Skip to contentSkip to conversation

fs/buffer: simplify the code flow of LRU management algorithm

PATCH In review Minfei Huang 2 participants 4 messages Sep 10, 2015 – Oct 9, 2015 VFS

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 fs/buffer: simplify the code flow of LRU management algorithm

From: Minfei Huang <[email protected]>

There is a buffer_head lru list cache in local cpu to accelerate the speed. The LRU management algorithm is simple enough in bh_lru_install().

There are three situtaions we should deal with.
1) All/part of the lru cache is NULL.
2) The new buffer_head hitts the lru cache.
3) The new buffer_head does hit the lru cache.

We put the new buffer_head at the head of lru cache, then copy the buffer_head from the original lru cache, and drop the spare.

---
 fs/buffer.c | 32 ++++++++++++++++++++------------
 1 file changed, 20 insertions(+), 12 deletions(-)
fs/buffer.c +20 -12
index 1cf7a53..2139574 100644
@@ -1287,8 +1287,6 @@ static inline void check_irqs_on(void)
  */
 static void bh_lru_install(struct buffer_head *bh)
 {
-	struct buffer_head *evictee = NULL;
-
 	check_irqs_on();
 	bh_lru_lock();
 	if (__this_cpu_read(bh_lrus.bhs[0]) != bh) {
@@ -1302,25 +1300,35 @@ static void bh_lru_install(struct buffer_head *bh)
 			struct buffer_head *bh2 =
 				__this_cpu_read(bh_lrus.bhs[in]);
 
-			if (bh2 == bh) {
+			if (bh2 == NULL) {
+				/* Rest value in bh_lrus.bhs always is NULL */
+				break;
+			} else if (bh2 == bh) {
 				__brelse(bh2);
 			} else {
-				if (out >= BH_LRU_SIZE) {
-					BUG_ON(evictee != NULL);
-					evictee = bh2;
+				if (out == BH_LRU_SIZE) {
+					/*
+					 * this condition will be happened,
+					 * only if none of entry in
+					 * bh_lrus.bhs hits the new bh,
+					 * so the last bh should be released.
+					 */
+					BUG_ON(in != BH_LRU_SIZE - 1);
+					__brelse(bh2);
+					break;
 				} else {
 					bhs[out++] = bh2;
 				}
 			}
 		}
-		while (out < BH_LRU_SIZE)
-			bhs[out++] = NULL;
-		memcpy(this_cpu_ptr(&bh_lrus.bhs), bhs, sizeof(bhs));
+		/*
+		 * it is fine that the value out may be smaller than
+		 * BH_LRU_SIZE. The rest of the value in bh_lrus.bhs is NULL.
+		 */
+		memcpy(this_cpu_ptr(&bh_lrus.bhs), bhs,
+				sizeof(struct buffer_head *) * out);
 	}
 	bh_lru_unlock();
-
-	if (evictee)
-		__brelse(evictee);
 }
 
 /*
Signature · 1 line
2.1.0
#1 lore raw

Ping, Could you someone help to review this patch?

Thanks
Minfei

Quoted 86 lines · Minfei Huang

From: Minfei Huang <[email protected]>

There is a buffer_head lru list cache in local cpu to accelerate the speed. The LRU management algorithm is simple enough in bh_lru_install().

There are three situtaions we should deal with.
1) All/part of the lru cache is NULL.
2) The new buffer_head hitts the lru cache.
3) The new buffer_head does hit the lru cache.

We put the new buffer_head at the head of lru cache, then copy the buffer_head from the original lru cache, and drop the spare.

