Julia Evans

Understanding how killall works using strace

Right now I’m on a million-hour train ride from New York to Montreal. So I’m looking at the output of strace because, uh, strace is cool, and it is teaching me some things about how the command line tools I use all the time work.

What strace does is capture every single system call that gets called when executing a program. System calls are the interface between userspace programs and the kernel, so looking at the output from strace is a fun way to understand how Linux works, and what’s really involved in running a program.

For example! killall! I ran

strace killall ruby1.9.1 2> killall-log.

This starts with

execve("/usr/bin/killall", ["killall", "ruby1.9.1"], [/* 48 vars */]) = 0

Every time you run a program, execve gets called to start, so execve will always be the first line.

Then this happens A WHOLE BUNCH OF TIMES:

open("/proc/4526/stat", O_RDONLY)       = 3
fstat(3, {st_mode=S_IFREG|0444, st_size=0, ...}) = 0
mmap(NULL, 4096, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0) = 0x7febbb269000
read(3, "4526 (chrome) S 4521 2607 2607 0"..., 1024) = 374
close(3)                                = 0
munmap(0x7febbb269000, 4096)            = 0

with different PIDs.

What’s going on here is that it goes through every PID. To find the PIDs, it opens the /proc directory. There’s a directory in /proc for each PID.

bork@kiwi ~/w/homepage> ls /proc
1      1495   2408   2780   3278  8065         fb
10006  1498   2409   2782   3281  8066         filesystems
10152  1500   2410   2795   3283  8068         fs
10158  1504   2411   28     3317  8069         interrupts
1021   1513   2412   2802   35    8070         iomem

The system call that does this is:


Once it’s done that, then it iterates through all the PIDs, opens /proc/$PID/stat, and checks to see if the process has the right name. The kernel isn’t involved in seeing whether or not the process has the right name, so we don’t see that in the strace output.

Once it finds a PID that it wants to kill, it runs something like

kill(11510, SIGTERM)

to kill it. SIGTERM isn’t a very serious killing-y signal – it’s signal 15, and processes can ignore it or save their state before they stop. If you run killall -9, it will sent SIGKILL to all the matching processes and it will kill them dead.

This is really neat! I never thought of killall as having to do an exhaustive search through all PIDs before, but it makes sense.

After all of that, if there was something to kill, the only thing left is exit_group(0). man 2 exit_group tells me that this exits all threads in a process, and that this system call is called at the end .of every process

If we run killall blah, and there was no "blah" process to kill, instead we see:

write(2, "blah: no process found\n", 23blah: no process found) = 23

because it needs to write “no process found” to stderr.


I have learned a couple of new things, from people’s responses to this post!

If you want to see the library calls instead of the system calls, and want to see where it does the string comparisons, you can use ltrace!

For killall, finding python3 and killing it looks like:

__asprintf_chk(0x7fff195b6988, 1, 0x403fd9, 15499, 0x7f31d919e700) = 16
fopen("/proc/15499/stat", "r")              = 0x208f8f0
free(0x0208f8d0)                            = <void>
fscanf(0x208f8f0, 0x403fe7, 0x7fff195b71b0, 0x7f31d8fa5728, 0) = 1
fclose(0x208f8f0)                           = 0
strcmp("python3", "python3")                = 0
kill(15499, 15)                             = 0

And you can attach strace or ptrace to an already-running process to see what it’s up to. @zbrdge said that he sometimes uses it to see which files Apache is accessing during a HTTP request.

Day 46: Never Graduate A pandas cookbook