On Fri, Jan 15, 2021 at 04:26:21PM +0100, Alexander Potapenko wrote:
> > sysfs is "one value per file"
> What about existing interfaces that even export binary blobs through
> sysfs (e.g. /sys/firmware, /sys/boot_params)?
> What qualifies as a "value" in that case?

binary files are fine as the kernel is just a "pipe" through to the
hardware / firmware.  No translation or parsing happens in the kernel.
And that's NOT trace events :)

> > please put something like this in
> > tracefs, as there is no such rules there.  Or debugfs, but please, not
> > sysfs.
> Does tracefs have anything similar to sysfs_notify() or any other way
> to implement a poll() handler?

Don't know, look and see!

> Our main goal is to let users wait on poll(), so that they don't have
> to check the file for new contents every now and then. Is it possible
> with tracefs or debugfs?

debugfs, yes, you can do whatever you want.  tracefs probably has this,
otherwise how would trace tools work?  :)

> Also, for our goal debugfs isn't a particularly good fit, as Android
> kernels do not enable debugfs.

Do you care about Android kernels?  If so, yes, debugfs is not good.
And have you asked the Android kernel team about this?

> Not sure about tracefs, they seem to have it, need to check.

It should be there.

> Do you think it is viable to keep
> /sys/kernel/error_report/report_count in sysfs and use it for
> notifications, but move last_report somewhere else?

No, not at all, please keep it out of sysfs.

> (I'd probably prefer procfs, but it could be tracefs as well, if you
> find that better).

If it does not have to do with processes, it does not belong in procfs.

Again, this seems to fit in with tracing, so please use tracefs.

> This way it will still be possible to easily notify userspace about
> new reports, but the report itself won't have any atomicity
> guarantees. Users will have to check the report count to ensure it
> didn't change under their feet.

Again, use a file in tracefs.

thanks,

greg k-h