Re: [PATCH 3/3] xen: optimize xenbus driver for multiple concurrent xenstore accesses

[Juergen Gross <jgross@suse.com>]

On 11/01/17 16:29, Boris Ostrovsky wrote:
> 
>>>> +
>>>> +
>>>> +static bool test_reply(struct xb_req_data *req)
>>>> +{
>>>> +	if (req->state == xb_req_state_got_reply || !xenbus_ok())
>>>> +		return true;
>>>> +
>>>> +	/* Make sure to reread req->state each time. */
>>>> +	cpu_relax();
>>> I don't think I understand why this is needed.
>> I need a compiler barrier. Otherwise the compiler read req->state only
>> once outside the while loop.
> 
> 
> Then barrier() looks the right primitive to use here. cpu_relax(), while
> doing what you want, is intended for other purposes.

Hmm, yes, this sounds better.

>>
>>>> +
>>>> +	return false;
>>>> +}
>>>> +
>>>
>>>> +static void xs_send(struct xb_req_data *req, struct xsd_sockmsg *msg)
>>>>  {
>>>> -	mutex_lock(&xs_state.transaction_mutex);
>>>> -	atomic_inc(&xs_state.transaction_count);
>>>> -	mutex_unlock(&xs_state.transaction_mutex);
>>>> -}
>>>> +	bool notify;
>>>>  
>>>> -static void transaction_end(void)
>>>> -{
>>>> -	if (atomic_dec_and_test(&xs_state.transaction_count))
>>>> -		wake_up(&xs_state.transaction_wq);
>>>> -}
>>>> +	req->msg = *msg;
>>>> +	req->err = 0;
>>>> +	req->state = xb_req_state_queued;
>>>> +	init_waitqueue_head(&req->wq);
>>>>  
>>>> -static void transaction_suspend(void)
>>>> -{
>>>> -	mutex_lock(&xs_state.transaction_mutex);
>>>> -	wait_event(xs_state.transaction_wq,
>>>> -		   atomic_read(&xs_state.transaction_count) == 0);
>>>> -}
>>>> +	xs_request_enter(req);
>>>>  
>>>> -static void transaction_resume(void)
>>>> -{
>>>> -	mutex_unlock(&xs_state.transaction_mutex);
>>>> +	req->msg.req_id = xs_request_id++;
>>> Is it safe to do this without a lock?
>> You are right: I should move this to xs_request_enter() inside the
>> lock. I think I'll let return xs_request_enter() the request id.
> 
> 
> Then please move xs_request_id's declaration close to xs_state_lock's
> declaration (just like you are going to move the two other state variables)

Already done. :-)

>>
>>>> +static int xs_reboot_notify(struct notifier_block *nb,
>>>> +			    unsigned long code, void *unused)
>>>>  {
>>>> -	struct xs_stored_msg *msg;
>>>
>>>
>>>> +	struct xb_req_data *req;
>>>> +
>>>> +	mutex_lock(&xb_write_mutex);
>>>> +	list_for_each_entry(req, &xs_reply_list, list)
>>>> +		wake_up(&req->wq);
>>>> +	list_for_each_entry(req, &xb_write_list, list)
>>>> +		wake_up(&req->wq);
>>> We are waking up waiters here but there is not guarantee that waiting
>>> threads will have a chance to run, is there?
>> You are right. But this isn't the point. We want to avoid blocking a
>> reboot due to some needed thread waiting for xenstore. And this task
>> is being accomplished here.
> 
> 
> I think it's worth adding a comment mentioning this.

Okay.


Juergen