The Linux Kernel Implementation of Pipes and FIFOs

Linux Kernel Implementation of Pipes, FIFOs and
other Filesystems

Divye Kapoor
B.Tech (IDD) CSI – IV
061305

Userspace v/s Kernel Space
Conventional Operating Systems generally segregate virtual memory into
kernel space and user space
● New research OS’s (Singularity etc.) maintain a single
virtual address space for all processes and depend on
language VMs for maintaining process isolation.

Kernel space is strictly reserved for running the kernel, kernel extensions
(modules) and device drivers
Memory is generally not swapped out

User space is where generally all user processes run
Each process has its own virtual address space and memory may be
swapped out.

There is hardware support for providing the distinction in privilege levels in
user space and kernel space
For the x86 architecture processors, this is provided by privilege rings

Privilege Rings
Click to edit Master text styles
Second level
● Third level

● Fourth level

● Fifth level

General Context of the Filesystem

User Space Processes

System Calls –
open, read, write, stat
VFS (in Kernel etc.
Space)
It implements these calls and redirects to the appropriate
implementation requested.

procfs pipefs ext3 reiserfs

Virtual Filesystems Filesystems with backing disk
stores

Linux Virtual Filesystem (VFS)
It is the software layer in the kernel that provides a uniform
filesystem interface to userspace programs

It provides an abstraction within the kernel that allows for
transparent working with a variety of filesystems.

Thus it allows many different filesystem implementations to coexist
freely

Each pipe and FIFO is implemented as a “file” mounted on the pipefs
filesystem.
A FIFO is just a thin wrapper around a pipe

Inodes, Direntry and File
Inodes provide a method to access the actual data blocks allocated to
a file.
struct inode

Inodes are created in the context of a Direntry which represents
(conceptually) the directory in which the file resides with respect to
the root of the filesystem on which it is located.
struct direntry

Every file is represented in the kernel as an object of the file
structure. It requires a direntry and an inode provided to it.
struct file

Internal Structure of VFS
Click to edit Master text styles
Second level
● Third level

● Fourth level

● Fifth level

Filesystem Registration
Somewhere in the kernel load path, the inbuilt file systems are
loaded.

Filesystems that are implemented as modules register
themselves with the kernel (VFS) by using the following API at
module load time.

#include <linux/fs.h>

extern int register_filesystem(struct file_system_type *);
extern int unregister_filesystem(struct file_system_type *);

Filesystem Registration…
struct file_system_type {
const char *name; // “pipefs”
int fs_flags; // “mount flags”
int (*get_sb) (struct file_system_type *, int,
const char *, void *, struct vfsmount *); // “get superblock” on mount
void (*kill_sb) (struct super_block *); // “destroy superblock” on umount
struct module *owner; // which module is responsible for accesses
struct file_system_type * next; // VFS internal list of filesystem types
/* others */
};

 The superblock is created by the VFS whenever a device
of the appropriate filesystem type is mounted.

 The superblock provides the interface for VFS to access
filesystem specific functions.

get_sb()
The get_sb() function returns the superblock structure which
contains a pointer to the super_operations structure.

Functions specified in the super_operations specify how inodes are
handled by the specific filesystem implementation.
struct super_operations {
struct inode *(*alloc_inode)(struct super_block *sb);
void (*destroy_inode)(struct inode *);

void (*dirty_inode) (struct inode *);
int (*write_inode) (struct inode *, int);
void (*drop_inode) (struct inode *);
void (*delete_inode) (struct inode *);

/* lots of other members */

};

Where are we now?

DCache and
ICache for direct VFS
inode operations
struct
file_system_type

Superblock for pipefs

struct
super_operations
inode operations on
inodes not present in
the cache

The pipe() system call
SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
{
int fd[2];
int error;

error = do_pipe_flags(fd, flags); // Handle flags, create a file *
representing the pipe
if (!error) {
if (copy_to_user(fildes, fd, sizeof(fd))) { // copy to
userspace
sys_close(fd[0]); // if error – close and exit.
sys_close(fd[1]);
error = -EFAULT;
}
}
return error;
}

The pipe() system call…
Creates a directory entry
Creates an inode
Creates a file
Associates the directory entry and the inode with it
Creates 2 file * by opening it with O_RDONLY and O_WRONLY flags
It associates a file_operations structure with each descriptor
depending on the operations expected to be performed on it.
Installs the file *s in the file descriptor table of the process and
generates 2 file descriptors (of type int)
Returns the file descriptors in the fd[2] array.

