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/key.h>
12 #include <linux/nvme-tcp.h>
13 #include <linux/nvme-keyring.h>
17 #include <net/tls_prot.h>
18 #include <net/handshake.h>
19 #include <linux/blk-mq.h>
20 #include <crypto/hash.h>
21 #include <net/busy_poll.h>
22 #include <trace/events/sock.h>
27 struct nvme_tcp_queue;
29 /* Define the socket priority to use for connections were it is desirable
30 * that the NIC consider performing optimized packet processing or filtering.
31 * A non-zero value being sufficient to indicate general consideration of any
32 * possible optimization. Making it a module param allows for alternative
33 * values that may be unique for some NIC implementations.
35 static int so_priority;
36 module_param(so_priority, int, 0644);
37 MODULE_PARM_DESC(so_priority, "nvme tcp socket optimize priority");
40 * TLS handshake timeout
42 static int tls_handshake_timeout = 10;
43 #ifdef CONFIG_NVME_TCP_TLS
44 module_param(tls_handshake_timeout, int, 0644);
45 MODULE_PARM_DESC(tls_handshake_timeout,
46 "nvme TLS handshake timeout in seconds (default 10)");
49 #ifdef CONFIG_DEBUG_LOCK_ALLOC
50 /* lockdep can detect a circular dependency of the form
51 * sk_lock -> mmap_lock (page fault) -> fs locks -> sk_lock
52 * because dependencies are tracked for both nvme-tcp and user contexts. Using
53 * a separate class prevents lockdep from conflating nvme-tcp socket use with
54 * user-space socket API use.
56 static struct lock_class_key nvme_tcp_sk_key[2];
57 static struct lock_class_key nvme_tcp_slock_key[2];
59 static void nvme_tcp_reclassify_socket(struct socket *sock)
61 struct sock *sk = sock->sk;
63 if (WARN_ON_ONCE(!sock_allow_reclassification(sk)))
66 switch (sk->sk_family) {
68 sock_lock_init_class_and_name(sk, "slock-AF_INET-NVME",
69 &nvme_tcp_slock_key[0],
70 "sk_lock-AF_INET-NVME",
74 sock_lock_init_class_and_name(sk, "slock-AF_INET6-NVME",
75 &nvme_tcp_slock_key[1],
76 "sk_lock-AF_INET6-NVME",
84 static void nvme_tcp_reclassify_socket(struct socket *sock) { }
87 enum nvme_tcp_send_state {
88 NVME_TCP_SEND_CMD_PDU = 0,
89 NVME_TCP_SEND_H2C_PDU,
94 struct nvme_tcp_request {
95 struct nvme_request req;
97 struct nvme_tcp_queue *queue;
105 struct list_head entry;
106 struct llist_node lentry;
109 struct bio *curr_bio;
110 struct iov_iter iter;
115 enum nvme_tcp_send_state state;
118 enum nvme_tcp_queue_flags {
119 NVME_TCP_Q_ALLOCATED = 0,
121 NVME_TCP_Q_POLLING = 2,
124 enum nvme_tcp_recv_state {
125 NVME_TCP_RECV_PDU = 0,
130 struct nvme_tcp_ctrl;
131 struct nvme_tcp_queue {
133 struct work_struct io_work;
136 struct mutex queue_lock;
137 struct mutex send_mutex;
138 struct llist_head req_list;
139 struct list_head send_list;
145 size_t data_remaining;
146 size_t ddgst_remaining;
150 struct nvme_tcp_request *request;
153 size_t cmnd_capsule_len;
154 struct nvme_tcp_ctrl *ctrl;
160 struct ahash_request *rcv_hash;
161 struct ahash_request *snd_hash;
164 struct completion tls_complete;
166 struct page_frag_cache pf_cache;
168 void (*state_change)(struct sock *);
169 void (*data_ready)(struct sock *);
170 void (*write_space)(struct sock *);
173 struct nvme_tcp_ctrl {
174 /* read only in the hot path */
175 struct nvme_tcp_queue *queues;
176 struct blk_mq_tag_set tag_set;
178 /* other member variables */
179 struct list_head list;
180 struct blk_mq_tag_set admin_tag_set;
181 struct sockaddr_storage addr;
182 struct sockaddr_storage src_addr;
183 struct nvme_ctrl ctrl;
185 struct work_struct err_work;
186 struct delayed_work connect_work;
187 struct nvme_tcp_request async_req;
188 u32 io_queues[HCTX_MAX_TYPES];
191 static LIST_HEAD(nvme_tcp_ctrl_list);
192 static DEFINE_MUTEX(nvme_tcp_ctrl_mutex);
193 static struct workqueue_struct *nvme_tcp_wq;
194 static const struct blk_mq_ops nvme_tcp_mq_ops;
195 static const struct blk_mq_ops nvme_tcp_admin_mq_ops;
196 static int nvme_tcp_try_send(struct nvme_tcp_queue *queue);
198 static inline struct nvme_tcp_ctrl *to_tcp_ctrl(struct nvme_ctrl *ctrl)
200 return container_of(ctrl, struct nvme_tcp_ctrl, ctrl);
203 static inline int nvme_tcp_queue_id(struct nvme_tcp_queue *queue)
205 return queue - queue->ctrl->queues;
208 static inline bool nvme_tcp_tls(struct nvme_ctrl *ctrl)
210 if (!IS_ENABLED(CONFIG_NVME_TCP_TLS))
213 return ctrl->opts->tls;
216 static inline struct blk_mq_tags *nvme_tcp_tagset(struct nvme_tcp_queue *queue)
218 u32 queue_idx = nvme_tcp_queue_id(queue);
221 return queue->ctrl->admin_tag_set.tags[queue_idx];
222 return queue->ctrl->tag_set.tags[queue_idx - 1];
225 static inline u8 nvme_tcp_hdgst_len(struct nvme_tcp_queue *queue)
227 return queue->hdr_digest ? NVME_TCP_DIGEST_LENGTH : 0;
230 static inline u8 nvme_tcp_ddgst_len(struct nvme_tcp_queue *queue)
232 return queue->data_digest ? NVME_TCP_DIGEST_LENGTH : 0;
235 static inline void *nvme_tcp_req_cmd_pdu(struct nvme_tcp_request *req)
240 static inline void *nvme_tcp_req_data_pdu(struct nvme_tcp_request *req)
242 /* use the pdu space in the back for the data pdu */
243 return req->pdu + sizeof(struct nvme_tcp_cmd_pdu) -
244 sizeof(struct nvme_tcp_data_pdu);
247 static inline size_t nvme_tcp_inline_data_size(struct nvme_tcp_request *req)
249 if (nvme_is_fabrics(req->req.cmd))
250 return NVME_TCP_ADMIN_CCSZ;
251 return req->queue->cmnd_capsule_len - sizeof(struct nvme_command);
254 static inline bool nvme_tcp_async_req(struct nvme_tcp_request *req)
256 return req == &req->queue->ctrl->async_req;
259 static inline bool nvme_tcp_has_inline_data(struct nvme_tcp_request *req)
263 if (unlikely(nvme_tcp_async_req(req)))
264 return false; /* async events don't have a request */
266 rq = blk_mq_rq_from_pdu(req);
268 return rq_data_dir(rq) == WRITE && req->data_len &&
269 req->data_len <= nvme_tcp_inline_data_size(req);
272 static inline struct page *nvme_tcp_req_cur_page(struct nvme_tcp_request *req)
274 return req->iter.bvec->bv_page;
277 static inline size_t nvme_tcp_req_cur_offset(struct nvme_tcp_request *req)
279 return req->iter.bvec->bv_offset + req->iter.iov_offset;
282 static inline size_t nvme_tcp_req_cur_length(struct nvme_tcp_request *req)
284 return min_t(size_t, iov_iter_single_seg_count(&req->iter),
285 req->pdu_len - req->pdu_sent);
288 static inline size_t nvme_tcp_pdu_data_left(struct nvme_tcp_request *req)
290 return rq_data_dir(blk_mq_rq_from_pdu(req)) == WRITE ?
291 req->pdu_len - req->pdu_sent : 0;
294 static inline size_t nvme_tcp_pdu_last_send(struct nvme_tcp_request *req,
297 return nvme_tcp_pdu_data_left(req) <= len;
300 static void nvme_tcp_init_iter(struct nvme_tcp_request *req,
303 struct request *rq = blk_mq_rq_from_pdu(req);
309 if (rq->rq_flags & RQF_SPECIAL_PAYLOAD) {
310 vec = &rq->special_vec;
312 size = blk_rq_payload_bytes(rq);
315 struct bio *bio = req->curr_bio;
319 vec = __bvec_iter_bvec(bio->bi_io_vec, bio->bi_iter);
321 bio_for_each_bvec(bv, bio, bi) {
324 size = bio->bi_iter.bi_size;
325 offset = bio->bi_iter.bi_bvec_done;
328 iov_iter_bvec(&req->iter, dir, vec, nr_bvec, size);
329 req->iter.iov_offset = offset;
332 static inline void nvme_tcp_advance_req(struct nvme_tcp_request *req,
335 req->data_sent += len;
336 req->pdu_sent += len;
337 iov_iter_advance(&req->iter, len);
338 if (!iov_iter_count(&req->iter) &&
339 req->data_sent < req->data_len) {
340 req->curr_bio = req->curr_bio->bi_next;
341 nvme_tcp_init_iter(req, ITER_SOURCE);
345 static inline void nvme_tcp_send_all(struct nvme_tcp_queue *queue)
349 /* drain the send queue as much as we can... */
351 ret = nvme_tcp_try_send(queue);
355 static inline bool nvme_tcp_queue_more(struct nvme_tcp_queue *queue)
357 return !list_empty(&queue->send_list) ||
358 !llist_empty(&queue->req_list);
361 static inline void nvme_tcp_queue_request(struct nvme_tcp_request *req,
362 bool sync, bool last)
364 struct nvme_tcp_queue *queue = req->queue;
367 empty = llist_add(&req->lentry, &queue->req_list) &&
368 list_empty(&queue->send_list) && !queue->request;
371 * if we're the first on the send_list and we can try to send
372 * directly, otherwise queue io_work. Also, only do that if we
373 * are on the same cpu, so we don't introduce contention.
