Re: [PATCH v3 1/3] block/diskstats: more accurate approximation of io_ticks for slow disks

From: Konstantin Khlebnikov
Date: Wed Mar 25 2020 - 04:02:47 EST



On 25/03/2020 09.28, Konstantin Khlebnikov wrote:


On 25/03/2020 06.40, Ming Lei wrote:
On Tue, Mar 24, 2020 at 10:06:56PM +0800, Ming Lei wrote:
On Tue, Mar 24, 2020 at 09:39:40AM +0300, Konstantin Khlebnikov wrote:
Currently io_ticks is approximated by adding one at each start and end of
requests if jiffies counter has changed. This works perfectly for requests
shorter than a jiffy or if one of requests starts/ends at each jiffy.

If disk executes just one request at a time and they are longer than two
jiffies then only first and last jiffies will be accounted.

Fix is simple: at the end of request add up into io_ticks jiffies passed
since last update rather than just one jiffy.

Example: common HDD executes random read 4k requests around 12ms.

fio --name=test --filename=/dev/sdb --rw=randread --direct=1 --runtime=30 &
iostat -x 10 sdb

Note changes of iostat's "%util" 8,43% -> 99,99% before/after patch:

Before:

Device: rrqm/s wrqm/s r/s w/s rkB/s wkB/s avgrq-sz avgqu-sz await r_await w_await svctm %util
sdbÂÂÂÂÂÂÂÂÂÂÂÂÂÂ 0,00ÂÂÂÂ 0,00ÂÂ 82,60ÂÂÂ 0,00ÂÂ 330,40ÂÂÂÂ 0,00ÂÂÂÂ 8,00ÂÂÂÂ 0,96ÂÂ 12,09ÂÂ 12,09ÂÂÂ 0,00ÂÂ 1,02ÂÂ 8,43

After:

Device: rrqm/s wrqm/s r/s w/s rkB/s wkB/s avgrq-sz avgqu-sz await r_await w_await svctm %util
sdbÂÂÂÂÂÂÂÂÂÂÂÂÂÂ 0,00ÂÂÂÂ 0,00ÂÂ 82,50ÂÂÂ 0,00ÂÂ 330,00ÂÂÂÂ 0,00ÂÂÂÂ 8,00ÂÂÂÂ 1,00ÂÂ 12,10ÂÂ 12,10ÂÂÂ 0,00Â 12,12Â 99,99

For load estimation "%util" is not as useful as average queue length,
but it clearly shows how often disk queue is completely empty.

Fixes: 5b18b5a73760 ("block: delete part_round_stats and switch to less precise counting")
Signed-off-by: Konstantin Khlebnikov <khlebnikov@xxxxxxxxxxxxxx>
---
 Documentation/admin-guide/iostats.rst | 5 ++++-
 block/bio.c | 8 ++++----
 block/blk-core.c | 4 ++--
 include/linux/genhd.h | 2 +-
 4 files changed, 11 insertions(+), 8 deletions(-)

