Re: [PATCHSET v3 0/4] Submit ->readpages() IO as read-ahead

[Jens Axboe <axboe@kernel.dk>]

On 6/21/18 6:21 AM, Christoph Hellwig wrote:
> On Wed, Jun 20, 2018 at 04:28:54PM -0600, Jens Axboe wrote:
>> The assumption has been true for a decade or more. The only problem with
>> it is that we have this ->readpages() interface that is only used for
>> readahead. Hence it should be ->readahead() or similar, that avoid the
>> fragility argument.
> 
> Or just stop this whole bike shed.  ->readpages doesn't keep the pages
> it reads locked, so it fundamentally is a readahead style operation.
> 
> Now if we actually want do to anything with that information is another
> question that I already asked in reply to the first round of these
> patches.
> 
> Especially as e.g. nvme actually sets hints based on the flag, that
> if hardware actually ever looked at them might be doing the wrong
> thing (fortunately for us I doubt any nvme hardware actually looks
> at that flag).

In most of these cases, these types of hints don't translate uniformly
across devices. So whether we pass them on to the actual hardware or
not is a different matter. For now, as mentioned, all I care about is
getting the tracing working properly, and being able to functionally
differentiate between a read and a read-ahead read in the block layer.

-- 
Jens Axboe