McEs, A Hacker Life: Improving our memory debugging infrastructure

Friday, April 21, 2006

Improving our memory debugging infrastructure

Federico writes about two small fixes we need to improve the memory debugging output of valgrind.

I've been working on a different approach: Instead of asking glib to be gc-friendly and to use malloc instead of g_slice implementation, I wrote patches for glib to hint valgrind directly. What it means is that glib will continue to use g_slice's complex implementation, and no gc-friendly cleanups, but it will give valgrind hints when a slice of memory is "freed" either using g_slice_free or for example g_array_resize...

With this patch you neither need to set G_DEBUG=gc-friendly, nor G_SLICE=malloc. glib simply picks it up when it's running under valgrind (or on top of ;). The patch also implements a global flag called g_mem_in_checker, that will be set to TRUE when running under valgrind (or other memory checkers in the future). ORBit then can check both this flag and g_mem_gc_friendly to initialize its buffers before writing them out.

¶ 2:51 AM

Comments:

ORBit avoids clearing buffers as an optimization. Having to look at a flag would sort of defeat this optimization...

# posted by

Anonymous : April 21, 2006 6:42 AM

I don't think so. You are comparing checking a single integer with clearing a buffer of, I don't know, say, 200 bytes.

# posted by

behdad : April 21, 2006 6:46 AM

No, that's checking an integer vs. nothing.

# posted by

Anonymous : April 21, 2006 7:12 AM

Of course runtime checking has some overhead. Whether it's negligible or not is another question (it is.)

Alex said that the runtime check sort of defeats the idea of not clearing the buffers in the first place. I don't agree. Clearing the buffers is much heavier than the conditional which when wrapped in G_UNLIKELY will be essentially free.

# posted by

behdad : April 21, 2006 7:16 AM

I looked at the code and I changed my opinion. I remember the code as having scatted one-word ifdefed pad-writes all over the place, but now its just in one place, and it has a runtime check for it already (see giop_send_buffer_init).

# posted by

Anonymous : April 21, 2006 7:30 AM

Alex, Yes, I saw that check too, but it's not the only place. Check p_memwipe and p_memzero (p_free is irrelevant here.)

# posted by

behdad : April 21, 2006 7:32 AM

About Me

Twitter Updates