1 // SPDX-License-Identifier: GPL-2.0
3 * NVMe over Fabrics TCP host.
4 * Copyright (c) 2018 Lightbits Labs. All rights reserved.
6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
7 #include <linux/module.h>
8 #include <linux/init.h>
9 #include <linux/slab.h>
10 #include <linux/err.h>
11 #include <linux/nvme-tcp.h>
14 #include <linux/blk-mq.h>
15 #include <crypto/hash.h>
16 #include <net/busy_poll.h>
21 struct nvme_tcp_queue;
23 /* Define the socket priority to use for connections were it is desirable
24 * that the NIC consider performing optimized packet processing or filtering.
25 * A non-zero value being sufficient to indicate general consideration of any
26 * possible optimization. Making it a module param allows for alternative
27 * values that may be unique for some NIC implementations.
29 static int so_priority;
30 module_param(so_priority, int, 0644);
31 MODULE_PARM_DESC(so_priority, "nvme tcp socket optimize priority");
33 #ifdef CONFIG_DEBUG_LOCK_ALLOC
34 /* lockdep can detect a circular dependency of the form
35 * sk_lock -> mmap_lock (page fault) -> fs locks -> sk_lock
36 * because dependencies are tracked for both nvme-tcp and user contexts. Using
37 * a separate class prevents lockdep from conflating nvme-tcp socket use with
38 * user-space socket API use.
40 static struct lock_class_key nvme_tcp_sk_key[2];
41 static struct lock_class_key nvme_tcp_slock_key[2];
43 static void nvme_tcp_reclassify_socket(struct socket *sock)
45 struct sock *sk = sock->sk;
47 if (WARN_ON_ONCE(!sock_allow_reclassification(sk)))
50 switch (sk->sk_family) {
52 sock_lock_init_class_and_name(sk, "slock-AF_INET-NVME",
53 &nvme_tcp_slock_key[0],
54 "sk_lock-AF_INET-NVME",
58 sock_lock_init_class_and_name(sk, "slock-AF_INET6-NVME",
59 &nvme_tcp_slock_key[1],
60 "sk_lock-AF_INET6-NVME",
68 static void nvme_tcp_reclassify_socket(struct socket *sock) { }
71 enum nvme_tcp_send_state {
72 NVME_TCP_SEND_CMD_PDU = 0,
73 NVME_TCP_SEND_H2C_PDU,
78 struct nvme_tcp_request {
79 struct nvme_request req;
81 struct nvme_tcp_queue *queue;
89 struct list_head entry;
90 struct llist_node lentry;
99 enum nvme_tcp_send_state state;
102 enum nvme_tcp_queue_flags {
103 NVME_TCP_Q_ALLOCATED = 0,
105 NVME_TCP_Q_POLLING = 2,
108 enum nvme_tcp_recv_state {
109 NVME_TCP_RECV_PDU = 0,
114 struct nvme_tcp_ctrl;
115 struct nvme_tcp_queue {
117 struct work_struct io_work;
120 struct mutex queue_lock;
121 struct mutex send_mutex;
122 struct llist_head req_list;
123 struct list_head send_list;
130 size_t data_remaining;
131 size_t ddgst_remaining;
135 struct nvme_tcp_request *request;
139 size_t cmnd_capsule_len;
140 struct nvme_tcp_ctrl *ctrl;
146 struct ahash_request *rcv_hash;
147 struct ahash_request *snd_hash;
151 struct page_frag_cache pf_cache;
153 void (*state_change)(struct sock *);
154 void (*data_ready)(struct sock *);
155 void (*write_space)(struct sock *);
158 struct nvme_tcp_ctrl {
159 /* read only in the hot path */
160 struct nvme_tcp_queue *queues;
161 struct blk_mq_tag_set tag_set;
163 /* other member variables */
164 struct list_head list;
165 struct blk_mq_tag_set admin_tag_set;
166 struct sockaddr_storage addr;
167 struct sockaddr_storage src_addr;
168 struct nvme_ctrl ctrl;
170 struct work_struct err_work;
171 struct delayed_work connect_work;
172 struct nvme_tcp_request async_req;
173 u32 io_queues[HCTX_MAX_TYPES];
176 static LIST_HEAD(nvme_tcp_ctrl_list);
177 static DEFINE_MUTEX(nvme_tcp_ctrl_mutex);
178 static struct workqueue_struct *nvme_tcp_wq;
179 static const struct blk_mq_ops nvme_tcp_mq_ops;
180 static const struct blk_mq_ops nvme_tcp_admin_mq_ops;
181 static int nvme_tcp_try_send(struct nvme_tcp_queue *queue);
183 static inline struct nvme_tcp_ctrl *to_tcp_ctrl(struct nvme_ctrl *ctrl)
185 return container_of(ctrl, struct nvme_tcp_ctrl, ctrl);
188 static inline int nvme_tcp_queue_id(struct nvme_tcp_queue *queue)
190 return queue - queue->ctrl->queues;
193 static inline struct blk_mq_tags *nvme_tcp_tagset(struct nvme_tcp_queue *queue)
195 u32 queue_idx = nvme_tcp_queue_id(queue);
198 return queue->ctrl->admin_tag_set.tags[queue_idx];
199 return queue->ctrl->tag_set.tags[queue_idx - 1];
202 static inline u8 nvme_tcp_hdgst_len(struct nvme_tcp_queue *queue)
204 return queue->hdr_digest ? NVME_TCP_DIGEST_LENGTH : 0;
207 static inline u8 nvme_tcp_ddgst_len(struct nvme_tcp_queue *queue)
209 return queue->data_digest ? NVME_TCP_DIGEST_LENGTH : 0;
212 static inline size_t nvme_tcp_inline_data_size(struct nvme_tcp_queue *queue)
214 return queue->cmnd_capsule_len - sizeof(struct nvme_command);
217 static inline bool nvme_tcp_async_req(struct nvme_tcp_request *req)
219 return req == &req->queue->ctrl->async_req;
222 static inline bool nvme_tcp_has_inline_data(struct nvme_tcp_request *req)
226 if (unlikely(nvme_tcp_async_req(req)))
227 return false; /* async events don't have a request */
229 rq = blk_mq_rq_from_pdu(req);
231 return rq_data_dir(rq) == WRITE && req->data_len &&
232 req->data_len <= nvme_tcp_inline_data_size(req->queue);
235 static inline struct page *nvme_tcp_req_cur_page(struct nvme_tcp_request *req)
237 return req->iter.bvec->bv_page;
240 static inline size_t nvme_tcp_req_cur_offset(struct nvme_tcp_request *req)
242 return req->iter.bvec->bv_offset + req->iter.iov_offset;
245 static inline size_t nvme_tcp_req_cur_length(struct nvme_tcp_request *req)
247 return min_t(size_t, iov_iter_single_seg_count(&req->iter),
248 req->pdu_len - req->pdu_sent);
251 static inline size_t nvme_tcp_pdu_data_left(struct nvme_tcp_request *req)
253 return rq_data_dir(blk_mq_rq_from_pdu(req)) == WRITE ?
254 req->pdu_len - req->pdu_sent : 0;
257 static inline size_t nvme_tcp_pdu_last_send(struct nvme_tcp_request *req,
260 return nvme_tcp_pdu_data_left(req) <= len;
263 static void nvme_tcp_init_iter(struct nvme_tcp_request *req,
266 struct request *rq = blk_mq_rq_from_pdu(req);
272 if (rq->rq_flags & RQF_SPECIAL_PAYLOAD) {
273 vec = &rq->special_vec;
275 size = blk_rq_payload_bytes(rq);
278 struct bio *bio = req->curr_bio;
282 vec = __bvec_iter_bvec(bio->bi_io_vec, bio->bi_iter);
284 bio_for_each_bvec(bv, bio, bi) {
287 size = bio->bi_iter.bi_size;
288 offset = bio->bi_iter.bi_bvec_done;
291 iov_iter_bvec(&req->iter, dir, vec, nr_bvec, size);
292 req->iter.iov_offset = offset;
295 static inline void nvme_tcp_advance_req(struct nvme_tcp_request *req,
298 req->data_sent += len;
299 req->pdu_sent += len;
300 iov_iter_advance(&req->iter, len);
301 if (!iov_iter_count(&req->iter) &&
302 req->data_sent < req->data_len) {
303 req->curr_bio = req->curr_bio->bi_next;
304 nvme_tcp_init_iter(req, WRITE);
308 static inline void nvme_tcp_send_all(struct nvme_tcp_queue *queue)
312 /* drain the send queue as much as we can... */
314 ret = nvme_tcp_try_send(queue);
318 static inline bool nvme_tcp_queue_more(struct nvme_tcp_queue *queue)
320 return !list_empty(&queue->send_list) ||
321 !llist_empty(&queue->req_list) || queue->more_requests;
324 static inline void nvme_tcp_queue_request(struct nvme_tcp_request *req,
325 bool sync, bool last)
327 struct nvme_tcp_queue *queue = req->queue;
330 empty = llist_add(&req->lentry, &queue->req_list) &&
331 list_empty(&queue->send_list) && !queue->request;
334 * if we're the first on the send_list and we can try to send
335 * directly, otherwise queue io_work. Also, only do that if we
336 * are on the same cpu, so we don't introduce contention.
