Forza/misc
Forza/misc
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Btrfs: Allocator Hints: updates to the latency rr policy

Browse Source
This commit is contained in:
Forza 2025-04-18 16:26:15 +02:00
parent 11f410819e
commit 9224c3ede4
1 changed files with 173 additions and 63 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

236
Btrfs/Allocator Hints/btrfs_allocator_hints-6.12_v4.patch
View File

@ -1,7 +1,7 @@
From 5e49c78f38cc7f5b7ec012021c8422c1db98ef7e Mon Sep 17 00:00:00 2001
From: Goffredo Baroncelli <kreijack@inwind.it>
Date: Sun, 24 Oct 2021 17:31:04 +0200
Subject: [PATCH 01/25] btrfs: add flags to give an hint to the chunk allocator
Subject: [PATCH 01/26] btrfs: add flags to give an hint to the chunk allocator
Add the following flags to give an hint about which chunk should be
allocated in which a disk.
@ -50,7 +50,7 @@ index fc29d273845d84..71c6135dc7cfb2 100644
From 160344ae9ae37b32593adc43716172c37b0a734c Mon Sep 17 00:00:00 2001
From: Goffredo Baroncelli <kreijack@inwind.it>
Date: Sun, 24 Oct 2021 17:31:05 +0200
Subject: [PATCH 02/25] btrfs: export dev_item.type in
Subject: [PATCH 02/26] btrfs: export dev_item.type in
/sys/fs/btrfs/<uuid>/devinfo/<devid>/type
Signed-off-by: Goffredo Baroncelli <kreijack@inwind.it>
@ -91,7 +91,7 @@ index 03926ad467c919..fe07a7cbcf74c4 100644
From 29637f2e3a69fe77a8097bd772a8a7803b9ec576 Mon Sep 17 00:00:00 2001
From: Goffredo Baroncelli <kreijack@inwind.it>
Date: Sun, 24 Oct 2021 17:31:06 +0200
Subject: [PATCH 03/25] btrfs: change the DEV_ITEM 'type' field via sysfs
Subject: [PATCH 03/26] btrfs: change the DEV_ITEM 'type' field via sysfs
Signed-off-by: Kai Krakow <kai@kaishome.de>
---
@ -197,7 +197,7 @@ index 4481575dd70f35..7bb14d51bffc58 100644
From 970b99e160487e9765b6e7db9f8a89a96ce79811 Mon Sep 17 00:00:00 2001
From: Goffredo Baroncelli <kreijack@inwind.it>
Date: Sun, 24 Oct 2021 17:31:07 +0200
Subject: [PATCH 04/25] btrfs: add allocator_hint mode
Subject: [PATCH 04/26] btrfs: add allocator_hint mode
When this mode is enabled, the chunk allocation policy is modified as
follow.
@ -388,7 +388,7 @@ index 7bb14d51bffc58..f3c5437e270a22 100644
From 1c1f2e27d3055b7721468c6980479a043f48e2b3 Mon Sep 17 00:00:00 2001
From: Kai Krakow <kk@netactive.de>
Date: Thu, 27 Jun 2024 20:05:58 +0200
Subject: [PATCH 05/25] btrfs: add allocator_hint for no allocation preferred
Subject: [PATCH 05/26] btrfs: add allocator_hint for no allocation preferred
This is useful where you want to prevent new allocations of chunks on a
disk which is going to removed from the pool anyways, e.g. due to bad
@ -441,7 +441,7 @@ index 71c6135dc7cfb2..92bcc59b129a97 100644
From 82553effe6b655f97478b6d13df7ab0ecc192e58 Mon Sep 17 00:00:00 2001
From: Kai Krakow <kai@kaishome.de>
Date: Fri, 6 Dec 2024 00:55:31 +0100
Subject: [PATCH 06/25] btrfs: add allocator_hint to disable allocation
Subject: [PATCH 06/26] btrfs: add allocator_hint to disable allocation
completely
This is useful where you want to prevent new allocations of chunks to
@ -516,7 +516,7 @@ index 92bcc59b129a97..3db20734aacfc6 100644
From 10248db4c682397c83b99daa2de4ee0e587c0be2 Mon Sep 17 00:00:00 2001
From: Anand Jain <anand.jain@oracle.com>
Date: Thu, 2 Jan 2025 02:06:31 +0800
Subject: [PATCH 07/25] btrfs: simplify output formatting in
Subject: [PATCH 07/26] btrfs: simplify output formatting in
