Mike Dodd | 8cfa702 | 2010-11-17 11:12:26 -0800 | [diff] [blame] | 1 | /** |
| 2 | * @file oprofile.c |
| 3 | * Main driver code |
| 4 | * |
| 5 | * @remark Copyright 2002 OProfile authors |
| 6 | * @remark Read the file COPYING |
| 7 | * |
| 8 | * @author John Levon |
| 9 | * @author Philippe Elie |
| 10 | */ |
| 11 | |
| 12 | #include "oprofile.h" |
| 13 | #include "op_util.h" |
| 14 | #include "config.h" |
| 15 | |
| 16 | EXPORT_NO_SYMBOLS; |
| 17 | |
| 18 | MODULE_AUTHOR("John Levon (levon@movementarian.org)"); |
| 19 | MODULE_DESCRIPTION("Continuous Profiling Module"); |
| 20 | MODULE_LICENSE("GPL"); |
| 21 | |
| 22 | MODULE_PARM(allow_unload, "i"); |
| 23 | MODULE_PARM_DESC(allow_unload, "Allow module to be unloaded."); |
| 24 | #ifdef CONFIG_SMP |
| 25 | static int allow_unload; |
| 26 | #else |
| 27 | static int allow_unload = 1; |
| 28 | #endif |
| 29 | |
| 30 | /* sysctl settables */ |
| 31 | struct oprof_sysctl sysctl_parms; |
| 32 | /* some of the sys ctl settable variable needs to be copied to protect |
| 33 | * against user that try to change through /proc/sys/dev/oprofile/ * running |
| 34 | * parameters during profiling */ |
| 35 | struct oprof_sysctl sysctl; |
| 36 | |
| 37 | static enum oprof_state state __cacheline_aligned_in_smp = STOPPED; |
| 38 | |
| 39 | static int op_major; |
| 40 | |
| 41 | static volatile ulong oprof_opened __cacheline_aligned_in_smp; |
| 42 | static volatile ulong oprof_note_opened __cacheline_aligned_in_smp; |
| 43 | static DECLARE_WAIT_QUEUE_HEAD(oprof_wait); |
| 44 | |
| 45 | static u32 oprof_ready[NR_CPUS] __cacheline_aligned_in_smp; |
| 46 | struct _oprof_data oprof_data[NR_CPUS] __cacheline_aligned; |
| 47 | |
| 48 | struct op_note * note_buffer __cacheline_aligned_in_smp; |
| 49 | u32 note_pos __cacheline_aligned_in_smp; |
| 50 | |
| 51 | // the interrupt handler ops structure to use |
| 52 | static struct op_int_operations const * int_ops; |
| 53 | |
| 54 | static char const * op_version = PACKAGE " " VERSION; |
| 55 | |
| 56 | /* ---------------- interrupt entry routines ------------------ */ |
| 57 | |
| 58 | inline static int need_wakeup(uint cpu, struct _oprof_data * data) |
| 59 | { |
| 60 | return data->nextbuf >= (data->buf_size - data->buf_watermark) && !oprof_ready[cpu]; |
| 61 | } |
| 62 | |
| 63 | inline static void next_sample(struct _oprof_data * data) |
| 64 | { |
| 65 | if (unlikely(++data->nextbuf == data->buf_size)) |
| 66 | data->nextbuf = 0; |
| 67 | } |
| 68 | |
| 69 | inline static void evict_op_entry(uint cpu, struct _oprof_data * data, long irq_enabled) |
| 70 | { |
| 71 | next_sample(data); |
| 72 | if (likely(!need_wakeup(cpu, data))) |
| 73 | return; |
| 74 | |
| 75 | /* locking rationale : |
| 76 | * |
| 77 | * other CPUs are not a race concern since we synch on oprof_wait->lock. |
| 78 | * |
| 79 | * for the current CPU, we might have interrupted another user of e.g. |
