[PATCH] fs/proc: Per thread I/O accounting

[PATCH] fs/proc: Per thread I/O accounting

[Tomas Smetana]

Hello,
  here's an attempt to fix the bug #10702: The /proc/#/io doesn't aggregate
I/O statistics of all the process' threads -- just the main one. The patch
fixes this and also adds /proc/#/task/#/io file for per-thread accounting
which should make I/O accounting consistent with CPU stats.

Reference: http://bugzilla.kernel.org/show_bug.cgi?id=10702

Please cc: me in the answers.


Signed-off-by: Tomas Smetana <t.smetana@gmail.com>
---
 fs/proc/base.c |   48 ++++++++++++++++++++++++++++++++++++++++++++++++
 1 files changed, 48 insertions(+), 0 deletions(-)

diff --git a/fs/proc/base.c b/fs/proc/base.c
index 3b45537..5437ae3 100644
--- a/fs/proc/base.c
+++ b/fs/proc/base.c
@@ -2377,6 +2377,51 @@ static int proc_base_fill_cache(struct file *filp, void *dirent,
 #ifdef CONFIG_TASK_IO_ACCOUNTING
 static int proc_pid_io_accounting(struct task_struct *task, char *buffer)
 {
+#ifdef CONFIG_TASK_XACCT
+	unsigned long long rchar = 0, wchar = 0;
+	unsigned long long syscr = 0, syscw = 0;
+#endif
+	unsigned long long read_bytes = 0, write_bytes = 0;
+	unsigned long long cancelled_write_bytes = 0;
+	struct task_struct *t = task;
+
+	do {
+#ifdef CONFIG_TASK_XACCT
+		rchar += (unsigned long long)t->rchar;
+		wchar += (unsigned long long)t->wchar;
+		syscr += (unsigned long long)t->syscr;
+		syscw += (unsigned long long)t->syscw;
+#endif
+		read_bytes += (unsigned long long)t->ioac.read_bytes;
+		write_bytes += (unsigned long long)t->ioac.write_bytes;
+		cancelled_write_bytes +=
+		    (unsigned long long)t->ioac.cancelled_write_bytes;
+		t = next_thread(t);
+	} while (t != task);
+
+	return sprintf(buffer,
+#ifdef CONFIG_TASK_XACCT
+			"rchar: %llu\n"
+			"wchar: %llu\n"
+			"syscr: %llu\n"
+			"syscw: %llu\n"
+#endif
+			"read_bytes: %llu\n"
+			"write_bytes: %llu\n"
+			"cancelled_write_bytes: %llu\n",
+#ifdef CONFIG_TASK_XACCT
+			rchar,
+			wchar,
+			syscr,
+			syscw,
+#endif
+			read_bytes,
+			write_bytes,
+			cancelled_write_bytes);
+}
+
+static int proc_tid_io_accounting(struct task_struct *task, char *buffer)
+{
 	return sprintf(buffer,
 #ifdef CONFIG_TASK_XACCT
 			"rchar: %llu\n"
@@ -2796,6 +2841,9 @@ static const struct pid_entry tid_base_stuff[] = {
 #ifdef CONFIG_FAULT_INJECTION
 	REG("make-it-fail", S_IRUGO|S_IWUSR, fault_inject),
 #endif
+#ifdef CONFIG_TASK_IO_ACCOUNTING
+	INF("io",	S_IRUGO, tid_io_accounting)
+#endif
 };
 
 static int proc_tid_base_readdir(struct file * filp,
-- 
1.5.4.1

Re: [PATCH] fs/proc: Per thread I/O accounting

[Andrea Righi]

Tomas Smetana wrote:
> Hello,
>   here's an attempt to fix the bug #10702: The /proc/#/io doesn't aggregate
> I/O statistics of all the process' threads -- just the main one. The patch
> fixes this and also adds /proc/#/task/#/io file for per-thread accounting
> which should make I/O accounting consistent with CPU stats.
> 
> Reference: http://bugzilla.kernel.org/show_bug.cgi?id=10702
> 
> Please cc: me in the answers.

Hi Tomas,

have a look at distinct-tgid-tid-i-o-statistics.patch in -mm (also
available here http://lwn.net/Articles/283937/).

Your patch seems ok (except the rcu locking, see below), but it doesn't
account the i/o activity of a task if it creates a lot of short-lived
children doing i/o.

> 
> 
> Signed-off-by: Tomas Smetana <t.smetana@gmail.com>
> ---
>  fs/proc/base.c |   48 ++++++++++++++++++++++++++++++++++++++++++++++++
>  1 files changed, 48 insertions(+), 0 deletions(-)
> 
> diff --git a/fs/proc/base.c b/fs/proc/base.c
> index 3b45537..5437ae3 100644
> --- a/fs/proc/base.c
> +++ b/fs/proc/base.c
> @@ -2377,6 +2377,51 @@ static int proc_base_fill_cache(struct file *filp, void *dirent,
>  #ifdef CONFIG_TASK_IO_ACCOUNTING
>  static int proc_pid_io_accounting(struct task_struct *task, char *buffer)
>  {
> +#ifdef CONFIG_TASK_XACCT
> +	unsigned long long rchar = 0, wchar = 0;
> +	unsigned long long syscr = 0, syscw = 0;
> +#endif
> +	unsigned long long read_bytes = 0, write_bytes = 0;
> +	unsigned long long cancelled_write_bytes = 0;
> +	struct task_struct *t = task;
> +

rcu_read_lock();

> +	do {
> +#ifdef CONFIG_TASK_XACCT
> +		rchar += (unsigned long long)t->rchar;
> +		wchar += (unsigned long long)t->wchar;
> +		syscr += (unsigned long long)t->syscr;
> +		syscw += (unsigned long long)t->syscw;
> +#endif
> +		read_bytes += (unsigned long long)t->ioac.read_bytes;
> +		write_bytes += (unsigned long long)t->ioac.write_bytes;
> +		cancelled_write_bytes +=
> +		    (unsigned long long)t->ioac.cancelled_write_bytes;
> +		t = next_thread(t);
> +	} while (t != task);

rcu_read_unlock();

> +
> +	return sprintf(buffer,
> +#ifdef CONFIG_TASK_XACCT
> +			"rchar: %llu\n"
> +			"wchar: %llu\n"
> +			"syscr: %llu\n"
> +			"syscw: %llu\n"
> +#endif
> +			"read_bytes: %llu\n"
> +			"write_bytes: %llu\n"
> +			"cancelled_write_bytes: %llu\n",
> +#ifdef CONFIG_TASK_XACCT
> +			rchar,
> +			wchar,
> +			syscr,
> +			syscw,
> +#endif
> +			read_bytes,
> +			write_bytes,
> +			cancelled_write_bytes);
> +}
> +
> +static int proc_tid_io_accounting(struct task_struct *task, char *buffer)
> +{
>  	return sprintf(buffer,
>  #ifdef CONFIG_TASK_XACCT
>  			"rchar: %llu\n"
> @@ -2796,6 +2841,9 @@ static const struct pid_entry tid_base_stuff[] = {
>  #ifdef CONFIG_FAULT_INJECTION
>  	REG("make-it-fail", S_IRUGO|S_IWUSR, fault_inject),
>  #endif
> +#ifdef CONFIG_TASK_IO_ACCOUNTING
> +	INF("io",	S_IRUGO, tid_io_accounting)
> +#endif
>  };
>  
>  static int proc_tid_base_readdir(struct file * filp,

-Andrea