The BSP team noticed that there is spin/mutex lock issue on sh-sci when
CPUFREQ is used. The issue is that the notifier function may call
mutex_lock() while the spinlock is held, which can lead to a BUG().
This may happen if CPUFREQ is changed while another CPU calls
clk_get_rate().
Taking the spinlock was added to the notifier function in commit
e552de2413 ("sh-sci: add platform device private data"), to
protect the list of serial ports against modification during traversal.
At that time the Common Clock Framework didn't exist yet, and
clk_get_rate() just returned clk->rate without taking a mutex.
Note that since commit d535a2305f ("serial: sh-sci: Require a
device per port mapping."), there's no longer a list of serial ports to
traverse, and taking the spinlock became superfluous.
To fix the issue, just remove the cpufreq notifier:
1. The notifier doesn't work correctly: all it does is update stored
clock rates; it does not update the divider in the hardware.
The divider will only be updated when calling sci_set_termios().
I believe this was broken back in 2004, when the old
drivers/char/sh-sci.c driver (where the notifier did update the
divider) was replaced by drivers/serial/sh-sci.c (where the
notifier just updated port->uartclk).
Cfr. full-history-linux commits 6f8deaef2e9675d9 ("[PATCH] sh: port
sh-sci driver to the new API") and 3f73fe878dc9210a ("[PATCH]
Remove old sh-sci driver").
2. On modern SoCs, the sh-sci parent clock rate is no longer related
to the CPU clock rate anyway, so using a cpufreq notifier is
futile.
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Geert writes:
Summary:
- Clean up the naming of clocks in the sh-sci driver and its DT bindings,
- Add support for the optional external clock on (H)SCI(F), where this pin
can serve as a clock input,
- Add support for the optional clock sources for the Baud Rate
Generator for External Clock (BRG), as found on some SCIF variants
and on HSCIF.
All platforms that used to define an sci_fck clock have now switched to
the fck name. Remove the fallback code.
Signed-off-by: Laurent Pinchart <laurent.pinchart+renesas@ideasonboard.com>
Acked-by: Simon Horman <horms+renesas@verge.net.au>
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Acked-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Add support for using the Baud Rate Generator for External Clock (BRG), as
found on some SCIF and HSCIF variants, to provide the sampling clock.
This can improve baud rate range and accuracy.
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Acked-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Add support for using the SCIx clock pin "(H)SCK" as an external clock
input on (H)SCI(F), providing the sampling clock.
Note that this feature is not yet supported on the select SCIFA variants
that also have it (e.g. sh7723, sh7724, and r8a7740).
On (H)SCIF variants with an External Baud Rate Generator (BRG), the
BRG Clock Select Register must be configured for the external clock.
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Acked-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Refactor the clock and baud rate parameter code to ease adding support
for multiple sampling clock sources.
sci_scbrr_calc() now returns the bit rate error, so it can be compared
to the bit rate error using other sampling clock sources.
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Acked-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The "renesas,scif" compatible value is currently used for the SCIF
variant in all Renesas SoCs of the R-Car family. However, the variant
used in the R-Car family is not the common "SH-4(A)" variant, but a
derivative with added "Baud Rate Generator for External Clock" (BRG),
which is also present in sh7734.
Use the family-specific SCIF compatible values for R-Car Gen1, Gen2, and
Gen3 SoCs to differentiate. The "renesas,scif" compatible value can
still be used as a common denominator for SCIF variants with the
"SH-4(A)" register layout (i.e. ignoring the "Serial Extension Mode
Register" (SCEMR) and the new BRG-specific registers).
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Acked-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The "renesas,scif" compatible value is currently used for the SCIF
variant in all Renesas SoCs of the R-Car and RZ families. However, the
variant used in the RZ family is not the common "SH-4(A)" variant, but
the "SH-2(A) with FIFO data count register" variant, as it has the
"Serial Extension Mode Register" (SCEMR), just like on sh7203, sh7263,
sh7264, and sh7269.
Use the (already documented) SoC-specific "renesas,scif-r7s72100"
compatible value to differentiate. The "renesas,scif" compatible value
can still be used as a common denominator for SCIF variants with the
"SH-4(A)" register layout (i.e. ignoring the SCEMR register).
Note that currently both variants are treated the same, but this may
change if support for the SCEMR register is ever added.
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Reviewed-by: Laurent Pinchart <laurent.pinchart+renesas@ideasonboard.com>
Acked-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Store the encoded port and register types directly in of_device_id.data,
instead of using a pointer to a structure.
