I think it's the right fix regardless. We should do as little as possible
from irq contexts, and that holds _doubly_ true if it mucks with things
like the page cache.
> With aio in 2.5 we may want to change this property for the unpinning
> stage that would be better run asynchronously from irq handlers, but I
> wouldn't change that for 2.4 (at least until we're forced to ship aio in
> production on top 2.4, that cannot happen until a final user<->kernel API is
> registered somewhere).
I'd really really hate to have the IO make pages go away from irq context
in 2.5.x too. I think async IO should always be started and cleaned up in
a user context - there simply isn't any reason not to (the notion of doing
an exit() or execve() with IO still pending to now-dead-memory is rather
horrible in itself).
> I think the foundamental design mistake that leads to __free_pages to
> fail from irq, is that we allow an anonymous page to reach count 0 and to be
> still in the LRU (the count == 0 check in shrink_cache is the other side
> of the hack too). That's the real BUG, that breaks subtly the freelist
Agreed. We should NEVER free the pages from the irq.
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to firstname.lastname@example.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/