STRICT_ALIGNMENT is a remnant of OpenSSL code would cast pointers to
size_t* and load more than one byte at a time. Not all architectures
support unaligned access, so it did an alignment check and only enterred
this path if aligned or the underlying architecture didn't care.
This is UB. Unaligned casts in C are undefined on all architectures, so
we switch these to memcpy some time ago. Compilers can optimize memcpy
to the unaligned accesses we wanted. That left our modes logic as:
- If STRICT_ALIGNMENT is 1 and things are unaligned, work byte-by-byte.
- Otherwise, use the memcpy-based word-by-word code, which now works
independent of STRICT_ALIGNMENT.
Remove the first check to simplify things. On x86, x86_64, and aarch64,
STRICT_ALIGNMENT is zero and this is a no-op. ARM is more complex. Per
[0], ARMv7 and up support unaligned access. ARMv5 do not. ARMv6 does,
but can run in a mode where it looks more like ARMv5.
For ARMv7 and up, STRICT_ALIGNMENT should have been zero, but was one.
Thus this change should be an improvement for ARMv7 (right now unaligned
inputs lose bsaes-armv7). The Android NDK does not even support the
pre-ARMv7 ABI anymore[1]. Nonetheless, Cronet still supports ARMv6 as a
library. It builds with -march=armv6 which GCC interprets as supporting
unaligned access, so it too did not want this code.
For completeness, should anyone still care about ARMv5 or be building
with an overly permissive -march flag, GCC does appear unable to inline
the memcpy calls. However, GCC also does not interpret
(uintptr_t)ptr % sizeof(size_t) as an alignment assertion, so such
consumers have already been paying for the memcpy here and throughout
the library.
In general, C's arcane pointer rules mean we must resort to memcpy
often, so, realistically, we must require that the compiler optimize
memcpy well.
[0] https://medium.com/@iLevex/the-curious-case-of-unaligned-access-on-arm-5dd0ebe24965
[1] https://developer.android.com/ndk/guides/abis#armeabi
Change-Id: I3c7dea562adaeb663032e395499e69530dd8e145
Reviewed-on: https://boringssl-review.googlesource.com/c/34873
Reviewed-by: Adam Langley <agl@google.com>
There is a C implementation of gcm_ghash_4bit to pair with
gcm_gmult_4bit. It's even slightly faster per the numbers below (x86_64
OPENSSL_NO_ASM build), but, more importantly, we trim down the
combinatorial explosion of GCM implementations and free up complexity
budget for potentially using bsaes better in the future.
Old:
Did 2557000 AES-128-GCM (16 bytes) seal operations in 1000057us (2556854.3 ops/sec): 40.9 MB/s
Did 94000 AES-128-GCM (1350 bytes) seal operations in 1009613us (93105.0 ops/sec): 125.7 MB/s
Did 17000 AES-128-GCM (8192 bytes) seal operations in 1024768us (16589.1 ops/sec): 135.9 MB/s
Did 2511000 AES-256-GCM (16 bytes) seal operations in 1000196us (2510507.9 ops/sec): 40.2 MB/s
Did 84000 AES-256-GCM (1350 bytes) seal operations in 1000412us (83965.4 ops/sec): 113.4 MB/s
Did 15000 AES-256-GCM (8192 bytes) seal operations in 1046963us (14327.2 ops/sec): 117.4 MB/s
New:
Did 2739000 AES-128-GCM (16 bytes) seal operations in 1000322us (2738118.3 ops/sec): 43.8 MB/s
Did 100000 AES-128-GCM (1350 bytes) seal operations in 1008190us (99187.7 ops/sec): 133.9 MB/s
Did 17000 AES-128-GCM (8192 bytes) seal operations in 1006360us (16892.6 ops/sec): 138.4 MB/s
Did 2546000 AES-256-GCM (16 bytes) seal operations in 1000150us (2545618.2 ops/sec): 40.7 MB/s
Did 86000 AES-256-GCM (1350 bytes) seal operations in 1000970us (85916.7 ops/sec): 116.0 MB/s
Did 14850 AES-256-GCM (8192 bytes) seal operations in 1023459us (14509.6 ops/sec): 118.9 MB/s
While I'm here, tighten up some of the functions and align the ctr32 and
non-ctr32 paths.
Bug: 256
Change-Id: Id4df699cefc8630dd5a350d44f927900340f5e60
Reviewed-on: https://boringssl-review.googlesource.com/c/34869
Reviewed-by: Adam Langley <agl@google.com>
But for the ABI conversion bits, these are just leaf functions and don't
even need unwind tables. Just renumber the registers on Windows to only
used volatile ones.
