GSoC'11: DWARF2 call frame information

[Kostik Belousov]

On Tue, Mar 22, 2011 at 11:39:58PM +0800, Xingxing Pan wrote:
> 2011/3/22 Kostik Belousov <kostikbel at gmail.com>:
> > On Mon, Mar 21, 2011 at 08:32:04PM +0300, Chagin Dmitry wrote:
> >> On Mon, Mar 21, 2011 at 05:36:13PM +0800, Xingxing Pan wrote:
> >> > 2011/3/21 Chagin Dmitry <dchagin at freebsd.org>:
> >> > >> powerfull script.
> >> > >>
> >> > >> Xingxing Pan
> >> > >
> >> > > hmm, which script? I think enough amd64, i386 and amd64/ia32.
> >> > >
> >> > > I suggest to write a example before continuing the conversation
> >> > > about the GSoC. For example (bcopy || bzero) && cpu_switch.
> >> > > Is it ok for you?
> >> > >
> >> > > --
> >> > > Have fun!
> >> > > chd
> >> > >
> >> >
> >> > Hi Chargin,
> >> >
> >> > Thank you for your reply.
> >> > The followings shows how I try to add DWARF for bcopy.
> >> >
> >> > --- ../8.2.0/sys/i386/include/asm.h     2011-03-21 14:35:56.111973722 +0800
> >> > +++ asm.h       2011-03-21 15:25:31.564636162 +0800
> >> > @@ -71,7 +71,7 @@
> >> >
> >> >  #define _ENTRY(x)      _START_ENTRY; \
> >> >                         .globl CNAME(x); .type CNAME(x), at function; CNAME(x):
> >> > -#define        END(x)          .size x, . - x
> >> > +#define        END(x)          .cfi_endproc; .size x, . - x
> >> >
> >> >  #ifdef PROF
> >> >  #define        ALTENTRY(x)     _ENTRY(x); \
> >> > @@ -80,9 +80,10 @@
> >> >                         popl %ebp; \
> >> >                         jmp 9f
> >> >  #define        ENTRY(x)        _ENTRY(x); \
> >> > -                       pushl %ebp; movl %esp,%ebp; \
> >> > +                       .cfi_startproc; \
> >> > +                       pushl %ebp; .cfi_adjust_cfa_offset 4; movl
> >> > %esp,%ebp; .cfi_def_cfa_register %ebp; \
> >> >                         call PIC_PLT(HIDENAME(mcount)); \
> >> > -                       popl %ebp; \
> >> > +                       popl %ebp; .cfi_def_cfa %esp, 4; \
> >> >
> >> > --- bcopy.S     2011-03-21 15:51:26.804203809 +0800
> >> > +++ ../8.2.0/lib/libc/i386/string/bcopy.S       2011-03-21
> >> > 14:28:15.023069890 +0800
> >> > @@ -51,9 +51,7 @@ ENTRY(bcopy)
> >> >  #endif
> >> >  #endif
> >> >         pushl   %esi
> >> > -       .cfi_adjust_cfa_offset 4;
> >> >         pushl   %edi
> >> > -       .cfi_adjust_cfa_offset 4;
> >> >  #if defined(MEMCOPY) || defined(MEMMOVE)
> >> >         movl    12(%esp),%edi
> >> >         movl    16(%esp),%esi
> >> > @@ -77,9 +75,7 @@ ENTRY(bcopy)
> >> >         rep
> >> >         movsb
> >> >         popl    %edi
> >> > -       .cfi_adjust_cfa_offset -4;
> >> >         popl    %esi
> >> > -       .cfi_adjust_cfa_offset -4;
> >> >         ret
> >> >  1:
> >> >         addl    %ecx,%edi       /* copy backwards. */
> >> > @@ -98,9 +94,7 @@ ENTRY(bcopy)
> >> >         rep
> >> >         movsl
> >> >         popl    %edi
> >> > -       .cfi_adjust_cfa_offset -4;
> >> >         popl    %esi
> >> > -       .cfi_adjust_cfa_offset -4;
> >> >         cld
> >> >         ret
> >> >  #ifdef MEMCOPY
> >> >
> >> > But I don't know how to add DWARF for cpu_switch, because I have no
> >> > idea about the circumstance when we need to backtrace through this
> >> > function. Suppose there's a cpu switch like this,
> >> > threadA->kernel->threadB. Then should the expected backtrace has the
> >> > following result?
> >> >
> >> > threadB's stack
> >> > kernel's stack
> >> > threadA's stack
> >>
> >>
> >> hmm, ok. good, avoid cpu_switch.
> >> First of all, please, read style(9) man page.
> >> In the second, evaluate the proposed plan (discussed with kib@):
> >>
> >> 1) Annotate libc, msun, rtld, libthr (you)
> > 1a) Develop and implement a testing plan to verify the implementation.
> > 1b) consider doing full register tracking for assembler code.
> >
> >> 2) vdso (I'm)
> >> 3) Annotate signal trampolines (you, after vdso)
> >>
> >> And i'm going to understand what I need to do to start GSoC for you.
> >> Thanks!
> >>
> >>
> >> --
> >> Have fun!
> >> chd
> >
> >
> >
> 
> Hi Kostik,
> 
> I think the basic testing method can be using GDB to set breakpoint in
> functions and observing the backtrace result. GDB uses Expect. I can
> learn something from GDB's testsuite.
Sounds good.

> 
> AFAIK, CFA and return address are enough for unwinding. Dose full
> register tracking
> means to emit DWARF for all the registers's saving and restoring in
> the life time of the function?

Not only save and restore, but also for move around. I am mostly about
the syscall entry sequence on amd64, see the description of the
`syscall' instruction and handling of %rcx in libc sources.

Rarely used routines could be left aside.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 196 bytes
Desc: not available
Url : http://lists.freebsd.org/pipermail/freebsd-hackers/attachments/20110322/ec2ba5e2/attachment.pgp