1 // SPDX-License-Identifier: GPL-2.0
3 * NVMe over Fabrics TCP host.
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/blk-mq.h>
15 #include <crypto/hash.h>
16 #include <net/busy_poll.h>
21 struct nvme_tcp_queue;
23 enum nvme_tcp_send_state {
24 NVME_TCP_SEND_CMD_PDU = 0,
25 NVME_TCP_SEND_H2C_PDU,
30 struct nvme_tcp_request {
31 struct nvme_request req;
33 struct nvme_tcp_queue *queue;
38 struct list_head entry;
47 enum nvme_tcp_send_state state;
50 enum nvme_tcp_queue_flags {
51 NVME_TCP_Q_ALLOCATED = 0,
55 enum nvme_tcp_recv_state {
56 NVME_TCP_RECV_PDU = 0,
62 struct nvme_tcp_queue {
64 struct work_struct io_work;
68 struct list_head send_list;
74 size_t data_remaining;
75 size_t ddgst_remaining;
79 struct nvme_tcp_request *request;
82 size_t cmnd_capsule_len;
83 struct nvme_tcp_ctrl *ctrl;
89 struct ahash_request *rcv_hash;
90 struct ahash_request *snd_hash;
94 struct page_frag_cache pf_cache;
96 void (*state_change)(struct sock *);
97 void (*data_ready)(struct sock *);
98 void (*write_space)(struct sock *);
101 struct nvme_tcp_ctrl {
102 /* read only in the hot path */
103 struct nvme_tcp_queue *queues;
104 struct blk_mq_tag_set tag_set;
106 /* other member variables */
107 struct list_head list;
108 struct blk_mq_tag_set admin_tag_set;
109 struct sockaddr_storage addr;
110 struct sockaddr_storage src_addr;
111 struct nvme_ctrl ctrl;
113 struct work_struct err_work;
114 struct delayed_work connect_work;
115 struct nvme_tcp_request async_req;
116 u32 io_queues[HCTX_MAX_TYPES];
119 static LIST_HEAD(nvme_tcp_ctrl_list);
120 static DEFINE_MUTEX(nvme_tcp_ctrl_mutex);
121 static struct workqueue_struct *nvme_tcp_wq;
122 static struct blk_mq_ops nvme_tcp_mq_ops;
123 static struct blk_mq_ops nvme_tcp_admin_mq_ops;
125 static inline struct nvme_tcp_ctrl *to_tcp_ctrl(struct nvme_ctrl *ctrl)
127 return container_of(ctrl, struct nvme_tcp_ctrl, ctrl);
130 static inline int nvme_tcp_queue_id(struct nvme_tcp_queue *queue)
132 return queue - queue->ctrl->queues;
135 static inline struct blk_mq_tags *nvme_tcp_tagset(struct nvme_tcp_queue *queue)
137 u32 queue_idx = nvme_tcp_queue_id(queue);
140 return queue->ctrl->admin_tag_set.tags[queue_idx];
141 return queue->ctrl->tag_set.tags[queue_idx - 1];
144 static inline u8 nvme_tcp_hdgst_len(struct nvme_tcp_queue *queue)
146 return queue->hdr_digest ? NVME_TCP_DIGEST_LENGTH : 0;
149 static inline u8 nvme_tcp_ddgst_len(struct nvme_tcp_queue *queue)
151 return queue->data_digest ? NVME_TCP_DIGEST_LENGTH : 0;
154 static inline size_t nvme_tcp_inline_data_size(struct nvme_tcp_queue *queue)
156 return queue->cmnd_capsule_len - sizeof(struct nvme_command);
159 static inline bool nvme_tcp_async_req(struct nvme_tcp_request *req)
161 return req == &req->queue->ctrl->async_req;
164 static inline bool nvme_tcp_has_inline_data(struct nvme_tcp_request *req)
169 if (unlikely(nvme_tcp_async_req(req)))
170 return false; /* async events don't have a request */
172 rq = blk_mq_rq_from_pdu(req);
173 bytes = blk_rq_payload_bytes(rq);
175 return rq_data_dir(rq) == WRITE && bytes &&
176 bytes <= nvme_tcp_inline_data_size(req->queue);
179 static inline struct page *nvme_tcp_req_cur_page(struct nvme_tcp_request *req)
181 return req->iter.bvec->bv_page;
184 static inline size_t nvme_tcp_req_cur_offset(struct nvme_tcp_request *req)
186 return req->iter.bvec->bv_offset + req->iter.iov_offset;
189 static inline size_t nvme_tcp_req_cur_length(struct nvme_tcp_request *req)
191 return min_t(size_t, req->iter.bvec->bv_len - req->iter.iov_offset,
192 req->pdu_len - req->pdu_sent);
195 static inline size_t nvme_tcp_req_offset(struct nvme_tcp_request *req)
197 return req->iter.iov_offset;
200 static inline size_t nvme_tcp_pdu_data_left(struct nvme_tcp_request *req)
202 return rq_data_dir(blk_mq_rq_from_pdu(req)) == WRITE ?
203 req->pdu_len - req->pdu_sent : 0;
206 static inline size_t nvme_tcp_pdu_last_send(struct nvme_tcp_request *req,
209 return nvme_tcp_pdu_data_left(req) <= len;
212 static void nvme_tcp_init_iter(struct nvme_tcp_request *req,
215 struct request *rq = blk_mq_rq_from_pdu(req);
221 if (rq->rq_flags & RQF_SPECIAL_PAYLOAD) {
222 vec = &rq->special_vec;
224 size = blk_rq_payload_bytes(rq);
227 struct bio *bio = req->curr_bio;
229 vec = __bvec_iter_bvec(bio->bi_io_vec, bio->bi_iter);
230 nsegs = bio_segments(bio);
231 size = bio->bi_iter.bi_size;
232 offset = bio->bi_iter.bi_bvec_done;
235 iov_iter_bvec(&req->iter, dir, vec, nsegs, size);
236 req->iter.iov_offset = offset;
239 static inline void nvme_tcp_advance_req(struct nvme_tcp_request *req,
242 req->data_sent += len;
243 req->pdu_sent += len;
244 iov_iter_advance(&req->iter, len);
245 if (!iov_iter_count(&req->iter) &&
246 req->data_sent < req->data_len) {
247 req->curr_bio = req->curr_bio->bi_next;
248 nvme_tcp_init_iter(req, WRITE);
252 static inline void nvme_tcp_queue_request(struct nvme_tcp_request *req)
254 struct nvme_tcp_queue *queue = req->queue;
256 spin_lock(&queue->lock);
257 list_add_tail(&req->entry, &queue->send_list);
258 spin_unlock(&queue->lock);
260 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
263 static inline struct nvme_tcp_request *
264 nvme_tcp_fetch_request(struct nvme_tcp_queue *queue)
266 struct nvme_tcp_request *req;
268 spin_lock(&queue->lock);
269 req = list_first_entry_or_null(&queue->send_list,
270 struct nvme_tcp_request, entry);
272 list_del(&req->entry);
273 spin_unlock(&queue->lock);
278 static inline void nvme_tcp_ddgst_final(struct ahash_request *hash,
281 ahash_request_set_crypt(hash, NULL, (u8 *)dgst, 0);
282 crypto_ahash_final(hash);
285 static inline void nvme_tcp_ddgst_update(struct ahash_request *hash,
286 struct page *page, off_t off, size_t len)
288 struct scatterlist sg;
290 sg_init_marker(&sg, 1);
291 sg_set_page(&sg, page, len, off);
292 ahash_request_set_crypt(hash, &sg, NULL, len);
293 crypto_ahash_update(hash);
296 static inline void nvme_tcp_hdgst(struct ahash_request *hash,
297 void *pdu, size_t len)
299 struct scatterlist sg;
301 sg_init_one(&sg, pdu, len);
302 ahash_request_set_crypt(hash, &sg, pdu + len, len);
303 crypto_ahash_digest(hash);
306 static int nvme_tcp_verify_hdgst(struct nvme_tcp_queue *queue,
307 void *pdu, size_t pdu_len)
309 struct nvme_tcp_hdr *hdr = pdu;
313 if (unlikely(!(hdr->flags & NVME_TCP_F_HDGST))) {
314 dev_err(queue->ctrl->ctrl.device,
315 "queue %d: header digest flag is cleared\n",
316 nvme_tcp_queue_id(queue));
320 recv_digest = *(__le32 *)(pdu + hdr->hlen);
321 nvme_tcp_hdgst(queue->rcv_hash, pdu, pdu_len);
322 exp_digest = *(__le32 *)(pdu + hdr->hlen);
323 if (recv_digest != exp_digest) {
324 dev_err(queue->ctrl->ctrl.device,
325 "header digest error: recv %#x expected %#x\n",
326 le32_to_cpu(recv_digest), le32_to_cpu(exp_digest));
333 static int nvme_tcp_check_ddgst(struct nvme_tcp_queue *queue, void *pdu)
335 struct nvme_tcp_hdr *hdr = pdu;
336 u8 digest_len = nvme_tcp_hdgst_len(queue);
339 len = le32_to_cpu(hdr->plen) - hdr->hlen -
340 ((hdr->flags & NVME_TCP_F_HDGST) ? digest_len : 0);
342 if (unlikely(len && !(hdr->flags & NVME_TCP_F_DDGST))) {
343 dev_err(queue->ctrl->ctrl.device,
