Re: [PATCH] infiniband: fix a subtle race condition

[Jason Gunthorpe <jgg@mellanox.com>]

On Thu, Jun 14, 2018 at 10:01:08AM +0300, Leon Romanovsky wrote:
> On Wed, Jun 13, 2018 at 11:21:54PM -0700, Cong Wang wrote:
> > On Wed, Jun 13, 2018 at 10:34 PM, Leon Romanovsky <leon@kernel.org> wrote:
> > >
> > > Hi Cong,
> > >
> > > If the compiler optimizes the first line (mutex_lock) as you wrote,
> > > it will reuse "f" for the second line (mutex_unlock) too.
> >
> > Nope, check the assembly if you don't trust me, at least
> > my compiler always fetches ctx->file without this patch.
> >
> > I can show you the assembly code tomorrow (too late to
> > access my dev machine now).
> 
> I trust you, so don't need to check it however wanted to emphasize
> that your solution is compiler specific and not universally true.
> 
> >
> >
> > >
> > > You need to ensure that ucma_modify_id() doesn't run in parallel to
> > > anything that uses "ctx->file" directly and indirectly.
> > >
> >
> > Talk is easy, show me the code. :) I knew there is probably
> > some other race with this code even after my patch, possibly with
> > ->close() for example, but for this specific unlock warning, this patch
> > is sufficient. I can't solve all the races in one patch.
> 
> We do prefer complete solution once the problem is fully understood.
> 
> It looks like you are one step away from final patch. It will be conversion
> of mutex to be rwlock and separating between read (almost in all places)
> and write (in ucma_migrate_id) paths.

This was my brief reaction too, this code path almost certainly has a
use-after-free, and we should fix the concurrency between the two
places in some correct way..

Jason