Skip to contentSkip to conversation

fs: Make /proc/sys inodes be owned by global root.

PATCH In review Radoslaw Burny 3 participants 8 messages Nov 26, 2018 – Dec 1, 2018 1 message not in the archive

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 fs: Make /proc/sys inodes be owned by global root.

Due to a recent commit (d151ddc00498 - fs: Update i_[ug]id_(read|write) to translate relative to s_user_ns), inodes under /proc/sys have -1 written to their i_uid/i_gid members if a containing userns does not have entries for root in the uid/gid_map.

This wouldn't normally matter, because these values are not used for access checks. However, a later change (0bd23d09b874 - Don't modify inodes with a uid or gid unknown to the vfs) changes the kernel to prevent opens for write if the i_uid/i_gid field in the inode is -1, even if the /proc/sys-specific access checks would otherwise pass.

This causes a problem: in a userns without root mapping, even the namespace creator cannot write to e.g. /proc/sys/kernel/shmmax. This change fixes the problem by overriding i_uid/i_gid back to GLOBAL_ROOT_UID/GID.

Tested: Used a repro program that creates a user namespace without any mapping and stat'ed /proc/$PID/root/proc/sys/kernel/shmmax from outside. Before the change, it shows uid/gid of 65534, with the change it's 0.

---
 fs/proc/proc_sysctl.c | 4 ++++
 1 file changed, 4 insertions(+)
fs/proc/proc_sysctl.c +4 -0
index c5cbbdff3c3d..67379a389658 100644
@@ -499,6 +499,10 @@ static struct inode *proc_sys_make_inode(struct super_block *sb,
 
 	if (root->set_ownership)
 		root->set_ownership(head, table, &inode->i_uid, &inode->i_gid);
+	else {
+		inode->i_uid = GLOBAL_ROOT_UID;
+		inode->i_gid = GLOBAL_ROOT_GID;
+	}
 
 out:
 	return inode;
Signature · 1 line
2.20.0.rc0.387.gc7a69e6b6c-goog
#1 lore raw
Quoted 2 lines · Radoslaw Burny

Due to a recent commit (d151ddc00498 - fs: Update i_[ug]id_(read|write) to translate relative to s_user_ns),

Recent? This is commit is from 2014 and present upstream since v4.8. And the commit ID you mentioned in your commit log seems to be incorrect. I get:

81754357770ebd900801231e7bc8d151ddc00498a fs: Update i_[ug]id_(read|write) to translate relative to s_user_ns

Quoted 3 lines

inodes under /proc/sys have -1
written to their i_uid/i_gid members if a containing userns does not have entries for root in the uid/gid_map.

Thanks for the description of how to run into the issue described but is there also a practical use case today where this is happening? I ask as it would be good to know the severity of the issue in the real world today.

Quoted 10 lines

This wouldn't normally matter, because these values are not used for access checks. However, a later change (0bd23d09b874 - Don't modify inodes with a uid or gid unknown to the vfs) changes the kernel to prevent opens for write if the i_uid/i_gid field in the inode is -1, even if the /proc/sys-specific access checks would otherwise pass.

This causes a problem: in a userns without root mapping, even the namespace creator cannot write to e.g. /proc/sys/kernel/shmmax. This change fixes the problem by overriding i_uid/i_gid back to GLOBAL_ROOT_UID/GID.

We really need Seth and Eric to provide guidance here as they were the ones devising this long ago, but to me your solution seems backward. Why allow any namespace to muck with /proc/sys/ seettings?

Let's recall that this case was a corner case, and writeback was the biggest concern, and for that it was decided that you'd simply not get write access, and so its read only. Its not clear to me if things like proc were considered. For the regular file case the situation can be addressed with chown, however we can't chown proc files.

Quoted 3 lines

Tested: Used a repro program that creates a user namespace without any mapping and stat'ed /proc/$PID/root/proc/sys/kernel/shmmax from outside. Before the change, it shows uid/gid of 65534,

I thought you said it would be uid/gid -1 without your patch?

