This reverts commits 4bad58ebc8 (and
399f8dd9a8, which tried to fix it).
I do not believe these are correct, and I'm about to release 5.13, so am
reverting them out of an abundance of caution.
The locking is odd, and appears broken.
On the allocation side (in __sigqueue_alloc()), the locking is somewhat
straightforward: it depends on sighand->siglock. Since one caller
doesn't hold that lock, it further then tests 'sigqueue_flags' to avoid
the case with no locks held.
On the freeing side (in sigqueue_cache_or_free()), there is no locking
at all, and the logic instead depends on 'current' being a single
thread, and not able to race with itself.
To make things more exciting, there's also the data race between freeing
a signal and allocating one, which is handled by using WRITE_ONCE() and
READ_ONCE(), and being mutually exclusive wrt the initial state (ie
freeing will only free if the old state was NULL, while allocating will
obviously only use the value if it was non-NULL, so only one or the
other will actually act on the value).
However, while the free->alloc paths do seem mutually exclusive thanks
to just the data value dependency, it's not clear what the memory
ordering constraints are on it. Could writes from the previous
allocation possibly be delayed and seen by the new allocation later,
causing logical inconsistencies?
So it's all very exciting and unusual.
And in particular, it seems that the freeing side is incorrect in
depending on "current" being single-threaded. Yes, 'current' is a
single thread, but in the presense of asynchronous events even a single
thread can have data races.
And such asynchronous events can and do happen, with interrupts causing
signals to be flushed and thus free'd (for example - sending a
SIGCONT/SIGSTOP can happen from interrupt context, and can flush
previously queued process control signals).
So regardless of all the other questions about the memory ordering and
locking for this new cached allocation, the sigqueue_cache_or_free()
assumptions seem to be fundamentally incorrect.
It may be that people will show me the errors of my ways, and tell me
why this is all safe after all. We can reinstate it if so. But my
current belief is that the WRITE_ONCE() that sets the cached entry needs
to be a smp_store_release(), and the READ_ONCE() that finds a cached
entry needs to be a smp_load_acquire() to handle memory ordering
correctly.
And the sequence in sigqueue_cache_or_free() would need to either use a
lock or at least be interrupt-safe some way (perhaps by using something
like the percpu 'cmpxchg': it doesn't need to be SMP-safe, but like the
percpu operations it needs to be interrupt-safe).
Fixes: 399f8dd9a8 ("signal: Prevent sigqueue caching after task got released")
Fixes: 4bad58ebc8 ("signal: Allow tasks to cache one sigqueue struct")
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Christian Brauner <christian.brauner@ubuntu.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
b4b27b9eed