RE: [PATCH net-next 1/7] r8152: adjust rx_bottom

[Hayes Wang <hayeswang@realtek.com>]

 David Miller [mailto:davem@davemloft.net] 
> Sent: Sunday, January 25, 2015 2:44 PM
[...]
> What keeps rtl_start_rx() from running in parallel with
> r8152_submit_rx(), or any other accessor of the RX agg->list?

Forgive my poor English. I would try to describe them clearly.
The steps about the rx agg->list would be
1. carrier on or autoresume occurs.
2. Call rtl_start_rx().
3. Rx agg->list flows between device and tp->rx_done.
4. carrier off or autosuspend occurs.
5. call rtl_stop_rx().

The rtl_start_rx() would only be called when the linking
status is changed from off to on or the auto resume occurs.
And rtl_start_rx() would reinitialize the tp->rx_done and
all of the rx agg->list. After step 2, the rx agg->list
would flow between the usb host controller and the driver.
If r8152_submit_rx() is success, the driver wouldn't own the
rx agg->list until it is returned from the usb host controller.
If r8152_submit_rx() is fail, the driver would still own the
rx agg->list, and queue it to the tp->rx_done with spin lock
for next try.

If the status stays in step 3, only the rx_bottom() would submit
the rx agg. The rtl_start_rx() wouldn't be called suddenly,
unless the linking down or auto suspend occur first and linking
on or auto resume occur again. If linking down or auto suspend
occur, rtl_stop_rx() would be called (step 5). After this step,
rx_bottom() wouldn't submit rx, and all rx agg->list would stop
flowing. That is, the tp->rx_done and all rx agg->list wouldn't
be changed until the next rtl_start_rx() is called.

Therefore, the flow for each rx agg->list would be
a. submittd by rtl_start_rx().
b. goto step c if success, otherwise goto step d.
c. completed by usb host controller.
d. queued to tp->rx_done with spin lock.
e. dequeue from tp->rx_done with spin lock by rx_botoom().
f. goto step i if link down, otherwise goto step g.
g. submitted by rx_botoom().
h. goto step b.
i. goto step a if link on.

And the patch change the step g to g1.
g1. submitted by rx_botoom() if (!ret), otherwise goto step d.

Best Regards,
Hayes
> 
> You also keep using different terminology from me when
> discussing what lists do or do not need protection, and that
> is going to make it difficult for anyone to follow our
> conversation at all.
> 
> We're talking specifically about RX agg->list objects and
> whether access to them need synchronization or not.