This patch fixes two corner-cases of finding RTT control block.
The first one is when there was a partial match (even single byte) at
the end of loaded buffer (uint8_t buf[1024]), but this was not part of
full match. In that case `cb_offset` was not updated correctly and the
returned `*address` was lower by the legth of the partial match. In case
of searched 'SEGGER RTT' (the default control block ID) string, it was
enough to match `buf[1023] == 'S'`, which is quite likely to happen, and
the `*address` was offset by 1 (e.g. it was 0x20000fff instead of
0x20010000).
Updating (or even maintaining) `cb_offset` is not needed, as start
address of control block can be calculated based on memory address that
was loaded into `uint8_t buf[1024]`, the offset within this buffer and
the length of expected string.
The second issue is when control block is prepended with a byte that
matches first ID character, e.g. there is `SEGGER RTT` control block ID
is prepended by another `S`, making memory contents be `SSEGGER RTT`. In
that case there was no match found.
Fix that issue by making sure that tested byte is always compared with
first byte of expected control block ID.
While at it, change names of local variables to better describe their
meaning.
Signed-off-by: Marcin Niestroj <m.niestroj@emb.dev>
Change-Id: I12aa6e202bf12bedcbb888ab595751a2a2518a24
Reviewed-on: https://review.openocd.org/c/openocd/+/7429
Tested-by: jenkins
Reviewed-by: Antonio Borneo <borneo.antonio@gmail.com>
Cores armv8m, e.g. Cortex-M33, can be instantiated without the
optional Security Extension.
In this case, the secure registers are not present and when GDB
try accessing them it triggers a set of errors.
For armv8m cores without security extension, don't provide to GDB
the description of the secure registers.
Change-Id: I254478a4cf883e85b786df3f62c726b2f40d88d9
Signed-off-by: Antonio Borneo <borneo.antonio@gmail.com>
Reported-by: Torbjörn SVENSSON <torbjorn.svensson@foss.st.com>
Reviewed-on: https://review.openocd.org/c/openocd/+/7402
Tested-by: jenkins
Reviewed-by: Tomas Vanek <vanekt@fbl.cz>
With flag --enable-verbose, configure enables compiling some
conditional code that with new gcc triggers an error:
error: '%04x' directive output may be truncated writing
between 4 and 8 bytes into a region of size 5
[-Werror=format-truncation=]
Extend the buffer to contain the full 8 bytes of %04x on a 'int'
and change the limit in snprintf.
Skip the intermediate buffer 's[4]'.
Align the code to the coding style.
Change-Id: Ifc8a6e4686555578a7355a1f6049471fd5e31913
Signed-off-by: Antonio Borneo <borneo.antonio@gmail.com>
Reported-by: Karl Hammar <karl@aspodata.se>
Reported-by: Tommy Murphy <tommy_murphy@hotmail.com>
Fixes: https://sourceforge.net/p/openocd/tickets/376/
Reviewed-on: https://review.openocd.org/c/openocd/+/7403
Tested-by: jenkins
Reviewed-by: Tomas Vanek <vanekt@fbl.cz>
Flashing a SAMD21J17D was failing during NVM erase. The samd21
datasheet specifies that one cause of error conditions is executing an
NVM command while the previous command is still running. The solution
is to wait for INTFLAG.READY after a command is issued.
SAMD21J17A was not exhibiting this problem. Perhaps the later silicon
revision has slower NVM erase times.
Signed-off-by: Dan Stahlke <dan@stahlke.org>
Change-Id: I19745dae4d3fc6e3a7611dcac628e067cb41e0f0
Reviewed-on: https://review.openocd.org/c/openocd/+/7391
Tested-by: jenkins
Reviewed-by: Antonio Borneo <borneo.antonio@gmail.com>
Reviewed-by: Tomas Vanek <vanekt@fbl.cz>
The command 'vdebug mem_path' is reported in the documentation as
'{Config Command}', but the code sets mode = COMMAND_ANY.
