);
514 		list_add(&cmd->entry, &queue->resp_send_list);
515 		queue->send_list_len++;
516 	}
517 }
518 
nvmet_tcp_fetch_cmd(struct nvmet_tcp_queue * queue)519 static struct nvmet_tcp_cmd *nvmet_tcp_fetch_cmd(struct nvmet_tcp_queue *queue)
520 {
521 	queue->snd_cmd = list_first_entry_or_null(&queue->resp_send_list,
522 				struct nvmet_tcp_cmd, entry);
523 	if (!queue->snd_cmd) {
524 		nvmet_tcp_process_resp_list(queue);
525 		queue->snd_cmd =
526 			list_first_entry_or_null(&queue->resp_send_list,
527 					struct nvmet_tcp_cmd, entry);
528 		if (unlikely(!queue->snd_cmd))
529 			return NULL;
530 	}
531 
532 	list_del_init(&queue->snd_cmd->entry);
533 	queue->send_list_len--;
534 
535 	if (nvmet_tcp_need_data_out(queue->snd_cmd))
536 		nvmet_setup_c2h_data_pdu(queue->snd_cmd);
537 	else if (nvmet_tcp_need_data_in(queue->snd_cmd))
538 		nvmet_setup_r2t_pdu(queue->snd_cmd);
539 	else
540 		nvmet_setup_response_pdu(queue->snd_cmd);
541 
542 	return queue->snd_cmd;
543 }
544 
nvmet_tcp_queue_response(struct nvmet_req * req)545 static void nvmet_tcp_queue_response(struct nvmet_req *req)
546 {
547 	struct nvmet_tcp_cmd *cmd =
548 		container_of(req, struct nvmet_tcp_cmd, req);
549 	struct nvmet_tcp_queue	*queue = cmd->queue;
550 	enum nvmet_tcp_recv_state queue_state;
551 	struct nvmet_tcp_cmd *queue_cmd;
552 	struct nvme_sgl_desc *sgl;
553 	u32 len;
554 
555 	/* Pairs with store_release in nvmet_prepare_receive_pdu() */
556 	queue_state = smp_load_acquire(&queue->rcv_state);
557 	queue_cmd = READ_ONCE(queue->cmd);
558 
559 	if (unlikely(cmd == queue_cmd)) {
560 		sgl = &cmd->req.cmd->common.dptr.sgl;
561 		len = le32_to_cpu(sgl->length);
562 
563 		/*
564 		 * Wait for inline data before processing the response.
565 		 * Avoid using helpers, this might happen before
566 		 * nvmet_req_init is completed.
567 		 */
568 		if (queue_state == NVMET_TCP_RECV_PDU &&
569 		    len && len <= cmd->req.port->inline_data_size &&
570 		    nvme_is_write(cmd->req.cmd))
571 			return;
572 	}
573 
574 	llist_add(&cmd->lentry, &queue->resp_list);
575 	queue_work_on(queue_cpu(queue), nvmet_tcp_wq, &cmd->queue->io_work);
576 }
577 
nvmet_tcp_execute_request(struct nvmet_tcp_cmd * cmd)578 static void nvmet_tcp_execute_request(struct nvmet_tcp_cmd *cmd)
579 {
580 	if (unlikely(cmd->flags & NVMET_TCP_F_INIT_FAILED))
581 		nvmet_tcp_queue_response(&cmd->req);
582 	else
583 		cmd->req.execute(&cmd->req);
584 }
585 
nvmet_try_send_data_pdu(struct nvmet_tcp_cmd * cmd)586 static int nvmet_try_send_data_pdu(struct nvmet_tcp_cmd *cmd)
587 {
588 	struct msghdr msg = {
589 		.msg_flags = MSG_DONTWAIT | MSG_MORE | MSG_SPLICE_PAGES,
590 	};
591 	struct bio_vec bvec;
592 	u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
593 	int left = sizeof(*cmd->data_pdu) - cmd->offset + hdgst;
594 	int ret;
595 
596 	bvec_set_virt(&bvec, (void *)cmd->data_pdu + cmd->offset, left);
597 	iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, &bvec, 1, left);
598 	ret = sock_sendmsg(cmd->queue->sock, &msg);
599 	if (ret <= 0)
600 		return ret;
601 
602 	cmd->offset += ret;
603 	left -= ret;
604 
605 	if (left)
606 		return -EAGAIN;
607 
608 	cmd->state = NVMET_TCP_SEND_DATA;
609 	cmd->offset  = 0;
610 	return 1;
611 }
612 
nvmet_try_send_data(struct nvmet_tcp_cmd * cmd,bool last_in_batch)613 static int nvmet_try_send_data(struct nvmet_tcp_cmd *cmd, bool last_in_batch)
614 {
615 	struct nvmet_tcp_queue *queue = cmd->queue;
616 	int ret;
617 
618 	while (cmd->cur_sg) {
619 		struct msghdr msg = {
620 			.msg_flags = MSG_DONTWAIT | MSG_SPLICE_PAGES,
621 		};
622 		struct page *page = sg_page(cmd->cur_sg);
623 		struct bio_vec bvec;
624 		u32 left = cmd->cur_sg->length - cmd->offset;
625 
626 		if ((!last_in_batch && cmd->queue->send_list_len) ||
627 		    cmd->wbytes_done + left < cmd->req.transfer_len ||
628 		    queue->data_digest || !queue->nvme_sq.sqhd_disabled)
629 			msg.msg_flags |= MSG_MORE;
630 
631 		bvec_set_page(&bvec, page, left, cmd->offset);
632 		iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, &bvec, 1, left);
633 		ret = sock_sendmsg(cmd->queue->sock, &msg);
634 		if (ret <= 0)
635 			return ret;
636 
637 		cmd->offset += ret;
638 		cmd->wbytes_done += ret;
639 
640 		/* Done with sg?*/
641 		if (cmd->offset == cmd->cur_sg->length) {
642 			cmd->cur_sg = sg_next(cmd->cur_sg);
643 			cmd->offset = 0;
644 		}
645 	}
646 
647 	if (queue->data_digest) {
648 		cmd->state = NVMET_TCP_SEND_DDGST;
649 		cmd->offset = 0;
650 	} else {
651 		if (queue->nvme_sq.sqhd_disabled) {
652 			cmd->queue->snd_cmd = NULL;
653 			nvmet_tcp_put_cmd(cmd);
654 		} else {
655 			nvmet_setup_response_pdu(cmd);
656 		}
657 	}
658 
659 	if (queue->nvme_sq.sqhd_disabled)
660 		nvmet_tcp_free_cmd_buffers(cmd);
661 
662 	return 1;
663 
664 }
665 
nvmet_try_send_response(struct nvmet_tcp_cmd * cmd,bool last_in_batch)666 static int nvmet_try_send_response(struct nvmet_tcp_cmd *cmd,
667 		bool last_in_batch)
668 {
669 	struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_SPLICE_PAGES, };
670 	struct bio_vec bvec;
671 	u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
672 	int left = sizeof(*cmd->rsp_pdu) - cmd->offset + hdgst;
673 	int ret;
674 
675 	if (!last_in_batch && cmd->queue->send_list_len)
676 		msg.msg_flags |= MSG_MORE;
677 	else
678 		msg.msg_flags |= MSG_EOR;
679 
680 	bvec_set_virt(&bvec, (void *)cmd->rsp_pdu + cmd->offset, left);
681 	iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, &bvec, 1, left);
682 	ret = sock_sendmsg(cmd->queue->sock, &msg);
683 	if (ret <= 0)
684 		return ret;
685 	cmd->offset += ret;
686 	left -= ret;
687 
688 	if (left)
689 		return -EAGAIN;
690 
691 	nvmet_tcp_free_cmd_buffers(cmd);
692 	cmd->queue->snd_cmd = NULL;
