[PATCH 0/1 RFC] read_policy type heuristic

[PATCH 0/1 RFC] read_policy type heuristic

[Anand Jain]

POC patch proposes a read_policy type Heuristic to solve the problem of
poor read performance for the mirrored stripes on mixed device types.
Read_policy Heuristic has benefits of both pid and has no dependency on
block layer iostats. Instead, it uses statically defined device_type and
its latency.

This patch is a POC patch that has no performance analysis yet. I am
sending this to get some early feedback.

For the read_policy type Latency, please refer here [1]. The read_policy
type Latency works better overall, and it is self-tuning based on the IO.

[1]
https://lore.kernel.org/linux-btrfs/20210120123437.OVx7ybGaVfmOdZxtpp43qcB_ORHQQs5OzPSzr3ZUGbo@z/T/#u

Anand Jain (1):
  btrfs: add read_policy heuristic

 fs/btrfs/sysfs.c   |  2 +-
 fs/btrfs/volumes.c | 52 ++++++++++++++++++++++++++++++++++++++++++++++
 fs/btrfs/volumes.h |  2 ++
 3 files changed, 55 insertions(+), 1 deletion(-)

-- 
2.33.1

[PATCH 1/1 RFC] btrfs: add read_policy heuristic

[Anand Jain]

Read_policy type Heuristic provides the low latency device-type if
read_policy type PID fails. The read policy heuristic retains the
advantages of the PID-based read_policy and has no dependency on
block layer iostats. This policy can be a default policy.

This patch depends on the patches[1], which provides device types
in the order of their latency.
[1] 
 btrfs: create chunk device type aware 
 btrfs: keep device type in the struct btrfs_device

Usage:
  $ echo heuristic > /sys/fs/btrfs/4b74dc57-526c-42c0-882a-8be30e1b8933/read_policy

Signed-off-by: Anand Jain <anand.jain@oracle.com>
---
 fs/btrfs/sysfs.c   |  2 +-
 fs/btrfs/volumes.c | 52 ++++++++++++++++++++++++++++++++++++++++++++++
 fs/btrfs/volumes.h |  2 ++
 3 files changed, 55 insertions(+), 1 deletion(-)

diff --git a/fs/btrfs/sysfs.c b/fs/btrfs/sysfs.c
index d5d0717fd09a..501c216009fb 100644
--- a/fs/btrfs/sysfs.c
+++ b/fs/btrfs/sysfs.c
@@ -1169,7 +1169,7 @@ static bool strmatch(const char *buffer, const char *string)
 	return false;
 }
 
-static const char * const btrfs_read_policy_name[] = { "pid" };
+static const char * const btrfs_read_policy_name[] = { "pid", "heuristic" };
 
 static ssize_t btrfs_read_policy_show(struct kobject *kobj,
 				      struct kobj_attribute *a, char *buf)
diff --git a/fs/btrfs/volumes.c b/fs/btrfs/volumes.c
index ad053c2a63a5..7efdd1dabf7b 100644
--- a/fs/btrfs/volumes.c
+++ b/fs/btrfs/volumes.c
@@ -5896,6 +5896,54 @@ int btrfs_is_parity_mirror(struct btrfs_fs_info *fs_info, u64 logical, u64 len)
 	return ret;
 }
 
+static int btrfs_read_heuristic(struct btrfs_fs_info *fs_info,
+				struct map_lookup *map, int first,
+				int num_stripes)
+{
+	int cur;
+	int new;
+	int stripe;
+	int latency;
+	int curlatency;
+
+	/*
+	 * First get the preferred mirror as per current pid and check its
+	 * latency
+	 */
+	stripe = first + (current->pid % num_stripes);
+	latency = btrfs_dev_type_to_latency(map->stripes[stripe].dev->dev_type);
+
+	/* If the device selected already has the lowest latency then return */
+	if (latency == 0)
+		return stripe;
+
+	cur = stripe;
+	curlatency = latency;
+
+	/* Find if any other stripes have lower latency */
+	for (new = first; new < first + num_stripes; new++) {
+		struct btrfs_device *device = map->stripes[new].dev;
+		int newlatency;
+
+		newlatency = btrfs_dev_type_to_latency(device->dev_type);
+
+		if (newlatency < curlatency) {
+			cur = new;
+			curlatency = newlatency;
+		}
+	}
+
+	/*
+	 * If the PID chosen device also has the same latency, then use that
+	 * device.
+	 */
+	if (curlatency == latency)
+		return stripe;
+
+	return cur;
+}
+
+
 static int find_live_mirror(struct btrfs_fs_info *fs_info,
 			    struct map_lookup *map, int first,
 			    int dev_replace_is_ongoing)
@@ -5925,6 +5973,10 @@ static int find_live_mirror(struct btrfs_fs_info *fs_info,
 	case BTRFS_READ_POLICY_PID:
 		preferred_mirror = first + (current->pid % num_stripes);
 		break;
+	case BTRFS_READ_POLICY_HEURISTIC:
+		preferred_mirror = btrfs_read_heuristic(fs_info, map, first,
+							num_stripes);
+		break;
 	}
 
 	if (dev_replace_is_ongoing &&
diff --git a/fs/btrfs/volumes.h b/fs/btrfs/volumes.h
index 0fe20eb4087a..c8a73722bf60 100644
--- a/fs/btrfs/volumes.h
+++ b/fs/btrfs/volumes.h
@@ -277,6 +277,8 @@ enum btrfs_chunk_allocation_policy {
 enum btrfs_read_policy {
 	/* Use process PID to choose the stripe */
 	BTRFS_READ_POLICY_PID,
+	/* Find and use device with the lowest latency */
+	BTRFS_READ_POLICY_HEURISTIC,
 	BTRFS_NR_READ_POLICY,
 };
 
-- 
2.33.1