the file which is compiled with SSE disabled. The functions set up
the FPU context for kernel, and compiler optimizations which could
lead to use of XMM registers before the fpu_kern_enter(9) is called or
after fpu_kern_leave(9), panic the machine.
Discussed with: jmg
Sponsored by: The FreeBSD Foundation
MFC after: 1 week
context into memory for the kernel threads which called
fpu_kern_thread(9). This allows the fpu_kern_enter() callers to not
check for is_fpu_kern_thread() to get the optimization.
Apply the flag to padlock(4) and aesni(4). In aesni_cipher_process(),
do not leak FPU context state on error.
Sponsored by: The FreeBSD Foundation
MFC after: 1 week
context switch just to call the done callback... On my machine, this
improves geli/gzero decrypt performance by ~27% from 550MB/sec to
~700MB/sec...
MFC after: 3 days
regression manages to do it)... We use a packed struct to coerce
gcc/clang into producing unaligned loads (there is not packed pointer
attribute, otherwise this would be easier)...
use _storeu_ and _loadu_ when using the structure is overkill...
be better at using types properly... Since we allocate our own key
schedule and make sure it's aligned, use the __m128i type in various
arguments to functions...
clang ignores __aligned on prototypes and gcc errors on them, leave them
in comments to document that these function arguments are require to be
aligned...
about all that changes is movdqa -> movdqu from reading the diff of the
disassembly output...
Noticed by: symbolics at gmx.com
MFC after: 3 days
performance... Use SSE2 instructions for calculating the XTS tweek
factor... Let the compiler do more work and handle register allocation
by using intrinsics, now only the key schedule is in assembly...
Replace .byte hard coded instructions w/ the proper instructions now
that both clang and gcc support them...
On my machine, pulling the code to userland I saw performance go from
~150MB/sec to 2GB/sec in XTS mode. GELI on GNOP saw a more modest
increase of about 3x due to other system overhead (geom and
opencrypto)...
These changes allow almost full disk io rate w/ geli...
Reviewed by: -current, -security
Thanks to: Mike Hamburg for the XTS tweek algorithm
64bit and 32bit ABIs. As a side-effect, it enables AVX on capable
CPUs.
In particular:
- Query the CPU support for XSAVE, list of the supported extensions
and the required size of FPU save area. The hw.use_xsave tunable is
provided for disabling XSAVE, and hw.xsave_mask may be used to
select the enabled extensions.
- Remove the FPU save area from PCB and dynamically allocate the
(run-time sized) user save area on the top of the kernel stack,
right above the PCB. Reorganize the thread0 PCB initialization to
postpone it after BSP is queried for save area size.
- The dumppcb, stoppcbs and susppcbs now do not carry the FPU state as
well. FPU state is only useful for suspend, where it is saved in
dynamically allocated suspfpusave area.
- Use XSAVE and XRSTOR to save/restore FPU state, if supported and
enabled.
- Define new mcontext_t flag _MC_HASFPXSTATE, indicating that
mcontext_t has a valid pointer to out-of-struct extended FPU
state. Signal handlers are supplied with stack-allocated fpu
state. The sigreturn(2) and setcontext(2) syscall honour the flag,
allowing the signal handlers to inspect and manipilate extended
state in the interrupted context.
- The getcontext(2) never returns extended state, since there is no
place in the fixed-sized mcontext_t to place variable-sized save
area. And, since mcontext_t is embedded into ucontext_t, makes it
impossible to fix in a reasonable way. Instead of extending
getcontext(2) syscall, provide a sysarch(2) facility to query
extended FPU state.
- Add ptrace(2) support for getting and setting extended state; while
there, implement missed PT_I386_{GET,SET}XMMREGS for 32bit binaries.
- Change fpu_kern KPI to not expose struct fpu_kern_ctx layout to
consumers, making it opaque. Internally, struct fpu_kern_ctx now
contains a space for the extended state. Convert in-kernel consumers
of fpu_kern KPI both on i386 and amd64.
First version of the support for AVX was submitted by Tim Bird
<tim.bird am sony com> on behalf of Sony. This version was written
from scratch.
Tested by: pho (previous version), Yamagi Burmeister <lists yamagi org>
MFC after: 1 month
- Operate on uint64_t types when doing XORing, etc. instead of uint8_t.
- Don't bzero() temporary block for every AES block. Do it once for entire
data block.
- AES-NI is available only on little endian architectures. Simplify code
that takes block number from IV.
Benchmarks:
Memory-backed md(4) device, software AES-XTS, 4kB sector:
# dd if=/dev/md0.eli bs=1m
59.61MB/s
Memory-backed md(4) device, old AES-NI AES-XTS, 4kB sector:
# dd if=/dev/md0.eli bs=1m
97.29MB/s
Memory-backed md(4) device, new AES-NI AES-XTS, 4kB sector:
# dd if=/dev/md0.eli bs=1m
221.26MB/s
127% performance improvement between old and new code.
Harddisk, raw speed:
# dd if=/dev/ada0 bs=1m
137.63MB/s
Harddisk, software AES-XTS, 4kB sector:
# dd if=/dev/ada0.eli bs=1m
47.83MB/s (34% of raw disk speed)
Harddisk, old AES-NI AES-XTS, 4kB sector:
# dd if=/dev/ada0.eli bs=1m
68.33MB/s (49% of raw disk speed)
Harddisk, new AES-NI AES-XTS, 4kB sector:
# dd if=/dev/ada0.eli bs=1m
108.35MB/s (78% of raw disk speed)
58% performance improvement between old and new code.
As a side-note, GELI with AES-NI using AES-CBC can achive native disk speed.
MFC after: 3 days
The aeskeys_{amd64,i386}.S content was mostly obtained from OpenBSD,
no objections to the license from core.
Hardware provided by: Sentex Communications
Tested by: fabient, pho (previous versions)
MFC after: 1 month
