path: root/sys/fio/doc/vfn.hlp

.help vfn Jul84 "Virtual Filename Mapping"
.ce
\fBVirtual Filename Mapping Package\fR
.ce
Detailed Design
.ce
Doug Tody
.ce
July 1984
.sp 2
.NH
Introduction

    This document presents the detailed design of the filename mapping
code, used by FIO to map virtual filenames (VFN's) to host operating system
filenames (OSFN's) and back again.  A description of the filename mapping
algorithm is given in \fIThe Reference Manual for the IRAF System Interface\fR,
May 1984.  The purpose of this document is more to design the software than
to document the design, hence much is omitted.  The discussion concentrates
on those aspects of the problem which were least-understood at the time of
the design.

.sh
Primary Functions

.nf
	map vfn->osfn
	map osfn->vfn
.fi

.sh
Functions for accessing the vfnmap file

.nf
	open and optionally lock vfnmap file
	close and unlock vfnmap file

	add entry to vfnmap
	delete entry from vfnmap
	lookup entry in vfnmap
.fi

.sh
Mapping Functions

.nf
	extract OSDIR prefix
	extract LDIR prefix
	expand LDIR
	fold subdir into OSDIR
	encode filename via escape sequence encoding
	decode encoded filename
	squeeze filename
	map filename extension
.fi


.nh
VFN Virtual Filename Mapping Package

    The VFN package is used to map and unmap virtual filenames and to add and
delete virtual filenames from the VFN database.  A distinct open operation is
required for each vfn to be accessed.  Any number of vfn's may be simultaneously
open for reading, but only \fIone\fR vfn may be opened for writing.
The mapping file is not physically opened unless the escape sequence encoded
filename is degenerate.  It is intended that the vfn will be opened for only
a brief period of time to minimize the amount of time that the mapping file
is locked.  The mapping file is locked only if the vfn is degenerate and the
access mode is VFN_WRITE.  The recognized vfn access modes are VFN_READ,
VFN_WRITE, and VFN_UNMAP (for reading directories).


.ks
.nf
	vp =	 vfnopen (vfn, mode)
		vfnclose (vp, update)
	stat =    vfnmap (vp, osfn)
	stat =	  vfnadd (vp, osfn)
	stat =    vfndel (vp, osfn)
	stat =	vfnunmap (vp, osfn, vfn)

	stat =    fmapfn (vfn, osfn)	[=:vfnopen/RO,vfnmap,vfnclose]
.fi
.ke


A distinction is made between mapping the filename and opening and closing
the vfn to permit efficient and secure error recovery.  The mapping file is
not updated on disk until the physical file operation (create, delete, etc)
has succeeded.  If the operation fails \fBvfnclose\fR is called with NO_UPDATE
and the mapping file is not touched.  The the vfn was opened VFN_READ the
update flag is ignored.  No vfn disk data structures will be modified
if a vfn is closed with NO_UPDATE set.  If updating is enabled, ".zmd"
dependency files may be created or deleted, the mapping file may be created,
deleted, or updated.

The procedure \fBvfnmap\fR returns ERR if the vfn is degenerate but no entry
could be found in the mapping file, i.e., if the file does not exist.
A status value of OK does not, however, imply that the file exists.
\fBVfnadd\fR returns ERR if the vfn is degenerate and an entry already
exists in the mapping file.  If the status return is OK and the vfn is
degenerate then a new entry has been added to the mapping file.
\fBVfndel\fR returns ERR if the vfn is degenerate but no entry
could be found in the mapping file.  \fIOsfn\fR is returned as a packed string.
The output buffer should be dimensioned SZ_PATHNAME.

.nh
Semicode for Selected FIO Procedures

    The RO class procedures call FMAPFN to map the VFN of an existing file
into an OSFN.  These operations are straightforward since the vfn database
is not affected.