btrfs_read_policy_show
Refactor the logic in btrfs_read_policy_show() to streamline the
@ -562,7 +562,7 @@ index 3675d961b39a2a..cde47f1c11757f 100644
From 4a49a279c14d9003fd7d4865706bc78142bf1645 Mon Sep 17 00:00:00 2001
From: Anand Jain <anand.jain@oracle.com>
Date: Thu, 2 Jan 2025 02:06:30 +0800
Subject: [PATCH 08/25] btrfs: initialize fs_devices->fs_info earlier
Subject: [PATCH 08/26] btrfs: initialize fs_devices->fs_info earlier
Currently, fs_devices->fs_info is initialized in btrfs_init_devices_late(),
but this occurs too late for find_live_mirror(), which is invoked by
@ -606,7 +606,7 @@ index 99d2c60ac2bf3e..21cc02df8edf06 100644
From ccb29226710d52abbd737fd0b2f438022c045af4 Mon Sep 17 00:00:00 2001
From: Anand Jain <anand.jain@oracle.com>
Date: Thu, 2 Jan 2025 02:06:32 +0800
Subject: [PATCH 09/25] btrfs: add btrfs_read_policy_to_enum helper and
Subject: [PATCH 09/26] btrfs: add btrfs_read_policy_to_enum helper and
refactor read policy store
Introduce the `btrfs_read_policy_to_enum` helper function to simplify the
@ -683,7 +683,7 @@ index cde47f1c11757f..8540af0807648e 100644
From cf73e9084375ab73182d3a2d510e878a137a9664 Mon Sep 17 00:00:00 2001
From: Anand Jain <anand.jain@oracle.com>
Date: Thu, 2 Jan 2025 02:06:34 +0800
Subject: [PATCH 10/25] btrfs: add tracking of read blocks for read policy
Subject: [PATCH 10/26] btrfs: add tracking of read blocks for read policy
Add fs_devices::read_cnt_blocks to track read blocks, initialize it in
open_fs_devices() and clean it up in close_fs_devices().
@ -801,7 +801,7 @@ index f3c5437e270a22..91a2358b74c91f 100644
From 7070070e90e889d165590aa05f02e671d041d12c Mon Sep 17 00:00:00 2001
From: Kai Krakow <kai@kaishome.de>
Date: Mon, 16 Sep 2024 18:18:25 +0930
Subject: [PATCH 11/25] btrfs: introduce CONFIG_BTRFS_EXPERIMENTAL from 6.13
Subject: [PATCH 11/26] btrfs: introduce CONFIG_BTRFS_EXPERIMENTAL from 6.13
CONFIG_BTRFS_EXPERIMENTAL is needed by the RAID1 balancing patches but
we don't want to use the full scope of the 6.13 patch because it also
@ -838,7 +838,7 @@ index 4fb925e8c981d8..ead317f1eeb859 100644
From 3efa6c755e4ae0dc36f606b329b10587f24dcab3 Mon Sep 17 00:00:00 2001
From: Anand Jain <anand.jain@oracle.com>
Date: Thu, 2 Jan 2025 02:06:33 +0800
Subject: [PATCH 12/25] btrfs: handle value associated with read policy
Subject: [PATCH 12/26] btrfs: handle value associated with read policy
parameter
This change enables specifying additional configuration values alongside
@ -901,7 +901,7 @@ index 8540af0807648e..b0e624c0598f48 100644
From 687cdc03a694afb2236c7c87de458c519be771ea Mon Sep 17 00:00:00 2001
From: Anand Jain <anand.jain@oracle.com>
Date: Thu, 2 Jan 2025 02:06:35 +0800
Subject: [PATCH 13/25] btrfs: introduce round-robin read policy
Subject: [PATCH 13/26] btrfs: introduce round-robin read policy
This feature balances I/O across the striped devices when reading from
mirrored blocks.
@ -1130,7 +1130,7 @@ index 91a2358b74c91f..65d56bffc6ef8b 100644
From 328002ad27e90dc8ff6b7c2022711b6f0df74a01 Mon Sep 17 00:00:00 2001
From: Anand Jain <anand.jain@oracle.com>
Date: Thu, 2 Jan 2025 02:06:36 +0800
Subject: [PATCH 14/25] btrfs: add RAID1 preferred read device
Subject: [PATCH 14/26] btrfs: add RAID1 preferred read device
When there's stale data on a mirrored device, this feature lets you choose
which device to read from. Mainly used for testing.