375 if (queue->io_cpu == raw_smp_processor_id() &&
376 sync && empty && mutex_trylock(&queue->send_mutex)) {
377 nvme_tcp_send_all(queue);
378 mutex_unlock(&queue->send_mutex);
381 if (last && nvme_tcp_queue_more(queue))
382 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
385 static void nvme_tcp_process_req_list(struct nvme_tcp_queue *queue)
387 struct nvme_tcp_request *req;
388 struct llist_node *node;
390 for (node = llist_del_all(&queue->req_list); node; node = node->next) {
391 req = llist_entry(node, struct nvme_tcp_request, lentry);
392 list_add(&req->entry, &queue->send_list);
396 static inline struct nvme_tcp_request *
397 nvme_tcp_fetch_request(struct nvme_tcp_queue *queue)
399 struct nvme_tcp_request *req;
401 req = list_first_entry_or_null(&queue->send_list,
402 struct nvme_tcp_request, entry);
404 nvme_tcp_process_req_list(queue);
405 req = list_first_entry_or_null(&queue->send_list,
406 struct nvme_tcp_request, entry);
411 list_del(&req->entry);
415 static inline void nvme_tcp_ddgst_final(struct ahash_request *hash,
418 ahash_request_set_crypt(hash, NULL, (u8 *)dgst, 0);
419 crypto_ahash_final(hash);
422 static inline void nvme_tcp_ddgst_update(struct ahash_request *hash,
423 struct page *page, off_t off, size_t len)
425 struct scatterlist sg;
427 sg_init_table(&sg, 1);
428 sg_set_page(&sg, page, len, off);
429 ahash_request_set_crypt(hash, &sg, NULL, len);
430 crypto_ahash_update(hash);
433 static inline void nvme_tcp_hdgst(struct ahash_request *hash,
434 void *pdu, size_t len)
436 struct scatterlist sg;
438 sg_init_one(&sg, pdu, len);
439 ahash_request_set_crypt(hash, &sg, pdu + len, len);
440 crypto_ahash_digest(hash);
443 static int nvme_tcp_verify_hdgst(struct nvme_tcp_queue *queue,
444 void *pdu, size_t pdu_len)
446 struct nvme_tcp_hdr *hdr = pdu;
450 if (unlikely(!(hdr->flags & NVME_TCP_F_HDGST))) {
451 dev_err(queue->ctrl->ctrl.device,
452 "queue %d: header digest flag is cleared\n",
453 nvme_tcp_queue_id(queue));
457 recv_digest = *(__le32 *)(pdu + hdr->hlen);
458 nvme_tcp_hdgst(queue->rcv_hash, pdu, pdu_len);
459 exp_digest = *(__le32 *)(pdu + hdr->hlen);
460 if (recv_digest != exp_digest) {
461 dev_err(queue->ctrl->ctrl.device,
462 "header digest error: recv %#x expected %#x\n",
463 le32_to_cpu(recv_digest), le32_to_cpu(exp_digest));
470 static int nvme_tcp_check_ddgst(struct nvme_tcp_queue *queue, void *pdu)
472 struct nvme_tcp_hdr *hdr = pdu;
473 u8 digest_len = nvme_tcp_hdgst_len(queue);
476 len = le32_to_cpu(hdr->plen) - hdr->hlen -
477 ((hdr->flags & NVME_TCP_F_HDGST) ? digest_len : 0);
479 if (unlikely(len && !(hdr->flags & NVME_TCP_F_DDGST))) {
480 dev_err(queue->ctrl->ctrl.device,
481 "queue %d: data digest flag is cleared\n",
482 nvme_tcp_queue_id(queue));
485 crypto_ahash_init(queue->rcv_hash);
490 static void nvme_tcp_exit_request(struct blk_mq_tag_set *set,
491 struct request *rq, unsigned int hctx_idx)
493 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
495 page_frag_free(req->pdu);
498 static int nvme_tcp_init_request(struct blk_mq_tag_set *set,
499 struct request *rq, unsigned int hctx_idx,
500 unsigned int numa_node)
502 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(set->driver_data);
503 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
504 struct nvme_tcp_cmd_pdu *pdu;
505 int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
506 struct nvme_tcp_queue *queue = &ctrl->queues[queue_idx];
507 u8 hdgst = nvme_tcp_hdgst_len(queue);
509 req->pdu = page_frag_alloc(&queue->pf_cache,
510 sizeof(struct nvme_tcp_cmd_pdu) + hdgst,
511 GFP_KERNEL | __GFP_ZERO);
517 nvme_req(rq)->ctrl = &ctrl->ctrl;
518 nvme_req(rq)->cmd = &pdu->cmd;
523 static int nvme_tcp_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
524 unsigned int hctx_idx)
526 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(data);
527 struct nvme_tcp_queue *queue = &ctrl->queues[hctx_idx + 1];
529 hctx->driver_data = queue;
533 static int nvme_tcp_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
534 unsigned int hctx_idx)
536 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(data);
537 struct nvme_tcp_queue *queue = &ctrl->queues[0];
539 hctx->driver_data = queue;
543 static enum nvme_tcp_recv_state
544 nvme_tcp_recv_state(struct nvme_tcp_queue *queue)
546 return (queue->pdu_remaining) ? NVME_TCP_RECV_PDU :
547 (queue->ddgst_remaining) ? NVME_TCP_RECV_DDGST :
551 static void nvme_tcp_init_recv_ctx(struct nvme_tcp_queue *queue)
553 queue->pdu_remaining = sizeof(struct nvme_tcp_rsp_pdu) +
554 nvme_tcp_hdgst_len(queue);
555 queue->pdu_offset = 0;
556 queue->data_remaining = -1;
557 queue->ddgst_remaining = 0;
560 static void nvme_tcp_error_recovery(struct nvme_ctrl *ctrl)
562 if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_RESETTING))
565 dev_warn(ctrl->device, "starting error recovery\n");
566 queue_work(nvme_reset_wq, &to_tcp_ctrl(ctrl)->err_work);
569 static int nvme_tcp_process_nvme_cqe(struct nvme_tcp_queue *queue,
570 struct nvme_completion *cqe)
572 struct nvme_tcp_request *req;
575 rq = nvme_find_rq(nvme_tcp_tagset(queue), cqe->command_id);
577 dev_err(queue->ctrl->ctrl.device,
578 "got bad cqe.command_id %#x on queue %d\n",
579 cqe->command_id, nvme_tcp_queue_id(queue));
580 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
584 req = blk_mq_rq_to_pdu(rq);
585 if (req->status == cpu_to_le16(NVME_SC_SUCCESS))
586 req->status = cqe->status;
588 if (!nvme_try_complete_req(rq, req->status, cqe->result))
589 nvme_complete_rq(rq);
595 static int nvme_tcp_handle_c2h_data(struct nvme_tcp_queue *queue,
596 struct nvme_tcp_data_pdu *pdu)
600 rq = nvme_find_rq(nvme_tcp_tagset(queue), pdu->command_id);
602 dev_err(queue->ctrl->ctrl.device,
603 "got bad c2hdata.command_id %#x on queue %d\n",
604 pdu->command_id, nvme_tcp_queue_id(queue));
608 if (!blk_rq_payload_bytes(rq)) {
609 dev_err(queue->ctrl->ctrl.device,
610 "queue %d tag %#x unexpected data\n",
611 nvme_tcp_queue_id(queue), rq->tag);
615 queue->data_remaining = le32_to_cpu(pdu->data_length);
617 if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS &&
618 unlikely(!(pdu->hdr.flags & NVME_TCP_F_DATA_LAST))) {
619 dev_err(queue->ctrl->ctrl.device,
620 "queue %d tag %#x SUCCESS set but not last PDU\n",
621 nvme_tcp_queue_id(queue), rq->tag);
622 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
629 static int nvme_tcp_handle_comp(struct nvme_tcp_queue *queue,
630 struct nvme_tcp_rsp_pdu *pdu)
632 struct nvme_completion *cqe = &pdu->cqe;
636 * AEN requests are special as they don't time out and can
637 * survive any kind of queue freeze and often don't respond to
638 * aborts. We don't even bother to allocate a struct request
639 * for them but rather special case them here.
641 if (unlikely(nvme_is_aen_req(nvme_tcp_queue_id(queue),
643 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
646 ret = nvme_tcp_process_nvme_cqe(queue, cqe);
651 static void nvme_tcp_setup_h2c_data_pdu(struct nvme_tcp_request *req)
653 struct nvme_tcp_data_pdu *data = nvme_tcp_req_data_pdu(req);
654 struct nvme_tcp_queue *queue = req->queue;
655 struct request *rq = blk_mq_rq_from_pdu(req);
656 u32 h2cdata_sent = req->pdu_len;
657 u8 hdgst = nvme_tcp_hdgst_len(queue);
658 u8 ddgst = nvme_tcp_ddgst_len(queue);
660 req->state = NVME_TCP_SEND_H2C_PDU;
662 req->pdu_len = min(req->h2cdata_left, queue->maxh2cdata);
664 req->h2cdata_left -= req->pdu_len;
665 req->h2cdata_offset += h2cdata_sent;
667 memset(data, 0, sizeof(*data));
668 data->hdr.type = nvme_tcp_h2c_data;
669 if (!req->h2cdata_left)
670 data->hdr.flags = NVME_TCP_F_DATA_LAST;
671 if (queue->hdr_digest)
672 data->hdr.flags |= NVME_TCP_F_HDGST;
673 if (queue->data_digest)
674 data->hdr.flags |= NVME_TCP_F_DDGST;
675 data->hdr.hlen = sizeof(*data);
676 data->hdr.pdo = data->hdr.hlen + hdgst;
678 cpu_to_le32(data->hdr.hlen + hdgst + req->pdu_len + ddgst);
679 data->ttag = req->ttag;
680 data->command_id = nvme_cid(rq);
681 data->data_offset = cpu_to_le32(req->h2cdata_offset);
682 data->data_length = cpu_to_le32(req->pdu_len);
685 static int nvme_tcp_handle_r2t(struct nvme_tcp_queue *queue,
686 struct nvme_tcp_r2t_pdu *pdu)
688 struct nvme_tcp_request *req;
690 u32 r2t_length = le32_to_cpu(pdu->r2t_length);
691 u32 r2t_offset = le32_to_cpu(pdu->r2t_offset);
693 rq = nvme_find_rq(nvme_tcp_tagset(queue), pdu->command_id);