344 "queue %d: data digest flag is cleared\n",
345 nvme_tcp_queue_id(queue));
348 crypto_ahash_init(queue->rcv_hash);
353 static void nvme_tcp_exit_request(struct blk_mq_tag_set *set,
354 struct request *rq, unsigned int hctx_idx)
356 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
358 page_frag_free(req->pdu);
361 static int nvme_tcp_init_request(struct blk_mq_tag_set *set,
362 struct request *rq, unsigned int hctx_idx,
363 unsigned int numa_node)
365 struct nvme_tcp_ctrl *ctrl = set->driver_data;
366 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
367 int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
368 struct nvme_tcp_queue *queue = &ctrl->queues[queue_idx];
369 u8 hdgst = nvme_tcp_hdgst_len(queue);
371 req->pdu = page_frag_alloc(&queue->pf_cache,
372 sizeof(struct nvme_tcp_cmd_pdu) + hdgst,
373 GFP_KERNEL | __GFP_ZERO);
378 nvme_req(rq)->ctrl = &ctrl->ctrl;
383 static int nvme_tcp_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
384 unsigned int hctx_idx)
386 struct nvme_tcp_ctrl *ctrl = data;
387 struct nvme_tcp_queue *queue = &ctrl->queues[hctx_idx + 1];
389 hctx->driver_data = queue;
393 static int nvme_tcp_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
394 unsigned int hctx_idx)
396 struct nvme_tcp_ctrl *ctrl = data;
397 struct nvme_tcp_queue *queue = &ctrl->queues[0];
399 hctx->driver_data = queue;
403 static enum nvme_tcp_recv_state
404 nvme_tcp_recv_state(struct nvme_tcp_queue *queue)
406 return (queue->pdu_remaining) ? NVME_TCP_RECV_PDU :
407 (queue->ddgst_remaining) ? NVME_TCP_RECV_DDGST :
411 static void nvme_tcp_init_recv_ctx(struct nvme_tcp_queue *queue)
413 queue->pdu_remaining = sizeof(struct nvme_tcp_rsp_pdu) +
414 nvme_tcp_hdgst_len(queue);
415 queue->pdu_offset = 0;
416 queue->data_remaining = -1;
417 queue->ddgst_remaining = 0;
420 static void nvme_tcp_error_recovery(struct nvme_ctrl *ctrl)
422 if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_RESETTING))
425 queue_work(nvme_wq, &to_tcp_ctrl(ctrl)->err_work);
428 static int nvme_tcp_process_nvme_cqe(struct nvme_tcp_queue *queue,
429 struct nvme_completion *cqe)
433 rq = blk_mq_tag_to_rq(nvme_tcp_tagset(queue), cqe->command_id);
435 dev_err(queue->ctrl->ctrl.device,
436 "queue %d tag 0x%x not found\n",
437 nvme_tcp_queue_id(queue), cqe->command_id);
438 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
442 nvme_end_request(rq, cqe->status, cqe->result);
448 static int nvme_tcp_handle_c2h_data(struct nvme_tcp_queue *queue,
449 struct nvme_tcp_data_pdu *pdu)
453 rq = blk_mq_tag_to_rq(nvme_tcp_tagset(queue), pdu->command_id);
455 dev_err(queue->ctrl->ctrl.device,
456 "queue %d tag %#x not found\n",
457 nvme_tcp_queue_id(queue), pdu->command_id);
461 if (!blk_rq_payload_bytes(rq)) {
462 dev_err(queue->ctrl->ctrl.device,
463 "queue %d tag %#x unexpected data\n",
464 nvme_tcp_queue_id(queue), rq->tag);
468 queue->data_remaining = le32_to_cpu(pdu->data_length);
470 if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS &&
471 unlikely(!(pdu->hdr.flags & NVME_TCP_F_DATA_LAST))) {
472 dev_err(queue->ctrl->ctrl.device,
473 "queue %d tag %#x SUCCESS set but not last PDU\n",
474 nvme_tcp_queue_id(queue), rq->tag);
475 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
482 static int nvme_tcp_handle_comp(struct nvme_tcp_queue *queue,
483 struct nvme_tcp_rsp_pdu *pdu)
485 struct nvme_completion *cqe = &pdu->cqe;
489 * AEN requests are special as they don't time out and can
490 * survive any kind of queue freeze and often don't respond to
491 * aborts. We don't even bother to allocate a struct request
492 * for them but rather special case them here.
494 if (unlikely(nvme_tcp_queue_id(queue) == 0 &&
495 cqe->command_id >= NVME_AQ_BLK_MQ_DEPTH))
496 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
499 ret = nvme_tcp_process_nvme_cqe(queue, cqe);
504 static int nvme_tcp_setup_h2c_data_pdu(struct nvme_tcp_request *req,
505 struct nvme_tcp_r2t_pdu *pdu)
507 struct nvme_tcp_data_pdu *data = req->pdu;
508 struct nvme_tcp_queue *queue = req->queue;
509 struct request *rq = blk_mq_rq_from_pdu(req);
510 u8 hdgst = nvme_tcp_hdgst_len(queue);
511 u8 ddgst = nvme_tcp_ddgst_len(queue);
513 req->pdu_len = le32_to_cpu(pdu->r2t_length);
516 if (unlikely(req->data_sent + req->pdu_len > req->data_len)) {
517 dev_err(queue->ctrl->ctrl.device,
518 "req %d r2t len %u exceeded data len %u (%zu sent)\n",
519 rq->tag, req->pdu_len, req->data_len,
524 if (unlikely(le32_to_cpu(pdu->r2t_offset) < req->data_sent)) {
525 dev_err(queue->ctrl->ctrl.device,
526 "req %d unexpected r2t offset %u (expected %zu)\n",
527 rq->tag, le32_to_cpu(pdu->r2t_offset),
532 memset(data, 0, sizeof(*data));
533 data->hdr.type = nvme_tcp_h2c_data;
534 data->hdr.flags = NVME_TCP_F_DATA_LAST;
535 if (queue->hdr_digest)
536 data->hdr.flags |= NVME_TCP_F_HDGST;
537 if (queue->data_digest)
538 data->hdr.flags |= NVME_TCP_F_DDGST;
539 data->hdr.hlen = sizeof(*data);
540 data->hdr.pdo = data->hdr.hlen + hdgst;
542 cpu_to_le32(data->hdr.hlen + hdgst + req->pdu_len + ddgst);
543 data->ttag = pdu->ttag;
544 data->command_id = rq->tag;
545 data->data_offset = cpu_to_le32(req->data_sent);
546 data->data_length = cpu_to_le32(req->pdu_len);
550 static int nvme_tcp_handle_r2t(struct nvme_tcp_queue *queue,
551 struct nvme_tcp_r2t_pdu *pdu)
553 struct nvme_tcp_request *req;
557 rq = blk_mq_tag_to_rq(nvme_tcp_tagset(queue), pdu->command_id);
559 dev_err(queue->ctrl->ctrl.device,
560 "queue %d tag %#x not found\n",
561 nvme_tcp_queue_id(queue), pdu->command_id);
564 req = blk_mq_rq_to_pdu(rq);
566 ret = nvme_tcp_setup_h2c_data_pdu(req, pdu);
570 req->state = NVME_TCP_SEND_H2C_PDU;
573 nvme_tcp_queue_request(req);
578 static int nvme_tcp_recv_pdu(struct nvme_tcp_queue *queue, struct sk_buff *skb,
579 unsigned int *offset, size_t *len)
581 struct nvme_tcp_hdr *hdr;
582 char *pdu = queue->pdu;
583 size_t rcv_len = min_t(size_t, *len, queue->pdu_remaining);
586 ret = skb_copy_bits(skb, *offset,
587 &pdu[queue->pdu_offset], rcv_len);
591 queue->pdu_remaining -= rcv_len;
592 queue->pdu_offset += rcv_len;
595 if (queue->pdu_remaining)
599 if (queue->hdr_digest) {
600 ret = nvme_tcp_verify_hdgst(queue, queue->pdu, hdr->hlen);
606 if (queue->data_digest) {
607 ret = nvme_tcp_check_ddgst(queue, queue->pdu);
613 case nvme_tcp_c2h_data:
614 return nvme_tcp_handle_c2h_data(queue, (void *)queue->pdu);
616 nvme_tcp_init_recv_ctx(queue);
617 return nvme_tcp_handle_comp(queue, (void *)queue->pdu);
619 nvme_tcp_init_recv_ctx(queue);
620 return nvme_tcp_handle_r2t(queue, (void *)queue->pdu);
622 dev_err(queue->ctrl->ctrl.device,
623 "unsupported pdu type (%d)\n", hdr->type);
628 static inline void nvme_tcp_end_request(struct request *rq, u16 status)
630 union nvme_result res = {};
632 nvme_end_request(rq, cpu_to_le16(status << 1), res);
635 static int nvme_tcp_recv_data(struct nvme_tcp_queue *queue, struct sk_buff *skb,
636 unsigned int *offset, size_t *len)
638 struct nvme_tcp_data_pdu *pdu = (void *)queue->pdu;
639 struct nvme_tcp_request *req;
642 rq = blk_mq_tag_to_rq(nvme_tcp_tagset(queue), pdu->command_id);
