Stephen Hemminger [off-list ref] writes:

On Tue, 19 May 2020 23:45:23 +0200
"Ahmed S. Darwish" [off-list ref] wrote:

quoted

Sequence counters write paths are critical sections that must never be
preempted, and blocking, even for CONFIG_PREEMPTION=n, is not allowed.

Commit 5dbe7c178d3f ("net: fix kernel deadlock with interface rename and
netdev name retrieval.") handled a deadlock, observed with
CONFIG_PREEMPTION=n, where the devnet_rename seqcount read side was
infinitely spinning: it got scheduled after the seqcount write side
blocked inside its own critical section.

To fix that deadlock, among other issues, the commit added a
cond_resched() inside the read side section. While this will get the
non-preemptible kernel eventually unstuck, the seqcount reader is fully
exhausting its slice just spinning -- until TIF_NEED_RESCHED is set.

The fix is also still broken: if the seqcount reader belongs to a
real-time scheduling policy, it can spin forever and the kernel will
livelock.

Disabling preemption over the seqcount write side critical section will
not work: inside it are a number of GFP_KERNEL allocations and mutex
locking through the drivers/base/ :: device_rename() call chain.

From all the above, replace the seqcount with a rwsem.

Fixes: 5dbe7c178d3f (net: fix kernel deadlock with interface rename and netdev name retrieval.)
Fixes: 30e6c9fa93cf (net: devnet_rename_seq should be a seqcount)
Fixes: c91f6df2db49 (sockopt: Change getsockopt() of SO_BINDTODEVICE to return an interface name)
Cc: <redacted>
Signed-off-by: Ahmed S. Darwish <redacted>
Reviewed-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>

Have your performance tested this with 1000's of network devices?

No. We did not. -ENOTESTCASE

The reason seqcount logic is was done here was to achieve scaleablity
and a semaphore does not scale as well.

That still does not make the livelock magically going away. Just make a
reader with real-time priority preempt the writer and the system stops
dead. The net result is perfomance <= 0.

This was observed on RT kernels without a special 1000's of network
devices test case.

Just for the record: This is not a RT specific problem. You can
reproduce that w/o an RT kernel as well. Just run the reader with
real-time scheduling policy.

As much as you hate it from a performance POV the only sane rule of
programming is: Correctness first.

And this code clearly violates that rule.

Thanks,

        tglx