13f2304f8SSagi Grimberg // SPDX-License-Identifier: GPL-2.0 23f2304f8SSagi Grimberg /* 33f2304f8SSagi Grimberg * NVMe over Fabrics TCP host. 43f2304f8SSagi Grimberg * Copyright (c) 2018 Lightbits Labs. All rights reserved. 53f2304f8SSagi Grimberg */ 63f2304f8SSagi Grimberg #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 73f2304f8SSagi Grimberg #include <linux/module.h> 83f2304f8SSagi Grimberg #include <linux/init.h> 93f2304f8SSagi Grimberg #include <linux/slab.h> 103f2304f8SSagi Grimberg #include <linux/err.h> 113f2304f8SSagi Grimberg #include <linux/nvme-tcp.h> 123f2304f8SSagi Grimberg #include <net/sock.h> 133f2304f8SSagi Grimberg #include <net/tcp.h> 143f2304f8SSagi Grimberg #include <linux/blk-mq.h> 153f2304f8SSagi Grimberg #include <crypto/hash.h> 161a9460ceSSagi Grimberg #include <net/busy_poll.h> 173f2304f8SSagi Grimberg 183f2304f8SSagi Grimberg #include "nvme.h" 193f2304f8SSagi Grimberg #include "fabrics.h" 203f2304f8SSagi Grimberg 213f2304f8SSagi Grimberg struct nvme_tcp_queue; 223f2304f8SSagi Grimberg 239912ade3SWunderlich, Mark /* Define the socket priority to use for connections were it is desirable 249912ade3SWunderlich, Mark * that the NIC consider performing optimized packet processing or filtering. 259912ade3SWunderlich, Mark * A non-zero value being sufficient to indicate general consideration of any 269912ade3SWunderlich, Mark * possible optimization. Making it a module param allows for alternative 279912ade3SWunderlich, Mark * values that may be unique for some NIC implementations. 289912ade3SWunderlich, Mark */ 299912ade3SWunderlich, Mark static int so_priority; 309912ade3SWunderlich, Mark module_param(so_priority, int, 0644); 319912ade3SWunderlich, Mark MODULE_PARM_DESC(so_priority, "nvme tcp socket optimize priority"); 329912ade3SWunderlich, Mark 33841aee4dSChris Leech #ifdef CONFIG_DEBUG_LOCK_ALLOC 34841aee4dSChris Leech /* lockdep can detect a circular dependency of the form 35841aee4dSChris Leech * sk_lock -> mmap_lock (page fault) -> fs locks -> sk_lock 36841aee4dSChris Leech * because dependencies are tracked for both nvme-tcp and user contexts. Using 37841aee4dSChris Leech * a separate class prevents lockdep from conflating nvme-tcp socket use with 38841aee4dSChris Leech * user-space socket API use. 39841aee4dSChris Leech */ 40841aee4dSChris Leech static struct lock_class_key nvme_tcp_sk_key[2]; 41841aee4dSChris Leech static struct lock_class_key nvme_tcp_slock_key[2]; 42841aee4dSChris Leech 43841aee4dSChris Leech static void nvme_tcp_reclassify_socket(struct socket *sock) 44841aee4dSChris Leech { 45841aee4dSChris Leech struct sock *sk = sock->sk; 46841aee4dSChris Leech 47841aee4dSChris Leech if (WARN_ON_ONCE(!sock_allow_reclassification(sk))) 48841aee4dSChris Leech return; 49841aee4dSChris Leech 50841aee4dSChris Leech switch (sk->sk_family) { 51841aee4dSChris Leech case AF_INET: 52841aee4dSChris Leech sock_lock_init_class_and_name(sk, "slock-AF_INET-NVME", 53841aee4dSChris Leech &nvme_tcp_slock_key[0], 54841aee4dSChris Leech "sk_lock-AF_INET-NVME", 55841aee4dSChris Leech &nvme_tcp_sk_key[0]); 56841aee4dSChris Leech break; 57841aee4dSChris Leech case AF_INET6: 58841aee4dSChris Leech sock_lock_init_class_and_name(sk, "slock-AF_INET6-NVME", 59841aee4dSChris Leech &nvme_tcp_slock_key[1], 60841aee4dSChris Leech "sk_lock-AF_INET6-NVME", 61841aee4dSChris Leech &nvme_tcp_sk_key[1]); 62841aee4dSChris Leech break; 63841aee4dSChris Leech default: 64841aee4dSChris Leech WARN_ON_ONCE(1); 65841aee4dSChris Leech } 66841aee4dSChris Leech } 67841aee4dSChris Leech #else 68841aee4dSChris Leech static void nvme_tcp_reclassify_socket(struct socket *sock) { } 69841aee4dSChris Leech #endif 70841aee4dSChris Leech 713f2304f8SSagi Grimberg enum nvme_tcp_send_state { 723f2304f8SSagi Grimberg NVME_TCP_SEND_CMD_PDU = 0, 733f2304f8SSagi Grimberg NVME_TCP_SEND_H2C_PDU, 743f2304f8SSagi Grimberg NVME_TCP_SEND_DATA, 753f2304f8SSagi Grimberg NVME_TCP_SEND_DDGST, 763f2304f8SSagi Grimberg }; 773f2304f8SSagi Grimberg 783f2304f8SSagi Grimberg struct nvme_tcp_request { 793f2304f8SSagi Grimberg struct nvme_request req; 803f2304f8SSagi Grimberg void *pdu; 813f2304f8SSagi Grimberg struct nvme_tcp_queue *queue; 823f2304f8SSagi Grimberg u32 data_len; 833f2304f8SSagi Grimberg u32 pdu_len; 843f2304f8SSagi Grimberg u32 pdu_sent; 85c2700d28SVarun Prakash u32 h2cdata_left; 86c2700d28SVarun Prakash u32 h2cdata_offset; 873f2304f8SSagi Grimberg u16 ttag; 881ba2e507SDaniel Wagner __le16 status; 893f2304f8SSagi Grimberg struct list_head entry; 9015ec928aSSagi Grimberg struct llist_node lentry; 91a7273d40SChristoph Hellwig __le32 ddgst; 923f2304f8SSagi Grimberg 933f2304f8SSagi Grimberg struct bio *curr_bio; 943f2304f8SSagi Grimberg struct iov_iter iter; 953f2304f8SSagi Grimberg 963f2304f8SSagi Grimberg /* send state */ 973f2304f8SSagi Grimberg size_t offset; 983f2304f8SSagi Grimberg size_t data_sent; 993f2304f8SSagi Grimberg enum nvme_tcp_send_state state; 1003f2304f8SSagi Grimberg }; 1013f2304f8SSagi Grimberg 1023f2304f8SSagi Grimberg enum nvme_tcp_queue_flags { 1033f2304f8SSagi Grimberg NVME_TCP_Q_ALLOCATED = 0, 1043f2304f8SSagi Grimberg NVME_TCP_Q_LIVE = 1, 10572e5d757SSagi Grimberg NVME_TCP_Q_POLLING = 2, 1063f2304f8SSagi Grimberg }; 1073f2304f8SSagi Grimberg 1083f2304f8SSagi Grimberg enum nvme_tcp_recv_state { 1093f2304f8SSagi Grimberg NVME_TCP_RECV_PDU = 0, 1103f2304f8SSagi Grimberg NVME_TCP_RECV_DATA, 1113f2304f8SSagi Grimberg NVME_TCP_RECV_DDGST, 1123f2304f8SSagi Grimberg }; 1133f2304f8SSagi Grimberg 1143f2304f8SSagi Grimberg struct nvme_tcp_ctrl; 1153f2304f8SSagi Grimberg struct nvme_tcp_queue { 1163f2304f8SSagi Grimberg struct socket *sock; 1173f2304f8SSagi Grimberg struct work_struct io_work; 1183f2304f8SSagi Grimberg int io_cpu; 1193f2304f8SSagi Grimberg 1209ebbfe49SChao Leng struct mutex queue_lock; 121db5ad6b7SSagi Grimberg struct mutex send_mutex; 12215ec928aSSagi Grimberg struct llist_head req_list; 1233f2304f8SSagi Grimberg struct list_head send_list; 1243f2304f8SSagi Grimberg 1253f2304f8SSagi Grimberg /* recv state */ 1263f2304f8SSagi Grimberg void *pdu; 1273f2304f8SSagi Grimberg int pdu_remaining; 1283f2304f8SSagi Grimberg int pdu_offset; 1293f2304f8SSagi Grimberg size_t data_remaining; 1303f2304f8SSagi Grimberg size_t ddgst_remaining; 1311a9460ceSSagi Grimberg unsigned int nr_cqe; 1323f2304f8SSagi Grimberg 1333f2304f8SSagi Grimberg /* send state */ 1343f2304f8SSagi Grimberg struct nvme_tcp_request *request; 1353f2304f8SSagi Grimberg 136c2700d28SVarun Prakash u32 maxh2cdata; 1373f2304f8SSagi Grimberg size_t cmnd_capsule_len; 1383f2304f8SSagi Grimberg struct nvme_tcp_ctrl *ctrl; 1393f2304f8SSagi Grimberg unsigned long flags; 1403f2304f8SSagi Grimberg bool rd_enabled; 1413f2304f8SSagi Grimberg 1423f2304f8SSagi Grimberg bool hdr_digest; 1433f2304f8SSagi Grimberg bool data_digest; 1443f2304f8SSagi Grimberg struct ahash_request *rcv_hash; 1453f2304f8SSagi Grimberg struct ahash_request *snd_hash; 1463f2304f8SSagi Grimberg __le32 exp_ddgst; 1473f2304f8SSagi Grimberg __le32 recv_ddgst; 1483f2304f8SSagi Grimberg 1493f2304f8SSagi Grimberg struct page_frag_cache pf_cache; 1503f2304f8SSagi Grimberg 1513f2304f8SSagi Grimberg void (*state_change)(struct sock *); 1523f2304f8SSagi Grimberg void (*data_ready)(struct sock *); 1533f2304f8SSagi Grimberg void (*write_space)(struct sock *); 1543f2304f8SSagi Grimberg }; 1553f2304f8SSagi Grimberg 1563f2304f8SSagi Grimberg struct nvme_tcp_ctrl { 1573f2304f8SSagi Grimberg /* read only in the hot path */ 1583f2304f8SSagi Grimberg struct nvme_tcp_queue *queues; 1593f2304f8SSagi Grimberg struct blk_mq_tag_set tag_set; 1603f2304f8SSagi Grimberg 1613f2304f8SSagi Grimberg /* other member variables */ 1623f2304f8SSagi Grimberg struct list_head list; 1633f2304f8SSagi Grimberg struct blk_mq_tag_set admin_tag_set; 1643f2304f8SSagi Grimberg struct sockaddr_storage addr; 1653f2304f8SSagi Grimberg struct sockaddr_storage src_addr; 1663f2304f8SSagi Grimberg struct nvme_ctrl ctrl; 1673f2304f8SSagi Grimberg 1683f2304f8SSagi Grimberg struct work_struct err_work; 1693f2304f8SSagi Grimberg struct delayed_work connect_work; 1703f2304f8SSagi Grimberg struct nvme_tcp_request async_req; 17164861993SSagi Grimberg u32 io_queues[HCTX_MAX_TYPES]; 1723f2304f8SSagi Grimberg }; 1733f2304f8SSagi Grimberg 1743f2304f8SSagi Grimberg static LIST_HEAD(nvme_tcp_ctrl_list); 1753f2304f8SSagi Grimberg static DEFINE_MUTEX(nvme_tcp_ctrl_mutex); 1763f2304f8SSagi Grimberg static struct workqueue_struct *nvme_tcp_wq; 1776acbd961SRikard Falkeborn static const struct blk_mq_ops nvme_tcp_mq_ops; 1786acbd961SRikard Falkeborn static const struct blk_mq_ops nvme_tcp_admin_mq_ops; 179db5ad6b7SSagi Grimberg static int nvme_tcp_try_send(struct nvme_tcp_queue *queue); 1803f2304f8SSagi Grimberg 1813f2304f8SSagi Grimberg static inline struct nvme_tcp_ctrl *to_tcp_ctrl(struct nvme_ctrl *ctrl) 1823f2304f8SSagi Grimberg { 1833f2304f8SSagi Grimberg return container_of(ctrl, struct nvme_tcp_ctrl, ctrl); 1843f2304f8SSagi Grimberg } 1853f2304f8SSagi Grimberg 1863f2304f8SSagi Grimberg static inline int nvme_tcp_queue_id(struct nvme_tcp_queue *queue) 1873f2304f8SSagi Grimberg { 1883f2304f8SSagi Grimberg return queue - queue->ctrl->queues; 1893f2304f8SSagi Grimberg } 1903f2304f8SSagi Grimberg 1913f2304f8SSagi Grimberg static inline struct blk_mq_tags *nvme_tcp_tagset(struct nvme_tcp_queue *queue) 1923f2304f8SSagi Grimberg { 1933f2304f8SSagi Grimberg u32 queue_idx = nvme_tcp_queue_id(queue); 1943f2304f8SSagi Grimberg 1953f2304f8SSagi Grimberg if (queue_idx == 0) 1963f2304f8SSagi Grimberg return queue->ctrl->admin_tag_set.tags[queue_idx]; 1973f2304f8SSagi Grimberg return queue->ctrl->tag_set.tags[queue_idx - 1]; 1983f2304f8SSagi Grimberg } 1993f2304f8SSagi Grimberg 2003f2304f8SSagi Grimberg static inline u8 nvme_tcp_hdgst_len(struct nvme_tcp_queue *queue) 2013f2304f8SSagi Grimberg { 2023f2304f8SSagi Grimberg return queue->hdr_digest ? NVME_TCP_DIGEST_LENGTH : 0; 2033f2304f8SSagi Grimberg } 2043f2304f8SSagi Grimberg 2053f2304f8SSagi Grimberg static inline u8 nvme_tcp_ddgst_len(struct nvme_tcp_queue *queue) 2063f2304f8SSagi Grimberg { 2073f2304f8SSagi Grimberg return queue->data_digest ? NVME_TCP_DIGEST_LENGTH : 0; 2083f2304f8SSagi Grimberg } 2093f2304f8SSagi Grimberg 21053ee9e29SCaleb Sander static inline size_t nvme_tcp_inline_data_size(struct nvme_tcp_request *req) 2113f2304f8SSagi Grimberg { 21253ee9e29SCaleb Sander if (nvme_is_fabrics(req->req.cmd)) 21353ee9e29SCaleb Sander return NVME_TCP_ADMIN_CCSZ; 21453ee9e29SCaleb Sander return req->queue->cmnd_capsule_len - sizeof(struct nvme_command); 2153f2304f8SSagi Grimberg } 2163f2304f8SSagi Grimberg 2173f2304f8SSagi Grimberg static inline bool nvme_tcp_async_req(struct nvme_tcp_request *req) 2183f2304f8SSagi Grimberg { 2193f2304f8SSagi Grimberg return req == &req->queue->ctrl->async_req; 2203f2304f8SSagi Grimberg } 2213f2304f8SSagi Grimberg 2223f2304f8SSagi Grimberg static inline bool nvme_tcp_has_inline_data(struct nvme_tcp_request *req) 2233f2304f8SSagi Grimberg { 2243f2304f8SSagi Grimberg struct request *rq; 2253f2304f8SSagi Grimberg 2263f2304f8SSagi Grimberg if (unlikely(nvme_tcp_async_req(req))) 2273f2304f8SSagi Grimberg return false; /* async events don't have a request */ 2283f2304f8SSagi Grimberg 2293f2304f8SSagi Grimberg rq = blk_mq_rq_from_pdu(req); 2303f2304f8SSagi Grimberg 23125e5cb78SSagi Grimberg return rq_data_dir(rq) == WRITE && req->data_len && 23253ee9e29SCaleb Sander req->data_len <= nvme_tcp_inline_data_size(req); 2333f2304f8SSagi Grimberg } 2343f2304f8SSagi Grimberg 2353f2304f8SSagi Grimberg static inline struct page *nvme_tcp_req_cur_page(struct nvme_tcp_request *req) 2363f2304f8SSagi Grimberg { 2373f2304f8SSagi Grimberg return req->iter.bvec->bv_page; 2383f2304f8SSagi Grimberg } 2393f2304f8SSagi Grimberg 2403f2304f8SSagi Grimberg static inline size_t nvme_tcp_req_cur_offset(struct nvme_tcp_request *req) 2413f2304f8SSagi Grimberg { 2423f2304f8SSagi Grimberg return req->iter.bvec->bv_offset + req->iter.iov_offset; 2433f2304f8SSagi Grimberg } 2443f2304f8SSagi Grimberg 2453f2304f8SSagi Grimberg static inline size_t nvme_tcp_req_cur_length(struct nvme_tcp_request *req) 2463f2304f8SSagi Grimberg { 247ca1ff67dSSagi Grimberg return min_t(size_t, iov_iter_single_seg_count(&req->iter), 2483f2304f8SSagi Grimberg req->pdu_len - req->pdu_sent); 2493f2304f8SSagi Grimberg } 2503f2304f8SSagi Grimberg 2513f2304f8SSagi Grimberg static inline size_t nvme_tcp_pdu_data_left(struct nvme_tcp_request *req) 2523f2304f8SSagi Grimberg { 2533f2304f8SSagi Grimberg return rq_data_dir(blk_mq_rq_from_pdu(req)) == WRITE ? 2543f2304f8SSagi Grimberg req->pdu_len - req->pdu_sent : 0; 2553f2304f8SSagi Grimberg } 2563f2304f8SSagi Grimberg 2573f2304f8SSagi Grimberg static inline size_t nvme_tcp_pdu_last_send(struct nvme_tcp_request *req, 2583f2304f8SSagi Grimberg int len) 2593f2304f8SSagi Grimberg { 2603f2304f8SSagi Grimberg return nvme_tcp_pdu_data_left(req) <= len; 2613f2304f8SSagi Grimberg } 2623f2304f8SSagi Grimberg 2633f2304f8SSagi Grimberg static void nvme_tcp_init_iter(struct nvme_tcp_request *req, 2643f2304f8SSagi Grimberg unsigned int dir) 2653f2304f8SSagi Grimberg { 2663f2304f8SSagi Grimberg struct request *rq = blk_mq_rq_from_pdu(req); 2673f2304f8SSagi Grimberg struct bio_vec *vec; 2683f2304f8SSagi Grimberg unsigned int size; 2690dc9edafSSagi Grimberg int nr_bvec; 2703f2304f8SSagi Grimberg size_t offset; 2713f2304f8SSagi Grimberg 2723f2304f8SSagi Grimberg if (rq->rq_flags & RQF_SPECIAL_PAYLOAD) { 2733f2304f8SSagi Grimberg vec = &rq->special_vec; 2740dc9edafSSagi Grimberg nr_bvec = 1; 2753f2304f8SSagi Grimberg size = blk_rq_payload_bytes(rq); 2763f2304f8SSagi Grimberg offset = 0; 2773f2304f8SSagi Grimberg } else { 2783f2304f8SSagi Grimberg struct bio *bio = req->curr_bio; 2790dc9edafSSagi Grimberg struct bvec_iter bi; 2800dc9edafSSagi Grimberg struct bio_vec bv; 2813f2304f8SSagi Grimberg 2823f2304f8SSagi Grimberg vec = __bvec_iter_bvec(bio->bi_io_vec, bio->bi_iter); 2830dc9edafSSagi Grimberg nr_bvec = 0; 2840dc9edafSSagi Grimberg bio_for_each_bvec(bv, bio, bi) { 2850dc9edafSSagi Grimberg nr_bvec++; 2860dc9edafSSagi Grimberg } 2873f2304f8SSagi Grimberg size = bio->bi_iter.bi_size; 2883f2304f8SSagi Grimberg offset = bio->bi_iter.bi_bvec_done; 2893f2304f8SSagi Grimberg } 2903f2304f8SSagi Grimberg 2910dc9edafSSagi Grimberg iov_iter_bvec(&req->iter, dir, vec, nr_bvec, size); 2923f2304f8SSagi Grimberg req->iter.iov_offset = offset; 2933f2304f8SSagi Grimberg } 2943f2304f8SSagi Grimberg 2953f2304f8SSagi Grimberg static inline void nvme_tcp_advance_req(struct nvme_tcp_request *req, 2963f2304f8SSagi Grimberg int len) 2973f2304f8SSagi Grimberg { 2983f2304f8SSagi Grimberg req->data_sent += len; 2993f2304f8SSagi Grimberg req->pdu_sent += len; 3003f2304f8SSagi Grimberg iov_iter_advance(&req->iter, len); 3013f2304f8SSagi Grimberg if (!iov_iter_count(&req->iter) && 3023f2304f8SSagi Grimberg req->data_sent < req->data_len) { 3033f2304f8SSagi Grimberg req->curr_bio = req->curr_bio->bi_next; 3043f2304f8SSagi Grimberg nvme_tcp_init_iter(req, WRITE); 3053f2304f8SSagi Grimberg } 3063f2304f8SSagi Grimberg } 3073f2304f8SSagi Grimberg 3085c11f7d9SSagi Grimberg static inline void nvme_tcp_send_all(struct nvme_tcp_queue *queue) 3095c11f7d9SSagi Grimberg { 3105c11f7d9SSagi Grimberg int ret; 3115c11f7d9SSagi Grimberg 3125c11f7d9SSagi Grimberg /* drain the send queue as much as we can... */ 3135c11f7d9SSagi Grimberg do { 3145c11f7d9SSagi Grimberg ret = nvme_tcp_try_send(queue); 3155c11f7d9SSagi Grimberg } while (ret > 0); 3165c11f7d9SSagi Grimberg } 3175c11f7d9SSagi Grimberg 31870f437fbSKeith Busch static inline bool nvme_tcp_queue_more(struct nvme_tcp_queue *queue) 31970f437fbSKeith Busch { 32070f437fbSKeith Busch return !list_empty(&queue->send_list) || 3213770a42bSSagi Grimberg !llist_empty(&queue->req_list); 32270f437fbSKeith Busch } 32370f437fbSKeith Busch 324db5ad6b7SSagi Grimberg static inline void nvme_tcp_queue_request(struct nvme_tcp_request *req, 32586f0348aSSagi Grimberg bool sync, bool last) 3263f2304f8SSagi Grimberg { 3273f2304f8SSagi Grimberg struct nvme_tcp_queue *queue = req->queue; 328db5ad6b7SSagi Grimberg bool empty; 3293f2304f8SSagi Grimberg 33015ec928aSSagi Grimberg empty = llist_add(&req->lentry, &queue->req_list) && 33115ec928aSSagi Grimberg list_empty(&queue->send_list) && !queue->request; 3323f2304f8SSagi Grimberg 333db5ad6b7SSagi Grimberg /* 334db5ad6b7SSagi Grimberg * if we're the first on the send_list and we can try to send 335db5ad6b7SSagi Grimberg * directly, otherwise queue io_work. Also, only do that if we 336db5ad6b7SSagi Grimberg * are on the same cpu, so we don't introduce contention. 