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 enum nvme_tcp_send_state {
34 NVME_TCP_SEND_CMD_PDU = 0,
35 NVME_TCP_SEND_H2C_PDU,
40 struct nvme_tcp_request {
41 struct nvme_request req;
43 struct nvme_tcp_queue *queue;
49 struct list_head entry;
50 struct llist_node lentry;
59 enum nvme_tcp_send_state state;
62 enum nvme_tcp_queue_flags {
63 NVME_TCP_Q_ALLOCATED = 0,
65 NVME_TCP_Q_POLLING = 2,
68 enum nvme_tcp_recv_state {
69 NVME_TCP_RECV_PDU = 0,
75 struct nvme_tcp_queue {
77 struct work_struct io_work;
80 struct mutex queue_lock;
81 struct mutex send_mutex;
82 struct llist_head req_list;
83 struct list_head send_list;
90 size_t data_remaining;
91 size_t ddgst_remaining;
95 struct nvme_tcp_request *request;
98 size_t cmnd_capsule_len;
99 struct nvme_tcp_ctrl *ctrl;
105 struct ahash_request *rcv_hash;
106 struct ahash_request *snd_hash;
110 struct page_frag_cache pf_cache;
112 void (*state_change)(struct sock *);
113 void (*data_ready)(struct sock *);
114 void (*write_space)(struct sock *);
117 struct nvme_tcp_ctrl {
118 /* read only in the hot path */
119 struct nvme_tcp_queue *queues;
120 struct blk_mq_tag_set tag_set;
122 /* other member variables */
123 struct list_head list;
124 struct blk_mq_tag_set admin_tag_set;
125 struct sockaddr_storage addr;
126 struct sockaddr_storage src_addr;
127 struct nvme_ctrl ctrl;
129 struct work_struct err_work;
130 struct delayed_work connect_work;
131 struct nvme_tcp_request async_req;
132 u32 io_queues[HCTX_MAX_TYPES];
135 static LIST_HEAD(nvme_tcp_ctrl_list);
136 static DEFINE_MUTEX(nvme_tcp_ctrl_mutex);
137 static struct workqueue_struct *nvme_tcp_wq;
138 static const struct blk_mq_ops nvme_tcp_mq_ops;
139 static const struct blk_mq_ops nvme_tcp_admin_mq_ops;
140 static int nvme_tcp_try_send(struct nvme_tcp_queue *queue);
142 static inline struct nvme_tcp_ctrl *to_tcp_ctrl(struct nvme_ctrl *ctrl)
144 return container_of(ctrl, struct nvme_tcp_ctrl, ctrl);
147 static inline int nvme_tcp_queue_id(struct nvme_tcp_queue *queue)
149 return queue - queue->ctrl->queues;
152 static inline struct blk_mq_tags *nvme_tcp_tagset(struct nvme_tcp_queue *queue)
154 u32 queue_idx = nvme_tcp_queue_id(queue);
157 return queue->ctrl->admin_tag_set.tags[queue_idx];
158 return queue->ctrl->tag_set.tags[queue_idx - 1];
161 static inline u8 nvme_tcp_hdgst_len(struct nvme_tcp_queue *queue)
163 return queue->hdr_digest ? NVME_TCP_DIGEST_LENGTH : 0;
166 static inline u8 nvme_tcp_ddgst_len(struct nvme_tcp_queue *queue)
168 return queue->data_digest ? NVME_TCP_DIGEST_LENGTH : 0;
171 static inline size_t nvme_tcp_inline_data_size(struct nvme_tcp_queue *queue)
173 return queue->cmnd_capsule_len - sizeof(struct nvme_command);
176 static inline bool nvme_tcp_async_req(struct nvme_tcp_request *req)
178 return req == &req->queue->ctrl->async_req;
181 static inline bool nvme_tcp_has_inline_data(struct nvme_tcp_request *req)
185 if (unlikely(nvme_tcp_async_req(req)))
186 return false; /* async events don't have a request */
188 rq = blk_mq_rq_from_pdu(req);
190 return rq_data_dir(rq) == WRITE && req->data_len &&
191 req->data_len <= nvme_tcp_inline_data_size(req->queue);
194 static inline struct page *nvme_tcp_req_cur_page(struct nvme_tcp_request *req)
196 return req->iter.bvec->bv_page;
199 static inline size_t nvme_tcp_req_cur_offset(struct nvme_tcp_request *req)
201 return req->iter.bvec->bv_offset + req->iter.iov_offset;
204 static inline size_t nvme_tcp_req_cur_length(struct nvme_tcp_request *req)
206 return min_t(size_t, iov_iter_single_seg_count(&req->iter),
207 req->pdu_len - req->pdu_sent);
210 static inline size_t nvme_tcp_pdu_data_left(struct nvme_tcp_request *req)
212 return rq_data_dir(blk_mq_rq_from_pdu(req)) == WRITE ?
213 req->pdu_len - req->pdu_sent : 0;
216 static inline size_t nvme_tcp_pdu_last_send(struct nvme_tcp_request *req,
219 return nvme_tcp_pdu_data_left(req) <= len;
222 static void nvme_tcp_init_iter(struct nvme_tcp_request *req,
225 struct request *rq = blk_mq_rq_from_pdu(req);
231 if (rq->rq_flags & RQF_SPECIAL_PAYLOAD) {
232 vec = &rq->special_vec;
234 size = blk_rq_payload_bytes(rq);
237 struct bio *bio = req->curr_bio;
241 vec = __bvec_iter_bvec(bio->bi_io_vec, bio->bi_iter);
243 bio_for_each_bvec(bv, bio, bi) {
246 size = bio->bi_iter.bi_size;
247 offset = bio->bi_iter.bi_bvec_done;
250 iov_iter_bvec(&req->iter, dir, vec, nr_bvec, size);
251 req->iter.iov_offset = offset;
254 static inline void nvme_tcp_advance_req(struct nvme_tcp_request *req,
257 req->data_sent += len;
258 req->pdu_sent += len;
259 iov_iter_advance(&req->iter, len);
260 if (!iov_iter_count(&req->iter) &&
261 req->data_sent < req->data_len) {
262 req->curr_bio = req->curr_bio->bi_next;
263 nvme_tcp_init_iter(req, WRITE);
267 static inline void nvme_tcp_send_all(struct nvme_tcp_queue *queue)
271 /* drain the send queue as much as we can... */
273 ret = nvme_tcp_try_send(queue);
277 static inline void nvme_tcp_queue_request(struct nvme_tcp_request *req,
278 bool sync, bool last)
280 struct nvme_tcp_queue *queue = req->queue;
283 empty = llist_add(&req->lentry, &queue->req_list) &&
284 list_empty(&queue->send_list) && !queue->request;
287 * if we're the first on the send_list and we can try to send
288 * directly, otherwise queue io_work. Also, only do that if we
289 * are on the same cpu, so we don't introduce contention.
291 if (queue->io_cpu == raw_smp_processor_id() &&
292 sync && empty && mutex_trylock(&queue->send_mutex)) {
293 queue->more_requests = !last;
294 nvme_tcp_send_all(queue);
295 queue->more_requests = false;
296 mutex_unlock(&queue->send_mutex);
298 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
302 static void nvme_tcp_process_req_list(struct nvme_tcp_queue *queue)
304 struct nvme_tcp_request *req;
305 struct llist_node *node;
307 for (node = llist_del_all(&queue->req_list); node; node = node->next) {
308 req = llist_entry(node, struct nvme_tcp_request, lentry);
309 list_add(&req->entry, &queue->send_list);
313 static inline struct nvme_tcp_request *
314 nvme_tcp_fetch_request(struct nvme_tcp_queue *queue)
316 struct nvme_tcp_request *req;
318 req = list_first_entry_or_null(&queue->send_list,
319 struct nvme_tcp_request, entry);
321 nvme_tcp_process_req_list(queue);
322 req = list_first_entry_or_null(&queue->send_list,
323 struct nvme_tcp_request, entry);
328 list_del(&req->entry);
332 static inline void nvme_tcp_ddgst_final(struct ahash_request *hash,
335 ahash_request_set_crypt(hash, NULL, (u8 *)dgst, 0);
336 crypto_ahash_final(hash);
339 static inline void nvme_tcp_ddgst_update(struct ahash_request *hash,
340 struct page *page, off_t off, size_t len)
342 struct scatterlist sg;
344 sg_init_marker(&sg, 1);
345 sg_set_page(&sg, page, len, off);
346 ahash_request_set_crypt(hash, &sg, NULL, len);
347 crypto_ahash_update(hash);
350 static inline void nvme_tcp_hdgst(struct ahash_request *hash,
351 void *pdu, size_t len)
353 struct scatterlist sg;
355 sg_init_one(&sg, pdu, len);
356 ahash_request_set_crypt(hash, &sg, pdu + len, len);
357 crypto_ahash_digest(hash);
360 static int nvme_tcp_verify_hdgst(struct nvme_tcp_queue *queue,
361 void *pdu, size_t pdu_len)
363 struct nvme_tcp_hdr *hdr = pdu;
367 if (unlikely(!(hdr->flags & NVME_TCP_F_HDGST))) {
368 dev_err(queue->ctrl->ctrl.device,
369 "queue %d: header digest flag is cleared\n",
370 nvme_tcp_queue_id(queue));
374 recv_digest = *(__le32 *)(pdu + hdr->hlen);
375 nvme_tcp_hdgst(queue->rcv_hash, pdu, pdu_len);
376 exp_digest = *(__le32 *)(pdu + hdr->hlen);
377 if (recv_digest != exp_digest) {
378 dev_err(queue->ctrl->ctrl.device,
379 "header digest error: recv %#x expected %#x\n",
380 le32_to_cpu(recv_digest), le32_to_cpu(exp_digest));
387 static int nvme_tcp_check_ddgst(struct nvme_tcp_queue *queue, void *pdu)
389 struct nvme_tcp_hdr *hdr = pdu;
390 u8 digest_len = nvme_tcp_hdgst_len(queue);
393 len = le32_to_cpu(hdr->plen) - hdr->hlen -
394 ((hdr->flags & NVME_TCP_F_HDGST) ? digest_len : 0);
396 if (unlikely(len && !(hdr->flags & NVME_TCP_F_DDGST))) {
397 dev_err(queue->ctrl->ctrl.device,
