xref: /linux/fs/smb/client/transport.c (revision 89a312991dc6e638a36adc43ccb91dbc25504c04)

// SPDX-License-Identifier: LGPL-2.1
/*
 *
 *   Copyright (C) International Business Machines  Corp., 2002,2008
 *   Author(s): Steve French (sfrench@us.ibm.com)
 *   Jeremy Allison (jra@samba.org) 2006.
 *
 */

#include <linux/fs.h>
#include <linux/list.h>
#include <linux/gfp.h>
#include <linux/wait.h>
#include <linux/net.h>
#include <linux/delay.h>
#include <linux/freezer.h>
#include <linux/tcp.h>
#include <linux/bvec.h>
#include <linux/highmem.h>
#include <linux/uaccess.h>
#include <linux/processor.h>
#include <linux/mempool.h>
#include <linux/sched/signal.h>
#include <linux/task_io_accounting_ops.h>
#include <linux/task_work.h>
#include "cifsglob.h"
#include "cifsproto.h"
#include "cifs_debug.h"
#include "smb2proto.h"
#include "smbdirect.h"
#include "compress.h"

void
cifs_wake_up_task(struct TCP_Server_Info *server, struct mid_q_entry *mid)
{
	if (mid->mid_state == MID_RESPONSE_RECEIVED)
		mid->mid_state = MID_RESPONSE_READY;
	wake_up_process(mid->callback_data);
}

void __release_mid(struct TCP_Server_Info *server, struct mid_q_entry *midEntry)
{
#ifdef CONFIG_CIFS_STATS2
	__le16 command = server->vals->lock_cmd;
	__u16 smb_cmd = le16_to_cpu(midEntry->command);
	unsigned long now;
	unsigned long roundtrip_time;
#endif

	if (midEntry->resp_buf && (midEntry->wait_cancelled) &&
	    (midEntry->mid_state == MID_RESPONSE_RECEIVED ||
	     midEntry->mid_state == MID_RESPONSE_READY) &&
	    server->ops->handle_cancelled_mid)
		server->ops->handle_cancelled_mid(midEntry, server);

	midEntry->mid_state = MID_FREE;
	atomic_dec(&mid_count);
	if (midEntry->large_buf)
		cifs_buf_release(midEntry->resp_buf);
	else
		cifs_small_buf_release(midEntry->resp_buf);
#ifdef CONFIG_CIFS_STATS2
	now = jiffies;
	if (now < midEntry->when_alloc)
		cifs_server_dbg(VFS, "Invalid mid allocation time\n");
	roundtrip_time = now - midEntry->when_alloc;

	if (smb_cmd < NUMBER_OF_SMB2_COMMANDS) {
		if (atomic_read(&server->num_cmds[smb_cmd]) == 0) {
			server->slowest_cmd[smb_cmd] = roundtrip_time;
			server->fastest_cmd[smb_cmd] = roundtrip_time;
		} else {
			if (server->slowest_cmd[smb_cmd] < roundtrip_time)
				server->slowest_cmd[smb_cmd] = roundtrip_time;
			else if (server->fastest_cmd[smb_cmd] > roundtrip_time)
				server->fastest_cmd[smb_cmd] = roundtrip_time;
		}
		cifs_stats_inc(&server->num_cmds[smb_cmd]);
		server->time_per_cmd[smb_cmd] += roundtrip_time;
	}
	/*
	 * commands taking longer than one second (default) can be indications
	 * that something is wrong, unless it is quite a slow link or a very
	 * busy server. Note that this calc is unlikely or impossible to wrap
	 * as long as slow_rsp_threshold is not set way above recommended max
	 * value (32767 ie 9 hours) and is generally harmless even if wrong
	 * since only affects debug counters - so leaving the calc as simple
	 * comparison rather than doing multiple conversions and overflow
	 * checks
	 */
	if ((slow_rsp_threshold != 0) &&
	    time_after(now, midEntry->when_alloc + (slow_rsp_threshold * HZ)) &&
	    (midEntry->command != command)) {
		/*
		 * smb2slowcmd[NUMBER_OF_SMB2_COMMANDS] counts by command
		 * NB: le16_to_cpu returns unsigned so can not be negative below
		 */
		if (smb_cmd < NUMBER_OF_SMB2_COMMANDS)
			cifs_stats_inc(&server->smb2slowcmd[smb_cmd]);

		trace_smb3_slow_rsp(smb_cmd, midEntry->mid, midEntry->pid,
			       midEntry->when_sent, midEntry->when_received);
		if (cifsFYI & CIFS_TIMER) {
			pr_debug("slow rsp: cmd %d mid %llu A: 0x%lx S: 0x%lx R: 0x%lx\n",
				 midEntry->command, midEntry->mid,
				 now - midEntry->when_alloc,
				 now - midEntry->when_sent,
				 now - midEntry->when_received);
		}
	}
#endif
	put_task_struct(midEntry->creator);

	mempool_free(midEntry, &cifs_mid_pool);
}

void
delete_mid(struct TCP_Server_Info *server, struct mid_q_entry *mid)
{
	spin_lock(&server->mid_queue_lock);

	if (!mid->deleted_from_q) {
		list_del_init(&mid->qhead);
		mid->deleted_from_q = true;
	}
	spin_unlock(&server->mid_queue_lock);

	release_mid(server, mid);
}

/*
 * smb_send_kvec - send an array of kvecs to the server
 * @server:	Server to send the data to
 * @smb_msg:	Message to send
 * @sent:	amount of data sent on socket is stored here
 *
 * Our basic "send data to server" function. Should be called with srv_mutex
 * held. The caller is responsible for handling the results.
 */
int
smb_send_kvec(struct TCP_Server_Info *server, struct msghdr *smb_msg,
	      size_t *sent)
{
	int rc = 0;
	int retries = 0;
	struct socket *ssocket = server->ssocket;