337db5ad6b7SSagi Grimberg */ 338bb833370SSagi Grimberg if (queue->io_cpu == raw_smp_processor_id() && 339db5ad6b7SSagi Grimberg sync && empty && mutex_trylock(&queue->send_mutex)) { 3405c11f7d9SSagi Grimberg nvme_tcp_send_all(queue); 341db5ad6b7SSagi Grimberg mutex_unlock(&queue->send_mutex); 3423f2304f8SSagi Grimberg } 34370f437fbSKeith Busch 34470f437fbSKeith Busch if (last && nvme_tcp_queue_more(queue)) 34570f437fbSKeith Busch queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work); 346db5ad6b7SSagi Grimberg } 3473f2304f8SSagi Grimberg 34815ec928aSSagi Grimberg static void nvme_tcp_process_req_list(struct nvme_tcp_queue *queue) 34915ec928aSSagi Grimberg { 35015ec928aSSagi Grimberg struct nvme_tcp_request *req; 35115ec928aSSagi Grimberg struct llist_node *node; 35215ec928aSSagi Grimberg 35315ec928aSSagi Grimberg for (node = llist_del_all(&queue->req_list); node; node = node->next) { 35415ec928aSSagi Grimberg req = llist_entry(node, struct nvme_tcp_request, lentry); 35515ec928aSSagi Grimberg list_add(&req->entry, &queue->send_list); 35615ec928aSSagi Grimberg } 35715ec928aSSagi Grimberg } 35815ec928aSSagi Grimberg 3593f2304f8SSagi Grimberg static inline struct nvme_tcp_request * 3603f2304f8SSagi Grimberg nvme_tcp_fetch_request(struct nvme_tcp_queue *queue) 3613f2304f8SSagi Grimberg { 3623f2304f8SSagi Grimberg struct nvme_tcp_request *req; 3633f2304f8SSagi Grimberg 3643f2304f8SSagi Grimberg req = list_first_entry_or_null(&queue->send_list, 3653f2304f8SSagi Grimberg struct nvme_tcp_request, entry); 36615ec928aSSagi Grimberg if (!req) { 36715ec928aSSagi Grimberg nvme_tcp_process_req_list(queue); 36815ec928aSSagi Grimberg req = list_first_entry_or_null(&queue->send_list, 36915ec928aSSagi Grimberg struct nvme_tcp_request, entry); 37015ec928aSSagi Grimberg if (unlikely(!req)) 37115ec928aSSagi Grimberg return NULL; 37215ec928aSSagi Grimberg } 3733f2304f8SSagi Grimberg 37415ec928aSSagi Grimberg list_del(&req->entry); 3753f2304f8SSagi Grimberg return req; 3763f2304f8SSagi Grimberg } 3773f2304f8SSagi Grimberg 378a7273d40SChristoph Hellwig static inline void nvme_tcp_ddgst_final(struct ahash_request *hash, 379a7273d40SChristoph Hellwig __le32 *dgst) 3803f2304f8SSagi Grimberg { 3813f2304f8SSagi Grimberg ahash_request_set_crypt(hash, NULL, (u8 *)dgst, 0); 3823f2304f8SSagi Grimberg crypto_ahash_final(hash); 3833f2304f8SSagi Grimberg } 3843f2304f8SSagi Grimberg 3853f2304f8SSagi Grimberg static inline void nvme_tcp_ddgst_update(struct ahash_request *hash, 3863f2304f8SSagi Grimberg struct page *page, off_t off, size_t len) 3873f2304f8SSagi Grimberg { 3883f2304f8SSagi Grimberg struct scatterlist sg; 3893f2304f8SSagi Grimberg 3903f2304f8SSagi Grimberg sg_init_marker(&sg, 1); 3913f2304f8SSagi Grimberg sg_set_page(&sg, page, len, off); 3923f2304f8SSagi Grimberg ahash_request_set_crypt(hash, &sg, NULL, len); 3933f2304f8SSagi Grimberg crypto_ahash_update(hash); 3943f2304f8SSagi Grimberg } 3953f2304f8SSagi Grimberg 3963f2304f8SSagi Grimberg static inline void nvme_tcp_hdgst(struct ahash_request *hash, 3973f2304f8SSagi Grimberg void *pdu, size_t len) 3983f2304f8SSagi Grimberg { 3993f2304f8SSagi Grimberg struct scatterlist sg; 4003f2304f8SSagi Grimberg 4013f2304f8SSagi Grimberg sg_init_one(&sg, pdu, len); 4023f2304f8SSagi Grimberg ahash_request_set_crypt(hash, &sg, pdu + len, len); 4033f2304f8SSagi Grimberg crypto_ahash_digest(hash); 4043f2304f8SSagi Grimberg } 4053f2304f8SSagi Grimberg 4063f2304f8SSagi Grimberg static int nvme_tcp_verify_hdgst(struct nvme_tcp_queue *queue, 4073f2304f8SSagi Grimberg void *pdu, size_t pdu_len) 4083f2304f8SSagi Grimberg { 4093f2304f8SSagi Grimberg struct nvme_tcp_hdr *hdr = pdu; 4103f2304f8SSagi Grimberg __le32 recv_digest; 4113f2304f8SSagi Grimberg __le32 exp_digest; 4123f2304f8SSagi Grimberg 4133f2304f8SSagi Grimberg if (unlikely(!(hdr->flags & NVME_TCP_F_HDGST))) { 4143f2304f8SSagi Grimberg dev_err(queue->ctrl->ctrl.device, 4153f2304f8SSagi Grimberg "queue %d: header digest flag is cleared\n", 4163f2304f8SSagi Grimberg nvme_tcp_queue_id(queue)); 4173f2304f8SSagi Grimberg return -EPROTO; 4183f2304f8SSagi Grimberg } 4193f2304f8SSagi Grimberg 4203f2304f8SSagi Grimberg recv_digest = *(__le32 *)(pdu + hdr->hlen); 4213f2304f8SSagi Grimberg nvme_tcp_hdgst(queue->rcv_hash, pdu, pdu_len); 4223f2304f8SSagi Grimberg exp_digest = *(__le32 *)(pdu + hdr->hlen); 4233f2304f8SSagi Grimberg if (recv_digest != exp_digest) { 4243f2304f8SSagi Grimberg dev_err(queue->ctrl->ctrl.device, 4253f2304f8SSagi Grimberg "header digest error: recv %#x expected %#x\n", 4263f2304f8SSagi Grimberg le32_to_cpu(recv_digest), le32_to_cpu(exp_digest)); 4273f2304f8SSagi Grimberg return -EIO; 4283f2304f8SSagi Grimberg } 4293f2304f8SSagi Grimberg 4303f2304f8SSagi Grimberg return 0; 4313f2304f8SSagi Grimberg } 4323f2304f8SSagi Grimberg 4333f2304f8SSagi Grimberg static int nvme_tcp_check_ddgst(struct nvme_tcp_queue *queue, void *pdu) 4343f2304f8SSagi Grimberg { 4353f2304f8SSagi Grimberg struct nvme_tcp_hdr *hdr = pdu; 4363f2304f8SSagi Grimberg u8 digest_len = nvme_tcp_hdgst_len(queue); 4373f2304f8SSagi Grimberg u32 len; 4383f2304f8SSagi Grimberg 4393f2304f8SSagi Grimberg len = le32_to_cpu(hdr->plen) - hdr->hlen - 4403f2304f8SSagi Grimberg ((hdr->flags & NVME_TCP_F_HDGST) ? digest_len : 0); 4413f2304f8SSagi Grimberg 4423f2304f8SSagi Grimberg if (unlikely(len && !(hdr->flags & NVME_TCP_F_DDGST))) { 4433f2304f8SSagi Grimberg dev_err(queue->ctrl->ctrl.device, 4443f2304f8SSagi Grimberg "queue %d: data digest flag is cleared\n", 4453f2304f8SSagi Grimberg nvme_tcp_queue_id(queue)); 4463f2304f8SSagi Grimberg return -EPROTO; 4473f2304f8SSagi Grimberg } 4483f2304f8SSagi Grimberg crypto_ahash_init(queue->rcv_hash); 4493f2304f8SSagi Grimberg 4503f2304f8SSagi Grimberg return 0; 4513f2304f8SSagi Grimberg } 4523f2304f8SSagi Grimberg 4533f2304f8SSagi Grimberg static void nvme_tcp_exit_request(struct blk_mq_tag_set *set, 4543f2304f8SSagi Grimberg struct request *rq, unsigned int hctx_idx) 4553f2304f8SSagi Grimberg { 4563f2304f8SSagi Grimberg struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq); 4573f2304f8SSagi Grimberg 4583f2304f8SSagi Grimberg page_frag_free(req->pdu); 4593f2304f8SSagi Grimberg } 4603f2304f8SSagi Grimberg 4613f2304f8SSagi Grimberg static int nvme_tcp_init_request(struct blk_mq_tag_set *set, 4623f2304f8SSagi Grimberg struct request *rq, unsigned int hctx_idx, 4633f2304f8SSagi Grimberg unsigned int numa_node) 4643f2304f8SSagi Grimberg { 46506427ca0SChristoph Hellwig struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(set->driver_data); 4663f2304f8SSagi Grimberg struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq); 467f4b9e6c9SKeith Busch struct nvme_tcp_cmd_pdu *pdu; 4683f2304f8SSagi Grimberg int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0; 4693f2304f8SSagi Grimberg struct nvme_tcp_queue *queue = &ctrl->queues[queue_idx]; 4703f2304f8SSagi Grimberg u8 hdgst = nvme_tcp_hdgst_len(queue); 4713f2304f8SSagi Grimberg 4723f2304f8SSagi Grimberg req->pdu = page_frag_alloc(&queue->pf_cache, 4733f2304f8SSagi Grimberg sizeof(struct nvme_tcp_cmd_pdu) + hdgst, 4743f2304f8SSagi Grimberg GFP_KERNEL | __GFP_ZERO); 4753f2304f8SSagi Grimberg if (!req->pdu) 4763f2304f8SSagi Grimberg return -ENOMEM; 4773f2304f8SSagi Grimberg 478f4b9e6c9SKeith Busch pdu = req->pdu; 4793f2304f8SSagi Grimberg req->queue = queue; 4803f2304f8SSagi Grimberg nvme_req(rq)->ctrl = &ctrl->ctrl; 481f4b9e6c9SKeith Busch nvme_req(rq)->cmd = &pdu->cmd; 4823f2304f8SSagi Grimberg 4833f2304f8SSagi Grimberg return 0; 4843f2304f8SSagi Grimberg } 4853f2304f8SSagi Grimberg 4863f2304f8SSagi Grimberg static int nvme_tcp_init_hctx(struct blk_mq_hw_ctx *hctx, void *data, 4873f2304f8SSagi Grimberg unsigned int hctx_idx) 4883f2304f8SSagi Grimberg { 48906427ca0SChristoph Hellwig struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(data); 4903f2304f8SSagi Grimberg struct nvme_tcp_queue *queue = &ctrl->queues[hctx_idx + 1]; 4913f2304f8SSagi Grimberg 4923f2304f8SSagi Grimberg hctx->driver_data = queue; 4933f2304f8SSagi Grimberg return 0; 4943f2304f8SSagi Grimberg } 4953f2304f8SSagi Grimberg 4963f2304f8SSagi Grimberg static int nvme_tcp_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data, 4973f2304f8SSagi Grimberg unsigned int hctx_idx) 4983f2304f8SSagi Grimberg { 49906427ca0SChristoph Hellwig struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(data); 5003f2304f8SSagi Grimberg struct nvme_tcp_queue *queue = &ctrl->queues[0]; 5013f2304f8SSagi Grimberg 5023f2304f8SSagi Grimberg hctx->driver_data = queue; 5033f2304f8SSagi Grimberg return 0; 5043f2304f8SSagi Grimberg } 5053f2304f8SSagi Grimberg 5063f2304f8SSagi Grimberg static enum nvme_tcp_recv_state 5073f2304f8SSagi Grimberg nvme_tcp_recv_state(struct nvme_tcp_queue *queue) 5083f2304f8SSagi Grimberg { 5093f2304f8SSagi Grimberg return (queue->pdu_remaining) ? NVME_TCP_RECV_PDU : 5103f2304f8SSagi Grimberg (queue->ddgst_remaining) ? NVME_TCP_RECV_DDGST : 5113f2304f8SSagi Grimberg NVME_TCP_RECV_DATA; 5123f2304f8SSagi Grimberg } 5133f2304f8SSagi Grimberg 5143f2304f8SSagi Grimberg static void nvme_tcp_init_recv_ctx(struct nvme_tcp_queue *queue) 5153f2304f8SSagi Grimberg { 5163f2304f8SSagi Grimberg queue->pdu_remaining = sizeof(struct nvme_tcp_rsp_pdu) + 5173f2304f8SSagi Grimberg nvme_tcp_hdgst_len(queue); 5183f2304f8SSagi Grimberg queue->pdu_offset = 0; 5193f2304f8SSagi Grimberg queue->data_remaining = -1; 5203f2304f8SSagi Grimberg queue->ddgst_remaining = 0; 5213f2304f8SSagi Grimberg } 5223f2304f8SSagi Grimberg 5233f2304f8SSagi Grimberg static void nvme_tcp_error_recovery(struct nvme_ctrl *ctrl) 5243f2304f8SSagi Grimberg { 5253f2304f8SSagi Grimberg if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_RESETTING)) 5263f2304f8SSagi Grimberg return; 5273f2304f8SSagi Grimberg 528236187c4SSagi Grimberg dev_warn(ctrl->device, "starting error recovery\n"); 52997b2512aSNigel Kirkland queue_work(nvme_reset_wq, &to_tcp_ctrl(ctrl)->err_work); 5303f2304f8SSagi Grimberg } 5313f2304f8SSagi Grimberg 5323f2304f8SSagi Grimberg static int nvme_tcp_process_nvme_cqe(struct nvme_tcp_queue *queue, 5333f2304f8SSagi Grimberg struct nvme_completion *cqe) 5343f2304f8SSagi Grimberg { 5351ba2e507SDaniel Wagner struct nvme_tcp_request *req; 5363f2304f8SSagi Grimberg struct request *rq; 5373f2304f8SSagi Grimberg 538e7006de6SSagi Grimberg rq = nvme_find_rq(nvme_tcp_tagset(queue), cqe->command_id); 5393f2304f8SSagi Grimberg if (!rq) { 5403f2304f8SSagi Grimberg dev_err(queue->ctrl->ctrl.device, 541e7006de6SSagi Grimberg "got bad cqe.command_id %#x on queue %d\n", 542e7006de6SSagi Grimberg cqe->command_id, nvme_tcp_queue_id(queue)); 5433f2304f8SSagi Grimberg nvme_tcp_error_recovery(&queue->ctrl->ctrl); 5443f2304f8SSagi Grimberg return -EINVAL; 5453f2304f8SSagi Grimberg } 5463f2304f8SSagi Grimberg 5471ba2e507SDaniel Wagner req = blk_mq_rq_to_pdu(rq); 5481ba2e507SDaniel Wagner if (req->status == cpu_to_le16(NVME_SC_SUCCESS)) 5491ba2e507SDaniel Wagner req->status = cqe->status; 5501ba2e507SDaniel Wagner 5511ba2e507SDaniel Wagner if (!nvme_try_complete_req(rq, req->status, cqe->result)) 552ff029451SChristoph Hellwig nvme_complete_rq(rq); 5531a9460ceSSagi Grimberg queue->nr_cqe++; 5543f2304f8SSagi Grimberg 5553f2304f8SSagi Grimberg return 0; 5563f2304f8SSagi Grimberg } 5573f2304f8SSagi Grimberg 5583f2304f8SSagi Grimberg static int nvme_tcp_handle_c2h_data(struct nvme_tcp_queue *queue, 5593f2304f8SSagi Grimberg struct nvme_tcp_data_pdu *pdu) 5603f2304f8SSagi Grimberg { 5613f2304f8SSagi Grimberg struct request *rq; 5623f2304f8SSagi Grimberg 563e7006de6SSagi Grimberg rq = nvme_find_rq(nvme_tcp_tagset(queue), pdu->command_id); 5643f2304f8SSagi Grimberg if (!rq) { 5653f2304f8SSagi Grimberg dev_err(queue->ctrl->ctrl.device, 566e7006de6SSagi Grimberg "got bad c2hdata.command_id %#x on queue %d\n", 567e7006de6SSagi Grimberg pdu->command_id, nvme_tcp_queue_id(queue)); 5683f2304f8SSagi Grimberg return -ENOENT; 5693f2304f8SSagi Grimberg } 5703f2304f8SSagi Grimberg 5713f2304f8SSagi Grimberg if (!blk_rq_payload_bytes(rq)) { 5723f2304f8SSagi Grimberg dev_err(queue->ctrl->ctrl.device, 5733f2304f8SSagi Grimberg "queue %d tag %#x unexpected data\n", 5743f2304f8SSagi Grimberg nvme_tcp_queue_id(queue), rq->tag); 5753f2304f8SSagi Grimberg return -EIO; 5763f2304f8SSagi Grimberg } 5773f2304f8SSagi Grimberg 5783f2304f8SSagi Grimberg queue->data_remaining = le32_to_cpu(pdu->data_length); 5793f2304f8SSagi Grimberg 580602d674cSSagi Grimberg if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS && 581602d674cSSagi Grimberg unlikely(!(pdu->hdr.flags & NVME_TCP_F_DATA_LAST))) { 582602d674cSSagi Grimberg dev_err(queue->ctrl->ctrl.device, 583602d674cSSagi Grimberg "queue %d tag %#x SUCCESS set but not last PDU\n", 584602d674cSSagi Grimberg nvme_tcp_queue_id(queue), rq->tag); 585602d674cSSagi Grimberg nvme_tcp_error_recovery(&queue->ctrl->ctrl); 586602d674cSSagi Grimberg return -EPROTO; 587602d674cSSagi Grimberg } 588602d674cSSagi Grimberg 5893f2304f8SSagi Grimberg return 0; 5903f2304f8SSagi Grimberg } 5913f2304f8SSagi Grimberg 5923f2304f8SSagi Grimberg static int nvme_tcp_handle_comp(struct nvme_tcp_queue *queue, 5933f2304f8SSagi Grimberg struct nvme_tcp_rsp_pdu *pdu) 5943f2304f8SSagi Grimberg { 5953f2304f8SSagi Grimberg struct nvme_completion *cqe = &pdu->cqe; 5963f2304f8SSagi Grimberg int ret = 0; 5973f2304f8SSagi Grimberg 5983f2304f8SSagi Grimberg /* 5993f2304f8SSagi Grimberg * AEN requests are special as they don't time out and can 6003f2304f8SSagi Grimberg * survive any kind of queue freeze and often don't respond to 6013f2304f8SSagi Grimberg * aborts. We don't even bother to allocate a struct request 6023f2304f8SSagi Grimberg * for them but rather special case them here. 6033f2304f8SSagi Grimberg */ 60458a8df67SIsrael Rukshin if (unlikely(nvme_is_aen_req(nvme_tcp_queue_id(queue), 60558a8df67SIsrael Rukshin cqe->command_id))) 6063f2304f8SSagi Grimberg nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status, 6073f2304f8SSagi Grimberg &cqe->result); 6083f2304f8SSagi Grimberg else 6093f2304f8SSagi Grimberg ret = nvme_tcp_process_nvme_cqe(queue, cqe); 6103f2304f8SSagi Grimberg 6113f2304f8SSagi Grimberg return ret; 6123f2304f8SSagi Grimberg } 6133f2304f8SSagi Grimberg 614c2700d28SVarun Prakash static void nvme_tcp_setup_h2c_data_pdu(struct nvme_tcp_request *req) 6153f2304f8SSagi Grimberg { 6163f2304f8SSagi Grimberg struct nvme_tcp_data_pdu *data = req->pdu; 6173f2304f8SSagi Grimberg struct nvme_tcp_queue *queue = req->queue; 6183f2304f8SSagi Grimberg struct request *rq = blk_mq_rq_from_pdu(req); 619c2700d28SVarun Prakash u32 h2cdata_sent = req->pdu_len; 6203f2304f8SSagi Grimberg u8 hdgst = nvme_tcp_hdgst_len(queue); 6213f2304f8SSagi Grimberg u8 ddgst = nvme_tcp_ddgst_len(queue); 6223f2304f8SSagi Grimberg 6231d3ef9c3SVarun Prakash req->state = NVME_TCP_SEND_H2C_PDU; 6241d3ef9c3SVarun Prakash req->offset = 0; 625c2700d28SVarun Prakash req->pdu_len = min(req->h2cdata_left, queue->maxh2cdata); 6263f2304f8SSagi Grimberg req->pdu_sent = 0; 627c2700d28SVarun Prakash req->h2cdata_left -= req->pdu_len; 628c2700d28SVarun Prakash req->h2cdata_offset += h2cdata_sent; 6293f2304f8SSagi Grimberg 6303f2304f8SSagi Grimberg memset(data, 0, sizeof(*data)); 6313f2304f8SSagi Grimberg data->hdr.type = nvme_tcp_h2c_data; 632c2700d28SVarun Prakash if (!req->h2cdata_left) 6333f2304f8SSagi Grimberg data->hdr.flags = NVME_TCP_F_DATA_LAST; 6343f2304f8SSagi Grimberg if (queue->hdr_digest) 6353f2304f8SSagi Grimberg data->hdr.flags |= NVME_TCP_F_HDGST; 6363f2304f8SSagi Grimberg if (queue->data_digest) 6373f2304f8SSagi Grimberg data->hdr.flags |= NVME_TCP_F_DDGST; 6383f2304f8SSagi Grimberg data->hdr.hlen = sizeof(*data); 6393f2304f8SSagi Grimberg data->hdr.pdo = data->hdr.hlen + hdgst; 6403f2304f8SSagi Grimberg data->hdr.plen = 6413f2304f8SSagi Grimberg cpu_to_le32(data->hdr.hlen + hdgst + req->pdu_len + ddgst); 642c2700d28SVarun Prakash data->ttag = req->ttag; 643e7006de6SSagi Grimberg data->command_id = nvme_cid(rq); 644c2700d28SVarun Prakash data->data_offset = cpu_to_le32(req->h2cdata_offset); 6453f2304f8SSagi Grimberg data->data_length = cpu_to_le32(req->pdu_len); 6463f2304f8SSagi Grimberg } 6473f2304f8SSagi Grimberg 6483f2304f8SSagi Grimberg static int nvme_tcp_handle_r2t(struct nvme_tcp_queue *queue, 6493f2304f8SSagi Grimberg struct nvme_tcp_r2t_pdu *pdu) 6503f2304f8SSagi Grimberg { 6513f2304f8SSagi Grimberg struct nvme_tcp_request *req; 6523f2304f8SSagi Grimberg struct request *rq; 6531d3ef9c3SVarun Prakash u32 r2t_length = le32_to_cpu(pdu->r2t_length); 654c2700d28SVarun Prakash u32 r2t_offset = le32_to_cpu(pdu->r2t_offset); 6553f2304f8SSagi Grimberg 656e7006de6SSagi Grimberg rq = nvme_find_rq(nvme_tcp_tagset(queue), pdu->command_id); 6573f2304f8SSagi Grimberg if (!rq) { 6583f2304f8SSagi Grimberg dev_err(queue->ctrl->ctrl.device, 659e7006de6SSagi Grimberg "got bad r2t.command_id %#x on queue %d\n", 660e7006de6SSagi Grimberg pdu->command_id, nvme_tcp_queue_id(queue)); 6613f2304f8SSagi Grimberg return -ENOENT; 6623f2304f8SSagi Grimberg } 6633f2304f8SSagi Grimberg req = blk_mq_rq_to_pdu(rq); 6643f2304f8SSagi Grimberg 6651d3ef9c3SVarun Prakash if (unlikely(!r2t_length)) { 6661d3ef9c3SVarun Prakash dev_err(queue->ctrl->ctrl.device, 6671d3ef9c3SVarun Prakash "req %d r2t len is %u, probably a bug...