Re: [PATCH v2] workingset: ensure memcg is valid for recency check

From: Yosry Ahmed
Date: Fri Aug 18 2023 - 12:32:18 EST

On Fri, Aug 18, 2023 at 9:24 AM Nhat Pham <nphamcs@xxxxxxxxx> wrote:
>
> On Fri, Aug 18, 2023 at 7:57 AM Yosry Ahmed <yosryahmed@xxxxxxxxxx> wrote:
> >
> > On Fri, Aug 18, 2023 at 6:49 AM Johannes Weiner <hannes@xxxxxxxxxxx> wrote:
> > >
> > > On Thu, Aug 17, 2023 at 05:12:17PM -0600, Yu Zhao wrote:
> > > > On Thu, Aug 17, 2023 at 4:50 PM Yosry Ahmed <yosryahmed@xxxxxxxxxx> wrote:
> > > > >
> > > > > On Thu, Aug 17, 2023 at 3:43 PM Nhat Pham <nphamcs@xxxxxxxxx> wrote:
> > > > > >
> > > > > > On Thu, Aug 17, 2023 at 1:50 PM Yosry Ahmed <yosryahmed@xxxxxxxxxx> wrote:
> > > > > > >
> > > > > > > On Thu, Aug 17, 2023 at 12:01 PM Nhat Pham <nphamcs@xxxxxxxxx> wrote:
> > > > > > > >
> > > > > > > > In eviction recency check, we are currently not holding a local
> > > > > > > > reference to the memcg that the refaulted folio belonged to when it was
> > > > > > > > evicted. This could cause serious memcg lifetime issues, for e.g in the
> > > > > > > > memcg hierarchy traversal done in mem_cgroup_get_nr_swap_pages(). This
> > > > > > > > has occurred in production:
> > > > > > > >
> > > > > > > > [ 155757.793456] BUG: kernel NULL pointer dereference, address: 00000000000000c0
> > > > > > > > [ 155757.807568] #PF: supervisor read access in kernel mode
> > > > > > > > [ 155757.818024] #PF: error_code(0x0000) - not-present page
> > > > > > > > [ 155757.828482] PGD 401f77067 P4D 401f77067 PUD 401f76067 PMD 0
> > > > > > > > [ 155757.839985] Oops: 0000 [#1] SMP
> > > > > > > > [ 155757.846444] CPU: 7 PID: 1380944 Comm: ThriftSrv-pri3- Kdump: loaded Tainted: G S 6.4.3-0_fbk1_rc0_594_g8d0cbcaa67ba #1
> > > > > > > > [ 155757.870808] Hardware name: Wiwynn Twin Lakes MP/Twin Lakes Passive MP, BIOS YMM16 05/24/2021
> > > > > > > > [ 155757.887870] RIP: 0010:mem_cgroup_get_nr_swap_pages+0x3d/0xb0
> > > > > > > > [ 155757.899377] Code: 29 19 4a 02 48 39 f9 74 63 48 8b 97 c0 00 00 00 48 8b b7 58 02 00 00 48 2b b7 c0 01 00 00 48 39 f0 48 0f 4d c6 48 39 d1 74 42 <48> 8b b2 c0 00 00 00 48 8b ba 58 02 00 00 48 2b ba c0 01 00 00 48
> > > > > > > > [ 155757.937125] RSP: 0018:ffffc9002ecdfbc8 EFLAGS: 00010286
> > > > > > > > [ 155757.947755] RAX: 00000000003a3b1c RBX: 000007ffffffffff RCX: ffff888280183000
> > > > > > > > [ 155757.962202] RDX: 0000000000000000 RSI: 0007ffffffffffff RDI: ffff888bbc2d1000
> > > > > > > > [ 155757.976648] RBP: 0000000000000001 R08: 000000000000000b R09: ffff888ad9cedba0
> > > > > > > > [ 155757.991094] R10: ffffea0039c07900 R11: 0000000000000010 R12: ffff888b23a7b000
> > > > > > > > [ 155758.005540] R13: 0000000000000000 R14: ffff888bbc2d1000 R15: 000007ffffc71354
> > > > > > > > [ 155758.019991] FS: 00007f6234c68640(0000) GS:ffff88903f9c0000(0000) knlGS:0000000000000000
