Re: scripts/gdb: issues when loading modules after lx-symbols

[Stefano Garzarella]

On Mon, Oct 05, 2020 at 11:45:41AM +0200, Jan Kiszka wrote:
> On 05.10.20 11:29, Stefano Garzarella wrote:
> > On Mon, Oct 05, 2020 at 10:33:30AM +0200, Jan Kiszka wrote:
> >> On 05.10.20 10:14, Stefano Garzarella wrote:
> >>> On Sun, Oct 04, 2020 at 08:52:37PM +0200, Jan Kiszka wrote:
> >>>> On 01.10.20 16:31, Stefano Garzarella wrote:
> >>>>> Hi,
> >>>>> I had some issues with gdb scripts and kernel modules in Linux 5.9-rc7.
> >>>>>
> >>>>> If the modules are already loaded, and I do 'lx-symbols', all work fine.
> >>>>> But, if I load a kernel module after 'lx-symbols', I had this issue:
> >>>>>
> >>>>> [ 5093.393940] invalid opcode: 0000 [#1] SMP PTI
> >>>>> [ 5093.395134] CPU: 0 PID: 576 Comm: modprobe Not tainted 
> >>>>> 5.9.0-rc7-ste-00010-gf0b671d9608d-dirty #2
> >>>>> [ 5093.397566] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 
> >>>>> 1.13.0-2.fc32 04/01/2014
> >>>>> [ 5093.400761] RIP: 0010:do_init_module+0x1/0x270
> >>>>> [ 5093.402553] Code: ff ff e9 cf fe ff ff 0f 0b 49 c7 c4 f2 ff ff ff e9 
> >>>>> c1 fe ff ff e8 5f b2 65 00 66 66 2e 0f 1f 84 00 00 00 00 00 0f 1f 40 00 
> >>>>> cc <1f> 44 00 00 55 ba 10 00 00 00 be c0 0c 00 00 48 89 e5 41 56 41 55
> >>>>> [ 5093.409505] RSP: 0018:ffffc90000563d18 EFLAGS: 00010246
> >>>>> [ 5093.412056] RAX: 0000000000000000 RBX: ffffffffc010a0c0 RCX: 
> >>>>> 0000000000004ee3
> >>>>> [ 5093.414472] RDX: 0000000000004ee2 RSI: ffffea0001efe188 RDI: 
> >>>>> ffffffffc010a0c0
> >>>>> [ 5093.416349] RBP: ffffc90000563e50 R08: 0000000000000000 R09: 
> >>>>> 0000000000000002
> >>>>> [ 5093.418044] R10: 0000000000000096 R11: 00000000000008a4 R12: 
> >>>>> ffff88807a0d1280
> >>>>> [ 5093.424721] R13: ffffffffc010a110 R14: ffff88807a0d1300 R15: 
> >>>>> ffffc90000563e70
> >>>>> [ 5093.427138] FS:  00007f018f632740(0000) GS:ffff88807dc00000(0000) 
> >>>>> knlGS:0000000000000000
> >>>>> [ 5093.430037] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> >>>>> [ 5093.432279] CR2: 000055fbe282b239 CR3: 000000007922a006 CR4: 
> >>>>> 0000000000170ef0
> >>>>> [ 5093.435096] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 
> >>>>> 0000000000000000
> >>>>> [ 5093.436765] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 
> >>>>> 0000000000000400
> >>>>> [ 5093.439689] Call Trace:
> >>>>> [ 5093.440954]  ? load_module+0x24b6/0x27d0
> >>>>> [ 5093.443212]  ? __kernel_read+0xd6/0x150
> >>>>> [ 5093.445140]  __do_sys_finit_module+0xd3/0xf0
> >>>>> [ 5093.446877]  __x64_sys_finit_module+0x1a/0x20
> >>>>> [ 5093.449098]  do_syscall_64+0x38/0x50
> >>>>> [ 5093.450877]  entry_SYSCALL_64_after_hwframe+0x44/0xa9
> >>>>> [ 5093.456153] RIP: 0033:0x7f018f75c43d
> >>>>> [ 5093.457728] Code: 00 c3 66 2e 0f 1f 84 00 00 00 00 00 90 f3 0f 1e fa 
> >>>>> 48 89 f8 48 89 f7 48 89 d6 48 89 ca 4d 89 c2 4d 89 c8 4c 8b 4c 24 08 0f 
> >>>>> 05 <48> 3d 01 f0 ff ff 73 01 c3 48 8b 0d 2b 6a 0c 00 f7 d8 64 89 01 48
> >>>>> [ 5093.466349] RSP: 002b:00007ffd7f080368 EFLAGS: 00000246 ORIG_RAX: 
> >>>>> 0000000000000139
> >>>>> [ 5093.470613] RAX: ffffffffffffffda RBX: 0000557e5c96f9c0 RCX: 
> >>>>> 00007f018f75c43d
> >>>>> [ 5093.474747] RDX: 0000000000000000 RSI: 0000557e5c964288 RDI: 
> >>>>> 0000000000000003
> >>>>> [ 5093.478049] RBP: 0000000000040000 R08: 0000000000000000 R09: 
