mkfs.erofs: should MAX_BLOCK_SIZE be tied to build host page size?

[Gao Xiang]

Hi Matthew,

On 2026/2/24 23:09, Matthew Lear wrote:
> Hi,
> I'm writing to ask if there is scope to change how mkfs.erofs handles
> its max block size ceiling.
> 
> MAX_BLOCK_SIZE currently defaults to the page size of the system where
> erofs-utils is built, not where it executes. This means a that compiled
> mkfs.erofs built on a 4K page host permanently inherits the build
> host's 4K block size limit, even when running on a system with a 16K
> page kernel (e.g. like our 2712 Raspberry Pi5 kernel):
> 
> $ getconf PAGE_SIZE
> 16384
> $ mkfs.erofs -b 16384 test.img /lib
> <E> erofs: invalid block size 16384
> 
> Passing MAX_BLOCK_SIZE=16384 at erofs-utils configure time results in a
> binary that successfully creates 16K block images, regardless of the
> page size of the host it's executing on. The restriction seems to be a
> build-time constant rather than something which has runtime
> justification.
> 
> Also, since modern kernels now support large block sizes (LBS) where
> block size > page size, this seems to add weight to removal of the
> inherited host page size cap.

Yes, but EROFS doesn't support LBS.. yet.

> 
> Without this, users on 16K page kernels are forced into the experimental
> sub-page compressed block path:
> 
> erofs: (device mmcblk0p4): EXPERIMENTAL EROFS subpage compressed block
> support in use. Use at your own risk!
> 
> If I build erofs-utils on a 4K host and use its mkfs for a 16K target,
> the binary rejects -b 16384. I need to configure with
> MAX_BLOCK_SIZE=16384 in order to use -b 16384 otherwise I'm forced to
> produce 4K-block images that trigger the experimental sub-page warning
> on the target device. This isn't a problem, but it does feel like the
> utility should decouple its validation logic from the host's
> page size.
> 
> The baked-in behaviour also affects official distribution packages,
> which may typically be built on 4K page size build farms. So this is
> possibly a wider issue for the 16K page size ecosystem.
> 
> Is there scope to increase the default block size ceiling of mkfs.erofs
> to 16K (or perhaps increase it to 16K for aarch64 compilations)?

I think the reasonable maximum block size for EROFS should be no
more than 64k.

I think we could consider set the MAX_BLOCK_SIZE to 64k for
aarch64, but leave it to the page size if `-b` is unspecified.

Would you mind submitting a patch for this (checking the target
arch, if aarch64, set MAX_BLOCK_SIZE to 64k instead) since I've
already had a bunch of other stuffs.

Thanks,
Gao Xiang

> 
> Cheers,
> -- Matt