	*sent = 0;

	if (server->noblocksnd)
		smb_msg->msg_flags = MSG_DONTWAIT + MSG_NOSIGNAL;
	else
		smb_msg->msg_flags = MSG_NOSIGNAL;

	while (msg_data_left(smb_msg)) {
		/*
		 * If blocking send, we try 3 times, since each can block
		 * for 5 seconds. For nonblocking  we have to try more
		 * but wait increasing amounts of time allowing time for
		 * socket to clear.  The overall time we wait in either
		 * case to send on the socket is about 15 seconds.
		 * Similarly we wait for 15 seconds for a response from
		 * the server in SendReceive[2] for the server to send
		 * a response back for most types of requests (except
		 * SMB Write past end of file which can be slow, and
		 * blocking lock operations). NFS waits slightly longer
		 * than CIFS, but this can make it take longer for
		 * nonresponsive servers to be detected and 15 seconds
		 * is more than enough time for modern networks to
		 * send a packet.  In most cases if we fail to send
		 * after the retries we will kill the socket and
		 * reconnect which may clear the network problem.
		 *
		 * Even if regular signals are masked, EINTR might be
		 * propagated from sk_stream_wait_memory() to here when
		 * TIF_NOTIFY_SIGNAL is used for task work. For example,
		 * certain io_uring completions will use that. Treat
		 * having EINTR with pending task work the same as EAGAIN
		 * to avoid unnecessary reconnects.
		 */
		rc = sock_sendmsg(ssocket, smb_msg);
		if (rc == -EAGAIN || unlikely(rc == -EINTR && task_work_pending(current))) {
			retries++;
			if (retries >= 14 ||
			    (!server->noblocksnd && (retries > 2))) {
				cifs_server_dbg(VFS, "sends on sock %p stuck for 15 seconds\n",
					 ssocket);
				return -EAGAIN;
			}
			msleep(1 << retries);
			continue;
		}

		if (rc < 0)
			return rc;

		if (rc == 0) {
			/* should never happen, letting socket clear before
			   retrying is our only obvious option here */
			cifs_server_dbg(VFS, "tcp sent no data\n");
			msleep(500);
			continue;
		}

		/* send was at least partially successful */
		*sent += rc;
		retries = 0; /* in case we get ENOSPC on the next send */
	}
	return 0;
}

unsigned long
smb_rqst_len(struct TCP_Server_Info *server, struct smb_rqst *rqst)
{
	unsigned int i;
	struct kvec *iov;
	int nvec;
	unsigned long buflen = 0;

	if (!is_smb1(server) && rqst->rq_nvec >= 2 &&
	    rqst->rq_iov[0].iov_len == 4) {
		iov = &rqst->rq_iov[1];
		nvec = rqst->rq_nvec - 1;
	} else {
		iov = rqst->rq_iov;
		nvec = rqst->rq_nvec;
	}

	/* total up iov array first */
	for (i = 0; i < nvec; i++)
		buflen += iov[i].iov_len;

	buflen += iov_iter_count(&rqst->rq_iter);
	return buflen;
}

int __smb_send_rqst(struct TCP_Server_Info *server, int num_rqst,
		    struct smb_rqst *rqst)
{
	int rc;
	struct kvec *iov;
	int n_vec;
	unsigned int send_length = 0;
	unsigned int i, j;
	sigset_t mask, oldmask;
	size_t total_len = 0, sent, size;
	struct socket *ssocket = server->ssocket;
	struct msghdr smb_msg = {};
	__be32 rfc1002_marker;

	cifs_in_send_inc(server);
	if (cifs_rdma_enabled(server)) {
		/* return -EAGAIN when connecting or reconnecting */
		rc = -EAGAIN;
		if (server->smbd_conn)
			rc = smbd_send(server, num_rqst, rqst);
		goto smbd_done;
	}

	rc = -EAGAIN;
	if (ssocket == NULL)
		goto out;

	rc = -ERESTARTSYS;
	if (fatal_signal_pending(current)) {
		cifs_dbg(FYI, "signal pending before send request\n");
		goto out;
	}

	rc = 0;
	/* cork the socket */
	tcp_sock_set_cork(ssocket->sk, true);

	for (j = 0; j < num_rqst; j++)
		send_length += smb_rqst_len(server, &rqst[j]);
	rfc1002_marker = cpu_to_be32(send_length);

	/*
	 * We should not allow signals to interrupt the network send because
	 * any partial send will cause session reconnects thus increasing
	 * latency of system calls and overload a server with unnecessary
	 * requests.
	 */

	sigfillset(&mask);
	sigprocmask(SIG_BLOCK, &mask, &oldmask);

