Re: [PATCH v2] EDAC/ghes: Fix buffer overflow in ghes_edac_register()

[Robert Richter]

On 25.07.22 12:45:24, Robert Richter wrote:
> On 25.07.22 12:14:21, Borislav Petkov wrote:
> > On Mon, Jul 25, 2022 at 11:53:06AM +0200, Robert Richter wrote:
> > > The dimm->label buffer must be pre-initialized zero. This broke with:
> > > 
> > >  b9cae27728d1 EDAC/ghes: Scan the system once on driver init
> > > 
> > > since krealloc/krealloc_array() does not zero out the new allocated
> > > buffer for struct dimm_info in enumerate_dimms(). This uninitialized
> > > broken struct is then copied in ghes_edac_register() to the final
> > > dimm_info destination. Originally, before b9cae27728d1, the struct was
> > > zero initialized with kzalloc'ed by edac_mc_alloc() and directly used.
> > > Now, that a copy is created first, this copy must be also zero
> > > initialized.
> > > 
> > > IMO this is the proper fix:
> > 
> > Maybe, but this needs fixing too:
> > 
> > 	/* both strings must be non-zero */
> > 	if (bank && *bank && device && *device)
> > 
> > Obviously one of those strings are zero coming from that BIOS...

I see now, what you mean here, may be change this:

	snprintf(dimm->label, sizeof(dimm->label), "%s %s",
		 (bank && *bank) ? bank : "N/A",
		 (device && *device) ? device : "N/A");

to:

	snprintf(dimm->label, sizeof(dimm->label), "%s%s%s",
		 (bank && *bank) ? bank : "",
		 (bank && device) ? " " : "",
		 (device && *device) ? device : "");

It keeps the default assignment from edac_mc_alloc_dimms() but changes
we label if one of bank or device is given.

-Robert

> 
> But the label is pre-initialized in edac_mc_alloc_dimms():
> 
> 	p = dimm->label;
> 	n = snprintf(p, len, "mc#%u", mci->mc_idx);
> 
> You check if the label is emtpy when copying it in
> ghes_edac_register():
> 
> 	if (strlen(src->label))
> 		memcpy(dst->label, src->label, sizeof(src->label));
> 
> So if there is nothing that comes from the bios, this default label
> string from edac_mc_alloc_dimms() will be used.
> 
> If you write "N/A" to the label instead, the sysfs dimm_label values
> wont be unique any longer between dimms, which might break existing
> applications.
> 
> -Robert
> 
> > 
> > -- 
> > Regards/Gruss,
> >     Boris.
> > 
> > https://people.kernel.org/tglx/notes-about-netiquette