tcp.c

   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * NVMe over Fabrics TCP host.
   4 * Copyright (c) 2018 Lightbits Labs. All rights reserved.
   5 */
   6#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   7#include <linux/module.h>
   8#include <linux/init.h>
   9#include <linux/slab.h>
  10#include <linux/err.h>
  11#include <linux/crc32.h>
  12#include <linux/nvme-tcp.h>
  13#include <linux/nvme-keyring.h>
  14#include <net/sock.h>
  15#include <net/tcp.h>
  16#include <net/tls.h>
  17#include <net/tls_prot.h>
  18#include <net/handshake.h>
  19#include <linux/blk-mq.h>
  20#include <net/busy_poll.h>
  21#include <trace/events/sock.h>
  22
  23#include "nvme.h"
  24#include "fabrics.h"
  25
  26struct nvme_tcp_queue;
  27
  28/* Define the socket priority to use for connections were it is desirable
  29 * that the NIC consider performing optimized packet processing or filtering.
  30 * A non-zero value being sufficient to indicate general consideration of any
  31 * possible optimization.  Making it a module param allows for alternative
  32 * values that may be unique for some NIC implementations.
  33 */
  34static int so_priority;
  35module_param(so_priority, int, 0644);
  36MODULE_PARM_DESC(so_priority, "nvme tcp socket optimize priority");
  37
  38/*
  39 * Use the unbound workqueue for nvme_tcp_wq, then we can set the cpu affinity
  40 * from sysfs.
  41 */
  42static bool wq_unbound;
  43module_param(wq_unbound, bool, 0644);
  44MODULE_PARM_DESC(wq_unbound, "Use unbound workqueue for nvme-tcp IO context (default false)");
  45
  46/*
  47 * TLS handshake timeout
  48 */
  49static int tls_handshake_timeout = 10;
  50#ifdef CONFIG_NVME_TCP_TLS
  51module_param(tls_handshake_timeout, int, 0644);
  52MODULE_PARM_DESC(tls_handshake_timeout,
  53		 "nvme TLS handshake timeout in seconds (default 10)");
  54#endif
  55
  56static atomic_t nvme_tcp_cpu_queues[NR_CPUS];
  57
  58#ifdef CONFIG_DEBUG_LOCK_ALLOC
  59/* lockdep can detect a circular dependency of the form
  60 *   sk_lock -> mmap_lock (page fault) -> fs locks -> sk_lock
  61 * because dependencies are tracked for both nvme-tcp and user contexts. Using
  62 * a separate class prevents lockdep from conflating nvme-tcp socket use with
  63 * user-space socket API use.
  64 */
  65static struct lock_class_key nvme_tcp_sk_key[2];
  66static struct lock_class_key nvme_tcp_slock_key[2];
  67
  68static void nvme_tcp_reclassify_socket(struct socket *sock)
  69{
  70	struct sock *sk = sock->sk;
  71
  72	if (WARN_ON_ONCE(!sock_allow_reclassification(sk)))
  73		return;
  74
  75	switch (sk->sk_family) {
  76	case AF_INET:
  77		sock_lock_init_class_and_name(sk, "slock-AF_INET-NVME",
  78					      &nvme_tcp_slock_key[0],
  79					      "sk_lock-AF_INET-NVME",
  80					      &nvme_tcp_sk_key[0]);
  81		break;
  82	case AF_INET6:
  83		sock_lock_init_class_and_name(sk, "slock-AF_INET6-NVME",
  84					      &nvme_tcp_slock_key[1],
  85					      "sk_lock-AF_INET6-NVME",
  86					      &nvme_tcp_sk_key[1]);
  87		break;
  88	default:
  89		WARN_ON_ONCE(1);
  90	}
  91}
  92#else
  93static void nvme_tcp_reclassify_socket(struct socket *sock) { }
  94#endif
  95
  96enum nvme_tcp_send_state {
  97	NVME_TCP_SEND_CMD_PDU = 0,
  98	NVME_TCP_SEND_H2C_PDU,
  99	NVME_TCP_SEND_DATA,
 100	NVME_TCP_SEND_DDGST,
 101};
 102
 103struct nvme_tcp_request {
 104	struct nvme_request	req;
 105	void			*pdu;
 106	struct nvme_tcp_queue	*queue;
 107	u32			data_len;
 108	u32			pdu_len;
 109	u32			pdu_sent;
 110	u32			h2cdata_left;
 111	u32			h2cdata_offset;
 112	u16			ttag;
 113	__le16			status;
 114	struct list_head	entry;
 115	struct llist_node	lentry;
 116	__le32			ddgst;
 117
 118	struct bio		*curr_bio;
 119	struct iov_iter		iter;
 120
 121	/* send state */
 122	size_t			offset;
 123	size_t			data_sent;
 124	enum nvme_tcp_send_state state;
 125};
 126
 127enum nvme_tcp_queue_flags {
 128	NVME_TCP_Q_ALLOCATED	= 0,
 129	NVME_TCP_Q_LIVE		= 1,
 130	NVME_TCP_Q_POLLING	= 2,
 131	NVME_TCP_Q_IO_CPU_SET	= 3,
 132};
 133
 134enum nvme_tcp_recv_state {
 135	NVME_TCP_RECV_PDU = 0,
 136	NVME_TCP_RECV_DATA,
 137	NVME_TCP_RECV_DDGST,
 138};
 139
 140struct nvme_tcp_ctrl;
 141struct nvme_tcp_queue {
 142	struct socket		*sock;
 143	struct work_struct	io_work;
 144	int			io_cpu;
 145
 146	struct mutex		queue_lock;
 147	struct mutex		send_mutex;
 148	struct llist_head	req_list;
 149	struct list_head	send_list;
 150
 151	/* recv state */
 152	void			*pdu;
 153	int			pdu_remaining;
 154	int			pdu_offset;
 155	size_t			data_remaining;
 156	size_t			ddgst_remaining;
 157	unsigned int		nr_cqe;
 158
 159	/* send state */
 160	struct nvme_tcp_request *request;
 161
 162	u32			maxh2cdata;
 163	size_t			cmnd_capsule_len;
 164	struct nvme_tcp_ctrl	*ctrl;
 165	unsigned long		flags;
 166	bool			rd_enabled;
 167
 168	bool			hdr_digest;
 169	bool			data_digest;
 170	bool			tls_enabled;
 171	u32			rcv_crc;
 172	u32			snd_crc;
 173	__le32			exp_ddgst;
 174	__le32			recv_ddgst;
 175	struct completion       tls_complete;
 176	int                     tls_err;
 177	struct page_frag_cache	pf_cache;
 178
 179	void (*state_change)(struct sock *);
 180	void (*data_ready)(struct sock *);
 181	void (*write_space)(struct sock *);
 182};
 183
 184struct nvme_tcp_ctrl {
 185	/* read only in the hot path */
 186	struct nvme_tcp_queue	*queues;
 187	struct blk_mq_tag_set	tag_set;
 188
 189	/* other member variables */
 190	struct list_head	list;
 191	struct blk_mq_tag_set	admin_tag_set;
 192	struct sockaddr_storage addr;
 193	struct sockaddr_storage src_addr;
 194	struct nvme_ctrl	ctrl;
 195
 196	struct work_struct	err_work;
 197	struct delayed_work	connect_work;
 198	struct nvme_tcp_request async_req;
 199	u32			io_queues[HCTX_MAX_TYPES];
 200};
 201
 202static LIST_HEAD(nvme_tcp_ctrl_list);
 203static DEFINE_MUTEX(nvme_tcp_ctrl_mutex);
 204static struct workqueue_struct *nvme_tcp_wq;
 205static const struct blk_mq_ops nvme_tcp_mq_ops;
 206static const struct blk_mq_ops nvme_tcp_admin_mq_ops;
 207static int nvme_tcp_try_send(struct nvme_tcp_queue *queue);
 208
 209static inline struct nvme_tcp_ctrl *to_tcp_ctrl(struct nvme_ctrl *ctrl)
 210{
 211	return container_of(ctrl, struct nvme_tcp_ctrl, ctrl);
 212}
 213
 214static inline int nvme_tcp_queue_id(struct nvme_tcp_queue *queue)
 215{
 216	return queue - queue->ctrl->queues;
 217}
 218
 219static inline bool nvme_tcp_recv_pdu_supported(enum nvme_tcp_pdu_type type)
 220{
 221	switch (type) {
 222	case nvme_tcp_c2h_term:
 223	case nvme_tcp_c2h_data:
 224	case nvme_tcp_r2t:
 225	case nvme_tcp_rsp:
 226		return true;
 227	default:
 228		return false;
 229	}
 230}
 231
 232/*
 233 * Check if the queue is TLS encrypted
 234 */
 235static inline bool nvme_tcp_queue_tls(struct nvme_tcp_queue *queue)
 236{
 237	if (!IS_ENABLED(CONFIG_NVME_TCP_TLS))
 238		return 0;
 239
 240	return queue->tls_enabled;
 241}
 242
 243/*
 244 * Check if TLS is configured for the controller.
 245 */
 246static inline bool nvme_tcp_tls_configured(struct nvme_ctrl *ctrl)
 247{
 248	if (!IS_ENABLED(CONFIG_NVME_TCP_TLS))
 249		return 0;
 250
 251	return ctrl->opts->tls || ctrl->opts->concat;
 252}
 253
 254static inline struct blk_mq_tags *nvme_tcp_tagset(struct nvme_tcp_queue *queue)
 255{
 256	u32 queue_idx = nvme_tcp_queue_id(queue);
 257
 258	if (queue_idx == 0)
 259		return queue->ctrl->admin_tag_set.tags[queue_idx];
 260	return queue->ctrl->tag_set.tags[queue_idx - 1];
 261}
 262
 263static inline u8 nvme_tcp_hdgst_len(struct nvme_tcp_queue *queue)
 264{
 265	return queue->hdr_digest ? NVME_TCP_DIGEST_LENGTH : 0;
 266}
 267
 268static inline u8 nvme_tcp_ddgst_len(struct nvme_tcp_queue *queue)
 269{
 270	return queue->data_digest ? NVME_TCP_DIGEST_LENGTH : 0;
 271}
 272
 273static inline void *nvme_tcp_req_cmd_pdu(struct nvme_tcp_request *req)
 274{
 275	return req->pdu;
 276}
 277
 278static inline void *nvme_tcp_req_data_pdu(struct nvme_tcp_request *req)
 279{
 280	/* use the pdu space in the back for the data pdu */
 281	return req->pdu + sizeof(struct nvme_tcp_cmd_pdu) -
 282		sizeof(struct nvme_tcp_data_pdu);
 283}
 284
 285static inline size_t nvme_tcp_inline_data_size(struct nvme_tcp_request *req)
 286{
 287	if (nvme_is_fabrics(req->req.cmd))
 288		return NVME_TCP_ADMIN_CCSZ;
 289	return req->queue->cmnd_capsule_len - sizeof(struct nvme_command);
 290}
 291
 292static inline bool nvme_tcp_async_req(struct nvme_tcp_request *req)
 293{
 294	return req == &req->queue->ctrl->async_req;
 295}
 296
 297static inline bool nvme_tcp_has_inline_data(struct nvme_tcp_request *req)
 298{
 299	struct request *rq;
 300
 301	if (unlikely(nvme_tcp_async_req(req)))
 302		return false; /* async events don't have a request */
 303
 304	rq = blk_mq_rq_from_pdu(req);
 305
 306	return rq_data_dir(rq) == WRITE && req->data_len &&
 307		req->data_len <= nvme_tcp_inline_data_size(req);
 308}
 309
 310static inline struct page *nvme_tcp_req_cur_page(struct nvme_tcp_request *req)
 311{
 312	return req->iter.bvec->bv_page;
 313}
 314
 315static inline size_t nvme_tcp_req_cur_offset(struct nvme_tcp_request *req)
 316{
 317	return req->iter.bvec->bv_offset + req->iter.iov_offset;
 318}
 319
 320static inline size_t nvme_tcp_req_cur_length(struct nvme_tcp_request *req)
 321{
 322	return min_t(size_t, iov_iter_single_seg_count(&req->iter),
 323			req->pdu_len - req->pdu_sent);
 324}
 325
 326static inline size_t nvme_tcp_pdu_data_left(struct nvme_tcp_request *req)
 327{
 328	return rq_data_dir(blk_mq_rq_from_pdu(req)) == WRITE ?
 329			req->pdu_len - req->pdu_sent : 0;
 330}
 331
 332static inline size_t nvme_tcp_pdu_last_send(struct nvme_tcp_request *req,
 333		int len)
 334{
 335	return nvme_tcp_pdu_data_left(req) <= len;
 336}
 337
 338static void nvme_tcp_init_iter(struct nvme_tcp_request *req,
 339		unsigned int dir)
 340{
 341	struct request *rq = blk_mq_rq_from_pdu(req);
 342	struct bio_vec *vec;
 343	unsigned int size;
 344	int nr_bvec;
 345	size_t offset;
 346
 347	if (rq->rq_flags & RQF_SPECIAL_PAYLOAD) {
 348		vec = &rq->special_vec;
 349		nr_bvec = 1;
 350		size = blk_rq_payload_bytes(rq);
 351		offset = 0;
 352	} else {
 353		struct bio *bio = req->curr_bio;
 354		struct bvec_iter bi;
 355		struct bio_vec bv;
 356
 357		vec = __bvec_iter_bvec(bio->bi_io_vec, bio->bi_iter);
 358		nr_bvec = 0;
 359		bio_for_each_bvec(bv, bio, bi) {
 360			nr_bvec++;
 361		}
 362		size = bio->bi_iter.bi_size;
 363		offset = bio->bi_iter.bi_bvec_done;
 364	}
 365
 366	iov_iter_bvec(&req->iter, dir, vec, nr_bvec, size);
 367	req->iter.iov_offset = offset;
 368}
 369
 370static inline void nvme_tcp_advance_req(struct nvme_tcp_request *req,
 371		int len)
 372{
 373	req->data_sent += len;
 374	req->pdu_sent += len;
 375	iov_iter_advance(&req->iter, len);
 376	if (!iov_iter_count(&req->iter) &&
 377	    req->data_sent < req->data_len) {
 378		req->curr_bio = req->curr_bio->bi_next;
 379		nvme_tcp_init_iter(req, ITER_SOURCE);
 380	}
 381}
 382
 383static inline void nvme_tcp_send_all(struct nvme_tcp_queue *queue)
 384{
 385	int ret;
 386
 387	/* drain the send queue as much as we can... */
 388	do {
 389		ret = nvme_tcp_try_send(queue);
 390	} while (ret > 0);
 391}
 392
 393static inline bool nvme_tcp_queue_has_pending(struct nvme_tcp_queue *queue)
 394{
 395	return !list_empty(&queue->send_list) ||
 396		!llist_empty(&queue->req_list);
 397}
 398
 399static inline bool nvme_tcp_queue_more(struct nvme_tcp_queue *queue)
 400{
 401	return !nvme_tcp_queue_tls(queue) &&
 402		nvme_tcp_queue_has_pending(queue);
 403}
 404
 405static inline void nvme_tcp_queue_request(struct nvme_tcp_request *req,
 406		bool last)
 407{
 408	struct nvme_tcp_queue *queue = req->queue;
 409	bool empty;
 410
 411	empty = llist_add(&req->lentry, &queue->req_list) &&
 412		list_empty(&queue->send_list) && !queue->request;
 413
 414	/*
 415	 * if we're the first on the send_list and we can try to send
 416	 * directly, otherwise queue io_work. Also, only do that if we
 417	 * are on the same cpu, so we don't introduce contention.
