Messages in this thread | | | Date | Mon, 18 Jan 1999 21:46:40 GMT | From | "Stephen C. Tweedie" <> | Subject | Re: Removing swap lockmap... |
| |
Hi,
In article <87iue47gy4.fsf@atlas.CARNet.hr>, Zlatko Calusic <Zlatko.Calusic@CARNet.hr> writes:
> I removed swap lockmap all together and, to my surprise, I can't > produce any ill behaviour on my system, not even under very heavy > swapping (in low memory condition).
Just because you can't reproduce it doesn't mean it works perfectly. There was a very good reason why the swap lock map was still required until recently. The race condition it fixed wass an obscure one but still important. However, very recent VM changes make me wonder if it is still absolutely necessary.
The problem was that if we swapped out a page, we might sometimes remove the swap cache for the page before the IO was complete. If we can _guarantee_ that the swap cache will persist until after the IO is complete, then any future attempt to use that swap page will find that the page is locked and will wait for the IO to complete.
However, if in fact the swap cache for the page _ever_ gets removed before the IO completes, then a future read in of the page might start before the current write had completed. This has been observed in practice. The swap lock protects against this.
Now that we always keep the swap cache intact in mm/vmscan.c and only reclaim it in mm/filemap.c, we might in fact be safe omiting the swap lock. I'd be nervous about it without a _thorough_ audit of the code, though, as this particular race is hard to reproduce.
--Stephen
- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.rutgers.edu Please read the FAQ at http://www.tux.org/lkml/
| |