finding out whether a device supports ordered writes ahead of time
Currently ext3 and reiserfs submit bios with BIO_RW_BARRIER and when the device doesn't support it it returns EOPNOTUPP. This scheme doesn't work at all for XFS because our I/O submission path keeps around far too much state (XFS supports multi-page metadata buffers). From looking at the code it seems that we can assume such a submission will just work if q->ordered is not QUEUE_ORDERED_NONE. Is that a valid assumption? and if yes should we look directly at the queue or provide an assecor?
Quoted 7 lines · Christoph Hellwig
Currently ext3 and reiserfs submit bios with
BIO_RW_BARRIERand when the device doesn't support it it returns EOPNOTUPP. This scheme doesn't work at all for XFS because our I/O submission path keeps around far too much state (XFS supports multi-page metadata buffers). From looking at the code it seems that we can assume such a submission will just work ifq->orderedis notQUEUE_ORDERED_NONE. Is that a valid assumption? and if yes should we look directly at the queue or provide an assecor?
I think Jens currently has things set to trust the drive's advertisement of the cache flush feature. But I don't think we've looked hard for drives that advertise and then fail the barriers, it wouldn't surprise me if one existed.
What you could do is issue a barrier write during mount to test things. If that works any failed barrier later could be considered an io error. Don't use blkdev_issue_flush for this, since it will work on scsi drives that don't support the BIO_RW_BARRIER.
-chris