 418	 */
 419	if (queue->io_cpu == raw_smp_processor_id() &&
 420	    empty && mutex_trylock(&queue->send_mutex)) {
 421		nvme_tcp_send_all(queue);
 422		mutex_unlock(&queue->send_mutex);
 423	}
 424
 425	if (last && nvme_tcp_queue_has_pending(queue))
 426		queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
 427}
 428
 429static void nvme_tcp_process_req_list(struct nvme_tcp_queue *queue)
 430{
 431	struct nvme_tcp_request *req;
 432	struct llist_node *node;
 433
 434	for (node = llist_del_all(&queue->req_list); node; node = node->next) {
 435		req = llist_entry(node, struct nvme_tcp_request, lentry);
 436		list_add(&req->entry, &queue->send_list);
 437	}
 438}
 439
 440static inline struct nvme_tcp_request *
 441nvme_tcp_fetch_request(struct nvme_tcp_queue *queue)
 442{
 443	struct nvme_tcp_request *req;
 444
 445	req = list_first_entry_or_null(&queue->send_list,
 446			struct nvme_tcp_request, entry);
 447	if (!req) {
 448		nvme_tcp_process_req_list(queue);
 449		req = list_first_entry_or_null(&queue->send_list,
 450				struct nvme_tcp_request, entry);
 451		if (unlikely(!req))
 452			return NULL;
 453	}
 454
 455	list_del_init(&req->entry);
 456	init_llist_node(&req->lentry);
 457	return req;
 458}
 459
 460#define NVME_TCP_CRC_SEED (~0)
 461
 462static inline void nvme_tcp_ddgst_update(u32 *crcp,
 463		struct page *page, size_t off, size_t len)
 464{
 465	page += off / PAGE_SIZE;
 466	off %= PAGE_SIZE;
 467	while (len) {
 468		const void *vaddr = kmap_local_page(page);
 469		size_t n = min(len, (size_t)PAGE_SIZE - off);
 470
 471		*crcp = crc32c(*crcp, vaddr + off, n);
 472		kunmap_local(vaddr);
 473		page++;
 474		off = 0;
 475		len -= n;
 476	}
 477}
 478
 479static inline __le32 nvme_tcp_ddgst_final(u32 crc)
 480{
 481	return cpu_to_le32(~crc);
 482}
 483
 484static inline __le32 nvme_tcp_hdgst(const void *pdu, size_t len)
 485{
 486	return cpu_to_le32(~crc32c(NVME_TCP_CRC_SEED, pdu, len));
 487}
 488
 489static inline void nvme_tcp_set_hdgst(void *pdu, size_t len)
 490{
 491	*(__le32 *)(pdu + len) = nvme_tcp_hdgst(pdu, len);
 492}
 493
 494static int nvme_tcp_verify_hdgst(struct nvme_tcp_queue *queue,
 495		void *pdu, size_t pdu_len)
 496{
 497	struct nvme_tcp_hdr *hdr = pdu;
 498	__le32 recv_digest;
 499	__le32 exp_digest;
 500
 501	if (unlikely(!(hdr->flags & NVME_TCP_F_HDGST))) {
 502		dev_err(queue->ctrl->ctrl.device,
 503			"queue %d: header digest flag is cleared\n",
 504			nvme_tcp_queue_id(queue));
 505		return -EPROTO;
 506	}
 507
 508	recv_digest = *(__le32 *)(pdu + hdr->hlen);
 509	exp_digest = nvme_tcp_hdgst(pdu, pdu_len);
 510	if (recv_digest != exp_digest) {
 511		dev_err(queue->ctrl->ctrl.device,
 512			"header digest error: recv %#x expected %#x\n",
 513			le32_to_cpu(recv_digest), le32_to_cpu(exp_digest));
 514		return -EIO;
 515	}
 516
 517	return 0;
 518}
 519
 520static int nvme_tcp_check_ddgst(struct nvme_tcp_queue *queue, void *pdu)
 521{
 522	struct nvme_tcp_hdr *hdr = pdu;
 523	u8 digest_len = nvme_tcp_hdgst_len(queue);
 524	u32 len;
 525
 526	len = le32_to_cpu(hdr->plen) - hdr->hlen -
 527		((hdr->flags & NVME_TCP_F_HDGST) ? digest_len : 0);
 528
 529	if (unlikely(len && !(hdr->flags & NVME_TCP_F_DDGST))) {
 530		dev_err(queue->ctrl->ctrl.device,
 531			"queue %d: data digest flag is cleared\n",
 532		nvme_tcp_queue_id(queue));
 533		return -EPROTO;
 534	}
 535	queue->rcv_crc = NVME_TCP_CRC_SEED;
 536
 537	return 0;
 538}
 539
 540static void nvme_tcp_exit_request(struct blk_mq_tag_set *set,
 541		struct request *rq, unsigned int hctx_idx)
 542{
 543	struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
 544
 545	page_frag_free(req->pdu);
 546}
 547
 548static int nvme_tcp_init_request(struct blk_mq_tag_set *set,
 549		struct request *rq, unsigned int hctx_idx,
 550		unsigned int numa_node)
 551{
 552	struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(set->driver_data);
 553	struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
 554	struct nvme_tcp_cmd_pdu *pdu;
 555	int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
 556	struct nvme_tcp_queue *queue = &ctrl->queues[queue_idx];
 557	u8 hdgst = nvme_tcp_hdgst_len(queue);
 558
 559	req->pdu = page_frag_alloc(&queue->pf_cache,
 560		sizeof(struct nvme_tcp_cmd_pdu) + hdgst,
 561		GFP_KERNEL | __GFP_ZERO);
 562	if (!req->pdu)
 563		return -ENOMEM;
 564
 565	pdu = req->pdu;
 566	req->queue = queue;
 567	nvme_req(rq)->ctrl = &ctrl->ctrl;
 568	nvme_req(rq)->cmd = &pdu->cmd;
 569	init_llist_node(&req->lentry);
 570	INIT_LIST_HEAD(&req->entry);
 571
 572	return 0;
 573}
 574
 575static int nvme_tcp_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 576		unsigned int hctx_idx)
 577{
 578	struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(data);
 579	struct nvme_tcp_queue *queue = &ctrl->queues[hctx_idx + 1];
 580
 581	hctx->driver_data = queue;
 582	return 0;
 583}
 584
 585static int nvme_tcp_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 586		unsigned int hctx_idx)
 587{
 588	struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(data);
 589	struct nvme_tcp_queue *queue = &ctrl->queues[0];
 590
 591	hctx->driver_data = queue;
 592	return 0;
 593}
 594
 595static enum nvme_tcp_recv_state
 596nvme_tcp_recv_state(struct nvme_tcp_queue *queue)
 597{
 598	return  (queue->pdu_remaining) ? NVME_TCP_RECV_PDU :
 599		(queue->ddgst_remaining) ? NVME_TCP_RECV_DDGST :
 600		NVME_TCP_RECV_DATA;
 601}
 602
 603static void nvme_tcp_init_recv_ctx(struct nvme_tcp_queue *queue)
 604{
 605	queue->pdu_remaining = sizeof(struct nvme_tcp_rsp_pdu) +
 606				nvme_tcp_hdgst_len(queue);
 607	queue->pdu_offset = 0;
 608	queue->data_remaining = -1;
 609	queue->ddgst_remaining = 0;
 610}
 611
 612static void nvme_tcp_error_recovery(struct nvme_ctrl *ctrl)
 613{
 614	if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_RESETTING))
 615		return;
 616
 617	dev_warn(ctrl->device, "starting error recovery\n");
 618	queue_work(nvme_reset_wq, &to_tcp_ctrl(ctrl)->err_work);
 619}
 620
 621static int nvme_tcp_process_nvme_cqe(struct nvme_tcp_queue *queue,
 622		struct nvme_completion *cqe)
 623{
 624	struct nvme_tcp_request *req;
 625	struct request *rq;
 626
 627	rq = nvme_find_rq(nvme_tcp_tagset(queue), cqe->command_id);
 628	if (!rq) {
 629		dev_err(queue->ctrl->ctrl.device,
 630			"got bad cqe.command_id %#x on queue %d\n",
 631			cqe->command_id, nvme_tcp_queue_id(queue));
 632		nvme_tcp_error_recovery(&queue->ctrl->ctrl);
 633		return -EINVAL;
 634	}
 635
 636	req = blk_mq_rq_to_pdu(rq);
 637	if (req->status == cpu_to_le16(NVME_SC_SUCCESS))
 638		req->status = cqe->status;
 639
 640	if (!nvme_try_complete_req(rq, req->status, cqe->result))
 641		nvme_complete_rq(rq);
 642	queue->nr_cqe++;
 643
 644	return 0;
 645}
 646
 647static int nvme_tcp_handle_c2h_data(struct nvme_tcp_queue *queue,
 648		struct nvme_tcp_data_pdu *pdu)
 649{
 650	struct request *rq;
 651
 652	rq = nvme_find_rq(nvme_tcp_tagset(queue), pdu->command_id);
 653	if (!rq) {
 654		dev_err(queue->ctrl->ctrl.device,
 655			"got bad c2hdata.command_id %#x on queue %d\n",
 656			pdu->command_id, nvme_tcp_queue_id(queue));
 657		return -ENOENT;
 658	}
 659
 660	if (!blk_rq_payload_bytes(rq)) {
 661		dev_err(queue->ctrl->ctrl.device,
 662			"queue %d tag %#x unexpected data\n",
 663			nvme_tcp_queue_id(queue), rq->tag);
 664		return -EIO;
 665	}
 666
 667	queue->data_remaining = le32_to_cpu(pdu->data_length);
 668
 669	if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS &&
 670	    unlikely(!(pdu->hdr.flags & NVME_TCP_F_DATA_LAST))) {
 671		dev_err(queue->ctrl->ctrl.device,
 672			"queue %d tag %#x SUCCESS set but not last PDU\n",
 673			nvme_tcp_queue_id(queue), rq->tag);
 674		nvme_tcp_error_recovery(&queue->ctrl->ctrl);
 675		return -EPROTO;
 676	}
 677
 678	return 0;
 679}
 680
 681static int nvme_tcp_handle_comp(struct nvme_tcp_queue *queue,
 682		struct nvme_tcp_rsp_pdu *pdu)
 683{
 684	struct nvme_completion *cqe = &pdu->cqe;
 685	int ret = 0;
 686
 687	/*
 688	 * AEN requests are special as they don't time out and can
 689	 * survive any kind of queue freeze and often don't respond to
 690	 * aborts.  We don't even bother to allocate a struct request
 691	 * for them but rather special case them here.
 692	 */
 693	if (unlikely(nvme_is_aen_req(nvme_tcp_queue_id(queue),
 694				     cqe->command_id)))
 695		nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
 696				&cqe->result);
 697	else
 698		ret = nvme_tcp_process_nvme_cqe(queue, cqe);
 699
 700	return ret;
 701}
 702
 703static void nvme_tcp_setup_h2c_data_pdu(struct nvme_tcp_request *req)
 704{
 705	struct nvme_tcp_data_pdu *data = nvme_tcp_req_data_pdu(req);
 706	struct nvme_tcp_queue *queue = req->queue;
 707	struct request *rq = blk_mq_rq_from_pdu(req);
 708	u32 h2cdata_sent = req->pdu_len;
 709	u8 hdgst = nvme_tcp_hdgst_len(queue);
 710	u8 ddgst = nvme_tcp_ddgst_len(queue);
 711
 712	req->state = NVME_TCP_SEND_H2C_PDU;
 713	req->offset = 0;
 714	req->pdu_len = min(req->h2cdata_left, queue->maxh2cdata);
 715	req->pdu_sent = 0;
 716	req->h2cdata_left -= req->pdu_len;
 717	req->h2cdata_offset += h2cdata_sent;
 718
 719	memset(data, 0, sizeof(*data));
 720	data->hdr.type = nvme_tcp_h2c_data;
 721	if (!req->h2cdata_left)
 722		data->hdr.flags = NVME_TCP_F_DATA_LAST;
 723	if (queue->hdr_digest)
 724		data->hdr.flags |= NVME_TCP_F_HDGST;
 725	if (queue->data_digest)
 726		data->hdr.flags |= NVME_TCP_F_DDGST;
 727	data->hdr.hlen = sizeof(*data);
 728	data->hdr.pdo = data->hdr.hlen + hdgst;
 729	data->hdr.plen =
 730		cpu_to_le32(data->hdr.hlen + hdgst + req->pdu_len + ddgst);
 731	data->ttag = req->ttag;
 732	data->command_id = nvme_cid(rq);
 733	data->data_offset = cpu_to_le32(req->h2cdata_offset);
 734	data->data_length = cpu_to_le32(req->pdu_len);
 735}
 736
 737static int nvme_tcp_handle_r2t(struct nvme_tcp_queue *queue,
 738		struct nvme_tcp_r2t_pdu *pdu)
 739{
 740	struct nvme_tcp_request *req;
 741	struct request *rq;
 742	u32 r2t_length = le32_to_cpu(pdu->r2t_length);
 743	u32 r2t_offset = le32_to_cpu(pdu->r2t_offset);
 744
 745	rq = nvme_find_rq(nvme_tcp_tagset(queue), pdu->command_id);
 746	if (!rq) {
 747		dev_err(queue->ctrl->ctrl.device,
 748			"got bad r2t.command_id %#x on queue %d\n",
 749			pdu->command_id, nvme_tcp_queue_id(queue));
 750		return -ENOENT;
 751	}
 752	req = blk_mq_rq_to_pdu(rq);
 753
 754	if (unlikely(!r2t_length)) {
 755		dev_err(queue->ctrl->ctrl.device,
 756			"req %d r2t len is %u, probably a bug...\n",
 757			rq->tag, r2t_length);
 758		return -EPROTO;
 759	}
 760
 761	if (unlikely(req->data_sent + r2t_length > req->data_len)) {
 762		dev_err(queue->ctrl->ctrl.device,
 763			"req %d r2t len %u exceeded data len %u (%zu sent)\n",
 764			rq->tag, r2t_length, req->data_len, req->data_sent);
 765		return -EPROTO;
 766	}
 767
 768	if (unlikely(r2t_offset < req->data_sent)) {
 769		dev_err(queue->ctrl->ctrl.device,
 770			"req %d unexpected r2t offset %u (expected %zu)\n",
 771			rq->tag, r2t_offset, req->data_sent);
 772		return -EPROTO;
 773	}
 774
 775	if (llist_on_list(&req->lentry) ||
 776	    !list_empty(&req->entry)) {
 777		dev_err(queue->ctrl->ctrl.device,
 778			"req %d unexpected r2t while processing request\n",
 779			rq->tag);
 780		return -EPROTO;
 781	}
 782
 783	req->pdu_len = 0;
 784	req->h2cdata_left = r2t_length;
 785	req->h2cdata_offset = r2t_offset;
 786	req->ttag = pdu->ttag;
 787
 788	nvme_tcp_setup_h2c_data_pdu(req);
 789
 790	llist_add(&req->lentry, &queue->req_list);
 791	queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
 792
 793	return 0;
 794}
 795
 796static void nvme_tcp_handle_c2h_term(struct nvme_tcp_queue *queue,
 797		struct nvme_tcp_term_pdu *pdu)
 798{
 799	u16 fes;
 800	const char *msg;
 801	u32 plen = le32_to_cpu(pdu->hdr.plen);
 802
 803	static const char * const msg_table[] = {
 804		[NVME_TCP_FES_INVALID_PDU_HDR] = "Invalid PDU Header Field",
 805		[NVME_TCP_FES_PDU_SEQ_ERR] = "PDU Sequence Error",
 806		[NVME_TCP_FES_HDR_DIGEST_ERR] = "Header Digest Error",
 807		[NVME_TCP_FES_DATA_OUT_OF_RANGE] = "Data Transfer Out Of Range",
 808		[NVME_TCP_FES_DATA_LIMIT_EXCEEDED] = "Data Transfer Limit Exceeded",
 809		[NVME_TCP_FES_UNSUPPORTED_PARAM] = "Unsupported Parameter",
 810	};
 811
 812	if (plen < NVME_TCP_MIN_C2HTERM_PLEN ||
 813	    plen > NVME_TCP_MAX_C2HTERM_PLEN) {
 814		dev_err(queue->ctrl->ctrl.device,
 815			"Received a malformed C2HTermReq PDU (plen = %u)\n",
 816			plen);
 817		return;
 818	}
 819
 820	fes = le16_to_cpu(pdu->fes);
 821	if (fes && fes < ARRAY_SIZE(msg_table))
 822		msg = msg_table[fes];
 823	else
 824		msg = "Unknown";
 825
 826	dev_err(queue->ctrl->ctrl.device,
 827		"Received C2HTermReq (FES = %s)\n", msg);
 828}
 829
 830static int nvme_tcp_recv_pdu(struct nvme_tcp_queue *queue, struct sk_buff *skb,
 831		unsigned int *offset, size_t *len)
 832{
 833	struct nvme_tcp_hdr *hdr;
 834	char *pdu = queue->pdu;
 835	size_t rcv_len = min_t(size_t, *len, queue->pdu_remaining);
 836	int ret;
 837
 838	ret = skb_copy_bits(skb, *offset,
 839		&pdu[queue->pdu_offset], rcv_len);
 840	if (unlikely(ret))
 841		return ret;
 842
 843	queue->pdu_remaining -= rcv_len;
 844	queue->pdu_offset += rcv_len;
 845	*offset += rcv_len;
 846	*len -= rcv_len;
 847	if (queue->pdu_remaining)
 848		return 0;
 849
 850	hdr = queue->pdu;
 851	if (unlikely(hdr->hlen != sizeof(struct nvme_tcp_rsp_pdu))) {
 852		if (!nvme_tcp_recv_pdu_supported(hdr->type))
 853			goto unsupported_pdu;
 854
 855		dev_err(queue->ctrl->ctrl.device,
 856			"pdu type %d has unexpected header length (%d)\n",
 857			hdr->type, hdr->hlen);
 858		return -EPROTO;
 859	}
 860
 861	if (unlikely(hdr->type == nvme_tcp_c2h_term)) {
 862		/*
 863		 * C2HTermReq never includes Header or Data digests.
