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 bool nvme_tcp_queue_more(struct nvme_tcp_queue *queue)
279 return !list_empty(&queue->send_list) ||
280 !llist_empty(&queue->req_list) || queue->more_requests;
283 static inline void nvme_tcp_queue_request(struct nvme_tcp_request *req,
284 bool sync, bool last)
286 struct nvme_tcp_queue *queue = req->queue;
289 empty = llist_add(&req->lentry, &queue->req_list) &&
290 list_empty(&queue->send_list) && !queue->request;
293 * if we're the first on the send_list and we can try to send
294 * directly, otherwise queue io_work. Also, only do that if we
295 * are on the same cpu, so we don't introduce contention.
297 if (queue->io_cpu == raw_smp_processor_id() &&
298 sync && empty && mutex_trylock(&queue->send_mutex)) {
299 queue->more_requests = !last;
300 nvme_tcp_send_all(queue);
301 queue->more_requests = false;
302 mutex_unlock(&queue->send_mutex);
305 if (last && nvme_tcp_queue_more(queue))
306 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
309 static void nvme_tcp_process_req_list(struct nvme_tcp_queue *queue)
311 struct nvme_tcp_request *req;
312 struct llist_node *node;
314 for (node = llist_del_all(&queue->req_list); node; node = node->next) {
315 req = llist_entry(node, struct nvme_tcp_request, lentry);
316 list_add(&req->entry, &queue->send_list);
320 static inline struct nvme_tcp_request *
321 nvme_tcp_fetch_request(struct nvme_tcp_queue *queue)
323 struct nvme_tcp_request *req;
325 req = list_first_entry_or_null(&queue->send_list,
326 struct nvme_tcp_request, entry);
328 nvme_tcp_process_req_list(queue);
329 req = list_first_entry_or_null(&queue->send_list,
330 struct nvme_tcp_request, entry);
335 list_del(&req->entry);
339 static inline void nvme_tcp_ddgst_final(struct ahash_request *hash,
342 ahash_request_set_crypt(hash, NULL, (u8 *)dgst, 0);
343 crypto_ahash_final(hash);
346 static inline void nvme_tcp_ddgst_update(struct ahash_request *hash,
347 struct page *page, off_t off, size_t len)
349 struct scatterlist sg;
351 sg_init_marker(&sg, 1);
352 sg_set_page(&sg, page, len, off);
353 ahash_request_set_crypt(hash, &sg, NULL, len);
354 crypto_ahash_update(hash);
357 static inline void nvme_tcp_hdgst(struct ahash_request *hash,
358 void *pdu, size_t len)
360 struct scatterlist sg;
362 sg_init_one(&sg, pdu, len);
363 ahash_request_set_crypt(hash, &sg, pdu + len, len);
364 crypto_ahash_digest(hash);
367 static int nvme_tcp_verify_hdgst(struct nvme_tcp_queue *queue,
368 void *pdu, size_t pdu_len)
370 struct nvme_tcp_hdr *hdr = pdu;
374 if (unlikely(!(hdr->flags & NVME_TCP_F_HDGST))) {
375 dev_err(queue->ctrl->ctrl.device,
376 "queue %d: header digest flag is cleared\n",
377 nvme_tcp_queue_id(queue));
381 recv_digest = *(__le32 *)(pdu + hdr->hlen);
382 nvme_tcp_hdgst(queue->rcv_hash, pdu, pdu_len);
383 exp_digest = *(__le32 *)(pdu + hdr->hlen);
384 if (recv_digest != exp_digest) {
385 dev_err(queue->ctrl->ctrl.device,
386 "header digest error: recv %#x expected %#x\n",
387 le32_to_cpu(recv_digest), le32_to_cpu(exp_digest));
394 static int nvme_tcp_check_ddgst(struct nvme_tcp_queue *queue, void *pdu)
396 struct nvme_tcp_hdr *hdr = pdu;
397 u8 digest_len = nvme_tcp_hdgst_len(queue);
400 len = le32_to_cpu(hdr->plen) - hdr->hlen -
401 ((hdr->flags & NVME_TCP_F_HDGST) ? digest_len : 0);
403 if (unlikely(len && !(hdr->flags & NVME_TCP_F_DDGST))) {
404 dev_err(queue->ctrl->ctrl.device,
405 "queue %d: data digest flag is cleared\n",
406 nvme_tcp_queue_id(queue));
409 crypto_ahash_init(queue->rcv_hash);
414 static void nvme_tcp_exit_request(struct blk_mq_tag_set *set,
415 struct request *rq, unsigned int hctx_idx)
417 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
419 page_frag_free(req->pdu);
422 static int nvme_tcp_init_request(struct blk_mq_tag_set *set,
423 struct request *rq, unsigned int hctx_idx,
424 unsigned int numa_node)
426 struct nvme_tcp_ctrl *ctrl = set->driver_data;
427 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
428 struct nvme_tcp_cmd_pdu *pdu;
429 int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
430 struct nvme_tcp_queue *queue = &ctrl->queues[queue_idx];
431 u8 hdgst = nvme_tcp_hdgst_len(queue);
433 req->pdu = page_frag_alloc(&queue->pf_cache,
434 sizeof(struct nvme_tcp_cmd_pdu) + hdgst,
435 GFP_KERNEL | __GFP_ZERO);
441 nvme_req(rq)->ctrl = &ctrl->ctrl;
442 nvme_req(rq)->cmd = &pdu->cmd;
447 static int nvme_tcp_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
448 unsigned int hctx_idx)
450 struct nvme_tcp_ctrl *ctrl = data;
451 struct nvme_tcp_queue *queue = &ctrl->queues[hctx_idx + 1];
453 hctx->driver_data = queue;
457 static int nvme_tcp_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
458 unsigned int hctx_idx)
460 struct nvme_tcp_ctrl *ctrl = data;
461 struct nvme_tcp_queue *queue = &ctrl->queues[0];
463 hctx->driver_data = queue;
467 static enum nvme_tcp_recv_state
468 nvme_tcp_recv_state(struct nvme_tcp_queue *queue)
470 return (queue->pdu_remaining) ? NVME_TCP_RECV_PDU :
471 (queue->ddgst_remaining) ? NVME_TCP_RECV_DDGST :
475 static void nvme_tcp_init_recv_ctx(struct nvme_tcp_queue *queue)
477 queue->pdu_remaining = sizeof(struct nvme_tcp_rsp_pdu) +
478 nvme_tcp_hdgst_len(queue);
479 queue->pdu_offset = 0;
480 queue->data_remaining = -1;
481 queue->ddgst_remaining = 0;
484 static void nvme_tcp_error_recovery(struct nvme_ctrl *ctrl)
486 if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_RESETTING))
489 dev_warn(ctrl->device, "starting error recovery\n");
490 queue_work(nvme_reset_wq, &to_tcp_ctrl(ctrl)->err_work);
493 static int nvme_tcp_process_nvme_cqe(struct nvme_tcp_queue *queue,
494 struct nvme_completion *cqe)
496 struct nvme_tcp_request *req;
499 rq = nvme_find_rq(nvme_tcp_tagset(queue), cqe->command_id);
501 dev_err(queue->ctrl->ctrl.device,
502 "got bad cqe.command_id %#x on queue %d\n",
503 cqe->command_id, nvme_tcp_queue_id(queue));
504 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
508 req = blk_mq_rq_to_pdu(rq);
509 if (req->status == cpu_to_le16(NVME_SC_SUCCESS))
510 req->status = cqe->status;
512 if (!nvme_try_complete_req(rq, req->status, cqe->result))
513 nvme_complete_rq(rq);
519 static int nvme_tcp_handle_c2h_data(struct nvme_tcp_queue *queue,
520 struct nvme_tcp_data_pdu *pdu)
524 rq = nvme_find_rq(nvme_tcp_tagset(queue), pdu->command_id);
526 dev_err(queue->ctrl->ctrl.device,
527 "got bad c2hdata.command_id %#x on queue %d\n",
528 pdu->command_id, nvme_tcp_queue_id(queue));
532 if (!blk_rq_payload_bytes(rq)) {
533 dev_err(queue->ctrl->ctrl.device,
534 "queue %d tag %#x unexpected data\n",
535 nvme_tcp_queue_id(queue), rq->tag);
539 queue->data_remaining = le32_to_cpu(pdu->data_length);
541 if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS &&
542 unlikely(!(pdu->hdr.flags & NVME_TCP_F_DATA_LAST))) {
543 dev_err(queue->ctrl->ctrl.device,
544 "queue %d tag %#x SUCCESS set but not last PDU\n",
545 nvme_tcp_queue_id(queue), rq->tag);
546 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
553 static int nvme_tcp_handle_comp(struct nvme_tcp_queue *queue,
554 struct nvme_tcp_rsp_pdu *pdu)
556 struct nvme_completion *cqe = &pdu->cqe;
560 * AEN requests are special as they don't time out and can
561 * survive any kind of queue freeze and often don't respond to
562 * aborts. We don't even bother to allocate a struct request
563 * for them but rather special case them here.