@ -1276,7 +1276,7 @@ index 65d56bffc6ef8b..d8075ad17a6d3a 100644
From 5084cf69a0e706dfcae5e594d915e46a124fb25c Mon Sep 17 00:00:00 2001
From: Anand Jain <anand.jain@oracle.com>
Date: Thu, 2 Jan 2025 02:06:37 +0800
Subject: [PATCH 15/25] btrfs: expose experimental mode in module information
Subject: [PATCH 15/26] btrfs: expose experimental mode in module information
Commit c9c49e8f157e ("btrfs: split out CONFIG_BTRFS_EXPERIMENTAL from
CONFIG_BTRFS_DEBUG") introduces a way to enable or disable experimental
@ -1307,7 +1307,7 @@ index c64d0713412231..4742bb2af601a7 100644
From fd9d23cf84c07baec0ba5d4bbd9ecd4c0e671e47 Mon Sep 17 00:00:00 2001
From: Anand Jain <anand.jain@oracle.com>
Date: Thu, 2 Jan 2025 02:06:38 +0800
Subject: [PATCH 16/25] btrfs: enable read policy configuration via modprobe
Subject: [PATCH 16/26] btrfs: enable read policy configuration via modprobe
parameter
This update allows configuring the `read_policy` methods using a
@ -1454,7 +1454,7 @@ index a2a0af8f6a9f94..f61844fc2da9ab 100644
From 77f79e1f0d91253b9a2aa0ff975bf34ecf3d243e Mon Sep 17 00:00:00 2001
From: Anand Jain <anand.jain@oracle.com>
Date: Thu, 2 Jan 2025 02:06:39 +0800
Subject: [PATCH 17/25] btrfs: modload to print read policy status
Subject: [PATCH 17/26] btrfs: modload to print read policy status
Modified the Btrfs loading message to include the read policy status
if the experimental feature is enabled.
@ -1490,7 +1490,7 @@ index 448db8974cda70..ea5ff01881d706 100644
From ea9e632401927e9c38ae4b3e505fff377535f58b Mon Sep 17 00:00:00 2001
From: Anand Jain <anand.jain@oracle.com>
Date: Fri, 11 Oct 2024 10:49:17 +0800
Subject: [PATCH 18/25] btrfs: use the path with the lowest latency for RAID1
Subject: [PATCH 18/26] btrfs: use the path with the lowest latency for RAID1
reads
This feature aims to direct the read I/O to the device with the lowest
@ -1605,7 +1605,7 @@ index d8075ad17a6d3a..6c1f219f83b388 100644
From 680350c9732c58e321968974868836bf13ec5c96 Mon Sep 17 00:00:00 2001
From: Kai Krakow <kai@kaishome.de>
Date: Wed, 9 Apr 2025 14:07:18 +0200
Subject: [PATCH 19/25] btrfs: move latency-based selection into helper
Subject: [PATCH 19/26] btrfs: move latency-based selection into helper
Signed-off-by: Kai Krakow <kai@kaishome.de>
---
@ -1688,7 +1688,7 @@ index a36c2bfa339785..c2f235a02a79ea 100644
From 1f255624630f889fbd9e268b8d7a77f5ed68fa8c Mon Sep 17 00:00:00 2001
From: Kai Krakow <kai@kaishome.de>
Date: Wed, 9 Apr 2025 15:21:14 +0200
Subject: [PATCH 20/25] btrfs: fix btrfs_read_rr to use the actual number of
Subject: [PATCH 20/26] btrfs: fix btrfs_read_rr to use the actual number of
stripes
While num_stripes is identical to index at the end of the loop, index
@ -1719,25 +1719,31 @@ index c2f235a02a79ea..63384cd731ded2 100644
return ret_stripe;
}
From cbe1e71a4bb32092f0fe1cc251c2455bb8a37a78 Mon Sep 17 00:00:00 2001
From c26c5bdfbeea36dd89fcbefe1c86561a5113869a Mon Sep 17 00:00:00 2001
From: Kai Krakow <kai@kaishome.de>
Date: Tue, 15 Apr 2025 09:04:57 +0200
Subject: [PATCH 21/25] btrfs: create a helper instead of open coding device
Subject: [PATCH 21/26] btrfs: create a helper instead of open coding device
latency calculation
Signed-off-by: Kai Krakow <kai@kaishome.de>
---
fs/btrfs/volumes.c | 27 ++++++++++++++-------------
1 file changed, 14 insertions(+), 13 deletions(-)
fs/btrfs/volumes.c | 38 +++++++++++++++++++++-----------------
1 file changed, 21 insertions(+), 17 deletions(-)
diff --git a/fs/btrfs/volumes.c b/fs/btrfs/volumes.c
index 63384cd731ded2..7d47cb2e0b0411 100644
index 63384cd731ded2..14baa1b391a936 100644
--- a/fs/btrfs/volumes.c
+++ b/fs/btrfs/volumes.c
@@ -6007,6 +6007,18 @@ static int btrfs_read_preferred(struct btrfs_chunk_map *map, int first,
@@ -6007,14 +6007,29 @@ static int btrfs_read_preferred(struct btrfs_chunk_map *map, int first,
return first;
}
+/*
+ * btrfs_device_read_latency
+ *
+ * Compute the average latency of the device by dividing total latency by
+ * number of IOs.
