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 the socket priority to use for connections were it is desirable
23 * that the NIC consider performing optimized packet processing or filtering.
24 * A non-zero value being sufficient to indicate general consideration of any
25 * possible optimization. Making it a module param allows for alternative
26 * values that may be unique for some NIC implementations.
28 static int so_priority;
29 module_param(so_priority, int, 0644);
30 MODULE_PARM_DESC(so_priority, "nvmet tcp socket optimize priority");
32 #define NVMET_TCP_RECV_BUDGET 8
33 #define NVMET_TCP_SEND_BUDGET 8
34 #define NVMET_TCP_IO_WORK_BUDGET 64
36 enum nvmet_tcp_send_state {
37 NVMET_TCP_SEND_DATA_PDU,
41 NVMET_TCP_SEND_RESPONSE
44 enum nvmet_tcp_recv_state {
52 NVMET_TCP_F_INIT_FAILED = (1 << 0),
55 struct nvmet_tcp_cmd {
56 struct nvmet_tcp_queue *queue;
59 struct nvme_tcp_cmd_pdu *cmd_pdu;
60 struct nvme_tcp_rsp_pdu *rsp_pdu;
61 struct nvme_tcp_data_pdu *data_pdu;
62 struct nvme_tcp_r2t_pdu *r2t_pdu;
71 struct msghdr recv_msg;
75 struct list_head entry;
76 struct llist_node lentry;
80 struct scatterlist *cur_sg;
81 enum nvmet_tcp_send_state state;
87 enum nvmet_tcp_queue_state {
88 NVMET_TCP_Q_CONNECTING,
90 NVMET_TCP_Q_DISCONNECTING,
93 struct nvmet_tcp_queue {
95 struct nvmet_tcp_port *port;
96 struct work_struct io_work;
98 struct nvmet_cq nvme_cq;
99 struct nvmet_sq nvme_sq;
102 struct nvmet_tcp_cmd *cmds;
103 unsigned int nr_cmds;
104 struct list_head free_list;
105 struct llist_head resp_list;
106 struct list_head resp_send_list;
108 struct nvmet_tcp_cmd *snd_cmd;
113 enum nvmet_tcp_recv_state rcv_state;
114 struct nvmet_tcp_cmd *cmd;
115 union nvme_tcp_pdu pdu;
120 struct ahash_request *snd_hash;
121 struct ahash_request *rcv_hash;
123 spinlock_t state_lock;
124 enum nvmet_tcp_queue_state state;
126 struct sockaddr_storage sockaddr;
127 struct sockaddr_storage sockaddr_peer;
128 struct work_struct release_work;
131 struct list_head queue_list;
133 struct nvmet_tcp_cmd connect;
135 struct page_frag_cache pf_cache;
137 void (*data_ready)(struct sock *);
138 void (*state_change)(struct sock *);
139 void (*write_space)(struct sock *);
142 struct nvmet_tcp_port {
144 struct work_struct accept_work;
145 struct nvmet_port *nport;
146 struct sockaddr_storage addr;
148 void (*data_ready)(struct sock *);
151 static DEFINE_IDA(nvmet_tcp_queue_ida);
152 static LIST_HEAD(nvmet_tcp_queue_list);
153 static DEFINE_MUTEX(nvmet_tcp_queue_mutex);
155 static struct workqueue_struct *nvmet_tcp_wq;
156 static struct nvmet_fabrics_ops nvmet_tcp_ops;
157 static void nvmet_tcp_free_cmd(struct nvmet_tcp_cmd *c);
158 static void nvmet_tcp_finish_cmd(struct nvmet_tcp_cmd *cmd);
160 static inline u16 nvmet_tcp_cmd_tag(struct nvmet_tcp_queue *queue,
161 struct nvmet_tcp_cmd *cmd)
163 return cmd - queue->cmds;
166 static inline bool nvmet_tcp_has_data_in(struct nvmet_tcp_cmd *cmd)
168 return nvme_is_write(cmd->req.cmd) &&
169 cmd->rbytes_done < cmd->req.transfer_len;
172 static inline bool nvmet_tcp_need_data_in(struct nvmet_tcp_cmd *cmd)
174 return nvmet_tcp_has_data_in(cmd) && !cmd->req.cqe->status;
177 static inline bool nvmet_tcp_need_data_out(struct nvmet_tcp_cmd *cmd)
179 return !nvme_is_write(cmd->req.cmd) &&
180 cmd->req.transfer_len > 0 &&
181 !cmd->req.cqe->status;
184 static inline bool nvmet_tcp_has_inline_data(struct nvmet_tcp_cmd *cmd)
186 return nvme_is_write(cmd->req.cmd) && cmd->pdu_len &&
190 static inline struct nvmet_tcp_cmd *
191 nvmet_tcp_get_cmd(struct nvmet_tcp_queue *queue)
193 struct nvmet_tcp_cmd *cmd;
195 cmd = list_first_entry_or_null(&queue->free_list,
196 struct nvmet_tcp_cmd, entry);
199 list_del_init(&cmd->entry);
201 cmd->rbytes_done = cmd->wbytes_done = 0;
209 static inline void nvmet_tcp_put_cmd(struct nvmet_tcp_cmd *cmd)
211 if (unlikely(cmd == &cmd->queue->connect))
214 list_add_tail(&cmd->entry, &cmd->queue->free_list);
217 static inline u8 nvmet_tcp_hdgst_len(struct nvmet_tcp_queue *queue)
219 return queue->hdr_digest ? NVME_TCP_DIGEST_LENGTH : 0;
222 static inline u8 nvmet_tcp_ddgst_len(struct nvmet_tcp_queue *queue)
224 return queue->data_digest ? NVME_TCP_DIGEST_LENGTH : 0;
227 static inline void nvmet_tcp_hdgst(struct ahash_request *hash,
228 void *pdu, size_t len)
230 struct scatterlist sg;
232 sg_init_one(&sg, pdu, len);
233 ahash_request_set_crypt(hash, &sg, pdu + len, len);
234 crypto_ahash_digest(hash);
237 static int nvmet_tcp_verify_hdgst(struct nvmet_tcp_queue *queue,
238 void *pdu, size_t len)
240 struct nvme_tcp_hdr *hdr = pdu;
244 if (unlikely(!(hdr->flags & NVME_TCP_F_HDGST))) {
245 pr_err("queue %d: header digest enabled but no header digest\n",
250 recv_digest = *(__le32 *)(pdu + hdr->hlen);
251 nvmet_tcp_hdgst(queue->rcv_hash, pdu, len);
252 exp_digest = *(__le32 *)(pdu + hdr->hlen);
