Re: fadvise interferes with readahead

[Claudio Freire]

On Tue, Nov 20, 2012 at 5:04 AM, Fengguang Wu <fengguang.wu@xxxxxxxxx> wrote:
> Yes. The kernel readahead code by design will outperform simple
> fadvise in the case of clustered random reads. Imagine the access
> pattern 1, 3, 2, 6, 4, 9. fadvise will trigger 6 IOs literally. While
> kernel readahead will likely trigger 3 IOs for 1, 3, 2-9. Because on
> the page miss for 2, it will detect the existence of history page 1
> and do readahead properly. For hard disks, it's mainly the number of
> IOs that matters. So even if kernel readahead loses some opportunities
> to do async IO and possibly loads some extra pages that will never be
> used, it still manges to perform much better.
>
>> The fix would lay in fadvise, I think. It should update readahead
>> tracking structures. Alternatively, one could try to do it in
>> do_generic_file_read, updating readahead on !PageUptodate or even on
>> page cache hits. I really don't have the expertise or time to go
>> modifying, building and testing the supposedly quite simple patch that
>> would fix this. It's mostly about the testing, in fact. So if someone
>> can comment or try by themselves, I guess it would really benefit
>> those relying on fadvise to fix this behavior.
>
> One possible solution is to try the context readahead at fadvise time
> to check the existence of history pages and do readahead accordingly.
>
> However it will introduce *real interferences* between kernel
> readahead and user prefetching. The original scheme is, once user
> space starts its own informed prefetching, kernel readahead will
> automatically stand out of the way.

I understand that would seem like a reasonable design, but in this
particular case it doesn't seem to be. I propose that in most cases it
doesn't really work well as a design decision, to make fadvise work as
direct I/O. Precisely because fadvise is supposed to be a hint to let
the kernel make better decisions, and not a request to make the kernel
stop making decisions.

Any interference so introduced wouldn't be any worse than the
interference introduced by readahead over reads. I agree, if fadvise
were to trigger readahead, it could be bad for applications that don't
read what they say the will. But if cache hits were to simply update
readahead state, it would only mean that read calls behave the same
regardless of fadvise calls. I think that's worth pursuing.

I ought to try to prepare a patch for this to illustrate my point. Not
sure I'll be able to though.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/