When jbd2 tries to get write access to one buffer, and if this buffer
is under writeback with BH_Shadow flag, jbd2 will wait until this buffer
has been written to disk, but sometimes the time taken to wait may be
much long, especially disk capacity is almost full.
Here add a proc entry "force-copy", if its value is not zero, jbd2 will
always do meta buffer copy-cout, then we can eliminate the unnecessary
wating time here, and reduce long tail latency for buffered-write.
I construct such test case below:
$cat offline.fio
; fio-rand-RW.job for fiotest
[global]
name=fio-rand-RW
filename=fio-rand-RW
rw=randrw
rwmixread=60
rwmixwrite=40
bs=4K
direct=0
numjobs=4
time_based=1
runtime=900
[file1]
size=60G
ioengine=sync
iodepth=16
$cat online.fio
; fio-seq-write.job for fiotest
[global]
name=fio-seq-write
filename=fio-seq-write
rw=write
bs=256K
direct=0
numjobs=1
time_based=1
runtime=60
[file1]
rate=50m
size=10G
ioengine=sync
iodepth=16
With this patch:
$cat /proc/fs/jbd2/sda5-8/force_copy
0
online fio almost always get such long tail latency:
Jobs: 1 (f=1), 0B/s-0B/s: [W(1)][100.0%][w=50.0MiB/s][w=200 IOPS][eta
00m:00s]
file1: (groupid=0, jobs=1): err= 0: pid=17855: Thu Nov 15 09:45:57 2018
write: IOPS=200, BW=50.0MiB/s (52.4MB/s)(3000MiB/60001msec)
clat (usec): min=135, max=4086.6k, avg=867.21, stdev=50338.22
lat (usec): min=139, max=4086.6k, avg=871.16, stdev=50338.22
clat percentiles (usec):
| 1.00th=[ 141], 5.00th=[ 143], 10.00th=[ 145],
| 20.00th=[ 147], 30.00th=[ 147], 40.00th=[ 149],
| 50.00th=[ 149], 60.00th=[ 151], 70.00th=[ 153],
| 80.00th=[ 155], 90.00th=[ 159], 95.00th=[ 163],
| 99.00th=[ 255], 99.50th=[ 273], 99.90th=[ 429],
| 99.95th=[ 441], 99.99th=[3640656]
$cat /proc/fs/jbd2/sda5-8/force_copy
1
online fio latency is much better.
Jobs: 1 (f=1), 0B/s-0B/s: [W(1)][100.0%][w=50.0MiB/s][w=200 IOPS][eta
00m:00s]
file1: (groupid=0, jobs=1): err= 0: pid=8084: Thu Nov 15 09:31:15 2018
write: IOPS=200, BW=50.0MiB/s (52.4MB/s)(3000MiB/60001msec)
clat (usec): min=137, max=545, avg=151.35, stdev=16.22
lat (usec): min=140, max=548, avg=155.31, stdev=16.65
clat percentiles (usec):
| 1.00th=[ 143], 5.00th=[ 145], 10.00th=[ 145], 20.00th=[
147],
| 30.00th=[ 147], 40.00th=[ 147], 50.00th=[ 149], 60.00th=[
149],
| 70.00th=[ 151], 80.00th=[ 155], 90.00th=[ 157], 95.00th=[
161],
| 99.00th=[ 239], 99.50th=[ 269], 99.90th=[ 420], 99.95th=[
429],
| 99.99th=[ 537]
As to the cost: because we'll always need to copy meta buffer, will
consume minor cpu time and some memory(at most 32MB for 128MB journal
size).