695 dev_err(queue->ctrl->ctrl.device,
696 "got bad r2t.command_id %#x on queue %d\n",
697 pdu->command_id, nvme_tcp_queue_id(queue));
700 req = blk_mq_rq_to_pdu(rq);
702 if (unlikely(!r2t_length)) {
703 dev_err(queue->ctrl->ctrl.device,
704 "req %d r2t len is %u, probably a bug...\n",
705 rq->tag, r2t_length);
709 if (unlikely(req->data_sent + r2t_length > req->data_len)) {
710 dev_err(queue->ctrl->ctrl.device,
711 "req %d r2t len %u exceeded data len %u (%zu sent)\n",
712 rq->tag, r2t_length, req->data_len, req->data_sent);
716 if (unlikely(r2t_offset < req->data_sent)) {
717 dev_err(queue->ctrl->ctrl.device,
718 "req %d unexpected r2t offset %u (expected %zu)\n",
719 rq->tag, r2t_offset, req->data_sent);
724 req->h2cdata_left = r2t_length;
725 req->h2cdata_offset = r2t_offset;
726 req->ttag = pdu->ttag;
728 nvme_tcp_setup_h2c_data_pdu(req);
729 nvme_tcp_queue_request(req, false, true);
734 static int nvme_tcp_recv_pdu(struct nvme_tcp_queue *queue, struct sk_buff *skb,
735 unsigned int *offset, size_t *len)
737 struct nvme_tcp_hdr *hdr;
738 char *pdu = queue->pdu;
739 size_t rcv_len = min_t(size_t, *len, queue->pdu_remaining);
742 ret = skb_copy_bits(skb, *offset,
743 &pdu[queue->pdu_offset], rcv_len);
747 queue->pdu_remaining -= rcv_len;
748 queue->pdu_offset += rcv_len;
751 if (queue->pdu_remaining)
755 if (queue->hdr_digest) {
756 ret = nvme_tcp_verify_hdgst(queue, queue->pdu, hdr->hlen);
762 if (queue->data_digest) {
763 ret = nvme_tcp_check_ddgst(queue, queue->pdu);
769 case nvme_tcp_c2h_data:
770 return nvme_tcp_handle_c2h_data(queue, (void *)queue->pdu);
772 nvme_tcp_init_recv_ctx(queue);
773 return nvme_tcp_handle_comp(queue, (void *)queue->pdu);
775 nvme_tcp_init_recv_ctx(queue);
776 return nvme_tcp_handle_r2t(queue, (void *)queue->pdu);
778 dev_err(queue->ctrl->ctrl.device,
779 "unsupported pdu type (%d)\n", hdr->type);
784 static inline void nvme_tcp_end_request(struct request *rq, u16 status)
786 union nvme_result res = {};
788 if (!nvme_try_complete_req(rq, cpu_to_le16(status << 1), res))
789 nvme_complete_rq(rq);
792 static int nvme_tcp_recv_data(struct nvme_tcp_queue *queue, struct sk_buff *skb,
793 unsigned int *offset, size_t *len)
795 struct nvme_tcp_data_pdu *pdu = (void *)queue->pdu;
797 nvme_cid_to_rq(nvme_tcp_tagset(queue), pdu->command_id);
798 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
803 recv_len = min_t(size_t, *len, queue->data_remaining);
807 if (!iov_iter_count(&req->iter)) {
808 req->curr_bio = req->curr_bio->bi_next;
811 * If we don`t have any bios it means that controller
812 * sent more data than we requested, hence error
814 if (!req->curr_bio) {
815 dev_err(queue->ctrl->ctrl.device,
816 "queue %d no space in request %#x",
817 nvme_tcp_queue_id(queue), rq->tag);
818 nvme_tcp_init_recv_ctx(queue);
821 nvme_tcp_init_iter(req, ITER_DEST);
824 /* we can read only from what is left in this bio */
825 recv_len = min_t(size_t, recv_len,
826 iov_iter_count(&req->iter));
828 if (queue->data_digest)
829 ret = skb_copy_and_hash_datagram_iter(skb, *offset,
830 &req->iter, recv_len, queue->rcv_hash);
832 ret = skb_copy_datagram_iter(skb, *offset,
833 &req->iter, recv_len);
835 dev_err(queue->ctrl->ctrl.device,
836 "queue %d failed to copy request %#x data",
837 nvme_tcp_queue_id(queue), rq->tag);
843 queue->data_remaining -= recv_len;
846 if (!queue->data_remaining) {
847 if (queue->data_digest) {
848 nvme_tcp_ddgst_final(queue->rcv_hash, &queue->exp_ddgst);
849 queue->ddgst_remaining = NVME_TCP_DIGEST_LENGTH;
851 if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
852 nvme_tcp_end_request(rq,
853 le16_to_cpu(req->status));
856 nvme_tcp_init_recv_ctx(queue);
863 static int nvme_tcp_recv_ddgst(struct nvme_tcp_queue *queue,
864 struct sk_buff *skb, unsigned int *offset, size_t *len)
866 struct nvme_tcp_data_pdu *pdu = (void *)queue->pdu;
867 char *ddgst = (char *)&queue->recv_ddgst;
868 size_t recv_len = min_t(size_t, *len, queue->ddgst_remaining);
869 off_t off = NVME_TCP_DIGEST_LENGTH - queue->ddgst_remaining;
872 ret = skb_copy_bits(skb, *offset, &ddgst[off], recv_len);
876 queue->ddgst_remaining -= recv_len;
879 if (queue->ddgst_remaining)
882 if (queue->recv_ddgst != queue->exp_ddgst) {
883 struct request *rq = nvme_cid_to_rq(nvme_tcp_tagset(queue),
885 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
887 req->status = cpu_to_le16(NVME_SC_DATA_XFER_ERROR);
889 dev_err(queue->ctrl->ctrl.device,
890 "data digest error: recv %#x expected %#x\n",
891 le32_to_cpu(queue->recv_ddgst),
892 le32_to_cpu(queue->exp_ddgst));
895 if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
896 struct request *rq = nvme_cid_to_rq(nvme_tcp_tagset(queue),
898 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
900 nvme_tcp_end_request(rq, le16_to_cpu(req->status));
904 nvme_tcp_init_recv_ctx(queue);
908 static int nvme_tcp_recv_skb(read_descriptor_t *desc, struct sk_buff *skb,
909 unsigned int offset, size_t len)
911 struct nvme_tcp_queue *queue = desc->arg.data;
912 size_t consumed = len;
915 if (unlikely(!queue->rd_enabled))
919 switch (nvme_tcp_recv_state(queue)) {
920 case NVME_TCP_RECV_PDU:
921 result = nvme_tcp_recv_pdu(queue, skb, &offset, &len);
923 case NVME_TCP_RECV_DATA:
924 result = nvme_tcp_recv_data(queue, skb, &offset, &len);
926 case NVME_TCP_RECV_DDGST:
927 result = nvme_tcp_recv_ddgst(queue, skb, &offset, &len);
933 dev_err(queue->ctrl->ctrl.device,
934 "receive failed: %d\n", result);
935 queue->rd_enabled = false;
936 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
944 static void nvme_tcp_data_ready(struct sock *sk)
946 struct nvme_tcp_queue *queue;
948 trace_sk_data_ready(sk);
950 read_lock_bh(&sk->sk_callback_lock);
951 queue = sk->sk_user_data;
952 if (likely(queue && queue->rd_enabled) &&
953 !test_bit(NVME_TCP_Q_POLLING, &queue->flags))
954 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
955 read_unlock_bh(&sk->sk_callback_lock);
958 static void nvme_tcp_write_space(struct sock *sk)
960 struct nvme_tcp_queue *queue;
962 read_lock_bh(&sk->sk_callback_lock);
963 queue = sk->sk_user_data;
964 if (likely(queue && sk_stream_is_writeable(sk))) {
965 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
966 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
968 read_unlock_bh(&sk->sk_callback_lock);
971 static void nvme_tcp_state_change(struct sock *sk)
973 struct nvme_tcp_queue *queue;
975 read_lock_bh(&sk->sk_callback_lock);
976 queue = sk->sk_user_data;
980 switch (sk->sk_state) {
986 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
989 dev_info(queue->ctrl->ctrl.device,
990 "queue %d socket state %d\n",
991 nvme_tcp_queue_id(queue), sk->sk_state);
994 queue->state_change(sk);
996 read_unlock_bh(&sk->sk_callback_lock);
999 static inline void nvme_tcp_done_send_req(struct nvme_tcp_queue *queue)
1001 queue->request = NULL;
1004 static void nvme_tcp_fail_request(struct nvme_tcp_request *req)
1006 if (nvme_tcp_async_req(req)) {
1007 union nvme_result res = {};
1009 nvme_complete_async_event(&req->queue->ctrl->ctrl,
1010 cpu_to_le16(NVME_SC_HOST_PATH_ERROR), &res);
1012 nvme_tcp_end_request(blk_mq_rq_from_pdu(req),
1013 NVME_SC_HOST_PATH_ERROR);
1017 static int nvme_tcp_try_send_data(struct nvme_tcp_request *req)
1019 struct nvme_tcp_queue *queue = req->queue;
1020 int req_data_len = req->data_len;
1021 u32 h2cdata_left = req->h2cdata_left;
1024 struct bio_vec bvec;
1025 struct msghdr msg = {
1026 .msg_flags = MSG_DONTWAIT | MSG_SPLICE_PAGES,
1028 struct page *page = nvme_tcp_req_cur_page(req);
1029 size_t offset = nvme_tcp_req_cur_offset(req);
1030 size_t len = nvme_tcp_req_cur_length(req);
1031 bool last = nvme_tcp_pdu_last_send(req, len);
1032 int req_data_sent = req->data_sent;
1035 if (last && !queue->data_digest && !nvme_tcp_queue_more(queue))
1036 msg.msg_flags |= MSG_EOR;
1038 msg.msg_flags |= MSG_MORE;
1040 if (!sendpage_ok(page))
1041 msg.msg_flags &= ~MSG_SPLICE_PAGES;
1043 bvec_set_page(&bvec, page, len, offset);
1044 iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, &bvec, 1, len);
1045 ret = sock_sendmsg(queue->sock, &msg);
1049 if (queue->data_digest)
1050 nvme_tcp_ddgst_update(queue->snd_hash, page,
1054 * update the request iterator except for the last payload send
1055 * in the request where we don't want to modify it as we may
1056 * compete with the RX path completing the request.