+ */
+static u64 btrfs_device_read_latency(struct btrfs_device *device)
+{
+ u64 read_wait = part_stat_read(device->bdev, nsecs[READ]);
@ -1753,7 +1759,16 @@ index 63384cd731ded2..7d47cb2e0b0411 100644
/*
* btrfs_best_stripe
*
@@ -6022,22 +6034,11 @@ static void btrfs_best_stripe(struct btrfs_fs_info *fs_info,
* Select a stripe for reading using the average latency:
- *
- * 1. Compute the average latency of the device by dividing total latency
- * by number of IOs.
- * 2. Store minimum latency and selected stripe in best_wait / best_stripe.
+ * Store minimum latency and selected stripe in best_wait / best_stripe.
*
* Will always find at least one stripe.
*/
@@ -6022,22 +6037,11 @@ static void btrfs_best_stripe(struct btrfs_fs_info *fs_info,
struct btrfs_chunk_map *map, int first,
int num_stripes, u64 *best_wait, int *best_stripe)
{
@ -1779,10 +1794,10 @@ index 63384cd731ded2..7d47cb2e0b0411 100644
*best_wait = avg_wait;
*best_stripe = index;
From 61994a4b9cb1e5cdaaba1276f95317a71a26a755 Mon Sep 17 00:00:00 2001
From e54bf3f05a0c202c8637206b9a4bb03b1e5fe42f Mon Sep 17 00:00:00 2001
From: Kai Krakow <kai@kaishome.de>
Date: Tue, 15 Apr 2025 01:28:06 +0200
Subject: [PATCH 22/25] btrfs: add filtering by latency to btrfs_read_rr
Subject: [PATCH 22/26] btrfs: add filtering by latency to btrfs_read_rr
This introduces a new parameter to btrfs_read_rr to select whether we
filter for latency. In case the caller passes latency, we return -1 if
@ -1794,10 +1809,10 @@ Signed-off-by: Kai Krakow <kai@kaishome.de>
1 file changed, 17 insertions(+), 3 deletions(-)
diff --git a/fs/btrfs/volumes.c b/fs/btrfs/volumes.c
index 7d47cb2e0b0411..2e2d7059895d9a 100644
index 14baa1b391a936..ff6fd21aaa3226 100644
--- a/fs/btrfs/volumes.c
+++ b/fs/btrfs/volumes.c
@@ -6091,7 +6091,8 @@ static int btrfs_cmp_devid(const void *a, const void *b)
@@ -6094,7 +6094,8 @@ static int btrfs_cmp_devid(const void *a, const void *b)
* The calculated stripe index is then used to select the corresponding device
* from the list of devices, which is ordered by devid.
