Re: [PATCH] dmabuf: Add the capability to expose DMA-BUF stats in sysfs

[Greg KH]

On Thu, Dec 10, 2020 at 11:27:27AM +0100, Daniel Vetter wrote:
> On Thu, Dec 10, 2020 at 11:10:45AM +0100, Greg KH wrote:
> > On Thu, Dec 10, 2020 at 10:58:50AM +0100, Christian König wrote:
> > > In general a good idea, but I have a few concern/comments here.
> > > 
> > > Am 10.12.20 um 05:43 schrieb Hridya Valsaraju:
> > > > This patch allows statistics to be enabled for each DMA-BUF in
> > > > sysfs by enabling the config CONFIG_DMABUF_SYSFS_STATS.
> > > > 
> > > > The following stats will be exposed by the interface:
> > > > 
> > > > /sys/kernel/dmabuf/<inode_number>/exporter_name
> > > > /sys/kernel/dmabuf/<inode_number>/size
> > > > /sys/kernel/dmabuf/<inode_number>/dev_map_info
> > > > 
> > > > The inode_number is unique for each DMA-BUF and was added earlier [1]
> > > > in order to allow userspace to track DMA-BUF usage across different
> > > > processes.
> > > > 
> > > > Currently, this information is exposed in
> > > > /sys/kernel/debug/dma_buf/bufinfo.
> > > > However, since debugfs is considered unsafe to be mounted in production,
> > > > it is being duplicated in sysfs.
> > > 
> > > Mhm, this makes it part of the UAPI. What is the justification for this?
> > > 
> > > In other words do we really need those debug information in a production
> > > environment?
> > 
> > Production environments seem to want to know who is using up memory :)
> 
> This only shows shared memory, so it does smell a lot like $specific_issue
> and we're designing a narrow solution for that and then have to carry it
> forever.

I think the "issue" is that this was a feature from ion that people
"missed" in the dmabuf move.  Taking away the ability to see what kind
of allocations were being made didn't make a lot of debugging tools
happy :(

But Hridya knows more, she's been dealing with the transition for a long
time now.

> E.g. why is the list of attachments not a sysfs link? That's how we
> usually expose struct device * pointers in sysfs to userspace, not as a
> list of things.

These aren't struct devices, so I don't understand the objection here.
Where else could these go in sysfs?

> Furthermore we don't have the exporter device covered anywhere, how is
> that tracked? Yes Android just uses ion for all shared buffers, but that's
> not how all of linux userspace works.

Do we have the exporter device link in the dmabuf interface?  If so,
great, let's use that, but for some reason I didn't think it was there.

> Then I guess there's the mmaps, you can fish them out of procfs. A tool
> which collects all that information might be useful, just as demonstration
> of how this is all supposed to be used.

There's a script somewhere that does this today, again, Hridya knows
more.

> There's also some things to make sure we're at least having thought about
> how other things fit in here. E.d. dma_resv attached to the dma-buf
> matters in general a lot. It doesn't matter on Android because
> everything's pinned all the time anyway.
> 
> Also I thought sysfs was one value one file, dumping an entire list into
> dev_info_map with properties we'll need to extend (once you care about
> dma_resv you also want to know which attachments are dynamic) does not
> smell like sysfs design at all.

sysfs is one value per file, what is being exported that is larger than
that here?  Did I miss something on review?

thanks,

greg k-h