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>
20 #define NVMET_TCP_DEF_INLINE_DATA_SIZE (4 * PAGE_SIZE)
22 #define NVMET_TCP_RECV_BUDGET 8
23 #define NVMET_TCP_SEND_BUDGET 8
24 #define NVMET_TCP_IO_WORK_BUDGET 64
26 enum nvmet_tcp_send_state {
27 NVMET_TCP_SEND_DATA_PDU,
31 NVMET_TCP_SEND_RESPONSE
34 enum nvmet_tcp_recv_state {
42 NVMET_TCP_F_INIT_FAILED = (1 << 0),
45 struct nvmet_tcp_cmd {
46 struct nvmet_tcp_queue *queue;
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;
61 struct msghdr recv_msg;
65 struct list_head entry;
66 struct llist_node lentry;
70 struct scatterlist *cur_sg;
71 enum nvmet_tcp_send_state state;
77 enum nvmet_tcp_queue_state {
78 NVMET_TCP_Q_CONNECTING,
80 NVMET_TCP_Q_DISCONNECTING,
83 struct nvmet_tcp_queue {
85 struct nvmet_tcp_port *port;
86 struct work_struct io_work;
88 struct nvmet_cq nvme_cq;
89 struct nvmet_sq nvme_sq;
92 struct nvmet_tcp_cmd *cmds;
94 struct list_head free_list;
95 struct llist_head resp_list;
96 struct list_head resp_send_list;
98 struct nvmet_tcp_cmd *snd_cmd;
103 enum nvmet_tcp_recv_state rcv_state;
104 struct nvmet_tcp_cmd *cmd;
105 union nvme_tcp_pdu pdu;
110 struct ahash_request *snd_hash;
111 struct ahash_request *rcv_hash;
113 spinlock_t state_lock;
114 enum nvmet_tcp_queue_state state;
116 struct sockaddr_storage sockaddr;
117 struct sockaddr_storage sockaddr_peer;
118 struct work_struct release_work;
121 struct list_head queue_list;
123 struct nvmet_tcp_cmd connect;
125 struct page_frag_cache pf_cache;
127 void (*data_ready)(struct sock *);
128 void (*state_change)(struct sock *);
129 void (*write_space)(struct sock *);
132 struct nvmet_tcp_port {
134 struct work_struct accept_work;
135 struct nvmet_port *nport;
136 struct sockaddr_storage addr;
138 void (*data_ready)(struct sock *);
141 static DEFINE_IDA(nvmet_tcp_queue_ida);
142 static LIST_HEAD(nvmet_tcp_queue_list);
143 static DEFINE_MUTEX(nvmet_tcp_queue_mutex);
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);
150 static inline u16 nvmet_tcp_cmd_tag(struct nvmet_tcp_queue *queue,
151 struct nvmet_tcp_cmd *cmd)
153 return cmd - queue->cmds;
156 static inline bool nvmet_tcp_has_data_in(struct nvmet_tcp_cmd *cmd)
158 return nvme_is_write(cmd->req.cmd) &&
159 cmd->rbytes_done < cmd->req.transfer_len;
162 static inline bool nvmet_tcp_need_data_in(struct nvmet_tcp_cmd *cmd)
164 return nvmet_tcp_has_data_in(cmd) && !cmd->req.cqe->status;
167 static inline bool nvmet_tcp_need_data_out(struct nvmet_tcp_cmd *cmd)
169 return !nvme_is_write(cmd->req.cmd) &&
170 cmd->req.transfer_len > 0 &&
171 !cmd->req.cqe->status;
174 static inline bool nvmet_tcp_has_inline_data(struct nvmet_tcp_cmd *cmd)
176 return nvme_is_write(cmd->req.cmd) && cmd->pdu_len &&
180 static inline struct nvmet_tcp_cmd *
181 nvmet_tcp_get_cmd(struct nvmet_tcp_queue *queue)
183 struct nvmet_tcp_cmd *cmd;
185 cmd = list_first_entry_or_null(&queue->free_list,
186 struct nvmet_tcp_cmd, entry);
189 list_del_init(&cmd->entry);
191 cmd->rbytes_done = cmd->wbytes_done = 0;
199 static inline void nvmet_tcp_put_cmd(struct nvmet_tcp_cmd *cmd)
201 if (unlikely(cmd == &cmd->queue->connect))
204 list_add_tail(&cmd->entry, &cmd->queue->free_list);
207 static inline u8 nvmet_tcp_hdgst_len(struct nvmet_tcp_queue *queue)
209 return queue->hdr_digest ? NVME_TCP_DIGEST_LENGTH : 0;
212 static inline u8 nvmet_tcp_ddgst_len(struct nvmet_tcp_queue *queue)
214 return queue->data_digest ? NVME_TCP_DIGEST_LENGTH : 0;
217 static inline void nvmet_tcp_hdgst(struct ahash_request *hash,
218 void *pdu, size_t len)
220 struct scatterlist sg;
222 sg_init_one(&sg, pdu, len);
223 ahash_request_set_crypt(hash, &sg, pdu + len, len);
224 crypto_ahash_digest(hash);
227 static int nvmet_tcp_verify_hdgst(struct nvmet_tcp_queue *queue,
228 void *pdu, size_t len)
230 struct nvme_tcp_hdr *hdr = pdu;
234 if (unlikely(!(hdr->flags & NVME_TCP_F_HDGST))) {
235 pr_err("queue %d: header digest enabled but no header digest\n",
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));
253 static int nvmet_tcp_check_ddgst(struct nvmet_tcp_queue *queue, void *pdu)
255 struct nvme_tcp_hdr *hdr = pdu;
256 u8 digest_len = nvmet_tcp_hdgst_len(queue);
259 len = le32_to_cpu(hdr->plen) - hdr->hlen -
260 (hdr->flags & NVME_TCP_F_HDGST ? digest_len : 0);
