Re: [PATCH mm-unstable v1] mm: add a total mapcount for large folios

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Thu, Aug 10, 2023 at 11:48:27AM +0100, Ryan Roberts wrote:
> > For PTE-mapped THP, it might be a bit bigger noise, although I doubt it is
> > really significant (judging from my experience on managing PageAnonExclusive
> > using set_bit/test_bit/clear_bit when (un)mapping anon pages).
> > 
> > As folio_add_file_rmap_range() indicates, for PTE-mapped THPs we should be
> > batching where possible (and Ryan is working on some more rmap batching). 
> 
> Yes, I've just posted [1] which batches the rmap removal. That would allow you
> to convert the per-page atomic_dec() into a (usually) single per-large-folio
> atomic_sub().
> 
> [1] https://lore.kernel.org/linux-mm/20230810103332.3062143-1-ryan.roberts@xxxxxxx/

Right, that'll definitely make more sense, thanks for the link; I'd be very
happy to read more later (finally I got some free time recently..).  But
then does it mean David's patch can be attached at the end instead of
proposed separately and early?

I was asking mostly because I read it as a standalone patch first, and
honestly I don't know the effect.  It's based on not only the added atomic
ops itself, but also the field changes.

For example, this patch moves Hugh's _nr_pages_mapped into the 2nd tail
page, I think it means for any rmap change of any small page of a huge one
we'll need to start touching one more 64B cacheline on x86.  I really have
no idea what does it mean for especially a large SMP: see 292648ac5cf1 on
why I had an impression of that.  But I've no enough experience or clue to
prove it a problem either, maybe would be interesting to measure the time
needed for some pte-mapped loops?  E.g., something like faulting in a thp,
then measure the split (by e.g. mprotect() at offset 1M on a 4K?) time it
takes before/after this patch.

When looking at this, I actually found one thing that is slightly
confusing, not directly relevant to your patch, but regarding the reuse of
tail page 1 on offset 24 bytes.  Current it's Hugh's _nr_pages_mapped,
and you're proposing to replace it with the total mapcount:

        atomic_t   _nr_pages_mapped;     /*    88     4 */

Now my question is.. isn't byte 24 of tail page 1 used for keeping a
poisoned mapping?  See prep_compound_tail() where it has:

	p->mapping = TAIL_MAPPING;

While here mapping is, afaict, also using offset 24 of the tail page 1:

        struct address_space * mapping;  /*    24     8 */

I hope I did a wrong math somewhere, though.

-- 
Peter Xu




[Index of Archives]     [Kernel Newbies]     [Security]     [Netfilter]     [Bugtraq]     [Linux FS]     [Yosemite Forum]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux RAID]     [Samba]     [Video 4 Linux]     [Device Mapper]     [Linux Resources]

  Powered by Linux