338 if (queue->io_cpu == raw_smp_processor_id() &&
339 sync && empty && mutex_trylock(&queue->send_mutex)) {
340 queue->more_requests = !last;
341 nvme_tcp_send_all(queue);
342 queue->more_requests = false;
343 mutex_unlock(&queue->send_mutex);
346 if (last && nvme_tcp_queue_more(queue))
347 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
350 static void nvme_tcp_process_req_list(struct nvme_tcp_queue *queue)
352 struct nvme_tcp_request *req;
353 struct llist_node *node;
355 for (node = llist_del_all(&queue->req_list); node; node = node->next) {
356 req = llist_entry(node, struct nvme_tcp_request, lentry);
357 list_add(&req->entry, &queue->send_list);
361 static inline struct nvme_tcp_request *
362 nvme_tcp_fetch_request(struct nvme_tcp_queue *queue)
364 struct nvme_tcp_request *req;
366 req = list_first_entry_or_null(&queue->send_list,
367 struct nvme_tcp_request, entry);
369 nvme_tcp_process_req_list(queue);
370 req = list_first_entry_or_null(&queue->send_list,
371 struct nvme_tcp_request, entry);
376 list_del(&req->entry);
380 static inline void nvme_tcp_ddgst_final(struct ahash_request *hash,
383 ahash_request_set_crypt(hash, NULL, (u8 *)dgst, 0);
384 crypto_ahash_final(hash);
387 static inline void nvme_tcp_ddgst_update(struct ahash_request *hash,
388 struct page *page, off_t off, size_t len)
390 struct scatterlist sg;
392 sg_init_marker(&sg, 1);
393 sg_set_page(&sg, page, len, off);
394 ahash_request_set_crypt(hash, &sg, NULL, len);
395 crypto_ahash_update(hash);
398 static inline void nvme_tcp_hdgst(struct ahash_request *hash,
399 void *pdu, size_t len)
401 struct scatterlist sg;
403 sg_init_one(&sg, pdu, len);
404 ahash_request_set_crypt(hash, &sg, pdu + len, len);
405 crypto_ahash_digest(hash);
408 static int nvme_tcp_verify_hdgst(struct nvme_tcp_queue *queue,
409 void *pdu, size_t pdu_len)
411 struct nvme_tcp_hdr *hdr = pdu;
415 if (unlikely(!(hdr->flags & NVME_TCP_F_HDGST))) {
416 dev_err(queue->ctrl->ctrl.device,
417 "queue %d: header digest flag is cleared\n",
418 nvme_tcp_queue_id(queue));
422 recv_digest = *(__le32 *)(pdu + hdr->hlen);
423 nvme_tcp_hdgst(queue->rcv_hash, pdu, pdu_len);
424 exp_digest = *(__le32 *)(pdu + hdr->hlen);
425 if (recv_digest != exp_digest) {
426 dev_err(queue->ctrl->ctrl.device,
427 "header digest error: recv %#x expected %#x\n",
428 le32_to_cpu(recv_digest), le32_to_cpu(exp_digest));
435 static int nvme_tcp_check_ddgst(struct nvme_tcp_queue *queue, void *pdu)
437 struct nvme_tcp_hdr *hdr = pdu;
438 u8 digest_len = nvme_tcp_hdgst_len(queue);
441 len = le32_to_cpu(hdr->plen) - hdr->hlen -
442 ((hdr->flags & NVME_TCP_F_HDGST) ? digest_len : 0);
444 if (unlikely(len && !(hdr->flags & NVME_TCP_F_DDGST))) {
445 dev_err(queue->ctrl->ctrl.device,
446 "queue %d: data digest flag is cleared\n",
447 nvme_tcp_queue_id(queue));
450 crypto_ahash_init(queue->rcv_hash);
455 static void nvme_tcp_exit_request(struct blk_mq_tag_set *set,
456 struct request *rq, unsigned int hctx_idx)
458 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
460 page_frag_free(req->pdu);
463 static int nvme_tcp_init_request(struct blk_mq_tag_set *set,
464 struct request *rq, unsigned int hctx_idx,
465 unsigned int numa_node)
467 struct nvme_tcp_ctrl *ctrl = set->driver_data;
468 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
469 struct nvme_tcp_cmd_pdu *pdu;
470 int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
471 struct nvme_tcp_queue *queue = &ctrl->queues[queue_idx];
472 u8 hdgst = nvme_tcp_hdgst_len(queue);
474 req->pdu = page_frag_alloc(&queue->pf_cache,
475 sizeof(struct nvme_tcp_cmd_pdu) + hdgst,
476 GFP_KERNEL | __GFP_ZERO);
482 nvme_req(rq)->ctrl = &ctrl->ctrl;
483 nvme_req(rq)->cmd = &pdu->cmd;
488 static int nvme_tcp_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
489 unsigned int hctx_idx)
491 struct nvme_tcp_ctrl *ctrl = data;
492 struct nvme_tcp_queue *queue = &ctrl->queues[hctx_idx + 1];
494 hctx->driver_data = queue;
498 static int nvme_tcp_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
499 unsigned int hctx_idx)
501 struct nvme_tcp_ctrl *ctrl = data;
502 struct nvme_tcp_queue *queue = &ctrl->queues[0];
504 hctx->driver_data = queue;
508 static enum nvme_tcp_recv_state
509 nvme_tcp_recv_state(struct nvme_tcp_queue *queue)
511 return (queue->pdu_remaining) ? NVME_TCP_RECV_PDU :
512 (queue->ddgst_remaining) ? NVME_TCP_RECV_DDGST :
516 static void nvme_tcp_init_recv_ctx(struct nvme_tcp_queue *queue)
518 queue->pdu_remaining = sizeof(struct nvme_tcp_rsp_pdu) +
519 nvme_tcp_hdgst_len(queue);
520 queue->pdu_offset = 0;
521 queue->data_remaining = -1;
522 queue->ddgst_remaining = 0;
525 static void nvme_tcp_error_recovery(struct nvme_ctrl *ctrl)
527 if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_RESETTING))
530 dev_warn(ctrl->device, "starting error recovery\n");
531 queue_work(nvme_reset_wq, &to_tcp_ctrl(ctrl)->err_work);
534 static int nvme_tcp_process_nvme_cqe(struct nvme_tcp_queue *queue,
535 struct nvme_completion *cqe)
537 struct nvme_tcp_request *req;
540 rq = nvme_find_rq(nvme_tcp_tagset(queue), cqe->command_id);
542 dev_err(queue->ctrl->ctrl.device,
543 "got bad cqe.command_id %#x on queue %d\n",
544 cqe->command_id, nvme_tcp_queue_id(queue));
545 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
549 req = blk_mq_rq_to_pdu(rq);
550 if (req->status == cpu_to_le16(NVME_SC_SUCCESS))
551 req->status = cqe->status;
553 if (!nvme_try_complete_req(rq, req->status, cqe->result))
554 nvme_complete_rq(rq);
560 static int nvme_tcp_handle_c2h_data(struct nvme_tcp_queue *queue,
561 struct nvme_tcp_data_pdu *pdu)
565 rq = nvme_find_rq(nvme_tcp_tagset(queue), pdu->command_id);
567 dev_err(queue->ctrl->ctrl.device,
568 "got bad c2hdata.command_id %#x on queue %d\n",
569 pdu->command_id, nvme_tcp_queue_id(queue));
573 if (!blk_rq_payload_bytes(rq)) {
574 dev_err(queue->ctrl->ctrl.device,
575 "queue %d tag %#x unexpected data\n",
576 nvme_tcp_queue_id(queue), rq->tag);
580 queue->data_remaining = le32_to_cpu(pdu->data_length);
582 if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS &&
583 unlikely(!(pdu->hdr.flags & NVME_TCP_F_DATA_LAST))) {
584 dev_err(queue->ctrl->ctrl.device,
585 "queue %d tag %#x SUCCESS set but not last PDU\n",
586 nvme_tcp_queue_id(queue), rq->tag);
587 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
594 static int nvme_tcp_handle_comp(struct nvme_tcp_queue *queue,
595 struct nvme_tcp_rsp_pdu *pdu)
597 struct nvme_completion *cqe = &pdu->cqe;
601 * AEN requests are special as they don't time out and can
602 * survive any kind of queue freeze and often don't respond to
603 * aborts. We don't even bother to allocate a struct request
604 * for them but rather special case them here.
606 if (unlikely(nvme_is_aen_req(nvme_tcp_queue_id(queue),
608 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
611 ret = nvme_tcp_process_nvme_cqe(queue, cqe);
616 static void nvme_tcp_setup_h2c_data_pdu(struct nvme_tcp_request *req)
618 struct nvme_tcp_data_pdu *data = req->pdu;
619 struct nvme_tcp_queue *queue = req->queue;
620 struct request *rq = blk_mq_rq_from_pdu(req);
621 u32 h2cdata_sent = req->pdu_len;
622 u8 hdgst = nvme_tcp_hdgst_len(queue);
623 u8 ddgst = nvme_tcp_ddgst_len(queue);
625 req->state = NVME_TCP_SEND_H2C_PDU;
627 req->pdu_len = min(req->h2cdata_left, queue->maxh2cdata);
629 req->h2cdata_left -= req->pdu_len;
630 req->h2cdata_offset += h2cdata_sent;
632 memset(data, 0, sizeof(*data));
633 data->hdr.type = nvme_tcp_h2c_data;
634 if (!req->h2cdata_left)
635 data->hdr.flags = NVME_TCP_F_DATA_LAST;
636 if (queue->hdr_digest)
637 data->hdr.flags |= NVME_TCP_F_HDGST;
638 if (queue->data_digest)
639 data->hdr.flags |= NVME_TCP_F_DDGST;
640 data->hdr.hlen = sizeof(*data);
641 data->hdr.pdo = data->hdr.hlen + hdgst;
643 cpu_to_le32(data->hdr.hlen + hdgst + req->pdu_len + ddgst);
644 data->ttag = req->ttag;
645 data->command_id = nvme_cid(rq);
646 data->data_offset = cpu_to_le32(req->h2cdata_offset);
647 data->data_length = cpu_to_le32(req->pdu_len);
650 static int nvme_tcp_handle_r2t(struct nvme_tcp_queue *queue,
651 struct nvme_tcp_r2t_pdu *pdu)
653 struct nvme_tcp_request *req;
655 u32 r2t_length = le32_to_cpu(pdu->r2t_length);
