Re: [PATCH v4 00/14] forcealign for xfs

[John Garry <john.g.garry@oracle.com>]

On 14/11/2024 12:48, Long Li wrote:
> On Wed, Sep 18, 2024 at 11:12:47AM +0100, John Garry wrote:
>> On 17/09/2024 23:27, Dave Chinner wrote:
>>>> # xfs_bmap -vvp  mnt/file
>>>> mnt/file:
>>>> EXT: FILE-OFFSET      BLOCK-RANGE      AG AG-OFFSET        TOTAL FLAGS
>>>>     0: [0..15]:         384..399          0 (384..399)          16 010000
>>>>     1: [16..31]:        400..415          0 (400..415)          16 000000
>>>>     2: [32..127]:       416..511          0 (416..511)          96 010000
>>>>     3: [128..255]:      256..383          0 (256..383)         128 000000
>>>> FLAG Values:
>>>>      0010000 Unwritten preallocated extent
>>>>
>>>> Here we have unaligned extents wrt extsize.
>>>>
>>>> The sub-alloc unit zeroing would solve that - is that what you would still
>>>> advocate (to solve that issue)?
>>> Yes, I thought that was already implemented for force-align with the
>>> DIO code via the extsize zero-around changes in the iomap code. Why
>>> isn't that zero-around code ensuring the correct extent layout here?
>> I just have not included the extsize zero-around changes here. They were
>> just grouped with the atomic writes support, as they were added specifically
>> for the atomic writes support. Indeed - to me at least - it is strange that
>> the DIO code changes are required for XFS forcealign implementation. And,
>> even if we use extsize zero-around changes for DIO path, what about buffered
>> IO?
> 
> I've been reviewing and testing the XFS atomic write patch series. Since
> there haven't been any new responses to the previous discussions on this
> issue, I'd like to inquire about the buffered IO problem with force-aligned
> files, which is a scenario we might encounter.
> 
> Consider a case where the file supports force-alignment with a 64K extent size,
> and the system page size is 4K. Take the following commands as an example:
> 
> xfs_io  -c "pwrite 64k 64k" mnt/file
> xfs_io  -c "pwrite 8k 8k" mnt/file
> 
> If unaligned unwritten extents are not permitted, we need to zero out the
> sub-allocation units for ranges [0, 8K] and [16K, 64K] to prevent stale
> data. 

How does this prevent stale data? Just zeroing will ensure aligned 
extents. Unless iomap is provided a mapping for the fully aligned extent.

> While this can be handled relatively easily in direct I/O scenarios,
> it presents significant challenges in buffered I/O operations. The main
> difficulty arises because the extent size (64K) is larger than the page
> size (4K), and our current code base has substantial limitations in handling
> such cases.

What is the limitation exactly?

> 
> Any thoughts on this?

TBH, the buffered IO case has not been considered too much.

The sub-extent zeroing was intended for atomic writes > 1x FSB and we 
only care about DIO there.

Thanks,
John