Skip to contentSkip to conversation

fanotify permission events on virtual filesystem

DISCUSSION Jan Kara 3 participants 6 messages Mar 20, 2019 – Apr 1, 2019 VFS

lore
Share & tools
lore

Jan Kara

author reviewer
#0 lore raw

Hello,

recently, one of our customers has reported a deadlock with fanotify. The analysis has shown that a process has put (likely by mistake) FAN_OPEN_PERM mark on /proc and / filesystem. That resulted in a deadlock like follows:

process 1: process 2: process 3:
open("/proc/process 2/maps")
- blocks waiting for response to
FAN_OPEN_PERM event

exec(2)
__do_execve_file()
- grabs current->signal->cred_guard_mutex do_open_execat()
- blocks waiting for response to FAN_OPEN_PERM event

reads fanotify event generated by process 1
create_fd()
dentry_open()
proc_maps_open()
blocks on
cred_guard_mutex of process 2.

Now this is not the only case where you can setup fanotify permissions events so that your listener deadlocks but I'd argue that this case is especially nasty and it is unrealistic to expect from userspace that it would be able to implement fanotify listener in such a way that is deadlock-free for proc filesystem since the lock dependencies there are very different. So how about we just forbid placing marks with fanotify permission events on proc? Any other virtual filesystem we should exclude?

Honza

Signature · 2 lines
Jan Kara <[email protected]>
SUSE Labs, CR
#1 lore raw
Quoted 27 lines · Jan Kara

Hello,

recently, one of our customers has reported a deadlock with fanotify. The analysis has shown that a process has put (likely by mistake) FAN_OPEN_PERM mark on /proc and / filesystem. That resulted in a deadlock like follows:

process 1: process 2: process 3:
open("/proc/process 2/maps")
- blocks waiting for response to
FAN_OPEN_PERM event

exec(2)
__do_execve_file()
- grabs current->signal->cred_guard_mutex do_open_execat()
- blocks waiting for response to FAN_OPEN_PERM event

reads fanotify event generated by process 1
create_fd()
dentry_open()
proc_maps_open()
blocks on
cred_guard_mutex of process 2.

Quoted 8 lines

Now this is not the only case where you can setup fanotify permissions events so that your listener deadlocks but I'd argue that this case is especially nasty and it is unrealistic to expect from userspace that it would be able to implement fanotify listener in such a way that is deadlock-free for proc filesystem since the lock dependencies there are very different. So how about we just forbid placing marks with fanotify permission events on proc? Any other virtual filesystem we should exclude?

I bet if we forbid placing marks on /proc, some apps would break. I always though that allowing O_PATH in event_f_flags can make sense for some apps.

What if instead of forbiding marks of /proc, we would force those marks to use O_PATH for fd creation. Some of the functionality will remain. Apps are less likely to break. Deadlocks will be less likely, although maybe still possible.

Note that the new FAN_REPORT_FID listener already excludes marking most virtual filesystems for lack of s_export_op.

Thanks,
Amir.

author reviewer
#2 lore raw
Quoted 34 lines · Amir Goldstein
Quoted 24 lines · Jan Kara

recently, one of our customers has reported a deadlock with fanotify. The analysis has shown that a process has put (likely by mistake) FAN_OPEN_PERM mark on /proc and / filesystem. That resulted in a deadlock like follows:

process 1: process 2: process 3:
open("/proc/process 2/maps")
- blocks waiting for response to
FAN_OPEN_PERM event

exec(2)
__do_execve_file()
- grabs current->signal->cred_guard_mutex do_open_execat()
- blocks waiting for response to FAN_OPEN_PERM event

reads fanotify event generated by process 1
create_fd()
dentry_open()
proc_maps_open()
blocks on
cred_guard_mutex of process 2.

