bpf: Try harder when allocating memory for large maps

It has been observed that sometimes a higher order memory allocation for BPF maps fails when there is no obvious memory pressure in a system.

E.g. the map (BPF_MAP_TYPE_LRU_HASH, key=38, value=56, max_elems=524288) could not be created due to vmalloc unable to allocate 75497472B, when the system's memory consumption (in MB) was the following:

    Total: 3942 Used: 837 (21.24%) Free: 138 Buffers: 239 Cached: 2727

Later analysis [1] by Michal Hocko showed that the vmalloc was not trying to reclaim memory from the page cache and was failing prematurely due to __GFP_NORETRY.

Considering dcda9b0471 ("mm, tree wide: replace __GFP_REPEAT by __GFP_RETRY_MAYFAIL with more useful semantic") and [1], we can replace __GFP_NORETRY with __GFP_RETRY_MAYFAIL, as it won't invoke OOM killer and will try harder to fulfil allocation requests.

Unfortunately, replacing the body of the BPF map memory allocation function with the kvmalloc_node helper function is not an option at this point in time, given 1) kmalloc is non-optional for higher order allocations, and 2) passing __GFP_RETRY_MAYFAIL to the kmalloc would stress the slab allocator too much for large requests.

The change has been tested with the workloads mentioned above and by observing oom_kill value from /proc/vmstat.

[1]: https://lore.kernel.org/bpf/[email protected]/

Acked-by Yonghong Song <[email protected]>

Signed-off-by Martynas Pumputis <[email protected]>

---
 kernel/bpf/syscall.c | 22 +++++++++++++++-------
 1 file changed, 15 insertions(+), 7 deletions(-)

kernel/bpf/syscall.c +15 -7

index 62f6bced3a3c..afca36f53c49 100644
@@ -136,21 +136,29 @@ static struct bpf_map *find_and_alloc_map(union bpf_attr *attr)
 
 void *bpf_map_area_alloc(size_t size, int numa_node)
 {
-	/* We definitely need __GFP_NORETRY, so OOM killer doesn't
-	 * trigger under memory pressure as we really just want to
-	 * fail instead.
+	/* We really just want to fail instead of triggering OOM killer
+	 * under memory pressure, therefore we set __GFP_NORETRY to kmalloc,
+	 * which is used for lower order allocation requests.
+	 *
+	 * It has been observed that higher order allocation requests done by
+	 * vmalloc with __GFP_NORETRY being set might fail due to not trying
+	 * to reclaim memory from the page cache, thus we set
+	 * __GFP_RETRY_MAYFAIL to avoid such situations.
 	 */
-	const gfp_t flags = __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO;
+
+	const gfp_t flags = __GFP_NOWARN | __GFP_ZERO;
 	void *area;
 
 	if (size <= (PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER)) {
-		area = kmalloc_node(size, GFP_USER | flags, numa_node);
+		area = kmalloc_node(size, GFP_USER | __GFP_NORETRY | flags,
+				    numa_node);
 		if (area != NULL)
 			return area;
 	}
 
-	return __vmalloc_node_flags_caller(size, numa_node, GFP_KERNEL | flags,
-					   __builtin_return_address(0));
+	return __vmalloc_node_flags_caller(size, numa_node,
+					   GFP_KERNEL | __GFP_RETRY_MAYFAIL |
+					   flags, __builtin_return_address(0));
 }
 
 void bpf_map_area_free(void *area)

Signature · 1 line

2.21.0

Quoted 18 lines · Martynas Pumputis

It has been observed that sometimes a higher order memory allocation for BPF maps fails when there is no obvious memory pressure in a system.

E.g. the map (BPF_MAP_TYPE_LRU_HASH, key=38, value=56, max_elems=524288) could not be created due to vmalloc unable to allocate 75497472B, when the system's memory consumption (in MB) was the following:

    Total: 3942 Used: 837 (21.24%) Free: 138 Buffers: 239 Cached: 2727

Later analysis [1] by Michal Hocko showed that the vmalloc was not trying to reclaim memory from the page cache and was failing prematurely due to __GFP_NORETRY.

Considering dcda9b0471 ("mm, tree wide: replace __GFP_REPEAT by __GFP_RETRY_MAYFAIL with more useful semantic") and [1], we can replace __GFP_NORETRY with __GFP_RETRY_MAYFAIL, as it won't invoke OOM killer and will try harder to fulfil allocation requests.

Quoted 5 lines

Unfortunately, replacing the body of the BPF map memory allocation function with the kvmalloc_node helper function is not an option at this point in time, given 1) kmalloc is non-optional for higher order allocations, and 2) passing __GFP_RETRY_MAYFAIL to the kmalloc would stress the slab allocator too much for large requests.

Thanks for extending the changelog!

Quoted 5 lines

The change has been tested with the workloads mentioned above and by observing oom_kill value from /proc/vmstat.

[1]: https://lore.kernel.org/bpf/[email protected]/

Quoted 2 lines

Acked-by Yonghong Song <[email protected]>

Signed-off-by Martynas Pumputis <[email protected]>

The patch looks good to me from the allocator usage POV. I wish there was a good way to give you a util function to use rather than opencoding but this is the only place with this semantic I have seen and I am not sure it is generic enough. Let's see what the future has to tell us.