Re: [PATCH] might_sleep() improvements

[Nick Piggin <piggin@cyberone.com.au>]

Linus Torvalds wrote:

>Nick Piggin wrote:
>
>>I think these should be pushed down to where the sleeping
>>actually happens if possible.
>>
>
>No, that ends up doing the wrong thing for most of the really common cases.
>
>In particular, most of the memory allocation functions very seldom actually
>sleep. After all, they'll find plenty of free memory (or easily freeable
>memory) without having to wait for any pageouts or anything like that.
>
>Yet the bug is there - the call _could_ have slept.
>
>So "might_sleep()" really does what the name suggests: it is used to say
>that a particular case _may_ sleep, even if it ends up being unlikely.
>
>Because what we're after is not a bug actually happening, but a latent bug
>that has been hidden by the fact that it happens so rarely in practice.
>
>This is why "might_sleep()" should happen as early as possible, and not
>get pushed down.
>
>

Yes I see. I agree. I thought some could be pushed down further without
losing info. I was mainly worried about adding the might_sleep_if
function.