Title: CS6900 Independent study
1CS6900 Independent study
- Bolin Hsu
- ltbolin_hsu_at_yahoo.comgt
2Introduction
- Relationship of Linux kernel and Linux device
drivers - Implementation of a Linux device driver
- How to write your own device driver
- Bibliography
3Concepts
4Objective
- Let the users access the device by file system
calls -
- fdopen(/dev/myrd,O_RDWR)
-
- write(fd,buffer,sizeof(buffer))
-
- read(fd,buffer,N)
-
- close(fd)
5Challenges
- How to connect the device file /dev/myrd to the
device - How to make the system calls access the device
6Linking of device file and the device
7Everything in UNIX is a file
- User interact with the device through filesystem
functions - open, close, read, write, ioctl, etc
- A device file or a filesystem
- Two interfaces to the device
- Interface to user device file
- Interface to kernel service routine
8Device file concept
- Device file type
- Character (inode.type 3)
- Block (inode.type 4)
- Device file number
- Major device type (2 floppy disk 3 hard disk,
etc) - Minor device number
- Examples
- brw-rw---- 1 root disk 3, 0 May 19 2004 /dev/had
9Creating the device file
- Create our ram disk device file
- Type is block
- Major number is 100
- Minor number is 0
- mknod /dev/ b 100 0
- Another option is to create device during device
driver initialization.
10Connecting system calls to the device driver
11How Linux kernel uses device drivers
User applications
Libraries
User mode
Kernel mode
File subsystem
Buffer/page cache
Block device driver
Character device driver
Hardware control
12Block I/O and Page I/O
- Two fundamental I/O type
- Block I/O bread()
- Page I/O brw_page()
- Block I/O read/write a single block. Kernel uses
it to handle file system superblock and inode. - Page I/O read/write a page in a file. Used in
read/write system calls.
13Page/buffer cache
- A cache is a RAM area
- The cache holds data which usually should be on a
block device. - I/O from the cache is faster than the device.
- Kernel can merge and schedule device I/O to
improve system performance.
14Sector, block, and buffer cache
- A sector is the basic data transfer unit of a
block device - A block is a collection of adjacent sectors
transferred in one I/O request - Every block device file has its own block size
- Every block has a corresponding buffer cache
15How device request uses buffer cache
request
Buffer head
Buffer head
Buffer head
request
16Device driver API calling sequence
open
close
read
write
device driver API
buffer cache calls
block device switch table
open
release
request
ioctl
check_media_change
Interrupt service ruotine
revalidate
Interrupt vector
device interrupt
17How read() works
- The read() system call searches the page in the
page cache. If the page is in the cache and
valid, it is copied to the user buffer. If no
valid page is found in the cache, the process
issues a request for the page and suspend itself. - When the page becomes available later, the kernel
wakes up the suspended process. The process then
copies the page to the user buffer.
18Trace read() system call
- These functions are called in order
- read()
- sys_read()
- def_blk_fops.read generic_file_read()
- do_generic_file_read()if the page is found in
the page cache, it is copied to the user buffer.
If it is not in the cache, call next function.
This function could suspend the calling process. - def_blk_aops.readpage blkdev_readpage()
- block_read_full_page()
- blkdev_get_block()
19How write() works
- The write() system call copies the user buffer to
the page cache and returns. - Unlike read(), the write() system call doesnt
wait for the transfer between page cache and the
device to complete. There is no data dependency
in write operation.
20Trace write() system call
- These functions are called in order
- write()
- sys_write()
- def_blk_fops.write generic_file_write()
- def_blk_aops.prepare_write blkdev_prepare_write()
def_blk_aops.commit_write blkdev_commit_write() - block_prepare_write()block_commit_write()
21How open() works
- open() makes the file subsystem read the inode of
the file from the device. - The kernel then checks the inode type and knows
it is a device file. The kernel then initializes
the inode structure in RAM to contain file
operations defined in the device driver. The
kernel uses the major number of the device to
find the file operations. - From this point on, file operation calls to this
inode will cause the execution of the
corresponding device driver functions.
22Trace open() system call
- These functions are called in order
- open()
- sys_open()
- file_open()
- dengry_open()
- def_blk_fops.open blkdev_open()
- do_open()
- get_blkfops()retrieves blkdevsMAJOR.bdops,
stores in bdev.bd_op
23The request queue of the device
- Requests to a block device is maintained in a
request queue of the device. - BLK_DEFAULT_QUEUE(MAJOR) finds the request queue
by the major number. - A request function from the device driver handles
the request. - The request function is registered when the queue
is initializedblk_init_queue(BLK_DEFAULT_QUEUE(M
AJOR),request_function)
24Implementation
25Whats in a device driver
- A device driver is like a server. It has
- A set of functions to services
- Functions to publish the services to the kernel
- Some internal data structures for bookkeeping
purposes
26In-kernel or kernel module
- Two ways of adding device driver to the kernel
- Compile into the kernel
- Loadable kernel module
- Benefit of kernel module
- No need to rebuild kernel
- No need to reboot system
27Block device interface to Linux kernel
- /usr/src/linux/fs/block_dev.c
- struct file_operations def_blk_fops
- open blkdev_open,
- release blkdev_close,
- llseek block_llseek,
- read generic_file_read,
- write generic_file_write,
- mmap generic_file_mmap,
- fsync block_fsync,
- ioctl blkdev_ioctl,
-
- // Warning nonstandard gcc structure
initialization
28File system interface of the block device
- /usr/src/linux/include/linux/fs.h
- struct block_device_operations
- int (open) (struct inode , struct file
) - int (release) (struct inode , struct
file ) - int (ioctl) (struct inode , struct file
, unsigned, unsigned long) - int (check_media_change) (kdev_t)
- int (revalidate) (kdev_t)
- struct module owner
- // file system hooks for the block devices
29Important global array
- read_ahead // how many blockblk_size // device
sizeblksize_size // block sizehardsect_size //
sector size - These arrays are indexed by major number
- Kernel fetches information about the device from
these arrays
30Initializing the device driver
- Done in init_module
- Allocate memory
- Initialize the device drivers slots in global
arrays - Initialize the request queue of the device driver
- Register the device driver (API)
31Initialization example
- Int Init_module(void)
- ...