.ks
.nf
	access, fchdir, finfo, fpath, fprot:	RO operations
	falloc, open/NF, fmkcopy:		RW=ADD procedures
	delete					RW=DEL procedure
	rename					RW=DEL+ADD
.fi
.ke	


.nf
# FALLOC -- Create a new file and allocate uninitialized storage.  Open/NF and
# make copy are similar operations hence the semicode is not shown.

procedure falloc (vfn, size)

begin
	# Map filename and determine if a file already exists with the
	# same name.
	vp = vfnopen (vfn, VFN_WRITE)			# LOCK
	if (vfnadd (vp, osfn) == ERR)
	    existing_file = yes
	else {
	    call zfacss to see if file exists
	    existing_file = yes if file exists
	}

	# If file exists and clobber is enabled, try to delete the file.
	# If filename is degenerate, entry is either already in mapping file
	# (if file exists), or has been added.

	if (existing_file)
	    iferr {
		if (file clobber enabled)
		    delete file
		else
		    error ("falloc would clobber file 'vfn'")
	    } then {
		vfnclose (vp, NO_UPDATE)
		erract (EA_ERROR)
	    }

	# Allocate the new file and update the filename mapping database.

	call ZFALOC to allocate the file
	if (failure) {
	    vfnclose (vp, NO_UPDATE)
	    error ("cannot allocate file 'vfn'")
	} else
	    vfnclose (vp, UPDATE)			# UNLOCK
end


# DELETE -- Delete a file and all subfiles.  A subfile is a file which is
# logically part of the parent file but which is physically a separate file
# at the kernel level.  An example is the pixel storage file associated with
# an image.  Whenever a file is deleted all subfiles must be deleted as well.
# The subfiles need not reside in the same directory as the main file.
# Subfile information is maintained in a separate, "invisible" file for each
# file having subfiles.  The subfile list file has the same vfn as the main
# file with the extension ".sfl" appended.  If the vfn already had an extension
# it is retained in the root of the new filename.  For example, the vfn of the
# subfile list file for "data.db" would be "data.db.sfl".

procedure delete (vfn)

begin
	# Delete the main file
	fdelpf (vfn)

	# Delete any subfiles.  Print warning message if a subfile appears
	# in the list but cannot be deleted.
	
	ifnoerr (fd = fsf_open (vfn, READ_ONLY)) {
	    while (getline (fd, subfilename, SZ_FNAME) != EOF)
		iferr (fdelpf (subfilename))
		    call erract (EA_WARN)
	    close (fd)
	}
end


# FDELPF -- Delete a single physical file.  Check if the file is protected
# and do not try to delete the file if it is protected.  If file cannot be
# deleted, determine why and print appropriate error message, and do not update
# the mapping file.

procedure fdelpf (vfn)

begin
	vp = vfnopen (vfn, VFN_WRITE)			# LOCK
	if (vfndel (vp, osfn) == ERR) {
	    vfnclose (vp, NO_UPDATE)
	    error ("attempt to delete a nonexistent file (vfn)")
	} 

	call ZFPROT to check for file protection
	if (file is protected) {
	    vfnclose (vp, NO_UPDATE)
	    error ("attempt to delete a protected file (vfn)")
	}
		
	call ZFDELE to delete the file
	if (failure) {
	    vfnclose (vp, NO_UPDATE)
	    call ZFACCS to determine if file exists
	    if (no such file)
		error ("attempt to delete a nonexistent file (vfn)")
	    else
		error ("cannot delete file 'vfn'")
	}

	vfnclose (vp, UPDATE)				# UNLOCK
end


# RENAME -- Rename a file.  A file may be renamed within a single directory
# or may be moved to another directory by the rename operation.  Note that
# we may only have one VFN opened for writing at a time.

procedure rename (oldvfn, newvfn)

begin
	# Delete old filename from VFN database.
	vp = vfnopen (oldvfn, VFN_WRITE)
	if (vfndel (vp, oldosfn) == ERR) {
	    vfnclose (vp, NO_UPDATE)
	    error ("attempt to rename a nonexistent file (vfn)")
	} else
	    vfnclose (vp, UPDATE)

	# Add new filename to VFN database.
	vp = vfnopen (newvfn, VFN_WRITE)
	if (vfnadd (vp, newosfn) == ERR) {
	    vfnclose (vp, NO_UPDATE)
	    error ("cannot create new file 'vfn'")
	} else
	    vfnclose (vp, UPDATE)
	
