Re: [PATCH] mm/memcg: use get_page() for device private pages in mc_handle_swap_pte()

[Miaohe Lin <linmiaohe@huawei.com>]

On 2023/7/15 11:56, Matthew Wilcox wrote:
> On Sat, Jul 15, 2023 at 11:28:02AM +0800, Miaohe Lin wrote:
>> When page table locked is held, the page can't be freed from under us.
> 
> But the page isn't mapped into the page table ... there's a swap entry
> in the page table, so I don't think your logic holds.
> 

IIUC, device_private_entry will hold one page refcnt when it's set to page table.
And there's similar code in do_swap_page():

  vm_fault_t do_swap_page(struct vm_fault *vmf)
    if (unlikely(non_swap_entry(entry))) {
      if (is_device_private_entry(entry))
        /*
         * Get a page reference while we know the page can't be
         * freed.
         */
        get_page(vmf->page);
        pte_unmap_unlock(vmf->pte, vmf->ptl);
        ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
        put_page(vmf->page);
    ...

If my logic doesn't hold, do_swap_page() will need to fix the code. Or am I miss something?

Thanks Matthew.