262 if (unlikely(len && !(hdr->flags & NVME_TCP_F_DDGST))) {
263 pr_err("queue %d: data digest flag is cleared\n", queue->idx);
270 static void nvmet_tcp_unmap_pdu_iovec(struct nvmet_tcp_cmd *cmd)
272 struct scatterlist *sg;
275 sg = &cmd->req.sg[cmd->sg_idx];
277 for (i = 0; i < cmd->nr_mapped; i++)
278 kunmap(sg_page(&sg[i]));
281 static void nvmet_tcp_map_pdu_iovec(struct nvmet_tcp_cmd *cmd)
283 struct kvec *iov = cmd->iov;
284 struct scatterlist *sg;
285 u32 length, offset, sg_offset;
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];
295 u32 iov_len = min_t(u32, length, sg->length - sg_offset);
297 iov->iov_base = kmap(sg_page(sg)) + sg->offset + sg_offset;
298 iov->iov_len = iov_len;
305 iov_iter_kvec(&cmd->recv_msg.msg_iter, READ, cmd->iov,
306 cmd->nr_mapped, cmd->pdu_len);
309 static void nvmet_tcp_fatal_error(struct nvmet_tcp_queue *queue)
311 queue->rcv_state = NVMET_TCP_RECV_ERR;
312 if (queue->nvme_sq.ctrl)
313 nvmet_ctrl_fatal_error(queue->nvme_sq.ctrl);
315 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
318 static int nvmet_tcp_map_data(struct nvmet_tcp_cmd *cmd)
320 struct nvme_sgl_desc *sgl = &cmd->req.cmd->common.dptr.sgl;
321 u32 len = le32_to_cpu(sgl->length);
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;
331 if (len > cmd->req.port->inline_data_size)
332 return NVME_SC_SGL_INVALID_OFFSET | NVME_SC_DNR;
335 cmd->req.transfer_len += len;
337 cmd->req.sg = sgl_alloc(len, GFP_KERNEL, &cmd->req.sg_cnt);
339 return NVME_SC_INTERNAL;
340 cmd->cur_sg = cmd->req.sg;
342 if (nvmet_tcp_has_data_in(cmd)) {
343 cmd->iov = kmalloc_array(cmd->req.sg_cnt,
344 sizeof(*cmd->iov), GFP_KERNEL);
351 sgl_free(cmd->req.sg);
352 return NVME_SC_INTERNAL;
355 static void nvmet_tcp_ddgst(struct ahash_request *hash,
356 struct nvmet_tcp_cmd *cmd)
358 ahash_request_set_crypt(hash, cmd->req.sg,
359 (void *)&cmd->exp_ddgst, cmd->req.transfer_len);
360 crypto_ahash_digest(hash);
363 static void nvmet_setup_c2h_data_pdu(struct nvmet_tcp_cmd *cmd)
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);
371 cmd->state = NVMET_TCP_SEND_DATA_PDU;
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;
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);
385 if (queue->data_digest) {
386 pdu->hdr.flags |= NVME_TCP_F_DDGST;
387 nvmet_tcp_ddgst(queue->snd_hash, cmd);
390 if (cmd->queue->hdr_digest) {
391 pdu->hdr.flags |= NVME_TCP_F_HDGST;
392 nvmet_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
396 static void nvmet_setup_r2t_pdu(struct nvmet_tcp_cmd *cmd)
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);
403 cmd->state = NVMET_TCP_SEND_R2T;
405 pdu->hdr.type = nvme_tcp_r2t;
407 pdu->hdr.hlen = sizeof(*pdu);
409 pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst);
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));
421 static void nvmet_setup_response_pdu(struct nvmet_tcp_cmd *cmd)
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);
428 cmd->state = NVMET_TCP_SEND_RESPONSE;
430 pdu->hdr.type = nvme_tcp_rsp;
432 pdu->hdr.hlen = sizeof(*pdu);
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));
441 static void nvmet_tcp_process_resp_list(struct nvmet_tcp_queue *queue)
443 struct llist_node *node;
445 node = llist_del_all(&queue->resp_list);
450 struct nvmet_tcp_cmd *cmd = llist_entry(node,
451 struct nvmet_tcp_cmd, lentry);
453 list_add(&cmd->entry, &queue->resp_send_list);
455 queue->send_list_len++;
459 static struct nvmet_tcp_cmd *nvmet_tcp_fetch_cmd(struct nvmet_tcp_queue *queue)
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);
466 list_first_entry_or_null(&queue->resp_send_list,
467 struct nvmet_tcp_cmd, entry);
468 if (unlikely(!queue->snd_cmd))
472 list_del_init(&queue->snd_cmd->entry);
473 queue->send_list_len--;
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);
480 nvmet_setup_response_pdu(queue->snd_cmd);
482 return queue->snd_cmd;
485 static void nvmet_tcp_queue_response(struct nvmet_req *req)
487 struct nvmet_tcp_cmd *cmd =
488 container_of(req, struct nvmet_tcp_cmd, req);
489 struct nvmet_tcp_queue *queue = cmd->queue;
491 llist_add(&cmd->lentry, &queue->resp_list);
492 queue_work_on(cmd->queue->cpu, nvmet_tcp_wq, &cmd->queue->io_work);
495 static int nvmet_try_send_data_pdu(struct nvmet_tcp_cmd *cmd)
