File systems

Gaël Thomas

Device and device driver

Device and device driver

  • Device = hardware component other than CPU and memory

  • Device driver = software allowing access to a device

    • 1 data structure giving the status of the device
    • 1 input / output function allowing access to the device
    • The driver is usually found in the kernel

Devices in UNIX

  • A device is identified by a number called dev

    • Most significant bits (major): driver number

      • For example: 8 = ssd hard drive driver
  • Least significant bits (minor): device number

    • For example: 0 = disk 1, 1 = disk 1 / part 1, 2 = disk 1 / part 2
  • The kernel contains a table which associates a driver number with the driver (access function + status)

2 types of peripherals

  • “character” devices

    • Read / write byte by byte
    • Generally access via MMIO or input / output bus

    blocks the CPU during the I/O operation

    • Keyboard, printer, sound card …
  • “block” devices

    • Read / write by data blocks (typically 512 bytes)
    • The device is therefore seen as an array of blocks
    • Usually access via DMA

    does not block the CPU during the I / O operation

    • Hard disk, DVD player …

Block devices in xv6

  • A single block device driver in xv6

    • Manages virtio hard disks (emulated by Qemu)
    • Function virtio_disk_rw() in virtio.c
  • virtio_disk_rw() takes two parameters:

    • a boolean, write, to tell if it is a read or a write

    • a buf (buf.h) structure

      • access to block blockno from disk dev
      • data read or written
        • If write == 0, the output of virtio_disk_rw, data = data read
        • If write == 1, the input of virtio_disk_rw, data = data to write

Principle of the virtio_disk_rw algorithm

  • virtio_disk_rw mainly performs the following actions:
    • Setup the DMA data transfer:

      • From disk to memory on a read
      • From memory to disk on a write
    • Sleep the process with the sleep function (see lecture #4)

      switch to another ready process

  • Once the transfer is complete
    1. The virtio disk generates an interrupt
    2. The interrupt is handled by the virtio_disk_intr function
    3. virtio_disk_intr calls wakeup to wake up the sleeping process

The I / O cache

  • Disk access is very slow compared to memory access

    • Hard disk drive: several milliseconds
    • SSD disk: x10, hundreds of microseconds
    • NVMe disk: x100, microseconds
    • Memory: x100, dozens of nanoseconds
  • I/O cache improves the performance of block type devices

    • Keeps frequently or recently used blocks in memory
    • Managed by the operating system kernel

Principle of an I/O cache

  • The system manages a set of buffers in memory

  • To read a block (read operation)

    • If the block is not yet in the cache
      1. Remove an unused buffer from the cache
      2. Copy the contents of the disk block to this buffer
    • Otherwise, simply return the buffer associated with the block
  • To modify a block (write operation)

    1. Read the block (call the read operation)
    2. Modify the contents of the buffer in memory
    3. Mark buffer as modified (written to disk later)

The xv6 buffer cache

  • buffer cache = xv6 I/O cache (bio.c)

    • Made up of a finite set of buf structures
    • Each buf structure is associated with a block of a disk
      • A buf can be valid if its block’s data has been read, invalid otherwise
    • Each buf has a reference counter to avoid eviction while still in use

How the buffer cache works: buffer management (1/3)

  • The buf structures form a circular double linked list, the head is the most recently used block
  • struct buf* bget(uint dev, uint blkno): return a locked buffer associated with (dev, blkno)
    • If there is already an buffer associated with (dev,blkno)
      • Increment the reference counter of the buffer
      • Lock the buffer
      • Return the buffer
    • Otherwise
      • Search for a buffer with counter == 0
      • Associate the buffer with (dev, blkno)
      • And then, same as above

How the buffer cache works: read buffer (2/3)

  • struct buf* bread(uint dev, uint blkno)
    • Goal: return a locked buffer for this block in the valid state
      1. Call bget() to find a buffer for this block
      2. If the buffer is invalid, call virtio_disk_rw()
  • void bwrite(struct buf* b)
    • Call virtio_disk_rw() to write the buffer data to the disk