656 u32 r2t_offset = le32_to_cpu(pdu->r2t_offset);
658 rq = nvme_find_rq(nvme_tcp_tagset(queue), pdu->command_id);
660 dev_err(queue->ctrl->ctrl.device,
661 "got bad r2t.command_id %#x on queue %d\n",
662 pdu->command_id, nvme_tcp_queue_id(queue));
665 req = blk_mq_rq_to_pdu(rq);
667 if (unlikely(!r2t_length)) {
668 dev_err(queue->ctrl->ctrl.device,
669 "req %d r2t len is %u, probably a bug...\n",
670 rq->tag, r2t_length);
674 if (unlikely(req->data_sent + r2t_length > req->data_len)) {
675 dev_err(queue->ctrl->ctrl.device,
676 "req %d r2t len %u exceeded data len %u (%zu sent)\n",
677 rq->tag, r2t_length, req->data_len, req->data_sent);
681 if (unlikely(r2t_offset < req->data_sent)) {
682 dev_err(queue->ctrl->ctrl.device,
683 "req %d unexpected r2t offset %u (expected %zu)\n",
684 rq->tag, r2t_offset, req->data_sent);
689 req->h2cdata_left = r2t_length;
690 req->h2cdata_offset = r2t_offset;
691 req->ttag = pdu->ttag;
693 nvme_tcp_setup_h2c_data_pdu(req);
694 nvme_tcp_queue_request(req, false, true);
699 static int nvme_tcp_recv_pdu(struct nvme_tcp_queue *queue, struct sk_buff *skb,
700 unsigned int *offset, size_t *len)
702 struct nvme_tcp_hdr *hdr;
703 char *pdu = queue->pdu;
704 size_t rcv_len = min_t(size_t, *len, queue->pdu_remaining);
707 ret = skb_copy_bits(skb, *offset,
708 &pdu[queue->pdu_offset], rcv_len);
712 queue->pdu_remaining -= rcv_len;
713 queue->pdu_offset += rcv_len;
716 if (queue->pdu_remaining)
720 if (queue->hdr_digest) {
721 ret = nvme_tcp_verify_hdgst(queue, queue->pdu, hdr->hlen);
727 if (queue->data_digest) {
728 ret = nvme_tcp_check_ddgst(queue, queue->pdu);
734 case nvme_tcp_c2h_data:
735 return nvme_tcp_handle_c2h_data(queue, (void *)queue->pdu);
737 nvme_tcp_init_recv_ctx(queue);
738 return nvme_tcp_handle_comp(queue, (void *)queue->pdu);
740 nvme_tcp_init_recv_ctx(queue);
741 return nvme_tcp_handle_r2t(queue, (void *)queue->pdu);
743 dev_err(queue->ctrl->ctrl.device,
744 "unsupported pdu type (%d)\n", hdr->type);
749 static inline void nvme_tcp_end_request(struct request *rq, u16 status)
751 union nvme_result res = {};
753 if (!nvme_try_complete_req(rq, cpu_to_le16(status << 1), res))
754 nvme_complete_rq(rq);
757 static int nvme_tcp_recv_data(struct nvme_tcp_queue *queue, struct sk_buff *skb,
758 unsigned int *offset, size_t *len)
760 struct nvme_tcp_data_pdu *pdu = (void *)queue->pdu;
762 nvme_cid_to_rq(nvme_tcp_tagset(queue), pdu->command_id);
763 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
768 recv_len = min_t(size_t, *len, queue->data_remaining);
772 if (!iov_iter_count(&req->iter)) {
773 req->curr_bio = req->curr_bio->bi_next;
776 * If we don`t have any bios it means that controller
777 * sent more data than we requested, hence error
779 if (!req->curr_bio) {
780 dev_err(queue->ctrl->ctrl.device,
781 "queue %d no space in request %#x",
782 nvme_tcp_queue_id(queue), rq->tag);
783 nvme_tcp_init_recv_ctx(queue);
786 nvme_tcp_init_iter(req, READ);
789 /* we can read only from what is left in this bio */
790 recv_len = min_t(size_t, recv_len,
791 iov_iter_count(&req->iter));
793 if (queue->data_digest)
794 ret = skb_copy_and_hash_datagram_iter(skb, *offset,
795 &req->iter, recv_len, queue->rcv_hash);
797 ret = skb_copy_datagram_iter(skb, *offset,
798 &req->iter, recv_len);
800 dev_err(queue->ctrl->ctrl.device,
801 "queue %d failed to copy request %#x data",
802 nvme_tcp_queue_id(queue), rq->tag);
808 queue->data_remaining -= recv_len;
811 if (!queue->data_remaining) {
812 if (queue->data_digest) {
813 nvme_tcp_ddgst_final(queue->rcv_hash, &queue->exp_ddgst);
814 queue->ddgst_remaining = NVME_TCP_DIGEST_LENGTH;
816 if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
817 nvme_tcp_end_request(rq,
818 le16_to_cpu(req->status));
821 nvme_tcp_init_recv_ctx(queue);
828 static int nvme_tcp_recv_ddgst(struct nvme_tcp_queue *queue,
829 struct sk_buff *skb, unsigned int *offset, size_t *len)
831 struct nvme_tcp_data_pdu *pdu = (void *)queue->pdu;
832 char *ddgst = (char *)&queue->recv_ddgst;
833 size_t recv_len = min_t(size_t, *len, queue->ddgst_remaining);
834 off_t off = NVME_TCP_DIGEST_LENGTH - queue->ddgst_remaining;
837 ret = skb_copy_bits(skb, *offset, &ddgst[off], recv_len);
841 queue->ddgst_remaining -= recv_len;
844 if (queue->ddgst_remaining)
847 if (queue->recv_ddgst != queue->exp_ddgst) {
848 struct request *rq = nvme_cid_to_rq(nvme_tcp_tagset(queue),
850 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
852 req->status = cpu_to_le16(NVME_SC_DATA_XFER_ERROR);
854 dev_err(queue->ctrl->ctrl.device,
855 "data digest error: recv %#x expected %#x\n",
856 le32_to_cpu(queue->recv_ddgst),
857 le32_to_cpu(queue->exp_ddgst));
860 if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
861 struct request *rq = nvme_cid_to_rq(nvme_tcp_tagset(queue),
863 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
865 nvme_tcp_end_request(rq, le16_to_cpu(req->status));
869 nvme_tcp_init_recv_ctx(queue);
873 static int nvme_tcp_recv_skb(read_descriptor_t *desc, struct sk_buff *skb,
874 unsigned int offset, size_t len)
876 struct nvme_tcp_queue *queue = desc->arg.data;
877 size_t consumed = len;
881 switch (nvme_tcp_recv_state(queue)) {
882 case NVME_TCP_RECV_PDU:
883 result = nvme_tcp_recv_pdu(queue, skb, &offset, &len);
885 case NVME_TCP_RECV_DATA:
886 result = nvme_tcp_recv_data(queue, skb, &offset, &len);
888 case NVME_TCP_RECV_DDGST:
889 result = nvme_tcp_recv_ddgst(queue, skb, &offset, &len);
895 dev_err(queue->ctrl->ctrl.device,
896 "receive failed: %d\n", result);
897 queue->rd_enabled = false;
898 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
906 static void nvme_tcp_data_ready(struct sock *sk)
908 struct nvme_tcp_queue *queue;
910 read_lock_bh(&sk->sk_callback_lock);
911 queue = sk->sk_user_data;
912 if (likely(queue && queue->rd_enabled) &&
913 !test_bit(NVME_TCP_Q_POLLING, &queue->flags))
914 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
915 read_unlock_bh(&sk->sk_callback_lock);
918 static void nvme_tcp_write_space(struct sock *sk)
920 struct nvme_tcp_queue *queue;
922 read_lock_bh(&sk->sk_callback_lock);
923 queue = sk->sk_user_data;
924 if (likely(queue && sk_stream_is_writeable(sk))) {
925 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
926 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
928 read_unlock_bh(&sk->sk_callback_lock);
931 static void nvme_tcp_state_change(struct sock *sk)
933 struct nvme_tcp_queue *queue;
935 read_lock_bh(&sk->sk_callback_lock);
936 queue = sk->sk_user_data;
940 switch (sk->sk_state) {
946 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
949 dev_info(queue->ctrl->ctrl.device,
950 "queue %d socket state %d\n",
951 nvme_tcp_queue_id(queue), sk->sk_state);
954 queue->state_change(sk);
956 read_unlock_bh(&sk->sk_callback_lock);
959 static inline void nvme_tcp_done_send_req(struct nvme_tcp_queue *queue)
961 queue->request = NULL;
964 static void nvme_tcp_fail_request(struct nvme_tcp_request *req)
966 if (nvme_tcp_async_req(req)) {
967 union nvme_result res = {};
969 nvme_complete_async_event(&req->queue->ctrl->ctrl,
970 cpu_to_le16(NVME_SC_HOST_PATH_ERROR), &res);
972 nvme_tcp_end_request(blk_mq_rq_from_pdu(req),
973 NVME_SC_HOST_PATH_ERROR);
977 static int nvme_tcp_try_send_data(struct nvme_tcp_request *req)
979 struct nvme_tcp_queue *queue = req->queue;
980 int req_data_len = req->data_len;
981 u32 h2cdata_left = req->h2cdata_left;
984 struct page *page = nvme_tcp_req_cur_page(req);
985 size_t offset = nvme_tcp_req_cur_offset(req);
986 size_t len = nvme_tcp_req_cur_length(req);
987 bool last = nvme_tcp_pdu_last_send(req, len);
988 int req_data_sent = req->data_sent;
989 int ret, flags = MSG_DONTWAIT;
991 if (last && !queue->data_digest && !nvme_tcp_queue_more(queue))
994 flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST;
996 if (sendpage_ok(page)) {
997 ret = kernel_sendpage(queue->sock, page, offset, len,
1000 ret = sock_no_sendpage(queue->sock, page, offset, len,
1006 if (queue->data_digest)
1007 nvme_tcp_ddgst_update(queue->snd_hash, page,
1011 * update the request iterator except for the last payload send
1012 * in the request where we don't want to modify it as we may
1013 * compete with the RX path completing the request.
