VNODE(9) BSD Kernel Manual VNODE(9)
vnode - an overview of vnodes
A vnode is an object in kernel memory that speaks the UNIX file interface
(open, read, write, close, readdir, etc.). Vnodes can represent files,
directories, FIFOs, domain sockets, block devices, character devices.
Each vnode has a set of methods which start with string 'VOP_'. These
methods include VOP_OPEN, VOP_READ, VOP_WRITE, VOP_RENAME, VOP_CLOSE,
VOP_MKDIR. Many of these methods correspond closely to the equivalent
filesystem call - open, read, write, rename, etc. Each filesystem (FFS,
NFS, etc.) provides implementations for these methods.
The Virtual Filesystem (VFS) library maintains a pool of vnodes. Filesys-
tems cannot allocate their own vnodes; they must use the functions pro-
vided by the VFS to create and manage vnodes.
When a client of the VFS requests a new vnode, the vnode allocation code
can reuse an old vnode object that is no longer in use. Whether a vnode
is in use is tracked by the vnode reference count (v_usecount). By con-
vention, each open file handle holds a reference as do VM objects backed
by files. A vnode with a reference count of 1 or more will not be de-
allocated or re-used to point to a different file. So, if you want to en-
sure that your vnode doesn't become a different file under you, you
better be sure you have a reference to it. A vnode that points to a valid
file and has a reference count of 1 or more is called "active".
When a vnode's reference count drops to zero, it becomes "inactive", that
is, a candidate for reuse. An "inactive" vnode still refers to a valid
file and one can try to reactivate it using vget(9) (this is used a lot
by caches).
Before the VFS can reuse an inactive vnode to refer to another file, it
must clean all information pertaining to the old file. A cleaned out
vnode is called a "reclaimed" vnode.
To support forceable unmounts and the revoke(2) system call, the VFS may
"reclaim" a vnode with a positive reference count. The "reclaimed" vnode
is given to the dead filesystem, which returns errors for most opera-
tions. The reclaimed vnode will not be re-used for another file until its
reference count hits zero.
The getnewvnode(9) system call allocates a vnode from the pool, possibly
reusing an "inactive" vnode, and returns it to the caller. The vnode re-
turned has a reference count (v_usecount) of 1.
The vref(9) call increments the reference count on the vnode. It may only
be on a vnode with reference count of 1 or greater. The vrele(9) and
vput(9) calls decrement the reference count. In addition, the vput(9)
call also releases the vnode lock.
The vget(9) call, when used on an inactive vnode, will make the vnode
"active" by bumping the reference count to one. When called on an active
vnode, vget increases the reference count by one. However, if the vnode
is being reclaimed concurrently, then vget will fail and return an error.
The vgone(9) and vgonel(9) orchestrate the reclamation of a vnode. They
can be called on both active and inactive vnodes.
When transitioning a vnode to the "reclaimed" state, the VFS will call
VOP_RECLAIM(9) method. Filesystems use this method to free any
filesystem-specific data they attached to the vnode.
The vnode actually has three different types of lock: the vnode lock, the
vnode interlock, and the vnode reclamation lock (VXLOCK).
The vnode lock and its consistent use accomplishes the following:
• It keeps a locked vnode from changing across certain pairs of VOP_
calls, thus preserving cached data. For example, it keeps the direc-
tory from changing between a VOP_LOOKUP call and a VOP_CREATE. The
VOP_LOOKUP call makes sure the name doesn't already exist in the
directory and finds free room in the directory for the new entry. The
VOP_CREATE can then go ahead and create the file without checking if
it already exists or looking for free space.
• Some filesystems rely on it to ensure that only one "thread" at a
time is calling VOP_ vnode operations on a given file or directory.
Otherwise, the filesystem's behavior is undefined.
• On rare occasions, code will hold the vnode lock so that a series of
VOP_ operations occurs as an atomic unit. (Of course, this doesn't
work with network filesystems like NFSv2 that don't have any notion
of bundling a bunch of operations into an atomic unit.)
• While the vnode lock is held, the vnode will not be reclaimed.
There is a discipline to using the vnode lock. Some VOP_ operations re-
quire that the vnode lock is held before being called. A description of
this rather arcane locking discipline is in sys/kern/vnode_if.src.
The vnode lock is acquired by calling vn_lock(9) and released by calling
VOP_UNLOCK(9).
A process is allowed to sleep while holding the vnode lock.
The implementation of the vnode lock is the responsibility of the indivi-
dual filesystems. Not all filesystems implement it.
To prevent deadlocks, when acquiring locks on multiple vnodes, the lock
of parent directory must be acquired before the lock on the child direc-
tory.
The vnode interlock (vp->v_interlock) is a spinlock. It is useful on
multi-processor systems for acquiring a quick exclusive lock on the con-
tents of the vnode. It MUST NOT be held while sleeping. (What fields does
it cover? What about splbio/interrupt issues?)
Operations on this lock are a no-op on uniprocessor systems.
The vnode reclamation lock (VXLOCK) is used to prevent multiple processes
from entering the vnode reclamation code. It is also used as a flag to
indicate that reclamation is in progress. The VXWANT flag is set by
processes that wish to be woken up when reclamation is finished.
The vwaitforio(9) call is used to wait for all outstanding write I/Os as-
sociated with a vnode to complete.
The vnode capability, v_id, is a 32-bit version number on the vnode.
Every time a vnode is reassigned to a new file, the vnode capability is
changed. This is used by code that wishes to keep pointers to vnodes but
doesn't want to hold a reference (e.g., caches). The code keeps both a
vnode * and a copy of the capability. The code can later compare the
vnode's capability to its copy and see if the vnode still points to the
same file.
Note: for this to work, memory assigned to hold a struct vnode can only
be used for another purpose when all pointers to it have disappeared.
Since the vnode pool has no way of knowing when all pointers have disap-
peared, it never frees memory it has allocated for vnodes.
Most of the fields of the vnode structure should be treated as opaque and
only manipulated through the proper APIs. This section describes the
fields that are manipulated directly.
The v_flag attribute contains random flags related to various functions.
They are summarized in table ...
The v_tag attribute indicates what filesystem the vnode belongs to. Very
little code actually uses this attribute and its use is deprecated. Pro-
grammers should seriously consider using more object-oriented approaches
(e.g. function tables). There is no safe way of defining new v_tags for
loadable filesystems. The v_tag attribute is read-only.
The v_type attribute indicates what type of file (e.g. directory, regu-
lar, FIFO) this vnode is. This is used by the generic code for various
checks. For example, the read(2) system call returns an error when a read
is attempted on a directory.
The v_data attribute allows a filesystem to attach a piece of file system
specific memory to the vnode. This contains information about the file
that is specific to the filesystem.
The v_numoutput attribute indicates the number of pending synchronous and
asynchronous writes on the vnode. It does not track the number of dirty
buffers attached to the vnode. The attribute is used by code like fsync
to wait for all writes to complete before returning to the user. This at-
tribute must be manipulated at splbio().
The v_writecount attribute tracks the number of write calls pending on
the vnode.
The vast majority of vnode functions may not be called from interrupt
context. The exceptions are bgetvp and brelvp. The following fields of
the vnode are manipulated at interrupt level: v_numoutput, v_holdcnt,
v_dirtyblkhd, v_cleanblkhd, v_bioflag, v_freelist, and v_synclist. Any
access to these fields should be protected by splbio.
This document first appeared in OpenBSD 2.9.
MirBSD #10-current February 22, 2001 2