644 dev_err(queue->ctrl->ctrl.device,
645 "queue %d tag %#x not found\n",
646 nvme_tcp_queue_id(queue), pdu->command_id);
649 req = blk_mq_rq_to_pdu(rq);
654 recv_len = min_t(size_t, *len, queue->data_remaining);
658 if (!iov_iter_count(&req->iter)) {
659 req->curr_bio = req->curr_bio->bi_next;
662 * If we don`t have any bios it means that controller
663 * sent more data than we requested, hence error
665 if (!req->curr_bio) {
666 dev_err(queue->ctrl->ctrl.device,
667 "queue %d no space in request %#x",
668 nvme_tcp_queue_id(queue), rq->tag);
669 nvme_tcp_init_recv_ctx(queue);
672 nvme_tcp_init_iter(req, READ);
675 /* we can read only from what is left in this bio */
676 recv_len = min_t(size_t, recv_len,
677 iov_iter_count(&req->iter));
679 if (queue->data_digest)
680 ret = skb_copy_and_hash_datagram_iter(skb, *offset,
681 &req->iter, recv_len, queue->rcv_hash);
683 ret = skb_copy_datagram_iter(skb, *offset,
684 &req->iter, recv_len);
686 dev_err(queue->ctrl->ctrl.device,
687 "queue %d failed to copy request %#x data",
688 nvme_tcp_queue_id(queue), rq->tag);
694 queue->data_remaining -= recv_len;
697 if (!queue->data_remaining) {
698 if (queue->data_digest) {
699 nvme_tcp_ddgst_final(queue->rcv_hash, &queue->exp_ddgst);
700 queue->ddgst_remaining = NVME_TCP_DIGEST_LENGTH;
702 if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
703 nvme_tcp_end_request(rq, NVME_SC_SUCCESS);
706 nvme_tcp_init_recv_ctx(queue);
713 static int nvme_tcp_recv_ddgst(struct nvme_tcp_queue *queue,
714 struct sk_buff *skb, unsigned int *offset, size_t *len)
716 struct nvme_tcp_data_pdu *pdu = (void *)queue->pdu;
717 char *ddgst = (char *)&queue->recv_ddgst;
718 size_t recv_len = min_t(size_t, *len, queue->ddgst_remaining);
719 off_t off = NVME_TCP_DIGEST_LENGTH - queue->ddgst_remaining;
722 ret = skb_copy_bits(skb, *offset, &ddgst[off], recv_len);
726 queue->ddgst_remaining -= recv_len;
729 if (queue->ddgst_remaining)
732 if (queue->recv_ddgst != queue->exp_ddgst) {
733 dev_err(queue->ctrl->ctrl.device,
734 "data digest error: recv %#x expected %#x\n",
735 le32_to_cpu(queue->recv_ddgst),
736 le32_to_cpu(queue->exp_ddgst));
740 if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
741 struct request *rq = blk_mq_tag_to_rq(nvme_tcp_tagset(queue),
744 nvme_tcp_end_request(rq, NVME_SC_SUCCESS);
748 nvme_tcp_init_recv_ctx(queue);
752 static int nvme_tcp_recv_skb(read_descriptor_t *desc, struct sk_buff *skb,
753 unsigned int offset, size_t len)
755 struct nvme_tcp_queue *queue = desc->arg.data;
756 size_t consumed = len;
760 switch (nvme_tcp_recv_state(queue)) {
761 case NVME_TCP_RECV_PDU:
762 result = nvme_tcp_recv_pdu(queue, skb, &offset, &len);
764 case NVME_TCP_RECV_DATA:
765 result = nvme_tcp_recv_data(queue, skb, &offset, &len);
767 case NVME_TCP_RECV_DDGST:
768 result = nvme_tcp_recv_ddgst(queue, skb, &offset, &len);
774 dev_err(queue->ctrl->ctrl.device,
775 "receive failed: %d\n", result);
776 queue->rd_enabled = false;
777 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
785 static void nvme_tcp_data_ready(struct sock *sk)
787 struct nvme_tcp_queue *queue;
789 read_lock(&sk->sk_callback_lock);
790 queue = sk->sk_user_data;
791 if (likely(queue && queue->rd_enabled))
792 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
793 read_unlock(&sk->sk_callback_lock);
796 static void nvme_tcp_write_space(struct sock *sk)
798 struct nvme_tcp_queue *queue;
800 read_lock_bh(&sk->sk_callback_lock);
801 queue = sk->sk_user_data;
802 if (likely(queue && sk_stream_is_writeable(sk))) {
803 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
804 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
806 read_unlock_bh(&sk->sk_callback_lock);
809 static void nvme_tcp_state_change(struct sock *sk)
811 struct nvme_tcp_queue *queue;
813 read_lock(&sk->sk_callback_lock);
814 queue = sk->sk_user_data;
818 switch (sk->sk_state) {
825 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
828 dev_info(queue->ctrl->ctrl.device,
829 "queue %d socket state %d\n",
830 nvme_tcp_queue_id(queue), sk->sk_state);
833 queue->state_change(sk);
835 read_unlock(&sk->sk_callback_lock);
838 static inline void nvme_tcp_done_send_req(struct nvme_tcp_queue *queue)
840 queue->request = NULL;
843 static void nvme_tcp_fail_request(struct nvme_tcp_request *req)
845 nvme_tcp_end_request(blk_mq_rq_from_pdu(req), NVME_SC_DATA_XFER_ERROR);
848 static int nvme_tcp_try_send_data(struct nvme_tcp_request *req)
850 struct nvme_tcp_queue *queue = req->queue;
853 struct page *page = nvme_tcp_req_cur_page(req);
854 size_t offset = nvme_tcp_req_cur_offset(req);
855 size_t len = nvme_tcp_req_cur_length(req);
856 bool last = nvme_tcp_pdu_last_send(req, len);
857 int ret, flags = MSG_DONTWAIT;
859 if (last && !queue->data_digest)
864 /* can't zcopy slab pages */
865 if (unlikely(PageSlab(page))) {
866 ret = sock_no_sendpage(queue->sock, page, offset, len,
869 ret = kernel_sendpage(queue->sock, page, offset, len,
875 nvme_tcp_advance_req(req, ret);
876 if (queue->data_digest)
877 nvme_tcp_ddgst_update(queue->snd_hash, page,
880 /* fully successful last write*/
881 if (last && ret == len) {
882 if (queue->data_digest) {
883 nvme_tcp_ddgst_final(queue->snd_hash,
885 req->state = NVME_TCP_SEND_DDGST;
888 nvme_tcp_done_send_req(queue);
896 static int nvme_tcp_try_send_cmd_pdu(struct nvme_tcp_request *req)
898 struct nvme_tcp_queue *queue = req->queue;
899 struct nvme_tcp_cmd_pdu *pdu = req->pdu;
900 bool inline_data = nvme_tcp_has_inline_data(req);
901 int flags = MSG_DONTWAIT | (inline_data ? MSG_MORE : MSG_EOR);
902 u8 hdgst = nvme_tcp_hdgst_len(queue);
903 int len = sizeof(*pdu) + hdgst - req->offset;
906 if (queue->hdr_digest && !req->offset)
907 nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
909 ret = kernel_sendpage(queue->sock, virt_to_page(pdu),
910 offset_in_page(pdu) + req->offset, len, flags);
911 if (unlikely(ret <= 0))
917 req->state = NVME_TCP_SEND_DATA;
918 if (queue->data_digest)
919 crypto_ahash_init(queue->snd_hash);
920 nvme_tcp_init_iter(req, WRITE);
922 nvme_tcp_done_send_req(queue);
931 static int nvme_tcp_try_send_data_pdu(struct nvme_tcp_request *req)
933 struct nvme_tcp_queue *queue = req->queue;
934 struct nvme_tcp_data_pdu *pdu = req->pdu;
935 u8 hdgst = nvme_tcp_hdgst_len(queue);
936 int len = sizeof(*pdu) - req->offset + hdgst;
939 if (queue->hdr_digest && !req->offset)
940 nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
942 ret = kernel_sendpage(queue->sock, virt_to_page(pdu),
943 offset_in_page(pdu) + req->offset, len,
944 MSG_DONTWAIT | MSG_MORE);
945 if (unlikely(ret <= 0))
950 req->state = NVME_TCP_SEND_DATA;
951 if (queue->data_digest)
952 crypto_ahash_init(queue->snd_hash);
954 nvme_tcp_init_iter(req, WRITE);
962 static int nvme_tcp_try_send_ddgst(struct nvme_tcp_request *req)
964 struct nvme_tcp_queue *queue = req->queue;
966 struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_EOR };
968 .iov_base = &req->ddgst + req->offset,
969 .iov_len = NVME_TCP_DIGEST_LENGTH - req->offset
972 ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
973 if (unlikely(ret <= 0))
