[ add tools@xxxxxxxxxx for bugzilla bridge mention ] Linux regression tracking (Thorsten Leemhuis) wrote: > On 24.09.24 04:54, Dan Williams wrote: > > Dan Williams wrote: > >> Greg Kroah-Hartman wrote: > >> [..] > >>> > >>> This is odd. > >>> > >>> Does the latest 6.10.y release also show this problem? > >>> > >>> I can't duplicate this here, and it's the first I've heard of it (given > >>> that USB mice are pretty popular, I would suspect others would have hit > >>> it as well...) > >> > >> Sorry for missing this earlier. One thought is that userspace has a > >> dependency on uevent_show() flushing device probing. In other words the > >> side effect of taking the device_lock() in uevent_show() is that udev > >> might enjoy some occasions where the reading the uevent flushes probing > >> before the udev rule runs. With this change, uevent_show() no longer > >> waits for any inflight probes to complete. > >> > >> One idea to fix this problem is to create a special case sysfs attribute > >> type that takes the device_lock() before kernfs_get_active() to avoid > >> the deadlock on attribute teardown. > >> > >> I'll take a look. Thanks for forwarding the report Thorsten! > > > > Ok, the following boots and passes the CXL unit tests, would appreciate > > if the reporter can give this a try: > > Somehow I apparently became a "bugzilla-man-in-the-middle interface" yet > again... But whatever! ¯\_(ツ)_/¯ Oh, sorry about that Thorsten, for some reason I thought we were living in this new world that Konstantin mentions where mailing-list to bugzilla communication is automated. Hopefully that arrives soon. I very much appreciate the help here and thank you for the hard work you put into taking care of regressions! > > To forward the latest comment from the ticket: > > """ > --- Comment #11 from brmails+k@xxxxxxxxxxx --- > Good news! > > I think the proposed patch by Dan Williams fixes the issue. > > I have tested it with v6.6.52 and v6.10.11. I haven't been able to > recreate the > issue with those modified kernels even once. > > The patch can be applied to v6.11.0 and v6.10.11 out of the box. For > v6.6.52 I > had to slightly modify it as the line > > #define SYSFS_GROUP_INVISIBLE 020000 > > doesn't exist in /include/linux/sysfs.h in v6.6.52 hence the patch > looking for > that line fails on that file. > But after adjusting the patch accordingly, the patch works fine on > v6.6.52 and > the issue is gone with the patched version of v6.6.52, just like 6.10.11. > > So, I assume that the fix / patch proposed by Dan Williams works as intended > resolving the issue I had. > > Thanks again for forwarding the bug report and for the quick fix! > """ Great news, I'll revise the patch a bit to add some clarifying documentation. Specifically, that the new DEVICE_ATTR_RW_LOCKED() is only for exceptional corner cases like uevent_show(), and that most usages of "dev->driver" in a sysfs attribute are best served by being declared as part of "drv_groups".
