Re: [PATCH v3 2/3] mm: replace vma_start_write() with vma_start_write_killable()

[Suren Baghdasaryan <surenb@google.com>]

On Thu, Feb 26, 2026 at 9:43 AM Liam R. Howlett <Liam.Howlett@oracle.com> wrote:
>
> * Suren Baghdasaryan <surenb@google.com> [260226 02:06]:
> > Now that we have vma_start_write_killable() we can replace most of the
> > vma_start_write() calls with it, improving reaction time to the kill
> > signal.
> >
> > There are several places which are left untouched by this patch:
> >
> > 1. free_pgtables() because function should free page tables even if a
> > fatal signal is pending.
> >
> > 2. process_vma_walk_lock(), which requires changes in its callers and
> > will be handled in the next patch.
> >
> > 3. userfaultd code, where some paths calling vma_start_write() can
> > handle EINTR and some can't without a deeper code refactoring.
> >
> > 4. mpol_rebind_mm() which is used by cpusset controller for migrations
> > and operates on a remote mm. Incomplete operations here would result
> > in an inconsistent cgroup state.
> >
> > 5. vm_flags_{set|mod|clear} require refactoring that involves moving
> > vma_start_write() out of these functions and replacing it with
> > vma_assert_write_locked(), then callers of these functions should
> > lock the vma themselves using vma_start_write_killable() whenever
> > possible.
> >
> > Suggested-by: Matthew Wilcox <willy@infradead.org>
> > Signed-off-by: Suren Baghdasaryan <surenb@google.com>
> > Reviewed-by: Ritesh Harjani (IBM) <ritesh.list@gmail.com> # powerpc
>
> Some nits below, but lgtm.
>
> Reviewed-by: Liam R. Howlett <Liam.Howlett@oracle.com>

Thanks!

