diff mbox

bnx2fc: Fix hung task messages when a cleanup response is not received during abort.

Message ID 20171115150606.10994-1-chad.dupuis@cavium.com (mailing list archive)
State Accepted
Headers show

Commit Message

Dupuis, Chad Nov. 15, 2017, 3:06 p.m. UTC
If a cleanup task is not responded to while we are in bnx2fc_abts_cleanup, it
will hang the SCSI error handler since we use wait_for_completion instead of
wait_for_completion_timeout.  So, use wait_for_completion_timeout so that we
don't hang the SCSI error handler thread forever.

Fixes the call trace:

[183373.131468] INFO: task scsi_eh_16:110146 blocked for more than 120 seconds.
[183373.131469] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
[183373.131470] scsi_eh_16      D ffff88103f2fca14     0 110146      2 0x00000080
[183373.131472]  ffff880855e77cb0 0000000000000046 ffff881050654e70 ffff880855e77fd8
[183373.131474]  ffff880855e77fd8 ffff880855e77fd8 ffff881050654e70 ffff88103f2fcb48
[183373.131475]  ffff88103f2fcb50 7fffffffffffffff ffff881050654e70 ffff88103f2fca14
[183373.131477] Call Trace:
[183373.131479]  [<ffffffff8168b579>] schedule+0x29/0x70
[183373.131481]  [<ffffffff81688fc9>] schedule_timeout+0x239/0x2d0
[183373.131486]  [<ffffffff8142821e>] ? __dev_printk+0x3e/0x90
[183373.131487]  [<ffffffff814282cd>] ? dev_printk+0x5d/0x80
[183373.131490]  [<ffffffff8168b956>] wait_for_completion+0x116/0x170
[183373.131492]  [<ffffffff810c4ec0>] ? wake_up_state+0x20/0x20
[183373.131494]  [<ffffffffa048c234>] bnx2fc_abts_cleanup+0x3d/0x62 [bnx2fc]
[183373.131497]  [<ffffffffa0483a80>] bnx2fc_eh_abort+0x470/0x580 [bnx2fc]
[183373.131500]  [<ffffffff814570af>] scsi_error_handler+0x59f/0x8b0
[183373.131501]  [<ffffffff81456b10>] ? scsi_eh_get_sense+0x250/0x250
[183373.131503]  [<ffffffff810b052f>] kthread+0xcf/0xe0
[183373.131505]  [<ffffffff810b0460>] ? kthread_create_on_node+0x140/0x140
[183373.131507]  [<ffffffff81696418>] ret_from_fork+0x58/0x90
[183373.131509]  [<ffffffff810b0460>] ? kthread_create_on_node+0x140/0x140

Signed-off-by: Chad Dupuis <chad.dupuis@cavium.com>
---
 drivers/scsi/bnx2fc/bnx2fc_io.c | 40 ++++++++++++++++++++++++++++++++--------
 1 file changed, 32 insertions(+), 8 deletions(-)