565 if (unlikely(nvme_is_aen_req(nvme_tcp_queue_id(queue),
567 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
570 ret = nvme_tcp_process_nvme_cqe(queue, cqe);
575 static int nvme_tcp_setup_h2c_data_pdu(struct nvme_tcp_request *req,
576 struct nvme_tcp_r2t_pdu *pdu)
578 struct nvme_tcp_data_pdu *data = req->pdu;
579 struct nvme_tcp_queue *queue = req->queue;
580 struct request *rq = blk_mq_rq_from_pdu(req);
581 u8 hdgst = nvme_tcp_hdgst_len(queue);
582 u8 ddgst = nvme_tcp_ddgst_len(queue);
584 req->pdu_len = le32_to_cpu(pdu->r2t_length);
587 if (unlikely(!req->pdu_len)) {
588 dev_err(queue->ctrl->ctrl.device,
589 "req %d r2t len is %u, probably a bug...\n",
590 rq->tag, req->pdu_len);
594 if (unlikely(req->data_sent + req->pdu_len > req->data_len)) {
595 dev_err(queue->ctrl->ctrl.device,
596 "req %d r2t len %u exceeded data len %u (%zu sent)\n",
597 rq->tag, req->pdu_len, req->data_len,
602 if (unlikely(le32_to_cpu(pdu->r2t_offset) < req->data_sent)) {
603 dev_err(queue->ctrl->ctrl.device,
604 "req %d unexpected r2t offset %u (expected %zu)\n",
605 rq->tag, le32_to_cpu(pdu->r2t_offset),
610 memset(data, 0, sizeof(*data));
611 data->hdr.type = nvme_tcp_h2c_data;
612 data->hdr.flags = NVME_TCP_F_DATA_LAST;
613 if (queue->hdr_digest)
614 data->hdr.flags |= NVME_TCP_F_HDGST;
615 if (queue->data_digest)
616 data->hdr.flags |= NVME_TCP_F_DDGST;
617 data->hdr.hlen = sizeof(*data);
618 data->hdr.pdo = data->hdr.hlen + hdgst;
620 cpu_to_le32(data->hdr.hlen + hdgst + req->pdu_len + ddgst);
621 data->ttag = pdu->ttag;
622 data->command_id = nvme_cid(rq);
623 data->data_offset = pdu->r2t_offset;
624 data->data_length = cpu_to_le32(req->pdu_len);
628 static int nvme_tcp_handle_r2t(struct nvme_tcp_queue *queue,
629 struct nvme_tcp_r2t_pdu *pdu)
631 struct nvme_tcp_request *req;
635 rq = nvme_find_rq(nvme_tcp_tagset(queue), pdu->command_id);
637 dev_err(queue->ctrl->ctrl.device,
638 "got bad r2t.command_id %#x on queue %d\n",
639 pdu->command_id, nvme_tcp_queue_id(queue));
642 req = blk_mq_rq_to_pdu(rq);
644 ret = nvme_tcp_setup_h2c_data_pdu(req, pdu);
648 req->state = NVME_TCP_SEND_H2C_PDU;
651 nvme_tcp_queue_request(req, false, true);
656 static int nvme_tcp_recv_pdu(struct nvme_tcp_queue *queue, struct sk_buff *skb,
657 unsigned int *offset, size_t *len)
659 struct nvme_tcp_hdr *hdr;
660 char *pdu = queue->pdu;
661 size_t rcv_len = min_t(size_t, *len, queue->pdu_remaining);
664 ret = skb_copy_bits(skb, *offset,
665 &pdu[queue->pdu_offset], rcv_len);
669 queue->pdu_remaining -= rcv_len;
670 queue->pdu_offset += rcv_len;
673 if (queue->pdu_remaining)
677 if (queue->hdr_digest) {
678 ret = nvme_tcp_verify_hdgst(queue, queue->pdu, hdr->hlen);
684 if (queue->data_digest) {
685 ret = nvme_tcp_check_ddgst(queue, queue->pdu);
691 case nvme_tcp_c2h_data:
692 return nvme_tcp_handle_c2h_data(queue, (void *)queue->pdu);
694 nvme_tcp_init_recv_ctx(queue);
695 return nvme_tcp_handle_comp(queue, (void *)queue->pdu);
697 nvme_tcp_init_recv_ctx(queue);
698 return nvme_tcp_handle_r2t(queue, (void *)queue->pdu);
700 dev_err(queue->ctrl->ctrl.device,
701 "unsupported pdu type (%d)\n", hdr->type);
706 static inline void nvme_tcp_end_request(struct request *rq, u16 status)
708 union nvme_result res = {};
710 if (!nvme_try_complete_req(rq, cpu_to_le16(status << 1), res))
711 nvme_complete_rq(rq);
714 static int nvme_tcp_recv_data(struct nvme_tcp_queue *queue, struct sk_buff *skb,
715 unsigned int *offset, size_t *len)
717 struct nvme_tcp_data_pdu *pdu = (void *)queue->pdu;
719 nvme_cid_to_rq(nvme_tcp_tagset(queue), pdu->command_id);
720 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
725 recv_len = min_t(size_t, *len, queue->data_remaining);
729 if (!iov_iter_count(&req->iter)) {
730 req->curr_bio = req->curr_bio->bi_next;
733 * If we don`t have any bios it means that controller
734 * sent more data than we requested, hence error
736 if (!req->curr_bio) {
737 dev_err(queue->ctrl->ctrl.device,
738 "queue %d no space in request %#x",
739 nvme_tcp_queue_id(queue), rq->tag);
740 nvme_tcp_init_recv_ctx(queue);
743 nvme_tcp_init_iter(req, READ);
746 /* we can read only from what is left in this bio */
747 recv_len = min_t(size_t, recv_len,
748 iov_iter_count(&req->iter));
750 if (queue->data_digest)
751 ret = skb_copy_and_hash_datagram_iter(skb, *offset,
752 &req->iter, recv_len, queue->rcv_hash);
754 ret = skb_copy_datagram_iter(skb, *offset,
755 &req->iter, recv_len);
757 dev_err(queue->ctrl->ctrl.device,
758 "queue %d failed to copy request %#x data",
759 nvme_tcp_queue_id(queue), rq->tag);
765 queue->data_remaining -= recv_len;
768 if (!queue->data_remaining) {
769 if (queue->data_digest) {
770 nvme_tcp_ddgst_final(queue->rcv_hash, &queue->exp_ddgst);
771 queue->ddgst_remaining = NVME_TCP_DIGEST_LENGTH;
773 if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
774 nvme_tcp_end_request(rq,
775 le16_to_cpu(req->status));
778 nvme_tcp_init_recv_ctx(queue);
785 static int nvme_tcp_recv_ddgst(struct nvme_tcp_queue *queue,
786 struct sk_buff *skb, unsigned int *offset, size_t *len)
788 struct nvme_tcp_data_pdu *pdu = (void *)queue->pdu;
789 char *ddgst = (char *)&queue->recv_ddgst;
790 size_t recv_len = min_t(size_t, *len, queue->ddgst_remaining);
791 off_t off = NVME_TCP_DIGEST_LENGTH - queue->ddgst_remaining;
794 ret = skb_copy_bits(skb, *offset, &ddgst[off], recv_len);
798 queue->ddgst_remaining -= recv_len;
801 if (queue->ddgst_remaining)
804 if (queue->recv_ddgst != queue->exp_ddgst) {
805 struct request *rq = nvme_cid_to_rq(nvme_tcp_tagset(queue),
807 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
809 req->status = cpu_to_le16(NVME_SC_DATA_XFER_ERROR);
811 dev_err(queue->ctrl->ctrl.device,
812 "data digest error: recv %#x expected %#x\n",
813 le32_to_cpu(queue->recv_ddgst),
814 le32_to_cpu(queue->exp_ddgst));
817 if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
818 struct request *rq = nvme_cid_to_rq(nvme_tcp_tagset(queue),
820 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
822 nvme_tcp_end_request(rq, le16_to_cpu(req->status));
826 nvme_tcp_init_recv_ctx(queue);
830 static int nvme_tcp_recv_skb(read_descriptor_t *desc, struct sk_buff *skb,
831 unsigned int offset, size_t len)
833 struct nvme_tcp_queue *queue = desc->arg.data;
834 size_t consumed = len;
838 switch (nvme_tcp_recv_state(queue)) {
839 case NVME_TCP_RECV_PDU:
840 result = nvme_tcp_recv_pdu(queue, skb, &offset, &len);
842 case NVME_TCP_RECV_DATA:
843 result = nvme_tcp_recv_data(queue, skb, &offset, &len);
845 case NVME_TCP_RECV_DDGST:
846 result = nvme_tcp_recv_ddgst(queue, skb, &offset, &len);
852 dev_err(queue->ctrl->ctrl.device,
853 "receive failed: %d\n", result);
854 queue->rd_enabled = false;
855 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
863 static void nvme_tcp_data_ready(struct sock *sk)
865 struct nvme_tcp_queue *queue;
867 read_lock_bh(&sk->sk_callback_lock);
868 queue = sk->sk_user_data;
869 if (likely(queue && queue->rd_enabled) &&
870 !test_bit(NVME_TCP_Q_POLLING, &queue->flags))
871 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
872 read_unlock_bh(&sk->sk_callback_lock);
875 static void nvme_tcp_write_space(struct sock *sk)
877 struct nvme_tcp_queue *queue;
879 read_lock_bh(&sk->sk_callback_lock);
880 queue = sk->sk_user_data;
881 if (likely(queue && sk_stream_is_writeable(sk))) {
882 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
883 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
885 read_unlock_bh(&sk->sk_callback_lock);
888 static void nvme_tcp_state_change(struct sock *sk)
890 struct nvme_tcp_queue *queue;
892 read_lock_bh(&sk->sk_callback_lock);
893 queue = sk->sk_user_data;
897 switch (sk->sk_state) {
903 nvme_tcp_error_recovery(&queue->ctrl->ctrl);
906 dev_info(queue->ctrl->ctrl.device,
907 "queue %d socket state %d\n",
908 nvme_tcp_queue_id(queue), sk->sk_state);
911 queue->state_change(sk);
913 read_unlock_bh(&sk->sk_callback_lock);
916 static inline void nvme_tcp_done_send_req(struct nvme_tcp_queue *queue)
918 queue->request = NULL;
921 static void nvme_tcp_fail_request(struct nvme_tcp_request *req)
923 nvme_tcp_end_request(blk_mq_rq_from_pdu(req), NVME_SC_HOST_PATH_ERROR);
926 static int nvme_tcp_try_send_data(struct nvme_tcp_request *req)
928 struct nvme_tcp_queue *queue = req->queue;
931 struct page *page = nvme_tcp_req_cur_page(req);
932 size_t offset = nvme_tcp_req_cur_offset(req);
933 size_t len = nvme_tcp_req_cur_length(req);
934 bool last = nvme_tcp_pdu_last_send(req, len);
935 int ret, flags = MSG_DONTWAIT;
937 if (last && !queue->data_digest && !nvme_tcp_queue_more(queue))
940 flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST;
942 if (sendpage_ok(page)) {
943 ret = kernel_sendpage(queue->sock, page, offset, len,
946 ret = sock_no_sendpage(queue->sock, page, offset, len,
952 if (queue->data_digest)
953 nvme_tcp_ddgst_update(queue->snd_hash, page,
957 * update the request iterator except for the last payload send
958 * in the request where we don't want to modify it as we may
959 * compete with the RX path completing the request.
