Re: [PATCH v7 11/14] xfs: add xfs_file_dio_write_atomic()

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

On 23/04/2025 06:18, Luis Chamberlain wrote:
> On Tue, Apr 22, 2025 at 07:08:32AM +0100, John Garry wrote:
>> On 21/04/2025 22:18, Luis Chamberlain wrote:
>>>> /*
>>>> +	 * The retry mechanism is based on the ->iomap_begin method returning
>>>> +	 * -ENOPROTOOPT, which would be when the REQ_ATOMIC-based write is not
>>>> +	 * possible. The REQ_ATOMIC-based method typically not be possible if
>>>> +	 * the write spans multiple extents or the disk blocks are misaligned.
>>>> +	 */
>>>> +	if (ret == -ENOPROTOOPT && dops == &xfs_direct_write_iomap_ops) {
>>> Based on feedback from LSFMM, due to the performance variaibility this
>>> can introduce, it sounded like some folks would like to opt-in to not
>>> have a software fallback and just require an error out.
>>>> Could an option be added to not allow the software fallback?
>>
>> I still don't see the use in this.
> 
> Its not the use, its the concern for underdeterminism in performance.

Sure, we don't offer RT performance guarantees, but what does?

> 
>> So consider userspace wants to write something atomically and we fail as a
>> HW-based atomic write is not possible.
> 
> Sounds like a terrible predicant for those that want hw atomics and
> reliability for it.

Well from our MySQL testing performance is good.

> 
>> What is userspace going to do next?
> 
> It would seem that would depend on their analysis on the number of
> software fallbacks where a software atomic based solution is used and
> the impact on performance.

sorry, but I don't understand this

> 
>> I heard something like "if HW-based atomics are not possible, then something
>> has not been configured properly for the FS" - that something would be
>> extent granularity and alignment, but we don't have a method to ensure this.
>> That is the whole point of having a FS fallback.
> 
> We do with LBS.

Sure, but not everyone wants LBS

> Its perfectly deterministic to be aligned with a sector
> size matching the block size, even for metadata writes.
> 
>>> If so, then I think the next patch would also need updating.
>>>
>>> Or are you suggesting that without the software fallback atomic writes
>>> greater than fs block size are not possible?
>>
>> Yes, as XFS has no method to guarantee extent granularity and alignment.
> 
> Ah, I think the documentation for this featuer should make this clear,
> it was not clear up to this point in patch review.
> 

ok, that can be added