Re: [openib-general] Re: [PATCH][RFC][0/4] InfiniBand userspace verbs implementation

[Andrea Arcangeli <andrea@suse.de>]

On Wed, May 11, 2005 at 04:12:41PM -0400, William Jordan wrote:
> If I am reading you correctly, you are saying that mlock currently
> prevents pages from migrating around to unfragment memory, but
> get_user_pages does not prevent this? If this is the case, this could

This is not the case. Infact get_user_pages is a stronger pin than
mlock. But if you call it by hand and you plan to write to the page, you
have to use the "write=1" flag, this is fundamental if you want to write
to the physical page from userland while it's being tracked by IB dma.

In short you should not use mlock and you should use only
get_user_pages(write=1).

If the problem appears again even after the last fix for the COW I did
last year, than it means we've another yet another bug to fix.

Using mlock for this is unnecessary. mlock is a "virtual" pin and it
provides weaker guarantees than what you need. You need _physical_ pin
and get_user_pages(write=1) is the only one that will give it to you.

write=0 is ok too if you're never ever going to write to the page with
the cpu from userland.

In the old days there was the concept that get_user_pages wasn't a
"pte-pin", but that was infact broken in the way COW was working with threads,
but this is fixed now that is really a "pte-pin" again (like in 2.2
which never had the corruption cow bug!) even though the pte may
temporarily be set to swapcache or null. In current 2.6 you're
guaranteed that despite the pte may be temporarly be set to not-present,
the next minor fault will bring into memory the very same physical page
that was there before. At least unless you map the thing writeprotect
(i.e. write=0) and you write to it from userland.. ;).
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/