961 if (req->data_sent + ret < req->data_len)
962 nvme_tcp_advance_req(req, ret);
964 /* fully successful last send in current PDU */
965 if (last && ret == len) {
966 if (queue->data_digest) {
967 nvme_tcp_ddgst_final(queue->snd_hash,
969 req->state = NVME_TCP_SEND_DDGST;
972 nvme_tcp_done_send_req(queue);
980 static int nvme_tcp_try_send_cmd_pdu(struct nvme_tcp_request *req)
982 struct nvme_tcp_queue *queue = req->queue;
983 struct nvme_tcp_cmd_pdu *pdu = req->pdu;
984 bool inline_data = nvme_tcp_has_inline_data(req);
985 u8 hdgst = nvme_tcp_hdgst_len(queue);
986 int len = sizeof(*pdu) + hdgst - req->offset;
987 int flags = MSG_DONTWAIT;
990 if (inline_data || nvme_tcp_queue_more(queue))
991 flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST;
995 if (queue->hdr_digest && !req->offset)
996 nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
998 ret = kernel_sendpage(queue->sock, virt_to_page(pdu),
999 offset_in_page(pdu) + req->offset, len, flags);
1000 if (unlikely(ret <= 0))
1006 req->state = NVME_TCP_SEND_DATA;
1007 if (queue->data_digest)
1008 crypto_ahash_init(queue->snd_hash);
1010 nvme_tcp_done_send_req(queue);
1019 static int nvme_tcp_try_send_data_pdu(struct nvme_tcp_request *req)
1021 struct nvme_tcp_queue *queue = req->queue;
1022 struct nvme_tcp_data_pdu *pdu = req->pdu;
1023 u8 hdgst = nvme_tcp_hdgst_len(queue);
1024 int len = sizeof(*pdu) - req->offset + hdgst;
1027 if (queue->hdr_digest && !req->offset)
1028 nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu));
1030 ret = kernel_sendpage(queue->sock, virt_to_page(pdu),
1031 offset_in_page(pdu) + req->offset, len,
1032 MSG_DONTWAIT | MSG_MORE | MSG_SENDPAGE_NOTLAST);
1033 if (unlikely(ret <= 0))
1038 req->state = NVME_TCP_SEND_DATA;
1039 if (queue->data_digest)
1040 crypto_ahash_init(queue->snd_hash);
1048 static int nvme_tcp_try_send_ddgst(struct nvme_tcp_request *req)
1050 struct nvme_tcp_queue *queue = req->queue;
1052 struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
1054 .iov_base = &req->ddgst + req->offset,
1055 .iov_len = NVME_TCP_DIGEST_LENGTH - req->offset
1058 if (nvme_tcp_queue_more(queue))
1059 msg.msg_flags |= MSG_MORE;
1061 msg.msg_flags |= MSG_EOR;
1063 ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
1064 if (unlikely(ret <= 0))
1067 if (req->offset + ret == NVME_TCP_DIGEST_LENGTH) {
1068 nvme_tcp_done_send_req(queue);
1076 static int nvme_tcp_try_send(struct nvme_tcp_queue *queue)
1078 struct nvme_tcp_request *req;
1081 if (!queue->request) {
1082 queue->request = nvme_tcp_fetch_request(queue);
1083 if (!queue->request)
1086 req = queue->request;
1088 if (req->state == NVME_TCP_SEND_CMD_PDU) {
1089 ret = nvme_tcp_try_send_cmd_pdu(req);
1092 if (!nvme_tcp_has_inline_data(req))
1096 if (req->state == NVME_TCP_SEND_H2C_PDU) {
1097 ret = nvme_tcp_try_send_data_pdu(req);
1102 if (req->state == NVME_TCP_SEND_DATA) {
1103 ret = nvme_tcp_try_send_data(req);
1108 if (req->state == NVME_TCP_SEND_DDGST)
1109 ret = nvme_tcp_try_send_ddgst(req);
1111 if (ret == -EAGAIN) {
1113 } else if (ret < 0) {
1114 dev_err(queue->ctrl->ctrl.device,
1115 "failed to send request %d\n", ret);
1116 if (ret != -EPIPE && ret != -ECONNRESET)
1117 nvme_tcp_fail_request(queue->request);
1118 nvme_tcp_done_send_req(queue);
1123 static int nvme_tcp_try_recv(struct nvme_tcp_queue *queue)
1125 struct socket *sock = queue->sock;
1126 struct sock *sk = sock->sk;
1127 read_descriptor_t rd_desc;
1130 rd_desc.arg.data = queue;
1134 consumed = sock->ops->read_sock(sk, &rd_desc, nvme_tcp_recv_skb);
1139 static void nvme_tcp_io_work(struct work_struct *w)
1141 struct nvme_tcp_queue *queue =
1142 container_of(w, struct nvme_tcp_queue, io_work);
1143 unsigned long deadline = jiffies + msecs_to_jiffies(1);
1146 bool pending = false;
1149 if (mutex_trylock(&queue->send_mutex)) {
1150 result = nvme_tcp_try_send(queue);
1151 mutex_unlock(&queue->send_mutex);
1154 else if (unlikely(result < 0))
1158 result = nvme_tcp_try_recv(queue);
1161 else if (unlikely(result < 0))
1167 } while (!time_after(jiffies, deadline)); /* quota is exhausted */
1169 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
1172 static void nvme_tcp_free_crypto(struct nvme_tcp_queue *queue)
1174 struct crypto_ahash *tfm = crypto_ahash_reqtfm(queue->rcv_hash);
1176 ahash_request_free(queue->rcv_hash);
1177 ahash_request_free(queue->snd_hash);
1178 crypto_free_ahash(tfm);
1181 static int nvme_tcp_alloc_crypto(struct nvme_tcp_queue *queue)
1183 struct crypto_ahash *tfm;
1185 tfm = crypto_alloc_ahash("crc32c", 0, CRYPTO_ALG_ASYNC);
1187 return PTR_ERR(tfm);
1189 queue->snd_hash = ahash_request_alloc(tfm, GFP_KERNEL);
1190 if (!queue->snd_hash)
1192 ahash_request_set_callback(queue->snd_hash, 0, NULL, NULL);
1194 queue->rcv_hash = ahash_request_alloc(tfm, GFP_KERNEL);
1195 if (!queue->rcv_hash)
1197 ahash_request_set_callback(queue->rcv_hash, 0, NULL, NULL);
1201 ahash_request_free(queue->snd_hash);
1203 crypto_free_ahash(tfm);
1207 static void nvme_tcp_free_async_req(struct nvme_tcp_ctrl *ctrl)
1209 struct nvme_tcp_request *async = &ctrl->async_req;
1211 page_frag_free(async->pdu);
1214 static int nvme_tcp_alloc_async_req(struct nvme_tcp_ctrl *ctrl)
1216 struct nvme_tcp_queue *queue = &ctrl->queues[0];
1217 struct nvme_tcp_request *async = &ctrl->async_req;
1218 u8 hdgst = nvme_tcp_hdgst_len(queue);
1220 async->pdu = page_frag_alloc(&queue->pf_cache,
1221 sizeof(struct nvme_tcp_cmd_pdu) + hdgst,
1222 GFP_KERNEL | __GFP_ZERO);
1226 async->queue = &ctrl->queues[0];
1230 static void nvme_tcp_free_queue(struct nvme_ctrl *nctrl, int qid)
