mm: fs: invalidate bh_lrus for only cold path

The kernel test robot reported the regression of fio.write_iops[1] with
commit 8cc621d ("mm: fs: invalidate BH LRU during page migration").

Since lru_add_drain is called frequently, invalidate bh_lrus there could
increase bh_lrus cache miss ratio, which needs more IO in the end.

This patch moves the bh_lrus invalidation from the hot path( e.g.,
zap_page_range, pagevec_release) to cold path(i.e., lru_add_drain_all,
lru_cache_disable).

Zhengjun Xing confirmed
 "I test the patch, the regression reduced to -2.9%"

[1] https://lore.kernel.org/lkml/20210520083144.GD14190@xsang-OptiPlex-9020/
[2] 8cc621d, mm: fs: invalidate BH LRU during page migration

Link: https://lkml.kernel.org/r/[email protected]
Signed-off-by: Minchan Kim <[email protected]>
Reported-by: kernel test robot <[email protected]>
Reviewed-by: Chris Goldsworthy <[email protected]>
Tested-by: "Xing, Zhengjun" <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>

Author: minchank

fs/buffer.c

@@ -1425,12 +1425,16 @@ void invalidate_bh_lrus(void)
 }
 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
 
-void invalidate_bh_lrus_cpu(int cpu)
+/*
+ * It's called from workqueue context so we need a bh_lru_lock to close
+ * the race with preemption/irq.
+ */
+void invalidate_bh_lrus_cpu(void)
 {
 	struct bh_lru *b;
 
 	bh_lru_lock();
-	b = per_cpu_ptr(&bh_lrus, cpu);
+	b = this_cpu_ptr(&bh_lrus);
 	__invalidate_bh_lrus(b);
 	bh_lru_unlock();
 }

include/linux/buffer_head.h

@@ -194,7 +194,7 @@ void __breadahead_gfp(struct block_device *, sector_t block, unsigned int size,
 struct buffer_head *__bread_gfp(struct block_device *,
 				sector_t block, unsigned size, gfp_t gfp);
 void invalidate_bh_lrus(void);
-void invalidate_bh_lrus_cpu(int cpu);
+void invalidate_bh_lrus_cpu(void);
 bool has_bh_in_lru(int cpu, void *dummy);
 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags);
 void free_buffer_head(struct buffer_head * bh);
@@ -408,7 +408,7 @@ static inline int inode_has_buffers(struct inode *inode) { return 0; }
 static inline void invalidate_inode_buffers(struct inode *inode) {}
 static inline int remove_inode_buffers(struct inode *inode) { return 1; }
 static inline int sync_mapping_buffers(struct address_space *mapping) { return 0; }
-static inline void invalidate_bh_lrus_cpu(int cpu) {}
+static inline void invalidate_bh_lrus_cpu(void) {}
 static inline bool has_bh_in_lru(int cpu, void *dummy) { return false; }
 #define buffer_heads_over_limit 0
 

mm/swap.c

@@ -620,7 +620,6 @@ void lru_add_drain_cpu(int cpu)
 		pagevec_lru_move_fn(pvec, lru_lazyfree_fn);
 
 	activate_page_drain(cpu);
-	invalidate_bh_lrus_cpu(cpu);
 }
 
 /**
@@ -703,6 +702,20 @@ void lru_add_drain(void)
 	local_unlock(&lru_pvecs.lock);
 }
 
+/*
+ * It's called from per-cpu workqueue context in SMP case so
+ * lru_add_drain_cpu and invalidate_bh_lrus_cpu should run on
+ * the same cpu. It shouldn't be a problem in !SMP case since
+ * the core is only one and the locks will disable preemption.
+ */
+static void lru_add_and_bh_lrus_drain(void)
+{
+	local_lock(&lru_pvecs.lock);
+	lru_add_drain_cpu(smp_processor_id());
+	local_unlock(&lru_pvecs.lock);
+	invalidate_bh_lrus_cpu();
+}
+
 void lru_add_drain_cpu_zone(struct zone *zone)
 {
 	local_lock(&lru_pvecs.lock);
@@ -717,7 +730,7 @@ static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
 
 static void lru_add_drain_per_cpu(struct work_struct *dummy)
 {
-	lru_add_drain();
+	lru_add_and_bh_lrus_drain();
 }
 
 /*
@@ -858,7 +871,7 @@ void lru_cache_disable(void)
 	 */
 	__lru_add_drain_all(true);
 #else
-	lru_add_drain();
+	lru_add_and_bh_lrus_drain();
 #endif
 }