Re: perf does not resolve plt symbols from libstdc++ right (.plt.sec problem)

[H.J. Lu]

On Mon, Mar 29, 2021 at 2:38 AM Richard Biener <rguenther@xxxxxxx> wrote:
>
> On Mon, 29 Mar 2021, Jiri Slaby wrote:
>
> > Any ideas on this?
> >
> > On 11. 01. 21, 7:31, Jiri Slaby wrote:
> > > Hi,
> > >
> > > this e-mails is a follow-up of my report at:
> > > https://bugzilla.suse.com/show_bug.cgi?id=1180681
> > >
> > > There is a problem with *@plt symbols in some libraries, they are unresolved
> > > by perf (memcmp@plt in this case):
> > >  >     0.26%  main2    /usr/lib64/libstdc++.so.6.0.28            0xa51a0
> > >             l [.] 0x00000000000a51a0
> > >
> > > On the other hand, plt symbols in other libraries are fine (memset@plt in
> > > this case):
> > >  >     0.17%  main2    /usr/lib64/libantlr4-runtime.so.4.8       0x4ed10
> > >             l [.] memset@plt
> > >
> > > I dumped memcmp's .plt.rela entries in perf:
> > > /usr/lib64/libantlr4-runtime.so.4.8: 154th addr=4e9d0 plt_off=4e020 hdr=10
> > > entry=10
> > > /usr/lib64/libstdc++.so.6.0.28: 772th addr=a1070 plt_off=9e020 hdr=10
> > > entry=10
> > >
> > > The difference (offset) of stdc++'s memcmp is 0xa51a0 (correct) - 0xa1070
> > > (perf's computed) = 0x4130.
> > >
> > > The problem is perf assumes nth entry of .plt.rela to correspond to nth
> > > function in .plt, but memcmp is in .plt.sec in libstdc++.so:
> > >
> > >  >Relocation section '.rela.plt' at offset 0x97900 contains 1018 entries:
> > >  >     Offset             Info             Type               Symbol's
> > > Value  Symbol's Name + Addend
> > >  > ...
> > >  > 00000000001dc838  0000007800000007 R_X86_64_JUMP_SLOT
> > > 0000000000000000 memcmp@GLIBC_2.2.5 + 0
> > >
> > > Perf does this with the rela entries:
> > > https://github.com/torvalds/linux/blob/f5e6c330254ae691f6d7befe61c786eb5056007e/tools/perf/util/symbol-elf.c#L385
> > >
> > > It takes a symbol index from sym.r_info. Then it resolves its name from
> > > .dynsym, appending "@plt" to it. Then this name is added to perf's symbol
> > > table along with address which is computed as .rela.plt index multiplied by
> > > entry size (shdr_plt.sh_entsize) plus plt header (shdr_plt.sh_entsize on
> > > x86_64 too).
> > >
> > > And from this comes (almost) the offset above:
> > >  >$ objdump -h /usr/lib64/libstdc++.so.6|grep -E ' .plt(\.sec)? '
> > >  >  12 .plt          00003fb0  000000000009e020  000000000009e020
> > > 0009e020  2**4
> > >  >  14 .plt.sec      00003fa0  00000000000a2160  00000000000a2160
> > > 000a2160  2**4
> > >
> > > 0xa2160-0x9e020 = 0x4140. I assume the 0x10 difference is that perf adds
> > > shdr_plt.sh_entsize (0x10) to the offset to skip the first .plt entry
> > > (header).
> > >
> > > Richard writes:
> > > ======
> > > .plt.sec is IIRC the "second" (sec) PLT entry - the one that will be used on
> > > the second call (and on).  This is used / emitted for ELF object
> > > instrumented for Intel CET.  The details escape me for the moment but I hope
> > > the x86 ABI documents this (and the constraints) in detail.
>
> I just checked and the x86_64 psABI doesn't say anything about .plt.sec

The second PLT is documented in section 13.2 Dynamic Linking in x86-64
psABI.  Please see elf_x86_64_get_synthetic_symtab in binutils for PLT symbol
processing.

-- 
H.J.