1232 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1233 struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1235 if (!test_and_clear_bit(NVME_TCP_Q_ALLOCATED, &queue->flags))
1238 if (queue->hdr_digest || queue->data_digest)
1239 nvme_tcp_free_crypto(queue);
1241 sock_release(queue->sock);
1243 mutex_destroy(&queue->send_mutex);
1244 mutex_destroy(&queue->queue_lock);
1247 static int nvme_tcp_init_connection(struct nvme_tcp_queue *queue)
1249 struct nvme_tcp_icreq_pdu *icreq;
1250 struct nvme_tcp_icresp_pdu *icresp;
1251 struct msghdr msg = {};
1253 bool ctrl_hdgst, ctrl_ddgst;
1256 icreq = kzalloc(sizeof(*icreq), GFP_KERNEL);
1260 icresp = kzalloc(sizeof(*icresp), GFP_KERNEL);
1266 icreq->hdr.type = nvme_tcp_icreq;
1267 icreq->hdr.hlen = sizeof(*icreq);
1269 icreq->hdr.plen = cpu_to_le32(icreq->hdr.hlen);
1270 icreq->pfv = cpu_to_le16(NVME_TCP_PFV_1_0);
1271 icreq->maxr2t = 0; /* single inflight r2t supported */
1272 icreq->hpda = 0; /* no alignment constraint */
1273 if (queue->hdr_digest)
1274 icreq->digest |= NVME_TCP_HDR_DIGEST_ENABLE;
1275 if (queue->data_digest)
1276 icreq->digest |= NVME_TCP_DATA_DIGEST_ENABLE;
1278 iov.iov_base = icreq;
1279 iov.iov_len = sizeof(*icreq);
1280 ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
1284 memset(&msg, 0, sizeof(msg));
1285 iov.iov_base = icresp;
1286 iov.iov_len = sizeof(*icresp);
1287 ret = kernel_recvmsg(queue->sock, &msg, &iov, 1,
1288 iov.iov_len, msg.msg_flags);
1293 if (icresp->hdr.type != nvme_tcp_icresp) {
1294 pr_err("queue %d: bad type returned %d\n",
1295 nvme_tcp_queue_id(queue), icresp->hdr.type);
1299 if (le32_to_cpu(icresp->hdr.plen) != sizeof(*icresp)) {
1300 pr_err("queue %d: bad pdu length returned %d\n",
1301 nvme_tcp_queue_id(queue), icresp->hdr.plen);
1305 if (icresp->pfv != NVME_TCP_PFV_1_0) {
1306 pr_err("queue %d: bad pfv returned %d\n",
1307 nvme_tcp_queue_id(queue), icresp->pfv);
1311 ctrl_ddgst = !!(icresp->digest & NVME_TCP_DATA_DIGEST_ENABLE);
1312 if ((queue->data_digest && !ctrl_ddgst) ||
1313 (!queue->data_digest && ctrl_ddgst)) {
1314 pr_err("queue %d: data digest mismatch host: %s ctrl: %s\n",
1315 nvme_tcp_queue_id(queue),
1316 queue->data_digest ? "enabled" : "disabled",
1317 ctrl_ddgst ? "enabled" : "disabled");
1321 ctrl_hdgst = !!(icresp->digest & NVME_TCP_HDR_DIGEST_ENABLE);
1322 if ((queue->hdr_digest && !ctrl_hdgst) ||
1323 (!queue->hdr_digest && ctrl_hdgst)) {
1324 pr_err("queue %d: header digest mismatch host: %s ctrl: %s\n",
1325 nvme_tcp_queue_id(queue),
1326 queue->hdr_digest ? "enabled" : "disabled",
1327 ctrl_hdgst ? "enabled" : "disabled");
1331 if (icresp->cpda != 0) {
1332 pr_err("queue %d: unsupported cpda returned %d\n",
1333 nvme_tcp_queue_id(queue), icresp->cpda);
1345 static bool nvme_tcp_admin_queue(struct nvme_tcp_queue *queue)
1347 return nvme_tcp_queue_id(queue) == 0;
1350 static bool nvme_tcp_default_queue(struct nvme_tcp_queue *queue)
1352 struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1353 int qid = nvme_tcp_queue_id(queue);
1355 return !nvme_tcp_admin_queue(queue) &&
1356 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT];
1359 static bool nvme_tcp_read_queue(struct nvme_tcp_queue *queue)
1361 struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1362 int qid = nvme_tcp_queue_id(queue);
1364 return !nvme_tcp_admin_queue(queue) &&
1365 !nvme_tcp_default_queue(queue) &&
1366 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT] +
1367 ctrl->io_queues[HCTX_TYPE_READ];
1370 static bool nvme_tcp_poll_queue(struct nvme_tcp_queue *queue)
1372 struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1373 int qid = nvme_tcp_queue_id(queue);
1375 return !nvme_tcp_admin_queue(queue) &&
1376 !nvme_tcp_default_queue(queue) &&
1377 !nvme_tcp_read_queue(queue) &&
1378 qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT] +
1379 ctrl->io_queues[HCTX_TYPE_READ] +
1380 ctrl->io_queues[HCTX_TYPE_POLL];
1383 static void nvme_tcp_set_queue_io_cpu(struct nvme_tcp_queue *queue)
1385 struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1386 int qid = nvme_tcp_queue_id(queue);
1389 if (nvme_tcp_default_queue(queue))
1391 else if (nvme_tcp_read_queue(queue))
1392 n = qid - ctrl->io_queues[HCTX_TYPE_DEFAULT] - 1;
1393 else if (nvme_tcp_poll_queue(queue))
1394 n = qid - ctrl->io_queues[HCTX_TYPE_DEFAULT] -
1395 ctrl->io_queues[HCTX_TYPE_READ] - 1;
1396 queue->io_cpu = cpumask_next_wrap(n - 1, cpu_online_mask, -1, false);
1399 static int nvme_tcp_alloc_queue(struct nvme_ctrl *nctrl,
1400 int qid, size_t queue_size)
1402 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1403 struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1404 int ret, rcv_pdu_size;
1406 mutex_init(&queue->queue_lock);
1408 init_llist_head(&queue->req_list);
1409 INIT_LIST_HEAD(&queue->send_list);
1410 mutex_init(&queue->send_mutex);
1411 INIT_WORK(&queue->io_work, nvme_tcp_io_work);
1412 queue->queue_size = queue_size;
1415 queue->cmnd_capsule_len = nctrl->ioccsz * 16;
1417 queue->cmnd_capsule_len = sizeof(struct nvme_command) +
1418 NVME_TCP_ADMIN_CCSZ;
1420 ret = sock_create(ctrl->addr.ss_family, SOCK_STREAM,
1421 IPPROTO_TCP, &queue->sock);
1423 dev_err(nctrl->device,
1424 "failed to create socket: %d\n", ret);
1425 goto err_destroy_mutex;
1428 /* Single syn retry */
1429 tcp_sock_set_syncnt(queue->sock->sk, 1);
1431 /* Set TCP no delay */
1432 tcp_sock_set_nodelay(queue->sock->sk);
1435 * Cleanup whatever is sitting in the TCP transmit queue on socket
1436 * close. This is done to prevent stale data from being sent should
1437 * the network connection be restored before TCP times out.