The code of the commands sets some value that is only used during
the init phase, so the documentation is correct.
Change mode of command 'vdebug mem_path' to COMMAND_CONFIG.
Change-Id: Icb940fe382cbc75015273b35dcc8a88fc2a7d0ac
Signed-off-by: Antonio Borneo <borneo.antonio@gmail.com>
Reviewed-on: https://review.openocd.org/c/openocd/+/7395
Tested-by: jenkins
Reviewed-by: Jacek Wuwer <jacekmw8@gmail.com>
When target is stopped in WFI/WFE or is in an infinite loop, the
sampled PC will always return the same value.
Command 'profile' requires that distance between min and max PC
should be at least 2, which is not the case for constant PC, and
incorrectly enforces the check through as assert().
Move the code that reads the optional parameters 'start' and 'end'
and check the gap 'end - start' before running the profile.
For self-computed min and max, increase max (or decrease min) to
match the required constraint.
Drop the assert().
Change-Id: I2be8df8568ce8c889923888c492e4f7ce354b16b
Signed-off-by: Antonio Borneo <borneo.antonio@gmail.com>
Fixes: https://sourceforge.net/p/openocd/tickets/370/
Reviewed-on: https://review.openocd.org/c/openocd/+/7400
Tested-by: jenkins
The implementation of command 'monitor profile' has few
issues:
- the address_space is a signed int, so cannot wrap-around on
space over INT_MAX;
- max address is incremented without check for overflow;
- assert() used on errors instead of returning error codes;
- only handles 32 bits PC;
- output file created and left empty on error.
This patch fixes the first two issues, as a wider fix would be too
invasive and should be postponed in a following series.
Change-Id: Id8ead3f6db0fd5730682a0d1638f11836d06a632
Signed-off-by: Antonio Borneo <borneo.antonio@gmail.com>
Fixes: https://sourceforge.net/p/openocd/tickets/370/
Reviewed-on: https://review.openocd.org/c/openocd/+/7394
Tested-by: jenkins
Clang complains about the variable 'orig_dfsr' that can be used
uninitialized both in cortex_a_read_cpu_memory() and in
cortex_a_write_cpu_memory().
The issue is caused by an incorrect error path that used to jump
through 'goto out'. The code after the label 'out' is specific to
handle the case of an error during memory R/W; it is incorrect to
jump there to handle an error during the initialization that
precedes the memory R/W.
Replace the 'goto out' with 'return retval'.
Remove the label 'out' that is now unused.
Change-Id: Ib4b140221d1c1b63419de109579bde8b63fc2e8c
Signed-off-by: Antonio Borneo <borneo.antonio@gmail.com>
Reviewed-on: https://review.openocd.org/c/openocd/+/7393
Tested-by: jenkins
Commit 35a503b08d ("arm_adi_v5: add ap refcount and add get/put
around ap use") modifies the examine functions of mem_ap, cortex_m,
cortex_a and aarch64 by calling dap_put_ap() and then looking again
for the mem-ap and calling dap_get_ap().
This causes an issue if the system is irresponsive and the examine
fails and left the AP pointer to NULL. If the system was already
examined the NULL pointer will cause a SIGSEGV.
Commit b6dad912b8 ("target/cortex_m: prevent segmentation fault
in cortex_m_poll()") proposes a fix for one specific case and only
on cortex_m.
Modify all the examine functions by skipping look-up for the AP if
it was already set in a previous examine; the target's AP is not
supposed to change during runtime.
Remove the partial fix for cortex_m as it is not needed anymore.
