quoted text > Steven Rostedt wrote:
>
> > On Wed, 20 Aug 2008, Eran Liberty wrote:
> >
> >
> >
> > > Steven Rostedt wrote:
> > >
> > >
> > > > On Wed, 20 Aug 2008, Steven Rostedt wrote:
> > > >
> > > >
> > > >
> > > > > On Wed, 20 Aug 2008, Benjamin Herrenschmidt wrote:
> > > > >
> > > > >
> > > > >
> > > > > > Found the problem (or at least -a- problem), it's a gcc bug.
> > > > > >
> > > > > > Well, first I must say the code generated by -pg is just plain
> > > > > > horrible :-)
> > > > > >
> > > > > > Appart from that, look at the exit of, for example, __d_lookup, as
> > > > > > generated by gcc when ftrace is enabled:
> > > > > >
> > > > > > c00c0498: 38 60 00 00 li r3,0
> > > > > > c00c049c: 81 61 00 00 lwz r11,0(r1)
> > > > > > c00c04a0: 80 0b 00 04 lwz r0,4(r11)
> > > > > > c00c04a4: 7d 61 5b 78 mr r1,r11
> > > > > > c00c04a8: bb 0b ff e0 lmw r24,-32(r11)
> > > > > > c00c04ac: 7c 08 03 a6 mtlr r0
> > > > > > c00c04b0: 4e 80 00 20 blr
> > > > > >
> > > > > > As you can see, it restores r1 -before- it pops r24..r31 off
> > > > > > the stack ! I let you imagine what happens if an interrupt happens
> > > > > > just in between those two instructions (mr and lmw). We don't do
> > > > > > redzones on our ABI, so basically, the registers end up corrupted
> > > > > > by the interrupt.
> > > > > >
> > > > > >
> > > > > Ouch! You've disassembled this without -pg too, and it does not
> have this
> > > > > bug? What version of gcc do you have?
> > > > >
> > > > >
> > > > >
> > > > I have:
> > > > gcc (Debian 4.3.1-2) 4.3.1
> > > >
> > > > c00c64c8: 81 61 00 00 lwz r11,0(r1)
> > > > c00c64cc: 7f 83 e3 78 mr r3,r28
> > > > c00c64d0: 80 0b 00 04 lwz r0,4(r11)
> > > > c00c64d4: ba eb ff dc lmw r23,-36(r11)
> > > > c00c64d8: 7d 61 5b 78 mr r1,r11
> > > > c00c64dc: 7c 08 03 a6 mtlr r0
> > > > c00c64e0: 4e 80 00 20 blr
> > > >
> > > >
> > > > My version looks fine. I'm thinking that this is a separate issue
> than what
> > > > Eran is seeing.
> > > >
> > > > Eran, can you do an "objdump -dr vmlinux" and search for __d_lookup,
> and
> > > > print out the end of the function dump.
> > > >
> > > > Thanks,
> > > >
> > > > -- Steve
> > > >
> > > >
> > > >
> > > >
> > > >
> > > powerpc-linux-gnu-objdump -dr --start-address=0xc00bb584 vmlinux | head
> -n 100
> > >
> > > vmlinux: file format elf32-powerpc
> > >
> > > Disassembly of section .text:
> > >
> > > c00bb584 <__d_lookup>:
> > >
> > >
> >
> > [...]
> >
> >
> >
> > > c00bb670: 41 9e 00 50 beq- cr7,c00bb6c0 <__d_lookup+0x13c>
> > > c00bb674: 83 de 00 00 lwz r30,0(r30)
> > > c00bb678: 2f 9e 00 00 cmpwi cr7,r30,0
> > > c00bb67c: 40 9e ff 98 bne+ cr7,c00bb614 <__d_lookup+0x90>
> > > c00bb680: 38 60 00 00 li r3,0
> > > c00bb684: 81 61 00 00 lwz r11,0(r1)
> > > c00bb688: 80 0b 00 04 lwz r0,4(r11)
> > > c00bb68c: 7d 61 5b 78 mr r1,r11
> > >
> > >
> >
> > [ BUG HERE IF INTERRUPT HAPPENS ]
> >
> >
> >
> > > c00bb690: bb 0b ff e0 lmw r24,-32(r11)
> > > c00bb694: 7c 08 03 a6 mtlr r0
> > > c00bb698: 4e 80 00 20 blr
> > >
> > >
> >
> > Yep, you have the same bug in your compiler.
> >
> > -- Steve
> >
> >
> Hmm... so whats now?
>
> Is there a way to prove this scenario is indeed the one that caused the
> opps?
Manually edit the broken binary to change the order of the restore and
see if the problem disappears. That will keep everything else
constant.