253 if (recv_digest != exp_digest) {
254 pr_err("queue %d: header digest error: recv %#x expected %#x\n",
255 queue->idx, le32_to_cpu(recv_digest),
256 le32_to_cpu(exp_digest));
263 static int nvmet_tcp_check_ddgst(struct nvmet_tcp_queue *queue, void *pdu)
265 struct nvme_tcp_hdr *hdr = pdu;
266 u8 digest_len = nvmet_tcp_hdgst_len(queue);
269 len = le32_to_cpu(hdr->plen) - hdr->hlen -
270 (hdr->flags & NVME_TCP_F_HDGST ? digest_len : 0);
272 if (unlikely(len && !(hdr->flags & NVME_TCP_F_DDGST))) {
273 pr_err("queue %d: data digest flag is cleared\n", queue->idx);
280 static void nvmet_tcp_unmap_pdu_iovec(struct nvmet_tcp_cmd *cmd)
282 struct scatterlist *sg;
285 sg = &cmd->req.sg[cmd->sg_idx];
287 for (i = 0; i < cmd->nr_mapped; i++)
288 kunmap(sg_page(&sg[i]));
291 static void nvmet_tcp_map_pdu_iovec(struct nvmet_tcp_cmd *cmd)
293 struct kvec *iov = cmd->iov;
294 struct scatterlist *sg;
295 u32 length, offset, sg_offset;
297 length = cmd->pdu_len;
298 cmd->nr_mapped = DIV_ROUND_UP(length, PAGE_SIZE);
299 offset = cmd->rbytes_done;
300 cmd->sg_idx = DIV_ROUND_UP(offset, PAGE_SIZE);
301 sg_offset = offset % PAGE_SIZE;
302 sg = &cmd->req.sg[cmd->sg_idx];
305 u32 iov_len = min_t(u32, length, sg->length - sg_offset);
307 iov->iov_base = kmap(sg_page(sg)) + sg->offset + sg_offset;
308 iov->iov_len = iov_len;
315 iov_iter_kvec(&cmd->recv_msg.msg_iter, READ, cmd->iov,
316 cmd->nr_mapped, cmd->pdu_len);
319 static void nvmet_tcp_fatal_error(struct nvmet_tcp_queue *queue)
321 queue->rcv_state = NVMET_TCP_RECV_ERR;
322 if (queue->nvme_sq.ctrl)
323 nvmet_ctrl_fatal_error(queue->nvme_sq.ctrl);
325 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
328 static int nvmet_tcp_map_data(struct nvmet_tcp_cmd *cmd)
330 struct nvme_sgl_desc *sgl = &cmd->req.cmd->common.dptr.sgl;
331 u32 len = le32_to_cpu(sgl->length);
336 if (sgl->type == ((NVME_SGL_FMT_DATA_DESC << 4) |
337 NVME_SGL_FMT_OFFSET)) {
338 if (!nvme_is_write(cmd->req.cmd))
339 return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
341 if (len > cmd->req.port->inline_data_size)
342 return NVME_SC_SGL_INVALID_OFFSET | NVME_SC_DNR;
345 cmd->req.transfer_len += len;
347 cmd->req.sg = sgl_alloc(len, GFP_KERNEL, &cmd->req.sg_cnt);
349 return NVME_SC_INTERNAL;
350 cmd->cur_sg = cmd->req.sg;
352 if (nvmet_tcp_has_data_in(cmd)) {
353 cmd->iov = kmalloc_array(cmd->req.sg_cnt,
354 sizeof(*cmd->iov), GFP_KERNEL);
361 sgl_free(cmd->req.sg);
362 return NVME_SC_INTERNAL;
365 static void nvmet_tcp_ddgst(struct ahash_request *hash,
366 struct nvmet_tcp_cmd *cmd)
368 ahash_request_set_crypt(hash, cmd->req.sg,
369 (void *)&cmd->exp_ddgst, cmd->req.transfer_len);
370 crypto_ahash_digest(hash);
373 static void nvmet_setup_c2h_data_pdu(struct nvmet_tcp_cmd *cmd)
375 struct nvme_tcp_data_pdu *pdu = cmd->data_pdu;
376 struct nvmet_tcp_queue *queue = cmd->queue;
377 u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
378 u8 ddgst = nvmet_tcp_ddgst_len(cmd->queue);
381 cmd->state = NVMET_TCP_SEND_DATA_PDU;
383 pdu->hdr.type = nvme_tcp_c2h_data;
384 pdu->hdr.flags = NVME_TCP_F_DATA_LAST | (queue->nvme_sq.sqhd_disabled ?
385 NVME_TCP_F_DATA_SUCCESS : 0);
386 pdu->hdr.hlen = sizeof(*pdu);
387 pdu->hdr.pdo = pdu->hdr.hlen + hdgst;
389 cpu_to_le32(pdu->hdr.hlen + hdgst +
390 cmd->req.transfer_len + ddgst);
391 pdu->command_id = cmd->req.cqe->command_id;
392 pdu->data_length = cpu_to_le32(cmd->req.transfer_len);
393 pdu->data_offset = cpu_to_le32(cmd->wbytes_done);
395 if (queue->data_digest) {
396 pdu->hdr.flags |= NVME_TCP_F_DDGST;
397 nvmet_tcp_ddgst(queue->snd_hash, cmd);
400 if (cmd->queue->hdr_digest) {
401 pdu->hdr.flags |= NVME_TCP_F_HDGST;
402 nvmet_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
406 static void nvmet_setup_r2t_pdu(struct nvmet_tcp_cmd *cmd)
408 struct nvme_tcp_r2t_pdu *pdu = cmd->r2t_pdu;
409 struct nvmet_tcp_queue *queue = cmd->queue;
410 u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
413 cmd->state = NVMET_TCP_SEND_R2T;
415 pdu->hdr.type = nvme_tcp_r2t;
417 pdu->hdr.hlen = sizeof(*pdu);
419 pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst);
421 pdu->command_id = cmd->req.cmd->common.command_id;
422 pdu->ttag = nvmet_tcp_cmd_tag(cmd->queue, cmd);
423 pdu->r2t_length = cpu_to_le32(cmd->req.transfer_len - cmd->rbytes_done);
424 pdu->r2t_offset = cpu_to_le32(cmd->rbytes_done);
425 if (cmd->queue->hdr_digest) {
426 pdu->hdr.flags |= NVME_TCP_F_HDGST;
427 nvmet_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
431 static void nvmet_setup_response_pdu(struct nvmet_tcp_cmd *cmd)
433 struct nvme_tcp_rsp_pdu *pdu = cmd->rsp_pdu;
434 struct nvmet_tcp_queue *queue = cmd->queue;
435 u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
438 cmd->state = NVMET_TCP_SEND_RESPONSE;
440 pdu->hdr.type = nvme_tcp_rsp;
442 pdu->hdr.hlen = sizeof(*pdu);
