unshare 시스템 호출

문서 범위와 변경 이력

이 문서는 `unshare()` system call의 개요, 필요성, 사용법, interface 명세, 설계·구현과 시험 방법을 설명합니다.

문서 버전 0.1은 Janak Desai가 2006년 1월 11일 작성한 최초 문서입니다.

unshare system call
===================

This document describes the new system call, unshare(). The document
provides an overview of the feature, why it is needed, how it can
be used, its interface specification, design, implementation and
how it can be tested.

Change Log
----------
version 0.1  Initial document, Janak Desai (janak@us.ibm.com), Jan 11, 2006

Contents
--------
	1) Overview
	2) Benefits
	3) Cost
	4) Requirements
	5) Functional Specification
	6) High Level Design
	7) Low Level Design
	8) Test Specification
	9) Future Work

1. 개요

전통적인 운영체제 커널은 프로세스 안의 여러 실행 context를 thread로 추상화하고 이를 유지할 별도 resource와 메커니즘을 둡니다. Linux는 프로세스와 thread를 구분하지 않고 프로세스가 resource를 공유하도록 해 추가 구조 없이 전통적인 thread 동작을 구현합니다.

이 방식은 단순할 뿐 아니라 애플리케이션이 전통적인 thread의 전부 공유 또는 전부 비공유 제약을 벗어나게 합니다. `clone`으로 thread를 만들 때 어떤 resource를 서로 공유할지 선택할 수 있습니다.

`unshare()`는 생성 시 공유했던 resource를 실행 중인 thread가 선택적으로 분리할 수 있게 Linux thread model에 추가된 primitive입니다. Al Viro가 2000년 8월 Linux-Kernel mailing list의 POSIX thread 논의에서 구상했습니다.

새 프로세스를 만들지 않고 공유 resource를 제어하려는 애플리케이션에 유용하며, process/thread를 virtual machine으로 보는 Linux primitive 집합을 자연스럽게 확장합니다.

1) Overview
-----------

Most legacy operating system kernels support an abstraction of threads
as multiple execution contexts within a process. These kernels provide
special resources and mechanisms to maintain these "threads". The Linux
kernel, in a clever and simple manner, does not make distinction
between processes and "threads". The kernel allows processes to share
resources and thus they can achieve legacy "threads" behavior without
requiring additional data structures and mechanisms in the kernel. The
power of implementing threads in this manner comes not only from
its simplicity but also from allowing application programmers to work
outside the confinement of all-or-nothing shared resources of legacy
threads. On Linux, at the time of thread creation using the clone system
call, applications can selectively choose which resources to share
between threads.

unshare() system call adds a primitive to the Linux thread model that
allows threads to selectively 'unshare' any resources that were being
shared at the time of their creation. unshare() was conceptualized by
Al Viro in the August of 2000, on the Linux-Kernel mailing list, as part
of the discussion on POSIX threads on Linux.  unshare() augments the
usefulness of Linux threads for applications that would like to control
shared resources without creating a new process. unshare() is a natural
addition to the set of available primitives on Linux that implement
the concept of process/thread as a virtual machine.

2. 이점과 보안 context별 namespace

PAM 같은 큰 애플리케이션 framework는 process resource의 공유 상태를 바꾸기 위해 새 프로세스를 만들기 어려울 수 있어 `unshare()`가 유용합니다. fork나 clone으로 만든 프로세스는 기본적으로 namespace를 공유하므로 thread를 쓰지 않는 프로그램도 기본 namespace에서 분리할 필요가 있다면 이 기능을 활용할 수 있습니다.

`unshare()`와 per-process namespace로 polyinstantiated directory를 구현할 수 있습니다. 사용자별 또는 보안 context별 `/tmp`, `/var/tmp`, home directory 인스턴스가 해당 디렉터리에서 작업하는 사용자 프로세스를 격리합니다.

