1 // SPDX-License-Identifier: GPL-2.0
3 * NVMe over Fabrics TCP target.
4 * Copyright (c) 2018 Lightbits Labs. All rights reserved.
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>
14 #include <linux/inet.h>
15 #include <linux/llist.h>
16 #include <crypto/hash.h>
17 #include <trace/events/sock.h>
21 #define NVMET_TCP_DEF_INLINE_DATA_SIZE (4 * PAGE_SIZE)
23 static int param_store_val(const char *str, int *val, int min, int max)
27 ret = kstrtoint(str, 10, &new_val);
31 if (new_val < min || new_val > max)
38 static int set_params(const char *str, const struct kernel_param *kp)
40 return param_store_val(str, kp->arg, 0, INT_MAX);
43 static const struct kernel_param_ops set_param_ops = {
48 /* Define the socket priority to use for connections were it is desirable
49 * that the NIC consider performing optimized packet processing or filtering.
50 * A non-zero value being sufficient to indicate general consideration of any
51 * possible optimization. Making it a module param allows for alternative
52 * values that may be unique for some NIC implementations.
54 static int so_priority;
55 device_param_cb(so_priority, &set_param_ops, &so_priority, 0644);
56 MODULE_PARM_DESC(so_priority, "nvmet tcp socket optimize priority: Default 0");
58 /* Define a time period (in usecs) that io_work() shall sample an activated
59 * queue before determining it to be idle. This optional module behavior
60 * can enable NIC solutions that support socket optimized packet processing
61 * using advanced interrupt moderation techniques.
63 static int idle_poll_period_usecs;
64 device_param_cb(idle_poll_period_usecs, &set_param_ops,
65 &idle_poll_period_usecs, 0644);
66 MODULE_PARM_DESC(idle_poll_period_usecs,
67 "nvmet tcp io_work poll till idle time period in usecs: Default 0");
69 #define NVMET_TCP_RECV_BUDGET 8
70 #define NVMET_TCP_SEND_BUDGET 8
71 #define NVMET_TCP_IO_WORK_BUDGET 64
73 enum nvmet_tcp_send_state {
74 NVMET_TCP_SEND_DATA_PDU,
78 NVMET_TCP_SEND_RESPONSE
81 enum nvmet_tcp_recv_state {
89 NVMET_TCP_F_INIT_FAILED = (1 << 0),
92 struct nvmet_tcp_cmd {
93 struct nvmet_tcp_queue *queue;
96 struct nvme_tcp_cmd_pdu *cmd_pdu;
97 struct nvme_tcp_rsp_pdu *rsp_pdu;
98 struct nvme_tcp_data_pdu *data_pdu;
99 struct nvme_tcp_r2t_pdu *r2t_pdu;
107 struct msghdr recv_msg;
111 struct list_head entry;
112 struct llist_node lentry;
116 struct scatterlist *cur_sg;
117 enum nvmet_tcp_send_state state;
123 enum nvmet_tcp_queue_state {
124 NVMET_TCP_Q_CONNECTING,
126 NVMET_TCP_Q_DISCONNECTING,
129 struct nvmet_tcp_queue {
131 struct nvmet_tcp_port *port;
132 struct work_struct io_work;
133 struct nvmet_cq nvme_cq;
134 struct nvmet_sq nvme_sq;
137 struct nvmet_tcp_cmd *cmds;
138 unsigned int nr_cmds;
139 struct list_head free_list;
140 struct llist_head resp_list;
141 struct list_head resp_send_list;
143 struct nvmet_tcp_cmd *snd_cmd;
148 enum nvmet_tcp_recv_state rcv_state;
149 struct nvmet_tcp_cmd *cmd;
150 union nvme_tcp_pdu pdu;
155 struct ahash_request *snd_hash;
156 struct ahash_request *rcv_hash;
158 unsigned long poll_end;
160 spinlock_t state_lock;
161 enum nvmet_tcp_queue_state state;
163 struct sockaddr_storage sockaddr;
164 struct sockaddr_storage sockaddr_peer;
165 struct work_struct release_work;
168 struct list_head queue_list;
170 struct nvmet_tcp_cmd connect;
172 struct page_frag_cache pf_cache;
174 void (*data_ready)(struct sock *);
175 void (*state_change)(struct sock *);
176 void (*write_space)(struct sock *);
179 struct nvmet_tcp_port {
181 struct work_struct accept_work;
182 struct nvmet_port *nport;
183 struct sockaddr_storage addr;
184 void (*data_ready)(struct sock *);
187 static DEFINE_IDA(nvmet_tcp_queue_ida);
188 static LIST_HEAD(nvmet_tcp_queue_list);
189 static DEFINE_MUTEX(nvmet_tcp_queue_mutex);
191 static struct workqueue_struct *nvmet_tcp_wq;
192 static const struct nvmet_fabrics_ops nvmet_tcp_ops;
193 static void nvmet_tcp_free_cmd(struct nvmet_tcp_cmd *c);
194 static void nvmet_tcp_free_cmd_buffers(struct nvmet_tcp_cmd *cmd);
196 static inline u16 nvmet_tcp_cmd_tag(struct nvmet_tcp_queue *queue,
197 struct nvmet_tcp_cmd *cmd)
199 if (unlikely(!queue->nr_cmds)) {
200 /* We didn't allocate cmds yet, send 0xffff */
204 return cmd - queue->cmds;
207 static inline bool nvmet_tcp_has_data_in(struct nvmet_tcp_cmd *cmd)
209 return nvme_is_write(cmd->req.cmd) &&
210 cmd->rbytes_done < cmd->req.transfer_len;
213 static inline bool nvmet_tcp_need_data_in(struct nvmet_tcp_cmd *cmd)
215 return nvmet_tcp_has_data_in(cmd) && !cmd->req.cqe->status;
218 static inline bool nvmet_tcp_need_data_out(struct nvmet_tcp_cmd *cmd)
220 return !nvme_is_write(cmd->req.cmd) &&
221 cmd->req.transfer_len > 0 &&
222 !cmd->req.cqe->status;
225 static inline bool nvmet_tcp_has_inline_data(struct nvmet_tcp_cmd *cmd)