693 	nvmet_tcp_put_cmd(cmd);
694 	return 1;
695 }
696 
nvmet_try_send_r2t(struct nvmet_tcp_cmd * cmd,bool last_in_batch)697 static int nvmet_try_send_r2t(struct nvmet_tcp_cmd *cmd, bool last_in_batch)
698 {
699 	struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_SPLICE_PAGES, };
700 	struct bio_vec bvec;
701 	u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
702 	int left = sizeof(*cmd->r2t_pdu) - cmd->offset + hdgst;
703 	int ret;
704 
705 	if (!last_in_batch && cmd->queue->send_list_len)
706 		msg.msg_flags |= MSG_MORE;
707 	else
708 		msg.msg_flags |= MSG_EOR;
709 
710 	bvec_set_virt(&bvec, (void *)cmd->r2t_pdu + cmd->offset, left);
711 	iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, &bvec, 1, left);
712 	ret = sock_sendmsg(cmd->queue->sock, &msg);
713 	if (ret <= 0)
714 		return ret;
715 	cmd->offset += ret;
716 	left -= ret;
717 
718 	if (left)
719 		return -EAGAIN;
720 
721 	cmd->queue->snd_cmd = NULL;
722 	return 1;
723 }
724 
nvmet_try_send_ddgst(struct nvmet_tcp_cmd * cmd,bool last_in_batch)725 static int nvmet_try_send_ddgst(struct nvmet_tcp_cmd *cmd, bool last_in_batch)
726 {
727 	struct nvmet_tcp_queue *queue = cmd->queue;
728 	int left = NVME_TCP_DIGEST_LENGTH - cmd->offset;
729 	struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
730 	struct kvec iov = {
731 		.iov_base = (u8 *)&cmd->exp_ddgst + cmd->offset,
732 		.iov_len = left
733 	};
734 	int ret;
735 
736 	if (!last_in_batch && cmd->queue->send_list_len)
737 		msg.msg_flags |= MSG_MORE;
738 	else
739 		msg.msg_flags |= MSG_EOR;
740 
741 	ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
742 	if (unlikely(ret <= 0))
743 		return ret;
744 
745 	cmd->offset += ret;
746 	left -= ret;
747 
748 	if (left)
749 		return -EAGAIN;
750 
751 	if (queue->nvme_sq.sqhd_disabled) {
752 		cmd->queue->snd_cmd = NULL;
753 		nvmet_tcp_put_cmd(cmd);
754 	} else {
755 		nvmet_setup_response_pdu(cmd);
756 	}
757 	return 1;
758 }
759 
nvmet_tcp_try_send_one(struct nvmet_tcp_queue * queue,bool last_in_batch)760 static int nvmet_tcp_try_send_one(struct nvmet_tcp_queue *queue,
761 		bool last_in_batch)
762 {
763 	struct nvmet_tcp_cmd *cmd = queue->snd_cmd;
764 	int ret = 0;
765 
766 	if (!cmd || queue->state == NVMET_TCP_Q_DISCONNECTING) {
767 		cmd = nvmet_tcp_fetch_cmd(queue);
768 		if (unlikely(!cmd))
769 			return 0;
770 	}
771 
772 	if (cmd->state == NVMET_TCP_SEND_DATA_PDU) {
773 		ret = nvmet_try_send_data_pdu(cmd);
774 		if (ret <= 0)
775 			goto done_send;
776 	}
777 
778 	if (cmd->state == NVMET_TCP_SEND_DATA) {
779 		ret = nvmet_try_send_data(cmd, last_in_batch);
780 		if (ret <= 0)
781 			goto done_send;
782 	}
783 
784 	if (cmd->state == NVMET_TCP_SEND_DDGST) {
785 		ret = nvmet_try_send_ddgst(cmd, last_in_batch);
786 		if (ret <= 0)
787 			goto done_send;
788 	}
789 
790 	if (cmd->state == NVMET_TCP_SEND_R2T) {
791 		ret = nvmet_try_send_r2t(cmd, last_in_batch);
792 		if (ret <= 0)
793 			goto done_send;
794 	}
795 
796 	if (cmd->state == NVMET_TCP_SEND_RESPONSE)
797 		ret = nvmet_try_send_response(cmd, last_in_batch);
798 
799 done_send:
800 	if (ret < 0) {
801 		if (ret == -EAGAIN)
802 			return 0;
803 		return ret;
804 	}
805 
806 	return 1;
807 }
808 
nvmet_tcp_try_send(struct nvmet_tcp_queue * queue,int budget,int * sends)809 static int nvmet_tcp_try_send(struct nvmet_tcp_queue *queue,
810 		int budget, int *sends)
811 {
812 	int i, ret = 0;
813 
814 	for (i = 0; i < budget; i++) {
815 		ret = nvmet_tcp_try_send_one(queue, i == budget - 1);
816 		if (unlikely(ret < 0)) {
817 			nvmet_tcp_socket_error(queue, ret);
818 			goto done;
819 		} else if (ret == 0) {
820 			break;
821 		}
822 		(*sends)++;
823 	}
824 done:
825 	return ret;
826 }
827 
nvmet_prepare_receive_pdu(struct nvmet_tcp_queue * queue)828 static void nvmet_prepare_receive_pdu(struct nvmet_tcp_queue *queue)
829 {
830 	queue->offset = 0;
831 	queue->left = sizeof(struct nvme_tcp_hdr);
832 	WRITE_ONCE(queue->cmd, NULL);
833 	/* Ensure rcv_state is visible only after queue->cmd is set */
834 	smp_store_release(&queue->rcv_state, NVMET_TCP_RECV_PDU);
835 }
836 
nvmet_tcp_free_crypto(struct nvmet_tcp_queue * queue)837 static void nvmet_tcp_free_crypto(struct nvmet_tcp_queue *queue)
838 {
839 	struct crypto_ahash *tfm = crypto_ahash_reqtfm(queue->rcv_hash);
840 
841 	ahash_request_free(queue->rcv_hash);
842 	ahash_request_free(queue->snd_hash);
843 	crypto_free_ahash(tfm);
844 }
845 
nvmet_tcp_alloc_crypto(struct nvmet_tcp_queue * queue)846 static int nvmet_tcp_alloc_crypto(struct nvmet_tcp_queue *queue)
847 {
848 	struct crypto_ahash *tfm;
849 
850 	tfm = crypto_alloc_ahash("crc32c", 0, CRYPTO_ALG_ASYNC);
851 	if (IS_ERR(tfm))
852 		return PTR_ERR(tfm);
853 
854 	queue->snd_hash = ahash_request_alloc(tfm, GFP_KERNEL);
855 	if (!queue->snd_hash)
856 		goto free_tfm;
857 	ahash_request_set_callback(queue->snd_hash, 0, NULL, NULL);
858 
859 	queue->rcv_hash = ahash_request_alloc(tfm, GFP_KERNEL);
860 	if (!queue->rcv_hash)
861 		goto free_snd_hash;
862 	ahash_request_set_callback(queue->rcv_hash, 0, NULL, NULL);
863 
864 	return 0;
865 free_snd_hash:
866 	ahash_request_free(queue->snd_hash);
867 free_tfm:
868 	crypto_free_ahash(tfm);
869 	return -ENOMEM;
870 }
871 
872 
nvmet_tcp_handle_icreq(struct nvmet_tcp_queue * queue)873 static int nvmet_tcp_handle_icreq(struct nvmet_tcp_queue *queue)
874 {
875 	struct nvme_tcp_icreq_pdu *icreq = &queue->pdu.icreq;
876 	struct nvme_tcp_icresp_pdu *icresp = &queue->pdu.icresp;
877 	struct msghdr msg = {};
878 	struct kvec iov;
879 	int ret;
880 
881 	if (le32_to_cpu(icreq->hdr.plen) != sizeof(struct nvme_tcp_icreq_pdu)) {
882 		pr_err("bad nvme-tcp pdu length (%d)\n",
883 			le32_to_cpu(icreq->hdr.plen));
