Merge tag 'cap-checkpoint-restore-v5.9' of git://git.kernel.org/pub/scm/linux/kernel/git/brauner/linux
Pull checkpoint-restore updates from Christian Brauner:
"This enables unprivileged checkpoint/restore of processes.
Given that this work has been going on for quite some time the first
sentence in this summary is hopefully more exciting than the actual
final code changes required. Unprivileged checkpoint/restore has seen
a frequent increase in interest over the last two years and has thus
been one of the main topics for the combined containers &
checkpoint/restore microconference since at least 2018 (cf. [1]).
Here are just the three most frequent use-cases that were brought forward:
- The JVM developers are integrating checkpoint/restore into a Java
VM to significantly decrease the startup time.
- In high-performance computing environment a resource manager will
typically be distributing jobs where users are always running as
non-root. Long-running and "large" processes with significant
startup times are supposed to be checkpointed and restored with
CRIU.
- Container migration as a non-root user.
In all of these scenarios it is either desirable or required to run
without CAP_SYS_ADMIN. The userspace implementation of
checkpoint/restore CRIU already has the pull request for supporting
unprivileged checkpoint/restore up (cf. [2]).
To enable unprivileged checkpoint/restore a new dedicated capability
CAP_CHECKPOINT_RESTORE is introduced. This solution has last been
discussed in 2019 in a talk by Google at Linux Plumbers (cf. [1]
"Update on Task Migration at Google Using CRIU") with Adrian and
Nicolas providing the implementation now over the last months. In
essence, this allows the CRIU binary to be installed with the
CAP_CHECKPOINT_RESTORE vfs capability set thereby enabling
unprivileged users to restore processes.
To make this possible the following permissions are altered:
- Selecting a specific PID via clone3() set_tid relaxed from userns
CAP_SYS_ADMIN to CAP_CHECKPOINT_RESTORE.
- Selecting a specific PID via /proc/sys/kernel/ns_last_pid relaxed
from userns CAP_SYS_ADMIN to CAP_CHECKPOINT_RESTORE.
- Accessing /proc/pid/map_files relaxed from init userns
CAP_SYS_ADMIN to init userns CAP_CHECKPOINT_RESTORE.
- Changing /proc/self/exe from userns CAP_SYS_ADMIN to userns
CAP_CHECKPOINT_RESTORE.
Of these four changes the /proc/self/exe change deserves a few words
because the reasoning behind even restricting /proc/self/exe changes
in the first place is just full of historical quirks and tracking this
down was a questionable version of fun that I'd like to spare others.
In short, it is trivial to change /proc/self/exe as an unprivileged
user, i.e. without userns CAP_SYS_ADMIN right now. Either via ptrace()
or by simply intercepting the elf loader in userspace during exec.
Nicolas was nice enough to even provide a POC for the latter (cf. [3])
to illustrate this fact.
The original patchset which introduced PR_SET_MM_MAP had no
permissions around changing the exe link. They too argued that it is
trivial to spoof the exe link already which is true. The argument
brought up against this was that the Tomoyo LSM uses the exe link in
tomoyo_manager() to detect whether the calling process is a policy
manager. This caused changing the exe links to be guarded by userns
CAP_SYS_ADMIN.
All in all this rather seems like a "better guard it with something
rather than nothing" argument which imho doesn't qualify as a great
security policy. Again, because spoofing the exe link is possible for
the calling process so even if this were security relevant it was
broken back then and would be broken today. So technically, dropping
all permissions around changing the exe link would probably be
possible and would send a clearer message to any userspace that relies
on /proc/self/exe for security reasons that they should stop doing
this but for now we're only relaxing the exe link permissions from
userns CAP_SYS_ADMIN to userns CAP_CHECKPOINT_RESTORE.
There's a final uapi change in here. Changing the exe link used to
accidently return EINVAL when the caller lacked the necessary
permissions instead of the more correct EPERM. This pr contains a
commit fixing this. I assume that userspace won't notice or care and
if they do I will revert this commit. But since we are changing the
permissions anyway it seems like a good opportunity to try this fix.
With these changes merged unprivileged checkpoint/restore will be
possible and has already been tested by various users"
[1] LPC 2018
1. "Task Migration at Google Using CRIU"
https://www.youtube.com/watch?v=yI_1cuhoDgA&t=12095
2. "Securely Migrating Untrusted Workloads with CRIU"
https://www.youtube.com/watch?v=yI_1cuhoDgA&t=14400
LPC 2019
1. "CRIU and the PID dance"
https://www.youtube.com/watch?v=LN2CUgp8deo&list=PLVsQ_xZBEyN30ZA3Pc9MZMFzdjwyz26dO&index=9&t=2m48s
2. "Update on Task Migration at Google Using CRIU"
https://www.youtube.com/watch?v=LN2CUgp8deo&list=PLVsQ_xZBEyN30ZA3Pc9MZMFzdjwyz26dO&index=9&t=1h2m8s
[2] https://github.com/checkpoint-restore/criu/pull/1155
[3] https://github.com/nviennot/run_as_exe
* tag 'cap-checkpoint-restore-v5.9' of git://git.kernel.org/pub/scm/linux/kernel/git/brauner/linux:
selftests: add clone3() CAP_CHECKPOINT_RESTORE test
prctl: exe link permission error changed from -EINVAL to -EPERM
prctl: Allow local CAP_CHECKPOINT_RESTORE to change /proc/self/exe
proc: allow access in init userns for map_files with CAP_CHECKPOINT_RESTORE
pid_namespace: use checkpoint_restore_ns_capable() for ns_last_pid
pid: use checkpoint_restore_ns_capable() for set_tid
capabilities: Introduce CAP_CHECKPOINT_RESTORE
