I am not sure, why I found it during SG debugging. But, I noticed that
even when req_queued list was empty, there were some request in
request_list having queued flag true. If I run test second time, it
first removes all request from request_list and hence busy_slot was
wrongly incremented.
Cc: <stable@vger.kernel.org>
Signed-off-by: Pratyush Anand <pratyush.anand@st.com>
Signed-off-by: Felipe Balbi <balbi@ti.com>
To work with scatter gather properly, fixes have been done in number of
functions. I will explain requirement of each fixes one by one.
start_slot: used to retrieve all request of SG during cleanup
dwc3_gadget_giveback: We need to skip link TRB if it was one of the
intermediate TRB of SG.
dwc3_prepare_one_trb: We need to track all submitted TRBs during
cleanup. Since, all TRBs would be serially allocated, so we can just
keep starting slot info and we can always find rest of them. We need to
pass sg node number, so that we cab appropriately program ISOC_FIRST/ISOC,
Chain etc.
dwc3_prepare_trbs: last_one should be set when it is last node
of SG as well as last node of request_list.
__dwc3_cleanup_done_trbs: It has been prepared after re-factorization of
dwc3_cleanup_done_reqs. It is called for each TRB of SG.
Signed-off-by: Pratyush Anand <pratyush.anand@st.com>
Signed-off-by: Felipe Balbi <balbi@ti.com>
UPDATE_TRANSFER does not need any parameters. So, no need to prepare it.
Signed-off-by: Pratyush Anand <pratyush.anand@st.com>
Signed-off-by: Felipe Balbi <balbi@ti.com>
When we reach to link trb, we just need to increase free_slot and then
calculate TRB. Return is not correct, as it will cause wrong TRB DMA
address to fetch in case of update transfer.
Cc: <stable@vger.kernel.org>
Signed-off-by: Pratyush Anand <pratyush.anand@st.com>
Signed-off-by: Felipe Balbi <balbi@ti.com>
There were still some corner cases where isoc transfer was not able to
restart, specially when missed isoc does not happen , and in fact gadget does
not queue any new request during giveback.
Cleanup function calls giveback first, which provides a way to queue
another request to gadget. But gadget did not had any data. So , it did
not call ep_queue. To twist it further, gadget did not queue till
cleanup for last queued TRB is called. If we ever reach this scenario,
we must call END TRANSFER, so that we receive a new xfernotready with
information about current microframe number.
Also insure that there is no request submitted to core when issuing END
TRANSFER.
Cc: <stable@vger.kernel.org> # v3.8
Signed-off-by: Pratyush Anand <pratyush.anand@st.com>
Signed-off-by: Felipe Balbi <balbi@ti.com>
Its better to return from each if condition as they are mutually
exclusive.
Signed-off-by: Pratyush Anand <pratyush.anand@st.com>
Signed-off-by: Felipe Balbi <balbi@ti.com>
There are two reasons to generate missed isoc.
1. when the host does not poll for all the data.
2. because of application-side delays that prevent all the data from
being transferred in programmed microframe.
Current code was able to handle first case only. This patch handles
scenario 2 as well.Scenario 2 sometime may occur with complex gadget
application, however it can be easily reproduced for testing purpose as
follows:
a. use isoc binterval as 1 in f_sourcesink.
b. use pattern=0
c. introduce a delay of 150us deliberately in source_sink_complete, so
that after few frames it lands into scenario 2.
d. now run testusb 16 (isoc in test). You will notice that if this
patch is not applied then isoc transfer is not able to recover after
first missed.
Current patch's approach is as under:
If missed isoc occurs and there is no request queued then issue END
TRANSFER, so that core generates next xfernotready and we will issue a
fresh START TRANSFER.
If there are still queued request then wait, do not issue either END or
UPDATE TRANSFER, just attach next request in request_list during giveback.
If any future queued request is successfully transferred then we will issue
UPDATE TRANSFER for all request in the request_list.
Cc: <stable@vger.kernel.org> # v3.6 v3.7 v3.8
Signed-off-by: Pratyush Anand <pratyush.anand@st.com>
Signed-off-by: Felipe Balbi <balbi@ti.com>
Synopsys says:
The HIRD Threshold field must be set to ‘0’ when the device core is
operating in super speed mode.
This patch implements above statement.
