On Thu, Aug 20, 2020 at 9:13 PM Eric Dumazet [off-list ref] wrote:

On 8/19/20 11:48 PM, Xin Long wrote:

quoted

With disabling bh in the whole sctp_get_port_local(), when
snum == 0 and too many ports have been used, the do-while
loop will take the cpu for a long time and cause cpu stuck:

  [ ] watchdog: BUG: soft lockup - CPU#11 stuck for 22s!
  [ ] RIP: 0010:native_queued_spin_lock_slowpath+0x4de/0x940
  [ ] Call Trace:
  [ ]  _raw_spin_lock+0xc1/0xd0
  [ ]  sctp_get_port_local+0x527/0x650 [sctp]
  [ ]  sctp_do_bind+0x208/0x5e0 [sctp]
  [ ]  sctp_autobind+0x165/0x1e0 [sctp]
  [ ]  sctp_connect_new_asoc+0x355/0x480 [sctp]
  [ ]  __sctp_connect+0x360/0xb10 [sctp]

There's no need to disable bh in the whole function of
sctp_get_port_local. So fix this cpu stuck by removing
local_bh_disable() called at the beginning, and using
spin_lock_bh() instead.

The same thing was actually done for inet_csk_get_port() in
Commit ea8add2b1903 ("tcp/dccp: better use of ephemeral
ports in bind()").

Thanks to Marcelo for pointing the buggy code out.

Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2")
Reported-by: Ying Xu <redacted>
Signed-off-by: Xin Long <lucien.xin@gmail.com>
---

Any reason you chose to not use a cond_resched() then ?

Clearly this function needs to yield, not only BH, but to other threads.

Right, it explains why the cpu stuck is gone, but sctp_get_port_local()
still use the cpu quite a lot:


Thanks, will post v2.