\n", 6681d3ef9c3SVarun Prakash rq->tag, r2t_length); 6691d3ef9c3SVarun Prakash return -EPROTO; 6701d3ef9c3SVarun Prakash } 6713f2304f8SSagi Grimberg 6721d3ef9c3SVarun Prakash if (unlikely(req->data_sent + r2t_length > req->data_len)) { 6731d3ef9c3SVarun Prakash dev_err(queue->ctrl->ctrl.device, 6741d3ef9c3SVarun Prakash "req %d r2t len %u exceeded data len %u (%zu sent)\n", 6751d3ef9c3SVarun Prakash rq->tag, r2t_length, req->data_len, req->data_sent); 6761d3ef9c3SVarun Prakash return -EPROTO; 6771d3ef9c3SVarun Prakash } 6783f2304f8SSagi Grimberg 679c2700d28SVarun Prakash if (unlikely(r2t_offset < req->data_sent)) { 6801d3ef9c3SVarun Prakash dev_err(queue->ctrl->ctrl.device, 6811d3ef9c3SVarun Prakash "req %d unexpected r2t offset %u (expected %zu)\n", 682c2700d28SVarun Prakash rq->tag, r2t_offset, req->data_sent); 6831d3ef9c3SVarun Prakash return -EPROTO; 6841d3ef9c3SVarun Prakash } 6851d3ef9c3SVarun Prakash 686c2700d28SVarun Prakash req->pdu_len = 0; 687c2700d28SVarun Prakash req->h2cdata_left = r2t_length; 688c2700d28SVarun Prakash req->h2cdata_offset = r2t_offset; 689c2700d28SVarun Prakash req->ttag = pdu->ttag; 690c2700d28SVarun Prakash 691c2700d28SVarun Prakash nvme_tcp_setup_h2c_data_pdu(req); 69286f0348aSSagi Grimberg nvme_tcp_queue_request(req, false, true); 6933f2304f8SSagi Grimberg 6943f2304f8SSagi Grimberg return 0; 6953f2304f8SSagi Grimberg } 6963f2304f8SSagi Grimberg 6973f2304f8SSagi Grimberg static int nvme_tcp_recv_pdu(struct nvme_tcp_queue *queue, struct sk_buff *skb, 6983f2304f8SSagi Grimberg unsigned int *offset, size_t *len) 6993f2304f8SSagi Grimberg { 7003f2304f8SSagi Grimberg struct nvme_tcp_hdr *hdr; 7013f2304f8SSagi Grimberg char *pdu = queue->pdu; 7023f2304f8SSagi Grimberg size_t rcv_len = min_t(size_t, *len, queue->pdu_remaining); 7033f2304f8SSagi Grimberg int ret; 7043f2304f8SSagi Grimberg 7053f2304f8SSagi Grimberg ret = skb_copy_bits(skb, *offset, 7063f2304f8SSagi Grimberg &pdu[queue->pdu_offset], rcv_len); 7073f2304f8SSagi Grimberg if (unlikely(ret)) 7083f2304f8SSagi Grimberg return ret; 7093f2304f8SSagi Grimberg 7103f2304f8SSagi Grimberg queue->pdu_remaining -= rcv_len; 7113f2304f8SSagi Grimberg queue->pdu_offset += rcv_len; 7123f2304f8SSagi Grimberg *offset += rcv_len; 7133f2304f8SSagi Grimberg *len -= rcv_len; 7143f2304f8SSagi Grimberg if (queue->pdu_remaining) 7153f2304f8SSagi Grimberg return 0; 7163f2304f8SSagi Grimberg 7173f2304f8SSagi Grimberg hdr = queue->pdu; 7183f2304f8SSagi Grimberg if (queue->hdr_digest) { 7193f2304f8SSagi Grimberg ret = nvme_tcp_verify_hdgst(queue, queue->pdu, hdr->hlen); 7203f2304f8SSagi Grimberg if (unlikely(ret)) 7213f2304f8SSagi Grimberg return ret; 7223f2304f8SSagi Grimberg } 7233f2304f8SSagi Grimberg 7243f2304f8SSagi Grimberg 7253f2304f8SSagi Grimberg if (queue->data_digest) { 7263f2304f8SSagi Grimberg ret = nvme_tcp_check_ddgst(queue, queue->pdu); 7273f2304f8SSagi Grimberg if (unlikely(ret)) 7283f2304f8SSagi Grimberg return ret; 7293f2304f8SSagi Grimberg } 7303f2304f8SSagi Grimberg 7313f2304f8SSagi Grimberg switch (hdr->type) { 7323f2304f8SSagi Grimberg case nvme_tcp_c2h_data: 7336be18260SSagi Grimberg return nvme_tcp_handle_c2h_data(queue, (void *)queue->pdu); 7343f2304f8SSagi Grimberg case nvme_tcp_rsp: 7353f2304f8SSagi Grimberg nvme_tcp_init_recv_ctx(queue); 7366be18260SSagi Grimberg return nvme_tcp_handle_comp(queue, (void *)queue->pdu); 7373f2304f8SSagi Grimberg case nvme_tcp_r2t: 7383f2304f8SSagi Grimberg nvme_tcp_init_recv_ctx(queue); 7396be18260SSagi Grimberg return nvme_tcp_handle_r2t(queue, (void *)queue->pdu); 7403f2304f8SSagi Grimberg default: 7413f2304f8SSagi Grimberg dev_err(queue->ctrl->ctrl.device, 7423f2304f8SSagi Grimberg "unsupported pdu type (%d)\n", hdr->type); 7433f2304f8SSagi Grimberg return -EINVAL; 7443f2304f8SSagi Grimberg } 7453f2304f8SSagi Grimberg } 7463f2304f8SSagi Grimberg 747988aef9eSChristoph Hellwig static inline void nvme_tcp_end_request(struct request *rq, u16 status) 748602d674cSSagi Grimberg { 749602d674cSSagi Grimberg union nvme_result res = {}; 750602d674cSSagi Grimberg 7512eb81a33SChristoph Hellwig if (!nvme_try_complete_req(rq, cpu_to_le16(status << 1), res)) 752ff029451SChristoph Hellwig nvme_complete_rq(rq); 753602d674cSSagi Grimberg } 754602d674cSSagi Grimberg 7553f2304f8SSagi Grimberg static int nvme_tcp_recv_data(struct nvme_tcp_queue *queue, struct sk_buff *skb, 7563f2304f8SSagi Grimberg unsigned int *offset, size_t *len) 7573f2304f8SSagi Grimberg { 7583f2304f8SSagi Grimberg struct nvme_tcp_data_pdu *pdu = (void *)queue->pdu; 7593b01a9d0SSagi Grimberg struct request *rq = 760e7006de6SSagi Grimberg nvme_cid_to_rq(nvme_tcp_tagset(queue), pdu->command_id); 7613b01a9d0SSagi Grimberg struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq); 7623f2304f8SSagi Grimberg 7633f2304f8SSagi Grimberg while (true) { 7643f2304f8SSagi Grimberg int recv_len, ret; 7653f2304f8SSagi Grimberg 7663f2304f8SSagi Grimberg recv_len = min_t(size_t, *len, queue->data_remaining); 7673f2304f8SSagi Grimberg if (!recv_len) 7683f2304f8SSagi Grimberg break; 7693f2304f8SSagi Grimberg 7703f2304f8SSagi Grimberg if (!iov_iter_count(&req->iter)) { 7713f2304f8SSagi Grimberg req->curr_bio = req->curr_bio->bi_next; 7723f2304f8SSagi Grimberg 7733f2304f8SSagi Grimberg /* 7743f2304f8SSagi Grimberg * If we don`t have any bios it means that controller 7753f2304f8SSagi Grimberg * sent more data than we requested, hence error 7763f2304f8SSagi Grimberg */ 7773f2304f8SSagi Grimberg if (!req->curr_bio) { 7783f2304f8SSagi Grimberg dev_err(queue->ctrl->ctrl.device, 7793f2304f8SSagi Grimberg "queue %d no space in request %#x", 7803f2304f8SSagi Grimberg nvme_tcp_queue_id(queue), rq->tag); 7813f2304f8SSagi Grimberg nvme_tcp_init_recv_ctx(queue); 7823f2304f8SSagi Grimberg return -EIO; 7833f2304f8SSagi Grimberg } 7843f2304f8SSagi Grimberg nvme_tcp_init_iter(req, READ); 7853f2304f8SSagi Grimberg } 7863f2304f8SSagi Grimberg 7873f2304f8SSagi Grimberg /* we can read only from what is left in this bio */ 7883f2304f8SSagi Grimberg recv_len = min_t(size_t, recv_len, 7893f2304f8SSagi Grimberg iov_iter_count(&req->iter)); 7903f2304f8SSagi Grimberg 7913f2304f8SSagi Grimberg if (queue->data_digest) 7923f2304f8SSagi Grimberg ret = skb_copy_and_hash_datagram_iter(skb, *offset, 7933f2304f8SSagi Grimberg &req->iter, recv_len, queue->rcv_hash); 7943f2304f8SSagi Grimberg else 7953f2304f8SSagi Grimberg ret = skb_copy_datagram_iter(skb, *offset, 7963f2304f8SSagi Grimberg &req->iter, recv_len); 7973f2304f8SSagi Grimberg if (ret) { 7983f2304f8SSagi Grimberg dev_err(queue->ctrl->ctrl.device, 7993f2304f8SSagi Grimberg "queue %d failed to copy request %#x data", 8003f2304f8SSagi Grimberg nvme_tcp_queue_id(queue), rq->tag); 8013f2304f8SSagi Grimberg return ret; 8023f2304f8SSagi Grimberg } 8033f2304f8SSagi Grimberg 8043f2304f8SSagi Grimberg *len -= recv_len; 8053f2304f8SSagi Grimberg *offset += recv_len; 8063f2304f8SSagi Grimberg queue->data_remaining -= recv_len; 8073f2304f8SSagi Grimberg } 8083f2304f8SSagi Grimberg 8093f2304f8SSagi Grimberg if (!queue->data_remaining) { 8103f2304f8SSagi Grimberg if (queue->data_digest) { 8113f2304f8SSagi Grimberg nvme_tcp_ddgst_final(queue->rcv_hash, &queue->exp_ddgst); 8123f2304f8SSagi Grimberg queue->ddgst_remaining = NVME_TCP_DIGEST_LENGTH; 8133f2304f8SSagi Grimberg } else { 8141a9460ceSSagi Grimberg if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) { 8151ba2e507SDaniel Wagner nvme_tcp_end_request(rq, 8161ba2e507SDaniel Wagner le16_to_cpu(req->status)); 8171a9460ceSSagi Grimberg queue->nr_cqe++; 8181a9460ceSSagi Grimberg } 8193f2304f8SSagi Grimberg nvme_tcp_init_recv_ctx(queue); 8203f2304f8SSagi Grimberg } 8213f2304f8SSagi Grimberg } 8223f2304f8SSagi Grimberg 8233f2304f8SSagi Grimberg return 0; 8243f2304f8SSagi Grimberg } 8253f2304f8SSagi Grimberg 8263f2304f8SSagi Grimberg static int nvme_tcp_recv_ddgst(struct nvme_tcp_queue *queue, 8273f2304f8SSagi Grimberg struct sk_buff *skb, unsigned int *offset, size_t *len) 8283f2304f8SSagi Grimberg { 829602d674cSSagi Grimberg struct nvme_tcp_data_pdu *pdu = (void *)queue->pdu; 8303f2304f8SSagi Grimberg char *ddgst = (char *)&queue->recv_ddgst; 8313f2304f8SSagi Grimberg size_t recv_len = min_t(size_t, *len, queue->ddgst_remaining); 8323f2304f8SSagi Grimberg off_t off = NVME_TCP_DIGEST_LENGTH - queue->ddgst_remaining; 8333f2304f8SSagi Grimberg int ret; 8343f2304f8SSagi Grimberg 8353f2304f8SSagi Grimberg ret = skb_copy_bits(skb, *offset, &ddgst[off], recv_len); 8363f2304f8SSagi Grimberg if (unlikely(ret)) 8373f2304f8SSagi Grimberg return ret; 8383f2304f8SSagi Grimberg 8393f2304f8SSagi Grimberg queue->ddgst_remaining -= recv_len; 8403f2304f8SSagi Grimberg *offset += recv_len; 8413f2304f8SSagi Grimberg *len -= recv_len; 8423f2304f8SSagi Grimberg if (queue->ddgst_remaining) 8433f2304f8SSagi Grimberg return 0; 8443f2304f8SSagi Grimberg 8453f2304f8SSagi Grimberg if (queue->recv_ddgst != queue->exp_ddgst) { 8461ba2e507SDaniel Wagner struct request *rq = nvme_cid_to_rq(nvme_tcp_tagset(queue), 8471ba2e507SDaniel Wagner pdu->command_id); 8481ba2e507SDaniel Wagner struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq); 8491ba2e507SDaniel Wagner 8501ba2e507SDaniel Wagner req->status = cpu_to_le16(NVME_SC_DATA_XFER_ERROR); 8511ba2e507SDaniel Wagner 8523f2304f8SSagi Grimberg dev_err(queue->ctrl->ctrl.device, 8533f2304f8SSagi Grimberg "data digest error: recv %#x expected %#x\n", 8543f2304f8SSagi Grimberg le32_to_cpu(queue->recv_ddgst), 8553f2304f8SSagi Grimberg le32_to_cpu(queue->exp_ddgst)); 8563f2304f8SSagi Grimberg } 8573f2304f8SSagi Grimberg 858602d674cSSagi Grimberg if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) { 859e7006de6SSagi Grimberg struct request *rq = nvme_cid_to_rq(nvme_tcp_tagset(queue), 860602d674cSSagi Grimberg pdu->command_id); 8611ba2e507SDaniel Wagner struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq); 862602d674cSSagi Grimberg 8631ba2e507SDaniel Wagner nvme_tcp_end_request(rq, le16_to_cpu(req->status)); 8641a9460ceSSagi Grimberg queue->nr_cqe++; 865602d674cSSagi Grimberg } 866602d674cSSagi Grimberg 8673f2304f8SSagi Grimberg nvme_tcp_init_recv_ctx(queue); 8683f2304f8SSagi Grimberg return 0; 8693f2304f8SSagi Grimberg } 8703f2304f8SSagi Grimberg 8713f2304f8SSagi Grimberg static int nvme_tcp_recv_skb(read_descriptor_t *desc, struct sk_buff *skb, 8723f2304f8SSagi Grimberg unsigned int offset, size_t len) 8733f2304f8SSagi Grimberg { 8743f2304f8SSagi Grimberg struct nvme_tcp_queue *queue = desc->arg.data; 8753f2304f8SSagi Grimberg size_t consumed = len; 8763f2304f8SSagi Grimberg int result; 8773f2304f8SSagi Grimberg 8783f2304f8SSagi Grimberg while (len) { 8793f2304f8SSagi Grimberg switch (nvme_tcp_recv_state(queue)) { 8803f2304f8SSagi Grimberg case NVME_TCP_RECV_PDU: 8813f2304f8SSagi Grimberg result = nvme_tcp_recv_pdu(queue, skb, &offset, &len); 8823f2304f8SSagi Grimberg break; 8833f2304f8SSagi Grimberg case NVME_TCP_RECV_DATA: 8843f2304f8SSagi Grimberg result = nvme_tcp_recv_data(queue, skb, &offset, &len); 8853f2304f8SSagi Grimberg break; 8863f2304f8SSagi Grimberg case NVME_TCP_RECV_DDGST: 8873f2304f8SSagi Grimberg result = nvme_tcp_recv_ddgst(queue, skb, &offset, &len); 8883f2304f8SSagi Grimberg break; 8893f2304f8SSagi Grimberg default: 8903f2304f8SSagi Grimberg result = -EFAULT; 8913f2304f8SSagi Grimberg } 8923f2304f8SSagi Grimberg if (result) { 8933f2304f8SSagi Grimberg dev_err(queue->ctrl->ctrl.device, 8943f2304f8SSagi Grimberg "receive failed: %d\n", result); 8953f2304f8SSagi Grimberg queue->rd_enabled = false; 8963f2304f8SSagi Grimberg nvme_tcp_error_recovery(&queue->ctrl->ctrl); 8973f2304f8SSagi Grimberg return result; 8983f2304f8SSagi Grimberg } 8993f2304f8SSagi Grimberg } 9003f2304f8SSagi Grimberg 9013f2304f8SSagi Grimberg return consumed; 9023f2304f8SSagi Grimberg } 9033f2304f8SSagi Grimberg 9043f2304f8SSagi Grimberg static void nvme_tcp_data_ready(struct sock *sk) 9053f2304f8SSagi Grimberg { 9063f2304f8SSagi Grimberg struct nvme_tcp_queue *queue; 9073f2304f8SSagi Grimberg 908386e5e6eSSagi Grimberg read_lock_bh(&sk->sk_callback_lock); 9093f2304f8SSagi Grimberg queue = sk->sk_user_data; 91072e5d757SSagi Grimberg if (likely(queue && queue->rd_enabled) && 91172e5d757SSagi Grimberg !test_bit(NVME_TCP_Q_POLLING, &queue->flags)) 9123f2304f8SSagi Grimberg queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work); 913386e5e6eSSagi Grimberg read_unlock_bh(&sk->sk_callback_lock); 9143f2304f8SSagi Grimberg } 9153f2304f8SSagi Grimberg 9163f2304f8SSagi Grimberg static void nvme_tcp_write_space(struct sock *sk) 9173f2304f8SSagi Grimberg { 9183f2304f8SSagi Grimberg struct nvme_tcp_queue *queue; 9193f2304f8SSagi Grimberg 9203f2304f8SSagi Grimberg read_lock_bh(&sk->sk_callback_lock); 9213f2304f8SSagi Grimberg queue = sk->sk_user_data; 9223f2304f8SSagi Grimberg if (likely(queue && sk_stream_is_writeable(sk))) { 9233f2304f8SSagi Grimberg clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 9243f2304f8SSagi Grimberg queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work); 9253f2304f8SSagi Grimberg } 9263f2304f8SSagi Grimberg read_unlock_bh(&sk->sk_callback_lock); 9273f2304f8SSagi Grimberg } 9283f2304f8SSagi Grimberg 9293f2304f8SSagi Grimberg static void nvme_tcp_state_change(struct sock *sk) 9303f2304f8SSagi Grimberg { 9313f2304f8SSagi Grimberg struct nvme_tcp_queue *queue; 9323f2304f8SSagi Grimberg 9338b73b45dSSagi Grimberg read_lock_bh(&sk->sk_callback_lock); 9343f2304f8SSagi Grimberg queue = sk->sk_user_data; 9353f2304f8SSagi Grimberg if (!queue) 9363f2304f8SSagi Grimberg goto done; 9373f2304f8SSagi Grimberg 9383f2304f8SSagi Grimberg switch (sk->sk_state) { 9393f2304f8SSagi Grimberg case TCP_CLOSE: 9403f2304f8SSagi Grimberg case TCP_CLOSE_WAIT: 9413f2304f8SSagi Grimberg case TCP_LAST_ACK: 9423f2304f8SSagi Grimberg case TCP_FIN_WAIT1: 9433f2304f8SSagi Grimberg case TCP_FIN_WAIT2: 9443f2304f8SSagi Grimberg nvme_tcp_error_recovery(&queue->ctrl->ctrl); 9453f2304f8SSagi Grimberg break; 9463f2304f8SSagi Grimberg default: 9473f2304f8SSagi Grimberg dev_info(queue->ctrl->ctrl.device, 9483f2304f8SSagi Grimberg "queue %d socket state %d\n", 9493f2304f8SSagi Grimberg nvme_tcp_queue_id(queue), sk->sk_state); 9503f2304f8SSagi Grimberg } 9513f2304f8SSagi Grimberg 9523f2304f8SSagi Grimberg queue->state_change(sk); 9533f2304f8SSagi Grimberg done: 9548b73b45dSSagi Grimberg read_unlock_bh(&sk->sk_callback_lock); 9553f2304f8SSagi Grimberg } 9563f2304f8SSagi Grimberg 9573f2304f8SSagi Grimberg static inline void nvme_tcp_done_send_req(struct nvme_tcp_queue *queue) 9583f2304f8SSagi Grimberg { 9593f2304f8SSagi Grimberg queue->request = NULL; 9603f2304f8SSagi Grimberg } 9613f2304f8SSagi Grimberg 9623f2304f8SSagi Grimberg static void nvme_tcp_fail_request(struct nvme_tcp_request *req) 9633f2304f8SSagi Grimberg { 96463573807SSagi Grimberg if (nvme_tcp_async_req(req)) { 96563573807SSagi Grimberg union nvme_result res = {}; 96663573807SSagi Grimberg 96763573807SSagi Grimberg nvme_complete_async_event(&req->queue->ctrl->ctrl, 96863573807SSagi Grimberg cpu_to_le16(NVME_SC_HOST_PATH_ERROR), &res); 96963573807SSagi Grimberg } else { 97063573807SSagi Grimberg nvme_tcp_end_request(blk_mq_rq_from_pdu(req), 97163573807SSagi Grimberg NVME_SC_HOST_PATH_ERROR); 97263573807SSagi Grimberg } 9733f2304f8SSagi Grimberg } 9743f2304f8SSagi Grimberg 9753f2304f8SSagi Grimberg static int nvme_tcp_try_send_data(struct nvme_tcp_request *req) 9763f2304f8SSagi Grimberg { 9773f2304f8SSagi Grimberg struct nvme_tcp_queue *queue = req->queue; 97825e1f67eSSagi Grimberg int req_data_len = req->data_len; 979c2700d28SVarun Prakash u32 h2cdata_left = req->h2cdata_left; 9803f2304f8SSagi Grimberg 9813f2304f8SSagi Grimberg while (true) { 9823f2304f8SSagi Grimberg struct page *page = nvme_tcp_req_cur_page(req); 9833f2304f8SSagi Grimberg size_t offset = nvme_tcp_req_cur_offset(req); 9843f2304f8SSagi Grimberg size_t len = nvme_tcp_req_cur_length(req); 9853f2304f8SSagi Grimberg bool last = nvme_tcp_pdu_last_send(req, len); 98625e1f67eSSagi Grimberg int req_data_sent = req->data_sent; 9873f2304f8SSagi Grimberg int ret, flags = MSG_DONTWAIT; 9883f2304f8SSagi Grimberg 989122e5b9fSSagi Grimberg if (last && !queue->data_digest && !nvme_tcp_queue_more(queue)) 9903f2304f8SSagi Grimberg flags |= MSG_EOR; 9913f2304f8SSagi Grimberg else 9925bb052d7SSagi Grimberg flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST; 9933f2304f8SSagi Grimberg 9947d4194abSColy Li if (sendpage_ok(page)) { 9957d4194abSColy Li ret = kernel_sendpage(queue->sock, page, offset, len, 99637c15219SMikhail Skorzhinskii flags); 99737c15219SMikhail Skorzhinskii } else { 9987d4194abSColy Li ret = sock_no_sendpage(queue->sock, page, offset, len, 99937c15219SMikhail Skorzhinskii flags); 100037c15219SMikhail Skorzhinskii } 10013f2304f8SSagi Grimberg if (ret <= 0) 10023f2304f8SSagi Grimberg return ret; 10033f2304f8SSagi Grimberg 10043f2304f8SSagi Grimberg if (queue->data_digest) 10053f2304f8SSagi Grimberg nvme_tcp_ddgst_update(queue->snd_hash, page, 10063f2304f8SSagi Grimberg offset, ret); 10073f2304f8SSagi Grimberg 1008e371af03SSagi Grimberg /* 1009e371af03SSagi Grimberg * update the request iterator except for the last payload send 1010e371af03SSagi Grimberg * in the request where we don't want to modify it as we may 1011e371af03SSagi Grimberg * compete with the RX path completing the request. 