On Tue, Aug 5, 2025 at 7:18 AM Vlastimil Babka <vbabka@xxxxxxx> wrote:
>
> On 8/5/25 1:15 AM, Suren Baghdasaryan wrote:
> > Utilize per-vma locks to stabilize vma after lookup without taking
> > mmap_lock during PROCMAP_QUERY ioctl execution. If vma lock is
> > contended, we fall back to mmap_lock but take it only momentarily
> > to lock the vma and release the mmap_lock. In a very unlikely case
> > of vm_refcnt overflow, this fall back path will fail and ioctl is
> > done under mmap_lock protection.
> >
> > This change is designed to reduce mmap_lock contention and prevent
> > PROCMAP_QUERY ioctl calls from blocking address space updates.
> >
> > Signed-off-by: Suren Baghdasaryan <surenb@xxxxxxxxxx>
> > ---
> > fs/proc/task_mmu.c | 81 +++++++++++++++++++++++++++++++++++++---------
> > 1 file changed, 65 insertions(+), 16 deletions(-)
> >
> > diff --git a/fs/proc/task_mmu.c b/fs/proc/task_mmu.c
> > index 843577aa7a32..1d06ecdbef6f 100644
> > --- a/fs/proc/task_mmu.c
> > +++ b/fs/proc/task_mmu.c
> > @@ -517,28 +517,78 @@ static int pid_maps_open(struct inode *inode, struct file *file)
> > PROCMAP_QUERY_VMA_FLAGS \
> > )
> >
> > -static int query_vma_setup(struct mm_struct *mm)
> > +#ifdef CONFIG_PER_VMA_LOCK
> > +
> > +static int query_vma_setup(struct proc_maps_locking_ctx *lock_ctx)
> > {
> > - return mmap_read_lock_killable(mm);
> > + lock_ctx->locked_vma = NULL;
> > + lock_ctx->mmap_locked = false;
> > +
> > + return 0;
> > }
> >
> > -static void query_vma_teardown(struct mm_struct *mm, struct vm_area_struct *vma)
> > +static void query_vma_teardown(struct proc_maps_locking_ctx *lock_ctx)
> > {
> > - mmap_read_unlock(mm);
> > + if (lock_ctx->mmap_locked)
> > + mmap_read_unlock(lock_ctx->mm);
> > + else
> > + unlock_vma(lock_ctx);
> > }
> >
> > -static struct vm_area_struct *query_vma_find_by_addr(struct mm_struct *mm, unsigned long addr)
> > +static struct vm_area_struct *query_vma_find_by_addr(struct proc_maps_locking_ctx *lock_ctx,
> > + unsigned long addr)
> > {
> > - return find_vma(mm, addr);
> > + struct vm_area_struct *vma;
> > + struct vma_iterator vmi;
> >
>
> Hm I think we can reach here with lock_ctx->mmap_locked being true via
> "goto next_vma" in query_matching_vma(). In that case we should just
> "return find_vma()" and doing the below is wrong, no?
Ah, you are quite right. I should handle mmap_locked differently in
query_vma_find_by_addr(). I will post the fix shortly.
>
> > + unlock_vma(lock_ctx);
> > + rcu_read_lock();
> > + vma_iter_init(&vmi, lock_ctx->mm, addr);
> > + vma = lock_next_vma(lock_ctx->mm, &vmi, addr);
> > + rcu_read_unlock();
> > +
> > + if (!IS_ERR_OR_NULL(vma)) {
> > + lock_ctx->locked_vma = vma;
> > + } else if (PTR_ERR(vma) == -EAGAIN) {
> > + /* Fallback to mmap_lock on vma->vm_refcnt overflow */
> > + mmap_read_lock(lock_ctx->mm);
> > + vma = find_vma(lock_ctx->mm, addr);
> > + lock_ctx->mmap_locked = true;
> > + }
> > +
> > + return vma;
> > }
> >
