For hugetlb backed jobs/VMs it's critical to understand the numa
information for the memory backing these jobs to deliver optimal
performance.
Currently this technically can be queried from /proc/self/numa_maps, but
there are significant issues with that. Namely:
1. Memory can be mapped or unmapped.
2. numa_maps are per process and need to be aggregated across all
processes in the cgroup. For shared memory this is more involved as
the userspace needs to make sure it doesn't double count shared
mappings.
3. I believe querying numa_maps needs to hold the mmap_lock which adds
to the contention on this lock.
For these reasons I propose simply adding hugetlb.*.numa_stat file,
which shows the numa information of the cgroup similarly to
memory.numa_stat.
On cgroup-v2:
cat /sys/fs/cgroup/unified/test/hugetlb.2MB.numa_stat
total=2097152 N0=2097152 N1=0
On cgroup-v1:
cat /sys/fs/cgroup/hugetlb/test/hugetlb.2MB.numa_stat
total=2097152 N0=2097152 N1=0
hierarichal_total=2097152 N0=2097152 N1=0
This patch was tested manually by allocating hugetlb memory and querying
the hugetlb.*.numa_stat file of the cgroup and its parents.

Cc: Mike Kravetz <[email protected]>
Cc: Andrew Morton <[email protected]>
Cc: Shuah Khan <[email protected]>
Cc: Miaohe Lin <[email protected]>
Cc: Oscar Salvador <[email protected]>
Cc: Michal Hocko <[email protected]>
Cc: Muchun Song <[email protected]>
Cc: David Rientjes <[email protected]>
Cc: Shakeel Butt <[email protected]>
Cc: Jue Wang <[email protected]>
Cc: Yang Yao <[email protected]>
Cc: Joanna Li <[email protected]>
Cc: Cannon Matthews <[email protected]>
Cc: [email protected]
Cc: [email protected]
Signed-off-by: Mina Almasry <[email protected]>
---
Changes in v8:
- Do not use unsigned long *usage to WRITE_ONCE() new usage values.
Changes in v7:
- Converted back usage to unsigned long + READ_ONCE/WRITE_ONCE rather
than atomic_long_t
Changes in v6:
- Changed usage from unsigned long to atomic_long_t
Changes in v5:
- Fixed commit message typo.
- Fixed per node usage documentation to be in pages.
- Removed unnecessary h_cg check.
Changes in v4:
- Removed unnecessary braces.
- usage is now counted in pages instead of bytes.
- Reverted unneeded changes to write_to_hugetlbfs.c
Changes in v3:
- Fixed typos (sorry!)
- Used conventional locations for cgroups mount points in docs/commit
message.
- Updated docs.
- Handle kzalloc_node failure, and proper deallocation of per node data.
- Use struct_size() to calculate the struct size.
- Use nr_node_ids instead of MAX_NUMNODES.
- Updated comments per multi-line comment pattern.
Changes in v2:
- Fix warning Reported-by: kernel test robot <[email protected]>
---
.../admin-guide/cgroup-v1/hugetlb.rst | 4 +
Documentation/admin-guide/cgroup-v2.rst | 5 +
include/linux/hugetlb.h | 4 +-
include/linux/hugetlb_cgroup.h | 7 +
mm/hugetlb_cgroup.c | 133 ++++++++++++++++--
5 files changed, 141 insertions(+), 12 deletions(-)
diff --git a/Documentation/admin-guide/cgroup-v1/hugetlb.rst b/Documentation/admin-guide/cgroup-v1/hugetlb.rst
index 338f2c7d7a1c..0fa724d82abb 100644
--- a/Documentation/admin-guide/cgroup-v1/hugetlb.rst
+++ b/Documentation/admin-guide/cgroup-v1/hugetlb.rst
@@ -29,12 +29,14 @@ Brief summary of control files::
hugetlb.<hugepagesize>.max_usage_in_bytes # show max "hugepagesize" hugetlb usage recorded
hugetlb.<hugepagesize>.usage_in_bytes # show current usage for "hugepagesize" hugetlb
hugetlb.<hugepagesize>.failcnt # show the number of allocation failure due to HugeTLB usage limit
