| 1 | .de P1 |
| 2 | .DS |
| 3 | .. |
| 4 | .de P2 |
| 5 | .DE |
| 6 | .. |
| 7 | .de UL |
| 8 | .lg 0 |
| 9 | .if n .ul |
| 10 | \%\&\\$3\f3\\$1\fR\&\\$2 |
| 11 | .lg |
| 12 | .. |
| 13 | .de UC |
| 14 | \&\\$3\s-1\\$1\\s0\&\\$2 |
| 15 | .. |
| 16 | .de IT |
| 17 | .lg 0 |
| 18 | .if n .ul |
| 19 | \%\&\\$3\f2\\$1\fR\&\\$2 |
| 20 | .lg |
| 21 | .. |
| 22 | .de SP |
| 23 | .sp \\$1 |
| 24 | .. |
| 25 | .hw device |
| 26 | .TL |
| 27 | UNIX Implementation |
| 28 | .AU "MH 2C-523" 2394 |
| 29 | K. Thompson |
| 30 | .AI |
| 31 | .MH |
| 32 | .AB |
| 33 | This paper describes in high-level terms the |
| 34 | implementation of the resident |
| 35 | .UX |
| 36 | kernel. |
| 37 | This discussion is broken into three parts. |
| 38 | The first part describes |
| 39 | how the |
| 40 | .UX |
| 41 | system views processes, users, and programs. |
| 42 | The second part describes the I/O system. |
| 43 | The last part describes the |
| 44 | .UX |
| 45 | file system. |
| 46 | .AE |
| 47 | .NH |
| 48 | INTRODUCTION |
| 49 | .PP |
| 50 | The |
| 51 | .UX |
| 52 | kernel consists of about 10,000 |
| 53 | lines of C code and about 1,000 lines of assembly code. |
| 54 | The assembly code can be further broken down into |
| 55 | 200 lines included for |
| 56 | the sake of efficiency |
| 57 | (they could have been written in C) |
| 58 | and 800 lines to perform hardware |
| 59 | functions not possible in C. |
| 60 | .PP |
| 61 | This code represents 5 to 10 percent of what has |
| 62 | been lumped into the broad expression |
| 63 | ``the |
| 64 | .UX |
| 65 | operating system.'' |
| 66 | The kernel is the only |
| 67 | .UX |
| 68 | code that |
| 69 | cannot be substituted by a user to his |
| 70 | own liking. |
| 71 | For this reason, |
| 72 | the kernel should make as few real |
| 73 | decisions as possible. |
| 74 | This does not mean to allow the user |
| 75 | a million options to do the same thing. |
| 76 | Rather, it means to allow only one way to |
| 77 | do one thing, |
| 78 | but have that way be the least-common divisor |
| 79 | of all the options that might have been provided. |
| 80 | .PP |
| 81 | What is or is not implemented in the kernel |
| 82 | represents both a great responsibility and a great power. |
| 83 | It is a soap-box platform on |
| 84 | ``the way things should be done.'' |
| 85 | Even so, if |
| 86 | ``the way'' is too radical, |
| 87 | no one will follow it. |
| 88 | Every important decision was weighed |
| 89 | carefully. |
| 90 | Throughout, |
| 91 | simplicity has been substituted for efficiency. |
| 92 | Complex algorithms are used only if |
| 93 | their complexity can be localized. |
| 94 | .NH |
| 95 | PROCESS CONTROL |
| 96 | .PP |
| 97 | In the |
| 98 | .UX |
| 99 | system, |
| 100 | a user executes programs in an |
| 101 | environment called a user process. |
| 102 | When a system function is required, |
| 103 | the user process calls the system |
| 104 | as a subroutine. |
| 105 | At some point in this call, |
| 106 | there is a distinct switch of environments. |
| 107 | After this, |
| 108 | the process is said to be a system process. |
| 109 | In the normal definition of processes, |
| 110 | the user and system processes are different |
| 111 | phases of the same process |
| 112 | (they never execute simultaneously). |
| 113 | For protection, |
| 114 | each system process has its own stack. |
| 115 | .PP |
| 116 | The user process may execute |
| 117 | from a read-only text segment, |
| 118 | which is shared by all processes |
| 119 | executing the same code. |
| 120 | There is no |
| 121 | .IT functional |
| 122 | benefit |
| 123 | from shared-text segments. |
| 124 | An |
| 125 | .IT efficiency |
| 126 | benefit comes from the fact |
| 127 | that there is no need to swap read-only |
| 128 | segments out because the original |
| 129 | copy on secondary memory is still current. |
| 130 | This is a great benefit to interactive |
| 131 | programs that tend to be swapped while |
| 132 | waiting for terminal input. |
| 133 | Furthermore, |
| 134 | if two processes are |
| 135 | executing |
| 136 | simultaneously |
| 137 | from the same copy of a read-only segment, |
| 138 | only one copy needs to reside in |
| 139 | primary memory. |
| 140 | This is a secondary effect, |
| 141 | because |
| 142 | simultaneous execution of a program |
| 143 | is not common. |
| 144 | It is ironic that this effect, |
| 145 | which reduces the use of primary memory, |
| 146 | only comes into play when there is |
| 147 | an overabundance of primary memory, |
| 148 | that is, |
| 149 | when there is enough memory |
| 150 | to keep waiting processes loaded. |
| 151 | .PP |
| 152 | All current read-only text segments in the |
| 153 | system are maintained from the |
| 154 | .IT "text table" . |
| 155 | A text table entry holds the location of the |
| 156 | text segment on secondary memory. |
| 157 | If the segment is loaded, |
| 158 | that table also holds the primary memory location |
| 159 | and the count of the number of processes |
| 160 | sharing this entry. |
| 161 | When this count is reduced to zero, |
| 162 | the entry is freed along with any |
| 163 | primary and secondary memory holding the segment. |
| 164 | When a process first executes a shared-text segment, |
| 165 | a text table entry is allocated and the |
| 166 | segment is loaded onto secondary memory. |
| 167 | If a second process executes a text segment |
| 168 | that is already allocated, |
| 169 | the entry reference count is simply incremented. |
| 170 | .PP |
| 171 | A user process has some strictly private |
| 172 | read-write data |
| 173 | contained in its |
| 174 | data segment. |
| 175 | As far as possible, |
| 176 | the system does not use the user's |
| 177 | data segment to hold system data. |
| 178 | In particular, |
| 179 | there are no I/O buffers in the |
| 180 | user address space. |
| 181 | .PP |