This saves memory and simplifies the source code, especially when adding
more compatible entries later.
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Reviewed-by: Laurent Pinchart <laurent.pinchart+renesas@ideasonboard.com>
Acked-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Add register definitions for the Baud Rate Generator for External Clock
(BRG), as found in some SCIF and in HSCIF, including a new regtype for
the "SH-4(A)"-derived SCIF variant with BRG.
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Acked-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The maximum baud rate depends on the sampling rate.
HSCIF has a variable sampling rate and sets s->sampling_rate to zero,
hence use the minimum sampling rate of 8.
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Reviewed-by: Laurent Pinchart <laurent.pinchart+renesas@ideasonboard.com>
Acked-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
For low bit rates, the for-loop that reduces the divider returned by
sci_scbrr_calc() and picks the clock select value may terminate without
finding suitable values, leading to out-of-range divider and clock
select values.
sci_baud_calc_hscif() doesn't suffer from this problem, as it correctly
uses clamp().
Since there are only two relevant differences between HSCIF and other
variants w.r.t. bit rate configuration (fixed vs. variable sample rate,
and an additional factor of two), sci_scbrr_calc() and
sci_baud_calc_hscif() can be merged, fixing the issue with out-of-range
values.
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Acked-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
When assuming D = 0.5 and F = 0, maximizing the receive margin M is
equivalent to maximizing the sample rate N.
Hence there's no need to calculate the receive margin, as we can obtain
the same result by iterating over all possible sample rates in reverse
order, and skipping parameter sets that don't provide a lower bit rate
error.
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Acked-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The algorithm to find the best parameters for the requested bit rate
calculates the relative bit rate error, using "(br * scrate) / 1000".
For small "br * scrate", this has two problems:
- The quotient may be zero, leading to a division by zero error,
- This may introduce a large rounding error.
Switch from relative to absolute bit rate error calculation to fix this.
The default baud rate generator values can be removed, as there will
always be one set of values that gives the smallest absolute error.
Print the best set of values when debugging.
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Acked-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
If bps >= 1048576, the multiplication of the predivider and "bps" will
overflow, and both br and err will contain bogus values.
Skip the current and all higher clock select predividers when overflow
is detected. Simplify the calculations using intermediates while we're
at it.
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Acked-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Move the -1 offset of br to the assignment to *brr, so br cannot become
negative anymore, and update the clamp() call. Now all unsigned values
in sci_baud_calc_hscif() can become unsigned.
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Acked-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Blindly writing the default configuration value into the SCSCR register
may change the clock selection bits, breaking the serial console if the
current driver settings differ from the default settings.
Keep the current clock selection bits to prevent this from happening
on e.g. r8a7791/koelsch when support for the BRG will be added.
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Reviewed-by: Laurent Pinchart <laurent.pinchart+renesas@ideasonboard.com>
Acked-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
As F is assumed to be zero in the receive margin formula, frame_len is
not used. Remove it, together with the sci_baud_calc_frame_len() helper
function.
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Reviewed-by: Laurent Pinchart <laurent.pinchart+renesas@ideasonboard.com>
Acked-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
As no platform defines an interface clock the SCI driver always falls
back to a clock named "peripheral_clk".
- On SH platforms that clock is the base clock for the SCI functional
clock and has the same frequency,
- On ARM platforms that clock doesn't exist, and clk_get() will return
the default clock for the device.
We can thus make the functional clock mandatory and drop the interface
clock.
EPROBE_DEFER is handled for clocks that may be referenced from DT (i.e.
"fck", and the deprecated "sci_ick").
Cc: devicetree@vger.kernel.org
Signed-off-by: Laurent Pinchart <laurent.pinchart+renesas@ideasonboard.com>
Acked-by: Simon Horman <horms+renesas@verge.net.au>
[geert: Handle EPROBE_DEFER, reformat description, break long comment line]
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Acked-by: Rob Herring <robh@kernel.org>
Acked-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
This patch fixes an issue that the "length" of scatterlist should be
set using sg_dma_len(). Otherwise, a dmaengine driver cannot work
correctly if CONFIG_NEED_SG_DMA_LENGTH=y.
Fixes: 7b39d90184 (serial: sh-sci: Fix NULL pointer dereference if HIGHMEM is enabled)
Signed-off-by: Yoshihiro Shimoda <yoshihiro.shimoda.uh@renesas.com>
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Acked-by: Simon Horman <horms+renesas@verge.net.au>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Add support for obtaining DMA channel information from the device tree.