>
> > ---
> >  arch/powerpc/kvm/book3s_hv_uvmem.c |  5 +-
> >  mm/khugepaged.c                    |  5 +-
> >  mm/madvise.c                       |  4 +-
> >  mm/memory.c                        |  2 +
> >  mm/mempolicy.c                     |  8 ++-
> >  mm/mlock.c                         | 21 +++++--
> >  mm/mprotect.c                      |  4 +-
> >  mm/mremap.c                        |  4 +-
> >  mm/vma.c                           | 93 +++++++++++++++++++++---------
> >  mm/vma_exec.c                      |  6 +-
> >  10 files changed, 109 insertions(+), 43 deletions(-)
> >
> > diff --git a/arch/powerpc/kvm/book3s_hv_uvmem.c b/arch/powerpc/kvm/book3s_hv_uvmem.c
> > index 5fbb95d90e99..0a28b48a46b8 100644
> > --- a/arch/powerpc/kvm/book3s_hv_uvmem.c
> > +++ b/arch/powerpc/kvm/book3s_hv_uvmem.c
> > @@ -410,7 +410,10 @@ static int kvmppc_memslot_page_merge(struct kvm *kvm,
> >                       ret = H_STATE;
> >                       break;
> >               }
> > -             vma_start_write(vma);
> > +             if (vma_start_write_killable(vma)) {
> > +                     ret = H_STATE;
> > +                     break;
> > +             }
> >               /* Copy vm_flags to avoid partial modifications in ksm_madvise */
> >               vm_flags = vma->vm_flags;
> >               ret = ksm_madvise(vma, vma->vm_start, vma->vm_end,
> > diff --git a/mm/khugepaged.c b/mm/khugepaged.c
> > index 1dd3cfca610d..6c92e31ee5fb 100644
> > --- a/mm/khugepaged.c
> > +++ b/mm/khugepaged.c
> > @@ -1141,7 +1141,10 @@ static enum scan_result collapse_huge_page(struct mm_struct *mm, unsigned long a
> >       if (result != SCAN_SUCCEED)
> >               goto out_up_write;
> >       /* check if the pmd is still valid */
> > -     vma_start_write(vma);
> > +     if (vma_start_write_killable(vma)) {
> > +             result = SCAN_FAIL;
> > +             goto out_up_write;
> > +     }
> >       result = check_pmd_still_valid(mm, address, pmd);
> >       if (result != SCAN_SUCCEED)
> >               goto out_up_write;
> > diff --git a/mm/madvise.c b/mm/madvise.c
> > index c0370d9b4e23..ccdaea6b3b15 100644
> > --- a/mm/madvise.c
> > +++ b/mm/madvise.c
> > @@ -173,7 +173,9 @@ static int madvise_update_vma(vm_flags_t new_flags,
> >       madv_behavior->vma = vma;
> >
> >       /* vm_flags is protected by the mmap_lock held in write mode. */
> > -     vma_start_write(vma);
> > +     if (vma_start_write_killable(vma))
> > +             return -EINTR;
> > +
> >       vm_flags_reset(vma, new_flags);
> >       if (set_new_anon_name)
> >               return replace_anon_vma_name(vma, anon_name);
> > diff --git a/mm/memory.c b/mm/memory.c
> > index 07778814b4a8..691062154cf5 100644
> > --- a/mm/memory.c
> > +++ b/mm/memory.c
> > @@ -379,6 +379,8 @@ void free_pgd_range(struct mmu_gather *tlb,
> >   * page tables that should be removed.  This can differ from the vma mappings on
> >   * some archs that may have mappings that need to be removed outside the vmas.
> >   * Note that the prev->vm_end and next->vm_start are often used.
> > + * We don't use vma_start_write_killable() because page tables should be freed
> > + * even if the task is being killed.
> >   *
> >   * The vma_end differs from the pg_end when a dup_mmap() failed and the tree has
> >   * unrelated data to the mm_struct being torn down.
> > diff --git a/mm/mempolicy.c b/mm/mempolicy.c
> > index 0e5175f1c767..90939f5bde02 100644
> > --- a/mm/mempolicy.c
> > +++ b/mm/mempolicy.c
> > @@ -1784,7 +1784,8 @@ SYSCALL_DEFINE4(set_mempolicy_home_node, unsigned long, start, unsigned long, le
> >               return -EINVAL;
> >       if (end == start)
> >               return 0;
> > -     mmap_write_lock(mm);
> > +     if (mmap_write_lock_killable(mm))
> > +             return -EINTR;
> >       prev = vma_prev(&vmi);
> >       for_each_vma_range(vmi, vma, end) {
> >               /*
> > @@ -1801,13 +1802,16 @@ SYSCALL_DEFINE4(set_mempolicy_home_node, unsigned long, start, unsigned long, le
> >                       err = -EOPNOTSUPP;
> >                       break;
> >               }
> > +             if (vma_start_write_killable(vma)) {
> > +                     err = -EINTR;
> > +                     break;
> > +             }
> >               new = mpol_dup(old);
> >               if (IS_ERR(new)) {
> >                       err = PTR_ERR(new);
> >                       break;
> >               }
> >
> > -             vma_start_write(vma);
> >               new->home_node = home_node;
> >               err = mbind_range(&vmi, vma, &prev, start, end, new);
> >               mpol_put(new);
> > diff --git a/mm/mlock.c b/mm/mlock.c
> > index 2f699c3497a5..c562c77c3ee0 100644
> > --- a/mm/mlock.c
> > +++ b/mm/mlock.c
> > @@ -420,7 +420,7 @@ static int mlock_pte_range(pmd_t *pmd, unsigned long addr,
> >   * Called for mlock(), mlock2() and mlockall(), to set @vma VM_LOCKED;
> >   * called for munlock() and munlockall(), to clear VM_LOCKED from @vma.
> >   */
> > -static void mlock_vma_pages_range(struct vm_area_struct *vma,
> > +static int mlock_vma_pages_range(struct vm_area_struct *vma,
> >       unsigned long start, unsigned long end, vm_flags_t newflags)
> >  {
> >       static const struct mm_walk_ops mlock_walk_ops = {
> > @@ -441,7 +441,9 @@ static void mlock_vma_pages_range(struct vm_area_struct *vma,
> >        */
> >       if (newflags & VM_LOCKED)
> >               newflags |= VM_IO;
> > -     vma_start_write(vma);
> > +     if (vma_start_write_killable(vma))
> > +             return -EINTR;
> > +
> >       vm_flags_reset_once(vma, newflags);
> >
> >       lru_add_drain();
> > @@ -452,6 +454,7 @@ static void mlock_vma_pages_range(struct vm_area_struct *vma,
> >               newflags &= ~VM_IO;
> >               vm_flags_reset_once(vma, newflags);
> >       }
> > +     return 0;
> >  }
> >
> >  /*
> > @@ -501,10 +504,12 @@ static int mlock_fixup(struct vma_iterator *vmi, struct vm_area_struct *vma,
> >        */
> >       if ((newflags & VM_LOCKED) && (oldflags & VM_LOCKED)) {
> >               /* No work to do, and mlocking twice would be wrong */