1015 if (req_data_sent + ret < req_data_len)
1016 nvme_tcp_advance_req(req, ret);
1018 /* fully successful last send in current PDU */
1019 if (last && ret == len) {
1020 if (queue->data_digest) {
1021 nvme_tcp_ddgst_final(queue->snd_hash,
1023 req->state = NVME_TCP_SEND_DDGST;
1027 nvme_tcp_setup_h2c_data_pdu(req);
1029 nvme_tcp_done_send_req(queue);
1037 static int nvme_tcp_try_send_cmd_pdu(struct nvme_tcp_request *req)
1039 struct nvme_tcp_queue *queue = req->queue;
1040 struct nvme_tcp_cmd_pdu *pdu = req->pdu;
1041 bool inline_data = nvme_tcp_has_inline_data(req);
1042 u8 hdgst = nvme_tcp_hdgst_len(queue);
1043 int len = sizeof(*pdu) + hdgst - req->offset;
1044 int flags = MSG_DONTWAIT;
1047 if (inline_data || nvme_tcp_queue_more(queue))
1048 flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST;
1052 if (queue->hdr_digest && !req->offset)
1053 nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
1055 ret = kernel_sendpage(queue->sock, virt_to_page(pdu),
1056 offset_in_page(pdu) + req->offset, len, flags);
1057 if (unlikely(ret <= 0))
1063 req->state = NVME_TCP_SEND_DATA;
1064 if (queue->data_digest)
1065 crypto_ahash_init(queue->snd_hash);
1067 nvme_tcp_done_send_req(queue);
1076 static int nvme_tcp_try_send_data_pdu(struct nvme_tcp_request *req)
1078 struct nvme_tcp_queue *queue = req->queue;
1079 struct nvme_tcp_data_pdu *pdu = req->pdu;
1080 u8 hdgst = nvme_tcp_hdgst_len(queue);
1081 int len = sizeof(*pdu) - req->offset + hdgst;
1084 if (queue->hdr_digest && !req->offset)
1085 nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
1087 if (!req->h2cdata_left)
1088 ret = kernel_sendpage(queue->sock, virt_to_page(pdu),
1089 offset_in_page(pdu) + req->offset, len,
1090 MSG_DONTWAIT | MSG_MORE | MSG_SENDPAGE_NOTLAST);
1092 ret = sock_no_sendpage(queue->sock, virt_to_page(pdu),
1093 offset_in_page(pdu) + req->offset, len,
1094 MSG_DONTWAIT | MSG_MORE);
1095 if (unlikely(ret <= 0))
1100 req->state = NVME_TCP_SEND_DATA;
1101 if (queue->data_digest)
1102 crypto_ahash_init(queue->snd_hash);
1110 static int nvme_tcp_try_send_ddgst(struct nvme_tcp_request *req)
1112 struct nvme_tcp_queue *queue = req->queue;
1113 size_t offset = req->offset;
1114 u32 h2cdata_left = req->h2cdata_left;
1116 struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
1118 .iov_base = (u8 *)&req->ddgst + req->offset,
1119 .iov_len = NVME_TCP_DIGEST_LENGTH - req->offset
1122 if (nvme_tcp_queue_more(queue))
1123 msg.msg_flags |= MSG_MORE;
1125 msg.msg_flags |= MSG_EOR;
1127 ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
1128 if (unlikely(ret <= 0))
1131 if (offset + ret == NVME_TCP_DIGEST_LENGTH) {
1133 nvme_tcp_setup_h2c_data_pdu(req);
1135 nvme_tcp_done_send_req(queue);
1143 static int nvme_tcp_try_send(struct nvme_tcp_queue *queue)
1145 struct nvme_tcp_request *req;
1148 if (!queue->request) {
1149 queue->request = nvme_tcp_fetch_request(queue);
1150 if (!queue->request)
1153 req = queue->request;
1155 if (req->state == NVME_TCP_SEND_CMD_PDU) {
1156 ret = nvme_tcp_try_send_cmd_pdu(req);
1159 if (!nvme_tcp_has_inline_data(req))
1163 if (req->state == NVME_TCP_SEND_H2C_PDU) {
1164 ret = nvme_tcp_try_send_data_pdu(req);
1169 if (req->state == NVME_TCP_SEND_DATA) {
1170 ret = nvme_tcp_try_send_data(req);
1175 if (req->state == NVME_TCP_SEND_DDGST)
1176 ret = nvme_tcp_try_send_ddgst(req);
1178 if (ret == -EAGAIN) {
1180 } else if (ret < 0) {
1181 dev_err(queue->ctrl->ctrl.device,
1182 "failed to send request %d\n", ret);
1183 nvme_tcp_fail_request(queue->request);
1184 nvme_tcp_done_send_req(queue);
1189 static int nvme_tcp_try_recv(struct nvme_tcp_queue *queue)
1191 struct socket *sock = queue->sock;
1192 struct sock *sk = sock->sk;
1193 read_descriptor_t rd_desc;
1196 rd_desc.arg.data = queue;
1200 consumed = sock->ops->read_sock(sk, &rd_desc, nvme_tcp_recv_skb);
1205 static void nvme_tcp_io_work(struct work_struct *w)
1207 struct nvme_tcp_queue *queue =
1208 container_of(w, struct nvme_tcp_queue, io_work);
1209 unsigned long deadline = jiffies + msecs_to_jiffies(1);
1212 bool pending = false;
1215 if (mutex_trylock(&queue->send_mutex)) {
1216 result = nvme_tcp_try_send(queue);
1217 mutex_unlock(&queue->send_mutex);
1220 else if (unlikely(result < 0))
1224 result = nvme_tcp_try_recv(queue);
1227 else if (unlikely(result < 0))
1233 } while (!time_after(jiffies, deadline)); /* quota is exhausted */
1235 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
1238 static void nvme_tcp_free_crypto(struct nvme_tcp_queue *queue)
1240 struct crypto_ahash *tfm = crypto_ahash_reqtfm(queue->rcv_hash);
1242 ahash_request_free(queue->rcv_hash);
1243 ahash_request_free(queue->snd_hash);
1244 crypto_free_ahash(tfm);
1247 static int nvme_tcp_alloc_crypto(struct nvme_tcp_queue *queue)
1249 struct crypto_ahash *tfm;
1251 tfm = crypto_alloc_ahash("crc32c", 0, CRYPTO_ALG_ASYNC);
1253 return PTR_ERR(tfm);
1255 queue->snd_hash = ahash_request_alloc(tfm, GFP_KERNEL);
1256 if (!queue->snd_hash)
1258 ahash_request_set_callback(queue->snd_hash, 0, NULL, NULL);
1260 queue->rcv_hash = ahash_request_alloc(tfm, GFP_KERNEL);
1261 if (!queue->rcv_hash)
1263 ahash_request_set_callback(queue->rcv_hash, 0, NULL, NULL);
1267 ahash_request_free(queue->snd_hash);
1269 crypto_free_ahash(tfm);
1273 static void nvme_tcp_free_async_req(struct nvme_tcp_ctrl *ctrl)
1275 struct nvme_tcp_request *async = &ctrl->async_req;
1277 page_frag_free(async->pdu);
1280 static int nvme_tcp_alloc_async_req(struct nvme_tcp_ctrl *ctrl)
1282 struct nvme_tcp_queue *queue = &ctrl->queues[0];
1283 struct nvme_tcp_request *async = &ctrl->async_req;
1284 u8 hdgst = nvme_tcp_hdgst_len(queue);
1286 async->pdu = page_frag_alloc(&queue->pf_cache,
1287 sizeof(struct nvme_tcp_cmd_pdu) + hdgst,
1288 GFP_KERNEL | __GFP_ZERO);
1292 async->queue = &ctrl->queues[0];
1296 static void nvme_tcp_free_queue(struct nvme_ctrl *nctrl, int qid)
1299 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1300 struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1302 if (!test_and_clear_bit(NVME_TCP_Q_ALLOCATED, &queue->flags))
1305 if (queue->hdr_digest || queue->data_digest)
1306 nvme_tcp_free_crypto(queue);
1308 if (queue->pf_cache.va) {
1309 page = virt_to_head_page(queue->pf_cache.va);
1310 __page_frag_cache_drain(page, queue->pf_cache.pagecnt_bias);
1311 queue->pf_cache.va = NULL;
1313 sock_release(queue->sock);
1315 mutex_destroy(&queue->send_mutex);
1316 mutex_destroy(&queue->queue_lock);
1319 static int nvme_tcp_init_connection(struct nvme_tcp_queue *queue)
1321 struct nvme_tcp_icreq_pdu *icreq;
1322 struct nvme_tcp_icresp_pdu *icresp;
1323 struct msghdr msg = {};
1325 bool ctrl_hdgst, ctrl_ddgst;
1329 icreq = kzalloc(sizeof(*icreq), GFP_KERNEL);
1333 icresp = kzalloc(sizeof(*icresp), GFP_KERNEL);
1339 icreq->hdr.type = nvme_tcp_icreq;
1340 icreq->hdr.hlen = sizeof(*icreq);
1342 icreq->hdr.plen = cpu_to_le32(icreq->hdr.hlen);
1343 icreq->pfv = cpu_to_le16(NVME_TCP_PFV_1_0);
1344 icreq->maxr2t = 0; /* single inflight r2t supported */
1345 icreq->hpda = 0; /* no alignment constraint */
1346 if (queue->hdr_digest)
1347 icreq->digest |= NVME_TCP_HDR_DIGEST_ENABLE;
1348 if (queue->data_digest)
1349 icreq->digest |= NVME_TCP_DATA_DIGEST_ENABLE;
1351 iov.iov_base = icreq;
1352 iov.iov_len = sizeof(*icreq);
1353 ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
1357 memset(&msg, 0, sizeof(msg));
1358 iov.iov_base = icresp;
1359 iov.iov_len = sizeof(*icresp);
1360 ret = kernel_recvmsg(queue->sock, &msg, &iov, 1,
1361 iov.iov_len, msg.msg_flags);
1366 if (icresp->hdr.type != nvme_tcp_icresp) {
1367 pr_err("queue %d: bad type returned %d\n",
1368 nvme_tcp_queue_id(queue), icresp->hdr.type);
1372 if (le32_to_cpu(icresp->hdr.plen) != sizeof(*icresp)) {
1373 pr_err("queue %d: bad pdu length returned %d\n",
1374 nvme_tcp_queue_id(queue), icresp->hdr.plen);
1378 if (icresp->pfv != NVME_TCP_PFV_1_0) {
1379 pr_err("queue %d: bad pfv returned %d\n",
1380 nvme_tcp_queue_id(queue), icresp->pfv);
1384 ctrl_ddgst = !!(icresp->digest & NVME_TCP_DATA_DIGEST_ENABLE);
1385 if ((queue->data_digest && !ctrl_ddgst) ||
1386 (!queue->data_digest && ctrl_ddgst)) {
1387 pr_err("queue %d: data digest mismatch host: %s ctrl: %s\n",
1388 nvme_tcp_queue_id(queue),
1389 queue->data_digest ? "enabled" : "disabled",
1390 ctrl_ddgst ? "enabled" : "disabled");
1394 ctrl_hdgst = !!(icresp->digest & NVME_TCP_HDR_DIGEST_ENABLE);
1395 if ((queue->hdr_digest && !ctrl_hdgst) ||
1396 (!queue->hdr_digest && ctrl_hdgst)) {
1397 pr_err("queue %d: header digest mismatch host: %s ctrl: %s\n",
1398 nvme_tcp_queue_id(queue),
1399 queue->hdr_digest ? "enabled" : "disabled",
1400 ctrl_hdgst ? "enabled" : "disabled");
1404 if (icresp->cpda != 0) {
1405 pr_err("queue %d: unsupported cpda returned %d\n",
1406 nvme_tcp_queue_id(queue), icresp->cpda);
1410 maxh2cdata = le32_to_cpu(icresp->maxdata);
1411 if ((maxh2cdata % 4) || (maxh2cdata < NVME_TCP_MIN_MAXH2CDATA)) {
1412 pr_err("queue %d: invalid maxh2cdata returned %u\n",
1413 nvme_tcp_queue_id(queue), maxh2cdata);
1416 queue->maxh2cdata = maxh2cdata;
1426 static bool nvme_tcp_admin_queue(struct nvme_tcp_queue *queue)
1428 return nvme_tcp_queue_id(queue) == 0;
1431 static bool nvme_tcp_default_queue(struct nvme_tcp_queue *queue)
1433 struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1434 int qid = nvme_tcp_queue_id(queue);
1436 return !nvme_tcp_admin_queue(queue) &&
1437 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT];
1440 static bool nvme_tcp_read_queue(struct nvme_tcp_queue *queue)
1442 struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1443 int qid = nvme_tcp_queue_id(queue);
1445 return !nvme_tcp_admin_queue(queue) &&
1446 !nvme_tcp_default_queue(queue) &&
1447 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT] +
1448 ctrl->io_queues[HCTX_TYPE_READ];
1451 static bool nvme_tcp_poll_queue(struct nvme_tcp_queue *queue)
1453 struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1454 int qid = nvme_tcp_queue_id(queue);
1456 return !nvme_tcp_admin_queue(queue) &&
1457 !nvme_tcp_default_queue(queue) &&
1458 !nvme_tcp_read_queue(queue) &&
1459 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT] +
1460 ctrl->io_queues[HCTX_TYPE_READ] +
1461 ctrl->io_queues[HCTX_TYPE_POLL];
1464 static void nvme_tcp_set_queue_io_cpu(struct nvme_tcp_queue *queue)
1466 struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1467 int qid = nvme_tcp_queue_id(queue);
1470 if (nvme_tcp_default_queue(queue))
1472 else if (nvme_tcp_read_queue(queue))
1473 n = qid - ctrl->io_queues[HCTX_TYPE_DEFAULT] - 1;
1474 else if (nvme_tcp_poll_queue(queue))
1475 n = qid - ctrl->io_queues[HCTX_TYPE_DEFAULT] -
1476 ctrl->io_queues[HCTX_TYPE_READ] - 1;
1477 queue->io_cpu = cpumask_next_wrap(n - 1, cpu_online_mask, -1, false);
1480 static int nvme_tcp_alloc_queue(struct nvme_ctrl *nctrl,
1481 int qid, size_t queue_size)
1483 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1484 struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1485 int ret, rcv_pdu_size;
1487 mutex_init(&queue->queue_lock);
1489 init_llist_head(&queue->req_list);
1490 INIT_LIST_HEAD(&queue->send_list);
1491 mutex_init(&queue->send_mutex);
1492 INIT_WORK(&queue->io_work, nvme_tcp_io_work);
1493 queue->queue_size = queue_size;
1496 queue->cmnd_capsule_len = nctrl->ioccsz * 16;
1498 queue->cmnd_capsule_len = sizeof(struct nvme_command) +
1499 NVME_TCP_ADMIN_CCSZ;
1501 ret = sock_create(ctrl->addr.ss_family, SOCK_STREAM,
1502 IPPROTO_TCP, &queue->sock);
1504 dev_err(nctrl->device,
1505 "failed to create socket: %d\n", ret);
1506 goto err_destroy_mutex;
1509 nvme_tcp_reclassify_socket(queue->sock);
1511 /* Single syn retry */
1512 tcp_sock_set_syncnt(queue->sock->sk, 1);
1514 /* Set TCP no delay */
1515 tcp_sock_set_nodelay(queue->sock->sk);
1518 * Cleanup whatever is sitting in the TCP transmit queue on socket
1519 * close. This is done to prevent stale data from being sent should
1520 * the network connection be restored before TCP times out.
1522 sock_no_linger(queue->sock->sk);
1524 if (so_priority > 0)
1525 sock_set_priority(queue->sock->sk, so_priority);
1527 /* Set socket type of service */
1528 if (nctrl->opts->tos >= 0)
1529 ip_sock_set_tos(queue->sock->sk, nctrl->opts->tos);
1531 /* Set 10 seconds timeout for icresp recvmsg */
1532 queue->sock->sk->sk_rcvtimeo = 10 * HZ;
1534 queue->sock->sk->sk_allocation = GFP_ATOMIC;
1535 nvme_tcp_set_queue_io_cpu(queue);
1536 queue->request = NULL;
1537 queue->data_remaining = 0;
1538 queue->ddgst_remaining = 0;
1539 queue->pdu_remaining = 0;
1540 queue->pdu_offset = 0;
1541 sk_set_memalloc(queue->sock->sk);
1543 if (nctrl->opts->mask & NVMF_OPT_HOST_TRADDR) {
1544 ret = kernel_bind(queue->sock, (struct sockaddr *)&ctrl->src_addr,
1545 sizeof(ctrl->src_addr));
1547 dev_err(nctrl->device,
1548 "failed to bind queue %d socket %d\n",
1554 if (nctrl->opts->mask & NVMF_OPT_HOST_IFACE) {
1555 char *iface = nctrl->opts->host_iface;
1556 sockptr_t optval = KERNEL_SOCKPTR(iface);
1558 ret = sock_setsockopt(queue->sock, SOL_SOCKET, SO_BINDTODEVICE,
1559 optval, strlen(iface));
1561 dev_err(nctrl->device,
1562 "failed to bind to interface %s queue %d err %d\n",
1568 queue->hdr_digest = nctrl->opts->hdr_digest;
1569 queue->data_digest = nctrl->opts->data_digest;
1570 if (queue->hdr_digest || queue->data_digest) {
1571 ret = nvme_tcp_alloc_crypto(queue);
1573 dev_err(nctrl->device,
1574 "failed to allocate queue %d crypto\n", qid);
1579 rcv_pdu_size = sizeof(struct nvme_tcp_rsp_pdu) +
1580 nvme_tcp_hdgst_len(queue);
1581 queue->pdu = kmalloc(rcv_pdu_size, GFP_KERNEL);
1587 dev_dbg(nctrl->device, "connecting queue %d\n",
1588 nvme_tcp_queue_id(queue));
1590 ret = kernel_connect(queue->sock, (struct sockaddr *)&ctrl->addr,
1591 sizeof(ctrl->addr), 0);
1593 dev_err(nctrl->device,
1594 "failed to connect socket: %d\n", ret);
1598 ret = nvme_tcp_init_connection(queue);
1600 goto err_init_connect;
1602 queue->rd_enabled = true;
1603 set_bit(NVME_TCP_Q_ALLOCATED, &queue->flags);
1604 nvme_tcp_init_recv_ctx(queue);
1606 write_lock_bh(&queue->sock->sk->sk_callback_lock);
1607 queue->sock->sk->sk_user_data = queue;
1608 queue->state_change = queue->sock->sk->sk_state_change;
1609 queue->data_ready = queue->sock->sk->sk_data_ready;
1610 queue->write_space = queue->sock->sk->sk_write_space;
1611 queue->sock->sk->sk_data_ready = nvme_tcp_data_ready;
1612 queue->sock->sk->sk_state_change = nvme_tcp_state_change;
1613 queue->sock->sk->sk_write_space = nvme_tcp_write_space;
1614 #ifdef CONFIG_NET_RX_BUSY_POLL
1615 queue->sock->sk->sk_ll_usec = 1;
1617 write_unlock_bh(&queue->sock->sk->sk_callback_lock);
1622 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1626 if (queue->hdr_digest || queue->data_digest)
1627 nvme_tcp_free_crypto(queue);
1629 sock_release(queue->sock);
1632 mutex_destroy(&queue->send_mutex);
1633 mutex_destroy(&queue->queue_lock);
1637 static void nvme_tcp_restore_sock_calls(struct nvme_tcp_queue *queue)
1639 struct socket *sock = queue->sock;
1641 write_lock_bh(&sock->sk->sk_callback_lock);
1642 sock->sk->sk_user_data = NULL;
1643 sock->sk->sk_data_ready = queue->data_ready;
1644 sock->sk->sk_state_change = queue->state_change;
1645 sock->sk->sk_write_space = queue->write_space;
1646 write_unlock_bh(&sock->sk->sk_callback_lock);
1649 static void __nvme_tcp_stop_queue(struct nvme_tcp_queue *queue)
1651 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1652 nvme_tcp_restore_sock_calls(queue);
1653 cancel_work_sync(&queue->io_work);
1656 static void nvme_tcp_stop_queue(struct nvme_ctrl *nctrl, int qid)
1658 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1659 struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1661 mutex_lock(&queue->queue_lock);
1662 if (test_and_clear_bit(NVME_TCP_Q_LIVE, &queue->flags))
1663 __nvme_tcp_stop_queue(queue);
1664 mutex_unlock(&queue->queue_lock);
1667 static int nvme_tcp_start_queue(struct nvme_ctrl *nctrl, int idx)
1669 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1673 ret = nvmf_connect_io_queue(nctrl, idx);
1675 ret = nvmf_connect_admin_queue(nctrl);
1678 set_bit(NVME_TCP_Q_LIVE, &ctrl->queues[idx].flags);
1680 if (test_bit(NVME_TCP_Q_ALLOCATED, &ctrl->queues[idx].flags))
1681 __nvme_tcp_stop_queue(&ctrl->queues[idx]);
1682 dev_err(nctrl->device,
1683 "failed to connect queue: %d ret=%d\n", idx, ret);
1688 static struct blk_mq_tag_set *nvme_tcp_alloc_tagset(struct nvme_ctrl *nctrl,
1691 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1692 struct blk_mq_tag_set *set;
1696 set = &ctrl->admin_tag_set;
1697 memset(set, 0, sizeof(*set));
1698 set->ops = &nvme_tcp_admin_mq_ops;
1699 set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
1700 set->reserved_tags = NVMF_RESERVED_TAGS;
1701 set->numa_node = nctrl->numa_node;
1702 set->flags = BLK_MQ_F_BLOCKING;
1703 set->cmd_size = sizeof(struct nvme_tcp_request);
1704 set->driver_data = ctrl;
1705 set->nr_hw_queues = 1;
1706 set->timeout = NVME_ADMIN_TIMEOUT;
1708 set = &ctrl->tag_set;
1709 memset(set, 0, sizeof(*set));
1710 set->ops = &nvme_tcp_mq_ops;
1711 set->queue_depth = nctrl->sqsize + 1;
1712 set->reserved_tags = NVMF_RESERVED_TAGS;
1713 set->numa_node = nctrl->numa_node;
1714 set->flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_BLOCKING;
1715 set->cmd_size = sizeof(struct nvme_tcp_request);
1716 set->driver_data = ctrl;
1717 set->nr_hw_queues = nctrl->queue_count - 1;
1718 set->timeout = NVME_IO_TIMEOUT;
1719 set->nr_maps = nctrl->opts->nr_poll_queues ? HCTX_MAX_TYPES : 2;
1722 ret = blk_mq_alloc_tag_set(set);
1724 return ERR_PTR(ret);
1729 static void nvme_tcp_free_admin_queue(struct nvme_ctrl *ctrl)
1731 if (to_tcp_ctrl(ctrl)->async_req.pdu) {
1732 cancel_work_sync(&ctrl->async_event_work);
1733 nvme_tcp_free_async_req(to_tcp_ctrl(ctrl));
1734 to_tcp_ctrl(ctrl)->async_req.pdu = NULL;
1737 nvme_tcp_free_queue(ctrl, 0);
1740 static void nvme_tcp_free_io_queues(struct nvme_ctrl *ctrl)
1744 for (i = 1; i < ctrl->queue_count; i++)
1745 nvme_tcp_free_queue(ctrl, i);
1748 static void nvme_tcp_stop_io_queues(struct nvme_ctrl *ctrl)
1752 for (i = 1; i < ctrl->queue_count; i++)
1753 nvme_tcp_stop_queue(ctrl, i);
1756 static int nvme_tcp_start_io_queues(struct nvme_ctrl *ctrl)
1760 for (i = 1; i < ctrl->queue_count; i++) {
1761 ret = nvme_tcp_start_queue(ctrl, i);
1763 goto out_stop_queues;
1769 for (i--; i >= 1; i--)
1770 nvme_tcp_stop_queue(ctrl, i);
1774 static int nvme_tcp_alloc_admin_queue(struct nvme_ctrl *ctrl)
1778 ret = nvme_tcp_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
1782 ret = nvme_tcp_alloc_async_req(to_tcp_ctrl(ctrl));
1784 goto out_free_queue;
1789 nvme_tcp_free_queue(ctrl, 0);
1793 static int __nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
1797 for (i = 1; i < ctrl->queue_count; i++) {
1798 ret = nvme_tcp_alloc_queue(ctrl, i, ctrl->sqsize + 1);
1800 goto out_free_queues;
1806 for (i--; i >= 1; i--)
1807 nvme_tcp_free_queue(ctrl, i);
1812 static unsigned int nvme_tcp_nr_io_queues(struct nvme_ctrl *ctrl)
1814 unsigned int nr_io_queues;
1816 nr_io_queues = min(ctrl->opts->nr_io_queues, num_online_cpus());
1817 nr_io_queues += min(ctrl->opts->nr_write_queues, num_online_cpus());
1818 nr_io_queues += min(ctrl->opts->nr_poll_queues, num_online_cpus());
1820 return nr_io_queues;
1823 static void nvme_tcp_set_io_queues(struct nvme_ctrl *nctrl,
1824 unsigned int nr_io_queues)
1826 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1827 struct nvmf_ctrl_options *opts = nctrl->opts;
1829 if (opts->nr_write_queues && opts->nr_io_queues < nr_io_queues) {
1831 * separate read/write queues
1832 * hand out dedicated default queues only after we have
1833 * sufficient read queues.
