4 This document describes the new system call, unshare(). The document
5 provides an overview of the feature, why it is needed, how it can
6 be used, its interface specification, design, implementation and
11 version 0.1 Initial document, Janak Desai (janak@us.ibm.com), Jan 11, 2006
19 5) Functional Specification
28 Most legacy operating system kernels support an abstraction of threads
29 as multiple execution contexts within a process. These kernels provide
30 special resources and mechanisms to maintain these "threads". The Linux
31 kernel, in a clever and simple manner, does not make distinction
32 between processes and "threads". The kernel allows processes to share
33 resources and thus they can achieve legacy "threads" behavior without
34 requiring additional data structures and mechanisms in the kernel. The
35 power of implementing threads in this manner comes not only from
36 its simplicity but also from allowing application programmers to work
37 outside the confinement of all-or-nothing shared resources of legacy
38 threads. On Linux, at the time of thread creation using the clone system
39 call, applications can selectively choose which resources to share
42 unshare() system call adds a primitive to the Linux thread model that
43 allows threads to selectively 'unshare' any resources that were being
44 shared at the time of their creation. unshare() was conceptualized by
45 Al Viro in the August of 2000, on the Linux-Kernel mailing list, as part
46 of the discussion on POSIX threads on Linux. unshare() augments the
47 usefulness of Linux threads for applications that would like to control
48 shared resources without creating a new process. unshare() is a natural
49 addition to the set of available primitives on Linux that implement
50 the concept of process/thread as a virtual machine.
55 unshare() would be useful to large application frameworks such as PAM
56 where creating a new process to control sharing/unsharing of process
57 resources is not possible. Since namespaces are shared by default
58 when creating a new process using fork or clone, unshare() can benefit
59 even non-threaded applications if they have a need to disassociate
60 from default shared namespace. The following lists two use-cases
61 where unshare() can be used.
63 2.1 Per-security context namespaces
64 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
66 unshare() can be used to implement polyinstantiated directories using
67 the kernel's per-process namespace mechanism. Polyinstantiated directories,
68 such as per-user and/or per-security context instance of /tmp, /var/tmp or
69 per-security context instance of a user's home directory, isolate user
70 processes when working with these directories. Using unshare(), a PAM
71 module can easily setup a private namespace for a user at login.
72 Polyinstantiated directories are required for Common Criteria certification
73 with Labeled System Protection Profile, however, with the availability
74 of shared-tree feature in the Linux kernel, even regular Linux systems
75 can benefit from setting up private namespaces at login and
76 polyinstantiating /tmp, /var/tmp and other directories deemed
77 appropriate by system administrators.
79 2.2 unsharing of virtual memory and/or open files
80 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
82 Consider a client/server application where the server is processing
83 client requests by creating processes that share resources such as
84 virtual memory and open files. Without unshare(), the server has to
85 decide what needs to be shared at the time of creating the process
86 which services the request. unshare() allows the server an ability to
87 disassociate parts of the context during the servicing of the
88 request. For large and complex middleware application frameworks, this
89 ability to unshare() after the process was created can be very
95 In order to not duplicate code and to handle the fact that unshare()
96 works on an active task (as opposed to clone/fork working on a newly
97 allocated inactive task) unshare() had to make minor reorganizational
98 changes to copy_* functions utilized by clone/fork system call.
99 There is a cost associated with altering existing, well tested and
100 stable code to implement a new feature that may not get exercised
101 extensively in the beginning. However, with proper design and code
102 review of the changes and creation of an unshare() test for the LTP
103 the benefits of this new feature can exceed its cost.
108 unshare() reverses sharing that was done using clone(2) system call,
109 so unshare() should have a similar interface as clone(2). That is,
110 since flags in clone(int flags, void \*stack) specifies what should
111 be shared, similar flags in unshare(int flags) should specify
112 what should be unshared. Unfortunately, this may appear to invert
113 the meaning of the flags from the way they are used in clone(2).
114 However, there was no easy solution that was less confusing and that
115 allowed incremental context unsharing in future without an ABI change.
117 unshare() interface should accommodate possible future addition of
118 new context flags without requiring a rebuild of old applications.
119 If and when new context flags are added, unshare() design should allow
120 incremental unsharing of those resources on an as needed basis.
122 5) Functional Specification
123 ---------------------------
126 unshare - disassociate parts of the process execution context
131 int unshare(int flags);
134 unshare() allows a process to disassociate parts of its execution
135 context that are currently being shared with other processes. Part
136 of execution context, such as the namespace, is shared by default
137 when a new process is created using fork(2), while other parts,
138 such as the virtual memory, open file descriptors, etc, may be
139 shared by explicit request to share them when creating a process
142 The main use of unshare() is to allow a process to control its
143 shared execution context without creating a new process.
145 The flags argument specifies one or bitwise-or'ed of several of
146 the following constants.
149 If CLONE_FS is set, file system information of the caller
150 is disassociated from the shared file system information.
153 If CLONE_FILES is set, the file descriptor table of the
154 caller is disassociated from the shared file descriptor
158 If CLONE_NEWNS is set, the namespace of the caller is
159 disassociated from the shared namespace.
162 If CLONE_VM is set, the virtual memory of the caller is
163 disassociated from the shared virtual memory.