*/
@ -1807,7 +1822,7 @@ index 7d47cb2e0b0411..2e2d7059895d9a 100644
{
struct stripe_mirror stripes[BTRFS_RAID1_MAX_MIRRORS] = {0};
struct btrfs_device *device = map->stripes[first].dev;
@@ -6107,11 +6108,24 @@ static int btrfs_read_rr(struct btrfs_chunk_map *map, int first, int num_stripe)
@@ -6110,11 +6111,24 @@ static int btrfs_read_rr(struct btrfs_chunk_map *map, int first, int num_stripe)
fs_info->sectorsize_bits;
index = 0;
@ -1833,7 +1848,7 @@ index 7d47cb2e0b0411..2e2d7059895d9a 100644
sort(stripes, index, sizeof(struct stripe_mirror),
btrfs_cmp_devid, NULL);
@@ -6153,7 +6167,7 @@ static int find_live_mirror(struct btrfs_fs_info *fs_info,
@@ -6156,7 +6170,7 @@ static int find_live_mirror(struct btrfs_fs_info *fs_info,
break;
#ifdef CONFIG_BTRFS_EXPERIMENTAL
case BTRFS_READ_POLICY_RR:
@ -1843,10 +1858,10 @@ index 7d47cb2e0b0411..2e2d7059895d9a 100644
case BTRFS_READ_POLICY_DEVID:
preferred_mirror = btrfs_read_preferred(map, first, num_stripes);
From bd9761f9f70215bea4dd45789cbca084848da935 Mon Sep 17 00:00:00 2001
From 5eb1ebc30e9487b43e55604c139caffcaf4779cd Mon Sep 17 00:00:00 2001
From: Kai Krakow <kai@kaishome.de>
Date: Wed, 9 Apr 2025 15:59:59 +0200
Subject: [PATCH 23/25] btrfs: add hybrid latency-rr read policy
Subject: [PATCH 23/26] btrfs: add hybrid latency-rr read policy
This mode combines latency and round-robin modes by considering all
stripes within 120% of the minimum latency. It falls back to round-robin
@ -1905,10 +1920,10 @@ index fd096b83bb6c45..2014475af9716e 100644
u32 sectorsize = fs_devices->fs_info->sectorsize;
diff --git a/fs/btrfs/volumes.c b/fs/btrfs/volumes.c
index 2e2d7059895d9a..d3ab0e62c96689 100644
index ff6fd21aaa3226..f3fa68aa7fa446 100644
--- a/fs/btrfs/volumes.c
+++ b/fs/btrfs/volumes.c
@@ -6134,6 +6134,40 @@ static int btrfs_read_rr(struct btrfs_chunk_map *map, int first, int num_stripes
@@ -6137,6 +6137,40 @@ static int btrfs_read_rr(struct btrfs_chunk_map *map, int first, int num_stripes
return ret_stripe;
}
@ -1949,7 +1964,7 @@ index 2e2d7059895d9a..d3ab0e62c96689 100644
#endif
static int find_live_mirror(struct btrfs_fs_info *fs_info,
@@ -6176,6 +6210,10 @@ static int find_live_mirror(struct btrfs_fs_info *fs_info,
@@ -6179,6 +6213,10 @@ static int find_live_mirror(struct btrfs_fs_info *fs_info,
preferred_mirror = btrfs_read_fastest(fs_info, map, first,
num_stripes);
break;
@ -1974,10 +1989,10 @@ index 6c1f219f83b388..a6e8a722d9c742 100644
BTRFS_READ_POLICY_DEVID,
#endif
From ec0168f2a941c8c995f828a281f9b4eabd891466 Mon Sep 17 00:00:00 2001
From ef03ef7752a8d68310c1d31fdfad2f295c70a46d Mon Sep 17 00:00:00 2001
From: Kai Krakow <kai@kaishome.de>
Date: Tue, 15 Apr 2025 00:32:06 +0200
Subject: [PATCH 24/25] btrfs: add devinfo avg cumulative read latency to sysfs
Date: Wed, 16 Apr 2025 22:06:37 +0200
Subject: [PATCH 24/26] btrfs: add devinfo read stats to sysfs
Signed-off-by: Kai Krakow <kai@kaishome.de>
---
@ -1985,7 +2000,7 @@ Signed-off-by: Kai Krakow <kai@kaishome.de>
1 file changed, 24 insertions(+)
diff --git a/fs/btrfs/sysfs.c b/fs/btrfs/sysfs.c
index 2014475af9716e..adebb1324c9b1e 100644
index 2014475af9716e..d629ececa0b65e 100644
--- a/fs/btrfs/sysfs.c
+++ b/fs/btrfs/sysfs.c