> > > > > > > > [ 155758.036356] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> > > > > > > > [ 155758.048023] CR2: 00000000000000c0 CR3: 0000000a83eb8004 CR4: 00000000007706e0
> > > > > > > > [ 155758.062473] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
> > > > > > > > [ 155758.076924] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
> > > > > > > > [ 155758.091376] PKRU: 55555554
> > > > > > > > [ 155758.096957] Call Trace:
> > > > > > > > [ 155758.102016] <TASK>
> > > > > > > > [ 155758.106502] ? __die+0x78/0xc0
> > > > > > > > [ 155758.112793] ? page_fault_oops+0x286/0x380
> > > > > > > > [ 155758.121175] ? exc_page_fault+0x5d/0x110
> > > > > > > > [ 155758.129209] ? asm_exc_page_fault+0x22/0x30
> > > > > > > > [ 155758.137763] ? mem_cgroup_get_nr_swap_pages+0x3d/0xb0
> > > > > > > > [ 155758.148060] workingset_test_recent+0xda/0x1b0
> > > > > > > > [ 155758.157133] workingset_refault+0xca/0x1e0
> > > > > > > > [ 155758.165508] filemap_add_folio+0x4d/0x70
> > > > > > > > [ 155758.173538] page_cache_ra_unbounded+0xed/0x190
> > > > > > > > [ 155758.182919] page_cache_sync_ra+0xd6/0x1e0
> > > > > > > > [ 155758.191738] filemap_read+0x68d/0xdf0
> > > > > > > > [ 155758.199495] ? mlx5e_napi_poll+0x123/0x940
> > > > > > > > [ 155758.207981] ? __napi_schedule+0x55/0x90
> > > > > > > > [ 155758.216095] __x64_sys_pread64+0x1d6/0x2c0
> > > > > > > > [ 155758.224601] do_syscall_64+0x3d/0x80
> > > > > > > > [ 155758.232058] entry_SYSCALL_64_after_hwframe+0x46/0xb0
> > > > > > > > [ 155758.242473] RIP: 0033:0x7f62c29153b5
> > > > > > > > [ 155758.249938] Code: e8 48 89 75 f0 89 7d f8 48 89 4d e0 e8 b4 e6 f7 ff 41 89 c0 4c 8b 55 e0 48 8b 55 e8 48 8b 75 f0 8b 7d f8 b8 11 00 00 00 0f 05 <48> 3d 00 f0 ff ff 77 33 44 89 c7 48 89 45 f8 e8 e7 e6 f7 ff 48 8b
> > > > > > > > [ 155758.288005] RSP: 002b:00007f6234c5ffd0 EFLAGS: 00000293 ORIG_RAX: 0000000000000011
> > > > > > > > [ 155758.303474] RAX: ffffffffffffffda RBX: 00007f628c4e70c0 RCX: 00007f62c29153b5
> > > > > > > > [ 155758.318075] RDX: 000000000003c041 RSI: 00007f61d2986000 RDI: 0000000000000076
> > > > > > > > [ 155758.332678] RBP: 00007f6234c5fff0 R08: 0000000000000000 R09: 0000000064d5230c
> > > > > > > > [ 155758.347452] R10: 000000000027d450 R11: 0000000000000293 R12: 000000000003c041
> > > > > > > > [ 155758.362044] R13: 00007f61d2986000 R14: 00007f629e11b060 R15: 000000000027d450
> > > > > > > > [ 155758.376661] </TASK>
> > > > > > > >
> > > > > > > > This patch fixes the issue by getting a local reference inside
> > > > > > > > unpack_shadow().
> > > > > > > >
> > > > > > > > Fixes: f78dfc7b77d5 ("workingset: fix confusion around eviction vs refault container")
> > > > > > >
> > > > > > > Beyond mem_cgroup_get_nr_swap_pages(), we still use the eviction_memcg
> > > > > > > without grabbing a ref to it first in workingset_test_recent() (and in
> > > > > > > workingset_refault() before that) as well as lru_gen_test_recent().