166 On success, zero returned. On failure, -1 is returned and errno is
169 EPERM CLONE_NEWNS was specified by a non-root process (process
170 without CAP_SYS_ADMIN).
172 ENOMEM Cannot allocate sufficient memory to copy parts of caller's
173 context that need to be unshared.
175 EINVAL Invalid flag was specified as an argument.
178 The unshare() call is Linux-specific and should not be used
179 in programs intended to be portable.
187 Depending on the flags argument, the unshare() system call allocates
188 appropriate process context structures, populates it with values from
189 the current shared version, associates newly duplicated structures
190 with the current task structure and releases corresponding shared
191 versions. Helper functions of clone (copy_*) could not be used
192 directly by unshare() because of the following two reasons.
194 1) clone operates on a newly allocated not-yet-active task
195 structure, where as unshare() operates on the current active
196 task. Therefore unshare() has to take appropriate task_lock()
197 before associating newly duplicated context structures
199 2) unshare() has to allocate and duplicate all context structures
200 that are being unshared, before associating them with the
201 current task and releasing older shared structures. Failure
202 do so will create race conditions and/or oops when trying
203 to backout due to an error. Consider the case of unsharing
204 both virtual memory and namespace. After successfully unsharing
205 vm, if the system call encounters an error while allocating
206 new namespace structure, the error return code will have to
207 reverse the unsharing of vm. As part of the reversal the
208 system call will have to go back to older, shared, vm
209 structure, which may not exist anymore.
211 Therefore code from copy_* functions that allocated and duplicated
212 current context structure was moved into new dup_* functions. Now,
213 copy_* functions call dup_* functions to allocate and duplicate
214 appropriate context structures and then associate them with the
215 task structure that is being constructed. unshare() system call on
216 the other hand performs the following:
218 1) Check flags to force missing, but implied, flags
220 2) For each context structure, call the corresponding unshare()
221 helper function to allocate and duplicate a new context
222 structure, if the appropriate bit is set in the flags argument.
224 3) If there is no error in allocation and duplication and there
225 are new context structures then lock the current task structure,
226 associate new context structures with the current task structure,
227 and release the lock on the current task structure.
229 4) Appropriately release older, shared, context structures.
234 Implementation of unshare() can be grouped in the following 4 different
237 a) Reorganization of existing copy_* functions
239 b) unshare() system call service function
241 c) unshare() helper functions for each different process context
243 d) Registration of system call number for different architectures
245 7.1) Reorganization of copy_* functions
246 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
248 Each copy function such as copy_mm, copy_namespace, copy_files,
249 etc, had roughly two components. The first component allocated
250 and duplicated the appropriate structure and the second component
251 linked it to the task structure passed in as an argument to the copy
252 function. The first component was split into its own function.
253 These dup_* functions allocated and duplicated the appropriate
254 context structure. The reorganized copy_* functions invoked
255 their corresponding dup_* functions and then linked the newly
256 duplicated structures to the task structure with which the
257 copy function was called.
259 7.2) unshare() system call service function
260 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
263 Force implied flags. If CLONE_THREAD is set force CLONE_VM.
264 If CLONE_VM is set, force CLONE_SIGHAND. If CLONE_SIGHAND is
265 set and signals are also being shared, force CLONE_THREAD. If
266 CLONE_NEWNS is set, force CLONE_FS.
268 * For each context flag, invoke the corresponding unshare_*
269 helper routine with flags passed into the system call and a
270 reference to pointer pointing the new unshared structure
272 * If any new structures are created by unshare_* helper
273 functions, take the task_lock() on the current task,
274 modify appropriate context pointers, and release the
277 * For all newly unshared structures, release the corresponding
278 older, shared, structures.
280 7.3) unshare_* helper functions
281 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
283 For unshare_* helpers corresponding to CLONE_SYSVSEM, CLONE_SIGHAND,
284 and CLONE_THREAD, return -EINVAL since they are not implemented yet.
285 For others, check the flag value to see if the unsharing is
286 required for that structure. If it is, invoke the corresponding
287 dup_* function to allocate and duplicate the structure and return
293 Appropriately modify architecture specific code to register the
296 8) Test Specification
297 ---------------------
299 The test for unshare() should test the following:
301 1) Valid flags: Test to check that clone flags for signal and
302 signal handlers, for which unsharing is not implemented
305 2) Missing/implied flags: Test to make sure that if unsharing
306 namespace without specifying unsharing of filesystem, correctly
307 unshares both namespace and filesystem information.
309 3) For each of the four (namespace, filesystem, files and vm)
310 supported unsharing, verify that the system call correctly
311 unshares the appropriate structure. Verify that unsharing
312 them individually as well as in combination with each
313 other works as expected.
315 4) Concurrent execution: Use shared memory segments and futex on
316 an address in the shm segment to synchronize execution of
317 about 10 threads. Have a couple of threads execute execve,
318 a couple _exit and the rest unshare with different combination
319 of flags. Verify that unsharing is performed as expected and
320 that there are no oops or hangs.
325 The current implementation of unshare() does not allow unsharing of
326 signals and signal handlers. Signals are complex to begin with and
327 to unshare signals and/or signal handlers of a currently running
328 process is even more complex. If in the future there is a specific
329 need to allow unsharing of signals and/or signal handlers, it can
330 be incrementally added to unshare() without affecting legacy
331 applications using unshare().