[PATCH] Finish the task 'Convert mountlist_mtx to rwlock'

[Mark Johnston]

On Wed, Mar 11, 2015 at 09:14:49PM -0700, Davide Italiano wrote:
> On Wed, Mar 11, 2015 at 9:06 PM, Ryan Stone <rysto32 at gmail.com> wrote:
> > On Wed, Mar 11, 2015 at 9:10 AM, Tiwei Bie <btw at mail.ustc.edu.cn> wrote:
> >> Hi, Mateusz!
> >>
> >> I have finished the task: Convert mountlist_mtx to rwlock [1].
> >
> > My first comment is, are we sure that we actually want an rwlock here
> > instead of an rmlock?  An rmlock will offer much better performance in
> > workloads that mostly only take read locks, and rmlocks do not suffer
> > the priority inversion problems that rwlocks do.  From the description
> > on the wiki page, it sounds like an rmlock would be ideal here:
> >
> 
> Snippet:
> [...]
> -                       mtx_unlock(&mountlist_mtx);
> +                       rw_runlock(&mountlist_lock);
>                 mp->mnt_kern_flag |= MNTK_MWAIT;
>                 msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
>                 if (flags & MBF_MNTLSTLOCK)
> -                       mtx_lock(&mountlist_mtx);
> +                       rw_rlock(&mountlist_lock);
> [...]
> 
> My understanding is that readers are not allowed to sleep while
> holding an rmlock() so a drop-in replacement don't work in this case.

That's true of an rwlock as well though. And the mountlist lock is
dropped around the msleep call in this snippet.