884 		nvmet_tcp_fatal_error(queue);
885 		return -EPROTO;
886 	}
887 
888 	if (icreq->pfv != NVME_TCP_PFV_1_0) {
889 		pr_err("queue %d: bad pfv %d\n", queue->idx, icreq->pfv);
890 		return -EPROTO;
891 	}
892 
893 	if (icreq->hpda != 0) {
894 		pr_err("queue %d: unsupported hpda %d\n", queue->idx,
895 			icreq->hpda);
896 		return -EPROTO;
897 	}
898 
899 	queue->hdr_digest = !!(icreq->digest & NVME_TCP_HDR_DIGEST_ENABLE);
900 	queue->data_digest = !!(icreq->digest & NVME_TCP_DATA_DIGEST_ENABLE);
901 	if (queue->hdr_digest || queue->data_digest) {
902 		ret = nvmet_tcp_alloc_crypto(queue);
903 		if (ret)
904 			return ret;
905 	}
906 
907 	memset(icresp, 0, sizeof(*icresp));
908 	icresp->hdr.type = nvme_tcp_icresp;
909 	icresp->hdr.hlen = sizeof(*icresp);
910 	icresp->hdr.pdo = 0;
911 	icresp->hdr.plen = cpu_to_le32(icresp->hdr.hlen);
912 	icresp->pfv = cpu_to_le16(NVME_TCP_PFV_1_0);
913 	icresp->maxdata = cpu_to_le32(NVMET_TCP_MAXH2CDATA);
914 	icresp->cpda = 0;
915 	if (queue->hdr_digest)
916 		icresp->digest |= NVME_TCP_HDR_DIGEST_ENABLE;
917 	if (queue->data_digest)
918 		icresp->digest |= NVME_TCP_DATA_DIGEST_ENABLE;
919 
920 	iov.iov_base = icresp;
921 	iov.iov_len = sizeof(*icresp);
922 	ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
923 	if (ret < 0)
924 		return ret; /* queue removal will cleanup */
925 
926 	queue->state = NVMET_TCP_Q_LIVE;
927 	nvmet_prepare_receive_pdu(queue);
928 	return 0;
929 }
930 
nvmet_tcp_handle_req_failure(struct nvmet_tcp_queue * queue,struct nvmet_tcp_cmd * cmd,struct nvmet_req * req)931 static void nvmet_tcp_handle_req_failure(struct nvmet_tcp_queue *queue,
932 		struct nvmet_tcp_cmd *cmd, struct nvmet_req *req)
933 {
934 	size_t data_len = le32_to_cpu(req->cmd->common.dptr.sgl.length);
935 	int ret;
936 
937 	/*
938 	 * This command has not been processed yet, hence we are trying to
939 	 * figure out if there is still pending data left to receive. If
940 	 * we don't, we can simply prepare for the next pdu and bail out,
941 	 * otherwise we will need to prepare a buffer and receive the
942 	 * stale data before continuing forward.
943 	 */
944 	if (!nvme_is_write(cmd->req.cmd) || !data_len ||
945 	    data_len > cmd->req.port->inline_data_size) {
946 		nvmet_prepare_receive_pdu(queue);
947 		return;
948 	}
949 
950 	ret = nvmet_tcp_map_data(cmd);
951 	if (unlikely(ret)) {
952 		pr_err("queue %d: failed to map data\n", queue->idx);
953 		nvmet_tcp_fatal_error(queue);
954 		return;
955 	}
956 
957 	queue->rcv_state = NVMET_TCP_RECV_DATA;
958 	nvmet_tcp_build_pdu_iovec(cmd);
959 	cmd->flags |= NVMET_TCP_F_INIT_FAILED;
960 }
961 
nvmet_tcp_handle_h2c_data_pdu(struct nvmet_tcp_queue * queue)962 static int nvmet_tcp_handle_h2c_data_pdu(struct nvmet_tcp_queue *queue)
963 {
964 	struct nvme_tcp_data_pdu *data = &queue->pdu.data;
965 	struct nvmet_tcp_cmd *cmd;
966 	unsigned int exp_data_len;
967 
968 	if (likely(queue->nr_cmds)) {
969 		if (unlikely(data->ttag >= queue->nr_cmds)) {
970 			pr_err("queue %d: received out of bound ttag %u, nr_cmds %u\n",
971 				queue->idx, data->ttag, queue->nr_cmds);
972 			nvmet_tcp_fatal_error(queue);
973 			return -EPROTO;
974 		}
975 		cmd = &queue->cmds[data->ttag];
976 	} else {
977 		cmd = &queue->connect;
978 	}
979 
980 	if (le32_to_cpu(data->data_offset) != cmd->rbytes_done) {
981 		pr_err("ttag %u unexpected data offset %u (expected %u)\n",
982 			data->ttag, le32_to_cpu(data->data_offset),
983 			cmd->rbytes_done);
984 		/* FIXME: use path and transport errors */
985 		nvmet_tcp_fatal_error(queue);
986 		return -EPROTO;
987 	}
988 
989 	exp_data_len = le32_to_cpu(data->hdr.plen) -
990 			nvmet_tcp_hdgst_len(queue) -
991 			nvmet_tcp_ddgst_len(queue) -
992 			sizeof(*data);
993 
994 	cmd->pdu_len = le32_to_cpu(data->data_length);
995 	if (unlikely(cmd->pdu_len != exp_data_len ||
996 		     cmd->pdu_len == 0 ||
997 		     cmd->pdu_len > NVMET_TCP_MAXH2CDATA)) {
998 		pr_err("H2CData PDU len %u is invalid\n", cmd->pdu_len);
999 		/* FIXME: use proper transport errors */
1000 		nvmet_tcp_fatal_error(queue);
1001 		return -EPROTO;
1002 	}
1003 	cmd->pdu_recv = 0;
1004 	nvmet_tcp_build_pdu_iovec(cmd);
1005 	queue->cmd = cmd;
1006 	queue->rcv_state = NVMET_TCP_RECV_DATA;
1007 
1008 	return 0;
1009 }
1010 
nvmet_tcp_done_recv_pdu(struct nvmet_tcp_queue * queue)1011 static int nvmet_tcp_done_recv_pdu(struct nvmet_tcp_queue *queue)
1012 {
1013 	struct nvme_tcp_hdr *hdr = &queue->pdu.cmd.hdr;
1014 	struct nvme_command *nvme_cmd = &queue->pdu.cmd.cmd;
1015 	struct nvmet_req *req;
1016 	int ret;
1017 
1018 	if (unlikely(queue->state == NVMET_TCP_Q_CONNECTING)) {
1019 		if (hdr->type != nvme_tcp_icreq) {
1020 			pr_err("unexpected pdu type (%d) before icreq\n",
1021 				hdr->type);
1022 			nvmet_tcp_fatal_error(queue);
1023 			return -EPROTO;
1024 		}
1025 		return nvmet_tcp_handle_icreq(queue);
1026 	}
1027 
1028 	if (unlikely(hdr->type == nvme_tcp_icreq)) {
1029 		pr_err("queue %d: received icreq pdu in state %d\n",
1030 			queue->idx, queue->state);
1031 		nvmet_tcp_fatal_error(queue);
1032 		return -EPROTO;
1033 	}
1034 
1035 	if (hdr->type == nvme_tcp_h2c_data) {
1036 		ret = nvmet_tcp_handle_h2c_data_pdu(queue);
1037 		if (unlikely(ret))
1038 			return ret;
1039 		return 0;
1040 	}
1041 
1042 	queue->cmd = nvmet_tcp_get_cmd(queue);
1043 	if (unlikely(!queue->cmd)) {
1044 		/* This should never happen */
1045 		pr_err("queue %d: out of commands (%d) send_list_len: %d, opcode: %d",
1046 			queue->idx, queue->nr_cmds, queue->send_list_len,
1047 			nvme_cmd->common.opcode);
1048 		nvmet_tcp_fatal_error(queue);
1049 		return -ENOMEM;
1050 	}
1051 
1052 	req = &queue->cmd->req;
1053 	memcpy(req->cmd, nvme_cmd, sizeof(*nvme_cmd));
