For a yet unknown reason, both esp8266 and esp32 get stuck in a
reboot loop when thin archives are used.
As a workaround, disable thin archives for now if esp_wifi is used.
fixes#12258
* CPU files should already have 'CPU' defined by the board.
* Do not conditionally define CPU as it is not needed.
This is part of cleanup prior to moving the CPU/CPU_MODEL to
Makefile.features.
Instead of having a send buffer as member `esp_wifi` netdev, a local variable is used now as send buffer. This avoids the need for a locking mechanism and reduces the risk of deadlocks.
Receive call back function `_esp_wifi_rx_cb` is called from WiFi hardware driver with a pointer to a frame buffer that is allocated in the WiFi hardware driver. This frame buffer was freed immediately after copying its content to a single local receive buffer of the `esp_wifi` netdev. The local receive buffer remained occupied until the protocol stack had processed it. Further incoming packets were dropped. However, very often a number of subsequent WiFi frames are received at the same time before the first one is processed completely. Having the single local receive buffer to hold only one received frame, led to a number of lost packets, even at low network load. Therefore, a ringbuffer of rx_buf elements was introduced which doesn't store the frames directly but only references to the frame buffers allocated in WiFi hardware driver. Since there is enough memory to hold several frames, the frames buffers allocated in WiFi hardware driver aren't freed immediatly any longer but are kept until the frame is processed by the protocol stack. This results in a much less loss rate of packets.
Events of different type can be pending at the same time. Therefore it is not possible to use ascending identifiers for the presence of a pending event. Rather, each event type has to be represented by one bit. Thes bits ORed identify all types of pending events. In the esp_wifi_isr function all pending events are then handled in one call. Otherwise, some events might be lost.
Use the -gz option to compress ELF sections containing DWARF information.
This saves around 50% of disk space, without any side effects.
See https://gcc.gnu.org/onlinedocs/gcc-9.2.0/gcc/Debugging-Options.html#Debugging-Options
for more infomation on this option.
Some platforms have an outdated toolchain that does not support -gz so
the flag is blacklisted there. Even then, the results are quite impressive.
I used @cladmi's `buildtest` branch (https://github.com/cladmi/RIOT/tree/wip/du/buildtest)
with this change and compiled the `examples/default` application:
```
$ BUILD_IN_DOCKER=1 DOCKER="sudo docker" make -C examples/default buildtest-indocker
```
The size was obtained with:
```
$ find output -name "*.bin.bindirsize" -type f -exec tail -n1 '{}' \; | cut -f 1 | awk '{s+=$1} END {printf "%.0f", s}'
```
Results:
- Vanilla: 10328112 KB (~10GB).
- with -gz: 4982788 KB (~5GB).
This was inspired by #8496.
Module `newlib` is now used by default. Therefore, the separation of initialization of ctors and the newlibc is not needed any longer. Instead of calling `do_global_ctors` and `_init` separately, `__libc_init_array` is called. Explicit function `do_global_ctors` is removed.
Initializing the stdio file descriptors in global reent structure with newlib fake stdio file descriptors led to the problem that newlib stdio functions printf and puts were not working since they can't operate on these fake stdio file descriptors. Therefore, this initialization was removed. Now, the real stdio file descriptors as created automatically by newlib are used. Specific functions `printf`, `puts`, `getchar`and `putchar` are not required any longer and are removed now.
Modules newlib and newlib_syscalls_default are now used by default. Conditional compilations for MODULE_NEWLIB_SYSCALLS_DEFAULT as well as alternative code are removed completely.
printf and puts used ets_printf before. Unfortunately, ets_printf adds an additional \r for each \n which is not consistent with other RIOT platforms. As a result some automatic tests failed. Therefore, both functions write now character-wise directly to the UART interface.
Using a mutex for critical section handling with portENTER_CRITICAL and portEXIT_CRITICAL does not work for RIOT, as this function can also be called in the interrupt context. Therefore, the given mutex is not used. Instead, the basic default FreeRTOS mechanism for critical sections is used by simply disabling interrupts. Since context switches for the ESP32 are also based on interrupts, there is no possibility that another thread will enter the critical section once the interrupts are disabled.
The option value length of Ethernet addresses can be more than 6 byte in lwIP. Therefore, the max_len parameter is check to be greater than or equal to ETHERNET_ADDR_LEN.
Module esp_idf_heap is enabled in cpu/esp32/Makefile.dep depending on other modules. Since cpu/esp32/Makefile.dep is read after cpu/esp32/Makefile.include, the conditional definition of the linker options for the wrapper functions had to be moved from cpu/esp32/Makefile.include to cpu/esp32/Makefile.dep.
If module esp_idf_heap is used, the memory management functions _malloc_r, _realloc_r, _calloc_r and _free_r have to be overridden by wrapper functions to use the heap_* functions of module _esp_idf_heap. However, this can lead to multiple symbol errors for these functions for some applications. To solve this symbol conflict, _malloc_r, _realloc_r, _calloc_r and _free_r functions are renamed to __wrap_* and the linker options are extended by -Wl,-wrap option when module esp_idf_heap is used.
When standard C libraries are added to BASELIBS to group them together with all other modules, there are multiple definitions for the putchar function. The one that is defined writing to the UART as standard output and the one that is provided by the standard C libraries. To solve this symbol conflict, putchar and getchar functions that use the UART as standard output/input are renamed to __wrap_putchar and __wrap_getchar and the linker options are extended by -Wl,-wrap option.
When linking an application, symbol pthread_setcancelstate is not known in standard C libraries, even if the pthread module is linked. This is because the pthread module is grouped with all other modules, but not with the default C libraries when they are added to LINK_FLAGS. Therefore, standard C libraries have to be added also to BASELIBS to group them with all other modules.
