anolis: mm: support pre oom

ANBZ: #9079

Provide a universal and reliable rapid OOM resolution solution to assist
businesses in enhancing memory deployment density and improving the
stability of online business performance during high watermark operation.

Signed-off-by: Kaihao Bai <carlo.bai@linux.alibaba.com>
Reviewed-by: Xu Yu <xuyu@linux.alibaba.com>
Link: https://gitee.com/anolis/cloud-kernel/pulls/3223

include/linux/memcontrol.h

@@ -525,6 +525,10 @@ struct mem_cgroup {
 	bool allow_pgtable_bind;
 #endif
 
+#ifdef CONFIG_PRE_OOM
+	bool oom_offline;
+#endif
+
 #ifdef CONFIG_LRU_GEN
 	CK_KABI_USE(1, unsigned long mglru_batch_size)
 	CK_KABI_USE(2, unsigned long mglru_reclaim_pages)

include/linux/pre_oom.h

@@ -0,0 +1,36 @@
+#ifndef _LINUX_PRE_OOM_H
+#define _LINUX_PRE_OOM_H
+
+#include <linux/sched.h>
+
+#ifdef CONFIG_PRE_OOM
+
+#include <linux/types.h>
+#include <linux/jump_label.h>
+
+DECLARE_STATIC_KEY_FALSE(pre_oom_enabled_key);
+static inline bool pre_oom_enabled(void)
+{
+	return static_branch_unlikely(&pre_oom_enabled_key);
+}
+
+
+int pre_oom_enter(void);
+void pre_oom_leave(void);
+
+#else
+
+static inline bool pre_oom_enabled(void)
+{
+	return false;
+}
+
+static inline int pre_oom_enter(void)
+{
+	return 0;
+}
+
+static inline void pre_oom_leave(void) {}
+
+#endif /* CONFIG_PRE_OOM */
+#endif /* _LINUX_PRE_OOM_H */

include/linux/sched.h

@@ -976,7 +976,9 @@ struct task_struct {
 #ifdef CONFIG_IOMMU_SVA
 	unsigned			pasid_activated:1;
 #endif
-
+#ifdef CONFIG_PRE_OOM
+	unsigned			reclaim_stall:1;
+#endif
 	unsigned long			atomic_flags; /* Flags requiring atomic access. */
 
 	struct restart_block		restart_block;

kernel/exit.c

@@ -64,6 +64,7 @@
 #include <linux/rcuwait.h>
 #include <linux/compat.h>
 #include <linux/io_uring.h>
+#include <linux/pre_oom.h>
 #ifdef CONFIG_TEXT_UNEVICTABLE
 #include <linux/unevictable.h>
 #endif
@@ -859,6 +860,15 @@ void __noreturn do_exit(long code)
 
 	validate_creds_for_do_exit(tsk);
 
+#ifdef CONFIG_PRE_OOM
+	/*
+	 * Killed task has been stalled in reclaim path, release the semaphore
+	 * here.
+	 */
+	if (unlikely(tsk->reclaim_stall))
+		pre_oom_leave();
+#endif
+
 	check_stack_usage();
 	preempt_disable();
 	if (tsk->nr_dirtied)

mm/Kconfig

@@ -1032,4 +1032,10 @@ config PAGETABLE_SHARE
 
 	  If unsure, say N.
 
+config PRE_OOM
+	bool "Enable pre oom control"
+	depends on MEMCG
+	help
+	  This feature is used to ensure that higher priority tasks would not enter the direct
+	  reclaim path when applying for memory allocation.
 endmenu

mm/Makefile

@@ -138,3 +138,4 @@ obj-$(CONFIG_PAGECACHE_LIMIT) += pagecache_limit.o
 obj-$(CONFIG_HAVE_BOOTMEM_INFO_NODE) += bootmem_info.o
 obj-$(CONFIG_PGTABLE_BIND) += pgtable_bind.o
 obj-$(CONFIG_PAGETABLE_SHARE) += pgtable_share.o
+obj-$(CONFIG_PRE_OOM) += pre_oom.o

mm/memcontrol.c

@@ -68,6 +68,7 @@
 #include <net/ip.h>
 #include "slab.h"
 #include <linux/proc_fs.h>
+#include <linux/pre_oom.h>
 
