Re: [PATCH 2/2] loop: Better discard support for block devices

From: Evan Green
Date: Tue Nov 27 2018 - 18:34:47 EST


Hi Ming,

On Mon, Nov 26, 2018 at 6:55 PM Ming Lei <tom.leiming@xxxxxxxxx> wrote:
>
> On Tue, Nov 27, 2018 at 2:55 AM Evan Green <evgreen@xxxxxxxxxxxx> wrote:
> >
> > On Tue, Oct 30, 2018 at 4:06 PM Evan Green <evgreen@xxxxxxxxxxxx> wrote:
> > >
> > > If the backing device for a loop device is a block device,
>
> This shouldn't be a very common use case wrt. loop.

Yeah, I'm starting to gather that. Or maybe I'm just the first one to
mention it on the kernel lists ;) We've used this in our Chrome OS
installer, I believe for many years. Gwendal piped in with a few
reasons we do it this way on the cover letter, but in general I think
it allows us to have a unified set of functions to install to a file,
disk, or prepare an image that may have a different block size than
those on the running system.

>
> > > then mirror the discard properties of the underlying block
> > > device into the loop device. While in there, differentiate
> > > between REQ_OP_DISCARD and REQ_OP_WRITE_ZEROES, which are
> > > different for block devices, but which the loop device had
> > > just been lumping together.
> > >
> > > Signed-off-by: Evan Green <evgreen@xxxxxxxxxxxx>
> >
> > Any thoughts on this patch? This fixes issues for us when using a loop
> > device backed by a block device, where we see many logs like:
> >
> > [ 372.767286] print_req_error: I/O error, dev loop5, sector 88125696
>
> Seems not see any explanation about this IO error and the fix in your patch.
> Could you describe it a bit more?

Sure, I probably should have included more context with the series.

The loop device always reports that it supports discard, by setting up
the max_discard_sectors and max_write_zeroes_sectors in the blk queue.
When the loop device gets a discard or write-zeroes request, it turns
around and calls fallocate on the underlying device with the
PUNCH_HOLE flag. This makes sense when you're backed by a file and
hoping to just deallocate the space, but may fail when you're backed
by a block device that doesn't support discard, or doesn't write
zeroes to discarded sectors. Weirdly, lo_discard already had some code
for preserving EOPNOTSUPP, but then later the error is smashed into
EIO. Patch 1 pipes out EOPNOTSUPP properly, so it doesn't get squashed
into EIO.

Patch 2 reflects the discard characteristics of the underlying device
into the loop device. That way, if you're backed by a file or a block
device that does support discard, everything works great, and user
mode can even see and use the correct discard and write zero
granularities. If you're backed by a block device that does not
support discard, this is exposed to user mode, which then usually
avoids calling fallocate, and doesn't feel betrayed that their
requests are unexpectedly failing.

-Evan