The NMI can asynchronously interrupt the load/store exception handler. This
does occur frequently as the NMI handler code does invoke load/store
exceptions, and the load/store exception handler is heavly used. This was
corrupting the load/store exception handler saved state and thus randomly
corrupting registers a0 to a6 of the interruptee.
NMI Irq error paths call into printf which will use the newlib locks
but these can not be used within the NMI, a task switch can not occur
here. This case would be a terminal error path that is attempting to
write some debug message in the process, so just bail out of the
locking in this case. As a warning a ':' character is emitted and this
will typically prefix lines emitted in this context - it would be an
error for this path to be take in normal operation.
This change means that there is only one set of global stdin, stdout, and stderr
FILE streams shared by all the threads. This reduces memory usage and avoids
having to close these streams before threads exit. These streams still have a
lock to synchronise access.
* The mdns responder has been reworked to lower stack and memory usage. This is
a variation on the upstream code, it use malloc whereas the upstream code uses
pools. The high stack usage of the mdns responder was problem for
esp-open-rtos, so we might have to maintain the differences for now.
* Improved lwip core locking, and lock checking. Upstream improvements, that
need some added support from esp-open-rtos specific code. More core lock is
performed when calling from the esp-open-rtos code now, so a little safer. The
checking is not enforced, but projects might see warning messages and might
want to look into them.
* The esp-open-rtos lwip support has been sync'ed with the new freertos port
included with lwip. There are still some minor differences.
* A few lwip timer bugs have been resolved. This might help resolve some issues.
* Plus it picks up all the other upstream fixes and improvements.
* The default lwip stack has been lowered from 768 words to 480 words,
due to the reduced stack usage by the mdns responder.
It seems most common for the connection to the wider internet to be
via the station netif even if there is also an AP netif, so set the
default to the station netif. The lwip ipv4 route logic will still use
the AP netif if the destination is to that subnet.
The locks are using a bit of the limited ram. It's probably fine to combine some
of these to use the same mutex, and this patch does so for the set initialized
early. The file locks will still be separate, and dynamically created, so a
thread blocking on them will not deadlock all uses of newlib that need a lock.
Once scheduling has started, the startup stack is no longer usable, control can
never return to that stack. So bump up the heap end in the first task that runs
after scheduling has started. This gives back about 1k.
* pwm fix special state + debug print + IRAM interupt
* Special state = don't set timer, safer
* fix timer crash, cant divide by 0
* pwm dont start when duty is set
* reverse option
* fix low duty crash + comments
This patch gets the newlib standard stream descriptors playing well with the
lwip socket descriptors and the spiffs file descriptors. The LWIP_SOCKET_OFFSET
is now defined to be 3, rather than zero, to avoid clashing with the standard
stream descriptors, and the SPIFFS_FILEHDL_OFFSET is bumped up to start after
the lwip descriptors.
* bmp180: comment typo
* ds3231: minor code style fixes. Comment, on the week day start.
* Reverse engineered ets_timers.o Switch from FreeRTOS queue to task notification. Removed unknown/unused code. Rename sdk_ets_handler_isr to process_pending_timers. Add function for microseconds Simplify time to ticks conversion
* ets_timer: code for using the FreeRTOS timers, and remove from libs.
* libc: update to a recent newlib version.
* LwIP v2 support
* spi_write: use uint32_t for the page iteration counter. The page counter was using an uint8_t which seems unnecessary and might wrap.
* sysparam_get_bool: memcpy the binary values out.
* FreeRTOS 9.0.1
* Add an argument to ISRs. Disable interrupts while masking them.
* sysparam: reserve one more flash sector before placing the sysparams. This is to help work with recent SDKs that add a RF cal sector by default in the fifth last sector - just so the sysparam sectors do not jump around when using different SDK versions.
* upnp example: strip down the lwipopts.h file
* sysparam editor: accept newline to end a line of input.
* Add Wificfg. Uses the sysparam store to save the wifi configuration. Adds a basic http server for configuration.
* lwip: disable the tcpip_core_locking_input option. With this option enabled some lwip input processing occurs in the pp task which works well for some uses but some code uses a lot of stack (e.g. mdns) and will overflow the pp task stask, or might unnecessarily slow the critical pp task,
so disable this by default and if not already defined.
* sdk_cnx_add_rc: fix overflow of the table, when no match is found. Also adds source code for sdk_cnx_rc_search, adding a null pointer dereference. Check (that is not expected to be seen), and source code for sdk_cnx_remove_rc.
* http_get example: fix compilation with ipv6 disabled.
* lwip: update to master branch.
* wificfg: allow the AP channel to be 1.
* lwip: rework the tcp ooseq handling. It now accounts for the number of rx pool buffers used and the
available memory when deciding the number of ooseq buffers to retain. Enable the TCP Selective ACK support which appears to help a lot on lossy wifi when using the OOSEQ option.
* Update lwip, fixes losses of pcbs and associated problems.
* lwip: revise the tcp ooseq limit calulations. Was making a mess of the calculation sign. Also added a COPY_PP_RX_PBUFS define to include appropriate limits for this option.
* lwip: ooseq_max_bytes() - set a practical target memory size.
* lwip: revise ooseq handling. Modified for lwip backwards compatibility based on upstream feedback.
* lwip: update to the 2.0.3 release
* lwip: merge upstream master.
* libc: update to upstream master.
* lwip: fix building without TCP_QUEUE_OOSEQ
Also adds source code for sdk_cnx_rc_search, adding a null pointer dereference
check (that is not expected to be seen), and source code for sdk_cnx_remove_rc.
This is to help work with recent SDKs that add a RF cal sector by
default in the fifth last sector - just so the sysparam sectors do not
jump around when using different SDK versions.
The fiq NMI calls rand() from lmac:lmac.a:sdk_lmacTxFrame and the NMI
must not touch the newlib reent structure or enter critical regions
etc, so just use the simple and safe hwrand implementation as a
substitute.
Original work by @ourairquality
* Sysparam threadsafe and SPI access
* Sysparam test cases
* Fix for negative int8
* Sysparam getting bool without memory allocation. Bool tests.
* SPI flash refactoring.
* Extract common spiflash.c into core.
* Use spiflash.c in sysparam.
* Use memcpy in spiflash.c insted of hand-written version.
* Tests for spiflash.c
* Allow changing write function of stdout.
Required for stdout redirection.
Works on blocks, not chars - does _not_ use sdk_os_putc !
Should work even when linking with SPIFFS.
