Re: [PATCH v2 11/25] usb: gadget: f_tcm: Execute command on write completion

[Thinh Nguyen]

On Mon, Aug 29, 2022, Sebastian Andrzej Siewior wrote:
> On 2022-08-26 18:37:36 [+0000], Thinh Nguyen wrote:
> > On Fri, Aug 26, 2022, Sebastian Andrzej Siewior wrote:
> > > On 2022-07-18 18:27:12 [-0700], Thinh Nguyen wrote:
> > > > index 6fea80afe2d7..ec83f2f9a858 100644
> > > > --- a/drivers/usb/gadget/function/f_tcm.c
> > > > +++ b/drivers/usb/gadget/function/f_tcm.c
> > > > @@ -955,7 +949,7 @@ static void usbg_data_write_cmpl(struct usb_ep *ep, struct usb_request *req)
> > > >  				se_cmd->data_length);
> > > >  	}
> > > >  
> > > > -	complete(&cmd->write_complete);
> > > > +	target_execute_cmd(se_cmd);
> > > 
> > > usbg_data_write_cmpl() is invoked from interrupt service routing which
> > > may run with disabled interrupts. From looking at target_execute_cmd():
> > 
> > It will always be called with interrupts disabled as documented in
> > usb_request API.
> > 
> > > | void target_execute_cmd(struct se_cmd *cmd)
> > > | {
> > > …
> > > |         spin_lock_irq(&cmd->t_state_lock);
> > > …
> > > |         spin_unlock_irq(&cmd->t_state_lock);
> > > …
> > > | }
> > > 
> > > which means interrupts will remain open after leaving
> > > target_execute_cmd(). Now, why didn't the WARN_ONCE() in
> > > __handle_irq_event_percpu() trigger? Am I missing something?
> > > 
> > > >  	return;
> > > 
> > 
> > Since target_execute_cmd() is called in usbg_data_write_cmpl(),
> > interrupts are still disabled.
> 
> but you do realize that target_execute_cmd() will leave with enabled
> interrupts and this is not desired? I _think_ this was the reason why I
> ended up with the wait+complete construct instead of invoking this
> function directly.
> An _irqsave() in target_execute_cmd() would probably be all you need
> here.
> 

Ok. Maybe we should make a change in the target_execute_cmd() then. It
seems unreasonable to force the caller to workaround this such as the
wait+complete construct you did (and I don't recall we have changes in
place to know/guarantee that interrupts are enabled before executing
target_execute_cmd() previously either).

For the dwc3, we masked the interrupt at this point, so interrupt won't
be asserted here.

Thanks,
Thinh