Re: [RFC] per thread page reservation patch

[Nikita Danilov <nikita@xxxxxxxxxxxxx>]

Andrew Morton <akpm@xxxxxxxx> writes:

[...]

>
> Maybe I'm being thick, but I don't see how you can protect the reservation
> of an outer reserver in the above way:
>
> 	perthread_pages_reserve(10);
> 	...				/* current->private_pages_count = 10 */
> 		perthread_pages_reserve(10)	/* private_pages_count = 20 */
> 		use 5 pages			/* private_pages_count = 15 */
> 		perthread_pages_release(5);
>
> But how does the caller come up with the final "5"?

	wasreserved = perthread_pages_count();
	result = perthread_pages_reserve(estimate_reserve(), GFP_KERNEL);
	if (result != 0)
		return result;

    /* do something that consumes reservation */

	perthread_pages_release(perthread_pages_count() - wasreserved);

>
> Seems better to me if prethread_pages_reserve() were to return the initial
> value of private_pages_count, so the caller can do:
>
> 	old = perthread_pages_reserve(10);
> 		use 5 pages
> 	perthread_pages_release(old);
>
> or whatever.
>
> That kinda stinks too in a way, because both the outer and the inner
> callers need to overallocate pages on behalf of the worst case user in some
> deep call stack.
>
> And the whole idea is pretty flaky really - how can one precalculate how
> much memory an arbitrary md-on-dm-on-loop-on-md-on-NBD stack will want to
> use?  It really would be better if we could drop the whole patch and make
> reiser4 behave more sanely when its writepage is called with for_reclaim=1.

Reiser4 doesn't use this for ->writepage(), by the way. This is used by
tree balancing code to assure that balancing cannot get -ENOMEM in the
middle of tree modification, because undo is _so_ very complicated.

Nikita.
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/