disable asm

[zooko]

When building the bundled copy of Crypto++, we should just disable ASM always. This will result in a performance penalty on platforms that formerly were able to use the asm implementation, but it would also avoid occasional weird bugs (most recently #84) and would make it easier for us (especially me, who doesn't really know ASM) to understand what it is doing.

[sneves]

Unless Crypto++ speed is a bottleneck of pycryptopp's applications, I see nothing wrong with this.

The main loss here seems to be AES-NI support, which increases AES speed by quite a lot.

[daira]

Where is the code that selects AES-NI? I don't see it in any obvious place under ​https://tahoe-lafs.org/trac/pycryptopp/browser/git/src-cryptopp (rijndael.cpp/h, aes.h, cpu.cpp).

[sneves]

The bits of rijndael.cpp that use AES-NI are guarded by CRYPTOPP_BOOL_AESNI_INTRINSICS_AVAILABLE. This macro itself is defined in config.h, line 293.

[zooko]

See also #76.

[zooko]

I pushed a patch that defined CRYPTOPP_DISABLE_ASM to the debug-netbsd-rsa-2 branch:

​https://github.com/zooko/pycryptopp/commit/2a38bfd4b4ab714d5f624a4e505993b6aec4418b

And triggered a build of it on all builders:

​https://tahoe-lafs.org/buildbot-pycryptopp/waterfall?last_time=1365665120

As well as telling us if this changes any behavior of unit tests, this also will show the difference in benchmarks between the previous version and the CRYPTOPP_DISABLE_ASM version. (To see this you have to manually click and visually compare the previous and current benchmarks on each builder.)

[zooko]

Hm... So with CRYPTOPP_DISABLE_ASM, the quick self-test of SHA256 segfaults at startup, on the "atlas" ubuntu box and on midnightmagic's netbsd box:

​https://tahoe-lafs.org/buildbot-pycryptopp/builders/atlas1%20natty%20syslib/builds/133/steps/version/logs/valgrind

But only when they are linking against the system-provided libcryptopp instead of the embedded cryptopp. I guess it is really necessary that the CRYPTOPP_DISABLE_ASM macro be set the same when we build pycryptopp as when the libcryptopp.so that we're linking against was built. It is a safe bet (I guess) that it is never set for any system-provided libcryptopp.so, so I'll just make it so that our setup.py sets CRYPTOPP_DISABLE_ASM only when it is building the embedded Crypto++.

[zooko]

Okay, I committed that patch to ​https://github.com/zooko/pycryptopp/commits/85_disable_asm and triggered builds: ​https://tahoe-lafs.org/buildbot-pycryptopp/waterfall?last_time=1365755564

[zooko]

Posted about this to tahoe-dev: ​https://tahoe-lafs.org/pipermail/tahoe-dev/2013-April/008164.html

[zooko]

Someone please review this branch!

​https://github.com/zooko/pycryptopp/commits/85_disable_asm

[zooko]

I posted an update on the Crypto++ mailing list: ​​https://groups.google.com/forum/?fromgroups=#!topic/cryptopp-users/qGIdqp3MIgg

#85 appears to be on the way to fixing this bug, but it is somewhat dissatisfying since I don't know _exactly_ why #85 fixes it.

[daira]

I will review the branch.

[daira]

Reviewed; ​needs resolution of this comment.

[zooko]

Brian: please design-review the design proposed in the comments: ​https://github.com/zooko/pycryptopp/commit/5c2e0f729a25eee8f374ff79a9299a818c378bc2#commitcomment-3005081

[sneves]

• Latest patch (​https://github.com/zooko/pycryptopp/commits/85_disable_asm), is setting the (now unsupported) PYCRYPTOPP_DISABLE_EMBEDDED_CRYPTOPP environment var in pycryptopp.egg-info/stdeb.cfg
• In setup.py:35, 'supported' is misspelled.

