quoted

quoted

quoted

quoted

"Allison" == Allison Henderson [off-list ref] writes:

quoted

Does sb_issue_zeroout() use the SCSI "write same" feature in the
background?  That would avoid busying the CPU/controller/bus with
writing out zeroes, which might be expensive for a large file.

Allison> Hmm, that's a good question, I will dig into it and see if I
Allison> can find out.

This is a bit of an ongoing project.

Unfortunately WRITE SAME is quirk central as many drives only implement
block ranges corresponding to what RAID vendors told them they needed.
Many of the drives I tested have internal caps at 16 or 32MB and will
fail in interesting (i.e. not necessarily graceful) ways if given bigger
ranges.

I've been lobbying for a way for devices to report their WRITE SAME
limit for a while. That feature finally made it into the latest SBC3
draft. The changes required to support it went into the SCSI layer
during the last merge window. What remains is wiring this up to
blkdev_issue_zeroout(). I have some patches sitting in my queue that I
hope to get polished and submitted soon.

Anyway. From a filesystem perspective sb_issue_zeroout() interface is
definitely the way to go. WRITE SAME will eventually be called if the
device supports it.

-- 
Martin K. Petersen	Oracle Linux Engineering