| Age | Commit message (Collapse) | Author | Files | Lines |
|---|
|
These x86_64 vectorized implementations support AVX, AVX-2, and AVX512F.
The AVX-512F implementation is disabled on Skylake, due to throttling,
but it is quite fast on >= Cannonlake.
On the left is cycle counts on a Core i7 6700HQ using the AVX-2
codepath, comparing this implementation ("new") to the implementation in
the current crypto api ("old"). On the right are benchmarks on a Xeon
Gold 5120 using the AVX-512 codepath. The new implementation is faster
on all benchmarks.
AVX-2 AVX-512
--------- -----------
size old new size old new
---- ---- ---- ---- ---- ----
0 70 68 0 74 70
16 92 90 16 96 92
32 134 104 32 136 106
48 172 120 48 184 124
64 218 136 64 218 138
80 254 158 80 260 160
96 298 174 96 300 176
112 342 192 112 342 194
128 388 212 128 384 212
144 428 228 144 420 226
160 466 246 160 464 248
176 510 264 176 504 264
192 550 282 192 544 282
208 594 302 208 582 300
224 628 316 224 624 318
240 676 334 240 662 338
256 716 354 256 708 358
272 764 374 272 748 372
288 802 352 288 788 358
304 420 366 304 422 370
320 428 360 320 432 364
336 484 378 336 486 380
352 426 384 352 434 390
368 478 400 368 480 408
384 488 394 384 490 398
400 542 408 400 542 412
416 486 416 416 492 426
432 534 430 432 538 436
448 544 422 448 546 432
464 600 438 464 600 448
480 540 448 480 548 456
496 594 464 496 594 476
512 602 456 512 606 470
528 656 476 528 656 480
544 600 480 544 606 498
560 650 494 560 652 512
576 664 490 576 662 508
592 714 508 592 716 522
608 656 514 608 664 538
624 708 532 624 710 552
640 716 524 640 720 516
656 770 536 656 772 526
672 716 548 672 722 544
688 770 562 688 768 556
704 774 552 704 778 556
720 826 568 720 832 568
736 768 574 736 780 584
752 822 592 752 826 600
768 830 584 768 836 560
784 884 602 784 888 572
800 828 610 800 838 588
816 884 628 816 884 604
832 888 618 832 894 598
848 942 632 848 946 612
864 884 644 864 896 628
880 936 660 880 942 644
896 948 652 896 952 608
912 1000 664 912 1004 616
928 942 676 928 954 634
944 994 690 944 1000 646
960 1002 680 960 1008 646
976 1054 694 976 1062 658
992 1002 706 992 1012 674
1008 1052 720 1008 1058 690
This commit wires in the prior implementation from Andy, and makes the
following changes to be suitable for kernel land.
- Some cosmetic and structural changes, like renaming labels to
.Lname, constants, and other Linux conventions, as well as making
the code easy for us to maintain moving forward.
- CPU feature checking is done in C by the glue code.
- We avoid jumping into the middle of functions, to appease objtool,
and instead parameterize shared code.
- We maintain frame pointers so that stack traces make sense.
- We remove the dependency on the perl xlate code, which transforms
the output into things that assemblers we don't care about use.
Importantly, none of our changes affect the arithmetic or core code, but
just involve the differing environment of kernel space.
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Samuel Neves <sneves@dei.uc.pt>
Co-developed-by: Samuel Neves <sneves@dei.uc.pt>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
These two C implementations from Zinc -- a 32x32 one and a 64x64 one,
depending on the platform -- come from Andrew Moon's public domain
poly1305-donna portable code, modified for usage in the kernel. The
precomputation in the 32-bit version and the use of 64x64 multiplies in
the 64-bit version make these perform better than the code it replaces.
Moon's code is also very widespread and has received many eyeballs of
scrutiny.