Quoted 8 lines

Now this is not the only case where you can setup fanotify permissions events so that your listener deadlocks but I'd argue that this case is especially nasty and it is unrealistic to expect from userspace that it would be able to implement fanotify listener in such a way that is deadlock-free for proc filesystem since the lock dependencies there are very different. So how about we just forbid placing marks with fanotify permission events on proc? Any other virtual filesystem we should exclude?

Quoted 1 line

I bet if we forbid placing marks on /proc, some apps would break.

Well, I didn't mean all marks, just the permission ones. I'm not sure there are apps that place permission events on /proc...

Quoted 7 lines

I always though that allowing O_PATH in event_f_flags can make sense for some apps.

What if instead of forbiding marks of /proc, we would force those marks to use O_PATH for fd creation. Some of the functionality will remain. Apps are less likely to break. Deadlocks will be less likely, although maybe still possible.

Yes, that's another option. But if this is automatic, it is going to be confusing to potential users - report O_PATH fd if getting normal fd is dangerous isn't great. And also the deadlocks are there only for permission events so there's no strong reason to restrict normal ones.

Quoted 2 lines

Note that the new FAN_REPORT_FID listener already excludes marking most virtual filesystems for lack of s_export_op.

Yes, I know.

Honza

Signature · 2 lines
Jan Kara <[email protected]>
SUSE Labs, CR
#3 lore raw
Quoted 36 lines · Jan Kara
Quoted 34 lines · Amir Goldstein
Quoted 32 lines · Jan Kara

recently, one of our customers has reported a deadlock with fanotify. The analysis has shown that a process has put (likely by mistake) FAN_OPEN_PERM mark on /proc and / filesystem. That resulted in a deadlock like follows:

process 1: process 2: process 3:
open("/proc/process 2/maps")
- blocks waiting for response to
FAN_OPEN_PERM event

exec(2)
__do_execve_file()
- grabs current->signal->cred_guard_mutex do_open_execat()
- blocks waiting for response to FAN_OPEN_PERM event

reads fanotify event generated by process 1
create_fd()
dentry_open()
proc_maps_open()
blocks on
cred_guard_mutex of process 2.

Now this is not the only case where you can setup fanotify permissions events so that your listener deadlocks but I'd argue that this case is especially nasty and it is unrealistic to expect from userspace that it would be able to implement fanotify listener in such a way that is deadlock-free for proc filesystem since the lock dependencies there are very different. So how about we just forbid placing marks with fanotify permission events on proc? Any other virtual filesystem we should exclude?

Quoted 5 lines
Quoted 1 line

I bet if we forbid placing marks on /proc, some apps would break.

Well, I didn't mean all marks, just the permission ones. I'm not sure there are apps that place permission events on /proc...

Maybe not intentionally.
I once tested a few fanotify based AntiVirus solutions.
In some of them, setting an "Exclude path" on some mount point would cause mark to not be set on that path, but for one in particular, the mark was still being set on the mount so path pattern filtering was done after receiving the events.
I did not check whether /proc was blacklisted out of the box or if it could be marked/excluded from scan.
IMO, assuming that all AntiVirus vendors blacklist all virtual filesystems is an assumption that we need to validate.
[CC Marko from F-Secure for commenting on the above.]

Quoted 8 lines
Quoted 7 lines

I always though that allowing O_PATH in event_f_flags can make sense for some apps.

What if instead of forbiding marks of /proc, we would force those marks to use O_PATH for fd creation. Some of the functionality will remain. Apps are less likely to break. Deadlocks will be less likely, although maybe still possible.

Quoted 5 lines

Yes, that's another option. But if this is automatic, it is going to be confusing to potential users - report O_PATH fd if getting normal fd is dangerous isn't great. And also the deadlocks are there only for permission events so there's no strong reason to restrict normal ones.

Of course, we can do this hack only for reading permission events for virtual filesystems. The question is what would be more surprising to existing apps:
- marks for permission events no longer working on /proc
- fd from permission events no longer readable from /proc
- something completely different

Thanks,
Amir.

