Re: KMSAN and dtrace

On Fri, Nov 22, 2024 at 7:47 AM Alan Somers <asomers@freebsd.org> wrote:
>
> On Fri, Nov 22, 2024 at 7:07 AM Mark Johnston <markj@freebsd.org> wrote:
> >
> > On Thu, Nov 21, 2024 at 04:06:53PM -0700, Alan Somers wrote:
> > > Would it be possible to make dtrace with with KMSAN?  It would
> > > certainly make my life easier.  As it is, every time I try to kldload
> > > traceall, whether from the command line or in the loader, my VM
> > > infinite loops printing alerts like this:
> > >
> > > MSan: Uninitialized stack memory from isize64+0x2e
> > > #0 0xffffffff833d8f10 at __msan_warning+0x140
> > > #1 0xffffffff86ec9c03 at dtrace_disp_opnd+0xd3
> > > #2 0xffffffff86ebe552 at dtrace_disx86+0xc602
> > > #3 0xffffffff86eca98e at dtrace_instr_size+0xee
> > > #4 0xffffffff86d5ec27 at fbt_provide_module_function+0x957
> > > #5 0xffffffff83303eff at link_elf_each_function_nameval+0x56f
> > > #6 0xffffffff86d56cb3 at fbt_provide_module+0x423
> > > #7 0xffffffff86d56871 at fbt_linker_file_cb+0x41
> > > #8 0xffffffff830d9f5c at linker_file_foreach+0xdc
> > > #9 0xffffffff830d63dd at linker_load_module+0x413d
> > > #10 0xffffffff830e6116 at linker_load_dependencies+0x1136
> > > #11 0xffffffff853b70b8 at link_elf_load_file+0x65c8
> > > #12 0xffffffff830d542e at linker_load_module+0x318e
> > > #13 0xffffffff830e0611 at kern_kldload+0x5d1
> > > #14 0xffffffff830e0d74 at sys_kldload+0x1a4
> > > #15 0xffffffff84fe56f6 at amd64_syscall+0x706
> > > #16 0xffffffff84f0ef4b at fast_syscall_common+0xf8
> > >
> >
> > Here you go:
> >
> > https://cgit.freebsd.org/src/commit/?id=418d8f0dc269b314bba51de63869b20da1d9a76e
> > https://cgit.freebsd.org/src/commit/?id=cc3da1955c16df5eb0019e0fef810696b035b7cf
> >
> > This one might also be important if your test system is low on RAM,
> > since the KMSAN shadow map gobbles up quite a lot of memory:
> >
> > https://cgit.freebsd.org/src/commit/?id=5d12db2dafece9f6a0453c4a45c4abed6b1e15ec
> >
> > I haven't yet tried running through the full test suite, but this was
> > enough to load dtracell (quite slow under KMSAN) and run some simple
> > scripts.
>
> Wow, thanks!  I'll try it out later today.

It works now.  I can load the module and do some probes.  Other
probes, though, still trigger warnings.  For instance,
'fbt:zfs:zio_data_buf_alloc:entry {@z[stack()] = count();}' triggers
warnings like this:

MSan: Uninitialized stack memory in copyout():arg1, offset 24/368,
addr 0xfffffe00b68ae018, from w_stillcold+0x28
#0 0xffffffff8340009c at kmsan_report_hook+0x15c
#1 0xffffffff833dbc61 at kmsan_copyout+0x1f1
#2 0xffffffff87e800f0 at dtrace_ioctl+0x4420
#3 0xffffffff8297795f at devfs_ioctl+0x3ef
#4 0xffffffff8547c277 at VOP_IOCTL_APV+0x107
#5 0xffffffff8381f974 at vn_ioctl+0x7a4
#6 0xffffffff8297a0f6 at devfs_ioctl_f+0x186
#7 0xffffffff834f3c2b at kern_ioctl+0xc5b
#8 0xffffffff834f2dc0 at sys_ioctl+0x580
#9 0xffffffff84fe7836 at amd64_syscall+0x706
#10 0xffffffff84f1128b at fast_syscall_common+0xf8