Fix the exception handlers. #27
Conversation
|
cc @robamu |
jonathanpallant
changed the title
jonathanpallant
force-pushed
the
fix-undef-handler
branch
from
f0e4edb to
f1ad209
Compare
jonathanpallant
changed the title
|
OK, I think this is good to go. I learned a lot about exception handling today 🙃 |
jonathanpallant
changed the title
|
No, it's still wrong 😭 The undefined exception handler is supposed to also return to the faulting instruction, and we currently don't. This is because we push LR before we work out which mode we're in, and then pop the original LR of the stack with RFEFD. We need to save our state manually. |
| core::arch::global_asm!( | ||
| r#" | ||
| // fn bkpt_from_a32(); | ||
| .arm | ||
| .global bkpt_from_a32 | ||
| .type bkpt_from_a32, %function | ||
| bkpt_from_a32: | ||
| bkpt #0 | ||
| bx lr | ||
| .size bkpt_from_a32, . - bkpt_from_a32 | ||
| "# | ||
| ); |
There was a problem hiding this comment.
unrelated, but this looks like it could fit cortex-a::asm/cortex-r::asm. Perhaps as an unsafe function?
There was a problem hiding this comment.
Yes, but it doesn't help here because I need to know the precise address of the instruction that does the breakpoint.
It was looking at CPSR's Thumb bit, which tells you if the *handler* is in thumb mode, not the code that threw the fault. Change the test to validate the address of the failing function, to verify that we've got this right. Also fixes the issue of _asm_default_undefined_handler damaging r4. You have to save all the state first, then you can touch registers.
One for T32 and one for A32. We have to do this because the exception is non-recoverable (we return to the faulting instruction, which faults again).
I had to move the DFSR read to after the alignment check was disabled, otherwise it double-faulted. Also it turns out the versatileab tests were running with the Cortex-A linker script not the Cortex-R linker script, and the Cortex-R linker script had a bug in it that we previously missed (it wasn't calling the new exception handlers).
jonathanpallant
force-pushed
the
fix-undef-handler
branch
from
93d17c1 to
16ac516
Compare
|
😭 😭 😭 😭 OK, I think I now work out which mode (A32 or T32) tripped the undefined abort, without damaging any registers. This took a lot of work to debug, but it seems to work now? |
jonathanpallant
changed the title
|
Tentatively claiming that I am finished here. |
- The SVC asm trampoline can now be over-ridden - The Undefined, Prefetch and Abort handlers can either return never, or can return a new address to continue executing from when the handler is over - I tried to ensure the all the handlers are listed in the same order as the vector table, for consistency - Put every function in its own section in the hope that unused ones will be GC'd by the linker
More exception handling updates.
|
Reader, I was not finished here. But now I am. I swear. |
| @@ -1,22 +1,25 @@ | |||
| //! # Run-time support for Arm Cortex-A | |||
|
really nice work! |
Also clarify that both only apply to AArch32 systems. The terms 'Cortex-R' and 'Cortex-A' are particular brands of processor made by Arm, and there are processors under each of those brands that run in (or only run in) AArch64 mode.
|
Let's go! Before I fix something else. |
It was looking at CPSR's Thumb bit, which tells you if the handler is in thumb mode, not the code that threw the fault.
I changed the undef-exception test to validate the address of the failing function, to verify that we've got this right. I also split the prefetch-exception test into separate T32 and A32 versions, to check that we return back into the correct state. I didn't bother splitting the data abort test.
I also fixed the issue of _asm_default_undefined_handler damaging r4. You have to save all the state first, then you can touch registers.