Jira (PUP-11897) 100% usage of a CPU core when an exec command sends EOF

[Corey Hickey (Jira)]

Corey Hickey created an issue

Puppet / PUP-11897 100% usage of a CPU core when an exec command sends EOF Issue Type: Bug Affects Versions: PUP 8.0.1 Assignee: Unassigned Components: Types and Providers Created: 2023/06/22 2:56 PM Priority: Normal Reporter: Corey Hickey Puppet Version:{} Seen in 7.14.0, 7.23.0, and current git 61234be0e4d6. The code originates in d4c1582ed0c2a3e. Puppet Server Version: N/A; reproducible via puppet apply. ** OS Name/Version:{} Seen in AlmaLinux 8.7 and Debian Unstable. This is probably POSIX-specific; the affected code runs under if Puppet.features.posix? Describe your issue in as much detail as possible… When puppet applies an exec resource, and the command run by that exec returns an EOF on the pipe handling stdout and stdterr, puppet ignores the EOF. This results in puppet rapidly looping over these system calls: 1359029 wait4(1359100, 0x7ffce8537cb8, WNOHANG, NULL) = 0 1359029 getpid()                        = 1359029 1359029 pselect6(7, [6], [], [], {tv_sec=0, tv_nsec=100000000}, NULL) = 1 (in [6], left {tv_sec=0, tv_nsec=99998955}) 1359029 fcntl(6, F_GETFL)               = 0x800 (flags O_RDONLY|O_NONBLOCK) 1359029 read(6, "", 4096)               = 0 There is no blocking in the loop, so puppet consumes 100% of a CPU core until the child process exits. Describe steps to reproduce… Use a manifest like /tmp/test.pp $c1 = [     '/bin/bash', '-c',     'exec /bin/sleep 10 >/dev/null 2>&1', ] exec { 'c1':     command => $c1, } Then run: time puppet apply /tmp/test.pp Desired Behavior: The timing output should be like: real    0m13.554s user    0m2.857s sys    0m0.516s Real time is expected to be high, due to the 10 second sleep, but user time should be relatively low. Actual Behavior: The timing output is: real    0m13.486s user    0m11.048s sys    0m2.265s Note the high user time; this is because puppet is busy-looping while the child sleeps for 10 seconds. I will submit a patch for this. Add Comment

This message was sent by Atlassian Jira (v8.20.21#820021-sha1:38274c8)

[Corey Hickey (Jira)]

Corey Hickey commented on PUP-11897

Re: 100% usage of a CPU core when an exec command sends EOF There are different ways to construct a command which will return an EOF. Here are several more examples: $c2 = [     '/bin/bash', '-c',     'exec >/dev/null 2>&1; sleep 10', ] exec { 'c2':     command => $c2, }   $c3 = [     '/usr/bin/perl', '-e',     'close(STDOUT); close(STDERR); sleep 10' ] exec { 'c3':     command => $c3, }   $c4 = 'exec /bin/sleep 10 >/dev/null 2>&1' exec { 'c4':     command  => $c4,     provider => 'shell', }

Add Comment

[Corey Hickey (Jira)]

Corey Hickey commented on PUP-11897

Re: 100% usage of a CPU core when an exec command sends EOF

PR submitted to: https://github.com/puppetlabs/puppet/pull/9074

Add Comment

[Corey Hickey (Jira)]

Corey Hickey updated an issue

Puppet / PUP-11897

100% usage of a CPU core when an exec command sends EOF

Change By: Corey Hickey { * } Puppet Version: { * }{*}{*}

Seen in 7.14.0, 7.23.0, and current git 61234be0e4d6. The code originates in d4c1582ed0c2a3e.

*Puppet Server Version:* N/A; reproducible via {{{}puppet apply{}}}. ** { * } OS Name/Version: { * }{*}{*}

Seen in AlmaLinux 8.7 and Debian Unstable. This is probably POSIX-specific; the affected code runs under {{if Puppet.features.posix?}}

*Describe your issue in as much detail as possible…*

When puppet applies an {{exec}} resource, and the command run by that exec returns an EOF on the pipe handling stdout and stdterr, puppet ignores the EOF. This results in puppet rapidly looping over these system calls:

{code:java}

1359029 wait4(1359100, 0x7ffce8537cb8, WNOHANG, NULL) = 0 1359029 getpid()                        = 1359029 1359029 pselect6(7, [6], [], [], {tv_sec=0, tv_nsec=100000000}, NULL) = 1 (in [6], left {tv_sec=0, tv_nsec=99998955}) 1359029 fcntl(6, F_GETFL)               = 0x800 (flags O_RDONLY|O_NONBLOCK) 1359029 read(6, "", 4096)               = 0

{code}

There is no blocking in the loop, so puppet consumes 100% of a CPU core until the child process exits.

*Describe steps to reproduce…*

Use a manifest like {{/tmp/test.pp}}

{code:java} $c1 = [     '/bin/bash', '-c',

'exec /bin/sleep 10 >/dev/null 2>&1', ] exec { 'c1':     command => $c1, }

{code} Then run: {code:java} time puppet apply /tmp/test.pp{code} *Desired Behavior:*

The timing output should be like:

{code:java}

real    0m13.554s user    0m2.857s sys    0m0.516s

{code}

Real time is expected to be high, due to the 10 second sleep, but user time should be relatively low.

*Actual Behavior:* The timing output is: {code:java}

real    0m13.486s user    0m11.048s sys    0m2.265s

{code}

Note the high user time; this is because puppet is busy-looping while the child sleeps for 10 seconds. I will submit a patch for this.

Add Comment