> >>>>> 0000000000000000
> >>>>> [ 5093.481298] R10: 0000000000000003 R11: 0000000000000246 R12: 
> >>>>> 0000000000000000
> >>>>> [ 5093.483725] R13: 0000557e5c964288 R14: 0000557e5c96f950 R15: 
> >>>>> 0000557e5c9775c0
> >>>>> [ 5093.485778] Modules linked in: virtio_vdpa(+) vdpa sunrpc kvm_intel 
> >>>>> kvm irqbypass virtio_blk virtio_rng rng_core [last unloaded: 
> >>>>> virtio_vdpa]
> >>>>> [ 5093.488695] ---[ end trace 23712ecebc11f53c ]---
> >>>>>
> >>>>> Guest kernel: Linux 5.9-rc7
> >>>>> gdb: GNU gdb (GDB) Fedora 9.1-6.fc32
> >>>>> I tried with QEMU 4.2.1 and the latest master branch: same issue.
> >>>>>
> >>>>>
> >>>>> I did some digging, and skipping the gdb 'add-symbol-file' command in 
> >>>>> symbol.py
> >>>>> avoid the issue, but of course I don't have the symbols loaded:
> >>>>>
> >>>>>     diff --git a/scripts/gdb/linux/symbols.py 
> >>>>> b/scripts/gdb/linux/symbols.py
> >>>>>     index 1be9763cf8bb..eadfaa4d4907 100644
> >>>>>     --- a/scripts/gdb/linux/symbols.py
> >>>>>     +++ b/scripts/gdb/linux/symbols.py
> >>>>>     @@ -129,7 +129,7 @@ lx-symbols command."""
> >>>>>                      filename=module_file,
> >>>>>                      addr=module_addr,
> >>>>>                      sections=self._section_arguments(module))
> >>>>>     -            gdb.execute(cmdline, to_string=True)
> >>>>>     +            #gdb.execute(cmdline, to_string=True)
> >>>>>                  if module_name not in self.loaded_modules:
> >>>>>                      self.loaded_modules.append(module_name)
> >>>>>              else:
> >>>>>
> >>>>> I tried several modules and this happens every time after '(gdb) 
> >>>>> lx-symbols'.
> >>>>>
> >>>>> Do you have any hints?
> >>>>>
> >>>> I assume you are debugging a kernel inside QEMU/KVM, right?
> >>>
> >>> Right!
> >>>
> >>>>                                                             Does it work
> >>>> without -enable-kvm?
> >>>
> >>> Yes, disabling kvm it works.
> >>>
> >>
> >> OK, there it is, still...
> >> What may also "work" is going down to single core.
> > 
> > No, I tried with single core and kvm enabled and I have the same issue.
> > 
> >>
> >>>>
> >>>> Debugging guests in KVM mode at least was unstable for a long time. I
> >>>> avoided setting soft-BPs - which is what the script does for the sake of
> >>>> tracking modules loading -, falling back to hw-BPs, as I had no time to
> >>>> debug that further. /Maybe/ that's the issue here.
> >>>
> >>> Thanks for the suggestion, I'll try to have a look.
> >>>
> >>
> >> Would be great if this issue could finally be resolved. And then covered
> >> by the kvm-unit tests. Those still succeed, I think.
> > 
> > Yeah, I'm a bit busy, but I'll try to find a fix.
> > 
> > Just an update, I tried to follow your suggestion using hw-BPs, but
> > unfortunately the gdb python module doesn't provide an easy way to set
> > them, so I hacked a bit gdb forcing hw-BPs and with this patch applied
> > to gdb I don't see the issue anymore:
> > 
> > diff --git a/gdb/python/py-breakpoint.c b/gdb/python/py-breakpoint.c
> > index 7369c91ad9..df8ec92049 100644
> > --- a/gdb/python/py-breakpoint.c
> > +++ b/gdb/python/py-breakpoint.c
> > @@ -57,7 +57,7 @@ struct pybp_code
> >  static struct pybp_code pybp_codes[] =
> >  {
> >    { "BP_NONE", bp_none},
