Re: [RFC PATCH v6.6] maple_tree: Fix MA_STATE_PREALLOC flag in mas_preallocate()

[Suren Baghdasaryan <surenb@google.com>]

On Tue, Apr 29, 2025 at 12:48 AM Hailong Liu <hailong.liu@oppo.com> wrote:
>
> On 4/29/2025 9:47 AM, Liam R. Howlett wrote:
> > Temporarily clear the preallocation flag when explicitly requesting
> > allocations.  Pre-existing allocations are already counted against the
> > request through mas_node_count_gfp(), but the allocations will not
> > happen if the MA_STATE_PREALLOC flag is set.  This flag is meant to
> > avoid re-allocating in bulk allocation mode, and to detect issues with
> > preallocation calculations.
> >
> > The MA_STATE_PREALLOC flag should also always be set on zero allocations
> > so that detection of underflow allocations will print a WARN_ON() during
> > consumption.
> >
> > User visible effect of this flaw is a WARN_ON() followed by a null
> > pointer dereference when subsequent requests for larger number of nodes
> > is ignored, such as the vma merge retry in mmap_region() caused by
> > drivers altering the vma flags.
> >
> > Reported-by: Zhaoyang Huang <zhaoyang.huang@unisoc.com>
> > Reported-by: Hailong Liu <hailong.liu@oppo.com>
> > Fixes: 54a611b605901 ("Maple Tree: add new data structure")
> > Link: https://lore.kernel.org/all/1652f7eb-a51b-4fee-8058-c73af63bacd1@oppo.com/
> > Link: https://lore.kernel.org/all/20250428184058.1416274-1-Liam.Howlett@oracle.com/
> > Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
> > Cc: Suren Baghdasaryan <surenb@google.com>
> > Cc: Hailong Liu <hailong.liu@oppo.com>
> > Cc: zhangpeng.00@bytedance.com <zhangpeng.00@bytedance.com>
> > Cc: Steve Kang <Steve.Kang@unisoc.com>
> > Cc: Matthew Wilcox <willy@infradead.org>
> > Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
> > Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com>
> > ---
> >  lib/maple_tree.c | 14 +++++++++-----
> >  1 file changed, 9 insertions(+), 5 deletions(-)
> >
> > Only the MA_STATE_PREALLOC flag needs to be fixed to avoid the vma
> > iterator issue.  Doing the minimum change here mitigates risk in the
> > stable kernels.
> >
> > If this fixes the issue, I'll resend without the RFC and add Stable to
> > the Cc list.
> >
> > Thanks again, Hailong for the work on this issue.  Your testcase helped
> > me narrow the bug down in the end.  I appreciate all the work and
> > support provided by the Android partners, especially the involvement on
> > the mailing list!
> That's my pleasure. I also learned a lot from you :)
> >
> > diff --git a/lib/maple_tree.c b/lib/maple_tree.c
> > index 4eda949063602..d9975b870dadc 100644
> > --- a/lib/maple_tree.c
> > +++ b/lib/maple_tree.c
> > @@ -5508,7 +5508,7 @@ int mas_preallocate(struct ma_state *mas, void *entry, gfp_t gfp)
> >       /* At this point, we are at the leaf node that needs to be altered. */
> >       /* Exact fit, no nodes needed. */
> >       if (wr_mas.r_min == mas->index && wr_mas.r_max == mas->last)
> > -             return 0;
> > +             goto ask_zero;
> >
> >       mas_wr_end_piv(&wr_mas);
> >       node_size = mas_wr_new_end(&wr_mas);
> > @@ -5517,10 +5517,11 @@ int mas_preallocate(struct ma_state *mas, void *entry, gfp_t gfp)
> >       if (node_size == wr_mas.node_end) {
> >               /* reuse node */
> >               if (!mt_in_rcu(mas->tree))
> > -                     return 0;
> > +                     goto ask_zero;
> > +
> >               /* shifting boundary */
> >               if (wr_mas.offset_end - mas->offset == 1)
> > -                     return 0;
> > +                     goto ask_zero;
> >       }
> >
> >       if (node_size >= mt_slots[wr_mas.type]) {
> > @@ -5539,10 +5540,13 @@ int mas_preallocate(struct ma_state *mas, void *entry, gfp_t gfp)
> >
> >       /* node store, slot store needs one node */
> >  ask_now:
> > +     mas->mas_flags &= ~MA_STATE_PREALLOC;
> >       mas_node_count_gfp(mas, request, gfp);
> > -     mas->mas_flags |= MA_STATE_PREALLOC;
> > -     if (likely(!mas_is_err(mas)))
> > +     if (likely(!mas_is_err(mas))) {
> > +ask_zero:
> > +             mas->mas_flags |= MA_STATE_PREALLOC;
> >               return 0;
> > +     }
> >
> >       mas_set_alloc_req(mas, 0);
> >       ret = xa_err(mas->node);
> Passed local test, planned to aging test.

I also confirmed the issue is not reproducible with your patch. Thanks!

>
> Brs,
> Hailong.
>