#4 lore raw

Amir Goldstein <[email protected]>:

Quoted 15 lines · Jan Kara
Quoted 2 lines

Well, I didn't mean all marks, just the permission ones. I'm not sure there are apps that place permission events on /proc...

Maybe not intentionally.
I once tested a few fanotify based AntiVirus solutions.
In some of them, setting an "Exclude path" on some mount point would cause mark to not be set on that path, but for one in particular, the mark was still being set on the mount so path pattern filtering was done after receiving the events.
I did not check whether /proc was blacklisted out of the box or if it could be marked/excluded from scan.
IMO, assuming that all AntiVirus vendors blacklist all virtual filesystems is an assumption that we need to validate.
[CC Marko from F-Secure for commenting on the above.]

Yeah, we have learned by experimentation to not mark some file systems.

(Also, inspecting some /proc files *during* OPEN_PERM processing of a regular file can lead to deadlocks.)

Marko

author reviewer
#5 lore raw
In reply to Amir Goldstein
Quoted 69 lines · Amir Goldstein
Quoted 36 lines · Jan Kara
Quoted 34 lines · Amir Goldstein
Quoted 32 lines · Jan Kara

recently, one of our customers has reported a deadlock with fanotify. The analysis has shown that a process has put (likely by mistake) FAN_OPEN_PERM mark on /proc and / filesystem. That resulted in a deadlock like follows:

process 1: process 2: process 3:
open("/proc/process 2/maps")
- blocks waiting for response to
FAN_OPEN_PERM event

exec(2)
__do_execve_file()
- grabs current->signal->cred_guard_mutex do_open_execat()
- blocks waiting for response to FAN_OPEN_PERM event

reads fanotify event generated by process 1
create_fd()
dentry_open()
proc_maps_open()
blocks on
cred_guard_mutex of process 2.

Now this is not the only case where you can setup fanotify permissions events so that your listener deadlocks but I'd argue that this case is especially nasty and it is unrealistic to expect from userspace that it would be able to implement fanotify listener in such a way that is deadlock-free for proc filesystem since the lock dependencies there are very different. So how about we just forbid placing marks with fanotify permission events on proc? Any other virtual filesystem we should exclude?

Quoted 5 lines
Quoted 1 line

I bet if we forbid placing marks on /proc, some apps would break.

Well, I didn't mean all marks, just the permission ones. I'm not sure there are apps that place permission events on /proc...

Maybe not intentionally.
I once tested a few fanotify based AntiVirus solutions.
In some of them, setting an "Exclude path" on some mount point would cause mark to not be set on that path, but for one in particular, the mark was still being set on the mount so path pattern filtering was done after receiving the events.
I did not check whether /proc was blacklisted out of the box or if it could be marked/excluded from scan.
IMO, assuming that all AntiVirus vendors blacklist all virtual filesystems is an assumption that we need to validate.
[CC Marko from F-Secure for commenting on the above.]

Quoted 8 lines
Quoted 7 lines

I always though that allowing O_PATH in event_f_flags can make sense for some apps.

What if instead of forbiding marks of /proc, we would force those marks to use O_PATH for fd creation. Some of the functionality will remain. Apps are less likely to break. Deadlocks will be less likely, although maybe still possible.

Quoted 5 lines

Yes, that's another option. But if this is automatic, it is going to be confusing to potential users - report O_PATH fd if getting normal fd is dangerous isn't great. And also the deadlocks are there only for permission events so there's no strong reason to restrict normal ones.

Quoted 6 lines

Of course, we can do this hack only for reading permission events for virtual filesystems. The question is what would be more surprising to existing apps:
- marks for permission events no longer working on /proc
- fd from permission events no longer readable from /proc
- something completely different

I think that option 1) is at least very explicit failure and given what Marko wrote I'm willing to give that a try. If people start complaining, we'll probably have to revert and try either reporting O_PATH fd or even FAN_NOFD. It's a bit of a gamble I agree but current status quo isn't good either.

Honza

Signature · 2 lines
Jan Kara <[email protected]>
SUSE Labs, CR