	# Rename the physical file.
	call ZFRNAM to rename the file

	# Patch up VFN database if the rename operation fails.  If the rename
	# fails then most likely the OSFN's were short and no mapping file
	# access was involved (else we would have had an abort above), but
	# then the calls cost almost nothing so make them anyhow.

	if (rename fails) {
	    # Restore old filename.
	    vp = vfnopen (oldvfn, VFN_WRITE)
	    vfnadd (vp, oldosfn)
	    vfnclose (vp, UPDATE)

	    # Delete new filename.
	    vp = vfnopen (newvfn, VFN_WRITE)
	    vfndel (vp, newosfn)
	    vfnclose (vp, UPDATE)

	    error ("cannot rename file (oldvfn -> newvfn)")
	}
end
.fi

.nh
Locking and Concurrency Considerations

    A locking mechanism is necessary to prevent two or more processes from
simultaneously modifying a mapping file.  The dimensions of the problem are
as follows:

.ls
.ls [1]
Mutual exclusion must be guaranteed.  The period of time during which a process
opens and reads the mapping file, modifies it, and updates the file on disk
is the critical section.  The locking protocol must guarantee that only one
process can be in the critical section at a time.  A read-only access of the
mapping file is not a critical section, but we must guarantee that the file
is not in the process of being written when such a read occurs.
.le
.ls [2]
Deadlock must either be prevented or it must be detected and broken.
Deadlock will eventually occur if a process is permitted to simultaneously
access more than one mapping file.  Deadlock will occur if process A locks
directory D1 and process B locks D2, then B tries to lock D1 and A tries to
lock D2.
.le
.ls [3]
Lockout will occur if a process dies while in the critical section, thus
failing to remove the lock.
.le
.le


On a system which provides file locking, i.e., which forbids a process
access to a file which is open with write permission by another process,
the host OS guarantees mutual exclusion and protection from lockout.
Unfortunately many UNIX systems (and probably some other systems as well)
do not provide file locking.  The scheme discussed in this section is
awkward but provides secure locking on such systems.  The file locking
facilities discussed herein are designed to make use of host system file
locking if available.  The discussion is oriented towards the problems
of providing locking on systems which do not provide locking at the kernel
level, i.e., in \fBzopnbf\fR.

.nh 2
Mutual Exclusion

    Mutual exclusion can be guaranteed by use of a \fBsemaphore\fR.
The transportability requirement makes it very difficult to implement a
general semaphore, but a binary semaphore is possible using a null length
file in the same directory as the mapping file.  To implement a semaphore
we must test and set the lock all in the same operation, to prevent
interleaving of the operations by two processes simultaneously trying to
set a lock (i.e., process A tests for a lock and finds none, B tests for a
lock and finds none, A sets a lock, B sets a lock, and mutual exclusion is
violated).

A suitable binary semaphore can be implemented by \fIdeleting\fR the lock
file to set the lock, rather than by testing for the lock (no lock file)
and then creating the lock file to set the lock.  We assume that the delete
operation will return error for an attempt to delete a nonexistent file.
Thus if the lock file can be successfully deleted, the lock has been tested
and found to be absent and the directory has been locked, all in one
indivisible kernel operation.


.ks
.nf
	# Gain exclusive access to a file.  The file must have an
	# associated lockfile which is deleted while a process has
	# the file locked.

	while (delete (lockfile) == ERR)
	    ;


	# Give up exclusive access to a file.
	create (lockfile)
.fi
.ke


The above is a bit simplistic because the file itself may not exist,
in which case there will be no lockfile, and the process may not have
delete permission for the lockfile if there is one.  The point here is
that the OS kernel guarantees that only one process will be allowed
to successfully delete the lockfile, hence the deletion operation can
serve to gain exclusive access to a file.  The problem of lockout, wherein
the lockfile gets lost, is dealt with later.

Locking the directory is necessary whenever the mapping file is to be modified.
While it is not necessary to lock the directory to read the mapping file,
by not doing so we run the risk of trying to read while the file is being
written to (permissible on some systems, an error condition on others).
The simplest solution to this problem is to lock the file for all accesses,
including reads as well as writes.  The problem with this approach is that
it precludes read access on directories for which a process does not have
write permission (preventing generation of the lock file).  This is not
acceptable.  Our solution is to include a \fBchecksum\fR in the mapping file.
If the file exists but cannot be opened for reading and a lock exists on the
directory, we will wait until the lock is lifted to read the file.  If the
checksum is in error the read will be repeated until a valid checksum is
obtained.