444 pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst);
445 if (cmd->queue->hdr_digest) {
446 pdu->hdr.flags |= NVME_TCP_F_HDGST;
447 nvmet_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
451 static void nvmet_tcp_process_resp_list(struct nvmet_tcp_queue *queue)
453 struct llist_node *node;
455 node = llist_del_all(&queue->resp_list);
460 struct nvmet_tcp_cmd *cmd = llist_entry(node,
461 struct nvmet_tcp_cmd, lentry);
463 list_add(&cmd->entry, &queue->resp_send_list);
465 queue->send_list_len++;
469 static struct nvmet_tcp_cmd *nvmet_tcp_fetch_cmd(struct nvmet_tcp_queue *queue)
471 queue->snd_cmd = list_first_entry_or_null(&queue->resp_send_list,
472 struct nvmet_tcp_cmd, entry);
473 if (!queue->snd_cmd) {
474 nvmet_tcp_process_resp_list(queue);
476 list_first_entry_or_null(&queue->resp_send_list,
477 struct nvmet_tcp_cmd, entry);
478 if (unlikely(!queue->snd_cmd))
482 list_del_init(&queue->snd_cmd->entry);
483 queue->send_list_len--;
485 if (nvmet_tcp_need_data_out(queue->snd_cmd))
486 nvmet_setup_c2h_data_pdu(queue->snd_cmd);
487 else if (nvmet_tcp_need_data_in(queue->snd_cmd))
488 nvmet_setup_r2t_pdu(queue->snd_cmd);
490 nvmet_setup_response_pdu(queue->snd_cmd);
492 return queue->snd_cmd;
495 static void nvmet_tcp_queue_response(struct nvmet_req *req)
497 struct nvmet_tcp_cmd *cmd =
498 container_of(req, struct nvmet_tcp_cmd, req);
499 struct nvmet_tcp_queue *queue = cmd->queue;
501 llist_add(&cmd->lentry, &queue->resp_list);
502 queue_work_on(cmd->queue->cpu, nvmet_tcp_wq, &cmd->queue->io_work);
505 static int nvmet_try_send_data_pdu(struct nvmet_tcp_cmd *cmd)
507 u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
508 int left = sizeof(*cmd->data_pdu) - cmd->offset + hdgst;
511 ret = kernel_sendpage(cmd->queue->sock, virt_to_page(cmd->data_pdu),
512 offset_in_page(cmd->data_pdu) + cmd->offset,
513 left, MSG_DONTWAIT | MSG_MORE);
523 cmd->state = NVMET_TCP_SEND_DATA;
528 static int nvmet_try_send_data(struct nvmet_tcp_cmd *cmd, bool last_in_batch)
530 struct nvmet_tcp_queue *queue = cmd->queue;
533 while (cmd->cur_sg) {
534 struct page *page = sg_page(cmd->cur_sg);
535 u32 left = cmd->cur_sg->length - cmd->offset;
536 int flags = MSG_DONTWAIT;
538 if ((!last_in_batch && cmd->queue->send_list_len) ||
539 cmd->wbytes_done + left < cmd->req.transfer_len ||
540 queue->data_digest || !queue->nvme_sq.sqhd_disabled)
543 ret = kernel_sendpage(cmd->queue->sock, page, cmd->offset,
549 cmd->wbytes_done += ret;
552 if (cmd->offset == cmd->cur_sg->length) {
553 cmd->cur_sg = sg_next(cmd->cur_sg);
558 if (queue->data_digest) {
559 cmd->state = NVMET_TCP_SEND_DDGST;
562 if (queue->nvme_sq.sqhd_disabled) {
563 cmd->queue->snd_cmd = NULL;
564 nvmet_tcp_put_cmd(cmd);
566 nvmet_setup_response_pdu(cmd);
570 if (queue->nvme_sq.sqhd_disabled) {
572 sgl_free(cmd->req.sg);
579 static int nvmet_try_send_response(struct nvmet_tcp_cmd *cmd,
582 u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
583 int left = sizeof(*cmd->rsp_pdu) - cmd->offset + hdgst;
584 int flags = MSG_DONTWAIT;
587 if (!last_in_batch && cmd->queue->send_list_len)
592 ret = kernel_sendpage(cmd->queue->sock, virt_to_page(cmd->rsp_pdu),
593 offset_in_page(cmd->rsp_pdu) + cmd->offset, left, flags);
603 sgl_free(cmd->req.sg);
604 cmd->queue->snd_cmd = NULL;
605 nvmet_tcp_put_cmd(cmd);
609 static int nvmet_try_send_r2t(struct nvmet_tcp_cmd *cmd, bool last_in_batch)
611 u8 hdgst = nvmet_tcp_hdgst_len(cmd->queue);
612 int left = sizeof(*cmd->r2t_pdu) - cmd->offset + hdgst;
613 int flags = MSG_DONTWAIT;
616 if (!last_in_batch && cmd->queue->send_list_len)
621 ret = kernel_sendpage(cmd->queue->sock, virt_to_page(cmd->r2t_pdu),
622 offset_in_page(cmd->r2t_pdu) + cmd->offset, left, flags);
631 cmd->queue->snd_cmd = NULL;
635 static int nvmet_try_send_ddgst(struct nvmet_tcp_cmd *cmd, bool last_in_batch)
637 struct nvmet_tcp_queue *queue = cmd->queue;
638 struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
640 .iov_base = &cmd->exp_ddgst + cmd->offset,
641 .iov_len = NVME_TCP_DIGEST_LENGTH - cmd->offset
645 if (!last_in_batch && cmd->queue->send_list_len)
646 msg.msg_flags |= MSG_MORE;
648 ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
649 if (unlikely(ret <= 0))
654 if (queue->nvme_sq.sqhd_disabled) {
655 cmd->queue->snd_cmd = NULL;
656 nvmet_tcp_put_cmd(cmd);
658 nvmet_setup_response_pdu(cmd);
663 static int nvmet_tcp_try_send_one(struct nvmet_tcp_queue *queue,
666 struct nvmet_tcp_cmd *cmd = queue->snd_cmd;
669 if (!cmd || queue->state == NVMET_TCP_Q_DISCONNECTING) {
670 cmd = nvmet_tcp_fetch_cmd(queue);
675 if (cmd->state == NVMET_TCP_SEND_DATA_PDU) {
676 ret = nvmet_try_send_data_pdu(cmd);
681 if (cmd->state == NVMET_TCP_SEND_DATA) {
682 ret = nvmet_try_send_data(cmd, last_in_batch);
687 if (cmd->state == NVMET_TCP_SEND_DDGST) {
688 ret = nvmet_try_send_ddgst(cmd, last_in_batch);
