libata "DMA timeout" fix
The desired effect of a DMA timeout should be to throw an I/O error, but that doesn't appear to be happening.
Those seeing DMA timeout messages, please test this patch.
Kernel hacker note: James B recommended that I implement my own scsi_done() function, which duplicates the real scsi_done() but omits the scsi_delete_timer() call. This is probably the best long term fix, but doing so involves exporting several currently-private bits of SCSI mid-layer, which I would rather not do. Probably best to create a __scsi_done() inside the SCSI mid-layer, and call that.
Jeff, the only user of ->eh_strategy_handler() in any kernel
===== drivers/scsi/libata-core.c 1.19 vs edited =====
drivers/scsi/libata-core.c +8 -0
@@ -2130,6 +2130,14 @@ cmd->result =SAM_STAT_CHECK_CONDITION ; elseata_to_sense_error (qc ); + +/* hack alert! we need this to get past the +* first check in scsi_done(). libata is the +* -only- user of ->eh_strategy_handler() in +* any kernel tree, which exposes some incorrect +* assumptions in the SCSI layer. +*/ +scsi_add_timer (cmd,2000 *HZ ,NULL ); }else { cmd->result =SAM_STAT_GOOD ; }
Quoted 8 lines · Jeff Garzik
=====
drivers/scsi/libata-core.c1.19 vs edited =====drivers/scsi/libata-core.c +1
-0@@ -2130,6 +2130,14 @@ cmd->result =SAM_STAT_CHECK_CONDITION ; elseata_to_sense_error (qc ); +
drivers/scsi/libata-core.c line 2143
Quoted 10 lines
drivers/scsi/libata-core.c +7
-0+/* hack alert! we need this to get past the +* first check in scsi_done(). libata is the +* -only- user of ->eh_strategy_handler() in +* any kernel tree, which exposes some incorrect +* assumptions in the SCSI layer. +*/ +scsi_add_timer (cmd,2000 *HZ ,NULL ); }else { cmd->result =SAM_STAT_GOOD ; }
You can't do this. Supposing there command's delayed, the timer fires and then the command returns with a sense error? The done will go through automatically completing the command, but your strategy handler will still think it has a failed command to handle.
The correct fix is this, I think (uncompiled, but you get the idea):
===== libata-core.c 1.19 vs edited =====
libata-core.c renamed from drivers/scsi/libata-core.c +6 -0
@@ -1972,6 +1972,11 @@/* FIXME */ } +static void ata_eng_timeout_done (struct scsi_cmnd *cmnd) +{ +scsi_finish_command (cmnd); +} +/** * ata_eng_timeout - Handle timeout of queued command * @ap: Port on which timed-out command is active
@@ -2005,6 +2010,7 @@ goto out; } + qc->scsidone = ata_eng_timeout_done;switch (qc->tf.protocol) {case ATA_PROT_DMA_READ :case ATA_PROT_DMA_WRITE :
James
Quoted 18 lines · James Bottomley
Quoted 16 lines · Jeff Garzik
=====
drivers/scsi/libata-core.c1.19 vs edited =====drivers/scsi/libata-core.c +8
-0@@ -2130,6 +2130,14 @@ cmd->result =SAM_STAT_CHECK_CONDITION ; elseata_to_sense_error (qc ); + +/* hack alert! we need this to get past the +* first check in scsi_done(). libata is the +* -only- user of ->eh_strategy_handler() in +* any kernel tree, which exposes some incorrect +* assumptions in the SCSI layer. +*/ +scsi_add_timer (cmd,2000 *HZ ,NULL ); }else {
Quoted 8 lines
drivers/scsi/libata-core.c line 2143
Quoted 2 lines
drivers/scsi/libata-core.c +0
-0cmd->result =SAM_STAT_GOOD ; }You can't do this. Supposing there command's delayed, the timer fires and then the command returns with a sense error? The done will go through automatically completing the command, but your strategy handler will still think it has a failed command to handle.
hmmm, yeah that will be a problem iff we are not already in the strategy handler.
Quoted 1 line
The correct fix is this, I think (uncompiled, but you get the idea):
Yeah, that's much better. That function is not exported though ;-)
Jeff
Quoted 1 line · Jeff Garzik
Yeah, that's much better. That function is not exported though ;-)
I can fix that. It really is a necessary function for drivers doing their own strategy handler ... of which yours seems to be the only one.
James
Thanks, testing James's fix on one machine at the moment. Will have another six or so machines as a libata test farm tomorrow.
