[Bug 1859359] Re: xHCI and event ring handling

[Benjamin David Lunt]

The xHCI specification states that after processing the Event TRB,
software is to write the address of that TRB to the
xHC_INTERRUPTER_DEQUEUE.  QEMU currently checks this value written and
compares it to its own current pointer, which is now incremented to the
next available TRB, therefore not matching.  When finding that it does
not match, it sends another interrupt.

On receiving this interrupt, software will see this interrupt as a
mismatch in cycle bits and simply write the address of the last
processed Event TRB to the xHC_INTERRUPTER_DEQUEUE and return again.
QEMU will then again check the address and find again that it is a
mismatch, again firing the interrupt.  This causes an infinite loop and
will halt the USB.

I do believe this to be in error.

However, it is up to the majority, which seams to be a Linux based
majority, so if it works on Linux, if it isn't broken, why fix it.

As for the multiple segments in the Event Ring, this was more of a
request than a bug report.  Sorry for the miss representation on that
part.

Thank you,
Ben

-- 
You received this bug notification because you are a member of qemu-
devel-ml, which is subscribed to QEMU.
https://bugs.launchpad.net/bugs/1859359

Title:
  xHCI and event ring handling

Status in QEMU:
  Invalid

Bug description:
  I believe that the Event Ring handling in QEMU is not correct.  For
  example, an Event Ring may have multiple segments.  However, the code
  in xhci_write_event()
  (https://git.qemu.org/?p=qemu.git;a=blob;f=hw/usb/hcd-
  xhci.c;hb=HEAD#l645), starting with line 668, seems to only support a
  single segment.

  Also, QEMU is sending a spurious interrupt after the Guest writes to
  the ERDP register due to the fact that the address written does not
  match the current index.  This is because the index is incremented
  after sending the event.  The xHCI specification states in section
  5.5.2.3.3 "When software finishes processing an Event TRB, it will
  write the address of that Event TRB to the ERDP."

  Since xhci_write_event() has already incremented the index pointer
  (intr->er_ep_idx), the check at line 3098
  (https://git.qemu.org/?p=qemu.git;a=blob;f=hw/usb/hcd-
  xhci.c;hb=HEAD#l3090) no longer is valid.

  I have not studied QEMU's code enough yet to offer a patch.  However,
  this should be a simple fix.

  intr->er_ep_idx++;
  if (intr->er_ep_idx >= intr->er_table[intr->er_segment].er_size) {
    intr->er_ep_idx = 0;
    intr->er_segment++;
    if (intr->er_segment >= intr->er_table_size) {
      intr->er_segment = 0;
      intr->er_pcs = !intr->er_pcs;
    }
  }

  Being sure to incorporate this new segment member into the above code
  (possibly as shown) as well as change the lines at 665 to use the new
  segment member of the structure, and of course in the initialization
  portion of the event ring.

  As for the spurious interrupt at line 3101, a new member will need to
  be added to the code to keep track of the last inserted ED (TRB) into
  the Event Ring and then of course checking against this new member,
  not the now newly incremented member.

  I have sent an email to the author listed at the top of the file as
  well, not sure if this is proper bug reporting etiquette or not.

  Thank you.

To manage notifications about this bug go to:
https://bugs.launchpad.net/qemu/+bug/1859359/+subscriptions