| 182 | The user data segment has two growing boundaries. |
| 183 | One, increased automatically by the system |
| 184 | as a result of memory faults, |
| 185 | is used for a stack. |
| 186 | The second boundary is only grown (or shrunk) by |
| 187 | explicit requests. |
| 188 | The contents of newly allocated primary memory |
| 189 | is initialized to zero. |
| 190 | .PP |
| 191 | Also associated and swapped with |
| 192 | a process is a small fixed-size |
| 193 | system data segment. |
| 194 | This segment contains all |
| 195 | the data about the process |
| 196 | that the system needs only when the |
| 197 | process is active. |
| 198 | Examples of the kind of data contained |
| 199 | in the system data segment are: |
| 200 | saved central processor registers, |
| 201 | open file descriptors, |
| 202 | accounting information, |
| 203 | scratch data area, |
| 204 | and the stack for the system phase |
| 205 | of the process. |
| 206 | The system data segment is not |
| 207 | addressable from the user process |
| 208 | and is therefore protected. |
| 209 | .PP |
| 210 | Last, |
| 211 | there is a process table with |
| 212 | one entry per process. |
| 213 | This entry contains all the data |
| 214 | needed by the system when the process |
| 215 | is |
| 216 | .IT not |
| 217 | active. |
| 218 | Examples are |
| 219 | the process's name, |
| 220 | the location of the other segments, |
| 221 | and scheduling information. |
| 222 | The process table entry is allocated |
| 223 | when the process is created, and freed |
| 224 | when the process terminates. |
| 225 | This process entry is always directly |
| 226 | addressable by the kernel. |
| 227 | .PP |
| 228 | Figure 1 shows the relationships |
| 229 | between the various process control |
| 230 | data. |
| 231 | In a sense, |
| 232 | the process table is the |
| 233 | definition of all processes, |
| 234 | because |
| 235 | all the data associated with a process |
| 236 | may be accessed |
| 237 | starting from the process table entry. |
| 238 | .KS |
| 239 | .sp 2.44i |
| 240 | .sp 2v |
| 241 | .ce |
| 242 | Fig. 1\(emProcess control data structure. |
| 243 | .KE |
| 244 | .NH 2 |
| 245 | Process creation and program execution |
| 246 | .PP |
| 247 | Processes are created by the system primitive |
| 248 | .UL fork . |
| 249 | The newly created process (child) is a copy of the original process (parent). |
| 250 | There is no detectable sharing of primary memory between the two processes. |
| 251 | (Of course, |
| 252 | if the parent process was executing from a read-only |
| 253 | text segment, |
| 254 | the child will share the text segment.) |
| 255 | Copies of all writable data segments |
| 256 | are made for the child process. |
| 257 | Files that were open before the |
| 258 | .UL fork |
| 259 | are |
| 260 | truly shared after the |
| 261 | .UL fork . |
| 262 | The processes are informed as to their part in the |
| 263 | relationship to |
| 264 | allow them to select their own |
| 265 | (usually non-identical) |
| 266 | destiny. |
| 267 | The parent may |
| 268 | .UL wait |
| 269 | for the termination of |
| 270 | any of its children. |
| 271 | .PP |
| 272 | A process may |
| 273 | .UL exec |
| 274 | a file. |
| 275 | This consists of exchanging the current text and data |
| 276 | segments of the process for new text and data |
| 277 | segments specified in the file. |
| 278 | The old segments are lost. |
| 279 | Doing an |
| 280 | .UL exec |
| 281 | does |
| 282 | .IT not |
| 283 | change processes; |
| 284 | the process that did the |
| 285 | .UL exec |
| 286 | persists, |
| 287 | but |
| 288 | after the |
| 289 | .UL exec |
| 290 | it is executing a different program. |
| 291 | Files that were open |
| 292 | before the |
| 293 | .UL exec |
| 294 | remain open after the |
| 295 | .UL exec . |
| 296 | .PP |
| 297 | If a program, |
| 298 | say the first pass of a compiler, |
| 299 | wishes to overlay itself with another program, |
| 300 | say the second pass, |
| 301 | then it simply |
| 302 | .UL exec s |
| 303 | the second program. |
| 304 | This is analogous |
| 305 | to a ``goto.'' |
| 306 | If a program wishes to regain control |
| 307 | after |
| 308 | .UL exec ing |
| 309 | a second program, |
| 310 | it should |
| 311 | .UL fork |
| 312 | a child process, |
| 313 | have the child |
| 314 | .UL exec |
| 315 | the second program, and |
| 316 | have the parent |
| 317 | .UL wait |
| 318 | for the child. |
| 319 | This is analogous to a ``call.'' |
| 320 | Breaking up the call into a binding followed by |
| 321 | a transfer is similar to the subroutine linkage in |
| 322 | SL-5. |
| 323 | .[ |
| 324 | griswold hanson sl5 overview |
| 325 | .] |
| 326 | .NH 2 |
| 327 | Swapping |
| 328 | .PP |
| 329 | The major data associated with a process |
| 330 | (the user data segment, |
| 331 | the system data segment, and |
| 332 | the text segment) |
| 333 | are swapped to and from secondary |
| 334 | memory, as needed. |
| 335 | The user data segment and the system data segment |
| 336 | are kept in contiguous primary memory to reduce |
| 337 | swapping latency. |
| 338 | (When low-latency devices, such as bubbles, |
| 339 | .UC CCD s, |
| 340 | or scatter/gather devices, |
| 341 | are used, |
| 342 | this decision will have to be reconsidered.) |
| 343 | Allocation of both primary |
| 344 | and secondary memory is performed |
| 345 | by the same simple first-fit algorithm. |
| 346 | When a process grows, |
| 347 | a new piece of primary memory is allocated. |
| 348 | The contents of the old memory is copied to the new memory. |
| 349 | The old memory is freed |
| 350 | and the tables are updated. |
| 351 | If there is not enough primary memory, |
| 352 | secondary memory is allocated instead. |
| 353 | The process is swapped out onto the |
| 354 | secondary memory, |
| 355 | ready to be swapped in with |
| 356 | its new size. |
| 357 | .PP |
| 358 | One separate process in the kernel, |
| 359 | the swapping process, |
| 360 | simply swaps the other |
| 361 | processes in and out of primary memory. |
| 362 | It examines the |
| 363 | process table looking for a process |
| 364 | that is swapped out and is |
| 365 | ready to run. |
| 366 | It allocates primary memory for that |
| 367 | process and |
| 368 | reads its segments into |
| 369 | primary memory, where that process competes for the |
| 370 | central processor with other loaded processes. |
| 371 | If no primary memory is available, |
| 372 | the swapping process makes memory available |
| 373 | by examining the process table for processes |
| 374 | that can be swapped out. |
| 375 | It selects a process to swap out, |
| 376 | writes it to secondary memory, |
| 377 | frees the primary memory, |
| 378 | and then goes back to look for a process |
| 379 | to swap in. |
| 380 | .PP |
| 381 | Thus there are two specific algorithms |
| 382 | to the swapping process. |
| 383 | Which of the possibly many processes that |
| 384 | are swapped out is to be swapped in? |
| 385 | This is decided by secondary storage residence |
| 386 | time. |
| 387 | The one with the longest time out is swapped in first. |
| 388 | There is a slight penalty for larger processes. |
| 389 | Which of the possibly many processes that |
| 390 | are loaded is to be swapped out? |
| 391 | Processes that are waiting for slow events |
| 392 | (i.e., not currently running or waiting for |
| 393 | disk I/O) |
| 394 | are picked first, |
| 395 | by age in primary memory, |
| 396 | again with size penalties. |
| 397 | The other processes are examined |
| 398 | by the same age algorithm, |
| 399 | but are not taken out unless they are |
| 400 | at least of some age. |
| 401 | This adds |
| 402 | hysteresis to the swapping and |
| 403 | prevents total thrashing. |
| 404 | .PP |
| 405 | These swapping algorithms are the |
| 406 | most suspect in the system. |
| 407 | With limited primary memory, |
| 408 | these algorithms cause total swapping. |
| 409 | This is not bad in itself, because |
| 410 | the swapping does not impact the |
| 411 | execution of the resident processes. |
| 412 | However, if the swapping device must |
| 413 | also be used for file storage, |
| 414 | the swapping traffic severely |
| 415 | impacts the file system traffic. |
| 416 | It is exactly these small systems |
| 417 | that tend to double usage of limited disk |
| 418 | resources. |
| 419 | .NH 2 |
| 420 | Synchronization and scheduling |
| 421 | .PP |
| 422 | Process synchronization is accomplished by having processes |
| 423 | wait for events. |
| 424 | Events are represented by arbitrary integers. |
| 425 | By convention, |
| 426 | events are chosen to be addresses of |
| 427 | tables associated with those events. |
| 428 | For example, a process that is waiting for |
| 429 | any of its children to terminate will wait |
| 430 | for an event that is the address of |
| 431 | its own process table entry. |
| 432 | When a process terminates, |
| 433 | it signals the event represented by |
| 434 | its parent's process table entry. |
| 435 | Signaling an event on which no process |
| 436 | is waiting has no effect. |
| 437 | Similarly, |
| 438 | signaling an event on which many processes |
| 439 | are waiting will wake all of them up. |
| 440 | This differs considerably from |
| 441 | Dijkstra's P and V |
| 442 | synchronization operations, |
| 443 | .[ |
| 444 | dijkstra sequential processes 1968 |
| 445 | .] |
| 446 | in that |
| 447 | no memory is associated with events. |
| 448 | Thus there need be no allocation of events |
| 449 | prior to their use. |
| 450 | Events exist simply by being used. |
| 451 | .PP |
| 452 | On the negative side, |
| 453 | because there is no memory associated with events, |
| 454 | no notion of ``how much'' |
| 455 | can be signaled via the event mechanism. |
| 456 | For example, |
| 457 | processes that want memory might |
| 458 | wait on an event associated with |
| 459 | memory allocation. |
| 460 | When any amount of memory becomes available, |
| 461 | the event would be signaled. |
| 462 | All the competing processes would then wake |
| 463 | up to fight over the new memory. |
| 464 | (In reality, |
| 465 | the swapping process is the only process |
| 466 | that waits for primary memory to become available.) |
| 467 | .PP |
| 468 | If an event occurs |
| 469 | between the time a process decides |
| 470 | to wait for that event and the |
| 471 | time that process enters the wait state, |
| 472 | then |
| 473 | the process will wait on an event that has |
| 474 | already happened (and may never happen again). |
| 475 | This race condition happens because there is no memory associated with |
| 476 | the event to indicate that the event has occurred; |
| 477 | the only action of an event is to change a set of processes |
| 478 | from wait state to run state. |
| 479 | This problem is relieved largely |
| 480 | by the fact that process switching can |
| 481 | only occur in the kernel by explicit calls |
| 482 | to the event-wait mechanism. |
| 483 | If the event in question is signaled by another |
| 484 | process, |
| 485 | then there is no problem. |
| 486 | But if the event is signaled by a hardware |
| 487 | interrupt, |
| 488 | then special care must be taken. |
| 489 | These synchronization races pose the biggest |
| 490 | problem when |
| 491 | .UX |
| 492 | is adapted to multiple-processor configurations. |
| 493 | .[ |
| 494 | hawley meyer multiprocessing unix |
| 495 | .] |
| 496 | .PP |
| 497 | The event-wait code in the kernel |
| 498 | is like a co-routine linkage. |
| 499 | At any time, |
| 500 | all but one of the processes has called event-wait. |
| 501 | The remaining process is the one currently executing. |
| 502 | When it calls event-wait, |