Cc: <stable@vger.kernel.org> # v3.6 v3.7 v3.8
Acked-by: Paul Zimmerman <paulz@synopsys.com>
Signed-off-by: Pratyush Anand <pratyush.anand@st.com>
Signed-off-by: Felipe Balbi <balbi@ti.com>
dwc3_gadget_set_ep_config expects maxburst as incremented by 1. So, by
default initialize ep->maxburst to 1 for ep0.
Cc: <stable@vger.kernel.org>
Signed-off-by: Pratyush Anand <pratyush.anand@st.com>
Signed-off-by: Felipe Balbi <balbi@ti.com>
The devm_ functions allocate memory that is released when a driver
detaches. This makes the code smaller and a bit simpler.
Signed-off-by: Jingoo Han <jg1.han@samsung.com>
Signed-off-by: Felipe Balbi <balbi@ti.com>
Currently we're allocating an entire page to
serve as our event buffer. Provided our events
are 4 bytes long, it's very unlikely we will
even trigger 1k events at once.
Even in the worst case scenario where every
endpoint triggers one event and we still have
a couple of error events, that would still
be less than 40 events.
In order to cope with future versions of the
IP which could (or could not) increase the
amount of possible events to trigger
simultaneously, we're using an arbitrary size
of 64 events for our event buffer.
We're saving 3840 bytes by doing so.
Signed-off-by: Felipe Balbi <balbi@ti.com>
As with dwc_readl/writel, the global registers are specified as
offsets starting from the beginning of the xHCI address space,
but the memory region pointed to by dwc->regs already maps to
the start of the global addresses. Fix by offsetting each of the
regs relative to DWC3_GLOBALS_REGS_START.
Signed-off-by: Jack Pham <jackp@codeaurora.org>
Signed-off-by: Felipe Balbi <balbi@ti.com>
commit 392142 moved event buffer allocation out of dwc3_core_init() but
event buffer allocation uses the cached copy of hwparams to determine
the number of event buffers and the caching is done in dwc3_core_init.
So moved dwc3_cache_hwparams function before dwc3_alloc_event_buffers so
that dwc3_alloc_event_buffers sees the correct number of event buffers.
Signed-off-by: Kishon Vijay Abraham I <kishon@ti.com>
Signed-off-by: Felipe Balbi <balbi@ti.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
USB dwc3 patches from Felipe:
"usb: dwc3: patches for v3.8
We can finaly drop HAVE_CLK dependency from exynos glue layer
now that clk API provides no-op stubs when it's not linked
into the kernel.
We're also switching over event buffer allocation to devm_kzalloc()
and moving the allocation out of dwc3_core_init() so that can be
re-used when implementing PM support for v3.9.
After the introduction of PLATFORM_DEVID_AUTO, we can also drop the
homebrew platform device ID handling we had on dwc3 core and let
driver core take care of that for us.
Exynos glue layer learns about DeviceTree and drops platform_data
support completely."
commit 380f0d2 (usb: dwc3: core: switch event
buffer allocation to devm_kzalloc()) was incomplete
leaving a trailing kfree(evt) in an error exit
path.
Fix this problem by removing the trailing kfree(evt).
Cc: Julia Lawall <julia.lawall@lip6.fr>
Reported-by: Fengguang Wu <fengguang.wu@intel.com>
Signed-off-by: Felipe Balbi <balbi@ti.com>
We are removing plat data which was used till now to init and
exit phy. We no longer need this since dwc3-core takes care of
initializing and shutting-down the phy using usb_phy_init()
and usb_phy_shutdown().
Signed-off-by: Vivek Gautam <gautam.vivek@samsung.com>
Signed-off-by: Felipe Balbi <balbi@ti.com>
This patch adds support to parse probe data for
dwc3-exynos driver using device tree.
Signed-off-by: Vivek Gautam <gautam.vivek@samsung.com>
Signed-off-by: Felipe Balbi <balbi@ti.com>
The lockless implementation of the unique id is quite impressive (:P)
but dirver's core can handle it, we can remove it and make our code a
little smaller.
Cc: Anton Tikhomirov <av.tikhomirov@samsung.com>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Felipe Balbi <balbi@ti.com>
commit 93abe8e (clk: add non CONFIG_HAVE_CLK routines)
added clk API stubs when !defined(CONFIG_HAVE_CLK).
This allows us to remove the HAVE_CLK dependency from
Exynos' glue layer and let it compile always.
