as proposed on tech-kern@.
While all DMA memory used in the USB framework is mapped BUS_DMAMAP_COHERENT
(including data memory, which is memcpy'd from/to USB-private buffers),
the CPU can reorder loads or stores from/to main memory, causing the
controller to have an incoherent view of the DMA descriptors lists for a
short time. bus_dmamap_sync() should contain memory barriers that prevents
the CPU from reordering load/store. Note that BUS_DMAMAP_COHERENT is
still required for the DMA descriptor lists - these can't work properly
with software cache coherency.
This fixes "host controller process error/host controller halted" errors
I'm occasionally seeing with a umodem device on uhci on x86.
Thanks to Michael Lorenz for testing it on his O2, and Izumi Tsutsui on
his Cobalt Qube 2700.
host controller implementations, start with two little functions
which fake up string descriptors (which were inconststent, language
table fetching didn't interoperate with other code in the tree)
-> the current names are confusing (didn't change other drivers)
* fix invalid memory access in usbd_transfer (kern/24636)
-> needed for this driver
* fix USB HC detach race condition (kern/32011)
-> main patch needed for this driver, sc_dying changes in other drivers
not necessary but seem right to me
Patch from Matthew Orgass.
http://mail-index.netbsd.org/tech-kern/2007/06/26/0001.html
device gets removed. However, when the diagnostic check fails,
it is much better to complete the free operation than to abort
it, because this just causes an infinite loop.
exist in newer spec revisions, and is recommended to be set to 1.
So call it _MBO and also use it in the fake root hub descriptors, just
for sanity, even if nothing ever looks at it.
one for tasks of the host controllers. This is needed for drivers like
ural(4) that want to do synchronous USB transfers from the task handler.
Before the split timeouts could not be handled correctly as the task
thread was still blocked. From FreeBSD.
An xfer could be aborted twice (which means that the second abort might
access deallocated memory). This happened when an xfer timed out and
the timeout started an abort. While that abort was taking place the
xfer could be cancelled (usually by closing the pipe), causing a second
abort to begin.
This is now handled by having flags indicating the abort state of an xfer.
Hopefully this will fix the occasional crashes when printing.
* Only futz with the done list in the soft interrupt handler. In the hard
interrupt handler, do nothing except mask WDH and queue the soft interrupt.
This simplifies a bunch of code, removes two O(n^2) queue manipulations,
and gets rid of some really sketchy stuff around the queue head access.
* Use the auto-masking code at the end of the interrupt handler for both WDH
and RHSC interrupts. Again, this reduces the code a little, and avoids
multiple writebacks to the chip registers.
in the wanted bus_dma space. If an allocation fails during operation
(i.e. when hotplugging an umass device), fall back to using the reserve.
The amount can be configured as USB_MEM_RESERVE. The default value is
256k.
Ideally, there would be a way to steal pages in the desired area from
UVM, but that's far more complicated, and this is not intrusive, plus
it works.
jmcneill