693 if (cmd->state == NVMET_TCP_SEND_R2T) {
694 ret = nvmet_try_send_r2t(cmd, last_in_batch);
699 if (cmd->state == NVMET_TCP_SEND_RESPONSE)
700 ret = nvmet_try_send_response(cmd, last_in_batch);
712 static int nvmet_tcp_try_send(struct nvmet_tcp_queue *queue,
713 int budget, int *sends)
717 for (i = 0; i < budget; i++) {
718 ret = nvmet_tcp_try_send_one(queue, i == budget - 1);
727 static void nvmet_prepare_receive_pdu(struct nvmet_tcp_queue *queue)
730 queue->left = sizeof(struct nvme_tcp_hdr);
732 queue->rcv_state = NVMET_TCP_RECV_PDU;
735 static void nvmet_tcp_free_crypto(struct nvmet_tcp_queue *queue)
737 struct crypto_ahash *tfm = crypto_ahash_reqtfm(queue->rcv_hash);
739 ahash_request_free(queue->rcv_hash);
740 ahash_request_free(queue->snd_hash);
741 crypto_free_ahash(tfm);
744 static int nvmet_tcp_alloc_crypto(struct nvmet_tcp_queue *queue)
746 struct crypto_ahash *tfm;
748 tfm = crypto_alloc_ahash("crc32c", 0, CRYPTO_ALG_ASYNC);
752 queue->snd_hash = ahash_request_alloc(tfm, GFP_KERNEL);
753 if (!queue->snd_hash)
755 ahash_request_set_callback(queue->snd_hash, 0, NULL, NULL);
757 queue->rcv_hash = ahash_request_alloc(tfm, GFP_KERNEL);
758 if (!queue->rcv_hash)
760 ahash_request_set_callback(queue->rcv_hash, 0, NULL, NULL);
764 ahash_request_free(queue->snd_hash);
766 crypto_free_ahash(tfm);
771 static int nvmet_tcp_handle_icreq(struct nvmet_tcp_queue *queue)
773 struct nvme_tcp_icreq_pdu *icreq = &queue->pdu.icreq;
774 struct nvme_tcp_icresp_pdu *icresp = &queue->pdu.icresp;
775 struct msghdr msg = {};
779 if (le32_to_cpu(icreq->hdr.plen) != sizeof(struct nvme_tcp_icreq_pdu)) {
780 pr_err("bad nvme-tcp pdu length (%d)\n",
781 le32_to_cpu(icreq->hdr.plen));
782 nvmet_tcp_fatal_error(queue);
785 if (icreq->pfv != NVME_TCP_PFV_1_0) {
786 pr_err("queue %d: bad pfv %d\n", queue->idx, icreq->pfv);
790 if (icreq->hpda != 0) {
791 pr_err("queue %d: unsupported hpda %d\n", queue->idx,
796 queue->hdr_digest = !!(icreq->digest & NVME_TCP_HDR_DIGEST_ENABLE);
797 queue->data_digest = !!(icreq->digest & NVME_TCP_DATA_DIGEST_ENABLE);
798 if (queue->hdr_digest || queue->data_digest) {
799 ret = nvmet_tcp_alloc_crypto(queue);
804 memset(icresp, 0, sizeof(*icresp));
805 icresp->hdr.type = nvme_tcp_icresp;
806 icresp->hdr.hlen = sizeof(*icresp);
808 icresp->hdr.plen = cpu_to_le32(icresp->hdr.hlen);
809 icresp->pfv = cpu_to_le16(NVME_TCP_PFV_1_0);
810 icresp->maxdata = cpu_to_le32(0x400000); /* 16M arbitrary limit */
812 if (queue->hdr_digest)
813 icresp->digest |= NVME_TCP_HDR_DIGEST_ENABLE;
814 if (queue->data_digest)
815 icresp->digest |= NVME_TCP_DATA_DIGEST_ENABLE;
817 iov.iov_base = icresp;
818 iov.iov_len = sizeof(*icresp);
819 ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
823 queue->state = NVMET_TCP_Q_LIVE;
824 nvmet_prepare_receive_pdu(queue);
827 if (queue->hdr_digest || queue->data_digest)
828 nvmet_tcp_free_crypto(queue);
832 static void nvmet_tcp_handle_req_failure(struct nvmet_tcp_queue *queue,
833 struct nvmet_tcp_cmd *cmd, struct nvmet_req *req)
835 size_t data_len = le32_to_cpu(req->cmd->common.dptr.sgl.length);
838 if (!nvme_is_write(cmd->req.cmd) ||
839 data_len > cmd->req.port->inline_data_size) {
840 nvmet_prepare_receive_pdu(queue);
844 ret = nvmet_tcp_map_data(cmd);
846 pr_err("queue %d: failed to map data\n", queue->idx);
847 nvmet_tcp_fatal_error(queue);
851 queue->rcv_state = NVMET_TCP_RECV_DATA;
852 nvmet_tcp_map_pdu_iovec(cmd);
853 cmd->flags |= NVMET_TCP_F_INIT_FAILED;
856 static int nvmet_tcp_handle_h2c_data_pdu(struct nvmet_tcp_queue *queue)
858 struct nvme_tcp_data_pdu *data = &queue->pdu.data;
859 struct nvmet_tcp_cmd *cmd;
861 cmd = &queue->cmds[data->ttag];
863 if (le32_to_cpu(data->data_offset) != cmd->rbytes_done) {
864 pr_err("ttag %u unexpected data offset %u (expected %u)\n",
865 data->ttag, le32_to_cpu(data->data_offset),
867 /* FIXME: use path and transport errors */
868 nvmet_req_complete(&cmd->req,
869 NVME_SC_INVALID_FIELD | NVME_SC_DNR);
873 cmd->pdu_len = le32_to_cpu(data->data_length);
875 nvmet_tcp_map_pdu_iovec(cmd);
877 queue->rcv_state = NVMET_TCP_RECV_DATA;
882 static int nvmet_tcp_done_recv_pdu(struct nvmet_tcp_queue *queue)
884 struct nvme_tcp_hdr *hdr = &queue->pdu.cmd.hdr;
885 struct nvme_command *nvme_cmd = &queue->pdu.cmd.cmd;
886 struct nvmet_req *req;
889 if (unlikely(queue->state == NVMET_TCP_Q_CONNECTING)) {
890 if (hdr->type != nvme_tcp_icreq) {
891 pr_err("unexpected pdu type (%d) before icreq\n",
893 nvmet_tcp_fatal_error(queue);
896 return nvmet_tcp_handle_icreq(queue);
899 if (hdr->type == nvme_tcp_h2c_data) {
900 ret = nvmet_tcp_handle_h2c_data_pdu(queue);
906 queue->cmd = nvmet_tcp_get_cmd(queue);
907 if (unlikely(!queue->cmd)) {
908 /* This should never happen */
909 pr_err("queue %d: out of commands (%d) send_list_len: %d, opcode: %d",