| 80 | * runqueue_lock, deadlocking on SMP and racing on UP. So we check that IRQs |
| 81 | * were not disabled (corresponding to the irqsave/restores in __wake_up(). |
| 82 | * |
| 83 | * Note that this requires all spinlocks taken by the full wake_up path |
| 84 | * to have saved IRQs - otherwise we can interrupt whilst holding a spinlock |
| 85 | * taken from some non-wake_up() path and deadlock. Currently this means only |
| 86 | * oprof_wait->lock and runqueue_lock: all instances disable IRQs before |
| 87 | * taking the lock. |
| 88 | * |
| 89 | * This will mean that approaching the end of the buffer, a number of the |
| 90 | * evictions may fail to wake up the daemon. We simply hope this doesn't |
| 91 | * take long; a pathological case could cause buffer overflow. |
| 92 | * |
| 93 | * Note that we use oprof_ready as our flag for whether we have initiated a |
| 94 | * wake-up. Once the wake-up is received, the flag is reset as well as |
| 95 | * data->nextbuf, preventing multiple wakeups. |
| 96 | * |
| 97 | * On 2.2, a global waitqueue_lock is used, so we must check it's not held |
| 98 | * by the current CPU. We make sure that any users of the wait queue (i.e. |
| 99 | * us and the code for wait_event_interruptible()) disable interrupts so it's |
| 100 | * still safe to check IF_MASK. |
| 101 | */ |
| 102 | if (likely(irq_enabled)) { |
| 103 | oprof_ready[cpu] = 1; |
| 104 | wake_up(&oprof_wait); |
| 105 | } |
| 106 | } |
| 107 | |
| 108 | inline static void |
| 109 | fill_op_entry(struct op_sample * ops, long eip, pid_t pid, pid_t tgid, int ctr) |
| 110 | { |
| 111 | ops->eip = eip; |
| 112 | ops->pid = pid; |
| 113 | ops->tgid = tgid; |
| 114 | ops->counter = ctr; |
| 115 | } |
| 116 | |
| 117 | void op_do_profile(uint cpu, long eip, long irq_enabled, int ctr) |
| 118 | { |
| 119 | struct _oprof_data * data = &oprof_data[cpu]; |
| 120 | pid_t const pid = current->pid; |
| 121 | pid_t const tgid = op_get_tgid(); |
| 122 | struct op_sample * samples = &data->buffer[data->nextbuf]; |
| 123 | |
| 124 | data->nr_irq++; |
| 125 | |
| 126 | fill_op_entry(samples, eip, pid, tgid, ctr); |
| 127 | evict_op_entry(cpu, data, irq_enabled); |
| 128 | } |
| 129 | |
| 130 | /* ---------------- driver routines ------------------ */ |
| 131 | |
| 132 | /* only stop and start profiling interrupt when we are |
| 133 | * fully running ! |
| 134 | */ |
| 135 | static void stop_cpu_perfctr(int cpu) |
| 136 | { |
| 137 | if (state == RUNNING) |
| 138 | int_ops->stop_cpu(cpu); |
| 139 | } |
| 140 | |
| 141 | static void start_cpu_perfctr(int cpu) |
| 142 | { |
| 143 | if (state == RUNNING) |
| 144 | int_ops->start_cpu(cpu); |
| 145 | } |
| 146 | |
| 147 | spinlock_t note_lock __cacheline_aligned_in_smp = SPIN_LOCK_UNLOCKED; |
| 148 | /* which buffer nr. is waiting to be read ? */ |
| 149 | int cpu_buffer_waiting; |
| 150 | |
