Re: kern/54945 (nfs uvm_pagegetdirty panic)

[Andrew Doran <ad%netbsd.org@localhost>]

Hi,

On Fri, Feb 21, 2020 at 12:07:59AM +0900, Rin Okuyama wrote:
> I found that the test program in the previous message does not cause
> panic even with NFS on ibm4xx, if taking enough intervals manually
> between each steps in the program, i.e., write-wait-mmap-wait...
> 
> This result maybe implies that panic occurred only on ibm4xx, and not
> amd64 nor armv6 just because of timing problem; ibm4xx is too slow to
> finish asynchronous NFS IO and mark that page clean before read from
> mmap'ed address.
> 
> Here's UVMHIST for successful read from mmap'ed address:
> 
>     http://www.netbsd.org/~rin/uvmhist_20200220_success.txt
> 
> Thoughts?

Thank you for the detailed investigation.  I think this is unlikely to be a
pmap prolem, since the assertion is concerned with software state - CLEAN
and DIRTY are very definite states set by UVM.  The pmap only comes into
play if the page's state is UNKNOWN.  The NFS changes on yamt-pagecache are
unlikey to help either, as you discovered.  I need to think about this one
more.

Andrew