> > -             vma_start_write(vma);
> > +             ret = vma_start_write_killable(vma);
> > +             if (ret)
> > +                     goto out;
> >               vm_flags_reset(vma, newflags);
> >       } else {
> > -             mlock_vma_pages_range(vma, start, end, newflags);
> > +             ret = mlock_vma_pages_range(vma, start, end, newflags);
> >       }
> >  out:
> >       *prev = vma;
> > @@ -733,9 +738,13 @@ static int apply_mlockall_flags(int flags)
> >
> >               error = mlock_fixup(&vmi, vma, &prev, vma->vm_start, vma->vm_end,
> >                                   newflags);
> > -             /* Ignore errors, but prev needs fixing up. */
> > -             if (error)
> > +             /* Ignore errors except EINTR, but prev needs fixing up. */
> > +             if (error) {
> > +                     if (error == -EINTR)
> > +                             return error;
> > +
> >                       prev = vma;
> > +             }
> >               cond_resched();
> >       }
> >  out:
> > diff --git a/mm/mprotect.c b/mm/mprotect.c
> > index c0571445bef7..49dbb7156936 100644
> > --- a/mm/mprotect.c
> > +++ b/mm/mprotect.c
> > @@ -765,7 +765,9 @@ mprotect_fixup(struct vma_iterator *vmi, struct mmu_gather *tlb,
> >        * vm_flags and vm_page_prot are protected by the mmap_lock
> >        * held in write mode.
> >        */
> > -     vma_start_write(vma);
> > +     error = vma_start_write_killable(vma);
> > +     if (error < 0)
> > +             goto fail;
> >       vm_flags_reset_once(vma, newflags);
> >       if (vma_wants_manual_pte_write_upgrade(vma))
> >               mm_cp_flags |= MM_CP_TRY_CHANGE_WRITABLE;
> > diff --git a/mm/mremap.c b/mm/mremap.c
> > index 2be876a70cc0..aef1e5f373c7 100644
> > --- a/mm/mremap.c
> > +++ b/mm/mremap.c
> > @@ -1286,7 +1286,9 @@ static unsigned long move_vma(struct vma_remap_struct *vrm)
> >               return -ENOMEM;
> >
> >       /* We don't want racing faults. */
> > -     vma_start_write(vrm->vma);
> > +     err = vma_start_write_killable(vrm->vma);
> > +     if (err)
> > +             return err;
> >
> >       /* Perform copy step. */
> >       err = copy_vma_and_data(vrm, &new_vma);
> > diff --git a/mm/vma.c b/mm/vma.c
> > index bb4d0326fecb..9f2664f1d078 100644
> > --- a/mm/vma.c
> > +++ b/mm/vma.c
> > @@ -530,6 +530,13 @@ __split_vma(struct vma_iterator *vmi, struct vm_area_struct *vma,
> >       if (err)
> >               goto out_free_vmi;
> >
> > +     err = vma_start_write_killable(vma);
> > +     if (err)
> > +             goto out_free_mpol;
> > +     err = vma_start_write_killable(new);
> > +     if (err)
> > +             goto out_free_mpol;
> > +
> >       err = anon_vma_clone(new, vma, VMA_OP_SPLIT);
> >       if (err)
> >               goto out_free_mpol;
> > @@ -540,9 +547,6 @@ __split_vma(struct vma_iterator *vmi, struct vm_area_struct *vma,
> >       if (new->vm_ops && new->vm_ops->open)
> >               new->vm_ops->open(new);
> >
> > -     vma_start_write(vma);
> > -     vma_start_write(new);
> > -
> >       init_vma_prep(&vp, vma);
> >       vp.insert = new;
> >       vma_prepare(&vp);
> > @@ -895,16 +899,22 @@ static __must_check struct vm_area_struct *vma_merge_existing_range(
> >       }
> >
> >       /* No matter what happens, we will be adjusting middle. */
> > -     vma_start_write(middle);
> > +     err = vma_start_write_killable(middle);
> > +     if (err)
> > +             goto abort;
> >
> >       if (merge_right) {
> > -             vma_start_write(next);
> > +             err = vma_start_write_killable(next);
> > +             if (err)
> > +                     goto abort;
> >               vmg->target = next;
> >               sticky_flags |= (next->vm_flags & VM_STICKY);
> >       }
> >
> >       if (merge_left) {
> > -             vma_start_write(prev);
> > +             err = vma_start_write_killable(prev);
> > +             if (err)
> > +                     goto abort;
> >               vmg->target = prev;
> >               sticky_flags |= (prev->vm_flags & VM_STICKY);
> >       }
> > @@ -1155,10 +1165,12 @@ int vma_expand(struct vma_merge_struct *vmg)
> >       struct vm_area_struct *next = vmg->next;
> >       bool remove_next = false;
> >       vm_flags_t sticky_flags;
> > -     int ret = 0;
> > +     int ret;
> >
> >       mmap_assert_write_locked(vmg->mm);
> > -     vma_start_write(target);
> > +     ret = vma_start_write_killable(target);
> > +     if (ret)
> > +             return ret;
> >
> >       if (next && target != next && vmg->end == next->vm_end)
> >               remove_next = true;
> > @@ -1187,6 +1199,9 @@ int vma_expand(struct vma_merge_struct *vmg)
> >        * we don't need to account for vmg->give_up_on_mm here.
> >        */
> >       if (remove_next) {
> > +             ret = vma_start_write_killable(next);
> > +             if (ret)
> > +                     return ret;
> >               ret = dup_anon_vma(target, next, &anon_dup);
> >               if (ret)
> >                       return ret;
> > @@ -1197,10 +1212,8 @@ int vma_expand(struct vma_merge_struct *vmg)
> >                       return ret;
> >       }
> >
> > -     if (remove_next) {
> > -             vma_start_write(next);
> > +     if (remove_next)
> >               vmg->__remove_next = true;
> > -     }
> >       if (commit_merge(vmg))
> >               goto nomem;
> >
> > @@ -1233,6 +1246,7 @@ int vma_shrink(struct vma_iterator *vmi, struct vm_area_struct *vma,
> >              unsigned long start, unsigned long end, pgoff_t pgoff)
> >  {
> >       struct vma_prepare vp;
> > +     int err;
> >
> >       WARN_ON((vma->vm_start != start) && (vma->vm_end != end));
> >
> > @@ -1244,7 +1258,11 @@ int vma_shrink(struct vma_iterator *vmi, struct vm_area_struct *vma,
> >       if (vma_iter_prealloc(vmi, NULL))
> >               return -ENOMEM;
> >
> > -     vma_start_write(vma);
> > +     err = vma_start_write_killable(vma);
> > +     if (err) {
> > +             vma_iter_free(vmi);
> > +             return err;
> > +     }
>
> Could avoid allocating here by reordering the lock, but this is fine.

