Re: [PATCH v3 bpf-next/net 3/5] bpf: Introduce SK_BPF_MEMCG_FLAGS and SK_BPF_MEMCG_SOCK_ISOLATED.

On Wed, Aug 27, 2025 at 4:48 PM Martin KaFai Lau [off-list ref] wrote:

On 8/26/25 11:38 AM, Kuniyuki Iwashima wrote:

quoted

The main targets are BPF_CGROUP_INET_SOCK_CREATE and
BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB as demonstrated in the selftest.

Note that we do not support modifying the flag once sk->sk_memcg is set
especially because UDP charges memory under sk->sk_receive_queue.lock
instead of lock_sock().

[ ... ]

quoted

diff --git a/include/net/sock.h b/include/net/sock.h
index 63a6a48afb48..d41a2f8f8b30 100644
--- a/include/net/sock.h
+++ b/include/net/sock.h

@@ -2596,10 +2596,39 @@ static inline gfp_t gfp_memcg_charge(void)
      return in_softirq() ? GFP_ATOMIC : GFP_KERNEL;
  }

+#define SK_BPF_MEMCG_FLAG_MASK       (SK_BPF_MEMCG_FLAG_MAX - 1)
+#define SK_BPF_MEMCG_PTR_MASK        ~SK_BPF_MEMCG_FLAG_MASK
+
  #ifdef CONFIG_MEMCG
+static inline void mem_cgroup_sk_set_flags(struct sock *sk, unsigned short flags)
+{
+     unsigned long val = (unsigned long)sk->sk_memcg;
+
+     val |= flags;
+     sk->sk_memcg = (struct mem_cgroup *)val;
+}
+

[ ... ]

quoted

+static int sk_bpf_set_get_memcg_flags(struct sock *sk, int *optval, bool getopt)
+{
+     if (!sk_has_account(sk))
+             return -EOPNOTSUPP;
+
+     if (getopt) {
+             *optval = mem_cgroup_sk_get_flags(sk);
+             return 0;
+     }
+
+     if (sock_owned_by_user_nocheck(sk) && mem_cgroup_from_sk(sk))

I still don't see how this can stop some of the bh bpf prog to set/clear the bit
after mem_cgroup_sk_alloc() is done. For example, in BPF_SOCK_OPS_RETRANS_CB,
bh_lock_sock() is held but not owned by user, so the bpf prog can continue to
change the SK_BPF_MEMCG_FLAGS. What am I missing?

You are totally right.  I think I was just confused.

Passing some more args to __bpf_setsockopt() for caller context purpose is quite
ugly which I think you also mentioned/tried earlier.

Another option I was thinking (but didn't try) was extending void *ctx
from struct sock * to like below and silently cast ctx back to the original
one in the dedicated setsockopt proto.

struct {
    struct sock * sk;
    /* caller info here */
}

 May be it can check "if
(in_softirq() && mem_cgroup_from_sk(sk)) return -EBUSY" but looks quite tricky
together with the sock_owned_by_user_nocheck().

It seems this sk_memcg bit change is only needed for sock_ops hook and create
hook? sock_ops already has its only func_proto bpf_sock_ops_setsockopt().
bpf_sock_ops_setsockopt can directly call sk_bpf_set_get_memcg_flags() but limit
it to the BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB. Is it also safe to allow it for
BPF_SOCK_OPS_TCP_LISTEN_CB? For the create hook, create a new func proto which
can directly call sk_bpf_set_get_memcg_flags(). wdyt?

I prefer the new proto instead of a hacky/complicated approach.

quoted

+             return -EBUSY;
+
+     if (*optval <= 0 || *optval >= SK_BPF_MEMCG_FLAG_MAX)

It seems the bit cannot be cleared (the *optval <= 0 check). There could be
multiple bpf progs running in a cgroup which want to clear this bit. e.g. the
parent cgroup's prog may want to ensure this bit is not set.

I didn't think about the situation, so the set helper will need to
reset all flags (in case there could be multiple bit flags in the future)

static inline void mem_cgroup_sk_set_flags(struct sock *sk, unsigned short flags
{
    unsigned long val = (unsigned long)sk->sk_memcg;

    val &= SK_BPF_MEMCG_PTR_MASK;
    val |= flags;

    sk->sk_memcg = (struct mem_cgroup *)val;
}