- myrd_storage (char )kmalloc(210241024sizeof
(char), GFP_KERNEL) - ...
- read_aheadMAJOR_NR 2 // sectors
- blk_sizeMAJOR_NR kmalloc(sizeof(int),
GFP_KERNEL) - blk_sizeMAJOR_NR 2048 // device size
(kilobytes) - blksize_sizeMAJOR_NR kmalloc(sizeof(int),
GFP_KERNEL) - blksize_sizeMAJOR_NR 512 // block size
(bytes) - hardsect_sizeMAJOR_NR kmalloc(sizeof(int),
GFP_KERNEL) - hardsect_sizeMAJOR_NR 512 // sector size
(bytes) - blk_init_queue(BLK_DEFAULT_QUEUE(MAJOR_NR),
myrd_request) - register_blkdev(MAJOR_NR,"myrd",myrd_ops)
- register_disk(NULL, MKDEV(MAJOR_NR, 0), 1,
myrd_ops, 100) - ...
32Cleanup example
- void cleanup_module(void)
-
- ...
- fsync_dev(MKDEV(MAJOR_NR, 0))
- blk_cleanup_queue(BLK_DEFAULT_QUEUE(MAJOR_NR))
- kfree(hardsect_sizeMAJOR_NR)
- hardsect_sizeMAJOR_NR NULL
- kfree(blk_sizeMAJOR_NR)
- blk_sizeMAJOR_NR NULL
- kfree(blksize_sizeMAJOR_NR)
- blksize_sizeMAJOR_NR NULL
- kfree(myrd_storage)
- unregister_blkdev(MAJOR_NR,"myrd")
33Servicing a request
- The kernel passes to the request function the
beginning sector and the number of sectors to
read. - The request function can calculate the starting
location and size bystart_loc sector
SECTOR_SIZEsize nr_sectors SECTOR_SIZE - Assume the RAM disk stores the data in an byte
array data, the memory region to be copied is
datastart_loc to datastart_locsize-1
34Request routine example
- void myrd_request(request_queue_t q)
-
- ...
- while(1)
-
- INIT_REQUEST // this macro does a return when
queue is empty - location myrd_storage CURRENT-gtsector
MYRD_HARDSECT_SIZE - size CURRENT-gtcurrent_nr_sectors
MYRD_HARDSECT_SIZE - spin_lock(myrd_lock)
- if (CURRENT-gtcmd READ)
- memcpy(CURRENT-gtbuffer, location, size)
- else if (CURRENT-gtcmd WRITE)
- memcpy(location, CURRENT-gtbuffer, size)
- spin_unlock(myrd_lock)
- end_request(1) // update the queue and
requests -
-
35Adding the device driver into the Linux source
tree
- Modify the Config.in in the directory
- Let user choose Y/N/M for this questiontristate
Configure My RAM disk CONFIG_MYRD - Add your code to the makefile
- obj_(CONFIG_MYRD) myrd.o
- The above line adds the object to obj_m or obj_y,
depending on the answer in make config
36Roll your own device driver
37Helpful courses
- If you dont feel comfortable working with the
hardware, these two courses are very helpful - CS3432 Digital circuit design
- CS3434 Computer interfacing
38A checklist of tasks
- Create the device file
- Define the default file operations
- def_blk_ops
- def_chr_fops
- Initialize the global data arrays
- Register the device driver to the kernel
- Cleanup everything when you are done
39How do I drive a real device
- The derive driver interface to the user and the
kernel remain the same. - The request function will have to interact with
the device. This is usually done by reading and
writing to device registers. - If the device is slow, you may want to use the
interrupt driven approach. The process sets up an
I/O operation and suspends. When the I/O
completes, the device raise an interrupt request
and the interrupt service routine in the device
driver is called to handle the completed transfer.
40How do I drive character device
- Needs to supply the file operations defined by
the file system. - No cache is used. Thus no need of the request
function. - Needs to initialize the global arrays of the
kernel and register the driver. - Still needs to implement interrupt service
routine.
41Looking ahead the 2.6.x kernel
- The 2.6 kernel removed several macro used in this
presentation. - The 2.6 kernel changed some global bookkeeping
arrays. - The 2.6 kernel is preemptive.
- The nonstandard structure initialization was
removed in 2.6.x
42Other types of drivers
- Sometimes character and block driver models are
not suitable - Network interface card driver
- Bus driver
- PCI, SCSI, etc.
- Need to deal with packets like in networking.
- And more
43Suggestions to future work
- Build a file system on the ram disk
- Partition the data array into boot block, super
block, inodes, data blocks. - Supply mount() and unmount() system calls, plus
other system calls. - Port the RAM disk driver to kernel 2.6.x
44Bibliography
- Linux device drivers, 2ed Alessandro Rubini
Jonathan Corbet, OReilly, Sebastopol, CA - Understanding the Linux kernel Daniel P.l Bovet
Marco Cesati, OReilly, Sebastopol, CA
ltftp//ftp.oreilly.com/pub/examples/linux/drivers2
/book_pdf.rargt - The design of the UNIX operating system Maurice
J. Bach, Prentice Hall, Englewood Cliffs, New
Jersey - Writing a UNIX device driver, 2ed Janet I. Egan,
Thomas J. Teixeira, John Wiley Sons, Inc.