976 if (req->offset + ret == NVME_TCP_DIGEST_LENGTH) {
977 nvme_tcp_done_send_req(queue);
985 static int nvme_tcp_try_send(struct nvme_tcp_queue *queue)
987 struct nvme_tcp_request *req;
990 if (!queue->request) {
991 queue->request = nvme_tcp_fetch_request(queue);
995 req = queue->request;
997 if (req->state == NVME_TCP_SEND_CMD_PDU) {
998 ret = nvme_tcp_try_send_cmd_pdu(req);
1001 if (!nvme_tcp_has_inline_data(req))
1005 if (req->state == NVME_TCP_SEND_H2C_PDU) {
1006 ret = nvme_tcp_try_send_data_pdu(req);
1011 if (req->state == NVME_TCP_SEND_DATA) {
1012 ret = nvme_tcp_try_send_data(req);
1017 if (req->state == NVME_TCP_SEND_DDGST)
1018 ret = nvme_tcp_try_send_ddgst(req);
1025 static int nvme_tcp_try_recv(struct nvme_tcp_queue *queue)
1027 struct socket *sock = queue->sock;
1028 struct sock *sk = sock->sk;
1029 read_descriptor_t rd_desc;
1032 rd_desc.arg.data = queue;
1036 consumed = sock->ops->read_sock(sk, &rd_desc, nvme_tcp_recv_skb);
1041 static void nvme_tcp_io_work(struct work_struct *w)
1043 struct nvme_tcp_queue *queue =
1044 container_of(w, struct nvme_tcp_queue, io_work);
1045 unsigned long start = jiffies + msecs_to_jiffies(1);
1048 bool pending = false;
1051 result = nvme_tcp_try_send(queue);
1054 } else if (unlikely(result < 0)) {
1055 dev_err(queue->ctrl->ctrl.device,
1056 "failed to send request %d\n", result);
1057 if (result != -EPIPE)
1058 nvme_tcp_fail_request(queue->request);
1059 nvme_tcp_done_send_req(queue);
1063 result = nvme_tcp_try_recv(queue);
1070 } while (time_after(jiffies, start)); /* quota is exhausted */
1072 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
1075 static void nvme_tcp_free_crypto(struct nvme_tcp_queue *queue)
1077 struct crypto_ahash *tfm = crypto_ahash_reqtfm(queue->rcv_hash);
1079 ahash_request_free(queue->rcv_hash);
1080 ahash_request_free(queue->snd_hash);
1081 crypto_free_ahash(tfm);
1084 static int nvme_tcp_alloc_crypto(struct nvme_tcp_queue *queue)
1086 struct crypto_ahash *tfm;
1088 tfm = crypto_alloc_ahash("crc32c", 0, CRYPTO_ALG_ASYNC);
1090 return PTR_ERR(tfm);
1092 queue->snd_hash = ahash_request_alloc(tfm, GFP_KERNEL);
1093 if (!queue->snd_hash)
1095 ahash_request_set_callback(queue->snd_hash, 0, NULL, NULL);
1097 queue->rcv_hash = ahash_request_alloc(tfm, GFP_KERNEL);
1098 if (!queue->rcv_hash)
1100 ahash_request_set_callback(queue->rcv_hash, 0, NULL, NULL);
1104 ahash_request_free(queue->snd_hash);
1106 crypto_free_ahash(tfm);
1110 static void nvme_tcp_free_async_req(struct nvme_tcp_ctrl *ctrl)
1112 struct nvme_tcp_request *async = &ctrl->async_req;
1114 page_frag_free(async->pdu);
1117 static int nvme_tcp_alloc_async_req(struct nvme_tcp_ctrl *ctrl)
1119 struct nvme_tcp_queue *queue = &ctrl->queues[0];
1120 struct nvme_tcp_request *async = &ctrl->async_req;
1121 u8 hdgst = nvme_tcp_hdgst_len(queue);
1123 async->pdu = page_frag_alloc(&queue->pf_cache,
1124 sizeof(struct nvme_tcp_cmd_pdu) + hdgst,
1125 GFP_KERNEL | __GFP_ZERO);
1129 async->queue = &ctrl->queues[0];
1133 static void nvme_tcp_free_queue(struct nvme_ctrl *nctrl, int qid)
1135 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1136 struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1138 if (!test_and_clear_bit(NVME_TCP_Q_ALLOCATED, &queue->flags))
1141 if (queue->hdr_digest || queue->data_digest)
1142 nvme_tcp_free_crypto(queue);
1144 sock_release(queue->sock);
1148 static int nvme_tcp_init_connection(struct nvme_tcp_queue *queue)
1150 struct nvme_tcp_icreq_pdu *icreq;
1151 struct nvme_tcp_icresp_pdu *icresp;
1152 struct msghdr msg = {};
1154 bool ctrl_hdgst, ctrl_ddgst;
1157 icreq = kzalloc(sizeof(*icreq), GFP_KERNEL);
1161 icresp = kzalloc(sizeof(*icresp), GFP_KERNEL);
1167 icreq->hdr.type = nvme_tcp_icreq;
1168 icreq->hdr.hlen = sizeof(*icreq);
1170 icreq->hdr.plen = cpu_to_le32(icreq->hdr.hlen);
1171 icreq->pfv = cpu_to_le16(NVME_TCP_PFV_1_0);
1172 icreq->maxr2t = 0; /* single inflight r2t supported */
1173 icreq->hpda = 0; /* no alignment constraint */
1174 if (queue->hdr_digest)
1175 icreq->digest |= NVME_TCP_HDR_DIGEST_ENABLE;
1176 if (queue->data_digest)
1177 icreq->digest |= NVME_TCP_DATA_DIGEST_ENABLE;
1179 iov.iov_base = icreq;
1180 iov.iov_len = sizeof(*icreq);
1181 ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
1185 memset(&msg, 0, sizeof(msg));
1186 iov.iov_base = icresp;
1187 iov.iov_len = sizeof(*icresp);
1188 ret = kernel_recvmsg(queue->sock, &msg, &iov, 1,
1189 iov.iov_len, msg.msg_flags);
1194 if (icresp->hdr.type != nvme_tcp_icresp) {
1195 pr_err("queue %d: bad type returned %d\n",
1196 nvme_tcp_queue_id(queue), icresp->hdr.type);
1200 if (le32_to_cpu(icresp->hdr.plen) != sizeof(*icresp)) {
1201 pr_err("queue %d: bad pdu length returned %d\n",
1202 nvme_tcp_queue_id(queue), icresp->hdr.plen);
1206 if (icresp->pfv != NVME_TCP_PFV_1_0) {
1207 pr_err("queue %d: bad pfv returned %d\n",
1208 nvme_tcp_queue_id(queue), icresp->pfv);
1212 ctrl_ddgst = !!(icresp->digest & NVME_TCP_DATA_DIGEST_ENABLE);
1213 if ((queue->data_digest && !ctrl_ddgst) ||
1214 (!queue->data_digest && ctrl_ddgst)) {
1215 pr_err("queue %d: data digest mismatch host: %s ctrl: %s\n",
1216 nvme_tcp_queue_id(queue),
1217 queue->data_digest ? "enabled" : "disabled",
1218 ctrl_ddgst ? "enabled" : "disabled");
1222 ctrl_hdgst = !!(icresp->digest & NVME_TCP_HDR_DIGEST_ENABLE);
1223 if ((queue->hdr_digest && !ctrl_hdgst) ||
1224 (!queue->hdr_digest && ctrl_hdgst)) {
1225 pr_err("queue %d: header digest mismatch host: %s ctrl: %s\n",
1226 nvme_tcp_queue_id(queue),
1227 queue->hdr_digest ? "enabled" : "disabled",
1228 ctrl_hdgst ? "enabled" : "disabled");
1232 if (icresp->cpda != 0) {
1233 pr_err("queue %d: unsupported cpda returned %d\n",
1234 nvme_tcp_queue_id(queue), icresp->cpda);
1246 static int nvme_tcp_alloc_queue(struct nvme_ctrl *nctrl,
1247 int qid, size_t queue_size)
1249 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1250 struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1251 struct linger sol = { .l_onoff = 1, .l_linger = 0 };
1252 int ret, opt, rcv_pdu_size, n;
1255 INIT_LIST_HEAD(&queue->send_list);
1256 spin_lock_init(&queue->lock);
1257 INIT_WORK(&queue->io_work, nvme_tcp_io_work);
1258 queue->queue_size = queue_size;
1261 queue->cmnd_capsule_len = nctrl->ioccsz * 16;
1263 queue->cmnd_capsule_len = sizeof(struct nvme_command) +
1264 NVME_TCP_ADMIN_CCSZ;
1266 ret = sock_create(ctrl->addr.ss_family, SOCK_STREAM,
1267 IPPROTO_TCP, &queue->sock);
1269 dev_err(nctrl->device,
1270 "failed to create socket: %d\n", ret);
1274 /* Single syn retry */
1276 ret = kernel_setsockopt(queue->sock, IPPROTO_TCP, TCP_SYNCNT,
1277 (char *)&opt, sizeof(opt));
1279 dev_err(nctrl->device,
1280 "failed to set TCP_SYNCNT sock opt %d\n", ret);
1284 /* Set TCP no delay */
1286 ret = kernel_setsockopt(queue->sock, IPPROTO_TCP,
1287 TCP_NODELAY, (char *)&opt, sizeof(opt));
1289 dev_err(nctrl->device,
1290 "failed to set TCP_NODELAY sock opt %d\n", ret);
1295 * Cleanup whatever is sitting in the TCP transmit queue on socket
1296 * close. This is done to prevent stale data from being sent should
1297 * the network connection be restored before TCP times out.