1058 if (req_data_sent + ret < req_data_len)
1059 nvme_tcp_advance_req(req, ret);
1061 /* fully successful last send in current PDU */
1062 if (last && ret == len) {
1063 if (queue->data_digest) {
1064 nvme_tcp_ddgst_final(queue->snd_hash,
1066 req->state = NVME_TCP_SEND_DDGST;
1070 nvme_tcp_setup_h2c_data_pdu(req);
1072 nvme_tcp_done_send_req(queue);
1080 static int nvme_tcp_try_send_cmd_pdu(struct nvme_tcp_request *req)
1082 struct nvme_tcp_queue *queue = req->queue;
1083 struct nvme_tcp_cmd_pdu *pdu = nvme_tcp_req_cmd_pdu(req);
1084 struct bio_vec bvec;
1085 struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_SPLICE_PAGES, };
1086 bool inline_data = nvme_tcp_has_inline_data(req);
1087 u8 hdgst = nvme_tcp_hdgst_len(queue);
1088 int len = sizeof(*pdu) + hdgst - req->offset;
1091 if (inline_data || nvme_tcp_queue_more(queue))
1092 msg.msg_flags |= MSG_MORE;
1094 msg.msg_flags |= MSG_EOR;
1096 if (queue->hdr_digest && !req->offset)
1097 nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
1099 bvec_set_virt(&bvec, (void *)pdu + req->offset, len);
1100 iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, &bvec, 1, len);
1101 ret = sock_sendmsg(queue->sock, &msg);
1102 if (unlikely(ret <= 0))
1108 req->state = NVME_TCP_SEND_DATA;
1109 if (queue->data_digest)
1110 crypto_ahash_init(queue->snd_hash);
1112 nvme_tcp_done_send_req(queue);
1121 static int nvme_tcp_try_send_data_pdu(struct nvme_tcp_request *req)
1123 struct nvme_tcp_queue *queue = req->queue;
1124 struct nvme_tcp_data_pdu *pdu = nvme_tcp_req_data_pdu(req);
1125 struct bio_vec bvec;
1126 struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_MORE, };
1127 u8 hdgst = nvme_tcp_hdgst_len(queue);
1128 int len = sizeof(*pdu) - req->offset + hdgst;
1131 if (queue->hdr_digest && !req->offset)
1132 nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
1134 if (!req->h2cdata_left)
1135 msg.msg_flags |= MSG_SPLICE_PAGES;
1137 bvec_set_virt(&bvec, (void *)pdu + req->offset, len);
1138 iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, &bvec, 1, len);
1139 ret = sock_sendmsg(queue->sock, &msg);
1140 if (unlikely(ret <= 0))
1145 req->state = NVME_TCP_SEND_DATA;
1146 if (queue->data_digest)
1147 crypto_ahash_init(queue->snd_hash);
1155 static int nvme_tcp_try_send_ddgst(struct nvme_tcp_request *req)
1157 struct nvme_tcp_queue *queue = req->queue;
1158 size_t offset = req->offset;
1159 u32 h2cdata_left = req->h2cdata_left;
1161 struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
1163 .iov_base = (u8 *)&req->ddgst + req->offset,
1164 .iov_len = NVME_TCP_DIGEST_LENGTH - req->offset
1167 if (nvme_tcp_queue_more(queue))
1168 msg.msg_flags |= MSG_MORE;
1170 msg.msg_flags |= MSG_EOR;
1172 ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
1173 if (unlikely(ret <= 0))
1176 if (offset + ret == NVME_TCP_DIGEST_LENGTH) {
1178 nvme_tcp_setup_h2c_data_pdu(req);
1180 nvme_tcp_done_send_req(queue);
1188 static int nvme_tcp_try_send(struct nvme_tcp_queue *queue)
1190 struct nvme_tcp_request *req;
1191 unsigned int noreclaim_flag;
1194 if (!queue->request) {
1195 queue->request = nvme_tcp_fetch_request(queue);
1196 if (!queue->request)
1199 req = queue->request;
1201 noreclaim_flag = memalloc_noreclaim_save();
1202 if (req->state == NVME_TCP_SEND_CMD_PDU) {
1203 ret = nvme_tcp_try_send_cmd_pdu(req);
1206 if (!nvme_tcp_has_inline_data(req))
1210 if (req->state == NVME_TCP_SEND_H2C_PDU) {
1211 ret = nvme_tcp_try_send_data_pdu(req);
1216 if (req->state == NVME_TCP_SEND_DATA) {
1217 ret = nvme_tcp_try_send_data(req);
1222 if (req->state == NVME_TCP_SEND_DDGST)
1223 ret = nvme_tcp_try_send_ddgst(req);
1225 if (ret == -EAGAIN) {
1227 } else if (ret < 0) {
1228 dev_err(queue->ctrl->ctrl.device,
1229 "failed to send request %d\n", ret);
1230 nvme_tcp_fail_request(queue->request);
1231 nvme_tcp_done_send_req(queue);
1234 memalloc_noreclaim_restore(noreclaim_flag);
1238 static int nvme_tcp_try_recv(struct nvme_tcp_queue *queue)
1240 struct socket *sock = queue->sock;
1241 struct sock *sk = sock->sk;
1242 read_descriptor_t rd_desc;
1245 rd_desc.arg.data = queue;
1249 consumed = sock->ops->read_sock(sk, &rd_desc, nvme_tcp_recv_skb);
1254 static void nvme_tcp_io_work(struct work_struct *w)
1256 struct nvme_tcp_queue *queue =
1257 container_of(w, struct nvme_tcp_queue, io_work);
1258 unsigned long deadline = jiffies + msecs_to_jiffies(1);
1261 bool pending = false;
1264 if (mutex_trylock(&queue->send_mutex)) {
1265 result = nvme_tcp_try_send(queue);
1266 mutex_unlock(&queue->send_mutex);
1269 else if (unlikely(result < 0))
1273 result = nvme_tcp_try_recv(queue);
1276 else if (unlikely(result < 0))
1279 if (!pending || !queue->rd_enabled)
1282 } while (!time_after(jiffies, deadline)); /* quota is exhausted */
1284 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
1287 static void nvme_tcp_free_crypto(struct nvme_tcp_queue *queue)
1289 struct crypto_ahash *tfm = crypto_ahash_reqtfm(queue->rcv_hash);
1291 ahash_request_free(queue->rcv_hash);
1292 ahash_request_free(queue->snd_hash);
1293 crypto_free_ahash(tfm);
1296 static int nvme_tcp_alloc_crypto(struct nvme_tcp_queue *queue)
1298 struct crypto_ahash *tfm;
1300 tfm = crypto_alloc_ahash("crc32c", 0, CRYPTO_ALG_ASYNC);
1302 return PTR_ERR(tfm);
1304 queue->snd_hash = ahash_request_alloc(tfm, GFP_KERNEL);
1305 if (!queue->snd_hash)
1307 ahash_request_set_callback(queue->snd_hash, 0, NULL, NULL);
1309 queue->rcv_hash = ahash_request_alloc(tfm, GFP_KERNEL);
1310 if (!queue->rcv_hash)
1312 ahash_request_set_callback(queue->rcv_hash, 0, NULL, NULL);
1316 ahash_request_free(queue->snd_hash);
1318 crypto_free_ahash(tfm);
1322 static void nvme_tcp_free_async_req(struct nvme_tcp_ctrl *ctrl)
1324 struct nvme_tcp_request *async = &ctrl->async_req;
1326 page_frag_free(async->pdu);
1329 static int nvme_tcp_alloc_async_req(struct nvme_tcp_ctrl *ctrl)
1331 struct nvme_tcp_queue *queue = &ctrl->queues[0];
1332 struct nvme_tcp_request *async = &ctrl->async_req;
1333 u8 hdgst = nvme_tcp_hdgst_len(queue);
1335 async->pdu = page_frag_alloc(&queue->pf_cache,
1336 sizeof(struct nvme_tcp_cmd_pdu) + hdgst,
1337 GFP_KERNEL | __GFP_ZERO);
1341 async->queue = &ctrl->queues[0];
1345 static void nvme_tcp_free_queue(struct nvme_ctrl *nctrl, int qid)
1348 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1349 struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1350 unsigned int noreclaim_flag;
1352 if (!test_and_clear_bit(NVME_TCP_Q_ALLOCATED, &queue->flags))
1355 if (queue->hdr_digest || queue->data_digest)
1356 nvme_tcp_free_crypto(queue);
1358 if (queue->pf_cache.va) {
1359 page = virt_to_head_page(queue->pf_cache.va);
1360 __page_frag_cache_drain(page, queue->pf_cache.pagecnt_bias);
1361 queue->pf_cache.va = NULL;
1364 noreclaim_flag = memalloc_noreclaim_save();
1365 /* ->sock will be released by fput() */
1366 fput(queue->sock->file);
1368 memalloc_noreclaim_restore(noreclaim_flag);
1371 mutex_destroy(&queue->send_mutex);
1372 mutex_destroy(&queue->queue_lock);
1375 static int nvme_tcp_init_connection(struct nvme_tcp_queue *queue)
1377 struct nvme_tcp_icreq_pdu *icreq;
1378 struct nvme_tcp_icresp_pdu *icresp;
1379 char cbuf[CMSG_LEN(sizeof(char))] = {};
1381 struct msghdr msg = {};
1383 bool ctrl_hdgst, ctrl_ddgst;
1387 icreq = kzalloc(sizeof(*icreq), GFP_KERNEL);
1391 icresp = kzalloc(sizeof(*icresp), GFP_KERNEL);
1397 icreq->hdr.type = nvme_tcp_icreq;
1398 icreq->hdr.hlen = sizeof(*icreq);
1400 icreq->hdr.plen = cpu_to_le32(icreq->hdr.hlen);
1401 icreq->pfv = cpu_to_le16(NVME_TCP_PFV_1_0);
1402 icreq->maxr2t = 0; /* single inflight r2t supported */
1403 icreq->hpda = 0; /* no alignment constraint */
1404 if (queue->hdr_digest)
1405 icreq->digest |= NVME_TCP_HDR_DIGEST_ENABLE;
1406 if (queue->data_digest)
1407 icreq->digest |= NVME_TCP_DATA_DIGEST_ENABLE;
1409 iov.iov_base = icreq;
1410 iov.iov_len = sizeof(*icreq);
1411 ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
1413 pr_warn("queue %d: failed to send icreq, error %d\n",
1414 nvme_tcp_queue_id(queue), ret);
1418 memset(&msg, 0, sizeof(msg));
1419 iov.iov_base = icresp;
1420 iov.iov_len = sizeof(*icresp);
1421 if (nvme_tcp_tls(&queue->ctrl->ctrl)) {
1422 msg.msg_control = cbuf;
1423 msg.msg_controllen = sizeof(cbuf);
1425 ret = kernel_recvmsg(queue->sock, &msg, &iov, 1,
1426 iov.iov_len, msg.msg_flags);
1428 pr_warn("queue %d: failed to receive icresp, error %d\n",
1429 nvme_tcp_queue_id(queue), ret);
1433 if (nvme_tcp_tls(&queue->ctrl->ctrl)) {
1434 ctype = tls_get_record_type(queue->sock->sk,
1435 (struct cmsghdr *)cbuf);
1436 if (ctype != TLS_RECORD_TYPE_DATA) {
1437 pr_err("queue %d: unhandled TLS record %d\n",
1438 nvme_tcp_queue_id(queue), ctype);
1443 if (icresp->hdr.type != nvme_tcp_icresp) {
1444 pr_err("queue %d: bad type returned %d\n",
1445 nvme_tcp_queue_id(queue), icresp->hdr.type);
1449 if (le32_to_cpu(icresp->hdr.plen) != sizeof(*icresp)) {
1450 pr_err("queue %d: bad pdu length returned %d\n",
1451 nvme_tcp_queue_id(queue), icresp->hdr.plen);
1455 if (icresp->pfv != NVME_TCP_PFV_1_0) {
1456 pr_err("queue %d: bad pfv returned %d\n",
