On Fri, Jul 09, 2021 at 06:39:44AM +0000, HAGIO KAZUHITO(萩尾 一仁) wrote:
-----Original Message-----
> On Mon, Jun 21, 2021 at 06:02:51AM +0000, HAGIO KAZUHITO(萩尾 一仁) wrote:
> > -----Original Message-----
> > > +
> > > + if (read_pd(dd->dfd, offset, &pd)) {
> > > + /*
> > > + * Truncated page descriptor at most
> > > + * references full page.
> > > + */
> > > + expected_size += block_size;
> > > + goto next;
> > > + }
> > > +
> > > + if (pd.offset == 0) {
> > > + if (!first_empty_pd)
> > > + first_empty_pd = page_idx;
> > > + /*
> > > + * Incomplete pages at most use the
> > > + * whole page.
> > > + */
> > > + expected_size += block_size;
> > > + } else if (!pd.flags) {
> > > + /*
> > > + * Zero page has no compression flags.
> > > + */
> >
> > Non-compressed page also has no compression flags.
> >
> > So something like (pd.offset == dd->data_offset) is needed to determine
> > whether the page is zero page? although it's not exact without excluding
> > zero pages.
> >
>
> Hi Kazu,
>
> Yes, you're right. Thanks for spotting it.
>
> I've added a code path for a case when zero pages are not it's excluded.
> It's a no brainer. However, I've got some issues figuring out whether a
> page descriptor references zero page when we start to differentiate
> zero/non-zero uncompressed pages and calculate expected size accordingly.
>
> dd->data_offset points to the beginning of page descriptors and it
> doesn't help to find zero page:
>
> data_offset -> pd for pfn 0
> pd for pfn 1
> ...
> pd for pfn N
> ... <-some gap ???
> zero page
> pfn 0
Oh, you're right, I misread something.
There should be no gap between pd for pfn N and zero page, but anyway
we cannot get the number of pds in advance without counting the bitmap..
I don't know may be some parts of bitmap weren't flushed either. I'll
investigate that further why "valid_pages * descriptor size" is not
equal to offset from data_offset to zero page on an incomplete dump.
>
> The essense of what I've got so far is the following:
>
> if (page_idx == 0)
> first_page_offset = pd.offset;
>
> [...]
> } else if (!pd.flags) {
> /* The page is not compressed */
> if (dd->flags & ZERO_EXCLUDED) {
> /*
> * Zero page is located just before the first
> * dumpable pfn. The following pages are non-zero.
> */
> if (pd.offset >= first_page_offset)
> expected_size += block_size;
> } else {
> expected_size += block_size;
> }
> } else if (pd.flags) {
> [...]
>
> As you can see it would work but only if pd for pfn 0 doesn't reference
> zero page. Do you have any idea how can we reliably determine zero page
> offset?
It can be that pd for pfn0 points to zero page, especially considering
makedumpfile --split option. What I thought of are:
1. Two-pass check
Count the number of dumpable pages first, calculate the offset of zero page,
then check pd.offset.
Yeah, I also did think of two pass mode but wasn't sure if it's possible
to precalculate zero page otherwise :)
2. Online adjustment, i.e.
if (pd.offset > first_page_offset) // the first page is always regarded as zero
page
expected_size += block_size;
else if (pd.offset < first_page_offset) {
first_page_offset = pd.offset // should be zero page
expected_size += block_size; // for the first page missed
}
but this is a bit tricky and confusing.
Also, there is need to add one page size for zero page if ZERO_EXCLUDED
either way.
Personally I think that the former is better, but maybe depending on how
much time is consumed with large dumps.
Agreed, I'll try two pass mode if ZERO_EXCLUDED is enabled for
non-compressed pages. The single pass introduced in the series to read
page descriptors wasn't noticable on 10G dump.
Thanks,
Roman
Thanks,
Kazu