Created attachment 242561 [details]
Test case for illustration problem with process accounting

It seems I found situation when process accounting do not append records for 
terminated processes.


How reproduce (kernel versions 3.17-rc1 - 4.9-rc1):

Create empty file for accounting, call system call acct() with this
file, sleep for not less than one jiffy, create new process and exit this 
process. Now records for terminated processes does not append to accounting 
file. And this state keep until process accounting restarted.
Note, system call acct() in this procedure return successfully, with zero.
It is important for reproduce that after process accounting on with 
acct() no exit of some process during one tick happen (current jiffies 
must increment before some process exit). On my system this happen very rare, 
and problem reproduce almost always.

I wrote program test.c which implement described above steps. Then, program test
size of accounting file. If size remain zero, then it seems problem.


How it was found (and possible cause):

I investigated bug in program atop 1.26-2 (Monitor for system resources 
and process activity):
https://bugs.launchpad.net/ubuntu/+source/atop/+bug/1022865
https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=778598

This bug sometimes reproduce on my system. Atop sometimes crash with SIGFPE 
on system start or at midnight (when atop restart with cron). 
After some debugging I found that this atop crash happen whenever described 
problem with process accounting happen.

After a study of the kernel/acct.c it seems that I found source of problem. 
I think the problem is in function check_free_space(), that is used to check the
free space on filesystem with accounting file. 
Lines (kernel versions 3.17-rc1 - 4.9-rc1)

if (time_is_before_jiffies(acct->needcheck))
		goto out;

are used for testing, whether it is time to check free space. 
Variable acct->needcheck is used for keeping next time to check free space.
If condition is true, then it is not time to check free space, it branch to end 
of function. If condition is false, then it check free space, state of process 
accounting (acct->active) is changed accordingly, next time to check free space 
write to acct->needcheck. But it is necessary to use function 
"time_is_before_jiffies" then!?

Whenever process accounting is switched on with acct(), variable 
acct->needcheck is set to current jiffies, acct->active is set to zero 
(disabled). If between process accounting switched on and first exit of process 
jiffies did not increment, then branch with "goto" will not happen, free space 
will check, and if free space is present, accounting will activate 
(acct->active = 1).  If between process accounting switched on and first exit 
of process passed more than one jiffy, then jiffies will be greater than 
acct->needcheck, branch will happen with "goto", acct->active remains zero. 
From this moment, current jiffies will be greater than acct->needcheck always, 
and always acct->active equal 0, and records for terminated processes does not 
append to accounting file.
Such behaviour observe in kernel versions 3.17-rc1 - 4.9-rc1.

So I suppose that this problem may be solve in versions 3.17-rc1 - 4.9-rc1 by 
following patch:

----------
diff --git a/kernel/acct.c b/kernel/acct.c
index 74963d1..37f1dc6 100644
--- a/kernel/acct.c
+++ b/kernel/acct.c
@@ -99,7 +99,7 @@ static int check_free_space(struct bsd_acct_struct *acct)
 {
        struct kstatfs sbuf;
 
-       if (time_is_before_jiffies(acct->needcheck))
+       if (time_is_after_jiffies(acct->needcheck))
                goto out;
 
        /* May block */
----------


In kernel versions 3.3-rc1 - 3.16:

In kernel versions 3.3-rc1 - 3.16 activation of process accounting implemented 
differently, so delay between call acct(filename) and process termination do not
produce problem, and program test.c do not detect problem. But, it seems, using
function time_is_before_jiffies is not right similarly.
Another problem arise, if during work of process accounting happen that current 
jiffies is greater than acct->needcheck (for example, if between two consecutive 
process terminations happen interval greater than ACCT_TIMEOUT seconds). 
Then in lines:

	if (!file || time_is_before_jiffies(acct->needcheck))
		goto out;

always will branch with "goto" and acct->needcheck will not change. So free 
space will not check more, until accounting restart. It is not good.
Note, that in version 3.17-rc1 - 4.9-rc1 this problem is also present.

Therefore I suppose that this problem for kernel version 3.3-rc1 - 3.16 may be 
solve by following patch:

----------
diff --git a/kernel/acct.c b/kernel/acct.c
index 808a86f..591bdcd 100644
--- a/kernel/acct.c
+++ b/kernel/acct.c
@@ -107,7 +107,7 @@ static int check_free_space(struct bsd_acct_struct *acct, struct file *file)
 
        spin_lock(&acct_lock);
        res = acct->active;
-       if (!file || time_is_before_jiffies(acct->needcheck))
+       if (!file || time_is_after_jiffies(acct->needcheck))
                goto out;
        spin_unlock(&acct_lock);
----------


In kernel 3.3 another method use for define time to check 
free space (by timer). So I not found in these versions such problem.


Sorry if it is all my mistake.
And sorry for my bad English.

Dmitry