ci_udc only allocates a single QTD structure per EP. All data needs to be
extracted from the DTD prior to calling ci_ep_submit_next_request(), since
that fills the QTD with next transaction's parameters. Fix
handle_ep_complete() to extract the transaction (remaining) length before
kicking off the next transaction.
In practice, this only causes writes to UMS devices to fail for me. I may
have tested the final versions of my previous ci_udc patch only with
reads. More recently, I had patches applied locally that allocated a QTD
per USB request rather than per USB EP, although since that doesn't give
any performance benefit, I'm dropping those.
Fixes: 2813006fec ("usb: ci_udc: allow multiple buffer allocs per ep")
Signed-off-by: Stephen Warren <swarren@nvidia.com>
Modify ci_ep_alloc_request() to return a dynamically allocated request
object, rather than a singleton that's part of the endpoint. This
requires moving various state from the endpoint structure to the request
structure, since we need one copy per request.
The "fast bounce buffer" b_fast is removed by this change rather than
moved to the request object. Instead, we enhance the bounce buffer logic
in ci_bounce()/ci_debounce() to keep the bounce buffer around between
request submissions. This avoids the need to allocate an arbitrarily-
sized bounce buffer up-front, yet avoids incurring the allocation
overhead each time a request is submitted.
A future enhancement would be to actually submit multiple requests to HW
at once. The Linux driver shows that this is possible. That might improve
throughput (depending on the USB protocol in use), since USB could be
performing a transfer to one HW buffer in parallel with whatever SW
actions U-Boot performs on another buffer. However, I have not made this
change as part of this patch, in order to keep SW changes related to
buffer management separate from any change in the way the HW is
programmed.
Signed-off-by: Stephen Warren <swarren@nvidia.com>
Tegra's USB controller appears to be a variant of the ChipIdea
controller; perhaps derived from it, or simply a different version of
the IP core to what U-Boot supports today.
In this variant, at least the following difference are present:
- Some registers are moved about.
- Setup transaction completion is reported in a separate 'epsetupstat'
register, rather than in 'epstat' (which still exists, perhaps for
other transaction types).
- USB connection speed is reported in a separate 'hostpc1_devlc'
register, rather than 'portsc'.
- The registers used by ci_udc.c begin at offset 0x130 from the USB
register base, rather than offset 0x140. However, this is handled
by the associated EHCI controller driver, since the register address
is stored in controller.ctrl->hcor.
Introduce define CONFIG_CI_UDC_HAS_HOSTPC to indicate which variant of
the controller should be supported. The "HAS_HOSTPC" part of this name
mirrors the similar "has_hostpc" field used by the Linux EHCI controller
core to represent the presence/absence of the hostpc1_devlc register.
Signed-off-by: Stephen Warren <swarren@nvidia.com>
usb_gadget_register_driver() currently unconditionally programs PORTSC
to select a ULPI PHY. This is incorrect on at least the Tegra boards I
am testing with, which use a UTMI PHY for the OTG ports. Make the PHY
selection code conditional upon the specific EHCI controller that is in
use.
Ideally, I believe that the PHY initialization code should be part of
ehci_hcd_init() in the relevant EHCI controller driver, or some board-
specific function that ehci_hcd_init() calls.
For MX6, I'm not sure this PHY initialization code is correct even before
this patch, since ehci-mx6's ehci_hcd_init() already configures PORTSC to
a board-specific value, and it seems likely that the code in ci_udc.c is
incorrectly undoing this. Perhaps this is not an issue if the PHY
selection register bits aren't implemented on this instance of the MX6
USB controller?
ehci-mxs.c doens't appear to touch PORTSC, so this code is likely still
required there.
Signed-off-by: Stephen Warren <swarren@nvidia.com>
At least drivers/usb/gadget/storage_common.c expects that ep->req.actual
contain the number of bytes actually transferred. (At least in practice,
I observed it failing to work correctly unless this was the case).
However, ci_udc.c modifies ep->req.length instead. I assume that .length
is supposed to represent the allocated buffer size, whereas .actual is
supposed to represent the actual number of bytes transferred. In the OUT
transaction case, this may happen simply because the host sends a smaller
packet than the max possible size, which is quite legal. In the IN case,
transferring fewer bytes than requested could presumably happen as an
error.
Modify handle_ep_complete() to write to .actual rather than modifying
.length.
Signed-off-by: Stephen Warren <swarren@nvidia.com>
ci_ep_queue() currently only fills in the page0/page1 fields in the
queue item. If the buffer is larger than 4KiB (unaligned) or 8KiB
(page-aligned), then this prevents the HW from knowing where to write
the balance of the data.
Fix this by initializing all 5 pageN pointers, which allows up to
16KiB (potentially non-page-aligned) buffers.
Signed-off-by: Stephen Warren <swarren@nvidia.com>
The mv_udc is not marvell-specific anymore. The mv_udc is used to drive
generic ChipIdea CI13xxx series OTG cores, so rename the driver to ci_udc
instead.
Signed-off-by: Marek Vasut <marex@denx.de>
Cc: Eric Nelson <eric.nelson@boundarydevices.com>
Cc: Stefano Babic <sbabic@denx.de>