910 queue->idx, queue->nr_cmds, queue->send_list_len,
911 nvme_cmd->common.opcode);
912 nvmet_tcp_fatal_error(queue);
916 req = &queue->cmd->req;
917 memcpy(req->cmd, nvme_cmd, sizeof(*nvme_cmd));
919 if (unlikely(!nvmet_req_init(req, &queue->nvme_cq,
920 &queue->nvme_sq, &nvmet_tcp_ops))) {
921 pr_err("failed cmd %p id %d opcode %d, data_len: %d\n",
922 req->cmd, req->cmd->common.command_id,
923 req->cmd->common.opcode,
924 le32_to_cpu(req->cmd->common.dptr.sgl.length));
926 nvmet_tcp_handle_req_failure(queue, queue->cmd, req);
930 ret = nvmet_tcp_map_data(queue->cmd);
932 pr_err("queue %d: failed to map data\n", queue->idx);
933 if (nvmet_tcp_has_inline_data(queue->cmd))
934 nvmet_tcp_fatal_error(queue);
936 nvmet_req_complete(req, ret);
941 if (nvmet_tcp_need_data_in(queue->cmd)) {
942 if (nvmet_tcp_has_inline_data(queue->cmd)) {
943 queue->rcv_state = NVMET_TCP_RECV_DATA;
944 nvmet_tcp_map_pdu_iovec(queue->cmd);
948 nvmet_tcp_queue_response(&queue->cmd->req);
952 queue->cmd->req.execute(&queue->cmd->req);
954 nvmet_prepare_receive_pdu(queue);
958 static const u8 nvme_tcp_pdu_sizes[] = {
959 [nvme_tcp_icreq] = sizeof(struct nvme_tcp_icreq_pdu),
960 [nvme_tcp_cmd] = sizeof(struct nvme_tcp_cmd_pdu),
961 [nvme_tcp_h2c_data] = sizeof(struct nvme_tcp_data_pdu),
964 static inline u8 nvmet_tcp_pdu_size(u8 type)
968 return (idx < ARRAY_SIZE(nvme_tcp_pdu_sizes) &&
969 nvme_tcp_pdu_sizes[idx]) ?
970 nvme_tcp_pdu_sizes[idx] : 0;
973 static inline bool nvmet_tcp_pdu_valid(u8 type)
978 case nvme_tcp_h2c_data:
986 static int nvmet_tcp_try_recv_pdu(struct nvmet_tcp_queue *queue)
988 struct nvme_tcp_hdr *hdr = &queue->pdu.cmd.hdr;
991 struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
994 iov.iov_base = (void *)&queue->pdu + queue->offset;
995 iov.iov_len = queue->left;
996 len = kernel_recvmsg(queue->sock, &msg, &iov, 1,
997 iov.iov_len, msg.msg_flags);
998 if (unlikely(len < 0))
1001 queue->offset += len;
1006 if (queue->offset == sizeof(struct nvme_tcp_hdr)) {
1007 u8 hdgst = nvmet_tcp_hdgst_len(queue);
1009 if (unlikely(!nvmet_tcp_pdu_valid(hdr->type))) {
1010 pr_err("unexpected pdu type %d\n", hdr->type);
1011 nvmet_tcp_fatal_error(queue);
1015 if (unlikely(hdr->hlen != nvmet_tcp_pdu_size(hdr->type))) {
1016 pr_err("pdu %d bad hlen %d\n", hdr->type, hdr->hlen);
1020 queue->left = hdr->hlen - queue->offset + hdgst;
1024 if (queue->hdr_digest &&
1025 nvmet_tcp_verify_hdgst(queue, &queue->pdu, queue->offset)) {
1026 nvmet_tcp_fatal_error(queue); /* fatal */
1030 if (queue->data_digest &&
1031 nvmet_tcp_check_ddgst(queue, &queue->pdu)) {
1032 nvmet_tcp_fatal_error(queue); /* fatal */
1036 return nvmet_tcp_done_recv_pdu(queue);
1039 static void nvmet_tcp_prep_recv_ddgst(struct nvmet_tcp_cmd *cmd)
1041 struct nvmet_tcp_queue *queue = cmd->queue;
1043 nvmet_tcp_ddgst(queue->rcv_hash, cmd);
1045 queue->left = NVME_TCP_DIGEST_LENGTH;
1046 queue->rcv_state = NVMET_TCP_RECV_DDGST;
1049 static int nvmet_tcp_try_recv_data(struct nvmet_tcp_queue *queue)
1051 struct nvmet_tcp_cmd *cmd = queue->cmd;
1054 while (msg_data_left(&cmd->recv_msg)) {
1055 ret = sock_recvmsg(cmd->queue->sock, &cmd->recv_msg,
1056 cmd->recv_msg.msg_flags);
1060 cmd->pdu_recv += ret;
1061 cmd->rbytes_done += ret;
1064 nvmet_tcp_unmap_pdu_iovec(cmd);
1066 if (!(cmd->flags & NVMET_TCP_F_INIT_FAILED) &&
1067 cmd->rbytes_done == cmd->req.transfer_len) {
1068 if (queue->data_digest) {
1069 nvmet_tcp_prep_recv_ddgst(cmd);
1072 cmd->req.execute(&cmd->req);
1075 nvmet_prepare_receive_pdu(queue);
1079 static int nvmet_tcp_try_recv_ddgst(struct nvmet_tcp_queue *queue)
1081 struct nvmet_tcp_cmd *cmd = queue->cmd;
1083 struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
1085 .iov_base = (void *)&cmd->recv_ddgst + queue->offset,
1086 .iov_len = queue->left
1089 ret = kernel_recvmsg(queue->sock, &msg, &iov, 1,
1090 iov.iov_len, msg.msg_flags);
1091 if (unlikely(ret < 0))
1094 queue->offset += ret;
1099 if (queue->data_digest && cmd->exp_ddgst != cmd->recv_ddgst) {
1100 pr_err("queue %d: cmd %d pdu (%d) data digest error: recv %#x expected %#x\n",
1101 queue->idx, cmd->req.cmd->common.command_id,
1102 queue->pdu.cmd.hdr.type, le32_to_cpu(cmd->recv_ddgst),
1103 le32_to_cpu(cmd->exp_ddgst));
1104 nvmet_tcp_finish_cmd(cmd);
1105 nvmet_tcp_fatal_error(queue);
1110 if (!(cmd->flags & NVMET_TCP_F_INIT_FAILED) &&
1111 cmd->rbytes_done == cmd->req.transfer_len)
1112 cmd->req.execute(&cmd->req);
1115 nvmet_prepare_receive_pdu(queue);
1119 static int nvmet_tcp_try_recv_one(struct nvmet_tcp_queue *queue)
1123 if (unlikely(queue->rcv_state == NVMET_TCP_RECV_ERR))
1126 if (queue->rcv_state == NVMET_TCP_RECV_PDU) {
1127 result = nvmet_tcp_try_recv_pdu(queue);
1132 if (queue->rcv_state == NVMET_TCP_RECV_DATA) {
1133 result = nvmet_tcp_try_recv_data(queue);
1138 if (queue->rcv_state == NVMET_TCP_RECV_DDGST) {
