If we're to allow the buggy CPU workaround to fire when __ARM_NEON__ is set,
CRYPTO_is_NEON_capable also needs to be aware of it. Also add an API to export
this value out of BoringSSL, so we can get some metrics on how prevalent this
chip is.
BUG=chromium:606629
Change-Id: I97d65a47a6130689098b32ce45a8c57c468aa405
Reviewed-on: https://boringssl-review.googlesource.com/7796
Reviewed-by: Adam Langley <agl@google.com>
Whatever compiler settings AOSP is using warns that this is a GNU extension.
Change-Id: Ife395d2b206b607b14c713cbb5a94d479816dad0
Reviewed-on: https://boringssl-review.googlesource.com/7604
Reviewed-by: Steven Valdez <svaldez@google.com>
Reviewed-by: David Benjamin <davidben@google.com>
Some ARMv8 Android devices don't have AT_HWCAP2. This means, when running in
32-bit mode (ARM capability APIs on Linux are different between AArch32 and
AArch64), we can't discover the various nice instructions.
On a Nexus 6P, this gives a, uh, minor performance win when running in 32-bit
mode.
Before:
Did 1085000 AES-128-GCM (16 bytes) seal operations in 1000003us (1084996.7 ops/sec): 17.4 MB/s
Did 60000 AES-128-GCM (1350 bytes) seal operations in 1013416us (59205.7 ops/sec): 79.9 MB/s
Did 11000 AES-128-GCM (8192 bytes) seal operations in 1019778us (10786.7 ops/sec): 88.4 MB/s
Did 1009000 AES-256-GCM (16 bytes) seal operations in 1000650us (1008344.6 ops/sec): 16.1 MB/s
Did 49000 AES-256-GCM (1350 bytes) seal operations in 1015698us (48242.7 ops/sec): 65.1 MB/s
Did 9394 AES-256-GCM (8192 bytes) seal operations in 1071104us (8770.4 ops/sec): 71.8 MB/s
Did 1557000 SHA-1 (16 bytes) operations in 1000317us (1556506.6 ops/sec): 24.9 MB/s
Did 762000 SHA-1 (256 bytes) operations in 1000527us (761598.6 ops/sec): 195.0 MB/s
Did 45000 SHA-1 (8192 bytes) operations in 1013773us (44388.6 ops/sec): 363.6 MB/s
Did 1459000 SHA-256 (16 bytes) operations in 1000271us (1458604.7 ops/sec): 23.3 MB/s
Did 538000 SHA-256 (256 bytes) operations in 1000990us (537467.9 ops/sec): 137.6 MB/s
Did 26000 SHA-256 (8192 bytes) operations in 1008403us (25783.3 ops/sec): 211.2 MB/s
After:
Did 1890000 AES-128-GCM (16 bytes) seal operations in 1000068us (1889871.5 ops/sec): 30.2 MB/s
Did 509000 AES-128-GCM (1350 bytes) seal operations in 1000112us (508943.0 ops/sec): 687.1 MB/s
Did 110000 AES-128-GCM (8192 bytes) seal operations in 1007966us (109130.7 ops/sec): 894.0 MB/s
Did 1960000 AES-256-GCM (16 bytes) seal operations in 1000303us (1959406.3 ops/sec): 31.4 MB/s
Did 460000 AES-256-GCM (1350 bytes) seal operations in 1001873us (459140.0 ops/sec): 619.8 MB/s
Did 97000 AES-256-GCM (8192 bytes) seal operations in 1005337us (96485.1 ops/sec): 790.4 MB/s
Did 1927000 SHA-1 (16 bytes) operations in 1000429us (1926173.7 ops/sec): 30.8 MB/s
Did 1151000 SHA-1 (256 bytes) operations in 1000425us (1150511.0 ops/sec): 294.5 MB/s
Did 87000 SHA-1 (8192 bytes) operations in 1003089us (86732.1 ops/sec): 710.5 MB/s
Did 2357390 SHA-256 (16 bytes) operations in 1000116us (2357116.6 ops/sec): 37.7 MB/s
Did 1410000 SHA-256 (256 bytes) operations in 1000176us (1409751.9 ops/sec): 360.9 MB/s
Did 101000 SHA-256 (8192 bytes) operations in 1007007us (100297.2 ops/sec): 821.6 MB/s
BUG=chromium:596156
Change-Id: Iacc1f8d8a07e991d4615f2e12c5c54923fb31aa2
Reviewed-on: https://boringssl-review.googlesource.com/7507
Reviewed-by: David Benjamin <davidben@google.com>
This removes the thread-unsafe SIGILL-based detection and the
multi-consumer-hostile CRYPTO_set_NEON_capable API. (Changing
OPENSSL_armcap_P after initialization is likely to cause problems.)
The right way to detect ARM features on Linux is getauxval. On aarch64,
we should be able to rely on this, so use it straight. Split this out
into its own file. The #ifdefs in the old cpu-arm.c meant it shared all
but no code with its arm counterpart anyway.
Unfortunately, various versions of Android have different missing APIs, so, on
arm, we need a series of workarounds. Previously, we used a SIGILL fallback
based on OpenSSL's logic, but this is inherently not thread-safe. (SIGILL also
does not tell us if the OS knows how to save and restore NEON state.) Instead,
base the behavior on Android NDK's cpu-features library, what Chromium
currently uses with CRYPTO_set_NEON_capable:
- Android before API level 20 does not provide getauxval. Where missing,
we can read from /proc/self/auxv.
- On some versions of Android, /proc/self/auxv is also not readable, so
use /proc/cpuinfo's Features line.
- Linux only advertises optional features in /proc/cpuinfo. ARMv8 makes NEON
mandatory, so /proc/cpuinfo can't be used without additional effort.
Finally, we must blacklist a particular chip because the NEON unit is broken
(https://crbug.com/341598).
Unfortunately, this means CRYPTO_library_init now depends on /proc being
available, which will require some care with Chromium's sandbox. The
simplest solution is to just call CRYPTO_library_init before entering
the sandbox.
It's worth noting that Chromium's current EnsureOpenSSLInit function already
depends on /proc/cpuinfo to detect the broken CPU, by way of base::CPU.
android_getCpuFeatures also interally depends on it. We were already relying on
both of those being stateful and primed prior to entering the sandbox.
BUG=chromium:589200
Change-Id: Ic5d1c341aab5a614eb129d8aa5ada2809edd6af8
Reviewed-on: https://boringssl-review.googlesource.com/7506
Reviewed-by: David Benjamin <davidben@google.com>