In doing so, this switches to writing rdrand explicitly. perlasm already
knows how to manually encode it and our minimum assembler versions
surely cover rdrand by now anyway. Also add the .size directive. I'm not
sure what it's used for, but the other files have it.
(This isn't a generally reusable technique. The more complex functions
will need actual unwind codes.)
Bug: 259
Change-Id: I1d5669bcf8b6e34939885d78aea6f60597be1528
Reviewed-on: https://boringssl-review.googlesource.com/c/34867
Commit-Queue: Adam Langley <agl@google.com>
Reviewed-by: Adam Langley <agl@google.com>
This shrinks the bssl binary by about 8k.
Change-Id: I571f258ccf7032ae34db3f20904ad9cc81cca839
Reviewed-on: https://boringssl-review.googlesource.com/c/34866
Commit-Queue: David Benjamin <davidben@google.com>
Reviewed-by: Adam Langley <agl@google.com>
Change-Id: Ic23fc5fbec2c4f8df5d06f807c6bd2c5e1f0e99c
Reviewed-on: https://boringssl-review.googlesource.com/c/34865
Commit-Queue: David Benjamin <davidben@google.com>
Reviewed-by: Adam Langley <agl@google.com>
Fix some missing CFI bits.
Change-Id: I42114527f0ef8e03079d37a9f466d64a63a313f5
Reviewed-on: https://boringssl-review.googlesource.com/c/34864
Commit-Queue: David Benjamin <davidben@google.com>
Reviewed-by: Adam Langley <agl@google.com>
This was caught by an aarch64 ABI tester. aarch64 has the same
considerations around small arguments as x86_64 does. The aarch64
version of bn_mul_mont does not mask off the upper words of the
argument.
The x86_64 version does, so size_t is, strictly speaking, wrong for
aarch64, but bn_mul_mont already has an implicit size limit to support
its internal alloca, so this doesn't really make things worse than
before.
Change-Id: I39bffc8fdb2287e45a2d1f0d1b4bd5532bbf3868
Reviewed-on: https://boringssl-review.googlesource.com/c/34804
Commit-Queue: David Benjamin <davidben@google.com>
Reviewed-by: Adam Langley <agl@google.com>
Update-Note: There's some chance this'll break iOS since I was unable to
test it there. The iPad I have to test on is too new to run 32-bit code
at all.
Change-Id: I6593f91b67a5e8a82828237d3b69ed948b07922d
Reviewed-on: https://boringssl-review.googlesource.com/c/34725
Commit-Queue: David Benjamin <davidben@google.com>
Reviewed-by: Adam Langley <agl@google.com>
Unfortunately, due to most OpenSSL assembly using custom exception
handlers to unwind, most of our assembly doesn't work with
non-destructive unwind. For now, CHECK_ABI behaves like
CHECK_ABI_NO_UNWIND on Windows, and CHECK_ABI_SEH will test unwinding on
both platforms.
The tests do, however, work with the unwind-code-based assembly we
recently added, as well as the clmul-based GHASH which is also
code-based. Remove the ad-hoc SEH tests which intentionally hit memory
access exceptions, now that we can test unwind directly.
Now that we can test it, the next step is to implement SEH directives in
perlasm so writing these unwind codes is less of a chore.
Bug: 259
Change-Id: I23a57a22c5dc9fa4513f575f18192335779678a5
Reviewed-on: https://boringssl-review.googlesource.com/c/34784
Commit-Queue: David Benjamin <davidben@google.com>
Reviewed-by: Adam Langley <agl@google.com>
RAND_bytes rarely uses large enough inputs for bsaes to be worth it.
https://boringssl-review.googlesource.com/c/boringssl/+/33589 includes some
rough benchmarks of various bits here. Some observations:
- 8 blocks of bsaes costs roughly 6.5 blocks of vpaes. Note the comparison
isn't quite accurate because I'm measuring bsaes_ctr32_encrypt_blocks against
vpaes_encrypt and vpaes in CTR mode today must make do with a C loop. Even
assuming a cutoff of 6 rather than 7 blocks, it's rare to ask for 96 bytes
of entropy at a time.
- CTR-DRBG performs some stray block operations (ctr_drbg_update), which bsaes
is bad at without extra work to fold them into the CTR loop (not really worth
it).
- CTR-DRBG calculates a couple new key schedules every RAND_bytes call. We
don't currently have a constant-time bsaes key schedule. Unfortunately, even
plain vpaes loses to the current aes_nohw used by bsaes, but it's not
constant-time. Also taking CTR-DRBG out of the bsaes equation
- Machines without AES hardware (clients) are not going to be RNG-bound. It's
mostly servers pushing way too many CBC IVs that care. This means bsaes's
current side channel tradeoffs make even less sense here.
