raspberrypi/linux

Issuing low-speed packets when the root port is in full-speed mode
causes the root port to stop responding. Explicitly fail when
enqueuing URBs to a LS endpoint on a FS bus.

dwc_otg: Enable NAK holdoff for control split transactions

Certain low-speed devices take a very long time to complete a
data or status stage of a control transaction, producing NAK
responses until they complete internal processing - the USB2.0
spec limit is up to 500mS. This causes the same type of interrupt
storm as seen with USB-serial dongles prior to c8edb23.

In certain circumstances, usually while booting, this interrupt
storm could cause SD card timeouts.

…2, DLM

See: #321

See: #349

Replace magic number with its #define name

and enable CONFIG_JUMP_LABEL
See: #348

dwc_otg: root port and FIQ bugfixes

Fix several issues resulting in crashes or inconsistent state
if a Model A root port was disconnected.

- Clean up queue heads properly in kill_urbs_in_qh_list by
  removing the empty QHs from the schedule lists
- Set the halt status properly to prevent IRQ handlers from
  using freed memory
- Add fiq_split related cleanup for saved registers
- Make microframe scheduling reclaim host channels if
  active during a disconnect
- Abort URBs with -ESHUTDOWN status response, informing
  device drivers so they respond in a more correct fashion
  and don't try to resubmit URBs
- Prevent IRQ handlers from attempting to handle channel
  interrupts if the associated URB was dequeued (and the
  driver state was cleared)

A dwc_otg_urb would get leaked if the HCD enqueue function
failed for any reason. Free the URB at the appropriate points.

Failing to set up the FIQ correctly would result in
"IRQ 32: nobody cared" errors in dmesg.

dwc_otg: fix potential sleep while atomic during urb enqueue

Fixes a regression introduced with eb1b482. Kmalloc called from
dwc_otg_hcd_qtd_add / dwc_otg_hcd_qtd_create did not always have
the GPF_ATOMIC flag set. Force this flag when inside the larger
critical section.

The lack of memory barriers could (very rarely) result in
vchiu_queue_pop reading the next value before it had been written
(getting either NULL, or a value that had been popped once already).

The userdata for VCHIQ services created through the ioctl API is
a kmalloced structure. These objects were getting leaked, most
notably in vchiq_release(), where the service could be closed,
freed and removed from the service list before the wait-to-die
loop was entered.

This change adds a userdata termination callback, and implements
it in the case where USER_SERVICE_T is used for the service
userdata.

dwc_otg: fiq_split: use TTs with more granularity

This fixes certain issues with split transaction scheduling.

- Isochronous multi-packet OUT transactions now hog the TT until
  they are completed - this prevents hubs aborting transactions
  if they get a periodic start-split out-of-order
- Don't perform TT allocation on non-periodic endpoints - this
  allows simultaneous use of the TT's bulk/control and periodic
  transaction buffers

This commit will mainly affect USB audio playback.

dwc_otg: make channel halts with unknown state less damaging

If the IRQ received a channel halt interrupt through the FIQ
with no other bits set, the IRQ would not release the host
channel and never complete the URB.

Add catchall handling to treat as a transaction error and retry.

dwc_otg: prevent BUG() in TT allocation if hub address is > 16

A fixed-size array is used to track TT allocation. This was
previously set to 16 which caused a crash because
dwc_otg_hcd_allocate_port would read past the end of the array.

This was hit if a hub was plugged in which enumerated as addr > 16,
due to previous device resets or unplugs.

Also add #ifdef FIQ_DEBUG around hcd->hub_port_alloc[], which grows
to a large size if 128 hub addresses are supported. This field is
for debug only for tracking which frame an allocate happened in.

dwc_otg: prevent OOPSes during device disconnects

The dwc_otg_urb_enqueue function is thread-unsafe. In particular the
access of urb->hcpriv, usb_hcd_link_urb_to_ep, dwc_otg_urb->qtd and
friends does not occur within a critical section and so if a device
was unplugged during activity there was a high chance that the
usbcore hub_thread would try to disable the endpoint with partially-
formed entries in the URB queue. This would result in BUG() or null
pointer dereferences.

Fix so that access of urb->hcpriv, enqueuing to the hardware and
adding to usbcore endpoint URB lists is contained within a single
critical section.

dwc_otg: more robust interrupt handling

In the case of a transaction to a device that had previously aborted
due to an error, several interrupts are enabled to reset the error
count when a device responds. This has the side-effect of making the
FIQ thrash because the hardware will generate multiple instances of
a NAK on an IN bulk/interrupt endpoint and multiple instances of ACK
on an OUT bulk/interrupt endpoint. Make the FIQ mask and clear the
associated interrupts.

Additionally, on non-split transactions make sure that only unmasked
interrupts are cleared. This caused a hard-to-trigger but serious
race condition when you had the combination of an endpoint awaiting
error recovery and a transaction completed on an endpoint - due to
the sequencing and timing of interrupts generated by the dwc_otg core,
it was possible to confuse the IRQ handler.

The dwc_otg driver will unmask certain interrupts on a transaction
that previously halted in the error state in order to reset the
QTD error count. The various fine-grained interrupt handlers do not
consider that other interrupts besides themselves were unmasked.

By disabling the two other interrupts only ever enabled in DMA mode
for this purpose, we can avoid unnecessary function calls in the
IRQ handler. This will also prevent an unneccesary FIQ interrupt
from being generated if the FIQ is enabled.

The current timeout is being hit with some cards that complete successfully with a longer timeout.
The timeout is not handled well, and is believed to be a code path that causes corruption.
872a8ff suggests that crappy cards can take up to 3 seconds to respond