 864		 * Skip the checks.
 865		 */
 866		nvme_tcp_handle_c2h_term(queue, (void *)queue->pdu);
 867		return -EINVAL;
 868	}
 869
 870	if (queue->hdr_digest) {
 871		ret = nvme_tcp_verify_hdgst(queue, queue->pdu, hdr->hlen);
 872		if (unlikely(ret))
 873			return ret;
 874	}
 875
 876
 877	if (queue->data_digest) {
 878		ret = nvme_tcp_check_ddgst(queue, queue->pdu);
 879		if (unlikely(ret))
 880			return ret;
 881	}
 882
 883	switch (hdr->type) {
 884	case nvme_tcp_c2h_data:
 885		return nvme_tcp_handle_c2h_data(queue, (void *)queue->pdu);
 886	case nvme_tcp_rsp:
 887		nvme_tcp_init_recv_ctx(queue);
 888		return nvme_tcp_handle_comp(queue, (void *)queue->pdu);
 889	case nvme_tcp_r2t:
 890		nvme_tcp_init_recv_ctx(queue);
 891		return nvme_tcp_handle_r2t(queue, (void *)queue->pdu);
 892	default:
 893		goto unsupported_pdu;
 894	}
 895
 896unsupported_pdu:
 897	dev_err(queue->ctrl->ctrl.device,
 898		"unsupported pdu type (%d)\n", hdr->type);
 899	return -EINVAL;
 900}
 901
 902static inline void nvme_tcp_end_request(struct request *rq, u16 status)
 903{
 904	union nvme_result res = {};
 905
 906	if (!nvme_try_complete_req(rq, cpu_to_le16(status << 1), res))
 907		nvme_complete_rq(rq);
 908}
 909
 910static int nvme_tcp_recv_data(struct nvme_tcp_queue *queue, struct sk_buff *skb,
 911			      unsigned int *offset, size_t *len)
 912{
 913	struct nvme_tcp_data_pdu *pdu = (void *)queue->pdu;
 914	struct request *rq =
 915		nvme_cid_to_rq(nvme_tcp_tagset(queue), pdu->command_id);
 916	struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
 917
 918	while (true) {
 919		int recv_len, ret;
 920
 921		recv_len = min_t(size_t, *len, queue->data_remaining);
 922		if (!recv_len)
 923			break;
 924
 925		if (!iov_iter_count(&req->iter)) {
 926			req->curr_bio = req->curr_bio->bi_next;
 927
 928			/*
 929			 * If we don`t have any bios it means that controller
 930			 * sent more data than we requested, hence error
 931			 */
 932			if (!req->curr_bio) {
 933				dev_err(queue->ctrl->ctrl.device,
 934					"queue %d no space in request %#x",
 935					nvme_tcp_queue_id(queue), rq->tag);
 936				nvme_tcp_init_recv_ctx(queue);
 937				return -EIO;
 938			}
 939			nvme_tcp_init_iter(req, ITER_DEST);
 940		}
 941
 942		/* we can read only from what is left in this bio */
 943		recv_len = min_t(size_t, recv_len,
 944				iov_iter_count(&req->iter));
 945
 946		if (queue->data_digest)
 947			ret = skb_copy_and_crc32c_datagram_iter(skb, *offset,
 948				&req->iter, recv_len, &queue->rcv_crc);
 949		else
 950			ret = skb_copy_datagram_iter(skb, *offset,
 951					&req->iter, recv_len);
 952		if (ret) {
 953			dev_err(queue->ctrl->ctrl.device,
 954				"queue %d failed to copy request %#x data",
 955				nvme_tcp_queue_id(queue), rq->tag);
 956			return ret;
 957		}
 958
 959		*len -= recv_len;
 960		*offset += recv_len;
 961		queue->data_remaining -= recv_len;
 962	}
 963
 964	if (!queue->data_remaining) {
 965		if (queue->data_digest) {
 966			queue->exp_ddgst = nvme_tcp_ddgst_final(queue->rcv_crc);
 967			queue->ddgst_remaining = NVME_TCP_DIGEST_LENGTH;
 968		} else {
 969			if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
 970				nvme_tcp_end_request(rq,
 971						le16_to_cpu(req->status));
 972				queue->nr_cqe++;
 973			}
 974			nvme_tcp_init_recv_ctx(queue);
 975		}
 976	}
 977
 978	return 0;
 979}
 980
 981static int nvme_tcp_recv_ddgst(struct nvme_tcp_queue *queue,
 982		struct sk_buff *skb, unsigned int *offset, size_t *len)
 983{
 984	struct nvme_tcp_data_pdu *pdu = (void *)queue->pdu;
 985	char *ddgst = (char *)&queue->recv_ddgst;
 986	size_t recv_len = min_t(size_t, *len, queue->ddgst_remaining);
 987	off_t off = NVME_TCP_DIGEST_LENGTH - queue->ddgst_remaining;
 988	int ret;
 989
 990	ret = skb_copy_bits(skb, *offset, &ddgst[off], recv_len);
 991	if (unlikely(ret))
 992		return ret;
 993
 994	queue->ddgst_remaining -= recv_len;
 995	*offset += recv_len;
 996	*len -= recv_len;
 997	if (queue->ddgst_remaining)
 998		return 0;
 999
1000	if (queue->recv_ddgst != queue->exp_ddgst) {
1001		struct request *rq = nvme_cid_to_rq(nvme_tcp_tagset(queue),
1002					pdu->command_id);
1003		struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
1004
1005		req->status = cpu_to_le16(NVME_SC_DATA_XFER_ERROR);
1006
1007		dev_err(queue->ctrl->ctrl.device,
1008			"data digest error: recv %#x expected %#x\n",
1009			le32_to_cpu(queue->recv_ddgst),
1010			le32_to_cpu(queue->exp_ddgst));
1011	}
1012
1013	if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
1014		struct request *rq = nvme_cid_to_rq(nvme_tcp_tagset(queue),
1015					pdu->command_id);
1016		struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
1017
1018		nvme_tcp_end_request(rq, le16_to_cpu(req->status));
1019		queue->nr_cqe++;
1020	}
1021
1022	nvme_tcp_init_recv_ctx(queue);
1023	return 0;
1024}
1025
1026static int nvme_tcp_recv_skb(read_descriptor_t *desc, struct sk_buff *skb,
1027			     unsigned int offset, size_t len)
1028{
1029	struct nvme_tcp_queue *queue = desc->arg.data;
1030	size_t consumed = len;
1031	int result;
1032
1033	if (unlikely(!queue->rd_enabled))
1034		return -EFAULT;
1035
1036	while (len) {
1037		switch (nvme_tcp_recv_state(queue)) {
1038		case NVME_TCP_RECV_PDU:
1039			result = nvme_tcp_recv_pdu(queue, skb, &offset, &len);
1040			break;
1041		case NVME_TCP_RECV_DATA:
1042			result = nvme_tcp_recv_data(queue, skb, &offset, &len);
1043			break;
1044		case NVME_TCP_RECV_DDGST:
1045			result = nvme_tcp_recv_ddgst(queue, skb, &offset, &len);
1046			break;
1047		default:
1048			result = -EFAULT;
1049		}
1050		if (result) {
1051			dev_err(queue->ctrl->ctrl.device,
1052				"receive failed:  %d\n", result);
1053			queue->rd_enabled = false;
1054			nvme_tcp_error_recovery(&queue->ctrl->ctrl);
1055			return result;
1056		}
1057	}
1058
1059	return consumed;
1060}
1061
1062static void nvme_tcp_data_ready(struct sock *sk)
1063{
1064	struct nvme_tcp_queue *queue;
1065
1066	trace_sk_data_ready(sk);
1067
1068	read_lock_bh(&sk->sk_callback_lock);
1069	queue = sk->sk_user_data;
1070	if (likely(queue && queue->rd_enabled) &&
1071	    !test_bit(NVME_TCP_Q_POLLING, &queue->flags))
1072		queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
1073	read_unlock_bh(&sk->sk_callback_lock);
1074}
1075
1076static void nvme_tcp_write_space(struct sock *sk)
1077{
1078	struct nvme_tcp_queue *queue;
1079
1080	read_lock_bh(&sk->sk_callback_lock);
1081	queue = sk->sk_user_data;
1082	if (likely(queue && sk_stream_is_writeable(sk))) {
1083		clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1084		/* Ensure pending TLS partial records are retried */
1085		if (nvme_tcp_queue_tls(queue))
1086			queue->write_space(sk);
1087		queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
1088	}
1089	read_unlock_bh(&sk->sk_callback_lock);
1090}
1091
1092static void nvme_tcp_state_change(struct sock *sk)
1093{
1094	struct nvme_tcp_queue *queue;
1095
1096	read_lock_bh(&sk->sk_callback_lock);
1097	queue = sk->sk_user_data;
1098	if (!queue)
1099		goto done;
1100
1101	switch (sk->sk_state) {
1102	case TCP_CLOSE:
1103	case TCP_CLOSE_WAIT:
1104	case TCP_LAST_ACK:
1105	case TCP_FIN_WAIT1:
1106	case TCP_FIN_WAIT2:
1107		nvme_tcp_error_recovery(&queue->ctrl->ctrl);
1108		break;
1109	default:
1110		dev_info(queue->ctrl->ctrl.device,
1111			"queue %d socket state %d\n",
1112			nvme_tcp_queue_id(queue), sk->sk_state);
1113	}
1114
1115	queue->state_change(sk);
1116done:
1117	read_unlock_bh(&sk->sk_callback_lock);
1118}
1119
1120static inline void nvme_tcp_done_send_req(struct nvme_tcp_queue *queue)
1121{
1122	queue->request = NULL;
1123}
1124
1125static void nvme_tcp_fail_request(struct nvme_tcp_request *req)
1126{
1127	if (nvme_tcp_async_req(req)) {
1128		union nvme_result res = {};
1129
1130		nvme_complete_async_event(&req->queue->ctrl->ctrl,
1131				cpu_to_le16(NVME_SC_HOST_PATH_ERROR), &res);
1132	} else {
1133		nvme_tcp_end_request(blk_mq_rq_from_pdu(req),
1134				NVME_SC_HOST_PATH_ERROR);
1135	}
1136}
1137
1138static int nvme_tcp_try_send_data(struct nvme_tcp_request *req)
1139{
1140	struct nvme_tcp_queue *queue = req->queue;
1141	int req_data_len = req->data_len;
1142	u32 h2cdata_left = req->h2cdata_left;
1143
1144	while (true) {
1145		struct bio_vec bvec;
1146		struct msghdr msg = {
1147			.msg_flags = MSG_DONTWAIT | MSG_SPLICE_PAGES,
1148		};
1149		struct page *page = nvme_tcp_req_cur_page(req);
1150		size_t offset = nvme_tcp_req_cur_offset(req);
1151		size_t len = nvme_tcp_req_cur_length(req);
1152		bool last = nvme_tcp_pdu_last_send(req, len);
1153		int req_data_sent = req->data_sent;
1154		int ret;
1155
1156		if (last && !queue->data_digest && !nvme_tcp_queue_more(queue))
1157			msg.msg_flags |= MSG_EOR;
1158		else
1159			msg.msg_flags |= MSG_MORE;
1160
1161		if (!sendpages_ok(page, len, offset))
1162			msg.msg_flags &= ~MSG_SPLICE_PAGES;
1163
1164		bvec_set_page(&bvec, page, len, offset);
1165		iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, &bvec, 1, len);
1166		ret = sock_sendmsg(queue->sock, &msg);
1167		if (ret <= 0)
1168			return ret;
1169
1170		if (queue->data_digest)
1171			nvme_tcp_ddgst_update(&queue->snd_crc, page,
1172					offset, ret);
1173
1174		/*
1175		 * update the request iterator except for the last payload send
1176		 * in the request where we don't want to modify it as we may
1177		 * compete with the RX path completing the request.
