Re: Race condition observed between page migration and page fault handling on arm64 machines

From: Dev Jain
Date: Thu Aug 01 2024 - 05:38:55 EST



On 8/1/24 14:12, David Hildenbrand wrote:
On 01.08.24 10:16, Dev Jain wrote:
I and Ryan had a discussion and we thought it would be best to get feedback
from the community.

The migration mm selftest currently fails on arm64 for shared anon mappings,
due to the following race:

Do you mean MAP_SHARED|MAP_ANON or MAP_PRIVATE|MAP_ANON_fork? Because you note shmem below, I assume you mean MAP_SHARED|MAP_ANON


Yes.



Migration:                        Page fault:
try_to_migrate_one():                    handle_pte_fault():
1. Nuke the PTE                        PTE has been deleted => do_pte_missing()
2. Mark the PTE for migration                PTE has not been deleted but is just not "present" => do_swap_page()


In filemap_fault_recheck_pte_none() we recheck under PTL to make sure that a temporary pte_none() really was persistent pte_none() and not a temporary pte_none() under PTL.

Should we do something similar in do_fault()? I see that we already do something like that on the "!vma->vm_ops->fault" path.

But of course, there is a tradeoff between letting migration (temporarily) fail and grabbing the PTL during page faults.


To dampen the tradeoff, we could do this in shmem_fault() instead? But then, this would mean that we do this in all

kinds of vma->vm_ops->fault, only when we discover another reference count race condition :) Doing this in do_fault()

should solve this once and for all. In fact, do_pte_missing() may call do_anonymous_page() or do_fault(), and I just

noticed that the former already checks this using vmf_pte_changed().