| lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
|
Open Source and information security mailing list archives
| ||
|
Message-ID: <Y2llcRiDLHc2kg/N@cmpxchg.org>
Date: Mon, 7 Nov 2022 15:07:13 -0500
From: Johannes Weiner <hannes@...xchg.org>
To: Hugh Dickins <hughd@...gle.com>
Cc: Linus Torvalds <torvalds@...ux-foundation.org>,
Stephen Rothwell <sfr@...b.auug.org.au>,
Alexander Gordeev <agordeev@...ux.ibm.com>,
Peter Zijlstra <peterz@...radead.org>,
Will Deacon <will@...nel.org>,
Aneesh Kumar <aneesh.kumar@...ux.ibm.com>,
Nick Piggin <npiggin@...il.com>,
Heiko Carstens <hca@...ux.ibm.com>,
Vasily Gorbik <gor@...ux.ibm.com>,
Christian Borntraeger <borntraeger@...ux.ibm.com>,
Sven Schnelle <svens@...ux.ibm.com>,
Nadav Amit <nadav.amit@...il.com>,
Jann Horn <jannh@...gle.com>,
John Hubbard <jhubbard@...dia.com>, X86 ML <x86@...nel.org>,
Matthew Wilcox <willy@...radead.org>,
Andrew Morton <akpm@...ux-foundation.org>,
kernel list <linux-kernel@...r.kernel.org>,
Linux-MM <linux-mm@...ck.org>,
Andrea Arcangeli <aarcange@...hat.com>,
"Kirill A . Shutemov" <kirill.shutemov@...ux.intel.com>,
Joerg Roedel <jroedel@...e.de>,
Uros Bizjak <ubizjak@...il.com>,
Alistair Popple <apopple@...dia.com>,
linux-arch <linux-arch@...r.kernel.org>
Subject: Re: mm: delay rmap removal until after TLB flush
Hello,
On Sun, Nov 06, 2022 at 01:06:19PM -0800, Hugh Dickins wrote:
> lock_page_memcg (uncertain)
> ---------------------------
> Johannes, please help! Linus has got quite upset by lock_page_memcg(),
> its calls from mm/rmap.c anyway, and most particularly by the way
> in which it is called at the start of page_remove_rmap(), before
> anyone's critical atomic_add_negative(), yet its use is to guarantee
> the stability of page memcg while doing the stats updates, done only
> when atomic_add_negative() says so.
As you mentioned, the pte lock historically wasn't always taken on the
move side. And so the reason lock_page_memcg() covers the mapcount
update is that move_account() needs to atomically see either a) the
page is mapped and counted, or b) unmapped and uncounted. If we lock
after mapcountdec, move_account() could miss pending updates that need
transferred, and it would break the scheme breaks thusly:
memcg1->nr_mapped = 1
memcg2->nr_mapped = 0
page_remove_rmap: mem_cgroup_move_account():
if atomic_add_negative(page->mapcount):
lock_page_memcg()
if page->mapcount: // NOT TAKEN
memcg1->nr_mapped--
memcg2->nr_mapped++
page->memcg = memcg2
unlock_page_memcg()
lock_page_memcg()
page->memcg->nr_mapped-- // UNDERFLOW memcg2->nr_mapped
unlock_page_memcg()
> I do have one relevant insight on this. It (or its antecedents under
> other names) date from the days when we did "reparenting" of memcg
> charges from an LRU: and in those days the lock_page_memcg() before
> mapcount adjustment was vital, to pair with the uses of folio_mapped()
> or page_mapped() in mem_cgroup_move_account() - those "mapped" checks
> are precisely around the stats which the rmap functions affect.
>
> But nowadays mem_cgroup_move_account() is only called, with page table
> lock held, on matching pages found in a task's page table: so its
> "mapped" checks are redundant - I've sometimes thought in the past of
> removing them, but held back, because I always have the notion (not
> hope!) that "reparenting" may one day be brought back from the grave.
> I'm too out of touch with memcg to know where that notion stands today.
>
> I've gone through a multiverse of opinions on those lock_page_memcg()s
> in the last day: I currently believe that Linus is right, that the
> lock_page_memcg()s could and should be moved just before the stats
> updates. But I am not 100% certain of that - is there still some
> reason why it's important that the page memcg at the instant of the
> critical mapcount transition be kept unchanged until the stats are
> updated? I've tried running scenarios through my mind but given up.