1178		 */
1179		if (req_data_sent + ret < req_data_len)
1180			nvme_tcp_advance_req(req, ret);
1181
1182		/* fully successful last send in current PDU */
1183		if (last && ret == len) {
1184			if (queue->data_digest) {
1185				req->ddgst =
1186					nvme_tcp_ddgst_final(queue->snd_crc);
1187				req->state = NVME_TCP_SEND_DDGST;
1188				req->offset = 0;
1189			} else {
1190				if (h2cdata_left)
1191					nvme_tcp_setup_h2c_data_pdu(req);
1192				else
1193					nvme_tcp_done_send_req(queue);
1194			}
1195			return 1;
1196		}
1197	}
1198	return -EAGAIN;
1199}
1200
1201static int nvme_tcp_try_send_cmd_pdu(struct nvme_tcp_request *req)
1202{
1203	struct nvme_tcp_queue *queue = req->queue;
1204	struct nvme_tcp_cmd_pdu *pdu = nvme_tcp_req_cmd_pdu(req);
1205	struct bio_vec bvec;
1206	struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_SPLICE_PAGES, };
1207	bool inline_data = nvme_tcp_has_inline_data(req);
1208	u8 hdgst = nvme_tcp_hdgst_len(queue);
1209	int len = sizeof(*pdu) + hdgst - req->offset;
1210	int ret;
1211
1212	if (inline_data || nvme_tcp_queue_more(queue))
1213		msg.msg_flags |= MSG_MORE;
1214	else
1215		msg.msg_flags |= MSG_EOR;
1216
1217	if (queue->hdr_digest && !req->offset)
1218		nvme_tcp_set_hdgst(pdu, sizeof(*pdu));
1219
1220	bvec_set_virt(&bvec, (void *)pdu + req->offset, len);
1221	iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, &bvec, 1, len);
1222	ret = sock_sendmsg(queue->sock, &msg);
1223	if (unlikely(ret <= 0))
1224		return ret;
1225
1226	len -= ret;
1227	if (!len) {
1228		if (inline_data) {
1229			req->state = NVME_TCP_SEND_DATA;
1230			if (queue->data_digest)
1231				queue->snd_crc = NVME_TCP_CRC_SEED;
1232		} else {
1233			nvme_tcp_done_send_req(queue);
1234		}
1235		return 1;
1236	}
1237	req->offset += ret;
1238
1239	return -EAGAIN;
1240}
1241
1242static int nvme_tcp_try_send_data_pdu(struct nvme_tcp_request *req)
1243{
1244	struct nvme_tcp_queue *queue = req->queue;
1245	struct nvme_tcp_data_pdu *pdu = nvme_tcp_req_data_pdu(req);
1246	struct bio_vec bvec;
1247	struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_MORE, };
1248	u8 hdgst = nvme_tcp_hdgst_len(queue);
1249	int len = sizeof(*pdu) - req->offset + hdgst;
1250	int ret;
1251
1252	if (queue->hdr_digest && !req->offset)
1253		nvme_tcp_set_hdgst(pdu, sizeof(*pdu));
1254
1255	if (!req->h2cdata_left)
1256		msg.msg_flags |= MSG_SPLICE_PAGES;
1257
1258	bvec_set_virt(&bvec, (void *)pdu + req->offset, len);
1259	iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, &bvec, 1, len);
1260	ret = sock_sendmsg(queue->sock, &msg);
1261	if (unlikely(ret <= 0))
1262		return ret;
1263
1264	len -= ret;
1265	if (!len) {
1266		req->state = NVME_TCP_SEND_DATA;
1267		if (queue->data_digest)
1268			queue->snd_crc = NVME_TCP_CRC_SEED;
1269		return 1;
1270	}
1271	req->offset += ret;
1272
1273	return -EAGAIN;
1274}
1275
1276static int nvme_tcp_try_send_ddgst(struct nvme_tcp_request *req)
1277{
1278	struct nvme_tcp_queue *queue = req->queue;
1279	size_t offset = req->offset;
1280	u32 h2cdata_left = req->h2cdata_left;
1281	int ret;
1282	struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
1283	struct kvec iov = {
1284		.iov_base = (u8 *)&req->ddgst + req->offset,
1285		.iov_len = NVME_TCP_DIGEST_LENGTH - req->offset
1286	};
1287
1288	if (nvme_tcp_queue_more(queue))
1289		msg.msg_flags |= MSG_MORE;
1290	else
1291		msg.msg_flags |= MSG_EOR;
1292
1293	ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
1294	if (unlikely(ret <= 0))
1295		return ret;
1296
1297	if (offset + ret == NVME_TCP_DIGEST_LENGTH) {
1298		if (h2cdata_left)
1299			nvme_tcp_setup_h2c_data_pdu(req);
1300		else
1301			nvme_tcp_done_send_req(queue);
1302		return 1;
1303	}
1304
1305	req->offset += ret;
1306	return -EAGAIN;
1307}
1308
1309static int nvme_tcp_try_send(struct nvme_tcp_queue *queue)
1310{
1311	struct nvme_tcp_request *req;
1312	unsigned int noreclaim_flag;
1313	int ret = 1;
1314
1315	if (!queue->request) {
1316		queue->request = nvme_tcp_fetch_request(queue);
1317		if (!queue->request)
1318			return 0;
1319	}
1320	req = queue->request;
1321
1322	noreclaim_flag = memalloc_noreclaim_save();
1323	if (req->state == NVME_TCP_SEND_CMD_PDU) {
1324		ret = nvme_tcp_try_send_cmd_pdu(req);
1325		if (ret <= 0)
1326			goto done;
1327		if (!nvme_tcp_has_inline_data(req))
1328			goto out;
1329	}
1330
1331	if (req->state == NVME_TCP_SEND_H2C_PDU) {
1332		ret = nvme_tcp_try_send_data_pdu(req);
1333		if (ret <= 0)
1334			goto done;
1335	}
1336
1337	if (req->state == NVME_TCP_SEND_DATA) {
1338		ret = nvme_tcp_try_send_data(req);
1339		if (ret <= 0)
1340			goto done;
1341	}
1342
1343	if (req->state == NVME_TCP_SEND_DDGST)
1344		ret = nvme_tcp_try_send_ddgst(req);
1345done:
1346	if (ret == -EAGAIN) {
1347		ret = 0;
1348	} else if (ret < 0) {
1349		dev_err(queue->ctrl->ctrl.device,
1350			"failed to send request %d\n", ret);
1351		nvme_tcp_fail_request(queue->request);
1352		nvme_tcp_done_send_req(queue);
1353	}
1354out:
1355	memalloc_noreclaim_restore(noreclaim_flag);
1356	return ret;
1357}
1358
1359static int nvme_tcp_try_recv(struct nvme_tcp_queue *queue)
1360{
1361	struct socket *sock = queue->sock;
1362	struct sock *sk = sock->sk;
1363	read_descriptor_t rd_desc;
1364	int consumed;
1365
1366	rd_desc.arg.data = queue;
1367	rd_desc.count = 1;
1368	lock_sock(sk);
1369	queue->nr_cqe = 0;
1370	consumed = sock->ops->read_sock(sk, &rd_desc, nvme_tcp_recv_skb);
1371	release_sock(sk);
1372	return consumed == -EAGAIN ? 0 : consumed;
1373}
1374
1375static void nvme_tcp_io_work(struct work_struct *w)
1376{
1377	struct nvme_tcp_queue *queue =
1378		container_of(w, struct nvme_tcp_queue, io_work);
1379	unsigned long deadline = jiffies + msecs_to_jiffies(1);
1380
1381	do {
1382		bool pending = false;
1383		int result;
1384
1385		if (mutex_trylock(&queue->send_mutex)) {
1386			result = nvme_tcp_try_send(queue);
1387			mutex_unlock(&queue->send_mutex);
1388			if (result > 0)
1389				pending = true;
1390			else if (unlikely(result < 0))
1391				break;
1392		}
1393
1394		result = nvme_tcp_try_recv(queue);
1395		if (result > 0)
1396			pending = true;
1397		else if (unlikely(result < 0))
1398			return;
1399
1400		/* did we get some space after spending time in recv? */
1401		if (nvme_tcp_queue_has_pending(queue) &&
1402		    sk_stream_is_writeable(queue->sock->sk))
1403			pending = true;
1404
1405		if (!pending || !queue->rd_enabled)
1406			return;
1407
1408	} while (!time_after(jiffies, deadline)); /* quota is exhausted */
1409
1410	queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
1411}
1412
1413static void nvme_tcp_free_async_req(struct nvme_tcp_ctrl *ctrl)
1414{
1415	struct nvme_tcp_request *async = &ctrl->async_req;
1416
1417	page_frag_free(async->pdu);
1418}
1419
1420static int nvme_tcp_alloc_async_req(struct nvme_tcp_ctrl *ctrl)
1421{
1422	struct nvme_tcp_queue *queue = &ctrl->queues[0];
1423	struct nvme_tcp_request *async = &ctrl->async_req;
1424	u8 hdgst = nvme_tcp_hdgst_len(queue);
1425
1426	async->pdu = page_frag_alloc(&queue->pf_cache,
1427		sizeof(struct nvme_tcp_cmd_pdu) + hdgst,
1428		GFP_KERNEL | __GFP_ZERO);
1429	if (!async->pdu)
1430		return -ENOMEM;
1431
1432	async->queue = &ctrl->queues[0];
1433	return 0;
1434}
1435
1436static void nvme_tcp_free_queue(struct nvme_ctrl *nctrl, int qid)
1437{
1438	struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1439	struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1440	unsigned int noreclaim_flag;
1441
1442	if (!test_and_clear_bit(NVME_TCP_Q_ALLOCATED, &queue->flags))
1443		return;
1444
1445	page_frag_cache_drain(&queue->pf_cache);
1446
1447	noreclaim_flag = memalloc_noreclaim_save();
1448	/* ->sock will be released by fput() */
1449	fput(queue->sock->file);
1450	queue->sock = NULL;
1451	memalloc_noreclaim_restore(noreclaim_flag);
1452
1453	kfree(queue->pdu);
1454	mutex_destroy(&queue->send_mutex);
1455	mutex_destroy(&queue->queue_lock);
1456}
1457
1458static int nvme_tcp_init_connection(struct nvme_tcp_queue *queue)
1459{
1460	struct nvme_tcp_icreq_pdu *icreq;
1461	struct nvme_tcp_icresp_pdu *icresp;
1462	char cbuf[CMSG_LEN(sizeof(char))] = {};
1463	u8 ctype;
1464	struct msghdr msg = {};
1465	struct kvec iov;
1466	bool ctrl_hdgst, ctrl_ddgst;
1467	u32 maxh2cdata;
1468	int ret;
1469
1470	icreq = kzalloc(sizeof(*icreq), GFP_KERNEL);
1471	if (!icreq)
1472		return -ENOMEM;
1473
1474	icresp = kzalloc(sizeof(*icresp), GFP_KERNEL);
1475	if (!icresp) {
1476		ret = -ENOMEM;
1477		goto free_icreq;
1478	}
1479
1480	icreq->hdr.type = nvme_tcp_icreq;
1481	icreq->hdr.hlen = sizeof(*icreq);
1482	icreq->hdr.pdo = 0;
1483	icreq->hdr.plen = cpu_to_le32(icreq->hdr.hlen);
1484	icreq->pfv = cpu_to_le16(NVME_TCP_PFV_1_0);
1485	icreq->maxr2t = 0; /* single inflight r2t supported */
1486	icreq->hpda = 0; /* no alignment constraint */
1487	if (queue->hdr_digest)
1488		icreq->digest |= NVME_TCP_HDR_DIGEST_ENABLE;
1489	if (queue->data_digest)
1490		icreq->digest |= NVME_TCP_DATA_DIGEST_ENABLE;
1491
1492	iov.iov_base = icreq;
1493	iov.iov_len = sizeof(*icreq);
1494	ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
1495	if (ret < 0) {
1496		pr_warn("queue %d: failed to send icreq, error %d\n",
1497			nvme_tcp_queue_id(queue), ret);
1498		goto free_icresp;
1499	}
1500
1501	memset(&msg, 0, sizeof(msg));
1502	iov.iov_base = icresp;
1503	iov.iov_len = sizeof(*icresp);
1504	if (nvme_tcp_queue_tls(queue)) {
1505		msg.msg_control = cbuf;
1506		msg.msg_controllen = sizeof(cbuf);
1507	}
1508	msg.msg_flags = MSG_WAITALL;
1509	ret = kernel_recvmsg(queue->sock, &msg, &iov, 1,
1510			iov.iov_len, msg.msg_flags);
1511	if (ret >= 0 && ret < sizeof(*icresp))
1512		ret = -ECONNRESET;
1513	if (ret < 0) {
1514		pr_warn("queue %d: failed to receive icresp, error %d\n",
1515			nvme_tcp_queue_id(queue), ret);
1516		goto free_icresp;
1517	}
1518	ret = -ENOTCONN;
1519	if (nvme_tcp_queue_tls(queue)) {
1520		ctype = tls_get_record_type(queue->sock->sk,
1521					    (struct cmsghdr *)cbuf);
1522		if (ctype != TLS_RECORD_TYPE_DATA) {
1523			pr_err("queue %d: unhandled TLS record %d\n",
1524			       nvme_tcp_queue_id(queue), ctype);
1525			goto free_icresp;
1526		}
1527	}
1528	ret = -EINVAL;
1529	if (icresp->hdr.type != nvme_tcp_icresp) {
1530		pr_err("queue %d: bad type returned %d\n",
1531			nvme_tcp_queue_id(queue), icresp->hdr.type);
1532		goto free_icresp;
1533	}
1534
1535	if (le32_to_cpu(icresp->hdr.plen) != sizeof(*icresp)) {
1536		pr_err("queue %d: bad pdu length returned %d\n",
1537			nvme_tcp_queue_id(queue), icresp->hdr.plen);
1538		goto free_icresp;
1539	}
1540
1541	if (icresp->pfv != NVME_TCP_PFV_1_0) {
1542		pr_err("queue %d: bad pfv returned %d\n",
1543			nvme_tcp_queue_id(queue), icresp->pfv);
1544		goto free_icresp;
1545	}
1546
1547	ctrl_ddgst = !!(icresp->digest & NVME_TCP_DATA_DIGEST_ENABLE);
1548	if ((queue->data_digest && !ctrl_ddgst) ||
1549	    (!queue->data_digest && ctrl_ddgst)) {
1550		pr_err("queue %d: data digest mismatch host: %s ctrl: %s\n",
1551			nvme_tcp_queue_id(queue),
1552			queue->data_digest ? "enabled" : "disabled",
1553			ctrl_ddgst ? "enabled" : "disabled");
1554		goto free_icresp;
1555	}
1556
1557	ctrl_hdgst = !!(icresp->digest & NVME_TCP_HDR_DIGEST_ENABLE);
1558	if ((queue->hdr_digest && !ctrl_hdgst) ||
1559	    (!queue->hdr_digest && ctrl_hdgst)) {
1560		pr_err("queue %d: header digest mismatch host: %s ctrl: %s\n",
1561			nvme_tcp_queue_id(queue),
1562			queue->hdr_digest ? "enabled" : "disabled",
1563			ctrl_hdgst ? "enabled" : "disabled");
1564		goto free_icresp;
1565	}
1566
1567	if (icresp->cpda != 0) {
1568		pr_err("queue %d: unsupported cpda returned %d\n",
1569			nvme_tcp_queue_id(queue), icresp->cpda);
1570		goto free_icresp;
1571	}
1572
1573	maxh2cdata = le32_to_cpu(icresp->maxdata);
1574	if ((maxh2cdata % 4) || (maxh2cdata < NVME_TCP_MIN_MAXH2CDATA)) {
1575		pr_err("queue %d: invalid maxh2cdata returned %u\n",
1576		       nvme_tcp_queue_id(queue), maxh2cdata);
1577		goto free_icresp;
1578	}
1579	queue->maxh2cdata = maxh2cdata;
1580
1581	ret = 0;
1582free_icresp:
1583	kfree(icresp);
1584free_icreq:
1585	kfree(icreq);
1586	return ret;
1587}
1588
1589static bool nvme_tcp_admin_queue(struct nvme_tcp_queue *queue)
1590{
1591	return nvme_tcp_queue_id(queue) == 0;
1592}
1593
1594static bool nvme_tcp_default_queue(struct nvme_tcp_queue *queue)
1595{
1596	struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1597	int qid = nvme_tcp_queue_id(queue);
1598
1599	return !nvme_tcp_admin_queue(queue) &&
1600		qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT];
1601}
1602
1603static bool nvme_tcp_read_queue(struct nvme_tcp_queue *queue)
1604{
1605	struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1606	int qid = nvme_tcp_queue_id(queue);
1607
1608	return !nvme_tcp_admin_queue(queue) &&
1609		!nvme_tcp_default_queue(queue) &&
1610		qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT] +
1611			  ctrl->io_queues[HCTX_TYPE_READ];
1612}
1613
1614static bool nvme_tcp_poll_queue(struct nvme_tcp_queue *queue)
1615{
1616	struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1617	int qid = nvme_tcp_queue_id(queue);
1618
1619	return !nvme_tcp_admin_queue(queue) &&
1620		!nvme_tcp_default_queue(queue) &&
1621		!nvme_tcp_read_queue(queue) &&
1622		qid < 1 + ctrl->io_queues[HCTX_TYPE_DEFAULT] +
1623			  ctrl->io_queues[HCTX_TYPE_READ] +
1624			  ctrl->io_queues[HCTX_TYPE_POLL];
1625}
1626
1627/*
1628 * Track the number of queues assigned to each cpu using a global per-cpu
1629 * counter and select the least used cpu from the mq_map. Our goal is to spread
1630 * different controllers I/O threads across different cpu cores.
