On Thu, Sep 28, 2017 at 09:00:52PM -0700, Linus Torvalds wrote: > On Thu, Sep 28, 2017 at 8:32 PM, Kyle Sanderson <kyle.l...@gmail.com> wrote: > > Not sure if the stack is crap or not, but this looks like an RCU crash? > > > > https://i.imgur.com/sBnNe1p.jpg > > Hmm. Not the clearest picture, and the "Code:" line in particular is > missing the interesting part, but at a guess it's taking a fault in > put_cred(), which inlines to > > if (atomic_dec_and_test(&(cred)->usage)) > __put_cred(cred); > > and I think it's that "cred" pointer that may be NULL, which makes > "&(cred)->usage" be a NULL pointer too, and you get a page fault when > it tries to decrement the usage count. > > Now, it goes without saying that the cred pointer should never *be* > NULL on a filp that is on the RCU freeing list, because we always > initialize file->f_cred when we allocate a file to the current creds. > > So there's something odd going on. Possibly entirely unrelated memory > corruption. > > Nothing obvious stands out, I think we'd need to see more of a pattern > of the problem to see what is up.
Kyle, if this is reproducible, please build your kernel with CONFIG_DEBUG_OBJECTS_RCU_HEAD=y. This can detect the call_rcu() equivalent of a double free, and these double frees have sometimes resulted in output looking something like the .jpg along with the NULL-pointer problem that Linus suspects. Thanx, Paul