Re: [PATCH bpf-next v6 06/34] bpf: prepare for memcg-based memory accounting for bpf maps

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On 11/18/20 2:28 AM, Roman Gushchin wrote:
On Tue, Nov 17, 2020 at 05:11:00PM -0800, Alexei Starovoitov wrote:
On Tue, Nov 17, 2020 at 5:07 PM Roman Gushchin <guro@xxxxxx> wrote:
On Tue, Nov 17, 2020 at 04:46:34PM -0800, Roman Gushchin wrote:
On Wed, Nov 18, 2020 at 01:06:17AM +0100, Daniel Borkmann wrote:
On 11/17/20 4:40 AM, Roman Gushchin wrote:
In the absolute majority of cases if a process is making a kernel
allocation, it's memory cgroup is getting charged.

Bpf maps can be updated from an interrupt context and in such
case there is no process which can be charged. It makes the memory
accounting of bpf maps non-trivial.

Fortunately, after commit 4127c6504f25 ("mm: kmem: enable kernel
memcg accounting from interrupt contexts") and b87d8cefe43c
("mm, memcg: rework remote charging API to support nesting")
it's finally possible.

To do it, a pointer to the memory cgroup of the process which created
the map is saved, and this cgroup is getting charged for all
allocations made from an interrupt context.

Allocations made from a process context will be accounted in a usual way.

Signed-off-by: Roman Gushchin <guro@xxxxxx>
Acked-by: Song Liu <songliubraving@xxxxxx>
[...]
+#ifdef CONFIG_MEMCG_KMEM
+static __always_inline int __bpf_map_update_elem(struct bpf_map *map, void *key,
+                                          void *value, u64 flags)
+{
+ struct mem_cgroup *old_memcg;
+ bool in_interrupt;
+ int ret;
+
+ /*
+  * If update from an interrupt context results in a memory allocation,
+  * the memory cgroup to charge can't be determined from the context
+  * of the current task. Instead, we charge the memory cgroup, which
+  * contained a process created the map.
+  */
+ in_interrupt = in_interrupt();
+ if (in_interrupt)
+         old_memcg = set_active_memcg(map->memcg);
+
+ ret = map->ops->map_update_elem(map, key, value, flags);
+
+ if (in_interrupt)
+         set_active_memcg(old_memcg);
+
+ return ret;

Hmm, this approach here won't work, see also commit 09772d92cd5a ("bpf: avoid
retpoline for lookup/update/delete calls on maps") which removes the indirect
call, so the __bpf_map_update_elem() and therefore the set_active_memcg() is
not invoked for the vast majority of cases.

I see. Well, the first option is to move these calls into map-specific update
functions, but the list is relatively long:
   nsim_map_update_elem()
   cgroup_storage_update_elem()
   htab_map_update_elem()
   htab_percpu_map_update_elem()
   dev_map_update_elem()
   dev_map_hash_update_elem()
   trie_update_elem()
   cpu_map_update_elem()
   bpf_pid_task_storage_update_elem()
   bpf_fd_inode_storage_update_elem()
   bpf_fd_sk_storage_update_elem()
   sock_map_update_elem()
   xsk_map_update_elem()

Alternatively, we can set the active memcg for the whole duration of bpf
execution. It's simpler, but will add some overhead. Maybe we can somehow
mark programs calling into update helpers and skip all others?

Actually, this is problematic if a program updates several maps, because
in theory they can belong to different cgroups.
So it seems that the first option is the way to go. Do you agree?

May be instead of kmalloc_node() that is used by most of the map updates
introduce bpf_map_kmalloc_node() that takes a map pointer as an argument?
And do set_memcg inside?

I suspect it's not only kmalloc_node(), but if there will be 2-3 allocation
helpers, it sounds like a good idea to me! I'll try and get back with v7 soon.

Could this be baked into kmalloc*() API itself given we also need to pass in
__GFP_ACCOUNT everywhere, so we'd have a new API with additional argument where
we pass the memcg pointer to tell it directly where to account it for instead of
having to have the extra set_active_memcg() set/restore dance via BPF wrapper?
It seems there would be not much specifics on BPF itself and if it's more generic
it could also be used by other subsystems.

Thanks,
Daniel



[Index of Archives]     [Linux Samsung SoC]     [Linux Rockchip SoC]     [Linux Actions SoC]     [Linux for Synopsys ARC Processors]     [Linux NFS]     [Linux NILFS]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]


  Powered by Linux