Change-Id: I806ec3b1b02fcc76e141c8dd3a65044febbf0a8c
Signed-off-by: Antonio Borneo <borneo.antonio@gmail.com>
Fixes: 35a503b08d ("arm_adi_v5: add ap refcount and add get/put around ap use")
Reviewed-on: https://review.openocd.org/c/openocd/+/7392
Tested-by: jenkins
Reviewed-by: Tomas Vanek <vanekt@fbl.cz>
Without the fix, will see "Can not find free FP Comparator" error log
Change-Id: Id0d91cc02b7055e44d27507f9c05ccd48ff49838
Signed-off-by: Nima Palizban <n.palizban@gmail.com>
Fixes: fb43f1ff4e (target: Rework 'set' variable of break-/watchpoints)
Reviewed-on: https://review.openocd.org/c/openocd/+/7389
Tested-by: jenkins
Reviewed-by: Tomas Vanek <vanekt@fbl.cz>
Reviewed-by: Antonio Borneo <borneo.antonio@gmail.com>
One place described the tcl port as 5555, which was changed in 163bd86071
Reported on IRC.
Change-Id: If740a29443793d6a4d4f8c9db54f0fc8344a6c1c
Signed-off-by: Karl Palsson <karlp@tweak.net.au>
Reviewed-on: https://review.openocd.org/c/openocd/+/7385
Tested-by: jenkins
Reviewed-by: Antonio Borneo <borneo.antonio@gmail.com>
If multiple targets are specified as -rtos <rtos_type>, the
rtos_update_threads was called only if the last target was specified as
rtos, which is inconsistent with other checks of whether or not smp target
is an rtos one.
Signed-off-by: Evgeniy Naydanov <evgeniy.naydanov@syntacore.com>
Change-Id: Ie52bc6b6c8f841d31b9590fcbc44e985d3cba0eb
Reviewed-on: https://review.openocd.org/c/openocd/+/7244
Tested-by: jenkins
Reviewed-by: Antonio Borneo <borneo.antonio@gmail.com>
When OpenOCD receives a step-execution command from GDB and the target
is configured as rtos=hwthread, OpenOCD reconstructs the thread-info.
However, OpenOCD does not restore the thread id which is currently
selected by GDB. Due to this issue, OpenOCD sends the information of
wrong thread to GDB after the step execution.
This commit fixes the above issue by adding a code to save/restore the
thread id selected by GDB.
Signed-off-by: Koudai Iwahori <koudai@google.com>
Change-Id: I761a1141c04d48f1290e4f09baa7c7024f86f36a
Reviewed-on: https://review.openocd.org/c/openocd/+/7358
Tested-by: jenkins
Reviewed-by: Antonio Borneo <borneo.antonio@gmail.com>
When OpenOCD receives 'g' packet (read general registers) from GDB and
target is configured as rtos=hwthread, hwthread_get_thread_reg_list is
called. However, it does not check if the register valid or not. Due to
this issue, OpenOCD returns invalid register values to GDB.
This commit adds a validity check to hwthread_get_thread_reg_list. If
the register is not valid, it tries to read the register from the
target.
Signed-off-by: Koudai Iwahori <koudai@google.com>
Change-Id: Iad6424b62124271ec411b1dfc044b57dfc460280
Reviewed-on: https://review.openocd.org/c/openocd/+/7357
Tested-by: jenkins
Reviewed-by: Antonio Borneo <borneo.antonio@gmail.com>
Clang ignores that xds110_swd_write_reg() is always called with
bit SWD_CMD_RNW in 'cmd' set to zero.
It then complains that the local variable 'value' gets passed by
address to xds110_swd_queue_cmd() and in case of 'read request'
such stack address get stored for later use:
src/jtag/drivers/xds110.c:1363:1: warning: Address of
stack memory associated with local variable 'value' is
still referred to by the global variable 'xds110' upon
returning to the caller. This will be a dangling
reference [core.StackAddressEscape]
To both xds110_swd_write_reg() and xds110_swd_read_reg(), add an
assert() to inform Clang about the state of bit SWD_CMD_RNW.
