yield_cpu is a sample of a preempted lock holder that gets propagated
back through the queue. Queued waiters use this to yield to the
preempted lock holder without continually sampling the lock word (which
would defeat the purpose of MCS queueing by bouncing the cache line).
The problem is that yield_cpu can become stale. It can take some time to
be passed down the chain, and if any queued waiter gets preempted then
it will cease to propagate the yield_cpu to later waiters.
This can result in yielding to a CPU that no longer holds the lock,
which is bad, but particularly if it is currently in H_CEDE (idle),
then it appears to be preempted and some hypervisors (PowerVM) can
cause very long H_CONFER latencies waiting for H_CEDE wakeup. This
results in latency spikes and hard lockups on oversubscribed
partitions with lock contention.
This is a minimal fix. Before yielding to yield_cpu, sample the lock
word to confirm yield_cpu is still the owner, and bail out of it is not.
Thanks to a bunch of people who reported this and tracked down the
exact problem using tracepoints and dispatch trace logs.
Fixes: 28db61e207 ("powerpc/qspinlock: allow propagation of yield CPU down the queue")
Cc: stable@vger.kernel.org # v6.2+
Reported-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Reported-by: Laurent Dufour <ldufour@linux.ibm.com>
Reported-by: Shrikanth Hegde <sshegde@linux.vnet.ibm.com>
Debugged-by: "Nysal Jan K.A" <nysal@linux.ibm.com>
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Tested-by: Shrikanth Hegde <sshegde@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://msgid.link/20231016124305.139923-2-npiggin@gmail.com
f9bc9bbe8a