Comments

Laurence Oberman Nov. 15, 2017, 3:21 p.m. UTC | #1
On Wed, 2017-11-15 at 07:06 -0800, Chad Dupuis wrote:
> If a cleanup task is not responded to while we are in
> bnx2fc_abts_cleanup, it
> will hang the SCSI error handler since we use wait_for_completion
> instead of
> wait_for_completion_timeout.  So, use wait_for_completion_timeout so
> that we
> don't hang the SCSI error handler thread forever.
> 
> Fixes the call trace:
> 
> [183373.131468] INFO: task scsi_eh_16:110146 blocked for more than
> 120 seconds.
> [183373.131469] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs"
> disables this message.
> [183373.131470] scsi_eh_16      D ffff88103f2fca14     0
> 110146      2 0x00000080
> [183373.131472]  ffff880855e77cb0 0000000000000046 ffff881050654e70
> ffff880855e77fd8
> [183373.131474]  ffff880855e77fd8 ffff880855e77fd8 ffff881050654e70
> ffff88103f2fcb48
> [183373.131475]  ffff88103f2fcb50 7fffffffffffffff ffff881050654e70
> ffff88103f2fca14
> [183373.131477] Call Trace:
> [183373.131479]  [<ffffffff8168b579>] schedule+0x29/0x70
> [183373.131481]  [<ffffffff81688fc9>] schedule_timeout+0x239/0x2d0
> [183373.131486]  [<ffffffff8142821e>] ? __dev_printk+0x3e/0x90
> [183373.131487]  [<ffffffff814282cd>] ? dev_printk+0x5d/0x80
> [183373.131490]  [<ffffffff8168b956>] wait_for_completion+0x116/0x170
> [183373.131492]  [<ffffffff810c4ec0>] ? wake_up_state+0x20/0x20
> [183373.131494]  [<ffffffffa048c234>] bnx2fc_abts_cleanup+0x3d/0x62
> [bnx2fc]
> [183373.131497]  [<ffffffffa0483a80>] bnx2fc_eh_abort+0x470/0x580
> [bnx2fc]
> [183373.131500]  [<ffffffff814570af>] scsi_error_handler+0x59f/0x8b0
> [183373.131501]  [<ffffffff81456b10>] ? scsi_eh_get_sense+0x250/0x250
> [183373.131503]  [<ffffffff810b052f>] kthread+0xcf/0xe0
> [183373.131505]  [<ffffffff810b0460>] ?
> kthread_create_on_node+0x140/0x140
> [183373.131507]  [<ffffffff81696418>] ret_from_fork+0x58/0x90
> [183373.131509]  [<ffffffff810b0460>] ?
> kthread_create_on_node+0x140/0x140
> 
> Signed-off-by: Chad Dupuis <chad.dupuis@cavium.com>
> ---
>  drivers/scsi/bnx2fc/bnx2fc_io.c | 40
> ++++++++++++++++++++++++++++++++--------
>  1 file changed, 32 insertions(+), 8 deletions(-)
> 
> diff --git a/drivers/scsi/bnx2fc/bnx2fc_io.c
> b/drivers/scsi/bnx2fc/bnx2fc_io.c
> index 5b6153f23f01..2cbb5be98ecb 100644
> --- a/drivers/scsi/bnx2fc/bnx2fc_io.c
> +++ b/drivers/scsi/bnx2fc/bnx2fc_io.c
> @@ -1084,24 +1084,35 @@ static int bnx2fc_abts_cleanup(struct
> bnx2fc_cmd *io_req)
>  {
>  	struct bnx2fc_rport *tgt = io_req->tgt;
>  	int rc = SUCCESS;
> +	unsigned int time_left;
>  
>  	io_req->wait_for_comp = 1;
>  	bnx2fc_initiate_cleanup(io_req);
>  
>  	spin_unlock_bh(&tgt->tgt_lock);
>  
> -	wait_for_completion(&io_req->tm_done);
> -
> +	/*
> +	 * Can't wait forever on cleanup response lest we let the
> SCSI error
> +	 * handler wait forever
> +	 */
> +	time_left = wait_for_completion_timeout(&io_req->tm_done,
> +	    BNX2FC_FW_TIMEOUT);
>  	io_req->wait_for_comp = 0;
> +	if (!time_left)
> +		BNX2FC_IO_DBG(io_req, "%s(): Wait for cleanup timed
> out.\n",
> +		    __func__);
> +
>  	/*
> -	 * release the reference taken in eh_abort to allow the
> -	 * target to re-login after flushing IOs
> +	 * Release reference held by SCSI command the cleanup
> completion
> +	 * hits the BNX2FC_CLEANUP case in bnx2fc_process_cq_compl()
> and
> +	 * thus the SCSI command is not returnedi by
> bnx2fc_scsi_done().
>  	 */
>  	kref_put(&io_req->refcount, bnx2fc_cmd_release);
>  
>  	spin_lock_bh(&tgt->tgt_lock);
>  	return rc;
>  }
> +
>  /**
>   * bnx2fc_eh_abort - eh_abort_handler api to abort an outstanding
>   *			SCSI command
> @@ -1118,6 +1129,7 @@ int bnx2fc_eh_abort(struct scsi_cmnd *sc_cmd)
>  	struct fc_lport *lport;
>  	struct bnx2fc_rport *tgt;
>  	int rc;
> +	unsigned int time_left;
>  
>  	rc = fc_block_scsi_eh(sc_cmd);
>  	if (rc)
> @@ -1194,6 +1206,11 @@ int bnx2fc_eh_abort(struct scsi_cmnd *sc_cmd)
>  		if (cancel_delayed_work(&io_req->timeout_work))
>  			kref_put(&io_req->refcount,
>  				 bnx2fc_cmd_release); /* drop timer
> hold */
> +		/*
> +		 * We don't want to hold off the upper layer timer
> so simply
> +		 * cleanup the command and return that I/O was
> successfully
> +		 * aborted.
> +		 */
>  		rc = bnx2fc_abts_cleanup(io_req);
>  		/* This only occurs when an task abort was requested
> while ABTS
>  		   is in progress.  Setting the IO_CLEANUP flag will
> skip the
> @@ -1201,7 +1218,7 @@ int bnx2fc_eh_abort(struct scsi_cmnd *sc_cmd)
>  		   was a result from the ABTS request rather than
> the CLEANUP
>  		   request */
>  		set_bit(BNX2FC_FLAG_IO_CLEANUP,	&io_req-
> >req_flags);
> -		goto out;
> +		goto done;
>  	}
>  
>  	/* Cancel the current timer running on this io_req */
> @@ -1221,7 +1238,11 @@ int bnx2fc_eh_abort(struct scsi_cmnd *sc_cmd)
>  	}
>  	spin_unlock_bh(&tgt->tgt_lock);
>  
> -	wait_for_completion(&io_req->tm_done);
> +	/* Wait 2 * RA_TOV + 1 to be sure timeout function hasn't
> fired */
> +	time_left = wait_for_completion_timeout(&io_req->tm_done,
> +	    (2 * rp->r_a_tov + 1) * HZ);
> +	if (time_left)
> +		BNX2FC_IO_DBG(io_req, "Timed out in eh_abort waiting
> for tm_done");
>  
>  	spin_lock_bh(&tgt->tgt_lock);
>  	io_req->wait_for_comp = 0;
> @@ -1233,8 +1254,12 @@ int bnx2fc_eh_abort(struct scsi_cmnd *sc_cmd)
>  		/* Let the scsi-ml try to recover this command */
>  		printk(KERN_ERR PFX "abort failed, xid = 0x%x\n",
>  		       io_req->xid);
> +		/*
> +		 * Cleanup firmware residuals before returning
> control back
> +		 * to SCSI ML.
> +		 */
>  		rc = bnx2fc_abts_cleanup(io_req);
> -		goto out;
> +		goto done;
>  	} else {
>  		/*
>  		 * We come here even when there was a race condition
> @@ -1249,7 +1274,6 @@ int bnx2fc_eh_abort(struct scsi_cmnd *sc_cmd)
>  done:
>  	/* release the reference taken in eh_abort */
>  	kref_put(&io_req->refcount, bnx2fc_cmd_release);
> -out:
>  	spin_unlock_bh(&tgt->tgt_lock);
>  	return rc;
>  }

