On Thu 20-01-22 15:20:14, Christoph Hellwig wrote: > On Thu, Jan 13, 2022 at 11:44:24AM +0100, Jan Kara wrote: > > Maybe the most disputable thing in this locking chain seems to be splicing > > from sysfs files. That does not seem terribly useful and due to special > > locking and behavior of sysfs files it allows for creating interesting lock > > dependencies. OTOH maybe there is someone out there who (possibly > > inadvertedly through some library) ends up using splice on sysfs files so > > chances for userspace breakage, if we disable splice for sysfs, would be > > non-negligible. Hum, tough. > > People were using sendfile on sysfs files, that is why support for this > got added back after it was removed for a while as part of the set_fs() > removal. > > The real question for me is why do we need freeing and writer counts on > sysfs or any other pure in-memory file system to start with? We don't. But freezing of sysfs is not part of the locking chain. Only freezing of the filesystem holding loopback backing file is (let's call that fs F). And splice from sysfs to some file in F is what ties freezing of F with a file lock in sysfs... Honza -- Jan Kara <jack@xxxxxxxx> SUSE Labs, CR