This requires switching from the legacy sh_dmae_slave structures with
hardcoded channel numbers and the corresponding filter function to:
1. dma_request_slave_channel_compat(),
- On legacy platforms, dma_request_slave_channel_compat() uses
the passed DMA channel numbers that originate from platform
device data,
- On DT-based platforms, dma_request_slave_channel_compat() will
retrieve the information from DT.
2. and the generic dmaengine_slave_config() configuration method,
which requires filling in DMA register ports and slave bus widths.
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Acked-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Occasionally, DMA transaction completes _after_ DMA engine is stopped.
Verify if the transaction has not finished before forcing the engine to
stop and push the data
Signed-off-by: Muhammad Hamza Farooq <mfarooq@visteon.com>
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
When DMA packet completion and timer expiry take place at the same time,
do not terminate the DMA engine, leading by submission of new
descriptors, as the DMA communication hasn't necessarily stopped here.
Signed-off-by: Muhammad Hamza Farooq <mfarooq@visteon.com>
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
dmaengine_submit() will not start the DMA operation, it merely adds
it to the pending queue. If the queue is no longer running, it won't be
restarted until dma_async_issue_pending() is called.
Signed-off-by: Muhammad Hamza Farooq <mfarooq@visteon.com>
[geert: Add more description]
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Since the DMA engine is not stopped everytime rx_timer_fn is called, the
interrupts have to be redirected back to CPU only when incomplete DMA
transaction is handled
Signed-off-by: Muhammad Hamza Farooq <mfarooq@visteon.com>
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
This prevents DMA timer timeout that can trigger after the port has
been closed.
Signed-off-by: Aleksandar Mitev <amitev@visteon.com>
[geert: Move del_timer_sync() outside spinlock to avoid circular locking
dependency between rx_timer_fn() and del_timer_sync()]
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
There's no need to call sci_start_rx() from sci_request_dma() when DMA
setup fails, as sci_startup() will call sci_start_rx() anyway.
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
For DMA receive requests, the driver is only notified by DMA completion
after the whole DMA request has been transferred. If less data is
received, it will stay stuck until more data arrives. The driver
handles this by setting up a timer handler from the receive interrupt,
after reception of the first character.
Unlike SCIFA and SCIFB, SCIF and HSCIF don't issue receive interrupts on
reception of individual characters if a receive DMA request is in
progress, so the timer is never set up.
To fix receive DMA on SCIF and HSCIF, submit the receive DMA request
from the receive interrupt handler instead.
In some sense this is similar to the SCIFA/SCIFB behavior, where the
RDRQE (Rx Data Transfer Request Enable) bit is also set from the receive
interrupt handler.
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The receive DMA workqueue function work_fn_rx() handles two things:
1. Reception of a full buffer on completion of a receive DMA request,
2. Reception of a partial buffer on receive DMA time-out.
The workqueue is kicked by both the receive DMA completion handler, and
by a timer to handle DMA time-out.
As there are always two receive DMA requests active, it's possible that
the receive DMA completion handler is called a second time before the
workqueue function runs.
As the time-out handler re-enables the receive interrupt, an interrupt
may come in before time-out has been fully handled.
Move part 1 into the receive DMA completion handler, and move part 2
into the receive DMA time-out handler, to fix these race conditions.
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
This allows to:
- Remove forward declarations of static functions,
- Coalesce two sections protected by #ifdef CONFIG_SERIAL_SH_SCI_DMA,
- Avoid shuffling functions around in the near future,
- Avoid adding forward declarations in the near future.
No functional changes.
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
On receive DMA time-out, avoid calling sci_dma_rx_push() if no data was
transferred by the timed out DMA request.
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
If CONFIG_SERIAL_SH_SCI_DMA is enabled, the driver doesn't enable TIE
on SCIF or HSCIF. However, this driver may call sci_tx_interrupt()
in sci_er_interrupt(). After that, the driver cannot care of the
interrupt, and then "irq 109: nobody cared" happens on r8a7791/koelsch
board. This patch fixes the issue.
Signed-off-by: Yoshihiro Shimoda <yoshihiro.shimoda.uh@renesas.com>
[geert] Keep kicking tx when using PIO
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The error handler calls sci_rx_interrupt() to drain the receive FIFO if
an error condition happens.
However, if DMA is enabled on SCIFA or SCIFB, this will call
disable_irq_nosync() twice. Due to this imbalance, the receive interrupt
will never be re-enabled, and reception stops forever.
