2012/3/22 Martin K. Petersen [off-list ref]:

quoted

quoted

quoted

quoted

quoted

"Shaohua" == Shaohua Li [off-list ref] writes:

Shaohua> Enabling discard merge is required for device with slow discard
Shaohua> (and very helpful for raid),

Before RAID support there really wasn't much point. That's one of the
reasons the current discard merge code has survived despite being
completely broken.

Exactly.

As for how I'd like to handle contiguous discard merges going forward
here's a proof of concept. It's on top of my write same tree so you may
have to noodle a bit.

It is crucially important that we never merge a command that can't be
processed by the device. So I'd like to do something like this:

This is great to consolidate the check with this patch. But looks this
isn't enough. As far as I know the current discard merge is broken
because blk_update_request() can't handle such request because
we override the first BIO to add payload. I had a quick hack and post
out yesterday. It works in my SATA ssd, but I really didn't know if there
is something missing in SCSI side, can you share anything here?

Thanks,
Shaohua