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>
20 struct nvme_tcp_queue;
22 enum nvme_tcp_send_state {
23 NVME_TCP_SEND_CMD_PDU = 0,
24 NVME_TCP_SEND_H2C_PDU,
29 struct nvme_tcp_request {
30 struct nvme_request req;
32 struct nvme_tcp_queue *queue;
37 struct list_head entry;
46 enum nvme_tcp_send_state state;
49 enum nvme_tcp_queue_flags {
50 NVME_TCP_Q_ALLOCATED = 0,
54 enum nvme_tcp_recv_state {
55 NVME_TCP_RECV_PDU = 0,
61 struct nvme_tcp_queue {
63 struct work_struct io_work;
67 struct list_head send_list;
73 size_t data_remaining;
74 size_t ddgst_remaining;
77 struct nvme_tcp_request *request;
80 size_t cmnd_capsule_len;
81 struct nvme_tcp_ctrl *ctrl;
87 struct ahash_request *rcv_hash;
88 struct ahash_request *snd_hash;
92 struct page_frag_cache pf_cache;
94 void (*state_change)(struct sock *);
95 void (*data_ready)(struct sock *);
96 void (*write_space)(struct sock *);
99 struct nvme_tcp_ctrl {
100 /* read only in the hot path */
101 struct nvme_tcp_queue *queues;
102 struct blk_mq_tag_set tag_set;
104 /* other member variables */
105 struct list_head list;
106 struct blk_mq_tag_set admin_tag_set;
107 struct sockaddr_storage addr;
108 struct sockaddr_storage src_addr;
109 struct nvme_ctrl ctrl;
111 struct work_struct err_work;
112 struct delayed_work connect_work;
113 struct nvme_tcp_request async_req;
116 static LIST_HEAD(nvme_tcp_ctrl_list);
117 static DEFINE_MUTEX(nvme_tcp_ctrl_mutex);
118 static struct workqueue_struct *nvme_tcp_wq;
119 static struct blk_mq_ops nvme_tcp_mq_ops;
120 static struct blk_mq_ops nvme_tcp_admin_mq_ops;
122 static inline struct nvme_tcp_ctrl *to_tcp_ctrl(struct nvme_ctrl *ctrl)
124 return container_of(ctrl, struct nvme_tcp_ctrl, ctrl);
127 static inline int nvme_tcp_queue_id(struct nvme_tcp_queue *queue)
129 return queue - queue->ctrl->queues;
132 static inline struct blk_mq_tags *nvme_tcp_tagset(struct nvme_tcp_queue *queue)
134 u32 queue_idx = nvme_tcp_queue_id(queue);
137 return queue->ctrl->admin_tag_set.tags[queue_idx];
138 return queue->ctrl->tag_set.tags[queue_idx - 1];
141 static inline u8 nvme_tcp_hdgst_len(struct nvme_tcp_queue *queue)
143 return queue->hdr_digest ? NVME_TCP_DIGEST_LENGTH : 0;
146 static inline u8 nvme_tcp_ddgst_len(struct nvme_tcp_queue *queue)
148 return queue->data_digest ? NVME_TCP_DIGEST_LENGTH : 0;
151 static inline size_t nvme_tcp_inline_data_size(struct nvme_tcp_queue *queue)
153 return queue->cmnd_capsule_len - sizeof(struct nvme_command);
156 static inline bool nvme_tcp_async_req(struct nvme_tcp_request *req)
158 return req == &req->queue->ctrl->async_req;
161 static inline bool nvme_tcp_has_inline_data(struct nvme_tcp_request *req)
166 if (unlikely(nvme_tcp_async_req(req)))
167 return false; /* async events don't have a request */
169 rq = blk_mq_rq_from_pdu(req);
170 bytes = blk_rq_payload_bytes(rq);
172 return rq_data_dir(rq) == WRITE && bytes &&
173 bytes <= nvme_tcp_inline_data_size(req->queue);
176 static inline struct page *nvme_tcp_req_cur_page(struct nvme_tcp_request *req)
178 return req->iter.bvec->bv_page;
181 static inline size_t nvme_tcp_req_cur_offset(struct nvme_tcp_request *req)
183 return req->iter.bvec->bv_offset + req->iter.iov_offset;
186 static inline size_t nvme_tcp_req_cur_length(struct nvme_tcp_request *req)
188 return min_t(size_t, req->iter.bvec->bv_len - req->iter.iov_offset,
189 req->pdu_len - req->pdu_sent);
192 static inline size_t nvme_tcp_req_offset(struct nvme_tcp_request *req)
194 return req->iter.iov_offset;
197 static inline size_t nvme_tcp_pdu_data_left(struct nvme_tcp_request *req)
199 return rq_data_dir(blk_mq_rq_from_pdu(req)) == WRITE ?
200 req->pdu_len - req->pdu_sent : 0;
203 static inline size_t nvme_tcp_pdu_last_send(struct nvme_tcp_request *req,
206 return nvme_tcp_pdu_data_left(req) <= len;
209 static void nvme_tcp_init_iter(struct nvme_tcp_request *req,
212 struct request *rq = blk_mq_rq_from_pdu(req);
218 if (rq->rq_flags & RQF_SPECIAL_PAYLOAD) {
219 vec = &rq->special_vec;
221 size = blk_rq_payload_bytes(rq);
224 struct bio *bio = req->curr_bio;
226 vec = __bvec_iter_bvec(bio->bi_io_vec, bio->bi_iter);
227 nsegs = bio_segments(bio);
228 size = bio->bi_iter.bi_size;
229 offset = bio->bi_iter.bi_bvec_done;
232 iov_iter_bvec(&req->iter, dir, vec, nsegs, size);
233 req->iter.iov_offset = offset;
236 static inline void nvme_tcp_advance_req(struct nvme_tcp_request *req,
239 req->data_sent += len;
240 req->pdu_sent += len;
241 iov_iter_advance(&req->iter, len);
242 if (!iov_iter_count(&req->iter) &&
243 req->data_sent < req->data_len) {
244 req->curr_bio = req->curr_bio->bi_next;
245 nvme_tcp_init_iter(req, WRITE);
249 static inline void nvme_tcp_queue_request(struct nvme_tcp_request *req)
251 struct nvme_tcp_queue *queue = req->queue;
253 spin_lock(&queue->lock);
254 list_add_tail(&req->entry, &queue->send_list);
255 spin_unlock(&queue->lock);
257 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
260 static inline struct nvme_tcp_request *
261 nvme_tcp_fetch_request(struct nvme_tcp_queue *queue)
263 struct nvme_tcp_request *req;
265 spin_lock(&queue->lock);
266 req = list_first_entry_or_null(&queue->send_list,
267 struct nvme_tcp_request, entry);
269 list_del(&req->entry);
270 spin_unlock(&queue->lock);
275 static inline void nvme_tcp_ddgst_final(struct ahash_request *hash,
278 ahash_request_set_crypt(hash, NULL, (u8 *)dgst, 0);
279 crypto_ahash_final(hash);
282 static inline void nvme_tcp_ddgst_update(struct ahash_request *hash,
283 struct page *page, off_t off, size_t len)
285 struct scatterlist sg;
287 sg_init_marker(&sg, 1);
288 sg_set_page(&sg, page, len, off);
289 ahash_request_set_crypt(hash, &sg, NULL, len);
290 crypto_ahash_update(hash);
293 static inline void nvme_tcp_hdgst(struct ahash_request *hash,
294 void *pdu, size_t len)
296 struct scatterlist sg;
298 sg_init_one(&sg, pdu, len);
299 ahash_request_set_crypt(hash, &sg, pdu + len, len);
300 crypto_ahash_digest(hash);
303 static int nvme_tcp_verify_hdgst(struct nvme_tcp_queue *queue,
304 void *pdu, size_t pdu_len)
306 struct nvme_tcp_hdr *hdr = pdu;
310 if (unlikely(!(hdr->flags & NVME_TCP_F_HDGST))) {
311 dev_err(queue->ctrl->ctrl.device,
312 "queue %d: header digest flag is cleared\n",
313 nvme_tcp_queue_id(queue));
317 recv_digest = *(__le32 *)(pdu + hdr->hlen);
318 nvme_tcp_hdgst(queue->rcv_hash, pdu, pdu_len);
319 exp_digest = *(__le32 *)(pdu + hdr->hlen);
320 if (recv_digest != exp_digest) {
321 dev_err(queue->ctrl->ctrl.device,
322 "header digest error: recv %#x expected %#x\n",
323 le32_to_cpu(recv_digest), le32_to_cpu(exp_digest));
330 static int nvme_tcp_check_ddgst(struct nvme_tcp_queue *queue, void *pdu)
332 struct nvme_tcp_hdr *hdr = pdu;
333 u8 digest_len = nvme_tcp_hdgst_len(queue);
336 len = le32_to_cpu(hdr->plen) - hdr->hlen -
337 ((hdr->flags & NVME_TCP_F_HDGST) ? digest_len : 0);