PAM module은 로그인 때 사용자의 private namespace를 쉽게 설정할 수 있습니다. 이 기능은 Labeled System Protection Profile을 적용한 Common Criteria 인증에 필요하며, shared-tree 기능을 이용하면 일반 Linux 시스템도 로그인 때 private namespace와 관리자 지정 directory의 개별 인스턴스를 구성할 수 있습니다.

2) Benefits
-----------

unshare() would be useful to large application frameworks such as PAM
where creating a new process to control sharing/unsharing of process
resources is not possible. Since namespaces are shared by default
when creating a new process using fork or clone, unshare() can benefit
even non-threaded applications if they have a need to disassociate
from default shared namespace. The following lists two use-cases
where unshare() can be used.

2.1 Per-security context namespaces
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

unshare() can be used to implement polyinstantiated directories using
the kernel's per-process namespace mechanism. Polyinstantiated directories,
such as per-user and/or per-security context instance of /tmp, /var/tmp or
per-security context instance of a user's home directory, isolate user
processes when working with these directories. Using unshare(), a PAM
module can easily setup a private namespace for a user at login.
Polyinstantiated directories are required for Common Criteria certification
with Labeled System Protection Profile, however, with the availability
of shared-tree feature in the Linux kernel, even regular Linux systems
can benefit from setting up private namespaces at login and
polyinstantiating /tmp, /var/tmp and other directories deemed
appropriate by system administrators.

2.2 가상 메모리와 열린 파일 분리

server가 virtual memory와 open file 같은 resource를 공유하는 프로세스를 만들어 client 요청을 처리하는 애플리케이션을 생각할 수 있습니다. `unshare()`가 없으면 요청 처리 프로세스를 만들 때 무엇을 공유할지 미리 결정해야 합니다.

`unshare()`는 요청 처리 도중 context 일부를 분리할 수 있게 합니다. 생성 뒤에 공유 상태를 바꾸는 능력은 크고 복잡한 middleware framework에 특히 유용합니다.

2.2 unsharing of virtual memory and/or open files
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Consider a client/server application where the server is processing
client requests by creating processes that share resources such as
virtual memory and open files. Without unshare(), the server has to
decide what needs to be shared at the time of creating the process
which services the request. unshare() allows the server an ability to
disassociate parts of the context during the servicing of the
request. For large and complex middleware application frameworks, this
ability to unshare() after the process was created can be very
useful.

3. 구현 비용

`unshare()`는 새로 할당되어 아직 활성화되지 않은 태스크를 다루는 clone/fork와 달리 현재 활성 태스크에 작동합니다. 코드를 중복하지 않으면서 이 차이를 처리하기 위해 clone/fork가 쓰는 `copy_*` 함수에 작은 구조 변경이 필요했습니다.

검증되고 안정적인 기존 코드를 초기 사용량이 적을 수 있는 새 기능 때문에 바꾸는 비용이 있지만, 적절한 설계·code review와 LTP용 `unshare()` 시험을 갖추면 이점이 비용을 넘을 수 있습니다.

3) Cost
-------

In order to not duplicate code and to handle the fact that unshare()
works on an active task (as opposed to clone/fork working on a newly
allocated inactive task) unshare() had to make minor reorganizational
changes to copy_* functions utilized by clone/fork system call.
There is a cost associated with altering existing, well tested and
stable code to implement a new feature that may not get exercised
extensively in the beginning. However, with proper design and code
review of the changes and creation of an unshare() test for the LTP
the benefits of this new feature can exceed its cost.

4. Interface 요구사항

`unshare()`는 `clone(2)`이 만든 공유를 되돌리므로 비슷한 interface를 가져야 합니다. `clone(int flags, void *stack)`의 flag가 공유할 항목을 지정하듯 `unshare(int flags)`의 같은 종류 flag는 분리할 항목을 지정합니다. clone과 비교하면 flag 의미가 뒤집힌 것처럼 보이지만, ABI 변경 없이 향후 context를 점진적으로 분리할 수 있는 덜 혼란스러운 대안이 없었습니다.

interface는 새 context flag가 추가돼도 기존 애플리케이션을 rebuild하지 않도록 해야 합니다. 새 resource도 필요할 때 점진적으로 분리할 수 있어야 합니다.