Signed-off-by: Felipe Balbi <balbi@ti.com>
We call usb_phy_init() from dwc3_core_init() during
probe, so adding usb_phy_shutdown() to dwc3_core_exit()
while removing the device so we don't keep PHYs
turned on, consuming power, unnecessarily.
Signed-off-by: Vivek Gautam <gautam.vivek@samsung.com>
Signed-off-by: Felipe Balbi <balbi@ti.com>
If a USB transfer has already been started, meaning
we have already issued StartTransfer command to that
particular endpoint, DWC3_EP_BUSY flag has also
already been set.
When we try to cancel this transfer which is already
in controller's cache, we will not receive XferComplete
event and we must clear DWC3_EP_BUSY in order to allow
subsequent requests to be properly started.
The best place to clear that flag is right after issuing
DWC3_DEPCMD_ENDTRANSFER.
Cc: stable@vger.kernel.org # v3.4 v3.5 v3.6
Reported-by: Moiz Sonasath <m-sonasath@ti.com>
Signed-off-by: Felipe Balbi <balbi@ti.com>
This patch is in preparation for adding PM support
dwc3 driver. We want to re-use dwc3_core_init and
dwc3_core_exit() functions on resume() and suspend()
callbacks respectively.
Moving even buffer allocation away from dwc3_core_init()
will allow us to reuse the event buffer which was allocated
long ago on our probe() routine.
Signed-off-by: Felipe Balbi <balbi@ti.com>
The rest of the driver is using devm_kzalloc() where possible
and this patch is just making event buffer allocation follow
the example.
Signed-off-by: Felipe Balbi <balbi@ti.com>
This resolves the merge problems with:
drivers/usb/dwc3/gadget.c
drivers/usb/musb/tusb6010.c
that had been seen in linux-next.
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
usb: dwc3: patches for v3.7 merge window
Some much needed changes for our dwc3 driver. First there's a
rework on the ep0 handling due to some Silicon issue we uncovered
which affects all users of this IP core (there's a missing
XferNotReady(DATA) event in some conditions). This issue which
show up as a SETUP transfers which wouldn't complete ever and
we would fail TD 7.06 of the Link Layer Test from USB-IF and
Lecroy's USB3 Exerciser.
We also fix a long standing bug regarding EP0 enable sequencing
where we weren't setting a particular bit (Ignore Sequence
Number). Since we never saw any problems caused by that, it
didn't deserve being sent to stable tree.
On this pull request we also fix Burst Size initialization which
should be done only in SuperSpeed and we were mistakenly setting
Burst Size to the maximum value on non-SuperSpeed mode. Again,
since we never saw any problems caused by that, we're not sending
this patch to stable.
There's also a memory ordering fix regarding usage of bitmaps in
dwc3 driver.
You will also find some sparse warnings fix, a fix for missed
isochronous packets when the endpoint is already busy, and a
fix for synchronization delay on dwc3_stop_active_transfer().
usb: xceiv: patches for v3.7 merge window
nop xceiv got its own header to avoid polluting otg.h. It has also
learned to work as USB2 and USB3 phys so we can use it on USB3
controllers.
Together with those two changes to nop xceiv, we're adding basic
PHY support to dwc3 driver, this is to allow platforms which actually
have a SW-controllable PHY talk to them through dwc3 driver.
We're adding a new phy driver for the OMAP architecture. This driver
is for the PHY found in OMAP4 SoCs, and a new phy driver for the
marvell architecture. An extra phy driver - for Tegra SoCs - is now
moving from arch/arm/mach-tegra* to drivers/usb/phy.
Also here, there's the creation of <linux/usb/phy.h> which should be
used from now on for PHY drivers, even those which don't support
OTG.
We will be adding support for transceivers on
dwc3 driver but not all boards have controllable
transceivers.
For those which don't provide controllable transceivers
we will register nop transceivers.
Signed-off-by: Felipe Balbi <balbi@ti.com>
We will be adding support for transceivers on
dwc3 driver but not all boards have controllable
transceivers.
For those which don't provide controllable transceivers
we will register nop transceivers.
Note that once OMAP's transceiver drivers reach mainline,
this glue layer will change accordingly.
Signed-off-by: Felipe Balbi <balbi@ti.com>
We will be adding support for transceivers on
dwc3 driver but not all boards have controllable
transceivers.