339 if (unlikely(len && !(hdr->flags & NVME_TCP_F_DDGST))) {
340 dev_err(queue->ctrl->ctrl.device,
341 "queue %d: data digest flag is cleared\n",
342 nvme_tcp_queue_id(queue));
345 crypto_ahash_init(queue->rcv_hash);
350 static void nvme_tcp_exit_request(struct blk_mq_tag_set *set,
351 struct request *rq, unsigned int hctx_idx)
353 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
355 page_frag_free(req->pdu);
358 static int nvme_tcp_init_request(struct blk_mq_tag_set *set,
359 struct request *rq, unsigned int hctx_idx,
360 unsigned int numa_node)
362 struct nvme_tcp_ctrl *ctrl = set->driver_data;
363 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
364 int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
365 struct nvme_tcp_queue *queue = &ctrl->queues[queue_idx];
366 u8 hdgst = nvme_tcp_hdgst_len(queue);
368 req->pdu = page_frag_alloc(&queue->pf_cache,
369 sizeof(struct nvme_tcp_cmd_pdu) + hdgst,
370 GFP_KERNEL | __GFP_ZERO);
375 nvme_req(rq)->ctrl = &ctrl->ctrl;
380 static int nvme_tcp_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
381 unsigned int hctx_idx)
383 struct nvme_tcp_ctrl *ctrl = data;
384 struct nvme_tcp_queue *queue = &ctrl->queues[hctx_idx + 1];
386 hctx->driver_data = queue;
390 static int nvme_tcp_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
391 unsigned int hctx_idx)
393 struct nvme_tcp_ctrl *ctrl = data;
394 struct nvme_tcp_queue *queue = &ctrl->queues[0];
396 hctx->driver_data = queue;
400 static enum nvme_tcp_recv_state
401 nvme_tcp_recv_state(struct nvme_tcp_queue *queue)
403 return (queue->pdu_remaining) ? NVME_TCP_RECV_PDU :
404 (queue->ddgst_remaining) ? NVME_TCP_RECV_DDGST :
408 static void nvme_tcp_init_recv_ctx(struct nvme_tcp_queue *queue)
410 queue->pdu_remaining = sizeof(struct nvme_tcp_rsp_pdu) +
411 nvme_tcp_hdgst_len(queue);
412 queue->pdu_offset = 0;
413 queue->data_remaining = -1;
414 queue->ddgst_remaining = 0;
417 static void nvme_tcp_error_recovery(struct nvme_ctrl *ctrl)
419 if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_RESETTING))
422 queue_work(nvme_wq, &to_tcp_ctrl(ctrl)->err_work);
425 static int nvme_tcp_process_nvme_cqe(struct nvme_tcp_queue *queue,
426 struct nvme_completion *cqe)
430 rq = blk_mq_tag_to_rq(nvme_tcp_tagset(queue), cqe->command_id);
432 dev_err(queue->ctrl->ctrl.device,
433 "queue %d tag 0x%x not found\n",
434 nvme_tcp_queue_id(queue), cqe->command_id);
435 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
439 nvme_end_request(rq, cqe->status, cqe->result);
444 static int nvme_tcp_handle_c2h_data(struct nvme_tcp_queue *queue,
445 struct nvme_tcp_data_pdu *pdu)
449 rq = blk_mq_tag_to_rq(nvme_tcp_tagset(queue), pdu->command_id);
451 dev_err(queue->ctrl->ctrl.device,
452 "queue %d tag %#x not found\n",
453 nvme_tcp_queue_id(queue), pdu->command_id);
457 if (!blk_rq_payload_bytes(rq)) {
458 dev_err(queue->ctrl->ctrl.device,
459 "queue %d tag %#x unexpected data\n",
460 nvme_tcp_queue_id(queue), rq->tag);
464 queue->data_remaining = le32_to_cpu(pdu->data_length);
470 static int nvme_tcp_handle_comp(struct nvme_tcp_queue *queue,
471 struct nvme_tcp_rsp_pdu *pdu)
473 struct nvme_completion *cqe = &pdu->cqe;
477 * AEN requests are special as they don't time out and can
478 * survive any kind of queue freeze and often don't respond to
479 * aborts. We don't even bother to allocate a struct request
480 * for them but rather special case them here.
482 if (unlikely(nvme_tcp_queue_id(queue) == 0 &&
483 cqe->command_id >= NVME_AQ_BLK_MQ_DEPTH))
484 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
487 ret = nvme_tcp_process_nvme_cqe(queue, cqe);
492 static int nvme_tcp_setup_h2c_data_pdu(struct nvme_tcp_request *req,
493 struct nvme_tcp_r2t_pdu *pdu)
495 struct nvme_tcp_data_pdu *data = req->pdu;
496 struct nvme_tcp_queue *queue = req->queue;
497 struct request *rq = blk_mq_rq_from_pdu(req);
498 u8 hdgst = nvme_tcp_hdgst_len(queue);
499 u8 ddgst = nvme_tcp_ddgst_len(queue);
501 req->pdu_len = le32_to_cpu(pdu->r2t_length);
504 if (unlikely(req->data_sent + req->pdu_len > req->data_len)) {
505 dev_err(queue->ctrl->ctrl.device,
506 "req %d r2t len %u exceeded data len %u (%zu sent)\n",
507 rq->tag, req->pdu_len, req->data_len,
512 if (unlikely(le32_to_cpu(pdu->r2t_offset) < req->data_sent)) {
513 dev_err(queue->ctrl->ctrl.device,
514 "req %d unexpected r2t offset %u (expected %zu)\n",
515 rq->tag, le32_to_cpu(pdu->r2t_offset),
520 memset(data, 0, sizeof(*data));
521 data->hdr.type = nvme_tcp_h2c_data;
522 data->hdr.flags = NVME_TCP_F_DATA_LAST;
523 if (queue->hdr_digest)
524 data->hdr.flags |= NVME_TCP_F_HDGST;
525 if (queue->data_digest)
526 data->hdr.flags |= NVME_TCP_F_DDGST;
527 data->hdr.hlen = sizeof(*data);
528 data->hdr.pdo = data->hdr.hlen + hdgst;
530 cpu_to_le32(data->hdr.hlen + hdgst + req->pdu_len + ddgst);
531 data->ttag = pdu->ttag;
532 data->command_id = rq->tag;
533 data->data_offset = cpu_to_le32(req->data_sent);
534 data->data_length = cpu_to_le32(req->pdu_len);
538 static int nvme_tcp_handle_r2t(struct nvme_tcp_queue *queue,
539 struct nvme_tcp_r2t_pdu *pdu)
541 struct nvme_tcp_request *req;
545 rq = blk_mq_tag_to_rq(nvme_tcp_tagset(queue), pdu->command_id);
547 dev_err(queue->ctrl->ctrl.device,
548 "queue %d tag %#x not found\n",
549 nvme_tcp_queue_id(queue), pdu->command_id);
552 req = blk_mq_rq_to_pdu(rq);
554 ret = nvme_tcp_setup_h2c_data_pdu(req, pdu);
558 req->state = NVME_TCP_SEND_H2C_PDU;
561 nvme_tcp_queue_request(req);
566 static int nvme_tcp_recv_pdu(struct nvme_tcp_queue *queue, struct sk_buff *skb,
567 unsigned int *offset, size_t *len)
569 struct nvme_tcp_hdr *hdr;
570 char *pdu = queue->pdu;
571 size_t rcv_len = min_t(size_t, *len, queue->pdu_remaining);
574 ret = skb_copy_bits(skb, *offset,
575 &pdu[queue->pdu_offset], rcv_len);
579 queue->pdu_remaining -= rcv_len;
580 queue->pdu_offset += rcv_len;
583 if (queue->pdu_remaining)
587 if (queue->hdr_digest) {
588 ret = nvme_tcp_verify_hdgst(queue, queue->pdu, hdr->hlen);
594 if (queue->data_digest) {
595 ret = nvme_tcp_check_ddgst(queue, queue->pdu);
601 case nvme_tcp_c2h_data:
602 ret = nvme_tcp_handle_c2h_data(queue, (void *)queue->pdu);
605 nvme_tcp_init_recv_ctx(queue);
606 ret = nvme_tcp_handle_comp(queue, (void *)queue->pdu);
609 nvme_tcp_init_recv_ctx(queue);
610 ret = nvme_tcp_handle_r2t(queue, (void *)queue->pdu);
613 dev_err(queue->ctrl->ctrl.device,
614 "unsupported pdu type (%d)\n", hdr->type);
621 static int nvme_tcp_recv_data(struct nvme_tcp_queue *queue, struct sk_buff *skb,
622 unsigned int *offset, size_t *len)
624 struct nvme_tcp_data_pdu *pdu = (void *)queue->pdu;
625 struct nvme_tcp_request *req;
628 rq = blk_mq_tag_to_rq(nvme_tcp_tagset(queue), pdu->command_id);
630 dev_err(queue->ctrl->ctrl.device,
631 "queue %d tag %#x not found\n",
632 nvme_tcp_queue_id(queue), pdu->command_id);