1631 *
1632 * Note that the accounting is not 100% perfect, but we don't need to be, we're
1633 * simply putting our best effort to select the best candidate cpu core that we
1634 * find at any given point.
1635 */
1636static void nvme_tcp_set_queue_io_cpu(struct nvme_tcp_queue *queue)
1637{
1638	struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1639	struct blk_mq_tag_set *set = &ctrl->tag_set;
1640	int qid = nvme_tcp_queue_id(queue) - 1;
1641	unsigned int *mq_map = NULL;
1642	int cpu, min_queues = INT_MAX, io_cpu;
1643
1644	if (wq_unbound)
1645		goto out;
1646
1647	if (nvme_tcp_default_queue(queue))
1648		mq_map = set->map[HCTX_TYPE_DEFAULT].mq_map;
1649	else if (nvme_tcp_read_queue(queue))
1650		mq_map = set->map[HCTX_TYPE_READ].mq_map;
1651	else if (nvme_tcp_poll_queue(queue))
1652		mq_map = set->map[HCTX_TYPE_POLL].mq_map;
1653
1654	if (WARN_ON(!mq_map))
1655		goto out;
1656
1657	/* Search for the least used cpu from the mq_map */
1658	io_cpu = WORK_CPU_UNBOUND;
1659	for_each_online_cpu(cpu) {
1660		int num_queues = atomic_read(&nvme_tcp_cpu_queues[cpu]);
1661
1662		if (mq_map[cpu] != qid)
1663			continue;
1664		if (num_queues < min_queues) {
1665			io_cpu = cpu;
1666			min_queues = num_queues;
1667		}
1668	}
1669	if (io_cpu != WORK_CPU_UNBOUND) {
1670		queue->io_cpu = io_cpu;
1671		atomic_inc(&nvme_tcp_cpu_queues[io_cpu]);
1672		set_bit(NVME_TCP_Q_IO_CPU_SET, &queue->flags);
1673	}
1674out:
1675	dev_dbg(ctrl->ctrl.device, "queue %d: using cpu %d\n",
1676		qid, queue->io_cpu);
1677}
1678
1679static void nvme_tcp_tls_done(void *data, int status, key_serial_t pskid)
1680{
1681	struct nvme_tcp_queue *queue = data;
1682	struct nvme_tcp_ctrl *ctrl = queue->ctrl;
1683	int qid = nvme_tcp_queue_id(queue);
1684	struct key *tls_key;
1685
1686	dev_dbg(ctrl->ctrl.device, "queue %d: TLS handshake done, key %x, status %d\n",
1687		qid, pskid, status);
1688
1689	if (status) {
1690		queue->tls_err = -status;
1691		goto out_complete;
1692	}
1693
1694	tls_key = nvme_tls_key_lookup(pskid);
1695	if (IS_ERR(tls_key)) {
1696		dev_warn(ctrl->ctrl.device, "queue %d: Invalid key %x\n",
1697			 qid, pskid);
1698		queue->tls_err = -ENOKEY;
1699	} else {
1700		queue->tls_enabled = true;
1701		if (qid == 0)
1702			ctrl->ctrl.tls_pskid = key_serial(tls_key);
1703		key_put(tls_key);
1704		queue->tls_err = 0;
1705	}
1706
1707out_complete:
1708	complete(&queue->tls_complete);
1709}
1710
1711static int nvme_tcp_start_tls(struct nvme_ctrl *nctrl,
1712			      struct nvme_tcp_queue *queue,
1713			      key_serial_t pskid)
1714{
1715	int qid = nvme_tcp_queue_id(queue);
1716	int ret;
1717	struct tls_handshake_args args;
1718	unsigned long tmo = tls_handshake_timeout * HZ;
1719	key_serial_t keyring = nvme_keyring_id();
1720
1721	dev_dbg(nctrl->device, "queue %d: start TLS with key %x\n",
1722		qid, pskid);
1723	memset(&args, 0, sizeof(args));
1724	args.ta_sock = queue->sock;
1725	args.ta_done = nvme_tcp_tls_done;
1726	args.ta_data = queue;
1727	args.ta_my_peerids[0] = pskid;
1728	args.ta_num_peerids = 1;
1729	if (nctrl->opts->keyring)
1730		keyring = key_serial(nctrl->opts->keyring);
1731	args.ta_keyring = keyring;
1732	args.ta_timeout_ms = tls_handshake_timeout * 1000;
1733	queue->tls_err = -EOPNOTSUPP;
1734	init_completion(&queue->tls_complete);
1735	ret = tls_client_hello_psk(&args, GFP_KERNEL);
1736	if (ret) {
1737		dev_err(nctrl->device, "queue %d: failed to start TLS: %d\n",
1738			qid, ret);
1739		return ret;
1740	}
1741	ret = wait_for_completion_interruptible_timeout(&queue->tls_complete, tmo);
1742	if (ret <= 0) {
1743		if (ret == 0)
1744			ret = -ETIMEDOUT;
1745
1746		dev_err(nctrl->device,
1747			"queue %d: TLS handshake failed, error %d\n",
1748			qid, ret);
1749		tls_handshake_cancel(queue->sock->sk);
1750	} else {
1751		if (queue->tls_err) {
1752			dev_err(nctrl->device,
1753				"queue %d: TLS handshake complete, error %d\n",
1754				qid, queue->tls_err);
1755		} else {
1756			dev_dbg(nctrl->device,
1757				"queue %d: TLS handshake complete\n", qid);
1758		}
1759		ret = queue->tls_err;
1760	}
1761	return ret;
1762}
1763
1764static int nvme_tcp_alloc_queue(struct nvme_ctrl *nctrl, int qid,
1765				key_serial_t pskid)
1766{
1767	struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1768	struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1769	int ret, rcv_pdu_size;
1770	struct file *sock_file;
1771
1772	mutex_init(&queue->queue_lock);
1773	queue->ctrl = ctrl;
1774	init_llist_head(&queue->req_list);
1775	INIT_LIST_HEAD(&queue->send_list);
1776	mutex_init(&queue->send_mutex);
1777	INIT_WORK(&queue->io_work, nvme_tcp_io_work);
1778
1779	if (qid > 0)
1780		queue->cmnd_capsule_len = nctrl->ioccsz * 16;
1781	else
1782		queue->cmnd_capsule_len = sizeof(struct nvme_command) +
1783						NVME_TCP_ADMIN_CCSZ;
1784
1785	ret = sock_create_kern(current->nsproxy->net_ns,
1786			ctrl->addr.ss_family, SOCK_STREAM,
1787			IPPROTO_TCP, &queue->sock);
1788	if (ret) {
1789		dev_err(nctrl->device,
1790			"failed to create socket: %d\n", ret);
1791		goto err_destroy_mutex;
1792	}
1793
1794	sock_file = sock_alloc_file(queue->sock, O_CLOEXEC, NULL);
1795	if (IS_ERR(sock_file)) {
1796		ret = PTR_ERR(sock_file);
1797		goto err_destroy_mutex;
1798	}
1799
1800	sk_net_refcnt_upgrade(queue->sock->sk);
1801	nvme_tcp_reclassify_socket(queue->sock);
1802
1803	/* Single syn retry */
1804	tcp_sock_set_syncnt(queue->sock->sk, 1);
1805
1806	/* Set TCP no delay */
1807	tcp_sock_set_nodelay(queue->sock->sk);
1808
1809	/*
1810	 * Cleanup whatever is sitting in the TCP transmit queue on socket
1811	 * close. This is done to prevent stale data from being sent should
1812	 * the network connection be restored before TCP times out.
1813	 */
1814	sock_no_linger(queue->sock->sk);
1815
1816	if (so_priority > 0)
1817		sock_set_priority(queue->sock->sk, so_priority);
1818
1819	/* Set socket type of service */
1820	if (nctrl->opts->tos >= 0)
1821		ip_sock_set_tos(queue->sock->sk, nctrl->opts->tos);
1822
1823	/* Set 10 seconds timeout for icresp recvmsg */
1824	queue->sock->sk->sk_rcvtimeo = 10 * HZ;
1825
1826	queue->sock->sk->sk_allocation = GFP_ATOMIC;
1827	queue->sock->sk->sk_use_task_frag = false;
1828	queue->io_cpu = WORK_CPU_UNBOUND;
1829	queue->request = NULL;
1830	queue->data_remaining = 0;
1831	queue->ddgst_remaining = 0;
1832	queue->pdu_remaining = 0;
1833	queue->pdu_offset = 0;
1834	sk_set_memalloc(queue->sock->sk);
1835
1836	if (nctrl->opts->mask & NVMF_OPT_HOST_TRADDR) {
1837		ret = kernel_bind(queue->sock, (struct sockaddr_unsized *)&ctrl->src_addr,
1838			sizeof(ctrl->src_addr));
1839		if (ret) {
1840			dev_err(nctrl->device,
1841				"failed to bind queue %d socket %d\n",
1842				qid, ret);
1843			goto err_sock;
1844		}
1845	}
1846
1847	if (nctrl->opts->mask & NVMF_OPT_HOST_IFACE) {
1848		char *iface = nctrl->opts->host_iface;
1849		sockptr_t optval = KERNEL_SOCKPTR(iface);
1850
1851		ret = sock_setsockopt(queue->sock, SOL_SOCKET, SO_BINDTODEVICE,
1852				      optval, strlen(iface));
1853		if (ret) {
1854			dev_err(nctrl->device,
1855			  "failed to bind to interface %s queue %d err %d\n",
1856			  iface, qid, ret);
1857			goto err_sock;
1858		}
1859	}
1860
1861	queue->hdr_digest = nctrl->opts->hdr_digest;
1862	queue->data_digest = nctrl->opts->data_digest;
1863
1864	rcv_pdu_size = sizeof(struct nvme_tcp_rsp_pdu) +
1865			nvme_tcp_hdgst_len(queue);
1866	queue->pdu = kmalloc(rcv_pdu_size, GFP_KERNEL);
1867	if (!queue->pdu) {
1868		ret = -ENOMEM;
1869		goto err_sock;
1870	}
1871
1872	dev_dbg(nctrl->device, "connecting queue %d\n",
1873			nvme_tcp_queue_id(queue));
1874
1875	ret = kernel_connect(queue->sock, (struct sockaddr_unsized *)&ctrl->addr,
1876		sizeof(ctrl->addr), 0);
1877	if (ret) {
1878		dev_err(nctrl->device,
1879			"failed to connect socket: %d\n", ret);
1880		goto err_rcv_pdu;
1881	}
1882
1883	/* If PSKs are configured try to start TLS */
1884	if (nvme_tcp_tls_configured(nctrl) && pskid) {
1885		ret = nvme_tcp_start_tls(nctrl, queue, pskid);
1886		if (ret)
1887			goto err_init_connect;
1888	}
1889
1890	ret = nvme_tcp_init_connection(queue);
1891	if (ret)
1892		goto err_init_connect;
1893
1894	set_bit(NVME_TCP_Q_ALLOCATED, &queue->flags);
1895
1896	return 0;
1897
1898err_init_connect:
1899	kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1900err_rcv_pdu:
1901	kfree(queue->pdu);
1902err_sock:
1903	/* ->sock will be released by fput() */
1904	fput(queue->sock->file);
1905	queue->sock = NULL;
1906err_destroy_mutex:
1907	mutex_destroy(&queue->send_mutex);
1908	mutex_destroy(&queue->queue_lock);
1909	return ret;
1910}
1911
1912static void nvme_tcp_restore_sock_ops(struct nvme_tcp_queue *queue)
1913{
1914	struct socket *sock = queue->sock;
1915
1916	write_lock_bh(&sock->sk->sk_callback_lock);
1917	sock->sk->sk_user_data  = NULL;
1918	sock->sk->sk_data_ready = queue->data_ready;
1919	sock->sk->sk_state_change = queue->state_change;
1920	sock->sk->sk_write_space  = queue->write_space;
1921	write_unlock_bh(&sock->sk->sk_callback_lock);
1922}
1923
1924static void __nvme_tcp_stop_queue(struct nvme_tcp_queue *queue)
1925{
1926	kernel_sock_shutdown(queue->sock, SHUT_RDWR);
1927	nvme_tcp_restore_sock_ops(queue);
1928	cancel_work_sync(&queue->io_work);
1929}
1930
1931static void nvme_tcp_stop_queue_nowait(struct nvme_ctrl *nctrl, int qid)
1932{
1933	struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1934	struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1935
1936	if (!test_bit(NVME_TCP_Q_ALLOCATED, &queue->flags))
1937		return;
1938
1939	if (test_and_clear_bit(NVME_TCP_Q_IO_CPU_SET, &queue->flags))
1940		atomic_dec(&nvme_tcp_cpu_queues[queue->io_cpu]);
1941
1942	mutex_lock(&queue->queue_lock);
1943	if (test_and_clear_bit(NVME_TCP_Q_LIVE, &queue->flags))
1944		__nvme_tcp_stop_queue(queue);
1945	/* Stopping the queue will disable TLS */
1946	queue->tls_enabled = false;
1947	mutex_unlock(&queue->queue_lock);
1948}
1949
1950static void nvme_tcp_wait_queue(struct nvme_ctrl *nctrl, int qid)
1951{
1952	struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1953	struct nvme_tcp_queue *queue = &ctrl->queues[qid];
1954	int timeout = 100;
1955
1956	while (timeout > 0) {
1957		if (!test_bit(NVME_TCP_Q_ALLOCATED, &queue->flags) ||
1958		    !sk_wmem_alloc_get(queue->sock->sk))
1959			return;
1960		msleep(2);
1961		timeout -= 2;
1962	}
1963	dev_warn(nctrl->device,
1964		 "qid %d: timeout draining sock wmem allocation expired\n",
1965		 qid);
1966}
1967
1968static void nvme_tcp_stop_queue(struct nvme_ctrl *nctrl, int qid)
1969{
1970	nvme_tcp_stop_queue_nowait(nctrl, qid);
1971	nvme_tcp_wait_queue(nctrl, qid);
