2019-06-01 11:08:55 +03:00
// SPDX-License-Identifier: GPL-2.0-only
2006-10-02 13:18:06 +04:00
/*
* Copyright ( C ) 2006 IBM Corporation
*
* Author : Serge Hallyn < serue @ us . ibm . com >
*
2006-10-02 13:18:19 +04:00
* Jun 2006 - namespaces support
* OpenVZ , SWsoft Inc .
* Pavel Emelianov < xemul @ openvz . org >
2006-10-02 13:18:06 +04:00
*/
include cleanup: Update gfp.h and slab.h includes to prepare for breaking implicit slab.h inclusion from percpu.h
percpu.h is included by sched.h and module.h and thus ends up being
included when building most .c files. percpu.h includes slab.h which
in turn includes gfp.h making everything defined by the two files
universally available and complicating inclusion dependencies.
percpu.h -> slab.h dependency is about to be removed. Prepare for
this change by updating users of gfp and slab facilities include those
headers directly instead of assuming availability. As this conversion
needs to touch large number of source files, the following script is
used as the basis of conversion.
http://userweb.kernel.org/~tj/misc/slabh-sweep.py
The script does the followings.
* Scan files for gfp and slab usages and update includes such that
only the necessary includes are there. ie. if only gfp is used,
gfp.h, if slab is used, slab.h.
* When the script inserts a new include, it looks at the include
blocks and try to put the new include such that its order conforms
to its surrounding. It's put in the include block which contains
core kernel includes, in the same order that the rest are ordered -
alphabetical, Christmas tree, rev-Xmas-tree or at the end if there
doesn't seem to be any matching order.
* If the script can't find a place to put a new include (mostly
because the file doesn't have fitting include block), it prints out
an error message indicating which .h file needs to be added to the
file.
The conversion was done in the following steps.
1. The initial automatic conversion of all .c files updated slightly
over 4000 files, deleting around 700 includes and adding ~480 gfp.h
and ~3000 slab.h inclusions. The script emitted errors for ~400
files.
2. Each error was manually checked. Some didn't need the inclusion,
some needed manual addition while adding it to implementation .h or
embedding .c file was more appropriate for others. This step added
inclusions to around 150 files.
3. The script was run again and the output was compared to the edits
from #2 to make sure no file was left behind.
4. Several build tests were done and a couple of problems were fixed.
e.g. lib/decompress_*.c used malloc/free() wrappers around slab
APIs requiring slab.h to be added manually.
5. The script was run on all .h files but without automatically
editing them as sprinkling gfp.h and slab.h inclusions around .h
files could easily lead to inclusion dependency hell. Most gfp.h
inclusion directives were ignored as stuff from gfp.h was usually
wildly available and often used in preprocessor macros. Each
slab.h inclusion directive was examined and added manually as
necessary.
6. percpu.h was updated not to include slab.h.
7. Build test were done on the following configurations and failures
were fixed. CONFIG_GCOV_KERNEL was turned off for all tests (as my
distributed build env didn't work with gcov compiles) and a few
more options had to be turned off depending on archs to make things
build (like ipr on powerpc/64 which failed due to missing writeq).
* x86 and x86_64 UP and SMP allmodconfig and a custom test config.
* powerpc and powerpc64 SMP allmodconfig
* sparc and sparc64 SMP allmodconfig
* ia64 SMP allmodconfig
* s390 SMP allmodconfig
* alpha SMP allmodconfig
* um on x86_64 SMP allmodconfig
8. percpu.h modifications were reverted so that it could be applied as
a separate patch and serve as bisection point.
Given the fact that I had only a couple of failures from tests on step
6, I'm fairly confident about the coverage of this conversion patch.
If there is a breakage, it's likely to be something in one of the arch
headers which should be easily discoverable easily on most builds of
the specific arch.
Signed-off-by: Tejun Heo <tj@kernel.org>
Guess-its-ok-by: Christoph Lameter <cl@linux-foundation.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
2010-03-24 11:04:11 +03:00
# include <linux/slab.h>
2011-05-23 22:51:41 +04:00
# include <linux/export.h>
2006-10-02 13:18:06 +04:00
# include <linux/nsproxy.h>
2006-10-02 13:18:07 +04:00
# include <linux/init_task.h>
2006-12-08 13:37:56 +03:00
# include <linux/mnt_namespace.h>
2006-10-02 13:18:14 +04:00
# include <linux/utsname.h>
2006-12-08 13:37:59 +03:00
# include <linux/pid_namespace.h>
2007-09-27 09:04:26 +04:00
# include <net/net_namespace.h>
2008-02-08 15:18:22 +03:00
# include <linux/ipc_namespace.h>
ns: Introduce Time Namespace
Time Namespace isolates clock values.
The kernel provides access to several clocks CLOCK_REALTIME,
CLOCK_MONOTONIC, CLOCK_BOOTTIME, etc.
CLOCK_REALTIME
System-wide clock that measures real (i.e., wall-clock) time.
CLOCK_MONOTONIC
Clock that cannot be set and represents monotonic time since
some unspecified starting point.
CLOCK_BOOTTIME
Identical to CLOCK_MONOTONIC, except it also includes any time
that the system is suspended.
For many users, the time namespace means the ability to changes date and
time in a container (CLOCK_REALTIME). Providing per namespace notions of
CLOCK_REALTIME would be complex with a massive overhead, but has a dubious
value.
But in the context of checkpoint/restore functionality, monotonic and
boottime clocks become interesting. Both clocks are monotonic with
unspecified starting points. These clocks are widely used to measure time
slices and set timers. After restoring or migrating processes, it has to be
guaranteed that they never go backward. In an ideal case, the behavior of
these clocks should be the same as for a case when a whole system is
suspended. All this means that it is required to set CLOCK_MONOTONIC and
CLOCK_BOOTTIME clocks, which can be achieved by adding per-namespace
offsets for clocks.
A time namespace is similar to a pid namespace in the way how it is
created: unshare(CLONE_NEWTIME) system call creates a new time namespace,
but doesn't set it to the current process. Then all children of the process
will be born in the new time namespace, or a process can use the setns()
system call to join a namespace.
This scheme allows setting clock offsets for a namespace, before any
processes appear in it.
All available clone flags have been used, so CLONE_NEWTIME uses the highest
bit of CSIGNAL. It means that it can be used only with the unshare() and
the clone3() system calls.
[ tglx: Adjusted paragraph about clone3() to reality and massaged the
changelog a bit. ]
Co-developed-by: Dmitry Safonov <dima@arista.com>
Signed-off-by: Andrei Vagin <avagin@gmail.com>
Signed-off-by: Dmitry Safonov <dima@arista.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Link: https://criu.org/Time_namespace
Link: https://lists.openvz.org/pipermail/criu/2018-June/041504.html
Link: https://lore.kernel.org/r/20191112012724.250792-4-dima@arista.com
2019-11-12 04:26:52 +03:00
# include <linux/time_namespace.h>
2020-05-05 17:04:30 +03:00
# include <linux/fs_struct.h>
nsproxy: attach to namespaces via pidfds
For quite a while we have been thinking about using pidfds to attach to
namespaces. This patchset has existed for about a year already but we've
wanted to wait to see how the general api would be received and adopted.
Now that more and more programs in userspace have started using pidfds
for process management it's time to send this one out.
This patch makes it possible to use pidfds to attach to the namespaces
of another process, i.e. they can be passed as the first argument to the
setns() syscall. When only a single namespace type is specified the
semantics are equivalent to passing an nsfd. That means
setns(nsfd, CLONE_NEWNET) equals setns(pidfd, CLONE_NEWNET). However,
when a pidfd is passed, multiple namespace flags can be specified in the
second setns() argument and setns() will attach the caller to all the
specified namespaces all at once or to none of them. Specifying 0 is not
valid together with a pidfd.
Here are just two obvious examples:
setns(pidfd, CLONE_NEWPID | CLONE_NEWNS | CLONE_NEWNET);
setns(pidfd, CLONE_NEWUSER);
Allowing to also attach subsets of namespaces supports various use-cases
where callers setns to a subset of namespaces to retain privilege, perform
an action and then re-attach another subset of namespaces.
If the need arises, as Eric suggested, we can extend this patchset to
assume even more context than just attaching all namespaces. His suggestion
specifically was about assuming the process' root directory when
setns(pidfd, 0) or setns(pidfd, SETNS_PIDFD) is specified. For now, just
keep it flexible in terms of supporting subsets of namespaces but let's
wait until we have users asking for even more context to be assumed. At
that point we can add an extension.
The obvious example where this is useful is a standard container
manager interacting with a running container: pushing and pulling files
or directories, injecting mounts, attaching/execing any kind of process,
managing network devices all these operations require attaching to all
or at least multiple namespaces at the same time. Given that nowadays
most containers are spawned with all namespaces enabled we're currently
looking at at least 14 syscalls, 7 to open the /proc/<pid>/ns/<ns>
nsfds, another 7 to actually perform the namespace switch. With time
namespaces we're looking at about 16 syscalls.
(We could amortize the first 7 or 8 syscalls for opening the nsfds by
stashing them in each container's monitor process but that would mean
we need to send around those file descriptors through unix sockets
everytime we want to interact with the container or keep on-disk
state. Even in scenarios where a caller wants to join a particular
namespace in a particular order callers still profit from batching
other namespaces. That mostly applies to the user namespace but
all container runtimes I found join the user namespace first no matter
if it privileges or deprivileges the container similar to how unshare
behaves.)
With pidfds this becomes a single syscall no matter how many namespaces
are supposed to be attached to.
A decently designed, large-scale container manager usually isn't the
parent of any of the containers it spawns so the containers don't die
when it crashes or needs to update or reinitialize. This means that
for the manager to interact with containers through pids is inherently
racy especially on systems where the maximum pid number is not
significicantly bumped. This is even more problematic since we often spawn
and manage thousands or ten-thousands of containers. Interacting with a
container through a pid thus can become risky quite quickly. Especially
since we allow for an administrator to enable advanced features such as
syscall interception where we're performing syscalls in lieu of the
container. In all of those cases we use pidfds if they are available and
we pass them around as stable references. Using them to setns() to the
target process' namespaces is as reliable as using nsfds. Either the
target process is already dead and we get ESRCH or we manage to attach
to its namespaces but we can't accidently attach to another process'
namespaces. So pidfds lend themselves to be used with this api.
The other main advantage is that with this change the pidfd becomes the
only relevant token for most container interactions and it's the only
token we need to create and send around.
Apart from significiantly reducing the number of syscalls from double
digit to single digit which is a decent reason post-spectre/meltdown
this also allows to switch to a set of namespaces atomically, i.e.
either attaching to all the specified namespaces succeeds or we fail. If
we fail we haven't changed a single namespace. There are currently three
namespaces that can fail (other than for ENOMEM which really is not
very interesting since we then have other problems anyway) for
non-trivial reasons, user, mount, and pid namespaces. We can fail to
attach to a pid namespace if it is not our current active pid namespace
or a descendant of it. We can fail to attach to a user namespace because
we are multi-threaded or because our current mount namespace shares
filesystem state with other tasks, or because we're trying to setns()
to the same user namespace, i.e. the target task has the same user
namespace as we do. We can fail to attach to a mount namespace because
it shares filesystem state with other tasks or because we fail to lookup
the new root for the new mount namespace. In most non-pathological
scenarios these issues can be somewhat mitigated. But there are cases where
we're half-attached to some namespace and failing to attach to another one.
I've talked about some of these problem during the hallway track (something
only the pre-COVID-19 generation will remember) of Plumbers in Los Angeles
in 2018(?). Even if all these issues could be avoided with super careful
userspace coding it would be nicer to have this done in-kernel. Pidfds seem
to lend themselves nicely for this.
The other neat thing about this is that setns() becomes an actual
counterpart to the namespace bits of unshare().
Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
Reviewed-by: Serge Hallyn <serge@hallyn.com>
Cc: Eric W. Biederman <ebiederm@xmission.com>
Cc: Serge Hallyn <serge@hallyn.com>
Cc: Jann Horn <jannh@google.com>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Aleksa Sarai <cyphar@cyphar.com>
Link: https://lore.kernel.org/r/20200505140432.181565-3-christian.brauner@ubuntu.com
2020-05-05 17:04:31 +03:00
# include <linux/proc_fs.h>
2013-04-12 04:50:06 +04:00
# include <linux/proc_ns.h>
2010-03-08 04:48:52 +03:00
# include <linux/file.h>
# include <linux/syscalls.h>
2016-01-29 11:54:06 +03:00
# include <linux/cgroup.h>
2017-03-07 23:41:36 +03:00
# include <linux/perf_event.h>
2006-10-02 13:18:07 +04:00
2007-07-16 10:41:07 +04:00
static struct kmem_cache * nsproxy_cachep ;
2010-03-11 02:23:10 +03:00
struct nsproxy init_nsproxy = {
2013-08-22 22:39:16 +04:00
. count = ATOMIC_INIT ( 1 ) ,
. uts_ns = & init_uts_ns ,
2010-03-11 02:23:10 +03:00
# if defined(CONFIG_POSIX_MQUEUE) || defined(CONFIG_SYSVIPC)
2013-08-22 22:39:16 +04:00
. ipc_ns = & init_ipc_ns ,
2010-03-11 02:23:10 +03:00
# endif
2013-08-22 22:39:16 +04:00
. mnt_ns = NULL ,
. pid_ns_for_children = & init_pid_ns ,
2010-03-11 02:23:10 +03:00
# ifdef CONFIG_NET
2013-08-22 22:39:16 +04:00
. net_ns = & init_net ,
2010-03-11 02:23:10 +03:00
# endif
2016-01-29 11:54:06 +03:00
# ifdef CONFIG_CGROUPS
. cgroup_ns = & init_cgroup_ns ,
# endif
ns: Introduce Time Namespace
Time Namespace isolates clock values.
The kernel provides access to several clocks CLOCK_REALTIME,
CLOCK_MONOTONIC, CLOCK_BOOTTIME, etc.
CLOCK_REALTIME
System-wide clock that measures real (i.e., wall-clock) time.
CLOCK_MONOTONIC
Clock that cannot be set and represents monotonic time since
some unspecified starting point.
CLOCK_BOOTTIME
Identical to CLOCK_MONOTONIC, except it also includes any time
that the system is suspended.
For many users, the time namespace means the ability to changes date and
time in a container (CLOCK_REALTIME). Providing per namespace notions of
CLOCK_REALTIME would be complex with a massive overhead, but has a dubious
value.
But in the context of checkpoint/restore functionality, monotonic and
boottime clocks become interesting. Both clocks are monotonic with
unspecified starting points. These clocks are widely used to measure time
slices and set timers. After restoring or migrating processes, it has to be
guaranteed that they never go backward. In an ideal case, the behavior of
these clocks should be the same as for a case when a whole system is
suspended. All this means that it is required to set CLOCK_MONOTONIC and
CLOCK_BOOTTIME clocks, which can be achieved by adding per-namespace
offsets for clocks.
A time namespace is similar to a pid namespace in the way how it is
created: unshare(CLONE_NEWTIME) system call creates a new time namespace,
but doesn't set it to the current process. Then all children of the process
will be born in the new time namespace, or a process can use the setns()
system call to join a namespace.
This scheme allows setting clock offsets for a namespace, before any
processes appear in it.
All available clone flags have been used, so CLONE_NEWTIME uses the highest
bit of CSIGNAL. It means that it can be used only with the unshare() and
the clone3() system calls.
[ tglx: Adjusted paragraph about clone3() to reality and massaged the
changelog a bit. ]
Co-developed-by: Dmitry Safonov <dima@arista.com>
Signed-off-by: Andrei Vagin <avagin@gmail.com>
Signed-off-by: Dmitry Safonov <dima@arista.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Link: https://criu.org/Time_namespace
Link: https://lists.openvz.org/pipermail/criu/2018-June/041504.html
Link: https://lore.kernel.org/r/20191112012724.250792-4-dima@arista.com
2019-11-12 04:26:52 +03:00
# ifdef CONFIG_TIME_NS
. time_ns = & init_time_ns ,
. time_ns_for_children = & init_time_ns ,
# endif
2010-03-11 02:23:10 +03:00
} ;
2006-10-02 13:18:06 +04:00
2009-06-18 03:27:56 +04:00
static inline struct nsproxy * create_nsproxy ( void )
2006-10-02 13:18:06 +04:00
{
2009-06-18 03:27:56 +04:00
struct nsproxy * nsproxy ;
2006-10-02 13:18:06 +04:00
2009-06-18 03:27:56 +04:00
nsproxy = kmem_cache_alloc ( nsproxy_cachep , GFP_KERNEL ) ;
if ( nsproxy )
atomic_set ( & nsproxy - > count , 1 ) ;
return nsproxy ;
2006-10-02 13:18:06 +04:00
}
/*
2007-05-08 11:25:21 +04:00
* Create new nsproxy and all of its the associated namespaces .
* Return the newly created nsproxy . Do not attach this to the task ,
* leave it to the caller to do proper locking and attach it to task .
2006-10-02 13:18:06 +04:00
*/
2007-07-16 10:41:15 +04:00
static struct nsproxy * create_new_namespaces ( unsigned long flags ,
2012-07-26 15:02:49 +04:00
struct task_struct * tsk , struct user_namespace * user_ns ,
struct fs_struct * new_fs )
2006-10-02 13:18:06 +04:00
{
2007-05-08 11:25:21 +04:00
struct nsproxy * new_nsp ;
2007-07-16 10:41:06 +04:00
int err ;
2006-10-02 13:18:06 +04:00
2009-06-18 03:27:56 +04:00
new_nsp = create_nsproxy ( ) ;
2007-05-08 11:25:21 +04:00
if ( ! new_nsp )
return ERR_PTR ( - ENOMEM ) ;
2006-10-02 13:18:08 +04:00
2012-07-26 15:02:49 +04:00
new_nsp - > mnt_ns = copy_mnt_ns ( flags , tsk - > nsproxy - > mnt_ns , user_ns , new_fs ) ;
2007-07-16 10:41:06 +04:00
if ( IS_ERR ( new_nsp - > mnt_ns ) ) {
err = PTR_ERR ( new_nsp - > mnt_ns ) ;
2007-05-08 11:25:21 +04:00
goto out_ns ;
2007-07-16 10:41:06 +04:00
}
2007-05-08 11:25:21 +04:00
2012-07-26 15:02:49 +04:00
new_nsp - > uts_ns = copy_utsname ( flags , user_ns , tsk - > nsproxy - > uts_ns ) ;
2007-07-16 10:41:06 +04:00
if ( IS_ERR ( new_nsp - > uts_ns ) ) {
err = PTR_ERR ( new_nsp - > uts_ns ) ;
2007-05-08 11:25:21 +04:00
goto out_uts ;
2007-07-16 10:41:06 +04:00
}
2007-05-08 11:25:21 +04:00
2012-07-26 15:02:49 +04:00
new_nsp - > ipc_ns = copy_ipcs ( flags , user_ns , tsk - > nsproxy - > ipc_ns ) ;
2007-07-16 10:41:06 +04:00
if ( IS_ERR ( new_nsp - > ipc_ns ) ) {
err = PTR_ERR ( new_nsp - > ipc_ns ) ;
2007-05-08 11:25:21 +04:00
goto out_ipc ;
2007-07-16 10:41:06 +04:00
}
2007-05-08 11:25:21 +04:00
2013-08-22 22:39:16 +04:00
new_nsp - > pid_ns_for_children =
copy_pid_ns ( flags , user_ns , tsk - > nsproxy - > pid_ns_for_children ) ;
if ( IS_ERR ( new_nsp - > pid_ns_for_children ) ) {
err = PTR_ERR ( new_nsp - > pid_ns_for_children ) ;
2007-05-08 11:25:21 +04:00
goto out_pid ;
2007-07-16 10:41:06 +04:00
}
2007-05-08 11:25:21 +04:00
2016-01-29 11:54:06 +03:00
new_nsp - > cgroup_ns = copy_cgroup_ns ( flags , user_ns ,
tsk - > nsproxy - > cgroup_ns ) ;
if ( IS_ERR ( new_nsp - > cgroup_ns ) ) {
err = PTR_ERR ( new_nsp - > cgroup_ns ) ;
goto out_cgroup ;
}
2012-07-26 15:02:49 +04:00
new_nsp - > net_ns = copy_net_ns ( flags , user_ns , tsk - > nsproxy - > net_ns ) ;
2007-09-27 09:04:26 +04:00
if ( IS_ERR ( new_nsp - > net_ns ) ) {
err = PTR_ERR ( new_nsp - > net_ns ) ;
goto out_net ;
}
ns: Introduce Time Namespace
Time Namespace isolates clock values.