1012e371af03SSagi Grimberg */ 101325e1f67eSSagi Grimberg if (req_data_sent + ret < req_data_len) 1014e371af03SSagi Grimberg nvme_tcp_advance_req(req, ret); 1015e371af03SSagi Grimberg 1016e371af03SSagi Grimberg /* fully successful last send in current PDU */ 10173f2304f8SSagi Grimberg if (last && ret == len) { 10183f2304f8SSagi Grimberg if (queue->data_digest) { 10193f2304f8SSagi Grimberg nvme_tcp_ddgst_final(queue->snd_hash, 10203f2304f8SSagi Grimberg &req->ddgst); 10213f2304f8SSagi Grimberg req->state = NVME_TCP_SEND_DDGST; 10223f2304f8SSagi Grimberg req->offset = 0; 10233f2304f8SSagi Grimberg } else { 1024c2700d28SVarun Prakash if (h2cdata_left) 1025c2700d28SVarun Prakash nvme_tcp_setup_h2c_data_pdu(req); 1026c2700d28SVarun Prakash else 10273f2304f8SSagi Grimberg nvme_tcp_done_send_req(queue); 10283f2304f8SSagi Grimberg } 10293f2304f8SSagi Grimberg return 1; 10303f2304f8SSagi Grimberg } 10313f2304f8SSagi Grimberg } 10323f2304f8SSagi Grimberg return -EAGAIN; 10333f2304f8SSagi Grimberg } 10343f2304f8SSagi Grimberg 10353f2304f8SSagi Grimberg static int nvme_tcp_try_send_cmd_pdu(struct nvme_tcp_request *req) 10363f2304f8SSagi Grimberg { 10373f2304f8SSagi Grimberg struct nvme_tcp_queue *queue = req->queue; 10383f2304f8SSagi Grimberg struct nvme_tcp_cmd_pdu *pdu = req->pdu; 10393f2304f8SSagi Grimberg bool inline_data = nvme_tcp_has_inline_data(req); 10403f2304f8SSagi Grimberg u8 hdgst = nvme_tcp_hdgst_len(queue); 10413f2304f8SSagi Grimberg int len = sizeof(*pdu) + hdgst - req->offset; 10425bb052d7SSagi Grimberg int flags = MSG_DONTWAIT; 10433f2304f8SSagi Grimberg int ret; 10443f2304f8SSagi Grimberg 1045122e5b9fSSagi Grimberg if (inline_data || nvme_tcp_queue_more(queue)) 10465bb052d7SSagi Grimberg flags |= MSG_MORE | MSG_SENDPAGE_NOTLAST; 10475bb052d7SSagi Grimberg else 10485bb052d7SSagi Grimberg flags |= MSG_EOR; 10495bb052d7SSagi Grimberg 10503f2304f8SSagi Grimberg if (queue->hdr_digest && !req->offset) 10513f2304f8SSagi Grimberg nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu)); 10523f2304f8SSagi Grimberg 10533f2304f8SSagi Grimberg ret = kernel_sendpage(queue->sock, virt_to_page(pdu), 10543f2304f8SSagi Grimberg offset_in_page(pdu) + req->offset, len, flags); 10553f2304f8SSagi Grimberg if (unlikely(ret <= 0)) 10563f2304f8SSagi Grimberg return ret; 10573f2304f8SSagi Grimberg 10583f2304f8SSagi Grimberg len -= ret; 10593f2304f8SSagi Grimberg if (!len) { 10603f2304f8SSagi Grimberg if (inline_data) { 10613f2304f8SSagi Grimberg req->state = NVME_TCP_SEND_DATA; 10623f2304f8SSagi Grimberg if (queue->data_digest) 10633f2304f8SSagi Grimberg crypto_ahash_init(queue->snd_hash); 10643f2304f8SSagi Grimberg } else { 10653f2304f8SSagi Grimberg nvme_tcp_done_send_req(queue); 10663f2304f8SSagi Grimberg } 10673f2304f8SSagi Grimberg return 1; 10683f2304f8SSagi Grimberg } 10693f2304f8SSagi Grimberg req->offset += ret; 10703f2304f8SSagi Grimberg 10713f2304f8SSagi Grimberg return -EAGAIN; 10723f2304f8SSagi Grimberg } 10733f2304f8SSagi Grimberg 10743f2304f8SSagi Grimberg static int nvme_tcp_try_send_data_pdu(struct nvme_tcp_request *req) 10753f2304f8SSagi Grimberg { 10763f2304f8SSagi Grimberg struct nvme_tcp_queue *queue = req->queue; 10773f2304f8SSagi Grimberg struct nvme_tcp_data_pdu *pdu = req->pdu; 10783f2304f8SSagi Grimberg u8 hdgst = nvme_tcp_hdgst_len(queue); 10793f2304f8SSagi Grimberg int len = sizeof(*pdu) - req->offset + hdgst; 10803f2304f8SSagi Grimberg int ret; 10813f2304f8SSagi Grimberg 10823f2304f8SSagi Grimberg if (queue->hdr_digest && !req->offset) 10833f2304f8SSagi Grimberg nvme_tcp_hdgst(queue->snd_hash, pdu, sizeof(*pdu)); 10843f2304f8SSagi Grimberg 1085c2700d28SVarun Prakash if (!req->h2cdata_left) 10863f2304f8SSagi Grimberg ret = kernel_sendpage(queue->sock, virt_to_page(pdu), 10873f2304f8SSagi Grimberg offset_in_page(pdu) + req->offset, len, 10885bb052d7SSagi Grimberg MSG_DONTWAIT | MSG_MORE | MSG_SENDPAGE_NOTLAST); 1089c2700d28SVarun Prakash else 1090c2700d28SVarun Prakash ret = sock_no_sendpage(queue->sock, virt_to_page(pdu), 1091c2700d28SVarun Prakash offset_in_page(pdu) + req->offset, len, 1092c2700d28SVarun Prakash MSG_DONTWAIT | MSG_MORE); 10933f2304f8SSagi Grimberg if (unlikely(ret <= 0)) 10943f2304f8SSagi Grimberg return ret; 10953f2304f8SSagi Grimberg 10963f2304f8SSagi Grimberg len -= ret; 10973f2304f8SSagi Grimberg if (!len) { 10983f2304f8SSagi Grimberg req->state = NVME_TCP_SEND_DATA; 10993f2304f8SSagi Grimberg if (queue->data_digest) 11003f2304f8SSagi Grimberg crypto_ahash_init(queue->snd_hash); 11013f2304f8SSagi Grimberg return 1; 11023f2304f8SSagi Grimberg } 11033f2304f8SSagi Grimberg req->offset += ret; 11043f2304f8SSagi Grimberg 11053f2304f8SSagi Grimberg return -EAGAIN; 11063f2304f8SSagi Grimberg } 11073f2304f8SSagi Grimberg 11083f2304f8SSagi Grimberg static int nvme_tcp_try_send_ddgst(struct nvme_tcp_request *req) 11093f2304f8SSagi Grimberg { 11103f2304f8SSagi Grimberg struct nvme_tcp_queue *queue = req->queue; 1111ce7723e9SVarun Prakash size_t offset = req->offset; 1112c2700d28SVarun Prakash u32 h2cdata_left = req->h2cdata_left; 11133f2304f8SSagi Grimberg int ret; 1114122e5b9fSSagi Grimberg struct msghdr msg = { .msg_flags = MSG_DONTWAIT }; 11153f2304f8SSagi Grimberg struct kvec iov = { 1116d89b9f3bSVarun Prakash .iov_base = (u8 *)&req->ddgst + req->offset, 11173f2304f8SSagi Grimberg .iov_len = NVME_TCP_DIGEST_LENGTH - req->offset 11183f2304f8SSagi Grimberg }; 11193f2304f8SSagi Grimberg 1120122e5b9fSSagi Grimberg if (nvme_tcp_queue_more(queue)) 1121122e5b9fSSagi Grimberg msg.msg_flags |= MSG_MORE; 1122122e5b9fSSagi Grimberg else 1123122e5b9fSSagi Grimberg msg.msg_flags |= MSG_EOR; 1124122e5b9fSSagi Grimberg 11253f2304f8SSagi Grimberg ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len); 11263f2304f8SSagi Grimberg if (unlikely(ret <= 0)) 11273f2304f8SSagi Grimberg return ret; 11283f2304f8SSagi Grimberg 1129ce7723e9SVarun Prakash if (offset + ret == NVME_TCP_DIGEST_LENGTH) { 1130c2700d28SVarun Prakash if (h2cdata_left) 1131c2700d28SVarun Prakash nvme_tcp_setup_h2c_data_pdu(req); 1132c2700d28SVarun Prakash else 11333f2304f8SSagi Grimberg nvme_tcp_done_send_req(queue); 11343f2304f8SSagi Grimberg return 1; 11353f2304f8SSagi Grimberg } 11363f2304f8SSagi Grimberg 11373f2304f8SSagi Grimberg req->offset += ret; 11383f2304f8SSagi Grimberg return -EAGAIN; 11393f2304f8SSagi Grimberg } 11403f2304f8SSagi Grimberg 11413f2304f8SSagi Grimberg static int nvme_tcp_try_send(struct nvme_tcp_queue *queue) 11423f2304f8SSagi Grimberg { 11433f2304f8SSagi Grimberg struct nvme_tcp_request *req; 11443f2304f8SSagi Grimberg int ret = 1; 11453f2304f8SSagi Grimberg 11463f2304f8SSagi Grimberg if (!queue->request) { 11473f2304f8SSagi Grimberg queue->request = nvme_tcp_fetch_request(queue); 11483f2304f8SSagi Grimberg if (!queue->request) 11493f2304f8SSagi Grimberg return 0; 11503f2304f8SSagi Grimberg } 11513f2304f8SSagi Grimberg req = queue->request; 11523f2304f8SSagi Grimberg 11533f2304f8SSagi Grimberg if (req->state == NVME_TCP_SEND_CMD_PDU) { 11543f2304f8SSagi Grimberg ret = nvme_tcp_try_send_cmd_pdu(req); 11553f2304f8SSagi Grimberg if (ret <= 0) 11563f2304f8SSagi Grimberg goto done; 11573f2304f8SSagi Grimberg if (!nvme_tcp_has_inline_data(req)) 11583f2304f8SSagi Grimberg return ret; 11593f2304f8SSagi Grimberg } 11603f2304f8SSagi Grimberg 11613f2304f8SSagi Grimberg if (req->state == NVME_TCP_SEND_H2C_PDU) { 11623f2304f8SSagi Grimberg ret = nvme_tcp_try_send_data_pdu(req); 11633f2304f8SSagi Grimberg if (ret <= 0) 11643f2304f8SSagi Grimberg goto done; 11653f2304f8SSagi Grimberg } 11663f2304f8SSagi Grimberg 11673f2304f8SSagi Grimberg if (req->state == NVME_TCP_SEND_DATA) { 11683f2304f8SSagi Grimberg ret = nvme_tcp_try_send_data(req); 11693f2304f8SSagi Grimberg if (ret <= 0) 11703f2304f8SSagi Grimberg goto done; 11713f2304f8SSagi Grimberg } 11723f2304f8SSagi Grimberg 11733f2304f8SSagi Grimberg if (req->state == NVME_TCP_SEND_DDGST) 11743f2304f8SSagi Grimberg ret = nvme_tcp_try_send_ddgst(req); 11753f2304f8SSagi Grimberg done: 11765ff4e112SSagi Grimberg if (ret == -EAGAIN) { 11773f2304f8SSagi Grimberg ret = 0; 11785ff4e112SSagi Grimberg } else if (ret < 0) { 11795ff4e112SSagi Grimberg dev_err(queue->ctrl->ctrl.device, 11805ff4e112SSagi Grimberg "failed to send request %d\n", ret); 11815ff4e112SSagi Grimberg nvme_tcp_fail_request(queue->request); 11825ff4e112SSagi Grimberg nvme_tcp_done_send_req(queue); 11835ff4e112SSagi Grimberg } 11843f2304f8SSagi Grimberg return ret; 11853f2304f8SSagi Grimberg } 11863f2304f8SSagi Grimberg 11873f2304f8SSagi Grimberg static int nvme_tcp_try_recv(struct nvme_tcp_queue *queue) 11883f2304f8SSagi Grimberg { 118910407ec9SPotnuri Bharat Teja struct socket *sock = queue->sock; 119010407ec9SPotnuri Bharat Teja struct sock *sk = sock->sk; 11913f2304f8SSagi Grimberg read_descriptor_t rd_desc; 11923f2304f8SSagi Grimberg int consumed; 11933f2304f8SSagi Grimberg 11943f2304f8SSagi Grimberg rd_desc.arg.data = queue; 11953f2304f8SSagi Grimberg rd_desc.count = 1; 11963f2304f8SSagi Grimberg lock_sock(sk); 11971a9460ceSSagi Grimberg queue->nr_cqe = 0; 119810407ec9SPotnuri Bharat Teja consumed = sock->ops->read_sock(sk, &rd_desc, nvme_tcp_recv_skb); 11993f2304f8SSagi Grimberg release_sock(sk); 12003f2304f8SSagi Grimberg return consumed; 12013f2304f8SSagi Grimberg } 12023f2304f8SSagi Grimberg 12033f2304f8SSagi Grimberg static void nvme_tcp_io_work(struct work_struct *w) 12043f2304f8SSagi Grimberg { 12053f2304f8SSagi Grimberg struct nvme_tcp_queue *queue = 12063f2304f8SSagi Grimberg container_of(w, struct nvme_tcp_queue, io_work); 1207ddef2957SWunderlich, Mark unsigned long deadline = jiffies + msecs_to_jiffies(1); 12083f2304f8SSagi Grimberg 12093f2304f8SSagi Grimberg do { 12103f2304f8SSagi Grimberg bool pending = false; 12113f2304f8SSagi Grimberg int result; 12123f2304f8SSagi Grimberg 1213db5ad6b7SSagi Grimberg if (mutex_trylock(&queue->send_mutex)) { 12143f2304f8SSagi Grimberg result = nvme_tcp_try_send(queue); 1215db5ad6b7SSagi Grimberg mutex_unlock(&queue->send_mutex); 12165ff4e112SSagi Grimberg if (result > 0) 12173f2304f8SSagi Grimberg pending = true; 12185ff4e112SSagi Grimberg else if (unlikely(result < 0)) 12195ff4e112SSagi Grimberg break; 122070f437fbSKeith Busch } 12213f2304f8SSagi Grimberg 12223f2304f8SSagi Grimberg result = nvme_tcp_try_recv(queue); 12233f2304f8SSagi Grimberg if (result > 0) 12243f2304f8SSagi Grimberg pending = true; 1225761ad26cSSagi Grimberg else if (unlikely(result < 0)) 122639d06079SSagi Grimberg return; 12273f2304f8SSagi Grimberg 1228160f3549SSagi Grimberg if (!pending || !queue->rd_enabled) 12293f2304f8SSagi Grimberg return; 12303f2304f8SSagi Grimberg 1231ddef2957SWunderlich, Mark } while (!time_after(jiffies, deadline)); /* quota is exhausted */ 12323f2304f8SSagi Grimberg 12333f2304f8SSagi Grimberg queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work); 12343f2304f8SSagi Grimberg } 12353f2304f8SSagi Grimberg 12363f2304f8SSagi Grimberg static void nvme_tcp_free_crypto(struct nvme_tcp_queue *queue) 12373f2304f8SSagi Grimberg { 12383f2304f8SSagi Grimberg struct crypto_ahash *tfm = crypto_ahash_reqtfm(queue->rcv_hash); 12393f2304f8SSagi Grimberg 12403f2304f8SSagi Grimberg ahash_request_free(queue->rcv_hash); 12413f2304f8SSagi Grimberg ahash_request_free(queue->snd_hash); 12423f2304f8SSagi Grimberg crypto_free_ahash(tfm); 12433f2304f8SSagi Grimberg } 12443f2304f8SSagi Grimberg 12453f2304f8SSagi Grimberg static int nvme_tcp_alloc_crypto(struct nvme_tcp_queue *queue) 12463f2304f8SSagi Grimberg { 12473f2304f8SSagi Grimberg struct crypto_ahash *tfm; 12483f2304f8SSagi Grimberg 12493f2304f8SSagi Grimberg tfm = crypto_alloc_ahash("crc32c", 0, CRYPTO_ALG_ASYNC); 12503f2304f8SSagi Grimberg if (IS_ERR(tfm)) 12513f2304f8SSagi Grimberg return PTR_ERR(tfm); 12523f2304f8SSagi Grimberg 12533f2304f8SSagi Grimberg queue->snd_hash = ahash_request_alloc(tfm, GFP_KERNEL); 12543f2304f8SSagi Grimberg if (!queue->snd_hash) 12553f2304f8SSagi Grimberg goto free_tfm; 12563f2304f8SSagi Grimberg ahash_request_set_callback(queue->snd_hash, 0, NULL, NULL); 12573f2304f8SSagi Grimberg 12583f2304f8SSagi Grimberg queue->rcv_hash = ahash_request_alloc(tfm, GFP_KERNEL); 12593f2304f8SSagi Grimberg if (!queue->rcv_hash) 12603f2304f8SSagi Grimberg goto free_snd_hash; 12613f2304f8SSagi Grimberg ahash_request_set_callback(queue->rcv_hash, 0, NULL, NULL); 12623f2304f8SSagi Grimberg 12633f2304f8SSagi Grimberg return 0; 12643f2304f8SSagi Grimberg free_snd_hash: 12653f2304f8SSagi Grimberg ahash_request_free(queue->snd_hash); 12663f2304f8SSagi Grimberg free_tfm: 12673f2304f8SSagi Grimberg crypto_free_ahash(tfm); 12683f2304f8SSagi Grimberg return -ENOMEM; 12693f2304f8SSagi Grimberg } 12703f2304f8SSagi Grimberg 12713f2304f8SSagi Grimberg static void nvme_tcp_free_async_req(struct nvme_tcp_ctrl *ctrl) 12723f2304f8SSagi Grimberg { 12733f2304f8SSagi Grimberg struct nvme_tcp_request *async = &ctrl->async_req; 12743f2304f8SSagi Grimberg 12753f2304f8SSagi Grimberg page_frag_free(async->pdu); 12763f2304f8SSagi Grimberg } 12773f2304f8SSagi Grimberg 12783f2304f8SSagi Grimberg static int nvme_tcp_alloc_async_req(struct nvme_tcp_ctrl *ctrl) 12793f2304f8SSagi Grimberg { 12803f2304f8SSagi Grimberg struct nvme_tcp_queue *queue = &ctrl->queues[0]; 12813f2304f8SSagi Grimberg struct nvme_tcp_request *async = &ctrl->async_req; 12823f2304f8SSagi Grimberg u8 hdgst = nvme_tcp_hdgst_len(queue); 12833f2304f8SSagi Grimberg 12843f2304f8SSagi Grimberg async->pdu = page_frag_alloc(&queue->pf_cache, 12853f2304f8SSagi Grimberg sizeof(struct nvme_tcp_cmd_pdu) + hdgst, 12863f2304f8SSagi Grimberg GFP_KERNEL | __GFP_ZERO); 12873f2304f8SSagi Grimberg if (!async->pdu) 12883f2304f8SSagi Grimberg return -ENOMEM; 12893f2304f8SSagi Grimberg 12903f2304f8SSagi Grimberg async->queue = &ctrl->queues[0]; 12913f2304f8SSagi Grimberg return 0; 12923f2304f8SSagi Grimberg } 12933f2304f8SSagi Grimberg 12943f2304f8SSagi Grimberg static void nvme_tcp_free_queue(struct nvme_ctrl *nctrl, int qid) 12953f2304f8SSagi Grimberg { 1296a5053c92SMaurizio Lombardi struct page *page; 12973f2304f8SSagi Grimberg struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl); 12983f2304f8SSagi Grimberg struct nvme_tcp_queue *queue = &ctrl->queues[qid]; 12993f2304f8SSagi Grimberg 13003f2304f8SSagi Grimberg if (!test_and_clear_bit(NVME_TCP_Q_ALLOCATED, &queue->flags)) 13013f2304f8SSagi Grimberg return; 13023f2304f8SSagi Grimberg 13033f2304f8SSagi Grimberg if (queue->hdr_digest || queue->data_digest) 13043f2304f8SSagi Grimberg nvme_tcp_free_crypto(queue); 13053f2304f8SSagi Grimberg 1306a5053c92SMaurizio Lombardi if (queue->pf_cache.va) { 1307a5053c92SMaurizio Lombardi page = virt_to_head_page(queue->pf_cache.va); 1308a5053c92SMaurizio Lombardi __page_frag_cache_drain(page, queue->pf_cache.pagecnt_bias); 1309a5053c92SMaurizio Lombardi queue->pf_cache.va = NULL; 1310a5053c92SMaurizio Lombardi } 13113f2304f8SSagi Grimberg sock_release(queue->sock); 13123f2304f8SSagi Grimberg kfree(queue->pdu); 1313d48f92cdSKeith Busch mutex_destroy(&queue->send_mutex); 13149ebbfe49SChao Leng mutex_destroy(&queue->queue_lock); 13153f2304f8SSagi Grimberg } 13163f2304f8SSagi Grimberg 13173f2304f8SSagi Grimberg static int nvme_tcp_init_connection(struct nvme_tcp_queue *queue) 13183f2304f8SSagi Grimberg { 13193f2304f8SSagi Grimberg struct nvme_tcp_icreq_pdu *icreq; 13203f2304f8SSagi Grimberg struct nvme_tcp_icresp_pdu *icresp; 13213f2304f8SSagi Grimberg struct msghdr msg = {}; 13223f2304f8SSagi Grimberg struct kvec iov; 13233f2304f8SSagi Grimberg bool ctrl_hdgst, ctrl_ddgst; 1324c2700d28SVarun Prakash u32 maxh2cdata; 13253f2304f8SSagi Grimberg int ret; 13263f2304f8SSagi Grimberg 13273f2304f8SSagi Grimberg icreq = kzalloc(sizeof(*icreq), GFP_KERNEL); 13283f2304f8SSagi Grimberg if (!icreq) 13293f2304f8SSagi Grimberg return -ENOMEM; 13303f2304f8SSagi Grimberg 13313f2304f8SSagi Grimberg icresp = kzalloc(sizeof(*icresp), GFP_KERNEL); 13323f2304f8SSagi Grimberg if (!icresp) { 13333f2304f8SSagi Grimberg ret = -ENOMEM; 13343f2304f8SSagi Grimberg goto free_icreq; 13353f2304f8SSagi Grimberg } 13363f2304f8SSagi Grimberg 13373f2304f8SSagi Grimberg icreq->hdr.type = nvme_tcp_icreq; 13383f2304f8SSagi Grimberg icreq->hdr.hlen = sizeof(*icreq); 13393f2304f8SSagi Grimberg icreq->hdr.pdo = 0; 13403f2304f8SSagi Grimberg icreq->hdr.plen = cpu_to_le32(icreq->hdr.hlen); 13413f2304f8SSagi Grimberg icreq->pfv = cpu_to_le16(NVME_TCP_PFV_1_0); 13423f2304f8SSagi Grimberg icreq->maxr2t = 0; /* single inflight r2t supported */ 13433f2304f8SSagi Grimberg icreq->hpda = 0; /* no alignment constraint */ 13443f2304f8SSagi Grimberg if (queue->hdr_digest) 13453f2304f8SSagi Grimberg icreq->digest |= NVME_TCP_HDR_DIGEST_ENABLE; 13463f2304f8SSagi Grimberg if (queue->data_digest) 13473f2304f8SSagi Grimberg icreq->digest |= NVME_TCP_DATA_DIGEST_ENABLE; 13483f2304f8SSagi Grimberg 13493f2304f8SSagi Grimberg iov.iov_base = icreq; 13503f2304f8SSagi Grimberg iov.iov_len = sizeof(*icreq); 13513f2304f8SSagi Grimberg ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len); 13523f2304f8SSagi Grimberg if (ret < 0) 13533f2304f8SSagi Grimberg goto free_icresp; 13543f2304f8SSagi Grimberg 13553f2304f8SSagi Grimberg memset(&msg, 0, sizeof(msg)); 13563f2304f8SSagi Grimberg iov.iov_base = icresp; 13573f2304f8SSagi Grimberg iov.iov_len = sizeof(*icresp); 13583f2304f8SSagi Grimberg ret = kernel_recvmsg(queue->sock, &msg, &iov, 1, 13593f2304f8SSagi Grimberg iov.iov_len, msg.msg_flags); 13603f2304f8SSagi Grimberg if (ret < 0) 13613f2304f8SSagi Grimberg goto free_icresp; 13623f2304f8SSagi Grimberg 13633f2304f8SSagi Grimberg ret = -EINVAL; 13643f2304f8SSagi Grimberg if (icresp->hdr.type != nvme_tcp_icresp) { 13653f2304f8SSagi Grimberg pr_err("queue %d: bad type returned %d\n", 13663f2304f8SSagi Grimberg nvme_tcp_queue_id(queue), icresp->hdr.type); 13673f2304f8SSagi Grimberg goto free_icresp; 13683f2304f8SSagi Grimberg } 13693f2304f8SSagi Grimberg 13703f2304f8SSagi Grimberg if (le32_to_cpu(icresp->hdr.plen) != sizeof(*icresp)) { 13713f2304f8SSagi Grimberg pr_err("queue %d: bad pdu length returned %d\n", 13723f2304f8SSagi Grimberg nvme_tcp_queue_id(queue), icresp->hdr.plen); 13733f2304f8SSagi Grimberg goto free_icresp; 13743f2304f8SSagi Grimberg } 13753f2304f8SSagi Grimberg 13763f2304f8SSagi Grimberg if (icresp->pfv != NVME_TCP_PFV_1_0) { 13773f2304f8SSagi Grimberg pr_err("queue %d: bad pfv returned %d\n", 13783f2304f8SSagi Grimberg nvme_tcp_queue_id(queue), icresp->pfv); 13793f2304f8SSagi Grimberg goto free_icresp; 13803f2304f8SSagi Grimberg } 13813f2304f8SSagi Grimberg 13823f2304f8SSagi Grimberg ctrl_ddgst = !!