	/* Generate a rfc1002 marker */
	{
		struct kvec hiov = {
			.iov_base = &rfc1002_marker,
			.iov_len  = 4
		};
		iov_iter_kvec(&smb_msg.msg_iter, ITER_SOURCE, &hiov, 1, 4);
		rc = smb_send_kvec(server, &smb_msg, &sent);
		if (rc < 0)
			goto unmask;

		total_len += sent;
		send_length += 4;
	}

	cifs_dbg(FYI, "Sending smb: smb_len=%u\n", send_length);

	for (j = 0; j < num_rqst; j++) {
		iov = rqst[j].rq_iov;
		n_vec = rqst[j].rq_nvec;

		size = 0;
		for (i = 0; i < n_vec; i++) {
			dump_smb(iov[i].iov_base, iov[i].iov_len);
			size += iov[i].iov_len;
		}

		iov_iter_kvec(&smb_msg.msg_iter, ITER_SOURCE, iov, n_vec, size);

		rc = smb_send_kvec(server, &smb_msg, &sent);
		if (rc < 0)
			goto unmask;

		total_len += sent;

		if (iov_iter_count(&rqst[j].rq_iter) > 0) {
			smb_msg.msg_iter = rqst[j].rq_iter;
			rc = smb_send_kvec(server, &smb_msg, &sent);
			if (rc < 0)
				break;
			total_len += sent;
		}
	}

unmask:
	sigprocmask(SIG_SETMASK, &oldmask, NULL);

	/*
	 * If signal is pending but we have already sent the whole packet to
	 * the server we need to return success status to allow a corresponding
	 * mid entry to be kept in the pending requests queue thus allowing
	 * to handle responses from the server by the client.
	 *
	 * If only part of the packet has been sent there is no need to hide
	 * interrupt because the session will be reconnected anyway, so there
	 * won't be any response from the server to handle.
	 */

	if (signal_pending(current) && (total_len != send_length)) {
		cifs_dbg(FYI, "signal is pending after attempt to send\n");
		rc = -ERESTARTSYS;
	}

	/* uncork it */
	tcp_sock_set_cork(ssocket->sk, false);

	if ((total_len > 0) && (total_len != send_length)) {
		cifs_dbg(FYI, "partial send (wanted=%u sent=%zu): terminating session\n",
			 send_length, total_len);
		/*
		 * If we have only sent part of an SMB then the next SMB could
		 * be taken as the remainder of this one. We need to kill the
		 * socket so the server throws away the partial SMB
		 */
		cifs_signal_cifsd_for_reconnect(server, false);
		trace_smb3_partial_send_reconnect(server->current_mid,
						  server->conn_id, server->hostname);
	}
smbd_done:
	/*
	 * there's hardly any use for the layers above to know the
	 * actual error code here. All they should do at this point is
	 * to retry the connection and hope it goes away.
	 */
	if (rc < 0 && rc != -EINTR && rc != -EAGAIN) {
		cifs_server_dbg(VFS, "Error %d sending data on socket to server\n",
			 rc);
		rc = -ECONNABORTED;
		cifs_signal_cifsd_for_reconnect(server, false);
	} else if (rc > 0)
		rc = 0;
out:
	cifs_in_send_dec(server);
	return rc;
}

static int
smb_send_rqst(struct TCP_Server_Info *server, int num_rqst,
	      struct smb_rqst *rqst, int flags)
{
	struct smb2_transform_hdr tr_hdr;
	struct smb_rqst new_rqst[MAX_COMPOUND] = {};
	struct kvec iov = {
		.iov_base = &tr_hdr,
		.iov_len = sizeof(tr_hdr),
	};
	int rc;

	if (flags & CIFS_COMPRESS_REQ)
		return smb_compress(server, &rqst[0], __smb_send_rqst);

	if (!(flags & CIFS_TRANSFORM_REQ))
		return __smb_send_rqst(server, num_rqst, rqst);

	if (WARN_ON_ONCE(num_rqst > MAX_COMPOUND - 1))
		return smb_EIO1(smb_eio_trace_tx_max_compound, num_rqst);

	if (!server->ops->init_transform_rq) {
		cifs_server_dbg(VFS, "Encryption requested but transform callback is missing\n");
		return smb_EIO(smb_eio_trace_tx_need_transform);
	}

	new_rqst[0].rq_iov = &iov;
	new_rqst[0].rq_nvec = 1;

	rc = server->ops->init_transform_rq(server, num_rqst + 1,
					    new_rqst, rqst);
	if (!rc) {
		rc = __smb_send_rqst(server, num_rqst + 1, new_rqst);
		smb3_free_compound_rqst(num_rqst, &new_rqst[1]);
	}
	return rc;
}

static int
wait_for_free_credits(struct TCP_Server_Info *server, const int num_credits,
		      const int timeout, const int flags,
		      unsigned int *instance)
{
	long rc;
	int *credits;
	int optype;
	long int t;
	int scredits, in_flight;

	if (timeout < 0)
		t = MAX_JIFFY_OFFSET;
	else
		t = msecs_to_jiffies(timeout);

	optype = flags & CIFS_OP_MASK;

	*instance = 0;

	credits = server->ops->get_credits_field(server, optype);
	/* Since an echo is already inflight, no need to wait to send another */
	if (*credits <= 0 && optype == CIFS_ECHO_OP)
		return -EAGAIN;

	spin_lock(&server->req_lock);
	if ((flags & CIFS_TIMEOUT_MASK) == CIFS_NON_BLOCKING) {
		/* oplock breaks must not be held up */
		server->in_flight++;
		if (server->in_flight > server->max_in_flight)
			server->max_in_flight = server->in_flight;
		*credits -= 1;
		*instance = server->reconnect_instance;
		scredits = *credits;
		in_flight = server->in_flight;
		spin_unlock(&server->req_lock);

		trace_smb3_nblk_credits(server->current_mid,
				server->conn_id, server->hostname, scredits, -1, in_flight);
		cifs_dbg(FYI, "%s: remove %u credits total=%d\n",
				__func__, 1, scredits);