635 req = blk_mq_rq_to_pdu(rq);
640 recv_len = min_t(size_t, *len, queue->data_remaining);
644 if (!iov_iter_count(&req->iter)) {
645 req->curr_bio = req->curr_bio->bi_next;
648 * If we don`t have any bios it means that controller
649 * sent more data than we requested, hence error
651 if (!req->curr_bio) {
652 dev_err(queue->ctrl->ctrl.device,
653 "queue %d no space in request %#x",
654 nvme_tcp_queue_id(queue), rq->tag);
655 nvme_tcp_init_recv_ctx(queue);
658 nvme_tcp_init_iter(req, READ);
661 /* we can read only from what is left in this bio */
662 recv_len = min_t(size_t, recv_len,
663 iov_iter_count(&req->iter));
665 if (queue->data_digest)
666 ret = skb_copy_and_hash_datagram_iter(skb, *offset,
667 &req->iter, recv_len, queue->rcv_hash);
669 ret = skb_copy_datagram_iter(skb, *offset,
670 &req->iter, recv_len);
672 dev_err(queue->ctrl->ctrl.device,
673 "queue %d failed to copy request %#x data",
674 nvme_tcp_queue_id(queue), rq->tag);
680 queue->data_remaining -= recv_len;
683 if (!queue->data_remaining) {
684 if (queue->data_digest) {
685 nvme_tcp_ddgst_final(queue->rcv_hash, &queue->exp_ddgst);
686 queue->ddgst_remaining = NVME_TCP_DIGEST_LENGTH;
688 nvme_tcp_init_recv_ctx(queue);
695 static int nvme_tcp_recv_ddgst(struct nvme_tcp_queue *queue,
696 struct sk_buff *skb, unsigned int *offset, size_t *len)
698 char *ddgst = (char *)&queue->recv_ddgst;
699 size_t recv_len = min_t(size_t, *len, queue->ddgst_remaining);
700 off_t off = NVME_TCP_DIGEST_LENGTH - queue->ddgst_remaining;
703 ret = skb_copy_bits(skb, *offset, &ddgst[off], recv_len);
707 queue->ddgst_remaining -= recv_len;
710 if (queue->ddgst_remaining)
713 if (queue->recv_ddgst != queue->exp_ddgst) {
714 dev_err(queue->ctrl->ctrl.device,
715 "data digest error: recv %#x expected %#x\n",
716 le32_to_cpu(queue->recv_ddgst),
717 le32_to_cpu(queue->exp_ddgst));
721 nvme_tcp_init_recv_ctx(queue);
725 static int nvme_tcp_recv_skb(read_descriptor_t *desc, struct sk_buff *skb,
726 unsigned int offset, size_t len)
728 struct nvme_tcp_queue *queue = desc->arg.data;
729 size_t consumed = len;
733 switch (nvme_tcp_recv_state(queue)) {
734 case NVME_TCP_RECV_PDU:
735 result = nvme_tcp_recv_pdu(queue, skb, &offset, &len);
737 case NVME_TCP_RECV_DATA:
738 result = nvme_tcp_recv_data(queue, skb, &offset, &len);
740 case NVME_TCP_RECV_DDGST:
741 result = nvme_tcp_recv_ddgst(queue, skb, &offset, &len);
747 dev_err(queue->ctrl->ctrl.device,
748 "receive failed: %d\n", result);
749 queue->rd_enabled = false;
750 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
758 static void nvme_tcp_data_ready(struct sock *sk)
760 struct nvme_tcp_queue *queue;
762 read_lock(&sk->sk_callback_lock);
763 queue = sk->sk_user_data;
764 if (likely(queue && queue->rd_enabled))
765 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
766 read_unlock(&sk->sk_callback_lock);
769 static void nvme_tcp_write_space(struct sock *sk)
771 struct nvme_tcp_queue *queue;
773 read_lock_bh(&sk->sk_callback_lock);
774 queue = sk->sk_user_data;
775 if (likely(queue && sk_stream_is_writeable(sk))) {
776 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
777 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
779 read_unlock_bh(&sk->sk_callback_lock);
782 static void nvme_tcp_state_change(struct sock *sk)
784 struct nvme_tcp_queue *queue;
786 read_lock(&sk->sk_callback_lock);
787 queue = sk->sk_user_data;
791 switch (sk->sk_state) {
798 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
801 dev_info(queue->ctrl->ctrl.device,
802 "queue %d socket state %d\n",
803 nvme_tcp_queue_id(queue), sk->sk_state);
806 queue->state_change(sk);
808 read_unlock(&sk->sk_callback_lock);
811 static inline void nvme_tcp_done_send_req(struct nvme_tcp_queue *queue)
813 queue->request = NULL;
816 static void nvme_tcp_fail_request(struct nvme_tcp_request *req)
818 union nvme_result res = {};
820 nvme_end_request(blk_mq_rq_from_pdu(req),
821 cpu_to_le16(NVME_SC_DATA_XFER_ERROR), res);
824 static int nvme_tcp_try_send_data(struct nvme_tcp_request *req)
826 struct nvme_tcp_queue *queue = req->queue;
829 struct page *page = nvme_tcp_req_cur_page(req);
830 size_t offset = nvme_tcp_req_cur_offset(req);
831 size_t len = nvme_tcp_req_cur_length(req);
832 bool last = nvme_tcp_pdu_last_send(req, len);
833 int ret, flags = MSG_DONTWAIT;
835 if (last && !queue->data_digest)
840 ret = kernel_sendpage(queue->sock, page, offset, len, flags);
844 nvme_tcp_advance_req(req, ret);
845 if (queue->data_digest)
846 nvme_tcp_ddgst_update(queue->snd_hash, page,
849 /* fully successful last write*/
850 if (last && ret == len) {
851 if (queue->data_digest) {
852 nvme_tcp_ddgst_final(queue->snd_hash,
854 req->state = NVME_TCP_SEND_DDGST;
857 nvme_tcp_done_send_req(queue);
865 static int nvme_tcp_try_send_cmd_pdu(struct nvme_tcp_request *req)
867 struct nvme_tcp_queue *queue = req->queue;
868 struct nvme_tcp_cmd_pdu *pdu = req->pdu;
869 bool inline_data = nvme_tcp_has_inline_data(req);
870 int flags = MSG_DONTWAIT | (inline_data ? MSG_MORE : MSG_EOR);
871 u8 hdgst = nvme_tcp_hdgst_len(queue);
872 int len = sizeof(*pdu) + hdgst - req->offset;
875 if (queue->hdr_digest && !req->offset)
876 nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
878 ret = kernel_sendpage(queue->sock, virt_to_page(pdu),
879 offset_in_page(pdu) + req->offset, len, flags);
880 if (unlikely(ret <= 0))
886 req->state = NVME_TCP_SEND_DATA;
887 if (queue->data_digest)
888 crypto_ahash_init(queue->snd_hash);
889 nvme_tcp_init_iter(req, WRITE);
891 nvme_tcp_done_send_req(queue);
900 static int nvme_tcp_try_send_data_pdu(struct nvme_tcp_request *req)
902 struct nvme_tcp_queue *queue = req->queue;
903 struct nvme_tcp_data_pdu *pdu = req->pdu;
904 u8 hdgst = nvme_tcp_hdgst_len(queue);
905 int len = sizeof(*pdu) - req->offset + hdgst;
908 if (queue->hdr_digest && !req->offset)
909 nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
911 ret = kernel_sendpage(queue->sock, virt_to_page(pdu),
912 offset_in_page(pdu) + req->offset, len,
913 MSG_DONTWAIT | MSG_MORE);
914 if (unlikely(ret <= 0))
919 req->state = NVME_TCP_SEND_DATA;
920 if (queue->data_digest)
921 crypto_ahash_init(queue->snd_hash);
923 nvme_tcp_init_iter(req, WRITE);
931 static int nvme_tcp_try_send_ddgst(struct nvme_tcp_request *req)
933 struct nvme_tcp_queue *queue = req->queue;
935 struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_EOR };
937 .iov_base = &req->ddgst + req->offset,
938 .iov_len = NVME_TCP_DIGEST_LENGTH - req->offset
941 ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
942 if (unlikely(ret <= 0))
945 if (req->offset + ret == NVME_TCP_DIGEST_LENGTH) {
946 nvme_tcp_done_send_req(queue);
954 static int nvme_tcp_try_send(struct nvme_tcp_queue *queue)
956 struct nvme_tcp_request *req;
959 if (!queue->request) {
960 queue->request = nvme_tcp_fetch_request(queue);