398 "queue %d: data digest flag is cleared\n",
399 nvme_tcp_queue_id(queue));
402 crypto_ahash_init(queue->rcv_hash);
407 static void nvme_tcp_exit_request(struct blk_mq_tag_set *set,
408 struct request *rq, unsigned int hctx_idx)
410 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
412 page_frag_free(req->pdu);
415 static int nvme_tcp_init_request(struct blk_mq_tag_set *set,
416 struct request *rq, unsigned int hctx_idx,
417 unsigned int numa_node)
419 struct nvme_tcp_ctrl *ctrl = set->driver_data;
420 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
421 struct nvme_tcp_cmd_pdu *pdu;
422 int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
423 struct nvme_tcp_queue *queue = &ctrl->queues[queue_idx];
424 u8 hdgst = nvme_tcp_hdgst_len(queue);
426 req->pdu = page_frag_alloc(&queue->pf_cache,
427 sizeof(struct nvme_tcp_cmd_pdu) + hdgst,
428 GFP_KERNEL | __GFP_ZERO);
434 nvme_req(rq)->ctrl = &ctrl->ctrl;
435 nvme_req(rq)->cmd = &pdu->cmd;
440 static int nvme_tcp_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
441 unsigned int hctx_idx)
443 struct nvme_tcp_ctrl *ctrl = data;
444 struct nvme_tcp_queue *queue = &ctrl->queues[hctx_idx + 1];
446 hctx->driver_data = queue;
450 static int nvme_tcp_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
451 unsigned int hctx_idx)
453 struct nvme_tcp_ctrl *ctrl = data;
454 struct nvme_tcp_queue *queue = &ctrl->queues[0];
456 hctx->driver_data = queue;
460 static enum nvme_tcp_recv_state
461 nvme_tcp_recv_state(struct nvme_tcp_queue *queue)
463 return (queue->pdu_remaining) ? NVME_TCP_RECV_PDU :
464 (queue->ddgst_remaining) ? NVME_TCP_RECV_DDGST :
468 static void nvme_tcp_init_recv_ctx(struct nvme_tcp_queue *queue)
470 queue->pdu_remaining = sizeof(struct nvme_tcp_rsp_pdu) +
471 nvme_tcp_hdgst_len(queue);
472 queue->pdu_offset = 0;
473 queue->data_remaining = -1;
474 queue->ddgst_remaining = 0;
477 static void nvme_tcp_error_recovery(struct nvme_ctrl *ctrl)
479 if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_RESETTING))
482 dev_warn(ctrl->device, "starting error recovery\n");
483 queue_work(nvme_reset_wq, &to_tcp_ctrl(ctrl)->err_work);
486 static int nvme_tcp_process_nvme_cqe(struct nvme_tcp_queue *queue,
487 struct nvme_completion *cqe)
489 struct nvme_tcp_request *req;
492 rq = nvme_find_rq(nvme_tcp_tagset(queue), cqe->command_id);
494 dev_err(queue->ctrl->ctrl.device,
495 "got bad cqe.command_id %#x on queue %d\n",
496 cqe->command_id, nvme_tcp_queue_id(queue));
497 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
501 req = blk_mq_rq_to_pdu(rq);
502 if (req->status == cpu_to_le16(NVME_SC_SUCCESS))
503 req->status = cqe->status;
505 if (!nvme_try_complete_req(rq, req->status, cqe->result))
506 nvme_complete_rq(rq);
512 static int nvme_tcp_handle_c2h_data(struct nvme_tcp_queue *queue,
513 struct nvme_tcp_data_pdu *pdu)
517 rq = nvme_find_rq(nvme_tcp_tagset(queue), pdu->command_id);
519 dev_err(queue->ctrl->ctrl.device,
520 "got bad c2hdata.command_id %#x on queue %d\n",
521 pdu->command_id, nvme_tcp_queue_id(queue));
525 if (!blk_rq_payload_bytes(rq)) {
526 dev_err(queue->ctrl->ctrl.device,
527 "queue %d tag %#x unexpected data\n",
528 nvme_tcp_queue_id(queue), rq->tag);
532 queue->data_remaining = le32_to_cpu(pdu->data_length);
534 if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS &&
535 unlikely(!(pdu->hdr.flags & NVME_TCP_F_DATA_LAST))) {
536 dev_err(queue->ctrl->ctrl.device,
537 "queue %d tag %#x SUCCESS set but not last PDU\n",
538 nvme_tcp_queue_id(queue), rq->tag);
539 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
546 static int nvme_tcp_handle_comp(struct nvme_tcp_queue *queue,
547 struct nvme_tcp_rsp_pdu *pdu)
549 struct nvme_completion *cqe = &pdu->cqe;
553 * AEN requests are special as they don't time out and can
554 * survive any kind of queue freeze and often don't respond to
555 * aborts. We don't even bother to allocate a struct request
556 * for them but rather special case them here.
558 if (unlikely(nvme_is_aen_req(nvme_tcp_queue_id(queue),
560 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
563 ret = nvme_tcp_process_nvme_cqe(queue, cqe);
568 static int nvme_tcp_setup_h2c_data_pdu(struct nvme_tcp_request *req,
569 struct nvme_tcp_r2t_pdu *pdu)
571 struct nvme_tcp_data_pdu *data = req->pdu;
572 struct nvme_tcp_queue *queue = req->queue;
573 struct request *rq = blk_mq_rq_from_pdu(req);
574 u8 hdgst = nvme_tcp_hdgst_len(queue);
575 u8 ddgst = nvme_tcp_ddgst_len(queue);
577 req->pdu_len = le32_to_cpu(pdu->r2t_length);
580 if (unlikely(!req->pdu_len)) {
581 dev_err(queue->ctrl->ctrl.device,
582 "req %d r2t len is %u, probably a bug...\n",
583 rq->tag, req->pdu_len);
587 if (unlikely(req->data_sent + req->pdu_len > req->data_len)) {
588 dev_err(queue->ctrl->ctrl.device,
589 "req %d r2t len %u exceeded data len %u (%zu sent)\n",
590 rq->tag, req->pdu_len, req->data_len,
595 if (unlikely(le32_to_cpu(pdu->r2t_offset) < req->data_sent)) {
596 dev_err(queue->ctrl->ctrl.device,
597 "req %d unexpected r2t offset %u (expected %zu)\n",
598 rq->tag, le32_to_cpu(pdu->r2t_offset),
603 memset(data, 0, sizeof(*data));
604 data->hdr.type = nvme_tcp_h2c_data;
605 data->hdr.flags = NVME_TCP_F_DATA_LAST;
606 if (queue->hdr_digest)
607 data->hdr.flags |= NVME_TCP_F_HDGST;
608 if (queue->data_digest)
609 data->hdr.flags |= NVME_TCP_F_DDGST;
610 data->hdr.hlen = sizeof(*data);
611 data->hdr.pdo = data->hdr.hlen + hdgst;
613 cpu_to_le32(data->hdr.hlen + hdgst + req->pdu_len + ddgst);
614 data->ttag = pdu->ttag;
615 data->command_id = nvme_cid(rq);
616 data->data_offset = cpu_to_le32(req->data_sent);
617 data->data_length = cpu_to_le32(req->pdu_len);
621 static int nvme_tcp_handle_r2t(struct nvme_tcp_queue *queue,
622 struct nvme_tcp_r2t_pdu *pdu)
624 struct nvme_tcp_request *req;
628 rq = nvme_find_rq(nvme_tcp_tagset(queue), pdu->command_id);
630 dev_err(queue->ctrl->ctrl.device,
631 "got bad r2t.command_id %#x on queue %d\n",
632 pdu->command_id, nvme_tcp_queue_id(queue));
635 req = blk_mq_rq_to_pdu(rq);
637 ret = nvme_tcp_setup_h2c_data_pdu(req, pdu);
641 req->state = NVME_TCP_SEND_H2C_PDU;
644 nvme_tcp_queue_request(req, false, true);
649 static int nvme_tcp_recv_pdu(struct nvme_tcp_queue *queue, struct sk_buff *skb,
650 unsigned int *offset, size_t *len)
652 struct nvme_tcp_hdr *hdr;
653 char *pdu = queue->pdu;
654 size_t rcv_len = min_t(size_t, *len, queue->pdu_remaining);
657 ret = skb_copy_bits(skb, *offset,
658 &pdu[queue->pdu_offset], rcv_len);
662 queue->pdu_remaining -= rcv_len;
663 queue->pdu_offset += rcv_len;
666 if (queue->pdu_remaining)
670 if (queue->hdr_digest) {
671 ret = nvme_tcp_verify_hdgst(queue, queue->pdu, hdr->hlen);
677 if (queue->data_digest) {
678 ret = nvme_tcp_check_ddgst(queue, queue->pdu);
684 case nvme_tcp_c2h_data:
685 return nvme_tcp_handle_c2h_data(queue, (void *)queue->pdu);
687 nvme_tcp_init_recv_ctx(queue);
688 return nvme_tcp_handle_comp(queue, (void *)queue->pdu);
690 nvme_tcp_init_recv_ctx(queue);
691 return nvme_tcp_handle_r2t(queue, (void *)queue->pdu);
693 dev_err(queue->ctrl->ctrl.device,
694 "unsupported pdu type (%d)\n", hdr->type);
699 static inline void nvme_tcp_end_request(struct request *rq, u16 status)
701 union nvme_result res = {};
703 if (!nvme_try_complete_req(rq, cpu_to_le16(status << 1), res))
704 nvme_complete_rq(rq);
707 static int nvme_tcp_recv_data(struct nvme_tcp_queue *queue, struct sk_buff *skb,
708 unsigned int *offset, size_t *len)
710 struct nvme_tcp_data_pdu *pdu = (void *)queue->pdu;
712 nvme_cid_to_rq(nvme_tcp_tagset(queue), pdu->command_id);
713 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
718 recv_len = min_t(size_t, *len, queue->data_remaining);
722 if (!iov_iter_count(&req->iter)) {
723 req->curr_bio = req->curr_bio->bi_next;
726 * If we don`t have any bios it means that controller
727 * sent more data than we requested, hence error
729 if (!req->curr_bio) {
730 dev_err(queue->ctrl->ctrl.device,
731 "queue %d no space in request %#x",
732 nvme_tcp_queue_id(queue), rq->tag);
733 nvme_tcp_init_recv_ctx(queue);
736 nvme_tcp_init_iter(req, READ);