1139 result = nvmet_tcp_try_recv_ddgst(queue);
1146 if (result == -EAGAIN)
1153 static int nvmet_tcp_try_recv(struct nvmet_tcp_queue *queue,
1154 int budget, int *recvs)
1158 for (i = 0; i < budget; i++) {
1159 ret = nvmet_tcp_try_recv_one(queue);
1168 static void nvmet_tcp_schedule_release_queue(struct nvmet_tcp_queue *queue)
1170 spin_lock(&queue->state_lock);
1171 if (queue->state != NVMET_TCP_Q_DISCONNECTING) {
1172 queue->state = NVMET_TCP_Q_DISCONNECTING;
1173 schedule_work(&queue->release_work);
1175 spin_unlock(&queue->state_lock);
1178 static void nvmet_tcp_io_work(struct work_struct *w)
1180 struct nvmet_tcp_queue *queue =
1181 container_of(w, struct nvmet_tcp_queue, io_work);
1188 ret = nvmet_tcp_try_recv(queue, NVMET_TCP_RECV_BUDGET, &ops);
1191 } else if (ret < 0) {
1192 if (ret == -EPIPE || ret == -ECONNRESET)
1193 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1195 nvmet_tcp_fatal_error(queue);
1199 ret = nvmet_tcp_try_send(queue, NVMET_TCP_SEND_BUDGET, &ops);
1201 /* transmitted message/data */
1203 } else if (ret < 0) {
1204 if (ret == -EPIPE || ret == -ECONNRESET)
1205 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1207 nvmet_tcp_fatal_error(queue);
1211 } while (pending && ops < NVMET_TCP_IO_WORK_BUDGET);
1214 * We exahusted our budget, requeue our selves
1217 queue_work_on(queue->cpu, nvmet_tcp_wq, &queue->io_work);
1220 static int nvmet_tcp_alloc_cmd(struct nvmet_tcp_queue *queue,
1221 struct nvmet_tcp_cmd *c)
1223 u8 hdgst = nvmet_tcp_hdgst_len(queue);
1226 c->req.port = queue->port->nport;
1228 c->cmd_pdu = page_frag_alloc(&queue->pf_cache,
1229 sizeof(*c->cmd_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO);
1232 c->req.cmd = &c->cmd_pdu->cmd;
1234 c->rsp_pdu = page_frag_alloc(&queue->pf_cache,
1235 sizeof(*c->rsp_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO);
1238 c->req.cqe = &c->rsp_pdu->cqe;
1240 c->data_pdu = page_frag_alloc(&queue->pf_cache,
1241 sizeof(*c->data_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO);
1245 c->r2t_pdu = page_frag_alloc(&queue->pf_cache,
1246 sizeof(*c->r2t_pdu) + hdgst, GFP_KERNEL | __GFP_ZERO);
1250 c->recv_msg.msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
1252 list_add_tail(&c->entry, &queue->free_list);
1256 page_frag_free(c->data_pdu);
1258 page_frag_free(c->rsp_pdu);
1260 page_frag_free(c->cmd_pdu);
1264 static void nvmet_tcp_free_cmd(struct nvmet_tcp_cmd *c)
1266 page_frag_free(c->r2t_pdu);
1267 page_frag_free(c->data_pdu);
1268 page_frag_free(c->rsp_pdu);
1269 page_frag_free(c->cmd_pdu);
1272 static int nvmet_tcp_alloc_cmds(struct nvmet_tcp_queue *queue)
1274 struct nvmet_tcp_cmd *cmds;
1275 int i, ret = -EINVAL, nr_cmds = queue->nr_cmds;
1277 cmds = kcalloc(nr_cmds, sizeof(struct nvmet_tcp_cmd), GFP_KERNEL);
1281 for (i = 0; i < nr_cmds; i++) {
1282 ret = nvmet_tcp_alloc_cmd(queue, cmds + i);
1292 nvmet_tcp_free_cmd(cmds + i);
1298 static void nvmet_tcp_free_cmds(struct nvmet_tcp_queue *queue)
1300 struct nvmet_tcp_cmd *cmds = queue->cmds;
1303 for (i = 0; i < queue->nr_cmds; i++)
1304 nvmet_tcp_free_cmd(cmds + i);
1306 nvmet_tcp_free_cmd(&queue->connect);
1310 static void nvmet_tcp_restore_socket_callbacks(struct nvmet_tcp_queue *queue)
1312 struct socket *sock = queue->sock;
1314 write_lock_bh(&sock->sk->sk_callback_lock);
1315 sock->sk->sk_data_ready = queue->data_ready;
1316 sock->sk->sk_state_change = queue->state_change;
1317 sock->sk->sk_write_space = queue->write_space;
1318 sock->sk->sk_user_data = NULL;
1319 write_unlock_bh(&sock->sk->sk_callback_lock);
1322 static void nvmet_tcp_finish_cmd(struct nvmet_tcp_cmd *cmd)
1324 nvmet_req_uninit(&cmd->req);
1325 nvmet_tcp_unmap_pdu_iovec(cmd);
1327 sgl_free(cmd->req.sg);
1330 static void nvmet_tcp_uninit_data_in_cmds(struct nvmet_tcp_queue *queue)
1332 struct nvmet_tcp_cmd *cmd = queue->cmds;
1335 for (i = 0; i < queue->nr_cmds; i++, cmd++) {
1336 if (nvmet_tcp_need_data_in(cmd))
1337 nvmet_tcp_finish_cmd(cmd);
1340 if (!queue->nr_cmds && nvmet_tcp_need_data_in(&queue->connect)) {
1341 /* failed in connect */
1342 nvmet_tcp_finish_cmd(&queue->connect);
1346 static void nvmet_tcp_release_queue_work(struct work_struct *w)
1348 struct nvmet_tcp_queue *queue =
1349 container_of(w, struct nvmet_tcp_queue, release_work);
1351 mutex_lock(&nvmet_tcp_queue_mutex);
1352 list_del_init(&queue->queue_list);
1353 mutex_unlock(&nvmet_tcp_queue_mutex);
1355 nvmet_tcp_restore_socket_callbacks(queue);
1356 flush_work(&queue->io_work);
1358 nvmet_tcp_uninit_data_in_cmds(queue);
1359 nvmet_sq_destroy(&queue->nvme_sq);
1360 cancel_work_sync(&queue->io_work);
1361 sock_release(queue->sock);
1362 nvmet_tcp_free_cmds(queue);
1363 if (queue->hdr_digest || queue->data_digest)
1364 nvmet_tcp_free_crypto(queue);
1365 ida_simple_remove(&nvmet_tcp_queue_ida, queue->idx);