| 151 | static int is_ready(void) |
| 152 | { |
| 153 | uint cpu_nr; |
| 154 | for (cpu_nr = 0 ; cpu_nr < smp_num_cpus; cpu_nr++) { |
| 155 | if (oprof_ready[cpu_nr]) { |
| 156 | cpu_buffer_waiting = cpu_nr; |
| 157 | return 1; |
| 158 | } |
| 159 | } |
| 160 | return 0; |
| 161 | } |
| 162 | |
| 163 | inline static void up_and_check_note(void) |
| 164 | { |
| 165 | note_pos++; |
| 166 | if (likely(note_pos < (sysctl.note_size - OP_PRE_NOTE_WATERMARK(sysctl.note_size)) && !is_ready())) |
| 167 | return; |
| 168 | |
| 169 | /* if we reach the end of the buffer, just pin |
| 170 | * to the last entry until it is read. This loses |
| 171 | * notes, but we have no choice. */ |
| 172 | if (unlikely(note_pos == sysctl.note_size)) { |
| 173 | static int warned; |
| 174 | if (!warned) { |
| 175 | printk(KERN_WARNING "note buffer overflow: restart " |
| 176 | "oprofile with a larger note buffer.\n"); |
| 177 | warned = 1; |
| 178 | } |
| 179 | sysctl.nr_note_buffer_overflow++; |
| 180 | note_pos = sysctl.note_size - 1; |
| 181 | } |
| 182 | |
| 183 | /* we just use cpu 0 as a convenient one to wake up */ |
| 184 | oprof_ready[0] = 2; |
| 185 | oprof_wake_up(&oprof_wait); |
| 186 | } |
| 187 | |
| 188 | /* if holding note_lock */ |
| 189 | void __oprof_put_note(struct op_note * onote) |
| 190 | { |
| 191 | /* ignore note if we're not up and running fully */ |
| 192 | if (state != RUNNING) |
| 193 | return; |
| 194 | |
| 195 | memcpy(¬e_buffer[note_pos], onote, sizeof(struct op_note)); |
| 196 | up_and_check_note(); |
| 197 | } |
| 198 | |
| 199 | void oprof_put_note(struct op_note * onote) |
| 200 | { |
| 201 | spin_lock(¬e_lock); |
| 202 | __oprof_put_note(onote); |
| 203 | spin_unlock(¬e_lock); |
| 204 | } |
| 205 | |
| 206 | static ssize_t oprof_note_read(char * buf, size_t count, loff_t * ppos) |
| 207 | { |
| 208 | struct op_note * mybuf; |
| 209 | uint num; |
| 210 | ssize_t max; |
| 211 | |
| 212 | max = sizeof(struct op_note) * sysctl.note_size; |
| 213 | |
| 214 | if (*ppos || count != max) |
| 215 | return -EINVAL; |
| 216 | |
| 217 | mybuf = vmalloc(max); |
| 218 | if (!mybuf) |
| 219 | return -EFAULT; |
| 220 | |
| 221 | spin_lock(¬e_lock); |
| 222 | |
| 223 | num = note_pos; |
| 224 | |
| 225 | count = note_pos * sizeof(struct op_note); |
| 226 | |
| 227 | if (count) |
| 228 | memcpy(mybuf, note_buffer, count); |
| 229 | |
| 230 | note_pos = 0; |
| 231 | |
| 232 | spin_unlock(¬e_lock); |
| 233 | |
| 234 | if (count && copy_to_user(buf, mybuf, count)) |
| 235 | count = -EFAULT; |
| 236 | |
| 237 | vfree(mybuf); |
| 238 | return count; |
| 239 | } |
| 240 | |
| 241 | static int oprof_note_open(void) |
| 242 | { |
| 243 | if (test_and_set_bit(0, &oprof_note_opened)) |
| 244 | return -EBUSY; |
| 245 | INC_USE_COUNT_MAYBE; |
| 246 | return 0; |
| 247 | } |
| 248 | |
| 249 | static int oprof_note_release(void) |
| 250 | { |
| 251 | BUG_ON(!oprof_note_opened); |