Signed-off-by: Xiaoguang Wang <[email protected]>
v2:
add missing "force-copy" directory delete operation
---
fs/jbd2/journal.c | 58 ++++++++++++++++++++++++++++++++++++++++++++++++++++
include/linux/jbd2.h | 5 +++++
2 files changed, 63 insertions(+)
diff --git a/fs/jbd2/journal.c b/fs/jbd2/journal.c
index 8ef6b6daaa7a..7fdaf501610a 100644
--- a/fs/jbd2/journal.c
+++ b/fs/jbd2/journal.c
@@ -418,6 +418,9 @@ int jbd2_journal_write_metadata_buffer(transaction_t *transaction,
}
kunmap_atomic(mapped_data);
+ /* force copy-out */
+ if (need_copy_out == 0 && journal->j_force_copy)
+ need_copy_out = 1;
/*
* Do we need to do a data copy?
*/
@@ -1100,6 +1103,58 @@ static const struct file_operations jbd2_seq_info_fops = {
.release = jbd2_seq_info_release,
};
+static int jbd2_seq_force_copy_show(struct seq_file *m, void *v)
+{
+ journal_t *journal = m->private;
+
+ seq_printf(m, "%u\n", journal->j_force_copy);
+ return 0;
+}
+
+static int jbd2_seq_force_copy_open(struct inode *inode, struct file *filp)
+{
+ journal_t *journal = PDE_DATA(inode);
+
+ return single_open(filp, jbd2_seq_force_copy_show, journal);
+}
+
+/* Worst case buffer size needed for holding an integer. */
+#define PROC_NUMBUF 13
+
+static ssize_t jbd2_seq_force_copy_write(struct file *file,
+ const char __user *buf, size_t count, loff_t *offset)
+{
+ struct inode *inode = file_inode(file);
+ journal_t *journal = PDE_DATA(inode);
+ char buffer[PROC_NUMBUF];
+ unsigned int force_copy;
+ int err;
+
+ memset(buffer, 0, sizeof(buffer));
+ if (count > sizeof(buffer) - 1)
+ count = sizeof(buffer) - 1;
+ if (copy_from_user(buffer, buf, count)) {
+ err = -EFAULT;
+ goto out;
+ }
+
+ err = kstrtouint(strstrip(buffer), 0, &force_copy);
+ if (err)
+ goto out;
+ journal->j_force_copy = force_copy;
+out:
+ return err < 0 ? err : count;
+}
+
+static const struct file_operations jbd2_seq_force_copy_fops = {
+ .owner = THIS_MODULE,
+ .open = jbd2_seq_force_copy_open,
+ .read = seq_read,
+ .write = jbd2_seq_force_copy_write,
+ .llseek = seq_lseek,
+ .release = single_release,
+};
+
static struct proc_dir_entry *proc_jbd2_stats;
static void jbd2_stats_proc_init(journal_t *journal)
@@ -1108,12 +1163,15 @@ static void jbd2_stats_proc_init(journal_t *journal)
if (journal->j_proc_entry) {
proc_create_data("info", S_IRUGO, journal->j_proc_entry,
&jbd2_seq_info_fops, journal);
+ proc_create_data("force_copy", 0644, journal->j_proc_entry,
+ &jbd2_seq_force_copy_fops, journal);
}
}
static void jbd2_stats_proc_exit(journal_t *journal)
{
remove_proc_entry("info", journal->j_proc_entry);
+ remove_proc_entry("force_copy", journal->j_proc_entry);
remove_proc_entry(journal->j_devname, proc_jbd2_stats);
}
diff --git a/include/linux/jbd2.h b/include/linux/jbd2.h
index b708e5169d1d..7034888eac6b 100644
--- a/include/linux/jbd2.h
+++ b/include/linux/jbd2.h
@@ -1111,6 +1111,11 @@ struct journal_s
*/
struct transaction_stats_s j_stats;
+ /**
+ * @j_force_copy: if not zero, force to do buffer copy-out.
+ */
+ unsigned int j_force_copy;
+
/**
* @j_failed_commit: Failed journal commit ID.
*/
--
2.14.4.44.g2045bb6
hi,
Sorry, you can ignore this patch, this fixing method maybe not good.
Regards,
Xiaoguang Wang
> When jbd2 tries to get write access to one buffer, and if this buffer
> is under writeback with BH_Shadow flag, jbd2 will wait until this buffer
> has been written to disk, but sometimes the time taken to wait may be
> much long, especially disk capacity is almost full.