1299 ret = kernel_setsockopt(queue->sock, SOL_SOCKET, SO_LINGER,
1300 (char *)&sol, sizeof(sol));
1302 dev_err(nctrl->device,
1303 "failed to set SO_LINGER sock opt %d\n", ret);
1307 /* Set socket type of service */
1308 if (nctrl->opts->tos >= 0) {
1309 opt = nctrl->opts->tos;
1310 ret = kernel_setsockopt(queue->sock, SOL_IP, IP_TOS,
1311 (char *)&opt, sizeof(opt));
1313 dev_err(nctrl->device,
1314 "failed to set IP_TOS sock opt %d\n", ret);
1319 queue->sock->sk->sk_allocation = GFP_ATOMIC;
1323 n = (qid - 1) % num_online_cpus();
1324 queue->io_cpu = cpumask_next_wrap(n - 1, cpu_online_mask, -1, false);
1325 queue->request = NULL;
1326 queue->data_remaining = 0;
1327 queue->ddgst_remaining = 0;
1328 queue->pdu_remaining = 0;
1329 queue->pdu_offset = 0;
1330 sk_set_memalloc(queue->sock->sk);
1332 if (nctrl->opts->mask & NVMF_OPT_HOST_TRADDR) {
1333 ret = kernel_bind(queue->sock, (struct sockaddr *)&ctrl->src_addr,
1334 sizeof(ctrl->src_addr));
1336 dev_err(nctrl->device,
1337 "failed to bind queue %d socket %d\n",
1343 queue->hdr_digest = nctrl->opts->hdr_digest;
1344 queue->data_digest = nctrl->opts->data_digest;
1345 if (queue->hdr_digest || queue->data_digest) {
1346 ret = nvme_tcp_alloc_crypto(queue);
1348 dev_err(nctrl->device,
1349 "failed to allocate queue %d crypto\n", qid);
1354 rcv_pdu_size = sizeof(struct nvme_tcp_rsp_pdu) +
1355 nvme_tcp_hdgst_len(queue);
1356 queue->pdu = kmalloc(rcv_pdu_size, GFP_KERNEL);
1362 dev_dbg(nctrl->device, "connecting queue %d\n",
1363 nvme_tcp_queue_id(queue));
1365 ret = kernel_connect(queue->sock, (struct sockaddr *)&ctrl->addr,
1366 sizeof(ctrl->addr), 0);
1368 dev_err(nctrl->device,
1369 "failed to connect socket: %d\n", ret);
1373 ret = nvme_tcp_init_connection(queue);
1375 goto err_init_connect;
1377 queue->rd_enabled = true;
1378 set_bit(NVME_TCP_Q_ALLOCATED, &queue->flags);
1379 nvme_tcp_init_recv_ctx(queue);
1381 write_lock_bh(&queue->sock->sk->sk_callback_lock);
1382 queue->sock->sk->sk_user_data = queue;
1383 queue->state_change = queue->sock->sk->sk_state_change;
1384 queue->data_ready = queue->sock->sk->sk_data_ready;
1385 queue->write_space = queue->sock->sk->sk_write_space;
1386 queue->sock->sk->sk_data_ready = nvme_tcp_data_ready;
1387 queue->sock->sk->sk_state_change = nvme_tcp_state_change;
1388 queue->sock->sk->sk_write_space = nvme_tcp_write_space;
1389 queue->sock->sk->sk_ll_usec = 1;
1390 write_unlock_bh(&queue->sock->sk->sk_callback_lock);
1395 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1399 if (queue->hdr_digest || queue->data_digest)
1400 nvme_tcp_free_crypto(queue);
1402 sock_release(queue->sock);
1407 static void nvme_tcp_restore_sock_calls(struct nvme_tcp_queue *queue)
1409 struct socket *sock = queue->sock;
1411 write_lock_bh(&sock->sk->sk_callback_lock);
1412 sock->sk->sk_user_data = NULL;
1413 sock->sk->sk_data_ready = queue->data_ready;
1414 sock->sk->sk_state_change = queue->state_change;
1415 sock->sk->sk_write_space = queue->write_space;
1416 write_unlock_bh(&sock->sk->sk_callback_lock);
1419 static void __nvme_tcp_stop_queue(struct nvme_tcp_queue *queue)
1421 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1422 nvme_tcp_restore_sock_calls(queue);
1423 cancel_work_sync(&queue->io_work);
1426 static void nvme_tcp_stop_queue(struct nvme_ctrl *nctrl, int qid)
1428 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1429 struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1431 if (!test_and_clear_bit(NVME_TCP_Q_LIVE, &queue->flags))
1434 __nvme_tcp_stop_queue(queue);
1437 static int nvme_tcp_start_queue(struct nvme_ctrl *nctrl, int idx)
1439 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1443 ret = nvmf_connect_io_queue(nctrl, idx, false);
1445 ret = nvmf_connect_admin_queue(nctrl);
1448 set_bit(NVME_TCP_Q_LIVE, &ctrl->queues[idx].flags);
1450 if (test_bit(NVME_TCP_Q_ALLOCATED, &ctrl->queues[idx].flags))
1451 __nvme_tcp_stop_queue(&ctrl->queues[idx]);
1452 dev_err(nctrl->device,
1453 "failed to connect queue: %d ret=%d\n", idx, ret);
1458 static struct blk_mq_tag_set *nvme_tcp_alloc_tagset(struct nvme_ctrl *nctrl,
1461 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1462 struct blk_mq_tag_set *set;
1466 set = &ctrl->admin_tag_set;
1467 memset(set, 0, sizeof(*set));
1468 set->ops = &nvme_tcp_admin_mq_ops;
1469 set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
1470 set->reserved_tags = 2; /* connect + keep-alive */
1471 set->numa_node = NUMA_NO_NODE;
1472 set->cmd_size = sizeof(struct nvme_tcp_request);
1473 set->driver_data = ctrl;
1474 set->nr_hw_queues = 1;
1475 set->timeout = ADMIN_TIMEOUT;
1477 set = &ctrl->tag_set;
1478 memset(set, 0, sizeof(*set));
1479 set->ops = &nvme_tcp_mq_ops;
1480 set->queue_depth = nctrl->sqsize + 1;
1481 set->reserved_tags = 1; /* fabric connect */
1482 set->numa_node = NUMA_NO_NODE;
1483 set->flags = BLK_MQ_F_SHOULD_MERGE;
1484 set->cmd_size = sizeof(struct nvme_tcp_request);
1485 set->driver_data = ctrl;
1486 set->nr_hw_queues = nctrl->queue_count - 1;
1487 set->timeout = NVME_IO_TIMEOUT;
1488 set->nr_maps = nctrl->opts->nr_poll_queues ? HCTX_MAX_TYPES : 2;
1491 ret = blk_mq_alloc_tag_set(set);
1493 return ERR_PTR(ret);
1498 static void nvme_tcp_free_admin_queue(struct nvme_ctrl *ctrl)
1500 if (to_tcp_ctrl(ctrl)->async_req.pdu) {
1501 nvme_tcp_free_async_req(to_tcp_ctrl(ctrl));
1502 to_tcp_ctrl(ctrl)->async_req.pdu = NULL;
1505 nvme_tcp_free_queue(ctrl, 0);
1508 static void nvme_tcp_free_io_queues(struct nvme_ctrl *ctrl)
1512 for (i = 1; i < ctrl->queue_count; i++)
1513 nvme_tcp_free_queue(ctrl, i);
1516 static void nvme_tcp_stop_io_queues(struct nvme_ctrl *ctrl)
1520 for (i = 1; i < ctrl->queue_count; i++)
1521 nvme_tcp_stop_queue(ctrl, i);
1524 static int nvme_tcp_start_io_queues(struct nvme_ctrl *ctrl)
1528 for (i = 1; i < ctrl->queue_count; i++) {
1529 ret = nvme_tcp_start_queue(ctrl, i);
1531 goto out_stop_queues;
1537 for (i--; i >= 1; i--)
1538 nvme_tcp_stop_queue(ctrl, i);
1542 static int nvme_tcp_alloc_admin_queue(struct nvme_ctrl *ctrl)
1546 ret = nvme_tcp_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
1550 ret = nvme_tcp_alloc_async_req(to_tcp_ctrl(ctrl));
1552 goto out_free_queue;
1557 nvme_tcp_free_queue(ctrl, 0);
1561 static int __nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
1565 for (i = 1; i < ctrl->queue_count; i++) {
1566 ret = nvme_tcp_alloc_queue(ctrl, i,
1569 goto out_free_queues;
1575 for (i--; i >= 1; i--)
1576 nvme_tcp_free_queue(ctrl, i);
1581 static unsigned int nvme_tcp_nr_io_queues(struct nvme_ctrl *ctrl)
1583 unsigned int nr_io_queues;
1585 nr_io_queues = min(ctrl->opts->nr_io_queues, num_online_cpus());
1586 nr_io_queues += min(ctrl->opts->nr_write_queues, num_online_cpus());
1587 nr_io_queues += min(ctrl->opts->nr_poll_queues, num_online_cpus());
1589 return nr_io_queues;
1592 static void nvme_tcp_set_io_queues(struct nvme_ctrl *nctrl,
1593 unsigned int nr_io_queues)
1595 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1596 struct nvmf_ctrl_options *opts = nctrl->opts;
1598 if (opts->nr_write_queues && opts->nr_io_queues < nr_io_queues) {
1600 * separate read/write queues
1601 * hand out dedicated default queues only after we have
1602 * sufficient read queues.
