Re: [PATCH 1/3] uprobes: install_breakpoint() should fail if is_swbp_insn() == T

[Oleg Nesterov <oleg@redhat.com>]

On 05/30, Peter Zijlstra wrote:
>
> register's vma
> iteration is very careful not to have the same vma twice,

Hmm. I am wondering if it is careful enough...

Just in case, I think your patch is great. But it seems to me
there is another problem.

__find_next_vma_info() checks tmpvi->mm == vma->vm_mm to detect the
already visited mm/vma. However, afaics this can be false positive?

The caller, register_for_each_vma(), does mmput() and after that
this memory can be freed and re-used as another mm_struct.

I'll recheck this, but perhaps we need something like below?

Oleg.

--- x/kernel/events/uprobes.c
+++ x/kernel/events/uprobes.c
@@ -851,7 +851,6 @@ static int register_for_each_vma(struct 
 			list_del(&vi->probe_list);
 			kfree(vi);
 			up_write(&mm->mmap_sem);
-			mmput(mm);
 			continue;
 		}
 		vaddr = vma_address(vma, uprobe->offset);
@@ -860,7 +859,6 @@ static int register_for_each_vma(struct 
 			list_del(&vi->probe_list);
 			kfree(vi);
 			up_write(&mm->mmap_sem);
-			mmput(mm);
 			continue;
 		}
 
@@ -870,7 +868,6 @@ static int register_for_each_vma(struct 
 			remove_breakpoint(uprobe, mm, vi->vaddr);
 
 		up_write(&mm->mmap_sem);
-		mmput(mm);
 		if (is_register) {
 			if (ret && ret == -EEXIST)
 				ret = 0;
@@ -881,6 +878,7 @@ static int register_for_each_vma(struct 
 
 	list_for_each_entry_safe(vi, tmpvi, &try_list, probe_list) {
 		list_del(&vi->probe_list);
+		mmput(vi->mm)
 		kfree(vi);
 	}