| 252 | clear_bit(0, &oprof_note_opened); |
| 253 | DEC_USE_COUNT_MAYBE; |
| 254 | return 0; |
| 255 | } |
| 256 | |
| 257 | static int check_buffer_amount(int cpu_nr) |
| 258 | { |
| 259 | struct _oprof_data * data = &oprof_data[cpu_nr]; |
| 260 | int size = data->buf_size; |
| 261 | int num = data->nextbuf; |
| 262 | if (num < size - data->buf_watermark && oprof_ready[cpu_nr] != 2) { |
| 263 | printk(KERN_WARNING "oprofile: Detected overflow of size %d. " |
| 264 | "You must increase the module buffer size with\n" |
| 265 | "opcontrol --setup --bufer-size= or reduce the " |
| 266 | "interrupt frequency\n", num); |
| 267 | data->nr_buffer_overflow += num; |
| 268 | num = size; |
| 269 | } else |
| 270 | data->nextbuf = 0; |
| 271 | return num; |
| 272 | } |
| 273 | |
| 274 | static int copy_buffer(char * buf, int cpu_nr) |
| 275 | { |
| 276 | struct op_buffer_head head; |
| 277 | int ret = -EFAULT; |
| 278 | |
| 279 | stop_cpu_perfctr(cpu_nr); |
| 280 | |
| 281 | head.cpu_nr = cpu_nr; |
| 282 | head.count = check_buffer_amount(cpu_nr); |
| 283 | head.state = state; |
| 284 | |
| 285 | oprof_ready[cpu_nr] = 0; |
| 286 | |
| 287 | if (copy_to_user(buf, &head, sizeof(struct op_buffer_head))) |
| 288 | goto out; |
| 289 | |
| 290 | if (head.count) { |
| 291 | size_t const size = head.count * sizeof(struct op_sample); |
| 292 | if (copy_to_user(buf + sizeof(struct op_buffer_head), |
| 293 | oprof_data[cpu_nr].buffer, size)) |
| 294 | goto out; |
| 295 | ret = size + sizeof(struct op_buffer_head); |
| 296 | } else { |
| 297 | ret = sizeof(struct op_buffer_head); |
| 298 | } |
| 299 | |
| 300 | out: |
| 301 | start_cpu_perfctr(cpu_nr); |
| 302 | return ret; |
| 303 | } |
| 304 | |
| 305 | static ssize_t oprof_read(struct file * file, char * buf, size_t count, loff_t * ppos) |
| 306 | { |
| 307 | ssize_t max; |
| 308 | |
| 309 | if (!capable(CAP_SYS_PTRACE)) |
| 310 | return -EPERM; |
| 311 | |
| 312 | switch (MINOR(file->f_dentry->d_inode->i_rdev)) { |
| 313 | case 2: return oprof_note_read(buf, count, ppos); |
| 314 | case 0: break; |
| 315 | default: return -EINVAL; |
| 316 | } |
| 317 | |
| 318 | max = sizeof(struct op_buffer_head) + sizeof(struct op_sample) * sysctl.buf_size; |
| 319 | |
| 320 | if (*ppos || count != max) |
| 321 | return -EINVAL; |
| 322 | |
| 323 | switch (state) { |
| 324 | case RUNNING: |
| 325 | wait_event_interruptible(oprof_wait, is_ready()); |
| 326 | if (signal_pending(current)) |
| 327 | return -EINTR; |
| 328 | break; |
| 329 | |
| 330 | /* Non-obvious. If O_NONBLOCK is set, that means |
| 331 | * the daemon knows it has to quit and is asking |
| 332 | * for final buffer data. If it's not set, then we |
| 333 | * have just transitioned to STOPPING, and we must |
| 334 | * inform the daemon (which we can do just by a normal |
| 335 | * operation). |
| 336 | */ |
| 337 | case STOPPING: { |
| 338 | int cpu; |
