Closed hartwork closed 1 year ago
It has been my intention to enable use_pty
by default once the behavior of sudo with a pty is as close as possible to the behavior without one. With some recent bug fixes we may have reached that point. Some behavior will always be different, for example a command run in a pty will consume terminal input even if the command being run does not need it. There are probably still some minor issues (proxying signals and job control from one terminal to another adds complexity) but those bugs will be fixed as they are found. I will tentatively plan for sudo 1.9.14 to enable use_pty
by default unless there is a major problem found in the meantime.
@millert cool!
I wonder how much need there is for a new option no_pty
or some other way to temporarily disable use_pty
once use_pty
is enabled by default. How do you feel about that?
There is no need for a new option, users can just negate use_pty
, for example:
Defaults !use_pty
@millert I wasn't aware, great, thanks! :+1:
Sudo 1.9.14 is out and sets use_pty by default.
@millert thanks a lot — awesome!
For anyone looking for all related commits, this is what I found (in reverse order as git log
would do):
Clarify that use_pty is on by default starting with 1.9.14.
Sudo runs the command in a pty by default in 1.9.14 and above.
Add commented out example for disabling use_pty.
sudo 1.9.14
(the change log part)Enable the use_pty option by default for sudo 1.9.14.
This seems to have caused a regression on Arch Linux. The following command produces a "staircase" effect under 1.9.14.p1:
$ sudo pacman -S wget | tee wget.log
resolving dependencies...
looking for conflicting packages...
Packages (1) wget-1.21.4-1
Total Installed Size: 3.18 MiB
:: Proceed with installation? [Y/n]
checking keyring...
checking package integrity...
loading package files...
checking for file conflicts...
checking available disk space...
:: Processing package changes...
installing wget...
Optional dependencies for wget
ca-certificates: HTTPS downloads [installed]
:: Running post-transaction hooks...
(1/1) Arming ConditionNeedsUpdate...
There is no problem when rolling back to 1.9.13.p3. The problem also goes away when setting Defaults !use_pty
in sudoers.
I haven't found any other way to trigger it (yet). Maybe it's a bug in pacman? Any thoughts? Thanks.
That indicates that the terminal is not set to convert newlines into return + newline pairs. Specifically, that the ONLCR flag is not set in the c_oflag
field of struct termios
(or post-processing is disabled). You can see the value of this flag by running stty -a
from the shell and looking for onlcr
and opost
.
The most likely cause is that either 1) sudo is being run from a terminal with the ONLCR or OPOST flags off, or 2) pacman is changing the terminal mode itself. When sudo allocates a pty it copies most of the terminal modes from the user's terminal to the pty.
Thanks for the hints. onlcr
and opost
are both definitely on. That seems to point the finger at pacman. Although, tee
might also be playing a role here...will investigate.
@Toolybird Does it happen with just sudo pacman
or only with tee? If only with tee, does the same thing happen if you pipe the output to cat?
It doesn't happen with just sudo pacman
. It also doesn't happen with cat
, but in this case the pacman command doesn't execute properly (possibly because of the [Y/n] prompt)
I can reproduce this running arch in docker (actually podman). It also happens for:
sudo podman -S wget | cat
So it is not specific to tee. It only happens after Proceed with installation
is displayed, the earlier output is OK. The terminal settings on the pty look fine too. I'll try to debug this further tomorrow.
I've confirmed this is a sudo problem and just committed a fix which is dependent on a few other commits from today. Basically, when sudo runs a command that has its output piped or redirected it leaves the user's terminal settings unchanged until the program being run needs to read from the terminal. At that point, sudo sets the user's terminal (not the pty running the command) into raw mode so it can proxy input from the user's terminal to the pty. However, this has the effect of turning off output post-processing, leading to the stair-step output. The fix is to preserve the terminal's output flags if output is not to a terminal.
Just putting this out there so someone can find the answer without having to figure it out for themselves - this breaks kdesu/yast
I discovered this when trying to run Yast software management, I'd get a bouncing icon and then nothing.
Running the program from the CLI returns the following:
> /usr/lib/YaST2/bin/sw_single_wrapper
Don't need password!!
kf.su: [ /home/abuild/rpmbuild/BUILD/kdesu-5.108.0/src/stubprocess.cpp : 215 ] Unknown request: "ok"
I have run kwriteconfig5 --file kdesurc --group super-user-command --key super-user-command sudo
so that kdesu will use sudo rather than su, so that sudoers will be respected by KDE.
