Re: Further copy_from_user() discussion.

[Vadim Lobanov <vlobanov@speakeasy.net>]

On Thu, 14 Apr 2005, Catalin Marinas wrote:

> Vadim Lobanov <vlobanov@speakeasy.net> wrote:
> > 2. Would it be possible to eliminate the might_sleep() call in
> > copy_from_user()? It seems that, very soon after, the __copy_from_user()
> > macro does another might_sleep(), with very few instructions in between.
> > But there might be some trick here that I'm missing.
>
> might_sleep() is used for debugging the possible sleep while in an
> atomic operation. I think it is safe to check this for all the calls
> to copy_from_user(), no matter if the access is OK or not (memset
> being used in the latter case). The same is for
> __copy_from_user(). Anyway, if you don't enable
> CONFIG_DEBUG_SPINLOCK_SLEEP, the might_sleep() macro is empty.
>
> --
> Catalin
>

Thanks for the response.

I think I misspoke a bit in my email above. The intent was not to
eliminate all might_sleep() calls from the copy_from_user() code path;
but rather juggle the source around a bit so there is only one
might_sleep() call per each code path. Currently, in the default case,
it calls it twice.

By the way, is the following still true about might_sleep()?
http://kerneltrap.org/node/3440/10103

-Vadim Lobanov