Registering file_operations
const struct file_operations read_pipefifo_fops = {
.llseek = no_llseek,
.read = do_sync_read,
.aio_read = pipe_read,
.write = bad_pipe_w,
.poll = pipe_poll,
.unlocked_ioctl = pipe_ioctl,
.open = pipe_read_open,
.release = pipe_read_release,
.fasync = pipe_read_fasync,
};

File operations are associated with each created file * when first
opened
These operations tell the VFS how to perform various operations on
the file.

The pipe() system call…
The file created to represent the pipe is backed by buffers managed by
a pipe_inode_info object.
struct pipe_inode_info { // shown partially… : represents the pipe
wait_queue_head_t wait; // waiting queue for blocked writers
unsigned int readers;
unsigned int writers;
unsigned int waiting_writers;
struct inode *inode;
struct pipe_buffer bufs[PIPE_BUFFERS]; // backing data store – points to
actual pages.
};
PIPE_BUFFERS == 16

Atomicity guarantees:
PIPE_BUF is the maximum limit which is guaranteed for atomic
operations. PIPE_BUF = PAGE_SIZE (4096)
Atomicity is lost on page faults.

So, where are we?
User calls
pipe(fd[2])
VFS
do_pipe() in kernel

Install file
Creates pipefs file created on
descriptors in the
pipefs.
process
Pipe_inode_info

buf[16]
At initialization
time
Page 1 Page 2 Page 3 …

At syscall time

Read system call
SYSCALL_DEFINE3(read, unsigned int, fd, char __user *, buf, size_t, count)
{
struct file *file;
ssize_t ret = -EBADF;
int fput_needed;

file = fget_light(fd, &fput_needed); // get the file * associated with
the file descriptor

if (file) {
loff_t pos = file_pos_read(file);
ret = vfs_read(file, buf, count, &pos);
file_pos_write(file, pos);
fput_light(file, fput_needed);
}

return ret;
}

VFS Read
ssize_t vfs_read(struct file *file, char __user *buf, size_t count, loff_t *pos)
{
// verification code

if (ret >= 0) {
count = ret;
if (file->f_op->read) // The filesystem_operations
defined read function
ret = file->f_op->read(file, buf, count, pos);
else
ret = do_sync_read(file, buf, count, pos); // default fallback

// housekeeping code
}

return ret;
}

Pipe_read()
static ssize_t pipe_read(…) // simplified
{
struct file *filp = …;
struct inode *inode = …;
mutex_lock(&inode->i_mutex); // lock the pipe inode –
concurrency protection
for (;;) { // for each buffer
// do all the reading from the buffers.
// handle all the pending signals that might interrupt the system call.
// wake up all waiting writers that there might be more room
pipe_wait(pipe); // if have to block – release mutex and block
}
mutex_unlock(&inode->i_mutex);

// wake up waiting writers
// if bytes read is > 0 mark file as accessed.
}

Pipe_read()
Gets the file * for reading
Locks its associated inode
Accesses the pipe buffers and reads off them
Copies the read data to userspace in the passed char * buffer in
the read() system call.
Wakes up the writers that are blocked on the wait queue
Releases the lock on the inode

In case the read blocks, it releases the mutex and calls schedule()
which invokes the scheduler, allowing a context switch.

So where are we?
Read() system call

vfs_read()

Intermediate
functions

pipe_read() Schedule()

Pipe_inode_info
buf[16]

Access controlled
by the inode
Page 1 Page 2 Page 3 …
mutex

The Write System Call
SYSCALL_DEFINE3(write, unsigned int, fd, const char __user *, buf,
size_t, count)
{
struct file *file;
ssize_t ret = -EBADF;
int fput_needed;

file = fget_light(fd, &fput_needed); // get the kernel file * from the (int)
fd provided by the user
if (file) {
loff_t pos = file_pos_read(file);
ret = vfs_write(file, buf, count, &pos); // go to the VFS layer
file_pos_write(file, pos);
fput_light(file, fput_needed);
}

return ret;
}

The Write System Call…
Follows a similar code path as for the Read system call
Functions
write()
vfs_write()
pipe_write()

A point to note is that all copying of data to/from userspace is done
via vector operations and page faults destroy atomicity constraints.

Kernel Pages for temporary storage of data are allocated on demand
upto PIPE_BUFFERS pages.

Pages are “pre-faulted” in userspace to ensure that copying data
remains atomic (under certain size constraints).

FIFOs
FIFOs are created with the mkfifo() or mknod() calls

mkfifo(const char *pathname, mode_t mode ) – userspace call
mknod(…) system call
vfs_mknod(…) for a FIFO creation
●
creates and installs a node in the global filesystem at the specified
path and with the requested permissions.
●
The file node has an associated inode and directory entry just like
for a pipe.