I'm not sure yet what we should do for the rest of the bsaes mess, but it seems
clear that we want to stick with vpaes for the RNG.
Bug: 256
Change-Id: Iec8f13af232794afd007cb1065913e8117eeee24
Reviewed-on: https://boringssl-review.googlesource.com/c/34744
Reviewed-by: Adam Langley <agl@google.com>
Calling conventions must specify how to handle arguments smaller than a
machine word. Should the caller pad them up to a machine word size with
predictable values (zero/sign-extended), or should the callee tolerate
an arbitrary bit pattern?
Annoyingly, I found no text in either SysV or Win64 ABI documentation
describing any of this and resorted to experiment. The short answer is
that callees must tolerate an arbitrary bit pattern on x86_64, which
means we must test this. See the comment in abi_test::internal::ToWord
for the long answer.
CHECK_ABI now, if the type of the parameter is smaller than
crypto_word_t, fills the remaining bytes with 0xaa. This is so the
number is out of bounds for code expecting either zero or sign
extension. (Not that crypto assembly has any business seeing negative
numbers.)
Doing so reveals a bug in ecp_nistz256_ord_sqr_mont. The rep parameter
is typed int, but the code expected uint64_t. In practice, the compiler
will always compile this correctly because:
- On both Win64 and SysV, rep is a register parameter.
- The rep parameter is always a constant, so the compiler has no reason
to leave garbage in the upper half.
However, I was indeed able to get a bug out of GCC via:
uint64_t foo = (1ull << 63) | 2; // Some global the compiler can't
// prove constant.
ecp_nistz256_ord_sqr_mont(res, a, foo >> 1);
Were ecp_nistz256_ord_sqr_mont a true int-taking function, this would
act like ecp_nistz256_ord_sqr_mont(res, a, 1). Instead, it hung. Fix
this by having it take a full-width word.
This mess has several consequences:
- ABI testing now ideally needs a functional testing component to fully cover
this case. A bad input might merely produce the wrong answer. Still,
this is fairly effective as it will cause most code to either segfault
or loop forever. (Not the enc parameter to AES however...)
- We cannot freely change the type of assembly function prototypes. If the
prototype says int or unsigned, it must be ignoring the upper half and
thus "fixing" it to size_t cannot have handled the full range. (Unless
it was simply wrong of the parameter is already bounded.) If the
prototype says size_t, switching to int or unsigned will hit this type
of bug. The former is a safer failure mode though.
- The simplest path out of this mess: new assembly code should *only*
ever take word-sized parameters. This is not a tall order as the bad
parameters are usually ints that should have been size_t.
Calling conventions are hard.
Change-Id: If8254aff8953844679fbce4bd3e345e5e2fa5213
Reviewed-on: https://boringssl-review.googlesource.com/c/34627
Commit-Queue: David Benjamin <davidben@google.com>
Reviewed-by: Adam Langley <agl@google.com>
As part of this, move the CPU checks to C.
Change-Id: I17b701e1196c1ca116bbd23e0e669cf603ad464d
Reviewed-on: https://boringssl-review.googlesource.com/c/34626
Reviewed-by: Adam Langley <agl@google.com>
Commit-Queue: David Benjamin <davidben@google.com>
It's an assembly function, so types are a little meaningless, but
everything is passed through as BN_ULONG, so be consistent. Also
annotate all the RSAZ prototypes with sizes.
Change-Id: I32e59e896da39e79c30ce9db52652fd645a033b4
Reviewed-on: https://boringssl-review.googlesource.com/c/34625
Reviewed-by: Adam Langley <agl@google.com>
Commit-Queue: David Benjamin <davidben@google.com>
This is much less interesting (stack-based parameters, Windows and SysV
match, no SEH concerns as far as I can tell) than x86_64, but it was
easy to do and I'm more familiar with x86 than ARM, so it made a better
second architecture to make sure all the architecture ifdefs worked out.
Also fix a bug in the x86_64 direction flag code. It was shifting in the
wrong direction, making give 0 or 1<<20 rather than 0 or 1.
(Happily, x86_64 appears to be unique in having vastly different calling
conventions between OSs. x86 is the same between SysV and Windows, and
ARM had the good sense to specify a (mostly) common set of rules.)
Since a lot of the assembly functions use the same names and the tests
were written generically, merely dropping in a trampoline and
CallerState implementation gives us a bunch of ABI tests for free.
