[PATCH] holepunch: fix mmap_sem i_mutex deadlock

sys_madvise has down_write of mmap_sem, then madvise_remove calls
vmtruncate_range which takes i_mutex and i_alloc_sem: no, we can easily devise
deadlocks from that ordering.

madvise_remove drop mmap_sem while calling vmtruncate_range: luckily, since
madvise_remove doesn't split or merge vmas, it's easy to handle this case with
a NULL prev, without restructuring sys_madvise.  (Though sad to retake
mmap_sem when it's unlikely to be needed, and certainly down_read is
sufficient for MADV_REMOVE, unlike the other madvices.)

Signed-off-by: Hugh Dickins <hugh@veritas.com>
Cc: Miklos Szeredi <mszeredi@suse.cz>
Cc: Badari Pulavarty <pbadari@us.ibm.com>
Cc: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/madvise.c

@@ -160,8 +160,9 @@ static long madvise_remove(struct vm_area_struct *vma,
 {
 	struct address_space *mapping;
 	loff_t offset, endoff;
+	int error;
 
-	*prev = vma;
+	*prev = NULL;	/* tell sys_madvise we drop mmap_sem */
 
 	if (vma->vm_flags & (VM_LOCKED|VM_NONLINEAR|VM_HUGETLB))
 		return -EINVAL;
@@ -180,7 +181,12 @@ static long madvise_remove(struct vm_area_struct *vma,
 			+ ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 	endoff = (loff_t)(end - vma->vm_start - 1)
 			+ ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
-	return  vmtruncate_range(mapping->host, offset, endoff);
+
+	/* vmtruncate_range needs to take i_mutex and i_alloc_sem */
+	up_write(&current->mm->mmap_sem);
+	error = vmtruncate_range(mapping->host, offset, endoff);
+	down_write(&current->mm->mmap_sem);
+	return error;
 }
 
 static long
@@ -315,12 +321,15 @@ asmlinkage long sys_madvise(unsigned long start, size_t len_in, int behavior)
 		if (error)
 			goto out;
 		start = tmp;
-		if (start < prev->vm_end)
+		if (prev && start < prev->vm_end)
 			start = prev->vm_end;
 		error = unmapped_error;
 		if (start >= end)
 			goto out;
+		if (prev)
 			vma = prev->vm_next;
+		else	/* madvise_remove dropped mmap_sem */
+			vma = find_vma(current->mm, start);
 	}
 out:
 	up_write(&current->mm->mmap_sem);