Re: [PATCH] mm: readahead: remove redundant ra_pages in file_ra_state

[YingHang Zhu <casualfisher@gmail.com>]

On Fri, Oct 26, 2012 at 10:30 AM, Ni zhan Chen <nizhan.chen@gmail.com> wrote:
> On 10/26/2012 09:27 AM, Fengguang Wu wrote:
>>
>> On Fri, Oct 26, 2012 at 11:25:44AM +1100, Dave Chinner wrote:
>>>
>>> On Thu, Oct 25, 2012 at 10:58:26AM +0800, Fengguang Wu wrote:
>>>>
>>>> Hi Chen,
>>>>
>>>>> But how can bdi related ra_pages reflect different files' readahead
>>>>> window? Maybe these different files are sequential read, random read
>>>>> and so on.
>>>>
>>>> It's simple: sequential reads will get ra_pages readahead size while
>>>> random reads will not get readahead at all.
>>>>
>>>> Talking about the below chunk, it might hurt someone that explicitly
>>>> takes advantage of the behavior, however the ra_pages*2 seems more
>>>> like a hack than general solution to me: if the user will need
>>>> POSIX_FADV_SEQUENTIAL to double the max readahead window size for
>>>> improving IO performance, then why not just increase bdi->ra_pages and
>>>> benefit all reads? One may argue that it offers some differential
>>>> behavior to specific applications, however it may also present as a
>>>> counter-optimization: if the root already tuned bdi->ra_pages to the
>>>> optimal size, the doubled readahead size will only cost more memory
>>>> and perhaps IO latency.
>>>>
>>>> --- a/mm/fadvise.c
>>>> +++ b/mm/fadvise.c
>>>> @@ -87,7 +86,6 @@ SYSCALL_DEFINE(fadvise64_64)(int fd, loff_t offset,
>>>> loff_t len, int advice)
>>>>                  spin_unlock(&file->f_lock);
>>>>                  break;
>>>>          case POSIX_FADV_SEQUENTIAL:
>>>> -               file->f_ra.ra_pages = bdi->ra_pages * 2;
>>>
>>> I think we really have to reset file->f_ra.ra_pages here as it is
>>> not a set-and-forget value. e.g.  shrink_readahead_size_eio() can
>>> reduce ra_pages as a result of IO errors. Hence if you have had io
>>> errors, telling the kernel that you are now going to do  sequential
>>> IO should reset the readahead to the maximum ra_pages value
>>> supported....
>>
>> Good point!
>>
>> .... but wait .... this patch removes file->f_ra.ra_pages in all other
>> places too, so there will be no file->f_ra.ra_pages to be reset here...
>
>
> In his patch,
>
>
>  static void shrink_readahead_size_eio(struct file *filp,
>                                         struct file_ra_state *ra)
>  {
> -       ra->ra_pages /= 4;
> +       spin_lock(&filp->f_lock);
> +       filp->f_mode |= FMODE_RANDOM;
> +       spin_unlock(&filp->f_lock);
>
> As the example in comment above this function, the read maybe still
> sequential, and it will waste IO bandwith if modify to FMODE_RANDOM
> directly.
I've considered about this. On the first try I modified file_ra_state.size and
file_ra_state.async_size directly, like

file_ra_state.async_size = 0;
file_ra_state.size /= 4;

but as what I comment here, we can not
predict whether the bad sectors will trash the readahead window, maybe the
following sectors after current one are ok to go in normal readahead,
it's hard to know,
the current approach gives us a chance to slow down softly.

Thanks,
        Ying Zhu
>
>>
>> Thanks,
>> Fengguang
>>
>

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>