On Fri, 9 Oct 2020 at 16:45, Johannes Berg <johannes@xxxxxxxxxxxxxxxx> wrote: > Hi Johannes, Thank you for the review! > On Fri, 2020-10-09 at 07:09 +0200, Nicolai Stange wrote: > > Johannes Berg <johannes@xxxxxxxxxxxxxxxx> writes: > > > > > On Thu, 2020-10-08 at 15:59 +0000, David Laight wrote: > > > > From: Taehee Yoo > > > > > Sent: 08 October 2020 16:49 > > > > > > > > > > When debugfs file is opened, its module should not be removed until > > > > > it's closed. > > > > > Because debugfs internally uses the module's data. > > > > > So, it could access freed memory. > > > > > > > > > > In order to avoid panic, it just sets .owner to THIS_MODULE. > > > > > So that all modules will be held when its debugfs file is opened. > > > > > > > > Can't you fix it in common code? > > > > Probably not: it's the call to ->release() that's faulting in the Oops > > quoted in the cover letter and that one can't be protected by the > > core debugfs code, unfortunately. > > > > There's a comment in full_proxy_release(), which reads as > > > > /* > > * We must not protect this against removal races here: the > > * original releaser should be called unconditionally in order > > * not to leak any resources. Releasers must not assume that > > * ->i_private is still being meaningful here. > > */ > > Yeah, found that too now :-) > > > > Yeah I was just wondering that too - weren't the proxy_fops even already > > > intended to fix this? > > > > No, as far as file_operations are concerned, the proxy fops's intent was > > only to ensure that the memory the file_operations' ->owner resides in > > is still valid so that try_module_get() won't splat at file open > > (c.f. [1]). > > Right. > > > You're right that the default "full" proxy fops do prevent all > > file_operations but ->release() from getting invoked on removed files, > > but the motivation had not been to protect the file_operations > > themselves, but accesses to any stale data associated with removed files > > ([2]). > > :) > > I actually got this to work in a crazy way, I'll send something out but > I'm sure it's a better idea to add the .owner everywhere, but please > let's do it in fewer than hundreds of patches :-) > Okay, as you mentioned earlier in 001/117 patch thread, I will squash patches into per-driver/subsystem then send them as v2. Thanks a lot! Taehee
