lib: lmb: fix overflow in __lmb_alloc_base w/ large RAM

If a 32-bit system has 2GB of RAM, and the base address of that RAM is
2GB, then start+size will overflow a 32-bit value (to a value of 0).

__lmb_alloc_base is affected by this; it calculates the minimum of
(start+size of RAM) and max_addr. However, when start+size is 0, it
is always less than max_addr, which causes the value of max_addr not
to be taken into account when restricting the allocation's location.

Fix this by calculating start+size separately, and if that calculation
underflows, using -1 (interpreted as the max unsigned value) as the
value instead, and then taking the min of that and max_addr. Now that
start+size doesn't overflow, it's typically large, and max_addr
dominates the min() call, and is taken into account.

The user-visible symptom of this bug is that CONFIG_BOOTMAP_SZ is ignored
on Tegra124 systems with 2GB of RAM, which in turn causes the DT to be
relocated at the very end of RAM, which the ARM Linux kernel doesn't map
during early boot, and which causes boot failures. With this fix,
CONFIG_BOOTMAP_SZ correctly restricts the relocated DT to a much lower
address, and everything works.

Signed-off-by: Stephen Warren <swarren@nvidia.com>

lib/lmb.c

@@ -295,7 +295,10 @@ phys_addr_t __lmb_alloc_base(struct lmb *lmb, phys_size_t size, ulong align, phy
 		if (max_addr == LMB_ALLOC_ANYWHERE)
 			base = lmb_align_down(lmbbase + lmbsize - size, align);
 		else if (lmbbase < max_addr) {
-			base = min(lmbbase + lmbsize, max_addr);
+			base = lmbbase + lmbsize;
+			if (base < lmbbase)
+				base = -1;
+			base = min(base, max_addr);
 			base = lmb_align_down(base - size, align);
 		} else
 			continue;