		return 0;
	}

	while (1) {
		spin_unlock(&server->req_lock);

		spin_lock(&server->srv_lock);
		if (server->tcpStatus == CifsExiting) {
			spin_unlock(&server->srv_lock);
			return -ENOENT;
		}
		spin_unlock(&server->srv_lock);

		spin_lock(&server->req_lock);
		if (*credits < num_credits) {
			scredits = *credits;
			spin_unlock(&server->req_lock);

			cifs_num_waiters_inc(server);
			rc = wait_event_killable_timeout(server->request_q,
				has_credits(server, credits, num_credits), t);
			cifs_num_waiters_dec(server);
			if (!rc) {
				spin_lock(&server->req_lock);
				scredits = *credits;
				in_flight = server->in_flight;
				spin_unlock(&server->req_lock);

				trace_smb3_credit_timeout(server->current_mid,
						server->conn_id, server->hostname, scredits,
						num_credits, in_flight);
				cifs_server_dbg(VFS, "wait timed out after %d ms\n",
						timeout);
				return -EBUSY;
			}
			if (rc == -ERESTARTSYS)
				return -ERESTARTSYS;
			spin_lock(&server->req_lock);
		} else {
			/*
			 * For normal commands, reserve the last MAX_COMPOUND
			 * credits to compound requests.
			 * Otherwise these compounds could be permanently
			 * starved for credits by single-credit requests.
			 *
			 * To prevent spinning CPU, block this thread until
			 * there are >MAX_COMPOUND credits available.
			 * But only do this is we already have a lot of
			 * credits in flight to avoid triggering this check
			 * for servers that are slow to hand out credits on
			 * new sessions.
			 */
			if (!optype && num_credits == 1 &&
			    server->in_flight > 2 * MAX_COMPOUND &&
			    *credits <= MAX_COMPOUND) {
				spin_unlock(&server->req_lock);

				cifs_num_waiters_inc(server);
				rc = wait_event_killable_timeout(
					server->request_q,
					has_credits(server, credits,
						    MAX_COMPOUND + 1),
					t);
				cifs_num_waiters_dec(server);
				if (!rc) {
					spin_lock(&server->req_lock);
					scredits = *credits;
					in_flight = server->in_flight;
					spin_unlock(&server->req_lock);

					trace_smb3_credit_timeout(
							server->current_mid,
							server->conn_id, server->hostname,
							scredits, num_credits, in_flight);
					cifs_server_dbg(VFS, "wait timed out after %d ms\n",
							timeout);
					return -EBUSY;
				}
				if (rc == -ERESTARTSYS)
					return -ERESTARTSYS;
				spin_lock(&server->req_lock);
				continue;
			}

			/*
			 * Can not count locking commands against total
			 * as they are allowed to block on server.
			 */

			/* update # of requests on the wire to server */
			if ((flags & CIFS_TIMEOUT_MASK) != CIFS_BLOCKING_OP) {
				*credits -= num_credits;
				server->in_flight += num_credits;
				if (server->in_flight > server->max_in_flight)
					server->max_in_flight = server->in_flight;
				*instance = server->reconnect_instance;
			}
			scredits = *credits;
			in_flight = server->in_flight;
			spin_unlock(&server->req_lock);

			trace_smb3_waitff_credits(server->current_mid,
					server->conn_id, server->hostname, scredits,
					-(num_credits), in_flight);
			cifs_dbg(FYI, "%s: remove %u credits total=%d\n",
					__func__, num_credits, scredits);
			break;
		}
	}
	return 0;
}

int wait_for_free_request(struct TCP_Server_Info *server, const int flags,
			  unsigned int *instance)
{
	return wait_for_free_credits(server, 1, -1, flags,
				     instance);
}

static int
wait_for_compound_request(struct TCP_Server_Info *server, int num,
			  const int flags, unsigned int *instance)
{
	int *credits;
	int scredits, in_flight;

	credits = server->ops->get_credits_field(server, flags & CIFS_OP_MASK);

	spin_lock(&server->req_lock);
	scredits = *credits;
	in_flight = server->in_flight;

	if (*credits < num) {
		/*
		 * If the server is tight on resources or just gives us less
		 * credits for other reasons (e.g. requests are coming out of
		 * order and the server delays granting more credits until it
		 * processes a missing mid) and we exhausted most available
		 * credits there may be situations when we try to send
		 * a compound request but we don't have enough credits. At this
		 * point the client needs to decide if it should wait for
		 * additional credits or fail the request. If at least one
		 * request is in flight there is a high probability that the
		 * server will return enough credits to satisfy this compound
		 * request.
		 *
		 * Return immediately if no requests in flight since we will be
		 * stuck on waiting for credits.
		 */
		if (server->in_flight == 0) {
			spin_unlock(&server->req_lock);
			trace_smb3_insufficient_credits(server->current_mid,
					server->conn_id, server->hostname, scredits,
					num, in_flight);
			cifs_dbg(FYI, "%s: %d requests in flight, needed %d total=%d\n",
					__func__, in_flight, num, scredits);
			return -EDEADLK;
		}
	}
	spin_unlock(&server->req_lock);

	return wait_for_free_credits(server, num, 60000, flags,
				     instance);
}

int
cifs_wait_mtu_credits(struct TCP_Server_Info *server, size_t size,
		      size_t *num, struct cifs_credits *credits)
{
	*num = size;
	credits->value = 0;
	credits->instance = server->reconnect_instance;
	return 0;
}