1604 ctrl->io_queues[HCTX_TYPE_READ] = opts->nr_io_queues;
1605 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_READ];
1606 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
1607 min(opts->nr_write_queues, nr_io_queues);
1608 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
1611 * shared read/write queues
1612 * either no write queues were requested, or we don't have
1613 * sufficient queue count to have dedicated default queues.
1615 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
1616 min(opts->nr_io_queues, nr_io_queues);
1617 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
1620 if (opts->nr_poll_queues && nr_io_queues) {
1621 /* map dedicated poll queues only if we have queues left */
1622 ctrl->io_queues[HCTX_TYPE_POLL] =
1623 min(opts->nr_poll_queues, nr_io_queues);
1627 static int nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
1629 unsigned int nr_io_queues;
1632 nr_io_queues = nvme_tcp_nr_io_queues(ctrl);
1633 ret = nvme_set_queue_count(ctrl, &nr_io_queues);
1637 ctrl->queue_count = nr_io_queues + 1;
1638 if (ctrl->queue_count < 2)
1641 dev_info(ctrl->device,
1642 "creating %d I/O queues.\n", nr_io_queues);
1644 nvme_tcp_set_io_queues(ctrl, nr_io_queues);
1646 return __nvme_tcp_alloc_io_queues(ctrl);
1649 static void nvme_tcp_destroy_io_queues(struct nvme_ctrl *ctrl, bool remove)
1651 nvme_tcp_stop_io_queues(ctrl);
1653 blk_cleanup_queue(ctrl->connect_q);
1654 blk_mq_free_tag_set(ctrl->tagset);
1656 nvme_tcp_free_io_queues(ctrl);
1659 static int nvme_tcp_configure_io_queues(struct nvme_ctrl *ctrl, bool new)
1663 ret = nvme_tcp_alloc_io_queues(ctrl);
1668 ctrl->tagset = nvme_tcp_alloc_tagset(ctrl, false);
1669 if (IS_ERR(ctrl->tagset)) {
1670 ret = PTR_ERR(ctrl->tagset);
1671 goto out_free_io_queues;
1674 ctrl->connect_q = blk_mq_init_queue(ctrl->tagset);
1675 if (IS_ERR(ctrl->connect_q)) {
1676 ret = PTR_ERR(ctrl->connect_q);
1677 goto out_free_tag_set;
1680 blk_mq_update_nr_hw_queues(ctrl->tagset,
1681 ctrl->queue_count - 1);
1684 ret = nvme_tcp_start_io_queues(ctrl);
1686 goto out_cleanup_connect_q;
1690 out_cleanup_connect_q:
1692 blk_cleanup_queue(ctrl->connect_q);
1695 blk_mq_free_tag_set(ctrl->tagset);
1697 nvme_tcp_free_io_queues(ctrl);
1701 static void nvme_tcp_destroy_admin_queue(struct nvme_ctrl *ctrl, bool remove)
1703 nvme_tcp_stop_queue(ctrl, 0);
1705 blk_cleanup_queue(ctrl->admin_q);
1706 blk_mq_free_tag_set(ctrl->admin_tagset);
1708 nvme_tcp_free_admin_queue(ctrl);
1711 static int nvme_tcp_configure_admin_queue(struct nvme_ctrl *ctrl, bool new)
1715 error = nvme_tcp_alloc_admin_queue(ctrl);
1720 ctrl->admin_tagset = nvme_tcp_alloc_tagset(ctrl, true);
1721 if (IS_ERR(ctrl->admin_tagset)) {
1722 error = PTR_ERR(ctrl->admin_tagset);
1723 goto out_free_queue;
1726 ctrl->admin_q = blk_mq_init_queue(ctrl->admin_tagset);
1727 if (IS_ERR(ctrl->admin_q)) {
1728 error = PTR_ERR(ctrl->admin_q);
1729 goto out_free_tagset;
1733 error = nvme_tcp_start_queue(ctrl, 0);
1735 goto out_cleanup_queue;
1737 error = nvme_enable_ctrl(ctrl);
1739 goto out_stop_queue;
1741 error = nvme_init_identify(ctrl);
1743 goto out_stop_queue;
1748 nvme_tcp_stop_queue(ctrl, 0);
1751 blk_cleanup_queue(ctrl->admin_q);
1754 blk_mq_free_tag_set(ctrl->admin_tagset);
1756 nvme_tcp_free_admin_queue(ctrl);
1760 static void nvme_tcp_teardown_admin_queue(struct nvme_ctrl *ctrl,
1763 blk_mq_quiesce_queue(ctrl->admin_q);
1764 nvme_tcp_stop_queue(ctrl, 0);
1765 if (ctrl->admin_tagset) {
1766 blk_mq_tagset_busy_iter(ctrl->admin_tagset,
1767 nvme_cancel_request, ctrl);
1768 blk_mq_tagset_wait_completed_request(ctrl->admin_tagset);
1770 blk_mq_unquiesce_queue(ctrl->admin_q);
1771 nvme_tcp_destroy_admin_queue(ctrl, remove);
1774 static void nvme_tcp_teardown_io_queues(struct nvme_ctrl *ctrl,
1777 if (ctrl->queue_count <= 1)
1779 nvme_stop_queues(ctrl);
1780 nvme_tcp_stop_io_queues(ctrl);
1782 blk_mq_tagset_busy_iter(ctrl->tagset,
1783 nvme_cancel_request, ctrl);
1784 blk_mq_tagset_wait_completed_request(ctrl->tagset);
1787 nvme_start_queues(ctrl);
1788 nvme_tcp_destroy_io_queues(ctrl, remove);
1791 static void nvme_tcp_reconnect_or_remove(struct nvme_ctrl *ctrl)
1793 /* If we are resetting/deleting then do nothing */
1794 if (ctrl->state != NVME_CTRL_CONNECTING) {
1795 WARN_ON_ONCE(ctrl->state == NVME_CTRL_NEW ||
1796 ctrl->state == NVME_CTRL_LIVE);
1800 if (nvmf_should_reconnect(ctrl)) {
1801 dev_info(ctrl->device, "Reconnecting in %d seconds...\n",
1802 ctrl->opts->reconnect_delay);
1803 queue_delayed_work(nvme_wq, &to_tcp_ctrl(ctrl)->connect_work,
1804 ctrl->opts->reconnect_delay * HZ);
1806 dev_info(ctrl->device, "Removing controller...\n");
1807 nvme_delete_ctrl(ctrl);
1811 static int nvme_tcp_setup_ctrl(struct nvme_ctrl *ctrl, bool new)
1813 struct nvmf_ctrl_options *opts = ctrl->opts;
1816 ret = nvme_tcp_configure_admin_queue(ctrl, new);
1821 dev_err(ctrl->device, "icdoff is not supported!\n");
1825 if (opts->queue_size > ctrl->sqsize + 1)
1826 dev_warn(ctrl->device,
1827 "queue_size %zu > ctrl sqsize %u, clamping down\n",
1828 opts->queue_size, ctrl->sqsize + 1);
1830 if (ctrl->sqsize + 1 > ctrl->maxcmd) {
1831 dev_warn(ctrl->device,
1832 "sqsize %u > ctrl maxcmd %u, clamping down\n",
1833 ctrl->sqsize + 1, ctrl->maxcmd);
1834 ctrl->sqsize = ctrl->maxcmd - 1;
1837 if (ctrl->queue_count > 1) {
1838 ret = nvme_tcp_configure_io_queues(ctrl, new);