1457 nvme_tcp_queue_id(queue), icresp->pfv);
1461 ctrl_ddgst = !!(icresp->digest & NVME_TCP_DATA_DIGEST_ENABLE);
1462 if ((queue->data_digest && !ctrl_ddgst) ||
1463 (!queue->data_digest && ctrl_ddgst)) {
1464 pr_err("queue %d: data digest mismatch host: %s ctrl: %s\n",
1465 nvme_tcp_queue_id(queue),
1466 queue->data_digest ? "enabled" : "disabled",
1467 ctrl_ddgst ? "enabled" : "disabled");
1471 ctrl_hdgst = !!(icresp->digest & NVME_TCP_HDR_DIGEST_ENABLE);
1472 if ((queue->hdr_digest && !ctrl_hdgst) ||
1473 (!queue->hdr_digest && ctrl_hdgst)) {
1474 pr_err("queue %d: header digest mismatch host: %s ctrl: %s\n",
1475 nvme_tcp_queue_id(queue),
1476 queue->hdr_digest ? "enabled" : "disabled",
1477 ctrl_hdgst ? "enabled" : "disabled");
1481 if (icresp->cpda != 0) {
1482 pr_err("queue %d: unsupported cpda returned %d\n",
1483 nvme_tcp_queue_id(queue), icresp->cpda);
1487 maxh2cdata = le32_to_cpu(icresp->maxdata);
1488 if ((maxh2cdata % 4) || (maxh2cdata < NVME_TCP_MIN_MAXH2CDATA)) {
1489 pr_err("queue %d: invalid maxh2cdata returned %u\n",
1490 nvme_tcp_queue_id(queue), maxh2cdata);
1493 queue->maxh2cdata = maxh2cdata;
1503 static bool nvme_tcp_admin_queue(struct nvme_tcp_queue *queue)
1505 return nvme_tcp_queue_id(queue) == 0;
1508 static bool nvme_tcp_default_queue(struct nvme_tcp_queue *queue)
1510 struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1511 int qid = nvme_tcp_queue_id(queue);
1513 return !nvme_tcp_admin_queue(queue) &&
1514 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT];
1517 static bool nvme_tcp_read_queue(struct nvme_tcp_queue *queue)
1519 struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1520 int qid = nvme_tcp_queue_id(queue);
1522 return !nvme_tcp_admin_queue(queue) &&
1523 !nvme_tcp_default_queue(queue) &&
1524 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT] +
1525 ctrl->io_queues[HCTX_TYPE_READ];
1528 static bool nvme_tcp_poll_queue(struct nvme_tcp_queue *queue)
1530 struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1531 int qid = nvme_tcp_queue_id(queue);
1533 return !nvme_tcp_admin_queue(queue) &&
1534 !nvme_tcp_default_queue(queue) &&
1535 !nvme_tcp_read_queue(queue) &&
1536 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT] +
1537 ctrl->io_queues[HCTX_TYPE_READ] +
1538 ctrl->io_queues[HCTX_TYPE_POLL];
1541 static void nvme_tcp_set_queue_io_cpu(struct nvme_tcp_queue *queue)
1543 struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1544 int qid = nvme_tcp_queue_id(queue);
1547 if (nvme_tcp_default_queue(queue))
1549 else if (nvme_tcp_read_queue(queue))
1550 n = qid - ctrl->io_queues[HCTX_TYPE_DEFAULT] - 1;
1551 else if (nvme_tcp_poll_queue(queue))
1552 n = qid - ctrl->io_queues[HCTX_TYPE_DEFAULT] -
1553 ctrl->io_queues[HCTX_TYPE_READ] - 1;
1554 queue->io_cpu = cpumask_next_wrap(n - 1, cpu_online_mask, -1, false);
1557 static void nvme_tcp_tls_done(void *data, int status, key_serial_t pskid)
1559 struct nvme_tcp_queue *queue = data;
1560 struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1561 int qid = nvme_tcp_queue_id(queue);
1562 struct key *tls_key;
1564 dev_dbg(ctrl->ctrl.device, "queue %d: TLS handshake done, key %x, status %d\n",
1565 qid, pskid, status);
1568 queue->tls_err = -status;
1572 tls_key = key_lookup(pskid);
1573 if (IS_ERR(tls_key)) {
1574 dev_warn(ctrl->ctrl.device, "queue %d: Invalid key %x\n",
1576 queue->tls_err = -ENOKEY;
1578 ctrl->ctrl.tls_key = tls_key;
1583 complete(&queue->tls_complete);
1586 static int nvme_tcp_start_tls(struct nvme_ctrl *nctrl,
1587 struct nvme_tcp_queue *queue,
1590 int qid = nvme_tcp_queue_id(queue);
1592 struct tls_handshake_args args;
1593 unsigned long tmo = tls_handshake_timeout * HZ;
1594 key_serial_t keyring = nvme_keyring_id();
1596 dev_dbg(nctrl->device, "queue %d: start TLS with key %x\n",
1598 memset(&args, 0, sizeof(args));
1599 args.ta_sock = queue->sock;
1600 args.ta_done = nvme_tcp_tls_done;
1601 args.ta_data = queue;
1602 args.ta_my_peerids[0] = pskid;
1603 args.ta_num_peerids = 1;
1604 if (nctrl->opts->keyring)
1605 keyring = key_serial(nctrl->opts->keyring);
1606 args.ta_keyring = keyring;
1607 args.ta_timeout_ms = tls_handshake_timeout * 1000;
1608 queue->tls_err = -EOPNOTSUPP;
1609 init_completion(&queue->tls_complete);
1610 ret = tls_client_hello_psk(&args, GFP_KERNEL);
1612 dev_err(nctrl->device, "queue %d: failed to start TLS: %d\n",
1616 ret = wait_for_completion_interruptible_timeout(&queue->tls_complete, tmo);
1621 dev_err(nctrl->device,
1622 "queue %d: TLS handshake failed, error %d\n",
1624 tls_handshake_cancel(queue->sock->sk);
1626 dev_dbg(nctrl->device,
1627 "queue %d: TLS handshake complete, error %d\n",
1628 qid, queue->tls_err);
1629 ret = queue->tls_err;
1634 static int nvme_tcp_alloc_queue(struct nvme_ctrl *nctrl, int qid,
1637 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1638 struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1639 int ret, rcv_pdu_size;
1640 struct file *sock_file;
1642 mutex_init(&queue->queue_lock);
1644 init_llist_head(&queue->req_list);
1645 INIT_LIST_HEAD(&queue->send_list);
1646 mutex_init(&queue->send_mutex);
1647 INIT_WORK(&queue->io_work, nvme_tcp_io_work);
1650 queue->cmnd_capsule_len = nctrl->ioccsz * 16;
1652 queue->cmnd_capsule_len = sizeof(struct nvme_command) +
1653 NVME_TCP_ADMIN_CCSZ;
1655 ret = sock_create(ctrl->addr.ss_family, SOCK_STREAM,
1656 IPPROTO_TCP, &queue->sock);
1658 dev_err(nctrl->device,
1659 "failed to create socket: %d\n", ret);
1660 goto err_destroy_mutex;
1663 sock_file = sock_alloc_file(queue->sock, O_CLOEXEC, NULL);
1664 if (IS_ERR(sock_file)) {
1665 ret = PTR_ERR(sock_file);
1666 goto err_destroy_mutex;
1668 nvme_tcp_reclassify_socket(queue->sock);
1670 /* Single syn retry */
1671 tcp_sock_set_syncnt(queue->sock->sk, 1);
1673 /* Set TCP no delay */
1674 tcp_sock_set_nodelay(queue->sock->sk);
1677 * Cleanup whatever is sitting in the TCP transmit queue on socket
1678 * close. This is done to prevent stale data from being sent should
1679 * the network connection be restored before TCP times out.
1681 sock_no_linger(queue->sock->sk);
1683 if (so_priority > 0)
1684 sock_set_priority(queue->sock->sk, so_priority);
1686 /* Set socket type of service */
1687 if (nctrl->opts->tos >= 0)
1688 ip_sock_set_tos(queue->sock->sk, nctrl->opts->tos);
1690 /* Set 10 seconds timeout for icresp recvmsg */
1691 queue->sock->sk->sk_rcvtimeo = 10 * HZ;
1693 queue->sock->sk->sk_allocation = GFP_ATOMIC;
1694 queue->sock->sk->sk_use_task_frag = false;
1695 nvme_tcp_set_queue_io_cpu(queue);
1696 queue->request = NULL;
1697 queue->data_remaining = 0;
1698 queue->ddgst_remaining = 0;
1699 queue->pdu_remaining = 0;
1700 queue->pdu_offset = 0;
1701 sk_set_memalloc(queue->sock->sk);
1703 if (nctrl->opts->mask & NVMF_OPT_HOST_TRADDR) {
1704 ret = kernel_bind(queue->sock, (struct sockaddr *)&ctrl->src_addr,
1705 sizeof(ctrl->src_addr));
1707 dev_err(nctrl->device,
1708 "failed to bind queue %d socket %d\n",
1714 if (nctrl->opts->mask & NVMF_OPT_HOST_IFACE) {
1715 char *iface = nctrl->opts->host_iface;
1716 sockptr_t optval = KERNEL_SOCKPTR(iface);
1718 ret = sock_setsockopt(queue->sock, SOL_SOCKET, SO_BINDTODEVICE,
1719 optval, strlen(iface));
1721 dev_err(nctrl->device,
1722 "failed to bind to interface %s queue %d err %d\n",
1728 queue->hdr_digest = nctrl->opts->hdr_digest;
1729 queue->data_digest = nctrl->opts->data_digest;
1730 if (queue->hdr_digest || queue->data_digest) {
1731 ret = nvme_tcp_alloc_crypto(queue);
1733 dev_err(nctrl->device,
1734 "failed to allocate queue %d crypto\n", qid);
1739 rcv_pdu_size = sizeof(struct nvme_tcp_rsp_pdu) +
1740 nvme_tcp_hdgst_len(queue);
1741 queue->pdu = kmalloc(rcv_pdu_size, GFP_KERNEL);
1747 dev_dbg(nctrl->device, "connecting queue %d\n",
1748 nvme_tcp_queue_id(queue));
1750 ret = kernel_connect(queue->sock, (struct sockaddr *)&ctrl->addr,
1751 sizeof(ctrl->addr), 0);
1753 dev_err(nctrl->device,
1754 "failed to connect socket: %d\n", ret);
1758 /* If PSKs are configured try to start TLS */
1759 if (IS_ENABLED(CONFIG_NVME_TCP_TLS) && pskid) {
1760 ret = nvme_tcp_start_tls(nctrl, queue, pskid);
1762 goto err_init_connect;
1765 ret = nvme_tcp_init_connection(queue);
1767 goto err_init_connect;
1769 set_bit(NVME_TCP_Q_ALLOCATED, &queue->flags);
1774 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1778 if (queue->hdr_digest || queue->data_digest)
1779 nvme_tcp_free_crypto(queue);
1781 /* ->sock will be released by fput() */
1782 fput(queue->sock->file);
1785 mutex_destroy(&queue->send_mutex);
1786 mutex_destroy(&queue->queue_lock);
1790 static void nvme_tcp_restore_sock_ops(struct nvme_tcp_queue *queue)
1792 struct socket *sock = queue->sock;
1794 write_lock_bh(&sock->sk->sk_callback_lock);
1795 sock->sk->sk_user_data = NULL;
1796 sock->sk->sk_data_ready = queue->data_ready;
1797 sock->sk->sk_state_change = queue->state_change;
1798 sock->sk->sk_write_space = queue->write_space;
1799 write_unlock_bh(&sock->sk->sk_callback_lock);
1802 static void __nvme_tcp_stop_queue(struct nvme_tcp_queue *queue)
1804 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1805 nvme_tcp_restore_sock_ops(queue);