| 503 | a process whose event has been signaled |
| 504 | is selected and that process |
| 505 | returns from its call to event-wait. |
| 506 | .PP |
| 507 | Which of the runable processes is to run next? |
| 508 | Associated with each process is a priority. |
| 509 | The priority of a system process is assigned by the code |
| 510 | issuing the wait on an event. |
| 511 | This is roughly equivalent to the response |
| 512 | that one would expect on such an event. |
| 513 | Disk events have high priority, |
| 514 | teletype events are low, |
| 515 | and time-of-day events are very low. |
| 516 | (From observation, |
| 517 | the difference in system process priorities |
| 518 | has little or no performance impact.) |
| 519 | All user-process priorities are lower than the |
| 520 | lowest system priority. |
| 521 | User-process priorities are assigned |
| 522 | by an algorithm based on the |
| 523 | recent ratio of the amount of compute time to real time consumed |
| 524 | by the process. |
| 525 | A process that has used a lot of |
| 526 | compute time in the last real-time |
| 527 | unit is assigned a low user priority. |
| 528 | Because interactive processes are characterized |
| 529 | by low ratios of compute to real time, |
| 530 | interactive response is maintained without any |
| 531 | special arrangements. |
| 532 | .PP |
| 533 | The scheduling algorithm simply picks |
| 534 | the process with the highest priority, |
| 535 | thus |
| 536 | picking all system processes first and |
| 537 | user processes second. |
| 538 | The compute-to-real-time ratio is updated |
| 539 | every second. |
| 540 | Thus, |
| 541 | all other things being equal, |
| 542 | looping user processes will be |
| 543 | scheduled round-robin with a |
| 544 | 1-second quantum. |
| 545 | A high-priority process waking up will |
| 546 | preempt a running, low-priority process. |
| 547 | The scheduling algorithm has a very desirable |
| 548 | negative feedback character. |
| 549 | If a process uses its high priority |
| 550 | to hog the computer, |
| 551 | its priority will drop. |
| 552 | At the same time, if a low-priority |
| 553 | process is ignored for a long time, |
| 554 | its priority will rise. |
| 555 | .NH |
| 556 | I/O SYSTEM |
| 557 | .PP |
| 558 | The I/O system |
| 559 | is broken into two completely separate systems: |
| 560 | the block I/O system and the character I/O system. |
| 561 | In retrospect, |
| 562 | the names should have been ``structured I/O'' |
| 563 | and ``unstructured I/O,'' respectively; |
| 564 | while the term ``block I/O'' has some meaning, |
| 565 | ``character I/O'' is a complete misnomer. |
| 566 | .PP |
| 567 | Devices are characterized by a major device number, |
| 568 | a minor device number, and |
| 569 | a class (block or character). |
| 570 | For each class, |
| 571 | there is an array of entry points into the device drivers. |
| 572 | The major device number is used to index the array |
| 573 | when calling the code for a particular device driver. |
| 574 | The minor device number is passed to the |
| 575 | device driver as an argument. |
| 576 | The minor number has no significance other |
| 577 | than that attributed to it by the driver. |
| 578 | Usually, |
| 579 | the driver uses the minor number to access |
| 580 | one of several identical physical devices. |
| 581 | .PP |
| 582 | The use of the array of entry points |
| 583 | (configuration table) |
| 584 | as the only connection between the |
| 585 | system code and the device drivers is |
| 586 | very important. |
| 587 | Early versions of the system had a much |
| 588 | less formal connection with the drivers, |
| 589 | so that it was extremely hard to handcraft |
| 590 | differently configured systems. |
| 591 | Now it is possible to create new |
| 592 | device drivers in an average of a few hours. |
| 593 | The configuration table in most cases |
| 594 | is created automatically by a program |
| 595 | that reads the system's parts list. |
| 596 | .NH 2 |
| 597 | Block I/O system |
| 598 | .PP |
| 599 | The model block I/O device consists |
| 600 | of randomly addressed, secondary |
| 601 | memory blocks of 512 bytes each. |
| 602 | The blocks are uniformly addressed |
| 603 | 0, 1, .\|.\|. up to the size of the device. |
| 604 | The block device driver has the job of |
| 605 | emulating this model on a |
| 606 | physical device. |
| 607 | .PP |
| 608 | The block I/O devices are accessed |
| 609 | through a layer of buffering software. |
| 610 | The system maintains a list of buffers |
| 611 | (typically between 10 and 70) |
| 612 | each assigned a device name and |
| 613 | a device address. |
| 614 | This buffer pool constitutes a data cache |
| 615 | for the block devices. |
| 616 | On a read request, |
| 617 | the cache is searched for the desired block. |
| 618 | If the block is found, |
| 619 | the data are made available to the |
| 620 | requester without any physical I/O. |
| 621 | If the block is not in the cache, |
| 622 | the least recently used block in the cache is renamed, |
| 623 | the correct device driver is called to |
| 624 | fill up the renamed buffer, and then the |
| 625 | data are made available. |
| 626 | Write requests are handled in an analogous manner. |
| 627 | The correct buffer is found |
| 628 | and relabeled if necessary. |
| 629 | The write is performed simply by marking |
| 630 | the buffer as ``dirty.'' |
| 631 | The physical I/O is then deferred until |
| 632 | the buffer is renamed. |
| 633 | .PP |
| 634 | The benefits in reduction of physical I/O |
| 635 | of this scheme are substantial, |
| 636 | especially considering the file system implementation. |
| 637 | There are, |
| 638 | however, |
| 639 | some drawbacks. |
| 640 | The asynchronous nature of the |
| 641 | algorithm makes error reporting |
| 642 | and meaningful user error handling |
| 643 | almost impossible. |
