Re: [PATCH v2] dma-buf/fence: Take refcount on the module that owns the fence

[Daniel Vetter]

On Mon, Jun 25, 2018 at 09:21:15PM +0530, Akhil P Oommen wrote:
> 
> 
> On 6/25/2018 1:20 PM, Daniel Vetter wrote:
> > On Fri, Jun 22, 2018 at 11:08:48AM +0100, Chris Wilson wrote:
> > > Quoting Gustavo Padovan (2018-06-22 11:04:16)
> > > > Hi Akhil,
> > > > 
> > > > On Fri, 2018-06-22 at 15:10 +0530, Akhil P Oommen wrote:
> > > > > Each fence object holds function pointers of the module that
> > > > > initialized
> > > > > it. Allowing the module to unload before this fence's release is
> > > > > catastrophic. So, keep a refcount on the module until the fence is
> > > > > released.
> > > > > 
> > > > > Signed-off-by: Akhil P Oommen <akhilpo@xxxxxxxxxxxxxx>
> > > > > ---
> > > > > Changes in v2:
> > > > > - added description for the new function parameter.
> > > > > 
> > > > >   drivers/dma-buf/dma-fence.c | 16 +++++++++++++---
> > > > >   include/linux/dma-fence.h   | 10 ++++++++--
> > > > >   2 files changed, 21 insertions(+), 5 deletions(-)
> > > > > 
> > > > > diff --git a/drivers/dma-buf/dma-fence.c b/drivers/dma-buf/dma-
> > > > > fence.c
> > > > > index 4edb9fd..2aaa44e 100644
> > > > > --- a/drivers/dma-buf/dma-fence.c
> > > > > +++ b/drivers/dma-buf/dma-fence.c
> > > > > @@ -18,6 +18,7 @@
> > > > >    * more details.
> > > > >    */
> > > > > +#include <linux/module.h>
> > > > >   #include <linux/slab.h>
> > > > >   #include <linux/export.h>
> > > > >   #include <linux/atomic.h>
> > > > > @@ -168,6 +169,7 @@ void dma_fence_release(struct kref *kref)
> > > > >   {
> > > > >        struct dma_fence *fence =
> > > > >                container_of(kref, struct dma_fence, refcount);
> > > > > +     struct module *module = fence->owner;
> > > > >        trace_dma_fence_destroy(fence);
> > > > > @@ -178,6 +180,8 @@ void dma_fence_release(struct kref *kref)
> > > > >                fence->ops->release(fence);
> > > > >        else
> > > > >                dma_fence_free(fence);
> > > > > +
> > > > > +     module_put(module);
> > > > >   }
> > > > >   EXPORT_SYMBOL(dma_fence_release);
> > > > > @@ -541,6 +545,7 @@ struct default_wait_cb {
> > > > >   /**
> > > > >    * dma_fence_init - Initialize a custom fence.
> > > > > + * @module:  [in]    the module that calls this API
> > > > >    * @fence:   [in]    the fence to initialize
> > > > >    * @ops:     [in]    the dma_fence_ops for operations on this
> > > > > fence
> > > > >    * @lock:    [in]    the irqsafe spinlock to use for locking
> > > > > this fence
> > > > > @@ -556,8 +561,9 @@ struct default_wait_cb {
> > > > >    * to check which fence is later by simply using dma_fence_later.
> > > > >    */
> > > > >   void
> > > > > -dma_fence_init(struct dma_fence *fence, const struct dma_fence_ops
> > > > > *ops,
> > > > > -            spinlock_t *lock, u64 context, unsigned seqno)
> > > > > +_dma_fence_init(struct module *module, struct dma_fence *fence,
> > > > > +             const struct dma_fence_ops *ops, spinlock_t *lock,
> > > > > +             u64 context, unsigned seqno)
> > > > >   {
> > > > >        BUG_ON(!lock);
> > > > >        BUG_ON(!ops || !ops->wait || !ops->enable_signaling ||
> > > > > @@ -571,7 +577,11 @@ struct default_wait_cb {
> > > > >        fence->seqno = seqno;
> > > > >        fence->flags = 0UL;
> > > > >        fence->error = 0;
> > > > > +     fence->owner = module;
> > > > > +
> > > > > +     if (!try_module_get(module))
> > > > > +             fence->owner = NULL;
> > > > >        trace_dma_fence_init(fence);
> > > > >   }
> > > > > -EXPORT_SYMBOL(dma_fence_init);
> > > > > +EXPORT_SYMBOL(_dma_fence_init);
> > > > Do we still need to export the symbol, it won't be called from outside
> > > > anymore? Other than that looks good to me:
> > > There's a big drawback in that a module reference is often insufficient,
> > > and that a reference on the driver (or whatever is required for the
> > > lifetime of the fence) will already hold the module reference.
> > > 
> > > Considering that we want a few 100k fences in flight per second, is
> > > there no other way to only export a fence with a module reference?
> > We'd need to make the timeline a full-blown object (Maarten owes me one
> > for that design screw-up), and then we could stuff all these things in
> > there.
> > 
> > And I think that's the right fix, since try_module_get for every
> > dma_fence_init just ain't cool really :-)
> > -Daniel
> Thanks for the feedback, Daniel.
> I see your point, but I am not sure how much impact an extra refcounting
> would create considering the whole effort of setting up a new fence. Also,
> this refcounting is not required for built-in modules.
> 
> As of now, unloading a kernel module that uses fence_init() is an easy way
> to bring down the system. This patch simply fixes that. What you have
> suggested sounds like a non-trivial effort which someone who is more
> familiar with this code base can do a better job than me. Perhaps we can
> take this patch now to fix the issue at hand and later somebody else can
> share a more optimal solution. :)

Module unload is a developer-only feature for a reason. Given that I don't
think fixing this with a hack is the right approach. And dma_fence_init()
is supposed to be really fast.

Also note that you can fix this already for your own driver by simply
waiting for all pending dma_fences to get released, so I don't think it's
super-important to land this asap.

Yes the real fix is a bit more involved, but shouldn't be too hard to pull
off really.
-Daniel

> 
> @Gustavo & @Sumit, I would like the maintainers to take a decision here.
> 
> -Akhil.

> _______________________________________________
> dri-devel mailing list
> dri-devel@xxxxxxxxxxxxxxxxxxxxx
> https://lists.freedesktop.org/mailman/listinfo/dri-devel


-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch