memalign() does not work as expected

[Vojtech Horky]

This problem was discussed on ML (​initial post), information here is rather a summary.

In current mainline (1302), memalign() fails to satisfy a request even when there is evidently enough memory to satisfy it. The attached patch adds a memalign program to HelenOS that tries to allocate blocks of growing size using malloc() and memalign().

for (i = 0; i < 30; i++) {
   void *ptr = malloc(1 << i); // or memalign(1 << i, 32)
   // if ptr != NULL, access it and free it
   // print i if allocation was successful
}

The expected behavior is that both functions would be able to allocate blocks of (more or less) similar sizes. For example, when malloc() is able to allocate 8M, memalign() shall be able to allocate at least 4M.

However, as the attached screenshots prove, memalign() starts to return NULL much earlier than malloc() and moreover, this "upper boundary" is not consistent through repeated runs.

It is apparent that this is some problem inside libc memory allocator but the root cause is not known. One possible suspect is rounding/aligning inside allocation routines. The second patch fixes the memalign application somehow but I am not sure whether it really fixes the root cause and whether the rounding is not an overkill.

[Vojtech Horky]

memalign application demonstrating the problem

[Vojtech Horky]

Screenshot of unexpected behavior

[Vojtech Horky]

Screenshot with "fixed" allocator (i.e. expected behavior)

[Vojtech Horky]

Hint for fixing the memory allocator - reviewer wanted

[Jakub Jermář]

No need to have this fixed for 0.5.0, rescheduling for the next release.

[Jakub Jermář]

Fixed in mainline,1553.