| 644 | The cavalier approach to I/O error |
| 645 | handling in the |
| 646 | .UX |
| 647 | system is partly due to the asynchronous |
| 648 | nature of the block I/O system. |
| 649 | A second problem is in the delayed writes. |
| 650 | If the system stops unexpectedly, |
| 651 | it is almost certain that there is a |
| 652 | lot of logically complete, |
| 653 | but physically incomplete, |
| 654 | I/O in the buffers. |
| 655 | There is a system primitive to |
| 656 | flush all outstanding I/O activity |
| 657 | from the buffers. |
| 658 | Periodic use of this primitive helps, |
| 659 | but does not solve, the problem. |
| 660 | Finally, |
| 661 | the associativity in the buffers |
| 662 | can alter the physical I/O sequence |
| 663 | from that of the logical I/O sequence. |
| 664 | This means that there are times |
| 665 | when data structures on disk are inconsistent, |
| 666 | even though the software is careful |
| 667 | to perform I/O in the correct order. |
| 668 | On non-random devices, |
| 669 | notably magnetic tape, |
| 670 | the inversions of writes can be disastrous. |
| 671 | The problem with magnetic tapes is ``cured'' by |
| 672 | allowing only one outstanding write request |
| 673 | per drive. |
| 674 | .NH 2 |
| 675 | Character I/O system |
| 676 | .PP |
| 677 | The character I/O system consists of all |
| 678 | devices that do not fall into the block I/O model. |
| 679 | This includes the ``classical'' character devices |
| 680 | such as communications lines, paper tape, and |
| 681 | line printers. |
| 682 | It also includes magnetic tape and disks when |
| 683 | they are not used in a stereotyped way, |
| 684 | for example, 80-byte physical records on tape |
| 685 | and track-at-a-time disk copies. |
| 686 | In short, |
| 687 | the character I/O interface |
| 688 | means ``everything other than block.'' |
| 689 | I/O requests from the user are sent to the |
| 690 | device driver essentially unaltered. |
| 691 | The implementation of these requests is, of course, |
| 692 | up to the device driver. |
| 693 | There are guidelines and conventions |
| 694 | to help the implementation of |
| 695 | certain types of device drivers. |
| 696 | .NH 3 |
| 697 | Disk drivers |
| 698 | .PP |
| 699 | Disk drivers are implemented |
| 700 | with a queue of transaction records. |
| 701 | Each record holds a read/write flag, |
| 702 | a primary memory address, |
| 703 | a secondary memory address, and |
| 704 | a transfer byte count. |
| 705 | Swapping is accomplished by passing |
| 706 | such a record to the swapping device driver. |
| 707 | The block I/O interface is implemented by |
| 708 | passing such records with requests to |
| 709 | fill and empty system buffers. |
| 710 | The character I/O interface to the disk |
| 711 | drivers create a transaction record that |
| 712 | points directly into the user area. |
| 713 | The routine that creates this record also insures |
| 714 | that the user is not swapped during this |
| 715 | I/O transaction. |
| 716 | Thus by implementing the general disk driver, |
| 717 | it is possible to use the disk |
| 718 | as a block device, |
| 719 | a character device, and a swap device. |
| 720 | The only really disk-specific code in normal |
| 721 | disk drivers is the pre-sort of transactions to |
| 722 | minimize latency for a particular device, and |
| 723 | the actual issuing of the I/O request. |
| 724 | .NH 3 |
| 725 | Character lists |
| 726 | .PP |
| 727 | Real character-oriented devices may |
| 728 | be implemented using the common |
| 729 | code to handle character lists. |
| 730 | A character list is a queue of characters. |
| 731 | One routine puts a character on a queue. |
| 732 | Another gets a character from a queue. |
| 733 | It is also possible to ask how many |
| 734 | characters are currently on a queue. |
| 735 | Storage for all queues in the system comes |
| 736 | from a single common pool. |
| 737 | Putting a character on a queue will allocate |
| 738 | space from the common pool and link the |
| 739 | character onto the data structure defining the queue. |
| 740 | Getting a character from a queue returns |
| 741 | the corresponding space to the pool. |
| 742 | .PP |
| 743 | A typical character-output device |
| 744 | (paper tape punch, for example) |
| 745 | is implemented by passing characters |
| 746 | from the user onto a character queue until |
| 747 | some maximum number of characters is on the queue. |
| 748 | The I/O is prodded to start as |
| 749 | soon as there is anything on the queue |
| 750 | and, once started, |
| 751 | it is sustained by hardware completion interrupts. |
| 752 | Each time there is a completion interrupt, |
| 753 | the driver gets the next character from the queue |
| 754 | and sends it to the hardware. |
| 755 | The number of characters on the queue is checked and, |
| 756 | as the count falls through some intermediate level, |
| 757 | an event (the queue address) is signaled. |
| 758 | The process that is passing characters from |
| 759 | the user to the queue can be waiting on the event, and |
| 760 | refill the queue to its maximum |
| 761 | when the event occurs. |
| 762 | .PP |
| 763 | A typical character input device |
| 764 | (for example, a paper tape reader) |
| 765 | is handled in a very similar manner. |
| 766 | .PP |
| 767 | Another class of character devices is the terminals. |
| 768 | A terminal is represented by three |
| 769 | character queues. |
| 770 | There are two input queues (raw and canonical) |
| 771 | and an output queue. |
| 772 | Characters going to the output of a terminal |
| 773 | are handled by common code exactly as described |
| 774 | above. |
| 775 | The main difference is that there is also code |
| 776 | to interpret the output stream as |
| 777 | .UC ASCII |
| 778 | characters and to perform some translations, |
| 779 | e.g., escapes for deficient terminals. |