Change-Id: I15408c18d43e88cfa1c5c0634a8b268a150ed961
Reviewed-on: https://boringssl-review.googlesource.com/c/34624
Reviewed-by: Adam Langley <agl@google.com>
Commit-Queue: David Benjamin <davidben@google.com>
This doesn't cover all the functions used by Node, but it's the easy
bits. (EVP_PKEY_paramgen will be done separately as its a non-trivial
bit of machinery.)
Change-Id: I6501e99f9239ffcdcc57b961ebe85d0ad3965549
Reviewed-on: https://boringssl-review.googlesource.com/c/34544
Reviewed-by: Adam Langley <agl@google.com>
Commit-Queue: Adam Langley <agl@google.com>
We currently require clmul instructions for constant-time GHASH
on x86_64. Otherwise, it falls back to a variable-time 4-bit table
implementation. However, a significant proportion of clients lack these
instructions.
Inspired by vpaes, we can use pshufb and a slightly different order of
incorporating the bits to make a constant-time GHASH. This requires
SSSE3, which is very common. Benchmarking old machines we had on hand,
it appears to be a no-op on Sandy Bridge and a small slowdown for
Penryn.
Sandy Bridge (Intel Pentium CPU 987 @ 1.50GHz):
(Note: these numbers are before 16-byte-aligning the table. That was an
improvement on Penryn, so it's possible Sandy Bridge is now better.)
Before:
Did 4244750 AES-128-GCM (16 bytes) seal operations in 4015000us (1057222.9 ops/sec): 16.9 MB/s
Did 442000 AES-128-GCM (1350 bytes) seal operations in 4016000us (110059.8 ops/sec): 148.6 MB/s
Did 84000 AES-128-GCM (8192 bytes) seal operations in 4015000us (20921.5 ops/sec): 171.4 MB/s
Did 3349250 AES-256-GCM (16 bytes) seal operations in 4016000us (833976.6 ops/sec): 13.3 MB/s
Did 343500 AES-256-GCM (1350 bytes) seal operations in 4016000us (85532.9 ops/sec): 115.5 MB/s
Did 65250 AES-256-GCM (8192 bytes) seal operations in 4015000us (16251.6 ops/sec): 133.1 MB/s
After:
Did 4229250 AES-128-GCM (16 bytes) seal operations in 4016000us (1053100.1 ops/sec): 16.8 MB/s [-0.4%]
Did 442250 AES-128-GCM (1350 bytes) seal operations in 4016000us (110122.0 ops/sec): 148.7 MB/s [+0.1%]
Did 83500 AES-128-GCM (8192 bytes) seal operations in 4015000us (20797.0 ops/sec): 170.4 MB/s [-0.6%]
Did 3286500 AES-256-GCM (16 bytes) seal operations in 4016000us (818351.6 ops/sec): 13.1 MB/s [-1.9%]
Did 342750 AES-256-GCM (1350 bytes) seal operations in 4015000us (85367.4 ops/sec): 115.2 MB/s [-0.2%]
Did 65250 AES-256-GCM (8192 bytes) seal operations in 4016000us (16247.5 ops/sec): 133.1 MB/s [-0.0%]
Penryn (Intel Core 2 Duo CPU P8600 @ 2.40GHz):
Before:
Did 1179000 AES-128-GCM (16 bytes) seal operations in 1000139us (1178836.1 ops/sec): 18.9 MB/s
Did 97000 AES-128-GCM (1350 bytes) seal operations in 1006347us (96388.2 ops/sec): 130.1 MB/s
Did 18000 AES-128-GCM (8192 bytes) seal operations in 1028943us (17493.7 ops/sec): 143.3 MB/s
Did 977000 AES-256-GCM (16 bytes) seal operations in 1000197us (976807.6 ops/sec): 15.6 MB/s
Did 82000 AES-256-GCM (1350 bytes) seal operations in 1012434us (80992.9 ops/sec): 109.3 MB/s
Did 15000 AES-256-GCM (8192 bytes) seal operations in 1006528us (14902.7 ops/sec): 122.1 MB/s
After:
Did 1306000 AES-128-GCM (16 bytes) seal operations in 1000153us (1305800.2 ops/sec): 20.9 MB/s [+10.8%]
Did 94000 AES-128-GCM (1350 bytes) seal operations in 1009852us (93082.9 ops/sec): 125.7 MB/s [-3.4%]
Did 17000 AES-128-GCM (8192 bytes) seal operations in 1012096us (16796.8 ops/sec): 137.6 MB/s [-4.0%]
Did 1070000 AES-256-GCM (16 bytes) seal operations in 1000929us (1069006.9 ops/sec): 17.1 MB/s [+9.4%]
Did 79000 AES-256-GCM (1350 bytes) seal operations in 1002209us (78825.9 ops/sec): 106.4 MB/s [-2.7%]
Did 15000 AES-256-GCM (8192 bytes) seal operations in 1061489us (14131.1 ops/sec): 115.8 MB/s [-5.2%]
Change-Id: I1c3760a77af7bee4aee3745d1c648d9e34594afb
Reviewed-on: https://boringssl-review.googlesource.com/c/34267
Commit-Queue: David Benjamin <davidben@google.com>
Reviewed-by: Adam Langley <agl@google.com>
With 2fe0360a4e1b988e7b0aa0b4348bf55805512c09, we no longer use the
other member of this union so it can be removed.
