On 2/9/26 21:17, Martin KaFai Lau wrote:
> On 2/8/26 9:14 AM, Michal Luczaj wrote:
>> On 2/7/26 23:00, Kuniyuki Iwashima wrote:
>>> On Sat, Feb 7, 2026 at 6:35 AM Michal Luczaj <[email protected]> wrote:
>>>> This patch also happens to fix a deadlock that may occur when
>>>> bpf_iter_unix_seq_show()'s lock_sock_fast() takes the fast path and the
>>>> iter prog attempts to update a sockmap. Which ends up spinning at
>>>> sock_map_update_elem()'s bh_lock_sock():
>>>
>>> Hmm.. this seems to be a more general problem for
>>> bpf iter vs sockmap. bpf_iter_{tcp,udp}_seq_show() also
>>> hold lock_sock(), where this patch's solution does not help.
>>> We need to resolve this regardless of socket family.
>>
>> I don't see any deadlocks there. Note that I've mentioned lock_sock_fast()
>> fast path was a problem, not lock_sock().
>
> For the tcp/udp, I think the bpf_iter should be fine: lock_sock() in
> seq_show and bh_lock_sock() in map_update. It seems redundant though.
I wasn't sure what exactly you suspect of being redundant, so I did some
digging:
lock_sock() in tcp/udp iter is expected (among others?) by kfunc
bpf_sock_destroy(). Relevant commit 4ddbcb886268 ("bpf: Add
bpf_sock_destroy kfunc"),
https://lore.kernel.org/all/[email protected]/
In short: lock must be taken for synchronization of proto::diag_destroy().
Reasons for bh_lock_sock() during bpf sockmap update are explained in
commit 0126240f448d ("bpf: sockmap: Allow update from BPF"),
https://lore.kernel.org/bpf/[email protected]/
In short: socket shouldn't be allowed to change its state during the
update. BH lock because bpf can't sleep.
> From looking at may_update_sockmap(), other bpf progs (e.g., tc) can do
> map_update also. On those paths, I am not sure why
> sock_map_update_elem() does not need to check "!sock_owned_by_user(sk)".
> If it is indeed an issue, it probably needs to be addressed separately.
Since sockmap update can happen in a tracing prog, can you really expect a
socket to be always owned?
> It should also be helpful to be consistent with tcp/udp iter and use
> lock_sock() instead of lock_sock_fast() in bpf_iter_unix_seq_show().
OK, I'll tweak that in v3.