739 /* we can read only from what is left in this bio */
740 recv_len = min_t(size_t, recv_len,
741 iov_iter_count(&req->iter));
743 if (queue->data_digest)
744 ret = skb_copy_and_hash_datagram_iter(skb, *offset,
745 &req->iter, recv_len, queue->rcv_hash);
747 ret = skb_copy_datagram_iter(skb, *offset,
748 &req->iter, recv_len);
750 dev_err(queue->ctrl->ctrl.device,
751 "queue %d failed to copy request %#x data",
752 nvme_tcp_queue_id(queue), rq->tag);
758 queue->data_remaining -= recv_len;
761 if (!queue->data_remaining) {
762 if (queue->data_digest) {
763 nvme_tcp_ddgst_final(queue->rcv_hash, &queue->exp_ddgst);
764 queue->ddgst_remaining = NVME_TCP_DIGEST_LENGTH;
766 if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
767 nvme_tcp_end_request(rq,
768 le16_to_cpu(req->status));
771 nvme_tcp_init_recv_ctx(queue);
778 static int nvme_tcp_recv_ddgst(struct nvme_tcp_queue *queue,
779 struct sk_buff *skb, unsigned int *offset, size_t *len)
781 struct nvme_tcp_data_pdu *pdu = (void *)queue->pdu;
782 char *ddgst = (char *)&queue->recv_ddgst;
783 size_t recv_len = min_t(size_t, *len, queue->ddgst_remaining);
784 off_t off = NVME_TCP_DIGEST_LENGTH - queue->ddgst_remaining;
787 ret = skb_copy_bits(skb, *offset, &ddgst[off], recv_len);
791 queue->ddgst_remaining -= recv_len;
794 if (queue->ddgst_remaining)
797 if (queue->recv_ddgst != queue->exp_ddgst) {
798 struct request *rq = nvme_cid_to_rq(nvme_tcp_tagset(queue),
800 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
802 req->status = cpu_to_le16(NVME_SC_DATA_XFER_ERROR);
804 dev_err(queue->ctrl->ctrl.device,
805 "data digest error: recv %#x expected %#x\n",
806 le32_to_cpu(queue->recv_ddgst),
807 le32_to_cpu(queue->exp_ddgst));
810 if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
811 struct request *rq = nvme_cid_to_rq(nvme_tcp_tagset(queue),
813 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
815 nvme_tcp_end_request(rq, le16_to_cpu(req->status));
819 nvme_tcp_init_recv_ctx(queue);
823 static int nvme_tcp_recv_skb(read_descriptor_t *desc, struct sk_buff *skb,
824 unsigned int offset, size_t len)
826 struct nvme_tcp_queue *queue = desc->arg.data;
827 size_t consumed = len;
831 switch (nvme_tcp_recv_state(queue)) {
832 case NVME_TCP_RECV_PDU:
833 result = nvme_tcp_recv_pdu(queue, skb, &offset, &len);
835 case NVME_TCP_RECV_DATA:
836 result = nvme_tcp_recv_data(queue, skb, &offset, &len);
838 case NVME_TCP_RECV_DDGST:
839 result = nvme_tcp_recv_ddgst(queue, skb, &offset, &len);
845 dev_err(queue->ctrl->ctrl.device,
846 "receive failed: %d\n", result);
847 queue->rd_enabled = false;
848 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
856 static void nvme_tcp_data_ready(struct sock *sk)
858 struct nvme_tcp_queue *queue;
860 read_lock_bh(&sk->sk_callback_lock);
861 queue = sk->sk_user_data;
862 if (likely(queue && queue->rd_enabled) &&
863 !test_bit(NVME_TCP_Q_POLLING, &queue->flags))
864 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
865 read_unlock_bh(&sk->sk_callback_lock);
868 static void nvme_tcp_write_space(struct sock *sk)
870 struct nvme_tcp_queue *queue;
872 read_lock_bh(&sk->sk_callback_lock);
873 queue = sk->sk_user_data;
874 if (likely(queue && sk_stream_is_writeable(sk))) {
875 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
876 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
878 read_unlock_bh(&sk->sk_callback_lock);
881 static void nvme_tcp_state_change(struct sock *sk)
883 struct nvme_tcp_queue *queue;
885 read_lock_bh(&sk->sk_callback_lock);
886 queue = sk->sk_user_data;
890 switch (sk->sk_state) {
896 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
899 dev_info(queue->ctrl->ctrl.device,
900 "queue %d socket state %d\n",
901 nvme_tcp_queue_id(queue), sk->sk_state);
904 queue->state_change(sk);
906 read_unlock_bh(&sk->sk_callback_lock);
909 static inline bool nvme_tcp_queue_more(struct nvme_tcp_queue *queue)
911 return !list_empty(&queue->send_list) ||
912 !llist_empty(&queue->req_list) || queue->more_requests;
915 static inline void nvme_tcp_done_send_req(struct nvme_tcp_queue *queue)
917 queue->request = NULL;
920 static void nvme_tcp_fail_request(struct nvme_tcp_request *req)
922 nvme_tcp_end_request(blk_mq_rq_from_pdu(req), NVME_SC_HOST_PATH_ERROR);
925 static int nvme_tcp_try_send_data(struct nvme_tcp_request *req)
927 struct nvme_tcp_queue *queue = req->queue;
930 struct page *page = nvme_tcp_req_cur_page(req);
931 size_t offset = nvme_tcp_req_cur_offset(req);
932 size_t len = nvme_tcp_req_cur_length(req);
933 bool last = nvme_tcp_pdu_last_send(req, len);
934 int ret, flags = MSG_DONTWAIT;
936 if (last && !queue->data_digest && !nvme_tcp_queue_more(queue))
939 flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST;
941 if (sendpage_ok(page)) {
942 ret = kernel_sendpage(queue->sock, page, offset, len,
945 ret = sock_no_sendpage(queue->sock, page, offset, len,
951 if (queue->data_digest)
952 nvme_tcp_ddgst_update(queue->snd_hash, page,
955 /* fully successful last write*/
956 if (last && ret == len) {
957 if (queue->data_digest) {
958 nvme_tcp_ddgst_final(queue->snd_hash,
960 req->state = NVME_TCP_SEND_DDGST;
963 nvme_tcp_done_send_req(queue);
967 nvme_tcp_advance_req(req, ret);
972 static int nvme_tcp_try_send_cmd_pdu(struct nvme_tcp_request *req)
974 struct nvme_tcp_queue *queue = req->queue;
975 struct nvme_tcp_cmd_pdu *pdu = req->pdu;
976 bool inline_data = nvme_tcp_has_inline_data(req);
977 u8 hdgst = nvme_tcp_hdgst_len(queue);
978 int len = sizeof(*pdu) + hdgst - req->offset;
979 int flags = MSG_DONTWAIT;
982 if (inline_data || nvme_tcp_queue_more(queue))
983 flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST;
987 if (queue->hdr_digest && !req->offset)
988 nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
990 ret = kernel_sendpage(queue->sock, virt_to_page(pdu),
991 offset_in_page(pdu) + req->offset, len, flags);
992 if (unlikely(ret <= 0))
998 req->state = NVME_TCP_SEND_DATA;
999 if (queue->data_digest)
1000 crypto_ahash_init(queue->snd_hash);
1002 nvme_tcp_done_send_req(queue);
1011 static int nvme_tcp_try_send_data_pdu(struct nvme_tcp_request *req)
1013 struct nvme_tcp_queue *queue = req->queue;
1014 struct nvme_tcp_data_pdu *pdu = req->pdu;
1015 u8 hdgst = nvme_tcp_hdgst_len(queue);
1016 int len = sizeof(*pdu) - req->offset + hdgst;
1019 if (queue->hdr_digest && !req->offset)
1020 nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
1022 ret = kernel_sendpage(queue->sock, virt_to_page(pdu),
1023 offset_in_page(pdu) + req->offset, len,
1024 MSG_DONTWAIT | MSG_MORE | MSG_SENDPAGE_NOTLAST);
1025 if (unlikely(ret <= 0))
1030 req->state = NVME_TCP_SEND_DATA;
1031 if (queue->data_digest)
1032 crypto_ahash_init(queue->snd_hash);
1040 static int nvme_tcp_try_send_ddgst(struct nvme_tcp_request *req)
1042 struct nvme_tcp_queue *queue = req->queue;
1044 struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
1046 .iov_base = &req->ddgst + req->offset,
1047 .iov_len = NVME_TCP_DIGEST_LENGTH - req->offset
1050 if (nvme_tcp_queue_more(queue))
1051 msg.msg_flags |= MSG_MORE;
1053 msg.msg_flags |= MSG_EOR;
1055 ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
1056 if (unlikely(ret <= 0))
1059 if (req->offset + ret == NVME_TCP_DIGEST_LENGTH) {
1060 nvme_tcp_done_send_req(queue);
1068 static int nvme_tcp_try_send(struct nvme_tcp_queue *queue)
1070 struct nvme_tcp_request *req;
1073 if (!queue->request) {
1074 queue->request = nvme_tcp_fetch_request(queue);
1075 if (!queue->request)
1078 req = queue->request;
1080 if (req->state == NVME_TCP_SEND_CMD_PDU) {
1081 ret = nvme_tcp_try_send_cmd_pdu(req);
1084 if (!nvme_tcp_has_inline_data(req))
1088 if (req->state == NVME_TCP_SEND_H2C_PDU) {
1089 ret = nvme_tcp_try_send_data_pdu(req);
1094 if (req->state == NVME_TCP_SEND_DATA) {
1095 ret = nvme_tcp_try_send_data(req);
1100 if (req->state == NVME_TCP_SEND_DDGST)
1101 ret = nvme_tcp_try_send_ddgst(req);
1103 if (ret == -EAGAIN) {
1105 } else if (ret < 0) {
1106 dev_err(queue->ctrl->ctrl.device,
1107 "failed to send request %d\n", ret);
1108 if (ret != -EPIPE && ret != -ECONNRESET)