The kernel provides access to several clocks CLOCK_REALTIME,
CLOCK_MONOTONIC, CLOCK_BOOTTIME, etc.
CLOCK_REALTIME
System-wide clock that measures real (i.e., wall-clock) time.
CLOCK_MONOTONIC
Clock that cannot be set and represents monotonic time since
some unspecified starting point.
CLOCK_BOOTTIME
Identical to CLOCK_MONOTONIC, except it also includes any time
that the system is suspended.
For many users, the time namespace means the ability to changes date and
time in a container (CLOCK_REALTIME). Providing per namespace notions of
CLOCK_REALTIME would be complex with a massive overhead, but has a dubious
value.
But in the context of checkpoint/restore functionality, monotonic and
boottime clocks become interesting. Both clocks are monotonic with
unspecified starting points. These clocks are widely used to measure time
slices and set timers. After restoring or migrating processes, it has to be
guaranteed that they never go backward. In an ideal case, the behavior of
these clocks should be the same as for a case when a whole system is
suspended. All this means that it is required to set CLOCK_MONOTONIC and
CLOCK_BOOTTIME clocks, which can be achieved by adding per-namespace
offsets for clocks.
A time namespace is similar to a pid namespace in the way how it is
created: unshare(CLONE_NEWTIME) system call creates a new time namespace,
but doesn't set it to the current process. Then all children of the process
will be born in the new time namespace, or a process can use the setns()
system call to join a namespace.
This scheme allows setting clock offsets for a namespace, before any
processes appear in it.
All available clone flags have been used, so CLONE_NEWTIME uses the highest
bit of CSIGNAL. It means that it can be used only with the unshare() and
the clone3() system calls.
[ tglx: Adjusted paragraph about clone3() to reality and massaged the
changelog a bit. ]
Co-developed-by: Dmitry Safonov <dima@arista.com>
Signed-off-by: Andrei Vagin <avagin@gmail.com>
Signed-off-by: Dmitry Safonov <dima@arista.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Link: https://criu.org/Time_namespace
Link: https://lists.openvz.org/pipermail/criu/2018-June/041504.html
Link: https://lore.kernel.org/r/20191112012724.250792-4-dima@arista.com
2019-11-12 04:26:52 +03:00
new_nsp - > time_ns_for_children = copy_time_ns ( flags , user_ns ,
tsk - > nsproxy - > time_ns_for_children ) ;
if ( IS_ERR ( new_nsp - > time_ns_for_children ) ) {
err = PTR_ERR ( new_nsp - > time_ns_for_children ) ;
goto out_time ;
}
new_nsp - > time_ns = get_time_ns ( tsk - > nsproxy - > time_ns ) ;
2007-05-08 11:25:21 +04:00
return new_nsp ;
ns: Introduce Time Namespace
Time Namespace isolates clock values.
The kernel provides access to several clocks CLOCK_REALTIME,
CLOCK_MONOTONIC, CLOCK_BOOTTIME, etc.
CLOCK_REALTIME
System-wide clock that measures real (i.e., wall-clock) time.
CLOCK_MONOTONIC
Clock that cannot be set and represents monotonic time since
some unspecified starting point.
CLOCK_BOOTTIME
Identical to CLOCK_MONOTONIC, except it also includes any time
that the system is suspended.
For many users, the time namespace means the ability to changes date and
time in a container (CLOCK_REALTIME). Providing per namespace notions of
CLOCK_REALTIME would be complex with a massive overhead, but has a dubious
value.
But in the context of checkpoint/restore functionality, monotonic and
boottime clocks become interesting. Both clocks are monotonic with
unspecified starting points. These clocks are widely used to measure time
slices and set timers. After restoring or migrating processes, it has to be
guaranteed that they never go backward. In an ideal case, the behavior of
these clocks should be the same as for a case when a whole system is
suspended. All this means that it is required to set CLOCK_MONOTONIC and
CLOCK_BOOTTIME clocks, which can be achieved by adding per-namespace
offsets for clocks.
A time namespace is similar to a pid namespace in the way how it is
created: unshare(CLONE_NEWTIME) system call creates a new time namespace,
but doesn't set it to the current process. Then all children of the process
will be born in the new time namespace, or a process can use the setns()
system call to join a namespace.
This scheme allows setting clock offsets for a namespace, before any
processes appear in it.
All available clone flags have been used, so CLONE_NEWTIME uses the highest
bit of CSIGNAL. It means that it can be used only with the unshare() and
the clone3() system calls.
[ tglx: Adjusted paragraph about clone3() to reality and massaged the
changelog a bit. ]
Co-developed-by: Dmitry Safonov <dima@arista.com>
Signed-off-by: Andrei Vagin <avagin@gmail.com>
Signed-off-by: Dmitry Safonov <dima@arista.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Link: https://criu.org/Time_namespace
Link: https://lists.openvz.org/pipermail/criu/2018-June/041504.html
Link: https://lore.kernel.org/r/20191112012724.250792-4-dima@arista.com
2019-11-12 04:26:52 +03:00
out_time :
put_net ( new_nsp - > net_ns ) ;
2007-09-27 09:04:26 +04:00
out_net :
2016-01-29 11:54:06 +03:00
put_cgroup_ns ( new_nsp - > cgroup_ns ) ;
out_cgroup :
2013-08-22 22:39:16 +04:00
if ( new_nsp - > pid_ns_for_children )
put_pid_ns ( new_nsp - > pid_ns_for_children ) ;
2007-05-08 11:25:21 +04:00
out_pid :
if ( new_nsp - > ipc_ns )
put_ipc_ns ( new_nsp - > ipc_ns ) ;
out_ipc :
if ( new_nsp - > uts_ns )
put_uts_ns ( new_nsp - > uts_ns ) ;
out_uts :
if ( new_nsp - > mnt_ns )
put_mnt_ns ( new_nsp - > mnt_ns ) ;
out_ns :
2007-07-16 10:41:07 +04:00
kmem_cache_free ( nsproxy_cachep , new_nsp ) ;
2007-07-16 10:41:06 +04:00
return ERR_PTR ( err ) ;
2006-10-02 13:18:06 +04:00
}
/*
* called from clone . This now handles copy for nsproxy and all
* namespaces therein .
*/
2007-07-16 10:41:15 +04:00
int copy_namespaces ( unsigned long flags , struct task_struct * tsk )
2006-10-02 13:18:06 +04:00
{
struct nsproxy * old_ns = tsk - > nsproxy ;
2012-07-26 11:50:47 +04:00
struct user_namespace * user_ns = task_cred_xxx ( tsk , user_ns ) ;
2006-10-02 13:18:08 +04:00
struct nsproxy * new_ns ;
2006-10-02 13:18:06 +04:00
2013-03-10 04:15:23 +04:00
if ( likely ( ! ( flags & ( CLONE_NEWNS | CLONE_NEWUTS | CLONE_NEWIPC |
2016-01-29 11:54:06 +03:00
CLONE_NEWPID | CLONE_NEWNET |
ns: Introduce Time Namespace
Time Namespace isolates clock values.
The kernel provides access to several clocks CLOCK_REALTIME,
CLOCK_MONOTONIC, CLOCK_BOOTTIME, etc.
CLOCK_REALTIME
System-wide clock that measures real (i.e., wall-clock) time.
CLOCK_MONOTONIC
Clock that cannot be set and represents monotonic time since
some unspecified starting point.
CLOCK_BOOTTIME
Identical to CLOCK_MONOTONIC, except it also includes any time
that the system is suspended.
For many users, the time namespace means the ability to changes date and
time in a container (CLOCK_REALTIME). Providing per namespace notions of
CLOCK_REALTIME would be complex with a massive overhead, but has a dubious
value.
But in the context of checkpoint/restore functionality, monotonic and
boottime clocks become interesting. Both clocks are monotonic with
unspecified starting points. These clocks are widely used to measure time
slices and set timers. After restoring or migrating processes, it has to be
guaranteed that they never go backward. In an ideal case, the behavior of
these clocks should be the same as for a case when a whole system is
suspended. All this means that it is required to set CLOCK_MONOTONIC and
CLOCK_BOOTTIME clocks, which can be achieved by adding per-namespace
offsets for clocks.
A time namespace is similar to a pid namespace in the way how it is
created: unshare(CLONE_NEWTIME) system call creates a new time namespace,
but doesn't set it to the current process. Then all children of the process
will be born in the new time namespace, or a process can use the setns()
system call to join a namespace.
This scheme allows setting clock offsets for a namespace, before any
processes appear in it.
All available clone flags have been used, so CLONE_NEWTIME uses the highest
bit of CSIGNAL. It means that it can be used only with the unshare() and
the clone3() system calls.
[ tglx: Adjusted paragraph about clone3() to reality and massaged the
changelog a bit. ]
Co-developed-by: Dmitry Safonov <dima@arista.com>
Signed-off-by: Andrei Vagin <avagin@gmail.com>
Signed-off-by: Dmitry Safonov <dima@arista.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Link: https://criu.org/Time_namespace
Link: https://lists.openvz.org/pipermail/criu/2018-June/041504.html
Link: https://lore.kernel.org/r/20191112012724.250792-4-dima@arista.com
2019-11-12 04:26:52 +03:00
CLONE_NEWCGROUP | CLONE_NEWTIME ) ) ) ) {
2022-09-21 03:31:19 +03:00
if ( ( flags & CLONE_VM ) | |
likely ( old_ns - > time_ns_for_children = = old_ns - > time_ns ) ) {
ns: Introduce Time Namespace
Time Namespace isolates clock values.
The kernel provides access to several clocks CLOCK_REALTIME,
CLOCK_MONOTONIC, CLOCK_BOOTTIME, etc.
CLOCK_REALTIME
System-wide clock that measures real (i.e., wall-clock) time.
CLOCK_MONOTONIC
Clock that cannot be set and represents monotonic time since
some unspecified starting point.
CLOCK_BOOTTIME
Identical to CLOCK_MONOTONIC, except it also includes any time
that the system is suspended.
For many users, the time namespace means the ability to changes date and
time in a container (CLOCK_REALTIME). Providing per namespace notions of
CLOCK_REALTIME would be complex with a massive overhead, but has a dubious
value.
But in the context of checkpoint/restore functionality, monotonic and
boottime clocks become interesting. Both clocks are monotonic with
unspecified starting points. These clocks are widely used to measure time
slices and set timers. After restoring or migrating processes, it has to be
guaranteed that they never go backward. In an ideal case, the behavior of
these clocks should be the same as for a case when a whole system is
suspended. All this means that it is required to set CLOCK_MONOTONIC and
CLOCK_BOOTTIME clocks, which can be achieved by adding per-namespace
offsets for clocks.
A time namespace is similar to a pid namespace in the way how it is
created: unshare(CLONE_NEWTIME) system call creates a new time namespace,
but doesn't set it to the current process. Then all children of the process
will be born in the new time namespace, or a process can use the setns()
system call to join a namespace.
This scheme allows setting clock offsets for a namespace, before any
processes appear in it.
All available clone flags have been used, so CLONE_NEWTIME uses the highest
bit of CSIGNAL. It means that it can be used only with the unshare() and
the clone3() system calls.
[ tglx: Adjusted paragraph about clone3() to reality and massaged the
changelog a bit. ]
Co-developed-by: Dmitry Safonov <dima@arista.com>
Signed-off-by: Andrei Vagin <avagin@gmail.com>
Signed-off-by: Dmitry Safonov <dima@arista.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Link: https://criu.org/Time_namespace
Link: https://lists.openvz.org/pipermail/criu/2018-June/041504.html
Link: https://lore.kernel.org/r/20191112012724.250792-4-dima@arista.com
2019-11-12 04:26:52 +03:00
get_nsproxy ( old_ns ) ;
return 0 ;
}
} else if ( ! ns_capable ( user_ns , CAP_SYS_ADMIN ) )
2013-03-10 04:15:23 +04:00
return - EPERM ;
2008-04-29 12:01:00 +04:00
/*
* CLONE_NEWIPC must detach from the undolist : after switching
* to a new ipc namespace , the semaphore arrays from the old
* namespace are unreachable . In clone parlance , CLONE_SYSVSEM
* means share undolist with parent , so we must forbid using
* it along with CLONE_NEWIPC .
*/
2013-02-27 22:32:09 +04:00
if ( ( flags & ( CLONE_NEWIPC | CLONE_SYSVSEM ) ) = =
2020-11-15 21:00:54 +03:00
( CLONE_NEWIPC | CLONE_SYSVSEM ) )
2013-03-10 04:15:23 +04:00
return - EINVAL ;
2008-04-29 12:01:00 +04:00
2013-02-22 04:44:21 +04:00
new_ns = create_new_namespaces ( flags , tsk , user_ns , tsk - > fs ) ;
2013-03-10 04:15:23 +04:00
if ( IS_ERR ( new_ns ) )
return PTR_ERR ( new_ns ) ;
2006-12-08 13:37:59 +03:00
2022-09-21 03:31:19 +03:00
if ( ( flags & CLONE_VM ) = = 0 )
timens_on_fork ( new_ns , tsk ) ;
ns: Introduce Time Namespace
Time Namespace isolates clock values.
The kernel provides access to several clocks CLOCK_REALTIME,
CLOCK_MONOTONIC, CLOCK_BOOTTIME, etc.
CLOCK_REALTIME
System-wide clock that measures real (i.e., wall-clock) time.
CLOCK_MONOTONIC
Clock that cannot be set and represents monotonic time since
some unspecified starting point.
CLOCK_BOOTTIME
Identical to CLOCK_MONOTONIC, except it also includes any time
that the system is suspended.
For many users, the time namespace means the ability to changes date and
time in a container (CLOCK_REALTIME). Providing per namespace notions of
CLOCK_REALTIME would be complex with a massive overhead, but has a dubious
value.
But in the context of checkpoint/restore functionality, monotonic and
boottime clocks become interesting. Both clocks are monotonic with
unspecified starting points. These clocks are widely used to measure time
slices and set timers. After restoring or migrating processes, it has to be
guaranteed that they never go backward. In an ideal case, the behavior of
these clocks should be the same as for a case when a whole system is
suspended. All this means that it is required to set CLOCK_MONOTONIC and
CLOCK_BOOTTIME clocks, which can be achieved by adding per-namespace
offsets for clocks.
A time namespace is similar to a pid namespace in the way how it is
created: unshare(CLONE_NEWTIME) system call creates a new time namespace,
but doesn't set it to the current process. Then all children of the process
will be born in the new time namespace, or a process can use the setns()
system call to join a namespace.
This scheme allows setting clock offsets for a namespace, before any
processes appear in it.
All available clone flags have been used, so CLONE_NEWTIME uses the highest
bit of CSIGNAL. It means that it can be used only with the unshare() and
the clone3() system calls.
[ tglx: Adjusted paragraph about clone3() to reality and massaged the
changelog a bit. ]
Co-developed-by: Dmitry Safonov <dima@arista.com>
Signed-off-by: Andrei Vagin <avagin@gmail.com>
Signed-off-by: Dmitry Safonov <dima@arista.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Link: https://criu.org/Time_namespace
Link: https://lists.openvz.org/pipermail/criu/2018-June/041504.html
Link: https://lore.kernel.org/r/20191112012724.250792-4-dima@arista.com
2019-11-12 04:26:52 +03:00
2007-05-08 11:25:21 +04:00
tsk - > nsproxy = new_ns ;
2013-03-10 04:15:23 +04:00
return 0 ;
2006-10-02 13:18:06 +04:00
}
void free_nsproxy ( struct nsproxy * ns )
{
2006-12-08 13:37:59 +03:00
if ( ns - > mnt_ns )
put_mnt_ns ( ns - > mnt_ns ) ;
if ( ns - > uts_ns )
put_uts_ns ( ns - > uts_ns ) ;
if ( ns - > ipc_ns )
put_ipc_ns ( ns - > ipc_ns ) ;
2013-08-22 22:39:16 +04:00
if ( ns - > pid_ns_for_children )
put_pid_ns ( ns - > pid_ns_for_children ) ;
ns: Introduce Time Namespace
Time Namespace isolates clock values.
The kernel provides access to several clocks CLOCK_REALTIME,
CLOCK_MONOTONIC, CLOCK_BOOTTIME, etc.
CLOCK_REALTIME
System-wide clock that measures real (i.e., wall-clock) time.
CLOCK_MONOTONIC
Clock that cannot be set and represents monotonic time since
some unspecified starting point.
CLOCK_BOOTTIME
Identical to CLOCK_MONOTONIC, except it also includes any time
that the system is suspended.
For many users, the time namespace means the ability to changes date and
time in a container (CLOCK_REALTIME). Providing per namespace notions of
CLOCK_REALTIME would be complex with a massive overhead, but has a dubious
value.
But in the context of checkpoint/restore functionality, monotonic and
boottime clocks become interesting. Both clocks are monotonic with
unspecified starting points. These clocks are widely used to measure time
slices and set timers. After restoring or migrating processes, it has to be
guaranteed that they never go backward. In an ideal case, the behavior of
these clocks should be the same as for a case when a whole system is
suspended. All this means that it is required to set CLOCK_MONOTONIC and
CLOCK_BOOTTIME clocks, which can be achieved by adding per-namespace
offsets for clocks.
A time namespace is similar to a pid namespace in the way how it is
created: unshare(CLONE_NEWTIME) system call creates a new time namespace,
but doesn't set it to the current process. Then all children of the process
will be born in the new time namespace, or a process can use the setns()
system call to join a namespace.
This scheme allows setting clock offsets for a namespace, before any
processes appear in it.
All available clone flags have been used, so CLONE_NEWTIME uses the highest
bit of CSIGNAL. It means that it can be used only with the unshare() and
the clone3() system calls.
[ tglx: Adjusted paragraph about clone3() to reality and massaged the
changelog a bit. ]
Co-developed-by: Dmitry Safonov <dima@arista.com>
Signed-off-by: Andrei Vagin <avagin@gmail.com>
Signed-off-by: Dmitry Safonov <dima@arista.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Link: https://criu.org/Time_namespace
Link: https://lists.openvz.org/pipermail/criu/2018-June/041504.html
Link: https://lore.kernel.org/r/20191112012724.250792-4-dima@arista.com
2019-11-12 04:26:52 +03:00
if ( ns - > time_ns )
put_time_ns ( ns - > time_ns ) ;
if ( ns - > time_ns_for_children )
put_time_ns ( ns - > time_ns_for_children ) ;
2016-01-29 11:54:06 +03:00
put_cgroup_ns ( ns - > cgroup_ns ) ;
2007-09-27 09:04:26 +04:00
put_net ( ns - > net_ns ) ;
2007-07-16 10:41:07 +04:00
kmem_cache_free ( nsproxy_cachep , ns ) ;
2006-10-02 13:18:06 +04:00
}
2007-05-08 11:25:21 +04:00
/*
* Called from unshare . Unshare all the namespaces part of nsproxy .