.nh 2
Deadlock

    Deadlock can be avoided by the simple expedient of permitting a process to
lock only one directory at a time.  The only time a process needs to lock
more than one directory is when renaming a file with a long, degenerate name
from one directory to another.  Deadlock is unlikely but would certainly
occur at infrequent intervals.  Locking only one directory at a time is
inefficient (because separate references are needed to map the filename
and to edit the mapping file), but it does not matter since lock file
accesses are expected to be infrequent (few mapped filenames are degenerate).
Detection of and breaking of deadlock is possible but not worth the trouble.
Thus we shall avoid the problem of deadlock entirely by permitting a process
to lock only a single directory at a time, for only a brief period of time.

.nh 2
Lockout

    At this point we have a solution which guarantees mutual exclusion and the
avoidance of deadlock nearly 100% of the time.  The only problem remaining
is \fBlockout\fR.  It is not possible to prevent lockout since we cannot
guarantee that a process (or the computer) will not die while in a critical
section, preventing removal of the lock.

The obvious way to implement automatic recovery from lockout is to add a
provision for timeout.  While we cannot guarantee that the time spent
in a critical section will be less than some absolute amount (because of
variable load conditions, swapping, the time required to delete a very
large file, etc.), we can say that the time spent in a critical section will
rarely be larger than some number on the order of one second.  In a worst
case situation where several processes are heavily accessing a directory
it could take an arbitrarily long time for a particular process to gain a
lock on the directory, but this is very unlikely.

If a process times out while waiting we must either abort or proceed to break
the lock.  This may be done by creating a new lockfile as if the transaction
had been completed.  There is a hidden bug, however -- if two or
more processes timeout simultaneously, the following scenario might occur:


.kf
.nf
	A times out
	B times out
	A breaks the lock
	A enters wait loop and places a new lock,
	    entering the critical section
	B breaks the lock set by A
	B enters wait loop and places a new lock,
	    entering the critical section
	[...mutual exclusion is violated...]
.fi
.ke


No matter how unlikely this scenario might be, it prevents us from using the
simple technique to break the lock.  Breaking the lock appears to be another
critical section, so perhaps we can use another semaphore to protect the lock
(we ignore the complications of checking for write permission on the directory,
which should be dealt with when the lock is set).

Even if a semaphore is used concurrency
can still be a problem, as another process may timeout and break the lock
shortly after the first process has done so; this can happen because the
section between timeout and the test for permission to break the lock is
interruptable.  To get around this we apply an additional constraint
that the lock can only be broken if it has been in place for a specified
interval of time which is much larger than the timeout interval.  This suffices
to recover from a process crash and prevents two processes from breaking
the lock at almost the same time.


.ks
.nf
	# Try to set a lock on the directory.  If we timeout, try to get
	# permission to break the lock; only one process is permitted to
	# break the lock, and the lock can only be broken once in a
	# specified interval of time.  The timelock files are normally
	# created whenever the directory is locked.

	repeat {
	    while (delete (lockfile) == ERR)
		if (timeout)
		    if (delete (timelock1) != ERR) {
			get creation date of timelock2
			if (timelock2 is an old file) {
			    create (lockfile)
			    delete (timelock2); create (timelock2)
			    create (timelock1)
			} else
			    create (timelock1)
		    }
	} until (lock is established)

	# Back to normal.
	carry out transaction
	create (lockfile)
.fi
.ke


Lockout is still possible if the process or the computer dies in the interval
between deletion and creation of timelock1, but the chances of that happening
are very remote because the interval is short and it only occurs during
recovery from lockout.  An additional check should perhaps be provided to
detect this unlikely circumstance and break the lock without further ado
if timelock1 somehow gets permanently deleted.  The mapping file can be
checkpointed when this occurs to minimize the risk.