int wait_for_response(struct TCP_Server_Info *server, struct mid_q_entry *mid)
{
	unsigned int sleep_state = TASK_KILLABLE;
	int error;

	if (mid->sr_flags & CIFS_INTERRUPTIBLE_WAIT)
		sleep_state = TASK_INTERRUPTIBLE;

	error = wait_event_state(server->response_q,
				 mid->mid_state != MID_REQUEST_SUBMITTED &&
				 mid->mid_state != MID_RESPONSE_RECEIVED,
				 (sleep_state | TASK_FREEZABLE_UNSAFE));
	if (error < 0)
		return -ERESTARTSYS;

	return 0;
}

/*
 * Send a SMB request and set the callback function in the mid to handle
 * the result. Caller is responsible for dealing with timeouts.
 */
int
cifs_call_async(struct TCP_Server_Info *server, struct smb_rqst *rqst,
		mid_receive_t receive, mid_callback_t callback,
		mid_handle_t handle, void *cbdata, const int flags,
		const struct cifs_credits *exist_credits)
{
	int rc;
	struct mid_q_entry *mid;
	struct cifs_credits credits = { .value = 0, .instance = 0 };
	unsigned int instance;
	int optype;

	optype = flags & CIFS_OP_MASK;

	if ((flags & CIFS_HAS_CREDITS) == 0) {
		rc = wait_for_free_request(server, flags, &instance);
		if (rc)
			return rc;
		credits.value = 1;
		credits.instance = instance;
	} else
		instance = exist_credits->instance;

	cifs_server_lock(server);

	/*
	 * We can't use credits obtained from the previous session to send this
	 * request. Check if there were reconnects after we obtained credits and
	 * return -EAGAIN in such cases to let callers handle it.
	 */
	if (instance != server->reconnect_instance) {
		cifs_server_unlock(server);
		add_credits_and_wake_if(server, &credits, optype);
		return -EAGAIN;
	}

	mid = server->ops->setup_async_request(server, rqst);
	if (IS_ERR(mid)) {
		cifs_server_unlock(server);
		add_credits_and_wake_if(server, &credits, optype);
		return PTR_ERR(mid);
	}

	mid->sr_flags = flags;
	mid->receive = receive;
	mid->callback = callback;
	mid->callback_data = cbdata;
	mid->handle = handle;
	mid->mid_state = MID_REQUEST_SUBMITTED;

	/* put it on the pending_mid_q */
	spin_lock(&server->mid_queue_lock);
	list_add_tail(&mid->qhead, &server->pending_mid_q);
	spin_unlock(&server->mid_queue_lock);

	/*
	 * Need to store the time in mid before calling I/O. For call_async,
	 * I/O response may come back and free the mid entry on another thread.
	 */
	cifs_save_when_sent(mid);
	rc = smb_send_rqst(server, 1, rqst, flags);

	if (rc < 0) {
		revert_current_mid(server, mid->credits);
		server->sequence_number -= 2;
		delete_mid(server, mid);
	}

	cifs_server_unlock(server);

	if (rc == 0)
		return 0;

	add_credits_and_wake_if(server, &credits, optype);
	return rc;
}

int cifs_sync_mid_result(struct mid_q_entry *mid, struct TCP_Server_Info *server)
{
	int rc = 0;

	cifs_dbg(FYI, "%s: cmd=%d mid=%llu state=%d\n",
		 __func__, le16_to_cpu(mid->command), mid->mid, mid->mid_state);

	spin_lock(&server->mid_queue_lock);
	switch (mid->mid_state) {
	case MID_RESPONSE_READY:
		spin_unlock(&server->mid_queue_lock);
		return rc;
	case MID_RETRY_NEEDED:
		rc = -EAGAIN;
		break;
	case MID_RESPONSE_MALFORMED:
		rc = smb_EIO(smb_eio_trace_rx_sync_mid_malformed);
		break;
	case MID_SHUTDOWN:
		rc = -EHOSTDOWN;
		break;
	case MID_RC:
		rc = mid->mid_rc;
		break;
	default:
		if (mid->deleted_from_q == false) {
			list_del_init(&mid->qhead);
			mid->deleted_from_q = true;
		}
		spin_unlock(&server->mid_queue_lock);
		cifs_server_dbg(VFS, "%s: invalid mid state mid=%llu state=%d\n",
			 __func__, mid->mid, mid->mid_state);
		rc = smb_EIO1(smb_eio_trace_rx_sync_mid_invalid, mid->mid_state);
		goto sync_mid_done;
	}
	spin_unlock(&server->mid_queue_lock);

sync_mid_done:
	release_mid(server, mid);
	return rc;
}

static void
cifs_compound_callback(struct TCP_Server_Info *server, struct mid_q_entry *mid)
{
	struct cifs_credits credits = {
		.value = server->ops->get_credits(mid),
		.instance = server->reconnect_instance,
	};

	add_credits(server, &credits, mid->optype);

	if (mid->mid_state == MID_RESPONSE_RECEIVED)
		mid->mid_state = MID_RESPONSE_READY;
}

static void
cifs_compound_last_callback(struct TCP_Server_Info *server, struct mid_q_entry *mid)
{
	cifs_compound_callback(server, mid);
	cifs_wake_up_task(server, mid);
}

static void
cifs_cancelled_callback(struct TCP_Server_Info *server, struct mid_q_entry *mid)
{
	cifs_compound_callback(server, mid);
	release_mid(server, mid);
}

