Misaligned memory accesses cause hard faults on 32 bit M0 platform

[bradjc]

This one took me a while to track down, but things stopped working when I moved code from Linux to the Cortex M0 nRF51822 BLE chip. I tracked this back to db_qmm_falloc() and db_qmm_balloc(). Basically, if something tries to allocate a buffer of a size not a multiple of 4, when the alloc() functions try to write the size to the beginning of the allocated area, it causes a misaligned memory access:

*((db_int*)(mmp->last_back)) = size;

My quick and dirty workaround is to convert size to the next biggest multiple of four:

db_int blob = size+sizeof(db_int);
if (blob % 4 != 0) {
    blob += (4 - (blob % 4));
}
size = blob - sizeof(db_int);

and throw that near the beginning of the alloc() functions.

I have no idea if that actually fixes anything or just masks issues temporarily. I've ran one SELECT statement that I got to work with that change in there, but haven't done any more than that (I wanted to get this posted so I didn't forget). It's entirely possible that this is the best solution and that all later accesses are safe.

[graemedouglas]

I will add automatic alignment as an option (with #define). I've never tested with ARM processors and so alignment was never an issue. Thank you again.

[kazetora]

@bradjc I tried to compile this using arm-gcc for cortex m0, and the size is too big (about 100KB). It seems you made it work with cortex m0, what sort of compiler and options are you using?

[graemedouglas]

Hi @kazetora,

It looks to me as if this isn't the same issue (though you will have the same problem that @bradjc mentions here; copy and paste his code just before lines 61 and 112 in src/dbmm/db_query_mm.c).

If you haven't already, try not compiling with any print statements. The most recent merge should make this simple, or you can use db_ctconf.h and set #define DB_CTCONF_SETTING_FEATURE_ERROR_MESSAGES 0 (presently line 116).

That should make everything work.