To fix this, restrict draining the FIFO to PIO mode, and just call
sci_receive_chars() directly.
Inspired by a patch from Yoshihiro Shimoda
<yoshihiro.shimoda.uh@renesas.com>.
Reported-by: Yoshihiro Shimoda <yoshihiro.shimoda.uh@renesas.com>
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
This patch fixes an issue that this driver causes a NULL pointer
dereference in the following conditions:
- CONFIG_HIGHMEM and CONFIG_SERIAL_SH_SCI_DMA are enabled
- This driver runs on the sci_dma_rx_push()
This issue was caused by virt_to_page(buf) in the sci_request_dma()
because this driver didn't check if the "buf" was valid or not. So,
this patch uses the "buf" from dma_alloc_coherent() as is, not page.
This patch also fixes a WARNING issue in sci_rx_dma_release():
WARNING: CPU: 0 PID: 1328 at lib/dma-debug.c:1125 check_unmap+0x444/0x848()
rcar-dmac e6700000.dma-controller: DMA-API: device driver frees DMA memory with different CPU address [device address=0x000000006dd89000] [size=64 bytes] [cpu alloc address=0x000000016189c000] [cpu free address=0x0000000080000000]
WARNING: CPU: 1 PID: 1 at drivers/base/dma-mapping.c:334 dma_common_free_remap+0x48/0x6c()
trying to free invalid coherent area: (null)
Signed-off-by: Yoshihiro Shimoda <yoshihiro.shimoda.uh@renesas.com>
[geert] Rebased
[geert] Reworded
[geert] Dropped .rx_chunk, as it's always identical to .rx_buf[0]
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
There's no need to keep all buffer and DMA pointers on the stack.
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Switch from using tty_buffer_request_room() and looping over
tty_insert_flip_char() to tty_insert_flip_string().
Keep track of buffer overruns in the icount structure, like
serial_core.c does.
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Currently sci_dma_rx_push() has to find the active scatterlist itself,
but in some cases the caller already knows.
Hence let the caller pass the scatterlist, and introduce a helper to
find the active DMA request while we're at it.
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
During serial port shutdown, the DMA receive worker function may still
be called after the receive DMA cleanup function has been called.
Fix this race condition between work_fn_rx() and sci_rx_dma_release() by
acquiring the port's spinlock in sci_rx_dma_release().
This requires releasing the spinlock in work_fn_rx() before calling (any
function that may call) sci_rx_dma_release().
Terminate all active receive DMA descriptors to release them, and to
make sure no more completions come in.
Do the same in sci_tx_dma_release() for symmetry, although the serial
upper layer will no longer submit more data at this point of time.
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
There is a problem when the sci_dma_rx_complete() is processed
before cancel process of work_fn_rx() completes by rx_timer_fn().
This patch locks work_fn_rx().
Signed-off-by: Kazuya Mizuguchi <kazuya.mizuguchi.ks@renesas.com>
Signed-off-by: Yoshihiro Kaneko <ykaneko0929@gmail.com>
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Resubmission of DMA descriptors is explicitly forbidden by the DMA
engine API.
Hence pass DMA_CTRL_ACK to dmaengine_prep_slave_sg(), and prepare a new
DMA descriptor instead of reusing the old one.
Remove sci_port.desc_rx[], as there's no longer a need to access the
active descriptor.
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Simplify the error handling in sci_submit_rx() by
- Moving it to the end of the function,
- Just calling dmaengine_terminate_all() instead of calling
async_tx_ack() for all already submitted descriptors.
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
As dmaengine_prep_slave_sg() is called with the DMA_CTRL_ACK flag set
for DMA transmit requests, there's no need to explicitly acknowledge DMA
transmit requests in the DMA transmit completion callback.
Hence remove the call to async_tx_ack(), and remove the now unused
dma_async_tx_descriptor pointer in the sci_port structure.
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Convert the SCI driver from the SHDMAE-specific partial DMA transfer
handling to the generic dmaengine residual data framework.
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Replace open-coded
- calls to dma_async_tx_descriptor.tx_submit() by calls to the
dmaengine_submit() helper,
- dma_cookie_t comparisons by calls to dma_submit_error().
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The mapped transmit buffer is never unmapped. This leaks quite some
mappings, as the mapping is done in uart_ops.startup(), i.e. every time
the device is opened. Unmap the buffer on device close.
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Simplify the DMA transmit code by using dma_map_single() instead of
constantly modifying the single-entry scatterlist to match what's
currently being transmitted.
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
