On x86-64, this_cpu_cmpxchg() uses CMPXCHG without LOCK prefix which
means it is only safe for the local CPU and not for multiple CPUs.
Recently the commit 36df6e3dbd ("cgroup: make css_rstat_updated nmi
safe") make css_rstat_updated lockless and uses lockless list to allow
reentrancy. Since css_rstat_updated can invoked from process context,
IRQ and NMI, it uses this_cpu_cmpxchg() to select the winner which will
inset the lockless lnode into the global per-cpu lockless list.
However the commit missed one case where lockless node of a cgroup can
be accessed and modified by another CPU doing the flushing. Basically
llist_del_first_init() in css_process_update_tree().
On a cursory look, it can be questioned how css_process_update_tree()
can see a lockless node in global lockless list where the updater is at
this_cpu_cmpxchg() and before llist_add() call in css_rstat_updated().
This can indeed happen in the presence of IRQs/NMI.
Consider this scenario: Updater for cgroup stat C on CPU A in process
context is after llist_on_list() check and before this_cpu_cmpxchg() in
css_rstat_updated() where it get interrupted by IRQ/NMI. In the IRQ/NMI
context, a new updater calls css_rstat_updated() for same cgroup C and
successfully inserts rstatc_pcpu->lnode.
Now concurrently CPU B is running the flusher and it calls
llist_del_first_init() for CPU A and got rstatc_pcpu->lnode of cgroup C
which was added by the IRQ/NMI updater.
Now imagine CPU B calling init_llist_node() on cgroup C's
rstatc_pcpu->lnode of CPU A and on CPU A, the process context updater
calling this_cpu_cmpxchg(rstatc_pcpu->lnode) concurrently.
The CMPXCNG without LOCK on CPU A is not safe and thus we need LOCK
prefix.
In Meta's fleet running the kernel with the commit 36df6e3dbd, we are
observing on some machines the memcg stats are getting skewed by more
than the actual memory on the system. On close inspection, we noticed
that lockless node for a workload for specific CPU was in the bad state
and thus all the updates on that CPU for that cgroup was being lost.
To confirm if this skew was indeed due to this CMPXCHG without LOCK in
css_rstat_updated(), we created a repro (using AI) at [1] which shows
that CMPXCHG without LOCK creates almost the same lnode corruption as
seem in Meta's fleet and with LOCK CMPXCHG the issue does not
reproduces.
Link: http://lore.kernel.org/efiagdwmzfwpdzps74fvcwq3n4cs36q33ij7eebcpssactv3zu@se4hqiwxcfxq [1]
Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
Cc: stable@vger.kernel.org # v6.17+
Fixes: 36df6e3dbd ("cgroup: make css_rstat_updated nmi safe")
Signed-off-by: Tejun Heo <tj@kernel.org>
3309b63a22