Commit b910eaaa ("bpf: Fix NULL pointer dereference in bpf_get_local_storage()
helper") fixed a bug for bpf_get_local_storage() helper so different tasks
won't mess up with each other's percpu local storage.

The percpu data contains 8 slots so it can hold up to 8 contexts (same or
different tasks), for 8 different program runs, at the same time. This in
general is sufficient. But our internal testing showed the following warning
multiple times:

  [...]
  warning: WARNING: CPU: 13 PID: 41661 at include/linux/bpf-cgroup.h:193
     __cgroup_bpf_run_filter_sock_ops+0x13e/0x180
  RIP: 0010:__cgroup_bpf_run_filter_sock_ops+0x13e/0x180
  <IRQ>
   tcp_call_bpf.constprop.99+0x93/0xc0
   tcp_conn_request+0x41e/0xa50
   ? tcp_rcv_state_process+0x203/0xe00
   tcp_rcv_state_process+0x203/0xe00
   ? sk_filter_trim_cap+0xbc/0x210
   ? tcp_v6_inbound_md5_hash.constprop.41+0x44/0x160
   tcp_v6_do_rcv+0x181/0x3e0
   tcp_v6_rcv+0xc65/0xcb0
   ip6_protocol_deliver_rcu+0xbd/0x450
   ip6_input_finish+0x11/0x20
   ip6_input+0xb5/0xc0
   ip6_sublist_rcv_finish+0x37/0x50
   ip6_sublist_rcv+0x1dc/0x270
   ipv6_list_rcv+0x113/0x140
   __netif_receive_skb_list_core+0x1a0/0x210
   netif_receive_skb_list_internal+0x186/0x2a0
   gro_normal_list.part.170+0x19/0x40
   napi_complete_done+0x65/0x150
   mlx5e_napi_poll+0x1ae/0x680
   __napi_poll+0x25/0x120
   net_rx_action+0x11e/0x280
   __do_softirq+0xbb/0x271
   irq_exit_rcu+0x97/0xa0
   common_interrupt+0x7f/0xa0
   </IRQ>
   asm_common_interrupt+0x1e/0x40
  RIP: 0010:bpf_prog_1835a9241238291a_tw_egress+0x5/0xbac
   ? __cgroup_bpf_run_filter_skb+0x378/0x4e0
   ? do_softirq+0x34/0x70
   ? ip6_finish_output2+0x266/0x590
   ? ip6_finish_output+0x66/0xa0
   ? ip6_output+0x6c/0x130
   ? ip6_xmit+0x279/0x550
   ? ip6_dst_check+0x61/0xd0
  [...]

Using drgn [0] to dump the percpu buffer contents showed that on this CPU
slot 0 is still available, but slots 1-7 are occupied and those tasks in
slots 1-7 mostly don't exist any more. So we might have issues in
bpf_cgroup_storage_unset().

Further debugging confirmed that there is a bug in bpf_cgroup_storage_unset().
Currently, it tries to unset "current" slot with searching from the start.
So the following sequence is possible:

  1. A task is running and claims slot 0
  2. Running BPF program is done, and it checked slot 0 has the "task"
     and ready to reset it to NULL (not yet).
  3. An interrupt happens, another BPF program runs and it claims slot 1
     with the *same* task.
  4. The unset() in interrupt context releases slot 0 since it matches "task".
  5. Interrupt is done, the task in process context reset slot 0.

At the end, slot 1 is not reset and the same process can continue to occupy
slots 2-7 and finally, when the above step 1-5 is repeated again, step 3 BPF
program won't be able to claim an empty slot and a warning will be issued.

To fix the issue, for unset() function, we should traverse from the last slot
to the first. This way, the above issue can be avoided.

The same reverse traversal should also be done in bpf_get_local_storage() helper
itself. Otherwise, incorrect local storage may be returned to BPF program.

  [0] https://github.com/osandov/drgn

Fixes: b910eaaa ("bpf: Fix NULL pointer dereference in bpf_get_local_storage() helper")
Signed-off-by: Yonghong Song <yhs@fb.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Andrii Nakryiko <andrii@kernel.org>
Link: https://lore.kernel.org/bpf/20210810010413.1976277-1-yhs@fb.com