With multiple dax devices having the same node affinity, the kernel wrongly assigned
default_dram memory type to some devices after the memory hotplug operation. Fix this by
not clearing node_memory_types on the dax device remove.
The current kernel cleared node_memory_type on successful removal of a dax device.
But then we can have multiple dax devices with the same node affinity. Clearing the
node_memory_type results in assigning other dax devices to the default dram type when
we bring them online.
Signed-off-by: Aneesh Kumar K.V <[email protected]>
---
mm/memory-tiers.c | 37 +++++++++++++++++++++++++++++--------
1 file changed, 29 insertions(+), 8 deletions(-)
diff --git a/mm/memory-tiers.c b/mm/memory-tiers.c
index ba844fe9cc8c..c4bd6d052a33 100644
--- a/mm/memory-tiers.c
+++ b/mm/memory-tiers.c
@@ -27,9 +27,14 @@ struct demotion_nodes {
nodemask_t preferred;
};
+struct node_memory_type_map {
+ struct memory_dev_type *memtype;
+ int map_count;
+};
+
static DEFINE_MUTEX(memory_tier_lock);
static LIST_HEAD(memory_tiers);
-static struct memory_dev_type *node_memory_types[MAX_NUMNODES];
+static struct node_memory_type_map node_memory_types[MAX_NUMNODES];
static struct memory_dev_type *default_dram_type;
#ifdef CONFIG_MIGRATION
static int top_tier_adistance;
@@ -386,9 +391,19 @@ static inline void establish_demotion_targets(void) {}
static inline void __init_node_memory_type(int node, struct memory_dev_type *memtype)
{
- if (!node_memory_types[node]) {
- node_memory_types[node] = memtype;
- kref_get(&memtype->kref);
+ if (!node_memory_types[node].memtype)
+ node_memory_types[node].memtype = memtype;
+ /*
+ * for each device getting added in the same NUMA node
+ * with this specific memtype, bump the map count. We
+ * Only take memtype device reference once, so that
+ * changing a node memtype can be done by droping the
+ * only reference count taken here.
+ */
+
+ if (node_memory_types[node].memtype == memtype) {
+ if (!node_memory_types[node].map_count++)
+ kref_get(&memtype->kref);
}
}
@@ -406,7 +421,7 @@ static struct memory_tier *set_node_memory_tier(int node)
__init_node_memory_type(node, default_dram_type);
- memtype = node_memory_types[node];
+ memtype = node_memory_types[node].memtype;
node_set(node, memtype->nodes);
memtier = find_create_memory_tier(memtype);
if (!IS_ERR(memtier))
@@ -448,7 +463,7 @@ static bool clear_node_memory_tier(int node)
rcu_assign_pointer(pgdat->memtier, NULL);
synchronize_rcu();
- memtype = node_memory_types[node];
+ memtype = node_memory_types[node].memtype;
node_clear(node, memtype->nodes);
if (nodes_empty(memtype->nodes)) {
list_del_init(&memtype->tier_sibiling);
@@ -502,8 +517,14 @@ EXPORT_SYMBOL_GPL(init_node_memory_type);
void clear_node_memory_type(int node, struct memory_dev_type *memtype)
{
mutex_lock(&memory_tier_lock);
- if (node_memory_types[node] == memtype) {
- node_memory_types[node] = NULL;
+ if (node_memory_types[node].memtype == memtype)
+ node_memory_types[node].map_count--;
+ /*
+ * If we umapped all the attached devices to this node,
+ * clear the node memory type.
+ */
+ if (!node_memory_types[node].map_count) {
+ node_memory_types[node].memtype = NULL;
kref_put(&memtype->kref, release_memtype);
}
mutex_unlock(&memory_tier_lock);
--
2.37.2
On Fri, 26 Aug 2022 15:32:24 +0530 "Aneesh Kumar K.V" <[email protected]> wrote:
> With multiple dax devices having the same node affinity, the kernel wrongly assigned
> default_dram memory type to some devices after the memory hotplug operation. Fix this by
> not clearing node_memory_types on the dax device remove.
>
> The current kernel cleared node_memory_type on successful removal of a dax device.
> But then we can have multiple dax devices with the same node affinity. Clearing the
> node_memory_type results in assigning other dax devices to the default dram type when
> we bring them online.
