[<prev] [next>] [thread-next>] [day] [month] [year] [list]
Message-Id: <1460417755-18201-1-git-send-email-avagin@openvz.org>
Date: Mon, 11 Apr 2016 16:35:40 -0700
From: Andrey Vagin <avagin@...nvz.org>
To: linux-kernel@...r.kernel.org
Cc: Andrey Vagin <avagin@...nvz.org>, Oleg Nesterov <oleg@...hat.com>,
Andrew Morton <akpm@...ux-foundation.org>,
Cyrill Gorcunov <gorcunov@...nvz.org>,
Pavel Emelyanov <xemul@...allels.com>,
Roger Luethi <rl@...lgate.ch>, Arnd Bergmann <arnd@...db.de>,
Arnaldo Carvalho de Melo <acme@...nel.org>,
David Ahern <dsahern@...il.com>,
Andy Lutomirski <luto@...capital.net>,
Pavel Odintsov <pavel.odintsov@...il.com>
Subject: [PATCH 0/15] task_diag: add a new interface to get information about processes (v3)
Current interface is a bunch of files in /proc/PID. While this appears to be
simple and there are a number of problems with it.
* Lots of syscalls
At least three syscalls per each PID are required — open(), read(), and
close()
* Variety of formats
There are many different formats used by files in /proc/PID/ hierarchy.
Therefore, there is a need to write parser for each such format.
* Non-extendable formats
Some formats in /proc/PID are non-extendable. For example, /proc/PID/maps
last column (file name) is optional, therefore there is no way to add more
columns without breaking the format.
* Slow read due to extra info[edit]
Sometimes getting information is slow due to extra attributes that are not
always needed. For example, /proc/PID/smaps contains VmFlags field (which
can't be added to /proc/PID/maps, see previous item), but it also contains
page stats that take long time to generate.
$ time cat /proc/*/maps > /dev/null
real 0m0.061s
user 0m0.002s
sys 0m0.059s
$ time cat /proc/*/smaps > /dev/null
real 0m0.253s
user 0m0.004s
sys 0m0.247s
Proposed solution
-----------------
The proposed solution is the /proc/task_diag file, which operates based on the
following principles:
* Transactional: write request, read response
* Netlink message format (same as used by sock_diag; binary and extendable)
* Ability to specify a set of processes to get info about
* Optimal grouping of attributes
Any attribute in a group can't affect a response time
The user-kernel interface is encapsulated in include/uapi/linux/task_diag.h
A request is described by the task_diag_pid structure:
struct task_diag_pid {
__u64 show_flags; /* specify which information are required */
__u64 dump_stratagy; /* specify a group of processes */
__u32 pid;
};
dump_stratagy specifies a group of processes:
/* system wide strategies (the pid fiel is ignored) */
TASK_DIAG_DUMP_ALL - all processes
TASK_DIAG_DUMP_ALL_THREAD - all threads
/* per-process strategies */
TASK_DIAG_DUMP_CHILDREN - all children
TASK_DIAG_DUMP_THREAD - all threads
TASK_DIAG_DUMP_ONE - one process
show_flags specifies which information are required. If we set the
TASK_DIAG_SHOW_BASE flag, the response message will contain the TASK_DIAG_BASE
attribute which is described by the task_diag_base structure.
struct task_diag_base {
__u32 tgid;
__u32 pid;
__u32 ppid;
__u32 tpid;
__u32 sid;
__u32 pgid;
__u8 state;
char comm[TASK_DIAG_COMM_LEN];
};
In future, it can be extended by optional attributes. The request describes
which task properties are required and for which processes they are required
for.
A response can be divided into a few netlink packets. Each task is described
by a netlink message. If all information about a process doesn't fit into a
message, the TASK_DIAG_FLAG_CONT flag will be set and the next message will
continue describing the same process.
The task diag is much faster than the proc file system. We don't need to create
a new file descriptor for each task. We need to send a request and get a
response. It allows to get information for a few tasks for one request-response
iteration.
As for security, task_diag always works as procfs with hidepid = 2 (highest
level of security).
I have compared performance of procfs and task-diag for the
"ps ax -o pid,ppid" command.
ps uses /proc/PID/* files:
$ time ./ps/pscommand ax | wc -l
50089
real 0m1.596s
user 0m0.475s
sys 0m1.126s
ps uses the task_diag interface
$ time ./ps/pscommand ax | wc -l
50089
real 0m0.148s
user 0m0.069s
sys 0m0.086s
Read /proc/PID/stat for 30K tasks:
$ time ./task_proc_all > /dev/null
real 0m0.258s
user 0m0.019s
sys 0m0.232s
Get the same information via task_diag:
$ time ./task_diag_all > /dev/null
real 0m0.052s
user 0m0.013s
sys 0m0.036s
And here are statistics on syscalls which were called by each
command.