.nh 2
Rollback

    Unfortunately, automatic lockout detection and recovery brings with it
the possibility that the lock will be broken when a process takes an abnormally
long time to complete a transaction.  This might happen when a heavily loaded
system has begun swapping processes, or when a background job with a
very low priority accesses a directory.  We must be able to detect that the
lock has been broken and \fIrollback\fR the transaction, i.e., obtain a new
lock and try again, repeating the unsuccessful transaction.

Timeouts leading to improper breaking of the lock are not a problem if the
host system provides file locking for files opened for writing.  After placing
the lock on a directory a process will open the mapping file with readwrite
permission and all other processes will be locked out until the transaction
completes.  Unfortunately file locking is not provided on all systems (e.g.,
many versions of UNIX do not provide file locking).

Secure protection from a broken lock is difficult because if we check that
the lock is still in place and then perform the update, another process may
break the lock immediately after we check that the lock is in place and
before the update occurs.  About the best we can do is check the creation time
on timelock2 immediately before updating, updating only if the timelock has
not been touched since we created it at lock time.  If the lock has been
broken our timelock file will have been deleted and the transaction must be
rolled back.  If a lot of time remains on the lock we go ahead and perform
the update, otherwise a new timelock2 is written, providing a time equal to
the minimum lifetime of a lock in which to update the file.


.ks
.nf
	perform transaction upon MFD (in memory)

	# Determine if the lock is still in place and likely to remain
	# in place until the update is completed.

	repeat {
	    get creation date of timelock2
	    if (not the timelock we set at vfn_wait time)
		rollback transaction
	    else if (not much time left on lock)
		rollback transaction
	    else
		break
	}

	# Update and remove the lock.

	update the mapping file
	close (mapping file)

	get creation date of timelock2
	if (not our timelock)
	    bad news: warn user

	create (lockfile)
.fi
.ke

.nh 2
File Locking Facilties

    From the above code fragments it appears that the lockfile approach
to file locking will work on any machine on which it is an error to delete
a nonexistent file.  The next step is to encapsulate all this in file
locking primitives which will use the host OS file locking facilities if
any, otherwise the lockfile techniques we have developed.  A set of file
locking primitives are presented below.  These are low level routines
with fairly restrictive semantics, and are not intended to be used in other
than system code.


.ks
.nf
	time =	    osfn_lock (osfn)
	nsec =	osfn_timeleft (osfn, time)
	nsec =	  osfn_unlock (osfn, time)
.fi
.ke


A file is locked with the \fBosfn_lock\fR primitive, which returns when
it has successfully placed a lock on the file \fIosfn\fR.  The lock is
guaranteed to remain in place for at least \fItimeout\fR seconds, where
\fItimeout\fR is a system constant.
On some systems the file may not actually be locked until it is opened
with write access.  If the file does not exist or cannot be locked
\fBerror\fR is called.  If the file is already locked but the lock has
expired \fBosfn_lock\fR will break the old lock and return when it has
set a new one.

The primitive \fBosfn_timeleft\fR returns the number of seconds remaining
on the lock on file \fIosfn\fR.  ERR is returned if the file is no longer
locked or if the file is currently locked by another user.

A lock is removed with \fBosfn_unlock\fR.  The number of seconds remaining
on the lock at the time it was removed is returned as the function value.
ERR is returned if the file was no longer locked or had been locked by
another user when \fBosfn_unlock\fR was called.


.nf
# OSFN_LOCK -- Lock the named OSFN, i.e., gain exclusive write access
# to a file.  Only the process gaining the lock on a file may write
# to it, but there is no guarantee that another process may not read
# a locked file.  On some systems the file will not actually be locked
# until it is opened with write permission.  If multiple files exist
# in a directory with the same root but different extensions, only one
# can be locked at a time.

long procedure osfn_lock (osfn)

begin
	# Even if file locking is provided by the OS we must determine
	# if the file is write protected.  If the file is not write
	# protected but cannot be opened for writing our caller will
	# conclude that the file is locked by another process.

	if (file locking is handled by the OS)
	    if (file osfn is write protected)
		error ("no write permission on file 'osfn'")
	    else
		return (clktime)

	# Generate filenames.
	basename = osfn minus any extension
	lockfile  = strpak (basename // ".lok")
	timelock1 = strpak (basename // ".tl1")
	timelock2 = strpak (basename // ".tl2")