How the buffer cache works: write buffer (3/3)

  • void brelse(struct buf* b)
    • Release the lock associated with b
    • Decreases the reference counter
    • Move the buffer to the head of the list (most recently used) if it is unused

The log

Operation versus writing to disk

  • A write operation of a process often requires several block writes

    • File creation requires:

      • Allocation of a new file
      • Adding the name to a directory
    • Adding data to a file requires:

      • Writing new blocks to disk
      • Updating the file size
    • Deleting a file requires:

      • Deleting the data blocks from the file
      • Deleting the name from the directory

Consistency issues

  • The system can crash anytime

    Inconsistency if it stops in the middle of an operation

    • A name in a directory references a non-existent file
    • Data added to a file but size not updated
  • Operations must be propagated in the order in which they were performed

    Inconsistency if propagation in random order

    • Adding a file then deleting the file does not exist at the end
    • Deleting a file then adding the file exists at the end
    • Similarly, adding data then truncating (size should be 0)

Bad solutions

  • No cache when writing (directly propagate write operations)

    • Very inefficient because each write becomes very (very!) slow
  • Recovery in the case of a crash

    • Recovering a file system is slow

    • examples: FAT32 on Windows or ext2 on Linux

    • Recovering is not always possible

      a crash makes the filesystem unusable!

First idea: transactions

  • A transaction is a set of writes that is

    • Either fully executed
    • Or not executed at all
  • Principle of implementation

    • An operation (coherent set of writes) == a transaction
    • The writes of a transaction are first written to disk in a “pending” area
    • Once the operation is complete, the “pending” area is marked as valid (the transaction is complete)
    • Regularly (or in the event of a crash), validated writes in the pending zone are propagated to the file system

Second idea: log

  • To ensure that the entries are propagated in order in which they were executed, the pending zone is structured like a log

    • Each entry is added at the end of the log
    • The validated transactions of the pending zone are propagated to the file system in the order of the log (from the start of the log to the end)

Third idea: parallel log

  • Problems: Multiple processes may perform transactions in parallel

    • Parallel transaction writes are interleaved in the log

    How do you know which ones are validated?

  • Classic solution

    • If several transactions in parallel, all the operations are validated when the last one is completed
    • Advantage: easy to implement (count of the number of operations in parallel)
    • Disadvantage: risk of never validating if new operations continue to arrive

Log structure

  • The system technically manages two logs

    • One in memory called memory log

      • Contains only the list of modified block numbers
      • The content of the modified blocks is in the buffer cache
    • One on disk called disk log

      • Contains the list of modified block numbers and a copy of the blocks
      • Note: the block is propagated from the log to the filesystem later

The system can therefore manage up to 3 copies of a block

  • One on disk in the file system called disk block
  • One on disk in the log called disk log block
  • One in memory in the buffer cache called cached block

Log algorithm principle

  • Steps to modify block number n
    1. load the disk block in the buffer cache
    2. modification of the buffer (i.e. cached block)
    3. add n to the list of modified blocks in the memory log
  • At the end of an operation, steps to validate the transaction
    1. copy modified cached blocks to disk log
    2. copy the modified block list to disk log
    3. mark the transaction as validated
  • Later, to propagate the transaction
    1. copy disk log blocks to file system
    2. reset disk log and memory log

Using the log

  • Three functions in the log management interface (log.c)

    • begin_op() : start a transaction
    • end_op() : validate a transaction
    • log_write(struct buf* b) : add b to the transaction
  • To perform a logged operation, instead of calling directly bwrite (), we have to execute:

b = bread(...);
// Modify data of b

Implementation in xv6 (1/3)

  • void begin_op() : start a transaction

    • If log writing to disk in progress, wait
    • If the log is full, wait
    • Increments the number of pending operations (log.outstanding)
  • void end_op() : complete a transaction

