UNIX: System Design [282]

• no real design before implementation

• first development in 1969 at Bell Labs by Thompson and Ritchie

• Ritchie had worked on MULTICS

• UNIX is a pun on MULTICS

• file system and shell are similar

• implemented in C

• designed by programmers for programmers

• designed to be a time-sharing system

• not much layering (see earlier slide)

• shell programs combine ordinary programs

• pipes for redirection of input/output: % myprog <indata >outdata

UNIX: Processes [283]

• process is a program in execution

• new process created by fork

• parent can create a child

• child runs the same program as parent

• typically, child will execve a new program

• parent waits for a child's exit

• if a parent exits, then the child is a zombie

• fork allows a pipe between parent and child (see IPC)

• read from empty pipe or write to full pipe: block

• signal handles exceptional conditions (keyboard interrupt)

• process can ignore a signal or have a signal handler routine

• signals can be used to start and stop subprocesses on demand

UNIX: Process Control Block [284]

• called process structure

• process ID, priority, etc.

• array of PCBs defined at system linking time

• ready queue is doubly-linked list through the PCBs

• pointers to parent, youngest living child, other relatives

• normal execution is in user mode

• system calls switch to system mode

• system mode uses kernel stack for that process

UNIX: Scheduling [285]

• designed to benefit interactive processes

• small CPU time slices using a priority algorithm

• larger numbers indicate lower priority

• processes doing disk I/O are less than ``pzero''

• ordinary user processes have positive priorities

• less likely to run than any system process

• user processes can set precedence over one another using nice

• high CPU usage IMPLIES lower priority (more positive)

• process aging prevents starvation

• time-slice every 0.1 sec and recompute priorities every 1 second

• round-robin uses timeout which tells clock to interrupt

• reschedule and then set another timeout

• priority recomputation also uses timeout

UNIX: Events [286]

• relinquish CPU because of I/O or time slice expired

• or sleep waiting for some event

• argument is address of kernel data structure for that specific event

• system calls wakeup on all sleeping processes for that event

• wait for disk I/O to complete (sleep on address of buffer header)

• race condition:

• process decides to sleep (based on, say, flag)

• event occurs

• process calls sleep (but event will never occur)

• raise hardware processor priority for this critical section

• no interrupts and process can run until sleeping

UNIX: Memory Management [287]

• early system just swapped out processes if not enough memory

• PID0=scheduler process (swapper) wakes up every 4 secs

• swap out a process if:

• idle

• in main memory a long time

• large

• old

• swap in a process if:

• swapped out a long time

• small

• some UNIX systems still do this

• Berkeley UNIX uses demand paging and secondarily swapping

UNIX: Demand Paging [288]

• virtual memory

• swapping kept to minimum because more jobs in memory

• only parts of each process in memory

• list of free frames

• modification of second chance (clock) algorithm

• memory is swept linearly and repeatedly by software clock hand

• if frame is already free or in use (say for I/O), skip

• otherwise, go to page-table entry for frame

• if invalid, put frame on free list

• otherwise, mark as invalid but reclaimable

• clock hand implemented by pagedaemon

• runs only if free frames falls below threshold

• if hardware supports ref bit, then use it

• one pass of the clock turns bits off

• second pass checks bit and puts onto free list

UNIX: File System [289]

• file is a sequence of bytes

• files organized by directories

• file data linked by inodes (see earlier slide)

• system calls use a file descriptor as an argument

• kernel indexes into table of open files for current process

• each entry points to a file structure

• each file structure points to an inode

UNIX: I/O System [290]

• disk files and I/O devices treated as similarly as possible

• general device driver code

• specific device driver code for each device

• block devices

• disks and tapes

• array of entry points for drivers

• use a block buffer cache for block I/O

• or use a queue of pending transfers for raw device interfaces

• character devices

• terminals, line printers, etc.

• array of entry points for drivers

• C-lists are small blocks of characters

• write enqueues onto the list for the device

• interrupts cause dequeueing

• input is also interrupt driven

UNIX: IPC [291]

• not one of the strong points of UNIX

• pipe, shared files, messages, shared memory

• sockets

• stream: reliable and sequenced stream

• datagram: unreliable and unsequenced messages

• socket creates a socket and returns descriptor

• descriptor indexes into array of open ``files''

• bind assigns a name to a socket

• client-server model

• server creates socket and binds to well-known address

• client uses connect on well-known address

• server listens to say that it is ready for connections

• server accepts individual connection

• server usually forks a process to talk with client

• server goes back to listening

• server subprocess and client do read and write