[pnfs] pdflush/writepages race

[Tom Tucker]

I've got an interesting situation where the pdflush daemon is sitting in
blk_congestion_wait and the process generating the dirty pages is in
nfs_wait_on_write_congestion. This is all on a heavily modified 2.6.18
kernel.

No one makes progress and we're effectively hung. Everything seems to
work fine until pdflush gets involved. 

I'm curious how the BDI_write_congested bit is supposed to work. It
looks to me like it only ever gets cleared in nfs_writepages, but
nfs_writepages itself can block if it's set. I assume this is for
competing threads flushing pages. In any case, I'm looking to scavenge
cerebellum cycles where possible. 

If this rings any bells with anyone, please let me know...

Thanks,
Tom