>
> Here add a proc entry "force-copy", if its value is not zero, jbd2 will
> always do meta buffer copy-cout, then we can eliminate the unnecessary
> wating time here, and reduce long tail latency for buffered-write.
>
> I construct such test case below:
>
> $cat offline.fio
> ; fio-rand-RW.job for fiotest
>
> [global]
> name=fio-rand-RW
> filename=fio-rand-RW
> rw=randrw
> rwmixread=60
> rwmixwrite=40
> bs=4K
> direct=0
> numjobs=4
> time_based=1
> runtime=900
>
> [file1]
> size=60G
> ioengine=sync
> iodepth=16
>
> $cat online.fio
> ; fio-seq-write.job for fiotest
>
> [global]
> name=fio-seq-write
> filename=fio-seq-write
> rw=write
> bs=256K
> direct=0
> numjobs=1
> time_based=1
> runtime=60
>
> [file1]
> rate=50m
> size=10G
> ioengine=sync
> iodepth=16
>
> With this patch:
> $cat /proc/fs/jbd2/sda5-8/force_copy
> 0
>
> online fio almost always get such long tail latency:
>
> Jobs: 1 (f=1), 0B/s-0B/s: [W(1)][100.0%][w=50.0MiB/s][w=200 IOPS][eta
> 00m:00s]
> file1: (groupid=0, jobs=1): err= 0: pid=17855: Thu Nov 15 09:45:57 2018
> write: IOPS=200, BW=50.0MiB/s (52.4MB/s)(3000MiB/60001msec)
> clat (usec): min=135, max=4086.6k, avg=867.21, stdev=50338.22
> lat (usec): min=139, max=4086.6k, avg=871.16, stdev=50338.22
> clat percentiles (usec):
> | 1.00th=[ 141], 5.00th=[ 143], 10.00th=[ 145],
> | 20.00th=[ 147], 30.00th=[ 147], 40.00th=[ 149],
> | 50.00th=[ 149], 60.00th=[ 151], 70.00th=[ 153],
> | 80.00th=[ 155], 90.00th=[ 159], 95.00th=[ 163],
> | 99.00th=[ 255], 99.50th=[ 273], 99.90th=[ 429],
> | 99.95th=[ 441], 99.99th=[3640656]
>
> $cat /proc/fs/jbd2/sda5-8/force_copy
> 1
>
> online fio latency is much better.
>
> Jobs: 1 (f=1), 0B/s-0B/s: [W(1)][100.0%][w=50.0MiB/s][w=200 IOPS][eta
> 00m:00s]
> file1: (groupid=0, jobs=1): err= 0: pid=8084: Thu Nov 15 09:31:15 2018
> write: IOPS=200, BW=50.0MiB/s (52.4MB/s)(3000MiB/60001msec)
> clat (usec): min=137, max=545, avg=151.35, stdev=16.22
> lat (usec): min=140, max=548, avg=155.31, stdev=16.65
> clat percentiles (usec):
> | 1.00th=[ 143], 5.00th=[ 145], 10.00th=[ 145], 20.00th=[
> 147],
> | 30.00th=[ 147], 40.00th=[ 147], 50.00th=[ 149], 60.00th=[
> 149],
> | 70.00th=[ 151], 80.00th=[ 155], 90.00th=[ 157], 95.00th=[
> 161],
> | 99.00th=[ 239], 99.50th=[ 269], 99.90th=[ 420], 99.95th=[
> 429],
> | 99.99th=[ 537]
>
> As to the cost: because we'll always need to copy meta buffer, will
> consume minor cpu time and some memory(at most 32MB for 128MB journal
> size).