227 return nvme_is_write(cmd->req.cmd) && cmd->pdu_len &&
231 static inline struct nvmet_tcp_cmd *
232 nvmet_tcp_get_cmd(struct nvmet_tcp_queue *queue)
234 struct nvmet_tcp_cmd *cmd;
236 cmd = list_first_entry_or_null(&queue->free_list,
237 struct nvmet_tcp_cmd, entry);
240 list_del_init(&cmd->entry);
242 cmd->rbytes_done = cmd->wbytes_done = 0;
250 static inline void nvmet_tcp_put_cmd(struct nvmet_tcp_cmd *cmd)
252 if (unlikely(cmd == &cmd->queue->connect))
255 list_add_tail(&cmd->entry, &cmd->queue->free_list);
258 static inline int queue_cpu(struct nvmet_tcp_queue *queue)
260 return queue->sock->sk->sk_incoming_cpu;
263 static inline u8 nvmet_tcp_hdgst_len(struct nvmet_tcp_queue *queue)
265 return queue->hdr_digest ? NVME_TCP_DIGEST_LENGTH : 0;
268 static inline u8 nvmet_tcp_ddgst_len(struct nvmet_tcp_queue *queue)
270 return queue->data_digest ? NVME_TCP_DIGEST_LENGTH : 0;
273 static inline void nvmet_tcp_hdgst(struct ahash_request *hash,
274 void *pdu, size_t len)
276 struct scatterlist sg;
278 sg_init_one(&sg, pdu, len);
279 ahash_request_set_crypt(hash, &sg, pdu + len, len);
280 crypto_ahash_digest(hash);
283 static int nvmet_tcp_verify_hdgst(struct nvmet_tcp_queue *queue,
284 void *pdu, size_t len)
286 struct nvme_tcp_hdr *hdr = pdu;
290 if (unlikely(!(hdr->flags & NVME_TCP_F_HDGST))) {
291 pr_err("queue %d: header digest enabled but no header digest\n",
296 recv_digest = *(__le32 *)(pdu + hdr->hlen);
297 nvmet_tcp_hdgst(queue->rcv_hash, pdu, len);
298 exp_digest = *(__le32 *)(pdu + hdr->hlen);
299 if (recv_digest != exp_digest) {
300 pr_err("queue %d: header digest error: recv %#x expected %#x\n",
301 queue->idx, le32_to_cpu(recv_digest),
302 le32_to_cpu(exp_digest));
309 static int nvmet_tcp_check_ddgst(struct nvmet_tcp_queue *queue, void *pdu)
311 struct nvme_tcp_hdr *hdr = pdu;
312 u8 digest_len = nvmet_tcp_hdgst_len(queue);
315 len = le32_to_cpu(hdr->plen) - hdr->hlen -
316 (hdr->flags & NVME_TCP_F_HDGST ? digest_len : 0);
318 if (unlikely(len && !(hdr->flags & NVME_TCP_F_DDGST))) {
319 pr_err("queue %d: data digest flag is cleared\n", queue->idx);
326 static void nvmet_tcp_free_cmd_buffers(struct nvmet_tcp_cmd *cmd)
329 sgl_free(cmd->req.sg);
334 static void nvmet_tcp_build_pdu_iovec(struct nvmet_tcp_cmd *cmd)
336 struct bio_vec *iov = cmd->iov;
337 struct scatterlist *sg;
338 u32 length, offset, sg_offset;
341 length = cmd->pdu_len;
342 nr_pages = DIV_ROUND_UP(length, PAGE_SIZE);
343 offset = cmd->rbytes_done;
344 cmd->sg_idx = offset / PAGE_SIZE;
345 sg_offset = offset % PAGE_SIZE;
346 sg = &cmd->req.sg[cmd->sg_idx];
349 u32 iov_len = min_t(u32, length, sg->length - sg_offset);
351 bvec_set_page(iov, sg_page(sg), iov_len,
352 sg->offset + sg_offset);
360 iov_iter_bvec(&cmd->recv_msg.msg_iter, ITER_DEST, cmd->iov,
361 nr_pages, cmd->pdu_len);
364 static void nvmet_tcp_fatal_error(struct nvmet_tcp_queue *queue)
366 queue->rcv_state = NVMET_TCP_RECV_ERR;
367 if (queue->nvme_sq.ctrl)
368 nvmet_ctrl_fatal_error(queue->nvme_sq.ctrl);
370 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
373 static void nvmet_tcp_socket_error(struct nvmet_tcp_queue *queue, int status)
375 if (status == -EPIPE || status == -ECONNRESET)
376 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
378 nvmet_tcp_fatal_error(queue);
381 static int nvmet_tcp_map_data(struct nvmet_tcp_cmd *cmd)
383 struct nvme_sgl_desc *sgl = &cmd->req.cmd->common.dptr.sgl;
384 u32 len = le32_to_cpu(sgl->length);
389 if (sgl->type == ((NVME_SGL_FMT_DATA_DESC << 4) |
390 NVME_SGL_FMT_OFFSET)) {
391 if (!nvme_is_write(cmd->req.cmd))
392 return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
394 if (len > cmd->req.port->inline_data_size)
395 return NVME_SC_SGL_INVALID_OFFSET | NVME_SC_DNR;
398 cmd->req.transfer_len += len;
400 cmd->req.sg = sgl_alloc(len, GFP_KERNEL, &cmd->req.sg_cnt);
402 return NVME_SC_INTERNAL;
403 cmd->cur_sg = cmd->req.sg;
405 if (nvmet_tcp_has_data_in(cmd)) {
406 cmd->iov = kmalloc_array(cmd->req.sg_cnt,
407 sizeof(*cmd->iov), GFP_KERNEL);
414 nvmet_tcp_free_cmd_buffers(cmd);
415 return NVME_SC_INTERNAL;
418 static void nvmet_tcp_calc_ddgst(struct ahash_request *hash,
419 struct nvmet_tcp_cmd *cmd)
421 ahash_request_set_crypt(hash, cmd->req.sg,
422 (void *)&cmd->exp_ddgst, cmd->req.transfer_len);
423 crypto_ahash_digest(hash);
426 static void nvmet_setup_c2h_data_pdu(struct nvmet_tcp_cmd *cmd)
428 struct nvme_tcp_data_pdu *pdu = cmd->data_pdu;
429 struct nvmet_tcp_queue *queue = cmd->queue;
430 u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
431 u8 ddgst = nvmet_tcp_ddgst_len(cmd->queue);
434 cmd->state = NVMET_TCP_SEND_DATA_PDU;
436 pdu->hdr.type = nvme_tcp_c2h_data;
437 pdu->hdr.flags = NVME_TCP_F_DATA_LAST | (queue->nvme_sq.sqhd_disabled ?
