[PATCH] malloc: Improve checked_request2size

[Wilco Dijkstra]

Hi DJ,
 
> Wilco Dijkstra <Wilco.Dijkstra@arm.com> writes:
>> Change checked_request2size to return PTRDIFF_T+1 for huge inputs.  This
>> ensures large allocation request stay large and can't be confused with a
>> small allocation.
>
> Every call to checked_request2size() immediately checks for 0 and
> returns NULL if so.  I don't see an opportunity for confusion here, 

This is currently used in eg. __libc_calloc:

# define usize2tidx(x) csize2tidx (checked_request2size (x))

  size_t tc_idx = usize2tidx (bytes);
  if (__glibc_likely (tc_idx < mp_.tcache_bins))
     ... tcache code

Since csize2tidx does (x) - MINSIZE, we get 0 - MINSIZE which is huge,
so a huge input implies a huge index, and all is fine.

However if you change it to do this (from Cupertino's patch):

-  size_t tc_idx = usize2tidx (bytes);
-  if (__glibc_likely (tc_idx < mp_.tcache_bins))
+  size_t nb = checked_request2size (bytes);
+  if (nb < mp_.tcache_max_bytes)

Now a huge size is changed into nb = 0 and then 0 < mp_.tcache_max_bytes
means we enter the tcache code and treat it like a valid size...

So my patch avoids this confusion by keeping huge sizes always huge.

> comparing with 0 is far more efficient than comparing against a random
> 64-bit constant.

SIZE_MAX does the job too, so I'll do that in v2.

Cheers,
Wilco