Change-Id: Ideb7c47a72df0b420eb1e7d8c718e1cacb2129f5
Reviewed-on: https://boringssl-review.googlesource.com/c/34449
Commit-Queue: Adam Langley <agl@google.com>
Commit-Queue: David Benjamin <davidben@google.com>
Reviewed-by: David Benjamin <davidben@google.com>
Casting an unaligned pointer to uint64_t* is undefined, even on
platforms that support unaligned access. Additionally, dereferencing as
uint64_t violates strict aliasing rules. Instead, use memcpys which we
assume any sensible compiler can optimize. Also simplify the PULL64
business with the existing CRYPTO_bswap8.
This also removes the need for the
SHA512_BLOCK_CAN_MANAGE_UNALIGNED_DATA logic. The generic C code now
handles unaligned data and the assembly already can as well. (The only
problematic platform with assembly is old ARM, but sha512-armv4.pl
already handles this via an __ARM_ARCH__ check. See also OpenSSL's
version of this file which always defines
SHA512_BLOCK_CAN_MANAGE_UNALIGNED_DATA if SHA512_ASM is defined.)
Add unaligned tests to digest_test.cc, so we retain coverage of
unaligned EVP_MD inputs.
Change-Id: Idfd8586c64bab2a77292af2fa8eebbd193e57c7d
Reviewed-on: https://boringssl-review.googlesource.com/c/34444
Commit-Queue: Adam Langley <agl@google.com>
Reviewed-by: Adam Langley <agl@google.com>
The first attempt involved using Linux's support for hardware
breakpoints to detect when assembly code was run. However, this doesn't
work with SDE, which is a problem.
This version has the assembly code update a global flags variable when
it's run, but only in non-FIPS and non-debug builds.
Update-Note: Assembly files now pay attention to the NDEBUG preprocessor
symbol. Ensure the build passes the symbol in. (If release builds fail
to link due to missing BORINGSSL_function_hit, this is the cause.)
Change-Id: I6b7ced442b7a77d0b4ae148b00c351f68af89a6e
Reviewed-on: https://boringssl-review.googlesource.com/c/33384
Commit-Queue: Adam Langley <agl@google.com>
Reviewed-by: David Benjamin <davidben@google.com>
Change-Id: If28096e677104c6109e31e31a636fee82ef4ba11
Reviewed-on: https://boringssl-review.googlesource.com/c/34266
Commit-Queue: David Benjamin <davidben@google.com>
Reviewed-by: Adam Langley <agl@google.com>
CRYPTO_gcm128_encrypt should be paired with CRYPTO_gcm128_tag, not
CRYPTO_gcm128_finish.
Change-Id: Ia3023a196fe5b613e9309b5bac19ea849dbc33b7
Reviewed-on: https://boringssl-review.googlesource.com/c/34265
Commit-Queue: Adam Langley <agl@google.com>
Reviewed-by: Adam Langley <agl@google.com>
We expect the table to have a slightly nested structure, so just
generate it that way. Avoid risking strict aliasing problems. Thanks to
Brian Smith for pointing this out.
Change-Id: Ie21610c4afab07a610d914265079135dba17b3b7
Reviewed-on: https://boringssl-review.googlesource.com/c/34264
Commit-Queue: Adam Langley <agl@google.com>
Reviewed-by: Adam Langley <agl@google.com>
This involves fixing some bugs in aes_nohw_cbc_encrypt's annotations,
and working around a libunwind bug. In doing so, support .cfi_remember_state
and .cfi_restore_state in perlasm.
Change-Id: Iaedfe691356b0468327a6be0958d034dafa760e5
Reviewed-on: https://boringssl-review.googlesource.com/c/34189
Reviewed-by: Adam Langley <agl@google.com>
Commit-Queue: Adam Langley <agl@google.com>
This is in preparation for adding ABI tests to them.