We experienced this at a major customer and provided this patch after
working with Chad and backported to RHEL as a test kernel.

Its been stable while running a test kernel and the changes look good
ate least to me.

Reviewed-by: Laurence Oberman <loberman@redhat.com>
Tested-by:   Laurence Oberman <loberman@redhat.com>

Thanks
Laurence
Martin K. Petersen Nov. 15, 2017, 11:45 p.m. UTC | #2
Chad,

> If a cleanup task is not responded to while we are in
> bnx2fc_abts_cleanup, it will hang the SCSI error handler since we use
> wait_for_completion instead of wait_for_completion_timeout.  So, use
> wait_for_completion_timeout so that we don't hang the SCSI error
> handler thread forever.

Applied to 4.15/scsi-fixes. Thanks!
diff mbox

Patch

diff --git a/drivers/scsi/bnx2fc/bnx2fc_io.c b/drivers/scsi/bnx2fc/bnx2fc_io.c
index 5b6153f23f01..2cbb5be98ecb 100644
--- a/drivers/scsi/bnx2fc/bnx2fc_io.c
+++ b/drivers/scsi/bnx2fc/bnx2fc_io.c
@@ -1084,24 +1084,35 @@  static int bnx2fc_abts_cleanup(struct bnx2fc_cmd *io_req)
 {
 	struct bnx2fc_rport *tgt = io_req->tgt;
 	int rc = SUCCESS;
+	unsigned int time_left;
 
 	io_req->wait_for_comp = 1;
 	bnx2fc_initiate_cleanup(io_req);
 
 	spin_unlock_bh(&tgt->tgt_lock);
 