2007-06-24 04:16:25 +04:00
* On success , returns the new nsproxy .
2007-05-08 11:25:21 +04:00
*/
int unshare_nsproxy_namespaces ( unsigned long unshare_flags ,
2012-07-26 16:15:35 +04:00
struct nsproxy * * new_nsp , struct cred * new_cred , struct fs_struct * new_fs )
2007-05-08 11:25:21 +04:00
{
2012-07-26 15:02:49 +04:00
struct user_namespace * user_ns ;
2007-05-08 11:25:21 +04:00
int err = 0 ;
2007-07-16 10:41:01 +04:00
if ( ! ( unshare_flags & ( CLONE_NEWNS | CLONE_NEWUTS | CLONE_NEWIPC |
ns: Introduce Time Namespace
Time Namespace isolates clock values.
The kernel provides access to several clocks CLOCK_REALTIME,
CLOCK_MONOTONIC, CLOCK_BOOTTIME, etc.
CLOCK_REALTIME
System-wide clock that measures real (i.e., wall-clock) time.
CLOCK_MONOTONIC
Clock that cannot be set and represents monotonic time since
some unspecified starting point.
CLOCK_BOOTTIME
Identical to CLOCK_MONOTONIC, except it also includes any time
that the system is suspended.
For many users, the time namespace means the ability to changes date and
time in a container (CLOCK_REALTIME). Providing per namespace notions of
CLOCK_REALTIME would be complex with a massive overhead, but has a dubious
value.
But in the context of checkpoint/restore functionality, monotonic and
boottime clocks become interesting. Both clocks are monotonic with
unspecified starting points. These clocks are widely used to measure time
slices and set timers. After restoring or migrating processes, it has to be
guaranteed that they never go backward. In an ideal case, the behavior of
these clocks should be the same as for a case when a whole system is
suspended. All this means that it is required to set CLOCK_MONOTONIC and
CLOCK_BOOTTIME clocks, which can be achieved by adding per-namespace
offsets for clocks.
A time namespace is similar to a pid namespace in the way how it is
created: unshare(CLONE_NEWTIME) system call creates a new time namespace,
but doesn't set it to the current process. Then all children of the process
will be born in the new time namespace, or a process can use the setns()
system call to join a namespace.
This scheme allows setting clock offsets for a namespace, before any
processes appear in it.
All available clone flags have been used, so CLONE_NEWTIME uses the highest
bit of CSIGNAL. It means that it can be used only with the unshare() and
the clone3() system calls.
[ tglx: Adjusted paragraph about clone3() to reality and massaged the
changelog a bit. ]
Co-developed-by: Dmitry Safonov <dima@arista.com>
Signed-off-by: Andrei Vagin <avagin@gmail.com>
Signed-off-by: Dmitry Safonov <dima@arista.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Link: https://criu.org/Time_namespace
Link: https://lists.openvz.org/pipermail/criu/2018-June/041504.html
Link: https://lore.kernel.org/r/20191112012724.250792-4-dima@arista.com
2019-11-12 04:26:52 +03:00
CLONE_NEWNET | CLONE_NEWPID | CLONE_NEWCGROUP |
CLONE_NEWTIME ) ) )
2007-05-08 11:25:21 +04:00
return 0 ;
2012-07-26 16:15:35 +04:00
user_ns = new_cred ? new_cred - > user_ns : current_user_ns ( ) ;
if ( ! ns_capable ( user_ns , CAP_SYS_ADMIN ) )
2007-05-08 11:25:21 +04:00
return - EPERM ;
2012-07-26 15:02:49 +04:00
* new_nsp = create_new_namespaces ( unshare_flags , current , user_ns ,
2012-07-26 16:15:35 +04:00
new_fs ? new_fs : current - > fs ) ;
2007-10-19 10:39:45 +04:00
if ( IS_ERR ( * new_nsp ) ) {
2007-05-08 11:25:21 +04:00
err = PTR_ERR ( * new_nsp ) ;
2007-10-19 10:39:45 +04:00
goto out ;
}
out :
2007-05-08 11:25:21 +04:00
return err ;
}
2007-07-16 10:41:07 +04:00
2007-10-19 10:39:54 +04:00
void switch_task_namespaces ( struct task_struct * p , struct nsproxy * new )
{
struct nsproxy * ns ;
might_sleep ( ) ;
2014-02-04 07:13:49 +04:00
task_lock ( p ) ;
2007-10-19 10:39:54 +04:00
ns = p - > nsproxy ;
2014-02-04 07:13:49 +04:00
p - > nsproxy = new ;
task_unlock ( p ) ;
2007-10-19 10:39:54 +04:00
2020-11-15 21:00:54 +03:00
if ( ns )
put_nsproxy ( ns ) ;
2007-10-19 10:39:54 +04:00
}
void exit_task_namespaces ( struct task_struct * p )
{
switch_task_namespaces ( p , NULL ) ;
}
2022-09-21 03:31:19 +03:00
int exec_task_namespaces ( void )
{
struct task_struct * tsk = current ;
struct nsproxy * new ;
if ( tsk - > nsproxy - > time_ns_for_children = = tsk - > nsproxy - > time_ns )
return 0 ;
new = create_new_namespaces ( 0 , tsk , current_user_ns ( ) , tsk - > fs ) ;
if ( IS_ERR ( new ) )
return PTR_ERR ( new ) ;
timens_on_fork ( new , tsk ) ;
switch_task_namespaces ( tsk , new ) ;
return 0 ;
}
nsproxy: attach to namespaces via pidfds
For quite a while we have been thinking about using pidfds to attach to
namespaces. This patchset has existed for about a year already but we've
wanted to wait to see how the general api would be received and adopted.
Now that more and more programs in userspace have started using pidfds
for process management it's time to send this one out.
This patch makes it possible to use pidfds to attach to the namespaces
of another process, i.e. they can be passed as the first argument to the
setns() syscall. When only a single namespace type is specified the
semantics are equivalent to passing an nsfd. That means
setns(nsfd, CLONE_NEWNET) equals setns(pidfd, CLONE_NEWNET). However,
when a pidfd is passed, multiple namespace flags can be specified in the
second setns() argument and setns() will attach the caller to all the
specified namespaces all at once or to none of them. Specifying 0 is not
valid together with a pidfd.
Here are just two obvious examples:
setns(pidfd, CLONE_NEWPID | CLONE_NEWNS | CLONE_NEWNET);
setns(pidfd, CLONE_NEWUSER);
Allowing to also attach subsets of namespaces supports various use-cases
where callers setns to a subset of namespaces to retain privilege, perform
an action and then re-attach another subset of namespaces.
If the need arises, as Eric suggested, we can extend this patchset to
assume even more context than just attaching all namespaces. His suggestion
specifically was about assuming the process' root directory when
setns(pidfd, 0) or setns(pidfd, SETNS_PIDFD) is specified. For now, just
keep it flexible in terms of supporting subsets of namespaces but let's
wait until we have users asking for even more context to be assumed. At
that point we can add an extension.
The obvious example where this is useful is a standard container
manager interacting with a running container: pushing and pulling files
or directories, injecting mounts, attaching/execing any kind of process,
managing network devices all these operations require attaching to all
or at least multiple namespaces at the same time. Given that nowadays
most containers are spawned with all namespaces enabled we're currently
looking at at least 14 syscalls, 7 to open the /proc/<pid>/ns/<ns>
nsfds, another 7 to actually perform the namespace switch. With time
namespaces we're looking at about 16 syscalls.
(We could amortize the first 7 or 8 syscalls for opening the nsfds by
stashing them in each container's monitor process but that would mean
we need to send around those file descriptors through unix sockets
everytime we want to interact with the container or keep on-disk
state. Even in scenarios where a caller wants to join a particular
namespace in a particular order callers still profit from batching
other namespaces. That mostly applies to the user namespace but
all container runtimes I found join the user namespace first no matter
if it privileges or deprivileges the container similar to how unshare
behaves.)
With pidfds this becomes a single syscall no matter how many namespaces
are supposed to be attached to.
A decently designed, large-scale container manager usually isn't the
parent of any of the containers it spawns so the containers don't die
when it crashes or needs to update or reinitialize. This means that
for the manager to interact with containers through pids is inherently
racy especially on systems where the maximum pid number is not
significicantly bumped. This is even more problematic since we often spawn
and manage thousands or ten-thousands of containers. Interacting with a
container through a pid thus can become risky quite quickly. Especially
since we allow for an administrator to enable advanced features such as
syscall interception where we're performing syscalls in lieu of the
container. In all of those cases we use pidfds if they are available and
we pass them around as stable references. Using them to setns() to the
target process' namespaces is as reliable as using nsfds. Either the
target process is already dead and we get ESRCH or we manage to attach
to its namespaces but we can't accidently attach to another process'
namespaces. So pidfds lend themselves to be used with this api.
The other main advantage is that with this change the pidfd becomes the
only relevant token for most container interactions and it's the only
token we need to create and send around.
Apart from significiantly reducing the number of syscalls from double
digit to single digit which is a decent reason post-spectre/meltdown
this also allows to switch to a set of namespaces atomically, i.e.
either attaching to all the specified namespaces succeeds or we fail. If
we fail we haven't changed a single namespace. There are currently three
namespaces that can fail (other than for ENOMEM which really is not
very interesting since we then have other problems anyway) for
non-trivial reasons, user, mount, and pid namespaces. We can fail to
attach to a pid namespace if it is not our current active pid namespace
or a descendant of it. We can fail to attach to a user namespace because
we are multi-threaded or because our current mount namespace shares
filesystem state with other tasks, or because we're trying to setns()
to the same user namespace, i.e. the target task has the same user
namespace as we do. We can fail to attach to a mount namespace because
it shares filesystem state with other tasks or because we fail to lookup
the new root for the new mount namespace. In most non-pathological
scenarios these issues can be somewhat mitigated. But there are cases where
we're half-attached to some namespace and failing to attach to another one.
I've talked about some of these problem during the hallway track (something
only the pre-COVID-19 generation will remember) of Plumbers in Los Angeles
in 2018(?). Even if all these issues could be avoided with super careful
userspace coding it would be nicer to have this done in-kernel. Pidfds seem
to lend themselves nicely for this.
The other neat thing about this is that setns() becomes an actual
counterpart to the namespace bits of unshare().
Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
Reviewed-by: Serge Hallyn <serge@hallyn.com>
Cc: Eric W. Biederman <ebiederm@xmission.com>
Cc: Serge Hallyn <serge@hallyn.com>
Cc: Jann Horn <jannh@google.com>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Aleksa Sarai <cyphar@cyphar.com>
Link: https://lore.kernel.org/r/20200505140432.181565-3-christian.brauner@ubuntu.com
2020-05-05 17:04:31 +03:00
static int check_setns_flags ( unsigned long flags )
{
if ( ! flags | | ( flags & ~ ( CLONE_NEWNS | CLONE_NEWUTS | CLONE_NEWIPC |
2020-07-06 18:49:11 +03:00
CLONE_NEWNET | CLONE_NEWTIME | CLONE_NEWUSER |
CLONE_NEWPID | CLONE_NEWCGROUP ) ) )
nsproxy: attach to namespaces via pidfds
For quite a while we have been thinking about using pidfds to attach to
namespaces. This patchset has existed for about a year already but we've
wanted to wait to see how the general api would be received and adopted.
Now that more and more programs in userspace have started using pidfds
for process management it's time to send this one out.
This patch makes it possible to use pidfds to attach to the namespaces
of another process, i.e. they can be passed as the first argument to the
setns() syscall. When only a single namespace type is specified the
semantics are equivalent to passing an nsfd. That means
setns(nsfd, CLONE_NEWNET) equals setns(pidfd, CLONE_NEWNET). However,
when a pidfd is passed, multiple namespace flags can be specified in the
second setns() argument and setns() will attach the caller to all the
specified namespaces all at once or to none of them. Specifying 0 is not
valid together with a pidfd.
Here are just two obvious examples:
setns(pidfd, CLONE_NEWPID | CLONE_NEWNS | CLONE_NEWNET);
setns(pidfd, CLONE_NEWUSER);
Allowing to also attach subsets of namespaces supports various use-cases
where callers setns to a subset of namespaces to retain privilege, perform
an action and then re-attach another subset of namespaces.
If the need arises, as Eric suggested, we can extend this patchset to
assume even more context than just attaching all namespaces. His suggestion
specifically was about assuming the process' root directory when
setns(pidfd, 0) or setns(pidfd, SETNS_PIDFD) is specified. For now, just
keep it flexible in terms of supporting subsets of namespaces but let's
wait until we have users asking for even more context to be assumed. At
that point we can add an extension.
The obvious example where this is useful is a standard container
manager interacting with a running container: pushing and pulling files
or directories, injecting mounts, attaching/execing any kind of process,
managing network devices all these operations require attaching to all
or at least multiple namespaces at the same time. Given that nowadays
most containers are spawned with all namespaces enabled we're currently
looking at at least 14 syscalls, 7 to open the /proc/<pid>/ns/<ns>
nsfds, another 7 to actually perform the namespace switch. With time
namespaces we're looking at about 16 syscalls.
(We could amortize the first 7 or 8 syscalls for opening the nsfds by
stashing them in each container's monitor process but that would mean
we need to send around those file descriptors through unix sockets
everytime we want to interact with the container or keep on-disk
state. Even in scenarios where a caller wants to join a particular
namespace in a particular order callers still profit from batching
other namespaces. That mostly applies to the user namespace but
all container runtimes I found join the user namespace first no matter
if it privileges or deprivileges the container similar to how unshare
behaves.)
With pidfds this becomes a single syscall no matter how many namespaces
are supposed to be attached to.
A decently designed, large-scale container manager usually isn't the
parent of any of the containers it spawns so the containers don't die
when it crashes or needs to update or reinitialize. This means that
for the manager to interact with containers through pids is inherently
racy especially on systems where the maximum pid number is not
significicantly bumped. This is even more problematic since we often spawn
and manage thousands or ten-thousands of containers. Interacting with a
container through a pid thus can become risky quite quickly. Especially
since we allow for an administrator to enable advanced features such as
syscall interception where we're performing syscalls in lieu of the
container. In all of those cases we use pidfds if they are available and
we pass them around as stable references. Using them to setns() to the
target process' namespaces is as reliable as using nsfds. Either the
target process is already dead and we get ESRCH or we manage to attach
to its namespaces but we can't accidently attach to another process'
namespaces. So pidfds lend themselves to be used with this api.
The other main advantage is that with this change the pidfd becomes the
only relevant token for most container interactions and it's the only
token we need to create and send around.
Apart from significiantly reducing the number of syscalls from double
digit to single digit which is a decent reason post-spectre/meltdown
this also allows to switch to a set of namespaces atomically, i.e.
either attaching to all the specified namespaces succeeds or we fail. If
we fail we haven't changed a single namespace. There are currently three
namespaces that can fail (other than for ENOMEM which really is not
very interesting since we then have other problems anyway) for
non-trivial reasons, user, mount, and pid namespaces. We can fail to
attach to a pid namespace if it is not our current active pid namespace
or a descendant of it. We can fail to attach to a user namespace because
we are multi-threaded or because our current mount namespace shares
filesystem state with other tasks, or because we're trying to setns()
to the same user namespace, i.e. the target task has the same user
namespace as we do. We can fail to attach to a mount namespace because
it shares filesystem state with other tasks or because we fail to lookup
the new root for the new mount namespace. In most non-pathological
scenarios these issues can be somewhat mitigated. But there are cases where
we're half-attached to some namespace and failing to attach to another one.
I've talked about some of these problem during the hallway track (something
only the pre-COVID-19 generation will remember) of Plumbers in Los Angeles
in 2018(?). Even if all these issues could be avoided with super careful
userspace coding it would be nicer to have this done in-kernel. Pidfds seem
to lend themselves nicely for this.
The other neat thing about this is that setns() becomes an actual
counterpart to the namespace bits of unshare().
Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
Reviewed-by: Serge Hallyn <serge@hallyn.com>
Cc: Eric W. Biederman <ebiederm@xmission.com>
Cc: Serge Hallyn <serge@hallyn.com>
Cc: Jann Horn <jannh@google.com>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Aleksa Sarai <cyphar@cyphar.com>
Link: https://lore.kernel.org/r/20200505140432.181565-3-christian.brauner@ubuntu.com
2020-05-05 17:04:31 +03:00
return - EINVAL ;
# ifndef CONFIG_USER_NS
if ( flags & CLONE_NEWUSER )
return - EINVAL ;
# endif
# ifndef CONFIG_PID_NS
if ( flags & CLONE_NEWPID )
return - EINVAL ;
# endif
# ifndef CONFIG_UTS_NS
if ( flags & CLONE_NEWUTS )
return - EINVAL ;
# endif
# ifndef CONFIG_IPC_NS
if ( flags & CLONE_NEWIPC )
return - EINVAL ;
# endif
# ifndef CONFIG_CGROUPS
if ( flags & CLONE_NEWCGROUP )
return - EINVAL ;
# endif
# ifndef CONFIG_NET_NS
if ( flags & CLONE_NEWNET )
return - EINVAL ;
# endif
2020-07-06 18:49:11 +03:00
# ifndef CONFIG_TIME_NS
if ( flags & CLONE_NEWTIME )
return - EINVAL ;
# endif
nsproxy: attach to namespaces via pidfds
For quite a while we have been thinking about using pidfds to attach to
namespaces. This patchset has existed for about a year already but we've
wanted to wait to see how the general api would be received and adopted.
Now that more and more programs in userspace have started using pidfds
for process management it's time to send this one out.
This patch makes it possible to use pidfds to attach to the namespaces
of another process, i.e. they can be passed as the first argument to the
setns() syscall. When only a single namespace type is specified the
semantics are equivalent to passing an nsfd. That means
setns(nsfd, CLONE_NEWNET) equals setns(pidfd, CLONE_NEWNET). However,
when a pidfd is passed, multiple namespace flags can be specified in the
second setns() argument and setns() will attach the caller to all the
specified namespaces all at once or to none of them. Specifying 0 is not
valid together with a pidfd.
Here are just two obvious examples:
setns(pidfd, CLONE_NEWPID | CLONE_NEWNS | CLONE_NEWNET);
setns(pidfd, CLONE_NEWUSER);
Allowing to also attach subsets of namespaces supports various use-cases
where callers setns to a subset of namespaces to retain privilege, perform
an action and then re-attach another subset of namespaces.
If the need arises, as Eric suggested, we can extend this patchset to
assume even more context than just attaching all namespaces. His suggestion
specifically was about assuming the process' root directory when
setns(pidfd, 0) or setns(pidfd, SETNS_PIDFD) is specified. For now, just
keep it flexible in terms of supporting subsets of namespaces but let's
wait until we have users asking for even more context to be assumed. At
that point we can add an extension.
The obvious example where this is useful is a standard container
manager interacting with a running container: pushing and pulling files
or directories, injecting mounts, attaching/execing any kind of process,
managing network devices all these operations require attaching to all
or at least multiple namespaces at the same time. Given that nowadays
most containers are spawned with all namespaces enabled we're currently
looking at at least 14 syscalls, 7 to open the /proc/<pid>/ns/<ns>
nsfds, another 7 to actually perform the namespace switch. With time
namespaces we're looking at about 16 syscalls.