438 NVME_TCP_F_DATA_SUCCESS : 0);
439 pdu->hdr.hlen = sizeof(*pdu);
440 pdu->hdr.pdo = pdu->hdr.hlen + hdgst;
442 cpu_to_le32(pdu->hdr.hlen + hdgst +
443 cmd->req.transfer_len + ddgst);
444 pdu->command_id = cmd->req.cqe->command_id;
445 pdu->data_length = cpu_to_le32(cmd->req.transfer_len);
446 pdu->data_offset = cpu_to_le32(cmd->wbytes_done);
448 if (queue->data_digest) {
449 pdu->hdr.flags |= NVME_TCP_F_DDGST;
450 nvmet_tcp_calc_ddgst(queue->snd_hash, cmd);
453 if (cmd->queue->hdr_digest) {
454 pdu->hdr.flags |= NVME_TCP_F_HDGST;
455 nvmet_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
459 static void nvmet_setup_r2t_pdu(struct nvmet_tcp_cmd *cmd)
461 struct nvme_tcp_r2t_pdu *pdu = cmd->r2t_pdu;
462 struct nvmet_tcp_queue *queue = cmd->queue;
463 u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
466 cmd->state = NVMET_TCP_SEND_R2T;
468 pdu->hdr.type = nvme_tcp_r2t;
470 pdu->hdr.hlen = sizeof(*pdu);
472 pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst);
474 pdu->command_id = cmd->req.cmd->common.command_id;
475 pdu->ttag = nvmet_tcp_cmd_tag(cmd->queue, cmd);
476 pdu->r2t_length = cpu_to_le32(cmd->req.transfer_len - cmd->rbytes_done);
477 pdu->r2t_offset = cpu_to_le32(cmd->rbytes_done);
478 if (cmd->queue->hdr_digest) {
479 pdu->hdr.flags |= NVME_TCP_F_HDGST;
480 nvmet_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
484 static void nvmet_setup_response_pdu(struct nvmet_tcp_cmd *cmd)
486 struct nvme_tcp_rsp_pdu *pdu = cmd->rsp_pdu;
487 struct nvmet_tcp_queue *queue = cmd->queue;
488 u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
491 cmd->state = NVMET_TCP_SEND_RESPONSE;
493 pdu->hdr.type = nvme_tcp_rsp;
495 pdu->hdr.hlen = sizeof(*pdu);
497 pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst);
498 if (cmd->queue->hdr_digest) {
499 pdu->hdr.flags |= NVME_TCP_F_HDGST;
500 nvmet_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
504 static void nvmet_tcp_process_resp_list(struct nvmet_tcp_queue *queue)
506 struct llist_node *node;
507 struct nvmet_tcp_cmd *cmd;
509 for (node = llist_del_all(&queue->resp_list); node; node = node->next) {
510 cmd = llist_entry(node, struct nvmet_tcp_cmd, lentry);
511 list_add(&cmd->entry, &queue->resp_send_list);
512 queue->send_list_len++;
516 static struct nvmet_tcp_cmd *nvmet_tcp_fetch_cmd(struct nvmet_tcp_queue *queue)
518 queue->snd_cmd = list_first_entry_or_null(&queue->resp_send_list,
519 struct nvmet_tcp_cmd, entry);
520 if (!queue->snd_cmd) {
521 nvmet_tcp_process_resp_list(queue);
523 list_first_entry_or_null(&queue->resp_send_list,
524 struct nvmet_tcp_cmd, entry);
525 if (unlikely(!queue->snd_cmd))
529 list_del_init(&queue->snd_cmd->entry);
530 queue->send_list_len--;
532 if (nvmet_tcp_need_data_out(queue->snd_cmd))
533 nvmet_setup_c2h_data_pdu(queue->snd_cmd);
534 else if (nvmet_tcp_need_data_in(queue->snd_cmd))
535 nvmet_setup_r2t_pdu(queue->snd_cmd);
537 nvmet_setup_response_pdu(queue->snd_cmd);
539 return queue->snd_cmd;
542 static void nvmet_tcp_queue_response(struct nvmet_req *req)
544 struct nvmet_tcp_cmd *cmd =
545 container_of(req, struct nvmet_tcp_cmd, req);
546 struct nvmet_tcp_queue *queue = cmd->queue;
547 struct nvme_sgl_desc *sgl;
550 if (unlikely(cmd == queue->cmd)) {
551 sgl = &cmd->req.cmd->common.dptr.sgl;
552 len = le32_to_cpu(sgl->length);
555 * Wait for inline data before processing the response.
556 * Avoid using helpers, this might happen before
557 * nvmet_req_init is completed.
559 if (queue->rcv_state == NVMET_TCP_RECV_PDU &&
560 len && len <= cmd->req.port->inline_data_size &&
561 nvme_is_write(cmd->req.cmd))
565 llist_add(&cmd->lentry, &queue->resp_list);
566 queue_work_on(queue_cpu(queue), nvmet_tcp_wq, &cmd->queue->io_work);
569 static void nvmet_tcp_execute_request(struct nvmet_tcp_cmd *cmd)
571 if (unlikely(cmd->flags & NVMET_TCP_F_INIT_FAILED))
572 nvmet_tcp_queue_response(&cmd->req);
574 cmd->req.execute(&cmd->req);
577 static int nvmet_try_send_data_pdu(struct nvmet_tcp_cmd *cmd)
579 struct msghdr msg = {
580 .msg_flags = MSG_DONTWAIT | MSG_MORE | MSG_SPLICE_PAGES,
583 u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
584 int left = sizeof(*cmd->data_pdu) - cmd->offset + hdgst;
587 bvec_set_virt(&bvec, (void *)cmd->data_pdu + cmd->offset, left);
588 iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, &bvec, 1, left);
589 ret = sock_sendmsg(cmd->queue->sock, &msg);
599 cmd->state = NVMET_TCP_SEND_DATA;
604 static int nvmet_try_send_data(struct nvmet_tcp_cmd *cmd, bool last_in_batch)
606 struct nvmet_tcp_queue *queue = cmd->queue;
609 while (cmd->cur_sg) {
610 struct msghdr msg = {
611 .msg_flags = MSG_DONTWAIT | MSG_SPLICE_PAGES,
613 struct page *page = sg_page(cmd->cur_sg);
615 u32 left = cmd->cur_sg->length - cmd->offset;
617 if ((!last_in_batch && cmd->queue->send_list_len) ||
618 cmd->wbytes_done + left < cmd->req.transfer_len ||
619 queue->data_digest || !queue->nvme_sq.sqhd_disabled)
620 msg.msg_flags |= MSG_MORE;
622 bvec_set_page(&bvec, page, left, cmd->offset);
623 iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, &bvec, 1, left);
624 ret = sock_sendmsg(cmd->queue->sock, &msg);
629 cmd->wbytes_done += ret;
632 if (cmd->offset == cmd->cur_sg->length) {
633 cmd->cur_sg = sg_next(cmd->cur_sg);
638 if (queue->data_digest) {
639 cmd->state = NVMET_TCP_SEND_DDGST;
642 if (queue->nvme_sq.sqhd_disabled) {
643 cmd->queue->snd_cmd = NULL;
644 nvmet_tcp_put_cmd(cmd);
646 nvmet_setup_response_pdu(cmd);
650 if (queue->nvme_sq.sqhd_disabled)
651 nvmet_tcp_free_cmd_buffers(cmd);
657 static int nvmet_try_send_response(struct nvmet_tcp_cmd *cmd,
660 struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_SPLICE_PAGES, };
662 u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
