On Tue, Oct 12, 2021 at 11:32:25PM +0200, Vlastimil Babka wrote: > On 10/12/2021 10:43 PM, Darrick J. Wong wrote: > > On Tue, Oct 12, 2021 at 01:43:20PM -0700, Darrick J. Wong wrote: > >> On Sun, Oct 03, 2021 at 06:07:20PM -0700, David Rientjes wrote: > >>> On Thu, 30 Sep 2021, Rustam Kovhaev wrote: > >>> > >>>>>>> I think it's fair if something like XFS (not meant for tiny systems AFAIK?) > >>>>>>> excludes SLOB (meant for tiny systems). Clearly nobody tried to use these > >>>>>>> two together last 5 years anyway. > >>>>>> > >>>>>> +1 for adding Kconfig option, it seems like some things are not meant to > >>>>>> be together. > >>>>> > >>>>> But if we patch SLOB, we won't need it. > >>>> > >>>> OK, so we consider XFS on SLOB a supported configuration that might be > >>>> used and should be tested. > >>>> I'll look into maybe adding a config with CONFIG_SLOB and CONFIG_XFS_FS > >>>> to syzbot. > >>>> > >>>> It seems that we need to patch SLOB anyway, because any other code can > >>>> hit the very same issue. > >>>> > >>> > >>> It's probably best to introduce both (SLOB fix and Kconfig change for > >>> XFS), at least in the interim because the combo of XFS and SLOB could be > >>> broken in other ways. If syzbot doesn't complain with a patched kernel to > >>> allow SLOB to be used with XFS, then we could potentially allow them to be > >>> used together. > >>> > >>> (I'm not sure that this freeing issue is the *only* thing that is broken, > >>> nor that we have sufficient information to make that determination right > >>> now..) > >> > >> I audited the entire xfs (kernel) codebase and didn't find any other > >> usage errors. Thanks for the patch; I'll apply it to for-next. > > Which patch, the one that started this thread and uses kmem_cache_free() instead > of kfree()? I thought we said it's not the best way? It's probably better to fix slob to be able to tell that a kmem_free'd object actually belongs to a cache and should get freed that way, just like its larger sl[ua]b cousins. However, even if that does come to pass, anybody /else/ who wants to start(?) using XFS on a SLOB system will need this patch to fix the minor papercut. Now that I've checked the rest of the codebase, I don't find it reasonable to make XFS mutually exclusive with SLOB over two instances of slab cache misuse. Hence the RVB. :) --D > > Also, the obligatory > > > > Reviewed-by: Darrick J. Wong <djwong@xxxxxxxxxx> > > > > --D > > > >> > >> --D >