/*
 * cifs_pick_channel - pick an eligible channel for network operations
 *
 * @ses: session reference
 *
 * Select an eligible channel (not terminating and not marked as needing
 * reconnect), preferring the least loaded one. If no eligible channel is
 * found, fall back to the primary channel (index 0).
 *
 * Return: TCP_Server_Info pointer for the chosen channel, or NULL if @ses is
 * NULL.
 */
struct TCP_Server_Info *cifs_pick_channel(struct cifs_ses *ses)
{
	uint index = 0;
	unsigned int min_in_flight = UINT_MAX;
	struct TCP_Server_Info *server = NULL;
	int i, start, cur;

	if (!ses)
		return NULL;

	spin_lock(&ses->chan_lock);
	start = atomic_inc_return(&ses->chan_seq);
	for (i = 0; i < ses->chan_count; i++) {
		cur = (start + i) % ses->chan_count;
		server = ses->chans[cur].server;
		if (!server || server->terminate)
			continue;

		if (CIFS_CHAN_NEEDS_RECONNECT(ses, cur))
			continue;

		/*
		 * strictly speaking, we should pick up req_lock to read
		 * server->in_flight. But it shouldn't matter much here if we
		 * race while reading this data. The worst that can happen is
		 * that we could use a channel that's not least loaded. Avoiding
		 * taking the lock could help reduce wait time, which is
		 * important for this function
		 */
		if (server->in_flight < min_in_flight) {
			min_in_flight = server->in_flight;
			index = cur;
		}
	}

	server = ses->chans[index].server;
	spin_unlock(&ses->chan_lock);

	return server;
}

int
compound_send_recv(const unsigned int xid, struct cifs_ses *ses,
		   struct TCP_Server_Info *server,
		   const int flags, const int num_rqst, struct smb_rqst *rqst,
		   int *resp_buf_type, struct kvec *resp_iov)
{
	int i, j, optype, rc = 0;
	struct mid_q_entry *mid[MAX_COMPOUND];
	bool cancelled_mid[MAX_COMPOUND] = {false};
	struct cifs_credits credits[MAX_COMPOUND] = {
		{ .value = 0, .instance = 0 }
	};
	unsigned int instance;
	char *buf;

	optype = flags & CIFS_OP_MASK;

	for (i = 0; i < num_rqst; i++)
		resp_buf_type[i] = CIFS_NO_BUFFER;  /* no response buf yet */

	if (!ses || !ses->server || !server) {
		cifs_dbg(VFS, "Null session\n");
		return smb_EIO(smb_eio_trace_null_pointers);
	}

	spin_lock(&server->srv_lock);
	if (server->tcpStatus == CifsExiting) {
		spin_unlock(&server->srv_lock);
		return -ENOENT;
	}
	spin_unlock(&server->srv_lock);

	/*
	 * Wait for all the requests to become available.
	 * This approach still leaves the possibility to be stuck waiting for
	 * credits if the server doesn't grant credits to the outstanding
	 * requests and if the client is completely idle, not generating any
	 * other requests.
	 * This can be handled by the eventual session reconnect.
	 */
	rc = wait_for_compound_request(server, num_rqst, flags,
				       &instance);
	if (rc)
		return rc;

	for (i = 0; i < num_rqst; i++) {
		credits[i].value = 1;
		credits[i].instance = instance;
	}

	/*
	 * Make sure that we sign in the same order that we send on this socket
	 * and avoid races inside tcp sendmsg code that could cause corruption
	 * of smb data.
	 */

	cifs_server_lock(server);

	/*
	 * All the parts of the compound chain belong obtained credits from the
	 * same session. We can not use credits obtained from the previous
	 * session to send this request. Check if there were reconnects after
	 * we obtained credits and return -EAGAIN in such cases to let callers
	 * handle it.
	 */
	if (instance != server->reconnect_instance) {
		cifs_server_unlock(server);
		for (j = 0; j < num_rqst; j++)
			add_credits(server, &credits[j], optype);
		return -EAGAIN;
	}

	for (i = 0; i < num_rqst; i++) {
		mid[i] = server->ops->setup_request(ses, server, &rqst[i]);
		if (IS_ERR(mid[i])) {
			revert_current_mid(server, i);
			for (j = 0; j < i; j++)
				delete_mid(server, mid[j]);
			cifs_server_unlock(server);

			/* Update # of requests on wire to server */
			for (j = 0; j < num_rqst; j++)
				add_credits(server, &credits[j], optype);
			return PTR_ERR(mid[i]);
		}

		mid[i]->sr_flags = flags;
		mid[i]->mid_state = MID_REQUEST_SUBMITTED;
		mid[i]->optype = optype;
		/*
		 * Invoke callback for every part of the compound chain
		 * to calculate credits properly. Wake up this thread only when
		 * the last element is received.
		 */
		if (i < num_rqst - 1)
			mid[i]->callback = cifs_compound_callback;
		else
			mid[i]->callback = cifs_compound_last_callback;
	}
	rc = smb_send_rqst(server, num_rqst, rqst, flags);

	for (i = 0; i < num_rqst; i++)
		cifs_save_when_sent(mid[i]);

	if (rc < 0) {
		revert_current_mid(server, num_rqst);
		server->sequence_number -= 2;
	}

	cifs_server_unlock(server);

