Fix exception handler, add assert, reduce heap usage #4187
Conversation
|
Is it assumed that the cache is not disabled when crash happens? What about Exception 0, which usually indicates a cache disabled error? |
|
Ah, crud. One day I'm going to have to find a good reference manual for this chip. Doing the only exceptions I know how to generate on-demand (WDT, Load/Store) it seemed fine, but there are obviously others. I don't think I ever saw a way for I$ to be disabled. If that's possible, then either this patch should be killed or the strings should move into IRAM (defining a new, local ISTR() macro with the proper section attribute). That said, there are already PSTR()s in the existing routines (but for panics or WDT). Should they be undone? |
|
One option is to check if the cache is disabled when entering the exception handler. If it is, stop SPI operation, if any, and re-enable the cache. That's a bit tricky however, and I don't have hardware at the moment to test the idea. |
|
@igrr For the savings, I'm not sure it's worth the grief. Maybe just abandon this and then put in a new PR to remove the existing PSTR() uses? |
Move all exception strings to IRAM and out of both PMEM (illegal) and add output of any assert() failinf conditions. The exception handler may be called while the SPI interface is in a bad state. This means no PROGMEM reads are allowed, and all data and functions used must be in system RAM or IRAM. Add a new helper macro, ets_printf_P(), which places a constant string in IRAM and copies it to the stack before calling the real ets_printf(). This makes the code simpler to read as no unwieldy combinations of ets_putc/ets_printf/... are required to output anything. The old handler also mistakenly used PSTR() strings in some places, so fix those with this patch as well. Gives back ~180 bytes of heap to every sketch built as the exception handler is always included an application.
earlephilhower
force-pushed
the
exception_to_flash
branch
from
6b237c4
to
f6a3c32
Compare
|
@igrr This is a redesign/combo of this #4187 and the ASSERT changes of #4066. It removes any access to the SPI bus on an exception (there were quite a few spots), and instead makes sure everything needed to handle an exception is in IRAM. It also adds in the assert condition to an assert-fail exception output in a way that doesn't interfere with the ESP exception decoder. Overall, I think it's much cleaner than what was there, and it moves about 180 bytes out of the heap. I'll update #4066 to only remove those assert warnings, and not (as a side effect, almost) touch assertion handling. |
earlephilhower
changed the title
|
Latest merge dumps 'Assertion '0' failed' in the code for #4480, and prints prettier the actual assert condition in all cases, like |
Move all exception strings to IRAM and out of both PMEM (illegal) and add output of any assert() failinf conditions. The exception handler may be called while the SPI interface is in a bad state. This means no PROGMEM reads are allowed, and all data and functions used must be in system RAM or IRAM. Add a new helper macro, ets_printf_P(), which places a constant string in IRAM and copies it to the stack before calling the real ets_printf(). This makes the code simpler to read as no unwieldy combinations of ets_putc/ets_printf/... are required to output anything. The old handler also mistakenly used PSTR() strings in some places, so fix those with this patch as well. Gives back ~180 bytes of heap to every sketch built as the exception handler is always included an application. (cherry picked from commit 855b03c)
---edit---
Move all exception strings to IRAM and out of both PMEM (illegal) and add
output of any assert() failinf conditions.
The exception handler may be called while the SPI interface is in a bad
state. This means no PROGMEM reads are allowed, and all data and functions
used must be in system RAM or IRAM.
Add a new helper macro, ets_printf_P(), which places a constant string in
IRAM and copies it to the stack before calling the real ets_printf().
This makes the code simpler to read as no unwieldy combinations of
ets_putc/ets_printf/... are required to output anything.
The old handler also mistakenly used PSTR() strings in some places, so
fix those with this patch as well.
Gives back ~180 bytes of heap to every sketch built as the exception handler
is always included an application.
---old---
Move the constant strings for stack dumps and exception/reset cause
from RODATA into PROGMEM and adjust the printout functions accordingly.
Gives back ~128 bytes of heap to every sketch built as these are always
included.