Hi Dave,
On 2021-12-09 at 10:22:42 -0800, Dave Hansen wrote:
> On 12/2/21 11:32 PM, Pengfei Xu wrote:
> > +void populate_xstate_regs(void)
> > +{
> > + set_fpu_reg();
> > + set_avx2_ymm();
> > + set_avx512_opmask();
> > + set_pkru_xstate();
> > +}
>
> Pengfei, as I mentioned several times, XMM and YMM registers are not
> preserved across function calls. This only works by chance. The
> compiler is free to clobber them at basically any time between when they
> are populated and the XSAVE happens.
Thanks for suggestions!
Sorry, I read the link https://refspecs.linuxbase.org/elf/x86_64-abi-0.99.pdf
I have removed the XMM registers check as XMM registers are not preserved
across function calls in above doc mentioned:
"
Register | Usage | Preserved across function calls
%xmm0–%xmm1 | used to pass and return floating point arguments | No
%xmm2–%xmm7 | used to pass floating point arguments | No
%xmm8–%xmm15 | temporary registers | No
%mmx0–%mmx7 | temporary registers | No
"
It seems that the reason why the contents of the XMM register changed by chance
is:
As printf for example, after "printf" Glibc function:
do_lookup_x is triggered by printf:
"
0x00007ffff7fd5538 in do_lookup_x () from /lib64/ld-linux-x86-64.so.2 ->
0x00007ffff7fd55d7 <+1367>: e8 94 7b 01 00 callq 0x7ffff7fed170 <strcmp>
...
0x00007ffff7fee1af <+4159>: 90 nop
0x00007ffff7fee1b0 <+4160>: 66 0f ef c0 pxor %xmm0,%xmm0
0x00007ffff7fee1b4 <+4164>: 66 0f 6f 17 movdqa (%rdi),%xmm2
=> 0x00007ffff7fee1b8 <+4168>: 66 0f 6f 0e movdqa (%rsi),%xmm1
0x00007ffff7fee1bc <+4172>: 66 0f 74 c1 pcmpeqb %xmm1,%xmm0
" And after printf, xmm1 and xmm2 value was changed.
And seems some other glibc function will use XMM registers also.
YMM will not be preserved across function calls, but I'm not sure whether
some Glibc will use YMM.
I will remove YMM xstates check until there is some other better way for YMM
test or it's confirmed YMM will not be used by Glibc.
I think PKRU and AVX512 opmask will not be affected in the test.
PKRU xstate will not be affected by registers, and Glibc will not use AVX512
instructions.
How about the following changes:
Will remove set_avx2_ymm() and will only check XSAVE_MASK_FP, XSAVE_MASK_OPMASK
and XSAVE_MASK_PKRU xstates after signal handling and process switch,
xstates will be controlled by XSAVE_TEST_MASK:
"
#define XSAVE_TEST_MASK (XSAVE_MASK_FP | XSAVE_MASK_YMM | XSAVE_MASK_OPMASK \
| XSAVE_MASK_PKRU)
"
will change to
"
#define XSAVE_TEST_MASK (XSAVE_MASK_FP | XSAVE_MASK_OPMASK | XSAVE_MASK_PKRU)
"
FP will be used only by user space, and seems Glibc will not use FP in general.
This is the xstates position for FP, XMM, Header, YMM, AVX512 opmask and PKRU:
It could be saved by xsave instruction, and mask could control which part
will be saved(Header will be saved as mandatory):
FP (0-159 bytes)
XMM (160-415 bytes)
Reserved (416-511 bytes SDM vol1 13.4.1)
Header_used (512-527 bytes)
Headed_reserved (528-575 bytes must 00, otherwise xrstor will #GP)
YMM (Offset:CPUID.(EAX=0D,ECX=2).EBX Size:CPUID(EAX=0D,ECX=2).EAX)
AVX512 opmask (Offset:CPUID.(EAX=0D,ECX=5).EBX Size:CPUID(EAX=0D,ECX=5).EAX)
PKRU (Offset:CPUID.(EAX=0D,ECX=9).EBX Size:CPUID(EAX=0D,ECX=9).EAX)
Thanks!
BR.