Ack. I'll move it before vma_iter_prealloc().

>
> >
> >       init_vma_prep(&vp, vma);
> >       vma_prepare(&vp);
> > @@ -1434,7 +1452,9 @@ static int vms_gather_munmap_vmas(struct vma_munmap_struct *vms,
> >                       if (error)
> >                               goto end_split_failed;
> >               }
> > -             vma_start_write(next);
> > +             error = vma_start_write_killable(next);
> > +             if (error)
> > +                     goto munmap_gather_failed;
> >               mas_set(mas_detach, vms->vma_count++);
> >               error = mas_store_gfp(mas_detach, next, GFP_KERNEL);
> >               if (error)
> > @@ -1828,12 +1848,17 @@ static void vma_link_file(struct vm_area_struct *vma, bool hold_rmap_lock)
> >  static int vma_link(struct mm_struct *mm, struct vm_area_struct *vma)
> >  {
> >       VMA_ITERATOR(vmi, mm, 0);
> > +     int err;
> >
> >       vma_iter_config(&vmi, vma->vm_start, vma->vm_end);
> >       if (vma_iter_prealloc(&vmi, vma))
> >               return -ENOMEM;
> >
> > -     vma_start_write(vma);
> > +     err = vma_start_write_killable(vma);
> > +     if (err) {
> > +             vma_iter_free(&vmi);
> > +             return err;
> > +     }
>
> Ditto here, ordering would mean no freeing.