Change-Id: I7687c055ec71424b642e152f478723a930966e3a
Signed-off-by: Antonio Borneo <borneo.antonio@gmail.com>
Reviewed-on: https://review.openocd.org/c/openocd/+/7380
Tested-by: jenkins
The variable retval is assigned a value that is never used.
Scan-build reports:
Although the value stored to 'retval' is used in the
enclosing expression, the value is never actually read
from 'retval'.
Drop the dead assignment.
Change-Id: I11588dee748a55d52aa7f35bc1967b7df55af7fc
Signed-off-by: Antonio Borneo <borneo.antonio@gmail.com>
Reviewed-on: https://review.openocd.org/c/openocd/+/7379
Tested-by: jenkins
The command should now send the full 5 byte command length, which
includes the command tag (0x19) and the 4-byte baudrate word, instead
of only the last 3 bytes of the baudrate.
Signed-off-by: Nick Kraus <nick@nckraus.com>
Change-Id: Idd6e084efd7492489aa900cdbf08f540944041cb
Reviewed-on: https://review.openocd.org/c/openocd/+/7370
Tested-by: jenkins
Reviewed-by: Tomas Vanek <vanekt@fbl.cz>
Reviewed-by: Antonio Borneo <borneo.antonio@gmail.com>
Thread name loading was not correctly handled if a ThreadX task has a NULL
name.
Signed-off-by: Giulio Fieramosca <giulio@glgprograms.it>
Change-Id: I03071930182bc2585b61ce5d8c67491710883dd6
Reviewed-on: https://review.openocd.org/c/openocd/+/7328
Tested-by: jenkins
Reviewed-by: Tomas Vanek <vanekt@fbl.cz>
Reviewed-by: Antonio Borneo <borneo.antonio@gmail.com>
target_run_flash_async_algorithm expects the source_buffer to have
at least 2 words reserved for read and write pointers in addition to the
FIFO buffer. If the size of the data to be flashed is <= 8 bytes then
the flash function will fail with "corrupted fifo read pointer" error.
Ensure the allocated buffer is big enough to hold both FIFO buffer and
read/write pointers.
Change-Id: I09c22eaac517b8cfea8e0b463f5deb6b98afd267
Signed-off-by: Simon Smiganovski <simon.smiganovski@fruitcore.de>
Signed-off-by: Paul Fertser <fercerpav@gmail.com>
Reviewed-on: https://review.openocd.org/c/openocd/+/7342
Tested-by: jenkins
Reviewed-by: Tomas Vanek <vanekt@fbl.cz>
Reviewed-by: Antonio Borneo <borneo.antonio@gmail.com>
armv7m_start_algorithm() saves register values to arch_info->context.
armv7m_wait_algorithm() restores register values from arch_info->context.
Exclude registers with flag exist = false from both loops.
While on it refactor the register restore: introduce 'struct reg' pointer
and dereference it instead of numerous accesses by a full path
from armv7m pointer.
Change-Id: I1600084db84809ee13bcf8e7828b79f8c9ff9077
Signed-off-by: Tomas Vanek <vanekt@fbl.cz>
Reviewed-on: https://review.openocd.org/c/openocd/+/7276
Tested-by: jenkins
Reviewed-by: Antonio Borneo <borneo.antonio@gmail.com>
The speed coefficient for Raspberry Pi 2 was probably calibrated
for a scaled down clock frequency.
To prevent JTAG/SWD overclocking, use the value corresponding
to the 'official' maximum CPU clock.
Change-Id: Iaff58b092198dce6d6552c9d31d6a3ba4aaaa2d5
Signed-off-by: Tomas Vanek <vanekt@fbl.cz>
Reviewed-on: https://review.openocd.org/c/openocd/+/7305
Tested-by: jenkins
Reviewed-by: Jonathan Bell <jonathan@raspberrypi.com>
Clang spots a potential NULL pointer dereferencing that is instead
an incorrect use of an array of pointers:
src/rtos/hwthread.c:254:32: warning: Dereference of null pointer
[core.NullDereference]
(*rtos_reg_list)[j].number = (*reg_list)[i].number;
^~~~~~~~~~~~~~~~~~~~~
The error has not been spotted before because:
- this function is not called for the first core of the SMP node,
- for the other cores on Cortex-A it still returns valid register
value for the first 12 ARM registers, then it diverges.