	/*
	 * If sending failed for some reason or it is an oplock break that we
	 * will not receive a response to - return credits back
	 */
	if (rc < 0 || (flags & CIFS_NO_SRV_RSP)) {
		for (i = 0; i < num_rqst; i++)
			add_credits(server, &credits[i], optype);
		goto out;
	}

	/*
	 * At this point the request is passed to the network stack - we assume
	 * that any credits taken from the server structure on the client have
	 * been spent and we can't return them back. Once we receive responses
	 * we will collect credits granted by the server in the mid callbacks
	 * and add those credits to the server structure.
	 */

	/*
	 * Compounding is never used during session establish.
	 */
	spin_lock(&ses->ses_lock);
	if ((ses->ses_status == SES_NEW) || (optype & CIFS_NEG_OP) || (optype & CIFS_SESS_OP)) {
		spin_unlock(&ses->ses_lock);

		if (WARN_ON_ONCE(num_rqst != 1 || !resp_iov))
			return -EINVAL;

		cifs_server_lock(server);
		smb311_update_preauth_hash(ses, server, rqst[0].rq_iov, rqst[0].rq_nvec);
		cifs_server_unlock(server);

		spin_lock(&ses->ses_lock);
	}
	spin_unlock(&ses->ses_lock);

	for (i = 0; i < num_rqst; i++) {
		rc = wait_for_response(server, mid[i]);
		if (rc != 0)
			break;
	}
	if (rc != 0) {
		for (; i < num_rqst; i++) {
			cifs_server_dbg(FYI, "Cancelling wait for mid %llu cmd: %d\n",
				 mid[i]->mid, le16_to_cpu(mid[i]->command));
			send_cancel(ses, server, &rqst[i], mid[i], xid);
			spin_lock(&mid[i]->mid_lock);
			mid[i]->wait_cancelled = true;
			if (mid[i]->mid_state == MID_REQUEST_SUBMITTED ||
			    mid[i]->mid_state == MID_RESPONSE_RECEIVED) {
				mid[i]->callback = cifs_cancelled_callback;
				cancelled_mid[i] = true;
				credits[i].value = 0;
			}
			spin_unlock(&mid[i]->mid_lock);
		}
	}

	for (i = 0; i < num_rqst; i++) {
		if (rc < 0)
			goto out;

		rc = cifs_sync_mid_result(mid[i], server);
		if (rc != 0) {
			/* mark this mid as cancelled to not free it below */
			cancelled_mid[i] = true;
			goto out;
		}

		if (!mid[i]->resp_buf ||
		    mid[i]->mid_state != MID_RESPONSE_READY) {
			rc = smb_EIO1(smb_eio_trace_rx_mid_unready, mid[i]->mid_state);
			cifs_dbg(FYI, "Bad MID state?\n");
			goto out;
		}

		rc = server->ops->check_receive(mid[i], server,
						flags & CIFS_LOG_ERROR);

		if (resp_iov) {
			buf = (char *)mid[i]->resp_buf;
			resp_iov[i].iov_base = buf;
			resp_iov[i].iov_len = mid[i]->resp_buf_size;

			if (mid[i]->large_buf)
				resp_buf_type[i] = CIFS_LARGE_BUFFER;
			else
				resp_buf_type[i] = CIFS_SMALL_BUFFER;

			/* mark it so buf will not be freed by delete_mid */
			if ((flags & CIFS_NO_RSP_BUF) == 0)
				mid[i]->resp_buf = NULL;
		}
	}

	/*
	 * Compounding is never used during session establish.
	 */
	spin_lock(&ses->ses_lock);
	if ((ses->ses_status == SES_NEW) || (optype & CIFS_NEG_OP) || (optype & CIFS_SESS_OP)) {
		struct kvec iov = {
			.iov_base = resp_iov[0].iov_base,
			.iov_len = resp_iov[0].iov_len
		};
		spin_unlock(&ses->ses_lock);
		cifs_server_lock(server);
		smb311_update_preauth_hash(ses, server, &iov, 1);
		cifs_server_unlock(server);
		spin_lock(&ses->ses_lock);
	}
	spin_unlock(&ses->ses_lock);

out:
	/*
	 * This will dequeue all mids. After this it is important that the
	 * demultiplex_thread will not process any of these mids any further.
	 * This is prevented above by using a noop callback that will not
	 * wake this thread except for the very last PDU.
	 */
	for (i = 0; i < num_rqst; i++) {
		if (!cancelled_mid[i])
			delete_mid(server, mid[i]);
	}

	return rc;
}

int
cifs_send_recv(const unsigned int xid, struct cifs_ses *ses,
	       struct TCP_Server_Info *server,
	       struct smb_rqst *rqst, int *resp_buf_type, const int flags,
	       struct kvec *resp_iov)
{
	return compound_send_recv(xid, ses, server, flags, 1,
				  rqst, resp_buf_type, resp_iov);
}


