Re: [RFC PATCH] perf/x86: make perf callchain work without CONFIG_FRAME_POINTER

[Kairui Song]

On Sat, Apr 6, 2019 at 1:27 AM Josh Poimboeuf <jpoimboe@xxxxxxxxxx> wrote:
>
> On Sat, Apr 06, 2019 at 01:05:55AM +0800, Kairui Song wrote:
> > On Sat, Apr 6, 2019 at 12:57 AM Josh Poimboeuf <jpoimboe@xxxxxxxxxx> wrote:
> > >
> > > On Fri, Apr 05, 2019 at 11:13:02PM +0800, Kairui Song wrote:
> > > > Hi Josh, thanks for the review, I tried again, using latest upstream
> > > > kernel commit ea2cec24c8d429ee6f99040e4eb6c7ad627fe777:
> > > > # uname -a
> > > > Linux localhost.localdomain 5.1.0-rc3+ #29 SMP Fri Apr 5 22:53:05 CST
> > > > 2019 x86_64 x86_64 x86_64 GNU/Linux
> > > >
> > > > Having following config:
> > > > > CONFIG_UNWINDER_ORC=y
> > > > > # CONFIG_UNWINDER_FRAME_POINTER is not set
> > > > and CONFIG_FRAME_POINTER is off too.
> > > >
> > > > Then record something with perf (also latest upstream version):
> > > > ./perf record -g -e kmem:* -c 1
> > > >
> > > > Interrupt it, then view the output:
> > > > perf script | less
> > > >
> > > > Then I notice the stacktrace in kernle is incomplete like following.
> > > > Did I miss anything?
> > > > --------------
> > > > lvmetad   617 [000]    55.600786:                     kmem:kfree:
> > > > call_site=ffffffffb219e269 ptr=(nil)
> > > >         ffffffffb22b2d1c kfree+0x11c (/lib/modules/5.1.0-rc3+/build/vmlinux)
> > > >             7fba7e58fd0f __select+0x5f (/usr/lib64/libc-2.28.so)
> > > >
> > > > kworker/u2:5-rp   171 [000]    55.628529:
> > > > kmem:kmem_cache_alloc: call_site=ffffffffb20e963d
> > > > ptr=0xffffa07f39c581e0 bytes_req=80 bytes_alloc=80
> > > > gfp_flags=GFP_ATOMIC
> > > >         ffffffffb22b0dec kmem_cache_alloc+0x13c
> > > > (/lib/modules/5.1.0-rc3+/build/vmlinux)
> > > > -------------
> > > >
> > > > And for the patch, I debugged the problem, and found how it happend:
> > > > The reason is that we use following code for fetching the registers on
> > > > a trace point:
> > > > ...snip...
> > > > #define perf_arch_fetch_caller_regs(regs, __ip) { \
> > > > (regs)->ip = (__ip); \
> > > > (regs)->bp = caller_frame_pointer(); \
> > > > (regs)->cs = __KERNEL_CS;
> > > > ...snip...
> > >
> > > Thanks, I was able to recreate.  It only happens when unwinding from a
> > > tracepoint.  I haven't investigated yet, but
> > > perf_arch_fetch_caller_regs() looks highly suspect, since it's doing
> > > (regs)->bp = caller_frame_pointer(), even for ORC.
> > >
> > > My only explanation for how your patch works is that RBP just happens to
> > > point to somewhere higher on the stack, causing the unwinder to start at
> > > a semi-random location.  I suspect the real "fix" is that you're no
> > > longer passing the regs to unwind_start().
> > >
> >
> > Yes that's right. Simply not passing regs to unwind_start will let the
> > unwind start from the perf sample handling functions, and introduce a
> > lot of "noise", so I let it skipped the frames until it reached the
> > frame of the trace point. The regs->bp should still points to the
> > stack base of the function which get called in the tracepoint that
> > trigger perf sample, so let unwinder skip all the frames above it made
> > it work.
>
> Ah, now I think I understand, thanks.  perf_arch_fetch_caller_regs()
> puts it in regs->bp, and then perf_callchain_kernel() reads that value
> to tell the unwinder where to start dumping the stack trace.  I guess
> that explains why your patch works, though it still seems very odd that
> perf_arch_fetch_caller_regs() is using regs->bp to store the frame
> address.  Maybe regs->sp would be more appropriate.
>
> --
> Josh

Right, thanks for the comment. And after second thought there are some
other issues here in the patch indeed, it still won't fix the problem
when used with ebpf and tracepoint, I made some mistake about handling
the callchain with different ways, will rethink about this and post an
update later.


--
Best Regards,
Kairui Song