aboutsummaryrefslogtreecommitdiff
path: root/Documentation/usb/dwc3.rst
blob: f94a7ba16573f6f0bcb2026d2f12d50f6b40adae (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
===========
DWC3 driver
===========


TODO
~~~~

Please pick something while reading :)

- Convert interrupt handler to per-ep-thread-irq

  As it turns out some DWC3-commands ~1ms to complete. Currently we spin
  until the command completes which is bad.

  Implementation idea:

  - dwc core implements a demultiplexing irq chip for interrupts per
    endpoint. The interrupt numbers are allocated during probe and belong
    to the device. If MSI provides per-endpoint interrupt this dummy
    interrupt chip can be replaced with "real" interrupts.
  - interrupts are requested / allocated on usb_ep_enable() and removed on
    usb_ep_disable(). Worst case are 32 interrupts, the lower limit is two
    for ep0/1.
  - dwc3_send_gadget_ep_cmd() will sleep in wait_for_completion_timeout()
    until the command completes.
  - the interrupt handler is split into the following pieces:

    - primary handler of the device
      goes through every event and calls generic_handle_irq() for event
      it. On return from generic_handle_irq() in acknowledges the event
      counter so interrupt goes away (eventually).

    - threaded handler of the device
      none

    - primary handler of the EP-interrupt
      reads the event and tries to process it. Everything that requires
      sleeping is handed over to the Thread. The event is saved in an
      per-endpoint data-structure.
      We probably have to pay attention not to process events once we
      handed something to thread so we don't process event X prio Y
      where X > Y.

    - threaded handler of the EP-interrupt
      handles the remaining EP work which might sleep such as waiting
      for command completion.

  Latency:

   There should be no increase in latency since the interrupt-thread has a
   high priority and will be run before an average task in user land
   (except the user changed priorities).