Christoph Hellwig [off-list ref] wrote:

On Mon, Feb 15, 2021 at 10:46:23PM +0000, David Howells wrote:

quoted

The in_softirq() in netfs_rreq_terminated() works fine for the cache being
on a normal disk, as the completion handlers may get called in softirq
context, but for an NVMe drive, the completion handler may get called in
IRQ context.

Fix to use in_interrupt() instead of in_softirq() throughout the read
helpers, particularly when deciding whether to punt code that might sleep
off to a worker thread.

We must not use either check, as they all are unreliable especially
for PREEMPT-RT.

Is there a better way to do it?  The intent is to process the assessment phase
in the calling thread's context if possible rather than bumping over to a
worker thread.  For synchronous I/O, for example, that's done in the caller's
thread.  Maybe that's the answer - if it's known to be asynchronous, I have to
punt, but otherwise don't have to.

David