663 int left = sizeof(*cmd->rsp_pdu) - cmd->offset + hdgst;
666 if (!last_in_batch && cmd->queue->send_list_len)
667 msg.msg_flags |= MSG_MORE;
669 msg.msg_flags |= MSG_EOR;
671 bvec_set_virt(&bvec, (void *)cmd->rsp_pdu + cmd->offset, left);
672 iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, &bvec, 1, left);
673 ret = sock_sendmsg(cmd->queue->sock, &msg);
682 nvmet_tcp_free_cmd_buffers(cmd);
683 cmd->queue->snd_cmd = NULL;
684 nvmet_tcp_put_cmd(cmd);
688 static int nvmet_try_send_r2t(struct nvmet_tcp_cmd *cmd, bool last_in_batch)
690 struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_SPLICE_PAGES, };
692 u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
693 int left = sizeof(*cmd->r2t_pdu) - cmd->offset + hdgst;
696 if (!last_in_batch && cmd->queue->send_list_len)
697 msg.msg_flags |= MSG_MORE;
699 msg.msg_flags |= MSG_EOR;
701 bvec_set_virt(&bvec, (void *)cmd->r2t_pdu + cmd->offset, left);
702 iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, &bvec, 1, left);
703 ret = sock_sendmsg(cmd->queue->sock, &msg);
712 cmd->queue->snd_cmd = NULL;
716 static int nvmet_try_send_ddgst(struct nvmet_tcp_cmd *cmd, bool last_in_batch)
718 struct nvmet_tcp_queue *queue = cmd->queue;
719 int left = NVME_TCP_DIGEST_LENGTH - cmd->offset;
720 struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
722 .iov_base = (u8 *)&cmd->exp_ddgst + cmd->offset,
727 if (!last_in_batch && cmd->queue->send_list_len)
728 msg.msg_flags |= MSG_MORE;
730 msg.msg_flags |= MSG_EOR;
732 ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
733 if (unlikely(ret <= 0))
742 if (queue->nvme_sq.sqhd_disabled) {
743 cmd->queue->snd_cmd = NULL;
744 nvmet_tcp_put_cmd(cmd);
746 nvmet_setup_response_pdu(cmd);
751 static int nvmet_tcp_try_send_one(struct nvmet_tcp_queue *queue,
754 struct nvmet_tcp_cmd *cmd = queue->snd_cmd;
757 if (!cmd || queue->state == NVMET_TCP_Q_DISCONNECTING) {
758 cmd = nvmet_tcp_fetch_cmd(queue);
763 if (cmd->state == NVMET_TCP_SEND_DATA_PDU) {
764 ret = nvmet_try_send_data_pdu(cmd);
769 if (cmd->state == NVMET_TCP_SEND_DATA) {
770 ret = nvmet_try_send_data(cmd, last_in_batch);
775 if (cmd->state == NVMET_TCP_SEND_DDGST) {
776 ret = nvmet_try_send_ddgst(cmd, last_in_batch);
781 if (cmd->state == NVMET_TCP_SEND_R2T) {
782 ret = nvmet_try_send_r2t(cmd, last_in_batch);
787 if (cmd->state == NVMET_TCP_SEND_RESPONSE)
788 ret = nvmet_try_send_response(cmd, last_in_batch);
800 static int nvmet_tcp_try_send(struct nvmet_tcp_queue *queue,
801 int budget, int *sends)
805 for (i = 0; i < budget; i++) {
806 ret = nvmet_tcp_try_send_one(queue, i == budget - 1);
807 if (unlikely(ret < 0)) {
808 nvmet_tcp_socket_error(queue, ret);
810 } else if (ret == 0) {
819 static void nvmet_prepare_receive_pdu(struct nvmet_tcp_queue *queue)
822 queue->left = sizeof(struct nvme_tcp_hdr);
824 queue->rcv_state = NVMET_TCP_RECV_PDU;
827 static void nvmet_tcp_free_crypto(struct nvmet_tcp_queue *queue)
829 struct crypto_ahash *tfm = crypto_ahash_reqtfm(queue->rcv_hash);
831 ahash_request_free(queue->rcv_hash);
832 ahash_request_free(queue->snd_hash);
833 crypto_free_ahash(tfm);
836 static int nvmet_tcp_alloc_crypto(struct nvmet_tcp_queue *queue)
838 struct crypto_ahash *tfm;
840 tfm = crypto_alloc_ahash("crc32c", 0, CRYPTO_ALG_ASYNC);
844 queue->snd_hash = ahash_request_alloc(tfm, GFP_KERNEL);
845 if (!queue->snd_hash)
847 ahash_request_set_callback(queue->snd_hash, 0, NULL, NULL);
849 queue->rcv_hash = ahash_request_alloc(tfm, GFP_KERNEL);
850 if (!queue->rcv_hash)
852 ahash_request_set_callback(queue->rcv_hash, 0, NULL, NULL);
856 ahash_request_free(queue->snd_hash);
858 crypto_free_ahash(tfm);
863 static int nvmet_tcp_handle_icreq(struct nvmet_tcp_queue *queue)
865 struct nvme_tcp_icreq_pdu *icreq = &queue->pdu.icreq;
866 struct nvme_tcp_icresp_pdu *icresp = &queue->pdu.icresp;
867 struct msghdr msg = {};
871 if (le32_to_cpu(icreq->hdr.plen) != sizeof(struct nvme_tcp_icreq_pdu)) {
872 pr_err("bad nvme-tcp pdu length (%d)\n",
873 le32_to_cpu(icreq->hdr.plen));
874 nvmet_tcp_fatal_error(queue);
877 if (icreq->pfv != NVME_TCP_PFV_1_0) {
878 pr_err("queue %d: bad pfv %d\n", queue->idx, icreq->pfv);
882 if (icreq->hpda != 0) {
883 pr_err("queue %d: unsupported hpda %d\n", queue->idx,
888 queue->hdr_digest = !!(icreq->digest & NVME_TCP_HDR_DIGEST_ENABLE);
889 queue->data_digest = !!(icreq->digest & NVME_TCP_DATA_DIGEST_ENABLE);
890 if (queue->hdr_digest || queue->data_digest) {
891 ret = nvmet_tcp_alloc_crypto(queue);
896 memset(icresp, 0, sizeof(*icresp));
897 icresp->hdr.type = nvme_tcp_icresp;
898 icresp->hdr.hlen = sizeof(*icresp);
900 icresp->hdr.plen = cpu_to_le32(icresp->hdr.hlen);
901 icresp->pfv = cpu_to_le16(NVME_TCP_PFV_1_0);
902 icresp->maxdata = cpu_to_le32(0x400000); /* 16M arbitrary limit */
904 if (queue->hdr_digest)
905 icresp->digest |= NVME_TCP_HDR_DIGEST_ENABLE;
906 if (queue->data_digest)
907 icresp->digest |= NVME_TCP_DATA_DIGEST_ENABLE;
909 iov.iov_base = icresp;
910 iov.iov_len = sizeof(*icresp);
911 ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
915 queue->state = NVMET_TCP_Q_LIVE;
916 nvmet_prepare_receive_pdu(queue);
919 if (queue->hdr_digest || queue->data_digest)
920 nvmet_tcp_free_crypto(queue);
924 static void nvmet_tcp_handle_req_failure(struct nvmet_tcp_queue *queue,
925 struct nvmet_tcp_cmd *cmd, struct nvmet_req *req)
927 size_t data_len = le32_to_cpu(req->cmd->common.dptr.sgl.length);
931 * This command has not been processed yet, hence we are trying to
932 * figure out if there is still pending data left to receive. If
933 * we don't, we can simply prepare for the next pdu and bail out,
934 * otherwise we will need to prepare a buffer and receive the
935 * stale data before continuing forward.
