[PATCH] secretmem: optimize page_is_secretmem()
Mike Rapoport
rppt at kernel.org
Mon Apr 19 11:14:19 BST 2021
On Mon, Apr 19, 2021 at 11:40:56AM +0200, David Hildenbrand wrote:
> On 19.04.21 11:38, David Hildenbrand wrote:
> > On 19.04.21 11:36, Mike Rapoport wrote:
> > > On Mon, Apr 19, 2021 at 11:15:02AM +0200, David Hildenbrand wrote:
> > > > On 19.04.21 10:42, Mike Rapoport wrote:
> > > > > From: Mike Rapoport <rppt at linux.ibm.com>
> > > > >
> > > > > Kernel test robot reported -4.2% regression of will-it-scale.per_thread_ops
> > > > > due to commit "mm: introduce memfd_secret system call to create "secret"
> > > > > memory areas".
> > > > >
> > > > > The perf profile of the test indicated that the regression is caused by
> > > > > page_is_secretmem() called from gup_pte_range() (inlined by gup_pgd_range):
> > > > >
> > > > > 27.76 +2.5 30.23 perf-profile.children.cycles-pp.gup_pgd_range
> > > > > 0.00 +3.2 3.19 ± 2% perf-profile.children.cycles-pp.page_mapping
> > > > > 0.00 +3.7 3.66 ± 2% perf-profile.children.cycles-pp.page_is_secretmem
> > > > >
> > > > > Further analysis showed that the slow down happens because neither
> > > > > page_is_secretmem() nor page_mapping() are not inline and moreover,
> > > > > multiple page flags checks in page_mapping() involve calling
> > > > > compound_head() several times for the same page.
> > > > >
> > > > > Make page_is_secretmem() inline and replace page_mapping() with page flag
> > > > > checks that do not imply page-to-head conversion.
> > > > >
> > > > > Reported-by: kernel test robot <oliver.sang at intel.com>
> > > > > Signed-off-by: Mike Rapoport <rppt at linux.ibm.com>
> > > > > ---
> > > > >
> > > > > @Andrew,
> > > > > The patch is vs v5.12-rc7-mmots-2021-04-15-16-28, I'd appreciate if it would
> > > > > be added as a fixup to the memfd_secret series.
> > > > >
> > > > > include/linux/secretmem.h | 26 +++++++++++++++++++++++++-
> > > > > mm/secretmem.c | 12 +-----------
> > > > > 2 files changed, 26 insertions(+), 12 deletions(-)
> > > > >
> > > > > diff --git a/include/linux/secretmem.h b/include/linux/secretmem.h
> > > > > index 907a6734059c..b842b38cbeb1 100644
> > > > > --- a/include/linux/secretmem.h
> > > > > +++ b/include/linux/secretmem.h
> > > > > @@ -4,8 +4,32 @@
> > > > > #ifdef CONFIG_SECRETMEM
> > > > > +extern const struct address_space_operations secretmem_aops;
> > > > > +
> > > > > +static inline bool page_is_secretmem(struct page *page)
> > > > > +{
> > > > > + struct address_space *mapping;
> > > > > +
> > > > > + /*
> > > > > + * Using page_mapping() is quite slow because of the actual call
> > > > > + * instruction and repeated compound_head(page) inside the
> > > > > + * page_mapping() function.
> > > > > + * We know that secretmem pages are not compound and LRU so we can
> > > > > + * save a couple of cycles here.
> > > > > + */
> > > > > + if (PageCompound(page) || !PageLRU(page))
> > > > > + return false;
> > > >
> > > > I'd assume secretmem pages are rare in basically every setup out there. So
> > > > maybe throwing in a couple of likely()/unlikely() might make sense.
> > >
> > > I'd say we could do unlikely(page_is_secretmem()) at call sites. Here I can
> > > hardly estimate which pages are going to be checked.
> > > > > +
> > > > > + mapping = (struct address_space *)
> > > > > + ((unsigned long)page->mapping & ~PAGE_MAPPING_FLAGS);
> > > > > +
> > > >
> > > > Not sure if open-coding page_mapping is really a good idea here -- or even
> > > > necessary after the fast path above is in place. Anyhow, just my 2 cents.
> > >
> > > Well, most if the -4.2% of the performance regression kbuild reported were
> > > due to repeated compount_head(page) in page_mapping(). So the whole point
> > > of this patch is to avoid calling page_mapping().
> >
> > I would have thought the fast path "(PageCompound(page) ||
> > !PageLRU(page))" would already avoid calling page_mapping() in many cases.
>
> (and I do wonder if a generic page_mapping() optimization would make sense
> instead)
Not sure. Replacing page_mapping() with page_file_mapping() at the
call sites at fs/ and mm/ increased the defconfig image by nearly 2k
and page_file_mapping() is way simpler than page_mapping()
add/remove: 1/0 grow/shrink: 35/0 up/down: 1960/0 (1960)
Function old new delta
shrink_page_list 3414 3670 +256
__set_page_dirty_nobuffers 242 349 +107
check_move_unevictable_pages 904 987 +83
move_to_new_page 591 671 +80
shrink_active_list 912 970 +58
move_pages_to_lru 911 965 +54
migrate_pages 2500 2554 +54
shmem_swapin_page 1145 1197 +52
shmem_undo_range 1669 1719 +50
__test_set_page_writeback 620 670 +50
__set_page_dirty_buffers 187 237 +50
__pagevec_lru_add 757 807 +50
__munlock_pagevec 1155 1205 +50
__dump_page 1101 1151 +50
__cancel_dirty_page 182 232 +50
__remove_mapping 461 510 +49
rmap_walk_file 402 449 +47
isolate_movable_page 240 287 +47
test_clear_page_writeback 668 714 +46
page_cache_pipe_buf_try_steal 171 217 +46
page_endio 246 290 +44
page_file_mapping - 43 +43
__isolate_lru_page_prepare 254 297 +43
hugetlb_page_mapping_lock_write 39 81 +42
iomap_set_page_dirty 110 151 +41
clear_page_dirty_for_io 324 364 +40
wait_on_page_writeback_killable 118 157 +39
wait_on_page_writeback 105 144 +39
set_page_dirty 159 198 +39
putback_movable_page 32 71 +39
page_mkclean 172 211 +39
mark_buffer_dirty 176 215 +39
invalidate_inode_page 122 161 +39
delete_from_page_cache 139 178 +39
PageMovable 49 86 +37
isolate_migratepages_block 2843 2872 +29
Total: Before=17068648, After=17070608, chg +0.01%
> Willy can most probably give the best advise here :)
I think that's what folios are for :)
--
Sincerely yours,
Mike.
More information about the linux-riscv
mailing list