@@ -10,6 +10,9 @@
@ -2003,7 +2018,7 @@ index 2014475af9716e..adebb1324c9b1e 100644
BTRFS_ATTR_RW(devid, type, btrfs_devinfo_type_show, btrfs_devinfo_type_store);
+#ifdef CONFIG_BTRFS_EXPERIMENTAL
+static ssize_t btrfs_devinfo_avg_read_latency_show(struct kobject *kobj,
+static ssize_t btrfs_devinfo_read_stats_show(struct kobject *kobj,
+ struct kobj_attribute *a, char *buf)
+{
+ struct btrfs_device *device = container_of(kobj, struct btrfs_device,
@ -2015,9 +2030,9 @@ index 2014475af9716e..adebb1324c9b1e 100644
+ if (read_wait && read_ios && read_wait >= read_ios)
+ avg_wait = div_u64(read_wait, read_ios);
+
+ return scnprintf(buf, PAGE_SIZE, "cum %llu\n", avg_wait);
+ return scnprintf(buf, PAGE_SIZE, "ios %lu wait %llu avg %llu\n", read_ios, read_wait, avg_wait);
+}
+BTRFS_ATTR(devid, avg_read_latency, btrfs_devinfo_avg_read_latency_show);
+BTRFS_ATTR(devid, read_stats, btrfs_devinfo_read_stats_show);
+#endif
+
/*
@ -2027,44 +2042,139 @@ index 2014475af9716e..adebb1324c9b1e 100644
*/
static struct attribute *devid_attrs[] = {
+#ifdef CONFIG_BTRFS_EXPERIMENTAL
+ BTRFS_ATTR_PTR(devid, avg_read_latency),
+ BTRFS_ATTR_PTR(devid, read_stats),
+#endif
BTRFS_ATTR_PTR(devid, error_stats),
BTRFS_ATTR_PTR(devid, fsid),
BTRFS_ATTR_PTR(devid, in_fs_metadata),
From 6535b1149f58a0b2da7df22743e1eedfbc03b87f Mon Sep 17 00:00:00 2001
From a95354ce074f5e9ff53cc931044d2c224c204d32 Mon Sep 17 00:00:00 2001
From: Kai Krakow <kai@kaishome.de>
Date: Tue, 15 Apr 2025 04:42:16 +0200
Subject: [PATCH 25/25] btrfs: ignore latency early during the first IOs
Date: Wed, 16 Apr 2025 22:52:14 +0200
Subject: [PATCH 25/26] btrfs: add last IO age to sysfs read_stats
Devices may be slow in this early phase and create spikes which most
likely disqualifies them for reading for the rest of the system
lifetime.
Each time a stripe is going to be selected, increase a counter in each
possible stripe. After selecting a stripe, reset the counter to zero.
This way we can measure how long a stripe hasn't been selected. This
could be used to add a probabilistic read to devices that have been
slow in the past.
Signed-off-by: Kai Krakow <kai@kaishome.de>
---
fs/btrfs/volumes.c | 4 ++++
1 file changed, 4 insertions(+)
fs/btrfs/sysfs.c | 4 +++-
fs/btrfs/volumes.c | 21 ++++++++++++++++++---
fs/btrfs/volumes.h | 5 +++++
3 files changed, 26 insertions(+), 4 deletions(-)
diff --git a/fs/btrfs/sysfs.c b/fs/btrfs/sysfs.c
index d629ececa0b65e..0d87f4d13c1375 100644
--- a/fs/btrfs/sysfs.c
+++ b/fs/btrfs/sysfs.c
@@ -2192,7 +2192,9 @@ static ssize_t btrfs_devinfo_read_stats_show(struct kobject *kobj,
if (read_wait && read_ios && read_wait >= read_ios)
avg_wait = div_u64(read_wait, read_ios);
- return scnprintf(buf, PAGE_SIZE, "ios %lu wait %llu avg %llu\n", read_ios, read_wait, avg_wait);
+ return scnprintf(buf, PAGE_SIZE, "ios %lu wait %llu avg %llu age %llu\n",
+ read_ios, read_wait, avg_wait,
+ (u64)atomic64_read(&device->last_io_age));
}
BTRFS_ATTR(devid, read_stats, btrfs_devinfo_read_stats_show);
#endif
diff --git a/fs/btrfs/volumes.c b/fs/btrfs/volumes.c
index d3ab0e62c96689..72fd14c170393f 100644
index f3fa68aa7fa446..21bcc8529664d8 100644
--- a/fs/btrfs/volumes.c
+++ b/fs/btrfs/volumes.c
