Message ID | 20220613060723.197407-1-avagin@gmail.com (mailing list archive) |
---|---|
State | New |
Headers | show |
Series | [1/2] fs/exec: allow to unshare a time namespace on vfork+exec | expand |
On Sun, Jun 12, 2022 at 11:07:22PM -0700, Andrei Vagin wrote: > Right now, a new process can't be forked in another time namespace > if it shares mm with its parent. It is prohibited, because each time > namespace has its own vvar page that is mapped into a process address > space. > > When a process calls exec, it gets a new mm and so it could be "legal" > to switch time namespace in that case. This was not implemented and > now if we want to do this, we need to add another clone flag to not > break backward compatibility. > > We don't have any user requests to switch times on exec except the > vfork+exec combination, so there is no reason to add a new clone flag. > As for vfork+exec, this should be safe to allow switching timens with > the current clone flag. Right now, vfork (CLONE_VFORK | CLONE_VM) fails > if a child is forked into another time namespace. With this change, > vfork creates a new process in parent's timens, and the following exec > does the actual switch to the target time namespace. This seems like a very special case. None of the other namespaces do this, do they? How is CLONE_NEWTIME supposed to be used today? -Kees
On Sun, Jun 12, 2022 at 11:07:22PM -0700, Andrei Vagin wrote: > Right now, a new process can't be forked in another time namespace > if it shares mm with its parent. It is prohibited, because each time > namespace has its own vvar page that is mapped into a process address > space. > > When a process calls exec, it gets a new mm and so it could be "legal" > to switch time namespace in that case. This was not implemented and > now if we want to do this, we need to add another clone flag to not > break backward compatibility. > > We don't have any user requests to switch times on exec except the > vfork+exec combination, so there is no reason to add a new clone flag. > As for vfork+exec, this should be safe to allow switching timens with > the current clone flag. Right now, vfork (CLONE_VFORK | CLONE_VM) fails > if a child is forked into another time namespace. With this change, > vfork creates a new process in parent's timens, and the following exec > does the actual switch to the target time namespace. > > Suggested-by: Florian Weimer <fweimer@redhat.com> > Signed-off-by: Andrei Vagin <avagin@gmail.com> > --- Looks good, Acked-by: Christian Brauner (Microsoft) <brauner@kernel.org>
On Tue, Jun 14, 2022 at 02:14:35PM -0700, Kees Cook wrote: > On Sun, Jun 12, 2022 at 11:07:22PM -0700, Andrei Vagin wrote: > > Right now, a new process can't be forked in another time namespace > > if it shares mm with its parent. It is prohibited, because each time > > namespace has its own vvar page that is mapped into a process address > > space. > > > > When a process calls exec, it gets a new mm and so it could be "legal" > > to switch time namespace in that case. This was not implemented and > > now if we want to do this, we need to add another clone flag to not > > break backward compatibility. > > > > We don't have any user requests to switch times on exec except the > > vfork+exec combination, so there is no reason to add a new clone flag. > > As for vfork+exec, this should be safe to allow switching timens with > > the current clone flag. Right now, vfork (CLONE_VFORK | CLONE_VM) fails > > if a child is forked into another time namespace. With this change, > > vfork creates a new process in parent's timens, and the following exec > > does the actual switch to the target time namespace. > > This seems like a very special case. None of the other namespaces do > this, do they? > > How is CLONE_NEWTIME supposed to be used today? Time namespaces are similar to pid namespaces. If a process calls unshare(CLONE_NEWTIME) it will not change into a new time namespace. Only the children of the process will. You can also see this via /proc/<pid>/ns/time and /proc/<pid>/ns/time_for_children. After an unshare(CLONE_NEWTIME) /proc/<pid>/ns/time will be unchanged while /proc/<pid>/ns/time_for_children will reference a new time namespace. So if the process now calls fork() the child will be placed in a new time namespace. As Andrei correctly points out in the commit message each time namespace gets it's own vvar page mapped into the process address space. Consequently calls to clone*() with CLONE_VM will need to fail because it would alter the parent's mm as well. That includes vfork() which is roughly just CLONE_VM | CLONE_VFORK. fork() remains unaffected. So anything that implements a process launcher using vfork() needs to implement a fork() fallback after vfork failure() in case the original process has unshared a new time namespace. As posix spawn is implemented using vfork() we would force glibc to implement a fork() fallback and enforce the introducing of a lot of complexity to work around this. I think the proposal here makes sense and allows us to avoid introducing yet another clone flag. For vfork() it also makes sense because the calling process is suspended until exec or exit so the semantics between fork() and clone*() are sufficiently distinct to justify this difference. Iow, vfork() is distinctly targeted at enforcing an exec* call already anyway. Christian