497 u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
498 int left = sizeof(*cmd->data_pdu) - cmd->offset + hdgst;
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);
513 cmd->state = NVMET_TCP_SEND_DATA;
518 static int nvmet_try_send_data(struct nvmet_tcp_cmd *cmd, bool last_in_batch)
520 struct nvmet_tcp_queue *queue = cmd->queue;
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;
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)
533 ret = kernel_sendpage(cmd->queue->sock, page, cmd->offset,
539 cmd->wbytes_done += ret;
542 if (cmd->offset == cmd->cur_sg->length) {
543 cmd->cur_sg = sg_next(cmd->cur_sg);
548 if (queue->data_digest) {
549 cmd->state = NVMET_TCP_SEND_DDGST;
552 if (queue->nvme_sq.sqhd_disabled) {
553 cmd->queue->snd_cmd = NULL;
554 nvmet_tcp_put_cmd(cmd);
556 nvmet_setup_response_pdu(cmd);
560 if (queue->nvme_sq.sqhd_disabled) {
562 sgl_free(cmd->req.sg);
569 static int nvmet_try_send_response(struct nvmet_tcp_cmd *cmd,
572 u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
573 int left = sizeof(*cmd->rsp_pdu) - cmd->offset + hdgst;
574 int flags = MSG_DONTWAIT;
577 if (!last_in_batch && cmd->queue->send_list_len)
582 ret = kernel_sendpage(cmd->queue->sock, virt_to_page(cmd->rsp_pdu),
583 offset_in_page(cmd->rsp_pdu) + cmd->offset, left, flags);
593 sgl_free(cmd->req.sg);
594 cmd->queue->snd_cmd = NULL;
595 nvmet_tcp_put_cmd(cmd);
599 static int nvmet_try_send_r2t(struct nvmet_tcp_cmd *cmd, bool last_in_batch)
601 u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
602 int left = sizeof(*cmd->r2t_pdu) - cmd->offset + hdgst;
603 int flags = MSG_DONTWAIT;
606 if (!last_in_batch && cmd->queue->send_list_len)
611 ret = kernel_sendpage(cmd->queue->sock, virt_to_page(cmd->r2t_pdu),
612 offset_in_page(cmd->r2t_pdu) + cmd->offset, left, flags);
621 cmd->queue->snd_cmd = NULL;
625 static int nvmet_try_send_ddgst(struct nvmet_tcp_cmd *cmd)
627 struct nvmet_tcp_queue *queue = cmd->queue;
628 struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
630 .iov_base = &cmd->exp_ddgst + cmd->offset,
631 .iov_len = NVME_TCP_DIGEST_LENGTH - cmd->offset
635 ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
636 if (unlikely(ret <= 0))
641 if (queue->nvme_sq.sqhd_disabled) {
642 cmd->queue->snd_cmd = NULL;
643 nvmet_tcp_put_cmd(cmd);
645 nvmet_setup_response_pdu(cmd);
650 static int nvmet_tcp_try_send_one(struct nvmet_tcp_queue *queue,
653 struct nvmet_tcp_cmd *cmd = queue->snd_cmd;
656 if (!cmd || queue->state == NVMET_TCP_Q_DISCONNECTING) {
657 cmd = nvmet_tcp_fetch_cmd(queue);
662 if (cmd->state == NVMET_TCP_SEND_DATA_PDU) {
663 ret = nvmet_try_send_data_pdu(cmd);
668 if (cmd->state == NVMET_TCP_SEND_DATA) {
669 ret = nvmet_try_send_data(cmd, last_in_batch);
674 if (cmd->state == NVMET_TCP_SEND_DDGST) {
675 ret = nvmet_try_send_ddgst(cmd);
680 if (cmd->state == NVMET_TCP_SEND_R2T) {
681 ret = nvmet_try_send_r2t(cmd, last_in_batch);
686 if (cmd->state == NVMET_TCP_SEND_RESPONSE)
687 ret = nvmet_try_send_response(cmd, last_in_batch);
699 static int nvmet_tcp_try_send(struct nvmet_tcp_queue *queue,
700 int budget, int *sends)
704 for (i = 0; i < budget; i++) {
705 ret = nvmet_tcp_try_send_one(queue, i == budget - 1);
714 static void nvmet_prepare_receive_pdu(struct nvmet_tcp_queue *queue)
717 queue->left = sizeof(struct nvme_tcp_hdr);
719 queue->rcv_state = NVMET_TCP_RECV_PDU;
722 static void nvmet_tcp_free_crypto(struct nvmet_tcp_queue *queue)
724 struct crypto_ahash *tfm = crypto_ahash_reqtfm(queue->rcv_hash);
726 ahash_request_free(queue->rcv_hash);
727 ahash_request_free(queue->snd_hash);
728 crypto_free_ahash(tfm);
731 static int nvmet_tcp_alloc_crypto(struct nvmet_tcp_queue *queue)
733 struct crypto_ahash *tfm;
735 tfm = crypto_alloc_ahash("crc32c", 0, CRYPTO_ALG_ASYNC);
739 queue->snd_hash = ahash_request_alloc(tfm, GFP_KERNEL);
740 if (!queue->snd_hash)
742 ahash_request_set_callback(queue->snd_hash, 0, NULL, NULL);
744 queue->rcv_hash = ahash_request_alloc(tfm, GFP_KERNEL);
745 if (!queue->rcv_hash)
747 ahash_request_set_callback(queue->rcv_hash, 0, NULL, NULL);
751 ahash_request_free(queue->snd_hash);
753 crypto_free_ahash(tfm);
758 static int nvmet_tcp_handle_icreq(struct nvmet_tcp_queue *queue)
760 struct nvme_tcp_icreq_pdu *icreq = &queue->pdu.icreq;
