Re: [PATCH] pstore: Convert buf_lock to semaphore

[Arnd Bergmann]

On Sat, Dec 1, 2018 at 3:42 AM Kees Cook <keescook@xxxxxxxxxxxx> wrote:
> On Fri, Nov 30, 2018 at 2:52 PM Arnd Bergmann <arnd@xxxxxxxx> wrote:
> > On Fri, Nov 30, 2018 at 11:48 PM Kees Cook <keescook@xxxxxxxxxxxx> wrote:
> > >
> > > |BUG: sleeping function called from invalid context at kernel/sched/completion.c:99
> > > |in_atomic(): 1, irqs_disabled(): 1, pid: 2236, name: sig-xstate-bum
> > > |Preemption disabled at:
> > > |[<ffffffff99d60512>] pstore_dump+0x72/0x330
> > > |CPU: 26 PID: 2236 Comm: sig-xstate-bum Tainted: G      D           4.20.0-rc3 #45
> > > |Call Trace:
> > > | dump_stack+0x4f/0x6a
> > > | ___might_sleep.cold.91+0xd3/0xe4
> > > | __might_sleep+0x50/0x90
> > > | wait_for_completion+0x32/0x130
> > > | virt_efi_query_variable_info+0x14e/0x160
> > > | efi_query_variable_store+0x51/0x1a0
> > > | efivar_entry_set_safe+0xa3/0x1b0
> > > | efi_pstore_write+0x109/0x140
> > > | pstore_dump+0x11c/0x330
> > > | kmsg_dump+0xa4/0xd0
> > > | oops_exit+0x22/0x30
> > > ...
> > >
> > > Reported-by: Sebastian Andrzej Siewior <bigeasy@xxxxxxxxxxxxx>
> > > Fixes: 21b3ddd39fee ("efi: Don't use spinlocks for efi vars")
> > > Signed-off-by: Kees Cook <keescook@xxxxxxxxxxxx>
> >
> > Hmm, I've actually been working on a patch set recently to deprecate
> > all semaphores from the kernel and replace them with something
> > else as much as possible.
> >
> > Why can't this be a mutex instead?
>
> My understanding is that I can't use a mutex in interrupt context
> (Documentation/kernel-hacking/locking.rst) and pstore_dump() needs to
> handle being called from anywhere. I'm surprised it's managed to get
> away with using a spinlock for this long. :P

You are right that you can't take (or release) a mutex from interrupt
context. However, I don't think converting a spinlock to a semaphore
is going to help here either.

spinlock (or raw_spinlock) is generally the only thing that can be the
innermost lock that you take in any atomic context, and using
down_trylock doesn't make the context less atomic than it already is.

virt_efi_query_variable_info() however waits for a completion
and a semaphore, so that must not be called in atomic context.
Holding a semaphore instead of a spinlock is not going to help you
here, since the interrupt context means you might already be holding
arbitrary locks.

        Arnd