Re: [PATCH net v1] tls: fix hung task in tx_work_handler by using non-blocking sends

[Jakub Kicinski <kuba@kernel.org>]

On Sun, 01 Mar 2026 06:52:00 +0000 Jiayuan Chen wrote:
> > On Fri, 27 Feb 2026 14:32:31 +0800 Jiayuan Chen wrote:
> > > tx_work_handler calls tls_tx_records with flags=-1, which preserves
> > >  each record's original tx_flags but results in tcp_sendmsg_locked
> > >  using an infinite send timeout. When the peer is unresponsive and the
> > >  send buffer is full, tcp_sendmsg_locked blocks indefinitely in
> > >  sk_stream_wait_memory. This causes tls_sk_proto_close to hang in
> > >  cancel_delayed_work_sync waiting for tx_work_handler to finish,
> > >  leading to a hung task:
> > >  
> > >  INFO: task ...: blocked for more than ... seconds.
> > >  Call Trace:
> > >  cancel_delayed_work_sync
> > >  tls_sw_cancel_work_tx
> > >  tls_sk_proto_close
> > >  
> > >  A workqueue handler should never block indefinitely. Fix this by
> > >  introducing __tls_tx_records() with an extra_flags parameter that
> > >  gets OR'd into each record's tx_flags. tx_work_handler uses this to
> > >  pass MSG_DONTWAIT so tcp_sendmsg_locked returns -EAGAIN immediately
> > >  when the send buffer is full, without overwriting the original
> > >  per-record flags (MSG_MORE, MSG_NOSIGNAL, etc.). On -EAGAIN, the
> > >  existing reschedule mechanism retries after a short delay.
> > >  
> > >  Also consolidate the two identical reschedule paths (lock contention
> > >  and -EAGAIN) into one.
> > >   
> > It's not that simple. The default semantics for TCP sockets is that
> > queuing data and then calling close() is a legitimate thing to do
> > and the data should be sent cleanly, followed by a normal FIN in such
> > case.
> > 
> > Maybe we should explore trying to make sure we have enough wmem before
> > we start creating records. Get rid of the entire workqueue mess?  
> 
> Regarding wmem pre-check: the async crypto path is not triggered by
> wmem shortage — it's triggered when the crypto operation itself is
> asynchronous (e.g. cryptd fallback when SIMD is unavailable). At the
> time tls_do_encryption() returns -EINPROGRESS, wmem may be perfectly
> fine. The problem occurs later when tls_encrypt_done() fires and
> tx_work_handler tries to push the completed records — by that point
> the send buffer may have filled up. Since these are two different
> points in time, pre-checking wmem at record creation wouldn't help.

My recollection is that the work scheduling in the async encrypt path
is just a duct-tape fix for some old race. The sendmsg() paths should
normally wait for the async crypto to finish before returning to user
space.

> > Regarding your patch I think all callers passing -1 as flags are on 
> > the close path, you could have just added | DONTWAIT if the flags 
> > are -1.  
> 
> Regarding adding MSG_DONTWAIT unconditionally when flags == -1:
> tls_sw_release_resources_tx() also calls tls_tx_records(sk, -1).
> That's in the close path where we actually want to block and flush
> remaining records to honour the "close() should send data cleanly"
> semantics you mentioned. Making that non-blocking would cause data
> loss. So we do need to distinguish between the two callers, which
> is why I introduced __tls_tx_records() with the extra_flags parameter.

Possible, I didn't look very closely.

The extra_flags argument you're adding is extremely inelegant.