937 if (!nvme_is_write(cmd->req.cmd) || !data_len ||
938 data_len > cmd->req.port->inline_data_size) {
939 nvmet_prepare_receive_pdu(queue);
943 ret = nvmet_tcp_map_data(cmd);
945 pr_err("queue %d: failed to map data\n", queue->idx);
946 nvmet_tcp_fatal_error(queue);
950 queue->rcv_state = NVMET_TCP_RECV_DATA;
951 nvmet_tcp_build_pdu_iovec(cmd);
952 cmd->flags |= NVMET_TCP_F_INIT_FAILED;
955 static int nvmet_tcp_handle_h2c_data_pdu(struct nvmet_tcp_queue *queue)
957 struct nvme_tcp_data_pdu *data = &queue->pdu.data;
958 struct nvmet_tcp_cmd *cmd;
960 if (likely(queue->nr_cmds)) {
961 if (unlikely(data->ttag >= queue->nr_cmds)) {
962 pr_err("queue %d: received out of bound ttag %u, nr_cmds %u\n",
963 queue->idx, data->ttag, queue->nr_cmds);
964 nvmet_tcp_fatal_error(queue);
967 cmd = &queue->cmds[data->ttag];
969 cmd = &queue->connect;
972 if (le32_to_cpu(data->data_offset) != cmd->rbytes_done) {
973 pr_err("ttag %u unexpected data offset %u (expected %u)\n",
974 data->ttag, le32_to_cpu(data->data_offset),
976 /* FIXME: use path and transport errors */
977 nvmet_req_complete(&cmd->req,
978 NVME_SC_INVALID_FIELD | NVME_SC_DNR);
982 cmd->pdu_len = le32_to_cpu(data->data_length);
984 nvmet_tcp_build_pdu_iovec(cmd);
986 queue->rcv_state = NVMET_TCP_RECV_DATA;
991 static int nvmet_tcp_done_recv_pdu(struct nvmet_tcp_queue *queue)
993 struct nvme_tcp_hdr *hdr = &queue->pdu.cmd.hdr;
994 struct nvme_command *nvme_cmd = &queue->pdu.cmd.cmd;
995 struct nvmet_req *req;
998 if (unlikely(queue->state == NVMET_TCP_Q_CONNECTING)) {
999 if (hdr->type != nvme_tcp_icreq) {
1000 pr_err("unexpected pdu type (%d) before icreq\n",
1002 nvmet_tcp_fatal_error(queue);
1005 return nvmet_tcp_handle_icreq(queue);
1008 if (unlikely(hdr->type == nvme_tcp_icreq)) {
1009 pr_err("queue %d: received icreq pdu in state %d\n",
1010 queue->idx, queue->state);
1011 nvmet_tcp_fatal_error(queue);
1015 if (hdr->type == nvme_tcp_h2c_data) {
1016 ret = nvmet_tcp_handle_h2c_data_pdu(queue);
1022 queue->cmd = nvmet_tcp_get_cmd(queue);
1023 if (unlikely(!queue->cmd)) {
1024 /* This should never happen */
1025 pr_err("queue %d: out of commands (%d) send_list_len: %d, opcode: %d",
1026 queue->idx, queue->nr_cmds, queue->send_list_len,
1027 nvme_cmd->common.opcode);
1028 nvmet_tcp_fatal_error(queue);
1032 req = &queue->cmd->req;
1033 memcpy(req->cmd, nvme_cmd, sizeof(*nvme_cmd));
1035 if (unlikely(!nvmet_req_init(req, &queue->nvme_cq,
1036 &queue->nvme_sq, &nvmet_tcp_ops))) {
1037 pr_err("failed cmd %p id %d opcode %d, data_len: %d\n",
1038 req->cmd, req->cmd->common.command_id,
1039 req->cmd->common.opcode,
1040 le32_to_cpu(req->cmd->common.dptr.sgl.length));
1042 nvmet_tcp_handle_req_failure(queue, queue->cmd, req);
1046 ret = nvmet_tcp_map_data(queue->cmd);
1047 if (unlikely(ret)) {
1048 pr_err("queue %d: failed to map data\n", queue->idx);
1049 if (nvmet_tcp_has_inline_data(queue->cmd))
1050 nvmet_tcp_fatal_error(queue);
1052 nvmet_req_complete(req, ret);
1057 if (nvmet_tcp_need_data_in(queue->cmd)) {
1058 if (nvmet_tcp_has_inline_data(queue->cmd)) {
1059 queue->rcv_state = NVMET_TCP_RECV_DATA;
1060 nvmet_tcp_build_pdu_iovec(queue->cmd);
1064 nvmet_tcp_queue_response(&queue->cmd->req);
1068 queue->cmd->req.execute(&queue->cmd->req);
1070 nvmet_prepare_receive_pdu(queue);
1074 static const u8 nvme_tcp_pdu_sizes[] = {
1075 [nvme_tcp_icreq] = sizeof(struct nvme_tcp_icreq_pdu),
1076 [nvme_tcp_cmd] = sizeof(struct nvme_tcp_cmd_pdu),
1077 [nvme_tcp_h2c_data] = sizeof(struct nvme_tcp_data_pdu),
1080 static inline u8 nvmet_tcp_pdu_size(u8 type)
1084 return (idx < ARRAY_SIZE(nvme_tcp_pdu_sizes) &&
1085 nvme_tcp_pdu_sizes[idx]) ?
1086 nvme_tcp_pdu_sizes[idx] : 0;
1089 static inline bool nvmet_tcp_pdu_valid(u8 type)
1092 case nvme_tcp_icreq:
1094 case nvme_tcp_h2c_data:
1102 static int nvmet_tcp_try_recv_pdu(struct nvmet_tcp_queue *queue)
1104 struct nvme_tcp_hdr *hdr = &queue->pdu.cmd.hdr;
1107 struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
1110 iov.iov_base = (void *)&queue->pdu + queue->offset;
1111 iov.iov_len = queue->left;
1112 len = kernel_recvmsg(queue->sock, &msg, &iov, 1,
1113 iov.iov_len, msg.msg_flags);
1114 if (unlikely(len < 0))
1117 queue->offset += len;
1122 if (queue->offset == sizeof(struct nvme_tcp_hdr)) {
1123 u8 hdgst = nvmet_tcp_hdgst_len(queue);
1125 if (unlikely(!nvmet_tcp_pdu_valid(hdr->type))) {
1126 pr_err("unexpected pdu type %d\n", hdr->type);
1127 nvmet_tcp_fatal_error(queue);
1131 if (unlikely(hdr->hlen != nvmet_tcp_pdu_size(hdr->type))) {
1132 pr_err("pdu %d bad hlen %d\n", hdr->type, hdr->hlen);
1136 queue->left = hdr->hlen - queue->offset + hdgst;
1140 if (queue->hdr_digest &&
1141 nvmet_tcp_verify_hdgst(queue, &queue->pdu, hdr->hlen)) {
1142 nvmet_tcp_fatal_error(queue); /* fatal */
1146 if (queue->data_digest &&
1147 nvmet_tcp_check_ddgst(queue, &queue->pdu)) {
1148 nvmet_tcp_fatal_error(queue); /* fatal */
1152 return nvmet_tcp_done_recv_pdu(queue);
1155 static void nvmet_tcp_prep_recv_ddgst(struct nvmet_tcp_cmd *cmd)
1157 struct nvmet_tcp_queue *queue = cmd->queue;
1159 nvmet_tcp_calc_ddgst(queue->rcv_hash, cmd);
1161 queue->left = NVME_TCP_DIGEST_LENGTH;
1162 queue->rcv_state = NVMET_TCP_RECV_DDGST;
1165 static int nvmet_tcp_try_recv_data(struct nvmet_tcp_queue *queue)
1167 struct nvmet_tcp_cmd *cmd = queue->cmd;
1170 while (msg_data_left(&cmd->recv_msg)) {
1171 ret = sock_recvmsg(cmd->queue->sock, &cmd->recv_msg,
1172 cmd->recv_msg.msg_flags);
1176 cmd->pdu_recv += ret;
1177 cmd->rbytes_done += ret;
1180 if (queue->data_digest) {