Quoted 1 line

with the change it's 0.

Note that a good way to also test issues is with the lib/test_sysctl.c module and the tools/testing/selftests/sysctl/sysctl.sh script, so if you can device a test there, once we decide what to do that would be appreciated.

Luis

Quoted 23 lines
---
 fs/proc/proc_sysctl.c | 4 ++++
 1 file changed, 4 insertions(+)
fs/proc/proc_sysctl.c +4 -0
index c5cbbdff3c3d..67379a389658 100644
@@ -499,6 +499,10 @@ static struct inode *proc_sys_make_inode(struct super_block *sb,
 
 	if (root->set_ownership)
 		root->set_ownership(head, table, &inode->i_uid, &inode->i_gid);
+	else {
+		inode->i_uid = GLOBAL_ROOT_UID;
+		inode->i_gid = GLOBAL_ROOT_GID;
+	}
 
 out:
 	return inode;
Signature · 1 line
2.20.0.rc0.387.gc7a69e6b6c-goog
Reviewed-by
#2 lore raw
Quoted 10 lines · Luis Chamberlain
Quoted 2 lines · Radoslaw Burny

Due to a recent commit (d151ddc00498 - fs: Update i_[ug]id_(read|write) to translate relative to s_user_ns),

Recent? This is commit is from 2014 and present upstream since v4.8. And the commit ID you mentioned in your commit log seems to be incorrect. I get:

81754357770ebd900801231e7bc8d151ddc00498a fs: Update i_[ug]id_(read|write) to translate relative to s_user_ns

Quoted 8 lines
Quoted 3 lines

inodes under /proc/sys have -1
written to their i_uid/i_gid members if a containing userns does not have entries for root in the uid/gid_map.

Thanks for the description of how to run into the issue described but is there also a practical use case today where this is happening? I ask as it would be good to know the severity of the issue in the real world today.

People trying to run containers without a root user in the container. It atypical but something doable.

Quoted 6 lines
Quoted 6 lines

This wouldn't normally matter, because these values are not used for access checks. However, a later change (0bd23d09b874 - Don't modify inodes with a uid or gid unknown to the vfs) changes the kernel to prevent opens for write if the i_uid/i_gid field in the inode is -1, even if the /proc/sys-specific access checks would otherwise pass.

Quoted 8 lines
Quoted 4 lines

This causes a problem: in a userns without root mapping, even the namespace creator cannot write to e.g. /proc/sys/kernel/shmmax. This change fixes the problem by overriding i_uid/i_gid back to GLOBAL_ROOT_UID/GID.

We really need Seth and Eric to provide guidance here as they were the ones devising this long ago, but to me your solution seems backward. Why allow any namespace to muck with /proc/sys/ seettings?

There are many per namespace sysctls. Most of them are in the networking stack.

Quoted 11 lines

Let's recall that this case was a corner case, and writeback was the biggest concern, and for that it was decided that you'd simply not get write access, and so its read only. Its not clear to me if things like proc were considered. For the regular file case the situation can be addressed with chown, however we can't chown proc files.

Quoted 3 lines

Tested: Used a repro program that creates a user namespace without any mapping and stat'ed /proc/$PID/root/proc/sys/kernel/shmmax from outside. Before the change, it shows uid/gid of 65534,

I thought you said it would be uid/gid -1 without your patch?

It is INVALID_UID/INVALID_GID. It is an over simplifcation to call
them -1. As they are not a valid value and are never mapped in any
user namespace they are displayed as the overflow_uid or overflow_gid which is 65534 by default.

Quoted 6 lines
Quoted 1 line

with the change it's 0.

Note that a good way to also test issues is with the lib/test_sysctl.c module and the tools/testing/selftests/sysctl/sysctl.sh script, so if you can device a test there, once we decide what to do that would be appreciated.