1054 
1055 	if (unlikely(!nvmet_req_init(req, &queue->nvme_cq,
1056 			&queue->nvme_sq, &nvmet_tcp_ops))) {
1057 		pr_err("failed cmd %p id %d opcode %d, data_len: %d\n",
1058 			req->cmd, req->cmd->common.command_id,
1059 			req->cmd->common.opcode,
1060 			le32_to_cpu(req->cmd->common.dptr.sgl.length));
1061 
1062 		nvmet_tcp_handle_req_failure(queue, queue->cmd, req);
1063 		return 0;
1064 	}
1065 
1066 	ret = nvmet_tcp_map_data(queue->cmd);
1067 	if (unlikely(ret)) {
1068 		pr_err("queue %d: failed to map data\n", queue->idx);
1069 		if (nvmet_tcp_has_inline_data(queue->cmd))
1070 			nvmet_tcp_fatal_error(queue);
1071 		else
1072 			nvmet_req_complete(req, ret);
1073 		ret = -EAGAIN;
1074 		goto out;
1075 	}
1076 
1077 	if (nvmet_tcp_need_data_in(queue->cmd)) {
1078 		if (nvmet_tcp_has_inline_data(queue->cmd)) {
1079 			queue->rcv_state = NVMET_TCP_RECV_DATA;
1080 			nvmet_tcp_build_pdu_iovec(queue->cmd);
1081 			return 0;
1082 		}
1083 		/* send back R2T */
1084 		nvmet_tcp_queue_response(&queue->cmd->req);
1085 		goto out;
1086 	}
1087 
1088 	queue->cmd->req.execute(&queue->cmd->req);
1089 out:
1090 	nvmet_prepare_receive_pdu(queue);
1091 	return ret;
1092 }
1093 
1094 static const u8 nvme_tcp_pdu_sizes[] = {
1095 	[nvme_tcp_icreq]	= sizeof(struct nvme_tcp_icreq_pdu),
1096 	[nvme_tcp_cmd]		= sizeof(struct nvme_tcp_cmd_pdu),
1097 	[nvme_tcp_h2c_data]	= sizeof(struct nvme_tcp_data_pdu),
1098 };
1099 
nvmet_tcp_pdu_size(u8 type)1100 static inline u8 nvmet_tcp_pdu_size(u8 type)
1101 {
1102 	size_t idx = type;
1103 
1104 	return (idx < ARRAY_SIZE(nvme_tcp_pdu_sizes) &&
1105 		nvme_tcp_pdu_sizes[idx]) ?
1106 			nvme_tcp_pdu_sizes[idx] : 0;
1107 }
1108 
nvmet_tcp_pdu_valid(u8 type)1109 static inline bool nvmet_tcp_pdu_valid(u8 type)
1110 {
1111 	switch (type) {
1112 	case nvme_tcp_icreq:
1113 	case nvme_tcp_cmd:
1114 	case nvme_tcp_h2c_data:
1115 		/* fallthru */
1116 		return true;
1117 	}
1118 
1119 	return false;
1120 }
1121 
nvmet_tcp_try_recv_pdu(struct nvmet_tcp_queue * queue)1122 static int nvmet_tcp_try_recv_pdu(struct nvmet_tcp_queue *queue)
1123 {
1124 	struct nvme_tcp_hdr *hdr = &queue->pdu.cmd.hdr;
1125 	int len;
1126 	struct kvec iov;
1127 	struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
1128 
1129 recv:
1130 	iov.iov_base = (void *)&queue->pdu + queue->offset;
1131 	iov.iov_len = queue->left;
1132 	len = kernel_recvmsg(queue->sock, &msg, &iov, 1,
1133 			iov.iov_len, msg.msg_flags);
1134 	if (unlikely(len < 0))
1135 		return len;
1136 
1137 	queue->offset += len;
1138 	queue->left -= len;
1139 	if (queue->left)
1140 		return -EAGAIN;
1141 
1142 	if (queue->offset == sizeof(struct nvme_tcp_hdr)) {
1143 		u8 hdgst = nvmet_tcp_hdgst_len(queue);
1144 
1145 		if (unlikely(!nvmet_tcp_pdu_valid(hdr->type))) {
1146 			pr_err("unexpected pdu type %d\n", hdr->type);
1147 			nvmet_tcp_fatal_error(queue);
1148 			return -EIO;
1149 		}
1150 
1151 		if (unlikely(hdr->hlen != nvmet_tcp_pdu_size(hdr->type))) {
1152 			pr_err("pdu %d bad hlen %d\n", hdr->type, hdr->hlen);
1153 			return -EIO;
1154 		}
1155 
1156 		queue->left = hdr->hlen - queue->offset + hdgst;
1157 		goto recv;
1158 	}
1159 
1160 	if (queue->hdr_digest &&
1161 	    nvmet_tcp_verify_hdgst(queue, &queue->pdu, hdr->hlen)) {
1162 		nvmet_tcp_fatal_error(queue); /* fatal */
1163 		return -EPROTO;
1164 	}
1165 
1166 	if (queue->data_digest &&
1167 	    nvmet_tcp_check_ddgst(queue, &queue->pdu)) {
1168 		nvmet_tcp_fatal_error(queue); /* fatal */
1169 		return -EPROTO;
1170 	}
1171 
1172 	return nvmet_tcp_done_recv_pdu(queue);
1173 }
1174 
nvmet_tcp_prep_recv_ddgst(struct nvmet_tcp_cmd * cmd)1175 static void nvmet_tcp_prep_recv_ddgst(struct nvmet_tcp_cmd *cmd)
1176 {
1177 	struct nvmet_tcp_queue *queue = cmd->queue;
1178 
1179 	nvmet_tcp_calc_ddgst(queue->rcv_hash, cmd);
1180 	queue->offset = 0;
1181 	queue->left = NVME_TCP_DIGEST_LENGTH;
1182 	queue->rcv_state = NVMET_TCP_RECV_DDGST;
1183 }
1184 
nvmet_tcp_try_recv_data(struct nvmet_tcp_queue * queue)1185 static int nvmet_tcp_try_recv_data(struct nvmet_tcp_queue *queue)
1186 {
1187 	struct nvmet_tcp_cmd  *cmd = queue->cmd;
1188 	int ret;
1189 
1190 	while (msg_data_left(&cmd->recv_msg)) {
1191 		ret = sock_recvmsg(cmd->queue->sock, &cmd->recv_msg,
1192 			cmd->recv_msg.msg_flags);
1193 		if (ret <= 0)
1194 			return ret;
1195 
1196 		cmd->pdu_recv += ret;
1197 		cmd->rbytes_done += ret;
1198 	}
1199 
1200 	if (queue->data_digest) {
1201 		nvmet_tcp_prep_recv_ddgst(cmd);
1202 		return 0;
1203 	}
1204 
1205 	if (cmd->rbytes_done == cmd->req.transfer_len)
1206 		nvmet_tcp_execute_request(cmd);
1207 
1208 	nvmet_prepare_receive_pdu(queue);
1209 	return 0;
1210 }
1211 
nvmet_tcp_try_recv_ddgst(struct nvmet_tcp_queue * queue)1212 static int nvmet_tcp_try_recv_ddgst(struct nvmet_tcp_queue *queue)
1213 {
1214 	struct nvmet_tcp_cmd *cmd = queue->cmd;
1215 	int ret;
1216 	struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
1217 	struct kvec iov = {
1218 		.iov_base = (void *)&cmd->recv_ddgst + queue->offset,
1219 		.iov_len = queue->left
1220 	};
1221 
1222 	ret = kernel_recvmsg(queue->sock, &msg, &iov, 1,
1223 			iov.iov_len, msg.msg_flags);
1224 	if (unlikely(ret < 0))
1225 		return ret;
1226 
1227 	queue->offset += ret;
1228 	queue->left -= ret;
1229 	if (queue->left)
1230 		return -EAGAIN;
1231 
1232 	if (queue->data_digest && cmd->exp_ddgst != cmd->recv_ddgst) {
1233 		pr_err("queue %d: cmd %d pdu (%d) data digest error: recv %#x expected %#x\n",
1234 			queue->idx, cmd->req.cmd->common.command_id,
1235 			queue->pdu.cmd.hdr.type, le32_to_cpu(cmd->recv_ddgst),
1236 			le32_to_cpu(cmd->exp_ddgst));
1237 		nvmet_req_uninit(&cmd->req);
1238 		nvmet_tcp_free_cmd_buffers(cmd);