1843 if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_LIVE)) {
1844 /* state change failure is ok if we're in DELETING state */
1845 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING);
1850 nvme_start_ctrl(ctrl);
1854 if (ctrl->queue_count > 1)
1855 nvme_tcp_destroy_io_queues(ctrl, new);
1857 nvme_tcp_stop_queue(ctrl, 0);
1858 nvme_tcp_destroy_admin_queue(ctrl, new);
1862 static void nvme_tcp_reconnect_ctrl_work(struct work_struct *work)
1864 struct nvme_tcp_ctrl *tcp_ctrl = container_of(to_delayed_work(work),
1865 struct nvme_tcp_ctrl, connect_work);
1866 struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
1868 ++ctrl->nr_reconnects;
1870 if (nvme_tcp_setup_ctrl(ctrl, false))
1873 dev_info(ctrl->device, "Successfully reconnected (%d attempt)\n",
1874 ctrl->nr_reconnects);
1876 ctrl->nr_reconnects = 0;
1881 dev_info(ctrl->device, "Failed reconnect attempt %d\n",
1882 ctrl->nr_reconnects);
1883 nvme_tcp_reconnect_or_remove(ctrl);
1886 static void nvme_tcp_error_recovery_work(struct work_struct *work)
1888 struct nvme_tcp_ctrl *tcp_ctrl = container_of(work,
1889 struct nvme_tcp_ctrl, err_work);
1890 struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
1892 nvme_stop_keep_alive(ctrl);
1893 nvme_tcp_teardown_io_queues(ctrl, false);
1894 /* unquiesce to fail fast pending requests */
1895 nvme_start_queues(ctrl);
1896 nvme_tcp_teardown_admin_queue(ctrl, false);
1898 if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
1899 /* state change failure is ok if we're in DELETING state */
1900 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING);
1904 nvme_tcp_reconnect_or_remove(ctrl);
1907 static void nvme_tcp_teardown_ctrl(struct nvme_ctrl *ctrl, bool shutdown)
1909 cancel_work_sync(&to_tcp_ctrl(ctrl)->err_work);
1910 cancel_delayed_work_sync(&to_tcp_ctrl(ctrl)->connect_work);
1912 nvme_tcp_teardown_io_queues(ctrl, shutdown);
1914 nvme_shutdown_ctrl(ctrl);
1916 nvme_disable_ctrl(ctrl);
1917 nvme_tcp_teardown_admin_queue(ctrl, shutdown);
1920 static void nvme_tcp_delete_ctrl(struct nvme_ctrl *ctrl)
1922 nvme_tcp_teardown_ctrl(ctrl, true);
1925 static void nvme_reset_ctrl_work(struct work_struct *work)
1927 struct nvme_ctrl *ctrl =
1928 container_of(work, struct nvme_ctrl, reset_work);
1930 nvme_stop_ctrl(ctrl);
1931 nvme_tcp_teardown_ctrl(ctrl, false);
1933 if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
1934 /* state change failure is ok if we're in DELETING state */
1935 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING);
1939 if (nvme_tcp_setup_ctrl(ctrl, false))
1945 ++ctrl->nr_reconnects;
1946 nvme_tcp_reconnect_or_remove(ctrl);
1949 static void nvme_tcp_free_ctrl(struct nvme_ctrl *nctrl)
1951 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1953 if (list_empty(&ctrl->list))
1956 mutex_lock(&nvme_tcp_ctrl_mutex);
1957 list_del(&ctrl->list);
1958 mutex_unlock(&nvme_tcp_ctrl_mutex);
1960 nvmf_free_options(nctrl->opts);
1962 kfree(ctrl->queues);
1966 static void nvme_tcp_set_sg_null(struct nvme_command *c)
1968 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1972 sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
1973 NVME_SGL_FMT_TRANSPORT_A;
1976 static void nvme_tcp_set_sg_inline(struct nvme_tcp_queue *queue,
1977 struct nvme_command *c, u32 data_len)
1979 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1981 sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
1982 sg->length = cpu_to_le32(data_len);
1983 sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
1986 static void nvme_tcp_set_sg_host_data(struct nvme_command *c,
1989 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1992 sg->length = cpu_to_le32(data_len);
1993 sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
1994 NVME_SGL_FMT_TRANSPORT_A;
1997 static void nvme_tcp_submit_async_event(struct nvme_ctrl *arg)
1999 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(arg);
2000 struct nvme_tcp_queue *queue = &ctrl->queues[0];
2001 struct nvme_tcp_cmd_pdu *pdu = ctrl->async_req.pdu;
2002 struct nvme_command *cmd = &pdu->cmd;
2003 u8 hdgst = nvme_tcp_hdgst_len(queue);
2005 memset(pdu, 0, sizeof(*pdu));
2006 pdu->hdr.type = nvme_tcp_cmd;
2007 if (queue->hdr_digest)
2008 pdu->hdr.flags |= NVME_TCP_F_HDGST;
2009 pdu->hdr.hlen = sizeof(*pdu);
2010 pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst);
2012 cmd->common.opcode = nvme_admin_async_event;
2013 cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
2014 cmd->common.flags |= NVME_CMD_SGL_METABUF;
2015 nvme_tcp_set_sg_null(cmd);
2017 ctrl->async_req.state = NVME_TCP_SEND_CMD_PDU;
2018 ctrl->async_req.offset = 0;
2019 ctrl->async_req.curr_bio = NULL;
2020 ctrl->async_req.data_len = 0;
2022 nvme_tcp_queue_request(&ctrl->async_req);
2025 static enum blk_eh_timer_return
2026 nvme_tcp_timeout(struct request *rq, bool reserved)
2028 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2029 struct nvme_tcp_ctrl *ctrl = req->queue->ctrl;
2030 struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2032 dev_warn(ctrl->ctrl.device,
2033 "queue %d: timeout request %#x type %d\n",
2034 nvme_tcp_queue_id(req->queue), rq->tag, pdu->hdr.type);
2036 if (ctrl->ctrl.state != NVME_CTRL_LIVE) {
2038 * Teardown immediately if controller times out while starting
2039 * or we are already started error recovery. all outstanding
2040 * requests are completed on shutdown, so we return BLK_EH_DONE.