Thanks, I added this as a fix against "mm/demotion/dax/kmem: set node's
abstract distance to MEMTIER_DEFAULT_DAX_ADISTANCE".
Reworked this patch to apply, reworked all the subsequent patches as
a result of applying that, checked that this patch as-sent cleanly reverts
when all are applied, checked that everything compiled at each step of
the resulting series.
"Aneesh Kumar K.V" <[email protected]> writes:
> With multiple dax devices having the same node affinity, the kernel wrongly assigned
> default_dram memory type to some devices after the memory hotplug operation. Fix this by
> not clearing node_memory_types on the dax device remove.
Sorry for late reply.
Just for confirmation. There are multiple dax devices in one NUMA node?
If you can show the bug reproducing steps, that will make it even easier
to understand.
Best Regards,
Huang, Ying
> The current kernel cleared node_memory_type on successful removal of a dax device.
> But then we can have multiple dax devices with the same node affinity. Clearing the
> node_memory_type results in assigning other dax devices to the default dram type when
> we bring them online.
>
> Signed-off-by: Aneesh Kumar K.V <[email protected]>
> ---
> mm/memory-tiers.c | 37 +++++++++++++++++++++++++++++--------
> 1 file changed, 29 insertions(+), 8 deletions(-)
>
> diff --git a/mm/memory-tiers.c b/mm/memory-tiers.c
> index ba844fe9cc8c..c4bd6d052a33 100644
> --- a/mm/memory-tiers.c
> +++ b/mm/memory-tiers.c
> @@ -27,9 +27,14 @@ struct demotion_nodes {
> nodemask_t preferred;
> };
>
> +struct node_memory_type_map {
> + struct memory_dev_type *memtype;
> + int map_count;
> +};
> +
> static DEFINE_MUTEX(memory_tier_lock);
> static LIST_HEAD(memory_tiers);
> -static struct memory_dev_type *node_memory_types[MAX_NUMNODES];
> +static struct node_memory_type_map node_memory_types[MAX_NUMNODES];
> static struct memory_dev_type *default_dram_type;
> #ifdef CONFIG_MIGRATION
> static int top_tier_adistance;
> @@ -386,9 +391,19 @@ static inline void establish_demotion_targets(void) {}
>
> static inline void __init_node_memory_type(int node, struct memory_dev_type *memtype)
> {
> - if (!node_memory_types[node]) {
> - node_memory_types[node] = memtype;
> - kref_get(&memtype->kref);
> + if (!node_memory_types[node].memtype)
> + node_memory_types[node].memtype = memtype;
> + /*
> + * for each device getting added in the same NUMA node
> + * with this specific memtype, bump the map count. We
> + * Only take memtype device reference once, so that
> + * changing a node memtype can be done by droping the
> + * only reference count taken here.
> + */
> +
> + if (node_memory_types[node].memtype == memtype) {
> + if (!node_memory_types[node].map_count++)
> + kref_get(&memtype->kref);
> }
> }
>
> @@ -406,7 +421,7 @@ static struct memory_tier *set_node_memory_tier(int node)
>
> __init_node_memory_type(node, default_dram_type);
>
> - memtype = node_memory_types[node];
> + memtype = node_memory_types[node].memtype;
> node_set(node, memtype->nodes);
> memtier = find_create_memory_tier(memtype);
> if (!IS_ERR(memtier))
> @@ -448,7 +463,7 @@ static bool clear_node_memory_tier(int node)
>
> rcu_assign_pointer(pgdat->memtier, NULL);
> synchronize_rcu();
> - memtype = node_memory_types[node];
> + memtype = node_memory_types[node].memtype;
> node_clear(node, memtype->nodes);
> if (nodes_empty(memtype->nodes)) {
> list_del_init(&memtype->tier_sibiling);
> @@ -502,8 +517,14 @@ EXPORT_SYMBOL_GPL(init_node_memory_type);
> void clear_node_memory_type(int node, struct memory_dev_type *memtype)
> {
> mutex_lock(&memory_tier_lock);
> - if (node_memory_types[node] == memtype) {
> - node_memory_types[node] = NULL;
> + if (node_memory_types[node].memtype == memtype)
> + node_memory_types[node].map_count--;
> + /*
> + * If we umapped all the attached devices to this node,
> + * clear the node memory type.