+ hugetlb.<hugepagesize>.numa_stat # show the numa information of the hugetlb memory charged to this cgroup
For a system supporting three hugepage sizes (64k, 32M and 1G), the control
files include::
hugetlb.1GB.limit_in_bytes
hugetlb.1GB.max_usage_in_bytes
+ hugetlb.1GB.numa_stat
hugetlb.1GB.usage_in_bytes
hugetlb.1GB.failcnt
hugetlb.1GB.rsvd.limit_in_bytes
@@ -43,6 +45,7 @@ files include::
hugetlb.1GB.rsvd.failcnt
hugetlb.64KB.limit_in_bytes
hugetlb.64KB.max_usage_in_bytes
+ hugetlb.64KB.numa_stat
hugetlb.64KB.usage_in_bytes
hugetlb.64KB.failcnt
hugetlb.64KB.rsvd.limit_in_bytes
@@ -51,6 +54,7 @@ files include::
hugetlb.64KB.rsvd.failcnt
hugetlb.32MB.limit_in_bytes
hugetlb.32MB.max_usage_in_bytes
+ hugetlb.32MB.numa_stat
hugetlb.32MB.usage_in_bytes
hugetlb.32MB.failcnt
hugetlb.32MB.rsvd.limit_in_bytes
diff --git a/Documentation/admin-guide/cgroup-v2.rst b/Documentation/admin-guide/cgroup-v2.rst
index 4d8c27eca96b..356847f8f008 100644
--- a/Documentation/admin-guide/cgroup-v2.rst
+++ b/Documentation/admin-guide/cgroup-v2.rst
@@ -2252,6 +2252,11 @@ HugeTLB Interface Files
are local to the cgroup i.e. not hierarchical. The file modified event
generated on this file reflects only the local events.
+ hugetlb.<hugepagesize>.numa_stat
+ Similar to memory.numa_stat, it shows the numa information of the
+ hugetlb pages of <hugepagesize> in this cgroup. Only active in
+ use hugetlb pages are included. The per-node values are in bytes.
+
Misc
----
diff --git a/include/linux/hugetlb.h b/include/linux/hugetlb.h
index 1faebe1cd0ed..0445faaa636e 100644
--- a/include/linux/hugetlb.h
+++ b/include/linux/hugetlb.h
@@ -613,8 +613,8 @@ struct hstate {
#endif
#ifdef CONFIG_CGROUP_HUGETLB
/* cgroup control files */
- struct cftype cgroup_files_dfl[7];
- struct cftype cgroup_files_legacy[9];
+ struct cftype cgroup_files_dfl[8];
+ struct cftype cgroup_files_legacy[10];
#endif
char name[HSTATE_NAME_LEN];
};
diff --git a/include/linux/hugetlb_cgroup.h b/include/linux/hugetlb_cgroup.h
index c137396129db..0f6cd28558d7 100644
--- a/include/linux/hugetlb_cgroup.h
+++ b/include/linux/hugetlb_cgroup.h
@@ -36,6 +36,11 @@ enum hugetlb_memory_event {
HUGETLB_NR_MEMORY_EVENTS,
};
+struct hugetlb_cgroup_per_node {
+ /* hugetlb usage in pages over all hstates. */
+ unsigned long usage[HUGE_MAX_HSTATE];
+};
+
struct hugetlb_cgroup {
struct cgroup_subsys_state css;
@@ -57,6 +62,8 @@ struct hugetlb_cgroup {
/* Handle for "hugetlb.events.local" */
struct cgroup_file events_local_file[HUGE_MAX_HSTATE];
+
+ struct hugetlb_cgroup_per_node *nodeinfo[];
};
static inline struct hugetlb_cgroup *
diff --git a/mm/hugetlb_cgroup.c b/mm/hugetlb_cgroup.c
index 5383023d0cca..f590e6e14fe6 100644
--- a/mm/hugetlb_cgroup.c
+++ b/mm/hugetlb_cgroup.c
@@ -126,29 +126,58 @@ static void hugetlb_cgroup_init(struct hugetlb_cgroup *h_cgroup,
}
}
+static void hugetlb_cgroup_free(struct hugetlb_cgroup *h_cgroup)
+{
+ int node;
+
+ for_each_node(node)
+ kfree(h_cgroup->nodeinfo[node]);
+ kfree(h_cgroup);
+}
+
static struct cgroup_subsys_state *
hugetlb_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
{
struct hugetlb_cgroup *parent_h_cgroup = hugetlb_cgroup_from_css(parent_css);
struct hugetlb_cgroup *h_cgroup;
+ int node;
+
+ h_cgroup = kzalloc(struct_size(h_cgroup, nodeinfo, nr_node_ids),
+ GFP_KERNEL);
- h_cgroup = kzalloc(sizeof(*h_cgroup), GFP_KERNEL);
if (!h_cgroup)
return ERR_PTR(-ENOMEM);
if (!parent_h_cgroup)
root_h_cgroup = h_cgroup;
+ /*
+ * TODO: this routine can waste much memory for nodes which will
+ * never be onlined. It's better to use memory hotplug callback
+ * function.