/*
 * Discard any remaining data in the current SMB. To do this, we borrow the
 * current bigbuf.
 */
int
cifs_discard_remaining_data(struct TCP_Server_Info *server)
{
	unsigned int rfclen = server->pdu_size;
	size_t remaining = rfclen - server->total_read;

	while (remaining > 0) {
		ssize_t length;

		length = cifs_discard_from_socket(server,
				min_t(size_t, remaining,
				      CIFSMaxBufSize + MAX_HEADER_SIZE(server)));
		if (length < 0)
			return length;
		server->total_read += length;
		remaining -= length;
	}

	return 0;
}

static int
__cifs_readv_discard(struct TCP_Server_Info *server, struct mid_q_entry *mid,
		     bool malformed)
{
	int length;

	length = cifs_discard_remaining_data(server);
	dequeue_mid(server, mid, malformed);
	mid->resp_buf = server->smallbuf;
	server->smallbuf = NULL;
	return length;
}

static int
cifs_readv_discard(struct TCP_Server_Info *server, struct mid_q_entry *mid)
{
	struct cifs_io_subrequest *rdata = mid->callback_data;

	return  __cifs_readv_discard(server, mid, rdata->result);
}

int
cifs_readv_receive(struct TCP_Server_Info *server, struct mid_q_entry *mid)
{
	int length, len;
	unsigned int data_offset, data_len, end_off;
	struct cifs_io_subrequest *rdata = mid->callback_data;
	char *buf = server->smallbuf;
	unsigned int buflen = server->pdu_size;
	bool use_rdma_mr = false;

	cifs_dbg(FYI, "%s: mid=%llu offset=%llu bytes=%zu\n",
		 __func__, mid->mid, rdata->subreq.start, rdata->subreq.len);

	/*
	 * read the rest of READ_RSP header (sans Data array), or whatever we
	 * can if there's not enough data. At this point, we've read down to
	 * the Mid.
	 */
	len = min_t(unsigned int, buflen, server->vals->read_rsp_size) -
							HEADER_SIZE(server) + 1;

	length = cifs_read_from_socket(server,
				       buf + HEADER_SIZE(server) - 1, len);
	if (length < 0)
		return length;
	server->total_read += length;

	if (server->ops->is_session_expired &&
	    server->ops->is_session_expired(buf)) {
		cifs_reconnect(server, true);
		return -1;
	}

	if (server->ops->is_status_pending &&
	    server->ops->is_status_pending(buf, server)) {
		cifs_discard_remaining_data(server);
		return -1;
	}

	/* set up first two iov for signature check and to get credits */
	rdata->iov[0].iov_base = buf;
	rdata->iov[0].iov_len = server->total_read;
	cifs_dbg(FYI, "0: iov_base=%p iov_len=%zu\n",
		 rdata->iov[0].iov_base, rdata->iov[0].iov_len);

	/* Was the SMB read successful? */
	rdata->result = server->ops->map_error(buf, false);
	if (rdata->result != 0) {
		cifs_dbg(FYI, "%s: server returned error %d\n",
			 __func__, rdata->result);
		/* normal error on read response */
		return __cifs_readv_discard(server, mid, false);
	}

	/* Is there enough to get to the rest of the READ_RSP header? */
	if (server->total_read < server->vals->read_rsp_size) {
		cifs_dbg(FYI, "%s: server returned short header. got=%u expected=%zu\n",
			 __func__, server->total_read,
			 server->vals->read_rsp_size);
		rdata->result = smb_EIO2(smb_eio_trace_read_rsp_short,
					 server->total_read, server->vals->read_rsp_size);
		return cifs_readv_discard(server, mid);
	}

	data_offset = server->ops->read_data_offset(buf);
	if (data_offset < server->total_read) {
		/*
		 * win2k8 sometimes sends an offset of 0 when the read
		 * is beyond the EOF. Treat it as if the data starts just after
		 * the header.
		 */
		cifs_dbg(FYI, "%s: data offset (%u) inside read response header\n",
			 __func__, data_offset);
		data_offset = server->total_read;
	} else if (data_offset > MAX_CIFS_SMALL_BUFFER_SIZE) {
		/* data_offset is beyond the end of smallbuf */
		cifs_dbg(FYI, "%s: data offset (%u) beyond end of smallbuf\n",
			 __func__, data_offset);
		rdata->result = smb_EIO1(smb_eio_trace_read_overlarge,
					 data_offset);
		return cifs_readv_discard(server, mid);
	}

	cifs_dbg(FYI, "%s: total_read=%u data_offset=%u\n",
		 __func__, server->total_read, data_offset);

	len = data_offset - server->total_read;
	if (len > 0) {
		/* read any junk before data into the rest of smallbuf */
		length = cifs_read_from_socket(server,
					       buf + server->total_read, len);
		if (length < 0)
			return length;
		server->total_read += length;
		rdata->iov[0].iov_len = server->total_read;
	}

	/* how much data is in the response? */
#ifdef CONFIG_CIFS_SMB_DIRECT
	use_rdma_mr = rdata->mr;
#endif
	data_len = server->ops->read_data_length(buf, use_rdma_mr);
	if (!use_rdma_mr) {
		if (check_add_overflow(data_offset, data_len, &end_off) ||
		    end_off > buflen) {
			/* data_len is corrupt -- discard frame */
			rdata->result = smb_EIO2(smb_eio_trace_read_rsp_malformed,
						 end_off, buflen);
			return cifs_readv_discard(server, mid);
		}
	}

#ifdef CONFIG_CIFS_SMB_DIRECT
	if (rdata->mr)
		length = data_len; /* An RDMA read is already done. */
	else
#endif
		length = cifs_read_iter_from_socket(server, &rdata->subreq.io_iter,
						    data_len);
	if (length > 0)
		rdata->got_bytes += length;
	server->total_read += length;

	cifs_dbg(FYI, "total_read=%u buflen=%u remaining=%u\n",
		 server->total_read, buflen, data_len);

	/* discard anything left over */
	if (server->total_read < buflen)
		return cifs_readv_discard(server, mid);

	dequeue_mid(server, mid, false);
	mid->resp_buf = server->smallbuf;
	server->smallbuf = NULL;
	return length;
}