964 req = queue->request;
966 if (req->state == NVME_TCP_SEND_CMD_PDU) {
967 ret = nvme_tcp_try_send_cmd_pdu(req);
970 if (!nvme_tcp_has_inline_data(req))
974 if (req->state == NVME_TCP_SEND_H2C_PDU) {
975 ret = nvme_tcp_try_send_data_pdu(req);
980 if (req->state == NVME_TCP_SEND_DATA) {
981 ret = nvme_tcp_try_send_data(req);
986 if (req->state == NVME_TCP_SEND_DDGST)
987 ret = nvme_tcp_try_send_ddgst(req);
994 static int nvme_tcp_try_recv(struct nvme_tcp_queue *queue)
996 struct sock *sk = queue->sock->sk;
997 read_descriptor_t rd_desc;
1000 rd_desc.arg.data = queue;
1003 consumed = tcp_read_sock(sk, &rd_desc, nvme_tcp_recv_skb);
1008 static void nvme_tcp_io_work(struct work_struct *w)
1010 struct nvme_tcp_queue *queue =
1011 container_of(w, struct nvme_tcp_queue, io_work);
1012 unsigned long start = jiffies + msecs_to_jiffies(1);
1015 bool pending = false;
1018 result = nvme_tcp_try_send(queue);
1021 } else if (unlikely(result < 0)) {
1022 dev_err(queue->ctrl->ctrl.device,
1023 "failed to send request %d\n", result);
1024 if (result != -EPIPE)
1025 nvme_tcp_fail_request(queue->request);
1026 nvme_tcp_done_send_req(queue);
1030 result = nvme_tcp_try_recv(queue);
1037 } while (time_after(jiffies, start)); /* quota is exhausted */
1039 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
1042 static void nvme_tcp_free_crypto(struct nvme_tcp_queue *queue)
1044 struct crypto_ahash *tfm = crypto_ahash_reqtfm(queue->rcv_hash);
1046 ahash_request_free(queue->rcv_hash);
1047 ahash_request_free(queue->snd_hash);
1048 crypto_free_ahash(tfm);
1051 static int nvme_tcp_alloc_crypto(struct nvme_tcp_queue *queue)
1053 struct crypto_ahash *tfm;
1055 tfm = crypto_alloc_ahash("crc32c", 0, CRYPTO_ALG_ASYNC);
1057 return PTR_ERR(tfm);
1059 queue->snd_hash = ahash_request_alloc(tfm, GFP_KERNEL);
1060 if (!queue->snd_hash)
1062 ahash_request_set_callback(queue->snd_hash, 0, NULL, NULL);
1064 queue->rcv_hash = ahash_request_alloc(tfm, GFP_KERNEL);
1065 if (!queue->rcv_hash)
1067 ahash_request_set_callback(queue->rcv_hash, 0, NULL, NULL);
1071 ahash_request_free(queue->snd_hash);
1073 crypto_free_ahash(tfm);
1077 static void nvme_tcp_free_async_req(struct nvme_tcp_ctrl *ctrl)
1079 struct nvme_tcp_request *async = &ctrl->async_req;
1081 page_frag_free(async->pdu);
1084 static int nvme_tcp_alloc_async_req(struct nvme_tcp_ctrl *ctrl)
1086 struct nvme_tcp_queue *queue = &ctrl->queues[0];
1087 struct nvme_tcp_request *async = &ctrl->async_req;
1088 u8 hdgst = nvme_tcp_hdgst_len(queue);
1090 async->pdu = page_frag_alloc(&queue->pf_cache,
1091 sizeof(struct nvme_tcp_cmd_pdu) + hdgst,
1092 GFP_KERNEL | __GFP_ZERO);
1096 async->queue = &ctrl->queues[0];
1100 static void nvme_tcp_free_queue(struct nvme_ctrl *nctrl, int qid)
1102 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1103 struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1105 if (!test_and_clear_bit(NVME_TCP_Q_ALLOCATED, &queue->flags))
1108 if (queue->hdr_digest || queue->data_digest)
1109 nvme_tcp_free_crypto(queue);
1111 sock_release(queue->sock);
1115 static int nvme_tcp_init_connection(struct nvme_tcp_queue *queue)
1117 struct nvme_tcp_icreq_pdu *icreq;
1118 struct nvme_tcp_icresp_pdu *icresp;
1119 struct msghdr msg = {};
1121 bool ctrl_hdgst, ctrl_ddgst;
1124 icreq = kzalloc(sizeof(*icreq), GFP_KERNEL);
1128 icresp = kzalloc(sizeof(*icresp), GFP_KERNEL);
1134 icreq->hdr.type = nvme_tcp_icreq;
1135 icreq->hdr.hlen = sizeof(*icreq);
1137 icreq->hdr.plen = cpu_to_le32(icreq->hdr.hlen);
1138 icreq->pfv = cpu_to_le16(NVME_TCP_PFV_1_0);
1139 icreq->maxr2t = 0; /* single inflight r2t supported */
1140 icreq->hpda = 0; /* no alignment constraint */
1141 if (queue->hdr_digest)
1142 icreq->digest |= NVME_TCP_HDR_DIGEST_ENABLE;
1143 if (queue->data_digest)
1144 icreq->digest |= NVME_TCP_DATA_DIGEST_ENABLE;
1146 iov.iov_base = icreq;
1147 iov.iov_len = sizeof(*icreq);
1148 ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
1152 memset(&msg, 0, sizeof(msg));
1153 iov.iov_base = icresp;
1154 iov.iov_len = sizeof(*icresp);
1155 ret = kernel_recvmsg(queue->sock, &msg, &iov, 1,
1156 iov.iov_len, msg.msg_flags);
1161 if (icresp->hdr.type != nvme_tcp_icresp) {
1162 pr_err("queue %d: bad type returned %d\n",
1163 nvme_tcp_queue_id(queue), icresp->hdr.type);
1167 if (le32_to_cpu(icresp->hdr.plen) != sizeof(*icresp)) {
1168 pr_err("queue %d: bad pdu length returned %d\n",
1169 nvme_tcp_queue_id(queue), icresp->hdr.plen);
1173 if (icresp->pfv != NVME_TCP_PFV_1_0) {
1174 pr_err("queue %d: bad pfv returned %d\n",
1175 nvme_tcp_queue_id(queue), icresp->pfv);
1179 ctrl_ddgst = !!(icresp->digest & NVME_TCP_DATA_DIGEST_ENABLE);
1180 if ((queue->data_digest && !ctrl_ddgst) ||
1181 (!queue->data_digest && ctrl_ddgst)) {
1182 pr_err("queue %d: data digest mismatch host: %s ctrl: %s\n",
1183 nvme_tcp_queue_id(queue),
1184 queue->data_digest ? "enabled" : "disabled",
1185 ctrl_ddgst ? "enabled" : "disabled");
1189 ctrl_hdgst = !!(icresp->digest & NVME_TCP_HDR_DIGEST_ENABLE);
1190 if ((queue->hdr_digest && !ctrl_hdgst) ||
1191 (!queue->hdr_digest && ctrl_hdgst)) {
1192 pr_err("queue %d: header digest mismatch host: %s ctrl: %s\n",
1193 nvme_tcp_queue_id(queue),
1194 queue->hdr_digest ? "enabled" : "disabled",
1195 ctrl_hdgst ? "enabled" : "disabled");
1199 if (icresp->cpda != 0) {
1200 pr_err("queue %d: unsupported cpda returned %d\n",
1201 nvme_tcp_queue_id(queue), icresp->cpda);
1213 static int nvme_tcp_alloc_queue(struct nvme_ctrl *nctrl,
1214 int qid, size_t queue_size)
1216 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1217 struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1218 struct linger sol = { .l_onoff = 1, .l_linger = 0 };
1219 int ret, opt, rcv_pdu_size, n;
1222 INIT_LIST_HEAD(&queue->send_list);
1223 spin_lock_init(&queue->lock);
1224 INIT_WORK(&queue->io_work, nvme_tcp_io_work);
1225 queue->queue_size = queue_size;
1228 queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
1230 queue->cmnd_capsule_len = sizeof(struct nvme_command) +
1231 NVME_TCP_ADMIN_CCSZ;
1233 ret = sock_create(ctrl->addr.ss_family, SOCK_STREAM,
1234 IPPROTO_TCP, &queue->sock);
1236 dev_err(ctrl->ctrl.device,
1237 "failed to create socket: %d\n", ret);
1241 /* Single syn retry */
1243 ret = kernel_setsockopt(queue->sock, IPPROTO_TCP, TCP_SYNCNT,
1244 (char *)&opt, sizeof(opt));
1246 dev_err(ctrl->ctrl.device,
1247 "failed to set TCP_SYNCNT sock opt %d\n", ret);
1251 /* Set TCP no delay */
1253 ret = kernel_setsockopt(queue->sock, IPPROTO_TCP,
1254 TCP_NODELAY, (char *)&opt, sizeof(opt));
1256 dev_err(ctrl->ctrl.device,
1257 "failed to set TCP_NODELAY sock opt %d\n", ret);
1262 * Cleanup whatever is sitting in the TCP transmit queue on socket
1263 * close. This is done to prevent stale data from being sent should
1264 * the network connection be restored before TCP times out.