Also Valgrind does not spot any issue at runtime.
Address the array correctly.
While there, use DIV_ROUND_UP() macro for the computation.
Change-Id: Ib87e60e0edfd9671091f5dcfa9aedaf1aed800d1
Signed-off-by: Antonio Borneo <borneo.antonio@gmail.com>
Reviewed-on: https://review.openocd.org/c/openocd/+/7337
Tested-by: jenkins
Reviewed-by: Tomas Vanek <vanekt@fbl.cz>
Commit 0cedf10f8f ("Remove duplicate of a counter in
hwthread_update_threads") introduced a code bug.
In the second foreach_smp_target() loop, variable "threads_found"
gets passed to routine hwthread_fill_thread(). By removing the
counting of threads_found from the second loop, the
incorrect thread counter value gets passed to hwthread_fill_thread().
Change-Id: Ie89e53ccd28bb72b6838ef2f12106a1fe8d00994
Suggested-by: Daniel Goehring <dgoehrin@os.amperecomputing.com>
Signed-off-by: Tomas Vanek <vanekt@fbl.cz>
Reviewed-on: https://review.openocd.org/c/openocd/+/7307
Tested-by: jenkins
Reviewed-by: Antonio Borneo <borneo.antonio@gmail.com>
The autodetection of CMSIS-DAP v2 bulk interface is tricky
as not all adapters conform CMSIS-DAP specs.
If an interface has a string descriptor containing CMSIS-DAP,
then OpenOCD did not insisted on the correct interface class
LIBUSB_CLASS_VENDOR_SPEC.
However the relaxed test caused false autodetection of v2 bulk
interface on some CMSIS-DAP v1 adapters with an additional serial
interface with the string descriptor stupidly containing
CMSIS-DAP text.
Make the test less relaxed, refuse autodetection of the interfaces
with the class number of well known functions including CDC and MSC.
Link: https://sourceforge.net/p/openocd/tickets/368/
Change-Id: I917cb257eb42aab93560cc39c61ec35a60ce52e3
Signed-off-by: Tomas Vanek <vanekt@fbl.cz>
Reviewed-on: https://review.openocd.org/c/openocd/+/7279
Tested-by: jenkins
Reviewed-by: SilverFox <yyjdelete@126.com>
Reviewed-by: Antonio Borneo <borneo.antonio@gmail.com>
The variable retval is assigned a value that is never used, as it
is reassigned few lines below.
Drop the dead assignment.
Change-Id: Id4e9134408fab3e04936d36e95724bf8d3ab55aa
Signed-off-by: Antonio Borneo <borneo.antonio@gmail.com>
Reviewed-on: https://review.openocd.org/c/openocd/+/7304
Tested-by: jenkins
Clang get confused by initializing the array uint16_t lock_word[],
casting it to (uint8_t *), then accessing the second element of
the uint8_t pointer.
src/target/dsp5680xx.c:2046:41: warning: The left operand of '<<'
is a garbage value [core.UndefinedBinaryOperatorResult]
uint16_t tmp = (buffer[0] | (buffer[1] << 8));
~~~~~~~~~ ^
Fix it by replacing the array with a single uint16_t.
The code is still depending on host endianness; no fix for this is
proposed.
Change-Id: I16dfd60cab117dd145aeecf10d9593574ff233a2
Signed-off-by: Antonio Borneo <borneo.antonio@gmail.com>
Reviewed-on: https://review.openocd.org/c/openocd/+/7302
Tested-by: jenkins
The function esirisc_jtag_recv() can be called with argument
num_in_fields = 0, for example as consequence of calling
esirisc_jtag_continue().
