On POWER platforms, the hypervisor can notify the guest kernel about dynamic
changes in the cpu-numa associativity (VPHN topology update). Hence the
cpu-to-node mappings that we got from the firmware during boot, may no longer
be valid after such updates. This is handled using the arch_update_cpu_topology()
hook in the scheduler, and the sched-domains are rebuilt according to the new
mappings.
But unfortunately, at the moment, CPU hotplug ignores these updated mappings
and instead queries the firmware for the cpu-to-numa relationships and uses
them during CPU online. So the kernel can end up assigning wrong NUMA nodes
to CPUs during subsequent CPU hotplug online operations (after booting).
Further, a particularly problematic scenario can result from this bug:
On POWER platforms, the SMT mode can be switched between 1, 2, 4 (and even 8)
threads per core. The switch to Single-Threaded (ST) mode is performed by
offlining all except the first CPU thread in each core. Switching back to
SMT mode involves onlining those other threads back, in each core.
Now consider this scenario:
1. During boot, the kernel gets the cpu-to-node mappings from the firmware
and assigns the CPUs to NUMA nodes appropriately, during CPU online.
2. Later on, the hypervisor updates the cpu-to-node mappings dynamically and
communicates this update to the kernel. The kernel in turn updates its
cpu-to-node associations and rebuilds its sched domains. Everything is
fine so far.
3. Now, the user switches the machine from SMT to ST mode (say, by running
ppc64_cpu --smt=1). This involves offlining all except 1 thread in each
core.
4. The user then tries to switch back from ST to SMT mode (say, by running
ppc64_cpu --smt=4), and this involves onlining those threads back. Since
CPU hotplug ignores the new mappings, it queries the firmware and tries to
associate the newly onlined sibling threads to the old NUMA nodes. This
results in sibling threads within the same core getting associated with
different NUMA nodes, which is incorrect.
The scheduler's build-sched-domains code gets thoroughly confused with this
and enters an infinite loop and causes soft-lockups, as explained in detail
in commit 3be7db6ab (powerpc: VPHN topology change updates all siblings).
So to fix this, use the numa_cpu_lookup_table to remember the updated
cpu-to-node mappings, and use them during CPU hotplug online operations.
Further, we also need to ensure that all threads in a core are assigned to a
common NUMA node, irrespective of whether all those threads were online during
the topology update. To achieve this, we take care not to use cpu_sibling_mask()
since it is not hotplug invariant. Instead, we use cpu_first_sibling_thread()
and set up the mappings manually using the 'threads_per_core' value for that
particular platform. This helps us ensure that we don't hit this bug with any
combination of CPU hotplug and SMT mode switching.
Cc: stable@vger.kernel.org
Signed-off-by: Srivatsa S. Bhat <srivatsa.bhat@linux.vnet.ibm.com>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
d4edc5b6c4