diff --git a/Documentation/admin-guide/iostats.rst b/Documentation/admin-guide/iostats.rst
index df5b8345c41d..9b14b0c2c9c4 100644
--- a/Documentation/admin-guide/iostats.rst
+++ b/Documentation/admin-guide/iostats.rst
@@ -100,7 +100,7 @@ Field 10 -- # of milliseconds spent doing I/Os (unsigned int)
ÂÂÂÂÂ Since 5.0 this field counts jiffies when at least one request was
ÂÂÂÂÂ started or completed. If request runs more than 2 jiffies then some
-ÂÂÂ I/O time will not be accounted unless there are other requests.
+ÂÂÂ I/O time might be not accounted in case of concurrent requests.
 Field 11 -- weighted # of milliseconds spent doing I/Os (unsigned int)
ÂÂÂÂÂ This field is incremented at each I/O start, I/O completion, I/O
@@ -143,6 +143,9 @@ are summed (possibly overflowing the unsigned long variable they are
 summed to) and the result given to the user. There is no convenient
 user interface for accessing the per-CPU counters themselves.
+Since 4.19 request times are measured with nanoseconds precision and
+truncated to milliseconds before showing in this interface.
+
 Disks vs Partitions
 -------------------
diff --git a/block/bio.c b/block/bio.c
index 0985f3422556..b1053eb7af37 100644
--- a/block/bio.c
+++ b/block/bio.c
@@ -1762,14 +1762,14 @@ void bio_check_pages_dirty(struct bio *bio)
ÂÂÂÂÂ schedule_work(&bio_dirty_work);
 }
-void update_io_ticks(struct hd_struct *part, unsigned long now)
+void update_io_ticks(struct hd_struct *part, unsigned long now, bool end)
 {
ÂÂÂÂÂ unsigned long stamp;
 again:
ÂÂÂÂÂ stamp = READ_ONCE(part->stamp);
ÂÂÂÂÂ if (unlikely(stamp != now)) {
ÂÂÂÂÂÂÂÂÂ if (likely(cmpxchg(&part->stamp, stamp, now) == stamp)) {
-ÂÂÂÂÂÂÂÂÂÂÂ __part_stat_add(part, io_ticks, 1);
+ÂÂÂÂÂÂÂÂÂÂÂ __part_stat_add(part, io_ticks, end ? now - stamp : 1);
ÂÂÂÂÂÂÂÂÂ }
ÂÂÂÂÂ }
ÂÂÂÂÂ if (part->partno) {
@@ -1785,7 +1785,7 @@ void generic_start_io_acct(struct request_queue *q, int op,
ÂÂÂÂÂ part_stat_lock();
-ÂÂÂ update_io_ticks(part, jiffies);
+ÂÂÂ update_io_ticks(part, jiffies, false);
ÂÂÂÂÂ part_stat_inc(part, ios[sgrp]);
ÂÂÂÂÂ part_stat_add(part, sectors[sgrp], sectors);
ÂÂÂÂÂ part_inc_in_flight(q, part, op_is_write(op));
@@ -1803,7 +1803,7 @@ void generic_end_io_acct(struct request_queue *q, int req_op,
ÂÂÂÂÂ part_stat_lock();
-ÂÂÂ update_io_ticks(part, now);
+ÂÂÂ update_io_ticks(part, now, true);
ÂÂÂÂÂ part_stat_add(part, nsecs[sgrp], jiffies_to_nsecs(duration));
ÂÂÂÂÂ part_stat_add(part, time_in_queue, duration);
ÂÂÂÂÂ part_dec_in_flight(q, part, op_is_write(req_op));
diff --git a/block/blk-core.c b/block/blk-core.c
index abfdcf81a228..4401b30a1751 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -1337,7 +1337,7 @@ void blk_account_io_done(struct request *req, u64 now)
ÂÂÂÂÂÂÂÂÂ part_stat_lock();
ÂÂÂÂÂÂÂÂÂ part = req->part;
-ÂÂÂÂÂÂÂ update_io_ticks(part, jiffies);
+ÂÂÂÂÂÂÂ update_io_ticks(part, jiffies, true);
ÂÂÂÂÂÂÂÂÂ part_stat_inc(part, ios[sgrp]);
ÂÂÂÂÂÂÂÂÂ part_stat_add(part, nsecs[sgrp], now - req->start_time_ns);
ÂÂÂÂÂÂÂÂÂ part_stat_add(part, time_in_queue, nsecs_to_jiffies64(now - req->start_time_ns));
@@ -1379,7 +1379,7 @@ void blk_account_io_start(struct request *rq, bool new_io)
ÂÂÂÂÂÂÂÂÂ rq->part = part;
ÂÂÂÂÂ }
-ÂÂÂ update_io_ticks(part, jiffies);
+ÂÂÂ update_io_ticks(part, jiffies, false);
ÂÂÂÂÂ part_stat_unlock();
 }
diff --git a/include/linux/genhd.h b/include/linux/genhd.h
index d5c75df64bba..f1066f10b062 100644
--- a/include/linux/genhd.h
+++ b/include/linux/genhd.h
@@ -467,7 +467,7 @@ static inline void free_part_info(struct hd_struct *part)
ÂÂÂÂÂ kfree(part->info);
 }
-void update_io_ticks(struct hd_struct *part, unsigned long now);
+void update_io_ticks(struct hd_struct *part, unsigned long now, bool end);
 /* block/genhd.c */
 extern void device_add_disk(struct device *parent, struct gendisk *disk,