> + */
> + if (!node_memory_types[node].map_count) {
> + node_memory_types[node].memtype = NULL;
> kref_put(&memtype->kref, release_memtype);
> }
> mutex_unlock(&memory_tier_lock);
On 9/1/22 11:45 AM, Huang, Ying wrote:
> "Aneesh Kumar K.V" <[email protected]> writes:
>
>> With multiple dax devices having the same node affinity, the kernel wrongly assigned
>> default_dram memory type to some devices after the memory hotplug operation. Fix this by
>> not clearing node_memory_types on the dax device remove.
>
> Sorry for late reply.
>
> Just for confirmation. There are multiple dax devices in one NUMA node?
>
> If you can show the bug reproducing steps, that will make it even easier
> to understand.
>
NUMA nodes are assigned per region and you can have multiple devdax namespace with same NUMA node affinity.
dax0.1 and dax0.2 are examples. To recreate you can follow the below steps
root@ubuntu-guest:/sys/devices/system/node/node3# ls
compact cpumap meminfo memory34 memory36 memory38 memory41 memory43 memory45 memory47 memory50 memory52 memory54 numastat uevent
cpulist distance memory33 memory35 memory37 memory39 memory42 memory44 memory46 memory49 memory51 memory53 memory55 subsystem vmstat
root@ubuntu-guest:/sys/devices/system/node/node3#
root@ubuntu-guest:/sys/devices/system/node/node3# for mem in memory*; do echo 0 > $mem/online; done
root@ubuntu-guest:/sys/devices/system/node/node3# cd /sys/bus/dax/drivers
root@ubuntu-guest:/sys/bus/dax/drivers# echo dax0.0 > kmem/unbind
root@ubuntu-guest:/sys/bus/dax/drivers# cd /sys/devices/system/node/node3/
root@ubuntu-guest:/sys/devices/system/node/node3# ls
compact cpumap meminfo memory42 memory44 memory46 memory49 memory51 memory53 memory55 subsystem vmstat
cpulist distance memory41 memory43 memory45 memory47 memory50 memory52 memory54 numastat uevent
root@ubuntu-guest:/sys/devices/system/node/node3# for mem in memory*; do echo 1 > $mem/online; done
-aneesh
Aneesh Kumar K V <[email protected]> writes:
> On 9/1/22 11:45 AM, Huang, Ying wrote:
>> "Aneesh Kumar K.V" <[email protected]> writes:
>>
>>> With multiple dax devices having the same node affinity, the kernel wrongly assigned
>>> default_dram memory type to some devices after the memory hotplug operation. Fix this by
>>> not clearing node_memory_types on the dax device remove.
>>
>> Sorry for late reply.
>>
>> Just for confirmation. There are multiple dax devices in one NUMA node?
>>
>> If you can show the bug reproducing steps, that will make it even easier
>> to understand.
>>
>
> NUMA nodes are assigned per region and you can have multiple devdax namespace with same NUMA node affinity.
I think that the patch description will be clearer if we added the above
line into it.
> dax0.1 and dax0.2 are examples. To recreate you can follow the below steps
>
>
> root@ubuntu-guest:/sys/devices/system/node/node3# ls
> compact cpumap meminfo memory34 memory36 memory38 memory41 memory43 memory45 memory47 memory50 memory52 memory54 numastat uevent
> cpulist distance memory33 memory35 memory37 memory39 memory42 memory44 memory46 memory49 memory51 memory53 memory55 subsystem vmstat
> root@ubuntu-guest:/sys/devices/system/node/node3#
> root@ubuntu-guest:/sys/devices/system/node/node3# for mem in memory*; do echo 0 > $mem/online; done
> root@ubuntu-guest:/sys/devices/system/node/node3# cd /sys/bus/dax/drivers
> root@ubuntu-guest:/sys/bus/dax/drivers# echo dax0.0 > kmem/unbind
> root@ubuntu-guest:/sys/bus/dax/drivers# cd /sys/devices/system/node/node3/
> root@ubuntu-guest:/sys/devices/system/node/node3# ls
> compact cpumap meminfo memory42 memory44 memory46 memory49 memory51 memory53 memory55 subsystem vmstat
> cpulist distance memory41 memory43 memory45 memory47 memory50 memory52 memory54 numastat uevent
> root@ubuntu-guest:/sys/devices/system/node/node3# for mem in memory*; do echo 1 > $mem/online; done
Thanks a lot for detailed information.
Best Regards,
Huang, Ying