Use of memalign can trigger fragmentation issues such as:

// Internally, this needs to find a free block quite bit larger than s.
// Once the free region is found, any unaligned "padding" immediately
// before and after the block is marked free, so that the allocation
// takes only s bytes (plus malloc header overhead).
p = memalign(a, s);
// If there's little fragmentation so far, this allocation is likely
// located immediately after p.
p2 = malloc(x);
free(p);
// In theory, this should return the same value for p. However, the hole
// left by the free() call is only s in size (plus malloc header overhead)
// whereas memalign searches for a larger block in order to guarantee it
// can adjust the returned pointer to the alignment requirements. Hence,
// the pointer returned, if any, won't be p. If there's little or no space
// left after p2, this allocation will fail.
p = memalign(a, s);

In practice, this issue occurs when running the "dfu" command repeatedly
on NVIDIA Tegra boards, since DFU allocates a large 32M data buffer, and
then initializes the USB controller. If this is the first time USB has
been used in the U-Boot session, this causes a probe of the USB driver,
which causes various allocations, including a strdup() of a GPIO name
when requesting the VBUS GPIO. When DFU is torn down, the USB driver
is left probed, and hence its memory is left allocated. If "dfu" is
executed again, allocation of the 32M data buffer fails as described
above.

In practice, there is a memory hole exactly large enough to hold the 32M
data buffer than DFU needs. However, memalign() can't know that in a
general way. Given that, it's particularly annoying that the allocation
fails!

The issue is that memalign() tries to allocate something larger to
guarantee the ability to align the returned pointer. This patch modifies
memalign() so that if the "general case" over-sized allocation fails,
another allocation is attempted, of the exact size the user desired. If
that allocation just happens to be aligned in the way the user wants,
(and in the case described above, it will be, since the free memory
region is located where a previous identical allocation was located),
the pointer can be returned.

This patch is somewhat related to 806bd245

 "dfu: don't keep
freeing/reallocating". That patch worked around the issue by removing
repeated free/memalign within a single execution of "dfu". However,
the same technique can't be applied across multiple invocations, since
there's no reason to keep the DFU buffer allocated while DFU isn't
running. This patch addresses the root-cause a bit more directly.

This problem highlights some of the disadvantages of dynamic allocation
and deferred probing of devices.

This patch isn't checkpatch-clean, since it conforms to the existing
coding style in dlmalloc.c, which is different to the rest of U-Boot.

Signed-off-by: Stephen Warren <swarren@nvidia.com>
Reviewed-by: Tom Rini <trini@konsulko.com>
Acked-by: Lukasz Majewski <l.majewski@samsung.com>