1239 		nvmet_tcp_fatal_error(queue);
1240 		ret = -EPROTO;
1241 		goto out;
1242 	}
1243 
1244 	if (cmd->rbytes_done == cmd->req.transfer_len)
1245 		nvmet_tcp_execute_request(cmd);
1246 
1247 	ret = 0;
1248 out:
1249 	nvmet_prepare_receive_pdu(queue);
1250 	return ret;
1251 }
1252 
nvmet_tcp_try_recv_one(struct nvmet_tcp_queue * queue)1253 static int nvmet_tcp_try_recv_one(struct nvmet_tcp_queue *queue)
1254 {
1255 	int result = 0;
1256 
1257 	if (unlikely(queue->rcv_state == NVMET_TCP_RECV_ERR))
1258 		return 0;
1259 
1260 	if (queue->rcv_state == NVMET_TCP_RECV_PDU) {
1261 		result = nvmet_tcp_try_recv_pdu(queue);
1262 		if (result != 0)
1263 			goto done_recv;
1264 	}
1265 
1266 	if (queue->rcv_state == NVMET_TCP_RECV_DATA) {
1267 		result = nvmet_tcp_try_recv_data(queue);
1268 		if (result != 0)
1269 			goto done_recv;
1270 	}
1271 
1272 	if (queue->rcv_state == NVMET_TCP_RECV_DDGST) {
1273 		result = nvmet_tcp_try_recv_ddgst(queue);
1274 		if (result != 0)
1275 			goto done_recv;
1276 	}
1277 
1278 done_recv:
1279 	if (result < 0) {
1280 		if (result == -EAGAIN)
1281 			return 0;
1282 		return result;
1283 	}
1284 	return 1;
1285 }
1286 
nvmet_tcp_try_recv(struct nvmet_tcp_queue * queue,int budget,int * recvs)1287 static int nvmet_tcp_try_recv(struct nvmet_tcp_queue *queue,
1288 		int budget, int *recvs)
1289 {
1290 	int i, ret = 0;
1291 
1292 	for (i = 0; i < budget; i++) {
1293 		ret = nvmet_tcp_try_recv_one(queue);
1294 		if (unlikely(ret < 0)) {
1295 			nvmet_tcp_socket_error(queue, ret);
1296 			goto done;
1297 		} else if (ret == 0) {
1298 			break;
1299 		}
1300 		(*recvs)++;
1301 	}
1302 done:
1303 	return ret;
1304 }
1305 
nvmet_tcp_schedule_release_queue(struct nvmet_tcp_queue * queue)1306 static void nvmet_tcp_schedule_release_queue(struct nvmet_tcp_queue *queue)
1307 {
1308 	spin_lock(&queue->state_lock);
1309 	if (queue->state != NVMET_TCP_Q_DISCONNECTING) {
1310 		queue->state = NVMET_TCP_Q_DISCONNECTING;
1311 		queue_work(nvmet_wq, &queue->release_work);
1312 	}
1313 	spin_unlock(&queue->state_lock);
1314 }
1315 
nvmet_tcp_arm_queue_deadline(struct nvmet_tcp_queue * queue)1316 static inline void nvmet_tcp_arm_queue_deadline(struct nvmet_tcp_queue *queue)
1317 {
1318 	queue->poll_end = jiffies + usecs_to_jiffies(idle_poll_period_usecs);
1319 }
1320 
nvmet_tcp_check_queue_deadline(struct nvmet_tcp_queue * queue,int ops)1321 static bool nvmet_tcp_check_queue_deadline(struct nvmet_tcp_queue *queue,
1322 		int ops)
1323 {
1324 	if (!idle_poll_period_usecs)
1325 		return false;
1326 
1327 	if (ops)
1328 		nvmet_tcp_arm_queue_deadline(queue);
1329 
1330 	return !time_after(jiffies, queue->poll_end);
1331 }
1332 
nvmet_tcp_io_work(struct work_struct * w)1333 static void nvmet_tcp_io_work(struct work_struct *w)
1334 {
1335 	struct nvmet_tcp_queue *queue =
1336 		container_of(w, struct nvmet_tcp_queue, io_work);
1337 	bool pending;
1338 	int ret, ops = 0;
1339 
1340 	do {
1341 		pending = false;
1342 
1343 		ret = nvmet_tcp_try_recv(queue, NVMET_TCP_RECV_BUDGET, &ops);
1344 		if (ret > 0)
1345 			pending = true;
1346 		else if (ret < 0)
1347 			return;
1348 
1349 		ret = nvmet_tcp_try_send(queue, NVMET_TCP_SEND_BUDGET, &ops);
1350 		if (ret > 0)
1351 			pending = true;
1352 		else if (ret < 0)
1353 			return;
1354 
1355 	} while (pending && ops < NVMET_TCP_IO_WORK_BUDGET);
1356 
1357 	/*
1358 	 * Requeue the worker if idle deadline period is in progress or any
1359 	 * ops activity was recorded during the do-while loop above.
1360 	 */
1361 	if (nvmet_tcp_check_queue_deadline(queue, ops) || pending)
1362 		queue_work_on(queue_cpu(queue), nvmet_tcp_wq, &queue->io_work);
1363 }
1364 
nvmet_tcp_alloc_cmd(struct nvmet_tcp_queue * queue,struct nvmet_tcp_cmd * c)1365 static int nvmet_tcp_alloc_cmd(struct nvmet_tcp_queue *queue,
1366 		struct nvmet_tcp_cmd *c)
1367 {
1368 	u8 hdgst = nvmet_tcp_hdgst_len(queue);
1369 
1370 	c->queue = queue;
1371 	c->req.port = queue->port->nport;
1372 
1373 	c->cmd_pdu = page_frag_alloc(&queue->pf_cache,
1374 			sizeof(*c->cmd_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO);
1375 	if (!c->cmd_pdu)
1376 		return -ENOMEM;
1377 	c->req.cmd = &c->cmd_pdu->cmd;
1378 
1379 	c->rsp_pdu = page_frag_alloc(&queue->pf_cache,
1380 			sizeof(*c->rsp_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO);
1381 	if (!c->rsp_pdu)
1382 		goto out_free_cmd;
1383 	c->req.cqe = &c->rsp_pdu->cqe;
1384 
1385 	c->data_pdu = page_frag_alloc(&queue->pf_cache,
1386 			sizeof(*c->data_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO);
1387 	if (!c->data_pdu)
1388 		goto out_free_rsp;
1389 
1390 	c->r2t_pdu = page_frag_alloc(&queue->pf_cache,
1391 			sizeof(*c->r2t_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO);
1392 	if (!c->r2t_pdu)
1393 		goto out_free_data;
1394 
1395 	c->recv_msg.msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
1396 
1397 	list_add_tail(&c->entry, &queue->free_list);
1398 
1399 	return 0;
1400 out_free_data:
1401 	page_frag_free(c->data_pdu);
1402 out_free_rsp:
1403 	page_frag_free(c->rsp_pdu);
1404 out_free_cmd:
1405 	page_frag_free(c->cmd_pdu);
1406 	return -ENOMEM;
1407 }
1408 
nvmet_tcp_free_cmd(struct nvmet_tcp_cmd * c)1409 static void nvmet_tcp_free_cmd(struct nvmet_tcp_cmd *c)
1410 {
1411 	page_frag_free(c->r2t_pdu);
1412 	page_frag_free(c->data_pdu);
1413 	page_frag_free(c->rsp_pdu);
1414 	page_frag_free(c->cmd_pdu);
1415 }
1416 
nvmet_tcp_alloc_cmds(struct nvmet_tcp_queue * queue)1417 static int nvmet_tcp_alloc_cmds(struct nvmet_tcp_queue *queue)
1418 {
1419 	struct nvmet_tcp_cmd *cmds;
1420 	int i, ret = -EINVAL, nr_cmds = queue->nr_cmds;
1421 
1422 	cmds = kcalloc(nr_cmds, sizeof(struct nvmet_tcp_cmd), GFP_KERNEL);
1423 	if (!cmds)