(We could amortize the first 7 or 8 syscalls for opening the nsfds by
stashing them in each container's monitor process but that would mean
we need to send around those file descriptors through unix sockets
everytime we want to interact with the container or keep on-disk
state. Even in scenarios where a caller wants to join a particular
namespace in a particular order callers still profit from batching
other namespaces. That mostly applies to the user namespace but
all container runtimes I found join the user namespace first no matter
if it privileges or deprivileges the container similar to how unshare
behaves.)
With pidfds this becomes a single syscall no matter how many namespaces
are supposed to be attached to.
A decently designed, large-scale container manager usually isn't the
parent of any of the containers it spawns so the containers don't die
when it crashes or needs to update or reinitialize. This means that
for the manager to interact with containers through pids is inherently
racy especially on systems where the maximum pid number is not
significicantly bumped. This is even more problematic since we often spawn
and manage thousands or ten-thousands of containers. Interacting with a
container through a pid thus can become risky quite quickly. Especially
since we allow for an administrator to enable advanced features such as
syscall interception where we're performing syscalls in lieu of the
container. In all of those cases we use pidfds if they are available and
we pass them around as stable references. Using them to setns() to the
target process' namespaces is as reliable as using nsfds. Either the
target process is already dead and we get ESRCH or we manage to attach
to its namespaces but we can't accidently attach to another process'
namespaces. So pidfds lend themselves to be used with this api.
The other main advantage is that with this change the pidfd becomes the
only relevant token for most container interactions and it's the only
token we need to create and send around.
Apart from significiantly reducing the number of syscalls from double
digit to single digit which is a decent reason post-spectre/meltdown
this also allows to switch to a set of namespaces atomically, i.e.
either attaching to all the specified namespaces succeeds or we fail. If
we fail we haven't changed a single namespace. There are currently three
namespaces that can fail (other than for ENOMEM which really is not
very interesting since we then have other problems anyway) for
non-trivial reasons, user, mount, and pid namespaces. We can fail to
attach to a pid namespace if it is not our current active pid namespace
or a descendant of it. We can fail to attach to a user namespace because
we are multi-threaded or because our current mount namespace shares
filesystem state with other tasks, or because we're trying to setns()
to the same user namespace, i.e. the target task has the same user
namespace as we do. We can fail to attach to a mount namespace because
it shares filesystem state with other tasks or because we fail to lookup
the new root for the new mount namespace. In most non-pathological
scenarios these issues can be somewhat mitigated. But there are cases where
we're half-attached to some namespace and failing to attach to another one.
I've talked about some of these problem during the hallway track (something
only the pre-COVID-19 generation will remember) of Plumbers in Los Angeles
in 2018(?). Even if all these issues could be avoided with super careful
userspace coding it would be nicer to have this done in-kernel. Pidfds seem
to lend themselves nicely for this.
The other neat thing about this is that setns() becomes an actual
counterpart to the namespace bits of unshare().
Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
Reviewed-by: Serge Hallyn <serge@hallyn.com>
Cc: Eric W. Biederman <ebiederm@xmission.com>
Cc: Serge Hallyn <serge@hallyn.com>
Cc: Jann Horn <jannh@google.com>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Aleksa Sarai <cyphar@cyphar.com>
Link: https://lore.kernel.org/r/20200505140432.181565-3-christian.brauner@ubuntu.com
2020-05-05 17:04:31 +03:00
return 0 ;
}
2020-05-05 17:04:30 +03:00
static void put_nsset ( struct nsset * nsset )
{
unsigned flags = nsset - > flags ;
if ( flags & CLONE_NEWUSER )
put_cred ( nsset_cred ( nsset ) ) ;
nsproxy: attach to namespaces via pidfds
For quite a while we have been thinking about using pidfds to attach to
namespaces. This patchset has existed for about a year already but we've
wanted to wait to see how the general api would be received and adopted.
Now that more and more programs in userspace have started using pidfds
for process management it's time to send this one out.
This patch makes it possible to use pidfds to attach to the namespaces
of another process, i.e. they can be passed as the first argument to the
setns() syscall. When only a single namespace type is specified the
semantics are equivalent to passing an nsfd. That means
setns(nsfd, CLONE_NEWNET) equals setns(pidfd, CLONE_NEWNET). However,
when a pidfd is passed, multiple namespace flags can be specified in the
second setns() argument and setns() will attach the caller to all the
specified namespaces all at once or to none of them. Specifying 0 is not
valid together with a pidfd.
Here are just two obvious examples:
setns(pidfd, CLONE_NEWPID | CLONE_NEWNS | CLONE_NEWNET);
setns(pidfd, CLONE_NEWUSER);
Allowing to also attach subsets of namespaces supports various use-cases
where callers setns to a subset of namespaces to retain privilege, perform
an action and then re-attach another subset of namespaces.
If the need arises, as Eric suggested, we can extend this patchset to
assume even more context than just attaching all namespaces. His suggestion
specifically was about assuming the process' root directory when
setns(pidfd, 0) or setns(pidfd, SETNS_PIDFD) is specified. For now, just
keep it flexible in terms of supporting subsets of namespaces but let's
wait until we have users asking for even more context to be assumed. At
that point we can add an extension.
The obvious example where this is useful is a standard container
manager interacting with a running container: pushing and pulling files
or directories, injecting mounts, attaching/execing any kind of process,
managing network devices all these operations require attaching to all
or at least multiple namespaces at the same time. Given that nowadays
most containers are spawned with all namespaces enabled we're currently
looking at at least 14 syscalls, 7 to open the /proc/<pid>/ns/<ns>
nsfds, another 7 to actually perform the namespace switch. With time
namespaces we're looking at about 16 syscalls.
(We could amortize the first 7 or 8 syscalls for opening the nsfds by
stashing them in each container's monitor process but that would mean
we need to send around those file descriptors through unix sockets
everytime we want to interact with the container or keep on-disk
state. Even in scenarios where a caller wants to join a particular
namespace in a particular order callers still profit from batching
other namespaces. That mostly applies to the user namespace but
all container runtimes I found join the user namespace first no matter
if it privileges or deprivileges the container similar to how unshare
behaves.)
With pidfds this becomes a single syscall no matter how many namespaces
are supposed to be attached to.
A decently designed, large-scale container manager usually isn't the
parent of any of the containers it spawns so the containers don't die
when it crashes or needs to update or reinitialize. This means that
for the manager to interact with containers through pids is inherently
racy especially on systems where the maximum pid number is not
significicantly bumped. This is even more problematic since we often spawn
and manage thousands or ten-thousands of containers. Interacting with a
container through a pid thus can become risky quite quickly. Especially
since we allow for an administrator to enable advanced features such as
syscall interception where we're performing syscalls in lieu of the
container. In all of those cases we use pidfds if they are available and
we pass them around as stable references. Using them to setns() to the
target process' namespaces is as reliable as using nsfds. Either the
target process is already dead and we get ESRCH or we manage to attach
to its namespaces but we can't accidently attach to another process'
namespaces. So pidfds lend themselves to be used with this api.
The other main advantage is that with this change the pidfd becomes the
only relevant token for most container interactions and it's the only
token we need to create and send around.
Apart from significiantly reducing the number of syscalls from double
digit to single digit which is a decent reason post-spectre/meltdown
this also allows to switch to a set of namespaces atomically, i.e.
either attaching to all the specified namespaces succeeds or we fail. If
we fail we haven't changed a single namespace. There are currently three
namespaces that can fail (other than for ENOMEM which really is not
very interesting since we then have other problems anyway) for
non-trivial reasons, user, mount, and pid namespaces. We can fail to
attach to a pid namespace if it is not our current active pid namespace
or a descendant of it. We can fail to attach to a user namespace because
we are multi-threaded or because our current mount namespace shares
filesystem state with other tasks, or because we're trying to setns()
to the same user namespace, i.e. the target task has the same user
namespace as we do. We can fail to attach to a mount namespace because
it shares filesystem state with other tasks or because we fail to lookup
the new root for the new mount namespace. In most non-pathological
scenarios these issues can be somewhat mitigated. But there are cases where
we're half-attached to some namespace and failing to attach to another one.
I've talked about some of these problem during the hallway track (something
only the pre-COVID-19 generation will remember) of Plumbers in Los Angeles
in 2018(?). Even if all these issues could be avoided with super careful
userspace coding it would be nicer to have this done in-kernel. Pidfds seem
to lend themselves nicely for this.
The other neat thing about this is that setns() becomes an actual
counterpart to the namespace bits of unshare().
Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
Reviewed-by: Serge Hallyn <serge@hallyn.com>
Cc: Eric W. Biederman <ebiederm@xmission.com>
Cc: Serge Hallyn <serge@hallyn.com>
Cc: Jann Horn <jannh@google.com>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Aleksa Sarai <cyphar@cyphar.com>
Link: https://lore.kernel.org/r/20200505140432.181565-3-christian.brauner@ubuntu.com
2020-05-05 17:04:31 +03:00
/*
* We only created a temporary copy if we attached to more than just
* the mount namespace .
*/
if ( nsset - > fs & & ( flags & CLONE_NEWNS ) & & ( flags & ~ CLONE_NEWNS ) )
free_fs_struct ( nsset - > fs ) ;
2020-05-05 17:04:30 +03:00
if ( nsset - > nsproxy )
free_nsproxy ( nsset - > nsproxy ) ;
}
nsproxy: attach to namespaces via pidfds
For quite a while we have been thinking about using pidfds to attach to
namespaces. This patchset has existed for about a year already but we've
wanted to wait to see how the general api would be received and adopted.
Now that more and more programs in userspace have started using pidfds
for process management it's time to send this one out.
This patch makes it possible to use pidfds to attach to the namespaces
of another process, i.e. they can be passed as the first argument to the
setns() syscall. When only a single namespace type is specified the
semantics are equivalent to passing an nsfd. That means
setns(nsfd, CLONE_NEWNET) equals setns(pidfd, CLONE_NEWNET). However,
when a pidfd is passed, multiple namespace flags can be specified in the
second setns() argument and setns() will attach the caller to all the
specified namespaces all at once or to none of them. Specifying 0 is not
valid together with a pidfd.
Here are just two obvious examples:
setns(pidfd, CLONE_NEWPID | CLONE_NEWNS | CLONE_NEWNET);
setns(pidfd, CLONE_NEWUSER);
Allowing to also attach subsets of namespaces supports various use-cases
where callers setns to a subset of namespaces to retain privilege, perform
an action and then re-attach another subset of namespaces.
If the need arises, as Eric suggested, we can extend this patchset to
assume even more context than just attaching all namespaces. His suggestion
specifically was about assuming the process' root directory when
setns(pidfd, 0) or setns(pidfd, SETNS_PIDFD) is specified. For now, just
keep it flexible in terms of supporting subsets of namespaces but let's
wait until we have users asking for even more context to be assumed. At
that point we can add an extension.
The obvious example where this is useful is a standard container
manager interacting with a running container: pushing and pulling files
or directories, injecting mounts, attaching/execing any kind of process,
managing network devices all these operations require attaching to all
or at least multiple namespaces at the same time. Given that nowadays
most containers are spawned with all namespaces enabled we're currently
looking at at least 14 syscalls, 7 to open the /proc/<pid>/ns/<ns>
nsfds, another 7 to actually perform the namespace switch. With time
namespaces we're looking at about 16 syscalls.
(We could amortize the first 7 or 8 syscalls for opening the nsfds by
stashing them in each container's monitor process but that would mean
we need to send around those file descriptors through unix sockets
everytime we want to interact with the container or keep on-disk
state. Even in scenarios where a caller wants to join a particular
namespace in a particular order callers still profit from batching
other namespaces. That mostly applies to the user namespace but
all container runtimes I found join the user namespace first no matter
if it privileges or deprivileges the container similar to how unshare
behaves.)
With pidfds this becomes a single syscall no matter how many namespaces
are supposed to be attached to.
A decently designed, large-scale container manager usually isn't the
parent of any of the containers it spawns so the containers don't die
when it crashes or needs to update or reinitialize. This means that
for the manager to interact with containers through pids is inherently
racy especially on systems where the maximum pid number is not
significicantly bumped. This is even more problematic since we often spawn
and manage thousands or ten-thousands of containers. Interacting with a
container through a pid thus can become risky quite quickly. Especially
since we allow for an administrator to enable advanced features such as
syscall interception where we're performing syscalls in lieu of the
container. In all of those cases we use pidfds if they are available and
we pass them around as stable references. Using them to setns() to the
target process' namespaces is as reliable as using nsfds. Either the
target process is already dead and we get ESRCH or we manage to attach
to its namespaces but we can't accidently attach to another process'
namespaces. So pidfds lend themselves to be used with this api.
The other main advantage is that with this change the pidfd becomes the
only relevant token for most container interactions and it's the only
token we need to create and send around.
Apart from significiantly reducing the number of syscalls from double
digit to single digit which is a decent reason post-spectre/meltdown
this also allows to switch to a set of namespaces atomically, i.e.
either attaching to all the specified namespaces succeeds or we fail. If
we fail we haven't changed a single namespace. There are currently three
namespaces that can fail (other than for ENOMEM which really is not
very interesting since we then have other problems anyway) for
non-trivial reasons, user, mount, and pid namespaces. We can fail to
attach to a pid namespace if it is not our current active pid namespace
or a descendant of it. We can fail to attach to a user namespace because
we are multi-threaded or because our current mount namespace shares
filesystem state with other tasks, or because we're trying to setns()
to the same user namespace, i.e. the target task has the same user
namespace as we do. We can fail to attach to a mount namespace because
it shares filesystem state with other tasks or because we fail to lookup
the new root for the new mount namespace. In most non-pathological
scenarios these issues can be somewhat mitigated. But there are cases where
we're half-attached to some namespace and failing to attach to another one.
I've talked about some of these problem during the hallway track (something
only the pre-COVID-19 generation will remember) of Plumbers in Los Angeles
in 2018(?). Even if all these issues could be avoided with super careful
userspace coding it would be nicer to have this done in-kernel. Pidfds seem
to lend themselves nicely for this.
The other neat thing about this is that setns() becomes an actual
counterpart to the namespace bits of unshare().
Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
Reviewed-by: Serge Hallyn <serge@hallyn.com>
Cc: Eric W. Biederman <ebiederm@xmission.com>
Cc: Serge Hallyn <serge@hallyn.com>
Cc: Jann Horn <jannh@google.com>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Aleksa Sarai <cyphar@cyphar.com>
Link: https://lore.kernel.org/r/20200505140432.181565-3-christian.brauner@ubuntu.com
2020-05-05 17:04:31 +03:00
static int prepare_nsset ( unsigned flags , struct nsset * nsset )
2020-05-05 17:04:30 +03:00
{
struct task_struct * me = current ;
nsset - > nsproxy = create_new_namespaces ( 0 , me , current_user_ns ( ) , me - > fs ) ;
if ( IS_ERR ( nsset - > nsproxy ) )
return PTR_ERR ( nsset - > nsproxy ) ;
nsproxy: attach to namespaces via pidfds
For quite a while we have been thinking about using pidfds to attach to
namespaces. This patchset has existed for about a year already but we've
wanted to wait to see how the general api would be received and adopted.
Now that more and more programs in userspace have started using pidfds
for process management it's time to send this one out.
This patch makes it possible to use pidfds to attach to the namespaces
of another process, i.e. they can be passed as the first argument to the
setns() syscall. When only a single namespace type is specified the
semantics are equivalent to passing an nsfd. That means
setns(nsfd, CLONE_NEWNET) equals setns(pidfd, CLONE_NEWNET). However,
when a pidfd is passed, multiple namespace flags can be specified in the
second setns() argument and setns() will attach the caller to all the
specified namespaces all at once or to none of them. Specifying 0 is not
valid together with a pidfd.
Here are just two obvious examples:
setns(pidfd, CLONE_NEWPID | CLONE_NEWNS | CLONE_NEWNET);
setns(pidfd, CLONE_NEWUSER);
Allowing to also attach subsets of namespaces supports various use-cases
where callers setns to a subset of namespaces to retain privilege, perform
an action and then re-attach another subset of namespaces.
If the need arises, as Eric suggested, we can extend this patchset to
assume even more context than just attaching all namespaces. His suggestion
specifically was about assuming the process' root directory when
setns(pidfd, 0) or setns(pidfd, SETNS_PIDFD) is specified. For now, just
keep it flexible in terms of supporting subsets of namespaces but let's
wait until we have users asking for even more context to be assumed. At
that point we can add an extension.
The obvious example where this is useful is a standard container
manager interacting with a running container: pushing and pulling files
or directories, injecting mounts, attaching/execing any kind of process,
managing network devices all these operations require attaching to all
or at least multiple namespaces at the same time. Given that nowadays
most containers are spawned with all namespaces enabled we're currently
looking at at least 14 syscalls, 7 to open the /proc/<pid>/ns/<ns>
nsfds, another 7 to actually perform the namespace switch. With time
namespaces we're looking at about 16 syscalls.
(We could amortize the first 7 or 8 syscalls for opening the nsfds by
stashing them in each container's monitor process but that would mean
we need to send around those file descriptors through unix sockets
everytime we want to interact with the container or keep on-disk
state. Even in scenarios where a caller wants to join a particular
namespace in a particular order callers still profit from batching
other namespaces. That mostly applies to the user namespace but
all container runtimes I found join the user namespace first no matter
if it privileges or deprivileges the container similar to how unshare
behaves.)
With pidfds this becomes a single syscall no matter how many namespaces
are supposed to be attached to.
A decently designed, large-scale container manager usually isn't the
parent of any of the containers it spawns so the containers don't die
when it crashes or needs to update or reinitialize. This means that
for the manager to interact with containers through pids is inherently
racy especially on systems where the maximum pid number is not
significicantly bumped. This is even more problematic since we often spawn
and manage thousands or ten-thousands of containers. Interacting with a
container through a pid thus can become risky quite quickly. Especially
since we allow for an administrator to enable advanced features such as
syscall interception where we're performing syscalls in lieu of the
container. In all of those cases we use pidfds if they are available and
we pass them around as stable references. Using them to setns() to the
target process' namespaces is as reliable as using nsfds. Either the
target process is already dead and we get ESRCH or we manage to attach
to its namespaces but we can't accidently attach to another process'
namespaces. So pidfds lend themselves to be used with this api.
The other main advantage is that with this change the pidfd becomes the
only relevant token for most container interactions and it's the only
token we need to create and send around.
Apart from significiantly reducing the number of syscalls from double
digit to single digit which is a decent reason post-spectre/meltdown
this also allows to switch to a set of namespaces atomically, i.e.
either attaching to all the specified namespaces succeeds or we fail. If
we fail we haven't changed a single namespace. There are currently three
namespaces that can fail (other than for ENOMEM which really is not
very interesting since we then have other problems anyway) for
non-trivial reasons, user, mount, and pid namespaces. We can fail to
attach to a pid namespace if it is not our current active pid namespace
or a descendant of it. We can fail to attach to a user namespace because
we are multi-threaded or because our current mount namespace shares
filesystem state with other tasks, or because we're trying to setns()
to the same user namespace, i.e. the target task has the same user
namespace as we do. We can fail to attach to a mount namespace because
it shares filesystem state with other tasks or because we fail to lookup
the new root for the new mount namespace. In most non-pathological
scenarios these issues can be somewhat mitigated. But there are cases where
we're half-attached to some namespace and failing to attach to another one.
