Build mode detection via NDEBUG has unintended side effects

[nmeum]

Commit 6ddf46b fixed a stack overflow on the interrupt stack in debug mode by increasing the size of that stack in debug build mode. The debug build mode is detected by the absence of the NDEBUG macro (which is defined in release mode):

osv/conf/release.mk

Line 1 in bd0a0e1

conf-opt = -O2 -DNDEBUG

This is hacky but at least for the arch-switch.hh header it works as this header is only included from core/, which is compiled with the mode-specific CFLAGS. However, for the arch-cpu.hh header this approach doesn't work at all because this header is included from osv/sched.hh which part of the public API:

osv/include/osv/sched.hh

Line 13 in 2a8035a

#include "arch-cpu.hh"

Therefore, every software which includes osv/sched.hh must also compile with the mode-specific CFLAGS (e.g., include NDEBUG in release mode), which isn't necessarily the case. As such, code in core/* and application code will disagree on sizeof(arch_thread) and this causes very hard to debug bugs since arch_thread is part of sched::thread:

osv/include/osv/sched.hh

Line 832 in 2a8035a

arch_thread _arch;

As a result, code in core/* and application code will disagree on struct member offsets beyond _arch, causing hard to debug memory corruptions.

---

Noticed this due to broken stat counters while rebasing #1338.

[wkozaczuk]

Good catch! I am the author of 6ddf46b, so I will try to fix it as soon as I have time.

But you are welcome to send a patch as well :-)

[nmeum]

But you are welcome to send a patch as well :-)

I am willing to send a patch, but presently I am unsure how to best fix this. We either need applications to pick up some sort of common set of CFLAGS or add a config.h file that defines BUILD_MODE_DEBUG and is included by arch-cpu.hh (causing the build to fail if it isn't found in the include path)? Also open for alternative ideas…