1424 		goto out;
1425 
1426 	for (i = 0; i < nr_cmds; i++) {
1427 		ret = nvmet_tcp_alloc_cmd(queue, cmds + i);
1428 		if (ret)
1429 			goto out_free;
1430 	}
1431 
1432 	queue->cmds = cmds;
1433 
1434 	return 0;
1435 out_free:
1436 	while (--i >= 0)
1437 		nvmet_tcp_free_cmd(cmds + i);
1438 	kfree(cmds);
1439 out:
1440 	return ret;
1441 }
1442 
nvmet_tcp_free_cmds(struct nvmet_tcp_queue * queue)1443 static void nvmet_tcp_free_cmds(struct nvmet_tcp_queue *queue)
1444 {
1445 	struct nvmet_tcp_cmd *cmds = queue->cmds;
1446 	int i;
1447 
1448 	for (i = 0; i < queue->nr_cmds; i++)
1449 		nvmet_tcp_free_cmd(cmds + i);
1450 
1451 	nvmet_tcp_free_cmd(&queue->connect);
1452 	kfree(cmds);
1453 }
1454 
nvmet_tcp_restore_socket_callbacks(struct nvmet_tcp_queue * queue)1455 static void nvmet_tcp_restore_socket_callbacks(struct nvmet_tcp_queue *queue)
1456 {
1457 	struct socket *sock = queue->sock;
1458 
1459 	write_lock_bh(&sock->sk->sk_callback_lock);
1460 	sock->sk->sk_data_ready =  queue->data_ready;
1461 	sock->sk->sk_state_change = queue->state_change;
1462 	sock->sk->sk_write_space = queue->write_space;
1463 	sock->sk->sk_user_data = NULL;
1464 	write_unlock_bh(&sock->sk->sk_callback_lock);
1465 }
1466 
nvmet_tcp_uninit_data_in_cmds(struct nvmet_tcp_queue * queue)1467 static void nvmet_tcp_uninit_data_in_cmds(struct nvmet_tcp_queue *queue)
1468 {
1469 	struct nvmet_tcp_cmd *cmd = queue->cmds;
1470 	int i;
1471 
1472 	for (i = 0; i < queue->nr_cmds; i++, cmd++) {
1473 		if (nvmet_tcp_need_data_in(cmd))
1474 			nvmet_req_uninit(&cmd->req);
1475 	}
1476 
1477 	if (!queue->nr_cmds && nvmet_tcp_need_data_in(&queue->connect)) {
1478 		/* failed in connect */
1479 		nvmet_req_uninit(&queue->connect.req);
1480 	}
1481 }
1482 
nvmet_tcp_free_cmd_data_in_buffers(struct nvmet_tcp_queue * queue)1483 static void nvmet_tcp_free_cmd_data_in_buffers(struct nvmet_tcp_queue *queue)
1484 {
1485 	struct nvmet_tcp_cmd *cmd = queue->cmds;
1486 	int i;
1487 
1488 	for (i = 0; i < queue->nr_cmds; i++, cmd++)
1489 		nvmet_tcp_free_cmd_buffers(cmd);
1490 	nvmet_tcp_free_cmd_buffers(&queue->connect);
1491 }
1492 
nvmet_tcp_release_queue_work(struct work_struct * w)1493 static void nvmet_tcp_release_queue_work(struct work_struct *w)
1494 {
1495 	struct page *page;
1496 	struct nvmet_tcp_queue *queue =
1497 		container_of(w, struct nvmet_tcp_queue, release_work);
1498 
1499 	mutex_lock(&nvmet_tcp_queue_mutex);
1500 	list_del_init(&queue->queue_list);
1501 	mutex_unlock(&nvmet_tcp_queue_mutex);
1502 
1503 	nvmet_tcp_restore_socket_callbacks(queue);
1504 	cancel_work_sync(&queue->io_work);
1505 	/* stop accepting incoming data */
1506 	queue->rcv_state = NVMET_TCP_RECV_ERR;
1507 
1508 	nvmet_tcp_uninit_data_in_cmds(queue);
1509 	nvmet_sq_destroy(&queue->nvme_sq);
1510 	cancel_work_sync(&queue->io_work);
1511 	nvmet_tcp_free_cmd_data_in_buffers(queue);
1512 	sock_release(queue->sock);
1513 	nvmet_tcp_free_cmds(queue);
1514 	if (queue->hdr_digest || queue->data_digest)
1515 		nvmet_tcp_free_crypto(queue);
1516 	ida_free(&nvmet_tcp_queue_ida, queue->idx);
1517 
1518 	page = virt_to_head_page(queue->pf_cache.va);
1519 	__page_frag_cache_drain(page, queue->pf_cache.pagecnt_bias);
1520 	kfree(queue);
1521 }
1522 
nvmet_tcp_data_ready(struct sock * sk)1523 static void nvmet_tcp_data_ready(struct sock *sk)
1524 {
1525 	struct nvmet_tcp_queue *queue;
1526 
1527 	trace_sk_data_ready(sk);
1528 
1529 	read_lock_bh(&sk->sk_callback_lock);
1530 	queue = sk->sk_user_data;
1531 	if (likely(queue))
1532 		queue_work_on(queue_cpu(queue), nvmet_tcp_wq, &queue->io_work);
1533 	read_unlock_bh(&sk->sk_callback_lock);
1534 }
1535 
nvmet_tcp_write_space(struct sock * sk)1536 static void nvmet_tcp_write_space(struct sock *sk)
1537 {
1538 	struct nvmet_tcp_queue *queue;
1539 
1540 	read_lock_bh(&sk->sk_callback_lock);
1541 	queue = sk->sk_user_data;
1542 	if (unlikely(!queue))
1543 		goto out;
1544 
1545 	if (unlikely(queue->state == NVMET_TCP_Q_CONNECTING)) {
1546 		queue->write_space(sk);
1547 		goto out;
1548 	}
1549 
1550 	if (sk_stream_is_writeable(sk)) {
1551 		clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1552 		queue_work_on(queue_cpu(queue), nvmet_tcp_wq, &queue->io_work);
1553 	}
1554 out:
1555 	read_unlock_bh(&sk->sk_callback_lock);
1556 }
1557 
nvmet_tcp_state_change(struct sock * sk)1558 static void nvmet_tcp_state_change(struct sock *sk)
1559 {
1560 	struct nvmet_tcp_queue *queue;
1561 
1562 	read_lock_bh(&sk->sk_callback_lock);
1563 	queue = sk->sk_user_data;
1564 	if (!queue)
1565 		goto done;
1566 
1567 	switch (sk->sk_state) {
1568 	case TCP_FIN_WAIT2:
1569 	case TCP_LAST_ACK:
1570 		break;
1571 	case TCP_FIN_WAIT1:
1572 	case TCP_CLOSE_WAIT:
1573 	case TCP_CLOSE:
1574 		/* FALLTHRU */
1575 		nvmet_tcp_schedule_release_queue(queue);
1576 		break;
1577 	default:
1578 		pr_warn("queue %d unhandled state %d\n",
1579 			queue->idx, sk->sk_state);
1580 	}
1581 done:
1582 	read_unlock_bh(&sk->sk_callback_lock);
1583 }
1584 
nvmet_tcp_set_queue_sock(struct nvmet_tcp_queue * queue)1585 static int nvmet_tcp_set_queue_sock(struct nvmet_tcp_queue *queue)
1586 {
1587 	struct socket *sock = queue->sock;
1588 	struct inet_sock *inet = inet_sk(sock->sk);
1589 	int ret;
1590 
1591 	ret = kernel_getsockname(sock,
1592 		(struct sockaddr *)&queue->sockaddr);
1593 	if (ret < 0)
1594 		return ret;
1595 
1596 	ret = kernel_getpeername(sock,
1597 		(struct sockaddr *)&queue->sockaddr_peer);
1598 	if (ret < 0)
1599 		return ret;
1600 
1601 	/*
1602 	 * Cleanup whatever is sitting in the TCP transmit queue on socket
1603 	 * close. This is done to prevent stale data from being sent should
1604 	 * the network connection be restored before TCP times out.