2042 flush_work(&ctrl->err_work);
2043 nvme_tcp_teardown_io_queues(&ctrl->ctrl, false);
2044 nvme_tcp_teardown_admin_queue(&ctrl->ctrl, false);
2048 dev_warn(ctrl->ctrl.device, "starting error recovery\n");
2049 nvme_tcp_error_recovery(&ctrl->ctrl);
2051 return BLK_EH_RESET_TIMER;
2054 static blk_status_t nvme_tcp_map_data(struct nvme_tcp_queue *queue,
2057 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2058 struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2059 struct nvme_command *c = &pdu->cmd;
2061 c->common.flags |= NVME_CMD_SGL_METABUF;
2063 if (rq_data_dir(rq) == WRITE && req->data_len &&
2064 req->data_len <= nvme_tcp_inline_data_size(queue))
2065 nvme_tcp_set_sg_inline(queue, c, req->data_len);
2067 nvme_tcp_set_sg_host_data(c, req->data_len);
2072 static blk_status_t nvme_tcp_setup_cmd_pdu(struct nvme_ns *ns,
2075 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2076 struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2077 struct nvme_tcp_queue *queue = req->queue;
2078 u8 hdgst = nvme_tcp_hdgst_len(queue), ddgst = 0;
2081 ret = nvme_setup_cmd(ns, rq, &pdu->cmd);
2085 req->state = NVME_TCP_SEND_CMD_PDU;
2090 req->data_len = blk_rq_payload_bytes(rq);
2091 req->curr_bio = rq->bio;
2093 if (rq_data_dir(rq) == WRITE &&
2094 req->data_len <= nvme_tcp_inline_data_size(queue))
2095 req->pdu_len = req->data_len;
2096 else if (req->curr_bio)
2097 nvme_tcp_init_iter(req, READ);
2099 pdu->hdr.type = nvme_tcp_cmd;
2101 if (queue->hdr_digest)
2102 pdu->hdr.flags |= NVME_TCP_F_HDGST;
2103 if (queue->data_digest && req->pdu_len) {
2104 pdu->hdr.flags |= NVME_TCP_F_DDGST;
2105 ddgst = nvme_tcp_ddgst_len(queue);
2107 pdu->hdr.hlen = sizeof(*pdu);
2108 pdu->hdr.pdo = req->pdu_len ? pdu->hdr.hlen + hdgst : 0;
2110 cpu_to_le32(pdu->hdr.hlen + hdgst + req->pdu_len + ddgst);
2112 ret = nvme_tcp_map_data(queue, rq);
2113 if (unlikely(ret)) {
2114 dev_err(queue->ctrl->ctrl.device,
2115 "Failed to map data (%d)\n", ret);
2122 static blk_status_t nvme_tcp_queue_rq(struct blk_mq_hw_ctx *hctx,
2123 const struct blk_mq_queue_data *bd)
2125 struct nvme_ns *ns = hctx->queue->queuedata;
2126 struct nvme_tcp_queue *queue = hctx->driver_data;
2127 struct request *rq = bd->rq;
2128 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2129 bool queue_ready = test_bit(NVME_TCP_Q_LIVE, &queue->flags);
2132 if (!nvmf_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
2133 return nvmf_fail_nonready_command(&queue->ctrl->ctrl, rq);
2135 ret = nvme_tcp_setup_cmd_pdu(ns, rq);
2139 blk_mq_start_request(rq);
2141 nvme_tcp_queue_request(req);
2146 static int nvme_tcp_map_queues(struct blk_mq_tag_set *set)
2148 struct nvme_tcp_ctrl *ctrl = set->driver_data;
2149 struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
2151 if (opts->nr_write_queues && ctrl->io_queues[HCTX_TYPE_READ]) {
2152 /* separate read/write queues */
2153 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2154 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2155 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2156 set->map[HCTX_TYPE_READ].nr_queues =
2157 ctrl->io_queues[HCTX_TYPE_READ];
2158 set->map[HCTX_TYPE_READ].queue_offset =
2159 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2161 /* shared read/write queues */
2162 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2163 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2164 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2165 set->map[HCTX_TYPE_READ].nr_queues =
2166 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2167 set->map[HCTX_TYPE_READ].queue_offset = 0;
2169 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
2170 blk_mq_map_queues(&set->map[HCTX_TYPE_READ]);
2172 if (opts->nr_poll_queues && ctrl->io_queues[HCTX_TYPE_POLL]) {
2173 /* map dedicated poll queues only if we have queues left */
2174 set->map[HCTX_TYPE_POLL].nr_queues =
2175 ctrl->io_queues[HCTX_TYPE_POLL];
2176 set->map[HCTX_TYPE_POLL].queue_offset =
2177 ctrl->io_queues[HCTX_TYPE_DEFAULT] +
2178 ctrl->io_queues[HCTX_TYPE_READ];
2179 blk_mq_map_queues(&set->map[HCTX_TYPE_POLL]);
2182 dev_info(ctrl->ctrl.device,
2183 "mapped %d/%d/%d default/read/poll queues.\n",
2184 ctrl->io_queues[HCTX_TYPE_DEFAULT],
2185 ctrl->io_queues[HCTX_TYPE_READ],
2186 ctrl->io_queues[HCTX_TYPE_POLL]);
2191 static int nvme_tcp_poll(struct blk_mq_hw_ctx *hctx)
2193 struct nvme_tcp_queue *queue = hctx->driver_data;
2194 struct sock *sk = queue->sock->sk;
2196 if (sk_can_busy_loop(sk) && skb_queue_empty(&sk->sk_receive_queue))
2197 sk_busy_loop(sk, true);
2198 nvme_tcp_try_recv(queue);
2199 return queue->nr_cqe;
2202 static struct blk_mq_ops nvme_tcp_mq_ops = {
2203 .queue_rq = nvme_tcp_queue_rq,
2204 .complete = nvme_complete_rq,
2205 .init_request = nvme_tcp_init_request,
2206 .exit_request = nvme_tcp_exit_request,
2207 .init_hctx = nvme_tcp_init_hctx,
2208 .timeout = nvme_tcp_timeout,
2209 .map_queues = nvme_tcp_map_queues,
2210 .poll = nvme_tcp_poll,
2213 static struct blk_mq_ops nvme_tcp_admin_mq_ops = {
2214 .queue_rq = nvme_tcp_queue_rq,
2215 .complete = nvme_complete_rq,
2216 .init_request = nvme_tcp_init_request,
2217 .exit_request = nvme_tcp_exit_request,
2218 .init_hctx = nvme_tcp_init_admin_hctx,
2219 .timeout = nvme_tcp_timeout,
2222 static const struct nvme_ctrl_ops nvme_tcp_ctrl_ops = {
2224 .module = THIS_MODULE,
2225 .flags = NVME_F_FABRICS,
2226 .reg_read32 = nvmf_reg_read32,
2227 .reg_read64 = nvmf_reg_read64,
2228 .reg_write32 = nvmf_reg_write32,
2229 .free_ctrl = nvme_tcp_free_ctrl,
2230 .submit_async_event = nvme_tcp_submit_async_event,
2231 .delete_ctrl = nvme_tcp_delete_ctrl,
2232 .get_address = nvmf_get_address,
2236 nvme_tcp_existing_controller(struct nvmf_ctrl_options *opts)
2238 struct nvme_tcp_ctrl *ctrl;
2241 mutex_lock(&nvme_tcp_ctrl_mutex);
2242 list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list) {
2243 found = nvmf_ip_options_match(&ctrl->ctrl, opts);
2247 mutex_unlock(&nvme_tcp_ctrl_mutex);
2252 static struct nvme_ctrl *nvme_tcp_create_ctrl(struct device *dev,
2253 struct nvmf_ctrl_options *opts)
2255 struct nvme_tcp_ctrl *ctrl;
2258 ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
2260 return ERR_PTR(-ENOMEM);
2262 INIT_LIST_HEAD(&ctrl->list);
2263 ctrl->ctrl.opts = opts;
2264 ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues +
2265 opts->nr_poll_queues + 1;
2266 ctrl->ctrl.sqsize = opts->queue_size - 1;
2267 ctrl->ctrl.kato = opts->kato;
2269 INIT_DELAYED_WORK(&ctrl->connect_work,
2270 nvme_tcp_reconnect_ctrl_work);
2271 INIT_WORK(&ctrl->err_work, nvme_tcp_error_recovery_work);
2272 INIT_WORK(&ctrl->ctrl.reset_work, nvme_reset_ctrl_work);
2274 if (!(opts->mask & NVMF_OPT_TRSVCID)) {
2276 kstrdup(__stringify(NVME_TCP_DISC_PORT), GFP_KERNEL);
2277 if (!opts->trsvcid) {
2281 opts->mask |= NVMF_OPT_TRSVCID;
2284 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2285 opts->traddr, opts->trsvcid, &ctrl->addr);
2287 pr_err("malformed address passed: %s:%s\n",
2288 opts->traddr, opts->trsvcid);
2292 if (opts->mask & NVMF_OPT_HOST_TRADDR) {
2293 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2294 opts->host_traddr, NULL, &ctrl->src_addr);
2296 pr_err("malformed src address passed: %s\n",
2302 if (!opts->duplicate_connect && nvme_tcp_existing_controller(opts)) {
2307 ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
2309 if (!ctrl->queues) {
2314 ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_tcp_ctrl_ops, 0);
2316 goto out_kfree_queues;
2318 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
2321 goto out_uninit_ctrl;
2324 ret = nvme_tcp_setup_ctrl(&ctrl->ctrl, true);
2326 goto out_uninit_ctrl;
2328 dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISp\n",
2329 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
2331 nvme_get_ctrl(&ctrl->ctrl);
2333 mutex_lock(&nvme_tcp_ctrl_mutex);
2334 list_add_tail(&ctrl->list, &nvme_tcp_ctrl_list);
2335 mutex_unlock(&nvme_tcp_ctrl_mutex);
2340 nvme_uninit_ctrl(&ctrl->ctrl);
2341 nvme_put_ctrl(&ctrl->ctrl);
2344 return ERR_PTR(ret);
2346 kfree(ctrl->queues);
2349 return ERR_PTR(ret);
2352 static struct nvmf_transport_ops nvme_tcp_transport = {
2354 .module = THIS_MODULE,
2355 .required_opts = NVMF_OPT_TRADDR,
2356 .allowed_opts = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2357 NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
2358 NVMF_OPT_HDR_DIGEST | NVMF_OPT_DATA_DIGEST |
2359 NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES |
2361 .create_ctrl = nvme_tcp_create_ctrl,
2364 static int __init nvme_tcp_init_module(void)
2366 nvme_tcp_wq = alloc_workqueue("nvme_tcp_wq",
2367 WQ_MEM_RECLAIM | WQ_HIGHPRI, 0);
2371 nvmf_register_transport(&nvme_tcp_transport);
2375 static void __exit nvme_tcp_cleanup_module(void)
2377 struct nvme_tcp_ctrl *ctrl;
2379 nvmf_unregister_transport(&nvme_tcp_transport);
2381 mutex_lock(&nvme_tcp_ctrl_mutex);
2382 list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list)
2383 nvme_delete_ctrl(&ctrl->ctrl);
2384 mutex_unlock(&nvme_tcp_ctrl_mutex);
2385 flush_workqueue(nvme_delete_wq);
2387 destroy_workqueue(nvme_tcp_wq);
2390 module_init(nvme_tcp_init_module);
2391 module_exit(nvme_tcp_cleanup_module);
2393 MODULE_LICENSE("GPL v2");