I've talked about some of these problem during the hallway track (something
only the pre-COVID-19 generation will remember) of Plumbers in Los Angeles
in 2018(?). Even if all these issues could be avoided with super careful
userspace coding it would be nicer to have this done in-kernel. Pidfds seem
to lend themselves nicely for this.
The other neat thing about this is that setns() becomes an actual
counterpart to the namespace bits of unshare().
Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
Reviewed-by: Serge Hallyn <serge@hallyn.com>
Cc: Eric W. Biederman <ebiederm@xmission.com>
Cc: Serge Hallyn <serge@hallyn.com>
Cc: Jann Horn <jannh@google.com>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Aleksa Sarai <cyphar@cyphar.com>
Link: https://lore.kernel.org/r/20200505140432.181565-3-christian.brauner@ubuntu.com
2020-05-05 17:04:31 +03:00
if ( flags & CLONE_NEWUSER )
2020-05-05 17:04:30 +03:00
nsset - > cred = prepare_creds ( ) ;
else
nsset - > cred = current_cred ( ) ;
if ( ! nsset - > cred )
goto out ;
nsproxy: attach to namespaces via pidfds
For quite a while we have been thinking about using pidfds to attach to
namespaces. This patchset has existed for about a year already but we've
wanted to wait to see how the general api would be received and adopted.
Now that more and more programs in userspace have started using pidfds
for process management it's time to send this one out.
This patch makes it possible to use pidfds to attach to the namespaces
of another process, i.e. they can be passed as the first argument to the
setns() syscall. When only a single namespace type is specified the
semantics are equivalent to passing an nsfd. That means
setns(nsfd, CLONE_NEWNET) equals setns(pidfd, CLONE_NEWNET). However,
when a pidfd is passed, multiple namespace flags can be specified in the
second setns() argument and setns() will attach the caller to all the
specified namespaces all at once or to none of them. Specifying 0 is not
valid together with a pidfd.
Here are just two obvious examples:
setns(pidfd, CLONE_NEWPID | CLONE_NEWNS | CLONE_NEWNET);
setns(pidfd, CLONE_NEWUSER);
Allowing to also attach subsets of namespaces supports various use-cases
where callers setns to a subset of namespaces to retain privilege, perform
an action and then re-attach another subset of namespaces.
If the need arises, as Eric suggested, we can extend this patchset to
assume even more context than just attaching all namespaces. His suggestion
specifically was about assuming the process' root directory when
setns(pidfd, 0) or setns(pidfd, SETNS_PIDFD) is specified. For now, just
keep it flexible in terms of supporting subsets of namespaces but let's
wait until we have users asking for even more context to be assumed. At
that point we can add an extension.
The obvious example where this is useful is a standard container
manager interacting with a running container: pushing and pulling files
or directories, injecting mounts, attaching/execing any kind of process,
managing network devices all these operations require attaching to all
or at least multiple namespaces at the same time. Given that nowadays
most containers are spawned with all namespaces enabled we're currently
looking at at least 14 syscalls, 7 to open the /proc/<pid>/ns/<ns>
nsfds, another 7 to actually perform the namespace switch. With time
namespaces we're looking at about 16 syscalls.
(We could amortize the first 7 or 8 syscalls for opening the nsfds by
stashing them in each container's monitor process but that would mean
we need to send around those file descriptors through unix sockets
everytime we want to interact with the container or keep on-disk
state. Even in scenarios where a caller wants to join a particular
namespace in a particular order callers still profit from batching
other namespaces. That mostly applies to the user namespace but
all container runtimes I found join the user namespace first no matter
if it privileges or deprivileges the container similar to how unshare
behaves.)
With pidfds this becomes a single syscall no matter how many namespaces
are supposed to be attached to.
A decently designed, large-scale container manager usually isn't the
parent of any of the containers it spawns so the containers don't die
when it crashes or needs to update or reinitialize. This means that
for the manager to interact with containers through pids is inherently
racy especially on systems where the maximum pid number is not
significicantly bumped. This is even more problematic since we often spawn
and manage thousands or ten-thousands of containers. Interacting with a
container through a pid thus can become risky quite quickly. Especially
since we allow for an administrator to enable advanced features such as
syscall interception where we're performing syscalls in lieu of the
container. In all of those cases we use pidfds if they are available and
we pass them around as stable references. Using them to setns() to the
target process' namespaces is as reliable as using nsfds. Either the
target process is already dead and we get ESRCH or we manage to attach
to its namespaces but we can't accidently attach to another process'
namespaces. So pidfds lend themselves to be used with this api.
The other main advantage is that with this change the pidfd becomes the
only relevant token for most container interactions and it's the only
token we need to create and send around.
Apart from significiantly reducing the number of syscalls from double
digit to single digit which is a decent reason post-spectre/meltdown
this also allows to switch to a set of namespaces atomically, i.e.
either attaching to all the specified namespaces succeeds or we fail. If
we fail we haven't changed a single namespace. There are currently three
namespaces that can fail (other than for ENOMEM which really is not
very interesting since we then have other problems anyway) for
non-trivial reasons, user, mount, and pid namespaces. We can fail to
attach to a pid namespace if it is not our current active pid namespace
or a descendant of it. We can fail to attach to a user namespace because
we are multi-threaded or because our current mount namespace shares
filesystem state with other tasks, or because we're trying to setns()
to the same user namespace, i.e. the target task has the same user
namespace as we do. We can fail to attach to a mount namespace because
it shares filesystem state with other tasks or because we fail to lookup
the new root for the new mount namespace. In most non-pathological
scenarios these issues can be somewhat mitigated. But there are cases where
we're half-attached to some namespace and failing to attach to another one.
I've talked about some of these problem during the hallway track (something
only the pre-COVID-19 generation will remember) of Plumbers in Los Angeles
in 2018(?). Even if all these issues could be avoided with super careful
userspace coding it would be nicer to have this done in-kernel. Pidfds seem
to lend themselves nicely for this.
The other neat thing about this is that setns() becomes an actual
counterpart to the namespace bits of unshare().
Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
Reviewed-by: Serge Hallyn <serge@hallyn.com>
Cc: Eric W. Biederman <ebiederm@xmission.com>
Cc: Serge Hallyn <serge@hallyn.com>
Cc: Jann Horn <jannh@google.com>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Aleksa Sarai <cyphar@cyphar.com>
Link: https://lore.kernel.org/r/20200505140432.181565-3-christian.brauner@ubuntu.com
2020-05-05 17:04:31 +03:00
/* Only create a temporary copy of fs_struct if we really need to. */
if ( flags = = CLONE_NEWNS ) {
2020-05-05 17:04:30 +03:00
nsset - > fs = me - > fs ;
nsproxy: attach to namespaces via pidfds
For quite a while we have been thinking about using pidfds to attach to
namespaces. This patchset has existed for about a year already but we've
wanted to wait to see how the general api would be received and adopted.
Now that more and more programs in userspace have started using pidfds
for process management it's time to send this one out.
This patch makes it possible to use pidfds to attach to the namespaces
of another process, i.e. they can be passed as the first argument to the
setns() syscall. When only a single namespace type is specified the
semantics are equivalent to passing an nsfd. That means
setns(nsfd, CLONE_NEWNET) equals setns(pidfd, CLONE_NEWNET). However,
when a pidfd is passed, multiple namespace flags can be specified in the
second setns() argument and setns() will attach the caller to all the
specified namespaces all at once or to none of them. Specifying 0 is not
valid together with a pidfd.
Here are just two obvious examples:
setns(pidfd, CLONE_NEWPID | CLONE_NEWNS | CLONE_NEWNET);
setns(pidfd, CLONE_NEWUSER);
Allowing to also attach subsets of namespaces supports various use-cases
where callers setns to a subset of namespaces to retain privilege, perform
an action and then re-attach another subset of namespaces.
If the need arises, as Eric suggested, we can extend this patchset to
assume even more context than just attaching all namespaces. His suggestion
specifically was about assuming the process' root directory when
setns(pidfd, 0) or setns(pidfd, SETNS_PIDFD) is specified. For now, just
keep it flexible in terms of supporting subsets of namespaces but let's
wait until we have users asking for even more context to be assumed. At
that point we can add an extension.
The obvious example where this is useful is a standard container
manager interacting with a running container: pushing and pulling files
or directories, injecting mounts, attaching/execing any kind of process,
managing network devices all these operations require attaching to all
or at least multiple namespaces at the same time. Given that nowadays
most containers are spawned with all namespaces enabled we're currently
looking at at least 14 syscalls, 7 to open the /proc/<pid>/ns/<ns>
nsfds, another 7 to actually perform the namespace switch. With time
namespaces we're looking at about 16 syscalls.
(We could amortize the first 7 or 8 syscalls for opening the nsfds by
stashing them in each container's monitor process but that would mean
we need to send around those file descriptors through unix sockets
everytime we want to interact with the container or keep on-disk
state. Even in scenarios where a caller wants to join a particular
namespace in a particular order callers still profit from batching
other namespaces. That mostly applies to the user namespace but
all container runtimes I found join the user namespace first no matter
if it privileges or deprivileges the container similar to how unshare
behaves.)
With pidfds this becomes a single syscall no matter how many namespaces
are supposed to be attached to.
A decently designed, large-scale container manager usually isn't the
parent of any of the containers it spawns so the containers don't die
when it crashes or needs to update or reinitialize. This means that
for the manager to interact with containers through pids is inherently
racy especially on systems where the maximum pid number is not
significicantly bumped. This is even more problematic since we often spawn
and manage thousands or ten-thousands of containers. Interacting with a
container through a pid thus can become risky quite quickly. Especially
since we allow for an administrator to enable advanced features such as
syscall interception where we're performing syscalls in lieu of the
container. In all of those cases we use pidfds if they are available and
we pass them around as stable references. Using them to setns() to the
target process' namespaces is as reliable as using nsfds. Either the
target process is already dead and we get ESRCH or we manage to attach
to its namespaces but we can't accidently attach to another process'
namespaces. So pidfds lend themselves to be used with this api.
The other main advantage is that with this change the pidfd becomes the
only relevant token for most container interactions and it's the only
token we need to create and send around.
Apart from significiantly reducing the number of syscalls from double
digit to single digit which is a decent reason post-spectre/meltdown
this also allows to switch to a set of namespaces atomically, i.e.
either attaching to all the specified namespaces succeeds or we fail. If
we fail we haven't changed a single namespace. There are currently three
namespaces that can fail (other than for ENOMEM which really is not
very interesting since we then have other problems anyway) for
non-trivial reasons, user, mount, and pid namespaces. We can fail to
attach to a pid namespace if it is not our current active pid namespace
or a descendant of it. We can fail to attach to a user namespace because
we are multi-threaded or because our current mount namespace shares
filesystem state with other tasks, or because we're trying to setns()
to the same user namespace, i.e. the target task has the same user
namespace as we do. We can fail to attach to a mount namespace because
it shares filesystem state with other tasks or because we fail to lookup
the new root for the new mount namespace. In most non-pathological
scenarios these issues can be somewhat mitigated. But there are cases where
we're half-attached to some namespace and failing to attach to another one.
I've talked about some of these problem during the hallway track (something
only the pre-COVID-19 generation will remember) of Plumbers in Los Angeles
in 2018(?). Even if all these issues could be avoided with super careful
userspace coding it would be nicer to have this done in-kernel. Pidfds seem
to lend themselves nicely for this.
The other neat thing about this is that setns() becomes an actual
counterpart to the namespace bits of unshare().
Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
Reviewed-by: Serge Hallyn <serge@hallyn.com>
Cc: Eric W. Biederman <ebiederm@xmission.com>
Cc: Serge Hallyn <serge@hallyn.com>
Cc: Jann Horn <jannh@google.com>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Aleksa Sarai <cyphar@cyphar.com>
Link: https://lore.kernel.org/r/20200505140432.181565-3-christian.brauner@ubuntu.com
2020-05-05 17:04:31 +03:00
} else if ( flags & CLONE_NEWNS ) {
nsset - > fs = copy_fs_struct ( me - > fs ) ;
if ( ! nsset - > fs )
goto out ;
}
2020-05-05 17:04:30 +03:00
nsproxy: attach to namespaces via pidfds
For quite a while we have been thinking about using pidfds to attach to
namespaces. This patchset has existed for about a year already but we've
wanted to wait to see how the general api would be received and adopted.
Now that more and more programs in userspace have started using pidfds
for process management it's time to send this one out.
This patch makes it possible to use pidfds to attach to the namespaces
of another process, i.e. they can be passed as the first argument to the
setns() syscall. When only a single namespace type is specified the
semantics are equivalent to passing an nsfd. That means
setns(nsfd, CLONE_NEWNET) equals setns(pidfd, CLONE_NEWNET). However,
when a pidfd is passed, multiple namespace flags can be specified in the
second setns() argument and setns() will attach the caller to all the
specified namespaces all at once or to none of them. Specifying 0 is not
valid together with a pidfd.
Here are just two obvious examples:
setns(pidfd, CLONE_NEWPID | CLONE_NEWNS | CLONE_NEWNET);
setns(pidfd, CLONE_NEWUSER);
Allowing to also attach subsets of namespaces supports various use-cases
where callers setns to a subset of namespaces to retain privilege, perform
an action and then re-attach another subset of namespaces.
If the need arises, as Eric suggested, we can extend this patchset to
assume even more context than just attaching all namespaces. His suggestion
specifically was about assuming the process' root directory when
setns(pidfd, 0) or setns(pidfd, SETNS_PIDFD) is specified. For now, just
keep it flexible in terms of supporting subsets of namespaces but let's
wait until we have users asking for even more context to be assumed. At
that point we can add an extension.
The obvious example where this is useful is a standard container
manager interacting with a running container: pushing and pulling files
or directories, injecting mounts, attaching/execing any kind of process,
managing network devices all these operations require attaching to all
or at least multiple namespaces at the same time. Given that nowadays
most containers are spawned with all namespaces enabled we're currently
looking at at least 14 syscalls, 7 to open the /proc/<pid>/ns/<ns>
nsfds, another 7 to actually perform the namespace switch. With time
namespaces we're looking at about 16 syscalls.
(We could amortize the first 7 or 8 syscalls for opening the nsfds by
stashing them in each container's monitor process but that would mean
we need to send around those file descriptors through unix sockets
everytime we want to interact with the container or keep on-disk
state. Even in scenarios where a caller wants to join a particular
namespace in a particular order callers still profit from batching
other namespaces. That mostly applies to the user namespace but
all container runtimes I found join the user namespace first no matter
if it privileges or deprivileges the container similar to how unshare
behaves.)
With pidfds this becomes a single syscall no matter how many namespaces
are supposed to be attached to.
A decently designed, large-scale container manager usually isn't the
parent of any of the containers it spawns so the containers don't die
when it crashes or needs to update or reinitialize. This means that
for the manager to interact with containers through pids is inherently
racy especially on systems where the maximum pid number is not
significicantly bumped. This is even more problematic since we often spawn
and manage thousands or ten-thousands of containers. Interacting with a
container through a pid thus can become risky quite quickly. Especially
since we allow for an administrator to enable advanced features such as
syscall interception where we're performing syscalls in lieu of the
container. In all of those cases we use pidfds if they are available and
we pass them around as stable references. Using them to setns() to the
target process' namespaces is as reliable as using nsfds. Either the
target process is already dead and we get ESRCH or we manage to attach
to its namespaces but we can't accidently attach to another process'
namespaces. So pidfds lend themselves to be used with this api.
The other main advantage is that with this change the pidfd becomes the
only relevant token for most container interactions and it's the only
token we need to create and send around.
Apart from significiantly reducing the number of syscalls from double
digit to single digit which is a decent reason post-spectre/meltdown
this also allows to switch to a set of namespaces atomically, i.e.
either attaching to all the specified namespaces succeeds or we fail. If
we fail we haven't changed a single namespace. There are currently three
namespaces that can fail (other than for ENOMEM which really is not
very interesting since we then have other problems anyway) for
non-trivial reasons, user, mount, and pid namespaces. We can fail to
attach to a pid namespace if it is not our current active pid namespace
or a descendant of it. We can fail to attach to a user namespace because
we are multi-threaded or because our current mount namespace shares
filesystem state with other tasks, or because we're trying to setns()
to the same user namespace, i.e. the target task has the same user
namespace as we do. We can fail to attach to a mount namespace because
it shares filesystem state with other tasks or because we fail to lookup
the new root for the new mount namespace. In most non-pathological
scenarios these issues can be somewhat mitigated. But there are cases where
we're half-attached to some namespace and failing to attach to another one.
I've talked about some of these problem during the hallway track (something
only the pre-COVID-19 generation will remember) of Plumbers in Los Angeles
in 2018(?). Even if all these issues could be avoided with super careful
userspace coding it would be nicer to have this done in-kernel. Pidfds seem
to lend themselves nicely for this.
The other neat thing about this is that setns() becomes an actual
counterpart to the namespace bits of unshare().
Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
Reviewed-by: Serge Hallyn <serge@hallyn.com>
Cc: Eric W. Biederman <ebiederm@xmission.com>
Cc: Serge Hallyn <serge@hallyn.com>
Cc: Jann Horn <jannh@google.com>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Aleksa Sarai <cyphar@cyphar.com>
Link: https://lore.kernel.org/r/20200505140432.181565-3-christian.brauner@ubuntu.com
2020-05-05 17:04:31 +03:00
nsset - > flags = flags ;
2020-05-05 17:04:30 +03:00
return 0 ;
out :
put_nsset ( nsset ) ;
return - ENOMEM ;
}
nsproxy: attach to namespaces via pidfds
For quite a while we have been thinking about using pidfds to attach to
namespaces. This patchset has existed for about a year already but we've
wanted to wait to see how the general api would be received and adopted.
Now that more and more programs in userspace have started using pidfds
for process management it's time to send this one out.
This patch makes it possible to use pidfds to attach to the namespaces
of another process, i.e. they can be passed as the first argument to the
setns() syscall. When only a single namespace type is specified the
semantics are equivalent to passing an nsfd. That means
setns(nsfd, CLONE_NEWNET) equals setns(pidfd, CLONE_NEWNET). However,
when a pidfd is passed, multiple namespace flags can be specified in the
second setns() argument and setns() will attach the caller to all the
specified namespaces all at once or to none of them. Specifying 0 is not
valid together with a pidfd.
Here are just two obvious examples:
setns(pidfd, CLONE_NEWPID | CLONE_NEWNS | CLONE_NEWNET);
setns(pidfd, CLONE_NEWUSER);
Allowing to also attach subsets of namespaces supports various use-cases
where callers setns to a subset of namespaces to retain privilege, perform
an action and then re-attach another subset of namespaces.
If the need arises, as Eric suggested, we can extend this patchset to
assume even more context than just attaching all namespaces. His suggestion
specifically was about assuming the process' root directory when
setns(pidfd, 0) or setns(pidfd, SETNS_PIDFD) is specified. For now, just
keep it flexible in terms of supporting subsets of namespaces but let's
wait until we have users asking for even more context to be assumed. At
that point we can add an extension.
The obvious example where this is useful is a standard container
manager interacting with a running container: pushing and pulling files
or directories, injecting mounts, attaching/execing any kind of process,
managing network devices all these operations require attaching to all
or at least multiple namespaces at the same time. Given that nowadays
most containers are spawned with all namespaces enabled we're currently
looking at at least 14 syscalls, 7 to open the /proc/<pid>/ns/<ns>
nsfds, another 7 to actually perform the namespace switch. With time
namespaces we're looking at about 16 syscalls.
