Re: [PATCH 02/12] locking/rwsem: Exit early when held by an anonymous owner

[Christoph Hellwig]

On Tue, Jan 14, 2020 at 01:17:45PM -0500, Waiman Long wrote:
> The owner field is just a pointer to the task structure with the lower 3
> bits served as flag bits. Setting owner to RWSEM_OWNER_UNKNOWN (-2) will
> stop optimistic spinning. So under what condition did the crash happen?

When running xfstests with all patches in this series except for this
one, IIRC in generic/114.

> Anyway, PeterZ is working on revising the percpu-rwsem implementation to
> more gracefully handle the frozen case. At the end, there will not be a
> need for the RWSEM_OWNER_UNKNOWN magic and it can be removed.

Well, this series relies on that value.  And I think it fundamentally
is the right thing to do for AIO, and potentially other I/O related
locking where we take a lock to synchronize access to data, then
do I/O and then eventually get an I/O completion from an interrupt.
Even thinking from the PREEMP_RT context we want to boost the
initial thread as long as we can, then do nothing when it is off
to I/O hardware (except maybe providing good diagnostics that the cause
for the latency is I/O), and then boost the thread that is handling
the completion.  Things like the i_dio_count hack can't provide that.