[Problem]
cpuid <-> nodeid mapping is firstly established at boot time. And
workqueue caches
the mapping in wq_numa_possible_cpumask in wq_numa_init() at boot time.
When doing node online/offline, cpuid <-> nodeid mapping is
established/destroyed,
which means, cpuid <-> nodeid mapping will change if node hotplug
happens. But
workqueue does not update wq_numa_possible_cpumask.
So here is the problem:
Assume we have the following cpuid <-> nodeid in the beginning:
Node | CPU
------------------------
node 0 | 0-14, 60-74
node 1 | 15-29, 75-89
node 2 | 30-44, 90-104
node 3 | 45-59, 105-119
and we hot-remove node2 and node3, it becomes:
Node | CPU
------------------------
node 0 | 0-14, 60-74
node 1 | 15-29, 75-89
and we hot-add node4 and node5, it becomes:
Node | CPU
------------------------
node 0 | 0-14, 60-74
node 1 | 15-29, 75-89
node 4 | 30-59
node 5 | 90-119
But in wq_numa_possible_cpumask, cpu30 is still mapped to node2, and
the like.
When a pool workqueue is initialized, if its cpumask belongs to a
node, its
pool->node will be mapped to that node. And memory used by this
workqueue will
also be allocated on that node.
static struct worker_pool *get_unbound_pool(const struct
workqueue_attrs *attrs){
...
/* if cpumask is contained inside a NUMA node, we belong to
that node */
if (wq_numa_enabled) {
for_each_node(node) {
if (cpumask_subset(pool->attrs->cpumask,
wq_numa_possible_cpumask[node])) {
pool->node = node;
break;
}
}
}
Since wq_numa_possible_cpumask is not updated, it could be mapped to
an offline node,
which will lead to memory allocation failure:
SLUB: Unable to allocate memory on node 2 (gfp=0x80d0)
cache: kmalloc-192, object size: 192, buffer size: 192, default
order: 1, min order: 0
node 0: slabs: 6172, objs: 259224, free: 245741
node 1: slabs: 3261, objs: 136962, free: 127656
It happens here:
create_worker(struct worker_pool *pool)
|--> worker = alloc_worker(pool->node);
static struct worker *alloc_worker(int node)
{
struct worker *worker;
worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
--> Here, useing the wrong node.
......
return worker;
}
[Solution]
There are four mappings in the kernel:
1. nodeid (logical node id) <-> pxm
2. apicid (physical cpu id) <-> nodeid
3. cpuid (logical cpu id) <-> apicid
4. cpuid (logical cpu id) <-> nodeid
1. pxm (proximity domain) is provided by ACPI firmware in SRAT, and
nodeid <-> pxm
mapping is setup at boot time. This mapping is persistent, won't
change.
2. apicid <-> nodeid mapping is setup using info in 1. The mapping is
setup at boot
time and CPU hotadd time, and cleared at CPU hotremove time. This
mapping is also
persistent.
3. cpuid <-> apicid mapping is setup at boot time and CPU hotadd
time. cpuid is
allocated, lower ids first, and released at CPU hotremove time,
reused for other
hotadded CPUs. So this mapping is not persistent.
4. cpuid <-> nodeid mapping is also setup at boot time and CPU hotadd
time, and
cleared at CPU hotremove time. As a result of 3, this mapping is
not persistent.
To fix this problem, we establish cpuid <-> nodeid mapping for all
the possible
cpus at boot time, and make it persistent. And according to
init_cpu_to_node(),
cpuid <-> nodeid mapping is based on apicid <-> nodeid mapping and
cpuid <-> apicid
mapping. So the key point is obtaining all cpus' apicid.
apicid can be obtained by _MAT (Multiple APIC Table Entry) method or
found in
MADT (Multiple APIC Description Table). So we finish the job in the
following steps:
1. Enable apic registeration flow to handle both enabled and disabled
cpus.
This is done by introducing an extra parameter to
generic_processor_info to let the
caller control if disabled cpus are ignored.
2. Introduce a new array storing all possible cpuid <-> apicid
mapping. And also modify
the way cpuid is calculated. Establish all possible cpuid <->
apicid mapping when
registering local apic. Store the mapping in this array.
3. Enable _MAT and MADT relative apis to return non-presnet or
disabled cpus' apicid.
This is also done by introducing an extra parameter to these apis
to let the caller
control if disabled cpus are ignored.
4. Establish all possible cpuid <-> nodeid mapping.
This is done via an additional acpi namespace walk for processors.
For previous discussion, please refer to:
https://lkml.org/lkml/2015/2/27/145
https://lkml.org/lkml/2015/3/25/989
https://lkml.org/lkml/2015/5/14/244
https://lkml.org/lkml/2015/7/7/200
https://lkml.org/lkml/2015/9/27/209
Change log v4 -> v5:
1. Remove useless code in patch 1.
2. Small improvement of commit message.
Change log v3 -> v4:
1. Fix the kernel panic at boot time. The cause is that I tried to
build zonelists
before per cpu areas were initialized.
Change log v2 -> v3:
1. Online memory-less nodes at boot time to map cpus of memory-less
nodes.
2. Build zonelists for memory-less nodes so that memory allocator
will fall
back to proper nodes automatically.
Change log v1 -> v2:
1. Split code movement and actual changes. Add patch 1.
2. Synchronize best near online node record when node hotplug
happens. In patch 2.
3. Fix some comment.
Gu Zheng (4):
x86, acpi, cpu-hotplug: Enable acpi to register all possible cpus at
boot time.
x86, acpi, cpu-hotplug: Introduce cpuid_to_apicid[] array to store
persistent cpuid <-> apicid mapping.
x86, acpi, cpu-hotplug: Enable MADT APIs to return disabled apicid.
x86, acpi, cpu-hotplug: Set persistent cpuid <-> nodeid mapping when
booting.
Tang Chen (1):
x86, memhp, numa: Online memory-less nodes at boot time.
arch/ia64/kernel/acpi.c | 2 +-
arch/x86/include/asm/mpspec.h | 1 +
arch/x86/kernel/acpi/boot.c | 8 ++-
arch/x86/kernel/apic/apic.c | 85 +++++++++++++++++++++++++----
arch/x86/mm/numa.c | 27 +++++-----
drivers/acpi/acpi_processor.c | 5 +-
drivers/acpi/bus.c | 3 ++
drivers/acpi/processor_core.c | 122
++++++++++++++++++++++++++++++++++--------
include/linux/acpi.h | 2 +
9 files changed, 204 insertions(+), 51 deletions(-)