1972}
1973
1974
1975static void nvme_tcp_setup_sock_ops(struct nvme_tcp_queue *queue)
1976{
1977	write_lock_bh(&queue->sock->sk->sk_callback_lock);
1978	queue->sock->sk->sk_user_data = queue;
1979	queue->state_change = queue->sock->sk->sk_state_change;
1980	queue->data_ready = queue->sock->sk->sk_data_ready;
1981	queue->write_space = queue->sock->sk->sk_write_space;
1982	queue->sock->sk->sk_data_ready = nvme_tcp_data_ready;
1983	queue->sock->sk->sk_state_change = nvme_tcp_state_change;
1984	queue->sock->sk->sk_write_space = nvme_tcp_write_space;
1985#ifdef CONFIG_NET_RX_BUSY_POLL
1986	queue->sock->sk->sk_ll_usec = 1;
1987#endif
1988	write_unlock_bh(&queue->sock->sk->sk_callback_lock);
1989}
1990
1991static int nvme_tcp_start_queue(struct nvme_ctrl *nctrl, int idx)
1992{
1993	struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
1994	struct nvme_tcp_queue *queue = &ctrl->queues[idx];
1995	int ret;
1996
1997	queue->rd_enabled = true;
1998	nvme_tcp_init_recv_ctx(queue);
1999	nvme_tcp_setup_sock_ops(queue);
2000
2001	if (idx) {
2002		nvme_tcp_set_queue_io_cpu(queue);
2003		ret = nvmf_connect_io_queue(nctrl, idx);
2004	} else
2005		ret = nvmf_connect_admin_queue(nctrl);
2006
2007	if (!ret) {
2008		set_bit(NVME_TCP_Q_LIVE, &queue->flags);
2009	} else {
2010		if (test_bit(NVME_TCP_Q_ALLOCATED, &queue->flags))
2011			__nvme_tcp_stop_queue(queue);
2012		dev_err(nctrl->device,
2013			"failed to connect queue: %d ret=%d\n", idx, ret);
2014	}
2015	return ret;
2016}
2017
2018static void nvme_tcp_free_admin_queue(struct nvme_ctrl *ctrl)
2019{
2020	if (to_tcp_ctrl(ctrl)->async_req.pdu) {
2021		cancel_work_sync(&ctrl->async_event_work);
2022		nvme_tcp_free_async_req(to_tcp_ctrl(ctrl));
2023		to_tcp_ctrl(ctrl)->async_req.pdu = NULL;
2024	}
2025
2026	nvme_tcp_free_queue(ctrl, 0);
2027}
2028
2029static void nvme_tcp_free_io_queues(struct nvme_ctrl *ctrl)
2030{
2031	int i;
2032
2033	for (i = 1; i < ctrl->queue_count; i++)
2034		nvme_tcp_free_queue(ctrl, i);
2035}
2036
2037static void nvme_tcp_stop_io_queues(struct nvme_ctrl *ctrl)
2038{
2039	int i;
2040
2041	for (i = 1; i < ctrl->queue_count; i++)
2042		nvme_tcp_stop_queue_nowait(ctrl, i);
2043	for (i = 1; i < ctrl->queue_count; i++)
2044		nvme_tcp_wait_queue(ctrl, i);
2045}
2046
2047static int nvme_tcp_start_io_queues(struct nvme_ctrl *ctrl,
2048				    int first, int last)
2049{
2050	int i, ret;
2051
2052	for (i = first; i < last; i++) {
2053		ret = nvme_tcp_start_queue(ctrl, i);
2054		if (ret)
2055			goto out_stop_queues;
2056	}
2057
2058	return 0;
2059
2060out_stop_queues:
2061	for (i--; i >= first; i--)
2062		nvme_tcp_stop_queue(ctrl, i);
2063	return ret;
2064}
2065
2066static int nvme_tcp_alloc_admin_queue(struct nvme_ctrl *ctrl)
2067{
2068	int ret;
2069	key_serial_t pskid = 0;
2070
2071	if (nvme_tcp_tls_configured(ctrl)) {
2072		if (ctrl->opts->tls_key)
2073			pskid = key_serial(ctrl->opts->tls_key);
2074		else if (ctrl->opts->tls) {
2075			pskid = nvme_tls_psk_default(ctrl->opts->keyring,
2076						      ctrl->opts->host->nqn,
2077						      ctrl->opts->subsysnqn);
2078			if (!pskid) {
2079				dev_err(ctrl->device, "no valid PSK found\n");
2080				return -ENOKEY;
2081			}
2082		}
2083	}
2084
2085	ret = nvme_tcp_alloc_queue(ctrl, 0, pskid);
2086	if (ret)
2087		return ret;
2088
2089	ret = nvme_tcp_alloc_async_req(to_tcp_ctrl(ctrl));
2090	if (ret)
2091		goto out_free_queue;
2092
2093	return 0;
2094
2095out_free_queue:
2096	nvme_tcp_free_queue(ctrl, 0);
2097	return ret;
2098}
2099
2100static int __nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
2101{
2102	int i, ret;
2103
2104	if (nvme_tcp_tls_configured(ctrl)) {
2105		if (ctrl->opts->concat) {
2106			/*
2107			 * The generated PSK is stored in the
2108			 * fabric options
2109			 */
2110			if (!ctrl->opts->tls_key) {
2111				dev_err(ctrl->device, "no PSK generated\n");
2112				return -ENOKEY;
2113			}
2114			if (ctrl->tls_pskid &&
2115			    ctrl->tls_pskid != key_serial(ctrl->opts->tls_key)) {
2116				dev_err(ctrl->device, "Stale PSK id %08x\n", ctrl->tls_pskid);
2117				ctrl->tls_pskid = 0;
2118			}
2119		} else if (!ctrl->tls_pskid) {
2120			dev_err(ctrl->device, "no PSK negotiated\n");
2121			return -ENOKEY;
2122		}
2123	}
2124
2125	for (i = 1; i < ctrl->queue_count; i++) {
2126		ret = nvme_tcp_alloc_queue(ctrl, i,
2127				ctrl->tls_pskid);
2128		if (ret)
2129			goto out_free_queues;
2130	}
2131
2132	return 0;
2133
2134out_free_queues:
2135	for (i--; i >= 1; i--)
2136		nvme_tcp_free_queue(ctrl, i);
2137
2138	return ret;
2139}
2140
2141static int nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
2142{
2143	unsigned int nr_io_queues;
2144	int ret;
2145
2146	nr_io_queues = nvmf_nr_io_queues(ctrl->opts);
2147	ret = nvme_set_queue_count(ctrl, &nr_io_queues);
2148	if (ret)
2149		return ret;
2150
2151	if (nr_io_queues == 0) {
2152		dev_err(ctrl->device,
2153			"unable to set any I/O queues\n");
2154		return -ENOMEM;
2155	}
2156
2157	ctrl->queue_count = nr_io_queues + 1;
2158	dev_info(ctrl->device,
2159		"creating %d I/O queues.\n", nr_io_queues);
2160
2161	nvmf_set_io_queues(ctrl->opts, nr_io_queues,
2162			   to_tcp_ctrl(ctrl)->io_queues);
2163	return __nvme_tcp_alloc_io_queues(ctrl);
2164}
2165
2166static int nvme_tcp_configure_io_queues(struct nvme_ctrl *ctrl, bool new)
2167{
2168	int ret, nr_queues;
2169
2170	ret = nvme_tcp_alloc_io_queues(ctrl);
2171	if (ret)
2172		return ret;
2173
2174	if (new) {
2175		ret = nvme_alloc_io_tag_set(ctrl, &to_tcp_ctrl(ctrl)->tag_set,
2176				&nvme_tcp_mq_ops,
2177				ctrl->opts->nr_poll_queues ? HCTX_MAX_TYPES : 2,
2178				sizeof(struct nvme_tcp_request));
2179		if (ret)
2180			goto out_free_io_queues;
2181	}
2182
2183	/*
2184	 * Only start IO queues for which we have allocated the tagset
2185	 * and limited it to the available queues. On reconnects, the
2186	 * queue number might have changed.
2187	 */
2188	nr_queues = min(ctrl->tagset->nr_hw_queues + 1, ctrl->queue_count);
2189	ret = nvme_tcp_start_io_queues(ctrl, 1, nr_queues);
2190	if (ret)
2191		goto out_cleanup_connect_q;
2192
2193	if (!new) {
2194		nvme_start_freeze(ctrl);
2195		nvme_unquiesce_io_queues(ctrl);
2196		if (!nvme_wait_freeze_timeout(ctrl, NVME_IO_TIMEOUT)) {
2197			/*
2198			 * If we timed out waiting for freeze we are likely to
2199			 * be stuck.  Fail the controller initialization just
2200			 * to be safe.
2201			 */
2202			ret = -ENODEV;
2203			nvme_unfreeze(ctrl);
2204			goto out_wait_freeze_timed_out;
2205		}
2206		blk_mq_update_nr_hw_queues(ctrl->tagset,
2207			ctrl->queue_count - 1);
2208		nvme_unfreeze(ctrl);
2209	}
2210
2211	/*
2212	 * If the number of queues has increased (reconnect case)
2213	 * start all new queues now.
2214	 */
2215	ret = nvme_tcp_start_io_queues(ctrl, nr_queues,
2216				       ctrl->tagset->nr_hw_queues + 1);
2217	if (ret)
2218		goto out_wait_freeze_timed_out;
2219
2220	return 0;
2221
2222out_wait_freeze_timed_out:
2223	nvme_quiesce_io_queues(ctrl);
2224	nvme_sync_io_queues(ctrl);
2225	nvme_tcp_stop_io_queues(ctrl);
2226out_cleanup_connect_q:
2227	nvme_cancel_tagset(ctrl);
2228	if (new)
2229		nvme_remove_io_tag_set(ctrl);
2230out_free_io_queues:
2231	nvme_tcp_free_io_queues(ctrl);
2232	return ret;
2233}
2234
2235static int nvme_tcp_configure_admin_queue(struct nvme_ctrl *ctrl, bool new)
2236{
2237	int error;
2238
2239	error = nvme_tcp_alloc_admin_queue(ctrl);
2240	if (error)
2241		return error;
2242
2243	if (new) {
2244		error = nvme_alloc_admin_tag_set(ctrl,
2245				&to_tcp_ctrl(ctrl)->admin_tag_set,
2246				&nvme_tcp_admin_mq_ops,
2247				sizeof(struct nvme_tcp_request));
2248		if (error)
2249			goto out_free_queue;
2250	}
2251
2252	error = nvme_tcp_start_queue(ctrl, 0);
2253	if (error)
2254		goto out_cleanup_tagset;
2255
2256	if (ctrl->opts->concat && !ctrl->tls_pskid)
2257		return 0;
2258
2259	error = nvme_enable_ctrl(ctrl);
2260	if (error)
2261		goto out_stop_queue;
2262
2263	nvme_unquiesce_admin_queue(ctrl);
2264
2265	error = nvme_init_ctrl_finish(ctrl, false);
2266	if (error)
2267		goto out_quiesce_queue;
2268
2269	return 0;
2270
2271out_quiesce_queue:
2272	nvme_quiesce_admin_queue(ctrl);
2273	blk_sync_queue(ctrl->admin_q);
2274out_stop_queue:
2275	nvme_tcp_stop_queue(ctrl, 0);
2276	nvme_cancel_admin_tagset(ctrl);
2277out_cleanup_tagset:
2278	if (new)
2279		nvme_remove_admin_tag_set(ctrl);
2280out_free_queue:
2281	nvme_tcp_free_admin_queue(ctrl);
2282	return error;
2283}
2284
2285static void nvme_tcp_teardown_admin_queue(struct nvme_ctrl *ctrl,
2286		bool remove)
2287{
2288	nvme_quiesce_admin_queue(ctrl);
2289	blk_sync_queue(ctrl->admin_q);
2290	nvme_tcp_stop_queue(ctrl, 0);
2291	nvme_cancel_admin_tagset(ctrl);
2292	if (remove) {
2293		nvme_unquiesce_admin_queue(ctrl);
2294		nvme_remove_admin_tag_set(ctrl);
2295	}
2296	nvme_tcp_free_admin_queue(ctrl);
2297	if (ctrl->tls_pskid) {
2298		dev_dbg(ctrl->device, "Wipe negotiated TLS_PSK %08x\n",
2299			ctrl->tls_pskid);
2300		ctrl->tls_pskid = 0;
2301	}
2302}
2303
2304static void nvme_tcp_teardown_io_queues(struct nvme_ctrl *ctrl,
2305		bool remove)
2306{
2307	if (ctrl->queue_count <= 1)
2308		return;
2309	nvme_quiesce_io_queues(ctrl);
2310	nvme_sync_io_queues(ctrl);
2311	nvme_tcp_stop_io_queues(ctrl);
2312	nvme_cancel_tagset(ctrl);
2313	if (remove) {
2314		nvme_unquiesce_io_queues(ctrl);
2315		nvme_remove_io_tag_set(ctrl);
2316	}
2317	nvme_tcp_free_io_queues(ctrl);
2318}
2319
2320static void nvme_tcp_reconnect_or_remove(struct nvme_ctrl *ctrl,
2321		int status)
2322{
2323	enum nvme_ctrl_state state = nvme_ctrl_state(ctrl);
2324
2325	/* If we are resetting/deleting then do nothing */
2326	if (state != NVME_CTRL_CONNECTING) {
2327		WARN_ON_ONCE(state == NVME_CTRL_NEW || state == NVME_CTRL_LIVE);
2328		return;
2329	}
2330
2331	if (nvmf_should_reconnect(ctrl, status)) {
2332		dev_info(ctrl->device, "Reconnecting in %d seconds...\n",
2333			ctrl->opts->reconnect_delay);
2334		queue_delayed_work(nvme_wq, &to_tcp_ctrl(ctrl)->connect_work,
2335				ctrl->opts->reconnect_delay * HZ);
2336	} else {
2337		dev_info(ctrl->device, "Removing controller (%d)...\n",
2338			 status);
2339		nvme_delete_ctrl(ctrl);
2340	}
2341}
2342
2343/*
2344 * The TLS key is set by secure concatenation after negotiation has been
2345 * completed on the admin queue. We need to revoke the key when:
2346 * - concatenation is enabled (otherwise it's a static key set by the user)
2347 * and
2348 * - the generated key is present in ctrl->tls_key (otherwise there's nothing
2349 *   to revoke)
2350 * and
2351 * - a valid PSK key ID has been set in ctrl->tls_pskid (otherwise TLS
2352 *   negotiation has not run).
