Commit 38303521 authored by Tony Camuso's avatar Tony Camuso Committed by Corey Minyard
Browse files

ipmi: move message error checking to avoid deadlock



V1->V2: in handle_one_rcv_msg, if data_size > 2, set requeue to zero and
        goto out instead of calling ipmi_free_msg.
        Kosuke Tatsukawa <tatsu@ab.jp.nec.com>

In the source stack trace below, function set_need_watch tries to
take out the same si_lock that was taken earlier by ipmi_thread.

ipmi_thread() [drivers/char/ipmi/ipmi_si_intf.c:995]
 smi_event_handler() [drivers/char/ipmi/ipmi_si_intf.c:765]
  handle_transaction_done() [drivers/char/ipmi/ipmi_si_intf.c:555]
   deliver_recv_msg() [drivers/char/ipmi/ipmi_si_intf.c:283]
    ipmi_smi_msg_received() [drivers/char/ipmi/ipmi_msghandler.c:4503]
     intf_err_seq() [drivers/char/ipmi/ipmi_msghandler.c:1149]
      smi_remove_watch() [drivers/char/ipmi/ipmi_msghandler.c:999]
       set_need_watch() [drivers/char/ipmi/ipmi_si_intf.c:1066]

Upstream commit e1891cff adds code to
ipmi_smi_msg_received() to call smi_remove_watch() via intf_err_seq()
and this seems to be causing the deadlock.

commit e1891cff
Author: Corey Minyard <cminyard@mvista.com>
Date:   Wed Oct 24 15:17:04 2018 -0500
    ipmi: Make the smi watcher be disabled immediately when not needed

The fix is to put all messages in the queue and move the message
checking code out of ipmi_smi_msg_received and into handle_one_recv_msg,
which processes the message checking after ipmi_thread releases its
locks.

Additionally,Kosuke Tatsukawa <tatsu@ab.jp.nec.com> reported that
handle_new_recv_msgs calls ipmi_free_msg when handle_one_rcv_msg returns
zero, so that the call to ipmi_free_msg in handle_one_rcv_msg introduced
another panic when "ipmitool sensor list" was run in a loop. He
submitted this part of the patch.

+free_msg:
+               requeue = 0;
+               goto out;

Reported by: Osamu Samukawa <osa-samukawa@tg.jp.nec.com>
Characterized by: Kosuke Tatsukawa <tatsu@ab.jp.nec.com>
Signed-off-by: default avatarTony Camuso <tcamuso@redhat.com>
Fixes: e1891cff ("ipmi: Make the smi watcher be disabled immediately when not needed")
Cc: stable@vger.kernel.org # 5.1
Signed-off-by: default avatarCorey Minyard <cminyard@mvista.com>
parent c4436c91
Loading
Loading
Loading
Loading
+57 −57
Original line number Diff line number Diff line
@@ -4218,7 +4218,53 @@ static int handle_one_recv_msg(struct ipmi_smi *intf,
	int chan;

	ipmi_debug_msg("Recv:", msg->rsp, msg->rsp_size);
	if (msg->rsp_size < 2) {

	if ((msg->data_size >= 2)
	    && (msg->data[0] == (IPMI_NETFN_APP_REQUEST << 2))
	    && (msg->data[1] == IPMI_SEND_MSG_CMD)
	    && (msg->user_data == NULL)) {

		if (intf->in_shutdown)
			goto free_msg;

		/*
		 * This is the local response to a command send, start
		 * the timer for these.  The user_data will not be
		 * NULL if this is a response send, and we will let
		 * response sends just go through.
		 */

		/*
		 * Check for errors, if we get certain errors (ones
		 * that mean basically we can try again later), we
		 * ignore them and start the timer.  Otherwise we
		 * report the error immediately.
		 */
		if ((msg->rsp_size >= 3) && (msg->rsp[2] != 0)
		    && (msg->rsp[2] != IPMI_NODE_BUSY_ERR)
		    && (msg->rsp[2] != IPMI_LOST_ARBITRATION_ERR)
		    && (msg->rsp[2] != IPMI_BUS_ERR)
		    && (msg->rsp[2] != IPMI_NAK_ON_WRITE_ERR)) {
			int ch = msg->rsp[3] & 0xf;
			struct ipmi_channel *chans;

			/* Got an error sending the message, handle it. */

			chans = READ_ONCE(intf->channel_list)->c;
			if ((chans[ch].medium == IPMI_CHANNEL_MEDIUM_8023LAN)
			    || (chans[ch].medium == IPMI_CHANNEL_MEDIUM_ASYNC))
				ipmi_inc_stat(intf, sent_lan_command_errs);
			else
				ipmi_inc_stat(intf, sent_ipmb_command_errs);
			intf_err_seq(intf, msg->msgid, msg->rsp[2]);
		} else
			/* The message was sent, start the timer. */
			intf_start_seq_timer(intf, msg->msgid);
free_msg:
		requeue = 0;
		goto out;

	} else if (msg->rsp_size < 2) {
		/* Message is too small to be correct. */
		dev_warn(intf->si_dev,
			 "BMC returned too small a message for netfn %x cmd %x, got %d bytes\n",
@@ -4475,51 +4521,6 @@ void ipmi_smi_msg_received(struct ipmi_smi *intf,
	unsigned long flags = 0; /* keep us warning-free. */
	int run_to_completion = intf->run_to_completion;

	if ((msg->data_size >= 2)
	    && (msg->data[0] == (IPMI_NETFN_APP_REQUEST << 2))
	    && (msg->data[1] == IPMI_SEND_MSG_CMD)
	    && (msg->user_data == NULL)) {

		if (intf->in_shutdown)
			goto free_msg;

		/*
		 * This is the local response to a command send, start
		 * the timer for these.  The user_data will not be
		 * NULL if this is a response send, and we will let
		 * response sends just go through.
		 */

		/*
		 * Check for errors, if we get certain errors (ones
		 * that mean basically we can try again later), we
		 * ignore them and start the timer.  Otherwise we
		 * report the error immediately.
		 */
		if ((msg->rsp_size >= 3) && (msg->rsp[2] != 0)
		    && (msg->rsp[2] != IPMI_NODE_BUSY_ERR)
		    && (msg->rsp[2] != IPMI_LOST_ARBITRATION_ERR)
		    && (msg->rsp[2] != IPMI_BUS_ERR)
		    && (msg->rsp[2] != IPMI_NAK_ON_WRITE_ERR)) {
			int ch = msg->rsp[3] & 0xf;
			struct ipmi_channel *chans;

			/* Got an error sending the message, handle it. */

			chans = READ_ONCE(intf->channel_list)->c;
			if ((chans[ch].medium == IPMI_CHANNEL_MEDIUM_8023LAN)
			    || (chans[ch].medium == IPMI_CHANNEL_MEDIUM_ASYNC))
				ipmi_inc_stat(intf, sent_lan_command_errs);
			else
				ipmi_inc_stat(intf, sent_ipmb_command_errs);
			intf_err_seq(intf, msg->msgid, msg->rsp[2]);
		} else
			/* The message was sent, start the timer. */
			intf_start_seq_timer(intf, msg->msgid);

free_msg:
		ipmi_free_smi_msg(msg);
	} else {
	/*
	 * To preserve message order, we keep a queue and deliver from
	 * a tasklet.
@@ -4530,7 +4531,6 @@ free_msg:
	if (!run_to_completion)
		spin_unlock_irqrestore(&intf->waiting_rcv_msgs_lock,
				       flags);
	}

	if (!run_to_completion)
		spin_lock_irqsave(&intf->xmit_msgs_lock, flags);