[GIT PULL] nvme fix for 4.16-rc6

[Keith Busch]

On Wed, Mar 21, 2018 at 03:44:32PM -0600, Jens Axboe wrote:
> On 2nd though, let's not. While it worked fine on one box, my other
> test box (that has a bunch of drives) is not very happy:
> 
> [   30.241598] nvme nvme2: pci function 0000:0b:00.0                            
> [   30.247205] nvme nvme3: pci function 0000:81:00.0                            
> [   30.252684] nvme nvme4: pci function 0000:82:00.0                            
> [   30.258144] nvme nvme5: pci function 0000:83:00.0                            
> [   30.263606] nvme nvme6: pci function 0000:84:00.0                            
> [   30.360555] nvme nvme3: could not set timestamp (8194)                       
> [   30.481649] nvme nvme6: Shutdown timeout set to 8 seconds                    
> [   38.790949] nvme nvme4: Device not ready; aborting initialisation            
> [   38.797857] nvme nvme4: Removing after probe failure status: -19             
> [   60.708816] nvme nvme3: I/O 363 QID 8 timeout, completion polled             
> [   60.708820] nvme nvme6: I/O 781 QID 7 timeout, completion polled             
> [   68.068772] nvme nvme2: I/O 769 QID 28 timeout, completion polled            
> [   91.108626] nvme nvme6: I/O 781 QID 7 timeout, completion polled             
> [   98.660581] nvme nvme2: I/O 769 QID 28 timeout, completion polled            
> [  121.702691] nvme nvme6: I/O 100 QID 7 timeout, completion polled             
> [  128.998648] nvme nvme3: I/O 387 QID 4 timeout, completion polled             
> [  152.038523] nvme nvme6: I/O 781 QID 7 timeout, completion polled             
> 
> This is just doing an fdisk -l after load. No interrupts triggering,
> looking at /proc/interrupts for the queues that timeout. The commands
> do complete eventually, but only because we poll the queue. Ignore
> the probe failure, that one is expected.
> 
> So that's a pretty horrific failure, about half (or more) of the
> devices simply don't work. For something being pushed aggressively
> at -rc6 time, I'd say your testing is lacking.
> 
> I'm going to drop it from my 4.16 queue, and don't queue it up for
> 4.17 before we figure out what's going on here.

Well hell, that is awful. Thank you for checking; patch dropped.

Admin queue interrupts appear to be working, otherwise it couldn't have
gotten that far. I've no immediate explanation for your results, so
going back to the drawing board to sort it out.