---
 fs/buffer.c | 32 ++++++++++++++++++++------------
 1 file changed, 20 insertions(+), 12 deletions(-)
fs/buffer.c +20 -12
index 1cf7a53..2139574 100644
@@ -1287,8 +1287,6 @@ static inline void check_irqs_on(void)
  */
 static void bh_lru_install(struct buffer_head *bh)
 {
-	struct buffer_head *evictee = NULL;
-
 	check_irqs_on();
 	bh_lru_lock();
 	if (__this_cpu_read(bh_lrus.bhs[0]) != bh) {
@@ -1302,25 +1300,35 @@ static void bh_lru_install(struct buffer_head *bh)
 			struct buffer_head *bh2 =
 				__this_cpu_read(bh_lrus.bhs[in]);
 
-			if (bh2 == bh) {
+			if (bh2 == NULL) {
+				/* Rest value in bh_lrus.bhs always is NULL */
+				break;
+			} else if (bh2 == bh) {
 				__brelse(bh2);
 			} else {
-				if (out >= BH_LRU_SIZE) {
-					BUG_ON(evictee != NULL);
-					evictee = bh2;
+				if (out == BH_LRU_SIZE) {
+					/*
+					 * this condition will be happened,
+					 * only if none of entry in
+					 * bh_lrus.bhs hits the new bh,
+					 * so the last bh should be released.
+					 */
+					BUG_ON(in != BH_LRU_SIZE - 1);
+					__brelse(bh2);
+					break;
 				} else {
 					bhs[out++] = bh2;
 				}
 			}
 		}
-		while (out < BH_LRU_SIZE)
-			bhs[out++] = NULL;
-		memcpy(this_cpu_ptr(&bh_lrus.bhs), bhs, sizeof(bhs));
+		/*
+		 * it is fine that the value out may be smaller than
+		 * BH_LRU_SIZE. The rest of the value in bh_lrus.bhs is NULL.
+		 */
+		memcpy(this_cpu_ptr(&bh_lrus.bhs), bhs,
+				sizeof(struct buffer_head *) * out);
 	}
 	bh_lru_unlock();
-
-	if (evictee)
-		__brelse(evictee);
 }
 
 /*
Signature · 7 lines
2.1.0

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/
#2 lore raw

why not change like this:

fs/buffer.c +12 -21
index 82283ab..d6769f1 100644
@@ -1287,40 +1287,31 @@ static inline void check_irqs_on(void)
  */
 static void bh_lru_install(struct buffer_head *bh)
 {
-       struct buffer_head *evictee = NULL;
+       struct buffer_head *old = NULL;
 
        check_irqs_on();
        bh_lru_lock();
        if (__this_cpu_read(bh_lrus.bhs[0]) != bh) {
-               struct buffer_head *bhs[BH_LRU_SIZE];
-               int in;
+               struct buffer_head *temp;
                int out = 0;
 
+               old = __this_cpu_read(bh_lrus.bhs[0]);
                get_bh(bh);
-               bhs[out++] = bh;
-               for (in = 0; in < BH_LRU_SIZE; in++) {
-                       struct buffer_head *bh2 =
-                               __this_cpu_read(bh_lrus.bhs[in]);
-
-                       if (bh2 == bh) {
-                               __brelse(bh2);
+               __this_cpu_write(bh_lrus.bhs[out++], bh);
+               for (; out < BH_LRU_SIZE; out++) {
+                       if (old == bh || old == NULL) {
+                               break;
                        } else {
-                               if (out >= BH_LRU_SIZE) {
-                                       BUG_ON(evictee != NULL);
-                                       evictee = bh2;
-                               } else {
-                                       bhs[out++] = bh2;
-                               }
+                               temp = __this_cpu_read(bh_lrus.bhs[out]);
+                               __this_cpu_write(bh_lrus.bhs[out], old);
+                               old = temp;
                        }
                }
-               while (out < BH_LRU_SIZE)
-                       bhs[out++] = NULL;
-               memcpy(this_cpu_ptr(&bh_lrus.bhs), bhs, sizeof(bhs));
        }
        bh_lru_unlock();
 
-       if (evictee)
-               __brelse(evictee);
+       if (old)
+               __brelse(old);
 }
 
 /*

more simple to understand and have better performance . am i understanding correctly ?

Quoted 99 lines · Minfei Huang

Ping, Could you someone help to review this patch?

