Re: UVM/genfs review

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

On Thu, Jan 15, 2009 at 11:52:42AM +0100, Juergen Hannken-Illjes wrote:

> On Thu, Jan 15, 2009 at 07:42:07PM +0900, YAMAMOTO Takashi wrote:
> > hi,
> > 
> > > So this deadlock is independent of snapshots, fstrans_* etc are not
> > > involved at all.
> > 
> > is wapbl_begin in putpages necessary even without snapshots?
> 
> No, the only reason for this wapbl_begin is the copy-on-write for
> snapshots where the block allocation has to be enclosed in wapbl_begin.
> 
> > btw, why did you exclude the pagedaemon from need_wapbl?
> 
> Because snapshots will not do copy-on-write for the pagedaemon and return
> an error instead.
> 
> Still thinking that the wapbl design decision to use an exclusive lock
> is wrong and a producer-consumer model would be moreappropriate...

I agree but even so opening a transaction may cause the caller to wait for
resources, which could cause a similar deadlock.

Andrew