On Tue, Jan 11, 2022 at 3:43 AM Christian König christian.koenig@amd.com wrote:
Am 11.01.22 um 12:16 schrieb Greg Kroah-Hartman:
On Tue, Jan 11, 2022 at 11:58:07AM +0100, Christian König wrote:
This is also not a problem due to the high number of DMA-BUF exports during launch time, as even a single export can be delayed for an unpredictable amount of time. We cannot eliminate DMA-BUF exports completely during app-launches and we are unfortunately seeing reports of the exporting process occasionally sleeping long enough to cause user-visible jankiness :(
We also looked at whether any optimizations are possible from the kernfs implementation side[1] but the semaphore is used quite extensively and it looks like the best way forward would be to remove sysfs creation/teardown from the DMA-BUF export/release path altogether. We have some ideas on how we can reduce the code-complexity in the current patch. If we manage to simplify it considerably, would the approach of offloading sysfs creation and teardown into a separate thread be acceptable Christian?
At bare minimum I suggest to use a work_struct instead of re-inventing that with kthread.
And then only put the exporting of buffers into the background and not the teardown.
Thank you for the guidance!
One worry I have here with doing this async that now userspace might have a dma-buf, but the sysfs entry does not yet exist, or the dma-buf is gone, but the sysfs entry still exists. That's a bit awkward wrt semantics.
Thank you all for your thoughts and guidance. You are correct that we will be trading accuracy for performance here. One precedence we could find was in the case of RSS accounting where SPLIT_RSS_COUNTING caused the accounting to have less overhead but also made it less accurate. If you would prefer that it not be the default case, we can make it configurable by putting it behind a config instead.
Also I'm pretty sure that if we can hit this, then other subsystems using kernfs have similar problems, so trying to fix this in kernfs with slightly more fine-grained locking sounds like a much more solid approach. The linked patch talks about how the big delays happen due to direct reclaim, and that might be limited to specific code paths that we need to look at? As-is this feels a bit much like papering over kernfs issues in hackish ways in sysfs users, instead of tackling the problem at its root.
Which is exactly my feeling as well, yes.
More and more people are using sysfs/kernfs now for things that it was never designed for (i.e. high-speed statistic gathering). That's not the fault of kernfs, it's the fault of people thinking it can be used for stuff like that :)
I'm starting to get the feeling that we should maybe have questioned adding sysfs files for each exported DMA-buf a bit more. Anyway, to late for that. We have to live with the consequences.
But delays like this is odd, tearing down sysfs attributes should normally _never_ be a fast-path that matters to system throughput. So offloading it to a workqueue makes sense as the attributes here are for objects that are on the fast-path.
That's what is puzzling me as well. As far as I understood Hridya tearing down things is not the problem, because during teardown we usually have a dying task where it's usually not much of a problem if the corpse is around for another few milliseconds until everything is cleaned up.
We have seen instances where the last reference to the buffer is not dropped by the dying process but by Surfaceflinger[1].
The issue happens during creation of the sysfs attribute and that's extremely odd because if this waits for reclaim then drivers will certainly wait for reclaim as well. See we need a few bytes for the sysfs attribute, but drivers usually need a few megabytes for the DMA-buf backing store before they can even export the DMA-buf.
We have been working off of traces collected from the devices of end users to analyze the issue and currently don't have sufficient information to understand why exactly direct reclaim affects sysfs the way we are seeing it on the traces. We are actively trying to reproduce the issue consistently to perform more experiments to understand it. The DMA-BUF system heap on the Android Common Kernel keeps a pool of pre-allocated pages in reserve and we are guessing that it could possibly be the reason why we have not seen similar issues with direct reclaim earlier. We will update the thread once we have more information.
We are also working on a leaner version of the patch that uses work_struct instead.
Regards, Hridya
[1] : https://source.android.com/devices/graphics/surfaceflinger-windowmanager
So something doesn't add up in the rational for this problem.
Regards, Christian.
thanks,
greg k-h