1439 sock_no_linger(queue->sock->sk);
1441 if (so_priority > 0)
1442 sock_set_priority(queue->sock->sk, so_priority);
1444 /* Set socket type of service */
1445 if (nctrl->opts->tos >= 0)
1446 ip_sock_set_tos(queue->sock->sk, nctrl->opts->tos);
1448 /* Set 10 seconds timeout for icresp recvmsg */
1449 queue->sock->sk->sk_rcvtimeo = 10 * HZ;
1451 queue->sock->sk->sk_allocation = GFP_ATOMIC;
1452 nvme_tcp_set_queue_io_cpu(queue);
1453 queue->request = NULL;
1454 queue->data_remaining = 0;
1455 queue->ddgst_remaining = 0;
1456 queue->pdu_remaining = 0;
1457 queue->pdu_offset = 0;
1458 sk_set_memalloc(queue->sock->sk);
1460 if (nctrl->opts->mask & NVMF_OPT_HOST_TRADDR) {
1461 ret = kernel_bind(queue->sock, (struct sockaddr *)&ctrl->src_addr,
1462 sizeof(ctrl->src_addr));
1464 dev_err(nctrl->device,
1465 "failed to bind queue %d socket %d\n",
1471 if (nctrl->opts->mask & NVMF_OPT_HOST_IFACE) {
1472 char *iface = nctrl->opts->host_iface;
1473 sockptr_t optval = KERNEL_SOCKPTR(iface);
1475 ret = sock_setsockopt(queue->sock, SOL_SOCKET, SO_BINDTODEVICE,
1476 optval, strlen(iface));
1478 dev_err(nctrl->device,
1479 "failed to bind to interface %s queue %d err %d\n",
1485 queue->hdr_digest = nctrl->opts->hdr_digest;
1486 queue->data_digest = nctrl->opts->data_digest;
1487 if (queue->hdr_digest || queue->data_digest) {
1488 ret = nvme_tcp_alloc_crypto(queue);
1490 dev_err(nctrl->device,
1491 "failed to allocate queue %d crypto\n", qid);
1496 rcv_pdu_size = sizeof(struct nvme_tcp_rsp_pdu) +
1497 nvme_tcp_hdgst_len(queue);
1498 queue->pdu = kmalloc(rcv_pdu_size, GFP_KERNEL);
1504 dev_dbg(nctrl->device, "connecting queue %d\n",
1505 nvme_tcp_queue_id(queue));
1507 ret = kernel_connect(queue->sock, (struct sockaddr *)&ctrl->addr,
1508 sizeof(ctrl->addr), 0);
1510 dev_err(nctrl->device,
1511 "failed to connect socket: %d\n", ret);
1515 ret = nvme_tcp_init_connection(queue);
1517 goto err_init_connect;
1519 queue->rd_enabled = true;
1520 set_bit(NVME_TCP_Q_ALLOCATED, &queue->flags);
1521 nvme_tcp_init_recv_ctx(queue);
1523 write_lock_bh(&queue->sock->sk->sk_callback_lock);
1524 queue->sock->sk->sk_user_data = queue;
1525 queue->state_change = queue->sock->sk->sk_state_change;
1526 queue->data_ready = queue->sock->sk->sk_data_ready;
1527 queue->write_space = queue->sock->sk->sk_write_space;
1528 queue->sock->sk->sk_data_ready = nvme_tcp_data_ready;
1529 queue->sock->sk->sk_state_change = nvme_tcp_state_change;
1530 queue->sock->sk->sk_write_space = nvme_tcp_write_space;
1531 #ifdef CONFIG_NET_RX_BUSY_POLL
1532 queue->sock->sk->sk_ll_usec = 1;
1534 write_unlock_bh(&queue->sock->sk->sk_callback_lock);
1539 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1543 if (queue->hdr_digest || queue->data_digest)
1544 nvme_tcp_free_crypto(queue);
1546 sock_release(queue->sock);
1549 mutex_destroy(&queue->send_mutex);
1550 mutex_destroy(&queue->queue_lock);
1554 static void nvme_tcp_restore_sock_calls(struct nvme_tcp_queue *queue)
1556 struct socket *sock = queue->sock;
1558 write_lock_bh(&sock->sk->sk_callback_lock);
1559 sock->sk->sk_user_data = NULL;
1560 sock->sk->sk_data_ready = queue->data_ready;
1561 sock->sk->sk_state_change = queue->state_change;
1562 sock->sk->sk_write_space = queue->write_space;
1563 write_unlock_bh(&sock->sk->sk_callback_lock);
1566 static void __nvme_tcp_stop_queue(struct nvme_tcp_queue *queue)
1568 kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1569 nvme_tcp_restore_sock_calls(queue);
1570 cancel_work_sync(&queue->io_work);
1573 static void nvme_tcp_stop_queue(struct nvme_ctrl *nctrl, int qid)
1575 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1576 struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1578 mutex_lock(&queue->queue_lock);
1579 if (test_and_clear_bit(NVME_TCP_Q_LIVE, &queue->flags))
1580 __nvme_tcp_stop_queue(queue);
1581 mutex_unlock(&queue->queue_lock);
1584 static int nvme_tcp_start_queue(struct nvme_ctrl *nctrl, int idx)
1586 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1590 ret = nvmf_connect_io_queue(nctrl, idx);
1592 ret = nvmf_connect_admin_queue(nctrl);
1595 set_bit(NVME_TCP_Q_LIVE, &ctrl->queues[idx].flags);
1597 if (test_bit(NVME_TCP_Q_ALLOCATED, &ctrl->queues[idx].flags))
1598 __nvme_tcp_stop_queue(&ctrl->queues[idx]);
1599 dev_err(nctrl->device,
1600 "failed to connect queue: %d ret=%d\n", idx, ret);
1605 static struct blk_mq_tag_set *nvme_tcp_alloc_tagset(struct nvme_ctrl *nctrl,
1608 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1609 struct blk_mq_tag_set *set;
1613 set = &ctrl->admin_tag_set;
1614 memset(set, 0, sizeof(*set));
1615 set->ops = &nvme_tcp_admin_mq_ops;
1616 set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
1617 set->reserved_tags = NVMF_RESERVED_TAGS;
1618 set->numa_node = nctrl->numa_node;
1619 set->flags = BLK_MQ_F_BLOCKING;
1620 set->cmd_size = sizeof(struct nvme_tcp_request);
1621 set->driver_data = ctrl;
1622 set->nr_hw_queues = 1;
1623 set->timeout = NVME_ADMIN_TIMEOUT;
1625 set = &ctrl->tag_set;
1626 memset(set, 0, sizeof(*set));
1627 set->ops = &nvme_tcp_mq_ops;
1628 set->queue_depth = nctrl->sqsize + 1;
1629 set->reserved_tags = NVMF_RESERVED_TAGS;
1630 set->numa_node = nctrl->numa_node;
1631 set->flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_BLOCKING;
1632 set->cmd_size = sizeof(struct nvme_tcp_request);
1633 set->driver_data = ctrl;
1634 set->nr_hw_queues = nctrl->queue_count - 1;
1635 set->timeout = NVME_IO_TIMEOUT;
1636 set->nr_maps = nctrl->opts->nr_poll_queues ? HCTX_MAX_TYPES : 2;
1639 ret = blk_mq_alloc_tag_set(set);
1641 return ERR_PTR(ret);
1646 static void nvme_tcp_free_admin_queue(struct nvme_ctrl *ctrl)
1648 if (to_tcp_ctrl(ctrl)->async_req.pdu) {
1649 cancel_work_sync(&ctrl->async_event_work);
1650 nvme_tcp_free_async_req(to_tcp_ctrl(ctrl));
1651 to_tcp_ctrl(ctrl)->async_req.pdu = NULL;
1654 nvme_tcp_free_queue(ctrl, 0);
1657 static void nvme_tcp_free_io_queues(struct nvme_ctrl *ctrl)
1661 for (i = 1; i < ctrl->queue_count; i++)
1662 nvme_tcp_free_queue(ctrl, i);
1665 static void nvme_tcp_stop_io_queues(struct nvme_ctrl *ctrl)
1669 for (i = 1; i < ctrl->queue_count; i++)
1670 nvme_tcp_stop_queue(ctrl, i);
1673 static int nvme_tcp_start_io_queues(struct nvme_ctrl *ctrl)
1677 for (i = 1; i < ctrl->queue_count; i++) {
1678 ret = nvme_tcp_start_queue(ctrl, i);
1680 goto out_stop_queues;
1686 for (i--; i >= 1; i--)
1687 nvme_tcp_stop_queue(ctrl, i);
1691 static int nvme_tcp_alloc_admin_queue(struct nvme_ctrl *ctrl)
1695 ret = nvme_tcp_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
1699 ret = nvme_tcp_alloc_async_req(to_tcp_ctrl(ctrl));
1701 goto out_free_queue;
1706 nvme_tcp_free_queue(ctrl, 0);
1710 static int __nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
1714 for (i = 1; i < ctrl->queue_count; i++) {
1715 ret = nvme_tcp_alloc_queue(ctrl, i,
1718 goto out_free_queues;
1724 for (i--; i >= 1; i--)
1725 nvme_tcp_free_queue(ctrl, i);
1730 static unsigned int nvme_tcp_nr_io_queues(struct nvme_ctrl *ctrl)
1732 unsigned int nr_io_queues;
1734 nr_io_queues = min(ctrl->opts->nr_io_queues, num_online_cpus());
1735 nr_io_queues += min(ctrl->opts->nr_write_queues, num_online_cpus());
1736 nr_io_queues += min(ctrl->opts->nr_poll_queues, num_online_cpus());
1738 return nr_io_queues;
1741 static void nvme_tcp_set_io_queues(struct nvme_ctrl *nctrl,
1742 unsigned int nr_io_queues)
1744 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1745 struct nvmf_ctrl_options *opts = nctrl->opts;
1747 if (opts->nr_write_queues && opts->nr_io_queues < nr_io_queues) {
1749 * separate read/write queues
1750 * hand out dedicated default queues only after we have
1751 * sufficient read queues.
1753 ctrl->io_queues[HCTX_TYPE_READ] = opts->nr_io_queues;
1754 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_READ];
1755 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
1756 min(opts->nr_write_queues, nr_io_queues);
1757 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
1760 * shared read/write queues
1761 * either no write queues were requested, or we don't have
1762 * sufficient queue count to have dedicated default queues.
