On Fri, 12 Jul 2019 13:39:35 +0100 Mel Gorman mgorman@suse.de wrote:
So although I still think that just failing the migration if we cannot invalidate buffer heads is a safer choice, just extending the private_lock protected section does not seem as bad as I was afraid.
That does not seem too bad and your revised patch looks functionally fine. I'd leave out the tracepoints though because a perf probe would have got roughly the same data and the tracepoint may be too specific to track another class of problem. Whether the tracepoint survives or not and with a changelog added;
Acked-by: Mel Gorman mgorman@techsingularity.net
Andrew, which version do you want to go with, the original version or this one that holds private_lock for slightly longer during migration?
The revised version looks much more appealing for a -stable backport. I expect any mild performance issues can be address in the usual fashion. My main concern is not to put a large performance regression into mainline and stable kernels. How confident are we that this is (will be) sufficiently tested from that point of view?