Okay, I think there are two options.
- If we don't want to codify the pte lock requirement on the move
side, then moving the lock_page_memcg() like that would break the
locking scheme, as per above.
- If we DO want to codify the pte lock requirement, we should just
remove the lock_page_memcg() altogether, as it's fully redundant.
I'm leaning toward the second option. If somebody brings back
reparenting they can bring the lock back along with it.
[ If it's even still necessary by then.
It's conceivable reparenting is brought back only for cgroup2, where
the race wouldn't matter because of the hierarchical stats. The
reparenting side wouldn't have to move page state to the parent -
it's already there. And whether rmap would see the dying child or
the parent doesn't matter much either: the parent will see the
update anyway, directly or recursively, and we likely don't care to
balance the books on a dying cgroup.
It's then just a matter of lifetime - which should be guaranteed
also, as long as the pte lock prevents an RCU quiescent state. ]
So how about something like below?
UNTESTED, just for illustration. This is cgroup1 code, which I haven't
looked at too closely in a while. If you can't spot an immediate hole
in it, I'd go ahead and test it and send a proper patch.
---
>From 88a32b1b5737630fb981114f6333d8fd057bd8e9 Mon Sep 17 00:00:00 2001
From: Johannes Weiner <hannes@...xchg.org>
Date: Mon, 7 Nov 2022 12:05:09 -0500
Subject: [PATCH] mm: remove lock_page_memcg() from rmap
rmap changes (mapping and unmapping) of a page currently take
lock_page_memcg() to serialize 1) update of the mapcount and the
cgroup mapped counter with 2) cgroup moving the page and updating the
old cgroup and the new cgroup counters based on page_mapped().
Before b2052564e66d ("mm: memcontrol: continue cache reclaim from
offlined groups"), we used to reassign all pages that could be found
on a cgroup's LRU list on deletion - something that rmap didn't
naturally serialize against. Since that commit, however, the only
pages that get moved are those mapped into page tables of a task
that's being migrated. In that case, the pte lock is always held (and
we know the page is mapped), which keeps rmap changes at bay already.
The additional lock_page_memcg() by rmap is redundant. Remove it.
NOT-Signed-off-by: Johannes Weiner <hannes@...xchg.org>
---
mm/memcontrol.c | 27 ++++++++++++---------------
mm/rmap.c | 30 ++++++++++++------------------
2 files changed, 24 insertions(+), 33 deletions(-)
diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index 2d8549ae1b30..f7716e9038e9 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -5666,7 +5666,10 @@ static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
* @from: mem_cgroup which the page is moved from.
* @to: mem_cgroup which the page is moved to. @from != @to.
*
- * The caller must make sure the page is not on LRU (isolate_page() is useful.)
+ * This function acquires folio_lock() and folio_lock_memcg(). The
+ * caller must exclude all other possible ways of accessing
+ * page->memcg, such as LRU isolation (to lock out isolation) and
+ * having the page mapped and pte-locked (to lock out rmap).
*
* This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
* from old cgroup.