1806 cancel_work_sync(&queue->io_work);
1809 static void nvme_tcp_stop_queue(struct nvme_ctrl *nctrl, int qid)
1811 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1812 struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1814 if (!test_bit(NVME_TCP_Q_ALLOCATED, &queue->flags))
1817 mutex_lock(&queue->queue_lock);
1818 if (test_and_clear_bit(NVME_TCP_Q_LIVE, &queue->flags))
1819 __nvme_tcp_stop_queue(queue);
1820 mutex_unlock(&queue->queue_lock);
1823 static void nvme_tcp_setup_sock_ops(struct nvme_tcp_queue *queue)
1825 write_lock_bh(&queue->sock->sk->sk_callback_lock);
1826 queue->sock->sk->sk_user_data = queue;
1827 queue->state_change = queue->sock->sk->sk_state_change;
1828 queue->data_ready = queue->sock->sk->sk_data_ready;
1829 queue->write_space = queue->sock->sk->sk_write_space;
1830 queue->sock->sk->sk_data_ready = nvme_tcp_data_ready;
1831 queue->sock->sk->sk_state_change = nvme_tcp_state_change;
1832 queue->sock->sk->sk_write_space = nvme_tcp_write_space;
1833 #ifdef CONFIG_NET_RX_BUSY_POLL
1834 queue->sock->sk->sk_ll_usec = 1;
1836 write_unlock_bh(&queue->sock->sk->sk_callback_lock);
1839 static int nvme_tcp_start_queue(struct nvme_ctrl *nctrl, int idx)
1841 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1842 struct nvme_tcp_queue *queue = &ctrl->queues[idx];
1845 queue->rd_enabled = true;
1846 nvme_tcp_init_recv_ctx(queue);
1847 nvme_tcp_setup_sock_ops(queue);
1850 ret = nvmf_connect_io_queue(nctrl, idx);
1852 ret = nvmf_connect_admin_queue(nctrl);
1855 set_bit(NVME_TCP_Q_LIVE, &queue->flags);
1857 if (test_bit(NVME_TCP_Q_ALLOCATED, &queue->flags))
1858 __nvme_tcp_stop_queue(queue);
1859 dev_err(nctrl->device,
1860 "failed to connect queue: %d ret=%d\n", idx, ret);
1865 static void nvme_tcp_free_admin_queue(struct nvme_ctrl *ctrl)
1867 if (to_tcp_ctrl(ctrl)->async_req.pdu) {
1868 cancel_work_sync(&ctrl->async_event_work);
1869 nvme_tcp_free_async_req(to_tcp_ctrl(ctrl));
1870 to_tcp_ctrl(ctrl)->async_req.pdu = NULL;
1873 nvme_tcp_free_queue(ctrl, 0);
1876 static void nvme_tcp_free_io_queues(struct nvme_ctrl *ctrl)
1880 for (i = 1; i < ctrl->queue_count; i++)
1881 nvme_tcp_free_queue(ctrl, i);
1884 static void nvme_tcp_stop_io_queues(struct nvme_ctrl *ctrl)
1888 for (i = 1; i < ctrl->queue_count; i++)
1889 nvme_tcp_stop_queue(ctrl, i);
1892 static int nvme_tcp_start_io_queues(struct nvme_ctrl *ctrl,
1893 int first, int last)
1897 for (i = first; i < last; i++) {
1898 ret = nvme_tcp_start_queue(ctrl, i);
1900 goto out_stop_queues;
1906 for (i--; i >= first; i--)
1907 nvme_tcp_stop_queue(ctrl, i);
1911 static int nvme_tcp_alloc_admin_queue(struct nvme_ctrl *ctrl)
1914 key_serial_t pskid = 0;
1916 if (nvme_tcp_tls(ctrl)) {
1917 if (ctrl->opts->tls_key)
1918 pskid = key_serial(ctrl->opts->tls_key);
1920 pskid = nvme_tls_psk_default(ctrl->opts->keyring,
1921 ctrl->opts->host->nqn,
1922 ctrl->opts->subsysnqn);
1924 dev_err(ctrl->device, "no valid PSK found\n");
1929 ret = nvme_tcp_alloc_queue(ctrl, 0, pskid);
1933 ret = nvme_tcp_alloc_async_req(to_tcp_ctrl(ctrl));
1935 goto out_free_queue;
1940 nvme_tcp_free_queue(ctrl, 0);
1944 static int __nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
1948 if (nvme_tcp_tls(ctrl) && !ctrl->tls_key) {
1949 dev_err(ctrl->device, "no PSK negotiated\n");
1952 for (i = 1; i < ctrl->queue_count; i++) {
1953 ret = nvme_tcp_alloc_queue(ctrl, i,
1954 key_serial(ctrl->tls_key));
1956 goto out_free_queues;
1962 for (i--; i >= 1; i--)
1963 nvme_tcp_free_queue(ctrl, i);
1968 static int nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
1970 unsigned int nr_io_queues;
1973 nr_io_queues = nvmf_nr_io_queues(ctrl->opts);
1974 ret = nvme_set_queue_count(ctrl, &nr_io_queues);
1978 if (nr_io_queues == 0) {
1979 dev_err(ctrl->device,
1980 "unable to set any I/O queues\n");
1984 ctrl->queue_count = nr_io_queues + 1;
1985 dev_info(ctrl->device,
1986 "creating %d I/O queues.\n", nr_io_queues);
1988 nvmf_set_io_queues(ctrl->opts, nr_io_queues,
1989 to_tcp_ctrl(ctrl)->io_queues);
1990 return __nvme_tcp_alloc_io_queues(ctrl);
1993 static void nvme_tcp_destroy_io_queues(struct nvme_ctrl *ctrl, bool remove)
1995 nvme_tcp_stop_io_queues(ctrl);
1997 nvme_remove_io_tag_set(ctrl);
1998 nvme_tcp_free_io_queues(ctrl);
2001 static int nvme_tcp_configure_io_queues(struct nvme_ctrl *ctrl, bool new)
2005 ret = nvme_tcp_alloc_io_queues(ctrl);
2010 ret = nvme_alloc_io_tag_set(ctrl, &to_tcp_ctrl(ctrl)->tag_set,
2012 ctrl->opts->nr_poll_queues ? HCTX_MAX_TYPES : 2,
2013 sizeof(struct nvme_tcp_request));
2015 goto out_free_io_queues;
2019 * Only start IO queues for which we have allocated the tagset
2020 * and limitted it to the available queues. On reconnects, the
2021 * queue number might have changed.
2023 nr_queues = min(ctrl->tagset->nr_hw_queues + 1, ctrl->queue_count);
2024 ret = nvme_tcp_start_io_queues(ctrl, 1, nr_queues);
2026 goto out_cleanup_connect_q;
2029 nvme_start_freeze(ctrl);
2030 nvme_unquiesce_io_queues(ctrl);
2031 if (!nvme_wait_freeze_timeout(ctrl, NVME_IO_TIMEOUT)) {
2033 * If we timed out waiting for freeze we are likely to
2034 * be stuck. Fail the controller initialization just
2038 nvme_unfreeze(ctrl);
2039 goto out_wait_freeze_timed_out;
2041 blk_mq_update_nr_hw_queues(ctrl->tagset,
2042 ctrl->queue_count - 1);
2043 nvme_unfreeze(ctrl);
2047 * If the number of queues has increased (reconnect case)
2048 * start all new queues now.
2050 ret = nvme_tcp_start_io_queues(ctrl, nr_queues,
2051 ctrl->tagset->nr_hw_queues + 1);
2053 goto out_wait_freeze_timed_out;
2057 out_wait_freeze_timed_out:
2058 nvme_quiesce_io_queues(ctrl);
2059 nvme_sync_io_queues(ctrl);
2060 nvme_tcp_stop_io_queues(ctrl);
2061 out_cleanup_connect_q:
2062 nvme_cancel_tagset(ctrl);
2064 nvme_remove_io_tag_set(ctrl);
2066 nvme_tcp_free_io_queues(ctrl);
2070 static void nvme_tcp_destroy_admin_queue(struct nvme_ctrl *ctrl, bool remove)
2072 nvme_tcp_stop_queue(ctrl, 0);
2074 nvme_remove_admin_tag_set(ctrl);
2075 nvme_tcp_free_admin_queue(ctrl);
2078 static int nvme_tcp_configure_admin_queue(struct nvme_ctrl *ctrl, bool new)
2082 error = nvme_tcp_alloc_admin_queue(ctrl);
2087 error = nvme_alloc_admin_tag_set(ctrl,
2088 &to_tcp_ctrl(ctrl)->admin_tag_set,
2089 &nvme_tcp_admin_mq_ops,
2090 sizeof(struct nvme_tcp_request));
2092 goto out_free_queue;
2095 error = nvme_tcp_start_queue(ctrl, 0);
2097 goto out_cleanup_tagset;
2099 error = nvme_enable_ctrl(ctrl);
2101 goto out_stop_queue;
2103 nvme_unquiesce_admin_queue(ctrl);
2105 error = nvme_init_ctrl_finish(ctrl, false);
2107 goto out_quiesce_queue;
2112 nvme_quiesce_admin_queue(ctrl);
2113 blk_sync_queue(ctrl->admin_q);
2115 nvme_tcp_stop_queue(ctrl, 0);
2116 nvme_cancel_admin_tagset(ctrl);
2119 nvme_remove_admin_tag_set(ctrl);
2121 nvme_tcp_free_admin_queue(ctrl);
2125 static void nvme_tcp_teardown_admin_queue(struct nvme_ctrl *ctrl,
2128 nvme_quiesce_admin_queue(ctrl);
2129 blk_sync_queue(ctrl->admin_q);
2130 nvme_tcp_stop_queue(ctrl, 0);
2131 nvme_cancel_admin_tagset(ctrl);
2133 nvme_unquiesce_admin_queue(ctrl);
2134 nvme_tcp_destroy_admin_queue(ctrl, remove);
2137 static void nvme_tcp_teardown_io_queues(struct nvme_ctrl *ctrl,
2140 if (ctrl->queue_count <= 1)
2142 nvme_quiesce_admin_queue(ctrl);
2143 nvme_quiesce_io_queues(ctrl);
2144 nvme_sync_io_queues(ctrl);
2145 nvme_tcp_stop_io_queues(ctrl);
2146 nvme_cancel_tagset(ctrl);
2148 nvme_unquiesce_io_queues(ctrl);
2149 nvme_tcp_destroy_io_queues(ctrl, remove);
2152 static void nvme_tcp_reconnect_or_remove(struct nvme_ctrl *ctrl)
2154 /* If we are resetting/deleting then do nothing */
2155 if (ctrl->state != NVME_CTRL_CONNECTING) {
2156 WARN_ON_ONCE(ctrl->state == NVME_CTRL_NEW ||
2157 ctrl->state == NVME_CTRL_LIVE);
2161 if (nvmf_should_reconnect(ctrl)) {
2162 dev_info(ctrl->device, "Reconnecting in %d seconds...\n",
2163 ctrl->opts->reconnect_delay);
2164 queue_delayed_work(nvme_wq, &to_tcp_ctrl(ctrl)->connect_work,
2165 ctrl->opts->reconnect_delay * HZ);
2167 dev_info(ctrl->device, "Removing controller...\n");
2168 nvme_delete_ctrl(ctrl);
2172 static int nvme_tcp_setup_ctrl(struct nvme_ctrl *ctrl, bool new)
2174 struct nvmf_ctrl_options *opts = ctrl->opts;
2177 ret = nvme_tcp_configure_admin_queue(ctrl, new);
2183 dev_err(ctrl->device, "icdoff is not supported!\n");
2187 if (!nvme_ctrl_sgl_supported(ctrl)) {
2189 dev_err(ctrl->device, "Mandatory sgls are not supported!\n");
2193 if (opts->queue_size > ctrl->sqsize + 1)
2194 dev_warn(ctrl->device,
2195 "queue_size %zu > ctrl sqsize %u, clamping down\n",
2196 opts->queue_size, ctrl->sqsize + 1);
2198 if (ctrl->sqsize + 1 > ctrl->maxcmd) {
2199 dev_warn(ctrl->device,
2200 "sqsize %u > ctrl maxcmd %u, clamping down\n",
2201 ctrl->sqsize + 1, ctrl->maxcmd);
2202 ctrl->sqsize = ctrl->maxcmd - 1;
2205 if (ctrl->queue_count > 1) {
2206 ret = nvme_tcp_configure_io_queues(ctrl, new);
2211 if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_LIVE)) {
2213 * state change failure is ok if we started ctrl delete,
2214 * unless we're during creation of a new controller to
2215 * avoid races with teardown flow.