761 struct nvme_tcp_icresp_pdu *icresp = &queue->pdu.icresp;
762 struct msghdr msg = {};
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);
772 if (icreq->pfv != NVME_TCP_PFV_1_0) {
773 pr_err("queue %d: bad pfv %d\n", queue->idx, icreq->pfv);
777 if (icreq->hpda != 0) {
778 pr_err("queue %d: unsupported hpda %d\n", queue->idx,
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);
791 memset(icresp, 0, sizeof(*icresp));
792 icresp->hdr.type = nvme_tcp_icresp;
793 icresp->hdr.hlen = sizeof(*icresp);
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(0xffff); /* FIXME: support r2t */
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;
804 iov.iov_base = icresp;
805 iov.iov_len = sizeof(*icresp);
806 ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
810 queue->state = NVMET_TCP_Q_LIVE;
811 nvmet_prepare_receive_pdu(queue);
814 if (queue->hdr_digest || queue->data_digest)
815 nvmet_tcp_free_crypto(queue);
819 static void nvmet_tcp_handle_req_failure(struct nvmet_tcp_queue *queue,
820 struct nvmet_tcp_cmd *cmd, struct nvmet_req *req)
822 size_t data_len = le32_to_cpu(req->cmd->common.dptr.sgl.length);
825 if (!nvme_is_write(cmd->req.cmd) ||
826 data_len > cmd->req.port->inline_data_size) {
827 nvmet_prepare_receive_pdu(queue);
831 ret = nvmet_tcp_map_data(cmd);
833 pr_err("queue %d: failed to map data\n", queue->idx);
834 nvmet_tcp_fatal_error(queue);
838 queue->rcv_state = NVMET_TCP_RECV_DATA;
839 nvmet_tcp_map_pdu_iovec(cmd);
840 cmd->flags |= NVMET_TCP_F_INIT_FAILED;
843 static int nvmet_tcp_handle_h2c_data_pdu(struct nvmet_tcp_queue *queue)
845 struct nvme_tcp_data_pdu *data = &queue->pdu.data;
846 struct nvmet_tcp_cmd *cmd;
848 cmd = &queue->cmds[data->ttag];
850 if (le32_to_cpu(data->data_offset) != cmd->rbytes_done) {
851 pr_err("ttag %u unexpected data offset %u (expected %u)\n",
852 data->ttag, le32_to_cpu(data->data_offset),
854 /* FIXME: use path and transport errors */
855 nvmet_req_complete(&cmd->req,
856 NVME_SC_INVALID_FIELD | NVME_SC_DNR);
860 cmd->pdu_len = le32_to_cpu(data->data_length);
862 nvmet_tcp_map_pdu_iovec(cmd);
864 queue->rcv_state = NVMET_TCP_RECV_DATA;
869 static int nvmet_tcp_done_recv_pdu(struct nvmet_tcp_queue *queue)
871 struct nvme_tcp_hdr *hdr = &queue->pdu.cmd.hdr;
872 struct nvme_command *nvme_cmd = &queue->pdu.cmd.cmd;
873 struct nvmet_req *req;
876 if (unlikely(queue->state == NVMET_TCP_Q_CONNECTING)) {
877 if (hdr->type != nvme_tcp_icreq) {
878 pr_err("unexpected pdu type (%d) before icreq\n",
880 nvmet_tcp_fatal_error(queue);
883 return nvmet_tcp_handle_icreq(queue);
886 if (hdr->type == nvme_tcp_h2c_data) {
887 ret = nvmet_tcp_handle_h2c_data_pdu(queue);
893 queue->cmd = nvmet_tcp_get_cmd(queue);
894 if (unlikely(!queue->cmd)) {
895 /* This should never happen */
896 pr_err("queue %d: out of commands (%d) send_list_len: %d, opcode: %d",
897 queue->idx, queue->nr_cmds, queue->send_list_len,
898 nvme_cmd->common.opcode);
899 nvmet_tcp_fatal_error(queue);
903 req = &queue->cmd->req;
904 memcpy(req->cmd, nvme_cmd, sizeof(*nvme_cmd));
906 if (unlikely(!nvmet_req_init(req, &queue->nvme_cq,
907 &queue->nvme_sq, &nvmet_tcp_ops))) {
908 pr_err("failed cmd %p id %d opcode %d, data_len: %d\n",
909 req->cmd, req->cmd->common.command_id,
910 req->cmd->common.opcode,
911 le32_to_cpu(req->cmd->common.dptr.sgl.length));
913 nvmet_tcp_handle_req_failure(queue, queue->cmd, req);
917 ret = nvmet_tcp_map_data(queue->cmd);
919 pr_err("queue %d: failed to map data\n", queue->idx);
920 if (nvmet_tcp_has_inline_data(queue->cmd))
921 nvmet_tcp_fatal_error(queue);
923 nvmet_req_complete(req, ret);
928 if (nvmet_tcp_need_data_in(queue->cmd)) {
929 if (nvmet_tcp_has_inline_data(queue->cmd)) {
930 queue->rcv_state = NVMET_TCP_RECV_DATA;
931 nvmet_tcp_map_pdu_iovec(queue->cmd);
935 nvmet_tcp_queue_response(&queue->cmd->req);
939 queue->cmd->req.execute(&queue->cmd->req);
941 nvmet_prepare_receive_pdu(queue);
945 static const u8 nvme_tcp_pdu_sizes[] = {
946 [nvme_tcp_icreq] = sizeof(struct nvme_tcp_icreq_pdu),
947 [nvme_tcp_cmd] = sizeof(struct nvme_tcp_cmd_pdu),
948 [nvme_tcp_h2c_data] = sizeof(struct nvme_tcp_data_pdu),
951 static inline u8 nvmet_tcp_pdu_size(u8 type)
955 return (idx < ARRAY_SIZE(nvme_tcp_pdu_sizes) &&
956 nvme_tcp_pdu_sizes[idx]) ?
