Some of the defaults were missing or unclear. In particular, I suspect
the defaults were documented assuming there were still module parameters
and taking the default module parameters into account. Now, the defaults
are the values that will get chosen when the params passed to
dwc2_hcd_init are all -1.
Signed-off-by: Matthijs Kooijman <matthijs@stdin.nl>
Acked-by: Paul Zimmerman <paulz@synopsys.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The HWCFG4 register stores the supported utmi width values (8, 16 or
both). This commit reads that value and validates the configured value
against that.
If no (valid) value is given, the parameter defaulted to 8 bits
previously. However, the documentation for dwc2_core_params_struct
suggests that the default should have been 16. Also, the pci bindings
explicitely set the value to 16, so this commit changes the default to
16 bits (if supported, 8 bits otherwise).
With the default changed, the value set in pci.c is changed to -1 to
make it autodetected as well.
Signed-off-by: Matthijs Kooijman <matthijs@stdin.nl>
Acked-by: Paul Zimmerman <paulz@synopsys.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Before, the hwcfg registers were read at device init time, but
interpreted at various parts in the code. This commit unpacks the hwcfg
register values into a struct with properly labeled variables at init
time, which makes all the other code using these values more consise and
easier to read. Some values that were previously stored in the hsotg
struct are now moved into this new struct as well.
In addition to the hwcfg registers, the contents of some fifo size
registers are also unpacked. The hwcfg registers are read-only, so they
can be safely stored. The fifo size registers are read-write registers,
but their power-on values are significant: they give the maximum depth
of the fifo they describe.
This commit mostly moves code, but also attempts to simplify some
expressions from (val >> shift) & (mask >> shift) to
(val & mask) >> shift.
Finally, all of the parameters read from the hardware are debug printed
after unpacking them, so a bunch of debug prints can be removed from
other places.
Signed-off-by: Matthijs Kooijman <matthijs@stdin.nl>
Acked-by: Paul Zimmerman <paulz@synopsys.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Bits 16-31 are reserved, so the old code just reads the whole register to
get bits 0-15, assuming the reserved bits would be 0 (which seems true
on current hardware, but who knows...).
This commit properly masks out the reserved bits when reading and
doesn't touch the reserved bits while writing.
Signed-off-by: Matthijs Kooijman <matthijs@stdin.nl>
Acked-by: Paul Zimmerman <paulz@synopsys.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
For calculating FIFO offsets, the sizes of preceding fifos need to be
known. For filling the GDFIFOCFG register, these fifo sizes were read
from hardware registers. However, these values were written to these
registers just a few lines before, so we can just use the values written
instead.
Signed-off-by: Matthijs Kooijman <matthijs@stdin.nl>
Acked-by: Paul Zimmerman <paulz@synopsys.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
For some reason, the value of the HPTXFSIZ register was built in the
ptxfsiz variable, while there was also a hptxfsiz variable availble.
Better just use that and remove the (now unused) ptxfsiz variable.
Signed-off-by: Matthijs Kooijman <matthijs@stdin.nl>
Acked-by: Paul Zimmerman <paulz@synopsys.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The value of the hcchar register is built from individual values by
shifting and masking. Before, the debug output extracted the individual
values out of the complete hcchar register again by doing the reverse.
This commit makes the debug output use the original values instead.
One debug message got removed, since it would always print a fixed value
of zero.
Signed-off-by: Matthijs Kooijman <matthijs@stdin.nl>
Acked-by: Paul Zimmerman <paulz@synopsys.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
This line extracted the available queue space without properly shifting
it. Since the code only cared wether it was zero or not, it worked as
expected without the shift, but adding shift makes the code cleaner.
While we're here, store the result in a helper variable that was already
declared to increase readability a bit more.
Signed-off-by: Matthijs Kooijman <matthijs@stdin.nl>
Acked-by: Paul Zimmerman <paulz@synopsys.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
This (read-only) register was read twice, storing it for later use the
second time. Now it is only read once, storing it right away.
Signed-off-by: Matthijs Kooijman <matthijs@stdin.nl>
Acked-by: Paul Zimmerman <paulz@synopsys.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Various register fields wider than one bit have constants defined for
their value. Previously, these registers would define the values as they
appear in the register, so shifted to the right to the position the
value appears in the register.
This commit changes those constants to their natural values (e.g, 0, 1,
2, etc.), as they are after shifting the register value to the right.
This also changes all relevant code to shift the values before comparing
them with constants.
This has the advantage that the values can be stored in smaller
variables (now they always require a u32) and makes the handling of
these values more consistent with other register fields that represent
natural numbers instead of enumerations (e.g., number of host channels).
Signed-off-by: Matthijs Kooijman <matthijs@stdin.nl>
Acked-by: Paul Zimmerman <paulz@synopsys.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Previously, the max_packet_count could be set to 1 << x, where x is the
number of bits available (width + 4 in the code). Since 1 << x requires
x + 1 bits to represent, this will not work. The real maximum value is
(1 << x) - 1. This value is already used the default when the set value
is invalid, but the upper limit for the set value was off-by-one.
This change makes the check the same as the one for max_transfer_size,
which was already correct.
Signed-off-by: Matthijs Kooijman <matthijs@stdin.nl>
Acked-by: Paul Zimmerman <paulz@synopsys.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
A generic set of FIFOSIZE_* constants is defined which applies to all
fifo size and offset registers. It is already used for both the
GNPTXFSIZ and HPTXFSIZ registers, but it applies to DPTXFSIZN as well.
Some of these also had specific constants defined. This patch removes
the specific constants and documents to use the generic constants.
