[v8,0/9] crypto: Remove VLA usage

Message

Kees Cook Aug. 7, 2018, 9:18 p.m. UTC

v8 cover letter:

I continue to hope this can land in v4.19, but I realize that's unlikely.
It would be nice, though, if some of the "trivial" patches could get taken
(e.g. cbc, xcbc, ccm VLA removals) so I don't have to keep repeating them.
*fingers crossed*

Series cover letter:

This is nearly the last of the VLA removals[1], but it's one of the
largest because crypto gets used in lots of places. After looking
through code, usage, reading the threads Gustavo started, and comparing
the use-cases to the other VLA removals that have landed in the kernel,
I think this series is likely the best way forward to shut the door on
VLAs forever.

For background, the crypto stack usage is for callers to do an immediate
bit of work that doesn't allocate new memory. This means that other VLA
removal techniques (like just using kmalloc) aren't workable, and the
next common technique is needed: examination of maximum stack usage and
the addition of sanity checks. This series does that, and in several
cases, these maximums were already implicit in the code.

This series is intended to land via the crypto tree for 4.19, though it
touches dm, networking, and a few other things as well, since there are
dependent patches (new crypto #defines being used, etc).

Thanks!

-Kees

[1] https://lkml.kernel.org/r/CA+55aFzCG-zNmZwX4A2FQpadafLfEzK6CC=[e-mail]

Changelog:

v8:
- remove "impossible condition" BUG-check in cbc.

v7:
- refresh and reorganize without ahash->shash conversions
- merge hash maximums for both shash and ahash

v6:
- make xcbc blocksize unconditional
- add ahash-to-shash conversion patch series to entirely remove
  AHASH_REQUEST_ON_STACK from the kernel

v5:
- limit AHASH_REQUEST_ON_STACK size only to non-async hash wrapping.
- sanity-check ahash reqsize only when doing shash wrapping.
- remove frame_warn changes in favor of shash conversions and other fixes.
- send ahash to shash conversion patches and other fixes separately.

v4:
- add back *_REQUEST_ON_STACK patches.
- programmatically find stack sizes for *_REQUEST_ON_STACK patches.
- whitelist some code that trips FRAME_WARN on 32-bit builds.
- fix alignment patches.

v3:
- drop *_REQUEST_ON_STACK patches. The rest of this series is pretty
  straight-forward, and I'd like to get them tested in -next while
  we continue to chip away at the *_REQUEST_ON_STACK VLA removal patches
  separately. "Part 2" will continue with those.

v2:
- use 512 instead of PAGE_SIZE / 8 to avoid bloat on large-page archs.
- swtich xcbc to "16" max universally.
- fix typo in bounds check for dm buffer size.
- examine actual reqsizes for skcipher and ahash instead of guessing.
- improve names and comments for alg maxes





Ard Biesheuvel (1):
  crypto: ccm: Remove VLA usage

Kees Cook (8):
  crypto: xcbc: Remove VLA usage
  crypto: cbc: Remove VLA usage
  crypto: hash: Remove VLA usage
  dm: Remove VLA usage from hashes
  crypto alg: Introduce generic max blocksize and alignmask
  crypto: qat: Remove VLA usage
  crypto: shash: Remove VLA usage in unaligned hashing
  crypto: skcipher: Remove VLA usage for SKCIPHER_REQUEST_ON_STACK

 crypto/ahash.c                           |  4 +--
 crypto/algapi.c                          |  7 ++++-
 crypto/algif_hash.c                      |  2 +-
 crypto/ccm.c                             |  9 ++++---
 crypto/shash.c                           | 33 ++++++++++++++----------
 crypto/xcbc.c                            |  8 +++---
 drivers/crypto/qat/qat_common/qat_algs.c |  8 ++++--
 drivers/md/dm-integrity.c                | 23 ++++++++++++-----
 drivers/md/dm-verity-fec.c               |  5 +++-
 include/crypto/algapi.h                  |  4 ++-
 include/crypto/cbc.h                     |  2 +-
 include/crypto/hash.h                    |  6 ++++-
 include/crypto/internal/skcipher.h       |  1 +
 include/crypto/skcipher.h                |  4 ++-
 include/linux/compiler-gcc.h             |  1 -
 15 files changed, 79 insertions(+), 38 deletions(-)

Comments

Herbert Xu Sept. 4, 2018, 5:19 a.m. UTC | #1

On Tue, Aug 07, 2018 at 02:18:34PM -0700, Kees Cook wrote:
> v8 cover letter:
> 
> I continue to hope this can land in v4.19, but I realize that's unlikely.
> It would be nice, though, if some of the "trivial" patches could get taken
> (e.g. cbc, xcbc, ccm VLA removals) so I don't have to keep repeating them.
> *fingers crossed*
> 
> Series cover letter:
> 
> This is nearly the last of the VLA removals[1], but it's one of the
> largest because crypto gets used in lots of places. After looking
> through code, usage, reading the threads Gustavo started, and comparing
> the use-cases to the other VLA removals that have landed in the kernel,
> I think this series is likely the best way forward to shut the door on
> VLAs forever.
> 
> For background, the crypto stack usage is for callers to do an immediate
> bit of work that doesn't allocate new memory. This means that other VLA
> removal techniques (like just using kmalloc) aren't workable, and the
> next common technique is needed: examination of maximum stack usage and
> the addition of sanity checks. This series does that, and in several
> cases, these maximums were already implicit in the code.
> 
> This series is intended to land via the crypto tree for 4.19, though it
> touches dm, networking, and a few other things as well, since there are
> dependent patches (new crypto #defines being used, etc).

I have applied patches 1-4 and 6-8.  I'd like to get an ack from
the dm folks regarding patch 5.  As to patch 9, please fix it so
it doesn't rely on the BUG_ON to catch things.

Thanks,

Kees Cook Sept. 4, 2018, 5:50 a.m. UTC | #2

On Mon, Sep 3, 2018 at 10:19 PM, Herbert Xu <[e-mail]> wrote:
> On Tue, Aug 07, 2018 at 02:18:34PM -0700, Kees Cook wrote:
>> v8 cover letter:
>>
>> I continue to hope this can land in v4.19, but I realize that's unlikely.
>> It would be nice, though, if some of the "trivial" patches could get taken
>> (e.g. cbc, xcbc, ccm VLA removals) so I don't have to keep repeating them.
>> *fingers crossed*
>>
>> Series cover letter:
>>
>> This is nearly the last of the VLA removals[1], but it's one of the
>> largest because crypto gets used in lots of places. After looking
>> through code, usage, reading the threads Gustavo started, and comparing
>> the use-cases to the other VLA removals that have landed in the kernel,
>> I think this series is likely the best way forward to shut the door on
>> VLAs forever.
>>
>> For background, the crypto stack usage is for callers to do an immediate
>> bit of work that doesn't allocate new memory. This means that other VLA
>> removal techniques (like just using kmalloc) aren't workable, and the
>> next common technique is needed: examination of maximum stack usage and
>> the addition of sanity checks. This series does that, and in several
>> cases, these maximums were already implicit in the code.
>>
>> This series is intended to land via the crypto tree for 4.19, though it
>> touches dm, networking, and a few other things as well, since there are
>> dependent patches (new crypto #defines being used, etc).
>
> I have applied patches 1-4 and 6-8.  I'd like to get an ack from
> the dm folks regarding patch 5.  As to patch 9, please fix it so
> it doesn't rely on the BUG_ON to catch things.

Great! Thanks very much. I'll get a patch prepared to plumb
crypto_skcipher_set_reqsize() failures.

-Kees