(icresp->digest & NVME_TCP_DATA_DIGEST_ENABLE); 13833f2304f8SSagi Grimberg if ((queue->data_digest && !ctrl_ddgst) || 13843f2304f8SSagi Grimberg (!queue->data_digest && ctrl_ddgst)) { 13853f2304f8SSagi Grimberg pr_err("queue %d: data digest mismatch host: %s ctrl: %s\n", 13863f2304f8SSagi Grimberg nvme_tcp_queue_id(queue), 13873f2304f8SSagi Grimberg queue->data_digest ? "enabled" : "disabled", 13883f2304f8SSagi Grimberg ctrl_ddgst ? "enabled" : "disabled"); 13893f2304f8SSagi Grimberg goto free_icresp; 13903f2304f8SSagi Grimberg } 13913f2304f8SSagi Grimberg 13923f2304f8SSagi Grimberg ctrl_hdgst = !!(icresp->digest & NVME_TCP_HDR_DIGEST_ENABLE); 13933f2304f8SSagi Grimberg if ((queue->hdr_digest && !ctrl_hdgst) || 13943f2304f8SSagi Grimberg (!queue->hdr_digest && ctrl_hdgst)) { 13953f2304f8SSagi Grimberg pr_err("queue %d: header digest mismatch host: %s ctrl: %s\n", 13963f2304f8SSagi Grimberg nvme_tcp_queue_id(queue), 13973f2304f8SSagi Grimberg queue->hdr_digest ? "enabled" : "disabled", 13983f2304f8SSagi Grimberg ctrl_hdgst ? "enabled" : "disabled"); 13993f2304f8SSagi Grimberg goto free_icresp; 14003f2304f8SSagi Grimberg } 14013f2304f8SSagi Grimberg 14023f2304f8SSagi Grimberg if (icresp->cpda != 0) { 14033f2304f8SSagi Grimberg pr_err("queue %d: unsupported cpda returned %d\n", 14043f2304f8SSagi Grimberg nvme_tcp_queue_id(queue), icresp->cpda); 14053f2304f8SSagi Grimberg goto free_icresp; 14063f2304f8SSagi Grimberg } 14073f2304f8SSagi Grimberg 1408c2700d28SVarun Prakash maxh2cdata = le32_to_cpu(icresp->maxdata); 1409c2700d28SVarun Prakash if ((maxh2cdata % 4) || (maxh2cdata < NVME_TCP_MIN_MAXH2CDATA)) { 1410c2700d28SVarun Prakash pr_err("queue %d: invalid maxh2cdata returned %u\n", 1411c2700d28SVarun Prakash nvme_tcp_queue_id(queue), maxh2cdata); 1412c2700d28SVarun Prakash goto free_icresp; 1413c2700d28SVarun Prakash } 1414c2700d28SVarun Prakash queue->maxh2cdata = maxh2cdata; 1415c2700d28SVarun Prakash 14163f2304f8SSagi Grimberg ret = 0; 14173f2304f8SSagi Grimberg free_icresp: 14183f2304f8SSagi Grimberg kfree(icresp); 14193f2304f8SSagi Grimberg free_icreq: 14203f2304f8SSagi Grimberg kfree(icreq); 14213f2304f8SSagi Grimberg return ret; 14223f2304f8SSagi Grimberg } 14233f2304f8SSagi Grimberg 142440510a63SSagi Grimberg static bool nvme_tcp_admin_queue(struct nvme_tcp_queue *queue) 142540510a63SSagi Grimberg { 142640510a63SSagi Grimberg return nvme_tcp_queue_id(queue) == 0; 142740510a63SSagi Grimberg } 142840510a63SSagi Grimberg 142940510a63SSagi Grimberg static bool nvme_tcp_default_queue(struct nvme_tcp_queue *queue) 143040510a63SSagi Grimberg { 143140510a63SSagi Grimberg struct nvme_tcp_ctrl *ctrl = queue->ctrl; 143240510a63SSagi Grimberg int qid = nvme_tcp_queue_id(queue); 143340510a63SSagi Grimberg 143440510a63SSagi Grimberg return !nvme_tcp_admin_queue(queue) && 143540510a63SSagi Grimberg qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT]; 143640510a63SSagi Grimberg } 143740510a63SSagi Grimberg 143840510a63SSagi Grimberg static bool nvme_tcp_read_queue(struct nvme_tcp_queue *queue) 143940510a63SSagi Grimberg { 144040510a63SSagi Grimberg struct nvme_tcp_ctrl *ctrl = queue->ctrl; 144140510a63SSagi Grimberg int qid = nvme_tcp_queue_id(queue); 144240510a63SSagi Grimberg 144340510a63SSagi Grimberg return !nvme_tcp_admin_queue(queue) && 144440510a63SSagi Grimberg !nvme_tcp_default_queue(queue) && 144540510a63SSagi Grimberg qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT] + 144640510a63SSagi Grimberg ctrl->io_queues[HCTX_TYPE_READ]; 144740510a63SSagi Grimberg } 144840510a63SSagi Grimberg 144940510a63SSagi Grimberg static bool nvme_tcp_poll_queue(struct nvme_tcp_queue *queue) 145040510a63SSagi Grimberg { 145140510a63SSagi Grimberg struct nvme_tcp_ctrl *ctrl = queue->ctrl; 145240510a63SSagi Grimberg int qid = nvme_tcp_queue_id(queue); 145340510a63SSagi Grimberg 145440510a63SSagi Grimberg return !nvme_tcp_admin_queue(queue) && 145540510a63SSagi Grimberg !nvme_tcp_default_queue(queue) && 145640510a63SSagi Grimberg !nvme_tcp_read_queue(queue) && 145740510a63SSagi Grimberg qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT] + 145840510a63SSagi Grimberg ctrl->io_queues[HCTX_TYPE_READ] + 145940510a63SSagi Grimberg ctrl->io_queues[HCTX_TYPE_POLL]; 146040510a63SSagi Grimberg } 146140510a63SSagi Grimberg 146240510a63SSagi Grimberg static void nvme_tcp_set_queue_io_cpu(struct nvme_tcp_queue *queue) 146340510a63SSagi Grimberg { 146440510a63SSagi Grimberg struct nvme_tcp_ctrl *ctrl = queue->ctrl; 146540510a63SSagi Grimberg int qid = nvme_tcp_queue_id(queue); 146640510a63SSagi Grimberg int n = 0; 146740510a63SSagi Grimberg 146840510a63SSagi Grimberg if (nvme_tcp_default_queue(queue)) 146940510a63SSagi Grimberg n = qid - 1; 147040510a63SSagi Grimberg else if (nvme_tcp_read_queue(queue)) 147140510a63SSagi Grimberg n = qid - ctrl->io_queues[HCTX_TYPE_DEFAULT] - 1; 147240510a63SSagi Grimberg else if (nvme_tcp_poll_queue(queue)) 147340510a63SSagi Grimberg n = qid - ctrl->io_queues[HCTX_TYPE_DEFAULT] - 147440510a63SSagi Grimberg ctrl->io_queues[HCTX_TYPE_READ] - 1; 147540510a63SSagi Grimberg queue->io_cpu = cpumask_next_wrap(n - 1, cpu_online_mask, -1, false); 147640510a63SSagi Grimberg } 147740510a63SSagi Grimberg 1478fb8745d0SChristoph Hellwig static int nvme_tcp_alloc_queue(struct nvme_ctrl *nctrl, int qid) 14793f2304f8SSagi Grimberg { 14803f2304f8SSagi Grimberg struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl); 14813f2304f8SSagi Grimberg struct nvme_tcp_queue *queue = &ctrl->queues[qid]; 14826ebf71baSChristoph Hellwig int ret, rcv_pdu_size; 14833f2304f8SSagi Grimberg 14849ebbfe49SChao Leng mutex_init(&queue->queue_lock); 14853f2304f8SSagi Grimberg queue->ctrl = ctrl; 148615ec928aSSagi Grimberg init_llist_head(&queue->req_list); 14873f2304f8SSagi Grimberg INIT_LIST_HEAD(&queue->send_list); 1488db5ad6b7SSagi Grimberg mutex_init(&queue->send_mutex); 14893f2304f8SSagi Grimberg INIT_WORK(&queue->io_work, nvme_tcp_io_work); 14903f2304f8SSagi Grimberg 14913f2304f8SSagi Grimberg if (qid > 0) 14929924b030SIsrael Rukshin queue->cmnd_capsule_len = nctrl->ioccsz * 16; 14933f2304f8SSagi Grimberg else 14943f2304f8SSagi Grimberg queue->cmnd_capsule_len = sizeof(struct nvme_command) + 14953f2304f8SSagi Grimberg NVME_TCP_ADMIN_CCSZ; 14963f2304f8SSagi Grimberg 14973f2304f8SSagi Grimberg ret = sock_create(ctrl->addr.ss_family, SOCK_STREAM, 14983f2304f8SSagi Grimberg IPPROTO_TCP, &queue->sock); 14993f2304f8SSagi Grimberg if (ret) { 15009924b030SIsrael Rukshin dev_err(nctrl->device, 15013f2304f8SSagi Grimberg "failed to create socket: %d\n", ret); 15029ebbfe49SChao Leng goto err_destroy_mutex; 15033f2304f8SSagi Grimberg } 15043f2304f8SSagi Grimberg 1505841aee4dSChris Leech nvme_tcp_reclassify_socket(queue->sock); 1506841aee4dSChris Leech 15073f2304f8SSagi Grimberg /* Single syn retry */ 1508557eadfcSChristoph Hellwig tcp_sock_set_syncnt(queue->sock->sk, 1); 15093f2304f8SSagi Grimberg 15103f2304f8SSagi Grimberg /* Set TCP no delay */ 151112abc5eeSChristoph Hellwig tcp_sock_set_nodelay(queue->sock->sk); 15123f2304f8SSagi Grimberg 15133f2304f8SSagi Grimberg /* 15143f2304f8SSagi Grimberg * Cleanup whatever is sitting in the TCP transmit queue on socket 15153f2304f8SSagi Grimberg * close. This is done to prevent stale data from being sent should 15163f2304f8SSagi Grimberg * the network connection be restored before TCP times out. 15173f2304f8SSagi Grimberg */ 1518c433594cSChristoph Hellwig sock_no_linger(queue->sock->sk); 15193f2304f8SSagi Grimberg 15206e434967SChristoph Hellwig if (so_priority > 0) 15216e434967SChristoph Hellwig sock_set_priority(queue->sock->sk, so_priority); 15229912ade3SWunderlich, Mark 1523bb13985dSIsrael Rukshin /* Set socket type of service */ 15246ebf71baSChristoph Hellwig if (nctrl->opts->tos >= 0) 15256ebf71baSChristoph Hellwig ip_sock_set_tos(queue->sock->sk, nctrl->opts->tos); 1526bb13985dSIsrael Rukshin 1527adc99fd3SSagi Grimberg /* Set 10 seconds timeout for icresp recvmsg */ 1528adc99fd3SSagi Grimberg queue->sock->sk->sk_rcvtimeo = 10 * HZ; 1529adc99fd3SSagi Grimberg 15303f2304f8SSagi Grimberg queue->sock->sk->sk_allocation = GFP_ATOMIC; 153140510a63SSagi Grimberg nvme_tcp_set_queue_io_cpu(queue); 15323f2304f8SSagi Grimberg queue->request = NULL; 15333f2304f8SSagi Grimberg queue->data_remaining = 0; 15343f2304f8SSagi Grimberg queue->ddgst_remaining = 0; 15353f2304f8SSagi Grimberg queue->pdu_remaining = 0; 15363f2304f8SSagi Grimberg queue->pdu_offset = 0; 15373f2304f8SSagi Grimberg sk_set_memalloc(queue->sock->sk); 15383f2304f8SSagi Grimberg 15399924b030SIsrael Rukshin if (nctrl->opts->mask & NVMF_OPT_HOST_TRADDR) { 15403f2304f8SSagi Grimberg ret = kernel_bind(queue->sock, (struct sockaddr *)&ctrl->src_addr, 15413f2304f8SSagi Grimberg sizeof(ctrl->src_addr)); 15423f2304f8SSagi Grimberg if (ret) { 15439924b030SIsrael Rukshin dev_err(nctrl->device, 15443f2304f8SSagi Grimberg "failed to bind queue %d socket %d\n", 15453f2304f8SSagi Grimberg qid, ret); 15463f2304f8SSagi Grimberg goto err_sock; 15473f2304f8SSagi Grimberg } 15483f2304f8SSagi Grimberg } 15493f2304f8SSagi Grimberg 15503ede8f72SMartin Belanger if (nctrl->opts->mask & NVMF_OPT_HOST_IFACE) { 15513ede8f72SMartin Belanger char *iface = nctrl->opts->host_iface; 15523ede8f72SMartin Belanger sockptr_t optval = KERNEL_SOCKPTR(iface); 15533ede8f72SMartin Belanger 15543ede8f72SMartin Belanger ret = sock_setsockopt(queue->sock, SOL_SOCKET, SO_BINDTODEVICE, 15553ede8f72SMartin Belanger optval, strlen(iface)); 15563ede8f72SMartin Belanger if (ret) { 15573ede8f72SMartin Belanger dev_err(nctrl->device, 15583ede8f72SMartin Belanger "failed to bind to interface %s queue %d err %d\n", 15593ede8f72SMartin Belanger iface, qid, ret); 15603ede8f72SMartin Belanger goto err_sock; 15613ede8f72SMartin Belanger } 15623ede8f72SMartin Belanger } 15633ede8f72SMartin Belanger 15643f2304f8SSagi Grimberg queue->hdr_digest = nctrl->opts->hdr_digest; 15653f2304f8SSagi Grimberg queue->data_digest = nctrl->opts->data_digest; 15663f2304f8SSagi Grimberg if (queue->hdr_digest || queue->data_digest) { 15673f2304f8SSagi Grimberg ret = nvme_tcp_alloc_crypto(queue); 15683f2304f8SSagi Grimberg if (ret) { 15699924b030SIsrael Rukshin dev_err(nctrl->device, 15703f2304f8SSagi Grimberg "failed to allocate queue %d crypto\n", qid); 15713f2304f8SSagi Grimberg goto err_sock; 15723f2304f8SSagi Grimberg } 15733f2304f8SSagi Grimberg } 15743f2304f8SSagi Grimberg 15753f2304f8SSagi Grimberg rcv_pdu_size = sizeof(struct nvme_tcp_rsp_pdu) + 15763f2304f8SSagi Grimberg nvme_tcp_hdgst_len(queue); 15773f2304f8SSagi Grimberg queue->pdu = kmalloc(rcv_pdu_size, GFP_KERNEL); 15783f2304f8SSagi Grimberg if (!queue->pdu) { 15793f2304f8SSagi Grimberg ret = -ENOMEM; 15803f2304f8SSagi Grimberg goto err_crypto; 15813f2304f8SSagi Grimberg } 15823f2304f8SSagi Grimberg 15839924b030SIsrael Rukshin dev_dbg(nctrl->device, "connecting queue %d\n", 15843f2304f8SSagi Grimberg nvme_tcp_queue_id(queue)); 15853f2304f8SSagi Grimberg 15863f2304f8SSagi Grimberg ret = kernel_connect(queue->sock, (struct sockaddr *)&ctrl->addr, 15873f2304f8SSagi Grimberg sizeof(ctrl->addr), 0); 15883f2304f8SSagi Grimberg if (ret) { 15899924b030SIsrael Rukshin dev_err(nctrl->device, 15903f2304f8SSagi Grimberg "failed to connect socket: %d\n", ret); 15913f2304f8SSagi Grimberg goto err_rcv_pdu; 15923f2304f8SSagi Grimberg } 15933f2304f8SSagi Grimberg 15943f2304f8SSagi Grimberg ret = nvme_tcp_init_connection(queue); 15953f2304f8SSagi Grimberg if (ret) 15963f2304f8SSagi Grimberg goto err_init_connect; 15973f2304f8SSagi Grimberg 15983f2304f8SSagi Grimberg queue->rd_enabled = true; 15993f2304f8SSagi Grimberg set_bit(NVME_TCP_Q_ALLOCATED, &queue->flags); 16003f2304f8SSagi Grimberg nvme_tcp_init_recv_ctx(queue); 16013f2304f8SSagi Grimberg 16023f2304f8SSagi Grimberg write_lock_bh(&queue->sock->sk->sk_callback_lock); 16033f2304f8SSagi Grimberg queue->sock->sk->sk_user_data = queue; 16043f2304f8SSagi Grimberg queue->state_change = queue->sock->sk->sk_state_change; 16053f2304f8SSagi Grimberg queue->data_ready = queue->sock->sk->sk_data_ready; 16063f2304f8SSagi Grimberg queue->write_space = queue->sock->sk->sk_write_space; 16073f2304f8SSagi Grimberg queue->sock->sk->sk_data_ready = nvme_tcp_data_ready; 16083f2304f8SSagi Grimberg queue->sock->sk->sk_state_change = nvme_tcp_state_change; 16093f2304f8SSagi Grimberg queue->sock->sk->sk_write_space = nvme_tcp_write_space; 1610ac1c4e18SSebastian Andrzej Siewior #ifdef CONFIG_NET_RX_BUSY_POLL 16111a9460ceSSagi Grimberg queue->sock->sk->sk_ll_usec = 1; 1612ac1c4e18SSebastian Andrzej Siewior #endif 16133f2304f8SSagi Grimberg write_unlock_bh(&queue->sock->sk->sk_callback_lock); 16143f2304f8SSagi Grimberg 16153f2304f8SSagi Grimberg return 0; 16163f2304f8SSagi Grimberg 16173f2304f8SSagi Grimberg err_init_connect: 16183f2304f8SSagi Grimberg kernel_sock_shutdown(queue->sock, SHUT_RDWR); 16193f2304f8SSagi Grimberg err_rcv_pdu: 16203f2304f8SSagi Grimberg kfree(queue->pdu); 16213f2304f8SSagi Grimberg err_crypto: 16223f2304f8SSagi Grimberg if (queue->hdr_digest || queue->data_digest) 16233f2304f8SSagi Grimberg nvme_tcp_free_crypto(queue); 16243f2304f8SSagi Grimberg err_sock: 16253f2304f8SSagi Grimberg sock_release(queue->sock); 16263f2304f8SSagi Grimberg queue->sock = NULL; 16279ebbfe49SChao Leng err_destroy_mutex: 1628d48f92cdSKeith Busch mutex_destroy(&queue->send_mutex); 16299ebbfe49SChao Leng mutex_destroy(&queue->queue_lock); 16303f2304f8SSagi Grimberg return ret; 16313f2304f8SSagi Grimberg } 16323f2304f8SSagi Grimberg 16333f2304f8SSagi Grimberg static void nvme_tcp_restore_sock_calls(struct nvme_tcp_queue *queue) 16343f2304f8SSagi Grimberg { 16353f2304f8SSagi Grimberg struct socket *sock = queue->sock; 16363f2304f8SSagi Grimberg 16373f2304f8SSagi Grimberg write_lock_bh(&sock->sk->sk_callback_lock); 16383f2304f8SSagi Grimberg sock->sk->sk_user_data = NULL; 16393f2304f8SSagi Grimberg sock->sk->sk_data_ready = queue->data_ready; 16403f2304f8SSagi Grimberg sock->sk->sk_state_change = queue->state_change; 16413f2304f8SSagi Grimberg sock->sk->sk_write_space = queue->write_space; 16423f2304f8SSagi Grimberg write_unlock_bh(&sock->sk->sk_callback_lock); 16433f2304f8SSagi Grimberg } 16443f2304f8SSagi Grimberg 16453f2304f8SSagi Grimberg static void __nvme_tcp_stop_queue(struct nvme_tcp_queue *queue) 16463f2304f8SSagi Grimberg { 16473f2304f8SSagi Grimberg kernel_sock_shutdown(queue->sock, SHUT_RDWR); 16483f2304f8SSagi Grimberg nvme_tcp_restore_sock_calls(queue); 16493f2304f8SSagi Grimberg cancel_work_sync(&queue->io_work); 16503f2304f8SSagi Grimberg } 16513f2304f8SSagi Grimberg 16523f2304f8SSagi Grimberg static void nvme_tcp_stop_queue(struct nvme_ctrl *nctrl, int qid) 16533f2304f8SSagi Grimberg { 16543f2304f8SSagi Grimberg struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl); 16553f2304f8SSagi Grimberg struct nvme_tcp_queue *queue = &ctrl->queues[qid]; 16563f2304f8SSagi Grimberg 16572bff487fSMaurizio Lombardi if (!test_bit(NVME_TCP_Q_ALLOCATED, &queue->flags)) 16582bff487fSMaurizio Lombardi return; 16592bff487fSMaurizio Lombardi 16609ebbfe49SChao Leng mutex_lock(&queue->queue_lock); 16619ebbfe49SChao Leng if (test_and_clear_bit(NVME_TCP_Q_LIVE, &queue->flags)) 16623f2304f8SSagi Grimberg __nvme_tcp_stop_queue(queue); 16639ebbfe49SChao Leng mutex_unlock(&queue->queue_lock); 16643f2304f8SSagi Grimberg } 16653f2304f8SSagi Grimberg 16663f2304f8SSagi Grimberg static int nvme_tcp_start_queue(struct nvme_ctrl *nctrl, int idx) 16673f2304f8SSagi Grimberg { 16683f2304f8SSagi Grimberg struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl); 16693f2304f8SSagi Grimberg int ret; 16703f2304f8SSagi Grimberg 16713f2304f8SSagi Grimberg if (idx) 1672be42a33bSKeith Busch ret = nvmf_connect_io_queue(nctrl, idx); 16733f2304f8SSagi Grimberg else 16743f2304f8SSagi Grimberg ret = nvmf_connect_admin_queue(nctrl); 16753f2304f8SSagi Grimberg 16763f2304f8SSagi Grimberg if (!ret) { 16773f2304f8SSagi Grimberg set_bit(NVME_TCP_Q_LIVE, &ctrl->queues[idx].flags); 16783f2304f8SSagi Grimberg } else { 1679f34e2589SSagi Grimberg if (test_bit(NVME_TCP_Q_ALLOCATED, &ctrl->queues[idx].flags)) 16803f2304f8SSagi Grimberg __nvme_tcp_stop_queue(&ctrl->queues[idx]); 16813f2304f8SSagi Grimberg dev_err(nctrl->device, 16823f2304f8SSagi Grimberg "failed to connect queue: %d ret=%d\n", idx, ret); 16833f2304f8SSagi Grimberg } 16843f2304f8SSagi Grimberg return ret; 16853f2304f8SSagi Grimberg } 16863f2304f8SSagi Grimberg 16873f2304f8SSagi Grimberg static void nvme_tcp_free_admin_queue(struct nvme_ctrl *ctrl) 16883f2304f8SSagi Grimberg { 16893f2304f8SSagi Grimberg if (to_tcp_ctrl(ctrl)->async_req.pdu) { 1690ceb1e087SDavid Milburn cancel_work_sync(&ctrl->async_event_work); 16913f2304f8SSagi Grimberg nvme_tcp_free_async_req(to_tcp_ctrl(ctrl)); 