(We could amortize the first 7 or 8 syscalls for opening the nsfds by
stashing them in each container's monitor process but that would mean
we need to send around those file descriptors through unix sockets
everytime we want to interact with the container or keep on-disk
state. Even in scenarios where a caller wants to join a particular
namespace in a particular order callers still profit from batching
other namespaces. That mostly applies to the user namespace but
all container runtimes I found join the user namespace first no matter
if it privileges or deprivileges the container similar to how unshare
behaves.)
With pidfds this becomes a single syscall no matter how many namespaces
are supposed to be attached to.
A decently designed, large-scale container manager usually isn't the
parent of any of the containers it spawns so the containers don't die
when it crashes or needs to update or reinitialize. This means that
for the manager to interact with containers through pids is inherently
racy especially on systems where the maximum pid number is not
significicantly bumped. This is even more problematic since we often spawn
and manage thousands or ten-thousands of containers. Interacting with a
container through a pid thus can become risky quite quickly. Especially
since we allow for an administrator to enable advanced features such as
syscall interception where we're performing syscalls in lieu of the
container. In all of those cases we use pidfds if they are available and
we pass them around as stable references. Using them to setns() to the
target process' namespaces is as reliable as using nsfds. Either the
target process is already dead and we get ESRCH or we manage to attach
to its namespaces but we can't accidently attach to another process'
namespaces. So pidfds lend themselves to be used with this api.
The other main advantage is that with this change the pidfd becomes the
only relevant token for most container interactions and it's the only
token we need to create and send around.
Apart from significiantly reducing the number of syscalls from double
digit to single digit which is a decent reason post-spectre/meltdown
this also allows to switch to a set of namespaces atomically, i.e.
either attaching to all the specified namespaces succeeds or we fail. If
we fail we haven't changed a single namespace. There are currently three
namespaces that can fail (other than for ENOMEM which really is not
very interesting since we then have other problems anyway) for
non-trivial reasons, user, mount, and pid namespaces. We can fail to
attach to a pid namespace if it is not our current active pid namespace
or a descendant of it. We can fail to attach to a user namespace because
we are multi-threaded or because our current mount namespace shares
filesystem state with other tasks, or because we're trying to setns()
to the same user namespace, i.e. the target task has the same user
namespace as we do. We can fail to attach to a mount namespace because
it shares filesystem state with other tasks or because we fail to lookup
the new root for the new mount namespace. In most non-pathological
scenarios these issues can be somewhat mitigated. But there are cases where
we're half-attached to some namespace and failing to attach to another one.
I've talked about some of these problem during the hallway track (something
only the pre-COVID-19 generation will remember) of Plumbers in Los Angeles
in 2018(?). Even if all these issues could be avoided with super careful
userspace coding it would be nicer to have this done in-kernel. Pidfds seem
to lend themselves nicely for this.
The other neat thing about this is that setns() becomes an actual
counterpart to the namespace bits of unshare().
Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
Reviewed-by: Serge Hallyn <serge@hallyn.com>
Cc: Eric W. Biederman <ebiederm@xmission.com>
Cc: Serge Hallyn <serge@hallyn.com>
Cc: Jann Horn <jannh@google.com>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Aleksa Sarai <cyphar@cyphar.com>
Link: https://lore.kernel.org/r/20200505140432.181565-3-christian.brauner@ubuntu.com
2020-05-05 17:04:31 +03:00
static inline int validate_ns ( struct nsset * nsset , struct ns_common * ns )
{
return ns - > ops - > install ( nsset , ns ) ;
}
/*
* This is the inverse operation to unshare ( ) .
* Ordering is equivalent to the standard ordering used everywhere else
* during unshare and process creation . The switch to the new set of
* namespaces occurs at the point of no return after installation of
* all requested namespaces was successful in commit_nsset ( ) .
*/
static int validate_nsset ( struct nsset * nsset , struct pid * pid )
{
int ret = 0 ;
unsigned flags = nsset - > flags ;
struct user_namespace * user_ns = NULL ;
struct pid_namespace * pid_ns = NULL ;
struct nsproxy * nsp ;
struct task_struct * tsk ;
/* Take a "snapshot" of the target task's namespaces. */
rcu_read_lock ( ) ;
tsk = pid_task ( pid , PIDTYPE_PID ) ;
if ( ! tsk ) {
rcu_read_unlock ( ) ;
return - ESRCH ;
}
if ( ! ptrace_may_access ( tsk , PTRACE_MODE_READ_REALCREDS ) ) {
rcu_read_unlock ( ) ;
return - EPERM ;
}
task_lock ( tsk ) ;
nsp = tsk - > nsproxy ;
if ( nsp )
get_nsproxy ( nsp ) ;
task_unlock ( tsk ) ;
if ( ! nsp ) {
rcu_read_unlock ( ) ;
return - ESRCH ;
}
# ifdef CONFIG_PID_NS
if ( flags & CLONE_NEWPID ) {
pid_ns = task_active_pid_ns ( tsk ) ;
if ( unlikely ( ! pid_ns ) ) {
rcu_read_unlock ( ) ;
ret = - ESRCH ;
goto out ;
}
get_pid_ns ( pid_ns ) ;
}
# endif
# ifdef CONFIG_USER_NS
if ( flags & CLONE_NEWUSER )
user_ns = get_user_ns ( __task_cred ( tsk ) - > user_ns ) ;
# endif
rcu_read_unlock ( ) ;
/*
* Install requested namespaces . The caller will have
* verified earlier that the requested namespaces are
* supported on this kernel . We don ' t report errors here
* if a namespace is requested that isn ' t supported .
*/
# ifdef CONFIG_USER_NS
if ( flags & CLONE_NEWUSER ) {
ret = validate_ns ( nsset , & user_ns - > ns ) ;
if ( ret )
goto out ;
}
# endif
if ( flags & CLONE_NEWNS ) {
ret = validate_ns ( nsset , from_mnt_ns ( nsp - > mnt_ns ) ) ;
if ( ret )
goto out ;
}
# ifdef CONFIG_UTS_NS
if ( flags & CLONE_NEWUTS ) {
ret = validate_ns ( nsset , & nsp - > uts_ns - > ns ) ;
if ( ret )
goto out ;
}
# endif
# ifdef CONFIG_IPC_NS
if ( flags & CLONE_NEWIPC ) {
ret = validate_ns ( nsset , & nsp - > ipc_ns - > ns ) ;
if ( ret )
goto out ;
}
# endif
# ifdef CONFIG_PID_NS
if ( flags & CLONE_NEWPID ) {
ret = validate_ns ( nsset , & pid_ns - > ns ) ;
if ( ret )
goto out ;
}
# endif
# ifdef CONFIG_CGROUPS
if ( flags & CLONE_NEWCGROUP ) {
ret = validate_ns ( nsset , & nsp - > cgroup_ns - > ns ) ;
if ( ret )
goto out ;
}
# endif
# ifdef CONFIG_NET_NS
if ( flags & CLONE_NEWNET ) {
ret = validate_ns ( nsset , & nsp - > net_ns - > ns ) ;
if ( ret )
goto out ;
}
# endif
2020-07-06 18:49:11 +03:00
# ifdef CONFIG_TIME_NS
if ( flags & CLONE_NEWTIME ) {
ret = validate_ns ( nsset , & nsp - > time_ns - > ns ) ;
if ( ret )
goto out ;
}
# endif
nsproxy: attach to namespaces via pidfds
For quite a while we have been thinking about using pidfds to attach to
namespaces. This patchset has existed for about a year already but we've
wanted to wait to see how the general api would be received and adopted.
Now that more and more programs in userspace have started using pidfds
for process management it's time to send this one out.
This patch makes it possible to use pidfds to attach to the namespaces
of another process, i.e. they can be passed as the first argument to the
setns() syscall. When only a single namespace type is specified the
semantics are equivalent to passing an nsfd. That means
setns(nsfd, CLONE_NEWNET) equals setns(pidfd, CLONE_NEWNET). However,
when a pidfd is passed, multiple namespace flags can be specified in the
second setns() argument and setns() will attach the caller to all the
specified namespaces all at once or to none of them. Specifying 0 is not
valid together with a pidfd.
Here are just two obvious examples:
setns(pidfd, CLONE_NEWPID | CLONE_NEWNS | CLONE_NEWNET);
setns(pidfd, CLONE_NEWUSER);
Allowing to also attach subsets of namespaces supports various use-cases
where callers setns to a subset of namespaces to retain privilege, perform
an action and then re-attach another subset of namespaces.
If the need arises, as Eric suggested, we can extend this patchset to
assume even more context than just attaching all namespaces. His suggestion
specifically was about assuming the process' root directory when
setns(pidfd, 0) or setns(pidfd, SETNS_PIDFD) is specified. For now, just
keep it flexible in terms of supporting subsets of namespaces but let's
wait until we have users asking for even more context to be assumed. At
that point we can add an extension.
The obvious example where this is useful is a standard container
manager interacting with a running container: pushing and pulling files
or directories, injecting mounts, attaching/execing any kind of process,
managing network devices all these operations require attaching to all
or at least multiple namespaces at the same time. Given that nowadays
most containers are spawned with all namespaces enabled we're currently
looking at at least 14 syscalls, 7 to open the /proc/<pid>/ns/<ns>
nsfds, another 7 to actually perform the namespace switch. With time
namespaces we're looking at about 16 syscalls.
(We could amortize the first 7 or 8 syscalls for opening the nsfds by
stashing them in each container's monitor process but that would mean
we need to send around those file descriptors through unix sockets
everytime we want to interact with the container or keep on-disk
state. Even in scenarios where a caller wants to join a particular
namespace in a particular order callers still profit from batching
other namespaces. That mostly applies to the user namespace but
all container runtimes I found join the user namespace first no matter
if it privileges or deprivileges the container similar to how unshare
behaves.)
With pidfds this becomes a single syscall no matter how many namespaces
are supposed to be attached to.
A decently designed, large-scale container manager usually isn't the
parent of any of the containers it spawns so the containers don't die
when it crashes or needs to update or reinitialize. This means that
for the manager to interact with containers through pids is inherently
racy especially on systems where the maximum pid number is not
significicantly bumped. This is even more problematic since we often spawn
and manage thousands or ten-thousands of containers. Interacting with a
container through a pid thus can become risky quite quickly. Especially
since we allow for an administrator to enable advanced features such as
syscall interception where we're performing syscalls in lieu of the
container. In all of those cases we use pidfds if they are available and
we pass them around as stable references. Using them to setns() to the
target process' namespaces is as reliable as using nsfds. Either the
target process is already dead and we get ESRCH or we manage to attach
to its namespaces but we can't accidently attach to another process'
namespaces. So pidfds lend themselves to be used with this api.
The other main advantage is that with this change the pidfd becomes the
only relevant token for most container interactions and it's the only
token we need to create and send around.
Apart from significiantly reducing the number of syscalls from double
digit to single digit which is a decent reason post-spectre/meltdown
this also allows to switch to a set of namespaces atomically, i.e.
either attaching to all the specified namespaces succeeds or we fail. If
we fail we haven't changed a single namespace. There are currently three
namespaces that can fail (other than for ENOMEM which really is not
very interesting since we then have other problems anyway) for
non-trivial reasons, user, mount, and pid namespaces. We can fail to
attach to a pid namespace if it is not our current active pid namespace
or a descendant of it. We can fail to attach to a user namespace because
we are multi-threaded or because our current mount namespace shares
filesystem state with other tasks, or because we're trying to setns()
to the same user namespace, i.e. the target task has the same user
namespace as we do. We can fail to attach to a mount namespace because
it shares filesystem state with other tasks or because we fail to lookup
the new root for the new mount namespace. In most non-pathological
scenarios these issues can be somewhat mitigated. But there are cases where
we're half-attached to some namespace and failing to attach to another one.
I've talked about some of these problem during the hallway track (something
only the pre-COVID-19 generation will remember) of Plumbers in Los Angeles
in 2018(?). Even if all these issues could be avoided with super careful
userspace coding it would be nicer to have this done in-kernel. Pidfds seem
to lend themselves nicely for this.
The other neat thing about this is that setns() becomes an actual
counterpart to the namespace bits of unshare().
Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
Reviewed-by: Serge Hallyn <serge@hallyn.com>
Cc: Eric W. Biederman <ebiederm@xmission.com>
Cc: Serge Hallyn <serge@hallyn.com>
Cc: Jann Horn <jannh@google.com>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Aleksa Sarai <cyphar@cyphar.com>
Link: https://lore.kernel.org/r/20200505140432.181565-3-christian.brauner@ubuntu.com
2020-05-05 17:04:31 +03:00
out :
if ( pid_ns )
put_pid_ns ( pid_ns ) ;
if ( nsp )
put_nsproxy ( nsp ) ;
put_user_ns ( user_ns ) ;
return ret ;
}
2020-05-05 17:04:30 +03:00
/*
* This is the point of no return . There are just a few namespaces
* that do some actual work here and it ' s sufficiently minimal that
* a separate ns_common operation seems unnecessary for now .
* Unshare is doing the same thing . If we ' ll end up needing to do
* more in a given namespace or a helper here is ultimately not
* exported anymore a simple commit handler for each namespace
* should be added to ns_common .
*/
static void commit_nsset ( struct nsset * nsset )
{
unsigned flags = nsset - > flags ;
struct task_struct * me = current ;
# ifdef CONFIG_USER_NS
if ( flags & CLONE_NEWUSER ) {
/* transfer ownership */
commit_creds ( nsset_cred ( nsset ) ) ;
nsset - > cred = NULL ;
}
# endif
nsproxy: attach to namespaces via pidfds
For quite a while we have been thinking about using pidfds to attach to
namespaces. This patchset has existed for about a year already but we've
wanted to wait to see how the general api would be received and adopted.
Now that more and more programs in userspace have started using pidfds
for process management it's time to send this one out.
This patch makes it possible to use pidfds to attach to the namespaces
of another process, i.e. they can be passed as the first argument to the
setns() syscall. When only a single namespace type is specified the
semantics are equivalent to passing an nsfd. That means
setns(nsfd, CLONE_NEWNET) equals setns(pidfd, CLONE_NEWNET). However,
when a pidfd is passed, multiple namespace flags can be specified in the
second setns() argument and setns() will attach the caller to all the
specified namespaces all at once or to none of them. Specifying 0 is not
valid together with a pidfd.
Here are just two obvious examples:
setns(pidfd, CLONE_NEWPID | CLONE_NEWNS | CLONE_NEWNET);
setns(pidfd, CLONE_NEWUSER);
Allowing to also attach subsets of namespaces supports various use-cases
where callers setns to a subset of namespaces to retain privilege, perform
an action and then re-attach another subset of namespaces.
If the need arises, as Eric suggested, we can extend this patchset to
assume even more context than just attaching all namespaces. His suggestion
specifically was about assuming the process' root directory when
setns(pidfd, 0) or setns(pidfd, SETNS_PIDFD) is specified. For now, just
keep it flexible in terms of supporting subsets of namespaces but let's
wait until we have users asking for even more context to be assumed. At
that point we can add an extension.
The obvious example where this is useful is a standard container
manager interacting with a running container: pushing and pulling files
or directories, injecting mounts, attaching/execing any kind of process,
managing network devices all these operations require attaching to all
or at least multiple namespaces at the same time. Given that nowadays
most containers are spawned with all namespaces enabled we're currently
looking at at least 14 syscalls, 7 to open the /proc/<pid>/ns/<ns>
nsfds, another 7 to actually perform the namespace switch. With time
namespaces we're looking at about 16 syscalls.
(We could amortize the first 7 or 8 syscalls for opening the nsfds by
stashing them in each container's monitor process but that would mean
we need to send around those file descriptors through unix sockets
everytime we want to interact with the container or keep on-disk
state. Even in scenarios where a caller wants to join a particular
namespace in a particular order callers still profit from batching
other namespaces. That mostly applies to the user namespace but
all container runtimes I found join the user namespace first no matter
if it privileges or deprivileges the container similar to how unshare
behaves.)
With pidfds this becomes a single syscall no matter how many namespaces
are supposed to be attached to.
A decently designed, large-scale container manager usually isn't the
parent of any of the containers it spawns so the containers don't die
when it crashes or needs to update or reinitialize. This means that
for the manager to interact with containers through pids is inherently
racy especially on systems where the maximum pid number is not
significicantly bumped. This is even more problematic since we often spawn
and manage thousands or ten-thousands of containers. Interacting with a
container through a pid thus can become risky quite quickly. Especially
since we allow for an administrator to enable advanced features such as
syscall interception where we're performing syscalls in lieu of the
container. In all of those cases we use pidfds if they are available and
we pass them around as stable references. Using them to setns() to the
target process' namespaces is as reliable as using nsfds. Either the
target process is already dead and we get ESRCH or we manage to attach
to its namespaces but we can't accidently attach to another process'
namespaces. So pidfds lend themselves to be used with this api.
The other main advantage is that with this change the pidfd becomes the
only relevant token for most container interactions and it's the only
token we need to create and send around.
Apart from significiantly reducing the number of syscalls from double
digit to single digit which is a decent reason post-spectre/meltdown
this also allows to switch to a set of namespaces atomically, i.e.
either attaching to all the specified namespaces succeeds or we fail. If
we fail we haven't changed a single namespace. There are currently three
namespaces that can fail (other than for ENOMEM which really is not
very interesting since we then have other problems anyway) for
non-trivial reasons, user, mount, and pid namespaces. We can fail to
attach to a pid namespace if it is not our current active pid namespace
or a descendant of it. We can fail to attach to a user namespace because
we are multi-threaded or because our current mount namespace shares
filesystem state with other tasks, or because we're trying to setns()
to the same user namespace, i.e. the target task has the same user
namespace as we do. We can fail to attach to a mount namespace because
it shares filesystem state with other tasks or because we fail to lookup
the new root for the new mount namespace. In most non-pathological
scenarios these issues can be somewhat mitigated. But there are cases where
we're half-attached to some namespace and failing to attach to another one.
I've talked about some of these problem during the hallway track (something
only the pre-COVID-19 generation will remember) of Plumbers in Los Angeles
in 2018(?). Even if all these issues could be avoided with super careful
userspace coding it would be nicer to have this done in-kernel. Pidfds seem
to lend themselves nicely for this.
The other neat thing about this is that setns() becomes an actual
counterpart to the namespace bits of unshare().
Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
Reviewed-by: Serge Hallyn <serge@hallyn.com>
Cc: Eric W. Biederman <ebiederm@xmission.com>
Cc: Serge Hallyn <serge@hallyn.com>
Cc: Jann Horn <jannh@google.com>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Aleksa Sarai <cyphar@cyphar.com>
Link: https://lore.kernel.org/r/20200505140432.181565-3-christian.brauner@ubuntu.com
2020-05-05 17:04:31 +03:00
/* We only need to commit if we have used a temporary fs_struct. */
if ( ( flags & CLONE_NEWNS ) & & ( flags & ~ CLONE_NEWNS ) ) {
set_fs_root ( me - > fs , & nsset - > fs - > root ) ;
set_fs_pwd ( me - > fs , & nsset - > fs - > pwd ) ;
}
2020-05-05 17:04:30 +03:00
# ifdef CONFIG_IPC_NS
if ( flags & CLONE_NEWIPC )
exit_sem ( me ) ;
# endif
2020-07-06 18:49:11 +03:00
# ifdef CONFIG_TIME_NS
if ( flags & CLONE_NEWTIME )
timens_commit ( me , nsset - > nsproxy - > time_ns ) ;
# endif
2020-05-05 17:04:30 +03:00
/* transfer ownership */
switch_task_namespaces ( me , nsset - > nsproxy ) ;
nsset - > nsproxy = NULL ;
}
nsproxy: attach to namespaces via pidfds
For quite a while we have been thinking about using pidfds to attach to
namespaces. This patchset has existed for about a year already but we've
wanted to wait to see how the general api would be received and adopted.
