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