    • Decrement the number of operations in progress, and if equal to 0:

      • Write memory log + cached blocks in disk log (write_log())
      • Mark committed disk log transaction (write_head())
      • Propagate writes from disk log to the filesystem (install_trans())
      • Delete logs in memory and on disk (write_head())

Implementation in xv6 (2/3)

  • void log_write(struct buf* b)
    • Goal: put the block associated with b in the log
    • Find an entry for the block in the log
      • If already in the log: absorb the log entry (i.e., do nothing: the block is already logged to be written)
      • If new to the log:
        1. Add block number to the memory log
        2. Increase the reference counter of the buffer b to prevent it from leaving the buffer cache

Implementation in xv6 (3/3)

  • After a crash, call install_trans() which propagates the writes from disk log to file system

    • In the worst case, writes that had already been performed are replayed
    • But at the end of the replay, the filesystem is in a consistent state

Partitions and file systems

File system

  • File system: defines the structure for storing files (often for a block type device)

    • UFS : Unix Files System (xv6, BSD)
    • ext : extended file system (Linux - ext4 nowadays)
    • NTFS : New Technology File System (Windows)
    • APFS : APple File System (MacOS)
    • FAT : File Allocation Table (Windows)
    • BTRFS : B-TRee File System (Linux)
    • and many others !

Principle of a file system

  • File = consistent set of data that can be read or written

  • Filesystem = associate names and files

    • Example : /etc/passwdroot:*:0:0:System Administrator...

    • Usually a special symbol is used as a separator for directories

      • / in UNIX systems, in Windows systems


  • A disk is often made up of several partitions

    • Partition = continuous area that contains a file system
  • Typical structure of a disk

    • First block: partition table

      • For example: Master Boot Record
    • Blocks 2 to x: kernel loader

      • In charge of loading the kernel of one of the partitions
      • For example: LILO, GRUB
    • Blocks x to y: partition 1

    • Blocks y to z: partition 2

    • etc…

Disk image

  • A file itself can contain the data of a complete disc

    • Called a disk image or a virtual disk
    • Typically used in virtualization
    • For example: xv6.img is the disk image used with the qemu emulator to start xv6

UFS/xv6 file system

Overall file system structure

  • Five large contiguous zones (in fs.h)

    • The super block describes the other areas
    • The journal contains the disk logs
    • The dinode table contains the metadata of the files (size, type like ordinary or directory …)
    • The table of free blocks indicates the free blocks
    • The data blocks area contains the data of the files


  • A file on disk consists of:

    • metadata called a dinode (fixed size, see fs.h)

      • file type (ordinary, directory, device)
      • file size
      • the list of the file data blocks
      • an indirection block (see following slides)
      • device number if device file
      • number of hard links to the file (reminder: a hard link is a name in a directory)
    • data blocks

      • these are the blocks that contain the content of the file

Data blocks of a file

  • A dinode directly lists the numbers of the first 12 blocks

    • the dinode.addrs [0] block contains bytes 0 to 511 of the file
    • the dinode.addrs [i] block contains the bytes i * 512 to i * 512 + 511
  • The indirection block contains the following block numbers

    • the indirection block number ind is given in dinode.addrs [12]
    • the ind [0] block contains bytes 12 * 512 to 12 * 512 + 511
  • Note: since a block is 512 bytes and a block number is coded out of 4 characters, a file has a maximum size of 12 + 512/4 blocks.

Adding a block to a file

  • To add a new block to a dinode dino (function bmap () in fs.h)

    1. Find a free block number in the table of free blocks \ (function balloc() in fs.h)
    2. Mark the occupied block (put its bit 1 in the table)
    3. Add the block number to the list of data blocks in dino
    • this addition may require to allocate an indirection block


  • A directory is a file of type T_DIR

  • Contains an array associating names and numbers of dinodes

    • inum: inode number
    • name: file name
  • Inode 1 is necessarily a directory: it is the root directory of the filesystem

  • Note: dinode.nlink gives the number of times a dinode is referenced from a directory

    file deleted when nlink equals to 0.