957 nvme_tcp_pdu_sizes[idx] : 0;
960 static inline bool nvmet_tcp_pdu_valid(u8 type)
965 case nvme_tcp_h2c_data:
973 static int nvmet_tcp_try_recv_pdu(struct nvmet_tcp_queue *queue)
975 struct nvme_tcp_hdr *hdr = &queue->pdu.cmd.hdr;
978 struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
981 iov.iov_base = (void *)&queue->pdu + queue->offset;
982 iov.iov_len = queue->left;
983 len = kernel_recvmsg(queue->sock, &msg, &iov, 1,
984 iov.iov_len, msg.msg_flags);
985 if (unlikely(len < 0))
988 queue->offset += len;
993 if (queue->offset == sizeof(struct nvme_tcp_hdr)) {
994 u8 hdgst = nvmet_tcp_hdgst_len(queue);
996 if (unlikely(!nvmet_tcp_pdu_valid(hdr->type))) {
997 pr_err("unexpected pdu type %d\n", hdr->type);
998 nvmet_tcp_fatal_error(queue);
1002 if (unlikely(hdr->hlen != nvmet_tcp_pdu_size(hdr->type))) {
1003 pr_err("pdu %d bad hlen %d\n", hdr->type, hdr->hlen);
1007 queue->left = hdr->hlen - queue->offset + hdgst;
1011 if (queue->hdr_digest &&
1012 nvmet_tcp_verify_hdgst(queue, &queue->pdu, queue->offset)) {
1013 nvmet_tcp_fatal_error(queue); /* fatal */
1017 if (queue->data_digest &&
1018 nvmet_tcp_check_ddgst(queue, &queue->pdu)) {
1019 nvmet_tcp_fatal_error(queue); /* fatal */
1023 return nvmet_tcp_done_recv_pdu(queue);
1026 static void nvmet_tcp_prep_recv_ddgst(struct nvmet_tcp_cmd *cmd)
1028 struct nvmet_tcp_queue *queue = cmd->queue;
1030 nvmet_tcp_ddgst(queue->rcv_hash, cmd);
1032 queue->left = NVME_TCP_DIGEST_LENGTH;
1033 queue->rcv_state = NVMET_TCP_RECV_DDGST;
1036 static int nvmet_tcp_try_recv_data(struct nvmet_tcp_queue *queue)
1038 struct nvmet_tcp_cmd *cmd = queue->cmd;
1041 while (msg_data_left(&cmd->recv_msg)) {
1042 ret = sock_recvmsg(cmd->queue->sock, &cmd->recv_msg,
1043 cmd->recv_msg.msg_flags);
1047 cmd->pdu_recv += ret;
1048 cmd->rbytes_done += ret;
1051 nvmet_tcp_unmap_pdu_iovec(cmd);
1053 if (!(cmd->flags & NVMET_TCP_F_INIT_FAILED) &&
1054 cmd->rbytes_done == cmd->req.transfer_len) {
1055 if (queue->data_digest) {
1056 nvmet_tcp_prep_recv_ddgst(cmd);
1059 cmd->req.execute(&cmd->req);
1062 nvmet_prepare_receive_pdu(queue);
1066 static int nvmet_tcp_try_recv_ddgst(struct nvmet_tcp_queue *queue)
1068 struct nvmet_tcp_cmd *cmd = queue->cmd;
1070 struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
1072 .iov_base = (void *)&cmd->recv_ddgst + queue->offset,
1073 .iov_len = queue->left
1076 ret = kernel_recvmsg(queue->sock, &msg, &iov, 1,
1077 iov.iov_len, msg.msg_flags);
1078 if (unlikely(ret < 0))
1081 queue->offset += ret;
1086 if (queue->data_digest && cmd->exp_ddgst != cmd->recv_ddgst) {
1087 pr_err("queue %d: cmd %d pdu (%d) data digest error: recv %#x expected %#x\n",
1088 queue->idx, cmd->req.cmd->common.command_id,
1089 queue->pdu.cmd.hdr.type, le32_to_cpu(cmd->recv_ddgst),
1090 le32_to_cpu(cmd->exp_ddgst));
1091 nvmet_tcp_finish_cmd(cmd);
1092 nvmet_tcp_fatal_error(queue);
1097 if (!(cmd->flags & NVMET_TCP_F_INIT_FAILED) &&
1098 cmd->rbytes_done == cmd->req.transfer_len)
1099 cmd->req.execute(&cmd->req);
1102 nvmet_prepare_receive_pdu(queue);
1106 static int nvmet_tcp_try_recv_one(struct nvmet_tcp_queue *queue)
1110 if (unlikely(queue->rcv_state == NVMET_TCP_RECV_ERR))
1113 if (queue->rcv_state == NVMET_TCP_RECV_PDU) {
1114 result = nvmet_tcp_try_recv_pdu(queue);
1119 if (queue->rcv_state == NVMET_TCP_RECV_DATA) {
1120 result = nvmet_tcp_try_recv_data(queue);
1125 if (queue->rcv_state == NVMET_TCP_RECV_DDGST) {
1126 result = nvmet_tcp_try_recv_ddgst(queue);
1133 if (result == -EAGAIN)
1140 static int nvmet_tcp_try_recv(struct nvmet_tcp_queue *queue,
1141 int budget, int *recvs)
1145 for (i = 0; i < budget; i++) {
1146 ret = nvmet_tcp_try_recv_one(queue);
1155 static void nvmet_tcp_schedule_release_queue(struct nvmet_tcp_queue *queue)
1157 spin_lock(&queue->state_lock);
1158 if (queue->state != NVMET_TCP_Q_DISCONNECTING) {
1159 queue->state = NVMET_TCP_Q_DISCONNECTING;
1160 schedule_work(&queue->release_work);
1162 spin_unlock(&queue->state_lock);
1165 static void nvmet_tcp_io_work(struct work_struct *w)
1167 struct nvmet_tcp_queue *queue =
1168 container_of(w, struct nvmet_tcp_queue, io_work);
1175 ret = nvmet_tcp_try_recv(queue, NVMET_TCP_RECV_BUDGET, &ops);
1178 } else if (ret < 0) {
1179 if (ret == -EPIPE || ret == -ECONNRESET)
1180 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1182 nvmet_tcp_fatal_error(queue);
1186 ret = nvmet_tcp_try_send(queue, NVMET_TCP_SEND_BUDGET, &ops);
1188 /* transmitted message/data */
1190 } else if (ret < 0) {
1191 if (ret == -EPIPE || ret == -ECONNRESET)
1192 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1194 nvmet_tcp_fatal_error(queue);
1198 } while (pending && ops < NVMET_TCP_IO_WORK_BUDGET);
1201 * We exahusted our budget, requeue our selves
1204 queue_work_on(queue->cpu, nvmet_tcp_wq, &queue->io_work);
1207 static int nvmet_tcp_alloc_cmd(struct nvmet_tcp_queue *queue,
1208 struct nvmet_tcp_cmd *c)
1210 u8 hdgst = nvmet_tcp_hdgst_len(queue);
1213 c->req.port = queue->port->nport;
1215 c->cmd_pdu = page_frag_alloc(&queue->pf_cache,