1266 ret = kernel_setsockopt(queue->sock, SOL_SOCKET, SO_LINGER,
1267 (char *)&sol, sizeof(sol));
1269 dev_err(ctrl->ctrl.device,
1270 "failed to set SO_LINGER sock opt %d\n", ret);
1274 queue->sock->sk->sk_allocation = GFP_ATOMIC;
1278 n = (qid - 1) % num_online_cpus();
1279 queue->io_cpu = cpumask_next_wrap(n - 1, cpu_online_mask, -1, false);
1280 queue->request = NULL;
1281 queue->data_remaining = 0;
1282 queue->ddgst_remaining = 0;
1283 queue->pdu_remaining = 0;
1284 queue->pdu_offset = 0;
1285 sk_set_memalloc(queue->sock->sk);
1287 if (ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR) {
1288 ret = kernel_bind(queue->sock, (struct sockaddr *)&ctrl->src_addr,
1289 sizeof(ctrl->src_addr));
1291 dev_err(ctrl->ctrl.device,
1292 "failed to bind queue %d socket %d\n",
1298 queue->hdr_digest = nctrl->opts->hdr_digest;
1299 queue->data_digest = nctrl->opts->data_digest;
1300 if (queue->hdr_digest || queue->data_digest) {
1301 ret = nvme_tcp_alloc_crypto(queue);
1303 dev_err(ctrl->ctrl.device,
1304 "failed to allocate queue %d crypto\n", qid);
1309 rcv_pdu_size = sizeof(struct nvme_tcp_rsp_pdu) +
1310 nvme_tcp_hdgst_len(queue);
1311 queue->pdu = kmalloc(rcv_pdu_size, GFP_KERNEL);
1317 dev_dbg(ctrl->ctrl.device, "connecting queue %d\n",
1318 nvme_tcp_queue_id(queue));
1320 ret = kernel_connect(queue->sock, (struct sockaddr *)&ctrl->addr,
1321 sizeof(ctrl->addr), 0);
1323 dev_err(ctrl->ctrl.device,
1324 "failed to connect socket: %d\n", ret);
1328 ret = nvme_tcp_init_connection(queue);
1330 goto err_init_connect;
1332 queue->rd_enabled = true;
1333 set_bit(NVME_TCP_Q_ALLOCATED, &queue->flags);
1334 nvme_tcp_init_recv_ctx(queue);
1336 write_lock_bh(&queue->sock->sk->sk_callback_lock);
1337 queue->sock->sk->sk_user_data = queue;
1338 queue->state_change = queue->sock->sk->sk_state_change;
1339 queue->data_ready = queue->sock->sk->sk_data_ready;
1340 queue->write_space = queue->sock->sk->sk_write_space;
1341 queue->sock->sk->sk_data_ready = nvme_tcp_data_ready;
1342 queue->sock->sk->sk_state_change = nvme_tcp_state_change;
1343 queue->sock->sk->sk_write_space = nvme_tcp_write_space;
1344 write_unlock_bh(&queue->sock->sk->sk_callback_lock);
1349 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1353 if (queue->hdr_digest || queue->data_digest)
1354 nvme_tcp_free_crypto(queue);
1356 sock_release(queue->sock);
1361 static void nvme_tcp_restore_sock_calls(struct nvme_tcp_queue *queue)
1363 struct socket *sock = queue->sock;
1365 write_lock_bh(&sock->sk->sk_callback_lock);
1366 sock->sk->sk_user_data = NULL;
1367 sock->sk->sk_data_ready = queue->data_ready;
1368 sock->sk->sk_state_change = queue->state_change;
1369 sock->sk->sk_write_space = queue->write_space;
1370 write_unlock_bh(&sock->sk->sk_callback_lock);
1373 static void __nvme_tcp_stop_queue(struct nvme_tcp_queue *queue)
1375 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1376 nvme_tcp_restore_sock_calls(queue);
1377 cancel_work_sync(&queue->io_work);
1380 static void nvme_tcp_stop_queue(struct nvme_ctrl *nctrl, int qid)
1382 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1383 struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1385 if (!test_and_clear_bit(NVME_TCP_Q_LIVE, &queue->flags))
1388 __nvme_tcp_stop_queue(queue);
1391 static int nvme_tcp_start_queue(struct nvme_ctrl *nctrl, int idx)
1393 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1397 ret = nvmf_connect_io_queue(nctrl, idx, false);
1399 ret = nvmf_connect_admin_queue(nctrl);
1402 set_bit(NVME_TCP_Q_LIVE, &ctrl->queues[idx].flags);
1404 __nvme_tcp_stop_queue(&ctrl->queues[idx]);
1405 dev_err(nctrl->device,
1406 "failed to connect queue: %d ret=%d\n", idx, ret);
1411 static struct blk_mq_tag_set *nvme_tcp_alloc_tagset(struct nvme_ctrl *nctrl,
1414 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1415 struct blk_mq_tag_set *set;
1419 set = &ctrl->admin_tag_set;
1420 memset(set, 0, sizeof(*set));
1421 set->ops = &nvme_tcp_admin_mq_ops;
1422 set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
1423 set->reserved_tags = 2; /* connect + keep-alive */
1424 set->numa_node = NUMA_NO_NODE;
1425 set->cmd_size = sizeof(struct nvme_tcp_request);
1426 set->driver_data = ctrl;
1427 set->nr_hw_queues = 1;
1428 set->timeout = ADMIN_TIMEOUT;
1430 set = &ctrl->tag_set;
1431 memset(set, 0, sizeof(*set));
1432 set->ops = &nvme_tcp_mq_ops;
1433 set->queue_depth = nctrl->sqsize + 1;
1434 set->reserved_tags = 1; /* fabric connect */
1435 set->numa_node = NUMA_NO_NODE;
1436 set->flags = BLK_MQ_F_SHOULD_MERGE;
1437 set->cmd_size = sizeof(struct nvme_tcp_request);
1438 set->driver_data = ctrl;
1439 set->nr_hw_queues = nctrl->queue_count - 1;
1440 set->timeout = NVME_IO_TIMEOUT;
1441 set->nr_maps = 2 /* default + read */;
1444 ret = blk_mq_alloc_tag_set(set);
1446 return ERR_PTR(ret);
1451 static void nvme_tcp_free_admin_queue(struct nvme_ctrl *ctrl)
1453 if (to_tcp_ctrl(ctrl)->async_req.pdu) {
1454 nvme_tcp_free_async_req(to_tcp_ctrl(ctrl));
1455 to_tcp_ctrl(ctrl)->async_req.pdu = NULL;
1458 nvme_tcp_free_queue(ctrl, 0);
1461 static void nvme_tcp_free_io_queues(struct nvme_ctrl *ctrl)
1465 for (i = 1; i < ctrl->queue_count; i++)
1466 nvme_tcp_free_queue(ctrl, i);
1469 static void nvme_tcp_stop_io_queues(struct nvme_ctrl *ctrl)
1473 for (i = 1; i < ctrl->queue_count; i++)
1474 nvme_tcp_stop_queue(ctrl, i);
1477 static int nvme_tcp_start_io_queues(struct nvme_ctrl *ctrl)
1481 for (i = 1; i < ctrl->queue_count; i++) {
1482 ret = nvme_tcp_start_queue(ctrl, i);
1484 goto out_stop_queues;
1490 for (i--; i >= 1; i--)
1491 nvme_tcp_stop_queue(ctrl, i);
1495 static int nvme_tcp_alloc_admin_queue(struct nvme_ctrl *ctrl)
1499 ret = nvme_tcp_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
1503 ret = nvme_tcp_alloc_async_req(to_tcp_ctrl(ctrl));
1505 goto out_free_queue;
1510 nvme_tcp_free_queue(ctrl, 0);
1514 static int nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
1518 for (i = 1; i < ctrl->queue_count; i++) {
1519 ret = nvme_tcp_alloc_queue(ctrl, i,
1522 goto out_free_queues;
1528 for (i--; i >= 1; i--)