> > > > > > >
> > > > > > > Wouldn't the fix go back even further? or am I misinterpreting the problem?
> > > > > > Hmm I don't see eviction_memcg being used outside of *_test_recent
> > > > > > (the rest just uses memcg = folio_memcg(folio), which if I'm not mistaken is
> > > > > > the memcg that is refaulting the folio into memory).
> > > > > >
> > > > > > Inside workingset_test_recent(), the only other place where eviction_memcg
> > > > > > is used is for mem_cgroup_lruvec. This function call won't crash whether
> > > > > > eviction_memcg is valid or not.
> > > > >
> > > > > If eviction_memcg is invalid because the memory was already freed, we
> > > > > are basically dereferencing garbage in mem_cgroup_lruvec() aren't we?
> > > > >
> > > > > > The crash only happens during
> > > > > > mem_cgroup_get_nr_swap_pages, which has an upward traversal from
> > > > > > eviction_memcg to root.
> > > > > >
> > > > > > Let me know if this does not make sense and/or is insufficient to ensure
> > > > > > safe upward traversal from eviction_memcg to root!
> > > > > > >
> > > > > > >
> > > > > > >
> > > > > > > > Signed-off-by: Nhat Pham <nphamcs@xxxxxxxxx>
> > > > > > > > Cc: stable@xxxxxxxxxxxxxxx
> > > > > > > > ---
> > > > > > > > mm/workingset.c | 65 ++++++++++++++++++++++++++++++++-----------------
> > > > > > > > 1 file changed, 43 insertions(+), 22 deletions(-)
> > > > > > > >
> > > > > > > > diff --git a/mm/workingset.c b/mm/workingset.c
> > > > > > > > index da58a26d0d4d..03cadad4e484 100644
> > > > > > > > --- a/mm/workingset.c
> > > > > > > > +++ b/mm/workingset.c
> > > > > > > > @@ -206,10 +206,19 @@ static void *pack_shadow(int memcgid, pg_data_t *pgdat, unsigned long eviction,
> > > > > > > > return xa_mk_value(eviction);
> > > > > > > > }
> > > > > > > >
> > > > > > > > -static void unpack_shadow(void *shadow, int *memcgidp, pg_data_t **pgdat,
> > > > > > > > - unsigned long *evictionp, bool *workingsetp)
> > > > > > > > +/*
> > > > > > > > + * Unpacks the stored fields of a shadow entry into the given pointers.
> > > > > > > > + *
> > > > > > > > + * The memcg pointer is only populated if the memcg recorded in the shadow
> > > > > > > > + * entry is valid. In this case, a reference to the memcg will be acquired,
> > > > > > > > + * and a corresponding mem_cgroup_put() will be needed when we no longer
> > > > > > > > + * need the memcg.
> > > > > > > > + */
> > > > > > > > +static void unpack_shadow(void *shadow, struct mem_cgroup **memcgp,
> > > > > > > > + pg_data_t **pgdat, unsigned long *evictionp, bool *workingsetp)
> > > > > > > > {
> > > > > > > > unsigned long entry = xa_to_value(shadow);
> > > > > > > > + struct mem_cgroup *memcg;
> > > > > > > > int memcgid, nid;
> > > > > > > > bool workingset;
> > > > > > > >
> > > > > > > > @@ -220,7 +229,24 @@ static void unpack_shadow(void *shadow, int *memcgidp, pg_data_t **pgdat,
> > > > > > > > memcgid = entry & ((1UL << MEM_CGROUP_ID_SHIFT) - 1);
> > > > > > > > entry >>= MEM_CGROUP_ID_SHIFT;
> > > > > > > >
> > > > > > > > - *memcgidp = memcgid;
> > > > > > > > + /*
> > > > > > > > + * Look up the memcg associated with the stored ID. It might
> > > > > > > > + * have been deleted since the folio's eviction.