Now that more and more programs in userspace have started using pidfds
for process management it's time to send this one out.
This patch makes it possible to use pidfds to attach to the namespaces
of another process, i.e. they can be passed as the first argument to the
setns() syscall. When only a single namespace type is specified the
semantics are equivalent to passing an nsfd. That means
setns(nsfd, CLONE_NEWNET) equals setns(pidfd, CLONE_NEWNET). However,
when a pidfd is passed, multiple namespace flags can be specified in the
second setns() argument and setns() will attach the caller to all the
specified namespaces all at once or to none of them. Specifying 0 is not
valid together with a pidfd.
Here are just two obvious examples:
setns(pidfd, CLONE_NEWPID | CLONE_NEWNS | CLONE_NEWNET);
setns(pidfd, CLONE_NEWUSER);
Allowing to also attach subsets of namespaces supports various use-cases
where callers setns to a subset of namespaces to retain privilege, perform
an action and then re-attach another subset of namespaces.
If the need arises, as Eric suggested, we can extend this patchset to
assume even more context than just attaching all namespaces. His suggestion
specifically was about assuming the process' root directory when
setns(pidfd, 0) or setns(pidfd, SETNS_PIDFD) is specified. For now, just
keep it flexible in terms of supporting subsets of namespaces but let's
wait until we have users asking for even more context to be assumed. At
that point we can add an extension.
The obvious example where this is useful is a standard container
manager interacting with a running container: pushing and pulling files
or directories, injecting mounts, attaching/execing any kind of process,
managing network devices all these operations require attaching to all
or at least multiple namespaces at the same time. Given that nowadays
most containers are spawned with all namespaces enabled we're currently
looking at at least 14 syscalls, 7 to open the /proc/<pid>/ns/<ns>
nsfds, another 7 to actually perform the namespace switch. With time
namespaces we're looking at about 16 syscalls.
(We could amortize the first 7 or 8 syscalls for opening the nsfds by
stashing them in each container's monitor process but that would mean
we need to send around those file descriptors through unix sockets
everytime we want to interact with the container or keep on-disk
state. Even in scenarios where a caller wants to join a particular
namespace in a particular order callers still profit from batching
other namespaces. That mostly applies to the user namespace but
all container runtimes I found join the user namespace first no matter
if it privileges or deprivileges the container similar to how unshare
behaves.)
With pidfds this becomes a single syscall no matter how many namespaces
are supposed to be attached to.
A decently designed, large-scale container manager usually isn't the
parent of any of the containers it spawns so the containers don't die
when it crashes or needs to update or reinitialize. This means that
for the manager to interact with containers through pids is inherently
racy especially on systems where the maximum pid number is not
significicantly bumped. This is even more problematic since we often spawn
and manage thousands or ten-thousands of containers. Interacting with a
container through a pid thus can become risky quite quickly. Especially
since we allow for an administrator to enable advanced features such as
syscall interception where we're performing syscalls in lieu of the
container. In all of those cases we use pidfds if they are available and
we pass them around as stable references. Using them to setns() to the
target process' namespaces is as reliable as using nsfds. Either the
target process is already dead and we get ESRCH or we manage to attach
to its namespaces but we can't accidently attach to another process'
namespaces. So pidfds lend themselves to be used with this api.
The other main advantage is that with this change the pidfd becomes the
only relevant token for most container interactions and it's the only
token we need to create and send around.
Apart from significiantly reducing the number of syscalls from double
digit to single digit which is a decent reason post-spectre/meltdown
this also allows to switch to a set of namespaces atomically, i.e.
either attaching to all the specified namespaces succeeds or we fail. If
we fail we haven't changed a single namespace. There are currently three
namespaces that can fail (other than for ENOMEM which really is not
very interesting since we then have other problems anyway) for
non-trivial reasons, user, mount, and pid namespaces. We can fail to
attach to a pid namespace if it is not our current active pid namespace
or a descendant of it. We can fail to attach to a user namespace because
we are multi-threaded or because our current mount namespace shares
filesystem state with other tasks, or because we're trying to setns()
to the same user namespace, i.e. the target task has the same user
namespace as we do. We can fail to attach to a mount namespace because
it shares filesystem state with other tasks or because we fail to lookup
the new root for the new mount namespace. In most non-pathological
scenarios these issues can be somewhat mitigated. But there are cases where
we're half-attached to some namespace and failing to attach to another one.
I've talked about some of these problem during the hallway track (something
only the pre-COVID-19 generation will remember) of Plumbers in Los Angeles
in 2018(?). Even if all these issues could be avoided with super careful
userspace coding it would be nicer to have this done in-kernel. Pidfds seem
to lend themselves nicely for this.
The other neat thing about this is that setns() becomes an actual
counterpart to the namespace bits of unshare().
Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
Reviewed-by: Serge Hallyn <serge@hallyn.com>
Cc: Eric W. Biederman <ebiederm@xmission.com>
Cc: Serge Hallyn <serge@hallyn.com>
Cc: Jann Horn <jannh@google.com>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Aleksa Sarai <cyphar@cyphar.com>
Link: https://lore.kernel.org/r/20200505140432.181565-3-christian.brauner@ubuntu.com
2020-05-05 17:04:31 +03:00
SYSCALL_DEFINE2 ( setns , int , fd , int , flags )
2010-03-08 04:48:52 +03:00
{
struct file * file ;
nsproxy: attach to namespaces via pidfds
For quite a while we have been thinking about using pidfds to attach to
namespaces. This patchset has existed for about a year already but we've
wanted to wait to see how the general api would be received and adopted.
Now that more and more programs in userspace have started using pidfds
for process management it's time to send this one out.
This patch makes it possible to use pidfds to attach to the namespaces
of another process, i.e. they can be passed as the first argument to the
setns() syscall. When only a single namespace type is specified the
semantics are equivalent to passing an nsfd. That means
setns(nsfd, CLONE_NEWNET) equals setns(pidfd, CLONE_NEWNET). However,
when a pidfd is passed, multiple namespace flags can be specified in the
second setns() argument and setns() will attach the caller to all the
specified namespaces all at once or to none of them. Specifying 0 is not
valid together with a pidfd.
Here are just two obvious examples:
setns(pidfd, CLONE_NEWPID | CLONE_NEWNS | CLONE_NEWNET);
setns(pidfd, CLONE_NEWUSER);
Allowing to also attach subsets of namespaces supports various use-cases
where callers setns to a subset of namespaces to retain privilege, perform
an action and then re-attach another subset of namespaces.
If the need arises, as Eric suggested, we can extend this patchset to
assume even more context than just attaching all namespaces. His suggestion
specifically was about assuming the process' root directory when
setns(pidfd, 0) or setns(pidfd, SETNS_PIDFD) is specified. For now, just
keep it flexible in terms of supporting subsets of namespaces but let's
wait until we have users asking for even more context to be assumed. At
that point we can add an extension.
The obvious example where this is useful is a standard container
manager interacting with a running container: pushing and pulling files
or directories, injecting mounts, attaching/execing any kind of process,
managing network devices all these operations require attaching to all
or at least multiple namespaces at the same time. Given that nowadays
most containers are spawned with all namespaces enabled we're currently
looking at at least 14 syscalls, 7 to open the /proc/<pid>/ns/<ns>
nsfds, another 7 to actually perform the namespace switch. With time
namespaces we're looking at about 16 syscalls.
(We could amortize the first 7 or 8 syscalls for opening the nsfds by
stashing them in each container's monitor process but that would mean
we need to send around those file descriptors through unix sockets
everytime we want to interact with the container or keep on-disk
state. Even in scenarios where a caller wants to join a particular
namespace in a particular order callers still profit from batching
other namespaces. That mostly applies to the user namespace but
all container runtimes I found join the user namespace first no matter
if it privileges or deprivileges the container similar to how unshare
behaves.)
With pidfds this becomes a single syscall no matter how many namespaces
are supposed to be attached to.
A decently designed, large-scale container manager usually isn't the
parent of any of the containers it spawns so the containers don't die
when it crashes or needs to update or reinitialize. This means that
for the manager to interact with containers through pids is inherently
racy especially on systems where the maximum pid number is not
significicantly bumped. This is even more problematic since we often spawn
and manage thousands or ten-thousands of containers. Interacting with a
container through a pid thus can become risky quite quickly. Especially
since we allow for an administrator to enable advanced features such as
syscall interception where we're performing syscalls in lieu of the
container. In all of those cases we use pidfds if they are available and
we pass them around as stable references. Using them to setns() to the
target process' namespaces is as reliable as using nsfds. Either the
target process is already dead and we get ESRCH or we manage to attach
to its namespaces but we can't accidently attach to another process'
namespaces. So pidfds lend themselves to be used with this api.
The other main advantage is that with this change the pidfd becomes the
only relevant token for most container interactions and it's the only
token we need to create and send around.
Apart from significiantly reducing the number of syscalls from double
digit to single digit which is a decent reason post-spectre/meltdown
this also allows to switch to a set of namespaces atomically, i.e.
either attaching to all the specified namespaces succeeds or we fail. If
we fail we haven't changed a single namespace. There are currently three
namespaces that can fail (other than for ENOMEM which really is not
very interesting since we then have other problems anyway) for
non-trivial reasons, user, mount, and pid namespaces. We can fail to
attach to a pid namespace if it is not our current active pid namespace
or a descendant of it. We can fail to attach to a user namespace because
we are multi-threaded or because our current mount namespace shares
filesystem state with other tasks, or because we're trying to setns()
to the same user namespace, i.e. the target task has the same user
namespace as we do. We can fail to attach to a mount namespace because
it shares filesystem state with other tasks or because we fail to lookup
the new root for the new mount namespace. In most non-pathological
scenarios these issues can be somewhat mitigated. But there are cases where
we're half-attached to some namespace and failing to attach to another one.
I've talked about some of these problem during the hallway track (something
only the pre-COVID-19 generation will remember) of Plumbers in Los Angeles
in 2018(?). Even if all these issues could be avoided with super careful
userspace coding it would be nicer to have this done in-kernel. Pidfds seem
to lend themselves nicely for this.
The other neat thing about this is that setns() becomes an actual
counterpart to the namespace bits of unshare().
Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
Reviewed-by: Serge Hallyn <serge@hallyn.com>
Cc: Eric W. Biederman <ebiederm@xmission.com>
Cc: Serge Hallyn <serge@hallyn.com>
Cc: Jann Horn <jannh@google.com>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Aleksa Sarai <cyphar@cyphar.com>
Link: https://lore.kernel.org/r/20200505140432.181565-3-christian.brauner@ubuntu.com
2020-05-05 17:04:31 +03:00
struct ns_common * ns = NULL ;
2020-05-05 17:04:30 +03:00
struct nsset nsset = { } ;
nsproxy: attach to namespaces via pidfds
For quite a while we have been thinking about using pidfds to attach to
namespaces. This patchset has existed for about a year already but we've
wanted to wait to see how the general api would be received and adopted.
Now that more and more programs in userspace have started using pidfds
for process management it's time to send this one out.
This patch makes it possible to use pidfds to attach to the namespaces
of another process, i.e. they can be passed as the first argument to the
setns() syscall. When only a single namespace type is specified the
semantics are equivalent to passing an nsfd. That means
setns(nsfd, CLONE_NEWNET) equals setns(pidfd, CLONE_NEWNET). However,
when a pidfd is passed, multiple namespace flags can be specified in the
second setns() argument and setns() will attach the caller to all the
specified namespaces all at once or to none of them. Specifying 0 is not
valid together with a pidfd.
Here are just two obvious examples:
setns(pidfd, CLONE_NEWPID | CLONE_NEWNS | CLONE_NEWNET);
setns(pidfd, CLONE_NEWUSER);
Allowing to also attach subsets of namespaces supports various use-cases
where callers setns to a subset of namespaces to retain privilege, perform
an action and then re-attach another subset of namespaces.
If the need arises, as Eric suggested, we can extend this patchset to
assume even more context than just attaching all namespaces. His suggestion
specifically was about assuming the process' root directory when
setns(pidfd, 0) or setns(pidfd, SETNS_PIDFD) is specified. For now, just
keep it flexible in terms of supporting subsets of namespaces but let's
wait until we have users asking for even more context to be assumed. At
that point we can add an extension.
The obvious example where this is useful is a standard container
manager interacting with a running container: pushing and pulling files
or directories, injecting mounts, attaching/execing any kind of process,
managing network devices all these operations require attaching to all
or at least multiple namespaces at the same time. Given that nowadays
most containers are spawned with all namespaces enabled we're currently
looking at at least 14 syscalls, 7 to open the /proc/<pid>/ns/<ns>
nsfds, another 7 to actually perform the namespace switch. With time
namespaces we're looking at about 16 syscalls.
(We could amortize the first 7 or 8 syscalls for opening the nsfds by
stashing them in each container's monitor process but that would mean
we need to send around those file descriptors through unix sockets
everytime we want to interact with the container or keep on-disk
state. Even in scenarios where a caller wants to join a particular
namespace in a particular order callers still profit from batching
other namespaces. That mostly applies to the user namespace but
all container runtimes I found join the user namespace first no matter
if it privileges or deprivileges the container similar to how unshare
behaves.)
With pidfds this becomes a single syscall no matter how many namespaces
are supposed to be attached to.
A decently designed, large-scale container manager usually isn't the
parent of any of the containers it spawns so the containers don't die
when it crashes or needs to update or reinitialize. This means that
for the manager to interact with containers through pids is inherently
racy especially on systems where the maximum pid number is not
significicantly bumped. This is even more problematic since we often spawn
and manage thousands or ten-thousands of containers. Interacting with a
container through a pid thus can become risky quite quickly. Especially
since we allow for an administrator to enable advanced features such as
syscall interception where we're performing syscalls in lieu of the
container. In all of those cases we use pidfds if they are available and
we pass them around as stable references. Using them to setns() to the
target process' namespaces is as reliable as using nsfds. Either the
target process is already dead and we get ESRCH or we manage to attach
to its namespaces but we can't accidently attach to another process'
namespaces. So pidfds lend themselves to be used with this api.
The other main advantage is that with this change the pidfd becomes the
only relevant token for most container interactions and it's the only
token we need to create and send around.
Apart from significiantly reducing the number of syscalls from double
digit to single digit which is a decent reason post-spectre/meltdown
this also allows to switch to a set of namespaces atomically, i.e.
either attaching to all the specified namespaces succeeds or we fail. If
we fail we haven't changed a single namespace. There are currently three
namespaces that can fail (other than for ENOMEM which really is not
very interesting since we then have other problems anyway) for
non-trivial reasons, user, mount, and pid namespaces. We can fail to
attach to a pid namespace if it is not our current active pid namespace
or a descendant of it. We can fail to attach to a user namespace because
we are multi-threaded or because our current mount namespace shares
filesystem state with other tasks, or because we're trying to setns()
to the same user namespace, i.e. the target task has the same user
namespace as we do. We can fail to attach to a mount namespace because
it shares filesystem state with other tasks or because we fail to lookup
the new root for the new mount namespace. In most non-pathological
scenarios these issues can be somewhat mitigated. But there are cases where
we're half-attached to some namespace and failing to attach to another one.
I've talked about some of these problem during the hallway track (something
only the pre-COVID-19 generation will remember) of Plumbers in Los Angeles
in 2018(?). Even if all these issues could be avoided with super careful
userspace coding it would be nicer to have this done in-kernel. Pidfds seem
to lend themselves nicely for this.
The other neat thing about this is that setns() becomes an actual
counterpart to the namespace bits of unshare().
Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
Reviewed-by: Serge Hallyn <serge@hallyn.com>
Cc: Eric W. Biederman <ebiederm@xmission.com>
Cc: Serge Hallyn <serge@hallyn.com>
Cc: Jann Horn <jannh@google.com>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Aleksa Sarai <cyphar@cyphar.com>
Link: https://lore.kernel.org/r/20200505140432.181565-3-christian.brauner@ubuntu.com
2020-05-05 17:04:31 +03:00
int err = 0 ;
2010-03-08 04:48:52 +03:00
nsproxy: attach to namespaces via pidfds
For quite a while we have been thinking about using pidfds to attach to
namespaces. This patchset has existed for about a year already but we've
wanted to wait to see how the general api would be received and adopted.
Now that more and more programs in userspace have started using pidfds
for process management it's time to send this one out.
This patch makes it possible to use pidfds to attach to the namespaces
of another process, i.e. they can be passed as the first argument to the
setns() syscall. When only a single namespace type is specified the
semantics are equivalent to passing an nsfd. That means
setns(nsfd, CLONE_NEWNET) equals setns(pidfd, CLONE_NEWNET). However,
when a pidfd is passed, multiple namespace flags can be specified in the
second setns() argument and setns() will attach the caller to all the
specified namespaces all at once or to none of them. Specifying 0 is not
valid together with a pidfd.
Here are just two obvious examples:
setns(pidfd, CLONE_NEWPID | CLONE_NEWNS | CLONE_NEWNET);
setns(pidfd, CLONE_NEWUSER);
Allowing to also attach subsets of namespaces supports various use-cases
where callers setns to a subset of namespaces to retain privilege, perform
an action and then re-attach another subset of namespaces.
If the need arises, as Eric suggested, we can extend this patchset to
assume even more context than just attaching all namespaces. His suggestion
specifically was about assuming the process' root directory when
setns(pidfd, 0) or setns(pidfd, SETNS_PIDFD) is specified. For now, just
keep it flexible in terms of supporting subsets of namespaces but let's
wait until we have users asking for even more context to be assumed. At
that point we can add an extension.
The obvious example where this is useful is a standard container
manager interacting with a running container: pushing and pulling files
or directories, injecting mounts, attaching/execing any kind of process,
managing network devices all these operations require attaching to all
or at least multiple namespaces at the same time. Given that nowadays
most containers are spawned with all namespaces enabled we're currently
looking at at least 14 syscalls, 7 to open the /proc/<pid>/ns/<ns>
nsfds, another 7 to actually perform the namespace switch. With time
namespaces we're looking at about 16 syscalls.
(We could amortize the first 7 or 8 syscalls for opening the nsfds by
stashing them in each container's monitor process but that would mean
we need to send around those file descriptors through unix sockets
everytime we want to interact with the container or keep on-disk
state. Even in scenarios where a caller wants to join a particular
namespace in a particular order callers still profit from batching
other namespaces. That mostly applies to the user namespace but
all container runtimes I found join the user namespace first no matter
if it privileges or deprivileges the container similar to how unshare
behaves.)
With pidfds this becomes a single syscall no matter how many namespaces
are supposed to be attached to.
A decently designed, large-scale container manager usually isn't the
parent of any of the containers it spawns so the containers don't die
when it crashes or needs to update or reinitialize. This means that
for the manager to interact with containers through pids is inherently
racy especially on systems where the maximum pid number is not
significicantly bumped. This is even more problematic since we often spawn
and manage thousands or ten-thousands of containers. Interacting with a
container through a pid thus can become risky quite quickly. Especially
since we allow for an administrator to enable advanced features such as
syscall interception where we're performing syscalls in lieu of the
container. In all of those cases we use pidfds if they are available and
we pass them around as stable references. Using them to setns() to the
target process' namespaces is as reliable as using nsfds. Either the
target process is already dead and we get ESRCH or we manage to attach
to its namespaces but we can't accidently attach to another process'
namespaces. So pidfds lend themselves to be used with this api.
