[PATCH 2/2] dmabuf/tracing: Add dma-buf trace events

Daniel Vetter daniel.vetter at ffwll.ch
Tue Aug 4 20:26:52 UTC 2020


On Tue, Aug 4, 2020 at 12:28 AM Al Viro <viro at zeniv.linux.org.uk> wrote:
>
> On Mon, Aug 03, 2020 at 09:22:53AM -0700, Suren Baghdasaryan wrote:
> > On Mon, Aug 3, 2020 at 9:12 AM Matthew Wilcox <willy at infradead.org> wrote:
> > >
> > > On Mon, Aug 03, 2020 at 09:00:00AM -0700, Suren Baghdasaryan wrote:
> > > > On Mon, Aug 3, 2020 at 8:41 AM Matthew Wilcox <willy at infradead.org> wrote:
> > > > >
> > > > > On Mon, Aug 03, 2020 at 02:47:19PM +0000, Kalesh Singh wrote:
> > > > > > +static void dma_buf_fd_install(int fd, struct file *filp)
> > > > > > +{
> > > > > > +     trace_dma_buf_fd_ref_inc(current, filp);
> > > > > > +}
> > > > >
> > > > > You're adding a new file_operation in order to just add a new tracepoint?
> > > > > NACK.
> > > >
> > > > Hi Matthew,
> > > > The plan is to attach a BPF to this tracepoint in order to track
> > > > dma-buf users. If you feel this is an overkill, what would you suggest
> > > > as an alternative?
> > >
> > > I'm sure BPF can attach to fd_install and filter on file->f_ops belonging
> > > to dma_buf, for example.
> >
> > Sounds like a workable solution. Will explore that direction. Thanks Matthew!
>
> No, it is not a solution at all.
>
> What kind of locking would you use?  With _any_ of those approaches.
>
> How would you use the information that is hopelessly out of date/incoherent/whatnot
> at the very moment you obtain it?
>
> IOW, what the hell is that horror for?  You do realize, for example, that there's
> such thing as dup(), right?  And dup2() as well.  And while we are at it, how
> do you keep track of removals, considering the fact that you can stick a file
> reference into SCM_RIGHTS datagram sent to yourself, close descriptors and an hour
> later pick that datagram, suddenly getting descriptor back?
>
> Besides, "I have no descriptors left" != "I can't be currently sitting in the middle
> of syscall on that sucker"; close() does *NOT* terminate ongoing operations.
>
> You are looking at the drastically wrong abstraction level.  Please, describe what
> it is that you are trying to achieve.

For added entertainment (since this is specifically about dma-buf) you
can stuff them into various gpu drivers, and convert to a native gpu
driver handle thing. That's actually the expected use case, first a
buffer sharing gets established with AF_UNIX, then both sides close
the dma-buf fd handle.

GPU drivers then internally cache the struct file so that we can hand
out the same (to avoid confusion when re-importing it on some other
driver), so for the case of dma-buf the "it's not actually an
installed fd anywhere for unlimited time" is actually the normal
use-case, not some odd corner.

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


More information about the dri-devel mailing list