1835 ctrl->io_queues[HCTX_TYPE_READ] = opts->nr_io_queues;
1836 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_READ];
1837 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
1838 min(opts->nr_write_queues, nr_io_queues);
1839 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
1842 * shared read/write queues
1843 * either no write queues were requested, or we don't have
1844 * sufficient queue count to have dedicated default queues.
1846 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
1847 min(opts->nr_io_queues, nr_io_queues);
1848 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
1851 if (opts->nr_poll_queues && nr_io_queues) {
1852 /* map dedicated poll queues only if we have queues left */
1853 ctrl->io_queues[HCTX_TYPE_POLL] =
1854 min(opts->nr_poll_queues, nr_io_queues);
1858 static int nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
1860 unsigned int nr_io_queues;
1863 nr_io_queues = nvme_tcp_nr_io_queues(ctrl);
1864 ret = nvme_set_queue_count(ctrl, &nr_io_queues);
1868 if (nr_io_queues == 0) {
1869 dev_err(ctrl->device,
1870 "unable to set any I/O queues\n");
1874 ctrl->queue_count = nr_io_queues + 1;
1875 dev_info(ctrl->device,
1876 "creating %d I/O queues.\n", nr_io_queues);
1878 nvme_tcp_set_io_queues(ctrl, nr_io_queues);
1880 return __nvme_tcp_alloc_io_queues(ctrl);
1883 static void nvme_tcp_destroy_io_queues(struct nvme_ctrl *ctrl, bool remove)
1885 nvme_tcp_stop_io_queues(ctrl);
1887 blk_mq_destroy_queue(ctrl->connect_q);
1888 blk_mq_free_tag_set(ctrl->tagset);
1890 nvme_tcp_free_io_queues(ctrl);
1893 static int nvme_tcp_configure_io_queues(struct nvme_ctrl *ctrl, bool new)
1897 ret = nvme_tcp_alloc_io_queues(ctrl);
1902 ctrl->tagset = nvme_tcp_alloc_tagset(ctrl, false);
1903 if (IS_ERR(ctrl->tagset)) {
1904 ret = PTR_ERR(ctrl->tagset);
1905 goto out_free_io_queues;
1908 ret = nvme_ctrl_init_connect_q(ctrl);
1910 goto out_free_tag_set;
1913 ret = nvme_tcp_start_io_queues(ctrl);
1915 goto out_cleanup_connect_q;
1918 nvme_start_queues(ctrl);
1919 if (!nvme_wait_freeze_timeout(ctrl, NVME_IO_TIMEOUT)) {
1921 * If we timed out waiting for freeze we are likely to
1922 * be stuck. Fail the controller initialization just
1926 goto out_wait_freeze_timed_out;
1928 blk_mq_update_nr_hw_queues(ctrl->tagset,
1929 ctrl->queue_count - 1);
1930 nvme_unfreeze(ctrl);
1935 out_wait_freeze_timed_out:
1936 nvme_stop_queues(ctrl);
1937 nvme_sync_io_queues(ctrl);
1938 nvme_tcp_stop_io_queues(ctrl);
1939 out_cleanup_connect_q:
1940 nvme_cancel_tagset(ctrl);
1942 blk_mq_destroy_queue(ctrl->connect_q);
1945 blk_mq_free_tag_set(ctrl->tagset);
1947 nvme_tcp_free_io_queues(ctrl);
1951 static void nvme_tcp_destroy_admin_queue(struct nvme_ctrl *ctrl, bool remove)
1953 nvme_tcp_stop_queue(ctrl, 0);
1955 blk_mq_destroy_queue(ctrl->admin_q);
1956 blk_mq_destroy_queue(ctrl->fabrics_q);
1957 blk_mq_free_tag_set(ctrl->admin_tagset);
1959 nvme_tcp_free_admin_queue(ctrl);
1962 static int nvme_tcp_configure_admin_queue(struct nvme_ctrl *ctrl, bool new)
1966 error = nvme_tcp_alloc_admin_queue(ctrl);
1971 ctrl->admin_tagset = nvme_tcp_alloc_tagset(ctrl, true);
1972 if (IS_ERR(ctrl->admin_tagset)) {
1973 error = PTR_ERR(ctrl->admin_tagset);
1974 goto out_free_queue;
1977 ctrl->fabrics_q = blk_mq_init_queue(ctrl->admin_tagset);
1978 if (IS_ERR(ctrl->fabrics_q)) {
1979 error = PTR_ERR(ctrl->fabrics_q);
1980 goto out_free_tagset;
1983 ctrl->admin_q = blk_mq_init_queue(ctrl->admin_tagset);
1984 if (IS_ERR(ctrl->admin_q)) {
1985 error = PTR_ERR(ctrl->admin_q);
1986 goto out_cleanup_fabrics_q;
1990 error = nvme_tcp_start_queue(ctrl, 0);
1992 goto out_cleanup_queue;
1994 error = nvme_enable_ctrl(ctrl);
1996 goto out_stop_queue;
1998 nvme_start_admin_queue(ctrl);
2000 error = nvme_init_ctrl_finish(ctrl);
2002 goto out_quiesce_queue;
2007 nvme_stop_admin_queue(ctrl);
2008 blk_sync_queue(ctrl->admin_q);
2010 nvme_tcp_stop_queue(ctrl, 0);
2011 nvme_cancel_admin_tagset(ctrl);
2014 blk_mq_destroy_queue(ctrl->admin_q);
2015 out_cleanup_fabrics_q:
2017 blk_mq_destroy_queue(ctrl->fabrics_q);
2020 blk_mq_free_tag_set(ctrl->admin_tagset);
2022 nvme_tcp_free_admin_queue(ctrl);
2026 static void nvme_tcp_teardown_admin_queue(struct nvme_ctrl *ctrl,
2029 nvme_stop_admin_queue(ctrl);
2030 blk_sync_queue(ctrl->admin_q);
2031 nvme_tcp_stop_queue(ctrl, 0);
2032 nvme_cancel_admin_tagset(ctrl);
2034 nvme_start_admin_queue(ctrl);
2035 nvme_tcp_destroy_admin_queue(ctrl, remove);
2038 static void nvme_tcp_teardown_io_queues(struct nvme_ctrl *ctrl,
2041 if (ctrl->queue_count <= 1)
2043 nvme_stop_admin_queue(ctrl);
2044 nvme_start_freeze(ctrl);
2045 nvme_stop_queues(ctrl);
2046 nvme_sync_io_queues(ctrl);
2047 nvme_tcp_stop_io_queues(ctrl);
2048 nvme_cancel_tagset(ctrl);
2050 nvme_start_queues(ctrl);
2051 nvme_tcp_destroy_io_queues(ctrl, remove);
2054 static void nvme_tcp_reconnect_or_remove(struct nvme_ctrl *ctrl)
2056 /* If we are resetting/deleting then do nothing */
2057 if (ctrl->state != NVME_CTRL_CONNECTING) {
2058 WARN_ON_ONCE(ctrl->state == NVME_CTRL_NEW ||
2059 ctrl->state == NVME_CTRL_LIVE);
2063 if (nvmf_should_reconnect(ctrl)) {
2064 dev_info(ctrl->device, "Reconnecting in %d seconds...\n",
2065 ctrl->opts->reconnect_delay);
2066 queue_delayed_work(nvme_wq, &to_tcp_ctrl(ctrl)->connect_work,
2067 ctrl->opts->reconnect_delay * HZ);
2069 dev_info(ctrl->device, "Removing controller...\n");
2070 nvme_delete_ctrl(ctrl);
2074 static int nvme_tcp_setup_ctrl(struct nvme_ctrl *ctrl, bool new)
2076 struct nvmf_ctrl_options *opts = ctrl->opts;
2079 ret = nvme_tcp_configure_admin_queue(ctrl, new);
2085 dev_err(ctrl->device, "icdoff is not supported!\n");
2089 if (!nvme_ctrl_sgl_supported(ctrl)) {
2091 dev_err(ctrl->device, "Mandatory sgls are not supported!\n");
2095 if (opts->queue_size > ctrl->sqsize + 1)
2096 dev_warn(ctrl->device,
2097 "queue_size %zu > ctrl sqsize %u, clamping down\n",
2098 opts->queue_size, ctrl->sqsize + 1);
2100 if (ctrl->sqsize + 1 > ctrl->maxcmd) {
2101 dev_warn(ctrl->device,
2102 "sqsize %u > ctrl maxcmd %u, clamping down\n",
2103 ctrl->sqsize + 1, ctrl->maxcmd);
2104 ctrl->sqsize = ctrl->maxcmd - 1;
2107 if (ctrl->queue_count > 1) {
2108 ret = nvme_tcp_configure_io_queues(ctrl, new);
2113 if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_LIVE)) {
2115 * state change failure is ok if we started ctrl delete,
2116 * unless we're during creation of a new controller to
2117 * avoid races with teardown flow.