1109 nvme_tcp_fail_request(queue->request);
1110 nvme_tcp_done_send_req(queue);
1115 static int nvme_tcp_try_recv(struct nvme_tcp_queue *queue)
1117 struct socket *sock = queue->sock;
1118 struct sock *sk = sock->sk;
1119 read_descriptor_t rd_desc;
1122 rd_desc.arg.data = queue;
1126 consumed = sock->ops->read_sock(sk, &rd_desc, nvme_tcp_recv_skb);
1131 static void nvme_tcp_io_work(struct work_struct *w)
1133 struct nvme_tcp_queue *queue =
1134 container_of(w, struct nvme_tcp_queue, io_work);
1135 unsigned long deadline = jiffies + msecs_to_jiffies(1);
1138 bool pending = false;
1141 if (mutex_trylock(&queue->send_mutex)) {
1142 result = nvme_tcp_try_send(queue);
1143 mutex_unlock(&queue->send_mutex);
1146 else if (unlikely(result < 0))
1149 pending = !llist_empty(&queue->req_list);
1151 result = nvme_tcp_try_recv(queue);
1154 else if (unlikely(result < 0))
1160 } while (!time_after(jiffies, deadline)); /* quota is exhausted */
1162 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
1165 static void nvme_tcp_free_crypto(struct nvme_tcp_queue *queue)
1167 struct crypto_ahash *tfm = crypto_ahash_reqtfm(queue->rcv_hash);
1169 ahash_request_free(queue->rcv_hash);
1170 ahash_request_free(queue->snd_hash);
1171 crypto_free_ahash(tfm);
1174 static int nvme_tcp_alloc_crypto(struct nvme_tcp_queue *queue)
1176 struct crypto_ahash *tfm;
1178 tfm = crypto_alloc_ahash("crc32c", 0, CRYPTO_ALG_ASYNC);
1180 return PTR_ERR(tfm);
1182 queue->snd_hash = ahash_request_alloc(tfm, GFP_KERNEL);
1183 if (!queue->snd_hash)
1185 ahash_request_set_callback(queue->snd_hash, 0, NULL, NULL);
1187 queue->rcv_hash = ahash_request_alloc(tfm, GFP_KERNEL);
1188 if (!queue->rcv_hash)
1190 ahash_request_set_callback(queue->rcv_hash, 0, NULL, NULL);
1194 ahash_request_free(queue->snd_hash);
1196 crypto_free_ahash(tfm);
1200 static void nvme_tcp_free_async_req(struct nvme_tcp_ctrl *ctrl)
1202 struct nvme_tcp_request *async = &ctrl->async_req;
1204 page_frag_free(async->pdu);
1207 static int nvme_tcp_alloc_async_req(struct nvme_tcp_ctrl *ctrl)
1209 struct nvme_tcp_queue *queue = &ctrl->queues[0];
1210 struct nvme_tcp_request *async = &ctrl->async_req;
1211 u8 hdgst = nvme_tcp_hdgst_len(queue);
1213 async->pdu = page_frag_alloc(&queue->pf_cache,
1214 sizeof(struct nvme_tcp_cmd_pdu) + hdgst,
1215 GFP_KERNEL | __GFP_ZERO);
1219 async->queue = &ctrl->queues[0];
1223 static void nvme_tcp_free_queue(struct nvme_ctrl *nctrl, int qid)
1225 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1226 struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1228 if (!test_and_clear_bit(NVME_TCP_Q_ALLOCATED, &queue->flags))
1231 if (queue->hdr_digest || queue->data_digest)
1232 nvme_tcp_free_crypto(queue);
1234 sock_release(queue->sock);
1236 mutex_destroy(&queue->send_mutex);
1237 mutex_destroy(&queue->queue_lock);
1240 static int nvme_tcp_init_connection(struct nvme_tcp_queue *queue)
1242 struct nvme_tcp_icreq_pdu *icreq;
1243 struct nvme_tcp_icresp_pdu *icresp;
1244 struct msghdr msg = {};
1246 bool ctrl_hdgst, ctrl_ddgst;
1249 icreq = kzalloc(sizeof(*icreq), GFP_KERNEL);
1253 icresp = kzalloc(sizeof(*icresp), GFP_KERNEL);
1259 icreq->hdr.type = nvme_tcp_icreq;
1260 icreq->hdr.hlen = sizeof(*icreq);
1262 icreq->hdr.plen = cpu_to_le32(icreq->hdr.hlen);
1263 icreq->pfv = cpu_to_le16(NVME_TCP_PFV_1_0);
1264 icreq->maxr2t = 0; /* single inflight r2t supported */
1265 icreq->hpda = 0; /* no alignment constraint */
1266 if (queue->hdr_digest)
1267 icreq->digest |= NVME_TCP_HDR_DIGEST_ENABLE;
1268 if (queue->data_digest)
1269 icreq->digest |= NVME_TCP_DATA_DIGEST_ENABLE;
1271 iov.iov_base = icreq;
1272 iov.iov_len = sizeof(*icreq);
1273 ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
1277 memset(&msg, 0, sizeof(msg));
1278 iov.iov_base = icresp;
1279 iov.iov_len = sizeof(*icresp);
1280 ret = kernel_recvmsg(queue->sock, &msg, &iov, 1,
1281 iov.iov_len, msg.msg_flags);
1286 if (icresp->hdr.type != nvme_tcp_icresp) {
1287 pr_err("queue %d: bad type returned %d\n",
1288 nvme_tcp_queue_id(queue), icresp->hdr.type);
1292 if (le32_to_cpu(icresp->hdr.plen) != sizeof(*icresp)) {
1293 pr_err("queue %d: bad pdu length returned %d\n",
1294 nvme_tcp_queue_id(queue), icresp->hdr.plen);
1298 if (icresp->pfv != NVME_TCP_PFV_1_0) {
1299 pr_err("queue %d: bad pfv returned %d\n",
1300 nvme_tcp_queue_id(queue), icresp->pfv);
1304 ctrl_ddgst = !!(icresp->digest & NVME_TCP_DATA_DIGEST_ENABLE);
1305 if ((queue->data_digest && !ctrl_ddgst) ||
1306 (!queue->data_digest && ctrl_ddgst)) {
1307 pr_err("queue %d: data digest mismatch host: %s ctrl: %s\n",
1308 nvme_tcp_queue_id(queue),
1309 queue->data_digest ? "enabled" : "disabled",
1310 ctrl_ddgst ? "enabled" : "disabled");
1314 ctrl_hdgst = !!(icresp->digest & NVME_TCP_HDR_DIGEST_ENABLE);
1315 if ((queue->hdr_digest && !ctrl_hdgst) ||
1316 (!queue->hdr_digest && ctrl_hdgst)) {
1317 pr_err("queue %d: header digest mismatch host: %s ctrl: %s\n",
1318 nvme_tcp_queue_id(queue),
1319 queue->hdr_digest ? "enabled" : "disabled",
1320 ctrl_hdgst ? "enabled" : "disabled");
1324 if (icresp->cpda != 0) {
1325 pr_err("queue %d: unsupported cpda returned %d\n",
1326 nvme_tcp_queue_id(queue), icresp->cpda);
1338 static bool nvme_tcp_admin_queue(struct nvme_tcp_queue *queue)
1340 return nvme_tcp_queue_id(queue) == 0;
1343 static bool nvme_tcp_default_queue(struct nvme_tcp_queue *queue)
1345 struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1346 int qid = nvme_tcp_queue_id(queue);
1348 return !nvme_tcp_admin_queue(queue) &&
1349 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT];
1352 static bool nvme_tcp_read_queue(struct nvme_tcp_queue *queue)
1354 struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1355 int qid = nvme_tcp_queue_id(queue);
1357 return !nvme_tcp_admin_queue(queue) &&
1358 !nvme_tcp_default_queue(queue) &&
1359 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT] +
1360 ctrl->io_queues[HCTX_TYPE_READ];
1363 static bool nvme_tcp_poll_queue(struct nvme_tcp_queue *queue)
1365 struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1366 int qid = nvme_tcp_queue_id(queue);
1368 return !nvme_tcp_admin_queue(queue) &&
1369 !nvme_tcp_default_queue(queue) &&
1370 !nvme_tcp_read_queue(queue) &&
1371 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT] +
1372 ctrl->io_queues[HCTX_TYPE_READ] +
1373 ctrl->io_queues[HCTX_TYPE_POLL];
1376 static void nvme_tcp_set_queue_io_cpu(struct nvme_tcp_queue *queue)
1378 struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1379 int qid = nvme_tcp_queue_id(queue);
1382 if (nvme_tcp_default_queue(queue))
1384 else if (nvme_tcp_read_queue(queue))
1385 n = qid - ctrl->io_queues[HCTX_TYPE_DEFAULT] - 1;
1386 else if (nvme_tcp_poll_queue(queue))
1387 n = qid - ctrl->io_queues[HCTX_TYPE_DEFAULT] -
1388 ctrl->io_queues[HCTX_TYPE_READ] - 1;
1389 queue->io_cpu = cpumask_next_wrap(n - 1, cpu_online_mask, -1, false);
1392 static int nvme_tcp_alloc_queue(struct nvme_ctrl *nctrl,
1393 int qid, size_t queue_size)
1395 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1396 struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1397 int ret, rcv_pdu_size;
1399 mutex_init(&queue->queue_lock);
1401 init_llist_head(&queue->req_list);
1402 INIT_LIST_HEAD(&queue->send_list);
1403 mutex_init(&queue->send_mutex);
1404 INIT_WORK(&queue->io_work, nvme_tcp_io_work);
1405 queue->queue_size = queue_size;
1408 queue->cmnd_capsule_len = nctrl->ioccsz * 16;
1410 queue->cmnd_capsule_len = sizeof(struct nvme_command) +
1411 NVME_TCP_ADMIN_CCSZ;
1413 ret = sock_create(ctrl->addr.ss_family, SOCK_STREAM,
1414 IPPROTO_TCP, &queue->sock);
1416 dev_err(nctrl->device,
1417 "failed to create socket: %d\n", ret);
1418 goto err_destroy_mutex;
1421 /* Single syn retry */
1422 tcp_sock_set_syncnt(queue->sock->sk, 1);
1424 /* Set TCP no delay */
1425 tcp_sock_set_nodelay(queue->sock->sk);
1428 * Cleanup whatever is sitting in the TCP transmit queue on socket
1429 * close. This is done to prevent stale data from being sent should
1430 * the network connection be restored before TCP times out.