There's a bit of interference between the x86 implementation, which
relies on internal details of the old scalar implementation. In the next
commit, the x86 implementation will be replaced with a faster one that
doesn't rely on this, so none of this matters much. But for now, to keep
this passing the tests, we inline the bits of the old implementation
that the x86 implementation relied on. Also, since we now support a
slightly larger key space, via the union, some offsets had to be fixed
up.
Nonce calculation was folded in with the emit function, to take
advantage of 64x64 arithmetic. However, Adiantum appeared to rely on no
nonce handling in emit, so this path was conditionalized. We also
introduced a new struct, poly1305_core_key, to represent the precise
amount of space that particular implementation uses.
Testing with kbench9000, depending on the CPU, the update function for
the 32x32 version has been improved by 4%-7%, and for the 64x64 by
19%-30%. The 32x32 gains are small, but I think there's great value in
having a parallel implementation to the 64x64 one so that the two can be
compared side-by-side as nice stand-alone units.
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
These are all functions which are invoked from elsewhere, so annotate
them as global using the new SYM_FUNC_START and their ENDPROC's by
SYM_FUNC_END.
Make sure ENTRY/ENDPROC is not defined on X86_64, given these were the
last users.
Signed-off-by: Jiri Slaby <jslaby@suse.cz>
Signed-off-by: Borislav Petkov <bp@suse.de>
Reviewed-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> [hibernate]
Reviewed-by: Boris Ostrovsky <boris.ostrovsky@oracle.com> [xen bits]
Acked-by: Herbert Xu <herbert@gondor.apana.org.au> [crypto]
Cc: Allison Randal <allison@lohutok.net>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Andy Shevchenko <andy@infradead.org>
Cc: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Cc: Armijn Hemel <armijn@tjaldur.nl>
Cc: Cao jin <caoj.fnst@cn.fujitsu.com>
Cc: Darren Hart <dvhart@infradead.org>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Enrico Weigelt <info@metux.net>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Herbert Xu <herbert@gondor.apana.org.au>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Jim Mattson <jmattson@google.com>
Cc: Joerg Roedel <joro@8bytes.org>
Cc: Josh Poimboeuf <jpoimboe@redhat.com>
Cc: Juergen Gross <jgross@suse.com>
Cc: Kate Stewart <kstewart@linuxfoundation.org>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: kvm ML <kvm@vger.kernel.org>
Cc: Len Brown <len.brown@intel.com>
Cc: linux-arch@vger.kernel.org
Cc: linux-crypto@vger.kernel.org
Cc: linux-efi <linux-efi@vger.kernel.org>
Cc: linux-efi@vger.kernel.org
Cc: linux-pm@vger.kernel.org
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Matt Fleming <matt@codeblueprint.co.uk>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Pavel Machek <pavel@ucw.cz>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: platform-driver-x86@vger.kernel.org
Cc: "Radim Krčmář" <rkrcmar@redhat.com>
Cc: Sean Christopherson <sean.j.christopherson@intel.com>
Cc: Stefano Stabellini <sstabellini@kernel.org>
Cc: "Steven Rostedt (VMware)" <rostedt@goodmis.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vitaly Kuznetsov <vkuznets@redhat.com>
Cc: Wanpeng Li <wanpengli@tencent.com>
Cc: Wei Huang <wei@redhat.com>
Cc: x86-ml <x86@kernel.org>
Cc: xen-devel@lists.xenproject.org
Cc: Xiaoyao Li <xiaoyao.li@linux.intel.com>
Link: https://lkml.kernel.org/r/20191011115108.12392-25-jslaby@suse.cz
|
|
Based on 1 normalized pattern(s):
this program is free software you can redistribute it and or modify
it under the terms of the gnu general public license as published by
the free software foundation either version 2 of the license or at
your option any later version
extracted by the scancode license scanner the SPDX license identifier
GPL-2.0-or-later
has been chosen to replace the boilerplate/reference in 3029 file(s).