+ */
+ for_each_node(node) {
+ /* Set node_to_alloc to -1 for offline nodes. */
+ int node_to_alloc =
+ node_state(node, N_NORMAL_MEMORY) ? node : -1;
+ h_cgroup->nodeinfo[node] =
+ kzalloc_node(sizeof(struct hugetlb_cgroup_per_node),
+ GFP_KERNEL, node_to_alloc);
+ if (!h_cgroup->nodeinfo[node])
+ goto fail_alloc_nodeinfo;
+ }
+
hugetlb_cgroup_init(h_cgroup, parent_h_cgroup);
return &h_cgroup->css;
+
+fail_alloc_nodeinfo:
+ hugetlb_cgroup_free(h_cgroup);
+ return ERR_PTR(-ENOMEM);
}
static void hugetlb_cgroup_css_free(struct cgroup_subsys_state *css)
{
- struct hugetlb_cgroup *h_cgroup;
-
- h_cgroup = hugetlb_cgroup_from_css(css);
- kfree(h_cgroup);
+ hugetlb_cgroup_free(hugetlb_cgroup_from_css(css));
}
/*
@@ -292,7 +321,17 @@ static void __hugetlb_cgroup_commit_charge(int idx, unsigned long nr_pages,
return;
__set_hugetlb_cgroup(page, h_cg, rsvd);
- return;
+ if (!rsvd) {
+ unsigned long usage =
+ h_cg->nodeinfo[page_to_nid(page)]->usage[idx];
+ /*
+ * This write is not atomic due to fetching usage and writing
+ * to it, but that's fine because we call this with
+ * hugetlb_lock held anyway.
+ */
+ WRITE_ONCE(h_cg->nodeinfo[page_to_nid(page)]->usage[idx],
+ usage + nr_pages);
+ }
}
void hugetlb_cgroup_commit_charge(int idx, unsigned long nr_pages,
@@ -331,8 +370,17 @@ static void __hugetlb_cgroup_uncharge_page(int idx, unsigned long nr_pages,
if (rsvd)
css_put(&h_cg->css);
-
- return;
+ else {
+ unsigned long usage =
+ h_cg->nodeinfo[page_to_nid(page)]->usage[idx];
+ /*
+ * This write is not atomic due to fetching usage and writing
+ * to it, but that's fine because we call this with
+ * hugetlb_lock held anyway.
+ */
+ WRITE_ONCE(h_cg->nodeinfo[page_to_nid(page)]->usage[idx],
+ usage - nr_pages);
+ }
}
void hugetlb_cgroup_uncharge_page(int idx, unsigned long nr_pages,
@@ -421,6 +469,59 @@ enum {
RES_RSVD_FAILCNT,
};
+static int hugetlb_cgroup_read_numa_stat(struct seq_file *seq, void *dummy)
+{
+ int nid;
+ struct cftype *cft = seq_cft(seq);
+ int idx = MEMFILE_IDX(cft->private);
+ bool legacy = MEMFILE_ATTR(cft->private);
+ struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(seq_css(seq));
+ struct cgroup_subsys_state *css;
+ unsigned long usage;
+
+ if (legacy) {
+ /* Add up usage across all nodes for the non-hierarchical total. */
+ usage = 0;
+ for_each_node_state(nid, N_MEMORY)
+ usage += READ_ONCE(h_cg->nodeinfo[nid]->usage[idx]);
+ seq_printf(seq, "total=%lu", usage * PAGE_SIZE);
+
+ /* Simply print the per-node usage for the non-hierarchical total. */
+ for_each_node_state(nid, N_MEMORY)
+ seq_printf(seq, " N%d=%lu", nid,
+ READ_ONCE(h_cg->nodeinfo[nid]->usage[idx]) *
+ PAGE_SIZE);
+ seq_putc(seq, '\n');
+ }
+
+ /*
+ * The hierarchical total is pretty much the value recorded by the
+ * counter, so use that.
+ */
+ seq_printf(seq, "%stotal=%lu", legacy ? "hierarichal_" : "",
+ page_counter_read(&h_cg->hugepage[idx]) * PAGE_SIZE);
+
+ /*
+ * For each node, transverse the css tree to obtain the hierarichal
+ * node usage.