1370 static void nvmet_tcp_data_ready(struct sock *sk)
1372 struct nvmet_tcp_queue *queue;
1374 read_lock_bh(&sk->sk_callback_lock);
1375 queue = sk->sk_user_data;
1377 queue_work_on(queue->cpu, nvmet_tcp_wq, &queue->io_work);
1378 read_unlock_bh(&sk->sk_callback_lock);
1381 static void nvmet_tcp_write_space(struct sock *sk)
1383 struct nvmet_tcp_queue *queue;
1385 read_lock_bh(&sk->sk_callback_lock);
1386 queue = sk->sk_user_data;
1387 if (unlikely(!queue))
1390 if (unlikely(queue->state == NVMET_TCP_Q_CONNECTING)) {
1391 queue->write_space(sk);
1395 if (sk_stream_is_writeable(sk)) {
1396 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1397 queue_work_on(queue->cpu, nvmet_tcp_wq, &queue->io_work);
1400 read_unlock_bh(&sk->sk_callback_lock);
1403 static void nvmet_tcp_state_change(struct sock *sk)
1405 struct nvmet_tcp_queue *queue;
1407 write_lock_bh(&sk->sk_callback_lock);
1408 queue = sk->sk_user_data;
1412 switch (sk->sk_state) {
1414 case TCP_CLOSE_WAIT:
1417 sk->sk_user_data = NULL;
1418 nvmet_tcp_schedule_release_queue(queue);
1421 pr_warn("queue %d unhandled state %d\n",
1422 queue->idx, sk->sk_state);
1425 write_unlock_bh(&sk->sk_callback_lock);
1428 static int nvmet_tcp_set_queue_sock(struct nvmet_tcp_queue *queue)
1430 struct socket *sock = queue->sock;
1431 struct inet_sock *inet = inet_sk(sock->sk);
1432 struct linger sol = { .l_onoff = 1, .l_linger = 0 };
1435 ret = kernel_getsockname(sock,
1436 (struct sockaddr *)&queue->sockaddr);
1440 ret = kernel_getpeername(sock,
1441 (struct sockaddr *)&queue->sockaddr_peer);
1446 * Cleanup whatever is sitting in the TCP transmit queue on socket
1447 * close. This is done to prevent stale data from being sent should
1448 * the network connection be restored before TCP times out.
1450 ret = kernel_setsockopt(sock, SOL_SOCKET, SO_LINGER,
1451 (char *)&sol, sizeof(sol));
1455 if (so_priority > 0) {
1456 ret = kernel_setsockopt(sock, SOL_SOCKET, SO_PRIORITY,
1457 (char *)&so_priority, sizeof(so_priority));
1462 /* Set socket type of service */
1463 if (inet->rcv_tos > 0) {
1464 int tos = inet->rcv_tos;
1466 ret = kernel_setsockopt(sock, SOL_IP, IP_TOS,
1467 (char *)&tos, sizeof(tos));
1472 write_lock_bh(&sock->sk->sk_callback_lock);
1473 sock->sk->sk_user_data = queue;
1474 queue->data_ready = sock->sk->sk_data_ready;
1475 sock->sk->sk_data_ready = nvmet_tcp_data_ready;
1476 queue->state_change = sock->sk->sk_state_change;
1477 sock->sk->sk_state_change = nvmet_tcp_state_change;
1478 queue->write_space = sock->sk->sk_write_space;
1479 sock->sk->sk_write_space = nvmet_tcp_write_space;
1480 write_unlock_bh(&sock->sk->sk_callback_lock);
1485 static int nvmet_tcp_alloc_queue(struct nvmet_tcp_port *port,
1486 struct socket *newsock)
1488 struct nvmet_tcp_queue *queue;
1491 queue = kzalloc(sizeof(*queue), GFP_KERNEL);
1495 INIT_WORK(&queue->release_work, nvmet_tcp_release_queue_work);
1496 INIT_WORK(&queue->io_work, nvmet_tcp_io_work);
1497 queue->sock = newsock;
1500 spin_lock_init(&queue->state_lock);
1501 queue->state = NVMET_TCP_Q_CONNECTING;
1502 INIT_LIST_HEAD(&queue->free_list);
1503 init_llist_head(&queue->resp_list);
1504 INIT_LIST_HEAD(&queue->resp_send_list);
1506 queue->idx = ida_simple_get(&nvmet_tcp_queue_ida, 0, 0, GFP_KERNEL);
1507 if (queue->idx < 0) {
1509 goto out_free_queue;
1512 ret = nvmet_tcp_alloc_cmd(queue, &queue->connect);
1514 goto out_ida_remove;
1516 ret = nvmet_sq_init(&queue->nvme_sq);
1518 goto out_free_connect;
1520 port->last_cpu = cpumask_next_wrap(port->last_cpu,
1521 cpu_online_mask, -1, false);
1522 queue->cpu = port->last_cpu;
1523 nvmet_prepare_receive_pdu(queue);
1525 mutex_lock(&nvmet_tcp_queue_mutex);
1526 list_add_tail(&queue->queue_list, &nvmet_tcp_queue_list);
1527 mutex_unlock(&nvmet_tcp_queue_mutex);
1529 ret = nvmet_tcp_set_queue_sock(queue);
1531 goto out_destroy_sq;
1533 queue_work_on(queue->cpu, nvmet_tcp_wq, &queue->io_work);
1537 mutex_lock(&nvmet_tcp_queue_mutex);
1538 list_del_init(&queue->queue_list);
1539 mutex_unlock(&nvmet_tcp_queue_mutex);
1540 nvmet_sq_destroy(&queue->nvme_sq);
1542 nvmet_tcp_free_cmd(&queue->connect);
1544 ida_simple_remove(&nvmet_tcp_queue_ida, queue->idx);
1550 static void nvmet_tcp_accept_work(struct work_struct *w)
1552 struct nvmet_tcp_port *port =
1553 container_of(w, struct nvmet_tcp_port, accept_work);
1554 struct socket *newsock;
1558 ret = kernel_accept(port->sock, &newsock, O_NONBLOCK);
1561 pr_warn("failed to accept err=%d\n", ret);
1564 ret = nvmet_tcp_alloc_queue(port, newsock);
1566 pr_err("failed to allocate queue\n");
1567 sock_release(newsock);
1572 static void nvmet_tcp_listen_data_ready(struct sock *sk)
1574 struct nvmet_tcp_port *port;
1576 read_lock_bh(&sk->sk_callback_lock);
1577 port = sk->sk_user_data;
1581 if (sk->sk_state == TCP_LISTEN)
1582 schedule_work(&port->accept_work);