1181 nvmet_tcp_prep_recv_ddgst(cmd);
1185 if (cmd->rbytes_done == cmd->req.transfer_len)
1186 nvmet_tcp_execute_request(cmd);
1188 nvmet_prepare_receive_pdu(queue);
1192 static int nvmet_tcp_try_recv_ddgst(struct nvmet_tcp_queue *queue)
1194 struct nvmet_tcp_cmd *cmd = queue->cmd;
1196 struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
1198 .iov_base = (void *)&cmd->recv_ddgst + queue->offset,
1199 .iov_len = queue->left
1202 ret = kernel_recvmsg(queue->sock, &msg, &iov, 1,
1203 iov.iov_len, msg.msg_flags);
1204 if (unlikely(ret < 0))
1207 queue->offset += ret;
1212 if (queue->data_digest && cmd->exp_ddgst != cmd->recv_ddgst) {
1213 pr_err("queue %d: cmd %d pdu (%d) data digest error: recv %#x expected %#x\n",
1214 queue->idx, cmd->req.cmd->common.command_id,
1215 queue->pdu.cmd.hdr.type, le32_to_cpu(cmd->recv_ddgst),
1216 le32_to_cpu(cmd->exp_ddgst));
1217 nvmet_req_uninit(&cmd->req);
1218 nvmet_tcp_free_cmd_buffers(cmd);
1219 nvmet_tcp_fatal_error(queue);
1224 if (cmd->rbytes_done == cmd->req.transfer_len)
1225 nvmet_tcp_execute_request(cmd);
1229 nvmet_prepare_receive_pdu(queue);
1233 static int nvmet_tcp_try_recv_one(struct nvmet_tcp_queue *queue)
1237 if (unlikely(queue->rcv_state == NVMET_TCP_RECV_ERR))
1240 if (queue->rcv_state == NVMET_TCP_RECV_PDU) {
1241 result = nvmet_tcp_try_recv_pdu(queue);
1246 if (queue->rcv_state == NVMET_TCP_RECV_DATA) {
1247 result = nvmet_tcp_try_recv_data(queue);
1252 if (queue->rcv_state == NVMET_TCP_RECV_DDGST) {
1253 result = nvmet_tcp_try_recv_ddgst(queue);
1260 if (result == -EAGAIN)
1267 static int nvmet_tcp_try_recv(struct nvmet_tcp_queue *queue,
1268 int budget, int *recvs)
1272 for (i = 0; i < budget; i++) {
1273 ret = nvmet_tcp_try_recv_one(queue);
1274 if (unlikely(ret < 0)) {
1275 nvmet_tcp_socket_error(queue, ret);
1277 } else if (ret == 0) {
1286 static void nvmet_tcp_schedule_release_queue(struct nvmet_tcp_queue *queue)
1288 spin_lock(&queue->state_lock);
1289 if (queue->state != NVMET_TCP_Q_DISCONNECTING) {
1290 queue->state = NVMET_TCP_Q_DISCONNECTING;
1291 queue_work(nvmet_wq, &queue->release_work);
1293 spin_unlock(&queue->state_lock);
1296 static inline void nvmet_tcp_arm_queue_deadline(struct nvmet_tcp_queue *queue)
1298 queue->poll_end = jiffies + usecs_to_jiffies(idle_poll_period_usecs);
1301 static bool nvmet_tcp_check_queue_deadline(struct nvmet_tcp_queue *queue,
1304 if (!idle_poll_period_usecs)
1308 nvmet_tcp_arm_queue_deadline(queue);
1310 return !time_after(jiffies, queue->poll_end);
1313 static void nvmet_tcp_io_work(struct work_struct *w)
1315 struct nvmet_tcp_queue *queue =
1316 container_of(w, struct nvmet_tcp_queue, io_work);
1323 ret = nvmet_tcp_try_recv(queue, NVMET_TCP_RECV_BUDGET, &ops);
1329 ret = nvmet_tcp_try_send(queue, NVMET_TCP_SEND_BUDGET, &ops);
1335 } while (pending && ops < NVMET_TCP_IO_WORK_BUDGET);
1338 * Requeue the worker if idle deadline period is in progress or any
1339 * ops activity was recorded during the do-while loop above.
1341 if (nvmet_tcp_check_queue_deadline(queue, ops) || pending)
1342 queue_work_on(queue_cpu(queue), nvmet_tcp_wq, &queue->io_work);
1345 static int nvmet_tcp_alloc_cmd(struct nvmet_tcp_queue *queue,
1346 struct nvmet_tcp_cmd *c)
1348 u8 hdgst = nvmet_tcp_hdgst_len(queue);
1351 c->req.port = queue->port->nport;
1353 c->cmd_pdu = page_frag_alloc(&queue->pf_cache,
1354 sizeof(*c->cmd_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO);
1357 c->req.cmd = &c->cmd_pdu->cmd;
1359 c->rsp_pdu = page_frag_alloc(&queue->pf_cache,
1360 sizeof(*c->rsp_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO);
1363 c->req.cqe = &c->rsp_pdu->cqe;
1365 c->data_pdu = page_frag_alloc(&queue->pf_cache,
1366 sizeof(*c->data_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO);
1370 c->r2t_pdu = page_frag_alloc(&queue->pf_cache,
1371 sizeof(*c->r2t_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO);
1375 c->recv_msg.msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
1377 list_add_tail(&c->entry, &queue->free_list);
1381 page_frag_free(c->data_pdu);
1383 page_frag_free(c->rsp_pdu);
1385 page_frag_free(c->cmd_pdu);
1389 static void nvmet_tcp_free_cmd(struct nvmet_tcp_cmd *c)
1391 page_frag_free(c->r2t_pdu);
1392 page_frag_free(c->data_pdu);
1393 page_frag_free(c->rsp_pdu);
1394 page_frag_free(c->cmd_pdu);
1397 static int nvmet_tcp_alloc_cmds(struct nvmet_tcp_queue *queue)
1399 struct nvmet_tcp_cmd *cmds;
1400 int i, ret = -EINVAL, nr_cmds = queue->nr_cmds;
1402 cmds = kcalloc(nr_cmds, sizeof(struct nvmet_tcp_cmd), GFP_KERNEL);
1406 for (i = 0; i < nr_cmds; i++) {
1407 ret = nvmet_tcp_alloc_cmd(queue, cmds + i);
1417 nvmet_tcp_free_cmd(cmds + i);
1423 static void nvmet_tcp_free_cmds(struct nvmet_tcp_queue *queue)
1425 struct nvmet_tcp_cmd *cmds = queue->cmds;
1428 for (i = 0; i < queue->nr_cmds; i++)
1429 nvmet_tcp_free_cmd(cmds + i);
1431 nvmet_tcp_free_cmd(&queue->connect);
1435 static void nvmet_tcp_restore_socket_callbacks(struct nvmet_tcp_queue *queue)
1437 struct socket *sock = queue->sock;
1439 write_lock_bh(&sock->sk->sk_callback_lock);
1440 sock->sk->sk_data_ready = queue->data_ready;
1441 sock->sk->sk_state_change = queue->state_change;
1442 sock->sk->sk_write_space = queue->write_space;
1443 sock->sk->sk_user_data = NULL;
1444 write_unlock_bh(&sock->sk->sk_callback_lock);
1447 static void nvmet_tcp_uninit_data_in_cmds(struct nvmet_tcp_queue *queue)
1449 struct nvmet_tcp_cmd *cmd = queue->cmds;
1452 for (i = 0; i < queue->nr_cmds; i++, cmd++) {
1453 if (nvmet_tcp_need_data_in(cmd))