1216 sizeof(*c->cmd_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO);
1219 c->req.cmd = &c->cmd_pdu->cmd;
1221 c->rsp_pdu = page_frag_alloc(&queue->pf_cache,
1222 sizeof(*c->rsp_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO);
1225 c->req.cqe = &c->rsp_pdu->cqe;
1227 c->data_pdu = page_frag_alloc(&queue->pf_cache,
1228 sizeof(*c->data_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO);
1232 c->r2t_pdu = page_frag_alloc(&queue->pf_cache,
1233 sizeof(*c->r2t_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO);
1237 c->recv_msg.msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
1239 list_add_tail(&c->entry, &queue->free_list);
1243 page_frag_free(c->data_pdu);
1245 page_frag_free(c->rsp_pdu);
1247 page_frag_free(c->cmd_pdu);
1251 static void nvmet_tcp_free_cmd(struct nvmet_tcp_cmd *c)
1253 page_frag_free(c->r2t_pdu);
1254 page_frag_free(c->data_pdu);
1255 page_frag_free(c->rsp_pdu);
1256 page_frag_free(c->cmd_pdu);
1259 static int nvmet_tcp_alloc_cmds(struct nvmet_tcp_queue *queue)
1261 struct nvmet_tcp_cmd *cmds;
1262 int i, ret = -EINVAL, nr_cmds = queue->nr_cmds;
1264 cmds = kcalloc(nr_cmds, sizeof(struct nvmet_tcp_cmd), GFP_KERNEL);
1268 for (i = 0; i < nr_cmds; i++) {
1269 ret = nvmet_tcp_alloc_cmd(queue, cmds + i);
1279 nvmet_tcp_free_cmd(cmds + i);
1285 static void nvmet_tcp_free_cmds(struct nvmet_tcp_queue *queue)
1287 struct nvmet_tcp_cmd *cmds = queue->cmds;
1290 for (i = 0; i < queue->nr_cmds; i++)
1291 nvmet_tcp_free_cmd(cmds + i);
1293 nvmet_tcp_free_cmd(&queue->connect);
1297 static void nvmet_tcp_restore_socket_callbacks(struct nvmet_tcp_queue *queue)
1299 struct socket *sock = queue->sock;
1301 write_lock_bh(&sock->sk->sk_callback_lock);
1302 sock->sk->sk_data_ready = queue->data_ready;
1303 sock->sk->sk_state_change = queue->state_change;
1304 sock->sk->sk_write_space = queue->write_space;
1305 sock->sk->sk_user_data = NULL;
1306 write_unlock_bh(&sock->sk->sk_callback_lock);
1309 static void nvmet_tcp_finish_cmd(struct nvmet_tcp_cmd *cmd)
1311 nvmet_req_uninit(&cmd->req);
1312 nvmet_tcp_unmap_pdu_iovec(cmd);
1314 sgl_free(cmd->req.sg);
1317 static void nvmet_tcp_uninit_data_in_cmds(struct nvmet_tcp_queue *queue)
1319 struct nvmet_tcp_cmd *cmd = queue->cmds;
1322 for (i = 0; i < queue->nr_cmds; i++, cmd++) {
1323 if (nvmet_tcp_need_data_in(cmd))
1324 nvmet_tcp_finish_cmd(cmd);
1327 if (!queue->nr_cmds && nvmet_tcp_need_data_in(&queue->connect)) {
1328 /* failed in connect */
1329 nvmet_tcp_finish_cmd(&queue->connect);
1333 static void nvmet_tcp_release_queue_work(struct work_struct *w)
1335 struct nvmet_tcp_queue *queue =
1336 container_of(w, struct nvmet_tcp_queue, release_work);
1338 mutex_lock(&nvmet_tcp_queue_mutex);
1339 list_del_init(&queue->queue_list);
1340 mutex_unlock(&nvmet_tcp_queue_mutex);
1342 nvmet_tcp_restore_socket_callbacks(queue);
1343 flush_work(&queue->io_work);
1345 nvmet_tcp_uninit_data_in_cmds(queue);
1346 nvmet_sq_destroy(&queue->nvme_sq);
1347 cancel_work_sync(&queue->io_work);
1348 sock_release(queue->sock);
1349 nvmet_tcp_free_cmds(queue);
1350 if (queue->hdr_digest || queue->data_digest)
1351 nvmet_tcp_free_crypto(queue);
1352 ida_simple_remove(&nvmet_tcp_queue_ida, queue->idx);
1357 static void nvmet_tcp_data_ready(struct sock *sk)
1359 struct nvmet_tcp_queue *queue;
1361 read_lock_bh(&sk->sk_callback_lock);
1362 queue = sk->sk_user_data;
1364 queue_work_on(queue->cpu, nvmet_tcp_wq, &queue->io_work);
1365 read_unlock_bh(&sk->sk_callback_lock);
1368 static void nvmet_tcp_write_space(struct sock *sk)
1370 struct nvmet_tcp_queue *queue;
1372 read_lock_bh(&sk->sk_callback_lock);
1373 queue = sk->sk_user_data;
1374 if (unlikely(!queue))
1377 if (unlikely(queue->state == NVMET_TCP_Q_CONNECTING)) {
1378 queue->write_space(sk);
1382 if (sk_stream_is_writeable(sk)) {
1383 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1384 queue_work_on(queue->cpu, nvmet_tcp_wq, &queue->io_work);
1387 read_unlock_bh(&sk->sk_callback_lock);
1390 static void nvmet_tcp_state_change(struct sock *sk)
1392 struct nvmet_tcp_queue *queue;
1394 write_lock_bh(&sk->sk_callback_lock);
1395 queue = sk->sk_user_data;
1399 switch (sk->sk_state) {
1401 case TCP_CLOSE_WAIT:
1404 sk->sk_user_data = NULL;
1405 nvmet_tcp_schedule_release_queue(queue);
1408 pr_warn("queue %d unhandled state %d\n",
1409 queue->idx, sk->sk_state);
1412 write_unlock_bh(&sk->sk_callback_lock);
1415 static int nvmet_tcp_set_queue_sock(struct nvmet_tcp_queue *queue)
1417 struct socket *sock = queue->sock;
1418 struct inet_sock *inet = inet_sk(sock->sk);
1419 struct linger sol = { .l_onoff = 1, .l_linger = 0 };
1422 ret = kernel_getsockname(sock,
1423 (struct sockaddr *)&queue->sockaddr);
1427 ret = kernel_getpeername(sock,
1428 (struct sockaddr *)&queue->sockaddr_peer);
1433 * Cleanup whatever is sitting in the TCP transmit queue on socket
1434 * close. This is done to prevent stale data from being sent should
1435 * the network connection be restored before TCP times out.