1584 read_unlock_bh(&sk->sk_callback_lock);
1587 static int nvmet_tcp_add_port(struct nvmet_port *nport)
1589 struct nvmet_tcp_port *port;
1590 __kernel_sa_family_t af;
1593 port = kzalloc(sizeof(*port), GFP_KERNEL);
1597 switch (nport->disc_addr.adrfam) {
1598 case NVMF_ADDR_FAMILY_IP4:
1601 case NVMF_ADDR_FAMILY_IP6:
1605 pr_err("address family %d not supported\n",
1606 nport->disc_addr.adrfam);
1611 ret = inet_pton_with_scope(&init_net, af, nport->disc_addr.traddr,
1612 nport->disc_addr.trsvcid, &port->addr);
1614 pr_err("malformed ip/port passed: %s:%s\n",
1615 nport->disc_addr.traddr, nport->disc_addr.trsvcid);
1619 port->nport = nport;
1620 port->last_cpu = -1;
1621 INIT_WORK(&port->accept_work, nvmet_tcp_accept_work);
1622 if (port->nport->inline_data_size < 0)
1623 port->nport->inline_data_size = NVMET_TCP_DEF_INLINE_DATA_SIZE;
1625 ret = sock_create(port->addr.ss_family, SOCK_STREAM,
1626 IPPROTO_TCP, &port->sock);
1628 pr_err("failed to create a socket\n");
1632 port->sock->sk->sk_user_data = port;
1633 port->data_ready = port->sock->sk->sk_data_ready;
1634 port->sock->sk->sk_data_ready = nvmet_tcp_listen_data_ready;
1637 ret = kernel_setsockopt(port->sock, IPPROTO_TCP,
1638 TCP_NODELAY, (char *)&opt, sizeof(opt));
1640 pr_err("failed to set TCP_NODELAY sock opt %d\n", ret);
1644 ret = kernel_setsockopt(port->sock, SOL_SOCKET, SO_REUSEADDR,
1645 (char *)&opt, sizeof(opt));
1647 pr_err("failed to set SO_REUSEADDR sock opt %d\n", ret);
1651 if (so_priority > 0) {
1652 ret = kernel_setsockopt(port->sock, SOL_SOCKET, SO_PRIORITY,
1653 (char *)&so_priority, sizeof(so_priority));
1655 pr_err("failed to set SO_PRIORITY sock opt %d\n", ret);
1660 ret = kernel_bind(port->sock, (struct sockaddr *)&port->addr,
1661 sizeof(port->addr));
1663 pr_err("failed to bind port socket %d\n", ret);
1667 ret = kernel_listen(port->sock, 128);
1669 pr_err("failed to listen %d on port sock\n", ret);
1674 pr_info("enabling port %d (%pISpc)\n",
1675 le16_to_cpu(nport->disc_addr.portid), &port->addr);
1680 sock_release(port->sock);
1686 static void nvmet_tcp_remove_port(struct nvmet_port *nport)
1688 struct nvmet_tcp_port *port = nport->priv;
1690 write_lock_bh(&port->sock->sk->sk_callback_lock);
1691 port->sock->sk->sk_data_ready = port->data_ready;
1692 port->sock->sk->sk_user_data = NULL;
1693 write_unlock_bh(&port->sock->sk->sk_callback_lock);
1694 cancel_work_sync(&port->accept_work);
1696 sock_release(port->sock);
1700 static void nvmet_tcp_delete_ctrl(struct nvmet_ctrl *ctrl)
1702 struct nvmet_tcp_queue *queue;
1704 mutex_lock(&nvmet_tcp_queue_mutex);
1705 list_for_each_entry(queue, &nvmet_tcp_queue_list, queue_list)
1706 if (queue->nvme_sq.ctrl == ctrl)
1707 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1708 mutex_unlock(&nvmet_tcp_queue_mutex);
1711 static u16 nvmet_tcp_install_queue(struct nvmet_sq *sq)
1713 struct nvmet_tcp_queue *queue =
1714 container_of(sq, struct nvmet_tcp_queue, nvme_sq);
1717 /* Let inflight controller teardown complete */
1718 flush_scheduled_work();
1721 queue->nr_cmds = sq->size * 2;
1722 if (nvmet_tcp_alloc_cmds(queue))
1723 return NVME_SC_INTERNAL;
1727 static void nvmet_tcp_disc_port_addr(struct nvmet_req *req,
1728 struct nvmet_port *nport, char *traddr)
1730 struct nvmet_tcp_port *port = nport->priv;
1732 if (inet_addr_is_any((struct sockaddr *)&port->addr)) {
1733 struct nvmet_tcp_cmd *cmd =
1734 container_of(req, struct nvmet_tcp_cmd, req);
1735 struct nvmet_tcp_queue *queue = cmd->queue;
1737 sprintf(traddr, "%pISc", (struct sockaddr *)&queue->sockaddr);
1739 memcpy(traddr, nport->disc_addr.traddr, NVMF_TRADDR_SIZE);
1743 static struct nvmet_fabrics_ops nvmet_tcp_ops = {
1744 .owner = THIS_MODULE,
1745 .type = NVMF_TRTYPE_TCP,
1747 .has_keyed_sgls = 0,
1748 .add_port = nvmet_tcp_add_port,
1749 .remove_port = nvmet_tcp_remove_port,
1750 .queue_response = nvmet_tcp_queue_response,
1751 .delete_ctrl = nvmet_tcp_delete_ctrl,
1752 .install_queue = nvmet_tcp_install_queue,
1753 .disc_traddr = nvmet_tcp_disc_port_addr,
1756 static int __init nvmet_tcp_init(void)
1760 nvmet_tcp_wq = alloc_workqueue("nvmet_tcp_wq", WQ_HIGHPRI, 0);
1764 ret = nvmet_register_transport(&nvmet_tcp_ops);
1770 destroy_workqueue(nvmet_tcp_wq);
1774 static void __exit nvmet_tcp_exit(void)
1776 struct nvmet_tcp_queue *queue;
1778 nvmet_unregister_transport(&nvmet_tcp_ops);
1780 flush_scheduled_work();
1781 mutex_lock(&nvmet_tcp_queue_mutex);
1782 list_for_each_entry(queue, &nvmet_tcp_queue_list, queue_list)
1783 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1784 mutex_unlock(&nvmet_tcp_queue_mutex);
1785 flush_scheduled_work();
1787 destroy_workqueue(nvmet_tcp_wq);
1790 module_init(nvmet_tcp_init);
1791 module_exit(nvmet_tcp_exit);
1793 MODULE_LICENSE("GPL v2");
1794 MODULE_ALIAS("nvmet-transport-3"); /* 3 == NVMF_TRTYPE_TCP */