Re: [PATCH v10 06/14] btrfs: optionally extend i_size in cow_file_range_inline()

[Omar Sandoval <osandov@osandov.com>]

On Tue, Aug 24, 2021 at 07:32:06AM +0800, Qu Wenruo wrote:
> 
> 
> On 2021/8/24 上午2:16, Omar Sandoval wrote:
> > On Sat, Aug 21, 2021 at 09:11:26AM +0800, Qu Wenruo wrote:
> > > 
> > > 
> > > On 2021/8/21 上午2:11, Omar Sandoval wrote:
> > > > On Fri, Aug 20, 2021 at 05:13:34PM +0800, Qu Wenruo wrote:
> > > > > 
> > > > > 
> > > > > On 2021/8/20 下午4:51, Nikolay Borisov wrote:
> > > > > > 
> > > > > > 
> > > > > > On 18.08.21 г. 0:06, Omar Sandoval wrote:
> > > > > > > From: Omar Sandoval <osandov@fb.com>
> > > > > > > 
> > > > > > > Currently, an inline extent is always created after i_size is extended
> > > > > > > from btrfs_dirty_pages(). However, for encoded writes, we only want to
> > > > > > > update i_size after we successfully created the inline extent.
> > > > > 
> > > > > To me, the idea of write first then update isize is just going to cause
> > > > > tons of inline extent related prblems.
> > > > > 
> > > > > The current example is falloc, which only update the isize after the
> > > > > falloc finishes.
> > > > > 
> > > > > This behavior has already bothered me quite a lot, as it can easily
> > > > > create mixed inline and regular extents.
> > > > 
> > > > Do you have an example of how this would happen? I have the inode and
> > > > extent bits locked during an encoded write, and I see that fallocate
> > > > does the same.
> > > 
> > > xfs_io -f -c "pwrite 0 1K" -c "sync" -c "falloc 0 4k" -c "pwrite 4k 4k"
> > > 
> > > The [0, 1K) will be written as inline without doubt.
> > > 
> > > Then we go to falloc, it will try to zero the range [1K, 4K), but it
> > > doesn't increase the isize.
> > > Thus the page [0, 4k) will still be written back as inline, since isize
> > > is still 1K.
> > > 
> > > Later [4K, 8K) will be written back as regular, causing mixed extents.
> > 
> > I'll have to read fallocate more closely to follow what's going on here
> > and figure out if it applies to encoded writes. Please help me out if
> > you see how this would be an issue with encoded writes.
> 
> This won't cause anything wrong, if the encoded writes follows the
> existing inline extents requirement (always at offset 0).
> 
> Otherwise, the read path could be affected to handle inlined extent at
> non-zero offset.
> 
> > 
> > > > > Can't we remember the old isize (with proper locking), enlarge isize
> > > > > (with holes filled), do the write.
> > > > > 
> > > > > If something wrong happened, we truncate the isize back to its old isize.
> > > > > 
> > > [...]
> > > > > > 
> > > > > > Urgh, just some days ago Qu was talking about how awkward it is to have
> > > > > > mixed extents in a file. And now, AFAIU, you are making them more likely
> > > > > > since now they can be created not just at the beginning of the file but
> > > > > > also after i_size write. While this won't be a problem in and of itself
> > > > > > it goes just the opposite way of us trying to shrink the possible cases
> > > > > > when we can have mixed extents.
> > > > > 
> > > > > Tree-checker should reject such inline extent at non-zero offset.
> > > > 
> > > > This change does not allow creating inline extents at a non-zero offset.
> > > > 
> > > > > > Qu what is your take on that?
> > > > > 
> > > > > My question is, why encoded write needs to bother the inline extents at all?
> > > > > 
> > > > > My intuition of such encoded write is, it should not create inline
> > > > > extents at all.
> > > > > 
> > > > > Or is there any special use-case involved for encoded write?
> > > > 
> > > > We create compressed inline extents with normal writes. We should be
> > > > able to send and receive them without converting them into regular
> > > > extents.
> > > > 
> > > But my first impression for any encoded write is that, they should work
> > > like DIO, thus everything should be sectorsize aligned.
> > > 
> > > Then why could they create inline extent? As inline extent can only be
> > > possible when the isize is smaller than sectorsize.
> > 
> > ENCODED_WRITE is not defined as "O_DIRECT, but encoded". It happens to
> > have some resemblance to O_DIRECT because we have alignment requirements
> > for new extents and because we bypass the page cache, but there's no
> > reason to copy arbitrary restrictions from O_DIRECT. If someone is using
> > ENCODED_WRITE to write compressed data, then they care about space
> > efficiency, so we should make efficient use of inline extents.
> > 
> Then as long as the inline extent requirement for 0 offset is still
> followed, I'll be fine with that.
> 
> But for non-zero offset inline extent? It looks like a much larger
> change, and may affect read path.
> 
> So I'd prefer we keep the 0 offset requirement for inline extent, and
> find a better way to work around.

Ah, okay. I didn't get rid of the 0 offset requirement and I have no
plans to. In fact, this patch kind of does the opposite: it gets rid of
the start parameter to cow_file_range_inline() because it doesn't make
sense for it to ever be anything other than 0 (and we're already
checking that start == 0 in the callers).