Some TI chips raise the DMA complete interrupt before the actual
transfer has been completed. The code tries to busy wait for a few
microseconds and if that fails it arms an hrtimer to recheck. So far
so good, but that has the following issue:
CPU 0 CPU1
start_next_transfer(RQ1);
DMA interrupt
if (premature_irq(RQ1))
if (!hrtimer_active(timer))
hrtimer_start(timer);
hrtimer expires
timer->state = CALLBACK_RUNNING;
timer->fn()
cppi41_recheck_tx_req()
complete_request(RQ1);
if (requests_pending())
start_next_transfer(RQ2);
DMA interrupt
if (premature_irq(RQ2))
if (!hrtimer_active(timer))
hrtimer_start(timer);
timer->state = INACTIVE;
The premature interrupt of request2 on CPU1 does not arm the timer and
therefor the request completion never happens because it checks for
!hrtimer_active(). hrtimer_active() evaluates:
timer->state != HRTIMER_STATE_INACTIVE
which of course evaluates to true in the above case as timer->state is
CALLBACK_RUNNING.
That's clearly documented:
* A timer is active, when it is enqueued into the rbtree or the
* callback function is running or it's in the state of being migrated
* to another cpu.
But that's not what the code wants to check. The code wants to check
whether the timer is queued, i.e. whether its armed and waiting for
expiry.
We have a helper function for this: hrtimer_is_queued(). This
evaluates:
timer->state & HRTIMER_STATE_QUEUED
So in the above case this evaluates to false and therefor forces the
DMA interrupt on CPU1 to call hrtimer_start().
Use hrtimer_is_queued() instead of hrtimer_active() and evrything is
good.
Reported-by: Torben Hohn <torbenh@linutronix.de>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: stable@vger.kernel.org
Signed-off-by: Felipe Balbi <balbi@ti.com>
Reprogramming the DMA after tear down is initiated leads to warning.
This is mainly seen with ISOCH since we do a delayed completion for
ISOCH transfers. In ISOCH transfers dma_completion should not reprogram
if the channel tear down is initiated.
Signed-off-by: George Cherian <george.cherian@ti.com>
Signed-off-by: Felipe Balbi <balbi@ti.com>
In case of ISOCH transfers the hrtimer workaround for the hardware issue
is not very reliable. Instead of checking musb_is_tx_fifo_empty() in hrtimer
routine, schedule a completion work and check the same in completion work.
Signed-off-by: George Cherian <george.cherian@ti.com>
Signed-off-by: Felipe Balbi <balbi@ti.com>
Enable CPPI to handle high bandwidth transfers, especially to support
webcam captures. Use a single bd to get the whole of the data in case of
high bandwidth transfers.
Signed-off-by: George Cherian <george.cherian@ti.com>
Signed-off-by: Felipe Balbi <balbi@ti.com>
The TX-complete interrupt of the CPPI41 on AM335x fires too early.
Adding a loop and counting how long it takes until the
MUSB_TXCSR_TXPKTRDY bit is cleared I see
FS:
|musb-hdrc musb-hdrc.0.auto: configure ep1/80 packet_sz=64, mode=0, dma_addr=0xadc54002, len=1514 is_tx=1
|cppi41_dma_callback() 74 loops
|musb-hdrc musb-hdrc.0.auto: configure ep1/80 packet_sz=64, mode=0, dma_addr=0xadcd8802, len=1514 is_tx=1
|cppi41_dma_callback() 66 loops
|musb-hdrc musb-hdrc.0.auto: configure ep1/80 packet_sz=64, mode=0, dma_addr=0xadcd8002, len=1514 is_tx=1
|cppi41_dma_callback() 136 loops
|musb-hdrc musb-hdrc.0.auto: configure ep1/80 packet_sz=64, mode=0, dma_addr=0xadf55802, len=1514 is_tx=1
|cppi41_dma_callback() 136 loops
avg: 110 - 150us
HS:
|musb-hdrc musb-hdrc.0.auto: configure ep1/80 packet_sz=512, mode=0, dma_addr=0xaca6f002, len=1514 is_tx=1
|cppi41_dma_callback() 0 loops
|musb-hdrc musb-hdrc.0.auto: configure ep1/80 packet_sz=512, mode=0, dma_addr=0xadd6f802, len=1514 is_tx=1
|cppi41_dma_callback() 2 loops
|musb-hdrc musb-hdrc.0.auto: configure ep1/80 packet_sz=512, mode=0, dma_addr=0xadd6f002, len=1514 is_tx=1
|cppi41_dma_callback() 13 loops
avg: 2us
for the same test case. One loop means a udelay(1). The delay seems to
depend on the packet size. On HS the bit is always cleared for small
packet sizes while on FS it is never the case, it mostly around 110us.