KDE frameworks got an update around the same time as this change rolled out, so the natural thought was that kdesu was broken, but it turns out to be this option.
Running echo 'Defaults !use_pty' | sudo tee -a /etc/sudoers
and then repeating the above instantly fixes it:
> /usr/lib/YaST2/bin/sw_single_wrapper
Don't need password!!
and the GUI appears and works as intended.
I intended to upstream this to KDE but the issue already exists there: https://bugs.kde.org/show_bug.cgi?id=452532
I will provide them this information also and encourage them to fix kdesu so that it it works properly with the secure option enabled.
Edit: A patch has already been merged to kdesu so that it will work with sudo's new default. Hooray! and thanks to Fabian Vogt for fixing it up for us all. Edit2: Merged already! Thanks again Fabian!
Since use_pty is the default, I have noticed some strange behavior using FreeBSD 13 on single-core systems. Sometimes (i.e. 1 out of 10 times) sudo seems to hang. top shows two sudo processes that together generate 100% load. It takes about a minute for the system to become responsive again and for sudo to continue running properly. I don't have this problem on systems with multiple cores.
The simplest solution is obviously to set "Default !use_pty". But it didn't leave me alone, which is why I spent some time debugging. Apparently there can be situations where the while loop to wait for the parent process in the exec_cmnd_pty function (exec_monitor.c) drives the child process to full load, thus blocking the parent process or rather making it very slow. I changed the wait time of nanosleep from 1µs to 10µs and now the problem doesn't seem to occur anymore.
Do you have any thoughts on this?
@koenigbj It's possible that FreeBSD doesn't really support such short sleep times. I just committed 0cb3e33 which replaces that loop with a read from the other end of a socketpair. That should remove the possibility of a tight CPU loop.
@millert Thanks a lot! 0cb3e33 works like a charm for me. I have now been able to invoke sudo a thousand times without it getting stuck.
In principle, FreeBSD supports such short sleep times, but the hardware must of course also play along. In my case we are talking about virtual machines and therefore there is a virtual clock driver, which could be the cause.
Hello there.
I noticed a regression due to enabling this option by default. (See #312)
In summary, if you run sudo in parallel with use_pty
enabled, it may break tty. https://github.com/go-task/task/issues/1369 has an aciinema and sample code that shows this behavior.
While I now understand the security implications that lead to this decision, it would be great if we could somehow fix this regression. Any suggestions are welcome.
Thanks!
Hello there. I noticed a regression due to enabling this option by default. (See #312) In summary, if you run sudo in parallel with
use_pty
enabled, it may break tty. go-task/task#1369 has an aciinema and sample code that shows this behavior.While I now understand the security implications that lead to this decision, it would be great if we could somehow fix this regression. Any suggestions are welcome.
Thanks!
Fixed by fabb626
Hello! :wave:
I became aware recently that sudo needs non-default option
use_pty
in/etc/sudoers
— a lineDefaults use_pty
— to no longer be vulnerable to privilege escalation viaTIOCSTI
and/or lesser-knownTIOCLINUX
command injection by default. For anyone curious to see the attack with their own eyes, ttyjack can be used.While Linux >=6.2.0 introduced a config switch
CONFIG_LEGACY_TIOCSTI
(that is defaulty
upstream,n
in Arch Linux) and a related sysctl variabledev.tty.legacy_tiocsti
to disableTIOCSTI
, there is no equivalent forTIOCLINUX
available and use of a PTY may be the best protection for both attacks and similar future ones in sudo and other software, because the code under execution no longer gets access to the outer controlling terminal.This timeline and CVE list potentially indicates how little known the attack vector is to this day, 12+ years after the introduction of
use_pty
by commit 6f05b565c3937e873715dd9e424f602e14bf1984 for sudo 1.8.0 and ~18 years after the first(?) public report about security aspects ofTIOCSTI
in 2005.As of today, sudo documents the issue in
man 5 sudoers
near optionuse_pty
saying:Is there a chance that option
use_pty
could be enabled by default in sudo?Thanks for your consideration! :pray:
Best, Sebastian