* Vlastimil Babka <vbabka@xxxxxxx> [230929 05:52]: > On 9/27/23 18:07, Liam R. Howlett wrote: > > When merging of the previous VMA fails after the vma iterator has been > > moved to the previous entry, the vma iterator must be advanced to ensure > > the caller takes the correct action on the next vma iterator event. Fix > > this by adding a vma_next() call to the error path. > > > > Users may experience higher CPU usage, most likely in very low memory > > situations. > > Maybe we could say explicitly that before this fix, vma_merge will be called > twice on the same vma, which to the best of our knowledge does not cause > anything worse than some wasted cycles because vma == prev, but it's fragile? I will modify this statement again in v3. > > > Link: https://urldefense.com/v3/__https://lore.kernel.org/linux-mm/CAG48ez12VN1JAOtTNMY*Y2YnsU45yL5giS-Qn=ejtiHpgJAbdQ@xxxxxxxxxxxxxx/__;Kw!!ACWV5N9M2RV99hQ!LhxAWtA9bZgQkMs8Egf7OLmMSj69FWYmfgxD-UoydFparflJmeHvdvKoQChX_kelOhqCP_SSnB1juSOrAg$ > > Closes: https://urldefense.com/v3/__https://lore.kernel.org/linux-mm/CAG48ez12VN1JAOtTNMY*Y2YnsU45yL5giS-Qn=ejtiHpgJAbdQ@xxxxxxxxxxxxxx/__;Kw!!ACWV5N9M2RV99hQ!LhxAWtA9bZgQkMs8Egf7OLmMSj69FWYmfgxD-UoydFparflJmeHvdvKoQChX_kelOhqCP_SSnB1juSOrAg$ > > Fixes: 18b098af2890 ("vma_merge: set vma iterator to correct position.") > > Cc: stable@xxxxxxxxxxxxxxx > > Cc: Jann Horn <jannh@xxxxxxxxxx> > > Signed-off-by: Liam R. Howlett <Liam.Howlett@xxxxxxxxxx> > > --- > > mm/mmap.c | 12 +++++++++--- > > 1 file changed, 9 insertions(+), 3 deletions(-) > > > > diff --git a/mm/mmap.c b/mm/mmap.c > > index b56a7f0c9f85..b5bc4ca9bdc4 100644 > > --- a/mm/mmap.c > > +++ b/mm/mmap.c > > @@ -968,14 +968,14 @@ struct vm_area_struct *vma_merge(struct vma_iterator *vmi, struct mm_struct *mm, > > vma_pgoff = curr->vm_pgoff; > > vma_start_write(curr); > > remove = curr; > > - err = dup_anon_vma(next, curr); > > + err = dup_anon_vma(next, curr, &anon_dup); > > } > > } > > } > > > > /* Error in anon_vma clone. */ > > if (err) > > - return NULL; > > + goto anon_vma_fail; > > > > if (vma_start < vma->vm_start || vma_end > vma->vm_end) > > vma_expanded = true; > > The vma_iter_config() actions done in this part are something we don't need > to undo? Oh, right. Thanks. This made me realise that my prealloc_fail path assumes there is a 'curr' vma that will cause the vma_next() to set the correct range. We actually may be 1 or 4, which means we are looking to add a VMA to the gap; in this case, vma_next() would go beyond where it was at the start of the function, and may cause issue since we do not return an error. So the current patch fixes the problem Jann discovered (and with any iterators), but the issue may exist in other error scenarios today or in the future. There also may be an issue with next merging failing and skipping a vma... Looking through the code, there are functions that do not match the entire vma iterator location due to trimming (mbind_range, mlock_fixup) so using the start/end that was passed in may not accurately represent the range passed in though the vma iterator. Although, those ranges do point to the correct location in the tree - they just may be smaller. All the callers have `addr` within the range of the vma iterator, so it would be safe to do vma_iter_set(vmi, addr); and vma_iter_load(vmi) to restore the location and range. Safest would be to save the vma iterator start and end location and restore those after a re-walk, but this seems unnecessary and would complicate backporting. > > > @@ -988,7 +988,7 @@ struct vm_area_struct *vma_merge(struct vma_iterator *vmi, struct mm_struct *mm, > > } > > > > if (vma_iter_prealloc(vmi, vma)) > > - return NULL; > > + goto prealloc_fail; > > > > > init_multi_vma_prep(&vp, vma, adjust, remove, remove2); > > VM_WARN_ON(vp.anon_vma && adjust && adjust->anon_vma && > > @@ -1016,6 +1016,12 @@ struct vm_area_struct *vma_merge(struct vma_iterator *vmi, struct mm_struct *mm, > > vma_complete(&vp, vmi, mm); > > khugepaged_enter_vma(res, vm_flags); > > return res; > > + > > +prealloc_fail: > > +anon_vma_fail: > > + if (merge_prev) > > + vma_next(vmi); > > + return NULL; > > } > > > > /* >
