https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84908
--- Comment #8 from Jason Vas Dias <jason.vas.dias at gmail dot com> ---
Thanks for the clarification, and I hope the kernel
developers stop compiling the mainline vDSO with
-mindirect-branch=thunk-extern -mindirect-branch-register
.
But there are still a few things I am trying to figure out :
o Why is the thunk entry reference & relocation inserted
for 6 switch clauses and not for 5 ?
o So do I understand correctly:
__x86_indirect_thunk_rax
is resolved and jumped to save %rax , jump to the address in %rcx,
and then restore %rax, and return ?
it is normally
A) generated for libraries / executable as required or
B) is in libgcc ?
If (A), then what triggers it and why is it not being generated for the
vDSO?
If (B), then where is the code in libgcc ? I can't find it.
I'd also like more details on why it is wrong to compile the vDSO with these
flags - it does work, and yes causes compilation problems like this one which
can be worked around by declaring all 'notrace static inline' entry points
in vclock_gettime.c to have the function attributes:
static inline
__attribute__((
indirect_branch("keep"),
function_return("keep")
))
int do_monotonic_raw( int clock, struct timespec *ts );
which effectively disables the effect of
-mindirect-branch=thunk-extern -mindirect-branch-register
for these functions ; also __vdso_clock_gettime itself MUST
be then declared:
__attribute__((
indirect_branch("keep"),
function_return("keep")
))
int __vdso_clock_gettime ( ... ) { ...
But does it cause any other issues ?
I did investigate changing the switch in __vdso_clock_gettime
to :
#define _GTOD_LABEL_PREFIX_ _vcg_
#define _SYMCAT_(_S1_,_S2_) _S1_##_S2_
#define _GTOD_CLK_LABEL_(_CLK_) _SYMCAT_(_GTOD_LABEL_PREFIX_,_CLK_)
notrace int __vdso_clock_gettime(clockid_t clock, struct timespec *ts)
{
static const void * clk_jmp_tbl[ MAX_CLOCKS ] = {
[ CLOCK_REALTIME ] =
&&_GTOD_CLK_LABEL_(CLOCK_REALTIME) ,
[ CLOCK_MONOTONIC ] =
&&_GTOD_CLK_LABEL_(CLOCK_MONOTONIC) ,
[ CLOCK_PROCESS_CPUTIME_ID ] = &&fallback ,
[ CLOCK_THREAD_CPUTIME_ID ] = &&fallback ,
[ CLOCK_MONOTONIC_RAW ] =
&&_GTOD_CLK_LABEL_(CLOCK_MONOTONIC_RAW) ,
[ CLOCK_REALTIME_COARSE ] =
&&_GTOD_CLK_LABEL_(CLOCK_REALTIME_COARSE) ,
[ CLOCK_MONOTONIC_COARSE ] =
&&_GTOD_CLK_LABEL_(CLOCK_MONOTONIC_COARSE) ,
[ CLOCK_BOOTTIME ] = &&fallback ,
[ CLOCK_BOOTTIME_ALARM ] = &&fallback ,
[ CLOCK_SGI_CYCLE ] = &&fallback ,
[ CLOCK_TAI ] = &&fallback , // unused clocks
[ 12 ] = &¬_supported ,
[ 13 ] = &¬_supported ,
[ 14 ] = &¬_supported ,
[ 15 ] = &¬_supported
};
goto *clk_jmp_tbl [ clock & 0xf ] ;
_GTOD_CLK_LABEL_(CLOCK_REALTIME) :
if (do_realtime(ts) == VCLOCK_NONE)
goto fallback;
goto return_ok;
_GTOD_CLK_LABEL_(CLOCK_MONOTONIC) :
if (do_monotonic(ts) == VCLOCK_NONE)
goto fallback;
goto return_ok;
_GTOD_CLK_LABEL_(CLOCK_MONOTONIC_RAW) :
if (do_monotonic_raw(ts) == VCLOCK_NONE)
goto fallback;
goto return_ok;
_GTOD_CLK_LABEL_(CLOCK_REALTIME_COARSE) :
do_realtime_coarse(ts);
goto return_ok;
_GTOD_CLK_LABEL_(CLOCK_MONOTONIC_COARSE):
do_monotonic_coarse(ts);
goto return_ok;
return_ok:
return 0;
not_supported:
return -1;
fallback:
return vdso_fallback_gettime(clock, ts);
}
Results in 16 dynamic relocations,
shown in objdump -R arch/x86/entry/vdso/vdso64.so.dbg output:
DYNAMIC RELOCATION RECORDS
OFFSET TYPE VALUE
00000000000004a0 R_X86_64_RELATIVE *ABS*+0x0000000000000c49
...
000000000000518 R_X86_64_RELATIVE *ABS*+0x0000000000000b31
what's the problem with dyn relocs ?
I can't understand why GCC decides to generate relocations
here - it knows all the information before hand - why
can't it insert code to compute
( ( LOAD_ADDRESS_OF(__vdso_clock_gettime) == %pc at entry) + label_offset )
and populate the clk_jmp_table in an initializer in the
above version of the function rather than generate
relocations ?
Thanks & Best Regards,
Jason