* Kees Cook: > On Sun, Jun 12, 2022 at 11:07:22PM -0700, Andrei Vagin wrote: >> Right now, a new process can't be forked in another time namespace >> if it shares mm with its parent. It is prohibited, because each time >> namespace has its own vvar page that is mapped into a process address >> space. >> >> When a process calls exec, it gets a new mm and so it could be "legal" >> to switch time namespace in that case. This was not implemented and >> now if we want to do this, we need to add another clone flag to not >> break backward compatibility. >> >> We don't have any user requests to switch times on exec except the >> vfork+exec combination, so there is no reason to add a new clone flag. >> As for vfork+exec, this should be safe to allow switching timens with >> the current clone flag. Right now, vfork (CLONE_VFORK | CLONE_VM) fails >> if a child is forked into another time namespace. With this change, >> vfork creates a new process in parent's timens, and the following exec >> does the actual switch to the target time namespace. > > This seems like a very special case. None of the other namespaces do > this, do they? I think this started with CLONE_NEWPID, which had a similar delayed effect with unshare: it happens only after fork, not for the current process image. I think it's just a limitation of the unshare interface. Some of the effects simply have to be delayed due to their nature. Thanks, Florian
On Wed, Jun 15, 2022 at 09:53:29AM +0200, Florian Weimer wrote: > * Kees Cook: > > > On Sun, Jun 12, 2022 at 11:07:22PM -0700, Andrei Vagin wrote: > >> Right now, a new process can't be forked in another time namespace > >> if it shares mm with its parent. It is prohibited, because each time > >> namespace has its own vvar page that is mapped into a process address > >> space. > >> > >> When a process calls exec, it gets a new mm and so it could be "legal" > >> to switch time namespace in that case. This was not implemented and > >> now if we want to do this, we need to add another clone flag to not > >> break backward compatibility. > >> > >> We don't have any user requests to switch times on exec except the > >> vfork+exec combination, so there is no reason to add a new clone flag. > >> As for vfork+exec, this should be safe to allow switching timens with > >> the current clone flag. Right now, vfork (CLONE_VFORK | CLONE_VM) fails > >> if a child is forked into another time namespace. With this change, > >> vfork creates a new process in parent's timens, and the following exec > >> does the actual switch to the target time namespace. > > > > This seems like a very special case. None of the other namespaces do > > this, do they? > > I think this started with CLONE_NEWPID, which had a similar delayed > effect with unshare: it happens only after fork, not for the current > process image. I think it's just a limitation of the unshare interface. > Some of the effects simply have to be delayed due to their nature. I tried to give more context in another mail wrt to time namespaces specifically. For pid namespaces one problem would be that it could end up confusing a process about its own pid. This was a more serious problem when the pid cache was still active in glibc; but fwiw systemd still has a pid cache afair. Christian
* Christian Brauner: > For pid namespaces one problem would be that it could end up confusing a > process about its own pid. This was a more serious problem when the pid > cache was still active in glibc; but fwiw systemd still has a pid cache > afair. Right. glibc still has a TID cache, mainly for use with recursive mutexes (where we need a 32-bit thread identifier and can't perform a system call on every locking operation for performance reasons). Assuming that a non-delayed CLONE_NEWPID would also change the TID underneath us, we'd have subtly broken recursive mutexes. vfork gets away with not updating the TID cache (which is shared with the parent process) because the parent process is suspended while the new subprocess is still running and has not execve'ed yet. Now one could argue that calling unshare automatically means that you must not call any glibc functions afterwards (similar to thread-creating clone), or at least that you cannot call any functions which are not async-signal-safe, but that does not match existing application practice. And I think we actually prefer that file servers call chroot after unshare(CLONE_FS), rather than trying to reimplement restricted pathname lookup in userspace. Thanks, Florian
On Wed, Jun 15, 2022 at 10:14:19AM +0200, Florian Weimer wrote: > * Christian Brauner: > > > For pid namespaces one problem would be that it could end up confusing a > > process about its own pid. This was a more serious problem when the pid > > cache was still active in glibc; but fwiw systemd still has a pid cache > > afair. > > Right. glibc still has a TID cache, mainly for use with recursive > mutexes (where we need a 32-bit thread identifier and can't perform a > system call on every locking operation for performance reasons). > Assuming that a non-delayed CLONE_NEWPID would also change the TID > underneath us, we'd have subtly broken recursive mutexes. Fwiw, you can't call CLONE_NEWPID with CLONE_THREAD. This guarantees that threads can send signals to each other and all threads within the same threadgroup can be reached via proc. It'd be awkward if you'd have a thread whose thread-group leader lives in an ancestor pidns. Even if you'd make whole threadgroup change pid namespaces immediately it would mean allocating new TGID and TIDs in the new pid namespaces - unless they are accidently not already allocated. > > vfork gets away with not updating the TID cache (which is shared with > the parent process) because the parent process is suspended while the > new subprocess is still running and has not execve'ed yet. > > Now one could argue that calling unshare automatically means that you > must not call any glibc functions afterwards (similar to thread-creating > clone), or at least that you cannot call any functions which are not > async-signal-safe, but that does not match existing application > practice. And I think we actually prefer that file servers call chroot Yeah, that'd be a rather subtle and risky change for pid namespaces.