1764 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
1765 min(opts->nr_io_queues, nr_io_queues);
1766 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
1769 if (opts->nr_poll_queues && nr_io_queues) {
1770 /* map dedicated poll queues only if we have queues left */
1771 ctrl->io_queues[HCTX_TYPE_POLL] =
1772 min(opts->nr_poll_queues, nr_io_queues);
1776 static int nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
1778 unsigned int nr_io_queues;
1781 nr_io_queues = nvme_tcp_nr_io_queues(ctrl);
1782 ret = nvme_set_queue_count(ctrl, &nr_io_queues);
1786 if (nr_io_queues == 0) {
1787 dev_err(ctrl->device,
1788 "unable to set any I/O queues\n");
1792 ctrl->queue_count = nr_io_queues + 1;
1793 dev_info(ctrl->device,
1794 "creating %d I/O queues.\n", nr_io_queues);
1796 nvme_tcp_set_io_queues(ctrl, nr_io_queues);
1798 return __nvme_tcp_alloc_io_queues(ctrl);
1801 static void nvme_tcp_destroy_io_queues(struct nvme_ctrl *ctrl, bool remove)
1803 nvme_tcp_stop_io_queues(ctrl);
1805 blk_cleanup_queue(ctrl->connect_q);
1806 blk_mq_free_tag_set(ctrl->tagset);
1808 nvme_tcp_free_io_queues(ctrl);
1811 static int nvme_tcp_configure_io_queues(struct nvme_ctrl *ctrl, bool new)
1815 ret = nvme_tcp_alloc_io_queues(ctrl);
1820 ctrl->tagset = nvme_tcp_alloc_tagset(ctrl, false);
1821 if (IS_ERR(ctrl->tagset)) {
1822 ret = PTR_ERR(ctrl->tagset);
1823 goto out_free_io_queues;
1826 ctrl->connect_q = blk_mq_init_queue(ctrl->tagset);
1827 if (IS_ERR(ctrl->connect_q)) {
1828 ret = PTR_ERR(ctrl->connect_q);
1829 goto out_free_tag_set;
1833 ret = nvme_tcp_start_io_queues(ctrl);
1835 goto out_cleanup_connect_q;
1838 nvme_start_queues(ctrl);
1839 if (!nvme_wait_freeze_timeout(ctrl, NVME_IO_TIMEOUT)) {
1841 * If we timed out waiting for freeze we are likely to
1842 * be stuck. Fail the controller initialization just
1846 goto out_wait_freeze_timed_out;
1848 blk_mq_update_nr_hw_queues(ctrl->tagset,
1849 ctrl->queue_count - 1);
1850 nvme_unfreeze(ctrl);
1855 out_wait_freeze_timed_out:
1856 nvme_stop_queues(ctrl);
1857 nvme_sync_io_queues(ctrl);
1858 nvme_tcp_stop_io_queues(ctrl);
1859 out_cleanup_connect_q:
1860 nvme_cancel_tagset(ctrl);
1862 blk_cleanup_queue(ctrl->connect_q);
1865 blk_mq_free_tag_set(ctrl->tagset);
1867 nvme_tcp_free_io_queues(ctrl);
1871 static void nvme_tcp_destroy_admin_queue(struct nvme_ctrl *ctrl, bool remove)
1873 nvme_tcp_stop_queue(ctrl, 0);
1875 blk_cleanup_queue(ctrl->admin_q);
1876 blk_cleanup_queue(ctrl->fabrics_q);
1877 blk_mq_free_tag_set(ctrl->admin_tagset);
1879 nvme_tcp_free_admin_queue(ctrl);
1882 static int nvme_tcp_configure_admin_queue(struct nvme_ctrl *ctrl, bool new)
1886 error = nvme_tcp_alloc_admin_queue(ctrl);
1891 ctrl->admin_tagset = nvme_tcp_alloc_tagset(ctrl, true);
1892 if (IS_ERR(ctrl->admin_tagset)) {
1893 error = PTR_ERR(ctrl->admin_tagset);
1894 goto out_free_queue;
1897 ctrl->fabrics_q = blk_mq_init_queue(ctrl->admin_tagset);
1898 if (IS_ERR(ctrl->fabrics_q)) {
1899 error = PTR_ERR(ctrl->fabrics_q);
1900 goto out_free_tagset;
1903 ctrl->admin_q = blk_mq_init_queue(ctrl->admin_tagset);
1904 if (IS_ERR(ctrl->admin_q)) {
1905 error = PTR_ERR(ctrl->admin_q);
1906 goto out_cleanup_fabrics_q;
1910 error = nvme_tcp_start_queue(ctrl, 0);
1912 goto out_cleanup_queue;
1914 error = nvme_enable_ctrl(ctrl);
1916 goto out_stop_queue;
1918 blk_mq_unquiesce_queue(ctrl->admin_q);
1920 error = nvme_init_ctrl_finish(ctrl);
1922 goto out_quiesce_queue;
1927 blk_mq_quiesce_queue(ctrl->admin_q);
1928 blk_sync_queue(ctrl->admin_q);
1930 nvme_tcp_stop_queue(ctrl, 0);
1931 nvme_cancel_admin_tagset(ctrl);
1934 blk_cleanup_queue(ctrl->admin_q);
1935 out_cleanup_fabrics_q:
1937 blk_cleanup_queue(ctrl->fabrics_q);
1940 blk_mq_free_tag_set(ctrl->admin_tagset);
1942 nvme_tcp_free_admin_queue(ctrl);
1946 static void nvme_tcp_teardown_admin_queue(struct nvme_ctrl *ctrl,
1949 blk_mq_quiesce_queue(ctrl->admin_q);
1950 blk_sync_queue(ctrl->admin_q);
1951 nvme_tcp_stop_queue(ctrl, 0);
1952 nvme_cancel_admin_tagset(ctrl);
1954 blk_mq_unquiesce_queue(ctrl->admin_q);
1955 nvme_tcp_destroy_admin_queue(ctrl, remove);
1958 static void nvme_tcp_teardown_io_queues(struct nvme_ctrl *ctrl,
1961 if (ctrl->queue_count <= 1)
1963 blk_mq_quiesce_queue(ctrl->admin_q);
1964 nvme_start_freeze(ctrl);
1965 nvme_stop_queues(ctrl);
1966 nvme_sync_io_queues(ctrl);
1967 nvme_tcp_stop_io_queues(ctrl);
1968 nvme_cancel_tagset(ctrl);
1970 nvme_start_queues(ctrl);
1971 nvme_tcp_destroy_io_queues(ctrl, remove);
1974 static void nvme_tcp_reconnect_or_remove(struct nvme_ctrl *ctrl)
1976 /* If we are resetting/deleting then do nothing */
1977 if (ctrl->state != NVME_CTRL_CONNECTING) {
1978 WARN_ON_ONCE(ctrl->state == NVME_CTRL_NEW ||
1979 ctrl->state == NVME_CTRL_LIVE);
1983 if (nvmf_should_reconnect(ctrl)) {
1984 dev_info(ctrl->device, "Reconnecting in %d seconds...\n",
1985 ctrl->opts->reconnect_delay);
1986 queue_delayed_work(nvme_wq, &to_tcp_ctrl(ctrl)->connect_work,
1987 ctrl->opts->reconnect_delay * HZ);
1989 dev_info(ctrl->device, "Removing controller...\n");
1990 nvme_delete_ctrl(ctrl);
1994 static int nvme_tcp_setup_ctrl(struct nvme_ctrl *ctrl, bool new)
1996 struct nvmf_ctrl_options *opts = ctrl->opts;
1999 ret = nvme_tcp_configure_admin_queue(ctrl, new);
2005 dev_err(ctrl->device, "icdoff is not supported!\n");
2009 if (!nvme_ctrl_sgl_supported(ctrl)) {
2011 dev_err(ctrl->device, "Mandatory sgls are not supported!\n");
2015 if (opts->queue_size > ctrl->sqsize + 1)
2016 dev_warn(ctrl->device,
2017 "queue_size %zu > ctrl sqsize %u, clamping down\n",
2018 opts->queue_size, ctrl->sqsize + 1);
2020 if (ctrl->sqsize + 1 > ctrl->maxcmd) {
2021 dev_warn(ctrl->device,
2022 "sqsize %u > ctrl maxcmd %u, clamping down\n",
2023 ctrl->sqsize + 1, ctrl->maxcmd);
2024 ctrl->sqsize = ctrl->maxcmd - 1;
2027 if (ctrl->queue_count > 1) {
2028 ret = nvme_tcp_configure_io_queues(ctrl, new);
2033 if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_LIVE)) {
2035 * state change failure is ok if we started ctrl delete,
2036 * unless we're during creation of a new controller to
2037 * avoid races with teardown flow.