For those which don't provide controllable transceivers
we will register nop transceivers.
Signed-off-by: Felipe Balbi <balbi@ti.com>
If xfernotready is received and there is no request in request_list then
REQUEST_PENDING flag must be set, so that next request in ep queue is executed.
In case of isoc transfer, if xfernotready is already elapsed and even first
request has not been queued to request_list, then issue END TRANSFER, so that
you can receive xfernotready again and can have notion of current microframe.
Signed-off-by: Pratyush Anand <pratyush.anand@st.com>
Signed-off-by: Felipe Balbi <balbi@ti.com>
Populate the resources for xhci afresh instead of directly using the
*struct resource* of core. *resource* structure has parent, sibling,
child pointers which should be filled only by resource API's. By
directly using the *resource* pointer of core in xhci, these parent,
sibling, child pointers are already populated even before
*platform_device_add* causing side effects.
Cc: stable@vger.kernel.org # v3.4, v3.5
Reported-by: Ruchika Kharwar <ruchika@ti.com>
Tested-by: Moiz Sonasath <m-sonasath@ti.com>
Signed-off-by: Kishon Vijay Abraham I <kishon@ti.com>
Signed-off-by: Felipe Balbi <balbi@ti.com>
In case of ep0 out, if length is not aligned to maxpacket size then we use
dwc->ep_bounce_addr for dma transfer and not request->dma. Since, we have
alreday done memcpy from dwc->ep0_bounce to request->buf, so we do not need to
issue cache sync function. In fact, cache sync function will bring wrong data
in request->buf from request->dma in this scenario.
So, cache sync function must not be executed in case of ep0 bounced.
Cc: <stable@vger.kernel.org> # v3.4 v3.5
Signed-off-by: Pratyush Anand <pratyush.anand@st.com>
Signed-off-by: Felipe Balbi <balbi@ti.com>
This commit removes USB_GADGET_DUALSPEED and USB_GADGET_SUPERSPEED
Kconfig options. Since now kernel allows many UDC drivers to be
compiled, those options may turn to no longer be valid. For
instance, if someone decides to build UDC that supports super
speed and UDC that supports high speed only, the latter will be
"assumed" to support super speed since USB_GADGET_SUPERSPEED will
be selected by the former.
The test of whether CONFIG_USB_GADGET_*SPEED was defined was just
an optimisation which removed otherwise dead code (ie. if UDC is
not dual speed, there is no need to handle cases that can happen
if speed is high). This commit removes those checks.
Signed-off-by: Michal Nazarewicz <mina86@mina86.com>
Signed-off-by: Felipe Balbi <balbi@ti.com>
When connection is established non-ss mode, endpoint.maxburst is correctly set
to 0, this means that current code will set maxburst bitfield on DEPCFG's
parameter 0 to the highest possible value (0x0f) which is wrong.
Even though this hasn't caused any issues so far (HW seems to ignore that when
not running in SS mode) we want to make sure maxburst bitfield is only set to
anything other than zero if we're running on SuperSpeed mode. In order to
achieve that, let's check for gadget's operating speed before setting maxburst
bitfield when issuing DEPCFG command.
[ balbi@ti.com : improved commit log a bit in order to clarify the situation ]
Signed-off-by: Chanho Park <chanho61.park@samsung.com>
Signed-off-by: Kyungmin Park <kyungmin.park@samsung.com>
Signed-off-by: Felipe Balbi <balbi@ti.com>
As a bitmap is used for free/used. As a device freed
all memory operations must be scheduled before the bitmap
is manipulated.
Signed-off-by: Oliver Neukum <oneukum@suse.de>
Signed-off-by: Felipe Balbi <balbi@ti.com>
In case some gadget driver tries to enable an endpoint
which is already enabled, we print a nice WARN so we
can track broken gadget drivers. The only problem is that
we're printing the WARN when we already changed endpoint's
name, which would result in endpoints named as:
ep1in-bulk-bulk-bulk-bulk-bulk-bulk-bulk
To prevent that, we will continue to print the WARN,
but do so before changing endpoint's name and return
early.
Signed-off-by: Felipe Balbi <balbi@ti.com>
dwc3_stop_active_transfer has been called from two places. After each
calling of this function , we should allow 100 us delay to synchronize
with interconnect.
It would be better if we put this delay in that function only, rather
than into calling routine of this function.
