Bug 62903 - Better handling of a thread limit forced by cgroups via systemd
Summary: Better handling of a thread limit forced by cgroups via systemd
Status: NEW
Alias: None
Product: Apache httpd-2
Classification: Unclassified
Component: Core (show other bugs)
Version: 2.4.29
Hardware: PC Linux
: P2 normal (vote)
Target Milestone: ---
Assignee: Apache HTTPD Bugs Mailing List
URL:
Keywords:
Depends on:
Blocks:
 
Reported: 2018-11-12 08:33 UTC by Erik Wasser
Modified: 2018-11-12 08:40 UTC (History)
1 user (show)



Attachments

Note You need to log in before you can comment on or make changes to this bug.
Description Erik Wasser 2018-11-12 08:33:24 UTC
Hi everyone,

this is mainly a copy of my posting on the httpd-users mailing list[1] because I think there is a bug somewhere deep down:

I'm stumbled upon some strange issues with the apache2. At first some facts about the environment there the apache2 was running:

* Virtuozzo 7 with a stripped down version of Ubuntu 18.04 as container OS.
* The thread/processes limit is set to 1500 for the container.
* The apache2 is using the `mpm_worker_module` with this configuration:

<IfModule mpm_worker_module>
     ServerLimit              16
     StartServers             3
     MinSpareThreads          75
     MaxSpareThreads          250
     ThreadLimit              64
     ThreadsPerChild          25
     MaxRequestWorkers        400
     MaxConnectionsPerChild   10000
</IfModule>

* I'm using apache2-2.4.29-1ubuntu4.4.

* I was dealing with some kind of benchmarking and I used this command `ab -kc 1000 -t 60 http://foo.invalid/` for it. The returned page was a simple static HTML page with not FastCGI/PHP or any other "fancy" stuff.

I was expecting some delay with the benchmarks because the number of connections `1000` were larger than the upper thread limit of the `mpm_worker_module`. This should not a problem at all because apache2 shows some good error messages if it runs into some process/thread limit.

But in this case, this is what I found in the `error.log` (removed the timestamp for readability):

… [mpm_worker:alert] [pid 2835:tid 140106655919040] (11)Resource temporarily unavailable: AH00282: apr_thread_create: unable to create worker thread
… [mpm_worker:alert] [pid 2898:tid 140106655919040] (11)Resource temporarily unavailable: AH00282: apr_thread_create: unable to create worker thread
… [mpm_worker:alert] [pid 2938:tid 140106655919040] (11)Resource temporarily unavailable: AH00282: apr_thread_create: unable to create worker thread

Or

… [mpm_worker:crit] [pid 25242:tid 140005187725056] (22)Invalid argument: AH03139: ap_queue_pop failed
… [mpm_worker:crit] [pid 25242:tid 140005187725056] (22)Invalid argument: AH03139: ap_queue_pop failed
… [mpm_worker:crit] [pid 25242:tid 140005187725056] (22)Invalid argument: AH03139: ap_queue_pop failed
… [mpm_worker:alert] [pid 25396:tid 140005187929856] (11)Resource temporarily unavailable: AH03142: apr_thread_create: unable to create worker thread

Or

… [core:error] [pid 401:tid 140005310155712] AH00546: no record of generation 0 of exiting child 25241
… [core:error] [pid 401:tid 140005310155712] AH00546: no record of generation 0 of exiting child 25396

And no other messages.

Sometimes the apache2 hangs with 100% CPU usage even after the ab testing was long over. Here the line from `htop`:

 > 5875 www-data  20  0  372M  3340  1068 S 198.  0.1  0:31.16 /usr/sbin/apache2 -k start

The load was going up to 20 even without any further requests. I've to restart the apache2 to fix this.

`strace` during the ab run shows this:

2835  clone(child_stack=0x7f6d09becf70, 
flags=CLONE_VM|CLONE_FS|CLONE_FILES|CLONE_SIGHAND|CLONE_THREAD|CLONE_SYSVSEM|CLONE_SETTLS|CLONE_PARENT_SETTID|CLONE_CHILD_CLEARTID,parent_tidptr=0x7f6d09bed9d0, tls=0x7f6d09bed700, child_tidptr=0x7f6d09bed9d0) = -1 EAGAIN (Resource temporarily unavailable)
2835  clone(child_stack=0x7f6d09becf70, flags=CLONE_VM|CLONE_FS|CLONE_FILES|CLONE_SIGHAND|CLONE_THREAD|CLONE_SYSVSEM|CLONE_SETTLS|CLONE_PARENT_SETTID|CLONE_CHILD_CLEARTID,parent_tidptr=0x7f6d09bed9d0, tls=0x7f6d09bed700, child_tidptr=0x7f6d09bed9d0) = -1 EAGAIN (Resource temporarily unavailable)
2835  clone(child_stack=0x7f6d09becf70, flags=CLONE_VM|CLONE_FS|CLONE_FILES|CLONE_SIGHAND|CLONE_THREAD|CLONE_SYSVSEM|CLONE_SETTLS|CLONE_PARENT_SETTID|CLONE_CHILD_CLEARTID,parent_tidptr=0x7f6d09bed9d0, tls=0x7f6d09bed700, child_tidptr=0x7f6d09bed9d0) = -1 EAGAIN (Resource temporarily unavailable)

So someone was holding back the apache2 from creating new threads. To make a long story short: It was systemd that limited the total processes of the apache2 cgroup to 225 threads:

$ systemctl status apache2 | grep Tasks
     Tasks: 194 (limit: 225)

Setting the systemd value `DefaultTasksMax=infinity` and everything was fine but 2 questions are still remaining:

1. The error message shows that apache2 is not ready to handle successfully a fail of the clone() call. I think this is a bug.

2. Can apache2 detect a task limit via cgroups on it's own to detect this kind of misconfiguration? In this case the max configured threads were larger than the limit of the cgroup. Is it possible to detect this and issue a warning to the logfiles? In other cases of limits like these the apache2 is more helpful with the given error messages.

[1]http://mail-archives.apache.org/mod_mbox/httpd-users/201810.mbox/%3C544ce613-3f9b-5f0f-c52e-2946af4396e9%40maxcluster.de%3E