sizeof(time_t) was previously switched from 4 to 8, ROM functions that use
time_t or dependent types (such as "struct stat") are no longer called due as
they still treat sizeof(time_t) as 4 (see commit
24c20d18).
However, there is a ROM callpath that was left out. If putchar is the first
stdio print related call, the call path will result in cantwrite() ->
__swsetup_r() -> __smakebuf_r() -> __swhatbuf_r() using the ROM "struct stat"
(where sizeof(time_t)==4).
Instead of removing all printf related ROM newlib functions (which will result
in increased binary size), this commit adds a workaround to setup the stdio
files before any print related calls occur.
This results in cantwrite() always returning false, thus the callpath described
above never being reached.
Closes https://github.com/espressif/esp-idf/issues/9269
...and all their callers.
With the upcoming switch from sizeof(time_t)==4 to sizeof(time_t)==8,
sizeof(struct stat) is also increasing.
A few newlib functions present in ROM allocate 'struct stat' on the
stack and call _fstat_r on this structure. The implementation of
fstat is provided in ESP-IDF. This implementation will often do
memset(st, 0, sizeof(*st)), where st is 'struct stat*', before setting
some fields of this structure. If IDF is built with sizeof(st)
different from sizeof(st) which ROM was built with, this will lead
to an out-of-bounds write and a stack corruption.
This commit removes problematic ROM functions from the linker script.
Here are the functions which allocate 'struct stat':
* _isatty_r (in ROM)
* __swhatbuf_r, called by __smakebuf_r, called by __swsetup_r and
__srefill_r (in ROM)
* _fseeko_r (not in ROM)
* glob2 (not in ROM)
* _gettemp (not in ROM)
As a result, these functions are used from libc.a, and use correct
size of 'stat' structure.
Closes https://github.com/espressif/esp-idf/issues/7980
This adds support for the retargetable locking implementation in
newlib 3. This feature will be enabled in the future toolchain builds.
With the present version of the toolchain, this code doesn't get used.
When _RETARGETABLE_LOCKING gets enabled, newlib locking implementation
will be modified as follows:
- Legacy ESP-specific _lock_xxx functions are preserved. This is done
because ROM copies of newlib in ESP32 and ESP32-S2 rely on these
functions through the function pointer table. Also there is some
code in IDF which still uses these locking functions.
- New __retarget_lock_xxx functions are introduced. Newlib expects
these functions to be provided by the system. These functions work
pretty much the same way as the ESP-specific _lock_xxx functions,
except one major difference: _lock_acquire receives the lock pointer
by value, and as such doesn't support lazy initialization.
- Static locks used by newlib are now explicitly initialized at
startup. Since it is unlikely that these static locks are used at
the same time, all compatible locks are set to point to the same
mutex. This saves a bit of RAM. Note that there are still many locks
not initialized statically, in particular those inside FILE
structures.
* Linker was choosing ROM symbols for these, which use integer soft-float
operations and are much slower.
* _divsf3() moved to IRAM to avoid regressions with any code that does
integer float division in IRAM interrupt handlers (+88 bytes IRAM)
* Thanks to michal for reporting:
https://esp32.com/viewtopic.php?f=14&t=10540&p=43367
Removes the need to know/guess the paths to these libraries. Once we are gcc 8 only, we
can remove -nostdlib and no additional arguments are needed for system libraries.
The catch is: any time IDF overrides a symbol in the toolchain sysroot, we need
an undefined linker marker to make sure this symbol is seen by linker.
1. separate rom include files and linkscript to esp_rom
2. modefiy "include rom/xxx.h" to "include esp32/rom/xxx.h"
3. Forward compatible
4. update mqtt