1605 	 */
1606 	sock_no_linger(sock->sk);
1607 
1608 	if (so_priority > 0)
1609 		sock_set_priority(sock->sk, so_priority);
1610 
1611 	/* Set socket type of service */
1612 	if (inet->rcv_tos > 0)
1613 		ip_sock_set_tos(sock->sk, inet->rcv_tos);
1614 
1615 	ret = 0;
1616 	write_lock_bh(&sock->sk->sk_callback_lock);
1617 	if (sock->sk->sk_state != TCP_ESTABLISHED) {
1618 		/*
1619 		 * If the socket is already closing, don't even start
1620 		 * consuming it
1621 		 */
1622 		ret = -ENOTCONN;
1623 	} else {
1624 		sock->sk->sk_user_data = queue;
1625 		queue->data_ready = sock->sk->sk_data_ready;
1626 		sock->sk->sk_data_ready = nvmet_tcp_data_ready;
1627 		queue->state_change = sock->sk->sk_state_change;
1628 		sock->sk->sk_state_change = nvmet_tcp_state_change;
1629 		queue->write_space = sock->sk->sk_write_space;
1630 		sock->sk->sk_write_space = nvmet_tcp_write_space;
1631 		if (idle_poll_period_usecs)
1632 			nvmet_tcp_arm_queue_deadline(queue);
1633 		queue_work_on(queue_cpu(queue), nvmet_tcp_wq, &queue->io_work);
1634 	}
1635 	write_unlock_bh(&sock->sk->sk_callback_lock);
1636 
1637 	return ret;
1638 }
1639 
nvmet_tcp_alloc_queue(struct nvmet_tcp_port * port,struct socket * newsock)1640 static int nvmet_tcp_alloc_queue(struct nvmet_tcp_port *port,
1641 		struct socket *newsock)
1642 {
1643 	struct nvmet_tcp_queue *queue;
1644 	int ret;
1645 
1646 	queue = kzalloc(sizeof(*queue), GFP_KERNEL);
1647 	if (!queue)
1648 		return -ENOMEM;
1649 
1650 	INIT_WORK(&queue->release_work, nvmet_tcp_release_queue_work);
1651 	INIT_WORK(&queue->io_work, nvmet_tcp_io_work);
1652 	queue->sock = newsock;
1653 	queue->port = port;
1654 	queue->nr_cmds = 0;
1655 	spin_lock_init(&queue->state_lock);
1656 	queue->state = NVMET_TCP_Q_CONNECTING;
1657 	INIT_LIST_HEAD(&queue->free_list);
1658 	init_llist_head(&queue->resp_list);
1659 	INIT_LIST_HEAD(&queue->resp_send_list);
1660 
1661 	queue->idx = ida_alloc(&nvmet_tcp_queue_ida, GFP_KERNEL);
1662 	if (queue->idx < 0) {
1663 		ret = queue->idx;
1664 		goto out_free_queue;
1665 	}
1666 
1667 	ret = nvmet_tcp_alloc_cmd(queue, &queue->connect);
1668 	if (ret)
1669 		goto out_ida_remove;
1670 
1671 	ret = nvmet_sq_init(&queue->nvme_sq);
1672 	if (ret)
1673 		goto out_free_connect;
1674 
1675 	nvmet_prepare_receive_pdu(queue);
1676 
1677 	mutex_lock(&nvmet_tcp_queue_mutex);
1678 	list_add_tail(&queue->queue_list, &nvmet_tcp_queue_list);
1679 	mutex_unlock(&nvmet_tcp_queue_mutex);
1680 
1681 	ret = nvmet_tcp_set_queue_sock(queue);
1682 	if (ret)
1683 		goto out_destroy_sq;
1684 
1685 	return 0;
1686 out_destroy_sq:
1687 	mutex_lock(&nvmet_tcp_queue_mutex);
1688 	list_del_init(&queue->queue_list);
1689 	mutex_unlock(&nvmet_tcp_queue_mutex);
1690 	nvmet_sq_destroy(&queue->nvme_sq);
1691 out_free_connect:
1692 	nvmet_tcp_free_cmd(&queue->connect);
1693 out_ida_remove:
1694 	ida_free(&nvmet_tcp_queue_ida, queue->idx);
1695 out_free_queue:
1696 	kfree(queue);
1697 	return ret;
1698 }
1699 
nvmet_tcp_accept_work(struct work_struct * w)1700 static void nvmet_tcp_accept_work(struct work_struct *w)
1701 {
1702 	struct nvmet_tcp_port *port =
1703 		container_of(w, struct nvmet_tcp_port, accept_work);
1704 	struct socket *newsock;
1705 	int ret;
1706 
1707 	while (true) {
1708 		ret = kernel_accept(port->sock, &newsock, O_NONBLOCK);
1709 		if (ret < 0) {
1710 			if (ret != -EAGAIN)
1711 				pr_warn("failed to accept err=%d\n", ret);
1712 			return;
1713 		}
1714 		ret = nvmet_tcp_alloc_queue(port, newsock);
1715 		if (ret) {
1716 			pr_err("failed to allocate queue\n");
1717 			sock_release(newsock);
1718 		}
1719 	}
1720 }
1721 
nvmet_tcp_listen_data_ready(struct sock * sk)1722 static void nvmet_tcp_listen_data_ready(struct sock *sk)
1723 {
1724 	struct nvmet_tcp_port *port;
1725 
1726 	trace_sk_data_ready(sk);
1727 
1728 	read_lock_bh(&sk->sk_callback_lock);
1729 	port = sk->sk_user_data;
1730 	if (!port)
1731 		goto out;
1732 
1733 	if (sk->sk_state == TCP_LISTEN)
1734 		queue_work(nvmet_wq, &port->accept_work);
1735 out:
1736 	read_unlock_bh(&sk->sk_callback_lock);
1737 }
1738 
nvmet_tcp_add_port(struct nvmet_port * nport)1739 static int nvmet_tcp_add_port(struct nvmet_port *nport)
1740 {
1741 	struct nvmet_tcp_port *port;
1742 	__kernel_sa_family_t af;
1743 	int ret;
1744 
1745 	port = kzalloc(sizeof(*port), GFP_KERNEL);
1746 	if (!port)
1747 		return -ENOMEM;
1748 
1749 	switch (nport->disc_addr.adrfam) {
1750 	case NVMF_ADDR_FAMILY_IP4:
1751 		af = AF_INET;
1752 		break;
1753 	case NVMF_ADDR_FAMILY_IP6:
1754 		af = AF_INET6;
1755 		break;
1756 	default:
1757 		pr_err("address family %d not supported\n",
1758 				nport->disc_addr.adrfam);
1759 		ret = -EINVAL;
1760 		goto err_port;
1761 	}
1762 
1763 	ret = inet_pton_with_scope(&init_net, af, nport->disc_addr.traddr,
1764 			nport->disc_addr.trsvcid, &port->addr);
1765 	if (ret) {
1766 		pr_err("malformed ip/port passed: %s:%s\n",
1767 			nport->disc_addr.traddr, nport->disc_addr.trsvcid);
1768 		goto err_port;
1769 	}
1770 
1771 	port->nport = nport;
1772 	INIT_WORK(&port->accept_work, nvmet_tcp_accept_work);
1773 	if (port->nport->inline_data_size < 0)
1774 		port->nport->inline_data_size = NVMET_TCP_DEF_INLINE_DATA_SIZE;
1775 
1776 	ret = sock_create(port->addr.ss_family, SOCK_STREAM,
1777 				IPPROTO_TCP, &port->sock);
1778 	if (ret) {
