Re: [PATCH] generic/795: add unaligned boundary test cases for WRITE_ZEROES

On 7/7/2026 1:10 PM, Christoph Hellwig wrote:
> On Tue, Jul 07, 2026 at 11:08:39AM +0800, Zhang Yi wrote:
>>> +# WRITE_ZEROES must leave the whole requested range backed by *written*
>>> +# (zeroed) extents while preserving the out-of-range bytes of the partial
>>> +# boundary units. The possible scenarios are: written_edges, hole_edges,
>>> +# unwritten_edges and delalloc_edges.
>>
>> unwritten_edges needs to distinguish between dirty and clean scenarios.
>> The expected result for clean should be tmp.zero, while the expected
>> result for dirty should be tmp.pattern.
> 
> Does it?  I'd expect everything in the range to be zeroed, and everything
> outside to be left alone, as the dirty state is just an imlementation
> detail not visible to the user.

Yes, although this detail is not perceptible to the user, the final
results of these two scenarios are different.

For clean unwritten extents, the portion beyond the range up to the
aligned block boundary should be zeroed; otherwise, stale data may be
exposed after unwritten-to-written conversion. So the result should
be tmp.zero.

For dirty unwritten extents, all data outside the inode boundary
should be left untouched, so the result should be tmp.pattern.

> Similar to other falloc operations,
> I'd expect the file system to write back any boundary block first
> to make the dirty state difference moot.
> 

In fact, in ext4, for dirty extents (including both dirty delalloc and
dirty unwritten extents), they are not written back first. Instead, the
unaligned boundary blocks are zeroed out directly, followed by a
synchronous write-back before the syscall returns, ensuring the extents
are converted to the written state. This point differs from the behavior
of XFS.

Thanks,
Yi.