16923f2304f8SSagi Grimberg to_tcp_ctrl(ctrl)->async_req.pdu = NULL; 16933f2304f8SSagi Grimberg } 16943f2304f8SSagi Grimberg 16953f2304f8SSagi Grimberg nvme_tcp_free_queue(ctrl, 0); 16963f2304f8SSagi Grimberg } 16973f2304f8SSagi Grimberg 16983f2304f8SSagi Grimberg static void nvme_tcp_free_io_queues(struct nvme_ctrl *ctrl) 16993f2304f8SSagi Grimberg { 17003f2304f8SSagi Grimberg int i; 17013f2304f8SSagi Grimberg 17023f2304f8SSagi Grimberg for (i = 1; i < ctrl->queue_count; i++) 17033f2304f8SSagi Grimberg nvme_tcp_free_queue(ctrl, i); 17043f2304f8SSagi Grimberg } 17053f2304f8SSagi Grimberg 17063f2304f8SSagi Grimberg static void nvme_tcp_stop_io_queues(struct nvme_ctrl *ctrl) 17073f2304f8SSagi Grimberg { 17083f2304f8SSagi Grimberg int i; 17093f2304f8SSagi Grimberg 17103f2304f8SSagi Grimberg for (i = 1; i < ctrl->queue_count; i++) 17113f2304f8SSagi Grimberg nvme_tcp_stop_queue(ctrl, i); 17123f2304f8SSagi Grimberg } 17133f2304f8SSagi Grimberg 171409035f86SDaniel Wagner static int nvme_tcp_start_io_queues(struct nvme_ctrl *ctrl, 171509035f86SDaniel Wagner int first, int last) 17163f2304f8SSagi Grimberg { 1717462b8b2dSChaitanya Kulkarni int i, ret; 17183f2304f8SSagi Grimberg 171909035f86SDaniel Wagner for (i = first; i < last; i++) { 17203f2304f8SSagi Grimberg ret = nvme_tcp_start_queue(ctrl, i); 17213f2304f8SSagi Grimberg if (ret) 17223f2304f8SSagi Grimberg goto out_stop_queues; 17233f2304f8SSagi Grimberg } 17243f2304f8SSagi Grimberg 17253f2304f8SSagi Grimberg return 0; 17263f2304f8SSagi Grimberg 17273f2304f8SSagi Grimberg out_stop_queues: 172809035f86SDaniel Wagner for (i--; i >= first; i--) 17293f2304f8SSagi Grimberg nvme_tcp_stop_queue(ctrl, i); 17303f2304f8SSagi Grimberg return ret; 17313f2304f8SSagi Grimberg } 17323f2304f8SSagi Grimberg 17333f2304f8SSagi Grimberg static int nvme_tcp_alloc_admin_queue(struct nvme_ctrl *ctrl) 17343f2304f8SSagi Grimberg { 17353f2304f8SSagi Grimberg int ret; 17363f2304f8SSagi Grimberg 1737fb8745d0SChristoph Hellwig ret = nvme_tcp_alloc_queue(ctrl, 0); 17383f2304f8SSagi Grimberg if (ret) 17393f2304f8SSagi Grimberg return ret; 17403f2304f8SSagi Grimberg 17413f2304f8SSagi Grimberg ret = nvme_tcp_alloc_async_req(to_tcp_ctrl(ctrl)); 17423f2304f8SSagi Grimberg if (ret) 17433f2304f8SSagi Grimberg goto out_free_queue; 17443f2304f8SSagi Grimberg 17453f2304f8SSagi Grimberg return 0; 17463f2304f8SSagi Grimberg 17473f2304f8SSagi Grimberg out_free_queue: 17483f2304f8SSagi Grimberg nvme_tcp_free_queue(ctrl, 0); 17493f2304f8SSagi Grimberg return ret; 17503f2304f8SSagi Grimberg } 17513f2304f8SSagi Grimberg 1752efb973b1SSagi Grimberg static int __nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl) 17533f2304f8SSagi Grimberg { 17543f2304f8SSagi Grimberg int i, ret; 17553f2304f8SSagi Grimberg 17563f2304f8SSagi Grimberg for (i = 1; i < ctrl->queue_count; i++) { 1757fb8745d0SChristoph Hellwig ret = nvme_tcp_alloc_queue(ctrl, i); 17583f2304f8SSagi Grimberg if (ret) 17593f2304f8SSagi Grimberg goto out_free_queues; 17603f2304f8SSagi Grimberg } 17613f2304f8SSagi Grimberg 17623f2304f8SSagi Grimberg return 0; 17633f2304f8SSagi Grimberg 17643f2304f8SSagi Grimberg out_free_queues: 17653f2304f8SSagi Grimberg for (i--; i >= 1; i--) 17663f2304f8SSagi Grimberg nvme_tcp_free_queue(ctrl, i); 17673f2304f8SSagi Grimberg 17683f2304f8SSagi Grimberg return ret; 17693f2304f8SSagi Grimberg } 17703f2304f8SSagi Grimberg 17713f2304f8SSagi Grimberg static unsigned int nvme_tcp_nr_io_queues(struct nvme_ctrl *ctrl) 17723f2304f8SSagi Grimberg { 1773873946f4SSagi Grimberg unsigned int nr_io_queues; 1774873946f4SSagi Grimberg 1775873946f4SSagi Grimberg nr_io_queues = min(ctrl->opts->nr_io_queues, num_online_cpus()); 1776873946f4SSagi Grimberg nr_io_queues += min(ctrl->opts->nr_write_queues, num_online_cpus()); 17771a9460ceSSagi Grimberg nr_io_queues += min(ctrl->opts->nr_poll_queues, num_online_cpus()); 1778873946f4SSagi Grimberg 1779873946f4SSagi Grimberg return nr_io_queues; 17803f2304f8SSagi Grimberg } 17813f2304f8SSagi Grimberg 178264861993SSagi Grimberg static void nvme_tcp_set_io_queues(struct nvme_ctrl *nctrl, 178364861993SSagi Grimberg unsigned int nr_io_queues) 178464861993SSagi Grimberg { 178564861993SSagi Grimberg struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl); 178664861993SSagi Grimberg struct nvmf_ctrl_options *opts = nctrl->opts; 178764861993SSagi Grimberg 178864861993SSagi Grimberg if (opts->nr_write_queues && opts->nr_io_queues < nr_io_queues) { 178964861993SSagi Grimberg /* 179064861993SSagi Grimberg * separate read/write queues 179164861993SSagi Grimberg * hand out dedicated default queues only after we have 179264861993SSagi Grimberg * sufficient read queues. 179364861993SSagi Grimberg */ 179464861993SSagi Grimberg ctrl->io_queues[HCTX_TYPE_READ] = opts->nr_io_queues; 179564861993SSagi Grimberg nr_io_queues -= ctrl->io_queues[HCTX_TYPE_READ]; 179664861993SSagi Grimberg ctrl->io_queues[HCTX_TYPE_DEFAULT] = 179764861993SSagi Grimberg min(opts->nr_write_queues, nr_io_queues); 179864861993SSagi Grimberg nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT]; 179964861993SSagi Grimberg } else { 180064861993SSagi Grimberg /* 180164861993SSagi Grimberg * shared read/write queues 180264861993SSagi Grimberg * either no write queues were requested, or we don't have 180364861993SSagi Grimberg * sufficient queue count to have dedicated default queues. 180464861993SSagi Grimberg */ 180564861993SSagi Grimberg ctrl->io_queues[HCTX_TYPE_DEFAULT] = 180664861993SSagi Grimberg min(opts->nr_io_queues, nr_io_queues); 180764861993SSagi Grimberg nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT]; 180864861993SSagi Grimberg } 18091a9460ceSSagi Grimberg 18101a9460ceSSagi Grimberg if (opts->nr_poll_queues && nr_io_queues) { 18111a9460ceSSagi Grimberg /* map dedicated poll queues only if we have queues left */ 18121a9460ceSSagi Grimberg ctrl->io_queues[HCTX_TYPE_POLL] = 18131a9460ceSSagi Grimberg min(opts->nr_poll_queues, nr_io_queues); 18141a9460ceSSagi Grimberg } 181564861993SSagi Grimberg } 181664861993SSagi Grimberg 1817efb973b1SSagi Grimberg static int nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl) 18183f2304f8SSagi Grimberg { 18193f2304f8SSagi Grimberg unsigned int nr_io_queues; 18203f2304f8SSagi Grimberg int ret; 18213f2304f8SSagi Grimberg 18223f2304f8SSagi Grimberg nr_io_queues = nvme_tcp_nr_io_queues(ctrl); 18233f2304f8SSagi Grimberg ret = nvme_set_queue_count(ctrl, &nr_io_queues); 18243f2304f8SSagi Grimberg if (ret) 18253f2304f8SSagi Grimberg return ret; 18263f2304f8SSagi Grimberg 1827664227fdSRuozhu Li if (nr_io_queues == 0) { 182872f57242SSagi Grimberg dev_err(ctrl->device, 182972f57242SSagi Grimberg "unable to set any I/O queues\n"); 183072f57242SSagi Grimberg return -ENOMEM; 183172f57242SSagi Grimberg } 18323f2304f8SSagi Grimberg 1833664227fdSRuozhu Li ctrl->queue_count = nr_io_queues + 1; 18343f2304f8SSagi Grimberg dev_info(ctrl->device, 18353f2304f8SSagi Grimberg "creating %d I/O queues.\n", nr_io_queues); 18363f2304f8SSagi Grimberg 183764861993SSagi Grimberg nvme_tcp_set_io_queues(ctrl, nr_io_queues); 183864861993SSagi Grimberg 1839efb973b1SSagi Grimberg return __nvme_tcp_alloc_io_queues(ctrl); 18403f2304f8SSagi Grimberg } 18413f2304f8SSagi Grimberg 18423f2304f8SSagi Grimberg static void nvme_tcp_destroy_io_queues(struct nvme_ctrl *ctrl, bool remove) 18433f2304f8SSagi Grimberg { 18443f2304f8SSagi Grimberg nvme_tcp_stop_io_queues(ctrl); 1845de777825SChristoph Hellwig if (remove) 1846de777825SChristoph Hellwig nvme_remove_io_tag_set(ctrl); 18473f2304f8SSagi Grimberg nvme_tcp_free_io_queues(ctrl); 18483f2304f8SSagi Grimberg } 18493f2304f8SSagi Grimberg 18503f2304f8SSagi Grimberg static int nvme_tcp_configure_io_queues(struct nvme_ctrl *ctrl, bool new) 18513f2304f8SSagi Grimberg { 185209035f86SDaniel Wagner int ret, nr_queues; 18533f2304f8SSagi Grimberg 1854efb973b1SSagi Grimberg ret = nvme_tcp_alloc_io_queues(ctrl); 18553f2304f8SSagi Grimberg if (ret) 18563f2304f8SSagi Grimberg return ret; 18573f2304f8SSagi Grimberg 18583f2304f8SSagi Grimberg if (new) { 1859de777825SChristoph Hellwig ret = nvme_alloc_io_tag_set(ctrl, &to_tcp_ctrl(ctrl)->tag_set, 1860de777825SChristoph Hellwig &nvme_tcp_mq_ops, 1861de777825SChristoph Hellwig BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_BLOCKING, 1862de777825SChristoph Hellwig sizeof(struct nvme_tcp_request)); 18632f7a7e5dSChristoph Hellwig if (ret) 18643f2304f8SSagi Grimberg goto out_free_io_queues; 18653f2304f8SSagi Grimberg } 18663f2304f8SSagi Grimberg 186709035f86SDaniel Wagner /* 186809035f86SDaniel Wagner * Only start IO queues for which we have allocated the tagset 186909035f86SDaniel Wagner * and limitted it to the available queues. On reconnects, the 187009035f86SDaniel Wagner * queue number might have changed. 187109035f86SDaniel Wagner */ 187209035f86SDaniel Wagner nr_queues = min(ctrl->tagset->nr_hw_queues + 1, ctrl->queue_count); 187309035f86SDaniel Wagner ret = nvme_tcp_start_io_queues(ctrl, 1, nr_queues); 18743f2304f8SSagi Grimberg if (ret) 18753f2304f8SSagi Grimberg goto out_cleanup_connect_q; 18763f2304f8SSagi Grimberg 18772875b0aeSSagi Grimberg if (!new) { 18789f27bd70SChristoph Hellwig nvme_unquiesce_io_queues(ctrl); 1879e5c01f4fSSagi Grimberg if (!nvme_wait_freeze_timeout(ctrl, NVME_IO_TIMEOUT)) { 1880e5c01f4fSSagi Grimberg /* 1881e5c01f4fSSagi Grimberg * If we timed out waiting for freeze we are likely to 1882e5c01f4fSSagi Grimberg * be stuck. Fail the controller initialization just 1883e5c01f4fSSagi Grimberg * to be safe. 1884e5c01f4fSSagi Grimberg */ 1885e5c01f4fSSagi Grimberg ret = -ENODEV; 1886e5c01f4fSSagi Grimberg goto out_wait_freeze_timed_out; 1887e5c01f4fSSagi Grimberg } 18882875b0aeSSagi Grimberg blk_mq_update_nr_hw_queues(ctrl->tagset, 18892875b0aeSSagi Grimberg ctrl->queue_count - 1); 18902875b0aeSSagi Grimberg nvme_unfreeze(ctrl); 18912875b0aeSSagi Grimberg } 18922875b0aeSSagi Grimberg 189309035f86SDaniel Wagner /* 189409035f86SDaniel Wagner * If the number of queues has increased (reconnect case) 189509035f86SDaniel Wagner * start all new queues now. 189609035f86SDaniel Wagner */ 189709035f86SDaniel Wagner ret = nvme_tcp_start_io_queues(ctrl, nr_queues, 189809035f86SDaniel Wagner ctrl->tagset->nr_hw_queues + 1); 189909035f86SDaniel Wagner if (ret) 190009035f86SDaniel Wagner goto out_wait_freeze_timed_out; 190109035f86SDaniel Wagner 19023f2304f8SSagi Grimberg return 0; 19033f2304f8SSagi Grimberg 1904e5c01f4fSSagi Grimberg out_wait_freeze_timed_out: 19059f27bd70SChristoph Hellwig nvme_quiesce_io_queues(ctrl); 190670a99574SChao Leng nvme_sync_io_queues(ctrl); 1907e5c01f4fSSagi Grimberg nvme_tcp_stop_io_queues(ctrl); 19083f2304f8SSagi Grimberg out_cleanup_connect_q: 190970a99574SChao Leng nvme_cancel_tagset(ctrl); 1910e85037a2SSagi Grimberg if (new) 1911de777825SChristoph Hellwig nvme_remove_io_tag_set(ctrl); 19123f2304f8SSagi Grimberg out_free_io_queues: 19133f2304f8SSagi Grimberg nvme_tcp_free_io_queues(ctrl); 19143f2304f8SSagi Grimberg return ret; 19153f2304f8SSagi Grimberg } 19163f2304f8SSagi Grimberg 19173f2304f8SSagi Grimberg static void nvme_tcp_destroy_admin_queue(struct nvme_ctrl *ctrl, bool remove) 19183f2304f8SSagi Grimberg { 19193f2304f8SSagi Grimberg nvme_tcp_stop_queue(ctrl, 0); 1920de777825SChristoph Hellwig if (remove) 1921de777825SChristoph Hellwig nvme_remove_admin_tag_set(ctrl); 19223f2304f8SSagi Grimberg nvme_tcp_free_admin_queue(ctrl); 19233f2304f8SSagi Grimberg } 19243f2304f8SSagi Grimberg 19253f2304f8SSagi Grimberg static int nvme_tcp_configure_admin_queue(struct nvme_ctrl *ctrl, bool new) 19263f2304f8SSagi Grimberg { 19273f2304f8SSagi Grimberg int error; 19283f2304f8SSagi Grimberg 19293f2304f8SSagi Grimberg error = nvme_tcp_alloc_admin_queue(ctrl); 19303f2304f8SSagi Grimberg if (error) 19313f2304f8SSagi Grimberg return error; 19323f2304f8SSagi Grimberg 19333f2304f8SSagi Grimberg if (new) { 1934de777825SChristoph Hellwig error = nvme_alloc_admin_tag_set(ctrl, 1935de777825SChristoph Hellwig &to_tcp_ctrl(ctrl)->admin_tag_set, 1936de777825SChristoph Hellwig &nvme_tcp_admin_mq_ops, BLK_MQ_F_BLOCKING, 1937de777825SChristoph Hellwig sizeof(struct nvme_tcp_request)); 19382f7a7e5dSChristoph Hellwig if (error) 19393f2304f8SSagi Grimberg goto out_free_queue; 19403f2304f8SSagi Grimberg } 19413f2304f8SSagi Grimberg 19423f2304f8SSagi Grimberg error = nvme_tcp_start_queue(ctrl, 0); 19433f2304f8SSagi Grimberg if (error) 1944de777825SChristoph Hellwig goto out_cleanup_tagset; 19453f2304f8SSagi Grimberg 1946c0f2f45bSSagi Grimberg error = nvme_enable_ctrl(ctrl); 19473f2304f8SSagi Grimberg if (error) 19483f2304f8SSagi Grimberg goto out_stop_queue; 19493f2304f8SSagi Grimberg 19509f27bd70SChristoph Hellwig nvme_unquiesce_admin_queue(ctrl); 1951e7832cb4SSagi Grimberg 195294cc781fSChristoph Hellwig error = nvme_init_ctrl_finish(ctrl, false); 19533f2304f8SSagi Grimberg if (error) 195470a99574SChao Leng goto out_quiesce_queue; 19553f2304f8SSagi Grimberg 19563f2304f8SSagi Grimberg return 0; 19573f2304f8SSagi Grimberg 195870a99574SChao Leng out_quiesce_queue: 19599f27bd70SChristoph Hellwig nvme_quiesce_admin_queue(ctrl); 196070a99574SChao Leng blk_sync_queue(ctrl->admin_q); 19613f2304f8SSagi Grimberg out_stop_queue: 19623f2304f8SSagi Grimberg nvme_tcp_stop_queue(ctrl, 0); 196370a99574SChao Leng nvme_cancel_admin_tagset(ctrl); 1964de777825SChristoph Hellwig out_cleanup_tagset: 19653f2304f8SSagi Grimberg if (new) 1966de777825SChristoph Hellwig nvme_remove_admin_tag_set(ctrl); 19673f2304f8SSagi Grimberg out_free_queue: 19683f2304f8SSagi Grimberg nvme_tcp_free_admin_queue(ctrl); 19693f2304f8SSagi Grimberg return error; 19703f2304f8SSagi Grimberg } 19713f2304f8SSagi Grimberg 19723f2304f8SSagi Grimberg static void nvme_tcp_teardown_admin_queue(struct nvme_ctrl *ctrl, 19733f2304f8SSagi Grimberg bool remove) 19743f2304f8SSagi Grimberg { 19759f27bd70SChristoph Hellwig nvme_quiesce_admin_queue(ctrl); 1976d6f66210SChao Leng blk_sync_queue(ctrl->admin_q); 19773f2304f8SSagi Grimberg nvme_tcp_stop_queue(ctrl, 0); 1978563c8158SChao Leng nvme_cancel_admin_tagset(ctrl); 1979e7832cb4SSagi Grimberg if (remove) 19809f27bd70SChristoph Hellwig nvme_unquiesce_admin_queue(ctrl); 19813f2304f8SSagi Grimberg nvme_tcp_destroy_admin_queue(ctrl, remove); 19823f2304f8SSagi Grimberg } 19833f2304f8SSagi Grimberg 19843f2304f8SSagi Grimberg static void nvme_tcp_teardown_io_queues(struct nvme_ctrl *ctrl, 19853f2304f8SSagi Grimberg bool remove) 19863f2304f8SSagi Grimberg { 19873f2304f8SSagi Grimberg if (ctrl->queue_count <= 1) 1988d6f66210SChao Leng return; 19899f27bd70SChristoph Hellwig nvme_quiesce_admin_queue(ctrl); 19902875b0aeSSagi Grimberg nvme_start_freeze(ctrl); 19919f27bd70SChristoph Hellwig nvme_quiesce_io_queues(ctrl); 1992d6f66210SChao Leng nvme_sync_io_queues(ctrl); 19933f2304f8SSagi Grimberg nvme_tcp_stop_io_queues(ctrl); 1994563c8158SChao Leng nvme_cancel_tagset(ctrl); 19953f2304f8SSagi Grimberg if (remove) 19969f27bd70SChristoph Hellwig nvme_unquiesce_io_queues(ctrl); 19973f2304f8SSagi Grimberg nvme_tcp_destroy_io_queues(ctrl, remove); 19983f2304f8SSagi Grimberg } 19993f2304f8SSagi Grimberg 20003f2304f8SSagi Grimberg static void nvme_tcp_reconnect_or_remove(struct nvme_ctrl *ctrl) 20013f2304f8SSagi Grimberg { 20023f2304f8SSagi Grimberg /* If we are resetting/deleting then do nothing */ 20033f2304f8SSagi Grimberg if (ctrl->state != NVME_CTRL_CONNECTING) { 20043f2304f8SSagi Grimberg WARN_ON_ONCE(ctrl->state == NVME_CTRL_NEW || 20053f2304f8SSagi Grimberg ctrl->state == NVME_CTRL_LIVE); 20063f2304f8SSagi Grimberg return; 20073f2304f8SSagi Grimberg } 20083f2304f8SSagi Grimberg 20093f2304f8SSagi Grimberg if (nvmf_should_reconnect(ctrl)) { 20103f2304f8SSagi Grimberg dev_info(ctrl->device, "Reconnecting in %d seconds...\n", 20113f2304f8SSagi Grimberg ctrl->opts->reconnect_delay); 20123f2304f8SSagi Grimberg queue_delayed_work(nvme_wq, &to_tcp_ctrl(ctrl)->connect_work, 20133f2304f8SSagi Grimberg ctrl->opts->reconnect_delay * HZ); 20143f2304f8SSagi Grimberg } else { 20153f2304f8SSagi Grimberg dev_info(ctrl->device, "Removing controller...\n"); 20163f2304f8SSagi Grimberg nvme_delete_ctrl(ctrl); 20173f2304f8SSagi Grimberg } 20183f2304f8SSagi Grimberg } 20193f2304f8SSagi Grimberg 20203f2304f8SSagi Grimberg static int nvme_tcp_setup_ctrl(struct nvme_ctrl *ctrl, bool new) 20213f2304f8SSagi Grimberg { 20223f2304f8SSagi Grimberg struct nvmf_ctrl_options *opts = ctrl->opts; 2023312910f4SColin Ian King int ret; 20243f2304f8SSagi Grimberg 20253f2304f8SSagi Grimberg ret = nvme_tcp_configure_admin_queue(ctrl, new); 20263f2304f8SSagi Grimberg if (ret) 20273f2304f8SSagi Grimberg return ret; 20283f2304f8SSagi Grimberg 20293f2304f8SSagi Grimberg if (ctrl->icdoff) { 2030522af60cSDan Carpenter ret = -EOPNOTSUPP; 20313f2304f8SSagi Grimberg dev_err(ctrl->device, "icdoff is not supported!\n"); 20323f2304f8SSagi Grimberg goto destroy_admin; 20333f2304f8SSagi Grimberg } 20343f2304f8SSagi Grimberg 20353b54064fSChaitanya Kulkarni if (!nvme_ctrl_sgl_supported(ctrl)) { 2036522af60cSDan Carpenter ret = -EOPNOTSUPP; 203773ffcefcSMax Gurtovoy dev_err(ctrl->device, "Mandatory sgls are not supported!