| 780 | Another common aspect of terminals is code |
| 781 | to insert real-time delay after certain control characters. |
| 782 | .PP |
| 783 | Input on terminals is a little different. |
| 784 | Characters are collected from the terminal and |
| 785 | placed on a raw input queue. |
| 786 | Some device-dependent code conversion and |
| 787 | escape interpretation is handled here. |
| 788 | When a line is complete in the raw queue, |
| 789 | an event is signaled. |
| 790 | The code catching this signal then copies a |
| 791 | line from the raw queue to a canonical queue |
| 792 | performing the character erase and line kill editing. |
| 793 | User read requests on terminals can be |
| 794 | directed at either the raw or canonical queues. |
| 795 | .NH 3 |
| 796 | Other character devices |
| 797 | .PP |
| 798 | Finally, |
| 799 | there are devices that fit no general category. |
| 800 | These devices are set up as character I/O drivers. |
| 801 | An example is a driver that reads and writes |
| 802 | unmapped primary memory as an I/O device. |
| 803 | Some devices are too |
| 804 | fast to be treated a character at time, |
| 805 | but do not fit the disk I/O mold. |
| 806 | Examples are fast communications lines and |
| 807 | fast line printers. |
| 808 | These devices either have their own buffers |
| 809 | or ``borrow'' block I/O buffers for a while and |
| 810 | then give them back. |
| 811 | .NH |
| 812 | THE FILE SYSTEM |
| 813 | .PP |
| 814 | In the |
| 815 | .UX |
| 816 | system, |
| 817 | a file is a (one-dimensional) array of bytes. |
| 818 | No other structure of files is implied by the |
| 819 | system. |
| 820 | Files are attached anywhere |
| 821 | (and possibly multiply) |
| 822 | onto a hierarchy of directories. |
| 823 | Directories are simply files that |
| 824 | users cannot write. |
| 825 | For a further discussion |
| 826 | of the external view of files and directories, |
| 827 | see Ref.\0 |
| 828 | .[ |
| 829 | ritchie thompson unix bstj 1978 |
| 830 | %Q This issue |
| 831 | .]. |
| 832 | .PP |
| 833 | The |
| 834 | .UX |
| 835 | file system is a disk data structure |
| 836 | accessed completely through |
| 837 | the block I/O system. |
| 838 | As stated before, |
| 839 | the canonical view of a ``disk'' is |
| 840 | a randomly addressable array of |
| 841 | 512-byte blocks. |
| 842 | A file system breaks the disk into |
| 843 | four self-identifying regions. |
| 844 | The first block (address 0) |
| 845 | is unused by the file system. |
| 846 | It is left aside for booting procedures. |
| 847 | The second block (address 1) |
| 848 | contains the so-called ``super-block.'' |
| 849 | This block, |
| 850 | among other things, |
| 851 | contains the size of the disk and |
| 852 | the boundaries of the other regions. |
| 853 | Next comes the i-list, |
| 854 | a list of file definitions. |
| 855 | Each file definition is |
| 856 | a 64-byte structure, called an i-node. |
| 857 | The offset of a particular i-node |
| 858 | within the i-list is called its i-number. |
| 859 | The combination of device name |
| 860 | (major and minor numbers) and i-number |
| 861 | serves to uniquely name a particular file. |
| 862 | After the i-list, |
| 863 | and to the end of the disk, |
| 864 | come free storage blocks that |
| 865 | are available for the contents of files. |
| 866 | .PP |
| 867 | The free space on a disk is maintained |
| 868 | by a linked list of available disk blocks. |
| 869 | Every block in this chain contains a disk address |
| 870 | of the next block in the chain. |
| 871 | The remaining space contains the address of up to |
| 872 | 50 disk blocks that are also free. |
| 873 | Thus with one I/O operation, |
| 874 | the system obtains 50 free blocks and a |
| 875 | pointer where to find more. |
| 876 | The disk allocation algorithms are |
| 877 | very straightforward. |
| 878 | Since all allocation is in fixed-size |
| 879 | blocks and there is strict accounting of |
| 880 | space, |
| 881 | there is no need to compact or garbage collect. |
| 882 | However, |
| 883 | as disk space becomes dispersed, |
| 884 | latency gradually increases. |
| 885 | Some installations choose to occasionally compact |
| 886 | disk space to reduce latency. |
| 887 | .PP |
| 888 | An i-node contains 13 disk addresses. |
| 889 | The first 10 of these addresses point directly at |
| 890 | the first 10 blocks of a file. |
| 891 | If a file is larger than 10 blocks (5,120 bytes), |
| 892 | then the eleventh address points at a block |
| 893 | that contains the addresses of the next 128 blocks of the file. |
| 894 | If the file is still larger than this |
| 895 | (70,656 bytes), |
| 896 | then the twelfth block points at up to 128 blocks, |
| 897 | each pointing to 128 blocks of the file. |
| 898 | Files yet larger |
| 899 | (8,459,264 bytes) |
| 900 | use the thirteenth address for a ``triple indirect'' address. |
| 901 | The algorithm ends here with the maximum file size |
| 902 | of 1,082,201,087 bytes. |
| 903 | .PP |
| 904 | A logical directory hierarchy is added |
| 905 | to this flat physical structure simply |
| 906 | by adding a new type of file, the directory. |
| 907 | A directory is accessed exactly as an ordinary file. |
| 908 | It contains 16-byte entries consisting of |
| 909 | a 14-byte name and an i-number. |
| 910 | The root of the hierarchy is at a known i-number |
| 911 | (\fIviz.,\fR 2). |
| 912 | The file system structure allows an arbitrary, directed graph |
| 913 | of directories with regular files linked in |
| 914 | at arbitrary places in this graph. |
| 915 | In fact, |
| 916 | very early |
| 917 | .UX |
| 918 | systems used such a structure. |
| 919 | Administration of such a structure became so |
| 920 | chaotic that later systems were restricted |
| 921 | to a directory tree. |
| 922 | Even now, |
| 923 | with regular files linked multiply |