+ */
+ for_each_node_state(nid, N_MEMORY) {
+ usage = 0;
+ rcu_read_lock();
+ css_for_each_descendant_pre(css, &h_cg->css) {
+ usage += READ_ONCE(hugetlb_cgroup_from_css(css)
+ ->nodeinfo[nid]
+ ->usage[idx]);
+ }
+ rcu_read_unlock();
+ seq_printf(seq, " N%d=%lu", nid, usage * PAGE_SIZE);
+ }
+
+ seq_putc(seq, '\n');
+
+ return 0;
+}
+
static u64 hugetlb_cgroup_read_u64(struct cgroup_subsys_state *css,
struct cftype *cft)
{
@@ -671,8 +772,14 @@ static void __init __hugetlb_cgroup_file_dfl_init(int idx)
events_local_file[idx]);
cft->flags = CFTYPE_NOT_ON_ROOT;
- /* NULL terminate the last cft */
+ /* Add the numa stat file */
cft = &h->cgroup_files_dfl[6];
+ snprintf(cft->name, MAX_CFTYPE_NAME, "%s.numa_stat", buf);
+ cft->seq_show = hugetlb_cgroup_read_numa_stat;
+ cft->flags = CFTYPE_NOT_ON_ROOT;
+
+ /* NULL terminate the last cft */
+ cft = &h->cgroup_files_dfl[7];
memset(cft, 0, sizeof(*cft));
WARN_ON(cgroup_add_dfl_cftypes(&hugetlb_cgrp_subsys,
@@ -742,8 +849,14 @@ static void __init __hugetlb_cgroup_file_legacy_init(int idx)
cft->write = hugetlb_cgroup_reset;
cft->read_u64 = hugetlb_cgroup_read_u64;
+ /* Add the numa stat file */
+ cft = &h->cgroup_files_dfl[8];
+ snprintf(cft->name, MAX_CFTYPE_NAME, "%s.numa_stat", buf);
+ cft->private = MEMFILE_PRIVATE(idx, 1);
+ cft->seq_show = hugetlb_cgroup_read_numa_stat;
+
/* NULL terminate the last cft */
- cft = &h->cgroup_files_legacy[8];
+ cft = &h->cgroup_files_legacy[9];
memset(cft, 0, sizeof(*cft));
WARN_ON(cgroup_add_legacy_cftypes(&hugetlb_cgrp_subsys,
--
2.34.0.rc2.393.gf8c9666880-goog
On Mon, Nov 22, 2021 at 4:10 PM Mina Almasry <[email protected]> wrote:
>
> For hugetlb backed jobs/VMs it's critical to understand the numa
> information for the memory backing these jobs to deliver optimal
> performance.
>
> Currently this technically can be queried from /proc/self/numa_maps, but
> there are significant issues with that. Namely:
> 1. Memory can be mapped or unmapped.
> 2. numa_maps are per process and need to be aggregated across all
> processes in the cgroup. For shared memory this is more involved as
> the userspace needs to make sure it doesn't double count shared
> mappings.
> 3. I believe querying numa_maps needs to hold the mmap_lock which adds
> to the contention on this lock.
>
> For these reasons I propose simply adding hugetlb.*.numa_stat file,
> which shows the numa information of the cgroup similarly to
> memory.numa_stat.
>
> On cgroup-v2:
> cat /sys/fs/cgroup/unified/test/hugetlb.2MB.numa_stat
> total=2097152 N0=2097152 N1=0
>
> On cgroup-v1:
> cat /sys/fs/cgroup/hugetlb/test/hugetlb.2MB.numa_stat
> total=2097152 N0=2097152 N1=0
> hierarichal_total=2097152 N0=2097152 N1=0
>
> This patch was tested manually by allocating hugetlb memory and querying
> the hugetlb.*.numa_stat file of the cgroup and its parents.
> 
> Cc: Mike Kravetz <[email protected]>
> Cc: Andrew Morton <[email protected]>
> Cc: Shuah Khan <[email protected]>
> Cc: Miaohe Lin <[email protected]>
> Cc: Oscar Salvador <[email protected]>
> Cc: Michal Hocko <[email protected]>
> Cc: Muchun Song <[email protected]>
> Cc: David Rientjes <[email protected]>
> Cc: Shakeel Butt <[email protected]>
> Cc: Jue Wang <[email protected]>
> Cc: Yang Yao <[email protected]>
> Cc: Joanna Li <[email protected]>
> Cc: Cannon Matthews <[email protected]>
> Cc: [email protected]
> Cc: [email protected]
>
> Signed-off-by: Mina Almasry <[email protected]>
>
Reviewed-by: Shakeel Butt <[email protected]>
On Tue, Nov 23, 2021 at 8:10 AM Mina Almasry <[email protected]> wrote:
>
> For hugetlb backed jobs/VMs it's critical to understand the numa
> information for the memory backing these jobs to deliver optimal
> performance.