1432 sock_no_linger(queue->sock->sk);
1434 if (so_priority > 0)
1435 sock_set_priority(queue->sock->sk, so_priority);
1437 /* Set socket type of service */
1438 if (nctrl->opts->tos >= 0)
1439 ip_sock_set_tos(queue->sock->sk, nctrl->opts->tos);
1441 /* Set 10 seconds timeout for icresp recvmsg */
1442 queue->sock->sk->sk_rcvtimeo = 10 * HZ;
1444 queue->sock->sk->sk_allocation = GFP_ATOMIC;
1445 nvme_tcp_set_queue_io_cpu(queue);
1446 queue->request = NULL;
1447 queue->data_remaining = 0;
1448 queue->ddgst_remaining = 0;
1449 queue->pdu_remaining = 0;
1450 queue->pdu_offset = 0;
1451 sk_set_memalloc(queue->sock->sk);
1453 if (nctrl->opts->mask & NVMF_OPT_HOST_TRADDR) {
1454 ret = kernel_bind(queue->sock, (struct sockaddr *)&ctrl->src_addr,
1455 sizeof(ctrl->src_addr));
1457 dev_err(nctrl->device,
1458 "failed to bind queue %d socket %d\n",
1464 if (nctrl->opts->mask & NVMF_OPT_HOST_IFACE) {
1465 char *iface = nctrl->opts->host_iface;
1466 sockptr_t optval = KERNEL_SOCKPTR(iface);
1468 ret = sock_setsockopt(queue->sock, SOL_SOCKET, SO_BINDTODEVICE,
1469 optval, strlen(iface));
1471 dev_err(nctrl->device,
1472 "failed to bind to interface %s queue %d err %d\n",
1478 queue->hdr_digest = nctrl->opts->hdr_digest;
1479 queue->data_digest = nctrl->opts->data_digest;
1480 if (queue->hdr_digest || queue->data_digest) {
1481 ret = nvme_tcp_alloc_crypto(queue);
1483 dev_err(nctrl->device,
1484 "failed to allocate queue %d crypto\n", qid);
1489 rcv_pdu_size = sizeof(struct nvme_tcp_rsp_pdu) +
1490 nvme_tcp_hdgst_len(queue);
1491 queue->pdu = kmalloc(rcv_pdu_size, GFP_KERNEL);
1497 dev_dbg(nctrl->device, "connecting queue %d\n",
1498 nvme_tcp_queue_id(queue));
1500 ret = kernel_connect(queue->sock, (struct sockaddr *)&ctrl->addr,
1501 sizeof(ctrl->addr), 0);
1503 dev_err(nctrl->device,
1504 "failed to connect socket: %d\n", ret);
1508 ret = nvme_tcp_init_connection(queue);
1510 goto err_init_connect;
1512 queue->rd_enabled = true;
1513 set_bit(NVME_TCP_Q_ALLOCATED, &queue->flags);
1514 nvme_tcp_init_recv_ctx(queue);
1516 write_lock_bh(&queue->sock->sk->sk_callback_lock);
1517 queue->sock->sk->sk_user_data = queue;
1518 queue->state_change = queue->sock->sk->sk_state_change;
1519 queue->data_ready = queue->sock->sk->sk_data_ready;
1520 queue->write_space = queue->sock->sk->sk_write_space;
1521 queue->sock->sk->sk_data_ready = nvme_tcp_data_ready;
1522 queue->sock->sk->sk_state_change = nvme_tcp_state_change;
1523 queue->sock->sk->sk_write_space = nvme_tcp_write_space;
1524 #ifdef CONFIG_NET_RX_BUSY_POLL
1525 queue->sock->sk->sk_ll_usec = 1;
1527 write_unlock_bh(&queue->sock->sk->sk_callback_lock);
1532 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1536 if (queue->hdr_digest || queue->data_digest)
1537 nvme_tcp_free_crypto(queue);
1539 sock_release(queue->sock);
1542 mutex_destroy(&queue->send_mutex);
1543 mutex_destroy(&queue->queue_lock);
1547 static void nvme_tcp_restore_sock_calls(struct nvme_tcp_queue *queue)
1549 struct socket *sock = queue->sock;
1551 write_lock_bh(&sock->sk->sk_callback_lock);
1552 sock->sk->sk_user_data = NULL;
1553 sock->sk->sk_data_ready = queue->data_ready;
1554 sock->sk->sk_state_change = queue->state_change;
1555 sock->sk->sk_write_space = queue->write_space;
1556 write_unlock_bh(&sock->sk->sk_callback_lock);
1559 static void __nvme_tcp_stop_queue(struct nvme_tcp_queue *queue)
1561 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1562 nvme_tcp_restore_sock_calls(queue);
1563 cancel_work_sync(&queue->io_work);
1566 static void nvme_tcp_stop_queue(struct nvme_ctrl *nctrl, int qid)
1568 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1569 struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1571 mutex_lock(&queue->queue_lock);
1572 if (test_and_clear_bit(NVME_TCP_Q_LIVE, &queue->flags))
1573 __nvme_tcp_stop_queue(queue);
1574 mutex_unlock(&queue->queue_lock);
1577 static int nvme_tcp_start_queue(struct nvme_ctrl *nctrl, int idx)
1579 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1583 ret = nvmf_connect_io_queue(nctrl, idx);
1585 ret = nvmf_connect_admin_queue(nctrl);
1588 set_bit(NVME_TCP_Q_LIVE, &ctrl->queues[idx].flags);
1590 if (test_bit(NVME_TCP_Q_ALLOCATED, &ctrl->queues[idx].flags))
1591 __nvme_tcp_stop_queue(&ctrl->queues[idx]);
1592 dev_err(nctrl->device,
1593 "failed to connect queue: %d ret=%d\n", idx, ret);
1598 static struct blk_mq_tag_set *nvme_tcp_alloc_tagset(struct nvme_ctrl *nctrl,
1601 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1602 struct blk_mq_tag_set *set;
1606 set = &ctrl->admin_tag_set;
1607 memset(set, 0, sizeof(*set));
1608 set->ops = &nvme_tcp_admin_mq_ops;
1609 set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
1610 set->reserved_tags = NVMF_RESERVED_TAGS;
1611 set->numa_node = nctrl->numa_node;
1612 set->flags = BLK_MQ_F_BLOCKING;
1613 set->cmd_size = sizeof(struct nvme_tcp_request);
1614 set->driver_data = ctrl;
1615 set->nr_hw_queues = 1;
1616 set->timeout = NVME_ADMIN_TIMEOUT;
1618 set = &ctrl->tag_set;
1619 memset(set, 0, sizeof(*set));
1620 set->ops = &nvme_tcp_mq_ops;
1621 set->queue_depth = nctrl->sqsize + 1;
1622 set->reserved_tags = NVMF_RESERVED_TAGS;
1623 set->numa_node = nctrl->numa_node;
1624 set->flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_BLOCKING;
1625 set->cmd_size = sizeof(struct nvme_tcp_request);
1626 set->driver_data = ctrl;
1627 set->nr_hw_queues = nctrl->queue_count - 1;
1628 set->timeout = NVME_IO_TIMEOUT;
1629 set->nr_maps = nctrl->opts->nr_poll_queues ? HCTX_MAX_TYPES : 2;
1632 ret = blk_mq_alloc_tag_set(set);
1634 return ERR_PTR(ret);
1639 static void nvme_tcp_free_admin_queue(struct nvme_ctrl *ctrl)
1641 if (to_tcp_ctrl(ctrl)->async_req.pdu) {
1642 cancel_work_sync(&ctrl->async_event_work);
1643 nvme_tcp_free_async_req(to_tcp_ctrl(ctrl));
1644 to_tcp_ctrl(ctrl)->async_req.pdu = NULL;
1647 nvme_tcp_free_queue(ctrl, 0);
1650 static void nvme_tcp_free_io_queues(struct nvme_ctrl *ctrl)
1654 for (i = 1; i < ctrl->queue_count; i++)
1655 nvme_tcp_free_queue(ctrl, i);
1658 static void nvme_tcp_stop_io_queues(struct nvme_ctrl *ctrl)
1662 for (i = 1; i < ctrl->queue_count; i++)
1663 nvme_tcp_stop_queue(ctrl, i);
1666 static int nvme_tcp_start_io_queues(struct nvme_ctrl *ctrl)
1670 for (i = 1; i < ctrl->queue_count; i++) {
1671 ret = nvme_tcp_start_queue(ctrl, i);
1673 goto out_stop_queues;
1679 for (i--; i >= 1; i--)
1680 nvme_tcp_stop_queue(ctrl, i);
1684 static int nvme_tcp_alloc_admin_queue(struct nvme_ctrl *ctrl)
1688 ret = nvme_tcp_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
1692 ret = nvme_tcp_alloc_async_req(to_tcp_ctrl(ctrl));
1694 goto out_free_queue;
1699 nvme_tcp_free_queue(ctrl, 0);
1703 static int __nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
1707 for (i = 1; i < ctrl->queue_count; i++) {
1708 ret = nvme_tcp_alloc_queue(ctrl, i,
1711 goto out_free_queues;
1717 for (i--; i >= 1; i--)
1718 nvme_tcp_free_queue(ctrl, i);
1723 static unsigned int nvme_tcp_nr_io_queues(struct nvme_ctrl *ctrl)
1725 unsigned int nr_io_queues;
1727 nr_io_queues = min(ctrl->opts->nr_io_queues, num_online_cpus());
1728 nr_io_queues += min(ctrl->opts->nr_write_queues, num_online_cpus());
1729 nr_io_queues += min(ctrl->opts->nr_poll_queues, num_online_cpus());
1731 return nr_io_queues;
1734 static void nvme_tcp_set_io_queues(struct nvme_ctrl *nctrl,
1735 unsigned int nr_io_queues)
1737 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1738 struct nvmf_ctrl_options *opts = nctrl->opts;
1740 if (opts->nr_write_queues && opts->nr_io_queues < nr_io_queues) {
1742 * separate read/write queues
1743 * hand out dedicated default queues only after we have
1744 * sufficient read queues.
