From:   Claudio Imbrenda <imbrenda@linux.ibm.com>
To:     linux-kernel@vger.kernel.org, linux-mm@kvack.org
Cc:     thuth@redhat.com, frankja@linux.ibm.com, borntraeger@de.ibm.com,
        Ulrich.Weigand@de.ibm.com, heiko.carstens@de.ibm.com,
        david@redhat.com, ultrachin@163.com, akpm@linux-foundation.org,
        vbabka@suse.cz, brookxu.cn@gmail.com, xiaoggchen@tencent.com,
        linuszeng@tencent.com, yihuilu@tencent.com, mhocko@suse.com,
        daniel.m.jordan@oracle.com, axboe@kernel.dk, legion@kernel.org,
        peterz@infradead.org, aarcange@redhat.com, christian@brauner.io,
        ebiederm@xmission.com, tglx@linutronix.de
Subject: [RFC v1 4/4] kernel/fork.c: process_mmput_async: stop OOM while freeing memory
Date:   Thu, 11 Nov 2021 10:50:08 +0100
Message-Id: <20211111095008.264412-6-imbrenda@linux.ibm.com>
In-Reply-To: <20211111095008.264412-1-imbrenda@linux.ibm.com>
References: <20211111095008.264412-1-imbrenda@linux.ibm.com>
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit
Precedence: bulk

This patch implements a simple OOM notifier to stop the OOM killer
while a mm is being reclaimed asynchronously using the
process_mmput_async syscall.

Tested on s390x.

Signed-off-by: Claudio Imbrenda <imbrenda@linux.ibm.com>
---
 kernel/fork.c | 28 ++++++++++++++++++++++++++++
 1 file changed, 28 insertions(+)

diff --git a/kernel/fork.c b/kernel/fork.c
index 0da39b76005c..7279209eb69c 100644
--- a/kernel/fork.c
+++ b/kernel/fork.c
@@ -118,6 +118,11 @@
  */
 #define MAX_THREADS FUTEX_TID_MASK
 
+/*
+ * Priority for the OOM notifier used in process_mmput_async
+ */
+#define PROCESS_MMPUT_ASYNC_OOM_NOTIFY_PRIORITY 70
+
 /*
  * Protected counters by write_lock_irq(&tasklist_lock)
  */
@@ -3203,13 +3208,27 @@ int sysctl_max_threads(struct ctl_table *table, int write,
 	return 0;
 }
 
+/* Prevent the OOM from being triggered while we are cleaning up asynchronously */
+static int mmput_async_oom_notifier(struct notifier_block *nb, unsigned long dummy, void *parm)
+{
+	/*
+	 * We cannot know the speed at which pages are being freed, so we
+	 * fake it and say it's at least one. This is already enough to
+	 * stop the OOM killer.
+	 */
+	*(unsigned long *)parm += PAGE_SIZE;
+	return NOTIFY_OK;
+}
+
 SYSCALL_DEFINE2(process_mmput_async, int, pidfd, unsigned int, flags)
 {
 #ifdef CONFIG_MMU
+	struct notifier_block oom_nb;
 	struct mm_struct *mm = NULL;
 	struct task_struct *task;
 	unsigned int tmp;
 	struct pid *pid;
+	int r;
 
 	if (flags)
 		return -EINVAL;
@@ -3280,8 +3299,17 @@ SYSCALL_DEFINE2(process_mmput_async, int, pidfd, unsigned int, flags)
 	if (atomic_read(&mm->mm_users))
 		panic("mm_users not 0 but trying to __mmput anyway!");
 
+	/*
+	 * Register an OOM notifier, to stop the OOM while we are
+	 * asynchronously freeing the mm.
+	 */
+	oom_nb.priority = PROCESS_MMPUT_ASYNC_OOM_NOTIFY_PRIORITY;
+	oom_nb.notifier_call = mmput_async_oom_notifier;
+	r = register_oom_notifier(&oom_nb);
 	/* Do the actual work */
 	__mmput(mm);
+	if (!r)
+		unregister_oom_notifier(&oom_nb);
 	/* And put the extra reference taken above */
 	mmdrop(mm);
 
-- 
2.31.1