\n"); 203873ffcefcSMax Gurtovoy goto destroy_admin; 203973ffcefcSMax Gurtovoy } 204073ffcefcSMax Gurtovoy 20413f2304f8SSagi Grimberg if (opts->queue_size > ctrl->sqsize + 1) 20423f2304f8SSagi Grimberg dev_warn(ctrl->device, 20433f2304f8SSagi Grimberg "queue_size %zu > ctrl sqsize %u, clamping down\n", 20443f2304f8SSagi Grimberg opts->queue_size, ctrl->sqsize + 1); 20453f2304f8SSagi Grimberg 20463f2304f8SSagi Grimberg if (ctrl->sqsize + 1 > ctrl->maxcmd) { 20473f2304f8SSagi Grimberg dev_warn(ctrl->device, 20483f2304f8SSagi Grimberg "sqsize %u > ctrl maxcmd %u, clamping down\n", 20493f2304f8SSagi Grimberg ctrl->sqsize + 1, ctrl->maxcmd); 20503f2304f8SSagi Grimberg ctrl->sqsize = ctrl->maxcmd - 1; 20513f2304f8SSagi Grimberg } 20523f2304f8SSagi Grimberg 20533f2304f8SSagi Grimberg if (ctrl->queue_count > 1) { 20543f2304f8SSagi Grimberg ret = nvme_tcp_configure_io_queues(ctrl, new); 20553f2304f8SSagi Grimberg if (ret) 20563f2304f8SSagi Grimberg goto destroy_admin; 20573f2304f8SSagi Grimberg } 20583f2304f8SSagi Grimberg 20593f2304f8SSagi Grimberg if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_LIVE)) { 2060bea54ef5SIsrael Rukshin /* 2061ecca390eSSagi Grimberg * state change failure is ok if we started ctrl delete, 2062bea54ef5SIsrael Rukshin * unless we're during creation of a new controller to 2063bea54ef5SIsrael Rukshin * avoid races with teardown flow. 2064bea54ef5SIsrael Rukshin */ 2065ecca390eSSagi Grimberg WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING && 2066ecca390eSSagi Grimberg ctrl->state != NVME_CTRL_DELETING_NOIO); 2067bea54ef5SIsrael Rukshin WARN_ON_ONCE(new); 20683f2304f8SSagi Grimberg ret = -EINVAL; 20693f2304f8SSagi Grimberg goto destroy_io; 20703f2304f8SSagi Grimberg } 20713f2304f8SSagi Grimberg 20723f2304f8SSagi Grimberg nvme_start_ctrl(ctrl); 20733f2304f8SSagi Grimberg return 0; 20743f2304f8SSagi Grimberg 20753f2304f8SSagi Grimberg destroy_io: 207670a99574SChao Leng if (ctrl->queue_count > 1) { 20779f27bd70SChristoph Hellwig nvme_quiesce_io_queues(ctrl); 207870a99574SChao Leng nvme_sync_io_queues(ctrl); 207970a99574SChao Leng nvme_tcp_stop_io_queues(ctrl); 208070a99574SChao Leng nvme_cancel_tagset(ctrl); 20813f2304f8SSagi Grimberg nvme_tcp_destroy_io_queues(ctrl, new); 208270a99574SChao Leng } 20833f2304f8SSagi Grimberg destroy_admin: 20849f27bd70SChristoph Hellwig nvme_quiesce_admin_queue(ctrl); 208570a99574SChao Leng blk_sync_queue(ctrl->admin_q); 20863f2304f8SSagi Grimberg nvme_tcp_stop_queue(ctrl, 0); 208770a99574SChao Leng nvme_cancel_admin_tagset(ctrl); 20883f2304f8SSagi Grimberg nvme_tcp_destroy_admin_queue(ctrl, new); 20893f2304f8SSagi Grimberg return ret; 20903f2304f8SSagi Grimberg } 20913f2304f8SSagi Grimberg 20923f2304f8SSagi Grimberg static void nvme_tcp_reconnect_ctrl_work(struct work_struct *work) 20933f2304f8SSagi Grimberg { 20943f2304f8SSagi Grimberg struct nvme_tcp_ctrl *tcp_ctrl = container_of(to_delayed_work(work), 20953f2304f8SSagi Grimberg struct nvme_tcp_ctrl, connect_work); 20963f2304f8SSagi Grimberg struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl; 20973f2304f8SSagi Grimberg 20983f2304f8SSagi Grimberg ++ctrl->nr_reconnects; 20993f2304f8SSagi Grimberg 21003f2304f8SSagi Grimberg if (nvme_tcp_setup_ctrl(ctrl, false)) 21013f2304f8SSagi Grimberg goto requeue; 21023f2304f8SSagi Grimberg 210356a77d26SColin Ian King dev_info(ctrl->device, "Successfully reconnected (%d attempt)\n", 21043f2304f8SSagi Grimberg ctrl->nr_reconnects); 21053f2304f8SSagi Grimberg 21063f2304f8SSagi Grimberg ctrl->nr_reconnects = 0; 21073f2304f8SSagi Grimberg 21083f2304f8SSagi Grimberg return; 21093f2304f8SSagi Grimberg 21103f2304f8SSagi Grimberg requeue: 21113f2304f8SSagi Grimberg dev_info(ctrl->device, "Failed reconnect attempt %d\n", 21123f2304f8SSagi Grimberg ctrl->nr_reconnects); 21133f2304f8SSagi Grimberg nvme_tcp_reconnect_or_remove(ctrl); 21143f2304f8SSagi Grimberg } 21153f2304f8SSagi Grimberg 21163f2304f8SSagi Grimberg static void nvme_tcp_error_recovery_work(struct work_struct *work) 21173f2304f8SSagi Grimberg { 21183f2304f8SSagi Grimberg struct nvme_tcp_ctrl *tcp_ctrl = container_of(work, 21193f2304f8SSagi Grimberg struct nvme_tcp_ctrl, err_work); 21203f2304f8SSagi Grimberg struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl; 21213f2304f8SSagi Grimberg 21223f2304f8SSagi Grimberg nvme_stop_keep_alive(ctrl); 2123ff9fc7ebSSagi Grimberg flush_work(&ctrl->async_event_work); 21243f2304f8SSagi Grimberg nvme_tcp_teardown_io_queues(ctrl, false); 21253f2304f8SSagi Grimberg /* unquiesce to fail fast pending requests */ 21269f27bd70SChristoph Hellwig nvme_unquiesce_io_queues(ctrl); 21273f2304f8SSagi Grimberg nvme_tcp_teardown_admin_queue(ctrl, false); 21289f27bd70SChristoph Hellwig nvme_unquiesce_admin_queue(ctrl); 21291f1a4f89SSagi Grimberg nvme_auth_stop(ctrl); 21303f2304f8SSagi Grimberg 21313f2304f8SSagi Grimberg if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) { 2132ecca390eSSagi Grimberg /* state change failure is ok if we started ctrl delete */ 2133ecca390eSSagi Grimberg WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING && 2134ecca390eSSagi Grimberg ctrl->state != NVME_CTRL_DELETING_NOIO); 21353f2304f8SSagi Grimberg return; 21363f2304f8SSagi Grimberg } 21373f2304f8SSagi Grimberg 21383f2304f8SSagi Grimberg nvme_tcp_reconnect_or_remove(ctrl); 21393f2304f8SSagi Grimberg } 21403f2304f8SSagi Grimberg 21413f2304f8SSagi Grimberg static void nvme_tcp_teardown_ctrl(struct nvme_ctrl *ctrl, bool shutdown) 21423f2304f8SSagi Grimberg { 21433f2304f8SSagi Grimberg nvme_tcp_teardown_io_queues(ctrl, shutdown); 21449f27bd70SChristoph Hellwig nvme_quiesce_admin_queue(ctrl); 2145*285b6e9bSChristoph Hellwig nvme_disable_ctrl(ctrl, shutdown); 21463f2304f8SSagi Grimberg nvme_tcp_teardown_admin_queue(ctrl, shutdown); 21473f2304f8SSagi Grimberg } 21483f2304f8SSagi Grimberg 21493f2304f8SSagi Grimberg static void nvme_tcp_delete_ctrl(struct nvme_ctrl *ctrl) 21503f2304f8SSagi Grimberg { 21513f2304f8SSagi Grimberg nvme_tcp_teardown_ctrl(ctrl, true); 21523f2304f8SSagi Grimberg } 21533f2304f8SSagi Grimberg 21543f2304f8SSagi Grimberg static void nvme_reset_ctrl_work(struct work_struct *work) 21553f2304f8SSagi Grimberg { 21563f2304f8SSagi Grimberg struct nvme_ctrl *ctrl = 21573f2304f8SSagi Grimberg container_of(work, struct nvme_ctrl, reset_work); 21583f2304f8SSagi Grimberg 21593f2304f8SSagi Grimberg nvme_stop_ctrl(ctrl); 21603f2304f8SSagi Grimberg nvme_tcp_teardown_ctrl(ctrl, false); 21613f2304f8SSagi Grimberg 21623f2304f8SSagi Grimberg if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) { 2163ecca390eSSagi Grimberg /* state change failure is ok if we started ctrl delete */ 2164ecca390eSSagi Grimberg WARN_ON_ONCE(ctrl->state != NVME_CTRL_DELETING && 2165ecca390eSSagi Grimberg ctrl->state != NVME_CTRL_DELETING_NOIO); 21663f2304f8SSagi Grimberg return; 21673f2304f8SSagi Grimberg } 21683f2304f8SSagi Grimberg 21693f2304f8SSagi Grimberg if (nvme_tcp_setup_ctrl(ctrl, false)) 21703f2304f8SSagi Grimberg goto out_fail; 21713f2304f8SSagi Grimberg 21723f2304f8SSagi Grimberg return; 21733f2304f8SSagi Grimberg 21743f2304f8SSagi Grimberg out_fail: 21753f2304f8SSagi Grimberg ++ctrl->nr_reconnects; 21763f2304f8SSagi Grimberg nvme_tcp_reconnect_or_remove(ctrl); 21773f2304f8SSagi Grimberg } 21783f2304f8SSagi Grimberg 2179f7f70f4aSRuozhu Li static void nvme_tcp_stop_ctrl(struct nvme_ctrl *ctrl) 2180f7f70f4aSRuozhu Li { 2181c4abd875SSagi Grimberg flush_work(&to_tcp_ctrl(ctrl)->err_work); 2182f7f70f4aSRuozhu Li cancel_delayed_work_sync(&to_tcp_ctrl(ctrl)->connect_work); 2183f7f70f4aSRuozhu Li } 2184f7f70f4aSRuozhu Li 21853f2304f8SSagi Grimberg static void nvme_tcp_free_ctrl(struct nvme_ctrl *nctrl) 21863f2304f8SSagi Grimberg { 21873f2304f8SSagi Grimberg struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl); 21883f2304f8SSagi Grimberg 21893f2304f8SSagi Grimberg if (list_empty(&ctrl->list)) 21903f2304f8SSagi Grimberg goto free_ctrl; 21913f2304f8SSagi Grimberg 21923f2304f8SSagi Grimberg mutex_lock(&nvme_tcp_ctrl_mutex); 21933f2304f8SSagi Grimberg list_del(&ctrl->list); 21943f2304f8SSagi Grimberg mutex_unlock(&nvme_tcp_ctrl_mutex); 21953f2304f8SSagi Grimberg 21963f2304f8SSagi Grimberg nvmf_free_options(nctrl->opts); 21973f2304f8SSagi Grimberg free_ctrl: 21983f2304f8SSagi Grimberg kfree(ctrl->queues); 21993f2304f8SSagi Grimberg kfree(ctrl); 22003f2304f8SSagi Grimberg } 22013f2304f8SSagi Grimberg 22023f2304f8SSagi Grimberg static void nvme_tcp_set_sg_null(struct nvme_command *c) 22033f2304f8SSagi Grimberg { 22043f2304f8SSagi Grimberg struct nvme_sgl_desc *sg = &c->common.dptr.sgl; 22053f2304f8SSagi Grimberg 22063f2304f8SSagi Grimberg sg->addr = 0; 22073f2304f8SSagi Grimberg sg->length = 0; 22083f2304f8SSagi Grimberg sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) | 22093f2304f8SSagi Grimberg NVME_SGL_FMT_TRANSPORT_A; 22103f2304f8SSagi Grimberg } 22113f2304f8SSagi Grimberg 22123f2304f8SSagi Grimberg static void nvme_tcp_set_sg_inline(struct nvme_tcp_queue *queue, 22133f2304f8SSagi Grimberg struct nvme_command *c, u32 data_len) 22143f2304f8SSagi Grimberg { 22153f2304f8SSagi Grimberg struct nvme_sgl_desc *sg = &c->common.dptr.sgl; 22163f2304f8SSagi Grimberg 22173f2304f8SSagi Grimberg sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff); 22183f2304f8SSagi Grimberg sg->length = cpu_to_le32(data_len); 22193f2304f8SSagi Grimberg sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET; 22203f2304f8SSagi Grimberg } 22213f2304f8SSagi Grimberg 22223f2304f8SSagi Grimberg static void nvme_tcp_set_sg_host_data(struct nvme_command *c, 22233f2304f8SSagi Grimberg u32 data_len) 22243f2304f8SSagi Grimberg { 22253f2304f8SSagi Grimberg struct nvme_sgl_desc *sg = &c->common.dptr.sgl; 22263f2304f8SSagi Grimberg 22273f2304f8SSagi Grimberg sg->addr = 0; 22283f2304f8SSagi Grimberg sg->length = cpu_to_le32(data_len); 22293f2304f8SSagi Grimberg sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) | 22303f2304f8SSagi Grimberg NVME_SGL_FMT_TRANSPORT_A; 22313f2304f8SSagi Grimberg } 22323f2304f8SSagi Grimberg 22333f2304f8SSagi Grimberg static void nvme_tcp_submit_async_event(struct nvme_ctrl *arg) 22343f2304f8SSagi Grimberg { 22353f2304f8SSagi Grimberg struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(arg); 22363f2304f8SSagi Grimberg struct nvme_tcp_queue *queue = &ctrl->queues[0]; 22373f2304f8SSagi Grimberg struct nvme_tcp_cmd_pdu *pdu = ctrl->async_req.pdu; 22383f2304f8SSagi Grimberg struct nvme_command *cmd = &pdu->cmd; 22393f2304f8SSagi Grimberg u8 hdgst = nvme_tcp_hdgst_len(queue); 22403f2304f8SSagi Grimberg 22413f2304f8SSagi Grimberg memset(pdu, 0, sizeof(*pdu)); 22423f2304f8SSagi Grimberg pdu->hdr.type = nvme_tcp_cmd; 22433f2304f8SSagi Grimberg if (queue->hdr_digest) 22443f2304f8SSagi Grimberg pdu->hdr.flags |= NVME_TCP_F_HDGST; 22453f2304f8SSagi Grimberg pdu->hdr.hlen = sizeof(*pdu); 22463f2304f8SSagi Grimberg pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst); 22473f2304f8SSagi Grimberg 22483f2304f8SSagi Grimberg cmd->common.opcode = nvme_admin_async_event; 22493f2304f8SSagi Grimberg cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH; 22503f2304f8SSagi Grimberg cmd->common.flags |= NVME_CMD_SGL_METABUF; 22513f2304f8SSagi Grimberg nvme_tcp_set_sg_null(cmd); 22523f2304f8SSagi Grimberg 22533f2304f8SSagi Grimberg ctrl->async_req.state = NVME_TCP_SEND_CMD_PDU; 22543f2304f8SSagi Grimberg ctrl->async_req.offset = 0; 22553f2304f8SSagi Grimberg ctrl->async_req.curr_bio = NULL; 22563f2304f8SSagi Grimberg ctrl->async_req.data_len = 0; 22573f2304f8SSagi Grimberg 225886f0348aSSagi Grimberg nvme_tcp_queue_request(&ctrl->async_req, true, true); 22593f2304f8SSagi Grimberg } 22603f2304f8SSagi Grimberg 2261236187c4SSagi Grimberg static void nvme_tcp_complete_timed_out(struct request *rq) 2262236187c4SSagi Grimberg { 2263236187c4SSagi Grimberg struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq); 2264236187c4SSagi Grimberg struct nvme_ctrl *ctrl = &req->queue->ctrl->ctrl; 2265236187c4SSagi Grimberg 2266236187c4SSagi Grimberg nvme_tcp_stop_queue(ctrl, nvme_tcp_queue_id(req->queue)); 226793ba75c9SChaitanya Kulkarni nvmf_complete_timed_out_request(rq); 2268236187c4SSagi Grimberg } 2269236187c4SSagi Grimberg 22709bdb4833SJohn Garry static enum blk_eh_timer_return nvme_tcp_timeout(struct request *rq) 22713f2304f8SSagi Grimberg { 22723f2304f8SSagi Grimberg struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq); 2273236187c4SSagi Grimberg struct nvme_ctrl *ctrl = &req->queue->ctrl->ctrl; 22743f2304f8SSagi Grimberg struct nvme_tcp_cmd_pdu *pdu = req->pdu; 22753f2304f8SSagi Grimberg 2276236187c4SSagi Grimberg dev_warn(ctrl->device, 22773f2304f8SSagi Grimberg "queue %d: timeout request %#x type %d\n", 227839d57757SSagi Grimberg nvme_tcp_queue_id(req->queue), rq->tag, pdu->hdr.type); 22793f2304f8SSagi Grimberg 2280236187c4SSagi Grimberg if (ctrl->state != NVME_CTRL_LIVE) { 228139d57757SSagi Grimberg /* 2282236187c4SSagi Grimberg * If we are resetting, connecting or deleting we should 2283236187c4SSagi Grimberg * complete immediately because we may block controller 2284236187c4SSagi Grimberg * teardown or setup sequence 2285236187c4SSagi Grimberg * - ctrl disable/shutdown fabrics requests 2286236187c4SSagi Grimberg * - connect requests 2287236187c4SSagi Grimberg * - initialization admin requests 2288236187c4SSagi Grimberg * - I/O requests that entered after unquiescing and 2289236187c4SSagi Grimberg * the controller stopped responding 2290236187c4SSagi Grimberg * 2291236187c4SSagi Grimberg * All other requests should be cancelled by the error 2292236187c4SSagi Grimberg * recovery work, so it's fine that we fail it here. 229339d57757SSagi Grimberg */ 2294236187c4SSagi Grimberg nvme_tcp_complete_timed_out(rq); 22953f2304f8SSagi Grimberg return BLK_EH_DONE; 22963f2304f8SSagi Grimberg } 22973f2304f8SSagi Grimberg 2298236187c4SSagi Grimberg /* 2299236187c4SSagi Grimberg * LIVE state should trigger the normal error recovery which will 2300236187c4SSagi Grimberg * handle completing this request. 2301236187c4SSagi Grimberg */ 2302236187c4SSagi Grimberg nvme_tcp_error_recovery(ctrl); 23033f2304f8SSagi Grimberg return BLK_EH_RESET_TIMER; 23043f2304f8SSagi Grimberg } 23053f2304f8SSagi Grimberg 23063f2304f8SSagi Grimberg static blk_status_t nvme_tcp_map_data(struct nvme_tcp_queue *queue, 23073f2304f8SSagi Grimberg struct request *rq) 23083f2304f8SSagi Grimberg { 23093f2304f8SSagi Grimberg struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq); 23103f2304f8SSagi Grimberg struct nvme_tcp_cmd_pdu *pdu = req->pdu; 23113f2304f8SSagi Grimberg struct nvme_command *c = &pdu->cmd; 23123f2304f8SSagi Grimberg 23133f2304f8SSagi Grimberg c->common.flags |= NVME_CMD_SGL_METABUF; 23143f2304f8SSagi Grimberg 231525e5cb78SSagi Grimberg if (!blk_rq_nr_phys_segments(rq)) 231625e5cb78SSagi Grimberg nvme_tcp_set_sg_null(c); 231725e5cb78SSagi Grimberg else if (rq_data_dir(rq) == WRITE && 231853ee9e29SCaleb Sander req->data_len <= nvme_tcp_inline_data_size(req)) 23193f2304f8SSagi Grimberg nvme_tcp_set_sg_inline(queue, c, req->data_len); 23203f2304f8SSagi Grimberg else 23213f2304f8SSagi Grimberg nvme_tcp_set_sg_host_data(c, req->data_len); 23223f2304f8SSagi Grimberg 23233f2304f8SSagi Grimberg return 0; 23243f2304f8SSagi Grimberg } 23253f2304f8SSagi Grimberg 23263f2304f8SSagi Grimberg static blk_status_t nvme_tcp_setup_cmd_pdu(struct nvme_ns *ns, 23273f2304f8SSagi Grimberg struct request *rq) 23283f2304f8SSagi Grimberg { 23293f2304f8SSagi Grimberg struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq); 23303f2304f8SSagi Grimberg struct nvme_tcp_cmd_pdu *pdu = req->pdu; 23313f2304f8SSagi Grimberg struct nvme_tcp_queue *queue = req->queue; 23323f2304f8SSagi Grimberg u8 hdgst = nvme_tcp_hdgst_len(queue), ddgst = 0; 23333f2304f8SSagi Grimberg blk_status_t ret; 23343f2304f8SSagi Grimberg 2335f4b9e6c9SKeith Busch ret = nvme_setup_cmd(ns, rq); 23363f2304f8SSagi Grimberg if (ret) 23373f2304f8SSagi Grimberg return ret; 23383f2304f8SSagi Grimberg 23393f2304f8SSagi Grimberg req->state = NVME_TCP_SEND_CMD_PDU; 23401ba2e507SDaniel Wagner req->status = cpu_to_le16(NVME_SC_SUCCESS); 23413f2304f8SSagi Grimberg req->offset = 0; 23423f2304f8SSagi Grimberg req->data_sent = 0; 23433f2304f8SSagi Grimberg req->pdu_len = 0; 23443f2304f8SSagi Grimberg req->pdu_sent = 0; 2345c2700d28SVarun Prakash req->h2cdata_left = 0; 234625e5cb78SSagi Grimberg req->data_len = blk_rq_nr_phys_segments(rq) ? 