In this case, num_in_bytes is zero and the allocation of the
variable-length array 'r' requires size zero.
src/target/esirisc_jtag.c:133:2: warning: Declared variable-length
array (VLA) has zero size [core.VLASize]
uint8_t r[num_in_bytes * 2];
^~~~~~~~~ ~~~~~~~~~~~~~~~~
Fix it by forcing size one when num_in_bytes is zero.
Change-Id: Id764c7b5ec4f5b3c18c7da650bbff39fc98ed049
Signed-off-by: Antonio Borneo <borneo.antonio@gmail.com>
Reviewed-on: https://review.openocd.org/c/openocd/+/7301
Tested-by: jenkins
Clang assumes that size could assume a value that is not 1 nor 2
nor 4. In such condition the buffer in t is allocated (size != 1)
and not initialized. This triggers an error:
src/target/openrisc/or1k_du_adv.c:655:14: warning: 2nd function
call argument is an uninitialized value [core.CallAndMessage]
crc_calc = adbg_compute_crc(crc_calc, data[i], 8);
^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Add the default case to cover other values of size.
After this fix, clang still complains on the same line, this time
misunderstanding the limits of the loop and considering that
buf_bswap16() only swaps the first 16 bits, thus passing not
initialized value data[2] to adbg_compute_crc()
Replace malloc() with calloc() to silent it.
Change-Id: I358d7fb2ebefd69255670641bd435b770762a301
Signed-off-by: Antonio Borneo <borneo.antonio@gmail.com>
Reviewed-on: https://review.openocd.org/c/openocd/+/7300
Tested-by: jenkins
All the converters functions:
h_u64_to_le()
h_u64_to_be()
h_u32_to_le()
h_u32_to_be()
h_u24_to_le()
h_u24_to_be()
h_u16_to_le()
h_u16_to_be()
have signed type in their prototype, while the function name and
all the current use cases pass an unsigned value.
Change the prototypes to use unsigned types.
Change-Id: I76dcfdd7912b81f60902184712b2907eae9843f7
Signed-off-by: Antonio Borneo <borneo.antonio@gmail.com>
Reviewed-on: https://review.openocd.org/c/openocd/+/7299
Tested-by: jenkins
If we just invented thread 1 to represent the current execution, we
need to make sure the RTOS object also claims it's the current thread
so that threadx_get_thread_reg_list() doesn't attempt to read a
thread control block at 0x00000001.
Signed-off-by: Ben McMorran <bemcmorr@microsoft.com>
Change-Id: I7f71e730d047858898297e4cb31db8e47e0c371c
Reviewed-on: https://review.openocd.org/c/openocd/+/7280
Tested-by: jenkins
Reviewed-by: Antonio Borneo <borneo.antonio@gmail.com>
The existing rp2040-core0.cfg configuration file was intended
for a special adapter which selects a SWD multidrop target on its own.
This means that rp2040-core0.cfg is totally unusable with a standard SWD
adapter.
To fix the problem, mark rp2040-core0.cfg as deprecated and
add rp2040.cfg, a basic config file with multidrop target selection.
Change-Id: I5194e42f529a2d9645481424b7c66ab61efa44ee
Signed-off-by: Tomas Vanek <vanekt@fbl.cz>
Reviewed-on: https://review.openocd.org/c/openocd/+/7275
Tested-by: jenkins
Reviewed-by: Jonathan Bell <jonathan@raspberrypi.com>
Reviewed-by: Antonio Borneo <borneo.antonio@gmail.com>
instrument "target/stm32x5x_common.cfg" used by both STM32L5x/U5x
to support HLA adapters like "interface/stlink.cfg" in non-secure mode
if the device switches to secure mode, the debug session will be
stopped immediately (with an explanatory message).