This testing has been performed with g_ether (musb as device) and using BULK
transfers.
INTR transfers are way more fun: during init the gadget sends a INT
packet to the host and cppi41 says "transfer done" shortly after. The
MUSB_TXCSR_TXPKTRDY bit is set even seconds later. The reason is that the host
did not try to receive it, it does so after the interface (on host side) has
been configured. Until this happens, that packet remains in musb's FIFO.
To fix this, two things are done:
- No DMA transfers for INT based endpoints. These transfer are usually
very small and rare so it is likely better to skip the DMA engine and
stuff the four bytes directly into the FIFO
- on HS we poll up to 25us and hope that bit goes away. If not we setup
a hrtimer to poll for it. The 140us delay is a rule of thumb. In FS
the command
| ping 10.10.10.10 -c1 -s65130
creates about 44 1514bytes transfers. About 19 of them need a second
timer to complete.
Reported-by: Bin Liu <b-liu@ti.com>
Cc: stable@vger.kernel.org
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Felipe Balbi <balbi@ti.com>
This patch moves most of the logic in cppi41_dma_callback() into
cppi41_trans_done() where it can be called from another function.
Instead of computing "transferred" (the number of bytes transferred in
the last transaction) in cppi41_trans_done() the member
"cppi41_channel->prog_len" is now set to 0 if the transfer as a whole
can be considered as done. If it is != 0 then the next iteration is
assumed.
This is a preparation for a workaround.
Cc: stable@vger.kernel.org
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Felipe Balbi <balbi@ti.com>
If everything (musb, cppi41, phy) is built-in then musb will start
without the dma engine printing only
|musb-hdrc musb-hdrc.0.auto: Falied to request rx1.
The reason for this is that the musb device structs are created & probed
before those of the cppi41 device. So the cppi41 device is probed too
late.
As a workaround for this allow the musb_cppi41 part to defer the probe
if everything is fine except for the missing DMA controller. In case of
another error we continue.
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Felipe Balbi <balbi@ti.com>
Add the missing unlock before return from function cppi41_dma_callback()
in the error handling case.
Signed-off-by: Wei Yongjun <yongjun_wei@trendmicro.com.cn>
Signed-off-by: Felipe Balbi <balbi@ti.com>
Since the musb-gadget code now calls the dma engine properly it is
possible to enable it for the TX path in device mode.
AM335x Advisory 1.0.13 says that we may lose the toggle bit on multiple
RX transfers. There is a workaround in host mode but none in device mode
and therefore RX transfers are disabled.
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Felipe Balbi <balbi@ti.com>
This driver is currently used by musb' cppi41 couter part. I may merge
both dma engine user of musb at some point but not just yet.
The driver seems to work in RX/TX mode in host mode, tested on mass
storage. I increaed the size of the TX / RX transfers and waited for the
core code to cancel a transfers and it seems to recover.
v2..3:
- use mall transfers on RX side and check data toggle.
- use rndis mode on tx side so we haveon interrupt for 4096 transfers.
- remove custom "transferred" hack and use dmaengine_tx_status() to
compute the total amount of data that has been transferred.
- cancel transfers and reclaim descriptors
v1..v2:
- RX path added
- dma mode 0 & 1 is working
- device tree nodes re-created.
Cc: Vinod Koul <vinod.koul@intel.com>
Cc: Dan Williams <djbw@fb.com>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Felipe Balbi <balbi@ti.com>