The mknod(…) system call handles creation of nodes in the filesystem of a variety
of things:
Regular Files
Block and Character devices
FIFOs and Sockets

FIFOs
When an open(…) system call takes place on the
filename:
long do_sys_open(int dfd, const char __user *filename, int flags, int mode) //
simplified
{
if (…) {
if (fd >= 0) {
struct file *f = do_filp_open(dfd, tmp, flags, mode); // open the fifo
if (IS_ERR(f)) {
// …
} else {
fsnotify_open(f->f_path.dentry);
fd_install(fd, f); // install it to the file descriptor table
}
}
}
return fd;
}

Do_filep_open()
Does a lot of processing at the VFS level and finally calls the .open handler
registered in the file_operations for the FIFO file.

Recap: Every file has an associated file_operations structure that
specifies how the file may be manipulated by VFS.

struct file_operations { // simplified
struct module *owner;
loff_t (*llseek) (struct file *, loff_t, int);
ssize_t (*read) (struct file *, char __user *, size_t, loff_t *);
ssize_t (*write) (struct file *, const char __user *, size_t, loff_t *);
int (*ioctl) (struct inode *, struct file *, unsigned int, unsigned long);
int (*mmap) (struct file *, struct vm_area_struct *);
int (*open) (struct inode *, struct file *);
};

fifo_open(…)
The file_operations .open member points to fifo_open(…)

fifo_open(…) does the following tasks:
Locks the inode of the file
Allocates a pipe_inode_info object for providing buffer space
●
This is the actual “pipe”
Depending on the open flags: O_RDONLY, O_WRONLY it allocates
a file pointer
Increments the reference counter of the number of
readers/writers on the pipe
Unlocks the inode and returns the file *

So, how did it happen again?
Open() system
mkfifo()
call

mknod() system
do_sys_open()
call

Intermediate
fd_install()
functions do_filep_open()
file * into the
Gets the file *
process
do_sys_mknod()
VFS
file_operations.op
node for FIFO en
created fifo_open() At fifo open
Handles all FIFO
time.
At FIFO creation restrictions
time

After open(…)
After the FIFO is successfully opened, it is equivalent to a pipe.

Reading and writing to it are governed by the pipe read/write
functions.

Effect of fork()
Forking a new process simply copies the file descriptor table and its
associated file * to the new process and updates the readers/writers
count as appropriate within the pipe_inode_info object.

Effect of close()
The file descriptor is flushed

The reader/writer count for the pipe is decremented

If the count is zero, deallocate the pipe buffers and get rid of the pipe
file object.

If it is a fifo, don’t get rid of the node entry, but get rid of the file
object in the kernel

FIFOs require an explicit remove/unlink(…) on the file node for their
removal from the filesystem.

Generalization to other filesystems
Other (device backed) filesystems go a bit further.

The buffers that are managed by the inodes are actually allocated
from a common buffer cache.

A read or write request to a buffer that is not found in the buffer
cache generates a device block IO request otherwise IO savings are
made.

Dirty buffers are periodically flushed to the IO devices and the cache
maintenance is done by the bdflush kernel daemon.

/proc
Proc is a virtual file system (procfs)

All nodes in the /proc filesystem are created from kernel space in an
on-demand manner.
Filesystem node
File
Direntry
Inodes

Reads/Writes/Opens of files on this device are handled by kernel
functions/modules wishing to export/import their data to/from
userspace.
These functions implement the appropriate API to communicate
with user processes via /proc

Conclusion
The linux implementation of pipes is very efficient and allows for easy
IPC.

FIFOs are implemented as wrappers around pipes

The /proc filesystem is created dynamically by the kernel at runtime
by suitable manipulation behind the scenes.

References
The linux kernel sources documentation
linux-source-2.6.27/Documentation/filesystems/vfs.txt
The actual linux kernel source code
linux-source-2.6.27/include/linux/pipe_fs_i.h
linux-source-2.6.27/fs/open.c
linux-source-2.6.27/fs/pipe.c
linux-source-2.6.27/fs/fifo.c
A tour of the Linux VFS by Michael K. Johnson. 1996
http://www.tldp.org/LDP/khg/HyperNews/get/fs/vfstour.html

Thank You

Any Questions?

The Linux Kernel Implementation of Pipes and FIFOs

More Related Content

What's hot (20)

Similar to The Linux Kernel Implementation of Pipes and FIFOs (20)

More from Divye Kapoor (6)

Recently uploaded (20)

The Linux Kernel Implementation of Pipes and FIFOs