	# If the lockfile can be deleted (usual case) then we have
	# little to do.
	if (delete (lockfile) == OK)
	    goto setlock_

	# If the lockfile cannot be deleted check that the file itself
	# exists and that we have delete permission on the directory.

	if (file 'osfn' does not exist)
	    error ("attempt to lock a nonexistent file (osfn)")
	if (no delete permission on directory)
	    error ("cannot delete file (lockfile)")

	# The file exists and all the necessary permissions are granted,
	# hence someone else has the file locked and we must wait.

	repeat {
	    for (nsec=0;  nsec < timeout_period;  nsec=nsec+1)
		if (delete (lockfile) == OK)
		    goto setlock_
	    if (delete (timelock1) == OK) {
		get creation date of timelock2
		if (timelock2 is an old file or does not exist) {
		    create (lockfile)
		    delete (timelock2); create (timelock2)
		    create (timelock1)
		} else
		    create (timelock1)
	    } else if (continual failure to delete timelock1)
		create (timelock1)
	}

setlock_
	delete (timelock2)
	create (timelock2)

	return (creation time of timelock2)
end


# OSFN_TIMELEFT -- Determine if a file is still locked, and if so, how
# much time remains on the lock.  TIME is the time value returned when
# the file was locked.  All time values are in units of seconds.

int procedure osfn_timeleft (osfn, time)

begin
	if (file locking is handled by the OS)
	    return (big number)

	basename  = osfn minus any extension
	lockfile  = strpak (basename // ".lok")
	timelock2 = strpak (basename // ".tl2")

	if (lockfile exists)
	    return (ERR)
	else if (cannot get file info on timelock2)
	    return (ERR)
	else if (timelock2.create_time != time)
	    return (ERR)
	else {
	    timeleft = max (0, timeout_period - (clktime - time)
	    return (timeleft)
	}
end


# OSFN_UNLOCK -- Release the lock on a file and return the number of
# seconds that were left on the lock.  ERR is returned if the file is
# no longer locked or if the lock is not the one originally placed
# on the file.

int procedure osfn_unlock (osfn, time)

begin
	timeleft = osfn_timeleft (osfn, time)

	if (timeleft != ERR) {
	    basename = osfn minus any extension
	    lockfile = strpak (basename // ".lok")
	    create (lockfile)
	}

	return (timeleft)
end
.fi

.nh
VFN Package Data Structures

    A process may have only a single VFN open with write permission at any
one time to eliminate the possibility of deadlock (section 4).  Any number
of VFN's may be open for read-only access, e.g., when recursively descending
a directory tree.  Most VFN accesses do not involve a reference to a mapping
file.  Since the mapping file is infrequently referenced, separate descriptors
are used for the VFN and the mapping file.  The VFN descriptor is called the
VFD and the mapping file descriptor the MFD.

The MFD is only allocated if a mapping file is referenced, i.e., if the OSFN
is long.  Before allocating a new MFD we must search the list of open VFN's
to see if the mapping file has already been opened and assigned a MFD.  Every
VFN must have its own VFD.  To prevent having to MALLOC a
VFD every time a filename is mapped, one VFD will always be allocated (after
the first file reference).  Thus, for a simple filename mapping where the
OSFN is short, no MALLOC or other kernel calls will be required, i.e., the only
expense will be the string operations required to map the filename.


.ks
.nf
# VFN Descriptor

struct vfd {
	struct	mfd *v_mfd		# ptr to mapping file descr.
	int	v_acmode		# access mode
	int	v_len_osdir		# length of v_osdir string
	int	v_len_root		# length of v_root string
	int	v_len_extn		# length of v_extn string
	char	v_vfn[33]		# original VFN, minus LDIR
	char	v_osdir[33]		# OS directory name
	char	v_root[33]		# encoded root filename
	char	v_extn[33]		# encoded and mapped extension
}
.fi
.ke


.ks
.nf
# Mapping File Descriptor.  The length of the descriptor is adjusted as 
# necessary to provide storage for the filename pairs.