1746 ctrl->io_queues[HCTX_TYPE_READ] = opts->nr_io_queues;
1747 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_READ];
1748 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
1749 min(opts->nr_write_queues, nr_io_queues);
1750 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
1753 * shared read/write queues
1754 * either no write queues were requested, or we don't have
1755 * sufficient queue count to have dedicated default queues.
1757 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
1758 min(opts->nr_io_queues, nr_io_queues);
1759 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
1762 if (opts->nr_poll_queues && nr_io_queues) {
1763 /* map dedicated poll queues only if we have queues left */
1764 ctrl->io_queues[HCTX_TYPE_POLL] =
1765 min(opts->nr_poll_queues, nr_io_queues);
1769 static int nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
1771 unsigned int nr_io_queues;
1774 nr_io_queues = nvme_tcp_nr_io_queues(ctrl);
1775 ret = nvme_set_queue_count(ctrl, &nr_io_queues);
1779 if (nr_io_queues == 0) {
1780 dev_err(ctrl->device,
1781 "unable to set any I/O queues\n");
1785 ctrl->queue_count = nr_io_queues + 1;
1786 dev_info(ctrl->device,
1787 "creating %d I/O queues.\n", nr_io_queues);
1789 nvme_tcp_set_io_queues(ctrl, nr_io_queues);
1791 return __nvme_tcp_alloc_io_queues(ctrl);
1794 static void nvme_tcp_destroy_io_queues(struct nvme_ctrl *ctrl, bool remove)
1796 nvme_tcp_stop_io_queues(ctrl);
1798 blk_cleanup_queue(ctrl->connect_q);
1799 blk_mq_free_tag_set(ctrl->tagset);
1801 nvme_tcp_free_io_queues(ctrl);
1804 static int nvme_tcp_configure_io_queues(struct nvme_ctrl *ctrl, bool new)
1808 ret = nvme_tcp_alloc_io_queues(ctrl);
1813 ctrl->tagset = nvme_tcp_alloc_tagset(ctrl, false);
1814 if (IS_ERR(ctrl->tagset)) {
1815 ret = PTR_ERR(ctrl->tagset);
1816 goto out_free_io_queues;
1819 ctrl->connect_q = blk_mq_init_queue(ctrl->tagset);
1820 if (IS_ERR(ctrl->connect_q)) {
1821 ret = PTR_ERR(ctrl->connect_q);
1822 goto out_free_tag_set;
1826 ret = nvme_tcp_start_io_queues(ctrl);
1828 goto out_cleanup_connect_q;
1831 nvme_start_queues(ctrl);
1832 if (!nvme_wait_freeze_timeout(ctrl, NVME_IO_TIMEOUT)) {
1834 * If we timed out waiting for freeze we are likely to
1835 * be stuck. Fail the controller initialization just
1839 goto out_wait_freeze_timed_out;
1841 blk_mq_update_nr_hw_queues(ctrl->tagset,
1842 ctrl->queue_count - 1);
1843 nvme_unfreeze(ctrl);
1848 out_wait_freeze_timed_out:
1849 nvme_stop_queues(ctrl);
1850 nvme_sync_io_queues(ctrl);
1851 nvme_tcp_stop_io_queues(ctrl);
1852 out_cleanup_connect_q:
1853 nvme_cancel_tagset(ctrl);
1855 blk_cleanup_queue(ctrl->connect_q);
1858 blk_mq_free_tag_set(ctrl->tagset);
1860 nvme_tcp_free_io_queues(ctrl);
1864 static void nvme_tcp_destroy_admin_queue(struct nvme_ctrl *ctrl, bool remove)
1866 nvme_tcp_stop_queue(ctrl, 0);
1868 blk_cleanup_queue(ctrl->admin_q);
1869 blk_cleanup_queue(ctrl->fabrics_q);
1870 blk_mq_free_tag_set(ctrl->admin_tagset);
1872 nvme_tcp_free_admin_queue(ctrl);
1875 static int nvme_tcp_configure_admin_queue(struct nvme_ctrl *ctrl, bool new)
1879 error = nvme_tcp_alloc_admin_queue(ctrl);
1884 ctrl->admin_tagset = nvme_tcp_alloc_tagset(ctrl, true);
1885 if (IS_ERR(ctrl->admin_tagset)) {
1886 error = PTR_ERR(ctrl->admin_tagset);
1887 goto out_free_queue;
1890 ctrl->fabrics_q = blk_mq_init_queue(ctrl->admin_tagset);
1891 if (IS_ERR(ctrl->fabrics_q)) {
1892 error = PTR_ERR(ctrl->fabrics_q);
1893 goto out_free_tagset;
1896 ctrl->admin_q = blk_mq_init_queue(ctrl->admin_tagset);
1897 if (IS_ERR(ctrl->admin_q)) {
1898 error = PTR_ERR(ctrl->admin_q);
1899 goto out_cleanup_fabrics_q;
1903 error = nvme_tcp_start_queue(ctrl, 0);
1905 goto out_cleanup_queue;
1907 error = nvme_enable_ctrl(ctrl);
1909 goto out_stop_queue;
1911 blk_mq_unquiesce_queue(ctrl->admin_q);
1913 error = nvme_init_ctrl_finish(ctrl);
1915 goto out_quiesce_queue;
1920 blk_mq_quiesce_queue(ctrl->admin_q);
1921 blk_sync_queue(ctrl->admin_q);
1923 nvme_tcp_stop_queue(ctrl, 0);
1924 nvme_cancel_admin_tagset(ctrl);
1927 blk_cleanup_queue(ctrl->admin_q);
1928 out_cleanup_fabrics_q:
1930 blk_cleanup_queue(ctrl->fabrics_q);
1933 blk_mq_free_tag_set(ctrl->admin_tagset);
1935 nvme_tcp_free_admin_queue(ctrl);
1939 static void nvme_tcp_teardown_admin_queue(struct nvme_ctrl *ctrl,
1942 blk_mq_quiesce_queue(ctrl->admin_q);
1943 blk_sync_queue(ctrl->admin_q);
1944 nvme_tcp_stop_queue(ctrl, 0);
1945 nvme_cancel_admin_tagset(ctrl);
1947 blk_mq_unquiesce_queue(ctrl->admin_q);
1948 nvme_tcp_destroy_admin_queue(ctrl, remove);
1951 static void nvme_tcp_teardown_io_queues(struct nvme_ctrl *ctrl,
1954 if (ctrl->queue_count <= 1)
1956 blk_mq_quiesce_queue(ctrl->admin_q);
1957 nvme_start_freeze(ctrl);
1958 nvme_stop_queues(ctrl);
1959 nvme_sync_io_queues(ctrl);
1960 nvme_tcp_stop_io_queues(ctrl);
1961 nvme_cancel_tagset(ctrl);
1963 nvme_start_queues(ctrl);
1964 nvme_tcp_destroy_io_queues(ctrl, remove);
1967 static void nvme_tcp_reconnect_or_remove(struct nvme_ctrl *ctrl)
1969 /* If we are resetting/deleting then do nothing */
1970 if (ctrl->state != NVME_CTRL_CONNECTING) {
1971 WARN_ON_ONCE(ctrl->state == NVME_CTRL_NEW ||
1972 ctrl->state == NVME_CTRL_LIVE);
1976 if (nvmf_should_reconnect(ctrl)) {
1977 dev_info(ctrl->device, "Reconnecting in %d seconds...\n",
1978 ctrl->opts->reconnect_delay);
1979 queue_delayed_work(nvme_wq, &to_tcp_ctrl(ctrl)->connect_work,
1980 ctrl->opts->reconnect_delay * HZ);
1982 dev_info(ctrl->device, "Removing controller...\n");
1983 nvme_delete_ctrl(ctrl);
1987 static int nvme_tcp_setup_ctrl(struct nvme_ctrl *ctrl, bool new)
1989 struct nvmf_ctrl_options *opts = ctrl->opts;
1992 ret = nvme_tcp_configure_admin_queue(ctrl, new);
1998 dev_err(ctrl->device, "icdoff is not supported!\n");
2002 if (!nvme_ctrl_sgl_supported(ctrl)) {
2004 dev_err(ctrl->device, "Mandatory sgls are not supported!\n");
2008 if (opts->queue_size > ctrl->sqsize + 1)
2009 dev_warn(ctrl->device,
2010 "queue_size %zu > ctrl sqsize %u, clamping down\n",
2011 opts->queue_size, ctrl->sqsize + 1);
2013 if (ctrl->sqsize + 1 > ctrl->maxcmd) {
2014 dev_warn(ctrl->device,
2015 "sqsize %u > ctrl maxcmd %u, clamping down\n",
2016 ctrl->sqsize + 1, ctrl->maxcmd);
2017 ctrl->sqsize = ctrl->maxcmd - 1;
2020 if (ctrl->queue_count > 1) {
2021 ret = nvme_tcp_configure_io_queues(ctrl, new);
2026 if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_LIVE)) {
2028 * state change failure is ok if we started ctrl delete,
2029 * unless we're during creation of a new controller to
2030 * avoid races with teardown flow.