[zooko]

Thank you very much for the review. Committed it to ​https://github.com/zooko/pycryptopp master:

• ​https://github.com/zooko/pycryptopp/commit/345be57fe19c65f239b0d25b48b640f70d3d0b33
• ​https://github.com/zooko/pycryptopp/commit/c20ccaaf1d395712f1d9a8eb0c58ff5afba5100e
• ​https://github.com/zooko/pycryptopp/commit/f42c836b76b9bbd3eee3f011a46dda7708079199
• ​https://github.com/zooko/pycryptopp/commit/f42c836b76b9bbd3eee3f011a46dda7708079199
• ​https://github.com/zooko/pycryptopp/commit/75f2898907becffda46d5ec07dad31013ce7cb4a

There's an open pull request to ​https://github.com/tahoe-lafs/pycryptopp master, which I'm hoping will stimulate Brian to review these patches.

[zooko]

This hasn't been applied to trunk yet.

[zooko]

We had already agreed to disable assembly optimizations in pycryptopp, because there seem to have been a lot of bugs in the optimized assembly code in the past, and because the added speed really makes no difference to our uses, as far as I know.

However, in order to explain and justify to other people (e.g. Debian packagers) why we are doing this, and why they should consider doing the same thing themselves, I just read through the entire history of issues in pycryptopp and classified whether they were runtime errors (and therefore potential security bugs) or build-time errors (therefore probably not), and whether they would have been avoided if we had been disabling assembly optimizations all along. Here are the results. They clearly show that we should disable the optimized assembly! About half of all the security-threatening bugs we've had would never have been an issue if we'd avoided assembly from the beginning.

By the way, in my opinion the author of Crypto++, Wei Dai, is an exceptionally skilled, careful, and experienced coder, and I would assume that if Crypto++ has had this many security-threatening bugs in its optimized assembly code, then other crypto libraries that also use optimized assembly code have also had at least as many.

bugs that cause run-time failures

(These bugs are potential security issues.)

• would have been avoided by DISABLE_ASM:
  • #24
  • #31
  • #45 (three *different* bugs in the assembly implementation)
  • #67
  • #84
  • #86

• unclear if it would have been avoided if we'd used DISABLE_ASM:
  • #65

• would not have been avoided by DISABLE_ASM:
  • #17
  • #44
  • #83

• would not have been avoided by DISABLE_ASM (but would have been avoided by using cffi instead of CPython API)
  • #19
  • #70
  • #80

• would have been avoided if we *didn't* use DISABLE_ASM! (A bug only in the non-ASM version!)
  • #66

bugs that cause deterministic build or compilation failures

(These bugs are *typically* not potential security issues but they can be, and in any case they are engineering/deployment issues.)

• would have been avoided by DISABLE_ASM:
  • #37
  • #96

• would not have been avoided by DISABLE_ASM:
  • #22
  • #23
  • #32
  • #39
  • #62
  • #77
  • #78

[daira]

What is the status of this ticket and ​https://github.com/tahoe-lafs/pycryptopp/pull/23 ?

[zooko]

Replying to daira:

What is the status of this ticket and ​https://github.com/tahoe-lafs/pycryptopp/pull/23 ?

I'm confused. I apparently thought that the patch had been merged into pycryptopp trunk. Obviously, it hasn't.

[daira]

Let's fix this before the Tahoe-LAFS 1.11 1.10.1 release. It's needed for Tahoe-LAFS to work on FreeBSD (​https://tahoe-lafs.org/trac/tahoe-lafs/ticket/2269), and it's a clear auditability improvement.

[zooko]

I agree we should fix this right away. What's the next step?

[daira]

I suspect it is to separate out the changes that are relevant to this ticket from ​https://github.com/tahoe-lafs/pycryptopp/pull/23 , and file another pull request.

[daira]

See ​https://github.com/tahoe-lafs/pycryptopp/pull/26, although it needs more work.