@@ -6007,12 +6007,16 @@ static int btrfs_read_preferred(struct btrfs_chunk_map *map, int first,
return first;
}
@@ -6192,6 +6192,13 @@ static int find_live_mirror(struct btrfs_fs_info *fs_info,
else
num_stripes = map->num_stripes;
+#define BTRFS_MIN_READ_IOS_FOR_VALID_LATENCY 100
+#ifdef CONFIG_BTRFS_EXPERIMENTAL
+ /* age each possible stripe by 1 IO */
+ for (int i = first; i < first + num_stripes; i++) {
+ atomic64_inc(&map->stripes[i].dev->last_io_age);
+ }
+#endif
+
switch (policy) {
default:
/* Shouldn't happen, just warn and use pid instead of failing */
@@ -6235,14 +6242,22 @@ static int find_live_mirror(struct btrfs_fs_info *fs_info,
for (tolerance = 0; tolerance < 2; tolerance++) {
if (map->stripes[preferred_mirror].dev->bdev &&
(tolerance || map->stripes[preferred_mirror].dev != srcdev))
- return preferred_mirror;
+ goto out;
for (i = first; i < first + num_stripes; i++) {
if (map->stripes[i].dev->bdev &&
- (tolerance || map->stripes[i].dev != srcdev))
- return i;
+ (tolerance || map->stripes[i].dev != srcdev)) {
+ preferred_mirror = i;
+ goto out;
+ }
}
}
+out:
+#ifdef CONFIG_BTRFS_EXPERIMENTAL
+ /* reset age of selected stripe */
+ atomic64_set(&map->stripes[preferred_mirror].dev->last_io_age, 0);
+#endif
+
/* we couldn't find one that doesn't fail. Just return something
* and the io error handling code will clean up eventually
*/
diff --git a/fs/btrfs/volumes.h b/fs/btrfs/volumes.h
index a6e8a722d9c742..f2807a7463bf17 100644
--- a/fs/btrfs/volumes.h
+++ b/fs/btrfs/volumes.h
@@ -197,6 +197,11 @@ struct btrfs_device {
/* Bandwidth limit for scrub, in bytes */
u64 scrub_speed_max;
+
+#ifdef CONFIG_BTRFS_EXPERIMENTAL
+ /* store an age of last read access */
+ atomic64_t last_io_age;
+#endif
};
/*
From 8adf66f59b2836c78db1f50c74f76b783764e84b Mon Sep 17 00:00:00 2001
From: Kai Krakow <kai@kaishome.de>
Date: Thu, 17 Apr 2025 00:26:03 +0200
Subject: [PATCH 26/26] btrfs: Probe read latency if device is 1000 IOs behind
its siblings
This should solve a problem where devices get "frozen" if their read
latency spiked in the past.
Signed-off-by: Kai Krakow <kai@kaishome.de>
---
fs/btrfs/volumes.c | 5 ++++-
1 file changed, 4 insertions(+), 1 deletion(-)
diff --git a/fs/btrfs/volumes.c b/fs/btrfs/volumes.c
index 21bcc8529664d8..6b1aededbb1505 100644
--- a/fs/btrfs/volumes.c
+++ b/fs/btrfs/volumes.c
@@ -6013,13 +6013,16 @@ static int btrfs_read_preferred(struct btrfs_chunk_map *map, int first,
* Compute the average latency of the device by dividing total latency by
* number of IOs.
*/
+#define BTRFS_MAX_AGE_FOR_VALID_LATENCY 1000
static u64 btrfs_device_read_latency(struct btrfs_device *device)
{
u64 read_wait = part_stat_read(device->bdev, nsecs[READ]);
unsigned long read_ios = part_stat_read(device->bdev, ios[READ]);
+ u64 last_io_age = (u64)atomic64_read(&device->last_io_age);
u64 avg_wait = 0;
+ if (read_ios < BTRFS_MIN_READ_IOS_FOR_VALID_LATENCY)
+ return 0;
+
if (read_wait && read_ios && read_wait >= read_ios)
- if (read_wait && read_ios && read_wait >= read_ios)
+ if (last_io_age < BTRFS_MAX_AGE_FOR_VALID_LATENCY
+ && read_wait && read_ios && read_wait >= read_ios)
avg_wait = div_u64(read_wait, read_ios);
return avg_wait;