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Allison Randal <allison@lohutok.net>
Cc: linux-spdx@vger.kernel.org
Link: https://lkml.kernel.org/r/20190527070032.746973796@linutronix.de
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
|
|
The x86_64 implementation of Poly1305 produces the wrong result on some
inputs because poly1305_4block_avx2() incorrectly assumes that when
partially reducing the accumulator, the bits carried from limb 'd4' to
limb 'h0' fit in a 32-bit integer. This is true for poly1305-generic
which processes only one block at a time. However, it's not true for
the AVX2 implementation, which processes 4 blocks at a time and
therefore can produce intermediate limbs about 4x larger.
Fix it by making the relevant calculations use 64-bit arithmetic rather
than 32-bit. Note that most of the carries already used 64-bit
arithmetic, but the d4 -> h0 carry was different for some reason.
To be safe I also made the same change to the corresponding SSE2 code,
though that only operates on 1 or 2 blocks at a time. I don't think
it's really needed for poly1305_block_sse2(), but it doesn't hurt
because it's already x86_64 code. It *might* be needed for
poly1305_2block_sse2(), but overflows aren't easy to reproduce there.
This bug was originally detected by my patches that improve testmgr to
fuzz algorithms against their generic implementation. But also add a
test vector which reproduces it directly (in the AVX2 case).
Fixes: b1ccc8f4b631 ("crypto: poly1305 - Add a four block AVX2 variant for x86_64")
Fixes: c70f4abef07a ("crypto: poly1305 - Add a SSE2 SIMD variant for x86_64")
Cc: <stable@vger.kernel.org> # v4.3+
Cc: Martin Willi <martin@strongswan.org>
Cc: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Eric Biggers <ebiggers@google.com>
Reviewed-by: Martin Willi <martin@strongswan.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
A lot of asm-optimized routines in arch/x86/crypto/ keep its
constants in .data. This is wrong, they should be on .rodata.
Mnay of these constants are the same in different modules.
For example, 128-bit shuffle mask 0x000102030405060708090A0B0C0D0E0F
exists in at least half a dozen places.
There is a way to let linker merge them and use just one copy.
The rules are as follows: mergeable objects of different sizes
should not share sections. You can't put them all in one .rodata
section, they will lose "mergeability".
GCC puts its mergeable constants in ".rodata.cstSIZE" sections,
or ".rodata.cstSIZE.<object_name>" if -fdata-sections is used.
This patch does the same:
.section .rodata.cst16.SHUF_MASK, "aM", @progbits, 16
It is important that all data in such section consists of
16-byte elements, not larger ones, and there are no implicit
use of one element from another.
When this is not the case, use non-mergeable section:
.section .rodata[.VAR_NAME], "a", @progbits
This reduces .data by ~15 kbytes:
text data bss dec hex filename
11097415 2705840 2630712 16433967 fac32f vmlinux-prev.o
11112095 2690672 2630712 16433479 fac147 vmlinux.o
Merged objects are visible in System.map:
ffffffff81a28810 r POLY
ffffffff81a28810 r POLY
ffffffff81a28820 r TWOONE
ffffffff81a28820 r TWOONE
ffffffff81a28830 r PSHUFFLE_BYTE_FLIP_MASK <- merged regardless of
ffffffff81a28830 r SHUF_MASK <------------- the name difference
ffffffff81a28830 r SHUF_MASK
ffffffff81a28830 r SHUF_MASK
..
ffffffff81a28d00 r K512 <- merged three identical 640-byte tables
ffffffff81a28d00 r K512
ffffffff81a28d00 r K512
Use of object names in section name suffixes is not strictly necessary,
but might help if someday link stage will use garbage collection
to eliminate unused sections (ld --gc-sections).