[zooko]

We applied the first two patches from ​https://github.com/tahoe-lafs/pycryptopp/pull/26 (​https://github.com/twopir/pycryptopp/commit/b4cb7086cba8ebfdc90fe63b1da879a7ce84c03f and ​https://github.com/twopir/pycryptopp/commit/cd7a209d10992b87bbbfd70846ee561b6bb218b9), asked Jeromy "whyrusleeping" Johnson to manually verify that this made the build go from failing to working on his Mac OS X 10.10.5 (see ​https://github.com/tahoe-lafs/pycryptopp/pull/27), and then ran benchmarks before and after the patch on my Google Chromebook Pixel 1 (Intel(R) Core(TM) i5-3427U CPU @ 1.80GHz). The benchmarks (shown below) showed as expected: no change for ECDSA, Ed25519, RSA, or hashlib, and small change in performance for AES, XSalsa20, and SHA256.

benchmarks of current trunk with asm enabled:

zooko@spark ~/playground/pycryptopp $ time python setup.py bench
running bench
<class 'pycryptopp.bench.bench_sigs.ECDSA256'>
generate key
best: 4.350e-02,   1th-best: 4.350e-02, mean: 4.350e-02,   1th-worst: 4.350e-02, worst: 4.350e-02 (of      1)
sign
best: 9.089e-01,   1th-best: 9.089e-01, mean: 9.089e-01,   1th-worst: 9.089e-01, worst: 9.089e-01 (of      1)
verify
best: 3.479e+00,   1th-best: 3.479e+00, mean: 3.479e+00,   1th-worst: 3.479e+00, worst: 3.479e+00 (of      1)

<class 'pycryptopp.bench.bench_sigs.Ed25519'>
generate key
best: 1.825e+00,   1th-best: 1.825e+00, mean: 1.825e+00,   1th-worst: 1.825e+00, worst: 1.825e+00 (of      1)
sign
best: 1.830e+00,   1th-best: 1.830e+00, mean: 1.830e+00,   1th-worst: 1.830e+00, worst: 1.830e+00 (of      1)
verify
best: 6.249e+00,   1th-best: 6.249e+00, mean: 6.249e+00,   1th-worst: 6.249e+00, worst: 6.249e+00 (of      1)

<class 'pycryptopp.bench.bench_sigs.RSA2048'>
generate key
best: 9.093e+01,   1th-best: 9.093e+01, mean: 9.093e+01,   1th-worst: 9.093e+01, worst: 9.093e+01 (of      1)
sign
best: 1.963e+00,   1th-best: 1.963e+00, mean: 1.963e+00,   1th-worst: 1.963e+00, worst: 1.963e+00 (of      1)
verify
best: 4.617e-02,   1th-best: 4.617e-02, mean: 4.617e-02,   1th-worst: 4.617e-02, worst: 4.617e-02 (of      1)

<class 'pycryptopp.bench.bench_sigs.RSA3248'>
generate key
best: 5.065e+02,   1th-best: 5.065e+02, mean: 5.065e+02,   1th-worst: 5.065e+02, worst: 5.065e+02 (of      1)
sign
best: 1.006e+01,   1th-best: 1.006e+01, mean: 1.006e+01,   1th-worst: 1.006e+01, worst: 1.006e+01 (of      1)
verify
best: 1.068e-01,   1th-best: 1.068e-01, mean: 1.068e-01,   1th-worst: 1.068e-01, worst: 1.068e-01 (of      1)

milliseconds per operation

<AES-128>
large (10000000 B)
best: 3.982e+00,   1th-best: 3.982e+00, mean: 3.982e+00,   1th-worst: 3.982e+00, worst: 3.982e+00 (of      1)

<AES-256>
large (10000000 B)
best: 5.422e+00,   1th-best: 5.422e+00, mean: 5.422e+00,   1th-worst: 5.422e+00, worst: 5.422e+00 (of      1)

<XSalsa20-256>
large (10000000 B)
best: 1.387e+00,   1th-best: 1.387e+00, mean: 1.387e+00,   1th-worst: 1.387e+00, worst: 1.387e+00 (of      1)

nanoseconds per byte crypted

<class 'pycryptopp.bench.bench_hashes.SHA256'>
large (10000000 B)
best: 5.136e+00,   1th-best: 5.136e+00, mean: 5.136e+00,   1th-worst: 5.136e+00, worst: 5.136e+00 (of      1)