4) Requirements
---------------

unshare() reverses sharing that was done using clone(2) system call,
so unshare() should have a similar interface as clone(2). That is,
since flags in clone(int flags, void \*stack) specifies what should
be shared, similar flags in unshare(int flags) should specify
what should be unshared. Unfortunately, this may appear to invert
the meaning of the flags from the way they are used in clone(2).
However, there was no easy solution that was less confusing and that
allowed incremental context unsharing in future without an ABI change.

unshare() interface should accommodate possible future addition of
new context flags without requiring a rebuild of old applications.
If and when new context flags are added, unshare() design should allow
incremental unsharing of those resources on an as needed basis.

5. 기능 명세

이름은 `unshare`이며 프로세스 실행 context 일부의 연결을 끊습니다.

#include <sched.h>

int unshare(int flags);

`unshare()`는 현재 다른 프로세스와 공유하는 실행 context 일부를 분리합니다. namespace처럼 fork 시 기본 공유되는 부분도 있고, virtual memory나 open file descriptor처럼 clone 시 명시적으로 요청해야 공유되는 부분도 있습니다. 주된 용도는 새 프로세스를 만들지 않고 현재 프로세스가 공유 context를 제어하는 것입니다.

성공하면 0, 실패하면 -1을 반환하고 errno를 설정합니다.

`unshare()`는 Linux 전용이므로 portable 프로그램에 사용하면 안 됩니다. 관련 호출은 `clone(2)`과 `fork(2)`입니다.

5) Functional Specification
---------------------------

NAME
	unshare - disassociate parts of the process execution context

SYNOPSIS
	#include <sched.h>

	int unshare(int flags);

DESCRIPTION
	unshare() allows a process to disassociate parts of its execution
	context that are currently being shared with other processes. Part
	of execution context, such as the namespace, is shared by default
	when a new process is created using fork(2), while other parts,
	such as the virtual memory, open file descriptors, etc, may be
	shared by explicit request to share them when creating a process
	using clone(2).

	The main use of unshare() is to allow a process to control its
	shared execution context without creating a new process.

	The flags argument specifies one or bitwise-or'ed of several of
	the following constants.

	CLONE_FS
		If CLONE_FS is set, file system information of the caller
		is disassociated from the shared file system information.

	CLONE_FILES
		If CLONE_FILES is set, the file descriptor table of the
		caller is disassociated from the shared file descriptor
		table.

	CLONE_NEWNS
		If CLONE_NEWNS is set, the namespace of the caller is
		disassociated from the shared namespace.

	CLONE_VM
		If CLONE_VM is set, the virtual memory of the caller is
		disassociated from the shared virtual memory.

RETURN VALUE
	On success, zero returned. On failure, -1 is returned and errno is

ERRORS
	EPERM	CLONE_NEWNS was specified by a non-root process (process
		without CAP_SYS_ADMIN).

	ENOMEM	Cannot allocate sufficient memory to copy parts of caller's
		context that need to be unshared.

	EINVAL	Invalid flag was specified as an argument.

CONFORMING TO
	The unshare() call is Linux-specific and  should  not be used
	in programs intended to be portable.

SEE ALSO
	clone(2), fork(2)

6. 상위 수준 설계

flag에 따라 `unshare()`는 적절한 process context 구조를 할당하고 현재 공유 구조의 값으로 채운 뒤, 새 복제본을 current task에 연결하고 이전 공유 구조를 해제합니다.

clone의 `copy_*` helper를 직접 사용할 수 없는 첫 이유는 clone은 새로 할당된 비활성 태스크를 다루지만 unshare는 현재 활성 태스크에 작동한다는 점입니다. 새 context를 연결할 때 알맞은 `task_lock()`이 필요합니다.

둘째, 현재 태스크에 새 구조를 연결하고 이전 공유 구조를 해제하기 전에 분리 대상 context 전체를 먼저 할당·복제해야 합니다. 그렇지 않으면 오류 rollback에서 경쟁 조건이나 oops가 생깁니다.