In doing so, update delocate.go so that OPENSSL_ia32cap_get is consistently
callable outside the module. Right now it's callable both inside and outside
normally, but not in FIPS mode because the function is generated. This is
needed for tests and the module to share headers that touch OPENSSL_ia32cap_P.
Change-Id: Idbc7d694acfb974e0b04adac907dab621e87de62
Reviewed-on: https://boringssl-review.googlesource.com/c/34188
Commit-Queue: Adam Langley <agl@google.com>
Reviewed-by: Adam Langley <agl@google.com>
This does not actually matter, but writing new CFI directives with the
tester seemed like fun. (It caught two typos, one intentional and one
accidental.)
Change-Id: Iff3e0358f2e56caa26079f658fa7a682772150a1
Reviewed-on: https://boringssl-review.googlesource.com/c/34185
Commit-Queue: David Benjamin <davidben@google.com>
Reviewed-by: Adam Langley <agl@google.com>
Bug: 181
Change-Id: Ica9299613d7fd1b803533b7e489b9ba8fe816a24
Reviewed-on: https://boringssl-review.googlesource.com/c/33968
Commit-Queue: David Benjamin <davidben@google.com>
Reviewed-by: Adam Langley <agl@google.com>
Postgres contains a “pqcrypto” module that showcases the worst of 90's
crypto, including Blowfish and CAST5 in CFB, CBC, and ECB modes. (Also,
64-bit keys for both of those.)
In order to minimise the patching needed to build Postgres, put these
things in decrepit.
Change-Id: I8390c5153dd7227eef07293a4363878d79df8b21
Reviewed-on: https://boringssl-review.googlesource.com/c/34044
Reviewed-by: Adam Langley <agl@google.com>
Reviewed-by: David Benjamin <davidben@google.com>
Commit-Queue: Adam Langley <agl@google.com>
Unwind testing will make CHECK_ABI much slower. The original
ptrace-based design is some 10,000x slower. I've found an alternate
design that's a mere 1,000x slower, but this probably warrants being
more straightforward. It also removes the weirdness where NDEBUG
controlled which tests were run.
While it does mean we need to write some extra tests for p256-x86_64.pl,
we otherwise do not directly unit test our assembly anyway. Usually we
test the public crypto APIs themselves. So, for most files, this isn't
actually extra work.
Bug: 181
Change-Id: I7cbb7f930c2ea6ae32a201da503dcd36844704f0
Reviewed-on: https://boringssl-review.googlesource.com/c/33965
Commit-Queue: David Benjamin <davidben@google.com>
Reviewed-by: Adam Langley <agl@google.com>
Prior to 82639e6f we used thread-local data for the PRNG state. That
change switched to using a mutex-protected pool instead in order to save
memory in heavily-threaded applications.
However, the pool mutex can get extremely hot in cases where the PRNG is
heavily used. 8e8f2504 was a short-term work around, but supporting both
modes is overly complex.
This change moves back to the state of the prior to 82639e6f. The best
way to review this is to diff the changed files against '82639e6f^' and
note that the only difference is a comment added in rand.c:
https://paste.googleplex.com/4997991748337664
Change-Id: I8febce089696fa6bc39f94f4a1e268127a8f78db
Reviewed-on: https://boringssl-review.googlesource.com/c/34024
Commit-Queue: David Benjamin <davidben@google.com>
Reviewed-by: David Benjamin <davidben@google.com>
While gdb can figure it out, libunwind requires CFI directives to
unwind a leaf function, even though the directives are trivial.
Adding them matches what GCC outputs, and likely gdb has many
heuristics that less complex tools (e.g. profilers) may not.
Bug: 181
Change-Id: I25c72152de33109a29710a828aeb99c608dd0470
Reviewed-on: https://boringssl-review.googlesource.com/c/33964
Commit-Queue: David Benjamin <davidben@google.com>
Reviewed-by: Adam Langley <agl@google.com>
This was also caught by the in-progress unwind tester. There are two
issues here. First, .cfi_endproc must come after ret to fully cover the
function. More importantly, this function is confused about whether it
has a frame pointer or not.
It looks like it does (movq %rsp, %rbp), and annotates accordingly, but
it does not actually use the frame pointer. It cannot. $y4 is rbp and
gets clobbered immediately after the preamble!
Remove this instruction and align the CFI annotations with a
frame-pointer-less function.
