SPI flash operations lock for shorter periods
Based on bug report here:
https://esp32.com/viewtopic.php?f=13&t=1489&p=6816#p6773
Long SPI flash erase (and possibly write/read) cycles block all tasks on both CPUs for an extended period, and can lead to WiFi dissassociation (and general priority inversion.)
This MR inserts preemption breaks into large operations (all erases, writes every 8KB, reads every 16KB).
Overhead of a single spi_flash_guart_start()/spi_flash_guard_end() cycle measured at approx 67us (assuming no preemption.)
See merge request !600
bugfix: ESP32 Core Dump reported issues
This MR fixes the following issues:
1. fixed path to esptool when there is no compiled python module for it
2. GDB inernal failure handling was added
See merge request !599
Error handling in NVS initialization
Previously nvs_flash_init worked under an assumption that there should always be at least one free page available. This is true during normal operation, but in some cases (such as when changing application partition table from a non-OTA to an OTA one), NVS partition may get truncated, which will cause empty pages to be lost.
This MR adds error checks for this condition, and updates code which calls `nvs_flash_init` to check for the return code.
For most examples, a simple `ESP_ERROR_CHECK` is added around `nvs_flash_init`. For NVS examples and the OTA example, more robust error handling is added.
This change also removes nvs_flash_init calls from examples which don't use NVS.
See merge request !582
idf_monitor: Small fixes (baud rate, EOL, /dev/tty.X on macOS, Ctrl-T on failure)
* "make monitor" now passed the configured baud rate.
Closes#436https://github.com/espressif/esp-idf/issues/436
* Pass toolchain prefix from sdkconfig into monitor tool
* Allow setting EOL in idf_monitor.py, use CRLF by default
* Detect if /dev/tty.X is used on macOS, warn and replace with /dev/cu.X
* If a build fails or gdb exits, ignore Ctrl-T (allowing Ctrl-T Ctrl-A/F to be same key sequence everywhere)
* Add a note about winpty on Windows.
Ref 02fdf8271d (commitcomment-21369196)
* Fix problems with Console.cancel() not existing in older pyserial
* Print more user-friendly symbols for "start of iram" and "start of flash"
See merge request !594
* Erase range operations allow preemption after each block or sector.
* Write operations allow preemption every 8KB of data.
* Reado operations allow preemption every 16KB of data.
This is to avoid confusion when idf_monitor prints the first symbol in each section, ie "WindowOverflow4" or similar,
when bootloader prints the section mapping address.
Closes#447https://github.com/espressif/esp-idf/issues/447
* "make monitor" not passed the configured baud rate
Closes#436https://github.com/espressif/esp-idf/issues/436
* Pass toolchain prefix from sdkconfig into monitor tool
* Allow setting EOL in idf_monitor.py, use CRLF by default
* Detect if /dev/tty.X is used on macOS, warn and replace with /dev/cu.X
* If a build fails or gdb exits, ignore Ctrl-T (allowing Ctrl-T Ctrl-A/F to be same key sequence everywhere)
* Add a note about winpty on Windows
Ref 02fdf8271d (commitcomment-21369196)
modify touch pad read function
The touch pad read function taks too much time within spin lock waiting for the "done bit". (about 7.6ms as we tested)
So we try to use a mutex on the touch read function and any other functions might change the "done bit".
See merge request !543
This issue is reported from https://github.com/espressif/esp-idf/issues/304.
We found that when we operate the hw command registers in I2C struct, sometimes the behaviour would be different in DEBUG/RELEASE optimisation level:
The code looks like this:
I2C[i2c_num]->command[p_i2c->cmd_idx].byte_num -= 0;
In DEBUG configuration:
I2C[i2c_num]->command[p_i2c->cmd_idx].byte_num -= 0;
400f3ab0: 3388 l32i.n a8, a3, 12
400f3ab2: 14c882 addi a8, a8, 20
400f3ab5: a08840 addx4 a8, a8, a4
400f3ab8: 0020c0 memw
400f3abb: 2898 l32i.n a9, a8, 8
400f3abd: 0020c0 memw
400f3ac0: 28b8 l32i.n a11, a8, 8
400f3ac2: 74a090 extui a10, a9, 0, 8
400f3ac5: 00af92 movi a9, 0xffffff00
400f3ac8: 109b90 and a9, a11, a9
400f3acb: 2099a0 or a9, a9, a10
400f3ace: 0020c0 memw
400f3ad1: 2899 s32i.n a9, a8, 8
In RELEASE configuration:
I2C[i2c_num]->command[p_i2c->cmd_idx].byte_num -= 0;
400f2ba2: 580572 l8ui a7, a5, 88
400f2ba5: 747070 extui a7, a7, 0, 8
400f2ba8: 0020c0 memw
400f2bab: 584572 s8i a7, a5, 88
Looks like the compiler will make it a 8bit operation after optimisation.