We spoke about this at LPC. And this is the correct behavioral change.

The problem is there is a default value for i_uid and i_gid that is correct in the general case. That default value is not corect for sysctl, because proc is weird. As the sysctl permission check in test_perm are all against GLOBAL_ROOT_UID and GLOBAL_ROOT_GID we did not notice that i_uid and i_gid were being set wrong.

So all this patch does is fix the default values i_uid and i_gid.

The commit comment seems worth cleaning up. But for the content of the code.

I expect when I have a few moments I will pick this change up.

Eric

Quoted 23 lines
Quoted 23 lines
---
 fs/proc/proc_sysctl.c | 4 ++++
 1 file changed, 4 insertions(+)
fs/proc/proc_sysctl.c +4 -0
index c5cbbdff3c3d..67379a389658 100644
@@ -499,6 +499,10 @@ static struct inode *proc_sys_make_inode(struct super_block *sb,
 
 	if (root->set_ownership)
 		root->set_ownership(head, table, &inode->i_uid, &inode->i_gid);
+	else {
+		inode->i_uid = GLOBAL_ROOT_UID;
+		inode->i_gid = GLOBAL_ROOT_GID;
+	}
 
 out:
 	return inode;
Signature · 1 line
2.20.0.rc0.387.gc7a69e6b6c-goog
#3 lore raw
Quoted 8 lines · Eric W. Biederman
Quoted 4 lines · Luis Chamberlain

Thanks for the description of how to run into the issue described but is there also a practical use case today where this is happening? I ask as it would be good to know the severity of the issue in the real world today.

People trying to run containers without a root user in the container. It atypical but something doable.

My question was if there are generic tools / propreitary tools which are doing this widely *today*. Or is this just a custom setup some folks use?

Quoted 8 lines

We spoke about this at LPC. And this is the correct behavioral change.

The problem is there is a default value for i_uid and i_gid that is correct in the general case. That default value is not corect for sysctl, because proc is weird. As the sysctl permission check in test_perm are all against GLOBAL_ROOT_UID and GLOBAL_ROOT_GID we did not notice that i_uid and i_gid were being set wrong.

Quoted 4 lines

So all this patch does is fix the default values i_uid and i_gid.

The commit comment seems worth cleaning up. But for the content of the code.

The logic seems sensible then, but are we implicating what a container does with its sysctl values onto the entire system? If so, sure, it seems you want this for networking purposes as there are a series of sysctl values a container may want to muck with, but are we sure we want the same for *all* sysctl entries?

Luis

#4 lore raw
Quoted 8 lines · Luis Chamberlain
Quoted 6 lines · Eric W. Biederman
Quoted 4 lines · Luis Chamberlain

Thanks for the description of how to run into the issue described but is there also a practical use case today where this is happening? I ask as it would be good to know the severity of the issue in the real world today.

Quoted 6 lines
Quoted 2 lines

People trying to run containers without a root user in the container. It atypical but something doable.

My question was if there are generic tools / propreitary tools which are doing this widely *today*. Or is this just a custom setup some folks use?

We will soon start using this setup at Google to harden our usage of CRIU. There are some more details in my LPC presentation:
https://linuxplumbersconf.org/event/2/contributions/210/

Although I don't know of specific tools using this setup, there was a kernel patch in 2017 to support such use case:
7c6d78148fa0 - prctl: Allow local CAP_SYS_ADMIN changing exe_file So, perhaps Virtuozzo people use a similar setup too?

Quoted 10 lines
Quoted 10 lines

We spoke about this at LPC. And this is the correct behavioral change.

The problem is there is a default value for i_uid and i_gid that is correct in the general case. That default value is not corect for sysctl, because proc is weird. As the sysctl permission check in test_perm are all against GLOBAL_ROOT_UID and GLOBAL_ROOT_GID we did not notice that i_uid and i_gid were being set wrong.

So all this patch does is fix the default values i_uid and i_gid.