-	wait_for_completion(&io_req->tm_done);
-
+	/*
+	 * Can't wait forever on cleanup response lest we let the SCSI error
+	 * handler wait forever
+	 */
+	time_left = wait_for_completion_timeout(&io_req->tm_done,
+	    BNX2FC_FW_TIMEOUT);
 	io_req->wait_for_comp = 0;
+	if (!time_left)
+		BNX2FC_IO_DBG(io_req, "%s(): Wait for cleanup timed out.\n",
+		    __func__);
+
 	/*
-	 * release the reference taken in eh_abort to allow the
-	 * target to re-login after flushing IOs
+	 * Release reference held by SCSI command the cleanup completion
+	 * hits the BNX2FC_CLEANUP case in bnx2fc_process_cq_compl() and
+	 * thus the SCSI command is not returnedi by bnx2fc_scsi_done().
 	 */
 	kref_put(&io_req->refcount, bnx2fc_cmd_release);
 
 	spin_lock_bh(&tgt->tgt_lock);
 	return rc;
 }
+
 /**
  * bnx2fc_eh_abort - eh_abort_handler api to abort an outstanding
  *			SCSI command
@@ -1118,6 +1129,7 @@  int bnx2fc_eh_abort(struct scsi_cmnd *sc_cmd)
 	struct fc_lport *lport;
 	struct bnx2fc_rport *tgt;
 	int rc;
+	unsigned int time_left;
 
 	rc = fc_block_scsi_eh(sc_cmd);
 	if (rc)
@@ -1194,6 +1206,11 @@  int bnx2fc_eh_abort(struct scsi_cmnd *sc_cmd)
 		if (cancel_delayed_work(&io_req->timeout_work))
 			kref_put(&io_req->refcount,
 				 bnx2fc_cmd_release); /* drop timer hold */
+		/*
+		 * We don't want to hold off the upper layer timer so simply
+		 * cleanup the command and return that I/O was successfully
+		 * aborted.
+		 */
 		rc = bnx2fc_abts_cleanup(io_req);
 		/* This only occurs when an task abort was requested while ABTS
 		   is in progress.  Setting the IO_CLEANUP flag will skip the
@@ -1201,7 +1218,7 @@  int bnx2fc_eh_abort(struct scsi_cmnd *sc_cmd)
 		   was a result from the ABTS request rather than the CLEANUP
 		   request */
 		set_bit(BNX2FC_FLAG_IO_CLEANUP,	&io_req->req_flags);
-		goto out;
+		goto done;
 	}
 
 	/* Cancel the current timer running on this io_req */
@@ -1221,7 +1238,11 @@  int bnx2fc_eh_abort(struct scsi_cmnd *sc_cmd)
 	}
 	spin_unlock_bh(&tgt->tgt_lock);
 
-	wait_for_completion(&io_req->tm_done);
+	/* Wait 2 * RA_TOV + 1 to be sure timeout function hasn't fired */
+	time_left = wait_for_completion_timeout(&io_req->tm_done,
+	    (2 * rp->r_a_tov + 1) * HZ);
+	if (time_left)
+		BNX2FC_IO_DBG(io_req, "Timed out in eh_abort waiting for tm_done");
 
 	spin_lock_bh(&tgt->tgt_lock);
 	io_req->wait_for_comp = 0;
@@ -1233,8 +1254,12 @@  int bnx2fc_eh_abort(struct scsi_cmnd *sc_cmd)
 		/* Let the scsi-ml try to recover this command */
 		printk(KERN_ERR PFX "abort failed, xid = 0x%x\n",
 		       io_req->xid);
+		/*
+		 * Cleanup firmware residuals before returning control back
+		 * to SCSI ML.
+		 */
 		rc = bnx2fc_abts_cleanup(io_req);
-		goto out;
+		goto done;
 	} else {
 		/*
 		 * We come here even when there was a race condition
@@ -1249,7 +1274,6 @@  int bnx2fc_eh_abort(struct scsi_cmnd *sc_cmd)
 done:
 	/* release the reference taken in eh_abort */
 	kref_put(&io_req->refcount, bnx2fc_cmd_release);
-out:
 	spin_unlock_bh(&tgt->tgt_lock);
 	return rc;
 }