| 924 | into arbitrary places in the tree, |
| 925 | accounting for space has become a problem. |
| 926 | It may become necessary to restrict the entire |
| 927 | structure to a tree, |
| 928 | and allow a new form of linking that |
| 929 | is subservient to the tree structure. |
| 930 | .PP |
| 931 | The file system allows |
| 932 | easy creation, |
| 933 | easy removal, |
| 934 | easy random accessing, |
| 935 | and very easy space allocation. |
| 936 | With most physical addresses confined |
| 937 | to a small contiguous section of disk, |
| 938 | it is also easy to dump, restore, and |
| 939 | check the consistency of the file system. |
| 940 | Large files suffer from indirect addressing, |
| 941 | but the cache prevents most of the implied physical I/O |
| 942 | without adding much execution. |
| 943 | The space overhead properties of this scheme are quite good. |
| 944 | For example, |
| 945 | on one particular file system, |
| 946 | there are 25,000 files containing 130M bytes of data-file content. |
| 947 | The overhead (i-node, indirect blocks, and last block breakage) |
| 948 | is about 11.5M bytes. |
| 949 | The directory structure to support these files |
| 950 | has about 1,500 directories containing 0.6M bytes of directory content |
| 951 | and about 0.5M bytes of overhead in accessing the directories. |
| 952 | Added up any way, |
| 953 | this comes out to less than a 10 percent overhead for actual |
| 954 | stored data. |
| 955 | Most systems have this much overhead in |
| 956 | padded trailing blanks alone. |
| 957 | .NH 2 |
| 958 | File system implementation |
| 959 | .PP |
| 960 | Because the i-node defines a file, |
| 961 | the implementation of the file system centers |
| 962 | around access to the i-node. |
| 963 | The system maintains a table of all active |
| 964 | i-nodes. |
| 965 | As a new file is accessed, |
| 966 | the system locates the corresponding i-node, |
| 967 | allocates an i-node table entry, and reads |
| 968 | the i-node into primary memory. |
| 969 | As in the buffer cache, |
| 970 | the table entry is considered to be the current |
| 971 | version of the i-node. |
| 972 | Modifications to the i-node are made to |
| 973 | the table entry. |
| 974 | When the last access to the i-node goes |
| 975 | away, |
| 976 | the table entry is copied back to the |
| 977 | secondary store i-list and the table entry is freed. |
| 978 | .PP |
| 979 | All I/O operations on files are carried out |
| 980 | with the aid of the corresponding i-node table entry. |
| 981 | The accessing of a file is a straightforward |
| 982 | implementation of the algorithms mentioned previously. |
| 983 | The user is not aware of i-nodes and i-numbers. |
| 984 | References to the file system are made in terms of |
| 985 | path names of the directory tree. |
| 986 | Converting a path name into an i-node table entry |
| 987 | is also straightforward. |
| 988 | Starting at some known i-node |
| 989 | (the root or the current directory of some process), |
| 990 | the next component of the path name is |
| 991 | searched by reading the directory. |
| 992 | This gives an i-number and an implied device |
| 993 | (that of the directory). |
| 994 | Thus the next i-node table entry can be accessed. |
| 995 | If that was the last component of the path name, |
| 996 | then this i-node is the result. |
| 997 | If not, |
| 998 | this i-node is the directory needed to look up |
| 999 | the next component of the path name, and the |
| 1000 | algorithm is repeated. |
| 1001 | .PP |
| 1002 | The user process accesses the file system with |
| 1003 | certain primitives. |
| 1004 | The most common of these are |
| 1005 | .UL open , |
| 1006 | .UL create , |
| 1007 | .UL read , |
| 1008 | .UL write , |
| 1009 | .UL seek , |
| 1010 | and |
| 1011 | .UL close . |
| 1012 | The data structures maintained are shown in Fig. 2. |
| 1013 | .KS |
| 1014 | .sp 22P |
| 1015 | .ce |
| 1016 | Fig. 2\(emFile system data structure. |
| 1017 | .KE |
| 1018 | In the system data segment associated with a user, |
| 1019 | there is room for some (usually between 10 and 50) open files. |
| 1020 | This open file table consists of pointers that can be used to access |
| 1021 | corresponding i-node table entries. |
| 1022 | Associated with each of these open files is |
| 1023 | a current I/O pointer. |
| 1024 | This is a byte offset of |
| 1025 | the next read/write operation on the file. |
| 1026 | The system treats each read/write request |
| 1027 | as random with an implied seek to the |
| 1028 | I/O pointer. |
| 1029 | The user usually thinks of the file as |
| 1030 | sequential with the I/O pointer |
| 1031 | automatically counting the number of bytes |
| 1032 | that have been read/written from the file. |
| 1033 | The user may, |
| 1034 | of course, |
| 1035 | perform random I/O by setting the I/O pointer |
| 1036 | before reads/writes. |
| 1037 | .PP |
| 1038 | With file sharing, |
| 1039 | it is necessary to allow related |
| 1040 | processes to share a common I/O pointer |
| 1041 | and yet have separate I/O pointers |
| 1042 | for independent processes |
| 1043 | that access the same file. |
| 1044 | With these two conditions, |
| 1045 | the I/O pointer cannot reside |
| 1046 | in the i-node table nor can |
| 1047 | it reside in the list of |
| 1048 | open files for the process. |
| 1049 | A new table |
| 1050 | (the open file table) |
| 1051 | was invented for the sole purpose |
| 1052 | of holding the I/O pointer. |
| 1053 | Processes that share the same open |
| 1054 | file |
| 1055 | (the result of |
| 1056 | .UL fork s) |
| 1057 | share a common open file table entry. |
| 1058 | A separate open of the same file will |
| 1059 | only share the i-node table entry, |
| 1060 | but will have distinct open file table entries. |
| 1061 | .PP |
| 1062 | The main file system primitives are implemented as follows. |