1454 nvmet_req_uninit(&cmd->req);
1457 if (!queue->nr_cmds && nvmet_tcp_need_data_in(&queue->connect)) {
1458 /* failed in connect */
1459 nvmet_req_uninit(&queue->connect.req);
1463 static void nvmet_tcp_free_cmd_data_in_buffers(struct nvmet_tcp_queue *queue)
1465 struct nvmet_tcp_cmd *cmd = queue->cmds;
1468 for (i = 0; i < queue->nr_cmds; i++, cmd++) {
1469 if (nvmet_tcp_need_data_in(cmd))
1470 nvmet_tcp_free_cmd_buffers(cmd);
1473 if (!queue->nr_cmds && nvmet_tcp_need_data_in(&queue->connect))
1474 nvmet_tcp_free_cmd_buffers(&queue->connect);
1477 static void nvmet_tcp_release_queue_work(struct work_struct *w)
1480 struct nvmet_tcp_queue *queue =
1481 container_of(w, struct nvmet_tcp_queue, release_work);
1483 mutex_lock(&nvmet_tcp_queue_mutex);
1484 list_del_init(&queue->queue_list);
1485 mutex_unlock(&nvmet_tcp_queue_mutex);
1487 nvmet_tcp_restore_socket_callbacks(queue);
1488 cancel_work_sync(&queue->io_work);
1489 /* stop accepting incoming data */
1490 queue->rcv_state = NVMET_TCP_RECV_ERR;
1492 nvmet_tcp_uninit_data_in_cmds(queue);
1493 nvmet_sq_destroy(&queue->nvme_sq);
1494 cancel_work_sync(&queue->io_work);
1495 nvmet_tcp_free_cmd_data_in_buffers(queue);
1496 sock_release(queue->sock);
1497 nvmet_tcp_free_cmds(queue);
1498 if (queue->hdr_digest || queue->data_digest)
1499 nvmet_tcp_free_crypto(queue);
1500 ida_free(&nvmet_tcp_queue_ida, queue->idx);
1502 page = virt_to_head_page(queue->pf_cache.va);
1503 __page_frag_cache_drain(page, queue->pf_cache.pagecnt_bias);
1507 static void nvmet_tcp_data_ready(struct sock *sk)
1509 struct nvmet_tcp_queue *queue;
1511 trace_sk_data_ready(sk);
1513 read_lock_bh(&sk->sk_callback_lock);
1514 queue = sk->sk_user_data;
1516 queue_work_on(queue_cpu(queue), nvmet_tcp_wq, &queue->io_work);
1517 read_unlock_bh(&sk->sk_callback_lock);
1520 static void nvmet_tcp_write_space(struct sock *sk)
1522 struct nvmet_tcp_queue *queue;
1524 read_lock_bh(&sk->sk_callback_lock);
1525 queue = sk->sk_user_data;
1526 if (unlikely(!queue))
1529 if (unlikely(queue->state == NVMET_TCP_Q_CONNECTING)) {
1530 queue->write_space(sk);
1534 if (sk_stream_is_writeable(sk)) {
1535 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1536 queue_work_on(queue_cpu(queue), nvmet_tcp_wq, &queue->io_work);
1539 read_unlock_bh(&sk->sk_callback_lock);
1542 static void nvmet_tcp_state_change(struct sock *sk)
1544 struct nvmet_tcp_queue *queue;
1546 read_lock_bh(&sk->sk_callback_lock);
1547 queue = sk->sk_user_data;
1551 switch (sk->sk_state) {
1556 case TCP_CLOSE_WAIT:
1559 nvmet_tcp_schedule_release_queue(queue);
1562 pr_warn("queue %d unhandled state %d\n",
1563 queue->idx, sk->sk_state);
1566 read_unlock_bh(&sk->sk_callback_lock);
1569 static int nvmet_tcp_set_queue_sock(struct nvmet_tcp_queue *queue)
1571 struct socket *sock = queue->sock;
1572 struct inet_sock *inet = inet_sk(sock->sk);
1575 ret = kernel_getsockname(sock,
1576 (struct sockaddr *)&queue->sockaddr);
1580 ret = kernel_getpeername(sock,
1581 (struct sockaddr *)&queue->sockaddr_peer);
1586 * Cleanup whatever is sitting in the TCP transmit queue on socket
1587 * close. This is done to prevent stale data from being sent should
1588 * the network connection be restored before TCP times out.
1590 sock_no_linger(sock->sk);
1592 if (so_priority > 0)
1593 sock_set_priority(sock->sk, so_priority);
1595 /* Set socket type of service */
1596 if (inet->rcv_tos > 0)
1597 ip_sock_set_tos(sock->sk, inet->rcv_tos);
1600 write_lock_bh(&sock->sk->sk_callback_lock);
1601 if (sock->sk->sk_state != TCP_ESTABLISHED) {
1603 * If the socket is already closing, don't even start
1608 sock->sk->sk_user_data = queue;
1609 queue->data_ready = sock->sk->sk_data_ready;
1610 sock->sk->sk_data_ready = nvmet_tcp_data_ready;
1611 queue->state_change = sock->sk->sk_state_change;
1612 sock->sk->sk_state_change = nvmet_tcp_state_change;
1613 queue->write_space = sock->sk->sk_write_space;
1614 sock->sk->sk_write_space = nvmet_tcp_write_space;
1615 if (idle_poll_period_usecs)
1616 nvmet_tcp_arm_queue_deadline(queue);
1617 queue_work_on(queue_cpu(queue), nvmet_tcp_wq, &queue->io_work);
1619 write_unlock_bh(&sock->sk->sk_callback_lock);
1624 static int nvmet_tcp_alloc_queue(struct nvmet_tcp_port *port,
1625 struct socket *newsock)
1627 struct nvmet_tcp_queue *queue;
1630 queue = kzalloc(sizeof(*queue), GFP_KERNEL);
1634 INIT_WORK(&queue->release_work, nvmet_tcp_release_queue_work);
1635 INIT_WORK(&queue->io_work, nvmet_tcp_io_work);
1636 queue->sock = newsock;
1639 spin_lock_init(&queue->state_lock);
1640 queue->state = NVMET_TCP_Q_CONNECTING;
1641 INIT_LIST_HEAD(&queue->free_list);
1642 init_llist_head(&queue->resp_list);
1643 INIT_LIST_HEAD(&queue->resp_send_list);
1645 queue->idx = ida_alloc(&nvmet_tcp_queue_ida, GFP_KERNEL);
1646 if (queue->idx < 0) {
1648 goto out_free_queue;
1651 ret = nvmet_tcp_alloc_cmd(queue, &queue->connect);
1653 goto out_ida_remove;
1655 ret = nvmet_sq_init(&queue->nvme_sq);
1657 goto out_free_connect;
1659 nvmet_prepare_receive_pdu(queue);
1661 mutex_lock(&nvmet_tcp_queue_mutex);
1662 list_add_tail(&queue->queue_list, &nvmet_tcp_queue_list);
1663 mutex_unlock(&nvmet_tcp_queue_mutex);
1665 ret = nvmet_tcp_set_queue_sock(queue);
1667 goto out_destroy_sq;
1671 mutex_lock(&nvmet_tcp_queue_mutex);
1672 list_del_init(&queue->queue_list);
1673 mutex_unlock(&nvmet_tcp_queue_mutex);
1674 nvmet_sq_destroy(&queue->nvme_sq);
1676 nvmet_tcp_free_cmd(&queue->connect);
1678 ida_free(&nvmet_tcp_queue_ida, queue->idx);
1684 static void nvmet_tcp_accept_work(struct work_struct *w)
