Re: [PATCH] EDAC/Intel: Fix shift-out-of-bounds when DIMM/NVDIMM is absent

[Kai-Heng Feng]

On Tue, May 16, 2023 at 8:53 PM Zhuo, Qiuxu <qiuxu.zhuo@xxxxxxxxx> wrote:
>
> > From: Kai-Heng Feng <kai.heng.feng@xxxxxxxxxxxxx>
> > ...
> > Subject: [PATCH] EDAC/Intel: Fix shift-out-of-bounds when DIMM/NVDIMM
> > is absent
> >
> > The following splat can be found on many systems equipped with EDAC:
> > [   13.875276] UBSAN: shift-out-of-bounds in
> > drivers/edac/skx_common.c:369:16
> > [   13.875279] shift exponent -66 is negative
> > [   13.875280] CPU: 11 PID: 519 Comm: systemd-udevd Not tainted 6.4.0-rc1+
> > #1
> > [   13.875282] Hardware name: HP HP Z4 G5 Workstation Desktop PC/8962,
> > BIOS U61 Ver. 01.01.15 04/19/2023
> > [   13.875283] Call Trace:
> > [   13.875285]  <TASK>
> > [   13.875287]  dump_stack_lvl+0x48/0x70
> > [   13.875295]  dump_stack+0x10/0x20
> > [   13.875297]  __ubsan_handle_shift_out_of_bounds+0x156/0x310
> > [   13.875302]  ? __kmem_cache_alloc_node+0x196/0x300
> > [   13.875307]  skx_get_dimm_info.cold+0xac/0x15d [i10nm_edac]
> > [   13.875312]  i10nm_get_dimm_config+0x240/0x360 [i10nm_edac]
> > [   13.875316]  ? kasprintf+0x4e/0x80
> > [   13.875321]  skx_register_mci+0x12b/0x1d0 [i10nm_edac]
> > [   13.875324]  ? __pfx_i10nm_get_dimm_config+0x10/0x10 [i10nm_edac]
> > [   13.875329]  i10nm_init+0x89f/0x1d10 [i10nm_edac]
> > [   13.875333]  ? __pfx_i10nm_init+0x10/0x10 [i10nm_edac]
> > [   13.875337]  do_one_initcall+0x46/0x240
> > [   13.875342]  ? kmalloc_trace+0x2a/0xb0
> > [   13.875346]  do_init_module+0x6a/0x280
> > [   13.875350]  load_module+0x2419/0x2500
> > [   13.875353]  ? security_kernel_post_read_file+0x5c/0x80
> > [   13.875358]  __do_sys_finit_module+0xcc/0x150
> > [   13.875360]  ? __do_sys_finit_module+0xcc/0x150
> > [   13.875363]  __x64_sys_finit_module+0x18/0x30
> > [   13.875365]  do_syscall_64+0x59/0x90
> > [   13.875368]  ? syscall_exit_to_user_mode+0x2a/0x50
> > [   13.875371]  ? do_syscall_64+0x69/0x90
> > [   13.875372]  ? do_syscall_64+0x69/0x90
> > [   13.875373]  ? do_syscall_64+0x69/0x90
> > [   13.875374]  ? do_syscall_64+0x69/0x90
> > [   13.875375]  ? syscall_exit_to_user_mode+0x2a/0x50
> > [   13.875376]  ? do_syscall_64+0x69/0x90
> > [   13.875377]  ? do_syscall_64+0x69/0x90
> > [   13.875378]  ? do_syscall_64+0x69/0x90
> > [   13.875379]  ? sysvec_call_function+0x4e/0xb0
> > [   13.875381]  entry_SYSCALL_64_after_hwframe+0x72/0xdc
> >
> > When a DIMM slot is empty, the read value of mtr can be 0xffffffff, therefore
>
> Looked like a buggy BIOS/hw that didn't set the mtr register.

If that's the case, I suspect the bug comes from Intel BIOS RC,
because the issue happens on different vendors' hardware.

>
> 1. Did you print the mtr register whose value was 0xffffffff?

Yes, 0xffffffff is the value. mcddrtcfg is also 0xffffffff.

> 2. Can you take a dmesg log with kernel "CONFIG_EDAC_DEBUG=y" enabled?
> 3. What was the CPU? Please take the output of "lscpu".

Both attached in Bugzlla [1].

> 4. Did you verify your patch that the issue was fixed on your systems?

I did, that's why I sent the patch to mailing list.

[1] https://bugzilla.kernel.org/show_bug.cgi?id=217453

Kai-Heng

>
> Thanks!
> -Qiuxu
>
> > the wrong "ranks" value creates shift-out-of-bounds error. The same issue
> > can be found on NVDIMM too.
> >
> > So only consider DIMM/NVDIMM is present when the value of
> > mtr/mcddrtcfg is not ~0.
> > ...