예를 들어 virtual memory와 namespace를 함께 분리할 때 VM 분리 뒤 namespace 할당이 실패하면 VM 분리를 되돌려야 합니다. 그러나 돌아갈 이전 공유 VM 구조가 이미 사라졌을 수 있습니다.

따라서 `copy_*`의 할당·복제 부분을 새 `dup_*` 함수로 옮겼습니다. `copy_*`는 `dup_*`로 구조를 만든 뒤 새 태스크에 연결하고, `unshare()`는 모든 복제가 성공한 뒤 한꺼번에 현재 태스크에 연결합니다.

6) High Level Design
--------------------

Depending on the flags argument, the unshare() system call allocates
appropriate process context structures, populates it with values from
the current shared version, associates newly duplicated structures
with the current task structure and releases corresponding shared
versions. Helper functions of clone (copy_*) could not be used
directly by unshare() because of the following two reasons.

  1) clone operates on a newly allocated not-yet-active task
     structure, where as unshare() operates on the current active
     task. Therefore unshare() has to take appropriate task_lock()
     before associating newly duplicated context structures

  2) unshare() has to allocate and duplicate all context structures
     that are being unshared, before associating them with the
     current task and releasing older shared structures. Failure
     do so will create race conditions and/or oops when trying
     to backout due to an error. Consider the case of unsharing
     both virtual memory and namespace. After successfully unsharing
     vm, if the system call encounters an error while allocating
     new namespace structure, the error return code will have to
     reverse the unsharing of vm. As part of the reversal the
     system call will have to go back to older, shared, vm
     structure, which may not exist anymore.

Therefore code from copy_* functions that allocated and duplicated
current context structure was moved into new dup_* functions. Now,
copy_* functions call dup_* functions to allocate and duplicate
appropriate context structures and then associate them with the
task structure that is being constructed. unshare() system call on
the other hand performs the following:

  1) Check flags to force missing, but implied, flags

  2) For each context structure, call the corresponding unshare()
     helper function to allocate and duplicate a new context
     structure, if the appropriate bit is set in the flags argument.

  3) If there is no error in allocation and duplication and there
     are new context structures then lock the current task structure,
     associate new context structures with the current task structure,
     and release the lock on the current task structure.

  4) Appropriately release older, shared, context structures.

7. 하위 수준 설계 개요

구현 작업은 기존 `copy_*` 함수 재구성, `unshare()` system call service, process context별 `unshare_*` helper, 아키텍처별 system call 번호 등록의 네 항목으로 나뉩니다.

7) Low Level Design
-------------------

Implementation of unshare() can be grouped in the following 4 different
items:

  a) Reorganization of existing copy_* functions

  b) unshare() system call service function

  c) unshare() helper functions for each different process context

  d) Registration of system call number for different architectures

7.1 copy_* 함수 재구성

`copy_mm`, `copy_namespace`, `copy_files` 같은 기존 copy 함수는 적절한 구조를 할당·복제하는 부분과 인자로 받은 task 구조에 연결하는 부분으로 이루어졌습니다.

첫 부분을 독립 `dup_*` 함수로 분리했습니다. 재구성된 `copy_*`는 대응 `dup_*`를 호출한 뒤 새 복제 구조를 자신이 받은 task에 연결합니다.

7.1) Reorganization of copy_* functions
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Each copy function such as copy_mm, copy_namespace, copy_files,
etc, had roughly two components. The first component allocated
and duplicated the appropriate structure and the second component
linked it to the task structure passed in as an argument to the copy
function. The first component was split into its own function.
These dup_* functions allocated and duplicated the appropriate
context structure. The reorganized copy_* functions invoked
their corresponding dup_* functions and then linked the newly
duplicated structures to the task structure with which the
copy function was called.

7.2 unshare service function

service는 먼저 flag를 검사하고 암시된 flag를 강제합니다. `CLONE_THREAD`이면 `CLONE_VM`, `CLONE_VM`이면 `CLONE_SIGHAND`, signal도 공유 중인 상태에서 `CLONE_SIGHAND`이면 `CLONE_THREAD`, `CLONE_NEWNS`이면 `CLONE_FS`를 함께 적용합니다.