1529 nvme_tcp_free_queue(ctrl, i);
1534 static unsigned int nvme_tcp_nr_io_queues(struct nvme_ctrl *ctrl)
1536 unsigned int nr_io_queues;
1538 nr_io_queues = min(ctrl->opts->nr_io_queues, num_online_cpus());
1539 nr_io_queues += min(ctrl->opts->nr_write_queues, num_online_cpus());
1541 return nr_io_queues;
1544 static int nvme_alloc_io_queues(struct nvme_ctrl *ctrl)
1546 unsigned int nr_io_queues;
1549 nr_io_queues = nvme_tcp_nr_io_queues(ctrl);
1550 ret = nvme_set_queue_count(ctrl, &nr_io_queues);
1554 ctrl->queue_count = nr_io_queues + 1;
1555 if (ctrl->queue_count < 2)
1558 dev_info(ctrl->device,
1559 "creating %d I/O queues.\n", nr_io_queues);
1561 return nvme_tcp_alloc_io_queues(ctrl);
1564 static void nvme_tcp_destroy_io_queues(struct nvme_ctrl *ctrl, bool remove)
1566 nvme_tcp_stop_io_queues(ctrl);
1568 if (ctrl->ops->flags & NVME_F_FABRICS)
1569 blk_cleanup_queue(ctrl->connect_q);
1570 blk_mq_free_tag_set(ctrl->tagset);
1572 nvme_tcp_free_io_queues(ctrl);
1575 static int nvme_tcp_configure_io_queues(struct nvme_ctrl *ctrl, bool new)
1579 ret = nvme_alloc_io_queues(ctrl);
1584 ctrl->tagset = nvme_tcp_alloc_tagset(ctrl, false);
1585 if (IS_ERR(ctrl->tagset)) {
1586 ret = PTR_ERR(ctrl->tagset);
1587 goto out_free_io_queues;
1590 if (ctrl->ops->flags & NVME_F_FABRICS) {
1591 ctrl->connect_q = blk_mq_init_queue(ctrl->tagset);
1592 if (IS_ERR(ctrl->connect_q)) {
1593 ret = PTR_ERR(ctrl->connect_q);
1594 goto out_free_tag_set;
1598 blk_mq_update_nr_hw_queues(ctrl->tagset,
1599 ctrl->queue_count - 1);
1602 ret = nvme_tcp_start_io_queues(ctrl);
1604 goto out_cleanup_connect_q;
1608 out_cleanup_connect_q:
1609 if (new && (ctrl->ops->flags & NVME_F_FABRICS))
1610 blk_cleanup_queue(ctrl->connect_q);
1613 blk_mq_free_tag_set(ctrl->tagset);
1615 nvme_tcp_free_io_queues(ctrl);
1619 static void nvme_tcp_destroy_admin_queue(struct nvme_ctrl *ctrl, bool remove)
1621 nvme_tcp_stop_queue(ctrl, 0);
1623 free_opal_dev(ctrl->opal_dev);
1624 blk_cleanup_queue(ctrl->admin_q);
1625 blk_mq_free_tag_set(ctrl->admin_tagset);
1627 nvme_tcp_free_admin_queue(ctrl);
1630 static int nvme_tcp_configure_admin_queue(struct nvme_ctrl *ctrl, bool new)
1634 error = nvme_tcp_alloc_admin_queue(ctrl);
1639 ctrl->admin_tagset = nvme_tcp_alloc_tagset(ctrl, true);
1640 if (IS_ERR(ctrl->admin_tagset)) {
1641 error = PTR_ERR(ctrl->admin_tagset);
1642 goto out_free_queue;
1645 ctrl->admin_q = blk_mq_init_queue(ctrl->admin_tagset);
1646 if (IS_ERR(ctrl->admin_q)) {
1647 error = PTR_ERR(ctrl->admin_q);
1648 goto out_free_tagset;
1652 error = nvme_tcp_start_queue(ctrl, 0);
1654 goto out_cleanup_queue;
1656 error = ctrl->ops->reg_read64(ctrl, NVME_REG_CAP, &ctrl->cap);
1658 dev_err(ctrl->device,
1659 "prop_get NVME_REG_CAP failed\n");
1660 goto out_stop_queue;
1663 ctrl->sqsize = min_t(int, NVME_CAP_MQES(ctrl->cap), ctrl->sqsize);
1665 error = nvme_enable_ctrl(ctrl, ctrl->cap);
1667 goto out_stop_queue;
1669 error = nvme_init_identify(ctrl);
1671 goto out_stop_queue;
1676 nvme_tcp_stop_queue(ctrl, 0);
1679 blk_cleanup_queue(ctrl->admin_q);
1682 blk_mq_free_tag_set(ctrl->admin_tagset);
1684 nvme_tcp_free_admin_queue(ctrl);
1688 static void nvme_tcp_teardown_admin_queue(struct nvme_ctrl *ctrl,
1691 blk_mq_quiesce_queue(ctrl->admin_q);
1692 nvme_tcp_stop_queue(ctrl, 0);
1693 blk_mq_tagset_busy_iter(ctrl->admin_tagset, nvme_cancel_request, ctrl);
1694 blk_mq_unquiesce_queue(ctrl->admin_q);
1695 nvme_tcp_destroy_admin_queue(ctrl, remove);
1698 static void nvme_tcp_teardown_io_queues(struct nvme_ctrl *ctrl,
1701 if (ctrl->queue_count <= 1)
1703 nvme_stop_queues(ctrl);
1704 nvme_tcp_stop_io_queues(ctrl);
1705 blk_mq_tagset_busy_iter(ctrl->tagset, nvme_cancel_request, ctrl);
1707 nvme_start_queues(ctrl);
1708 nvme_tcp_destroy_io_queues(ctrl, remove);
1711 static void nvme_tcp_reconnect_or_remove(struct nvme_ctrl *ctrl)
1713 /* If we are resetting/deleting then do nothing */
1714 if (ctrl->state != NVME_CTRL_CONNECTING) {
1715 WARN_ON_ONCE(ctrl->state == NVME_CTRL_NEW ||
1716 ctrl->state == NVME_CTRL_LIVE);
1720 if (nvmf_should_reconnect(ctrl)) {
1721 dev_info(ctrl->device, "Reconnecting in %d seconds...\n",
1722 ctrl->opts->reconnect_delay);
1723 queue_delayed_work(nvme_wq, &to_tcp_ctrl(ctrl)->connect_work,
1724 ctrl->opts->reconnect_delay * HZ);
1726 dev_info(ctrl->device, "Removing controller...\n");
1727 nvme_delete_ctrl(ctrl);
1731 static int nvme_tcp_setup_ctrl(struct nvme_ctrl *ctrl, bool new)
1733 struct nvmf_ctrl_options *opts = ctrl->opts;
1736 ret = nvme_tcp_configure_admin_queue(ctrl, new);
1741 dev_err(ctrl->device, "icdoff is not supported!\n");
1745 if (opts->queue_size > ctrl->sqsize + 1)
1746 dev_warn(ctrl->device,
1747 "queue_size %zu > ctrl sqsize %u, clamping down\n",
1748 opts->queue_size, ctrl->sqsize + 1);
1750 if (ctrl->sqsize + 1 > ctrl->maxcmd) {
1751 dev_warn(ctrl->device,
1752 "sqsize %u > ctrl maxcmd %u, clamping down\n",
1753 ctrl->sqsize + 1, ctrl->maxcmd);
1754 ctrl->sqsize = ctrl->maxcmd - 1;
1757 if (ctrl->queue_count > 1) {
1758 ret = nvme_tcp_configure_io_queues(ctrl, new);
1763 if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_LIVE)) {
1764 /* state change failure is ok if we're in DELETING state */
1765 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING);
1770 nvme_start_ctrl(ctrl);
1774 if (ctrl->queue_count > 1)
1775 nvme_tcp_destroy_io_queues(ctrl, new);
1777 nvme_tcp_stop_queue(ctrl, 0);
1778 nvme_tcp_destroy_admin_queue(ctrl, new);
1782 static void nvme_tcp_reconnect_ctrl_work(struct work_struct *work)
1784 struct nvme_tcp_ctrl *tcp_ctrl = container_of(to_delayed_work(work),
1785 struct nvme_tcp_ctrl, connect_work);
1786 struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
1788 ++ctrl->nr_reconnects;
1790 if (nvme_tcp_setup_ctrl(ctrl, false))
1793 dev_info(ctrl->device, "Successfully reconnected (%d attempt)\n",
1794 ctrl->nr_reconnects);
1796 ctrl->nr_reconnects = 0;
1801 dev_info(ctrl->device, "Failed reconnect attempt %d\n",
1802 ctrl->nr_reconnects);
1803 nvme_tcp_reconnect_or_remove(ctrl);