The other main advantage is that with this change the pidfd becomes the
only relevant token for most container interactions and it's the only
token we need to create and send around.
Apart from significiantly reducing the number of syscalls from double
digit to single digit which is a decent reason post-spectre/meltdown
this also allows to switch to a set of namespaces atomically, i.e.
either attaching to all the specified namespaces succeeds or we fail. If
we fail we haven't changed a single namespace. There are currently three
namespaces that can fail (other than for ENOMEM which really is not
very interesting since we then have other problems anyway) for
non-trivial reasons, user, mount, and pid namespaces. We can fail to
attach to a pid namespace if it is not our current active pid namespace
or a descendant of it. We can fail to attach to a user namespace because
we are multi-threaded or because our current mount namespace shares
filesystem state with other tasks, or because we're trying to setns()
to the same user namespace, i.e. the target task has the same user
namespace as we do. We can fail to attach to a mount namespace because
it shares filesystem state with other tasks or because we fail to lookup
the new root for the new mount namespace. In most non-pathological
scenarios these issues can be somewhat mitigated. But there are cases where
we're half-attached to some namespace and failing to attach to another one.
I've talked about some of these problem during the hallway track (something
only the pre-COVID-19 generation will remember) of Plumbers in Los Angeles
in 2018(?). Even if all these issues could be avoided with super careful
userspace coding it would be nicer to have this done in-kernel. Pidfds seem
to lend themselves nicely for this.
The other neat thing about this is that setns() becomes an actual
counterpart to the namespace bits of unshare().
Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
Reviewed-by: Serge Hallyn <serge@hallyn.com>
Cc: Eric W. Biederman <ebiederm@xmission.com>
Cc: Serge Hallyn <serge@hallyn.com>
Cc: Jann Horn <jannh@google.com>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Aleksa Sarai <cyphar@cyphar.com>
Link: https://lore.kernel.org/r/20200505140432.181565-3-christian.brauner@ubuntu.com
2020-05-05 17:04:31 +03:00
file = fget ( fd ) ;
if ( ! file )
return - EBADF ;
if ( proc_ns_file ( file ) ) {
ns = get_proc_ns ( file_inode ( file ) ) ;
if ( flags & & ( ns - > ops - > type ! = flags ) )
err = - EINVAL ;
flags = ns - > ops - > type ;
} else if ( ! IS_ERR ( pidfd_pid ( file ) ) ) {
err = check_setns_flags ( flags ) ;
} else {
2020-06-17 01:33:12 +03:00
err = - EINVAL ;
nsproxy: attach to namespaces via pidfds
For quite a while we have been thinking about using pidfds to attach to
namespaces. This patchset has existed for about a year already but we've
wanted to wait to see how the general api would be received and adopted.
Now that more and more programs in userspace have started using pidfds
for process management it's time to send this one out.
This patch makes it possible to use pidfds to attach to the namespaces
of another process, i.e. they can be passed as the first argument to the
setns() syscall. When only a single namespace type is specified the
semantics are equivalent to passing an nsfd. That means
setns(nsfd, CLONE_NEWNET) equals setns(pidfd, CLONE_NEWNET). However,
when a pidfd is passed, multiple namespace flags can be specified in the
second setns() argument and setns() will attach the caller to all the
specified namespaces all at once or to none of them. Specifying 0 is not
valid together with a pidfd.
Here are just two obvious examples:
setns(pidfd, CLONE_NEWPID | CLONE_NEWNS | CLONE_NEWNET);
setns(pidfd, CLONE_NEWUSER);
Allowing to also attach subsets of namespaces supports various use-cases
where callers setns to a subset of namespaces to retain privilege, perform
an action and then re-attach another subset of namespaces.
If the need arises, as Eric suggested, we can extend this patchset to
assume even more context than just attaching all namespaces. His suggestion
specifically was about assuming the process' root directory when
setns(pidfd, 0) or setns(pidfd, SETNS_PIDFD) is specified. For now, just
keep it flexible in terms of supporting subsets of namespaces but let's
wait until we have users asking for even more context to be assumed. At
that point we can add an extension.
The obvious example where this is useful is a standard container
manager interacting with a running container: pushing and pulling files
or directories, injecting mounts, attaching/execing any kind of process,
managing network devices all these operations require attaching to all
or at least multiple namespaces at the same time. Given that nowadays
most containers are spawned with all namespaces enabled we're currently
looking at at least 14 syscalls, 7 to open the /proc/<pid>/ns/<ns>
nsfds, another 7 to actually perform the namespace switch. With time
namespaces we're looking at about 16 syscalls.
(We could amortize the first 7 or 8 syscalls for opening the nsfds by
stashing them in each container's monitor process but that would mean
we need to send around those file descriptors through unix sockets
everytime we want to interact with the container or keep on-disk
state. Even in scenarios where a caller wants to join a particular
namespace in a particular order callers still profit from batching
other namespaces. That mostly applies to the user namespace but
all container runtimes I found join the user namespace first no matter
if it privileges or deprivileges the container similar to how unshare
behaves.)
With pidfds this becomes a single syscall no matter how many namespaces
are supposed to be attached to.
A decently designed, large-scale container manager usually isn't the
parent of any of the containers it spawns so the containers don't die
when it crashes or needs to update or reinitialize. This means that
for the manager to interact with containers through pids is inherently
racy especially on systems where the maximum pid number is not
significicantly bumped. This is even more problematic since we often spawn
and manage thousands or ten-thousands of containers. Interacting with a
container through a pid thus can become risky quite quickly. Especially
since we allow for an administrator to enable advanced features such as
syscall interception where we're performing syscalls in lieu of the
container. In all of those cases we use pidfds if they are available and
we pass them around as stable references. Using them to setns() to the
target process' namespaces is as reliable as using nsfds. Either the
target process is already dead and we get ESRCH or we manage to attach
to its namespaces but we can't accidently attach to another process'
namespaces. So pidfds lend themselves to be used with this api.
The other main advantage is that with this change the pidfd becomes the
only relevant token for most container interactions and it's the only
token we need to create and send around.
Apart from significiantly reducing the number of syscalls from double
digit to single digit which is a decent reason post-spectre/meltdown
this also allows to switch to a set of namespaces atomically, i.e.
either attaching to all the specified namespaces succeeds or we fail. If
we fail we haven't changed a single namespace. There are currently three
namespaces that can fail (other than for ENOMEM which really is not
very interesting since we then have other problems anyway) for
non-trivial reasons, user, mount, and pid namespaces. We can fail to
attach to a pid namespace if it is not our current active pid namespace
or a descendant of it. We can fail to attach to a user namespace because
we are multi-threaded or because our current mount namespace shares
filesystem state with other tasks, or because we're trying to setns()
to the same user namespace, i.e. the target task has the same user
namespace as we do. We can fail to attach to a mount namespace because
it shares filesystem state with other tasks or because we fail to lookup
the new root for the new mount namespace. In most non-pathological
scenarios these issues can be somewhat mitigated. But there are cases where
we're half-attached to some namespace and failing to attach to another one.
I've talked about some of these problem during the hallway track (something
only the pre-COVID-19 generation will remember) of Plumbers in Los Angeles
in 2018(?). Even if all these issues could be avoided with super careful
userspace coding it would be nicer to have this done in-kernel. Pidfds seem
to lend themselves nicely for this.
The other neat thing about this is that setns() becomes an actual
counterpart to the namespace bits of unshare().
Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
Reviewed-by: Serge Hallyn <serge@hallyn.com>
Cc: Eric W. Biederman <ebiederm@xmission.com>
Cc: Serge Hallyn <serge@hallyn.com>
Cc: Jann Horn <jannh@google.com>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Aleksa Sarai <cyphar@cyphar.com>
Link: https://lore.kernel.org/r/20200505140432.181565-3-christian.brauner@ubuntu.com
2020-05-05 17:04:31 +03:00
}
if ( err )
2010-03-08 04:48:52 +03:00
goto out ;
nsproxy: attach to namespaces via pidfds
For quite a while we have been thinking about using pidfds to attach to
namespaces. This patchset has existed for about a year already but we've
wanted to wait to see how the general api would be received and adopted.
Now that more and more programs in userspace have started using pidfds
for process management it's time to send this one out.
This patch makes it possible to use pidfds to attach to the namespaces
of another process, i.e. they can be passed as the first argument to the
setns() syscall. When only a single namespace type is specified the
semantics are equivalent to passing an nsfd. That means
setns(nsfd, CLONE_NEWNET) equals setns(pidfd, CLONE_NEWNET). However,
when a pidfd is passed, multiple namespace flags can be specified in the
second setns() argument and setns() will attach the caller to all the
specified namespaces all at once or to none of them. Specifying 0 is not
valid together with a pidfd.
Here are just two obvious examples:
setns(pidfd, CLONE_NEWPID | CLONE_NEWNS | CLONE_NEWNET);
setns(pidfd, CLONE_NEWUSER);
Allowing to also attach subsets of namespaces supports various use-cases
where callers setns to a subset of namespaces to retain privilege, perform
an action and then re-attach another subset of namespaces.
If the need arises, as Eric suggested, we can extend this patchset to
assume even more context than just attaching all namespaces. His suggestion
specifically was about assuming the process' root directory when
setns(pidfd, 0) or setns(pidfd, SETNS_PIDFD) is specified. For now, just
keep it flexible in terms of supporting subsets of namespaces but let's
wait until we have users asking for even more context to be assumed. At
that point we can add an extension.
The obvious example where this is useful is a standard container
manager interacting with a running container: pushing and pulling files
or directories, injecting mounts, attaching/execing any kind of process,
managing network devices all these operations require attaching to all
or at least multiple namespaces at the same time. Given that nowadays
most containers are spawned with all namespaces enabled we're currently
looking at at least 14 syscalls, 7 to open the /proc/<pid>/ns/<ns>
nsfds, another 7 to actually perform the namespace switch. With time
namespaces we're looking at about 16 syscalls.
(We could amortize the first 7 or 8 syscalls for opening the nsfds by
stashing them in each container's monitor process but that would mean
we need to send around those file descriptors through unix sockets
everytime we want to interact with the container or keep on-disk
state. Even in scenarios where a caller wants to join a particular
namespace in a particular order callers still profit from batching
other namespaces. That mostly applies to the user namespace but
all container runtimes I found join the user namespace first no matter
if it privileges or deprivileges the container similar to how unshare
behaves.)
With pidfds this becomes a single syscall no matter how many namespaces
are supposed to be attached to.
A decently designed, large-scale container manager usually isn't the
parent of any of the containers it spawns so the containers don't die
when it crashes or needs to update or reinitialize. This means that
for the manager to interact with containers through pids is inherently
racy especially on systems where the maximum pid number is not
significicantly bumped. This is even more problematic since we often spawn
and manage thousands or ten-thousands of containers. Interacting with a
container through a pid thus can become risky quite quickly. Especially
since we allow for an administrator to enable advanced features such as
syscall interception where we're performing syscalls in lieu of the
container. In all of those cases we use pidfds if they are available and
we pass them around as stable references. Using them to setns() to the
target process' namespaces is as reliable as using nsfds. Either the
target process is already dead and we get ESRCH or we manage to attach
to its namespaces but we can't accidently attach to another process'
namespaces. So pidfds lend themselves to be used with this api.
The other main advantage is that with this change the pidfd becomes the
only relevant token for most container interactions and it's the only
token we need to create and send around.
Apart from significiantly reducing the number of syscalls from double
digit to single digit which is a decent reason post-spectre/meltdown
this also allows to switch to a set of namespaces atomically, i.e.
either attaching to all the specified namespaces succeeds or we fail. If
we fail we haven't changed a single namespace. There are currently three
namespaces that can fail (other than for ENOMEM which really is not
very interesting since we then have other problems anyway) for
non-trivial reasons, user, mount, and pid namespaces. We can fail to
attach to a pid namespace if it is not our current active pid namespace
or a descendant of it. We can fail to attach to a user namespace because
we are multi-threaded or because our current mount namespace shares
filesystem state with other tasks, or because we're trying to setns()
to the same user namespace, i.e. the target task has the same user
namespace as we do. We can fail to attach to a mount namespace because
it shares filesystem state with other tasks or because we fail to lookup
the new root for the new mount namespace. In most non-pathological
scenarios these issues can be somewhat mitigated. But there are cases where
we're half-attached to some namespace and failing to attach to another one.
I've talked about some of these problem during the hallway track (something
only the pre-COVID-19 generation will remember) of Plumbers in Los Angeles
in 2018(?). Even if all these issues could be avoided with super careful
userspace coding it would be nicer to have this done in-kernel. Pidfds seem
to lend themselves nicely for this.
The other neat thing about this is that setns() becomes an actual
counterpart to the namespace bits of unshare().
Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
Reviewed-by: Serge Hallyn <serge@hallyn.com>
Cc: Eric W. Biederman <ebiederm@xmission.com>
Cc: Serge Hallyn <serge@hallyn.com>
Cc: Jann Horn <jannh@google.com>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Aleksa Sarai <cyphar@cyphar.com>
Link: https://lore.kernel.org/r/20200505140432.181565-3-christian.brauner@ubuntu.com
2020-05-05 17:04:31 +03:00
err = prepare_nsset ( flags , & nsset ) ;
2020-05-05 17:04:30 +03:00
if ( err )
2010-03-08 04:48:52 +03:00
goto out ;
nsproxy: attach to namespaces via pidfds
For quite a while we have been thinking about using pidfds to attach to
namespaces. This patchset has existed for about a year already but we've
wanted to wait to see how the general api would be received and adopted.
Now that more and more programs in userspace have started using pidfds
for process management it's time to send this one out.
This patch makes it possible to use pidfds to attach to the namespaces
of another process, i.e. they can be passed as the first argument to the
setns() syscall. When only a single namespace type is specified the
semantics are equivalent to passing an nsfd. That means
setns(nsfd, CLONE_NEWNET) equals setns(pidfd, CLONE_NEWNET). However,
when a pidfd is passed, multiple namespace flags can be specified in the
second setns() argument and setns() will attach the caller to all the
specified namespaces all at once or to none of them. Specifying 0 is not
valid together with a pidfd.
Here are just two obvious examples:
setns(pidfd, CLONE_NEWPID | CLONE_NEWNS | CLONE_NEWNET);
setns(pidfd, CLONE_NEWUSER);
Allowing to also attach subsets of namespaces supports various use-cases
where callers setns to a subset of namespaces to retain privilege, perform
an action and then re-attach another subset of namespaces.
If the need arises, as Eric suggested, we can extend this patchset to
assume even more context than just attaching all namespaces. His suggestion
specifically was about assuming the process' root directory when
setns(pidfd, 0) or setns(pidfd, SETNS_PIDFD) is specified. For now, just
keep it flexible in terms of supporting subsets of namespaces but let's
wait until we have users asking for even more context to be assumed. At
that point we can add an extension.
The obvious example where this is useful is a standard container
manager interacting with a running container: pushing and pulling files
or directories, injecting mounts, attaching/execing any kind of process,
managing network devices all these operations require attaching to all
or at least multiple namespaces at the same time. Given that nowadays
most containers are spawned with all namespaces enabled we're currently
looking at at least 14 syscalls, 7 to open the /proc/<pid>/ns/<ns>
nsfds, another 7 to actually perform the namespace switch. With time
namespaces we're looking at about 16 syscalls.
(We could amortize the first 7 or 8 syscalls for opening the nsfds by
stashing them in each container's monitor process but that would mean
we need to send around those file descriptors through unix sockets
everytime we want to interact with the container or keep on-disk
state. Even in scenarios where a caller wants to join a particular
namespace in a particular order callers still profit from batching
other namespaces. That mostly applies to the user namespace but
all container runtimes I found join the user namespace first no matter
if it privileges or deprivileges the container similar to how unshare
behaves.)
With pidfds this becomes a single syscall no matter how many namespaces
are supposed to be attached to.
A decently designed, large-scale container manager usually isn't the
parent of any of the containers it spawns so the containers don't die
when it crashes or needs to update or reinitialize. This means that
for the manager to interact with containers through pids is inherently
racy especially on systems where the maximum pid number is not
significicantly bumped. This is even more problematic since we often spawn
and manage thousands or ten-thousands of containers. Interacting with a
container through a pid thus can become risky quite quickly. Especially
since we allow for an administrator to enable advanced features such as
syscall interception where we're performing syscalls in lieu of the
container. In all of those cases we use pidfds if they are available and
we pass them around as stable references. Using them to setns() to the
target process' namespaces is as reliable as using nsfds. Either the
target process is already dead and we get ESRCH or we manage to attach
to its namespaces but we can't accidently attach to another process'
namespaces. So pidfds lend themselves to be used with this api.
The other main advantage is that with this change the pidfd becomes the
only relevant token for most container interactions and it's the only
token we need to create and send around.
Apart from significiantly reducing the number of syscalls from double
digit to single digit which is a decent reason post-spectre/meltdown
this also allows to switch to a set of namespaces atomically, i.e.
either attaching to all the specified namespaces succeeds or we fail. If
we fail we haven't changed a single namespace. There are currently three
namespaces that can fail (other than for ENOMEM which really is not
very interesting since we then have other problems anyway) for
non-trivial reasons, user, mount, and pid namespaces. We can fail to
attach to a pid namespace if it is not our current active pid namespace
or a descendant of it. We can fail to attach to a user namespace because
we are multi-threaded or because our current mount namespace shares
filesystem state with other tasks, or because we're trying to setns()
to the same user namespace, i.e. the target task has the same user
namespace as we do. We can fail to attach to a mount namespace because
it shares filesystem state with other tasks or because we fail to lookup
the new root for the new mount namespace. In most non-pathological
scenarios these issues can be somewhat mitigated. But there are cases where
we're half-attached to some namespace and failing to attach to another one.
I've talked about some of these problem during the hallway track (something
only the pre-COVID-19 generation will remember) of Plumbers in Los Angeles
in 2018(?). Even if all these issues could be avoided with super careful
userspace coding it would be nicer to have this done in-kernel. Pidfds seem
to lend themselves nicely for this.
The other neat thing about this is that setns() becomes an actual
counterpart to the namespace bits of unshare().
Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
Reviewed-by: Serge Hallyn <serge@hallyn.com>
Cc: Eric W. Biederman <ebiederm@xmission.com>
Cc: Serge Hallyn <serge@hallyn.com>
Cc: Jann Horn <jannh@google.com>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Aleksa Sarai <cyphar@cyphar.com>
Link: https://lore.kernel.org/r/20200505140432.181565-3-christian.brauner@ubuntu.com
2020-05-05 17:04:31 +03:00
if ( proc_ns_file ( file ) )
err = validate_ns ( & nsset , ns ) ;
else
err = validate_nsset ( & nsset , file - > private_data ) ;
2020-05-05 17:04:30 +03:00
if ( ! err ) {
commit_nsset ( & nsset ) ;
perf_event_namespaces ( current ) ;
2010-03-08 04:48:52 +03:00
}
2020-05-05 17:04:30 +03:00
put_nsset ( & nsset ) ;
2010-03-08 04:48:52 +03:00
out :
fput ( file ) ;
return err ;
}
2011-06-28 23:41:10 +04:00
int __init nsproxy_cache_init ( void )
2007-07-16 10:41:07 +04:00
{
2021-09-03 00:55:27 +03:00
nsproxy_cachep = KMEM_CACHE ( nsproxy , SLAB_PANIC | SLAB_ACCOUNT ) ;
2007-07-16 10:41:07 +04:00
return 0 ;
}