2217 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING &&
2218 ctrl->state != NVME_CTRL_DELETING_NOIO);
2224 nvme_start_ctrl(ctrl);
2228 if (ctrl->queue_count > 1) {
2229 nvme_quiesce_io_queues(ctrl);
2230 nvme_sync_io_queues(ctrl);
2231 nvme_tcp_stop_io_queues(ctrl);
2232 nvme_cancel_tagset(ctrl);
2233 nvme_tcp_destroy_io_queues(ctrl, new);
2236 nvme_stop_keep_alive(ctrl);
2237 nvme_tcp_teardown_admin_queue(ctrl, false);
2241 static void nvme_tcp_reconnect_ctrl_work(struct work_struct *work)
2243 struct nvme_tcp_ctrl *tcp_ctrl = container_of(to_delayed_work(work),
2244 struct nvme_tcp_ctrl, connect_work);
2245 struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
2247 ++ctrl->nr_reconnects;
2249 if (nvme_tcp_setup_ctrl(ctrl, false))
2252 dev_info(ctrl->device, "Successfully reconnected (%d attempt)\n",
2253 ctrl->nr_reconnects);
2255 ctrl->nr_reconnects = 0;
2260 dev_info(ctrl->device, "Failed reconnect attempt %d\n",
2261 ctrl->nr_reconnects);
2262 nvme_tcp_reconnect_or_remove(ctrl);
2265 static void nvme_tcp_error_recovery_work(struct work_struct *work)
2267 struct nvme_tcp_ctrl *tcp_ctrl = container_of(work,
2268 struct nvme_tcp_ctrl, err_work);
2269 struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
2271 nvme_stop_keep_alive(ctrl);
2272 flush_work(&ctrl->async_event_work);
2273 nvme_tcp_teardown_io_queues(ctrl, false);
2274 /* unquiesce to fail fast pending requests */
2275 nvme_unquiesce_io_queues(ctrl);
2276 nvme_tcp_teardown_admin_queue(ctrl, false);
2277 nvme_unquiesce_admin_queue(ctrl);
2278 nvme_auth_stop(ctrl);
2280 if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
2281 /* state change failure is ok if we started ctrl delete */
2282 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING &&
2283 ctrl->state != NVME_CTRL_DELETING_NOIO);
2287 nvme_tcp_reconnect_or_remove(ctrl);
2290 static void nvme_tcp_teardown_ctrl(struct nvme_ctrl *ctrl, bool shutdown)
2292 nvme_tcp_teardown_io_queues(ctrl, shutdown);
2293 nvme_quiesce_admin_queue(ctrl);
2294 nvme_disable_ctrl(ctrl, shutdown);
2295 nvme_tcp_teardown_admin_queue(ctrl, shutdown);
2298 static void nvme_tcp_delete_ctrl(struct nvme_ctrl *ctrl)
2300 nvme_tcp_teardown_ctrl(ctrl, true);
2303 static void nvme_reset_ctrl_work(struct work_struct *work)
2305 struct nvme_ctrl *ctrl =
2306 container_of(work, struct nvme_ctrl, reset_work);
2308 nvme_stop_ctrl(ctrl);
2309 nvme_tcp_teardown_ctrl(ctrl, false);
2311 if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
2312 /* state change failure is ok if we started ctrl delete */
2313 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING &&
2314 ctrl->state != NVME_CTRL_DELETING_NOIO);
2318 if (nvme_tcp_setup_ctrl(ctrl, false))
2324 ++ctrl->nr_reconnects;
2325 nvme_tcp_reconnect_or_remove(ctrl);
2328 static void nvme_tcp_stop_ctrl(struct nvme_ctrl *ctrl)
2330 flush_work(&to_tcp_ctrl(ctrl)->err_work);
2331 cancel_delayed_work_sync(&to_tcp_ctrl(ctrl)->connect_work);
2334 static void nvme_tcp_free_ctrl(struct nvme_ctrl *nctrl)
2336 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
2338 if (list_empty(&ctrl->list))
2341 mutex_lock(&nvme_tcp_ctrl_mutex);
2342 list_del(&ctrl->list);
2343 mutex_unlock(&nvme_tcp_ctrl_mutex);
2345 nvmf_free_options(nctrl->opts);
2347 kfree(ctrl->queues);
2351 static void nvme_tcp_set_sg_null(struct nvme_command *c)
2353 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2357 sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
2358 NVME_SGL_FMT_TRANSPORT_A;
2361 static void nvme_tcp_set_sg_inline(struct nvme_tcp_queue *queue,
2362 struct nvme_command *c, u32 data_len)
2364 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2366 sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
2367 sg->length = cpu_to_le32(data_len);
2368 sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
2371 static void nvme_tcp_set_sg_host_data(struct nvme_command *c,
2374 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2377 sg->length = cpu_to_le32(data_len);
2378 sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
2379 NVME_SGL_FMT_TRANSPORT_A;
2382 static void nvme_tcp_submit_async_event(struct nvme_ctrl *arg)
2384 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(arg);
2385 struct nvme_tcp_queue *queue = &ctrl->queues[0];
2386 struct nvme_tcp_cmd_pdu *pdu = ctrl->async_req.pdu;
2387 struct nvme_command *cmd = &pdu->cmd;
2388 u8 hdgst = nvme_tcp_hdgst_len(queue);
2390 memset(pdu, 0, sizeof(*pdu));
2391 pdu->hdr.type = nvme_tcp_cmd;
2392 if (queue->hdr_digest)
2393 pdu->hdr.flags |= NVME_TCP_F_HDGST;
2394 pdu->hdr.hlen = sizeof(*pdu);
2395 pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst);
2397 cmd->common.opcode = nvme_admin_async_event;
2398 cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
2399 cmd->common.flags |= NVME_CMD_SGL_METABUF;
2400 nvme_tcp_set_sg_null(cmd);
2402 ctrl->async_req.state = NVME_TCP_SEND_CMD_PDU;
2403 ctrl->async_req.offset = 0;
2404 ctrl->async_req.curr_bio = NULL;
2405 ctrl->async_req.data_len = 0;
2407 nvme_tcp_queue_request(&ctrl->async_req, true, true);
2410 static void nvme_tcp_complete_timed_out(struct request *rq)
2412 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2413 struct nvme_ctrl *ctrl = &req->queue->ctrl->ctrl;
2415 nvme_tcp_stop_queue(ctrl, nvme_tcp_queue_id(req->queue));
2416 nvmf_complete_timed_out_request(rq);
2419 static enum blk_eh_timer_return nvme_tcp_timeout(struct request *rq)
2421 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2422 struct nvme_ctrl *ctrl = &req->queue->ctrl->ctrl;
2423 struct nvme_tcp_cmd_pdu *pdu = nvme_tcp_req_cmd_pdu(req);
2424 u8 opc = pdu->cmd.common.opcode, fctype = pdu->cmd.fabrics.fctype;
2425 int qid = nvme_tcp_queue_id(req->queue);
2427 dev_warn(ctrl->device,
2428 "I/O tag %d (%04x) type %d opcode %#x (%s) QID %d timeout\n",
2429 rq->tag, nvme_cid(rq), pdu->hdr.type, opc,
2430 nvme_opcode_str(qid, opc, fctype), qid);
2432 if (ctrl->state != NVME_CTRL_LIVE) {
2434 * If we are resetting, connecting or deleting we should
2435 * complete immediately because we may block controller
2436 * teardown or setup sequence
2437 * - ctrl disable/shutdown fabrics requests
2438 * - connect requests
2439 * - initialization admin requests
2440 * - I/O requests that entered after unquiescing and
2441 * the controller stopped responding
2443 * All other requests should be cancelled by the error
2444 * recovery work, so it's fine that we fail it here.
2446 nvme_tcp_complete_timed_out(rq);
2451 * LIVE state should trigger the normal error recovery which will
2452 * handle completing this request.
2454 nvme_tcp_error_recovery(ctrl);
2455 return BLK_EH_RESET_TIMER;
2458 static blk_status_t nvme_tcp_map_data(struct nvme_tcp_queue *queue,
2461 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2462 struct nvme_tcp_cmd_pdu *pdu = nvme_tcp_req_cmd_pdu(req);
2463 struct nvme_command *c = &pdu->cmd;
2465 c->common.flags |= NVME_CMD_SGL_METABUF;
2467 if (!blk_rq_nr_phys_segments(rq))
2468 nvme_tcp_set_sg_null(c);
2469 else if (rq_data_dir(rq) == WRITE &&
2470 req->data_len <= nvme_tcp_inline_data_size(req))
2471 nvme_tcp_set_sg_inline(queue, c, req->data_len);
2473 nvme_tcp_set_sg_host_data(c, req->data_len);
2478 static blk_status_t nvme_tcp_setup_cmd_pdu(struct nvme_ns *ns,
2481 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2482 struct nvme_tcp_cmd_pdu *pdu = nvme_tcp_req_cmd_pdu(req);
2483 struct nvme_tcp_queue *queue = req->queue;
2484 u8 hdgst = nvme_tcp_hdgst_len(queue), ddgst = 0;
2487 ret = nvme_setup_cmd(ns, rq);
2491 req->state = NVME_TCP_SEND_CMD_PDU;
2492 req->status = cpu_to_le16(NVME_SC_SUCCESS);
2497 req->h2cdata_left = 0;
2498 req->data_len = blk_rq_nr_phys_segments(rq) ?