Justin
Quoted 45 lines · Jeff Garzik
Quoted 18 lines · James Bottomley
Quoted 16 lines · Jeff Garzik
=====
drivers/scsi/libata-core.c1.19 vs edited =====drivers/scsi/libata-core.c +8
-0@@ -2130,6 +2130,14 @@ cmd->result =SAM_STAT_CHECK_CONDITION ; elseata_to_sense_error (qc ); + +/* hack alert! we need this to get past the +* first check in scsi_done(). libata is the +* -only- user of ->eh_strategy_handler() in +* any kernel tree, which exposes some incorrect +* assumptions in the SCSI layer. +*/ +scsi_add_timer (cmd,2000 *HZ ,NULL ); }else {Quoted 8 lines
drivers/scsi/libata-core.c line 2143
Quoted 2 lines
drivers/scsi/libata-core.c +0
-0cmd->result =SAM_STAT_GOOD ; }You can't do this. Supposing there command's delayed, the timer fires and then the command returns with a sense error? The done will go through automatically completing the command, but your strategy handler will still think it has a failed command to handle.
hmmm, yeah that will be a problem iff we are not already in the strategy handler.
Quoted 1 line
The correct fix is this, I think (uncompiled, but you get the idea):
Yeah, that's much better. That function is not exported though ;-)
Jeff
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
Quoted 2 lines · Justin Cormack
Thanks, testing James's fix on one machine at the moment. Will have another six or so machines as a libata test farm tomorrow.
You'll need a few more changes... here's the final patch.
Jeff
# This is a BitKeeper generated patch for the following project:
# Project Name: Linux kernel tree
# This patch format is intended for GNU patch command version 2.5 or higher. # This patch includes the following deltas:
# ChangeSet 1.1682 -> 1.1683
# drivers/scsi/libata-core.c 1.19 -> 1.20
# drivers/scsi/scsi.c 1.136 -> 1.137
# drivers/scsi/scsi_priv.h 1.30 -> 1.31
# include/scsi/scsi_cmnd.h 1.4 -> 1.5
#
# The following is the BitKeeper ChangeSet Log
# --------------------------------------------
# 04/02/28 [email protected] 1.1683
# [libata] Use scsi_finish_command as completion function,
# in our error handling thread callback. # # This also exports scsi_finish_command in the SCSI layer. # # Thanks much to James Bottomley and his patience, as this solution # was figured out. # -------------------------------------------- # diff -Nru a/drivers/scsi/libata-core.c b/drivers/scsi/libata-core.c --- a/drivers/scsi/libata-core.c Sun Feb 29 12:30:47 2004 +++ b/drivers/scsi/libata-core.c Sun Feb 29 12:30:47 2004 @@ -2005,6 +2005,14 @@ goto out; }
+ /* hack alert! We cannot use the supplied completion
+ * function from inside the ->eh_strategy_handler() thread.
+ * libata is the only user of ->eh_strategy_handler() in
+ * any kernel, so the default scsi_done() assumes it is
+ * not being called from the SCSI EH.
+ */
+ qc->scsidone = scsi_finish_command;
+
switch (qc->tf.protocol) {
case ATA_PROT_DMA_READ:
case ATA_PROT_DMA_WRITE:
drivers/scsi/scsi.c +1 -0
@@ -847,6 +847,7 @@ cmd->done (cmd); } +EXPORT_SYMBOL (scsi_finish_command);/* * Function: scsi_adjust_queue_depth()
drivers/scsi/scsi_priv.h +0 -1
@@ -77,7 +77,6 @@extern int scsi_setup_command_freelist (struct Scsi_Host *shost);extern void scsi_destroy_command_freelist (struct Scsi_Host *shost);extern void scsi_done (struct scsi_cmnd *cmd); -extern void scsi_finish_command (struct scsi_cmnd *cmd);extern int scsi_retry_command (struct scsi_cmnd *cmd);extern int scsi_insert_special_req (struct scsi_request *sreq,int ); extern void scsi_init_cmd_from_req(struct scsi_cmnd *cmd,
include/scsi/scsi_cmnd.h +1 -0
@@ -159,5 +159,6 @@extern struct scsi_cmnd *scsi_get_command (struct scsi_device *,int );extern void scsi_put_command (struct scsi_cmnd *);extern void scsi_io_completion (struct scsi_cmnd *,unsigned int ,unsigned int ); +extern void scsi_finish_command (struct scsi_cmnd *cmd);#endif /* _SCSI_SCSI_CMND_H */