 #include <linux/uaccess.h>
 #ifdef CONFIG_TEXT_UNEVICTABLE
@@ -2751,11 +2752,13 @@ static void reclaim_wmark(struct mem_cgroup *memcg)
 	 * simply record the whole duration of reclaim_wmark work for the
 	 * overhead-accuracy trade-off.
 	 */
+	pre_oom_enter();
 	start = ktime_get_ns();
 	psi_memstall_enter(&pflags);
 	try_to_free_mem_cgroup_pages(memcg, nr_pages, GFP_KERNEL, true);
 	psi_memstall_leave(&pflags);
 	duration = ktime_get_ns() - start;
+	pre_oom_leave();
 
 	if (!css_tryget_online(&memcg->css))
 		return;
@@ -2791,10 +2794,12 @@ static unsigned long reclaim_high(struct mem_cgroup *memcg,
 
 		memcg_memory_event(memcg, MEMCG_HIGH);
 
+		pre_oom_enter();
 		psi_memstall_enter(&pflags);
 		nr_reclaimed += try_to_free_mem_cgroup_pages(memcg, nr_pages,
 							     gfp_mask, true);
 		psi_memstall_leave(&pflags);
+		pre_oom_leave();
 	} while ((memcg = parent_mem_cgroup(memcg)) &&
 		 !mem_cgroup_is_root(memcg));
 
@@ -3121,12 +3126,14 @@ retry:
 
 	memcg_memory_event(mem_over_limit, MEMCG_MAX);
 
+	pre_oom_enter();
 	memcg_lat_stat_start(&start);
 	psi_memstall_enter(&pflags);
 	nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
 						    gfp_mask, may_swap);
 	psi_memstall_leave(&pflags);
 	memcg_lat_stat_end(MEM_LAT_MEMCG_DIRECT_RECLAIM, start);
+	pre_oom_leave();
 
 	if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
 		goto retry;
@@ -6724,6 +6731,29 @@ static int memcg_pgtable_misplaced_write(struct cgroup_subsys_state *css,
 }
 #endif /* CONFIG_PGTABLE_BIND */
 
+#ifdef CONFIG_PRE_OOM
+static u64 memcg_oom_offline_read(struct cgroup_subsys_state *css,
+				   struct cftype *cft)
+{
+	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
+
+	return READ_ONCE(memcg->oom_offline);
+}
+
+static int memcg_oom_offline_write(struct cgroup_subsys_state *css,
+				    struct cftype *cft, u64 val)
+{
+	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
+
+	if (val)
+		memcg->oom_offline = true;
+	else
+		memcg->oom_offline = false;
+
+	return 0;
+}
+#endif /* CONFIG_PRE_OOM */
+
 #ifdef CONFIG_LRU_GEN
 static bool mglru_size_valid_check(struct mem_cgroup *memcg)
 {
@@ -7399,6 +7429,13 @@ static struct cftype mem_cgroup_legacy_files[] = {
 		.write_u64 = memcg_pgtable_misplaced_write,
 		.read_u64 = memcg_pgtable_misplaced_read,
 	},
+#endif
+#ifdef CONFIG_PRE_OOM
+	{
+		.name = "oom_offline",
+		.write_u64 = memcg_oom_offline_write,
+		.read_u64 = memcg_oom_offline_read,
+	},
 #endif
 	{ },	/* terminate */
 };

mm/page_alloc.c

@@ -76,6 +76,7 @@
 #include <linux/buffer_head.h>
 #include <linux/vmalloc.h>
 #include <linux/prezero.h>
+#include <linux/pre_oom.h>
 
 #include <asm/sections.h>
 #include <asm/tlbflush.h>
@@ -4741,6 +4742,7 @@ __perform_reclaim(gfp_t gfp_mask, unsigned int order,
 	cond_resched();
 