Signed-off-by: Pratyush Anand <pratyush.anand@st.com>
Reported-by: Michel Sanches <michel.sanches@st.com>
Signed-off-by: Felipe Balbi <balbi@ti.com>
When MISSED_ISOC is set, BUSY is also set. Since, we are handling
MISSED_ISOC as a separate case in third scenario, therefore handle only
BUSY but not MISSED_ISOC in second scenario.
Signed-off-by: Pratyush Anand <pratyush.anand@st.com>
Signed-off-by: Felipe Balbi <balbi@ti.com>
This patch fixes the following sparse warnings:
drivers/usb/dwc3/gadget.c:1096:7: warning: symbol 'ret' shadows an earlier one
drivers/usb/dwc3/gadget.c:1058:8: originally declared here
drivers/usb/dwc3/gadget.c:1100:16: warning: symbol 'dwc' shadows an earlier one
drivers/usb/dwc3/gadget.c:1057:15: originally declared here
drivers/usb/dwc3/gadget.c:1118:16: warning: symbol 'dwc' shadows an earlier one
drivers/usb/dwc3/gadget.c:1057:15: originally declared here
drivers/usb/dwc3/gadget.c:1800:19: warning: symbol 'dep' shadows an earlier one
drivers/usb/dwc3/gadget.c:1778:18: originally declared here
Also, fix the potential checkpatch errors around the if() loops that
this fix patch can create.
Signed-off-by: Moiz Sonasath <m-sonasath@ti.com>
Signed-off-by: Felipe Balbi <balbi@ti.com>
commit 68d3e66 (usb: dwc3: ep0: fix for possible early
delayed_status) added handling for early delayed status,
but the current code only works because so far delayed
status will always be on the IN direction.
This patch makes the code more robust by making sure that
we can handle all directions properly.
Signed-off-by: Felipe Balbi <balbi@ti.com>
Due to the late Silicon limitation found, we are
now pre-starting DATA phase's TRBs. If, still, we
get XferNotReady(DATA) we will ignore it unless
we're getting it for the wrong direction.
In that case we must keep the error case handling
plus add a ENDTRANSFER command to forcefully end
the Data TRB we started previously, then continue
to SetStall and so on.
Signed-off-by: Felipe Balbi <balbi@ti.com>
We uncovered a limitation of this core WRT to the
Link Layer Compliance Suite's TD7.06.
On that test, host will start a GetDescriptor(DEVICE)
standard request, but it will do so only on the
SETUP phase, meaning there will *NOT* be any DATA or
STATUS phases.
The idea of the test is to verify robustness of the
IP WRT framing errors, so the test will send a
sequence of different SETUP_DPs each with a different
framing error and the Suite expects us to be able to
receive all SETUP_DPs with no timeouts.
This core, has the ability to tell us which phase the
host is expecting before we start it. Whenever we
receive a TP or DP when no transfers are cached on
the internal IP's caches, the IP will generate a
XferNotReady event with status informing us (in case
of physical ep0/ep1) if it's related to DATA or STATUS
phases - SETUP phase is expected to be prestarted.
Because we're always waiting for XferNotReady
events for DATA and STATUS phases, we will never
be able to know that the Host wants to start another
SETUP phase instead, which will render us "not
compliant" with TD7.06.
In order to "fix" the problem we must not rely
on XferNotReady events for the DATA phase and try
to always pre-start DATA transfers on physical
endpoints 0 and 1. If host goes back to SETUP phase
from DATA phase we will receive a XferComplete for
that phase with TRB's status set to SETUP_PENDING,
which is only useful for printing a debugging log as
the core expects us to still go through to the STATUS
phase, initiate a CONTROL_STATUS TRB just so it
completes right away and, only then, we go back to
the pending SETUP phase.
SNPS has decided to modify the programming model of
the core so that on-demand DATA phases will not be
supported anymore. Note that this limitation does not
affect 2-stage transfers, meaning that if TD7.06 would
start a 2-stage transfer instead of a 3-stage transfer,
we would receive a "fake" XferNotReady(STATUS) which
would complete right after being initiated with
SETUP_PENDING status.
Other endpoints are also not affected, so we can still
use on-demand transfers on Bulk/Isoc/Interrupt endpoints.
Signed-off-by: Felipe Balbi <balbi@ti.com>