> > > > > > > > + *
> > > > > > > > + * Note that in rare events the ID could have been recycled
> > > > > > > > + * for a new cgroup that refaults a shared folio. This is
> > > > > > > > + * impossible to tell from the available data. However, this
> > > > > > > > + * should be a rare and limited disturbance, and activations
> > > > > > > > + * are always speculative anyway. Ultimately, it's the aging
> > > > > > > > + * algorithm's job to shake out the minimum access frequency
> > > > > > > > + * for the active cache.
> > > > > > > > + */
> > > > > > > > + memcg = mem_cgroup_from_id(memcgid);
> > > > > > > > + if (memcg && css_tryget(&memcg->css))
> > > > > > > > + *memcgp = memcg;
> > > > > > > > + else
> > > > > > > > + *memcgp = NULL;
> > > > > > > > +
> > > > > > > > *pgdat = NODE_DATA(nid);
> > > > > > > > *evictionp = entry;
> > > > > > > > *workingsetp = workingset;
> > > > > > > > @@ -262,15 +288,16 @@ static void *lru_gen_eviction(struct folio *folio)
> > > > > > > > static bool lru_gen_test_recent(void *shadow, bool file, struct lruvec **lruvec,
> > > > > > > > unsigned long *token, bool *workingset)
> > > > > > > > {
> > > > > > > > - int memcg_id;
> > > > > > > > unsigned long min_seq;
> > > > > > > > struct mem_cgroup *memcg;
> > > > > > > > struct pglist_data *pgdat;
> > > > > > > >
> > > > > > > > - unpack_shadow(shadow, &memcg_id, &pgdat, token, workingset);
> > > > > > > > + unpack_shadow(shadow, &memcg, &pgdat, token, workingset);
> > > > > > > > + if (!mem_cgroup_disabled() && !memcg)
> > > > > > > > + return false;
> > > > > > >
> > > > > > > +Yu Zhao
> > > > > > >
> > > > > > > There is a change of behavior here, right?
> > > > > > >
> > > > > > > The existing code will continue if !mem_cgroup_disabled() && !memcg is
> > > > > > > true, and mem_cgroup_lruvec() will return the lruvec of the root
> > > > > > > memcg. Now we are just returning false.
> > > > > > >
> > > > > > > Is this intentional?
> > > > > > Oh right, there is. Should have cc-ed Yu Zhao as well, my bad.
> > > > > > get_maintainers.pl isn't always sufficient I guess :)
> > > > > >
> > > > > > But yeah, this behavioral change is intentional.
> > > > > >
> > > > > > Correct me if I'm wrong of course, but it seems like MGLRU should
> > > > > > follow the same pattern here. That is, once we return from unpack_shadow,
> > > > > > the possible scenarios are the same as prescribed in workingset_test_recent:
> > > > > >
> > > > > > 1. If mem_cgroup is disabled, we can ignore this check.
> > > > > > 2. If mem_cgroup is enabled, then the only reason why we get NULL
> > > > > > memcg from unpack_shadow is if the eviction_memcg is no longer
> > > > > > valid. We should not try to get its lruvec, or substitute it with the
> > > > > > root memcg, but return false right away (i.e not recent).
> > > > > > >
> > > > >
> > > > > I will leave this for Yu :)
> > > >
> > > > Thanks, Yosry.
> > > >
> > > > Hi Nhat, it seems unnecessary to me to introduce a get/put into
> > > > lru_gen_test_recent() because it doesn't suffer from the bug this
> > > > patch tries to fix. In theory, the extra get/put can impact
> > > > performance, though admittedly the impact is unlikely to be
> > > > measurable. Regardless, the general practice is to fix the bug
> > > > locally, i.e., when the mem_cgroup_get_nr_swap_pages() path is taken,
> > > > rather than change the unrelated path. Thank you.
> > >
> > > Hey guys,
> > >
> > > I had suggested to have it in unpack_shadow() to keep things simple,
> > > and not further complicate the lifetime rules in this code. The
> > > tryget() is against a per-cpu counter, so it's not expensive.