2039 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING &&
2040 ctrl->state != NVME_CTRL_DELETING_NOIO);
2046 nvme_start_ctrl(ctrl);
2050 if (ctrl->queue_count > 1) {
2051 nvme_stop_queues(ctrl);
2052 nvme_sync_io_queues(ctrl);
2053 nvme_tcp_stop_io_queues(ctrl);
2054 nvme_cancel_tagset(ctrl);
2055 nvme_tcp_destroy_io_queues(ctrl, new);
2058 blk_mq_quiesce_queue(ctrl->admin_q);
2059 blk_sync_queue(ctrl->admin_q);
2060 nvme_tcp_stop_queue(ctrl, 0);
2061 nvme_cancel_admin_tagset(ctrl);
2062 nvme_tcp_destroy_admin_queue(ctrl, new);
2066 static void nvme_tcp_reconnect_ctrl_work(struct work_struct *work)
2068 struct nvme_tcp_ctrl *tcp_ctrl = container_of(to_delayed_work(work),
2069 struct nvme_tcp_ctrl, connect_work);
2070 struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
2072 ++ctrl->nr_reconnects;
2074 if (nvme_tcp_setup_ctrl(ctrl, false))
2077 dev_info(ctrl->device, "Successfully reconnected (%d attempt)\n",
2078 ctrl->nr_reconnects);
2080 ctrl->nr_reconnects = 0;
2085 dev_info(ctrl->device, "Failed reconnect attempt %d\n",
2086 ctrl->nr_reconnects);
2087 nvme_tcp_reconnect_or_remove(ctrl);
2090 static void nvme_tcp_error_recovery_work(struct work_struct *work)
2092 struct nvme_tcp_ctrl *tcp_ctrl = container_of(work,
2093 struct nvme_tcp_ctrl, err_work);
2094 struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
2096 nvme_stop_keep_alive(ctrl);
2097 nvme_tcp_teardown_io_queues(ctrl, false);
2098 /* unquiesce to fail fast pending requests */
2099 nvme_start_queues(ctrl);
2100 nvme_tcp_teardown_admin_queue(ctrl, false);
2101 blk_mq_unquiesce_queue(ctrl->admin_q);
2103 if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
2104 /* state change failure is ok if we started ctrl delete */
2105 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING &&
2106 ctrl->state != NVME_CTRL_DELETING_NOIO);
2110 nvme_tcp_reconnect_or_remove(ctrl);
2113 static void nvme_tcp_teardown_ctrl(struct nvme_ctrl *ctrl, bool shutdown)
2115 cancel_work_sync(&to_tcp_ctrl(ctrl)->err_work);
2116 cancel_delayed_work_sync(&to_tcp_ctrl(ctrl)->connect_work);
2118 nvme_tcp_teardown_io_queues(ctrl, shutdown);
2119 blk_mq_quiesce_queue(ctrl->admin_q);
2121 nvme_shutdown_ctrl(ctrl);
2123 nvme_disable_ctrl(ctrl);
2124 nvme_tcp_teardown_admin_queue(ctrl, shutdown);
2127 static void nvme_tcp_delete_ctrl(struct nvme_ctrl *ctrl)
2129 nvme_tcp_teardown_ctrl(ctrl, true);
2132 static void nvme_reset_ctrl_work(struct work_struct *work)
2134 struct nvme_ctrl *ctrl =
2135 container_of(work, struct nvme_ctrl, reset_work);
2137 nvme_stop_ctrl(ctrl);
2138 nvme_tcp_teardown_ctrl(ctrl, false);
2140 if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
2141 /* state change failure is ok if we started ctrl delete */
2142 WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING &&
2143 ctrl->state != NVME_CTRL_DELETING_NOIO);
2147 if (nvme_tcp_setup_ctrl(ctrl, false))
2153 ++ctrl->nr_reconnects;
2154 nvme_tcp_reconnect_or_remove(ctrl);
2157 static void nvme_tcp_free_ctrl(struct nvme_ctrl *nctrl)
2159 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
2161 if (list_empty(&ctrl->list))
2164 mutex_lock(&nvme_tcp_ctrl_mutex);
2165 list_del(&ctrl->list);
2166 mutex_unlock(&nvme_tcp_ctrl_mutex);
2168 nvmf_free_options(nctrl->opts);
2170 kfree(ctrl->queues);
2174 static void nvme_tcp_set_sg_null(struct nvme_command *c)
2176 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2180 sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
2181 NVME_SGL_FMT_TRANSPORT_A;
2184 static void nvme_tcp_set_sg_inline(struct nvme_tcp_queue *queue,
2185 struct nvme_command *c, u32 data_len)
2187 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2189 sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
2190 sg->length = cpu_to_le32(data_len);
2191 sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
2194 static void nvme_tcp_set_sg_host_data(struct nvme_command *c,
2197 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2200 sg->length = cpu_to_le32(data_len);
2201 sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
2202 NVME_SGL_FMT_TRANSPORT_A;
2205 static void nvme_tcp_submit_async_event(struct nvme_ctrl *arg)
2207 struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(arg);
2208 struct nvme_tcp_queue *queue = &ctrl->queues[0];
2209 struct nvme_tcp_cmd_pdu *pdu = ctrl->async_req.pdu;
2210 struct nvme_command *cmd = &pdu->cmd;
2211 u8 hdgst = nvme_tcp_hdgst_len(queue);
2213 memset(pdu, 0, sizeof(*pdu));
2214 pdu->hdr.type = nvme_tcp_cmd;
2215 if (queue->hdr_digest)
2216 pdu->hdr.flags |= NVME_TCP_F_HDGST;
2217 pdu->hdr.hlen = sizeof(*pdu);
2218 pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst);
2220 cmd->common.opcode = nvme_admin_async_event;
2221 cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
2222 cmd->common.flags |= NVME_CMD_SGL_METABUF;
2223 nvme_tcp_set_sg_null(cmd);
2225 ctrl->async_req.state = NVME_TCP_SEND_CMD_PDU;
2226 ctrl->async_req.offset = 0;
2227 ctrl->async_req.curr_bio = NULL;
2228 ctrl->async_req.data_len = 0;
2230 nvme_tcp_queue_request(&ctrl->async_req, true, true);
2233 static void nvme_tcp_complete_timed_out(struct request *rq)
2235 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2236 struct nvme_ctrl *ctrl = &req->queue->ctrl->ctrl;
2238 nvme_tcp_stop_queue(ctrl, nvme_tcp_queue_id(req->queue));
2239 if (blk_mq_request_started(rq) && !blk_mq_request_completed(rq)) {
2240 nvme_req(rq)->status = NVME_SC_HOST_ABORTED_CMD;
2241 blk_mq_complete_request(rq);
2245 static enum blk_eh_timer_return
2246 nvme_tcp_timeout(struct request *rq, bool reserved)
2248 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2249 struct nvme_ctrl *ctrl = &req->queue->ctrl->ctrl;
2250 struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2252 dev_warn(ctrl->device,
2253 "queue %d: timeout request %#x type %d\n",
2254 nvme_tcp_queue_id(req->queue), rq->tag, pdu->hdr.type);
2256 if (ctrl->state != NVME_CTRL_LIVE) {
2258 * If we are resetting, connecting or deleting we should
2259 * complete immediately because we may block controller
2260 * teardown or setup sequence
2261 * - ctrl disable/shutdown fabrics requests
2262 * - connect requests
2263 * - initialization admin requests
2264 * - I/O requests that entered after unquiescing and
2265 * the controller stopped responding
2267 * All other requests should be cancelled by the error
2268 * recovery work, so it's fine that we fail it here.
2270 nvme_tcp_complete_timed_out(rq);
2275 * LIVE state should trigger the normal error recovery which will
2276 * handle completing this request.
2278 nvme_tcp_error_recovery(ctrl);
2279 return BLK_EH_RESET_TIMER;
2282 static blk_status_t nvme_tcp_map_data(struct nvme_tcp_queue *queue,
2285 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2286 struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2287 struct nvme_command *c = &pdu->cmd;
2289 c->common.flags |= NVME_CMD_SGL_METABUF;
2291 if (!blk_rq_nr_phys_segments(rq))
2292 nvme_tcp_set_sg_null(c);
2293 else if (rq_data_dir(rq) == WRITE &&
2294 req->data_len <= nvme_tcp_inline_data_size(queue))
2295 nvme_tcp_set_sg_inline(queue, c, req->data_len);
2297 nvme_tcp_set_sg_host_data(c, req->data_len);
2302 static blk_status_t nvme_tcp_setup_cmd_pdu(struct nvme_ns *ns,
2305 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2306 struct nvme_tcp_cmd_pdu *pdu = req->pdu;
2307 struct nvme_tcp_queue *queue = req->queue;
2308 u8 hdgst = nvme_tcp_hdgst_len(queue), ddgst = 0;
2311 ret = nvme_setup_cmd(ns, rq);
2315 req->state = NVME_TCP_SEND_CMD_PDU;
2316 req->status = cpu_to_le16(NVME_SC_SUCCESS);
2321 req->data_len = blk_rq_nr_phys_segments(rq) ?