Bug: 181
Change-Id: I47b5f9798b3bcee1748e537b21c173d312a14b42
Reviewed-on: https://boringssl-review.googlesource.com/c/33947
Reviewed-by: Adam Langley <agl@google.com>
Commit-Queue: David Benjamin <davidben@google.com>
This was caught by in-progress work to test unwind information. It was
incorrect at two instructions: immediately before we jump to
.Lpoint_double_shortcut$x. This is needed because
ecp_nistz256_point_add$x tries to be clever about not unwinding the
stack frame in its tail call.
It's also unlikely that the SEH handlers in this file are correct at
this point, but that will be handled separately while overhauling
everything else here. (For Win64, probably the only ABI-compliant option
is to just properly unwind the stack frame. Without a custom handler,
Win64 unwind codes are very restrictive.)
Bug: 181
Change-Id: I9f576d868850312d6c14d1386f8fbfa85021b347
Reviewed-on: https://boringssl-review.googlesource.com/c/33946
Commit-Queue: David Benjamin <davidben@google.com>
Reviewed-by: Adam Langley <agl@google.com>
ecp_nistz256_point_add_affine does not support the doubling case and,
unlike ecp_nistz256_point_add which does a tail call, computes the wrong
answer. Note TestPointAdd in the unit tests skips this case.
This works fine because we only use ecp_nistz256_point_add_affine for
the g_scalar term, which is fully computed before the p_scalar term.
(Additionally it requires that the windowing pattern never hit the
doubling case for single multiplication.)
But this is not obvious from reading the multiplication functions, so
leave a comment at the call site to point this out.
Change-Id: I08882466d98030cdc882a5be9e702ee404e80cce
Reviewed-on: https://boringssl-review.googlesource.com/c/33945
Reviewed-by: Adam Langley <agl@google.com>
Commit-Queue: Adam Langley <agl@google.com>
The old points weren't even on the curve. I probably had no clue what I
was doing at the time when I generated them. Refresh them with a
checked-in generate script.
Change-Id: Ib4613fe922edcf45fc4ea49fc4c2cc23a9a2a9bd
Reviewed-on: https://boringssl-review.googlesource.com/c/33944
Reviewed-by: Adam Langley <agl@google.com>
Commit-Queue: David Benjamin <davidben@google.com>
We switched from thread-local storage to a mutex-pool in 82639e6f53
because, for highly-threaded processes, the memory used by all the
states could be quite large. I had judged that a mutex-pool should be
fine, but had underestimated the PRNG requirements of some of our jobs.
This change makes rand.c support using either thread-locals or a
mutex-pool. Thread-locals are used if fork-unsafe buffering is enabled.
While not strictly related to fork-safety, we already have the
fork-unsafe control, and it's already set by jobs that care a lot about
PRNG performance, so fits quite nicely here.
Change-Id: Iaf1e0171c70d4c8dbe1e42283ea13df5b613cb2d
Reviewed-on: https://boringssl-review.googlesource.com/c/31564
Commit-Queue: Adam Langley <agl@google.com>
Reviewed-by: David Benjamin <davidben@google.com>
Dear reader, I must apologize in advance. This CL contains the following:
- A new 256-line perlasm file with non-trivial perl bits and a dual-ABI
variadic function caller.
- C preprocessor gymnastics, with variadic macros and fun facts about
__VA_ARGS__'s behavior on empty argument lists.
- C++ template gymnastics, including variadic arguments, template
specialization, std::enable_if, and machinery to control template argument
deduction.
Enjoy.
This tests that our assembly functions correctly honor platform ABI
conventions. Right now this only tests callee-saved registers, but it should be
extendable to SEH/CFI unwind testing with single-step debugging APIs.
Register-checking does not involve anything funny and should be compatible with
SDE. (The future unwind testing is unlikely to be compatible.)
This CL adds support for x86_64 SysV and Win64 ABIs. ARM, AArch64, and x86 can
be added in the future. The testing is injected in two places. First, all the
assembly tests in p256-x86_64-test.cc are now instrumented. This is the
intended workflow and should capture all registers.
However, we currently do not unit-test our assembly much directly. We should do
that as follow-up work[0] but, in the meantime, I've also wrapped all of the GTest
main function in an ABI test. This is imperfect as ABI failures may be masked
by other stack frames, but it costs nothing[1] and is pretty reliable at
catching Win64 xmm register failures.
[0] An alternate strategy would be, in debug builds, unconditionally instrument
every assembly call in libcrypto. But the CHECK_ABI macro would be difficult to
replicate in pure C, and unwind testing may be too invasive for this. Still,
something to consider when we C++ libcrypto.
[1] When single-stepped unwind testing exists, it won't cost nothing. The
gtest_main.cc call will turn unwind testing off.