Signed-off-by: Denys Vlasenko <dvlasenk@redhat.com>
CC: Herbert Xu <herbert@gondor.apana.org.au>
CC: Josh Poimboeuf <jpoimboe@redhat.com>
CC: Xiaodong Liu <xiaodong.liu@intel.com>
CC: Megha Dey <megha.dey@intel.com>
CC: linux-crypto@vger.kernel.org
CC: x86@kernel.org
CC: linux-kernel@vger.kernel.org
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Extends the x86_64 Poly1305 authenticator by a function processing four
consecutive Poly1305 blocks in parallel using AVX2 instructions.
For large messages, throughput increases by ~15-45% compared to two
block SSE2:
testing speed of poly1305 (poly1305-simd)
test 0 ( 96 byte blocks, 16 bytes per update, 6 updates): 3809514 opers/sec, 365713411 bytes/sec
test 1 ( 96 byte blocks, 32 bytes per update, 3 updates): 5973423 opers/sec, 573448627 bytes/sec
test 2 ( 96 byte blocks, 96 bytes per update, 1 updates): 9446779 opers/sec, 906890803 bytes/sec
test 3 ( 288 byte blocks, 16 bytes per update, 18 updates): 1364814 opers/sec, 393066691 bytes/sec
test 4 ( 288 byte blocks, 32 bytes per update, 9 updates): 2045780 opers/sec, 589184697 bytes/sec
test 5 ( 288 byte blocks, 288 bytes per update, 1 updates): 3711946 opers/sec, 1069040592 bytes/sec
test 6 ( 1056 byte blocks, 32 bytes per update, 33 updates): 573686 opers/sec, 605812732 bytes/sec
test 7 ( 1056 byte blocks, 1056 bytes per update, 1 updates): 1647802 opers/sec, 1740079440 bytes/sec
test 8 ( 2080 byte blocks, 32 bytes per update, 65 updates): 292970 opers/sec, 609378224 bytes/sec
test 9 ( 2080 byte blocks, 2080 bytes per update, 1 updates): 943229 opers/sec, 1961916528 bytes/sec
test 10 ( 4128 byte blocks, 4128 bytes per update, 1 updates): 494623 opers/sec, 2041804569 bytes/sec
test 11 ( 8224 byte blocks, 8224 bytes per update, 1 updates): 254045 opers/sec, 2089271014 bytes/sec
testing speed of poly1305 (poly1305-simd)
test 0 ( 96 byte blocks, 16 bytes per update, 6 updates): 3826224 opers/sec, 367317552 bytes/sec
test 1 ( 96 byte blocks, 32 bytes per update, 3 updates): 5948638 opers/sec, 571069267 bytes/sec
test 2 ( 96 byte blocks, 96 bytes per update, 1 updates): 9439110 opers/sec, 906154627 bytes/sec
test 3 ( 288 byte blocks, 16 bytes per update, 18 updates): 1367756 opers/sec, 393913872 bytes/sec
test 4 ( 288 byte blocks, 32 bytes per update, 9 updates): 2056881 opers/sec, 592381958 bytes/sec
test 5 ( 288 byte blocks, 288 bytes per update, 1 updates): 3711153 opers/sec, 1068812179 bytes/sec
test 6 ( 1056 byte blocks, 32 bytes per update, 33 updates): 574940 opers/sec, 607136745 bytes/sec
test 7 ( 1056 byte blocks, 1056 bytes per update, 1 updates): 1948830 opers/sec, 2057964585 bytes/sec
test 8 ( 2080 byte blocks, 32 bytes per update, 65 updates): 293308 opers/sec, 610082096 bytes/sec
test 9 ( 2080 byte blocks, 2080 bytes per update, 1 updates): 1235224 opers/sec, 2569267792 bytes/sec
test 10 ( 4128 byte blocks, 4128 bytes per update, 1 updates): 684405 opers/sec, 2825226316 bytes/sec
test 11 ( 8224 byte blocks, 8224 bytes per update, 1 updates): 367101 opers/sec, 3019039446 bytes/sec
Benchmark results from a Core i5-4670T.
Signed-off-by: Martin Willi <martin@strongswan.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