Quoted 8 lines
Quoted 2 lines

The commit comment seems worth cleaning up. But for the content of the code.

The logic seems sensible then, but are we implicating what a container does with its sysctl values onto the entire system? If so, sure, it seems you want this for networking purposes as there are a series of sysctl values a container may want to muck with, but are we sure we want the same for *all* sysctl entries?

The point is that these sysctls do not affect the whole system, just an appropriate namespace.
For example, IPC-related files (e.g. shmmax) will always affect writing process's UTS namespace, regardless of /proc mountpoint that is used to access them:
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/ipc/ipc_sysctl.c?h=v4.20-rc4#n24

I presume the net-related sysctls that Eric was referring to have a similar behavior.

Quoted 2 lines

Luis

#5 lore raw
In reply to Luis Chamberlain
Quoted 24 lines · Luis Chamberlain
Quoted 8 lines · Eric W. Biederman
Quoted 4 lines · Luis Chamberlain

Thanks for the description of how to run into the issue described but is there also a practical use case today where this is happening? I ask as it would be good to know the severity of the issue in the real world today.

People trying to run containers without a root user in the container. It atypical but something doable.

My question was if there are generic tools / propreitary tools which are doing this widely *today*. Or is this just a custom setup some folks use?

Quoted 10 lines

We spoke about this at LPC. And this is the correct behavioral change.

The problem is there is a default value for i_uid and i_gid that is correct in the general case. That default value is not corect for sysctl, because proc is weird. As the sysctl permission check in test_perm are all against GLOBAL_ROOT_UID and GLOBAL_ROOT_GID we did not notice that i_uid and i_gid were being set wrong.

So all this patch does is fix the default values i_uid and i_gid.

Quoted 8 lines
Quoted 2 lines

The commit comment seems worth cleaning up. But for the content of the code.

The logic seems sensible then, but are we implicating what a container does with its sysctl values onto the entire system? If so, sure, it seems you want this for networking purposes as there are a series of sysctl values a container may want to muck with, but are we sure we want the same for *all* sysctl entries?

No. Please look at the patch again. It sets the default uid and gid for sysctl entries to 0. AKA GLOBAL_ROOT_UID and GLOBAL_ROOT_GID because there is a bug and they were not set to that value.

Those are the uids and gids that are tested agasint. It just happens you have to be in a weird configuration for this bug to become a problem.

Eric

#6 lore raw
Quoted 8 lines · Eric W. Biederman
Quoted 5 lines · Luis Chamberlain

The logic seems sensible then, but are we implicating what a container does with its sysctl values onto the entire system? If so, sure, it seems you want this for networking purposes as there are a series of sysctl values a container may want to muck with, but are we sure we want the same for *all* sysctl entries?

Quoted 6 lines

No. Please look at the patch again. It sets the default uid and gid for sysctl entries to 0. AKA GLOBAL_ROOT_UID and GLOBAL_ROOT_GID because there is a bug and they were not set to that value.

Those are the uids and gids that are tested agasint. It just happens you have to be in a weird configuration for this bug to become a problem.

Thanks, then provided the commit lot is modified:

Luis

A message that is not in the archive (sent off-list or not yet received): CAFkxGoM_rjciQ0sRh7Lhf_XfJu-g4Tth6Yo0L_YRVUaOnzjZuA@mail.gmail.com

#8 lore raw
Quoted 81 lines · Radoslaw Burny

On Tue, Nov 27, 2018 at 6:29 AM Eric W. Biederman <[email protected]> wrote:

Quoted 10 lines · Luis Chamberlain
Quoted 2 lines · Radoslaw Burny

Due to a recent commit (d151ddc00498 - fs: Update i_[ug]id_(read|write) to translate relative to s_user_ns),

Recent? This is commit is from 2014 and present upstream since v4.8. And the commit ID you mentioned in your commit log seems to be incorrect. I get:

81754357770ebd900801231e7bc8d151ddc00498a fs: Update i_[ug]id_(read|write) to translate relative to s_user_ns

Quoted 8 lines
Quoted 3 lines

inodes under /proc/sys have -1
written to their i_uid/i_gid members if a containing userns does not have entries for root in the uid/gid_map.