But the register value changes from 0x901 to 0x101.
After this 8-bit optimisation, the 11th bit changed from 1 to zero, which caused this error.
We are still trying to find out why that happens, because there might be some risk when operating the register struct.
This is a workaround to avoid "-=" operation on I2C register struct fields.
1. Add low speed channels for LEDC module.
2. Improve fade object allocate mechanism.
3. Improve ledc example, add 2 low speed channels in example.
4. Remove debug code
5. Improve the register bit field name of slow clock.
esp32: update wifi lib for some bugfix/features
1. add options to disable nvs
2. add station keepalive with softap
3. fix beacon rx start time not accurate issue
4. set wifi rx desc/buffer boundary check to internal memory boundary
See merge request !578
nvs_flash_init may return an error code in some cases, and applications
should check this error code (or at least assert on it being ESP_OK, to
make potential issues more immediately obvious).
This change modifies all the examples which use NVS to check the error
code. Most examples get a simple ESP_ERROR_CHECK assert, while NVS
examples, OTA example, and NVS unit tests get a more verbose check which
may be used in real applications.
Expand 'make monitor' support
New 'make monitor' idf_monitor tool for better monitor output. Running 'make monitor' will now:
* Automatically look up code addresses via addr2line and print function, source file, line number in terminal.
* Can reset the ESP32 by typing Ctrl-T Ctrl-R.
* Can run "make flash" by typing Ctrl-T Ctrl-F.
* Can run "make app-flash" by typing Ctrl-T Ctrl-A.
* If gdb stub starts, monitor will automatically run gdb and connect. When gdb exits, ESP32 resets and monitor resumes.
* Exit is still Ctrl-[
Have some more features I'd like to add (log output to file, crash dump support) but I think this is at the point of being useful.
See merge request !565
This change adds a check for the free page count to nvs_flash_init.
Under normal operation, NVS keeps at least one free page available,
except for transient states such as freeing up new page. Due to external
factors (such as NVS partition size reduction) this free page could be
lost, making NVS operation impossible. Previously this would cause an
error when performing any nvs_set operation or opening a new namespace.
With this change, an error is returned from nvs_flash_init to indicate
that NVS partition is in such a state.
make: don’t override TEST_COMPONENTS
When idf_monitor is launched from make, it gets modified `TEST_COMPONENTS`
variable which contains full paths to the test components instead of the
names. This causes `TEST_COMPONENT_PATHS` to be empty and the unit test
app gets built without any test components.
This change introduces an internal `TEST_COMPONENTS_LIST` variable which
gets set either from `COMPONENTS` if `TESTS_ALL` is 1, or from
`TEST_COMOPONENTS` otherwise.
Bootloader makefile is also fixed to avoid propagating `TESTS_ALL` to the
bootloader build step.
See merge request !569
bug fix: fix uart data bug when stop-bit is set as 2-bit mode
This is a workaround for a hardware bug.
When we want to use 2-stop-bit mode, we need to set stop-bit-1 and set dl1_en in rs485 register.
See merge request !544
1. add options to disable nvs
2. add station keepalive with softap
3. fix beacon rx start time not accurate issue
4. set wifi rx desc/buffer boundary check to internal memory boundary
The touch pad read function taks too much time within spin lock waiting for the "done bit". (about 7.6ms as we tested)
So we try to use a mutex on the touch read function and any other functions might change the "done bit".
When idf_monitor is launched from make, it gets modified TEST_COMPONENTS
variable which contains full paths to the test components instead of the
names. This causes TEST_COMPONENT_PATHS to be empty and the unit test
app gets built without any test components.
This change introduces an internal TEST_COMPONENTS_LIST variable which
gets set either from $(COMPONENTS) if TESTS_ALL is 1, and from
TEST_COMOPONENTS otherwise.
Bootloader makefile is also fixed to avoid propagating TESTS_ALL to the
bootloader build step.
