Re: [BUG/RFC] perf test fails on AMD CPUs

[Jiri Olsa <jolsa@redhat.com>]

On Mon, Aug 17, 2015 at 09:06:59AM -0700, Andy Lutomirski wrote:
> On Sun, Aug 16, 2015 at 9:36 PM, Borislav Petkov <bp@suse.de> wrote:
> > On Mon, Aug 17, 2015 at 12:29:56AM +0200, Jiri Olsa wrote:
> >> hi,
> >> 'perf test 18' is failing on systems with AMD processor.
> >
> > Hmm, still using that b0rked test box? :-)
> >
> > Also, which kernel?
> >
> > There have been substantial changes to the entry code recently. Although
> > I don't see anything being done differently on AMD there except
> > X86_BUG_SYSRET_SS_ATTRS but that should be unrelated.
> >
> >> The only reason I could find is that AMD does not set 'resume flag'
> >> in RFLAGS register the way the Intel CPU does.
> >>
> >> (simplified) test scenario:
> >>
> >>   - create breakpoint (on test_function) perf event with SIGIO signal
> >>     to be delivered any time the breakpoint is hit
> >>   - run test_function
> >>
> >>
> >> expected course of actions is:
> >>   1) CPU hits 'test_function'
> >>   2) DB exception is triggered, with RFLAGS.RF=0
> >>   3) DB exception handler sets regs->RFLAGS.RF=1 and perf handler
> >>      triggers irq_work pending work
> >>   4) DB exception executes iretd
> >>   5) irq_work interrupt is triggered, with RFLAGS.RF=1
> >>   6) irq_work interrupt calls kill_fasync with SIGIO signal
> >>   7) irq_work interrupt on return to userspace calls prepare_exit_to_usermode
> >>      which actually delivers the SIGIO signal
> >>   8) sigreturn syscall prepare registers to return to the
> >>      instruction from step 1) and sets RFLAGS.RF to the its original
> >>      value from step 5) (RFLAGS.RF=1)
> >>   9) CPU hits 'test_function' and DB exception is NOT triggered
> >>      due to RFLAGS.RF=1
> >>
> >> this is how I see it works on Intel
> >>
> >> But AMD gives me RFLAGS.RF=0 on step 5, which makes the step 9 to
> >> trigger the DB exception once again and makes the test fail.
> >
> > Adding Andy, he might have an idea. Leaving in the rest for reference.
> 
> Gee thanks :-p
> 
> Jiri, did you instrument the code and observe do_IRQ sees RF clear in
> its pt_regs?  Also, it might be worth checking that regs->ip in the
> irq_work matches regs->ip.

yep, thats what I saw.. once irq_work interrupt was triggered
the regs->ip was same as for the previous debug exception
but the RFLAGS.RF was 0

> 
> It's *possible* that I messed up and broke RF restore with
> opportunistic sysret, but the code looks correct:
> 
>         testq   $(X86_EFLAGS_RF|X86_EFLAGS_TF), %r11
>         jnz     opportunistic_sysret_failed

AFAICS the problematic paths did not hit syscalls

buuuuuut anyway, it looks like latest AMD firmware issue:

[root@amd-pike-07 ~]# cat /sys/devices/system/cpu/cpu0/microcode/version
0x6000822
[root@amd-pike-07 perf]# ./perf test 18
18: Test breakpoint overflow signal handler                  : Ok

[root@amd-pike-07 perf]# cat /sys/devices/system/cpu/cpu0/microcode/version
0x6000832
[root@amd-pike-07 perf]# ./perf test 18
18: Test breakpoint overflow signal handler                  : FAILED!


[root@amd-pike-07 ~]# cat /proc/cpuinfo 
processor       : 7
vendor_id       : AuthenticAMD
cpu family      : 21
model           : 2
model name      : AMD Opteron(tm) Processor 3380
stepping        : 0
microcode       : 0x6000832

SNIP


> >> AMD description of RF flag (SDM 3.1.6):
> >> =======================================
> >> Resume Flag (RF) Bit. Bit 16. The RF bit allows an instruction to be restarted following an
> >> instruction breakpoint resulting in a debug exception (#DB). This bit prevents multiple debug
> >> exceptions from occurring on the same instruction.
> >> The processor clears the RF bit after every instruction is successfully executed, except when the
> >> instruction is:
> >> •
> >> •
> >> An IRET that sets the RF bit.
> >> JMP, CALL, or INTn through a task gate.
> >> In both of the above cases, RF is not cleared to 0 until the next instruction successfully executes.
> >> When an exception occurs (or when a string instruction is interrupted), the processor normally sets
> >> RF=1 in the RFLAGS image saved on the interrupt stack. However, when a #DB exception occurs as a
> >> result of an instruction breakpoint, the processor clears the RF bit to 0 in the interrupt-stack RFLAGS
> >> image.
> 
> That's a little weird, I think.  Shouldn't RF be zero on #DB due to a
> *watchpoint* so that a watchpoint followed immediately by a breakpoint
> works?

the AMD description looked to be more vague (compared to Intels)

> 
> >> • For other cases, the value pushed for RF is the value that was in EFLAG.RF at the time the event handler was
> >> called. This includes:
> >> — Debug exceptions generated in response to instruction breakpoints
> >> — Hardware-generated interrupts arriving between instructions (including those arriving after the last
> >> iteration of a repeated string instruction)
> 
> This appears to be why it works on Intel.  Does AMD not do that?  We
> could probably work around this in software (by not using irq work for
> this), but yuck.

yep, but hopefuly it's the issue microcode ;-) Cc-ing guys from linux-firmware git

Sherry, Suravee, any idea?

thanks,
jirka