| 1063 | .UL \&open |
| 1064 | converts a file system path name into an i-node |
| 1065 | table entry. |
| 1066 | A pointer to the i-node table entry is placed in a |
| 1067 | newly created open file table entry. |
| 1068 | A pointer to the file table entry is placed in the |
| 1069 | system data segment for the process. |
| 1070 | .UL \&create |
| 1071 | first creates a new i-node entry, |
| 1072 | writes the i-number into a directory, and |
| 1073 | then builds the same structure as for an |
| 1074 | .UL open . |
| 1075 | .UL \&read |
| 1076 | and |
| 1077 | .UL write |
| 1078 | just access the i-node entry as described above. |
| 1079 | .UL \&seek |
| 1080 | simply manipulates the I/O pointer. |
| 1081 | No physical seeking is done. |
| 1082 | .UL \&close |
| 1083 | just frees the structures built by |
| 1084 | .UL open |
| 1085 | and |
| 1086 | .UL create . |
| 1087 | Reference counts are kept on the open file table entries and |
| 1088 | the i-node table entries to free these structures after |
| 1089 | the last reference goes away. |
| 1090 | .UL \&unlink |
| 1091 | simply decrements the count of the |
| 1092 | number of directories pointing at the given i-node. |
| 1093 | When the last reference to an i-node table entry |
| 1094 | goes away, |
| 1095 | if the i-node has no directories pointing to it, |
| 1096 | then the file is removed and the i-node is freed. |
| 1097 | This delayed removal of files prevents |
| 1098 | problems arising from removing active files. |
| 1099 | A file may be removed while still open. |
| 1100 | The resulting unnamed file vanishes |
| 1101 | when the file is closed. |
| 1102 | This is a method of obtaining temporary files. |
| 1103 | .PP |
| 1104 | There is a type of unnamed |
| 1105 | .UC FIFO |
| 1106 | file called a |
| 1107 | .UL pipe. |
| 1108 | Implementation of |
| 1109 | .UL pipe s |
| 1110 | consists of implied |
| 1111 | .UL seek s |
| 1112 | before each |
| 1113 | .UL read |
| 1114 | or |
| 1115 | .UL write |
| 1116 | in order to implement |
| 1117 | first-in-first-out. |
| 1118 | There are also checks and synchronization |
| 1119 | to prevent the |
| 1120 | writer from grossly outproducing the |
| 1121 | reader and to prevent the reader from |
| 1122 | overtaking the writer. |
| 1123 | .NH 2 |
| 1124 | Mounted file systems |
| 1125 | .PP |
| 1126 | The file system of a |
| 1127 | .UX |
| 1128 | system |
| 1129 | starts with some designated block device |
| 1130 | formatted as described above to contain |
| 1131 | a hierarchy. |
| 1132 | The root of this structure is the root of |
| 1133 | the |
| 1134 | .UX |
| 1135 | file system. |
| 1136 | A second formatted block device may be |
| 1137 | mounted |
| 1138 | at any leaf of |
| 1139 | the current hierarchy. |
| 1140 | This logically extends the current hierarchy. |
| 1141 | The implementation of |
| 1142 | mounting |
| 1143 | is trivial. |
| 1144 | A mount table is maintained containing |
| 1145 | pairs of designated leaf i-nodes and |
| 1146 | block devices. |
| 1147 | When converting a path name into an i-node, |
| 1148 | a check is made to see if the new i-node is a |
| 1149 | designated leaf. |
| 1150 | If it is, |
| 1151 | the i-node of the root |
| 1152 | of the block device replaces it. |
| 1153 | .PP |
| 1154 | Allocation of space for a file is taken |
| 1155 | from the free pool on the device on which the |
| 1156 | file lives. |
| 1157 | Thus a file system consisting of many |
| 1158 | mounted devices does not have a common pool of |
| 1159 | free secondary storage space. |
| 1160 | This separation of space on different |
| 1161 | devices is necessary to allow easy |
| 1162 | unmounting |
| 1163 | of a device. |
| 1164 | .NH 2 |
| 1165 | Other system functions |
| 1166 | .PP |
| 1167 | There are some other things that the system |
| 1168 | does for the user\-a |
| 1169 | little accounting, |
| 1170 | a little tracing/debugging, |
| 1171 | and a little access protection. |
| 1172 | Most of these things are not very |
| 1173 | well developed |
| 1174 | because our use of the system in computing science research |
| 1175 | does not need them. |
| 1176 | There are some features that are missed in some |
| 1177 | applications, for example, better inter-process communication. |
| 1178 | .PP |
| 1179 | The |
| 1180 | .UX |
| 1181 | kernel is an I/O multiplexer more than |
| 1182 | a complete operating system. |
| 1183 | This is as it should be. |
| 1184 | Because of this outlook, |
| 1185 | many features are |
| 1186 | found in most |
| 1187 | other operating systems that are missing from the |
| 1188 | .UX |
| 1189 | kernel. |
| 1190 | For example, |
| 1191 | the |
| 1192 | .UX |
| 1193 | kernel does not support |
| 1194 | file access methods, |
| 1195 | file disposition, |
| 1196 | file formats, |
| 1197 | file maximum size, |
| 1198 | spooling, |
| 1199 | command language, |
| 1200 | logical records, |
| 1201 | physical records, |
| 1202 | assignment of logical file names, |
| 1203 | logical file names, |
| 1204 | more than one character set, |
| 1205 | an operator's console, |
| 1206 | an operator, |
| 1207 | log-in, |
| 1208 | or log-out. |
| 1209 | Many of these things are symptoms rather than features. |
| 1210 | Many of these things are implemented |
| 1211 | in user software |
| 1212 | using the kernel as a tool. |
| 1213 | A good example of this is the command language. |
| 1214 | .[ |
| 1215 | bourne shell 1978 bstj |
| 1216 | %Q This issue |
| 1217 | .] |
| 1218 | Each user may have his own command language. |
| 1219 | Maintenance of such code is as easy as |
| 1220 | maintaining user code. |
| 1221 | The idea of implementing ``system'' code with general |
| 1222 | user primitives |
| 1223 | comes directly from |
| 1224 | .UC MULTICS . |
| 1225 | .[ |
| 1226 | organick multics 1972 |
| 1227 | .] |
| 1228 | .LP |
| 1229 | .[ |
| 1230 | $LIST$ |
| 1231 | .] |