1686 struct nvmet_tcp_port *port =
1687 container_of(w, struct nvmet_tcp_port, accept_work);
1688 struct socket *newsock;
1692 ret = kernel_accept(port->sock, &newsock, O_NONBLOCK);
1695 pr_warn("failed to accept err=%d\n", ret);
1698 ret = nvmet_tcp_alloc_queue(port, newsock);
1700 pr_err("failed to allocate queue\n");
1701 sock_release(newsock);
1706 static void nvmet_tcp_listen_data_ready(struct sock *sk)
1708 struct nvmet_tcp_port *port;
1710 trace_sk_data_ready(sk);
1712 read_lock_bh(&sk->sk_callback_lock);
1713 port = sk->sk_user_data;
1717 if (sk->sk_state == TCP_LISTEN)
1718 queue_work(nvmet_wq, &port->accept_work);
1720 read_unlock_bh(&sk->sk_callback_lock);
1723 static int nvmet_tcp_add_port(struct nvmet_port *nport)
1725 struct nvmet_tcp_port *port;
1726 __kernel_sa_family_t af;
1729 port = kzalloc(sizeof(*port), GFP_KERNEL);
1733 switch (nport->disc_addr.adrfam) {
1734 case NVMF_ADDR_FAMILY_IP4:
1737 case NVMF_ADDR_FAMILY_IP6:
1741 pr_err("address family %d not supported\n",
1742 nport->disc_addr.adrfam);
1747 ret = inet_pton_with_scope(&init_net, af, nport->disc_addr.traddr,
1748 nport->disc_addr.trsvcid, &port->addr);
1750 pr_err("malformed ip/port passed: %s:%s\n",
1751 nport->disc_addr.traddr, nport->disc_addr.trsvcid);
1755 port->nport = nport;
1756 INIT_WORK(&port->accept_work, nvmet_tcp_accept_work);
1757 if (port->nport->inline_data_size < 0)
1758 port->nport->inline_data_size = NVMET_TCP_DEF_INLINE_DATA_SIZE;
1760 ret = sock_create(port->addr.ss_family, SOCK_STREAM,
1761 IPPROTO_TCP, &port->sock);
1763 pr_err("failed to create a socket\n");
1767 port->sock->sk->sk_user_data = port;
1768 port->data_ready = port->sock->sk->sk_data_ready;
1769 port->sock->sk->sk_data_ready = nvmet_tcp_listen_data_ready;
1770 sock_set_reuseaddr(port->sock->sk);
1771 tcp_sock_set_nodelay(port->sock->sk);
1772 if (so_priority > 0)
1773 sock_set_priority(port->sock->sk, so_priority);
1775 ret = kernel_bind(port->sock, (struct sockaddr *)&port->addr,
1776 sizeof(port->addr));
1778 pr_err("failed to bind port socket %d\n", ret);
1782 ret = kernel_listen(port->sock, 128);
1784 pr_err("failed to listen %d on port sock\n", ret);
1789 pr_info("enabling port %d (%pISpc)\n",
1790 le16_to_cpu(nport->disc_addr.portid), &port->addr);
1795 sock_release(port->sock);
1801 static void nvmet_tcp_destroy_port_queues(struct nvmet_tcp_port *port)
1803 struct nvmet_tcp_queue *queue;
1805 mutex_lock(&nvmet_tcp_queue_mutex);
1806 list_for_each_entry(queue, &nvmet_tcp_queue_list, queue_list)
1807 if (queue->port == port)
1808 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1809 mutex_unlock(&nvmet_tcp_queue_mutex);
1812 static void nvmet_tcp_remove_port(struct nvmet_port *nport)
1814 struct nvmet_tcp_port *port = nport->priv;
1816 write_lock_bh(&port->sock->sk->sk_callback_lock);
1817 port->sock->sk->sk_data_ready = port->data_ready;
1818 port->sock->sk->sk_user_data = NULL;
1819 write_unlock_bh(&port->sock->sk->sk_callback_lock);
1820 cancel_work_sync(&port->accept_work);
1822 * Destroy the remaining queues, which are not belong to any
1825 nvmet_tcp_destroy_port_queues(port);
1827 sock_release(port->sock);
1831 static void nvmet_tcp_delete_ctrl(struct nvmet_ctrl *ctrl)
1833 struct nvmet_tcp_queue *queue;
1835 mutex_lock(&nvmet_tcp_queue_mutex);
1836 list_for_each_entry(queue, &nvmet_tcp_queue_list, queue_list)
1837 if (queue->nvme_sq.ctrl == ctrl)
1838 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1839 mutex_unlock(&nvmet_tcp_queue_mutex);
1842 static u16 nvmet_tcp_install_queue(struct nvmet_sq *sq)
1844 struct nvmet_tcp_queue *queue =
1845 container_of(sq, struct nvmet_tcp_queue, nvme_sq);
1848 /* Let inflight controller teardown complete */
1849 flush_workqueue(nvmet_wq);
1852 queue->nr_cmds = sq->size * 2;
1853 if (nvmet_tcp_alloc_cmds(queue))
1854 return NVME_SC_INTERNAL;
1858 static void nvmet_tcp_disc_port_addr(struct nvmet_req *req,
1859 struct nvmet_port *nport, char *traddr)
1861 struct nvmet_tcp_port *port = nport->priv;
1863 if (inet_addr_is_any((struct sockaddr *)&port->addr)) {
1864 struct nvmet_tcp_cmd *cmd =
1865 container_of(req, struct nvmet_tcp_cmd, req);
1866 struct nvmet_tcp_queue *queue = cmd->queue;
1868 sprintf(traddr, "%pISc", (struct sockaddr *)&queue->sockaddr);
1870 memcpy(traddr, nport->disc_addr.traddr, NVMF_TRADDR_SIZE);
1874 static const struct nvmet_fabrics_ops nvmet_tcp_ops = {
1875 .owner = THIS_MODULE,
1876 .type = NVMF_TRTYPE_TCP,
1878 .add_port = nvmet_tcp_add_port,
1879 .remove_port = nvmet_tcp_remove_port,
1880 .queue_response = nvmet_tcp_queue_response,
1881 .delete_ctrl = nvmet_tcp_delete_ctrl,
1882 .install_queue = nvmet_tcp_install_queue,
1883 .disc_traddr = nvmet_tcp_disc_port_addr,
1886 static int __init nvmet_tcp_init(void)
1890 nvmet_tcp_wq = alloc_workqueue("nvmet_tcp_wq",
1891 WQ_MEM_RECLAIM | WQ_HIGHPRI, 0);
1895 ret = nvmet_register_transport(&nvmet_tcp_ops);
1901 destroy_workqueue(nvmet_tcp_wq);
1905 static void __exit nvmet_tcp_exit(void)
1907 struct nvmet_tcp_queue *queue;
1909 nvmet_unregister_transport(&nvmet_tcp_ops);
1911 flush_workqueue(nvmet_wq);
1912 mutex_lock(&nvmet_tcp_queue_mutex);
1913 list_for_each_entry(queue, &nvmet_tcp_queue_list, queue_list)
1914 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1915 mutex_unlock(&nvmet_tcp_queue_mutex);
1916 flush_workqueue(nvmet_wq);
1918 destroy_workqueue(nvmet_tcp_wq);
1921 module_init(nvmet_tcp_init);
1922 module_exit(nvmet_tcp_exit);
1924 MODULE_LICENSE("GPL v2");
1925 MODULE_ALIAS("nvmet-transport-3"); /* 3 == NVMF_TRTYPE_TCP */