$ perf trace -s -o log -- ./task_proc_all > /dev/null
Summary of events:
task_proc_all (30781), 180785 events, 100.0%, 0.000 msec
syscall calls min avg max stddev
(msec) (msec) (msec) (%)
--------------- -------- --------- --------- --------- ------
read 30111 0.000 0.013 0.107 0.21%
write 1 0.008 0.008 0.008 0.00%
open 30111 0.007 0.012 0.145 0.24%
close 30112 0.004 0.011 0.110 0.20%
fstat 3 0.009 0.013 0.016 16.15%
mmap 8 0.011 0.020 0.027 11.24%
mprotect 4 0.019 0.023 0.028 8.33%
munmap 1 0.026 0.026 0.026 0.00%
brk 8 0.007 0.015 0.024 11.94%
ioctl 1 0.007 0.007 0.007 0.00%
access 1 0.019 0.019 0.019 0.00%
execve 1 0.000 0.000 0.000 0.00%
getdents 29 0.008 1.010 2.215 8.88%
arch_prctl 1 0.016 0.016 0.016 0.00%
openat 1 0.021 0.021 0.021 0.00%
$ perf trace -s -o log -- ./task_diag_all > /dev/null
Summary of events:
task_diag_all (30762), 717 events, 98.9%, 0.000 msec
syscall calls min avg max stddev
(msec) (msec) (msec) (%)
--------------- -------- --------- --------- --------- ------
read 2 0.000 0.008 0.016 100.00%
write 197 0.008 0.019 0.041 3.00%
open 2 0.023 0.029 0.036 22.45%
close 3 0.010 0.012 0.014 11.34%
fstat 3 0.012 0.044 0.106 70.52%
mmap 8 0.014 0.031 0.054 18.88%
mprotect 4 0.016 0.023 0.027 10.93%
munmap 1 0.022 0.022 0.022 0.00%
brk 1 0.040 0.040 0.040 0.00%
ioctl 1 0.011 0.011 0.011 0.00%
access 1 0.032 0.032 0.032 0.00%
getpid 1 0.012 0.012 0.012 0.00%
socket 1 0.032 0.032 0.032 0.00%
sendto 2 0.032 0.095 0.157 65.77%
recvfrom 129 0.009 0.235 0.418 2.45%
bind 1 0.018 0.018 0.018 0.00%
execve 1 0.000 0.000 0.000 0.00%
arch_prctl 1 0.012 0.012 0.012 0.00%
You can find the test programs from this experiment in tools/test/selftest/task_diag.
The idea of this functionality was suggested by Pavel Emelyanov (xemul@),
when he found that operations with /proc forms a significant part
of a checkpointing time.
Ten years ago there was attempt to add a netlink interface to access to /proc
information:
http://lwn.net/Articles/99600/
Links
-----
kernel: https://github.com/avagin/linux-task-diag
procps: https://github.com/avagin/procps-task-diag
wiki: https://criu.org/Task-diag
Changes from the first version:
-------------------------------
David Ahern implemented all required functionality to use task_diag in
perf.
Bellow you can find his results how it affects performance.
> Using the fork test command:
> 10,000 processes; 10k proc with 5 threads = 50,000 tasks
> reading /proc: 11.3 sec
> task_diag: 2.2 sec
>
> @7,440 tasks, reading /proc is at 0.77 sec and task_diag at 0.096
>
> 128 instances of sepcjbb, 80,000+ tasks:
> reading /proc: 32.1 sec
> task_diag: 3.9 sec
>
> So overall much snappier startup times.
Many thanks to David Ahern for the help with improving task_diag.
Changes from the second version:
--------------------------------
Use a proc transation file instead of the netlink interface.
Andy Lutomirski pointed out on security problems related to netlink sockets:
> Slightly off-topic, but this netlink is really rather bad as an
> example of how fds can be used as capabilities (in the real capability
> sense, not the Linux capabilities sense). You call socket and get a
> socket. That socket captures f_cred. Then you drop privs, and you
> assume that the socket you're holding on to retains the right to do
> certain things.
>
> This breaks pretty badly when, through things such as this patch set,
> existing code that creates netlink sockets suddenly starts capturing
> brand-new rights that didn't exist as part of a netlink socket before.
Cc: Oleg Nesterov <oleg@...hat.com>
Cc: Andrew Morton <akpm@...ux-foundation.org>
Cc: Cyrill Gorcunov <gorcunov@...nvz.org>
Cc: Pavel Emelyanov <xemul@...allels.com>
Cc: Roger Luethi <rl@...lgate.ch>
Cc: Arnd Bergmann <arnd@...db.de>
Cc: Arnaldo Carvalho de Melo <acme@...nel.org>
Cc: David Ahern <dsahern@...il.com>
Cc: Andy Lutomirski <luto@...capital.net>
Cc: Pavel Odintsov <pavel.odintsov@...il.com>
Signed-off-by: Andrey Vagin <avagin@...nvz.org>
--
2.1.0
Powered by blists - more mailing lists