1779 		pr_err("failed to create a socket\n");
1780 		goto err_port;
1781 	}
1782 
1783 	port->sock->sk->sk_user_data = port;
1784 	port->data_ready = port->sock->sk->sk_data_ready;
1785 	port->sock->sk->sk_data_ready = nvmet_tcp_listen_data_ready;
1786 	sock_set_reuseaddr(port->sock->sk);
1787 	tcp_sock_set_nodelay(port->sock->sk);
1788 	if (so_priority > 0)
1789 		sock_set_priority(port->sock->sk, so_priority);
1790 
1791 	ret = kernel_bind(port->sock, (struct sockaddr *)&port->addr,
1792 			sizeof(port->addr));
1793 	if (ret) {
1794 		pr_err("failed to bind port socket %d\n", ret);
1795 		goto err_sock;
1796 	}
1797 
1798 	ret = kernel_listen(port->sock, 128);
1799 	if (ret) {
1800 		pr_err("failed to listen %d on port sock\n", ret);
1801 		goto err_sock;
1802 	}
1803 
1804 	nport->priv = port;
1805 	pr_info("enabling port %d (%pISpc)\n",
1806 		le16_to_cpu(nport->disc_addr.portid), &port->addr);
1807 
1808 	return 0;
1809 
1810 err_sock:
1811 	sock_release(port->sock);
1812 err_port:
1813 	kfree(port);
1814 	return ret;
1815 }
1816 
nvmet_tcp_destroy_port_queues(struct nvmet_tcp_port * port)1817 static void nvmet_tcp_destroy_port_queues(struct nvmet_tcp_port *port)
1818 {
1819 	struct nvmet_tcp_queue *queue;
1820 
1821 	mutex_lock(&nvmet_tcp_queue_mutex);
1822 	list_for_each_entry(queue, &nvmet_tcp_queue_list, queue_list)
1823 		if (queue->port == port)
1824 			kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1825 	mutex_unlock(&nvmet_tcp_queue_mutex);
1826 }
1827 
nvmet_tcp_remove_port(struct nvmet_port * nport)1828 static void nvmet_tcp_remove_port(struct nvmet_port *nport)
1829 {
1830 	struct nvmet_tcp_port *port = nport->priv;
1831 
1832 	write_lock_bh(&port->sock->sk->sk_callback_lock);
1833 	port->sock->sk->sk_data_ready = port->data_ready;
1834 	port->sock->sk->sk_user_data = NULL;
1835 	write_unlock_bh(&port->sock->sk->sk_callback_lock);
1836 	cancel_work_sync(&port->accept_work);
1837 	/*
1838 	 * Destroy the remaining queues, which are not belong to any
1839 	 * controller yet.
1840 	 */
1841 	nvmet_tcp_destroy_port_queues(port);
1842 
1843 	sock_release(port->sock);
1844 	kfree(port);
1845 }
1846 
nvmet_tcp_delete_ctrl(struct nvmet_ctrl * ctrl)1847 static void nvmet_tcp_delete_ctrl(struct nvmet_ctrl *ctrl)
1848 {
1849 	struct nvmet_tcp_queue *queue;
1850 
1851 	mutex_lock(&nvmet_tcp_queue_mutex);
1852 	list_for_each_entry(queue, &nvmet_tcp_queue_list, queue_list)
1853 		if (queue->nvme_sq.ctrl == ctrl)
1854 			kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1855 	mutex_unlock(&nvmet_tcp_queue_mutex);
1856 }
1857 
nvmet_tcp_install_queue(struct nvmet_sq * sq)1858 static u16 nvmet_tcp_install_queue(struct nvmet_sq *sq)
1859 {
1860 	struct nvmet_tcp_queue *queue =
1861 		container_of(sq, struct nvmet_tcp_queue, nvme_sq);
1862 
1863 	if (sq->qid == 0) {
1864 		/* Let inflight controller teardown complete */
1865 		flush_workqueue(nvmet_wq);
1866 	}
1867 
1868 	queue->nr_cmds = sq->size * 2;
1869 	if (nvmet_tcp_alloc_cmds(queue)) {
1870 		queue->nr_cmds = 0;
1871 		return NVME_SC_INTERNAL;
1872 	}
1873 	return 0;
1874 }
1875 
nvmet_tcp_disc_port_addr(struct nvmet_req * req,struct nvmet_port * nport,char * traddr)1876 static void nvmet_tcp_disc_port_addr(struct nvmet_req *req,
1877 		struct nvmet_port *nport, char *traddr)
1878 {
1879 	struct nvmet_tcp_port *port = nport->priv;
1880 
1881 	if (inet_addr_is_any((struct sockaddr *)&port->addr)) {
1882 		struct nvmet_tcp_cmd *cmd =
1883 			container_of(req, struct nvmet_tcp_cmd, req);
1884 		struct nvmet_tcp_queue *queue = cmd->queue;
1885 
1886 		sprintf(traddr, "%pISc", (struct sockaddr *)&queue->sockaddr);
1887 	} else {
1888 		memcpy(traddr, nport->disc_addr.traddr, NVMF_TRADDR_SIZE);
1889 	}
1890 }
1891 
1892 static const struct nvmet_fabrics_ops nvmet_tcp_ops = {
1893 	.owner			= THIS_MODULE,
1894 	.type			= NVMF_TRTYPE_TCP,
1895 	.msdbd			= 1,
1896 	.add_port		= nvmet_tcp_add_port,
1897 	.remove_port		= nvmet_tcp_remove_port,
1898 	.queue_response		= nvmet_tcp_queue_response,
1899 	.delete_ctrl		= nvmet_tcp_delete_ctrl,
1900 	.install_queue		= nvmet_tcp_install_queue,
1901 	.disc_traddr		= nvmet_tcp_disc_port_addr,
1902 };
1903 
nvmet_tcp_init(void)1904 static int __init nvmet_tcp_init(void)
1905 {
1906 	int ret;
1907 
1908 	nvmet_tcp_wq = alloc_workqueue("nvmet_tcp_wq",
1909 				WQ_MEM_RECLAIM | WQ_HIGHPRI, 0);
1910 	if (!nvmet_tcp_wq)
1911 		return -ENOMEM;
1912 
1913 	ret = nvmet_register_transport(&nvmet_tcp_ops);
1914 	if (ret)
1915 		goto err;
1916 
1917 	return 0;
1918 err:
1919 	destroy_workqueue(nvmet_tcp_wq);
1920 	return ret;
1921 }
1922 
nvmet_tcp_exit(void)1923 static void __exit nvmet_tcp_exit(void)
1924 {
1925 	struct nvmet_tcp_queue *queue;
1926 
1927 	nvmet_unregister_transport(&nvmet_tcp_ops);
1928 
1929 	flush_workqueue(nvmet_wq);
1930 	mutex_lock(&nvmet_tcp_queue_mutex);
1931 	list_for_each_entry(queue, &nvmet_tcp_queue_list, queue_list)
1932 		kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1933 	mutex_unlock(&nvmet_tcp_queue_mutex);
1934 	flush_workqueue(nvmet_wq);
1935 
1936 	destroy_workqueue(nvmet_tcp_wq);
1937 	ida_destroy(&nvmet_tcp_queue_ida);
1938 }
1939 
1940 module_init(nvmet_tcp_init);
1941 module_exit(nvmet_tcp_exit);
1942 
1943 MODULE_LICENSE("GPL v2");
1944 MODULE_ALIAS("nvmet-transport-3"); /* 3 == NVMF_TRTYPE_TCP */
1945