2032 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING &&
2033 ctrl->state != NVME_CTRL_DELETING_NOIO);
2039 nvme_start_ctrl(ctrl);
2043 if (ctrl->queue_count > 1) {
2044 nvme_stop_queues(ctrl);
2045 nvme_sync_io_queues(ctrl);
2046 nvme_tcp_stop_io_queues(ctrl);
2047 nvme_cancel_tagset(ctrl);
2048 nvme_tcp_destroy_io_queues(ctrl, new);
2051 blk_mq_quiesce_queue(ctrl->admin_q);
2052 blk_sync_queue(ctrl->admin_q);
2053 nvme_tcp_stop_queue(ctrl, 0);
2054 nvme_cancel_admin_tagset(ctrl);
2055 nvme_tcp_destroy_admin_queue(ctrl, new);
2059 static void nvme_tcp_reconnect_ctrl_work(struct work_struct *work)
2061 struct nvme_tcp_ctrl *tcp_ctrl = container_of(to_delayed_work(work),
2062 struct nvme_tcp_ctrl, connect_work);
2063 struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
2065 ++ctrl->nr_reconnects;
2067 if (nvme_tcp_setup_ctrl(ctrl, false))
2070 dev_info(ctrl->device, "Successfully reconnected (%d attempt)\n",
2071 ctrl->nr_reconnects);
2073 ctrl->nr_reconnects = 0;
2078 dev_info(ctrl->device, "Failed reconnect attempt %d\n",
2079 ctrl->nr_reconnects);
2080 nvme_tcp_reconnect_or_remove(ctrl);
2083 static void nvme_tcp_error_recovery_work(struct work_struct *work)
2085 struct nvme_tcp_ctrl *tcp_ctrl = container_of(work,
2086 struct nvme_tcp_ctrl, err_work);
2087 struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
2089 nvme_stop_keep_alive(ctrl);
2090 nvme_tcp_teardown_io_queues(ctrl, false);
2091 /* unquiesce to fail fast pending requests */
2092 nvme_start_queues(ctrl);
2093 nvme_tcp_teardown_admin_queue(ctrl, false);
2094 blk_mq_unquiesce_queue(ctrl->admin_q);
2096 if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
2097 /* state change failure is ok if we started ctrl delete */
2098 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING &&
2099 ctrl->state != NVME_CTRL_DELETING_NOIO);
2103 nvme_tcp_reconnect_or_remove(ctrl);
2106 static void nvme_tcp_teardown_ctrl(struct nvme_ctrl *ctrl, bool shutdown)
2108 cancel_work_sync(&to_tcp_ctrl(ctrl)->err_work);
2109 cancel_delayed_work_sync(&to_tcp_ctrl(ctrl)->connect_work);
2111 nvme_tcp_teardown_io_queues(ctrl, shutdown);
2112 blk_mq_quiesce_queue(ctrl->admin_q);
2114 nvme_shutdown_ctrl(ctrl);
2116 nvme_disable_ctrl(ctrl);
2117 nvme_tcp_teardown_admin_queue(ctrl, shutdown);
2120 static void nvme_tcp_delete_ctrl(struct nvme_ctrl *ctrl)
2122 nvme_tcp_teardown_ctrl(ctrl, true);
2125 static void nvme_reset_ctrl_work(struct work_struct *work)
2127 struct nvme_ctrl *ctrl =
2128 container_of(work, struct nvme_ctrl, reset_work);
2130 nvme_stop_ctrl(ctrl);
2131 nvme_tcp_teardown_ctrl(ctrl, false);
2133 if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
2134 /* state change failure is ok if we started ctrl delete */
2135 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING &&
2136 ctrl->state != NVME_CTRL_DELETING_NOIO);
2140 if (nvme_tcp_setup_ctrl(ctrl, false))
2146 ++ctrl->nr_reconnects;
2147 nvme_tcp_reconnect_or_remove(ctrl);
2150 static void nvme_tcp_free_ctrl(struct nvme_ctrl *nctrl)
2152 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
2154 if (list_empty(&ctrl->list))
2157 mutex_lock(&nvme_tcp_ctrl_mutex);
2158 list_del(&ctrl->list);
2159 mutex_unlock(&nvme_tcp_ctrl_mutex);
2161 nvmf_free_options(nctrl->opts);
2163 kfree(ctrl->queues);
2167 static void nvme_tcp_set_sg_null(struct nvme_command *c)
2169 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2173 sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
2174 NVME_SGL_FMT_TRANSPORT_A;
2177 static void nvme_tcp_set_sg_inline(struct nvme_tcp_queue *queue,
2178 struct nvme_command *c, u32 data_len)
2180 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2182 sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
2183 sg->length = cpu_to_le32(data_len);
2184 sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
2187 static void nvme_tcp_set_sg_host_data(struct nvme_command *c,
2190 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2193 sg->length = cpu_to_le32(data_len);
2194 sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
2195 NVME_SGL_FMT_TRANSPORT_A;
2198 static void nvme_tcp_submit_async_event(struct nvme_ctrl *arg)
2200 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(arg);
2201 struct nvme_tcp_queue *queue = &ctrl->queues[0];
2202 struct nvme_tcp_cmd_pdu *pdu = ctrl->async_req.pdu;
2203 struct nvme_command *cmd = &pdu->cmd;
2204 u8 hdgst = nvme_tcp_hdgst_len(queue);
2206 memset(pdu, 0, sizeof(*pdu));
2207 pdu->hdr.type = nvme_tcp_cmd;
2208 if (queue->hdr_digest)
2209 pdu->hdr.flags |= NVME_TCP_F_HDGST;
2210 pdu->hdr.hlen = sizeof(*pdu);
2211 pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst);
2213 cmd->common.opcode = nvme_admin_async_event;
2214 cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
2215 cmd->common.flags |= NVME_CMD_SGL_METABUF;
2216 nvme_tcp_set_sg_null(cmd);
2218 ctrl->async_req.state = NVME_TCP_SEND_CMD_PDU;
2219 ctrl->async_req.offset = 0;
2220 ctrl->async_req.curr_bio = NULL;
2221 ctrl->async_req.data_len = 0;
2223 nvme_tcp_queue_request(&ctrl->async_req, true, true);
2226 static void nvme_tcp_complete_timed_out(struct request *rq)
2228 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2229 struct nvme_ctrl *ctrl = &req->queue->ctrl->ctrl;
2231 nvme_tcp_stop_queue(ctrl, nvme_tcp_queue_id(req->queue));
2232 if (blk_mq_request_started(rq) && !blk_mq_request_completed(rq)) {
2233 nvme_req(rq)->status = NVME_SC_HOST_ABORTED_CMD;
2234 blk_mq_complete_request(rq);
2238 static enum blk_eh_timer_return
2239 nvme_tcp_timeout(struct request *rq, bool reserved)
2241 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2242 struct nvme_ctrl *ctrl = &req->queue->ctrl->ctrl;
2243 struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2245 dev_warn(ctrl->device,
2246 "queue %d: timeout request %#x type %d\n",
2247 nvme_tcp_queue_id(req->queue), rq->tag, pdu->hdr.type);
2249 if (ctrl->state != NVME_CTRL_LIVE) {
2251 * If we are resetting, connecting or deleting we should
2252 * complete immediately because we may block controller
2253 * teardown or setup sequence
2254 * - ctrl disable/shutdown fabrics requests
2255 * - connect requests
2256 * - initialization admin requests
2257 * - I/O requests that entered after unquiescing and
2258 * the controller stopped responding
2260 * All other requests should be cancelled by the error
2261 * recovery work, so it's fine that we fail it here.
2263 nvme_tcp_complete_timed_out(rq);
2268 * LIVE state should trigger the normal error recovery which will
2269 * handle completing this request.
2271 nvme_tcp_error_recovery(ctrl);
2272 return BLK_EH_RESET_TIMER;
2275 static blk_status_t nvme_tcp_map_data(struct nvme_tcp_queue *queue,
2278 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2279 struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2280 struct nvme_command *c = &pdu->cmd;
2282 c->common.flags |= NVME_CMD_SGL_METABUF;
2284 if (!blk_rq_nr_phys_segments(rq))
2285 nvme_tcp_set_sg_null(c);
2286 else if (rq_data_dir(rq) == WRITE &&
2287 req->data_len <= nvme_tcp_inline_data_size(queue))
2288 nvme_tcp_set_sg_inline(queue, c, req->data_len);
2290 nvme_tcp_set_sg_host_data(c, req->data_len);
2295 static blk_status_t nvme_tcp_setup_cmd_pdu(struct nvme_ns *ns,
2298 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2299 struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2300 struct nvme_tcp_queue *queue = req->queue;
2301 u8 hdgst = nvme_tcp_hdgst_len(queue), ddgst = 0;
2304 ret = nvme_setup_cmd(ns, rq);
2308 req->state = NVME_TCP_SEND_CMD_PDU;
2309 req->status = cpu_to_le16(NVME_SC_SUCCESS);
2314 req->data_len = blk_rq_nr_phys_segments(rq) ?
