[0/5] Annotate allocation functions with alloc_size attribute

Message

Daniel Axtens Jan. 20, 2020, 7:43 a.m. UTC

Both gcc and clang support the 'alloc_size' function attribute. It tells
the compiler that a function returns a pointer to a certain amount of
memory.

This series tries applying that attribute to a number of our memory
allocation functions. This provides much more information to things that
use __builtin_object_size() (FORTIFY_SOURCE and some copy_to/from_user
stuff), as well as enhancing inlining opportunities where __builtin_mem* or
__builtin_str* are used.

With this series, FORTIFY_SOURCE picks up a bug in altera-stapl, which is
fixed in patch 1.

It also generates a bunch of warnings about times memory allocation
functions can be called with SIZE_MAX as the parameter. For example, from
patch 3:

drivers/staging/rts5208/rtsx_chip.c: In function ‘rtsx_write_cfg_seq’:
drivers/staging/rts5208/rtsx_chip.c:1453:7: warning: argument 1 value ‘18446744073709551615’ exceeds maximum object size 9223372036854775807 [-Walloc-size-larger-than=]
  data = vzalloc(array_size(dw_len, 4));
  ~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

The parameter to array_size is a size_t, but it is called with a signed
integer argument. If the argument is a negative integer, it will become a
very large positive number when cast to size_t. This could cause an
overflow, so array_size() will return SIZE_MAX _at compile time_. gcc then
notices that this value is too large for an allocation and throws a
warning.

I propose two ways to deal with this:

 - Individually go through and address these warnings, usualy by
   catching when struct_size/array_size etc are called with a signed
   type, and insert bounds checks or change the type where
   appropriate. Patch 3 is an example.

 - Patch 4: make kmalloc(_node) catch SIZE_MAX and return NULL early,
   preventing an annotated kmalloc-family allocation function from seeing
   SIZE_MAX as a parameter.

I'm not sure whether I like the idea of catching SIZE_MAX in the inlined
functions. Here are some pros and cons, and I'd be really interested to
hear feedback:

 * Making kmalloc return NULL early doesn't change _runtime_ behaviour:
   obviously no SIZE_MAX allocation will ever succeed. And it means we
   could have this feature earlier, which will help to catch issues like
   what we catch in altera-stapl.

 * However, it does mean we don't audit callsites where perhaps we should
   have stricter types or bounds-checking. It also doesn't cover any of the
   v*alloc functions.

Overall I think this is a meaningful strengthening of FORTIFY_SOURCE
and worth pursuing.

Daniel Axtens (5):
  altera-stapl: altera_get_note: prevent write beyond end of 'key'
  [RFC] kasan: kasan_test: hide allocation sizes from the compiler
  [RFC] staging: rts5208: make len a u16 in rtsx_write_cfg_seq
  [VERY RFC] mm: kmalloc(_node): return NULL immediately for SIZE_MAX
  [RFC] mm: annotate memory allocation functions with their sizes

 drivers/misc/altera-stapl/altera.c  | 12 ++++----
 drivers/staging/rts5208/rtsx_chip.c |  2 +-
 drivers/staging/rts5208/rtsx_chip.h |  2 +-
 include/linux/compiler_attributes.h |  6 ++++
 include/linux/kasan.h               | 12 ++++----
 include/linux/slab.h                | 44 +++++++++++++++++---------
 include/linux/vmalloc.h             | 26 ++++++++--------
 lib/test_kasan.c                    | 48 +++++++++++++++++++++--------
 8 files changed, 98 insertions(+), 54 deletions(-)