Ack. Will move.

>
> >       vma_iter_store_new(&vmi, vma);
> >       vma_link_file(vma, /* hold_rmap_lock= */false);
> >       mm->map_count++;
> > @@ -2215,9 +2240,8 @@ int mm_take_all_locks(struct mm_struct *mm)
> >        * is reached.
> >        */
> >       for_each_vma(vmi, vma) {
> > -             if (signal_pending(current))
> > +             if (signal_pending(current) || vma_start_write_killable(vma))
> >                       goto out_unlock;
> > -             vma_start_write(vma);
> >       }
> >
> >       vma_iter_init(&vmi, mm, 0);
> > @@ -2522,6 +2546,11 @@ static int __mmap_new_vma(struct mmap_state *map, struct vm_area_struct **vmap)
> >       if (!vma)
> >               return -ENOMEM;
> >
> > +     /* Lock the VMA since it is modified after insertion into VMA tree */
> > +     error = vma_start_write_killable(vma);
> > +     if (error)
> > +             goto free_vma;
> > +
>
> There's no way this is going to fail, right?

You are technically correct but this is only true because of how
vma_start_write_killable() is implemented, so it's a technical detail
of the implementation, which might change in the future...
To clarify, __vma_start_exclude_readers() always succeeds if the vma
is detached (which is the case here) because in that case it bails out
before calling rcuwait_wait_event().

>
> >       vma_iter_config(vmi, map->addr, map->end);
> >       vma_set_range(vma, map->addr, map->end, map->pgoff);
> >       vm_flags_init(vma, map->vm_flags);
> > @@ -2552,8 +2581,6 @@ static int __mmap_new_vma(struct mmap_state *map, struct vm_area_struct **vmap)
> >       WARN_ON_ONCE(!arch_validate_flags(map->vm_flags));
> >  #endif
> >
> > -     /* Lock the VMA since it is modified after insertion into VMA tree */
> > -     vma_start_write(vma);
> >       vma_iter_store_new(vmi, vma);
> >       map->mm->map_count++;
> >       vma_link_file(vma, map->hold_file_rmap_lock);
> > @@ -2864,6 +2891,7 @@ int do_brk_flags(struct vma_iterator *vmi, struct vm_area_struct *vma,
> >                unsigned long addr, unsigned long len, vm_flags_t vm_flags)
> >  {
> >       struct mm_struct *mm = current->mm;
> > +     int err = -ENOMEM;
> >
> >       /*
> >        * Check against address space limits by the changed size
> > @@ -2908,7 +2936,10 @@ int do_brk_flags(struct vma_iterator *vmi, struct vm_area_struct *vma,
> >       vma_set_range(vma, addr, addr + len, addr >> PAGE_SHIFT);
> >       vm_flags_init(vma, vm_flags);
> >       vma->vm_page_prot = vm_get_page_prot(vm_flags);
> > -     vma_start_write(vma);
> > +     if (vma_start_write_killable(vma)) {
> > +             err = -EINTR;
> > +             goto mas_store_fail;
> > +     }
>
> I'd rather have another label saying write lock failed.  Really, this
> will never fail though (besides syzbot..)

Sure, I'll add another label.

