On Wed, Sep 11, 2019 at 02:22:15AM +0000, Eric Wong wrote: > Btw, some of the changes in public-inbox to use EPOLLET / > EPOLLONESHOT would make it more sensitive to kernel bugs, > and missed wakeups could cause resource exhaustion... > > I just noticed this now: > https://lore.kernel.org/lkml/20190904175303.918854...@linuxfoundation.org/ > > And I remember I found another mm-compaction-related issue > several years back (early 3.x) which caused epoll wakeups to be > missed...
To give some more data points, downgrading to f4f0a3be still shows a number of /tmp/PerlIO* (deleted) entries, but the number of pipes stays the same over time. If I switch to the latest master, the number of broken pipes grows steadily following each git pull (in the hundreds after 10 minutes of running). I've downgraded lore.kernel.org to f4f0a3be for now to avoid the resource exhaustion. -K
