[CRIU] [PATCH 04/11 v3] criu: lazy-pages: replace page list with IOVs list

[Mike Rapoport]

On Tue, Nov 15, 2016 at 5:47 PM, Pavel Emelyanov <xemul at virtuozzo.com> wrote:
>
>>>> @@ -591,33 +627,33 @@ static int uffd_handle_page(struct lazy_pages_info *lpi, __u64 address,
>>>>
>>>>  static int handle_remaining_pages(struct lazy_pages_info *lpi, void *dest)
>>>>  {
>>>> -     struct uffd_pages_struct *uffd_pages;
>>>> -     int rc;
>>>> +     struct lazy_iovec *lazy_iov;
>>>> +     int nr_pages, i, err;
>>>> +     unsigned long addr;
>>>>
>>>> -     list_for_each_entry(uffd_pages, &lpi->pages, list) {
>>>> -             pr_debug("Checking remaining pages 0x%lx (flags 0x%x)\n",
>>>> -                      uffd_pages->addr, uffd_pages->flags);
>>>> -             if (uffd_pages->flags & UFFD_FLAG_SENT)
>>>> -                     continue;
>>>> +     lpi->pr.reset(&lpi->pr);
>>>>
>>>> -             rc = uffd_handle_page(lpi, uffd_pages->addr, dest);
>>>> -             if (rc < 0) {
>>>> -                     pr_err("Error during UFFD copy\n");
>>>> -                     return -1;
>>>> -             }
>>>> +     list_for_each_entry(lazy_iov, &lpi->iovs, l) {
>>>> +             nr_pages = lazy_iov->len / PAGE_SIZE;
>>>> +
>>>> +             for (i = 0; i < nr_pages; i++) {
>>>> +                     addr = lazy_iov->base + i * PAGE_SIZE;
>>>>
>>>> -             uffd_pages->flags |= UFFD_FLAG_SENT;
>>>> +                     err = uffd_handle_page(lpi, addr, dest);
>>>
>>> This place looks dangerous. The iovs can be quite large and requesting the whole
>>> IOV during single #PF handling may slow the restored tree down significantly.
>>
>> We never request more than one page at #PF time. The entire IOVs are
>> requested only after 5 secs timeout
>
> But here's the for (i = 0; i < nr_pages; i++) loop which does so, doesn't it?

These are graduate kernel-style small patches, and it seems I've
omitted some of them :)

So, at the end we'll get the following:
- at init time we create buffers large enough to hold largest IOV per process.
- when a #PF happens, we request exactly one page
- after there are no #PF's for 5 seconds, we start streaming the
remaining pages chunk by chunk


> -- Pavel



-- 
Sincerely yours,
Mike.