[LU-11887] OBD_FREE macro issues with OBD_DEBUG_MEMUSAGE and CONFIG_DEBUG_SLAB

Wed, 4 Sep 2019 16:36:05 +0000

There's some weirdness with the OBD_FREE family of macros, and at least one bug.

First item is more of a style thing, but:
free(), vfree(), and kfree() all accept null pointers without complaint. This simplifies code in exit paths in many places.

The OBD_FREE macros do in their default version, and this behavior is used (inconsistently), but if OBD_DEBUG_MEMUSAGE is on, they do not, because OBD_FREE_PRE asserts the pointer is not null.

We should fix that in OBD_DEBUG_MEMUSAGE;

Then, next:

There are several bugs with POISON_PTR.

If OBD_DEBUG_MEMUSAGE is on:
OBD_FREE poisons the PTR twice.

if CONFIG_DEBUG_SLAB is on, POISON_PTR is enabled...

And we call it after freeing the pointer in several places:
#define OBD_FREE(ptr, size) \
do { \
OBD_FREE_PRE(ptr, size, "kfreed"); \
kfree(ptr); \
POISON_PTR(ptr); \
} while (0)

#define OBD_FREE_LARGE(ptr, size) \
do { \
if (is_vmalloc_addr(ptr))

{ \ OBD_FREE_PRE(ptr, size, "vfreed"); \ vfree(ptr); \ POISON_PTR(ptr); \ }

else {

Yuck! I'll see about doing patches for this at some point.

Whamcloud Community JIRA

[LU-11887] OBD_FREE macro issues with OBD_DEBUG_MEMUSAGE and CONFIG_DEBUG_SLAB