Looks fine:

Reviewed-by: Ming Lei <ming.lei@xxxxxxxxxx>

BTW, there is still some gap(%65 vs. 99%) between this fix and the original
accounting(before applying Mike/Mikulas's 7 patches), and it might be
one thing to improve in future.

1) test, sda is single queue virtio-scsi, which is emulated by one HDD
image

2) fio test script:
fio --direct=1 --size=128G --bsrange=4k-4k \
ÂÂÂÂÂÂÂ --runtime=20 --numjobs=1 \
ÂÂÂÂÂÂÂ --ioengine=libaio --iodepth=16 \
ÂÂÂÂÂÂÂ --iodepth_batch_submit=16 \
ÂÂÂÂÂÂÂ --iodepth_batch_complete_min=16 \
ÂÂÂÂÂÂÂ --group_reporting=1 --filename=/dev/sda \
ÂÂÂÂÂÂÂ --name=seq-test --rw=read

3) result:
- v5.6-rc with this patch
Run status group 0 (all jobs):
ÂÂÂ READ: bw=79.4MiB/s (83.3MB/s), 79.4MiB/s-79.4MiB/s (83.3MB/s-83.3MB/s), io=155
88MiB (1665MB), run=20001-20001msec

Disk stats (read/write):
ÂÂ sda: ios=25039/0, merge=375596/0, ticks=18823/0, in_queue=4330, util=99.43%


- commit 112f158f66cb (which is previous commit of 5b18b5a73760)
Run status group 0 (all jobs):
ÂÂÂ READ: bw=81.4MiB/s (85.3MB/s), 81.4MiB/s-81.4MiB/s (85.3MB/s-85.3MB/s), io=166
28MiB (1707MB), run=20001-20001msec

Disk stats (read/write):
ÂÂ sda: ios=25749/0, merge=386236/0, ticks=17963/0, in_queue=12976, util=65.20%


Oh, no. Your result is opposite.

Well, 99.43% with this patch is actually more correct result:
every millisecond there is at least one request in disk.

Old code sampled in-flight at start and end of each request, not regularly every jiffy.
And accounted whole jiffy as inactive if in-flight currently is zero.
This way statistics was biased to samples where queue is empty.


Yep, my patch guarantees correct time only for queue-depth=1
it accounts jiffies between start -> end (and end -> end).
For queue-depth > 1 jiffies between start -> start are still lost.

Without exact counter for in-flight requests we cannot distinguish
cases start -> start and end -> start

All we could could is smoothing statistics even more and
account all short gaps as activity:

--- a/block/bio.c
+++ b/block/bio.c
@@ -1769,6 +1769,7 @@ void update_io_ticks(struct hd_struct *part, unsigned long now, bool end)
ÂÂÂÂÂÂÂ stamp = READ_ONCE(part->stamp);
ÂÂÂÂÂÂÂ if (unlikely(stamp != now)) {
ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ if (likely(cmpxchg(&part->stamp, stamp, now) == stamp)) {
+ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ end |= time_before(now, stamp + HZ / 100);
ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ __part_stat_add(part, io_ticks, end ? now - stamp : 1);
ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ }
ÂÂÂÂÂÂÂ }



Thanks,
Ming