@@ -5685,6 +5688,7 @@ static int mem_cgroup_move_account(struct page *page,
VM_BUG_ON(from == to);
VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
VM_BUG_ON(compound && !folio_test_large(folio));
+ VM_WARN_ON_ONCE(!folio_mapped(folio));
/*
* Prevent mem_cgroup_migrate() from looking at
@@ -5705,30 +5709,23 @@ static int mem_cgroup_move_account(struct page *page,
folio_memcg_lock(folio);
if (folio_test_anon(folio)) {
- if (folio_mapped(folio)) {
- __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
- __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
- if (folio_test_transhuge(folio)) {
- __mod_lruvec_state(from_vec, NR_ANON_THPS,
- -nr_pages);
- __mod_lruvec_state(to_vec, NR_ANON_THPS,
- nr_pages);
- }
+ __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
+ __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
+ if (folio_test_transhuge(folio)) {
+ __mod_lruvec_state(from_vec, NR_ANON_THPS, -nr_pages);
+ __mod_lruvec_state(to_vec, NR_ANON_THPS, nr_pages);
}
} else {
__mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
__mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
+ __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
+ __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
if (folio_test_swapbacked(folio)) {
__mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
__mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
}
- if (folio_mapped(folio)) {
- __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
- __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
- }
-
if (folio_test_dirty(folio)) {
struct address_space *mapping = folio_mapping(folio);
diff --git a/mm/rmap.c b/mm/rmap.c
index 2ec925e5fa6a..60c31375f274 100644
--- a/mm/rmap.c
+++ b/mm/rmap.c
@@ -1197,11 +1197,6 @@ void page_add_anon_rmap(struct page *page,
bool compound = flags & RMAP_COMPOUND;
bool first;
- if (unlikely(PageKsm(page)))
- lock_page_memcg(page);
- else
- VM_BUG_ON_PAGE(!PageLocked(page), page);
-
if (compound) {
atomic_t *mapcount;
VM_BUG_ON_PAGE(!PageLocked(page), page);
@@ -1217,19 +1212,19 @@ void page_add_anon_rmap(struct page *page,
if (first) {
int nr = compound ? thp_nr_pages(page) : 1;
/*
- * We use the irq-unsafe __{inc|mod}_zone_page_stat because
- * these counters are not modified in interrupt context, and
- * pte lock(a spinlock) is held, which implies preemption
- * disabled.
+ * We use the irq-unsafe __{inc|mod}_zone_page_stat
+ * because these counters are not modified in
+ * interrupt context, and pte lock(a spinlock) is
+ * held, which implies preemption disabled.
+ *
+ * The pte lock also stabilizes page->memcg wrt
+ * mem_cgroup_move_account().
*/
if (compound)
__mod_lruvec_page_state(page, NR_ANON_THPS, nr);
__mod_lruvec_page_state(page, NR_ANON_MAPPED, nr);
}
- if (unlikely(PageKsm(page)))
- unlock_page_memcg(page);
-
/* address might be in next vma when migration races vma_adjust */
else if (first)
__page_set_anon_rmap(page, vma, address,
@@ -1290,7 +1285,6 @@ void page_add_file_rmap(struct page *page,
int i, nr = 0;
VM_BUG_ON_PAGE(compound && !PageTransHuge(page), page);
- lock_page_memcg(page);
if (compound && PageTransHuge(page)) {
int nr_pages = thp_nr_pages(page);
@@ -1311,6 +1305,7 @@ void page_add_file_rmap(struct page *page,
if (nr == nr_pages && PageDoubleMap(page))
ClearPageDoubleMap(page);
+ /* The pte lock stabilizes page->memcg */
if (PageSwapBacked(page))
__mod_lruvec_page_state(page, NR_SHMEM_PMDMAPPED,
nr_pages);
@@ -1328,7 +1323,6 @@ void page_add_file_rmap(struct page *page,
out:
if (nr)
__mod_lruvec_page_state(page, NR_FILE_MAPPED, nr);
- unlock_page_memcg(page);
mlock_vma_page(page, vma, compound);
}
@@ -1356,6 +1350,7 @@ static void page_remove_file_rmap(struct page *page, bool compound)
}
if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
goto out;
+ /* The pte lock stabilizes page->memcg */
if (PageSwapBacked(page))
__mod_lruvec_page_state(page, NR_SHMEM_PMDMAPPED,
-nr_pages);
@@ -1423,8 +1418,6 @@ static void page_remove_anon_compound_rmap(struct page *page)
void page_remove_rmap(struct page *page,
struct vm_area_struct *vma, bool compound)
{
- lock_page_memcg(page);
-
if (!PageAnon(page)) {
page_remove_file_rmap(page, compound);
goto out;
@@ -1443,6 +1436,9 @@ void page_remove_rmap(struct page *page,
* We use the irq-unsafe __{inc|mod}_zone_page_stat because
* these counters are not modified in interrupt context, and
* pte lock(a spinlock) is held, which implies preemption disabled.
+ *
+ * The pte lock also stabilizes page->memcg wrt
+ * mem_cgroup_move_account().
*/
__dec_lruvec_page_state(page, NR_ANON_MAPPED);
@@ -1459,8 +1455,6 @@ void page_remove_rmap(struct page *page,
* faster for those pages still in swapcache.
*/
out:
- unlock_page_memcg(page);
-
munlock_vma_page(page, vma, compound);
}
--
2.38.1
Powered by blists - more mailing lists