 	/* We now go into synchronous reclaim */
+	pre_oom_enter();
 	cpuset_memory_pressure_bump();
 	memcg_lat_stat_start(&start);
 	psi_memstall_enter(&pflags);
@@ -4754,6 +4756,7 @@ __perform_reclaim(gfp_t gfp_mask, unsigned int order,
 	fs_reclaim_release(gfp_mask);
 	psi_memstall_leave(&pflags);
 	memcg_lat_stat_end(MEM_LAT_GLOBAL_DIRECT_RECLAIM, start);
+	pre_oom_leave();
 
 	cond_resched();
 
@@ -5029,7 +5032,6 @@ static inline struct page *
 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
 						struct alloc_context *ac)
 {
-	bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
 	const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
 	struct page *page = NULL;
 	unsigned int alloc_flags;
@@ -5041,6 +5043,7 @@ __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
 	unsigned int cpuset_mems_cookie;
 	unsigned int zonelist_iter_cookie;
 	int reserve_flags;
+	bool can_direct_reclaim;
 
 	/*
 	 * We also sanity check to catch abuse of atomic reserves being used by
@@ -5050,6 +5053,24 @@ __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
 				(__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
 		gfp_mask &= ~__GFP_ATOMIC;
 
+#ifdef CONFIG_PRE_OOM
+	/*
+	 * If Pre-OOM is enabled, the cgroup of QoS sensitive should avoid
+	 * direct reclaim and trigger OOM as soon as possible. Thus gfp_mask
+	 * should be reset here.
+	 */
+	if (pre_oom_enabled()) {
+		struct mem_cgroup *memcg;
+
+		memcg = get_mem_cgroup_from_mm(current->mm);
+		if (memcg && !memcg->oom_offline) {
+			gfp_mask &= ~__GFP_DIRECT_RECLAIM;
+			css_put(&memcg->css);
+		}
+	}
+#endif
+
+	can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
 restart:
 	compaction_retries = 0;
 	no_progress_loops = 0;
@@ -5168,7 +5189,7 @@ retry:
 
 	/* Caller is not willing to reclaim, we can't balance anything */
 	if (!can_direct_reclaim)
-		goto nopage;
+		goto oom;
 
 	/* Avoid recursion of direct reclaim */
 	if (current->flags & PF_MEMALLOC)
@@ -5222,6 +5243,7 @@ retry:
 	    check_retry_zonelist(zonelist_iter_cookie))
 		goto restart;
 
+oom:
 	/* Reclaim has failed us, start killing things */
 	page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
 	if (page)