각 context flag에 대해 system call flag와 새 구조 포인터를 받을 위치를 대응 `unshare_*` helper에 전달합니다.

helper가 새 구조를 만들었으면 current task에 `task_lock()`을 걸고 적절한 context pointer를 바꾼 뒤 lock을 풉니다. 마지막으로 새로 분리된 모든 항목의 이전 공유 구조를 해제합니다.

7.2) unshare() system call service function
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

       * Check flags
	 Force implied flags. If CLONE_THREAD is set force CLONE_VM.
	 If CLONE_VM is set, force CLONE_SIGHAND. If CLONE_SIGHAND is
	 set and signals are also being shared, force CLONE_THREAD. If
	 CLONE_NEWNS is set, force CLONE_FS.

       * For each context flag, invoke the corresponding unshare_*
	 helper routine with flags passed into the system call and a
	 reference to pointer pointing the new unshared structure

       * If any new structures are created by unshare_* helper
	 functions, take the task_lock() on the current task,
	 modify appropriate context pointers, and release the
         task lock.

       * For all newly unshared structures, release the corresponding
         older, shared, structures.

7.3 helper와 아키텍처 등록

문서 작성 시점에 `CLONE_SYSVSEM`, `CLONE_SIGHAND`, `CLONE_THREAD`에 대응하는 `unshare_*` helper는 구현되지 않아 `-EINVAL`을 반환합니다.

다른 helper는 해당 구조를 분리해야 하는지 flag를 검사하고, 필요하면 대응 `dup_*` 함수로 구조를 할당·복제해 pointer를 반환합니다.

마지막으로 새 system call을 등록하도록 아키텍처별 코드를 적절히 수정합니다.

7.3) unshare_* helper functions
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

For unshare_* helpers corresponding to CLONE_SYSVSEM, CLONE_SIGHAND,
and CLONE_THREAD, return -EINVAL since they are not implemented yet.
For others, check the flag value to see if the unsharing is
required for that structure. If it is, invoke the corresponding
dup_* function to allocate and duplicate the structure and return
a pointer to it.

7.4) Finally
~~~~~~~~~~~~

Appropriately modify architecture specific code to register the
new system call.

8. 시험 명세

8) Test Specification
---------------------

The test for unshare() should test the following:

  1) Valid flags: Test to check that clone flags for signal and
     signal handlers, for which unsharing is not implemented
     yet, return -EINVAL.

  2) Missing/implied flags: Test to make sure that if unsharing
     namespace without specifying unsharing of filesystem, correctly
     unshares both namespace and filesystem information.

  3) For each of the four (namespace, filesystem, files and vm)
     supported unsharing, verify that the system call correctly
     unshares the appropriate structure. Verify that unsharing
     them individually as well as in combination with each
     other works as expected.

  4) Concurrent execution: Use shared memory segments and futex on
     an address in the shm segment to synchronize execution of
     about 10 threads. Have a couple of threads execute execve,
     a couple _exit and the rest unshare with different combination
     of flags. Verify that unsharing is performed as expected and
     that there are no oops or hangs.

9. 향후 작업

현재 구현은 signal과 signal handler 분리를 허용하지 않습니다. signal 자체도 복잡하고 실행 중인 프로세스의 signal 상태를 분리하는 일은 더 복잡합니다.

향후 구체적인 필요가 생기면 기존 `unshare()` 애플리케이션에 영향을 주지 않고 이 기능을 점진적으로 추가할 수 있습니다.

9) Future Work
--------------

The current implementation of unshare() does not allow unsharing of
signals and signal handlers. Signals are complex to begin with and
to unshare signals and/or signal handlers of a currently running
process is even more complex. If in the future there is a specific
need to allow unsharing of signals and/or signal handlers, it can
be incrementally added to unshare() without affecting legacy
applications using unshare().