Change-Id: I6643b26445891fd46abfacac52bc024024c8d7f6
Reviewed-on: https://boringssl-review.googlesource.com/c/33764
Reviewed-by: Adam Langley <agl@google.com>
Reviewed-by: Adam Langley <alangley@gmail.com>
Commit-Queue: David Benjamin <davidben@google.com>
Valgrind's checking of uninitialised memory behaves very much like a
check for constant-time code: branches and memory indexes based on
uninitialised memory trigger warnings. Therefore, if we can tell
Valgrind that some secret is “uninitialised”, it'll give us a warning if
we do something non-constant-time with it.
This was the idea behind https://github.com/agl/ctgrind. But tricks like
that are no longer needed because Valgrind now comes with support for
marking regions of memory as defined or not. Therefore we can use that
API to check constant-time code.
This CL defines |CONSTTIME_SECRET| and |CONSTTIME_DECLASSIFY|, which are
no-ops unless the code is built with
|BORINGSSL_CONSTANT_TIME_VALIDATION| defined, which it isn't by default.
So this CL is a no-op itself so far. But it does show that a couple of
bits of constant-time time are, in fact, constant-time—seemingly even
when compiled with optimisations, which is nice.
The annotations in the RSA code are a) probably not marking all the
secrets as secret, and b) triggers warnings that are a little
interesting:
The anti-glitch check calls |BN_mod_exp_mont| which checks that the
input is less than the modulus. Of course, it is because the input is
the RSA plaintext that we just decrypted, but the plaintext is supposed
to be secret and so branching based on its contents isn't allows by
Valgrind. The answer isn't totally clear, but I've run out of time on
this for now.
Change-Id: I1608ed0b22d201e97595fafe46127159e02d5b1b
Reviewed-on: https://boringssl-review.googlesource.com/c/33504
Reviewed-by: Adam Langley <agl@google.com>
Reviewed-by: David Benjamin <davidben@google.com>
Commit-Queue: Adam Langley <agl@google.com>
We don't call it, so ship less code and reduce the number of places
where we must think about the bsaes -> aes_nohw fallback.
Bug: 256
Change-Id: I10ac2d70e18ec81e679631a9532c36d9edab1c6e
Reviewed-on: https://boringssl-review.googlesource.com/c/33586
Reviewed-by: Adam Langley <agl@google.com>
Commit-Queue: Adam Langley <agl@google.com>
This can break delocate with certain compiler settings.
Change-Id: I76cf0f780d0e967390feed754e39b0ab25068f42
Reviewed-on: https://boringssl-review.googlesource.com/c/33485
Commit-Queue: Adam Langley <alangley@gmail.com>
Commit-Queue: David Benjamin <davidben@google.com>
Reviewed-by: David Benjamin <davidben@google.com>
Other XOP code was removed already.
Change-Id: I0c457effebd22f89e722653b93905a0b2e3eb5c0
Reviewed-on: https://boringssl-review.googlesource.com/c/33424
Commit-Queue: David Benjamin <davidben@google.com>
Reviewed-by: David Benjamin <davidben@google.com>
Reviewed-by: Adam Langley <agl@google.com>
Put data in .rodata and, rather than adr, use the combination of adrp :pg_hi21:
and add :lo12:. Unfortunately, iOS uses different syntax, so we must add more
transforms to arm-xlate.pl.
Tested manually by:
1. Use Android NDK r19-beta1
2. Follow usual instructions to configure CMake for aarch64, but pass
-DCMAKE_EXE_LINKER_FLAGS="-fuse-ld=lld -Wl,-execute-only".
3. Build. Confirm with readelf -l tool/bssl that .text is not marked
readable.
4. Push the test binaries onto a Pixel 3. Test normally and with
--cpu={none,neon,crypto}. I had to pass --gtest_filter=-*Thread* to
crypto_test. There appears to be an issue with some runtime function
that's unrelated to our assembly.
No measurable performance difference.
Going forward, to support this, we will need to apply similar changes to
all other AArch64 assembly. This is relatively straightforward, but may
be a little finicky for dual-AArch32/AArch64 files (aesv8-armx.pl).
Update-Note: Assembly syntax is a mess. There's a decent chance some
assembler will get offend.
Change-Id: Ib59b921d4cce76584320fefd23e6bb7ebd4847eb
Reviewed-on: https://boringssl-review.googlesource.com/c/33245
Reviewed-by: Adam Langley <agl@google.com>
Commit-Queue: David Benjamin <davidben@google.com>
David Benjamin
Katrin Leinweber
Adam Langley
Jeremy Apthorp
Adam Langley
Brian Smith