mm/pre_oom.c

@@ -0,0 +1,158 @@
+// SPDX-License-Identifier: GPL-2.0
+#include <linux/init.h>
+#include <linux/fs.h>
+#include <linux/sysfs.h>
+#include <linux/kobject.h>
+#include <linux/mm.h>
+#include <linux/mm_types.h>
+#include <linux/slab.h>
+#include <linux/pre_oom.h>
+
+DEFINE_STATIC_KEY_FALSE(pre_oom_enabled_key);
+
+/*
+ * From 0 .. 3, which means the kernel can support up to
+ * num_online_cpus / oom_level tasks to reclaim memory.
+ */
+static int oom_level;
+static struct semaphore *sem;
+
+int pre_oom_enter(void)
+{
+	int result;
+
+	if (!pre_oom_enabled())
+		return 0;
+
+	result = down_killable(sem);
+	if (!result)
+		current->reclaim_stall = 1;
+
+	return result;
+}
+
+void pre_oom_leave(void)
+{
+	if (pre_oom_enabled() && current->reclaim_stall) {
+		current->reclaim_stall = 0;
+		up(sem);
+	}
+}
+
+static int adjust_oom_level(int level)
+{
+	unsigned long flags;
+	int count = num_online_cpus() / (level + 1);
+	int result = 0;
+
+	raw_spin_lock_irqsave(&sem->lock, flags);
+
+	/* There are no other tasks reclaiming memory */
+	if (sem->count == (num_online_cpus() / (oom_level + 1))) {
+		sem->count = count;
+		oom_level = level;
+	} else
+		result = -EPERM;
+
+	raw_spin_unlock_irqrestore(&sem->lock, flags);
+
+	return result;
+}
+
+#ifdef CONFIG_SYSFS
+static ssize_t pre_oom_enabled_show(struct kobject *kobj,
+				    struct kobj_attribute *attr, char *buf)
+{
+	return sprintf(buf, "%d\n", !!static_branch_unlikely(&pre_oom_enabled_key));
+}
+
+static ssize_t pre_oom_enabled_store(struct kobject *kobj,
+					   struct kobj_attribute *attr,
+					   const char *buf, size_t count)
+{
+	static DEFINE_MUTEX(mutex);
+	ssize_t ret = count;
+
+	mutex_lock(&mutex);
+
+	if (!strncmp(buf, "1", 1))
+		static_branch_enable(&pre_oom_enabled_key);
+	else if (!strncmp(buf, "0", 1))
+		static_branch_disable(&pre_oom_enabled_key);
+	else
+		ret = -EINVAL;
+
+	mutex_unlock(&mutex);
+	return ret;
+}
+
+static ssize_t pre_oom_level_show(struct kobject *kobj,
+				    struct kobj_attribute *attr, char *buf)
+{
+	return sprintf(buf, "%d\n", oom_level);
+}
+
+static ssize_t pre_oom_level_store(struct kobject *kobj,
+					   struct kobj_attribute *attr,
+					   const char *buf, size_t count)
+{
+	static DEFINE_MUTEX(mutex);
+	ssize_t ret = count;
+	unsigned long level;
+
+	ret = kstrtoul(buf, 10, &level);
+	if (ret)
+		return ret;
+
+	mutex_lock(&mutex);
+
+	if (level < 0 || level > 3)
+		ret = -EINVAL;
+
+	adjust_oom_level(level);
+
+	mutex_unlock(&mutex);
+	return count;
+
+}
+
+static struct kobj_attribute pre_oom_enabled_attr =
+		__ATTR(enabled, 0644, pre_oom_enabled_show,
+		       pre_oom_enabled_store);
+
+static struct kobj_attribute pre_oom_level_attr =
+		__ATTR(level, 0644, pre_oom_level_show,
+		       pre_oom_level_store);
+
+static struct attribute *pre_oom_attrs[] = {
+	&pre_oom_enabled_attr.attr,
+	&pre_oom_level_attr.attr,
+	NULL,
+};
+
+static const struct attribute_group pre_oom_attr_group = {
+	.attrs = pre_oom_attrs,
+	.name = "pre_oom",
+};
+#endif /* CONFIG_SYSFS */
+
+static int __init pre_oom_init(void)
+{
+#ifdef CONFIG_SYSFS
+	int err;
+
+	err = sysfs_create_group(mm_kobj, &pre_oom_attr_group);
+	if (err) {
+		pr_err("pre_oom: register sysfs failed\n");
+		return err;
+	}
+#endif
+	sem = kmalloc(sizeof(*sem), GFP_KERNEL);
+	if (!sem)
+		return -ENOMEM;
+
+	sema_init(sem, num_online_cpus());
+
+	return 0;
+}
+subsys_initcall(pre_oom_init);

mm/vmscan.c

@@ -61,6 +61,7 @@
 #include <linux/shmem_fs.h>
 #include <linux/ctype.h>
 #include <linux/debugfs.h>
+#include <linux/pre_oom.h>
 
 #include <asm/tlbflush.h>
 #include <asm/div64.h>
@@ -6964,6 +6965,7 @@ static int balance_pgdat(pg_data_t *pgdat, int order, int highest_zoneidx)
 	};
 
 	set_task_reclaim_state(current, &sc.reclaim_state);
+	pre_oom_enter();
 	psi_memstall_enter(&pflags);
 	__fs_reclaim_acquire();
 
@@ -7156,6 +7158,7 @@ out:
 	snapshot_refaults(NULL, pgdat);
 	__fs_reclaim_release();
 	psi_memstall_leave(&pflags);
+	pre_oom_leave();
 	set_task_reclaim_state(current, NULL);
 
 	/*