Thanks
Minfei

Quoted 86 lines · Minfei Huang

From: Minfei Huang <[email protected]>

There is a buffer_head lru list cache in local cpu to accelerate the speed. The LRU management algorithm is simple enough in bh_lru_install().

There are three situtaions we should deal with.
1) All/part of the lru cache is NULL.
2) The new buffer_head hitts the lru cache.
3) The new buffer_head does hit the lru cache.

We put the new buffer_head at the head of lru cache, then copy the buffer_head from the original lru cache, and drop the spare.

---
fs/buffer.c | 32 ++++++++++++++++++++------------

1 file changed, 20 insertions(+), 12 deletions(-)
fs/buffer.c +20 -12
index 1cf7a53..2139574 100644
@@ -1287,8 +1287,6 @@ static inline void check_irqs_on(void)
 */
 static void bh_lru_install(struct buffer_head *bh)
 {
-	struct buffer_head *evictee = NULL;
-
 	check_irqs_on();
 	bh_lru_lock();
 	if (__this_cpu_read(bh_lrus.bhs[0]) != bh) {
@@ -1302,25 +1300,35 @@ static void bh_lru_install(struct buffer_head *bh)
 			struct buffer_head *bh2 =
 				__this_cpu_read(bh_lrus.bhs[in]);
 
-			if (bh2 == bh) {
+			if (bh2 == NULL) {
+				/* Rest value in bh_lrus.bhs always is NULL */
+				break;
+			} else if (bh2 == bh) {
 				__brelse(bh2);
 			} else {
-				if (out >= BH_LRU_SIZE) {
-					BUG_ON(evictee != NULL);
-					evictee = bh2;
+				if (out == BH_LRU_SIZE) {
+					/*
+					 * this condition will be happened,
+					 * only if none of entry in
+					 * bh_lrus.bhs hits the new bh,
+					 * so the last bh should be released.
+					 */
+					BUG_ON(in != BH_LRU_SIZE - 1);
+					__brelse(bh2);
+					break;
 				} else {
 					bhs[out++] = bh2;
 				}
 			}
 		}
-		while (out < BH_LRU_SIZE)
-			bhs[out++] = NULL;
-		memcpy(this_cpu_ptr(&bh_lrus.bhs), bhs, sizeof(bhs));
+		/*
+		 * it is fine that the value out may be smaller than
+		 * BH_LRU_SIZE. The rest of the value in bh_lrus.bhs is NULL.
+		 */
+		memcpy(this_cpu_ptr(&bh_lrus.bhs), bhs,
+				sizeof(struct buffer_head *) * out);
 	}
 	bh_lru_unlock();
-
-	if (evictee)
-		__brelse(evictee);
 }
 
 /*
Signature · 7 lines
2.1.0

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

#3 lore raw
Quoted 34 lines · yalin wang

why not change like this:

fs/buffer.c +7 -10
index 82283ab..d6769f1 100644
@@ -1287,40 +1287,31 @@ static inline void check_irqs_on(void)
  */
 static void bh_lru_install(struct buffer_head *bh)
 {
-       struct buffer_head *evictee = NULL;
+       struct buffer_head *old = NULL;
 
        check_irqs_on();
        bh_lru_lock();
        if (__this_cpu_read(bh_lrus.bhs[0]) != bh) {
-               struct buffer_head *bhs[BH_LRU_SIZE];
-               int in;
+               struct buffer_head *temp;
                int out = 0;
 
+               old = __this_cpu_read(bh_lrus.bhs[0]);
                get_bh(bh);
-               bhs[out++] = bh;
-               for (in = 0; in < BH_LRU_SIZE; in++) {
-                       struct buffer_head *bh2 =
-                               __this_cpu_read(bh_lrus.bhs[in]);
-
-                       if (bh2 == bh) {
-                               __brelse(bh2);
+               __this_cpu_write(bh_lrus.bhs[out++], bh);
+               for (; out < BH_LRU_SIZE; out++) {
+                       if (old == bh || old == NULL) {
+                               break;

fs/buffer.c static inline void check_irqs_on(void) · line 1307

Quoted 10 lines
fs/buffer.c +3 -6
                        } else {
-                               if (out >= BH_LRU_SIZE) {
-                                       BUG_ON(evictee != NULL);
-                                       evictee = bh2;
-                               } else {
-                                       bhs[out++] = bh2;
-                               }
+                               temp = __this_cpu_read(bh_lrus.bhs[out]);
+                               __this_cpu_write(bh_lrus.bhs[out], old);
+                               old = temp;