> > >
> > > The NULL deref is evidence that while *some* cgroup members are still
> > > accessible once it's dead, not all of it is. There is no explicit
> > > guarantee from the cgroup code that anything BUT the tryget() is still
> > > valid against group that is under rcu freeing.
> > >
> > > Since it isn't expensive, let's keep it simple and robust, and prevent
> > > future bugs of the same class, by always ensuring the cgroup is alive
> > > before accessing random members. Especially in non-cgroup code.
> >
> > I looked at this again today with fresh eyes, and I want to go back to
> > what I initially said. Isn't RCU protection in this case enough to
> > keep the memcg "valid" (i.e accessible, not garbage)? The tryget is
> > not a lot of complexity or performance tax, but I want to really
> > understand what's happening here.
> >
> > Looking at the code again, this seems to be the sequence of events on
> > the cgroup side:
> > - css_put() puts the last reference invoking a call to css_release()
> > - css_release() queues css_release_work_fn()
> > - css_release() does some bookkeeping, makes some callbacks, and
> > queues css_free_rwork_fn() to run *after* an RCU grace period.
> > - css_free_rwork_fn() makes callbacks to free the memory, ultimately
> > freeing the memcg.
> >
> > On the memcg idr side, the removal sequence of events seem to be:
> > - mem_cgroup_id_put() will decrement the id ref and check if falls to 0
> > - If the id ref falls to 0, we call mem_cgroup_id_remove() *then* css_put()
> >
> > On the workingset_refault() side, the sequence of events seems to be:
> > - rcu_read_lock()
> > - memcg = mem_cgroup_from_id()
> > - ... // use memcg
> > - rcu_read_unlock()
> >
> > So technically, after holding the rcu read lock, if we find the memcg
> > in the idr, it must be valid, and it must not be freed until after the
> > rcu read section is completed. It's not just the cgroup internal
> > implementation, it's the contract between cgroup core and controllers
> > such as memcg.
> >
> > The memory controller expects a sequence of callbacks during freeing:
> > css_offline() -> css_released() -> css_free(). So memcg code is within
> > its right to access any fields of struct mem_cgroup that are not freed
> > by the css_offline() or css_released() until css_free() is called,
> > right?
>
>
> >
> >
> > Here is a guess / question, because I am not really familiar with
> > memory barriers and such, but is it at all possible that the actual
> > problem is reordering of instructions in mem_cgroup_id_put_many(),
> > such that we actually execute css_put() *before*
> > mem_cgroup_id_remove()?
> >
> > If this happens it seems possible for this to happen:
> >
> > cpu #1 cpu#2
> > css_put()
> > /* css_free_rwork_fn is queued */
> > rcu_read_lock()
> > mem_cgroup_from_id()
> > mem_cgroup_id_remove()
> > /* access memcg */
> >
> > If I understand correctly, if css_free_rwork_fn() is queued before the
> > rcu_read_lock in workingset_refault() begins, then it can be executed
> > during the rcu read section, and the memcg can be freed at any point
> > from under us. Perhaps what we need is memory barriers to ensure
> > correct ordering in mem_cgroup_id_put_many()? I am not sure if
> > rcu_read_lock() implies a barrier on the other side.
> >
> > Sorry if this is all off, I am just trying to understand what's going on.
> Ah that is wild. That does sound plausible.
> In this case, maybe something like this?
>
> mem_cgroup_id_remove(memcg);
> /*
> * Preventing css_put from happening before id removal due to
> * instruction reordering.
> *

This is redefining what smp_mb() is, probably unnecessary.

> * This guarantees that if a non-null memcg is acquired from ID within
> * an RCU read section, its css won't be freed for the
> * duration of this section.
> */
> smp_mb();
> /* Memcg ID pins CSS */
> css_put(&memcg->css);

I am not the best person to answer this question, ideally someone with
more understanding of memory barriers should chime in here to:
- Confirm my theory is correct.
- Confirm smp_mb() is the correct primitive to use. I am guessing
smp_wmb() is enough here.
- Confirm that we don't need an additional read barrier on the read
side, ideally rcu_read_lock() is enough, but I am not sure.