1806 static void nvme_tcp_error_recovery_work(struct work_struct *work)
1808 struct nvme_tcp_ctrl *tcp_ctrl = container_of(work,
1809 struct nvme_tcp_ctrl, err_work);
1810 struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
1812 nvme_stop_keep_alive(ctrl);
1813 nvme_tcp_teardown_io_queues(ctrl, false);
1814 /* unquiesce to fail fast pending requests */
1815 nvme_start_queues(ctrl);
1816 nvme_tcp_teardown_admin_queue(ctrl, false);
1818 if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
1819 /* state change failure is ok if we're in DELETING state */
1820 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING);
1824 nvme_tcp_reconnect_or_remove(ctrl);
1827 static void nvme_tcp_teardown_ctrl(struct nvme_ctrl *ctrl, bool shutdown)
1829 nvme_tcp_teardown_io_queues(ctrl, shutdown);
1831 nvme_shutdown_ctrl(ctrl);
1833 nvme_disable_ctrl(ctrl, ctrl->cap);
1834 nvme_tcp_teardown_admin_queue(ctrl, shutdown);
1837 static void nvme_tcp_delete_ctrl(struct nvme_ctrl *ctrl)
1839 nvme_tcp_teardown_ctrl(ctrl, true);
1842 static void nvme_reset_ctrl_work(struct work_struct *work)
1844 struct nvme_ctrl *ctrl =
1845 container_of(work, struct nvme_ctrl, reset_work);
1847 nvme_stop_ctrl(ctrl);
1848 nvme_tcp_teardown_ctrl(ctrl, false);
1850 if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
1851 /* state change failure is ok if we're in DELETING state */
1852 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING);
1856 if (nvme_tcp_setup_ctrl(ctrl, false))
1862 ++ctrl->nr_reconnects;
1863 nvme_tcp_reconnect_or_remove(ctrl);
1866 static void nvme_tcp_stop_ctrl(struct nvme_ctrl *ctrl)
1868 cancel_work_sync(&to_tcp_ctrl(ctrl)->err_work);
1869 cancel_delayed_work_sync(&to_tcp_ctrl(ctrl)->connect_work);
1872 static void nvme_tcp_free_ctrl(struct nvme_ctrl *nctrl)
1874 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1876 if (list_empty(&ctrl->list))
1879 mutex_lock(&nvme_tcp_ctrl_mutex);
1880 list_del(&ctrl->list);
1881 mutex_unlock(&nvme_tcp_ctrl_mutex);
1883 nvmf_free_options(nctrl->opts);
1885 kfree(ctrl->queues);
1889 static void nvme_tcp_set_sg_null(struct nvme_command *c)
1891 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1895 sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
1896 NVME_SGL_FMT_TRANSPORT_A;
1899 static void nvme_tcp_set_sg_inline(struct nvme_tcp_queue *queue,
1900 struct nvme_command *c, u32 data_len)
1902 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1904 sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
1905 sg->length = cpu_to_le32(data_len);
1906 sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
1909 static void nvme_tcp_set_sg_host_data(struct nvme_command *c,
1912 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1915 sg->length = cpu_to_le32(data_len);
1916 sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
1917 NVME_SGL_FMT_TRANSPORT_A;
1920 static void nvme_tcp_submit_async_event(struct nvme_ctrl *arg)
1922 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(arg);
1923 struct nvme_tcp_queue *queue = &ctrl->queues[0];
1924 struct nvme_tcp_cmd_pdu *pdu = ctrl->async_req.pdu;
1925 struct nvme_command *cmd = &pdu->cmd;
1926 u8 hdgst = nvme_tcp_hdgst_len(queue);
1928 memset(pdu, 0, sizeof(*pdu));
1929 pdu->hdr.type = nvme_tcp_cmd;
1930 if (queue->hdr_digest)
1931 pdu->hdr.flags |= NVME_TCP_F_HDGST;
1932 pdu->hdr.hlen = sizeof(*pdu);
1933 pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst);
1935 cmd->common.opcode = nvme_admin_async_event;
1936 cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1937 cmd->common.flags |= NVME_CMD_SGL_METABUF;
1938 nvme_tcp_set_sg_null(cmd);
1940 ctrl->async_req.state = NVME_TCP_SEND_CMD_PDU;
1941 ctrl->async_req.offset = 0;
1942 ctrl->async_req.curr_bio = NULL;
1943 ctrl->async_req.data_len = 0;
1945 nvme_tcp_queue_request(&ctrl->async_req);
1948 static enum blk_eh_timer_return
1949 nvme_tcp_timeout(struct request *rq, bool reserved)
1951 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
1952 struct nvme_tcp_ctrl *ctrl = req->queue->ctrl;
1953 struct nvme_tcp_cmd_pdu *pdu = req->pdu;
1955 dev_dbg(ctrl->ctrl.device,
1956 "queue %d: timeout request %#x type %d\n",
1957 nvme_tcp_queue_id(req->queue), rq->tag,
1960 if (ctrl->ctrl.state != NVME_CTRL_LIVE) {
1961 union nvme_result res = {};
1963 nvme_req(rq)->flags |= NVME_REQ_CANCELLED;
1964 nvme_end_request(rq, cpu_to_le16(NVME_SC_ABORT_REQ), res);
1968 /* queue error recovery */
1969 nvme_tcp_error_recovery(&ctrl->ctrl);
1971 return BLK_EH_RESET_TIMER;
1974 static blk_status_t nvme_tcp_map_data(struct nvme_tcp_queue *queue,
1977 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
1978 struct nvme_tcp_cmd_pdu *pdu = req->pdu;
1979 struct nvme_command *c = &pdu->cmd;
1981 c->common.flags |= NVME_CMD_SGL_METABUF;
1983 if (rq_data_dir(rq) == WRITE && req->data_len &&
1984 req->data_len <= nvme_tcp_inline_data_size(queue))
1985 nvme_tcp_set_sg_inline(queue, c, req->data_len);
1987 nvme_tcp_set_sg_host_data(c, req->data_len);
1992 static blk_status_t nvme_tcp_setup_cmd_pdu(struct nvme_ns *ns,
1995 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
1996 struct nvme_tcp_cmd_pdu *pdu = req->pdu;
1997 struct nvme_tcp_queue *queue = req->queue;
1998 u8 hdgst = nvme_tcp_hdgst_len(queue), ddgst = 0;
2001 ret = nvme_setup_cmd(ns, rq, &pdu->cmd);
2005 req->state = NVME_TCP_SEND_CMD_PDU;
2010 req->data_len = blk_rq_payload_bytes(rq);
2011 req->curr_bio = rq->bio;
2013 if (rq_data_dir(rq) == WRITE &&
2014 req->data_len <= nvme_tcp_inline_data_size(queue))
2015 req->pdu_len = req->data_len;
2016 else if (req->curr_bio)
2017 nvme_tcp_init_iter(req, READ);
2019 pdu->hdr.type = nvme_tcp_cmd;
2021 if (queue->hdr_digest)
2022 pdu->hdr.flags |= NVME_TCP_F_HDGST;
2023 if (queue->data_digest && req->pdu_len) {
2024 pdu->hdr.flags |= NVME_TCP_F_DDGST;
2025 ddgst = nvme_tcp_ddgst_len(queue);
2027 pdu->hdr.hlen = sizeof(*pdu);
2028 pdu->hdr.pdo = req->pdu_len ? pdu->hdr.hlen + hdgst : 0;
2030 cpu_to_le32(pdu->hdr.hlen + hdgst + req->pdu_len + ddgst);
2032 ret = nvme_tcp_map_data(queue, rq);