Note that the removed constants weren't actually used. Instead, most of
the related code uses hardcoded masks and shifts. But given that
subsequent patches will be moving that code around and introducing the
constants in the process, this patch leaves those untouched.
Also note that the GRXFSIZ register also contains a fifo size, but
there is no corresponding start address register (it is always the first
fifo in memory), the layout of the GRXFSIZ register is different and
cannot use the same constants.
Signed-off-by: Matthijs Kooijman <matthijs@stdin.nl>
Acked-by: Paul Zimmerman <paulz@synopsys.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Reorder the kernel doc comments for 'struct dwc2_core_params' to
match the ordering in the struct itself. Reorder the members of
'struct dwc2_qh' (and its kerneldoc comments) to minimize the
amount of structure padding.
Signed-off-by: Paul Zimmerman <paulz@synopsys.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
There was some code that cleared the dma_mask when dma was disabled in
the driver. Given that clearing the mask doesn't actually tell the usb
core we're not using dma, and a previous commit explicitely sets the
hcd->self.uses_dma value, it seems these values are unneeded and can
only potentially cause problems (when reloading a module, for example).
Signed-off-by: Matthijs Kooijman <matthijs@stdin.nl>
Acked-by: Paul Zimmerman <Paul.Zimmerman@synopsys.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
When dma is disabled inside dwc2 (because the hardware does not support
it, or the code was changed to disable it for testing), let the usb core
know about this by clearing hcd->self.uses_dma.
By default, the usb core assumes that dma is used when a dma_mask is
set, but this might not always match the dma_enable value in dwc2. To
prevent problems resulting from a mismatch, better to explicitely
disable dma in this case (though everything seemed to work with the
wrong value of uses_dma as well, probably only resulted in some unneeded
work).
Signed-off-by: Matthijs Kooijman <matthijs@stdin.nl>
Acked-by: Paul Zimmerman <Paul.Zimmerman@synopsys.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
If the platform or bus driver failed to setup a dma_mask, but the
hardware advertises support for DMA, before DMA would be enabled in
dwc2, but disabled in the usb core, making all connectivity break.
With this commit, the dwc2 driver will emit a warning and fall back to
slave mode in this case.
Note that since commit 642f2ec (staging: dwc2: Fix dma-enabled platform
devices using a default dma_mask) the platform bindings make sure a DMA
mask is always present, but having this check here anyway is probably a
good from a defensive programming standpoint (in case of changes to
platform.c or addition of new glue layers).
Signed-off-by: Matthijs Kooijman <matthijs@stdin.nl>
Acked-by: Paul Zimmerman <Paul.Zimmerman@synopsys.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The dwc2 driver sets the value of the DWC2 GAHBCFG register to 0x6,
which is GAHBCFG_HBSTLEN_INCR4. But different platforms may require
different values. In particular, the Broadcom 2835 SOC used in the
Raspberry Pi needs a value of 0x10, otherwise the DWC2 controller
stops working after a short period of heavy USB traffic.
So this patch adds another driver parameter named 'ahbcfg'. The
default value is 0x6. Any platform needing a different value should
add a DT attribute to set it.
This patch also removes the 'ahb_single' driver parameter, since
that bit can now be set using 'ahbcfg'.
This patch does not add DT support to platform.c, I will leave that
to whoever owns the first platform that needs a non-default value.
(Stephen?)
Signed-off-by: Paul Zimmerman <paulz@synopsys.com>
Tested-by: Stephen Warren <swarren@wwwdotorg.org>
Reviewed-by: Matthijs Kooijman <matthijs@stdin.nl>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Remove the use of bus_to_virt() and just fail the transfer if an
unaligned buffer with no virtual address is found. AFAIK that
can't happen anyway.
Also change setting of coherent DMA mask to the normal 32 bits.
31 bits was only needed when calling bus_to_virt() AFAICR.
Signed-off-by: Paul Zimmerman <paulz@synopsys.com>
Tested-by: Stephen Warren <swarren@wwwdotorg.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Change the non-aligned buffer debug messages to dev_vdbg(). Also
remove some duplicated debug output when the driver is loaded.
Signed-off-by: Paul Zimmerman <paulz@synopsys.com>
Tested-by: Stephen Warren <swarren@wwwdotorg.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The logic in dwc2_hcd_qtd_add() was a bit messy, and one of the
error exit paths was broken. Fix it up.
Signed-off-by: Paul Zimmerman <paulz@synopsys.com>
Tested-by: Stephen Warren <swarren@wwwdotorg.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The driver was lacking calls to usb_hcd_link_urb_to_ep(),
usb_hcd_unlink_urb_from_ep(), and usb_hcd_check_unlink_urb(). Add
those now.
Signed-off-by: Paul Zimmerman <paulz@synopsys.com>
Tested-by: Stephen Warren <swarren@wwwdotorg.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The parameters to dwc2_host_complete() didn't make much sense.
The 'context' parameter always came from the ->priv member of the
'dwc2_urb' parameter, and both of those always came from a struct
dwc2_qtd. So just pass in the struct dwc2_qtd instead.
This also allows us to null out the dwc2_qtd->urb member after it
is freed, which the calling code forgot to do in several places,
causing random driver crashes from dereferencing the freed pointer.
This also requires the calls to dwc2_hc_handle_tt_clear() to be
moved before the calls to dwc2_host_complete(), otherwise that
routine would do nothing because dwc2_qtd->urb has already been
freed.
Signed-off-by: Paul Zimmerman <paulz@synopsys.com>
Tested-by: Stephen Warren <swarren@wwwdotorg.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Matthijs Kooijman
Paul Zimmerman
Julien Delacou