Re: libata passthru: support PIO multi commands

From: Albert Lee
Date: Tue Jun 12 2007 - 22:57:42 EST

Alan Cox wrote:
>>ata_scsi_pass_thru() is not executed at ioctl submission time (block
>>queue submission time), but rather immediately before it is issued to
>>the drive. At that point you know the bus is idle, all other commands
>>have finished executing, and dev->multi_count is fresh not stale. The
>>code path goes from ata_scsi_pass_thru() to ata_qc_issue() without
>>releasing the spinlock, even.
> Think up to user space
> Poorusersapp set multicount to 4
> Evilproprietarytuningdaemon set multicount to 8
> Poorusersapp issue I/O
> at which point an error is indeed best.
>>But the last point is true -- we should error rather than just warn
>>there, AFAICS.
> Definitely. We've been asked "please do something stupid" and not even in
> a case where the requester may know better.

It looks like the ATA passthru commands contain more information than
what libata needs to execute a command.

e.g. protocol number:
libata could possibly infer the protocol from the command opcode.

e.g. multi_count:
libata caches dev->multi_count. Passing multi_count along with
each passthru command looks useless for libata.

e.g. t_dir:
libata could possible infer the direction from the command opcode
or from the protocol number (e.g. 4: PIO_IN / 5: PIO_OUT).

Due to the redundant info, there is possiblely inconsistency between
the parameters. e.g. t_dir vs protocol. e.g. command vs protocol.

It seems the "redundant" parameters are designed to allow stateless SATL
implementation: The application/passthru command tells the stateless SATL
implementation the protocol and the multi_count, etc. Then SATL just
follows the instruction blindly, even if asked to do something stupid.

Currently libata
- uses the passthru protocol number blindly
(even if the application issues a DMA command with wrong PIO protocol.)
- checks and warns about multi_count
- ignores t_dir, byte_block and so on.

Maybe we need a strategy to deal with incorrect passed-thru commands?
- check and reject if something wrong
- mimimal check and warn/ignore, if it doesn't hurt command execution


To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at
Please read the FAQ at