struct mfd {
	long	m_locktime		# clktime when lock set
	int	m_fd			# file descriptor
	int	m_nfiles		# number of files in map list
	int	m_lastop		# last operation performed
	int	m_modified		# was database modified
	char	m_vfnmap[]		# OSFN of mapping file
	int	m_checksum		# checksum of m_fnmap
	char	m_fnmap[nfiles*34*2]	# vfn/osfn pairs
}
.fi
.ke

.nh
Semicode for Parts of the VFN Package

.nf
# VFNOPEN -- Open that part of the VFN database associated with a particular
# VFN.  Allocate VFD descriptor, map but do not squeeze VFN to long OSFN.

pointer procedure vfnopen (vfn, mode)

begin
	if (first_time) {
	    permanently allocate a VFD
	    nvfn_open = 0
	    first_time = false
	}

	# Allocate and initialize VFD.
	if (no VFN's open) {
	    use preallocated VFD
	    increment count of open VFN's
	} else
	    allocate a new VFD

	call fbrkfn to break VFN into OSDIR, ROOT, and EXTN fields

	return (pointer to VFD)
end


# VFNCLOSE -- Close a VFN and optionally update the VFN database.  An update
# is performed only if the mapping file is open with write permission,
# a modify transaction has occurred, and updating is enabled.

procedure vfnclose (vp, update)

begin
	# If the mapping file was not used or if it was not modified we
	# just return the buffers and quit.

	mfp = vp.mfp
	if (mfp == NULL) {
	    if (nvfn_open > 1)
		mfree (vp, TY_STRUCT)
	    return
	} else if (mfp.m_modified == NO || update == NO_UPDATE) {
	    mfree (mfp, TY_STRUCT)
	    if (nvfn_open > 1)
		mfree (vp, TY_STRUCT)
	    return
	}

	# If we get here the mapping file is open with write permission,
	# a transaction has been performed which modified the database,
	# and we were called with updating enabled.  Rollback (repeat)
	# the transaction if the lock has been broken or if there is not
	# enough time remaining on the lock.

	while (osfn_timeleft (mfp.m_vfnmap, mfp.m_locktime) < xx) {
	    osfn_unlock (mfp.m_vfnmap, mfp.m_locktime)
	    switch (mfp.lastop) {
	    case VFN_ADD:
		vfnadd (vp, junkstr)
	    case VFN_DEL:
		vfndel (vp, junkstr)
	    }
	}

	# Update and close the mapping file.
	compute checksum and store in the mapping file
	rewrite mapping file to disk
	close (mapping file)

	if (osfn_unlock (mfp.m_vfnmap, mfp.m_locktime) == ERR)
	    warn ("broken file protect lock in directory 'vp.v_osdir'")

	mfree (mfp, TY_STRUCT)
	if (nvfn_open > 1)
	    mfree (vp, TY_STRUCT)
end


# VFNMAP -- Map and pack the VFN into an OSFN, but do not modify the
# database.  The mapping file is accessed only if the filename is
# degenerate.

int procedure vfnmap (vp, osfn)

begin
	# If the OSFN is short or long but still unique within directory,
	# then it is not necessary to access the mapping file.

	if (root is longer than permitted by host system) {
	    squeeze root
	    if (squeezed root filename is unique within directory) {
		concatenate and pack osfn
		return (OK)
	    }
	}

	# If we get here then the squeezed filename is degenerate, i.e.,
	# not unique within the directory.  It is necessary to read the
	# mapping file to learn what OSFN has been assigned to the file.

	mfp = allocate and init mapping file descriptor
	mfp.m_vfnmap = strpak (osdir // "zzvfnmap.vfn")

	# Open or create the mapping file.  Create must precede lock
	# as lock will abort if the file to be locked does not exist.
	# If opening existing file READ_WRITE, lock first to determine
	# if we have write perm on file, then keep trying to open file
	# until open succeeds (if OS level file locking is in use the
	# open will return ERR as long as another process has the
	# file open for writing).

	switch (vp.v_acmode) {
	case VFN_WRITE:
	    if (no mapping file created yet) {
		create a new mapping file
		time = osfn_lock (mfp.m_vfnmap)
	    } else {
		time = osfn_lock (mfp.m_vfnmap)
		repeat {
		    open mapping file for READ_WRITE access
		    sleep (1)
		} until (open succeeds)
	    }
	default:
	    open mapping file for READ_ONLY access
	}