2353 *
2354 * We cannot always revoke the key as nvme_tcp_alloc_admin_queue() is called
2355 * twice during secure concatenation, once on a 'normal' connection to run the
2356 * DH-HMAC-CHAP negotiation (which generates the key, so it _must not_ be set),
2357 * and once after the negotiation (which uses the key, so it _must_ be set).
2358 */
2359static bool nvme_tcp_key_revoke_needed(struct nvme_ctrl *ctrl)
2360{
2361	return ctrl->opts->concat && ctrl->opts->tls_key && ctrl->tls_pskid;
2362}
2363
2364static int nvme_tcp_setup_ctrl(struct nvme_ctrl *ctrl, bool new)
2365{
2366	struct nvmf_ctrl_options *opts = ctrl->opts;
2367	int ret;
2368
2369	ret = nvme_tcp_configure_admin_queue(ctrl, new);
2370	if (ret)
2371		return ret;
2372
2373	if (ctrl->opts->concat && !ctrl->tls_pskid) {
2374		/* See comments for nvme_tcp_key_revoke_needed() */
2375		dev_dbg(ctrl->device, "restart admin queue for secure concatenation\n");
2376		nvme_stop_keep_alive(ctrl);
2377		nvme_tcp_teardown_admin_queue(ctrl, false);
2378		ret = nvme_tcp_configure_admin_queue(ctrl, false);
2379		if (ret)
2380			goto destroy_admin;
2381	}
2382
2383	if (ctrl->icdoff) {
2384		ret = -EOPNOTSUPP;
2385		dev_err(ctrl->device, "icdoff is not supported!\n");
2386		goto destroy_admin;
2387	}
2388
2389	if (!nvme_ctrl_sgl_supported(ctrl)) {
2390		ret = -EOPNOTSUPP;
2391		dev_err(ctrl->device, "Mandatory sgls are not supported!\n");
2392		goto destroy_admin;
2393	}
2394
2395	if (opts->queue_size > ctrl->sqsize + 1)
2396		dev_warn(ctrl->device,
2397			"queue_size %zu > ctrl sqsize %u, clamping down\n",
2398			opts->queue_size, ctrl->sqsize + 1);
2399
2400	if (ctrl->sqsize + 1 > ctrl->maxcmd) {
2401		dev_warn(ctrl->device,
2402			"sqsize %u > ctrl maxcmd %u, clamping down\n",
2403			ctrl->sqsize + 1, ctrl->maxcmd);
2404		ctrl->sqsize = ctrl->maxcmd - 1;
2405	}
2406
2407	if (ctrl->queue_count > 1) {
2408		ret = nvme_tcp_configure_io_queues(ctrl, new);
2409		if (ret)
2410			goto destroy_admin;
2411	}
2412
2413	if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_LIVE)) {
2414		/*
2415		 * state change failure is ok if we started ctrl delete,
2416		 * unless we're during creation of a new controller to
2417		 * avoid races with teardown flow.
2418		 */
2419		enum nvme_ctrl_state state = nvme_ctrl_state(ctrl);
2420
2421		WARN_ON_ONCE(state != NVME_CTRL_DELETING &&
2422			     state != NVME_CTRL_DELETING_NOIO);
2423		WARN_ON_ONCE(new);
2424		ret = -EINVAL;
2425		goto destroy_io;
2426	}
2427
2428	nvme_start_ctrl(ctrl);
2429	return 0;
2430
2431destroy_io:
2432	if (ctrl->queue_count > 1) {
2433		nvme_quiesce_io_queues(ctrl);
2434		nvme_sync_io_queues(ctrl);
2435		nvme_tcp_stop_io_queues(ctrl);
2436		nvme_cancel_tagset(ctrl);
2437		if (new)
2438			nvme_remove_io_tag_set(ctrl);
2439		nvme_tcp_free_io_queues(ctrl);
2440	}
2441destroy_admin:
2442	nvme_stop_keep_alive(ctrl);
2443	nvme_tcp_teardown_admin_queue(ctrl, new);
2444	return ret;
2445}
2446
2447static void nvme_tcp_reconnect_ctrl_work(struct work_struct *work)
2448{
2449	struct nvme_tcp_ctrl *tcp_ctrl = container_of(to_delayed_work(work),
2450			struct nvme_tcp_ctrl, connect_work);
2451	struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
2452	int ret;
2453
2454	++ctrl->nr_reconnects;
2455
2456	ret = nvme_tcp_setup_ctrl(ctrl, false);
2457	if (ret)
2458		goto requeue;
2459
2460	dev_info(ctrl->device, "Successfully reconnected (attempt %d/%d)\n",
2461		 ctrl->nr_reconnects, ctrl->opts->max_reconnects);
2462
2463	ctrl->nr_reconnects = 0;
2464
2465	return;
2466
2467requeue:
2468	dev_info(ctrl->device, "Failed reconnect attempt %d/%d\n",
2469		 ctrl->nr_reconnects, ctrl->opts->max_reconnects);
2470	nvme_tcp_reconnect_or_remove(ctrl, ret);
2471}
2472
2473static void nvme_tcp_error_recovery_work(struct work_struct *work)
2474{
2475	struct nvme_tcp_ctrl *tcp_ctrl = container_of(work,
2476				struct nvme_tcp_ctrl, err_work);
2477	struct nvme_ctrl *ctrl = &tcp_ctrl->ctrl;
2478
2479	if (nvme_tcp_key_revoke_needed(ctrl))
2480		nvme_auth_revoke_tls_key(ctrl);
2481	nvme_stop_keep_alive(ctrl);
2482	flush_work(&ctrl->async_event_work);
2483	nvme_tcp_teardown_io_queues(ctrl, false);
2484	/* unquiesce to fail fast pending requests */
2485	nvme_unquiesce_io_queues(ctrl);
2486	nvme_tcp_teardown_admin_queue(ctrl, false);
2487	nvme_unquiesce_admin_queue(ctrl);
2488	nvme_auth_stop(ctrl);
2489
2490	if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
2491		/* state change failure is ok if we started ctrl delete */
2492		enum nvme_ctrl_state state = nvme_ctrl_state(ctrl);
2493
2494		WARN_ON_ONCE(state != NVME_CTRL_DELETING &&
2495			     state != NVME_CTRL_DELETING_NOIO);
2496		return;
2497	}
2498
2499	nvme_tcp_reconnect_or_remove(ctrl, 0);
2500}
2501
2502static void nvme_tcp_teardown_ctrl(struct nvme_ctrl *ctrl, bool shutdown)
2503{
2504	nvme_tcp_teardown_io_queues(ctrl, shutdown);
2505	nvme_quiesce_admin_queue(ctrl);
2506	nvme_disable_ctrl(ctrl, shutdown);
2507	nvme_tcp_teardown_admin_queue(ctrl, shutdown);
2508}
2509
2510static void nvme_tcp_delete_ctrl(struct nvme_ctrl *ctrl)
2511{
2512	nvme_tcp_teardown_ctrl(ctrl, true);
2513}
2514
2515static void nvme_reset_ctrl_work(struct work_struct *work)
2516{
2517	struct nvme_ctrl *ctrl =
2518		container_of(work, struct nvme_ctrl, reset_work);
2519	int ret;
2520
2521	if (nvme_tcp_key_revoke_needed(ctrl))
2522		nvme_auth_revoke_tls_key(ctrl);
2523	nvme_stop_ctrl(ctrl);
2524	nvme_tcp_teardown_ctrl(ctrl, false);
2525
2526	if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_CONNECTING)) {
2527		/* state change failure is ok if we started ctrl delete */
2528		enum nvme_ctrl_state state = nvme_ctrl_state(ctrl);
2529
2530		WARN_ON_ONCE(state != NVME_CTRL_DELETING &&
2531			     state != NVME_CTRL_DELETING_NOIO);
2532		return;
2533	}
2534
2535	ret = nvme_tcp_setup_ctrl(ctrl, false);
2536	if (ret)
2537		goto out_fail;
2538
2539	return;
2540
2541out_fail:
2542	++ctrl->nr_reconnects;
2543	nvme_tcp_reconnect_or_remove(ctrl, ret);
2544}
2545
2546static void nvme_tcp_stop_ctrl(struct nvme_ctrl *ctrl)
2547{
2548	flush_work(&to_tcp_ctrl(ctrl)->err_work);
2549	cancel_delayed_work_sync(&to_tcp_ctrl(ctrl)->connect_work);
2550}
2551
2552static void nvme_tcp_free_ctrl(struct nvme_ctrl *nctrl)
2553{
2554	struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
2555
2556	if (list_empty(&ctrl->list))
2557		goto free_ctrl;
2558
2559	mutex_lock(&nvme_tcp_ctrl_mutex);
2560	list_del(&ctrl->list);
2561	mutex_unlock(&nvme_tcp_ctrl_mutex);
2562
2563	nvmf_free_options(nctrl->opts);
2564free_ctrl:
2565	kfree(ctrl->queues);
2566	kfree(ctrl);
2567}
2568
2569static void nvme_tcp_set_sg_null(struct nvme_command *c)
2570{
2571	struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2572
2573	sg->addr = 0;
2574	sg->length = 0;
2575	sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
2576			NVME_SGL_FMT_TRANSPORT_A;
2577}
2578
2579static void nvme_tcp_set_sg_inline(struct nvme_tcp_queue *queue,
2580		struct nvme_command *c, u32 data_len)
2581{
2582	struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2583
2584	sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
2585	sg->length = cpu_to_le32(data_len);
2586	sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
2587}
2588
2589static void nvme_tcp_set_sg_host_data(struct nvme_command *c,
2590		u32 data_len)
2591{
2592	struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
2593
2594	sg->addr = 0;
2595	sg->length = cpu_to_le32(data_len);
2596	sg->type = (NVME_TRANSPORT_SGL_DATA_DESC << 4) |
2597			NVME_SGL_FMT_TRANSPORT_A;
2598}
2599
2600static void nvme_tcp_submit_async_event(struct nvme_ctrl *arg)
2601{
2602	struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(arg);
2603	struct nvme_tcp_queue *queue = &ctrl->queues[0];
2604	struct nvme_tcp_cmd_pdu *pdu = ctrl->async_req.pdu;
2605	struct nvme_command *cmd = &pdu->cmd;
2606	u8 hdgst = nvme_tcp_hdgst_len(queue);
2607
2608	memset(pdu, 0, sizeof(*pdu));
2609	pdu->hdr.type = nvme_tcp_cmd;
2610	if (queue->hdr_digest)
2611		pdu->hdr.flags |= NVME_TCP_F_HDGST;
2612	pdu->hdr.hlen = sizeof(*pdu);
2613	pdu->hdr.plen = cpu_to_le32(pdu->hdr.hlen + hdgst);
2614
2615	cmd->common.opcode = nvme_admin_async_event;
2616	cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
2617	cmd->common.flags |= NVME_CMD_SGL_METABUF;
2618	nvme_tcp_set_sg_null(cmd);
2619
2620	ctrl->async_req.state = NVME_TCP_SEND_CMD_PDU;
2621	ctrl->async_req.offset = 0;
2622	ctrl->async_req.curr_bio = NULL;
2623	ctrl->async_req.data_len = 0;
2624	init_llist_node(&ctrl->async_req.lentry);
2625	INIT_LIST_HEAD(&ctrl->async_req.entry);
2626
2627	nvme_tcp_queue_request(&ctrl->async_req, true);
2628}
2629
2630static void nvme_tcp_complete_timed_out(struct request *rq)
2631{
2632	struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2633	struct nvme_ctrl *ctrl = &req->queue->ctrl->ctrl;
2634
2635	nvme_tcp_stop_queue(ctrl, nvme_tcp_queue_id(req->queue));
2636	nvmf_complete_timed_out_request(rq);
2637}
2638
2639static enum blk_eh_timer_return nvme_tcp_timeout(struct request *rq)
2640{
2641	struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2642	struct nvme_ctrl *ctrl = &req->queue->ctrl->ctrl;
2643	struct nvme_tcp_cmd_pdu *pdu = nvme_tcp_req_cmd_pdu(req);
2644	struct nvme_command *cmd = &pdu->cmd;
2645	int qid = nvme_tcp_queue_id(req->queue);
2646
2647	dev_warn(ctrl->device,
2648		 "I/O tag %d (%04x) type %d opcode %#x (%s) QID %d timeout\n",
2649		 rq->tag, nvme_cid(rq), pdu->hdr.type, cmd->common.opcode,
2650		 nvme_fabrics_opcode_str(qid, cmd), qid);
2651
2652	if (nvme_ctrl_state(ctrl) != NVME_CTRL_LIVE) {
2653		/*
2654		 * If we are resetting, connecting or deleting we should
2655		 * complete immediately because we may block controller
2656		 * teardown or setup sequence
2657		 * - ctrl disable/shutdown fabrics requests
2658		 * - connect requests
2659		 * - initialization admin requests
2660		 * - I/O requests that entered after unquiescing and
2661		 *   the controller stopped responding
2662		 *
2663		 * All other requests should be cancelled by the error
2664		 * recovery work, so it's fine that we fail it here.
2665		 */
2666		nvme_tcp_complete_timed_out(rq);
2667		return BLK_EH_DONE;
2668	}
2669
2670	/*
2671	 * LIVE state should trigger the normal error recovery which will
2672	 * handle completing this request.
2673	 */
2674	nvme_tcp_error_recovery(ctrl);
2675	return BLK_EH_RESET_TIMER;
2676}
2677
2678static blk_status_t nvme_tcp_map_data(struct nvme_tcp_queue *queue,
2679			struct request *rq)
2680{
2681	struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2682	struct nvme_tcp_cmd_pdu *pdu = nvme_tcp_req_cmd_pdu(req);
2683	struct nvme_command *c = &pdu->cmd;
2684
2685	c->common.flags |= NVME_CMD_SGL_METABUF;
2686
2687	if (!blk_rq_nr_phys_segments(rq))
2688		nvme_tcp_set_sg_null(c);
2689	else if (rq_data_dir(rq) == WRITE &&
2690	    req->data_len <= nvme_tcp_inline_data_size(req))
2691		nvme_tcp_set_sg_inline(queue, c, req->data_len);
2692	else
2693		nvme_tcp_set_sg_host_data(c, req->data_len);
2694
2695	return 0;
2696}
2697
2698static blk_status_t nvme_tcp_setup_cmd_pdu(struct nvme_ns *ns,
2699		struct request *rq)
2700{
2701	struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2702	struct nvme_tcp_cmd_pdu *pdu = nvme_tcp_req_cmd_pdu(req);
2703	struct nvme_tcp_queue *queue = req->queue;
2704	u8 hdgst = nvme_tcp_hdgst_len(queue), ddgst = 0;
2705	blk_status_t ret;
2706
2707	ret = nvme_setup_cmd(ns, rq);
2708	if (ret)
2709		return ret;
2710
2711	req->state = NVME_TCP_SEND_CMD_PDU;
2712	req->status = cpu_to_le16(NVME_SC_SUCCESS);
2713	req->offset = 0;
2714	req->data_sent = 0;
2715	req->pdu_len = 0;
2716	req->pdu_sent = 0;
2717	req->h2cdata_left = 0;
2718	req->data_len = blk_rq_nr_phys_segments(rq) ?