1437 ret = kernel_setsockopt(sock, SOL_SOCKET, SO_LINGER,
1438 (char *)&sol, sizeof(sol));
1442 /* Set socket type of service */
1443 if (inet->rcv_tos > 0) {
1444 int tos = inet->rcv_tos;
1446 ret = kernel_setsockopt(sock, SOL_IP, IP_TOS,
1447 (char *)&tos, sizeof(tos));
1452 write_lock_bh(&sock->sk->sk_callback_lock);
1453 sock->sk->sk_user_data = queue;
1454 queue->data_ready = sock->sk->sk_data_ready;
1455 sock->sk->sk_data_ready = nvmet_tcp_data_ready;
1456 queue->state_change = sock->sk->sk_state_change;
1457 sock->sk->sk_state_change = nvmet_tcp_state_change;
1458 queue->write_space = sock->sk->sk_write_space;
1459 sock->sk->sk_write_space = nvmet_tcp_write_space;
1460 write_unlock_bh(&sock->sk->sk_callback_lock);
1465 static int nvmet_tcp_alloc_queue(struct nvmet_tcp_port *port,
1466 struct socket *newsock)
1468 struct nvmet_tcp_queue *queue;
1471 queue = kzalloc(sizeof(*queue), GFP_KERNEL);
1475 INIT_WORK(&queue->release_work, nvmet_tcp_release_queue_work);
1476 INIT_WORK(&queue->io_work, nvmet_tcp_io_work);
1477 queue->sock = newsock;
1480 spin_lock_init(&queue->state_lock);
1481 queue->state = NVMET_TCP_Q_CONNECTING;
1482 INIT_LIST_HEAD(&queue->free_list);
1483 init_llist_head(&queue->resp_list);
1484 INIT_LIST_HEAD(&queue->resp_send_list);
1486 queue->idx = ida_simple_get(&nvmet_tcp_queue_ida, 0, 0, GFP_KERNEL);
1487 if (queue->idx < 0) {
1489 goto out_free_queue;
1492 ret = nvmet_tcp_alloc_cmd(queue, &queue->connect);
1494 goto out_ida_remove;
1496 ret = nvmet_sq_init(&queue->nvme_sq);
1498 goto out_free_connect;
1500 port->last_cpu = cpumask_next_wrap(port->last_cpu,
1501 cpu_online_mask, -1, false);
1502 queue->cpu = port->last_cpu;
1503 nvmet_prepare_receive_pdu(queue);
1505 mutex_lock(&nvmet_tcp_queue_mutex);
1506 list_add_tail(&queue->queue_list, &nvmet_tcp_queue_list);
1507 mutex_unlock(&nvmet_tcp_queue_mutex);
1509 ret = nvmet_tcp_set_queue_sock(queue);
1511 goto out_destroy_sq;
1513 queue_work_on(queue->cpu, nvmet_tcp_wq, &queue->io_work);
1517 mutex_lock(&nvmet_tcp_queue_mutex);
1518 list_del_init(&queue->queue_list);
1519 mutex_unlock(&nvmet_tcp_queue_mutex);
1520 nvmet_sq_destroy(&queue->nvme_sq);
1522 nvmet_tcp_free_cmd(&queue->connect);
1524 ida_simple_remove(&nvmet_tcp_queue_ida, queue->idx);
1530 static void nvmet_tcp_accept_work(struct work_struct *w)
1532 struct nvmet_tcp_port *port =
1533 container_of(w, struct nvmet_tcp_port, accept_work);
1534 struct socket *newsock;
1538 ret = kernel_accept(port->sock, &newsock, O_NONBLOCK);
1541 pr_warn("failed to accept err=%d\n", ret);
1544 ret = nvmet_tcp_alloc_queue(port, newsock);
1546 pr_err("failed to allocate queue\n");
1547 sock_release(newsock);
1552 static void nvmet_tcp_listen_data_ready(struct sock *sk)
1554 struct nvmet_tcp_port *port;
1556 read_lock_bh(&sk->sk_callback_lock);
1557 port = sk->sk_user_data;
1561 if (sk->sk_state == TCP_LISTEN)
1562 schedule_work(&port->accept_work);
1564 read_unlock_bh(&sk->sk_callback_lock);
1567 static int nvmet_tcp_add_port(struct nvmet_port *nport)
1569 struct nvmet_tcp_port *port;
1570 __kernel_sa_family_t af;
1573 port = kzalloc(sizeof(*port), GFP_KERNEL);
1577 switch (nport->disc_addr.adrfam) {
1578 case NVMF_ADDR_FAMILY_IP4:
1581 case NVMF_ADDR_FAMILY_IP6:
1585 pr_err("address family %d not supported\n",
1586 nport->disc_addr.adrfam);
1591 ret = inet_pton_with_scope(&init_net, af, nport->disc_addr.traddr,
1592 nport->disc_addr.trsvcid, &port->addr);
1594 pr_err("malformed ip/port passed: %s:%s\n",
1595 nport->disc_addr.traddr, nport->disc_addr.trsvcid);
1599 port->nport = nport;
1600 port->last_cpu = -1;
1601 INIT_WORK(&port->accept_work, nvmet_tcp_accept_work);
1602 if (port->nport->inline_data_size < 0)
1603 port->nport->inline_data_size = NVMET_TCP_DEF_INLINE_DATA_SIZE;