2499 blk_rq_payload_bytes(rq) : 0;
2500 req->curr_bio = rq->bio;
2501 if (req->curr_bio && req->data_len)
2502 nvme_tcp_init_iter(req, rq_data_dir(rq));
2504 if (rq_data_dir(rq) == WRITE &&
2505 req->data_len <= nvme_tcp_inline_data_size(req))
2506 req->pdu_len = req->data_len;
2508 pdu->hdr.type = nvme_tcp_cmd;
2510 if (queue->hdr_digest)
2511 pdu->hdr.flags |= NVME_TCP_F_HDGST;
2512 if (queue->data_digest && req->pdu_len) {
2513 pdu->hdr.flags |= NVME_TCP_F_DDGST;
2514 ddgst = nvme_tcp_ddgst_len(queue);
2516 pdu->hdr.hlen = sizeof(*pdu);
2517 pdu->hdr.pdo = req->pdu_len ? pdu->hdr.hlen + hdgst : 0;
2519 cpu_to_le32(pdu->hdr.hlen + hdgst + req->pdu_len + ddgst);
2521 ret = nvme_tcp_map_data(queue, rq);
2522 if (unlikely(ret)) {
2523 nvme_cleanup_cmd(rq);
2524 dev_err(queue->ctrl->ctrl.device,
2525 "Failed to map data (%d)\n", ret);
2532 static void nvme_tcp_commit_rqs(struct blk_mq_hw_ctx *hctx)
2534 struct nvme_tcp_queue *queue = hctx->driver_data;
2536 if (!llist_empty(&queue->req_list))
2537 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
2540 static blk_status_t nvme_tcp_queue_rq(struct blk_mq_hw_ctx *hctx,
2541 const struct blk_mq_queue_data *bd)
2543 struct nvme_ns *ns = hctx->queue->queuedata;
2544 struct nvme_tcp_queue *queue = hctx->driver_data;
2545 struct request *rq = bd->rq;
2546 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2547 bool queue_ready = test_bit(NVME_TCP_Q_LIVE, &queue->flags);
2550 if (!nvme_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
2551 return nvme_fail_nonready_command(&queue->ctrl->ctrl, rq);
2553 ret = nvme_tcp_setup_cmd_pdu(ns, rq);
2557 nvme_start_request(rq);
2559 nvme_tcp_queue_request(req, true, bd->last);
2564 static void nvme_tcp_map_queues(struct blk_mq_tag_set *set)
2566 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(set->driver_data);
2568 nvmf_map_queues(set, &ctrl->ctrl, ctrl->io_queues);
2571 static int nvme_tcp_poll(struct blk_mq_hw_ctx *hctx, struct io_comp_batch *iob)
2573 struct nvme_tcp_queue *queue = hctx->driver_data;
2574 struct sock *sk = queue->sock->sk;
2576 if (!test_bit(NVME_TCP_Q_LIVE, &queue->flags))
2579 set_bit(NVME_TCP_Q_POLLING, &queue->flags);
2580 if (sk_can_busy_loop(sk) && skb_queue_empty_lockless(&sk->sk_receive_queue))
2581 sk_busy_loop(sk, true);
2582 nvme_tcp_try_recv(queue);
2583 clear_bit(NVME_TCP_Q_POLLING, &queue->flags);
2584 return queue->nr_cqe;
2587 static int nvme_tcp_get_address(struct nvme_ctrl *ctrl, char *buf, int size)
2589 struct nvme_tcp_queue *queue = &to_tcp_ctrl(ctrl)->queues[0];
2590 struct sockaddr_storage src_addr;
2593 len = nvmf_get_address(ctrl, buf, size);
2595 mutex_lock(&queue->queue_lock);
2597 if (!test_bit(NVME_TCP_Q_LIVE, &queue->flags))
2599 ret = kernel_getsockname(queue->sock, (struct sockaddr *)&src_addr);
2602 len--; /* strip trailing newline */
2603 len += scnprintf(buf + len, size - len, "%ssrc_addr=%pISc\n",
2604 (len) ? "," : "", &src_addr);
2607 mutex_unlock(&queue->queue_lock);
2612 static const struct blk_mq_ops nvme_tcp_mq_ops = {
2613 .queue_rq = nvme_tcp_queue_rq,
2614 .commit_rqs = nvme_tcp_commit_rqs,
2615 .complete = nvme_complete_rq,
2616 .init_request = nvme_tcp_init_request,
2617 .exit_request = nvme_tcp_exit_request,
2618 .init_hctx = nvme_tcp_init_hctx,
2619 .timeout = nvme_tcp_timeout,
2620 .map_queues = nvme_tcp_map_queues,
2621 .poll = nvme_tcp_poll,
2624 static const struct blk_mq_ops nvme_tcp_admin_mq_ops = {
2625 .queue_rq = nvme_tcp_queue_rq,
2626 .complete = nvme_complete_rq,
2627 .init_request = nvme_tcp_init_request,
2628 .exit_request = nvme_tcp_exit_request,
2629 .init_hctx = nvme_tcp_init_admin_hctx,
2630 .timeout = nvme_tcp_timeout,
2633 static const struct nvme_ctrl_ops nvme_tcp_ctrl_ops = {
2635 .module = THIS_MODULE,
2636 .flags = NVME_F_FABRICS | NVME_F_BLOCKING,
2637 .reg_read32 = nvmf_reg_read32,
2638 .reg_read64 = nvmf_reg_read64,
2639 .reg_write32 = nvmf_reg_write32,
2640 .free_ctrl = nvme_tcp_free_ctrl,
2641 .submit_async_event = nvme_tcp_submit_async_event,
2642 .delete_ctrl = nvme_tcp_delete_ctrl,
2643 .get_address = nvme_tcp_get_address,
2644 .stop_ctrl = nvme_tcp_stop_ctrl,
2648 nvme_tcp_existing_controller(struct nvmf_ctrl_options *opts)
2650 struct nvme_tcp_ctrl *ctrl;
2653 mutex_lock(&nvme_tcp_ctrl_mutex);
2654 list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list) {
2655 found = nvmf_ip_options_match(&ctrl->ctrl, opts);
2659 mutex_unlock(&nvme_tcp_ctrl_mutex);
2664 static struct nvme_ctrl *nvme_tcp_create_ctrl(struct device *dev,
2665 struct nvmf_ctrl_options *opts)
2667 struct nvme_tcp_ctrl *ctrl;
2670 ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
2672 return ERR_PTR(-ENOMEM);
2674 INIT_LIST_HEAD(&ctrl->list);
2675 ctrl->ctrl.opts = opts;
2676 ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues +
2677 opts->nr_poll_queues + 1;
2678 ctrl->ctrl.sqsize = opts->queue_size - 1;
2679 ctrl->ctrl.kato = opts->kato;
2681 INIT_DELAYED_WORK(&ctrl->connect_work,
2682 nvme_tcp_reconnect_ctrl_work);
2683 INIT_WORK(&ctrl->err_work, nvme_tcp_error_recovery_work);
2684 INIT_WORK(&ctrl->ctrl.reset_work, nvme_reset_ctrl_work);
2686 if (!(opts->mask & NVMF_OPT_TRSVCID)) {
2688 kstrdup(__stringify(NVME_TCP_DISC_PORT), GFP_KERNEL);
2689 if (!opts->trsvcid) {
2693 opts->mask |= NVMF_OPT_TRSVCID;
2696 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2697 opts->traddr, opts->trsvcid, &ctrl->addr);
2699 pr_err("malformed address passed: %s:%s\n",
2700 opts->traddr, opts->trsvcid);
2704 if (opts->mask & NVMF_OPT_HOST_TRADDR) {
2705 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2706 opts->host_traddr, NULL, &ctrl->src_addr);
2708 pr_err("malformed src address passed: %s\n",
2714 if (opts->mask & NVMF_OPT_HOST_IFACE) {
2715 if (!__dev_get_by_name(&init_net, opts->host_iface)) {
2716 pr_err("invalid interface passed: %s\n",
2723 if (!opts->duplicate_connect && nvme_tcp_existing_controller(opts)) {
2728 ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
2730 if (!ctrl->queues) {
2735 ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_tcp_ctrl_ops, 0);
2737 goto out_kfree_queues;
2739 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
2742 goto out_uninit_ctrl;
2745 ret = nvme_tcp_setup_ctrl(&ctrl->ctrl, true);
2747 goto out_uninit_ctrl;
2749 dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISp\n",
2750 nvmf_ctrl_subsysnqn(&ctrl->ctrl), &ctrl->addr);
2752 mutex_lock(&nvme_tcp_ctrl_mutex);
2753 list_add_tail(&ctrl->list, &nvme_tcp_ctrl_list);
2754 mutex_unlock(&nvme_tcp_ctrl_mutex);
2759 nvme_uninit_ctrl(&ctrl->ctrl);
2760 nvme_put_ctrl(&ctrl->ctrl);
2763 return ERR_PTR(ret);
2765 kfree(ctrl->queues);
2768 return ERR_PTR(ret);
2771 static struct nvmf_transport_ops nvme_tcp_transport = {
2773 .module = THIS_MODULE,
2774 .required_opts = NVMF_OPT_TRADDR,
2775 .allowed_opts = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2776 NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
2777 NVMF_OPT_HDR_DIGEST | NVMF_OPT_DATA_DIGEST |
2778 NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES |
2779 NVMF_OPT_TOS | NVMF_OPT_HOST_IFACE | NVMF_OPT_TLS |
2780 NVMF_OPT_KEYRING | NVMF_OPT_TLS_KEY,
2781 .create_ctrl = nvme_tcp_create_ctrl,
2784 static int __init nvme_tcp_init_module(void)
2786 BUILD_BUG_ON(sizeof(struct nvme_tcp_hdr) != 8);
2787 BUILD_BUG_ON(sizeof(struct nvme_tcp_cmd_pdu) != 72);
2788 BUILD_BUG_ON(sizeof(struct nvme_tcp_data_pdu) != 24);
2789 BUILD_BUG_ON(sizeof(struct nvme_tcp_rsp_pdu) != 24);
2790 BUILD_BUG_ON(sizeof(struct nvme_tcp_r2t_pdu) != 24);
2791 BUILD_BUG_ON(sizeof(struct nvme_tcp_icreq_pdu) != 128);
2792 BUILD_BUG_ON(sizeof(struct nvme_tcp_icresp_pdu) != 128);
2793 BUILD_BUG_ON(sizeof(struct nvme_tcp_term_pdu) != 24);
2795 nvme_tcp_wq = alloc_workqueue("nvme_tcp_wq",
2796 WQ_MEM_RECLAIM | WQ_HIGHPRI, 0);
2800 nvmf_register_transport(&nvme_tcp_transport);
2804 static void __exit nvme_tcp_cleanup_module(void)
2806 struct nvme_tcp_ctrl *ctrl;
2808 nvmf_unregister_transport(&nvme_tcp_transport);
2810 mutex_lock(&nvme_tcp_ctrl_mutex);
2811 list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list)
2812 nvme_delete_ctrl(&ctrl->ctrl);
2813 mutex_unlock(&nvme_tcp_ctrl_mutex);
2814 flush_workqueue(nvme_delete_wq);
2816 destroy_workqueue(nvme_tcp_wq);
2819 module_init(nvme_tcp_init_module);
2820 module_exit(nvme_tcp_cleanup_module);
2822 MODULE_LICENSE("GPL v2");