2719				blk_rq_payload_bytes(rq) : 0;
2720	req->curr_bio = rq->bio;
2721	if (req->curr_bio && req->data_len)
2722		nvme_tcp_init_iter(req, rq_data_dir(rq));
2723
2724	if (rq_data_dir(rq) == WRITE &&
2725	    req->data_len <= nvme_tcp_inline_data_size(req))
2726		req->pdu_len = req->data_len;
2727
2728	pdu->hdr.type = nvme_tcp_cmd;
2729	pdu->hdr.flags = 0;
2730	if (queue->hdr_digest)
2731		pdu->hdr.flags |= NVME_TCP_F_HDGST;
2732	if (queue->data_digest && req->pdu_len) {
2733		pdu->hdr.flags |= NVME_TCP_F_DDGST;
2734		ddgst = nvme_tcp_ddgst_len(queue);
2735	}
2736	pdu->hdr.hlen = sizeof(*pdu);
2737	pdu->hdr.pdo = req->pdu_len ? pdu->hdr.hlen + hdgst : 0;
2738	pdu->hdr.plen =
2739		cpu_to_le32(pdu->hdr.hlen + hdgst + req->pdu_len + ddgst);
2740
2741	ret = nvme_tcp_map_data(queue, rq);
2742	if (unlikely(ret)) {
2743		nvme_cleanup_cmd(rq);
2744		dev_err(queue->ctrl->ctrl.device,
2745			"Failed to map data (%d)\n", ret);
2746		return ret;
2747	}
2748
2749	return 0;
2750}
2751
2752static void nvme_tcp_commit_rqs(struct blk_mq_hw_ctx *hctx)
2753{
2754	struct nvme_tcp_queue *queue = hctx->driver_data;
2755
2756	if (!llist_empty(&queue->req_list))
2757		queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
2758}
2759
2760static blk_status_t nvme_tcp_queue_rq(struct blk_mq_hw_ctx *hctx,
2761		const struct blk_mq_queue_data *bd)
2762{
2763	struct nvme_ns *ns = hctx->queue->queuedata;
2764	struct nvme_tcp_queue *queue = hctx->driver_data;
2765	struct request *rq = bd->rq;
2766	struct nvme_tcp_request *req = blk_mq_rq_to_pdu(rq);
2767	bool queue_ready = test_bit(NVME_TCP_Q_LIVE, &queue->flags);
2768	blk_status_t ret;
2769
2770	if (!nvme_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
2771		return nvme_fail_nonready_command(&queue->ctrl->ctrl, rq);
2772
2773	ret = nvme_tcp_setup_cmd_pdu(ns, rq);
2774	if (unlikely(ret))
2775		return ret;
2776
2777	nvme_start_request(rq);
2778
2779	nvme_tcp_queue_request(req, bd->last);
2780
2781	return BLK_STS_OK;
2782}
2783
2784static void nvme_tcp_map_queues(struct blk_mq_tag_set *set)
2785{
2786	struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(set->driver_data);
2787
2788	nvmf_map_queues(set, &ctrl->ctrl, ctrl->io_queues);
2789}
2790
2791static int nvme_tcp_poll(struct blk_mq_hw_ctx *hctx, struct io_comp_batch *iob)
2792{
2793	struct nvme_tcp_queue *queue = hctx->driver_data;
2794	struct sock *sk = queue->sock->sk;
2795	int ret;
2796
2797	if (!test_bit(NVME_TCP_Q_LIVE, &queue->flags))
2798		return 0;
2799
2800	set_bit(NVME_TCP_Q_POLLING, &queue->flags);
2801	if (sk_can_busy_loop(sk) && skb_queue_empty_lockless(&sk->sk_receive_queue))
2802		sk_busy_loop(sk, true);
2803	ret = nvme_tcp_try_recv(queue);
2804	clear_bit(NVME_TCP_Q_POLLING, &queue->flags);
2805	return ret < 0 ? ret : queue->nr_cqe;
2806}
2807
2808static int nvme_tcp_get_address(struct nvme_ctrl *ctrl, char *buf, int size)
2809{
2810	struct nvme_tcp_queue *queue = &to_tcp_ctrl(ctrl)->queues[0];
2811	struct sockaddr_storage src_addr;
2812	int ret, len;
2813
2814	len = nvmf_get_address(ctrl, buf, size);
2815
2816	if (!test_bit(NVME_TCP_Q_LIVE, &queue->flags))
2817		return len;
2818
2819	mutex_lock(&queue->queue_lock);
2820
2821	ret = kernel_getsockname(queue->sock, (struct sockaddr *)&src_addr);
2822	if (ret > 0) {
2823		if (len > 0)
2824			len--; /* strip trailing newline */
2825		len += scnprintf(buf + len, size - len, "%ssrc_addr=%pISc\n",
2826				(len) ? "," : "", &src_addr);
2827	}
2828
2829	mutex_unlock(&queue->queue_lock);
2830
2831	return len;
2832}
2833
2834static const struct blk_mq_ops nvme_tcp_mq_ops = {
2835	.queue_rq	= nvme_tcp_queue_rq,
2836	.commit_rqs	= nvme_tcp_commit_rqs,
2837	.complete	= nvme_complete_rq,
2838	.init_request	= nvme_tcp_init_request,
2839	.exit_request	= nvme_tcp_exit_request,
2840	.init_hctx	= nvme_tcp_init_hctx,
2841	.timeout	= nvme_tcp_timeout,
2842	.map_queues	= nvme_tcp_map_queues,
2843	.poll		= nvme_tcp_poll,
2844};
2845
2846static const struct blk_mq_ops nvme_tcp_admin_mq_ops = {
2847	.queue_rq	= nvme_tcp_queue_rq,
2848	.complete	= nvme_complete_rq,
2849	.init_request	= nvme_tcp_init_request,
2850	.exit_request	= nvme_tcp_exit_request,
2851	.init_hctx	= nvme_tcp_init_admin_hctx,
2852	.timeout	= nvme_tcp_timeout,
2853};
2854
2855static const struct nvme_ctrl_ops nvme_tcp_ctrl_ops = {
2856	.name			= "tcp",
2857	.module			= THIS_MODULE,
2858	.flags			= NVME_F_FABRICS | NVME_F_BLOCKING,
2859	.reg_read32		= nvmf_reg_read32,
2860	.reg_read64		= nvmf_reg_read64,
2861	.reg_write32		= nvmf_reg_write32,
2862	.subsystem_reset	= nvmf_subsystem_reset,
2863	.free_ctrl		= nvme_tcp_free_ctrl,
2864	.submit_async_event	= nvme_tcp_submit_async_event,
2865	.delete_ctrl		= nvme_tcp_delete_ctrl,
2866	.get_address		= nvme_tcp_get_address,
2867	.stop_ctrl		= nvme_tcp_stop_ctrl,
2868	.get_virt_boundary	= nvmf_get_virt_boundary,
2869};
2870
2871static bool
2872nvme_tcp_existing_controller(struct nvmf_ctrl_options *opts)
2873{
2874	struct nvme_tcp_ctrl *ctrl;
2875	bool found = false;
2876
2877	mutex_lock(&nvme_tcp_ctrl_mutex);
2878	list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list) {
2879		found = nvmf_ip_options_match(&ctrl->ctrl, opts);
2880		if (found)
2881			break;
2882	}
2883	mutex_unlock(&nvme_tcp_ctrl_mutex);
2884
2885	return found;
2886}
2887
2888static struct nvme_tcp_ctrl *nvme_tcp_alloc_ctrl(struct device *dev,
2889		struct nvmf_ctrl_options *opts)
2890{
2891	struct nvme_tcp_ctrl *ctrl;
2892	int ret;
2893
2894	ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
2895	if (!ctrl)
2896		return ERR_PTR(-ENOMEM);
2897
2898	INIT_LIST_HEAD(&ctrl->list);
2899	ctrl->ctrl.opts = opts;
2900	ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues +
2901				opts->nr_poll_queues + 1;
2902	ctrl->ctrl.sqsize = opts->queue_size - 1;
2903	ctrl->ctrl.kato = opts->kato;
2904
2905	INIT_DELAYED_WORK(&ctrl->connect_work,
2906			nvme_tcp_reconnect_ctrl_work);
2907	INIT_WORK(&ctrl->err_work, nvme_tcp_error_recovery_work);
2908	INIT_WORK(&ctrl->ctrl.reset_work, nvme_reset_ctrl_work);
2909
2910	if (!(opts->mask & NVMF_OPT_TRSVCID)) {
2911		opts->trsvcid =
2912			kstrdup(__stringify(NVME_TCP_DISC_PORT), GFP_KERNEL);
2913		if (!opts->trsvcid) {
2914			ret = -ENOMEM;
2915			goto out_free_ctrl;
2916		}
2917		opts->mask |= NVMF_OPT_TRSVCID;
2918	}
2919
2920	ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2921			opts->traddr, opts->trsvcid, &ctrl->addr);
2922	if (ret) {
2923		pr_err("malformed address passed: %s:%s\n",
2924			opts->traddr, opts->trsvcid);
2925		goto out_free_ctrl;
2926	}
2927
2928	if (opts->mask & NVMF_OPT_HOST_TRADDR) {
2929		ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2930			opts->host_traddr, NULL, &ctrl->src_addr);
2931		if (ret) {
2932			pr_err("malformed src address passed: %s\n",
2933			       opts->host_traddr);
2934			goto out_free_ctrl;
2935		}
2936	}
2937
2938	if (opts->mask & NVMF_OPT_HOST_IFACE) {
2939		if (!__dev_get_by_name(&init_net, opts->host_iface)) {
2940			pr_err("invalid interface passed: %s\n",
2941			       opts->host_iface);
2942			ret = -ENODEV;
2943			goto out_free_ctrl;
2944		}
2945	}
2946
2947	if (!opts->duplicate_connect && nvme_tcp_existing_controller(opts)) {
2948		ret = -EALREADY;
2949		goto out_free_ctrl;
2950	}
2951
2952	ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
2953				GFP_KERNEL);
2954	if (!ctrl->queues) {
2955		ret = -ENOMEM;
2956		goto out_free_ctrl;
2957	}
2958
2959	ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_tcp_ctrl_ops, 0);
2960	if (ret)
2961		goto out_kfree_queues;
2962
2963	return ctrl;
2964out_kfree_queues:
2965	kfree(ctrl->queues);
2966out_free_ctrl:
2967	kfree(ctrl);
2968	return ERR_PTR(ret);
2969}
2970
2971static struct nvme_ctrl *nvme_tcp_create_ctrl(struct device *dev,
2972		struct nvmf_ctrl_options *opts)
2973{
2974	struct nvme_tcp_ctrl *ctrl;
2975	int ret;
2976
2977	ctrl = nvme_tcp_alloc_ctrl(dev, opts);
2978	if (IS_ERR(ctrl))
2979		return ERR_CAST(ctrl);
2980
2981	ret = nvme_add_ctrl(&ctrl->ctrl);
2982	if (ret)
2983		goto out_put_ctrl;
2984
2985	if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
2986		WARN_ON_ONCE(1);
2987		ret = -EINTR;
2988		goto out_uninit_ctrl;
2989	}
2990
2991	ret = nvme_tcp_setup_ctrl(&ctrl->ctrl, true);
2992	if (ret)
2993		goto out_uninit_ctrl;
2994
2995	dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISp, hostnqn: %s\n",
2996		nvmf_ctrl_subsysnqn(&ctrl->ctrl), &ctrl->addr, opts->host->nqn);
2997
2998	mutex_lock(&nvme_tcp_ctrl_mutex);
2999	list_add_tail(&ctrl->list, &nvme_tcp_ctrl_list);
3000	mutex_unlock(&nvme_tcp_ctrl_mutex);
3001
3002	return &ctrl->ctrl;
3003
3004out_uninit_ctrl:
3005	nvme_uninit_ctrl(&ctrl->ctrl);
3006out_put_ctrl:
3007	nvme_put_ctrl(&ctrl->ctrl);
3008	if (ret > 0)
3009		ret = -EIO;
3010	return ERR_PTR(ret);
3011}
3012
3013static struct nvmf_transport_ops nvme_tcp_transport = {
3014	.name		= "tcp",
3015	.module		= THIS_MODULE,
3016	.required_opts	= NVMF_OPT_TRADDR,
3017	.allowed_opts	= NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
3018			  NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
3019			  NVMF_OPT_HDR_DIGEST | NVMF_OPT_DATA_DIGEST |
3020			  NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES |
3021			  NVMF_OPT_TOS | NVMF_OPT_HOST_IFACE | NVMF_OPT_TLS |
3022			  NVMF_OPT_KEYRING | NVMF_OPT_TLS_KEY | NVMF_OPT_CONCAT,
3023	.create_ctrl	= nvme_tcp_create_ctrl,
3024};
3025
3026static int __init nvme_tcp_init_module(void)
3027{
3028	unsigned int wq_flags = WQ_MEM_RECLAIM | WQ_HIGHPRI | WQ_SYSFS;
3029	int cpu;
3030
3031	BUILD_BUG_ON(sizeof(struct nvme_tcp_hdr) != 8);
3032	BUILD_BUG_ON(sizeof(struct nvme_tcp_cmd_pdu) != 72);
3033	BUILD_BUG_ON(sizeof(struct nvme_tcp_data_pdu) != 24);
3034	BUILD_BUG_ON(sizeof(struct nvme_tcp_rsp_pdu) != 24);
3035	BUILD_BUG_ON(sizeof(struct nvme_tcp_r2t_pdu) != 24);
3036	BUILD_BUG_ON(sizeof(struct nvme_tcp_icreq_pdu) != 128);
3037	BUILD_BUG_ON(sizeof(struct nvme_tcp_icresp_pdu) != 128);
3038	BUILD_BUG_ON(sizeof(struct nvme_tcp_term_pdu) != 24);
3039
3040	if (wq_unbound)
3041		wq_flags |= WQ_UNBOUND;
3042
3043	nvme_tcp_wq = alloc_workqueue("nvme_tcp_wq", wq_flags, 0);
3044	if (!nvme_tcp_wq)
3045		return -ENOMEM;
3046
3047	for_each_possible_cpu(cpu)
3048		atomic_set(&nvme_tcp_cpu_queues[cpu], 0);
3049
3050	nvmf_register_transport(&nvme_tcp_transport);
3051	return 0;
3052}
3053
3054static void __exit nvme_tcp_cleanup_module(void)
3055{
3056	struct nvme_tcp_ctrl *ctrl;
3057
3058	nvmf_unregister_transport(&nvme_tcp_transport);
3059
3060	mutex_lock(&nvme_tcp_ctrl_mutex);
3061	list_for_each_entry(ctrl, &nvme_tcp_ctrl_list, list)
3062		nvme_delete_ctrl(&ctrl->ctrl);
3063	mutex_unlock(&nvme_tcp_ctrl_mutex);
3064	flush_workqueue(nvme_delete_wq);
3065
3066	destroy_workqueue(nvme_tcp_wq);
3067}
3068
3069module_init(nvme_tcp_init_module);
3070module_exit(nvme_tcp_cleanup_module);
3071
3072MODULE_DESCRIPTION("NVMe host TCP transport driver");
3073MODULE_LICENSE("GPL v2");