1605 ret = sock_create(port->addr.ss_family, SOCK_STREAM,
1606 IPPROTO_TCP, &port->sock);
1608 pr_err("failed to create a socket\n");
1612 port->sock->sk->sk_user_data = port;
1613 port->data_ready = port->sock->sk->sk_data_ready;
1614 port->sock->sk->sk_data_ready = nvmet_tcp_listen_data_ready;
1617 ret = kernel_setsockopt(port->sock, IPPROTO_TCP,
1618 TCP_NODELAY, (char *)&opt, sizeof(opt));
1620 pr_err("failed to set TCP_NODELAY sock opt %d\n", ret);
1624 ret = kernel_setsockopt(port->sock, SOL_SOCKET, SO_REUSEADDR,
1625 (char *)&opt, sizeof(opt));
1627 pr_err("failed to set SO_REUSEADDR sock opt %d\n", ret);
1631 ret = kernel_bind(port->sock, (struct sockaddr *)&port->addr,
1632 sizeof(port->addr));
1634 pr_err("failed to bind port socket %d\n", ret);
1638 ret = kernel_listen(port->sock, 128);
1640 pr_err("failed to listen %d on port sock\n", ret);
1645 pr_info("enabling port %d (%pISpc)\n",
1646 le16_to_cpu(nport->disc_addr.portid), &port->addr);
1651 sock_release(port->sock);
1657 static void nvmet_tcp_remove_port(struct nvmet_port *nport)
1659 struct nvmet_tcp_port *port = nport->priv;
1661 write_lock_bh(&port->sock->sk->sk_callback_lock);
1662 port->sock->sk->sk_data_ready = port->data_ready;
1663 port->sock->sk->sk_user_data = NULL;
1664 write_unlock_bh(&port->sock->sk->sk_callback_lock);
1665 cancel_work_sync(&port->accept_work);
1667 sock_release(port->sock);
1671 static void nvmet_tcp_delete_ctrl(struct nvmet_ctrl *ctrl)
1673 struct nvmet_tcp_queue *queue;
1675 mutex_lock(&nvmet_tcp_queue_mutex);
1676 list_for_each_entry(queue, &nvmet_tcp_queue_list, queue_list)
1677 if (queue->nvme_sq.ctrl == ctrl)
1678 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1679 mutex_unlock(&nvmet_tcp_queue_mutex);
1682 static u16 nvmet_tcp_install_queue(struct nvmet_sq *sq)
1684 struct nvmet_tcp_queue *queue =
1685 container_of(sq, struct nvmet_tcp_queue, nvme_sq);
1688 /* Let inflight controller teardown complete */
1689 flush_scheduled_work();
1692 queue->nr_cmds = sq->size * 2;
1693 if (nvmet_tcp_alloc_cmds(queue))
1694 return NVME_SC_INTERNAL;
1698 static void nvmet_tcp_disc_port_addr(struct nvmet_req *req,
1699 struct nvmet_port *nport, char *traddr)
1701 struct nvmet_tcp_port *port = nport->priv;
1703 if (inet_addr_is_any((struct sockaddr *)&port->addr)) {
1704 struct nvmet_tcp_cmd *cmd =
1705 container_of(req, struct nvmet_tcp_cmd, req);
1706 struct nvmet_tcp_queue *queue = cmd->queue;
1708 sprintf(traddr, "%pISc", (struct sockaddr *)&queue->sockaddr);
1710 memcpy(traddr, nport->disc_addr.traddr, NVMF_TRADDR_SIZE);
1714 static struct nvmet_fabrics_ops nvmet_tcp_ops = {
1715 .owner = THIS_MODULE,
1716 .type = NVMF_TRTYPE_TCP,
1718 .has_keyed_sgls = 0,
1719 .add_port = nvmet_tcp_add_port,
1720 .remove_port = nvmet_tcp_remove_port,
1721 .queue_response = nvmet_tcp_queue_response,
1722 .delete_ctrl = nvmet_tcp_delete_ctrl,
1723 .install_queue = nvmet_tcp_install_queue,
1724 .disc_traddr = nvmet_tcp_disc_port_addr,
1727 static int __init nvmet_tcp_init(void)
1731 nvmet_tcp_wq = alloc_workqueue("nvmet_tcp_wq", WQ_HIGHPRI, 0);
1735 ret = nvmet_register_transport(&nvmet_tcp_ops);
1741 destroy_workqueue(nvmet_tcp_wq);
1745 static void __exit nvmet_tcp_exit(void)
1747 struct nvmet_tcp_queue *queue;
1749 nvmet_unregister_transport(&nvmet_tcp_ops);
1751 flush_scheduled_work();
1752 mutex_lock(&nvmet_tcp_queue_mutex);
1753 list_for_each_entry(queue, &nvmet_tcp_queue_list, queue_list)
1754 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1755 mutex_unlock(&nvmet_tcp_queue_mutex);
1756 flush_scheduled_work();
1758 destroy_workqueue(nvmet_tcp_wq);
1761 module_init(nvmet_tcp_init);
1762 module_exit(nvmet_tcp_exit);
1764 MODULE_LICENSE("GPL v2");
1765 MODULE_ALIAS("nvmet-transport-3"); /* 3 == NVMF_TRTYPE_TCP */