capset

Now that you have been warned, some current kernel details. The structures are defined as follows.

#define _LINUX_CAPABILITY_VERSION_1  0x19980330
#define _LINUX_CAPABILITY_U32S_1     1

        /* V2 added in Linux 2.6.25; deprecated */
#define _LINUX_CAPABILITY_VERSION_2  0x20071026
#define _LINUX_CAPABILITY_U32S_2     2

        /* V3 added in Linux 2.6.26 */
#define _LINUX_CAPABILITY_VERSION_3  0x20080522
#define _LINUX_CAPABILITY_U32S_3     2

typedef struct __user_cap_header_struct {
   __u32 version;
   int pid;
} *cap_user_header_t;

typedef struct __user_cap_data_struct {
   __u32 effective;
   __u32 permitted;
   __u32 inheritable;
} *cap_user_data_t;

The effective, permitted, and inheritable fields are bit masks of the capabilities defined in capabilities(7). Note that the CAP_* values are bit indexes and need to be bit-shifted before ORing into the bit fields. To define the structures for passing to the system call, you have to use the struct __user_cap_header_struct and struct __user_cap_data_struct names because the typedefs are only pointers.

Kernels prior to 2.6.25 prefer 32-bit capabilities with version _LINUX_CAPABILITY_VERSION_1. Linux 2.6.25 added 64-bit capability sets, with version _LINUX_CAPABILITY_VERSION_2. There was, however, an API glitch, and Linux 2.6.26 added _LINUX_CAPABILITY_VERSION_3 to fix the problem.

Note that 64-bit capabilities use datap[0] and datap[1], whereas 32-bit capabilities use only datap[0].

On kernels that support file capabilities (VFS capabilities support), these system calls behave slightly differently. This support was added as an option in Linux 2.6.24, and became fixed (nonoptional) in Linux 2.6.33.

For capget() calls, one can probe the capabilities of any process by specifying its process ID with the hdrp->pid field value.

For details on the data, see capabilities(7).

VFS capabilities employ a file extended attribute (see xattr(7)) to allow capabilities to be attached to executables. This privilege model obsoletes kernel support for one process asynchronously setting the capabilities of another. That is, on kernels that have VFS capabilities support, when calling capset(), the only permitted values for hdrp->pid are 0 or, equivalently, the value returned by gettid(2).

On older kernels that do not provide VFS capabilities support capset() can, if the caller has the CAP_SETPCAP capability, be used to change not only the caller's own capabilities, but also the capabilities of other threads. The call operates on the capabilities of the thread specified by the pid field of hdrp when that is nonzero, or on the capabilities of the calling thread if pid is 0. If pid refers to a single-threaded process, then pid can be specified as a traditional process ID; operating on a thread of a multithreaded process requires a thread ID of the type returned by gettid(2). For capset(), pid can also be: -1, meaning perform the change on all threads except the caller and init(1); or a value less than -1, in which case the change is applied to all members of the process group whose ID is -pid.