<class 'pycryptopp.bench.bench_hashes.hashlibSHA256'>
large (10000000 B)
best: 4.798e+00,   1th-best: 4.798e+00, mean: 4.798e+00,   1th-worst: 4.798e+00, worst: 4.798e+00 (of      1)

nanoseconds per byte hashed

benchmarks with these two patches that disable asm:

running bench
<class 'pycryptopp.bench.bench_sigs.ECDSA256'>
generate key
best: 4.385e-02,   1th-best: 4.385e-02, mean: 4.385e-02,   1th-worst: 4.385e-02, worst: 4.385e-02 (of      1)
sign
best: 9.948e-01,   1th-best: 9.948e-01, mean: 9.948e-01,   1th-worst: 9.948e-01, worst: 9.948e-01 (of      1)
verify
best: 3.936e+00,   1th-best: 3.936e+00, mean: 3.936e+00,   1th-worst: 3.936e+00, worst: 3.936e+00 (of      1)

<class 'pycryptopp.bench.bench_sigs.Ed25519'>
generate key
best: 1.818e+00,   1th-best: 1.818e+00, mean: 1.818e+00,   1th-worst: 1.818e+00, worst: 1.818e+00 (of      1)
sign
best: 1.823e+00,   1th-best: 1.823e+00, mean: 1.823e+00,   1th-worst: 1.823e+00, worst: 1.823e+00 (of      1)
verify
best: 6.178e+00,   1th-best: 6.178e+00, mean: 6.178e+00,   1th-worst: 6.178e+00, worst: 6.178e+00 (of      1)

<class 'pycryptopp.bench.bench_sigs.RSA2048'>
generate key
best: 6.325e+01,   1th-best: 6.325e+01, mean: 6.325e+01,   1th-worst: 6.325e+01, worst: 6.325e+01 (of      1)
sign
best: 2.136e+00,   1th-best: 2.136e+00, mean: 2.136e+00,   1th-worst: 2.136e+00, worst: 2.136e+00 (of      1)
verify
best: 4.965e-02,   1th-best: 4.965e-02, mean: 4.965e-02,   1th-worst: 4.965e-02, worst: 4.965e-02 (of      1)

<class 'pycryptopp.bench.bench_sigs.RSA3248'>
generate key
best: 7.379e+02,   1th-best: 7.379e+02, mean: 7.379e+02,   1th-worst: 7.379e+02, worst: 7.379e+02 (of      1)
sign
best: 1.139e+01,   1th-best: 1.139e+01, mean: 1.139e+01,   1th-worst: 1.139e+01, worst: 1.139e+01 (of      1)
verify
best: 1.213e-01,   1th-best: 1.213e-01, mean: 1.213e-01,   1th-worst: 1.213e-01, worst: 1.213e-01 (of      1)

milliseconds per operation

<AES-128>
large (10000000 B)
best: 7.006e+00,   1th-best: 7.006e+00, mean: 7.006e+00,   1th-worst: 7.006e+00, worst: 7.006e+00 (of      1)

<AES-256>
large (10000000 B)
best: 9.376e+00,   1th-best: 9.376e+00, mean: 9.376e+00,   1th-worst: 9.376e+00, worst: 9.376e+00 (of      1)

<XSalsa20-256>
large (10000000 B)
best: 5.388e+00,   1th-best: 5.388e+00, mean: 5.388e+00,   1th-worst: 5.388e+00, worst: 5.388e+00 (of      1)

nanoseconds per byte crypted

<class 'pycryptopp.bench.bench_hashes.SHA256'>
large (10000000 B)
best: 6.213e+00,   1th-best: 6.213e+00, mean: 6.213e+00,   1th-worst: 6.213e+00, worst: 6.213e+00 (of      1)

<class 'pycryptopp.bench.bench_hashes.hashlibSHA256'>
large (10000000 B)
best: 4.839e+00,   1th-best: 4.839e+00, mean: 4.839e+00,   1th-worst: 4.839e+00, worst: 4.839e+00 (of      1)

nanoseconds per byte hashed

[zooko]

We landed ​https://github.com/tahoe-lafs/pycryptopp/pull/27.