Re: [RFC PATCH v2 1/2] mm/mremap: Fix uprobe anon page be overwritten when expanding vma during mremap

[Lorenzo Stoakes <lorenzo.stoakes@oracle.com>]

On Tue, May 27, 2025 at 05:17:15PM +0200, Oleg Nesterov wrote:
> On 05/27, Lorenzo Stoakes wrote:
> >
> > I am by the way assuming that uprobes work by installing a special PTE at
> > page offset 0 and only in the case where there is something installed here
> > do we need to worry.
>
> perhaps I misunderstood you but no...
>
> Basically, we have uprobe_register(struct inode *inode, loff_t offset, ...).
> If / when a process mmaps this inode/file and the (new) VMA includes this offset,
> we need to call uprobe_mmap() to install the breakpoint at the virtual address
> which corresponds to this offset.

OK this was based on the explanation given re: zero offset.

So presumably then this is determined by the perf_event_attr.probe_offset field?

OK well we can drop the vm_pgoff predicate and things should still work because,
after the move:

            A                          B
|-----------------------||--------------------------|
| VMA to be merged with || VMA being moved in place |
|-----------------------||--------------------------|

If A contains the relevant breakpoint(s), then they will stay in place, as
move_page_tables() won't overwrite them.

If B contains them, then the new mechanism that prevents uprobe_mmap() from
installing a new one (orphaning the old) will make everything right.

So we simply take my approach, and drop the vm_pgoff condition and everything
should work fine :)

Obviously let me know if you see any flaw in this.

>
> Oleg.
>

Cheers, Lorenzo