2315 blk_rq_payload_bytes(rq) : 0;
2316 req->curr_bio = rq->bio;
2317 if (req->curr_bio && req->data_len)
2318 nvme_tcp_init_iter(req, rq_data_dir(rq));
2320 if (rq_data_dir(rq) == WRITE &&
2321 req->data_len <= nvme_tcp_inline_data_size(queue))
2322 req->pdu_len = req->data_len;
2324 pdu->hdr.type = nvme_tcp_cmd;
2326 if (queue->hdr_digest)
2327 pdu->hdr.flags |= NVME_TCP_F_HDGST;
2328 if (queue->data_digest && req->pdu_len) {
2329 pdu->hdr.flags |= NVME_TCP_F_DDGST;
2330 ddgst = nvme_tcp_ddgst_len(queue);
2332 pdu->hdr.hlen = sizeof(*pdu);
2333 pdu->hdr.pdo = req->pdu_len ? pdu->hdr.hlen + hdgst : 0;
2335 cpu_to_le32(pdu->hdr.hlen + hdgst + req->pdu_len + ddgst);
2337 ret = nvme_tcp_map_data(queue, rq);
2338 if (unlikely(ret)) {
2339 nvme_cleanup_cmd(rq);
2340 dev_err(queue->ctrl->ctrl.device,
2341 "Failed to map data (%d)\n", ret);
2348 static void nvme_tcp_commit_rqs(struct blk_mq_hw_ctx *hctx)
2350 struct nvme_tcp_queue *queue = hctx->driver_data;
2352 if (!llist_empty(&queue->req_list))
2353 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
2356 static blk_status_t nvme_tcp_queue_rq(struct blk_mq_hw_ctx *hctx,
2357 const struct blk_mq_queue_data *bd)
2359 struct nvme_ns *ns = hctx->queue->queuedata;
2360 struct nvme_tcp_queue *queue = hctx->driver_data;
2361 struct request *rq = bd->rq;
2362 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2363 bool queue_ready = test_bit(NVME_TCP_Q_LIVE, &queue->flags);
2366 if (!nvme_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
2367 return nvme_fail_nonready_command(&queue->ctrl->ctrl, rq);
2369 ret = nvme_tcp_setup_cmd_pdu(ns, rq);
2373 blk_mq_start_request(rq);
2375 nvme_tcp_queue_request(req, true, bd->last);
2380 static int nvme_tcp_map_queues(struct blk_mq_tag_set *set)
2382 struct nvme_tcp_ctrl *ctrl = set->driver_data;
2383 struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
2385 if (opts->nr_write_queues && ctrl->io_queues[HCTX_TYPE_READ]) {
2386 /* separate read/write queues */
2387 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2388 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2389 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2390 set->map[HCTX_TYPE_READ].nr_queues =
2391 ctrl->io_queues[HCTX_TYPE_READ];
2392 set->map[HCTX_TYPE_READ].queue_offset =
2393 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2395 /* shared read/write queues */
2396 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2397 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2398 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2399 set->map[HCTX_TYPE_READ].nr_queues =
2400 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2401 set->map[HCTX_TYPE_READ].queue_offset = 0;
2403 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
2404 blk_mq_map_queues(&set->map[HCTX_TYPE_READ]);
2406 if (opts->nr_poll_queues && ctrl->io_queues[HCTX_TYPE_POLL]) {
2407 /* map dedicated poll queues only if we have queues left */
2408 set->map[HCTX_TYPE_POLL].nr_queues =
2409 ctrl->io_queues[HCTX_TYPE_POLL];
2410 set->map[HCTX_TYPE_POLL].queue_offset =
2411 ctrl->io_queues[HCTX_TYPE_DEFAULT] +
2412 ctrl->io_queues[HCTX_TYPE_READ];
2413 blk_mq_map_queues(&set->map[HCTX_TYPE_POLL]);
2416 dev_info(ctrl->ctrl.device,
2417 "mapped %d/%d/%d default/read/poll queues.\n",
2418 ctrl->io_queues[HCTX_TYPE_DEFAULT],
2419 ctrl->io_queues[HCTX_TYPE_READ],
2420 ctrl->io_queues[HCTX_TYPE_POLL]);
2425 static int nvme_tcp_poll(struct blk_mq_hw_ctx *hctx)
2427 struct nvme_tcp_queue *queue = hctx->driver_data;
2428 struct sock *sk = queue->sock->sk;
2430 if (!test_bit(NVME_TCP_Q_LIVE, &queue->flags))
2433 set_bit(NVME_TCP_Q_POLLING, &queue->flags);
2434 if (sk_can_busy_loop(sk) && skb_queue_empty_lockless(&sk->sk_receive_queue))
2435 sk_busy_loop(sk, true);
2436 nvme_tcp_try_recv(queue);
2437 clear_bit(NVME_TCP_Q_POLLING, &queue->flags);
2438 return queue->nr_cqe;
2441 static const struct blk_mq_ops nvme_tcp_mq_ops = {
2442 .queue_rq = nvme_tcp_queue_rq,
2443 .commit_rqs = nvme_tcp_commit_rqs,
2444 .complete = nvme_complete_rq,
2445 .init_request = nvme_tcp_init_request,
2446 .exit_request = nvme_tcp_exit_request,
2447 .init_hctx = nvme_tcp_init_hctx,
2448 .timeout = nvme_tcp_timeout,
2449 .map_queues = nvme_tcp_map_queues,
2450 .poll = nvme_tcp_poll,
2453 static const struct blk_mq_ops nvme_tcp_admin_mq_ops = {
2454 .queue_rq = nvme_tcp_queue_rq,
2455 .complete = nvme_complete_rq,
2456 .init_request = nvme_tcp_init_request,
2457 .exit_request = nvme_tcp_exit_request,
2458 .init_hctx = nvme_tcp_init_admin_hctx,
2459 .timeout = nvme_tcp_timeout,
2462 static const struct nvme_ctrl_ops nvme_tcp_ctrl_ops = {
2464 .module = THIS_MODULE,
2465 .flags = NVME_F_FABRICS,
2466 .reg_read32 = nvmf_reg_read32,
2467 .reg_read64 = nvmf_reg_read64,
2468 .reg_write32 = nvmf_reg_write32,
2469 .free_ctrl = nvme_tcp_free_ctrl,
2470 .submit_async_event = nvme_tcp_submit_async_event,
2471 .delete_ctrl = nvme_tcp_delete_ctrl,
2472 .get_address = nvmf_get_address,
2476 nvme_tcp_existing_controller(struct nvmf_ctrl_options *opts)
2478 struct nvme_tcp_ctrl *ctrl;
2481 mutex_lock(&nvme_tcp_ctrl_mutex);
2482 list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list) {
2483 found = nvmf_ip_options_match(&ctrl->ctrl, opts);
2487 mutex_unlock(&nvme_tcp_ctrl_mutex);
2492 static struct nvme_ctrl *nvme_tcp_create_ctrl(struct device *dev,
2493 struct nvmf_ctrl_options *opts)
2495 struct nvme_tcp_ctrl *ctrl;
2498 ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
2500 return ERR_PTR(-ENOMEM);
2502 INIT_LIST_HEAD(&ctrl->list);
2503 ctrl->ctrl.opts = opts;
2504 ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues +
2505 opts->nr_poll_queues + 1;
2506 ctrl->ctrl.sqsize = opts->queue_size - 1;
2507 ctrl->ctrl.kato = opts->kato;
2509 INIT_DELAYED_WORK(&ctrl->connect_work,
2510 nvme_tcp_reconnect_ctrl_work);
2511 INIT_WORK(&ctrl->err_work, nvme_tcp_error_recovery_work);
2512 INIT_WORK(&ctrl->ctrl.reset_work, nvme_reset_ctrl_work);
2514 if (!(opts->mask & NVMF_OPT_TRSVCID)) {
2516 kstrdup(__stringify(NVME_TCP_DISC_PORT), GFP_KERNEL);
2517 if (!opts->trsvcid) {
2521 opts->mask |= NVMF_OPT_TRSVCID;
2524 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2525 opts->traddr, opts->trsvcid, &ctrl->addr);
2527 pr_err("malformed address passed: %s:%s\n",
2528 opts->traddr, opts->trsvcid);
2532 if (opts->mask & NVMF_OPT_HOST_TRADDR) {
2533 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2534 opts->host_traddr, NULL, &ctrl->src_addr);
2536 pr_err("malformed src address passed: %s\n",
2542 if (opts->mask & NVMF_OPT_HOST_IFACE) {
2543 if (!__dev_get_by_name(&init_net, opts->host_iface)) {
2544 pr_err("invalid interface passed: %s\n",
2551 if (!opts->duplicate_connect && nvme_tcp_existing_controller(opts)) {
2556 ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
2558 if (!ctrl->queues) {
2563 ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_tcp_ctrl_ops, 0);
2565 goto out_kfree_queues;
2567 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
2570 goto out_uninit_ctrl;
2573 ret = nvme_tcp_setup_ctrl(&ctrl->ctrl, true);
2575 goto out_uninit_ctrl;
2577 dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISp\n",
2578 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
2580 mutex_lock(&nvme_tcp_ctrl_mutex);
2581 list_add_tail(&ctrl->list, &nvme_tcp_ctrl_list);
2582 mutex_unlock(&nvme_tcp_ctrl_mutex);
2587 nvme_uninit_ctrl(&ctrl->ctrl);
2588 nvme_put_ctrl(&ctrl->ctrl);
2591 return ERR_PTR(ret);
2593 kfree(ctrl->queues);
2596 return ERR_PTR(ret);
2599 static struct nvmf_transport_ops nvme_tcp_transport = {
2601 .module = THIS_MODULE,
2602 .required_opts = NVMF_OPT_TRADDR,
2603 .allowed_opts = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2604 NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
2605 NVMF_OPT_HDR_DIGEST | NVMF_OPT_DATA_DIGEST |
2606 NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES |
2607 NVMF_OPT_TOS | NVMF_OPT_HOST_IFACE,
2608 .create_ctrl = nvme_tcp_create_ctrl,
2611 static int __init nvme_tcp_init_module(void)
2613 nvme_tcp_wq = alloc_workqueue("nvme_tcp_wq",
2614 WQ_MEM_RECLAIM | WQ_HIGHPRI, 0);
2618 nvmf_register_transport(&nvme_tcp_transport);
2622 static void __exit nvme_tcp_cleanup_module(void)
2624 struct nvme_tcp_ctrl *ctrl;
2626 nvmf_unregister_transport(&nvme_tcp_transport);
2628 mutex_lock(&nvme_tcp_ctrl_mutex);
2629 list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list)
2630 nvme_delete_ctrl(&ctrl->ctrl);
2631 mutex_unlock(&nvme_tcp_ctrl_mutex);
2632 flush_workqueue(nvme_delete_wq);
2634 destroy_workqueue(nvme_tcp_wq);
2637 module_init(nvme_tcp_init_module);
2638 module_exit(nvme_tcp_cleanup_module);
2640 MODULE_LICENSE("GPL v2");