Re: Support for compressed inline extents

[Qu Wenruo <quwenruo.btrfs@gmx.com>]

On 2021/8/29 下午7:22, Forza wrote:
>
>
> On 2021-08-27 12:08, David Sterba wrote:
>> On Mon, Aug 23, 2021 at 04:23:29PM -0400, Zygo Blaxell wrote:
>>> On Mon, Aug 23, 2021 at 09:34:27PM +0200, Forza wrote:
>>>> Further up you showed that we can read encoded inlined data. What is
>>>> missing
>>>> for that we can read encoded inlined data that decode to >page_size
>>>> in size?
>>>
>>> In uncompress_inline():
>>>
>>>     // decoded length of extent on disk...
>>>     max_size = btrfs_file_extent_ram_bytes(leaf, item);
>>>
>>>     ...
>>>
>>>     // ...can never be more than one page because of this line(*)
>>>     max_size = min_t(unsigned long, PAGE_SIZE, max_size);
>>>
>>> There might be further constraints around this code (e.g. the caller
>>> only fills in structures for one page, or doesn't bother to call this
>>> function at all for offsets above PAGE_SIZE).
>>>
>>> All the restrictions would need to be removed in the kernel and support
>>> for reading multi-page inline extents added where necessary.  There
>>> would
>>> have to be an incompat bit on the filesystem to prevent old kernels from
>>> trying (and failing) to read longer inline extents.  The disk format is
>>> already technically capable of specifying a longer inline extent (up to
>>> min(UINT32_MAX, metadata_page_size)) but that was never the problem.
>>
>> Regarding the idea of compressed inline extents, I'm not much in favor
>> of increasing the limit beyond one page (or sector). The metadata space
>> is more precious and that's also the motivation behind low default
>> max_inline. Another thing is mixing data and metadata with potentially
>> different block group profiles.
>
> We already mix profiles today with inlining small extents. It is not a
> problem as most people use a better/higher redundancy for metadata than
> for data.
>
>> The inline files is IMO a nice little optimization and helps when the
>> size is below certain limit to avoid wasting data blocks and the
>> indirection.
>>
>
> To be fair, I think the benefit is that we inline instead of creating
> 4KiB extents for small amounts of data. This benefit would be true even
> if that data was compressed data and the compressed size was <2KiB.
>
> I do understand the earlier points that this would perhaps be a big
> thing to change, also incompatible with earlier kernels. Given that work
> the benefit is rather small.
>
> Perhaps if we are doing incompatible changes in the future, this could
> be considered at that time? One reference is what Qu wrote here about
> taking in more factors to consider for inlining?
> https://lore.kernel.org/linux-btrfs/d0dccd5e-c67f-a18d-8d6e-559504b5ee91@suse.com/


There are way more things to consider for inlining extent larger than
one sector.

- Reading the inline extent
   Now we must consider multiple pages other than just simply copying the
   data to the page.
   This also brings extra error paths which must be considered.

- Writing the extra pages of the inline extent
   Unlike current single sector inline, what if you overwrite the 2nd
   page of the inline extent?

   In theory, we should make the whole inline extents into one regular
   extent, then this means we have to read the whole inline extent out,
   re-marking them dirty.

   We don't have the exact facility to do that.
   The most similar one is defrag, but now we need to do that in write
   path too.

   Or you can treat them as regular extents, just cow the 2nd page?
   Then it completely wastes the extra space in metadata.

   Either way, it's way complex than you thought, and all my respect to
   all those guys solving the corner cases with compression and inline.


So my idea towards inline and compression is pretty simple, they are
good optimization, but not critical part.

Thus if they are affecting regular core functions to be more complex
than they should, then no.

And extending inline extent size beyond page size is exactly bringing
extra cost to the developers, with little and uncertain benefit.

Thanks,
Qu


>
>
> Thanks