Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
MIME-Version: 1.0
References: <20220327051853.57647-1-songmuchun@bytedance.com>
In-Reply-To: <20220327051853.57647-1-songmuchun@bytedance.com>
From:   Linus Torvalds <torvalds@linux-foundation.org>
Date:   Sun, 27 Mar 2022 14:08:33 -0700
Message-ID: <CAHk-=wh-mVrp3auBiK2GSMpuqS10Bbq_7fRa6+=zt-0LiF7O2A@mail.gmail.com>
Subject: Re: [PATCH 1/2] mm: kfence: fix missing objcg housekeeping for SLAB
To:     Muchun Song <songmuchun@bytedance.com>
Cc:     Alexander Potapenko <glider@google.com>,
        Marco Elver <elver@google.com>,
        Dmitry Vyukov <dvyukov@google.com>,
        Andrew Morton <akpm@linux-foundation.org>,
        Christoph Lameter <cl@linux.com>,
        Pekka Enberg <penberg@kernel.org>,
        David Rientjes <rientjes@google.com>,
        Joonsoo Kim <iamjoonsoo.kim@lge.com>,
        Vlastimil Babka <vbabka@suse.cz>, roman.gushchin@linux.dev,
        kasan-dev <kasan-dev@googlegroups.com>,
        Linux-MM <linux-mm@kvack.org>,
        Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
        syzbot <syzbot+f8c45ccc7d5d45fc5965@syzkaller.appspotmail.com>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Sat, Mar 26, 2022 at 10:19 PM Muchun Song <songmuchun@bytedance.com> wrote:
>
> The objcg is not cleared and put for kfence object when it is freed, which
> could lead to memory leak for struct obj_cgroup and wrong statistics of
> NR_SLAB_RECLAIMABLE_B or NR_SLAB_UNRECLAIMABLE_B.  Since the last freed
> object's objcg is not cleared, mem_cgroup_from_obj() could return the wrong
> memcg when this kfence object, which is not charged to any objcgs, is
> reallocated to other users.  A real word issue [1] is caused by this bug.

Good that this looks sorted out.

Patch 2/2 seems to still be up in the air. The patch not only causes
build errors, but it looks really very odd to me.

In particular, you do that loop with

                __SetPageSlab(&pages[i]);

in kfence_init_pool(), but that is *not* where you set the
MEMCG_DATA_OBJCGS, and instead do that virt_to_slab(addr) dance later.

That looks very odd to me. I think the two should go hand-in-hand,
since that __SetPageSlab() really is what makes it a slab thing, and I
think it should go together with setting the slab state correctly.

Finally, is there a syzbot report for that second problem?

Anyway, should I apply this PATCH 1/2 now directly as the solution for
the dentry issue, or should I wait for that second patch? They seem to
be related only indirectly, in that the problems were both introduced
by the same commit.

                     Linus