>
> >       if (vma_iter_store_gfp(vmi, vma, GFP_KERNEL))
> >               goto mas_store_fail;
> >
> > @@ -2928,7 +2959,7 @@ int do_brk_flags(struct vma_iterator *vmi, struct vm_area_struct *vma,
> >       vm_area_free(vma);
> >  unacct_fail:
> >       vm_unacct_memory(len >> PAGE_SHIFT);
> > -     return -ENOMEM;
> > +     return err;
> >  }
> >
> >  /**
> > @@ -3089,7 +3120,7 @@ int expand_upwards(struct vm_area_struct *vma, unsigned long address)
> >       struct mm_struct *mm = vma->vm_mm;
> >       struct vm_area_struct *next;
> >       unsigned long gap_addr;
> > -     int error = 0;
> > +     int error;
> >       VMA_ITERATOR(vmi, mm, vma->vm_start);
> >
> >       if (!(vma->vm_flags & VM_GROWSUP))
> > @@ -3126,12 +3157,14 @@ int expand_upwards(struct vm_area_struct *vma, unsigned long address)
> >
> >       /* We must make sure the anon_vma is allocated. */
> >       if (unlikely(anon_vma_prepare(vma))) {
> > -             vma_iter_free(&vmi);
> > -             return -ENOMEM;
> > +             error = -ENOMEM;
> > +             goto free;
> >       }
> >
> >       /* Lock the VMA before expanding to prevent concurrent page faults */
> > -     vma_start_write(vma);
> > +     error = vma_start_write_killable(vma);
> > +     if (error)
> > +             goto free;
> >       /* We update the anon VMA tree. */
> >       anon_vma_lock_write(vma->anon_vma);
> >
> > @@ -3160,6 +3193,7 @@ int expand_upwards(struct vm_area_struct *vma, unsigned long address)
> >               }
> >       }
> >       anon_vma_unlock_write(vma->anon_vma);
> > +free:
> >       vma_iter_free(&vmi);
> >       validate_mm(mm);
> >       return error;
> > @@ -3174,7 +3208,7 @@ int expand_downwards(struct vm_area_struct *vma, unsigned long address)
> >  {
> >       struct mm_struct *mm = vma->vm_mm;
> >       struct vm_area_struct *prev;
> > -     int error = 0;
> > +     int error;
> >       VMA_ITERATOR(vmi, mm, vma->vm_start);
> >
> >       if (!(vma->vm_flags & VM_GROWSDOWN))
> > @@ -3205,12 +3239,14 @@ int expand_downwards(struct vm_area_struct *vma, unsigned long address)
> >
> >       /* We must make sure the anon_vma is allocated. */
> >       if (unlikely(anon_vma_prepare(vma))) {
> > -             vma_iter_free(&vmi);
> > -             return -ENOMEM;
> > +             error = -ENOMEM;
> > +             goto free;
> >       }
> >
> >       /* Lock the VMA before expanding to prevent concurrent page faults */
> > -     vma_start_write(vma);
> > +     error = vma_start_write_killable(vma);
> > +     if (error)
> > +             goto free;
> >       /* We update the anon VMA tree. */
> >       anon_vma_lock_write(vma->anon_vma);
> >
> > @@ -3240,6 +3276,7 @@ int expand_downwards(struct vm_area_struct *vma, unsigned long address)
> >               }
> >       }
> >       anon_vma_unlock_write(vma->anon_vma);
> > +free:
> >       vma_iter_free(&vmi);
> >       validate_mm(mm);
> >       return error;
> > diff --git a/mm/vma_exec.c b/mm/vma_exec.c
> > index 8134e1afca68..a4addc2a8480 100644
> > --- a/mm/vma_exec.c
> > +++ b/mm/vma_exec.c
> > @@ -40,6 +40,7 @@ int relocate_vma_down(struct vm_area_struct *vma, unsigned long shift)
> >       struct vm_area_struct *next;
> >       struct mmu_gather tlb;
> >       PAGETABLE_MOVE(pmc, vma, vma, old_start, new_start, length);
> > +     int err;
> >
> >       BUG_ON(new_start > new_end);
> >
> > @@ -55,8 +56,9 @@ int relocate_vma_down(struct vm_area_struct *vma, unsigned long shift)
> >        * cover the whole range: [new_start, old_end)
> >        */
> >       vmg.target = vma;
> > -     if (vma_expand(&vmg))
> > -             return -ENOMEM;
> > +     err = vma_expand(&vmg);
> > +     if (err)
> > +             return err;
> >
> >       /*
> >        * move the page tables downwards, on failure we rely on
> > --
> > 2.53.0.414.gf7e9f6c205-goog
> >
> >
>