Change-Id: I645fdd55e3448ef82d0ddcc396f42fd7b2f39ac3
Signed-off-by: Tarek BOCHKATI <tarek.bouchkati@gmail.com>
Reported-by: Patrik Bachan <diggit@users.sourceforge.net>
Fixes: https://sourceforge.net/p/openocd/tickets/317/
Reviewed-on: https://review.openocd.org/c/openocd/+/6546
Reviewed-by: Tomas Vanek <vanekt@fbl.cz>
Reviewed-by: Antonio Borneo <borneo.antonio@gmail.com>
Tested-by: jenkins
`xtensa_do_step` is invoked from `xtensa_prepare_resume` to silently
step over BP/WP before resuming.
For example; in the case of WPs (DEBUGCAUSE_DB), in the current
implementation `xtensa_do_step` will generate one more
`TARGET_EVENT_HALTED` after the original one caused by WP itself.
This patch moves the halted event cb call after
the step is done successfully.
Signed-off-by: Erhan Kurubas <erhan.kurubas@espressif.com>
Change-Id: I9048e14fb316dc124847a42cfaefb1f76b5ce53e
Reviewed-on: https://review.openocd.org/c/openocd/+/7274
Tested-by: jenkins
Reviewed-by: Ian Thompson <ianst@cadence.com>
Reviewed-by: Antonio Borneo <borneo.antonio@gmail.com>
- Fix the declaration of riscv command 'set_mem_access'.
- Remove non existing riscv command 'set_scratch_ram'.
- Add riscv commands 'info', 'reset_delays'; copy the description
from the 'help' text.
- Don't add riscv commands 'set_prefer_sba' and 'test_sba_config_reg'
as they are marked as deprecated.
- Ensure that 'test_sba_config_reg' prints a deprecation warning
when used.
Change-Id: I39dc3aec4e7f13b69ac19685f1b593790acdde83
Signed-off-by: Antonio Borneo <borneo.antonio@gmail.com>
Signed-off-by: Jan Matyas <matyas@codasip.com>
Reviewed-on: https://review.openocd.org/c/openocd/+/7268
Reviewed-by: Tim Newsome <tim@sifive.com>
Tested-by: jenkins
Remove the @anchor{} tags that are not referenced in the
documentation.
Change-Id: Ia8e9f75afb08e08ef99d0c8fd82115d689e4a267
Signed-off-by: Antonio Borneo <borneo.antonio@gmail.com>
Reviewed-on: https://review.openocd.org/c/openocd/+/7270
Tested-by: jenkins
This GPIO driver is common to SoCs that have in-order ARM cores
(BCM2835) as well as superscalar (BCM2836-7) and speculative
out-of-order cores (BCM2711).
For BCM2837 and BCM2711, the processor can dual-issue stores and
is free to merge writes to peripheral memory for pages mapped
MT_NORMAL_NC, which is the default provided by /dev/[gpio]mem.
This can cause glitches (or missing edges) on GPIO pins when
toggled with no delay, as pipelined writes to the same address
can get arbitrarily squelched.
To prevent this happening, make sure the preceding write ops are
flushed outside the shareable domain by using a memory barrier.
Signed-off-by: Jonathan Bell <jonathan@raspberrypi.com>
Change-Id: I8805cc0911667bcb9b7f4ca340d7f4f1cb25d096
Reviewed-on: https://review.openocd.org/c/openocd/+/7258
Tested-by: jenkins
Reviewed-by: Antonio Borneo <borneo.antonio@gmail.com>
Marcin Niestroj
Paul Fertser
Antonio Borneo
Dan Stahlke
Nima Palizban
Karl Palsson
Evgeniy Naydanov
Koudai Iwahori
Marc Schink
Rocco Marco Guglielmi
Nick Kraus
Tomas Vanek
Giulio Fieramosca
Andreas Bolsch
Simon Smiganovski
George Voicu
Daniel Anselmi
Ben McMorran
Tarek BOCHKATI
Tarek BOCHKATI
Erhan Kurubas
Jonathan Bell