2119 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING &&
2120 ctrl->state != NVME_CTRL_DELETING_NOIO);
2126 nvme_start_ctrl(ctrl);
2130 if (ctrl->queue_count > 1) {
2131 nvme_stop_queues(ctrl);
2132 nvme_sync_io_queues(ctrl);
2133 nvme_tcp_stop_io_queues(ctrl);
2134 nvme_cancel_tagset(ctrl);
2135 nvme_tcp_destroy_io_queues(ctrl, new);
2138 nvme_stop_admin_queue(ctrl);
2139 blk_sync_queue(ctrl->admin_q);
2140 nvme_tcp_stop_queue(ctrl, 0);
2141 nvme_cancel_admin_tagset(ctrl);
2142 nvme_tcp_destroy_admin_queue(ctrl, new);
2146 static void nvme_tcp_reconnect_ctrl_work(struct work_struct *work)
2148 struct nvme_tcp_ctrl *tcp_ctrl = container_of(to_delayed_work(work),
2149 struct nvme_tcp_ctrl, connect_work);
2150 struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
2152 ++ctrl->nr_reconnects;
2154 if (nvme_tcp_setup_ctrl(ctrl, false))
2157 dev_info(ctrl->device, "Successfully reconnected (%d attempt)\n",
2158 ctrl->nr_reconnects);
2160 ctrl->nr_reconnects = 0;
2165 dev_info(ctrl->device, "Failed reconnect attempt %d\n",
2166 ctrl->nr_reconnects);
2167 nvme_tcp_reconnect_or_remove(ctrl);
2170 static void nvme_tcp_error_recovery_work(struct work_struct *work)
2172 struct nvme_tcp_ctrl *tcp_ctrl = container_of(work,
2173 struct nvme_tcp_ctrl, err_work);
2174 struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
2176 nvme_stop_keep_alive(ctrl);
2177 flush_work(&ctrl->async_event_work);
2178 nvme_tcp_teardown_io_queues(ctrl, false);
2179 /* unquiesce to fail fast pending requests */
2180 nvme_start_queues(ctrl);
2181 nvme_tcp_teardown_admin_queue(ctrl, false);
2182 nvme_start_admin_queue(ctrl);
2184 if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
2185 /* state change failure is ok if we started ctrl delete */
2186 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING &&
2187 ctrl->state != NVME_CTRL_DELETING_NOIO);
2191 nvme_tcp_reconnect_or_remove(ctrl);
2194 static void nvme_tcp_teardown_ctrl(struct nvme_ctrl *ctrl, bool shutdown)
2196 nvme_tcp_teardown_io_queues(ctrl, shutdown);
2197 nvme_stop_admin_queue(ctrl);
2199 nvme_shutdown_ctrl(ctrl);
2201 nvme_disable_ctrl(ctrl);
2202 nvme_tcp_teardown_admin_queue(ctrl, shutdown);
2205 static void nvme_tcp_delete_ctrl(struct nvme_ctrl *ctrl)
2207 nvme_tcp_teardown_ctrl(ctrl, true);
2210 static void nvme_reset_ctrl_work(struct work_struct *work)
2212 struct nvme_ctrl *ctrl =
2213 container_of(work, struct nvme_ctrl, reset_work);
2215 nvme_stop_ctrl(ctrl);
2216 nvme_tcp_teardown_ctrl(ctrl, false);
2218 if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
2219 /* state change failure is ok if we started ctrl delete */
2220 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING &&
2221 ctrl->state != NVME_CTRL_DELETING_NOIO);
2225 if (nvme_tcp_setup_ctrl(ctrl, false))
2231 ++ctrl->nr_reconnects;
2232 nvme_tcp_reconnect_or_remove(ctrl);
2235 static void nvme_tcp_stop_ctrl(struct nvme_ctrl *ctrl)
2237 cancel_work_sync(&to_tcp_ctrl(ctrl)->err_work);
2238 cancel_delayed_work_sync(&to_tcp_ctrl(ctrl)->connect_work);
2241 static void nvme_tcp_free_ctrl(struct nvme_ctrl *nctrl)
2243 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
2245 if (list_empty(&ctrl->list))
2248 mutex_lock(&nvme_tcp_ctrl_mutex);
2249 list_del(&ctrl->list);
2250 mutex_unlock(&nvme_tcp_ctrl_mutex);
2252 nvmf_free_options(nctrl->opts);
2254 kfree(ctrl->queues);
2258 static void nvme_tcp_set_sg_null(struct nvme_command *c)
2260 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2264 sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
2265 NVME_SGL_FMT_TRANSPORT_A;
2268 static void nvme_tcp_set_sg_inline(struct nvme_tcp_queue *queue,
2269 struct nvme_command *c, u32 data_len)
2271 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2273 sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
2274 sg->length = cpu_to_le32(data_len);
2275 sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
2278 static void nvme_tcp_set_sg_host_data(struct nvme_command *c,
2281 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2284 sg->length = cpu_to_le32(data_len);
2285 sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
2286 NVME_SGL_FMT_TRANSPORT_A;
2289 static void nvme_tcp_submit_async_event(struct nvme_ctrl *arg)
2291 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(arg);
2292 struct nvme_tcp_queue *queue = &ctrl->queues[0];
2293 struct nvme_tcp_cmd_pdu *pdu = ctrl->async_req.pdu;
2294 struct nvme_command *cmd = &pdu->cmd;
2295 u8 hdgst = nvme_tcp_hdgst_len(queue);
2297 memset(pdu, 0, sizeof(*pdu));
2298 pdu->hdr.type = nvme_tcp_cmd;
2299 if (queue->hdr_digest)
2300 pdu->hdr.flags |= NVME_TCP_F_HDGST;
2301 pdu->hdr.hlen = sizeof(*pdu);
2302 pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst);
2304 cmd->common.opcode = nvme_admin_async_event;
2305 cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
2306 cmd->common.flags |= NVME_CMD_SGL_METABUF;
2307 nvme_tcp_set_sg_null(cmd);
2309 ctrl->async_req.state = NVME_TCP_SEND_CMD_PDU;
2310 ctrl->async_req.offset = 0;
2311 ctrl->async_req.curr_bio = NULL;
2312 ctrl->async_req.data_len = 0;
2314 nvme_tcp_queue_request(&ctrl->async_req, true, true);
2317 static void nvme_tcp_complete_timed_out(struct request *rq)
2319 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2320 struct nvme_ctrl *ctrl = &req->queue->ctrl->ctrl;
2322 nvme_tcp_stop_queue(ctrl, nvme_tcp_queue_id(req->queue));
2323 nvmf_complete_timed_out_request(rq);
2326 static enum blk_eh_timer_return nvme_tcp_timeout(struct request *rq)
2328 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2329 struct nvme_ctrl *ctrl = &req->queue->ctrl->ctrl;
2330 struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2332 dev_warn(ctrl->device,
2333 "queue %d: timeout request %#x type %d\n",
2334 nvme_tcp_queue_id(req->queue), rq->tag, pdu->hdr.type);
2336 if (ctrl->state != NVME_CTRL_LIVE) {
2338 * If we are resetting, connecting or deleting we should
2339 * complete immediately because we may block controller
2340 * teardown or setup sequence
2341 * - ctrl disable/shutdown fabrics requests
2342 * - connect requests
2343 * - initialization admin requests
2344 * - I/O requests that entered after unquiescing and
2345 * the controller stopped responding
2347 * All other requests should be cancelled by the error
2348 * recovery work, so it's fine that we fail it here.
2350 nvme_tcp_complete_timed_out(rq);
2355 * LIVE state should trigger the normal error recovery which will
2356 * handle completing this request.
2358 nvme_tcp_error_recovery(ctrl);
2359 return BLK_EH_RESET_TIMER;
2362 static blk_status_t nvme_tcp_map_data(struct nvme_tcp_queue *queue,
2365 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2366 struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2367 struct nvme_command *c = &pdu->cmd;
2369 c->common.flags |= NVME_CMD_SGL_METABUF;
2371 if (!blk_rq_nr_phys_segments(rq))
2372 nvme_tcp_set_sg_null(c);
2373 else if (rq_data_dir(rq) == WRITE &&
2374 req->data_len <= nvme_tcp_inline_data_size(queue))
2375 nvme_tcp_set_sg_inline(queue, c, req->data_len);
2377 nvme_tcp_set_sg_host_data(c, req->data_len);
2382 static blk_status_t nvme_tcp_setup_cmd_pdu(struct nvme_ns *ns,
2385 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2386 struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2387 struct nvme_tcp_queue *queue = req->queue;
2388 u8 hdgst = nvme_tcp_hdgst_len(queue), ddgst = 0;
2391 ret = nvme_setup_cmd(ns, rq);
2395 req->state = NVME_TCP_SEND_CMD_PDU;
2396 req->status = cpu_to_le16(NVME_SC_SUCCESS);
2401 req->h2cdata_left = 0;
2402 req->data_len = blk_rq_nr_phys_segments(rq) ?