From path to inode

  • To find a dinode number from the path /e0/../en (see namex() in fs.c)
cur = 1
For i in  [0 .. n]
    Look for the association [inum, name] in the data blocks of
        the cur dinode such that name is ei
    cur = inum

File creation and deletion

  • To create the file f in the d directory (function create() in sysfile.c)

    1. Find a free inum dinode by finding an inode whose type is 0 in the dinode array (ialloc () in fs.h)
    2. Add the association [inum, f] to d
  • To delete the file f from the d directory (sys_unlink() function in sysfile.c)

    1. Delete the entry corresponding to f in d
    2. Decrement nlink from f and if nlink equals 0
    3. Delete data blocks from file f
    4. Remove the inode f (setting its type to 0)

xv6 I/O stack


  • inode = memory cache of a dinode

    • Enter the cache at open()

    • Can be evicted from cache from close()

    • Contains the fields of the dinode

    • + fields to know which dinode the inode corresponds to

      • Device number and dinode number
    • + fields required when the dinode is used

      • A lock to manage concurrent access
      • A counter giving the number of processes using the inode to know when the inode can be evicted from the cache
  • Inode table = table which contains the inodes

Main functions of inodes (1/3)

  • struct inode* iget(int dev, int inum)

    • Corresponds to open(): returns an inode associated with [dev, inum]

    • Increments the inode usage counter (non-evictable)

    • Do not lock the inode and do not read the inode from disk (optimization to avoid disc playback when creates a file)

      • inode.valid indicates whether the inode has been read from disk
  • void ilock(struct inode* ip)

    • Acquires a lock on the inode
    • Read inode from disk if not already read
  • void iunlock(struct inode* ip)

    • Release the lock on the inode

Main functions of inodes (2/3)

  • void itrunc(struct inode* ip)

    • Free all the blocks in the file (size 0)
  • void iupdate(struct inode* ip)

    • Copy the inode to the disk dinode (technically, via the I/O cache)

Main functions of inodes (3/3)

  • void iput(struct inode* ip)

    • Corresponds to close ()

    • Decreases the inode usage counter

    • If cpt drops to 0, the inode can be evicted from the cache and

      • If nlink is 0 (the inode is no longer referenced by a directory)

        • Delete data blocks from inode (itrunc)
        • Mark the inode as free (type = 0)
  • Note: if you delete a file from a directory (unlink()) while the file is still in use (open) by a process, the inode is not deleted: it will be when last close() when the reference counter drops to 0.

Open files

  • Multiple processes can open the same file

    • Each process has independent read / write permissions
    • Each process has a read cursor, which is independent of that of the other processes
  • A file structure opened by open () contains:

    • A pointer to an inode
    • Access permissions
    • A reading cursor

File descriptors

  • Each process has an ofile table of open files

    • A descriptor d is an index in this table
    • proc[i].ofile[d] points to an open file
    • proc[i].ofile[d].ip points to inode
  • Good to know

    • During a fork(), the parent and the child share the open files
    • So proc[parent].ofile[d] == proc[child].ofile[d]
    • And so, if the father reads, the child read cursor changes
    • Useful for setting up pipes

What you must remember

  • A device driver is just a function (virtio_disk_rw() for example)

  • Reads and writes are logged

    • Ensures file system consistency in the event of a crash
  • The kernel has an I/O cache

    • Is in memory, managed by the kernel
    • Allows to speed up I/O
  • A file system separates

    • The naming (directory) of the files (dinodes + data blocks)
    • The metadata (dinode) of the data blocks
  • A file descriptor is an index in the ofile table

    • proc->ofile[i] is an open file that references an inode