SCSI init discussion/SAN problem

[Evan Rempel <erempel@uvic.ca>]

I have a problem with the order that the SCSI subsystem attaches disk 
devices that shows up in a multipath environment.

My understanding is that during the finishing phase of the SCSI 
subsystem the partition table is read from the drive and the bare drive 
and each partition are registered with the kernel. Please correct me if 
I am wrong becuase I am not a kernel developer even at the tinkering level.

The problem shows up in a multipath environment where the same physical 
device has it's partition table read and then registered with the kernel 
*for each path on which it available*. I understand the requirement for 
the second (possibly more) devices registered with the kernel, and I 
want this behaviour to continue (how else would multipath work?).
The problem is that reading the partition table on each of the paths 
causes I/O to be generated to the physical disk on each of the paths.
For some disk controllers (any with active/passive controllers) this 
will initial a failover event from the active to the passive controller. 
This failover can take a few seconds, but multipathing may result in 
100's of such paths and failover events which make the boot time very 
long. I have a machine that takes close to 1hr to boot due to this behavior.

What I would like to have considered is the ability to get the serial 
number/WWName of the device prior to reading the partition table. If the 
serial number/WWName has already been registered under a different SCSI 
ID, then just use the partition table that was used to load the first 
instance. This will result in I/O only on the first path to each disk.

Another thing that might make things even better is to do something like 
the mp_prio utils of multipathing do and determine which paths is an 
active path, and only read the partition table from the active paths.
This may require a 2 pass device registration mechanism becuase it may 
be possible that none of the paths are active paths, meaning that the 
device did not get registered by the end of the device list. We would 
have to go back to the beginning of the list and for any device that was 
not yet registered with the kernel, read the serial number/WWName and 
partition table, register with the kernel and then determine if any of 
the other paths are for the same device to load them into the kernel.

I hope this is clear enough to start a dialog on how to change the scsi 
initialization faster for large systems on multipath hardware.

Evan Rempel
Senior Programmer Analyst
University of Victoria
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/