2403 blk_rq_payload_bytes(rq) : 0;
2404 req->curr_bio = rq->bio;
2405 if (req->curr_bio && req->data_len)
2406 nvme_tcp_init_iter(req, rq_data_dir(rq));
2408 if (rq_data_dir(rq) == WRITE &&
2409 req->data_len <= nvme_tcp_inline_data_size(queue))
2410 req->pdu_len = req->data_len;
2412 pdu->hdr.type = nvme_tcp_cmd;
2414 if (queue->hdr_digest)
2415 pdu->hdr.flags |= NVME_TCP_F_HDGST;
2416 if (queue->data_digest && req->pdu_len) {
2417 pdu->hdr.flags |= NVME_TCP_F_DDGST;
2418 ddgst = nvme_tcp_ddgst_len(queue);
2420 pdu->hdr.hlen = sizeof(*pdu);
2421 pdu->hdr.pdo = req->pdu_len ? pdu->hdr.hlen + hdgst : 0;
2423 cpu_to_le32(pdu->hdr.hlen + hdgst + req->pdu_len + ddgst);
2425 ret = nvme_tcp_map_data(queue, rq);
2426 if (unlikely(ret)) {
2427 nvme_cleanup_cmd(rq);
2428 dev_err(queue->ctrl->ctrl.device,
2429 "Failed to map data (%d)\n", ret);
2436 static void nvme_tcp_commit_rqs(struct blk_mq_hw_ctx *hctx)
2438 struct nvme_tcp_queue *queue = hctx->driver_data;
2440 if (!llist_empty(&queue->req_list))
2441 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
2444 static blk_status_t nvme_tcp_queue_rq(struct blk_mq_hw_ctx *hctx,
2445 const struct blk_mq_queue_data *bd)
2447 struct nvme_ns *ns = hctx->queue->queuedata;
2448 struct nvme_tcp_queue *queue = hctx->driver_data;
2449 struct request *rq = bd->rq;
2450 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2451 bool queue_ready = test_bit(NVME_TCP_Q_LIVE, &queue->flags);
2454 if (!nvme_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
2455 return nvme_fail_nonready_command(&queue->ctrl->ctrl, rq);
2457 ret = nvme_tcp_setup_cmd_pdu(ns, rq);
2461 blk_mq_start_request(rq);
2463 nvme_tcp_queue_request(req, true, bd->last);
2468 static int nvme_tcp_map_queues(struct blk_mq_tag_set *set)
2470 struct nvme_tcp_ctrl *ctrl = set->driver_data;
2471 struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
2473 if (opts->nr_write_queues && ctrl->io_queues[HCTX_TYPE_READ]) {
2474 /* separate read/write queues */
2475 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2476 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2477 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2478 set->map[HCTX_TYPE_READ].nr_queues =
2479 ctrl->io_queues[HCTX_TYPE_READ];
2480 set->map[HCTX_TYPE_READ].queue_offset =
2481 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2483 /* shared read/write queues */
2484 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2485 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2486 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2487 set->map[HCTX_TYPE_READ].nr_queues =
2488 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2489 set->map[HCTX_TYPE_READ].queue_offset = 0;
2491 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
2492 blk_mq_map_queues(&set->map[HCTX_TYPE_READ]);
2494 if (opts->nr_poll_queues && ctrl->io_queues[HCTX_TYPE_POLL]) {
2495 /* map dedicated poll queues only if we have queues left */
2496 set->map[HCTX_TYPE_POLL].nr_queues =
2497 ctrl->io_queues[HCTX_TYPE_POLL];
2498 set->map[HCTX_TYPE_POLL].queue_offset =
2499 ctrl->io_queues[HCTX_TYPE_DEFAULT] +
2500 ctrl->io_queues[HCTX_TYPE_READ];
2501 blk_mq_map_queues(&set->map[HCTX_TYPE_POLL]);
2504 dev_info(ctrl->ctrl.device,
2505 "mapped %d/%d/%d default/read/poll queues.\n",
2506 ctrl->io_queues[HCTX_TYPE_DEFAULT],
2507 ctrl->io_queues[HCTX_TYPE_READ],
2508 ctrl->io_queues[HCTX_TYPE_POLL]);
2513 static int nvme_tcp_poll(struct blk_mq_hw_ctx *hctx, struct io_comp_batch *iob)
2515 struct nvme_tcp_queue *queue = hctx->driver_data;
2516 struct sock *sk = queue->sock->sk;
2518 if (!test_bit(NVME_TCP_Q_LIVE, &queue->flags))
2521 set_bit(NVME_TCP_Q_POLLING, &queue->flags);
2522 if (sk_can_busy_loop(sk) && skb_queue_empty_lockless(&sk->sk_receive_queue))
2523 sk_busy_loop(sk, true);
2524 nvme_tcp_try_recv(queue);
2525 clear_bit(NVME_TCP_Q_POLLING, &queue->flags);
2526 return queue->nr_cqe;
2529 static const struct blk_mq_ops nvme_tcp_mq_ops = {
2530 .queue_rq = nvme_tcp_queue_rq,
2531 .commit_rqs = nvme_tcp_commit_rqs,
2532 .complete = nvme_complete_rq,
2533 .init_request = nvme_tcp_init_request,
2534 .exit_request = nvme_tcp_exit_request,
2535 .init_hctx = nvme_tcp_init_hctx,
2536 .timeout = nvme_tcp_timeout,
2537 .map_queues = nvme_tcp_map_queues,
2538 .poll = nvme_tcp_poll,
2541 static const struct blk_mq_ops nvme_tcp_admin_mq_ops = {
2542 .queue_rq = nvme_tcp_queue_rq,
2543 .complete = nvme_complete_rq,
2544 .init_request = nvme_tcp_init_request,
2545 .exit_request = nvme_tcp_exit_request,
2546 .init_hctx = nvme_tcp_init_admin_hctx,
2547 .timeout = nvme_tcp_timeout,
2550 static const struct nvme_ctrl_ops nvme_tcp_ctrl_ops = {
2552 .module = THIS_MODULE,
2553 .flags = NVME_F_FABRICS,
2554 .reg_read32 = nvmf_reg_read32,
2555 .reg_read64 = nvmf_reg_read64,
2556 .reg_write32 = nvmf_reg_write32,
2557 .free_ctrl = nvme_tcp_free_ctrl,
2558 .submit_async_event = nvme_tcp_submit_async_event,
2559 .delete_ctrl = nvme_tcp_delete_ctrl,
2560 .get_address = nvmf_get_address,
2561 .stop_ctrl = nvme_tcp_stop_ctrl,
2565 nvme_tcp_existing_controller(struct nvmf_ctrl_options *opts)
2567 struct nvme_tcp_ctrl *ctrl;
2570 mutex_lock(&nvme_tcp_ctrl_mutex);
2571 list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list) {
2572 found = nvmf_ip_options_match(&ctrl->ctrl, opts);
2576 mutex_unlock(&nvme_tcp_ctrl_mutex);
2581 static struct nvme_ctrl *nvme_tcp_create_ctrl(struct device *dev,
2582 struct nvmf_ctrl_options *opts)
2584 struct nvme_tcp_ctrl *ctrl;
2587 ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
2589 return ERR_PTR(-ENOMEM);
2591 INIT_LIST_HEAD(&ctrl->list);
2592 ctrl->ctrl.opts = opts;
2593 ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues +
2594 opts->nr_poll_queues + 1;
2595 ctrl->ctrl.sqsize = opts->queue_size - 1;
2596 ctrl->ctrl.kato = opts->kato;
2598 INIT_DELAYED_WORK(&ctrl->connect_work,
2599 nvme_tcp_reconnect_ctrl_work);
2600 INIT_WORK(&ctrl->err_work, nvme_tcp_error_recovery_work);
2601 INIT_WORK(&ctrl->ctrl.reset_work, nvme_reset_ctrl_work);
2603 if (!(opts->mask & NVMF_OPT_TRSVCID)) {
2605 kstrdup(__stringify(NVME_TCP_DISC_PORT), GFP_KERNEL);
2606 if (!opts->trsvcid) {
2610 opts->mask |= NVMF_OPT_TRSVCID;
2613 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2614 opts->traddr, opts->trsvcid, &ctrl->addr);
2616 pr_err("malformed address passed: %s:%s\n",
2617 opts->traddr, opts->trsvcid);
2621 if (opts->mask & NVMF_OPT_HOST_TRADDR) {
2622 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2623 opts->host_traddr, NULL, &ctrl->src_addr);
2625 pr_err("malformed src address passed: %s\n",
2631 if (opts->mask & NVMF_OPT_HOST_IFACE) {
2632 if (!__dev_get_by_name(&init_net, opts->host_iface)) {
2633 pr_err("invalid interface passed: %s\n",
2640 if (!opts->duplicate_connect && nvme_tcp_existing_controller(opts)) {
2645 ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
2647 if (!ctrl->queues) {
2652 ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_tcp_ctrl_ops, 0);
2654 goto out_kfree_queues;
2656 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
2659 goto out_uninit_ctrl;
2662 ret = nvme_tcp_setup_ctrl(&ctrl->ctrl, true);
2664 goto out_uninit_ctrl;
2666 dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISp\n",
2667 nvmf_ctrl_subsysnqn(&ctrl->ctrl), &ctrl->addr);
2669 mutex_lock(&nvme_tcp_ctrl_mutex);
2670 list_add_tail(&ctrl->list, &nvme_tcp_ctrl_list);
2671 mutex_unlock(&nvme_tcp_ctrl_mutex);
2676 nvme_uninit_ctrl(&ctrl->ctrl);
2677 nvme_put_ctrl(&ctrl->ctrl);
2680 return ERR_PTR(ret);
2682 kfree(ctrl->queues);
2685 return ERR_PTR(ret);
2688 static struct nvmf_transport_ops nvme_tcp_transport = {
2690 .module = THIS_MODULE,
2691 .required_opts = NVMF_OPT_TRADDR,
2692 .allowed_opts = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2693 NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
2694 NVMF_OPT_HDR_DIGEST | NVMF_OPT_DATA_DIGEST |
2695 NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES |
2696 NVMF_OPT_TOS | NVMF_OPT_HOST_IFACE,
2697 .create_ctrl = nvme_tcp_create_ctrl,
2700 static int __init nvme_tcp_init_module(void)
2702 nvme_tcp_wq = alloc_workqueue("nvme_tcp_wq",
2703 WQ_MEM_RECLAIM | WQ_HIGHPRI, 0);
2707 nvmf_register_transport(&nvme_tcp_transport);
2711 static void __exit nvme_tcp_cleanup_module(void)
2713 struct nvme_tcp_ctrl *ctrl;
2715 nvmf_unregister_transport(&nvme_tcp_transport);
2717 mutex_lock(&nvme_tcp_ctrl_mutex);
2718 list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list)
2719 nvme_delete_ctrl(&ctrl->ctrl);
2720 mutex_unlock(&nvme_tcp_ctrl_mutex);
2721 flush_workqueue(nvme_delete_wq);
2723 destroy_workqueue(nvme_tcp_wq);
2726 module_init(nvme_tcp_init_module);
2727 module_exit(nvme_tcp_cleanup_module);
2729 MODULE_LICENSE("GPL v2");