The following analysis is based on Mandrake 2.4.20-2mdk kernel, but the problems exists since 2.4.18 and probably earlier so it is unlikely to be a Mandrake-specific. I have pure IDE hardware; situation may be different on SCSI.
It appears that when do_generic_file_read queues readahead requests, it never ever runs tq_disk to trigger actual read. And in the worst case (when it is the first request in queue) it means that device queue remains plugged until next run_task_queue run. The effects are
- read-ahead may be delayed for as long as next read from device. In case of busy system doing much disk IO it is not as obvious (because tq_disk is run often); in case of single-user system running single-threaded aplication it makes read-ahead actually useless.
- it kills supermount. Supermount checks for media change on every file operation (sans actual read). For IDE ide_do_request blocks until queue is unplugged. In the worst case it happens first on next kupdated run i.e. 5 seconds. That explains terrible performance of supermount on CDs under some usage patterns (rpm /mnt/cdrom/*.rpm being the best example).
Comment at the end of generic_file_readahead suggets that it should unplug the queue:
/*
* If we tried to read ahead some pages,
* If we tried to read ahead asynchronously,
* Try to force unplug of the device in order to start an asynchronous
* read IO request.
but it never happens as far as I can tell.
Is it intended behaviour?
-andrey
P.S. Please Cc on replies as I am not on lkml
P.P.S. Juan, Chmouel, I have patch for yet another bug that makes IDE CD-ROMs usable with supermount again, need to verify it. Unfortunately it cannot be generalized for HDs or SCSI devices.
>>>>> "andrey" == Andrey Borzenkov <[email protected]> writes:
The last time that I looked, file_readahead was basically not used
in the whole kernel. Sorry for the delay.
Later, Juan.
andrey> The following analysis is based on Mandrake 2.4.20-2mdk kernel, but the problems exists since 2.4.18 and probably earlier so it is unlikely to be a Mandrake-specific. I have pure IDE hardware; situation may be different on SCSI.
andrey> It appears that when do_generic_file_read queues readahead requests, it never ever runs tq_disk to trigger actual read. And in the worst case (when it is the first request in queue) it means that device queue remains plugged until next run_task_queue run. The effects are
andrey> - read-ahead may be delayed for as long as next read from device. In case of busy system doing much disk IO it is not as obvious (because tq_disk is run often); in case of single-user system running single-threaded aplication it makes read-ahead actually useless.
andrey> - it kills supermount. Supermount checks for media change on every file operation (sans actual read). For IDE ide_do_request blocks until queue is unplugged. In the worst case it happens first on next kupdated run i.e. 5 seconds. That explains terrible performance of supermount on CDs under some usage patterns (rpm /mnt/cdrom/*.rpm being the best example).
andrey> Comment at the end of generic_file_readahead suggets that it should unplug the queue:
andrey> /*
andrey> * If we tried to read ahead some pages,
andrey> * If we tried to read ahead asynchronously,
andrey> * Try to force unplug of the device in order to start an asynchronous
andrey> * read IO request.
andrey> but it never happens as far as I can tell.
andrey> Is it intended behaviour?
andrey> -andrey
andrey> P.S. Please Cc on replies as I am not on lkml
andrey> P.P.S. Juan, Chmouel, I have patch for yet another bug that makes IDE CD-ROMs usable with supermount again, need to verify it. Unfortunately it cannot be generalized for HDs or SCSI devices.
--
In theory, practice and theory are the same, but in practice they
are different -- Larry McVoy