2322 blk_rq_payload_bytes(rq) : 0;
2323 req->curr_bio = rq->bio;
2324 if (req->curr_bio && req->data_len)
2325 nvme_tcp_init_iter(req, rq_data_dir(rq));
2327 if (rq_data_dir(rq) == WRITE &&
2328 req->data_len <= nvme_tcp_inline_data_size(queue))
2329 req->pdu_len = req->data_len;
2331 pdu->hdr.type = nvme_tcp_cmd;
2333 if (queue->hdr_digest)
2334 pdu->hdr.flags |= NVME_TCP_F_HDGST;
2335 if (queue->data_digest && req->pdu_len) {
2336 pdu->hdr.flags |= NVME_TCP_F_DDGST;
2337 ddgst = nvme_tcp_ddgst_len(queue);
2339 pdu->hdr.hlen = sizeof(*pdu);
2340 pdu->hdr.pdo = req->pdu_len ? pdu->hdr.hlen + hdgst : 0;
2342 cpu_to_le32(pdu->hdr.hlen + hdgst + req->pdu_len + ddgst);
2344 ret = nvme_tcp_map_data(queue, rq);
2345 if (unlikely(ret)) {
2346 nvme_cleanup_cmd(rq);
2347 dev_err(queue->ctrl->ctrl.device,
2348 "Failed to map data (%d)\n", ret);
2355 static void nvme_tcp_commit_rqs(struct blk_mq_hw_ctx *hctx)
2357 struct nvme_tcp_queue *queue = hctx->driver_data;
2359 if (!llist_empty(&queue->req_list))
2360 queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
2363 static blk_status_t nvme_tcp_queue_rq(struct blk_mq_hw_ctx *hctx,
2364 const struct blk_mq_queue_data *bd)
2366 struct nvme_ns *ns = hctx->queue->queuedata;
2367 struct nvme_tcp_queue *queue = hctx->driver_data;
2368 struct request *rq = bd->rq;
2369 struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2370 bool queue_ready = test_bit(NVME_TCP_Q_LIVE, &queue->flags);
2373 if (!nvme_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
2374 return nvme_fail_nonready_command(&queue->ctrl->ctrl, rq);
2376 ret = nvme_tcp_setup_cmd_pdu(ns, rq);
2380 blk_mq_start_request(rq);
2382 nvme_tcp_queue_request(req, true, bd->last);
2387 static int nvme_tcp_map_queues(struct blk_mq_tag_set *set)
2389 struct nvme_tcp_ctrl *ctrl = set->driver_data;
2390 struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
2392 if (opts->nr_write_queues && ctrl->io_queues[HCTX_TYPE_READ]) {
2393 /* separate read/write queues */
2394 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2395 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2396 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2397 set->map[HCTX_TYPE_READ].nr_queues =
2398 ctrl->io_queues[HCTX_TYPE_READ];
2399 set->map[HCTX_TYPE_READ].queue_offset =
2400 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2402 /* shared read/write queues */
2403 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2404 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2405 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2406 set->map[HCTX_TYPE_READ].nr_queues =
2407 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2408 set->map[HCTX_TYPE_READ].queue_offset = 0;
2410 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
2411 blk_mq_map_queues(&set->map[HCTX_TYPE_READ]);
2413 if (opts->nr_poll_queues && ctrl->io_queues[HCTX_TYPE_POLL]) {
2414 /* map dedicated poll queues only if we have queues left */
2415 set->map[HCTX_TYPE_POLL].nr_queues =
2416 ctrl->io_queues[HCTX_TYPE_POLL];
2417 set->map[HCTX_TYPE_POLL].queue_offset =
2418 ctrl->io_queues[HCTX_TYPE_DEFAULT] +
2419 ctrl->io_queues[HCTX_TYPE_READ];
2420 blk_mq_map_queues(&set->map[HCTX_TYPE_POLL]);
2423 dev_info(ctrl->ctrl.device,
2424 "mapped %d/%d/%d default/read/poll queues.\n",
2425 ctrl->io_queues[HCTX_TYPE_DEFAULT],
2426 ctrl->io_queues[HCTX_TYPE_READ],
2427 ctrl->io_queues[HCTX_TYPE_POLL]);
2432 static int nvme_tcp_poll(struct blk_mq_hw_ctx *hctx)
2434 struct nvme_tcp_queue *queue = hctx->driver_data;
2435 struct sock *sk = queue->sock->sk;
2437 if (!test_bit(NVME_TCP_Q_LIVE, &queue->flags))
2440 set_bit(NVME_TCP_Q_POLLING, &queue->flags);
2441 if (sk_can_busy_loop(sk) && skb_queue_empty_lockless(&sk->sk_receive_queue))
2442 sk_busy_loop(sk, true);
2443 nvme_tcp_try_recv(queue);
2444 clear_bit(NVME_TCP_Q_POLLING, &queue->flags);
2445 return queue->nr_cqe;
2448 static const struct blk_mq_ops nvme_tcp_mq_ops = {
2449 .queue_rq = nvme_tcp_queue_rq,
2450 .commit_rqs = nvme_tcp_commit_rqs,
2451 .complete = nvme_complete_rq,
2452 .init_request = nvme_tcp_init_request,
2453 .exit_request = nvme_tcp_exit_request,
2454 .init_hctx = nvme_tcp_init_hctx,
2455 .timeout = nvme_tcp_timeout,
2456 .map_queues = nvme_tcp_map_queues,
2457 .poll = nvme_tcp_poll,
2460 static const struct blk_mq_ops nvme_tcp_admin_mq_ops = {
2461 .queue_rq = nvme_tcp_queue_rq,
2462 .complete = nvme_complete_rq,
2463 .init_request = nvme_tcp_init_request,
2464 .exit_request = nvme_tcp_exit_request,
2465 .init_hctx = nvme_tcp_init_admin_hctx,
2466 .timeout = nvme_tcp_timeout,
2469 static const struct nvme_ctrl_ops nvme_tcp_ctrl_ops = {
2471 .module = THIS_MODULE,
2472 .flags = NVME_F_FABRICS,
2473 .reg_read32 = nvmf_reg_read32,
2474 .reg_read64 = nvmf_reg_read64,
2475 .reg_write32 = nvmf_reg_write32,
2476 .free_ctrl = nvme_tcp_free_ctrl,
2477 .submit_async_event = nvme_tcp_submit_async_event,
2478 .delete_ctrl = nvme_tcp_delete_ctrl,
2479 .get_address = nvmf_get_address,
2483 nvme_tcp_existing_controller(struct nvmf_ctrl_options *opts)
2485 struct nvme_tcp_ctrl *ctrl;
2488 mutex_lock(&nvme_tcp_ctrl_mutex);
2489 list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list) {
2490 found = nvmf_ip_options_match(&ctrl->ctrl, opts);
2494 mutex_unlock(&nvme_tcp_ctrl_mutex);
2499 static struct nvme_ctrl *nvme_tcp_create_ctrl(struct device *dev,
2500 struct nvmf_ctrl_options *opts)
2502 struct nvme_tcp_ctrl *ctrl;
2505 ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
2507 return ERR_PTR(-ENOMEM);
2509 INIT_LIST_HEAD(&ctrl->list);
2510 ctrl->ctrl.opts = opts;
2511 ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues +
2512 opts->nr_poll_queues + 1;
2513 ctrl->ctrl.sqsize = opts->queue_size - 1;
2514 ctrl->ctrl.kato = opts->kato;
2516 INIT_DELAYED_WORK(&ctrl->connect_work,
2517 nvme_tcp_reconnect_ctrl_work);
2518 INIT_WORK(&ctrl->err_work, nvme_tcp_error_recovery_work);
2519 INIT_WORK(&ctrl->ctrl.reset_work, nvme_reset_ctrl_work);
2521 if (!(opts->mask & NVMF_OPT_TRSVCID)) {
2523 kstrdup(__stringify(NVME_TCP_DISC_PORT), GFP_KERNEL);
2524 if (!opts->trsvcid) {
2528 opts->mask |= NVMF_OPT_TRSVCID;
2531 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2532 opts->traddr, opts->trsvcid, &ctrl->addr);
2534 pr_err("malformed address passed: %s:%s\n",
2535 opts->traddr, opts->trsvcid);
2539 if (opts->mask & NVMF_OPT_HOST_TRADDR) {
2540 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2541 opts->host_traddr, NULL, &ctrl->src_addr);
2543 pr_err("malformed src address passed: %s\n",
2549 if (opts->mask & NVMF_OPT_HOST_IFACE) {
2550 if (!__dev_get_by_name(&init_net, opts->host_iface)) {
2551 pr_err("invalid interface passed: %s\n",
2558 if (!opts->duplicate_connect && nvme_tcp_existing_controller(opts)) {
2563 ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
2565 if (!ctrl->queues) {
2570 ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_tcp_ctrl_ops, 0);
2572 goto out_kfree_queues;
2574 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
2577 goto out_uninit_ctrl;
2580 ret = nvme_tcp_setup_ctrl(&ctrl->ctrl, true);
2582 goto out_uninit_ctrl;
2584 dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISp\n",
2585 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
2587 mutex_lock(&nvme_tcp_ctrl_mutex);
2588 list_add_tail(&ctrl->list, &nvme_tcp_ctrl_list);
2589 mutex_unlock(&nvme_tcp_ctrl_mutex);
2594 nvme_uninit_ctrl(&ctrl->ctrl);
2595 nvme_put_ctrl(&ctrl->ctrl);
2598 return ERR_PTR(ret);
2600 kfree(ctrl->queues);
2603 return ERR_PTR(ret);
2606 static struct nvmf_transport_ops nvme_tcp_transport = {
2608 .module = THIS_MODULE,
2609 .required_opts = NVMF_OPT_TRADDR,
2610 .allowed_opts = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2611 NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
2612 NVMF_OPT_HDR_DIGEST | NVMF_OPT_DATA_DIGEST |
2613 NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES |
2614 NVMF_OPT_TOS | NVMF_OPT_HOST_IFACE,
2615 .create_ctrl = nvme_tcp_create_ctrl,
2618 static int __init nvme_tcp_init_module(void)
2620 nvme_tcp_wq = alloc_workqueue("nvme_tcp_wq",
2621 WQ_MEM_RECLAIM | WQ_HIGHPRI, 0);
2625 nvmf_register_transport(&nvme_tcp_transport);
2629 static void __exit nvme_tcp_cleanup_module(void)
2631 struct nvme_tcp_ctrl *ctrl;
2633 nvmf_unregister_transport(&nvme_tcp_transport);
2635 mutex_lock(&nvme_tcp_ctrl_mutex);
2636 list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list)
2637 nvme_delete_ctrl(&ctrl->ctrl);
2638 mutex_unlock(&nvme_tcp_ctrl_mutex);
2639 flush_workqueue(nvme_delete_wq);
2641 destroy_workqueue(nvme_tcp_wq);
2644 module_init(nvme_tcp_init_module);
2645 module_exit(nvme_tcp_cleanup_module);
2647 MODULE_LICENSE("GPL v2");