Thanks for the description of how to run into the issue described but is there also a practical use case today where this is happening? I ask as it would be good to know the severity of the issue in the real world today.

People trying to run containers without a root user in the container. It atypical but something doable.

Quoted 6 lines
Quoted 6 lines

This wouldn't normally matter, because these values are not used for access checks. However, a later change (0bd23d09b874 - Don't modify inodes with a uid or gid unknown to the vfs) changes the kernel to prevent opens for write if the i_uid/i_gid field in the inode is -1, even if the /proc/sys-specific access checks would otherwise pass.

Quoted 8 lines
Quoted 4 lines

This causes a problem: in a userns without root mapping, even the namespace creator cannot write to e.g. /proc/sys/kernel/shmmax. This change fixes the problem by overriding i_uid/i_gid back to GLOBAL_ROOT_UID/GID.

We really need Seth and Eric to provide guidance here as they were the ones devising this long ago, but to me your solution seems backward. Why allow any namespace to muck with /proc/sys/ seettings?

There are many per namespace sysctls. Most of them are in the networking stack.

Quoted 11 lines

Let's recall that this case was a corner case, and writeback was the biggest concern, and for that it was decided that you'd simply not get write access, and so its read only. Its not clear to me if things like proc were considered. For the regular file case the situation can be addressed with chown, however we can't chown proc files.

Quoted 3 lines

Tested: Used a repro program that creates a user namespace without any mapping and stat'ed /proc/$PID/root/proc/sys/kernel/shmmax from outside. Before the change, it shows uid/gid of 65534,

I thought you said it would be uid/gid -1 without your patch?

It is INVALID_UID/INVALID_GID. It is an over simplifcation to call them -1. As they are not a valid value and are never mapped in any user namespace they are displayed as the overflow_uid or overflow_gid which is 65534 by default.

Quoted 6 lines
Quoted 1 line

with the change it's 0.

Note that a good way to also test issues is with the lib/test_sysctl.c module and the tools/testing/selftests/sysctl/sysctl.sh script, so if you can device a test there, once we decide what to do that would be appreciated.

We spoke about this at LPC. And this is the correct behavioral change.

The problem is there is a default value for i_uid and i_gid that is correct in the general case. That default value is not corect for sysctl, because proc is weird. As the sysctl permission check in test_perm are all against GLOBAL_ROOT_UID and GLOBAL_ROOT_GID we did not notice that i_uid and i_gid were being set wrong.

So all this patch does is fix the default values i_uid and i_gid.

The commit comment seems worth cleaning up. But for the content of the code.

Quoted 8 lines

I expect when I have a few moments I will pick this change up.

Eric

Thanks, Eric. Should I send a v2 patch with an updated description, or can you just modify the description when applying this one?

I am absolutely swampped and moving at the moment. Can you please send a v2 with an updated description.

Thank you,
Eric

Quoted 24 lines
Quoted 24 lines
Quoted 24 lines

---
fs/proc/proc_sysctl.c | 4 ++++

1 file changed, 4 insertions(+)
fs/proc/proc_sysctl.c +4 -0
index c5cbbdff3c3d..67379a389658 100644
@@ -499,6 +499,10 @@ static struct inode *proc_sys_make_inode(struct super_block *sb,
 
 if (root->set_ownership)
 root->set_ownership(head, table, &inode->i_uid, &inode->i_gid);
+ else {
+ inode->i_uid = GLOBAL_ROOT_UID;
+ inode->i_gid = GLOBAL_ROOT_GID;
+ }
 

out:
return inode;

Signature · 1 line
2.20.0.rc0.387.gc7a69e6b6c-goog