2033 if (unlikely(ret)) {
2034 dev_err(queue->ctrl->ctrl.device,
2035 "Failed to map data (%d)\n", ret);
2042 static blk_status_t nvme_tcp_queue_rq(struct blk_mq_hw_ctx *hctx,
2043 const struct blk_mq_queue_data *bd)
2045 struct nvme_ns *ns = hctx->queue->queuedata;
2046 struct nvme_tcp_queue *queue = hctx->driver_data;
2047 struct request *rq = bd->rq;
2048 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2049 bool queue_ready = test_bit(NVME_TCP_Q_LIVE, &queue->flags);
2052 if (!nvmf_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
2053 return nvmf_fail_nonready_command(&queue->ctrl->ctrl, rq);
2055 ret = nvme_tcp_setup_cmd_pdu(ns, rq);
2059 blk_mq_start_request(rq);
2061 nvme_tcp_queue_request(req);
2066 static int nvme_tcp_map_queues(struct blk_mq_tag_set *set)
2068 struct nvme_tcp_ctrl *ctrl = set->driver_data;
2070 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2071 set->map[HCTX_TYPE_READ].nr_queues = ctrl->ctrl.opts->nr_io_queues;
2072 if (ctrl->ctrl.opts->nr_write_queues) {
2073 /* separate read/write queues */
2074 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2075 ctrl->ctrl.opts->nr_write_queues;
2076 set->map[HCTX_TYPE_READ].queue_offset =
2077 ctrl->ctrl.opts->nr_write_queues;
2079 /* mixed read/write queues */
2080 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2081 ctrl->ctrl.opts->nr_io_queues;
2082 set->map[HCTX_TYPE_READ].queue_offset = 0;
2084 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
2085 blk_mq_map_queues(&set->map[HCTX_TYPE_READ]);
2089 static struct blk_mq_ops nvme_tcp_mq_ops = {
2090 .queue_rq = nvme_tcp_queue_rq,
2091 .complete = nvme_complete_rq,
2092 .init_request = nvme_tcp_init_request,
2093 .exit_request = nvme_tcp_exit_request,
2094 .init_hctx = nvme_tcp_init_hctx,
2095 .timeout = nvme_tcp_timeout,
2096 .map_queues = nvme_tcp_map_queues,
2099 static struct blk_mq_ops nvme_tcp_admin_mq_ops = {
2100 .queue_rq = nvme_tcp_queue_rq,
2101 .complete = nvme_complete_rq,
2102 .init_request = nvme_tcp_init_request,
2103 .exit_request = nvme_tcp_exit_request,
2104 .init_hctx = nvme_tcp_init_admin_hctx,
2105 .timeout = nvme_tcp_timeout,
2108 static const struct nvme_ctrl_ops nvme_tcp_ctrl_ops = {
2110 .module = THIS_MODULE,
2111 .flags = NVME_F_FABRICS,
2112 .reg_read32 = nvmf_reg_read32,
2113 .reg_read64 = nvmf_reg_read64,
2114 .reg_write32 = nvmf_reg_write32,
2115 .free_ctrl = nvme_tcp_free_ctrl,
2116 .submit_async_event = nvme_tcp_submit_async_event,
2117 .delete_ctrl = nvme_tcp_delete_ctrl,
2118 .get_address = nvmf_get_address,
2119 .stop_ctrl = nvme_tcp_stop_ctrl,
2123 nvme_tcp_existing_controller(struct nvmf_ctrl_options *opts)
2125 struct nvme_tcp_ctrl *ctrl;
2128 mutex_lock(&nvme_tcp_ctrl_mutex);
2129 list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list) {
2130 found = nvmf_ip_options_match(&ctrl->ctrl, opts);
2134 mutex_unlock(&nvme_tcp_ctrl_mutex);
2139 static struct nvme_ctrl *nvme_tcp_create_ctrl(struct device *dev,
2140 struct nvmf_ctrl_options *opts)
2142 struct nvme_tcp_ctrl *ctrl;
2145 ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
2147 return ERR_PTR(-ENOMEM);
2149 INIT_LIST_HEAD(&ctrl->list);
2150 ctrl->ctrl.opts = opts;
2151 ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues + 1;
2152 ctrl->ctrl.sqsize = opts->queue_size - 1;
2153 ctrl->ctrl.kato = opts->kato;
2155 INIT_DELAYED_WORK(&ctrl->connect_work,
2156 nvme_tcp_reconnect_ctrl_work);
2157 INIT_WORK(&ctrl->err_work, nvme_tcp_error_recovery_work);
2158 INIT_WORK(&ctrl->ctrl.reset_work, nvme_reset_ctrl_work);
2160 if (!(opts->mask & NVMF_OPT_TRSVCID)) {
2162 kstrdup(__stringify(NVME_TCP_DISC_PORT), GFP_KERNEL);
2163 if (!opts->trsvcid) {
2167 opts->mask |= NVMF_OPT_TRSVCID;
2170 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2171 opts->traddr, opts->trsvcid, &ctrl->addr);
2173 pr_err("malformed address passed: %s:%s\n",
2174 opts->traddr, opts->trsvcid);
2178 if (opts->mask & NVMF_OPT_HOST_TRADDR) {
2179 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2180 opts->host_traddr, NULL, &ctrl->src_addr);
2182 pr_err("malformed src address passed: %s\n",
2188 if (!opts->duplicate_connect && nvme_tcp_existing_controller(opts)) {
2193 ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
2195 if (!ctrl->queues) {
2200 ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_tcp_ctrl_ops, 0);
2202 goto out_kfree_queues;
2204 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
2207 goto out_uninit_ctrl;
2210 ret = nvme_tcp_setup_ctrl(&ctrl->ctrl, true);
2212 goto out_uninit_ctrl;
2214 dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISp\n",
2215 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
2217 nvme_get_ctrl(&ctrl->ctrl);
2219 mutex_lock(&nvme_tcp_ctrl_mutex);
2220 list_add_tail(&ctrl->list, &nvme_tcp_ctrl_list);
2221 mutex_unlock(&nvme_tcp_ctrl_mutex);
2226 nvme_uninit_ctrl(&ctrl->ctrl);
2227 nvme_put_ctrl(&ctrl->ctrl);
2230 return ERR_PTR(ret);
2232 kfree(ctrl->queues);
2235 return ERR_PTR(ret);
2238 static struct nvmf_transport_ops nvme_tcp_transport = {
2240 .module = THIS_MODULE,
2241 .required_opts = NVMF_OPT_TRADDR,
2242 .allowed_opts = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2243 NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
2244 NVMF_OPT_HDR_DIGEST | NVMF_OPT_DATA_DIGEST |
2245 NVMF_OPT_NR_WRITE_QUEUES,
2246 .create_ctrl = nvme_tcp_create_ctrl,
2249 static int __init nvme_tcp_init_module(void)
2251 nvme_tcp_wq = alloc_workqueue("nvme_tcp_wq",
2252 WQ_MEM_RECLAIM | WQ_HIGHPRI, 0);
2256 nvmf_register_transport(&nvme_tcp_transport);
2260 static void __exit nvme_tcp_cleanup_module(void)
2262 struct nvme_tcp_ctrl *ctrl;
2264 nvmf_unregister_transport(&nvme_tcp_transport);
2266 mutex_lock(&nvme_tcp_ctrl_mutex);
2267 list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list)
2268 nvme_delete_ctrl(&ctrl->ctrl);
2269 mutex_unlock(&nvme_tcp_ctrl_mutex);
2270 flush_workqueue(nvme_delete_wq);
2272 destroy_workqueue(nvme_tcp_wq);
2275 module_init(nvme_tcp_init_module);
2276 module_exit(nvme_tcp_cleanup_module);
2278 MODULE_LICENSE("GPL v2");