>
> Currently this technically can be queried from /proc/self/numa_maps, but
> there are significant issues with that. Namely:
> 1. Memory can be mapped or unmapped.
> 2. numa_maps are per process and need to be aggregated across all
> processes in the cgroup. For shared memory this is more involved as
> the userspace needs to make sure it doesn't double count shared
> mappings.
> 3. I believe querying numa_maps needs to hold the mmap_lock which adds
> to the contention on this lock.
>
> For these reasons I propose simply adding hugetlb.*.numa_stat file,
> which shows the numa information of the cgroup similarly to
> memory.numa_stat.
>
> On cgroup-v2:
> cat /sys/fs/cgroup/unified/test/hugetlb.2MB.numa_stat
> total=2097152 N0=2097152 N1=0
>
> On cgroup-v1:
> cat /sys/fs/cgroup/hugetlb/test/hugetlb.2MB.numa_stat
> total=2097152 N0=2097152 N1=0
> hierarichal_total=2097152 N0=2097152 N1=0
>
> This patch was tested manually by allocating hugetlb memory and querying
> the hugetlb.*.numa_stat file of the cgroup and its parents.
> 
> Cc: Mike Kravetz <[email protected]>
> Cc: Andrew Morton <[email protected]>
> Cc: Shuah Khan <[email protected]>
> Cc: Miaohe Lin <[email protected]>
> Cc: Oscar Salvador <[email protected]>
> Cc: Michal Hocko <[email protected]>
> Cc: Muchun Song <[email protected]>
> Cc: David Rientjes <[email protected]>
> Cc: Shakeel Butt <[email protected]>
> Cc: Jue Wang <[email protected]>
> Cc: Yang Yao <[email protected]>
> Cc: Joanna Li <[email protected]>
> Cc: Cannon Matthews <[email protected]>
> Cc: [email protected]
> Cc: [email protected]
>
> Signed-off-by: Mina Almasry <[email protected]>
Reviewed-by: Muchun Song <[email protected]>
On 11/22/21 16:10, Mina Almasry wrote:
> For hugetlb backed jobs/VMs it's critical to understand the numa
> information for the memory backing these jobs to deliver optimal
> performance.
>
> Currently this technically can be queried from /proc/self/numa_maps, but
> there are significant issues with that. Namely:
> 1. Memory can be mapped or unmapped.
> 2. numa_maps are per process and need to be aggregated across all
> processes in the cgroup. For shared memory this is more involved as
> the userspace needs to make sure it doesn't double count shared
> mappings.
> 3. I believe querying numa_maps needs to hold the mmap_lock which adds
> to the contention on this lock.
>
> For these reasons I propose simply adding hugetlb.*.numa_stat file,
> which shows the numa information of the cgroup similarly to
> memory.numa_stat.
>
> On cgroup-v2:
> cat /sys/fs/cgroup/unified/test/hugetlb.2MB.numa_stat
> total=2097152 N0=2097152 N1=0
>
> On cgroup-v1:
> cat /sys/fs/cgroup/hugetlb/test/hugetlb.2MB.numa_stat
> total=2097152 N0=2097152 N1=0
> hierarichal_total=2097152 N0=2097152 N1=0
>
> This patch was tested manually by allocating hugetlb memory and querying
> the hugetlb.*.numa_stat file of the cgroup and its parents.
> 
> Cc: Mike Kravetz <[email protected]>
> Cc: Andrew Morton <[email protected]>
> Cc: Shuah Khan <[email protected]>
> Cc: Miaohe Lin <[email protected]>
> Cc: Oscar Salvador <[email protected]>
> Cc: Michal Hocko <[email protected]>
> Cc: Muchun Song <[email protected]>
> Cc: David Rientjes <[email protected]>
> Cc: Shakeel Butt <[email protected]>
> Cc: Jue Wang <[email protected]>
> Cc: Yang Yao <[email protected]>
> Cc: Joanna Li <[email protected]>
> Cc: Cannon Matthews <[email protected]>
> Cc: [email protected]
> Cc: [email protected]
>
> Signed-off-by: Mina Almasry <[email protected]>
>
Thanks for continuing to update, this version looks good.
Reviewed-by: Mike Kravetz <[email protected]>
--
Mike Kravetz