If we should copy the successive struct buffer_head, it is appropriate to use memcpy to copy a bunch of struct buffer_head.

Thanks
Minfei

Quoted 120 lines
fs/buffer.c +2 -5
                        }
                }
-               while (out < BH_LRU_SIZE)
-                       bhs[out++] = NULL;
-               memcpy(this_cpu_ptr(&bh_lrus.bhs), bhs, sizeof(bhs));
        }
        bh_lru_unlock();
 
-       if (evictee)
-               __brelse(evictee);
+       if (old)
+               __brelse(old);
 }
 
 /*

more simple to understand and have better performance . am i understanding correctly ?

Quoted 99 lines · Minfei Huang

Ping, Could you someone help to review this patch?

Thanks
Minfei

Quoted 86 lines · Minfei Huang

From: Minfei Huang <[email protected]>

There is a buffer_head lru list cache in local cpu to accelerate the speed. The LRU management algorithm is simple enough in bh_lru_install().

There are three situtaions we should deal with.
1) All/part of the lru cache is NULL.
2) The new buffer_head hitts the lru cache.
3) The new buffer_head does hit the lru cache.

We put the new buffer_head at the head of lru cache, then copy the buffer_head from the original lru cache, and drop the spare.

---
fs/buffer.c | 32 ++++++++++++++++++++------------

1 file changed, 20 insertions(+), 12 deletions(-)
fs/buffer.c +20 -12
index 1cf7a53..2139574 100644
@@ -1287,8 +1287,6 @@ static inline void check_irqs_on(void)
 */
 static void bh_lru_install(struct buffer_head *bh)
 {
-	struct buffer_head *evictee = NULL;
-
 	check_irqs_on();
 	bh_lru_lock();
 	if (__this_cpu_read(bh_lrus.bhs[0]) != bh) {
@@ -1302,25 +1300,35 @@ static void bh_lru_install(struct buffer_head *bh)
 			struct buffer_head *bh2 =
 				__this_cpu_read(bh_lrus.bhs[in]);
 
-			if (bh2 == bh) {
+			if (bh2 == NULL) {
+				/* Rest value in bh_lrus.bhs always is NULL */
+				break;
+			} else if (bh2 == bh) {
 				__brelse(bh2);
 			} else {
-				if (out >= BH_LRU_SIZE) {
-					BUG_ON(evictee != NULL);
-					evictee = bh2;
+				if (out == BH_LRU_SIZE) {
+					/*
+					 * this condition will be happened,
+					 * only if none of entry in
+					 * bh_lrus.bhs hits the new bh,
+					 * so the last bh should be released.
+					 */
+					BUG_ON(in != BH_LRU_SIZE - 1);
+					__brelse(bh2);
+					break;
 				} else {
 					bhs[out++] = bh2;
 				}
 			}
 		}
-		while (out < BH_LRU_SIZE)
-			bhs[out++] = NULL;
-		memcpy(this_cpu_ptr(&bh_lrus.bhs), bhs, sizeof(bhs));
+		/*
+		 * it is fine that the value out may be smaller than
+		 * BH_LRU_SIZE. The rest of the value in bh_lrus.bhs is NULL.
+		 */
+		memcpy(this_cpu_ptr(&bh_lrus.bhs), bhs,
+				sizeof(struct buffer_head *) * out);
 	}
 	bh_lru_unlock();
-
-	if (evictee)
-		__brelse(evictee);
 }
 
 /*
Signature · 7 lines
2.1.0

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/