mm: do not let vdso pages into LRU rotation

Hi,

I noticed that vdso pages are faulted and unmapped as if they were regular file pages. And I'm guessing this is so that the vdso mappings are able to use the generic COW code in memory.c.

However, it's a little unsettling that zap_pte_range() makes decisions based on PageAnon() and the page even reaches mark_page_accessed(), as that function makes several assumptions about the page being a regular LRU user page. It seems this isn't crashing today by sheer luck, but I am working on code that does when page_is_file_cache() returns garbage.

I'm using this hack to work around it:

mm/memory.c +2 -1

index c387430f06c3..f0537c500150 100644
@@ -1121,7 +1121,8 @@ again:
                                        set_page_dirty(page);
                                }
                                if (pte_young(ptent) &&
-                                   likely(!(vma->vm_flags & VM_SEQ_READ)))
+                                   likely(!(vma->vm_flags & VM_SEQ_READ)) &&
+                                   !PageReserved(page))
                                        mark_page_accessed(page);
                                rss[MM_FILEPAGES]--;
                        }

but I think we need a cleaner (and more robust) solution there to make it clearer that these pages are not regularly managed pages.

Could the VDSO be a VM_MIXEDMAP to keep the initial unmanaged pages out of the VM while allowing COW into regular anonymous pages?

Are there other requirements of the VDSO that I might be missing?

Quoted 2 lines · Johannes Weiner

Could the VDSO be a VM_MIXEDMAP to keep the initial unmanaged pages out of the VM while allowing COW into regular anonymous pages?

Probably. What are its limitations? We want ptrace to work on it, and mprotect needs to work and allow COW. access_process_vm should probably work, too.

Quoted 7 lines · Andy Lutomirski

Quoted 2 lines · Johannes Weiner

Could the VDSO be a VM_MIXEDMAP to keep the initial unmanaged pages out of the VM while allowing COW into regular anonymous pages?

Probably. What are its limitations? We want ptrace to work on it, and mprotect needs to work and allow COW. access_process_vm should probably work, too.

Thanks, that's good to know.

However, after looking at this a little longer, it appears this would need work in do_wp_page() to support non-page COW copying, then adding vm_ops->access and complicating ->fault in all VDSO implementations.

And it looks like - at least theoretically - drivers can inject non-VM pages into the page tables as well (comment above insert_page())

Given that this behavior has been around for a long time (the comment at the bottom of vm_normal_page is ancient), I'll probably go with a more conservative approach; add a comment to mark_page_accessed() and filter out non-VM pages in the function I'm going to call from it.