On Sun, 12 Jun 2022 23:07:22 -0700, Andrei Vagin wrote: > Right now, a new process can't be forked in another time namespace > if it shares mm with its parent. It is prohibited, because each time > namespace has its own vvar page that is mapped into a process address > space. > > When a process calls exec, it gets a new mm and so it could be "legal" > to switch time namespace in that case. This was not implemented and > now if we want to do this, we need to add another clone flag to not > break backward compatibility. > > [...] Applied to for-next/execve, thanks! [1/2] fs/exec: allow to unshare a time namespace on vfork+exec https://git.kernel.org/kees/c/133e2d3e81de [2/2] testing/timens: add a test for vfork+exit https://git.kernel.org/kees/c/6342140db660
diff --git a/fs/exec.c b/fs/exec.c index 0989fb8472a1..347e8f55bc2b 100644 --- a/fs/exec.c +++ b/fs/exec.c @@ -65,6 +65,7 @@ #include <linux/io_uring.h> #include <linux/syscall_user_dispatch.h> #include <linux/coredump.h> +#include <linux/time_namespace.h> #include <linux/uaccess.h> #include <asm/mmu_context.h> @@ -982,10 +983,12 @@ static int exec_mmap(struct mm_struct *mm) { struct task_struct *tsk; struct mm_struct *old_mm, *active_mm; + bool vfork; int ret; /* Notify parent that we're no longer interested in the old VM */ tsk = current; + vfork = !!tsk->vfork_done; old_mm = current->mm; exec_mm_release(tsk, old_mm); if (old_mm) @@ -1030,6 +1033,10 @@ static int exec_mmap(struct mm_struct *mm) tsk->mm->vmacache_seqnum = 0; vmacache_flush(tsk); task_unlock(tsk); + + if (vfork) + timens_on_fork(tsk->nsproxy, tsk); + if (old_mm) { mmap_read_unlock(old_mm); BUG_ON(active_mm != old_mm); diff --git a/kernel/fork.c b/kernel/fork.c index 9d44f2d46c69..9174146f6812 100644 --- a/kernel/fork.c +++ b/kernel/fork.c @@ -2033,8 +2033,11 @@ static __latent_entropy struct task_struct *copy_process( /* * If the new process will be in a different time namespace * do not allow it to share VM or a thread group with the forking task. + * + * On vfork, the child process enters the target time namespace only + * after exec. */ - if (clone_flags & (CLONE_THREAD | CLONE_VM)) { + if ((clone_flags & (CLONE_VM | CLONE_VFORK)) == CLONE_VM) { if (nsp->time_ns != nsp->time_ns_for_children) return ERR_PTR(-EINVAL); } diff --git a/kernel/nsproxy.c b/kernel/nsproxy.c index eec72ca962e2..b4cbb406bc28 100644 --- a/kernel/nsproxy.c +++ b/kernel/nsproxy.c @@ -179,7 +179,8 @@ int copy_namespaces(unsigned long flags, struct task_struct *tsk) if (IS_ERR(new_ns)) return PTR_ERR(new_ns); - timens_on_fork(new_ns, tsk); + if ((flags & CLONE_VM) == 0) + timens_on_fork(new_ns, tsk); tsk->nsproxy = new_ns; return 0;
Right now, a new process can't be forked in another time namespace if it shares mm with its parent. It is prohibited, because each time namespace has its own vvar page that is mapped into a process address space. When a process calls exec, it gets a new mm and so it could be "legal" to switch time namespace in that case. This was not implemented and now if we want to do this, we need to add another clone flag to not break backward compatibility. We don't have any user requests to switch times on exec except the vfork+exec combination, so there is no reason to add a new clone flag. As for vfork+exec, this should be safe to allow switching timens with the current clone flag. Right now, vfork (CLONE_VFORK | CLONE_VM) fails if a child is forked into another time namespace. With this change, vfork creates a new process in parent's timens, and the following exec does the actual switch to the target time namespace. Suggested-by: Florian Weimer <fweimer@redhat.com> Signed-off-by: Andrei Vagin <avagin@gmail.com> --- fs/exec.c | 7 +++++++ kernel/fork.c | 5 ++++- kernel/nsproxy.c | 3 ++- 3 files changed, 13 insertions(+), 2 deletions(-)