> Signed-off-by: Xiaoguang Wang <[email protected]>
>
> v2:
> add missing "force-copy" directory delete operation
> ---
> fs/jbd2/journal.c | 58 ++++++++++++++++++++++++++++++++++++++++++++++++++++
> include/linux/jbd2.h | 5 +++++
> 2 files changed, 63 insertions(+)
>
> diff --git a/fs/jbd2/journal.c b/fs/jbd2/journal.c
> index 8ef6b6daaa7a..7fdaf501610a 100644
> --- a/fs/jbd2/journal.c
> +++ b/fs/jbd2/journal.c
> @@ -418,6 +418,9 @@ int jbd2_journal_write_metadata_buffer(transaction_t *transaction,
> }
> kunmap_atomic(mapped_data);
>
> + /* force copy-out */
> + if (need_copy_out == 0 && journal->j_force_copy)
> + need_copy_out = 1;
> /*
> * Do we need to do a data copy?
> */
> @@ -1100,6 +1103,58 @@ static const struct file_operations jbd2_seq_info_fops = {
> .release = jbd2_seq_info_release,
> };
>
> +static int jbd2_seq_force_copy_show(struct seq_file *m, void *v)
> +{
> + journal_t *journal = m->private;
> +
> + seq_printf(m, "%u\n", journal->j_force_copy);
> + return 0;
> +}
> +
> +static int jbd2_seq_force_copy_open(struct inode *inode, struct file *filp)
> +{
> + journal_t *journal = PDE_DATA(inode);
> +
> + return single_open(filp, jbd2_seq_force_copy_show, journal);
> +}
> +
> +/* Worst case buffer size needed for holding an integer. */
> +#define PROC_NUMBUF 13
> +
> +static ssize_t jbd2_seq_force_copy_write(struct file *file,
> + const char __user *buf, size_t count, loff_t *offset)
> +{
> + struct inode *inode = file_inode(file);
> + journal_t *journal = PDE_DATA(inode);
> + char buffer[PROC_NUMBUF];
> + unsigned int force_copy;
> + int err;
> +
> + memset(buffer, 0, sizeof(buffer));
> + if (count > sizeof(buffer) - 1)
> + count = sizeof(buffer) - 1;
> + if (copy_from_user(buffer, buf, count)) {
> + err = -EFAULT;
> + goto out;
> + }
> +
> + err = kstrtouint(strstrip(buffer), 0, &force_copy);
> + if (err)
> + goto out;
> + journal->j_force_copy = force_copy;
> +out:
> + return err < 0 ? err : count;
> +}
> +
> +static const struct file_operations jbd2_seq_force_copy_fops = {
> + .owner = THIS_MODULE,
> + .open = jbd2_seq_force_copy_open,
> + .read = seq_read,
> + .write = jbd2_seq_force_copy_write,
> + .llseek = seq_lseek,
> + .release = single_release,
> +};
> +
> static struct proc_dir_entry *proc_jbd2_stats;
>
> static void jbd2_stats_proc_init(journal_t *journal)
> @@ -1108,12 +1163,15 @@ static void jbd2_stats_proc_init(journal_t *journal)
> if (journal->j_proc_entry) {
> proc_create_data("info", S_IRUGO, journal->j_proc_entry,
> &jbd2_seq_info_fops, journal);
> + proc_create_data("force_copy", 0644, journal->j_proc_entry,
> + &jbd2_seq_force_copy_fops, journal);
> }
> }
>
> static void jbd2_stats_proc_exit(journal_t *journal)
> {
> remove_proc_entry("info", journal->j_proc_entry);
> + remove_proc_entry("force_copy", journal->j_proc_entry);
> remove_proc_entry(journal->j_devname, proc_jbd2_stats);
> }
>
> diff --git a/include/linux/jbd2.h b/include/linux/jbd2.h
> index b708e5169d1d..7034888eac6b 100644
> --- a/include/linux/jbd2.h
> +++ b/include/linux/jbd2.h
> @@ -1111,6 +1111,11 @@ struct journal_s
> */
> struct transaction_stats_s j_stats;
>
> + /**
> + * @j_force_copy: if not zero, force to do buffer copy-out.
> + */
> + unsigned int j_force_copy;
> +
> /**
> * @j_failed_commit: Failed journal commit ID.
> */
>