234725e5cb78SSagi Grimberg blk_rq_payload_bytes(rq) : 0; 23483f2304f8SSagi Grimberg req->curr_bio = rq->bio; 2349e11e5116SSagi Grimberg if (req->curr_bio && req->data_len) 2350cb9b870fSSagi Grimberg nvme_tcp_init_iter(req, rq_data_dir(rq)); 23513f2304f8SSagi Grimberg 23523f2304f8SSagi Grimberg if (rq_data_dir(rq) == WRITE && 235353ee9e29SCaleb Sander req->data_len <= nvme_tcp_inline_data_size(req)) 23543f2304f8SSagi Grimberg req->pdu_len = req->data_len; 23553f2304f8SSagi Grimberg 23563f2304f8SSagi Grimberg pdu->hdr.type = nvme_tcp_cmd; 23573f2304f8SSagi Grimberg pdu->hdr.flags = 0; 23583f2304f8SSagi Grimberg if (queue->hdr_digest) 23593f2304f8SSagi Grimberg pdu->hdr.flags |= NVME_TCP_F_HDGST; 23603f2304f8SSagi Grimberg if (queue->data_digest && req->pdu_len) { 23613f2304f8SSagi Grimberg pdu->hdr.flags |= NVME_TCP_F_DDGST; 23623f2304f8SSagi Grimberg ddgst = nvme_tcp_ddgst_len(queue); 23633f2304f8SSagi Grimberg } 23643f2304f8SSagi Grimberg pdu->hdr.hlen = sizeof(*pdu); 23653f2304f8SSagi Grimberg pdu->hdr.pdo = req->pdu_len ? pdu->hdr.hlen + hdgst : 0; 23663f2304f8SSagi Grimberg pdu->hdr.plen = 23673f2304f8SSagi Grimberg cpu_to_le32(pdu->hdr.hlen + hdgst + req->pdu_len + ddgst); 23683f2304f8SSagi Grimberg 23693f2304f8SSagi Grimberg ret = nvme_tcp_map_data(queue, rq); 23703f2304f8SSagi Grimberg if (unlikely(ret)) { 237128a4cac4SMax Gurtovoy nvme_cleanup_cmd(rq); 23723f2304f8SSagi Grimberg dev_err(queue->ctrl->ctrl.device, 23733f2304f8SSagi Grimberg "Failed to map data (%d)\n", ret); 23743f2304f8SSagi Grimberg return ret; 23753f2304f8SSagi Grimberg } 23763f2304f8SSagi Grimberg 23773f2304f8SSagi Grimberg return 0; 23783f2304f8SSagi Grimberg } 23793f2304f8SSagi Grimberg 238086f0348aSSagi Grimberg static void nvme_tcp_commit_rqs(struct blk_mq_hw_ctx *hctx) 238186f0348aSSagi Grimberg { 238286f0348aSSagi Grimberg struct nvme_tcp_queue *queue = hctx->driver_data; 238386f0348aSSagi Grimberg 238486f0348aSSagi Grimberg if (!llist_empty(&queue->req_list)) 238586f0348aSSagi Grimberg queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work); 238686f0348aSSagi Grimberg } 238786f0348aSSagi Grimberg 23883f2304f8SSagi Grimberg static blk_status_t nvme_tcp_queue_rq(struct blk_mq_hw_ctx *hctx, 23893f2304f8SSagi Grimberg const struct blk_mq_queue_data *bd) 23903f2304f8SSagi Grimberg { 23913f2304f8SSagi Grimberg struct nvme_ns *ns = hctx->queue->queuedata; 23923f2304f8SSagi Grimberg struct nvme_tcp_queue *queue = hctx->driver_data; 23933f2304f8SSagi Grimberg struct request *rq = bd->rq; 23943f2304f8SSagi Grimberg struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq); 23953f2304f8SSagi Grimberg bool queue_ready = test_bit(NVME_TCP_Q_LIVE, &queue->flags); 23963f2304f8SSagi Grimberg blk_status_t ret; 23973f2304f8SSagi Grimberg 2398a9715744STao Chiu if (!nvme_check_ready(&queue->ctrl->ctrl, rq, queue_ready)) 2399a9715744STao Chiu return nvme_fail_nonready_command(&queue->ctrl->ctrl, rq); 24003f2304f8SSagi Grimberg 24013f2304f8SSagi Grimberg ret = nvme_tcp_setup_cmd_pdu(ns, rq); 24023f2304f8SSagi Grimberg if (unlikely(ret)) 24033f2304f8SSagi Grimberg return ret; 24043f2304f8SSagi Grimberg 24056887fc64SSagi Grimberg nvme_start_request(rq); 24063f2304f8SSagi Grimberg 240786f0348aSSagi Grimberg nvme_tcp_queue_request(req, true, bd->last); 24083f2304f8SSagi Grimberg 24093f2304f8SSagi Grimberg return BLK_STS_OK; 24103f2304f8SSagi Grimberg } 24113f2304f8SSagi Grimberg 2412a4e1d0b7SBart Van Assche static void nvme_tcp_map_queues(struct blk_mq_tag_set *set) 2413873946f4SSagi Grimberg { 241406427ca0SChristoph Hellwig struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(set->driver_data); 241564861993SSagi Grimberg struct nvmf_ctrl_options *opts = ctrl->ctrl.opts; 2416873946f4SSagi Grimberg 241764861993SSagi Grimberg if (opts->nr_write_queues && ctrl->io_queues[HCTX_TYPE_READ]) { 2418873946f4SSagi Grimberg /* separate read/write queues */ 2419873946f4SSagi Grimberg set->map[HCTX_TYPE_DEFAULT].nr_queues = 242064861993SSagi Grimberg ctrl->io_queues[HCTX_TYPE_DEFAULT]; 242164861993SSagi Grimberg set->map[HCTX_TYPE_DEFAULT].queue_offset = 0; 242264861993SSagi Grimberg set->map[HCTX_TYPE_READ].nr_queues = 242364861993SSagi Grimberg ctrl->io_queues[HCTX_TYPE_READ]; 2424873946f4SSagi Grimberg set->map[HCTX_TYPE_READ].queue_offset = 242564861993SSagi Grimberg ctrl->io_queues[HCTX_TYPE_DEFAULT]; 2426873946f4SSagi Grimberg } else { 242764861993SSagi Grimberg /* shared read/write queues */ 2428873946f4SSagi Grimberg set->map[HCTX_TYPE_DEFAULT].nr_queues = 242964861993SSagi Grimberg ctrl->io_queues[HCTX_TYPE_DEFAULT]; 243064861993SSagi Grimberg set->map[HCTX_TYPE_DEFAULT].queue_offset = 0; 243164861993SSagi Grimberg set->map[HCTX_TYPE_READ].nr_queues = 243264861993SSagi Grimberg ctrl->io_queues[HCTX_TYPE_DEFAULT]; 2433873946f4SSagi Grimberg set->map[HCTX_TYPE_READ].queue_offset = 0; 2434873946f4SSagi Grimberg } 2435873946f4SSagi Grimberg blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); 2436873946f4SSagi Grimberg blk_mq_map_queues(&set->map[HCTX_TYPE_READ]); 243764861993SSagi Grimberg 24381a9460ceSSagi Grimberg if (opts->nr_poll_queues && ctrl->io_queues[HCTX_TYPE_POLL]) { 24391a9460ceSSagi Grimberg /* map dedicated poll queues only if we have queues left */ 24401a9460ceSSagi Grimberg set->map[HCTX_TYPE_POLL].nr_queues = 24411a9460ceSSagi Grimberg ctrl->io_queues[HCTX_TYPE_POLL]; 24421a9460ceSSagi Grimberg set->map[HCTX_TYPE_POLL].queue_offset = 24431a9460ceSSagi Grimberg ctrl->io_queues[HCTX_TYPE_DEFAULT] + 24441a9460ceSSagi Grimberg ctrl->io_queues[HCTX_TYPE_READ]; 24451a9460ceSSagi Grimberg blk_mq_map_queues(&set->map[HCTX_TYPE_POLL]); 24461a9460ceSSagi Grimberg } 24471a9460ceSSagi Grimberg 244864861993SSagi Grimberg dev_info(ctrl->ctrl.device, 24491a9460ceSSagi Grimberg "mapped %d/%d/%d default/read/poll queues.\n", 245064861993SSagi Grimberg ctrl->io_queues[HCTX_TYPE_DEFAULT], 24511a9460ceSSagi Grimberg ctrl->io_queues[HCTX_TYPE_READ], 24521a9460ceSSagi Grimberg ctrl->io_queues[HCTX_TYPE_POLL]); 2453873946f4SSagi Grimberg } 2454873946f4SSagi Grimberg 24555a72e899SJens Axboe static int nvme_tcp_poll(struct blk_mq_hw_ctx *hctx, struct io_comp_batch *iob) 24561a9460ceSSagi Grimberg { 24571a9460ceSSagi Grimberg struct nvme_tcp_queue *queue = hctx->driver_data; 24581a9460ceSSagi Grimberg struct sock *sk = queue->sock->sk; 24591a9460ceSSagi Grimberg 2460f86e5bf8SSagi Grimberg if (!test_bit(NVME_TCP_Q_LIVE, &queue->flags)) 2461f86e5bf8SSagi Grimberg return 0; 2462f86e5bf8SSagi Grimberg 246372e5d757SSagi Grimberg set_bit(NVME_TCP_Q_POLLING, &queue->flags); 24643f926af3SEric Dumazet if (sk_can_busy_loop(sk) && skb_queue_empty_lockless(&sk->sk_receive_queue)) 24651a9460ceSSagi Grimberg sk_busy_loop(sk, true); 24661a9460ceSSagi Grimberg nvme_tcp_try_recv(queue); 246772e5d757SSagi Grimberg clear_bit(NVME_TCP_Q_POLLING, &queue->flags); 24681a9460ceSSagi Grimberg return queue->nr_cqe; 24691a9460ceSSagi Grimberg } 24701a9460ceSSagi Grimberg 247102c57a82SMartin Belanger static int nvme_tcp_get_address(struct nvme_ctrl *ctrl, char *buf, int size) 247202c57a82SMartin Belanger { 247302c57a82SMartin Belanger struct nvme_tcp_queue *queue = &to_tcp_ctrl(ctrl)->queues[0]; 247402c57a82SMartin Belanger struct sockaddr_storage src_addr; 247502c57a82SMartin Belanger int ret, len; 247602c57a82SMartin Belanger 247702c57a82SMartin Belanger len = nvmf_get_address(ctrl, buf, size); 247802c57a82SMartin Belanger 247902c57a82SMartin Belanger ret = kernel_getsockname(queue->sock, (struct sockaddr *)&src_addr); 248002c57a82SMartin Belanger if (ret > 0) { 248102c57a82SMartin Belanger if (len > 0) 248202c57a82SMartin Belanger len--; /* strip trailing newline */ 248302c57a82SMartin Belanger len += scnprintf(buf + len, size - len, "%ssrc_addr=%pISc\n", 248402c57a82SMartin Belanger (len) ? "," : "", &src_addr); 248502c57a82SMartin Belanger } 248602c57a82SMartin Belanger 248702c57a82SMartin Belanger return len; 248802c57a82SMartin Belanger } 248902c57a82SMartin Belanger 24906acbd961SRikard Falkeborn static const struct blk_mq_ops nvme_tcp_mq_ops = { 24913f2304f8SSagi Grimberg .queue_rq = nvme_tcp_queue_rq, 249286f0348aSSagi Grimberg .commit_rqs = nvme_tcp_commit_rqs, 24933f2304f8SSagi Grimberg .complete = nvme_complete_rq, 24943f2304f8SSagi Grimberg .init_request = nvme_tcp_init_request, 24953f2304f8SSagi Grimberg .exit_request = nvme_tcp_exit_request, 24963f2304f8SSagi Grimberg .init_hctx = nvme_tcp_init_hctx, 24973f2304f8SSagi Grimberg .timeout = nvme_tcp_timeout, 2498873946f4SSagi Grimberg .map_queues = nvme_tcp_map_queues, 24991a9460ceSSagi Grimberg .poll = nvme_tcp_poll, 25003f2304f8SSagi Grimberg }; 25013f2304f8SSagi Grimberg 25026acbd961SRikard Falkeborn static const struct blk_mq_ops nvme_tcp_admin_mq_ops = { 25033f2304f8SSagi Grimberg .queue_rq = nvme_tcp_queue_rq, 25043f2304f8SSagi Grimberg .complete = nvme_complete_rq, 25053f2304f8SSagi Grimberg .init_request = nvme_tcp_init_request, 25063f2304f8SSagi Grimberg .exit_request = nvme_tcp_exit_request, 25073f2304f8SSagi Grimberg .init_hctx = nvme_tcp_init_admin_hctx, 25083f2304f8SSagi Grimberg .timeout = nvme_tcp_timeout, 25093f2304f8SSagi Grimberg }; 25103f2304f8SSagi Grimberg 25113f2304f8SSagi Grimberg static const struct nvme_ctrl_ops nvme_tcp_ctrl_ops = { 25123f2304f8SSagi Grimberg .name = "tcp", 25133f2304f8SSagi Grimberg .module = THIS_MODULE, 25143f2304f8SSagi Grimberg .flags = NVME_F_FABRICS, 25153f2304f8SSagi Grimberg .reg_read32 = nvmf_reg_read32, 25163f2304f8SSagi Grimberg .reg_read64 = nvmf_reg_read64, 25173f2304f8SSagi Grimberg .reg_write32 = nvmf_reg_write32, 25183f2304f8SSagi Grimberg .free_ctrl = nvme_tcp_free_ctrl, 25193f2304f8SSagi Grimberg .submit_async_event = nvme_tcp_submit_async_event, 25203f2304f8SSagi Grimberg .delete_ctrl = nvme_tcp_delete_ctrl, 252102c57a82SMartin Belanger .get_address = nvme_tcp_get_address, 2522f7f70f4aSRuozhu Li .stop_ctrl = nvme_tcp_stop_ctrl, 25233f2304f8SSagi Grimberg }; 25243f2304f8SSagi Grimberg 25253f2304f8SSagi Grimberg static bool 25263f2304f8SSagi Grimberg nvme_tcp_existing_controller(struct nvmf_ctrl_options *opts) 25273f2304f8SSagi Grimberg { 25283f2304f8SSagi Grimberg struct nvme_tcp_ctrl *ctrl; 25293f2304f8SSagi Grimberg bool found = false; 25303f2304f8SSagi Grimberg 25313f2304f8SSagi Grimberg mutex_lock(&nvme_tcp_ctrl_mutex); 25323f2304f8SSagi Grimberg list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list) { 25333f2304f8SSagi Grimberg found = nvmf_ip_options_match(&ctrl->ctrl, opts); 25343f2304f8SSagi Grimberg if (found) 25353f2304f8SSagi Grimberg break; 25363f2304f8SSagi Grimberg } 25373f2304f8SSagi Grimberg mutex_unlock(&nvme_tcp_ctrl_mutex); 25383f2304f8SSagi Grimberg 25393f2304f8SSagi Grimberg return found; 25403f2304f8SSagi Grimberg } 25413f2304f8SSagi Grimberg 25423f2304f8SSagi Grimberg static struct nvme_ctrl *nvme_tcp_create_ctrl(struct device *dev, 25433f2304f8SSagi Grimberg struct nvmf_ctrl_options *opts) 25443f2304f8SSagi Grimberg { 25453f2304f8SSagi Grimberg struct nvme_tcp_ctrl *ctrl; 25463f2304f8SSagi Grimberg int ret; 25473f2304f8SSagi Grimberg 25483f2304f8SSagi Grimberg ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL); 25493f2304f8SSagi Grimberg if (!ctrl) 25503f2304f8SSagi Grimberg return ERR_PTR(-ENOMEM); 25513f2304f8SSagi Grimberg 25523f2304f8SSagi Grimberg INIT_LIST_HEAD(&ctrl->list); 25533f2304f8SSagi Grimberg ctrl->ctrl.opts = opts; 25541a9460ceSSagi Grimberg ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues + 25551a9460ceSSagi Grimberg opts->nr_poll_queues + 1; 25563f2304f8SSagi Grimberg ctrl->ctrl.sqsize = opts->queue_size - 1; 25573f2304f8SSagi Grimberg ctrl->ctrl.kato = opts->kato; 25583f2304f8SSagi Grimberg 25593f2304f8SSagi Grimberg INIT_DELAYED_WORK(&ctrl->connect_work, 25603f2304f8SSagi Grimberg nvme_tcp_reconnect_ctrl_work); 25613f2304f8SSagi Grimberg INIT_WORK(&ctrl->err_work, nvme_tcp_error_recovery_work); 25623f2304f8SSagi Grimberg INIT_WORK(&ctrl->ctrl.reset_work, nvme_reset_ctrl_work); 25633f2304f8SSagi Grimberg 25643f2304f8SSagi Grimberg if (!(opts->mask & NVMF_OPT_TRSVCID)) { 25653f2304f8SSagi Grimberg opts->trsvcid = 25663f2304f8SSagi Grimberg kstrdup(__stringify(NVME_TCP_DISC_PORT), GFP_KERNEL); 25673f2304f8SSagi Grimberg if (!opts->trsvcid) { 25683f2304f8SSagi Grimberg ret = -ENOMEM; 25693f2304f8SSagi Grimberg goto out_free_ctrl; 25703f2304f8SSagi Grimberg } 25713f2304f8SSagi Grimberg opts->mask |= NVMF_OPT_TRSVCID; 25723f2304f8SSagi Grimberg } 25733f2304f8SSagi Grimberg 25743f2304f8SSagi Grimberg ret = inet_pton_with_scope(&init_net, AF_UNSPEC, 25753f2304f8SSagi Grimberg opts->traddr, opts->trsvcid, &ctrl->addr); 25763f2304f8SSagi Grimberg if (ret) { 25773f2304f8SSagi Grimberg pr_err("malformed address passed: %s:%s\n", 25783f2304f8SSagi Grimberg opts->traddr, opts->trsvcid); 25793f2304f8SSagi Grimberg goto out_free_ctrl; 25803f2304f8SSagi Grimberg } 25813f2304f8SSagi Grimberg 25823f2304f8SSagi Grimberg if (opts->mask & NVMF_OPT_HOST_TRADDR) { 25833f2304f8SSagi Grimberg ret = inet_pton_with_scope(&init_net, AF_UNSPEC, 25843f2304f8SSagi Grimberg opts->host_traddr, NULL, &ctrl->src_addr); 25853f2304f8SSagi Grimberg if (ret) { 25863f2304f8SSagi Grimberg pr_err("malformed src address passed: %s\n", 25873f2304f8SSagi Grimberg opts->host_traddr); 25883f2304f8SSagi Grimberg goto out_free_ctrl; 25893f2304f8SSagi Grimberg } 25903f2304f8SSagi Grimberg } 25913f2304f8SSagi Grimberg 25923ede8f72SMartin Belanger if (opts->mask & NVMF_OPT_HOST_IFACE) { 25938b43ced6SPrabhakar Kushwaha if (!__dev_get_by_name(&init_net, opts->host_iface)) { 25943ede8f72SMartin Belanger pr_err("invalid interface passed: %s\n", 25953ede8f72SMartin Belanger opts->host_iface); 25963ede8f72SMartin Belanger ret = -ENODEV; 25973ede8f72SMartin Belanger goto out_free_ctrl; 25983ede8f72SMartin Belanger } 25993ede8f72SMartin Belanger } 26003ede8f72SMartin Belanger 26013f2304f8SSagi Grimberg if (!opts->duplicate_connect && nvme_tcp_existing_controller(opts)) { 26023f2304f8SSagi Grimberg ret = -EALREADY; 26033f2304f8SSagi Grimberg goto out_free_ctrl; 26043f2304f8SSagi Grimberg } 26053f2304f8SSagi Grimberg 2606873946f4SSagi Grimberg ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues), 26073f2304f8SSagi Grimberg GFP_KERNEL); 26083f2304f8SSagi Grimberg if (!ctrl->queues) { 26093f2304f8SSagi Grimberg ret = -ENOMEM; 26103f2304f8SSagi Grimberg goto out_free_ctrl; 26113f2304f8SSagi Grimberg } 26123f2304f8SSagi Grimberg 26133f2304f8SSagi Grimberg ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_tcp_ctrl_ops, 0); 26143f2304f8SSagi Grimberg if (ret) 26153f2304f8SSagi Grimberg goto out_kfree_queues; 26163f2304f8SSagi Grimberg 26173f2304f8SSagi Grimberg if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) { 26183f2304f8SSagi Grimberg WARN_ON_ONCE(1); 26193f2304f8SSagi Grimberg ret = -EINTR; 26203f2304f8SSagi Grimberg goto out_uninit_ctrl; 26213f2304f8SSagi Grimberg } 26223f2304f8SSagi Grimberg 26233f2304f8SSagi Grimberg ret = nvme_tcp_setup_ctrl(&ctrl->ctrl, true); 26243f2304f8SSagi Grimberg if (ret) 26253f2304f8SSagi Grimberg goto out_uninit_ctrl; 26263f2304f8SSagi Grimberg 26273f2304f8SSagi Grimberg dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISp\n", 2628e5ea42faSHannes Reinecke nvmf_ctrl_subsysnqn(&ctrl->ctrl), &ctrl->addr); 26293f2304f8SSagi Grimberg 26303f2304f8SSagi Grimberg mutex_lock(&nvme_tcp_ctrl_mutex); 26313f2304f8SSagi Grimberg list_add_tail(&ctrl->list, &nvme_tcp_ctrl_list); 26323f2304f8SSagi Grimberg mutex_unlock(&nvme_tcp_ctrl_mutex); 26333f2304f8SSagi Grimberg 26343f2304f8SSagi Grimberg return &ctrl->ctrl; 26353f2304f8SSagi Grimberg 26363f2304f8SSagi Grimberg out_uninit_ctrl: 26373f2304f8SSagi Grimberg nvme_uninit_ctrl(&ctrl->ctrl); 26383f2304f8SSagi Grimberg nvme_put_ctrl(&ctrl->ctrl); 26393f2304f8SSagi Grimberg if (ret > 0) 26403f2304f8SSagi Grimberg ret = -EIO; 26413f2304f8SSagi Grimberg return ERR_PTR(ret); 26423f2304f8SSagi Grimberg out_kfree_queues: 26433f2304f8SSagi Grimberg kfree(ctrl->queues); 26443f2304f8SSagi Grimberg out_free_ctrl: 26453f2304f8SSagi Grimberg kfree(ctrl); 26463f2304f8SSagi Grimberg return ERR_PTR(ret); 26473f2304f8SSagi Grimberg } 26483f2304f8SSagi Grimberg 26493f2304f8SSagi Grimberg static struct nvmf_transport_ops nvme_tcp_transport = { 26503f2304f8SSagi Grimberg .name = "tcp", 26513f2304f8SSagi Grimberg .module = THIS_MODULE, 26523f2304f8SSagi Grimberg .required_opts = NVMF_OPT_TRADDR, 26533f2304f8SSagi Grimberg .allowed_opts = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY | 26543f2304f8SSagi Grimberg NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO | 2655873946f4SSagi Grimberg NVMF_OPT_HDR_DIGEST | NVMF_OPT_DATA_DIGEST | 2656bb13985dSIsrael Rukshin NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES | 26573ede8f72SMartin Belanger NVMF_OPT_TOS | NVMF_OPT_HOST_IFACE, 26583f2304f8SSagi Grimberg .create_ctrl = nvme_tcp_create_ctrl, 26593f2304f8SSagi Grimberg }; 26603f2304f8SSagi Grimberg 26613f2304f8SSagi Grimberg static int __init nvme_tcp_init_module(void) 26623f2304f8SSagi Grimberg { 26633f2304f8SSagi Grimberg nvme_tcp_wq = alloc_workqueue("nvme_tcp_wq", 26643f2304f8SSagi Grimberg WQ_MEM_RECLAIM | WQ_HIGHPRI, 0); 26653f2304f8SSagi Grimberg if (!nvme_tcp_wq) 26663f2304f8SSagi Grimberg return -ENOMEM; 26673f2304f8SSagi Grimberg 26683f2304f8SSagi Grimberg nvmf_register_transport(&nvme_tcp_transport); 26693f2304f8SSagi Grimberg return 0; 26703f2304f8SSagi Grimberg } 26713f2304f8SSagi Grimberg 26723f2304f8SSagi Grimberg static void __exit nvme_tcp_cleanup_module(void) 26733f2304f8SSagi Grimberg { 26743f2304f8SSagi Grimberg struct nvme_tcp_ctrl *ctrl; 26753f2304f8SSagi Grimberg 26763f2304f8SSagi Grimberg nvmf_unregister_transport(&nvme_tcp_transport); 26773f2304f8SSagi Grimberg 26783f2304f8SSagi Grimberg mutex_lock(&nvme_tcp_ctrl_mutex); 26793f2304f8SSagi Grimberg list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list) 26803f2304f8SSagi Grimberg nvme_delete_ctrl(&ctrl->ctrl); 26813f2304f8SSagi Grimberg mutex_unlock(&nvme_tcp_ctrl_mutex); 26823f2304f8SSagi Grimberg flush_workqueue(nvme_delete_wq); 26833f2304f8SSagi Grimberg 26843f2304f8SSagi Grimberg destroy_workqueue(nvme_tcp_wq); 26853f2304f8SSagi Grimberg } 26863f2304f8SSagi Grimberg 26873f2304f8SSagi Grimberg module_init(nvme_tcp_init_module); 26883f2304f8SSagi Grimberg module_exit(nvme_tcp_cleanup_module); 26893f2304f8SSagi Grimberg 26903f2304f8SSagi Grimberg MODULE_LICENSE("GPL v2"); 2691