	# Read mapping file into descriptor.  Increase default size of
	# descriptor if necessary to read entire file.  Repeat the
	# read if the checksum is invalid, indicating that a write
	# was in progress when we read.

	maxch = default buffer size for the filename map
	repeat {
	    repeat {
		read maxch chars into mfp.m_checksum
		if (nchars_read >= maxch) {
		    increase size of descriptor
		    maxch = maxch + increase in storage
		}
	    } until (nchars_read < maxch)
	    compute checksum
	} until (checksum == mfp.m_checksum)

	if (nchars_read == EOF)
	    mfp.m_nfiles = 0
	else
	    mfp.m_nfiles = max (0, (nchars - SZ_INT) / SZ_FNMAP_PAIR)

	search mfp.m_fnmap for filename vp.vfn
	if (not found)
	    status = ERR
	else {
	    status = OK
	    pack osfn to output argument
	}

	if (access_mode != VFN_WRITE)
	    close mapping file

	return (status)
end


# VFNADD -- Map a VFN to an OSFN and add an entry for the VFN to the
# database if the OSFN is degenerate.

procedure vfnadd (vp, osfn)

begin
	# If VFNMAP does not return ERR then the file already exists.
	# We return ERR if the file already exists.

	if (vfnmap (vp, osfn) != ERR)
	    return (ERR)
	else if (short osfn)
	    return (OK)

	if (osfn is degenerate) {
	    generate a unique new_osfn
	    create degeneracy flag file osfn // ".zmd"
	    osfn = strpak (new_osfn)
	}

	add vfn,osfn pair to vp.mfp.m_fnmap
	mfp.m_lastop = VFN_ADD

	return (OK)
end


# VFNDEL -- Map a VFN to an OSFN and delete the entry for the VFN from
# the database if the OSFN is degenerate.  Do not delete the degeneracy
# flag file if no longer degenerate, because even though the OSFN is
# no longer degenerate the OSFN reflects the former degeneracy of the
# file, and we do not want to rename the file.

procedure vfnadd (vp, osfn)

begin
	# If VFNMAP returns ERR then the file does not exist.
	# We return ERR if the file does not exist.

	if (vfnmap (vp, osfn) == ERR)
	    return (ERR)
	else if (short osfn)
	    return (OK)

	delete vfn,osfn pair to vp.mfp.m_fnmap
	mfp.m_lastop = VFN_DEL

	return (OK)
end


# FBRKFN -- Transform a VFN into an OSDIR, an escape sequence encoded and
# extension mapped root OS filename ROOT, and an extension EXTN.  The root
# may be longer than permitted by the host OS, i.e., squeezing is not done
# here.

procedure fbrkfn (vfn, osdir, lenosdir, root, lenroot, extn, lenextn)

begin
	# If the VFN begins with an OSDIR prefix it is assumed to be an OSFN
	# and no mapping is performed.

	call ZFXDIR to extract osdir prefix, if any
	if (osdir prefix found) {
	    copy remainder of vfn to root
	    return
	}

	osdir = null_string
	root = null_string
	extn = null_string

	# Process the directory and filename fields.  In the case of a simple
	# filename the first pass performs the escape sequence encoding of the
	# filename directly into root, and we return after possibly mapping
	# the extension.

	repeat {
	    extract next field into root and extn with escape sequence encoding
	    if (delimiter == '$')
		if (osdir == null_string) {
		    osdir = recursively expand ldir
		    if (ldir not found)
			error ("logical directory 'ldir' not found")
		} else
		    error ("illegal $ delimiter in filename 'vfn'")
	    } else if (delimiter == '/')
		fold field, a subdirectory, into osdir
	} until (delimiter == EOS)

	# At this point we have osdir, root, and extn strings, any of which may
	# be null.  If more than one "." delimited extn string was encountered
	# during escape sequence encoding, or if the maximum extn length was
	# exceedd, then that extn will already have been incorporated into the
	# root.

	if (extn != null_string)
	    map filename extension
end