> > -  { "BP_BREAKPOINT", bp_breakpoint},
> > +  { "BP_BREAKPOINT", bp_hardware_breakpoint},
> >    { "BP_WATCHPOINT", bp_watchpoint},
> >    { "BP_HARDWARE_WATCHPOINT", bp_hardware_watchpoint},
> >    { "BP_READ_WATCHPOINT", bp_read_watchpoint},
> > @@ -383,7 +383,7 @@ bppy_get_location (PyObject *self, void *closure)
> > 
> >    BPPY_REQUIRE_VALID (obj);
> > 
> > -  if (obj->bp->type != bp_breakpoint)
> > +  if (obj->bp->type != bp_hardware_breakpoint)
> >      Py_RETURN_NONE;
> > 
> >    const char *str = event_location_to_string (obj->bp->location.get ());
> > @@ -730,7 +730,7 @@ bppy_init (PyObject *self, PyObject *args, PyObject 
> > *kwargs)
> >                                     "temporary","source", "function",
> >                                     "label", "line", "qualified", NULL };
> >    const char *spec = NULL;
> > -  enum bptype type = bp_breakpoint;
> > +  enum bptype type = bp_hardware_breakpoint;
> >    int access_type = hw_write;
> >    PyObject *internal = NULL;
> >    PyObject *temporary = NULL;
> > @@ -792,7 +792,7 @@ bppy_init (PyObject *self, PyObject *args, PyObject 
> > *kwargs)
> >      {
> >        switch (type)
> >         {
> > -       case bp_breakpoint:
> > +       case bp_hardware_breakpoint:
> >           {
> >             event_location_up location;
> >             symbol_name_match_type func_name_match_type
> > @@ -834,7 +834,7 @@ bppy_init (PyObject *self, PyObject *args, PyObject 
> > *kwargs)
> >             create_breakpoint (python_gdbarch,
> >                                location.get (), NULL, -1, NULL,
> >                                0,
> > -                              temporary_bp, bp_breakpoint,
> > +                              temporary_bp, bp_hardware_breakpoint,
> >                                0,
> >                                AUTO_BOOLEAN_TRUE,
> >                                ops,
> > @@ -1007,7 +1007,7 @@ gdbpy_breakpoint_created (struct breakpoint *bp)
> >    if (!user_breakpoint_p (bp) && bppy_pending_object == NULL)
> >      return;
> > 
> > -  if (bp->type != bp_breakpoint
> > +  if (bp->type != bp_hardware_breakpoint
> >        && bp->type != bp_watchpoint
> >        && bp->type != bp_hardware_watchpoint
> >        && bp->type != bp_read_watchpoint
> > 
> > Of course it is an hack, but it's a starting point :-)
> > 
> 
> There are two key differences with soft vs. hard BPs:
> 
>  - guest code modification to inject and remove INT3 (looking at your
>    panic, this might be the first thing to check)
>  - different exception vectors and their reflection to or filtering from
>    the guest
> 
> Both are similar in that the need to step over the intercepted
> instruction in order to resume - except that soft BP needs a
> remove-step-restore-INT3 cycle.

Thanks for the explanation!

> 
> You should try debugging that without the lx-symbols script, just by
> adding soft BPs and watching what happens to the guest, what becomes
> incorrectly visible to it. Report, and maybe KVM folks can jump in
> (adding the list).

It works well. Also using lx-symbols, without loading new modules in the
guest after it, I can debug the guest kernel with soft-BP.

The issue with soft-BP seems related to 'add-symbol-file' commands;
if I skip it in the python script, I don't have the panic.

I'll look deeper.

Stefano