Re: [PATCH] btrfs: also add stripe entries for NOCOW writes

[Josef Bacik <josef@toxicpanda.com>]

On Mon, Sep 23, 2024 at 04:58:34PM +0930, Qu Wenruo wrote:
> 
> 
> 在 2024/9/23 16:15, Johannes Thumshirn 写道:
> > From: Johannes Thumshirn <johannes.thumshirn@wdc.com>
> > 
> > NOCOW writes do not generate stripe_extent entries in the RAID stripe
> > tree, as the RAID stripe-tree feature initially was designed with a
> > zoned filesystem in mind and on a zoned filesystem, we do not allow NOCOW
> > writes. But the RAID stripe-tree feature is independent from the zoned
> > feature, so we must also allow NOCOW writes for zoned filesystems.
> > 
> > Signed-off-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
> 
> Sorry I'm going to repeat myself again, I still believe if we insert an
> RST entry at falloc() time, it will be more consistent with the non-RST
> code.
> 
> Yes, I known preallocated space will not need any read nor search RST
> entry, and we just fill the page cache with zero at read time.
> 
> But the point of proper (not just dummy) RST entry for the whole
> preallocated space is, we do not need to touch the RST entry anymore for
> NOCOW/PREALLOCATED write at all.
> 
> This makes the RST NOCOW/PREALLOC writes behavior to align with the
> non-RST code, which doesn't update any extent item, but only modify the
> file extent for PREALLOC writes.

I see what you're getting at here, but it creates a huge amount of problems
later down the line.

I prealloc an extent, I map that logical extent to a physical extent and then I
insert a RST entry for that mapping.  Now I rip out one of my disks, and now I
have to update RST entries for extents I'm not going to rewrite because they're
prealloc.

RST is a logical->physical mapping.  We do not need to update or insert anything
until we create that logical->physical mapping.  Keeping the rules consistent
across the different layers will make it easier to reason about and easier to
maintain.  Adding an index at endio time for NOCOW makes sense, we now have
created a thing on disk that we need to have a mapping for.  The same goes for
prealloc, adding an entry at prealloc time doesn't make logical sense as we
haven't yet instantiated that space on disk.  Thanks,

Josef