RUMP(3) BSD Library Functions Manual RUMP(3)NAMErump — The Rump Anykernel
LIBRARYrump Library (librump, -lrump)
SYNOPSIS
#include <rump/rump.h>
#include <rump/rump_syscalls.h>
DESCRIPTIONrump is part of the realization of a flexible anykernel architecture for
NetBSD. An anykernel architecture enables using kernel code in a number
of different kernel models. These models include, but are not limited
to, the original monolithic kernel, a microkernel server, or an exokernel
style application library. rump itself makes it possible to run unmodi‐
fied kernel components in a regular userspace process. Most of the time
"unmodified" means unmodified source code, but some architectures can
also execute unmodified kernel module binaries in userspace. Examples of
different use models are running file system drivers as userspace servers
(see p2k(3)) and being able to write standalone applications which under‐
stand file system images.
Regardless of the kernel model used, a rump kernel is a fullfledged ker‐
nel with its own virtual namespaces, including a file system hierarchy,
CPUs, TCP/UDP ports, device driver attachments and file descriptors.
This means that any modification to the system state on the host running
the rump kernel will not show up in the rump kernel and vice versa. A
rump kernel may also be significantly more lightweight than the host, and
might not include for example file system support at all.
Clients using services provided by rump kernels can exist either in the
same process as the rump kernel or in other processes. Local clients
access the rump kernel through direct function calls. They also natu‐
rally have access to the kernel memory space. This document is geared
towards local clients. For more information on remote clients, see
rump_sp(7). It is also possible to use unmodified application binaries
as remote clients with rumphijack(3).
A rump kernel is bootstrapped by calling rump_init(). Before bootstrap‐
ping the kernel, it is possible to control its functionality by setting
various environment variables:
RUMP_NCPU If set, indicates the number of virtual CPUs configured
into a rump kernel. The default is the number of host
CPUs. The number of virtual CPUs controls how many
threads can enter the rump kernel simultaneously.
RUMP_VERBOSE If set to non-zero, activates bootverbose.
RUMP_THREADS If set to 0, prevents the rump kernel from creating any
kernel threads. This is possible usually only for file
systems, as other subsystems depend on threads to work.
RUMP_MEMLIMIT If set, indicates how many bytes of memory a rump kernel
will allocate before attempting to purge caches. The
default is as much as the host allows.
RUMP_NVNODES Sets the value of the kern.maxvnodes sysctl node to the
indicated amount. Adjusting this may be useful for
example when testing vnode reclaim code paths. While
the same value can be set by means of sysctl, the env
variable is often more convenient for quick testing. As
expected, this option has effect only in rump kernels
which support VFS. The current default is 1024 vnodes.
A number of interfaces are available for requesting services from a rump
kernel. The most commonly used ones are the rump system calls. They are
exactly like regular system calls but with the exception that they target
the rump kernel of the current process instead of the host kernel. For
example, rump_sys_socket() takes the same parameters as socket() and will
open a socket in the rump kernel. The resulting file descriptor may be
used only in other rump system calls and will have undefined results if
passed to the host kernel.
Another set of interfaces specifically crafted for rump kernels are the
rump public calls. These calls reside in the rump_pub namespace. An
example is rump_pub_module_init() which initializes a prelinked kernel
module.
A rump kernel is constructed at build time by linking a set of libraries
with application level code. The mandatory libraries are the kernel base
(librump) and the rump hypercall library (librumpuser) which a rump ker‐
nel uses to request services from the host. Beyond that, there are three
factions which define the flavour of a rump kernel (librumpdev, librump‐
net and librumpvfs) and driver components which use features provided by
the base and factions. Notably, components may have interdependencies.
For example, a rump kernel providing a virtual IP router requires the
following components: rumpnet_netinet, rumpnet_net, rumpnet, rumpnet_vir‐
tif, rump, and rumpuser. A rump kernel providing an NFS client requires
the above and additionally rumpfs_nfs and rumpvfs.
In addition to defining the configuration at link time, it is also possi‐
ble to load components at runtime. There are two ways of doing this:
using dlopen() to link a shared library into a rump kernel and initializ‐
ing with rump_pub_module_init() or specifying a module on the file system
to rump_sys_modctl() and letting the rump kernel do the linking.
Notably, in the latter case debugging with symbols is not possible since
the host gdb does not know about symbols loaded by the rump kernel. Gen‐
erally speaking, dynamically loadable components must follow kernel mod‐
ule boundaries.
SEE ALSOrump_server(1), p2k(3), rump_etfs(3), rump_lwproc(3), rumpclient(3),
rumphijack(3), rumpuser(3), ukfs(3), rump_sp(7)
Antti Kantee, "Environmental Independence: BSD Kernel TCP/IP in
Userspace", Proceedings of AsiaBSDCon 2009, pp. 71-80, March 2009.
Antti Kantee, "Kernel Development in Userspace - The Rump Approach",
BSDCan 2009, May 2009.
Antti Kantee, "Rump File Systems: Kernel Code Reborn", Proceedings of the
2009 USENIX Annual Technical Conference, pp. 201-214, June 2009.
Arnaud Ysmal and Antti Kantee, "Fs-utils: File Systems Access Tools for
Userland", EuroBSDCon 2009, September 2009.
Antti Kantee, "Rump Device Drivers: Shine On You Kernel Diamond",
Proceedings of AsiaBSDCon 2010, pp. 75-84, March 2010.
HISTORYrump appeared as an experimental concept in NetBSD 5.0. The first stable
version was released in NetBSD 6.0.
AUTHORS
Antti Kantee ⟨pooka@iki.fi⟩
BSD March 25, 2011 BSD