| 339 | |
| 340 | if (!(file->f_flags & O_NONBLOCK)) |
| 341 | break; |
| 342 | |
| 343 | for (cpu = 0; cpu < smp_num_cpus; ++cpu) { |
| 344 | if (oprof_data[cpu].nextbuf) { |
| 345 | cpu_buffer_waiting = cpu; |
| 346 | oprof_ready[cpu] = 2; |
| 347 | break; |
| 348 | } |
| 349 | } |
| 350 | |
| 351 | if (cpu == smp_num_cpus) |
| 352 | return -EAGAIN; |
| 353 | |
| 354 | } |
| 355 | break; |
| 356 | |
| 357 | case STOPPED: BUG(); |
| 358 | } |
| 359 | |
| 360 | return copy_buffer(buf, cpu_buffer_waiting); |
| 361 | } |
| 362 | |
| 363 | |
| 364 | static int oprof_start(void); |
| 365 | static int oprof_stop(void); |
| 366 | |
| 367 | static int oprof_open(struct inode * ino, struct file * file) |
| 368 | { |
| 369 | int err; |
| 370 | |
| 371 | if (!capable(CAP_SYS_PTRACE)) |
| 372 | return -EPERM; |
| 373 | |
| 374 | switch (MINOR(file->f_dentry->d_inode->i_rdev)) { |
| 375 | case 1: return oprof_hash_map_open(); |
| 376 | case 2: return oprof_note_open(); |
| 377 | case 0: |
| 378 | /* make sure the other devices are open */ |
| 379 | if (is_map_ready()) |
| 380 | break; |
| 381 | default: |
| 382 | return -EINVAL; |
| 383 | } |
| 384 | |
| 385 | if (test_and_set_bit(0, &oprof_opened)) |
| 386 | return -EBUSY; |
| 387 | |
| 388 | err = oprof_start(); |
| 389 | if (err) |
| 390 | clear_bit(0, &oprof_opened); |
| 391 | return err; |
| 392 | } |
| 393 | |
| 394 | static int oprof_release(struct inode * ino, struct file * file) |
| 395 | { |
| 396 | switch (MINOR(file->f_dentry->d_inode->i_rdev)) { |
| 397 | case 1: return oprof_hash_map_release(); |
| 398 | case 2: return oprof_note_release(); |
| 399 | case 0: break; |
| 400 | default: return -EINVAL; |
| 401 | } |
| 402 | |
| 403 | BUG_ON(!oprof_opened); |
| 404 | |
| 405 | clear_bit(0, &oprof_opened); |
| 406 | |
| 407 | // FIXME: is this safe when I kill -9 the daemon ? |
| 408 | return oprof_stop(); |
| 409 | } |
| 410 | |
| 411 | static int oprof_mmap(struct file * file, struct vm_area_struct * vma) |
| 412 | { |
| 413 | if (MINOR(file->f_dentry->d_inode->i_rdev) == 1) |
| 414 | return oprof_hash_map_mmap(file, vma); |
| 415 | return -EINVAL; |
| 416 | } |
| 417 | |
| 418 | /* called under spinlock, cannot sleep */ |
| 419 | static void oprof_free_mem(uint num) |
| 420 | { |
| 421 | uint i; |
| 422 | for (i=0; i < num; i++) { |
| 423 | if (oprof_data[i].buffer) |
| 424 | vfree(oprof_data[i].buffer); |
| 425 | oprof_data[i].buffer = NULL; |
| 426 | } |
| 427 | vfree(note_buffer); |
| 428 | note_buffer = NULL; |
| 429 | } |
| 430 | |
| 431 | static int oprof_init_data(void) |
| 432 | { |
| 433 | uint i, notebufsize; |
| 434 | ulong buf_size; |
| 435 | struct _oprof_data * data; |
| 436 | |
| 437 | sysctl.nr_note_buffer_overflow = 0; |
| 438 | notebufsize = sizeof(struct op_note) * sysctl.note_size; |
| 439 | note_buffer = vmalloc(notebufsize); |
| 440 | if (!note_buffer) { |
| 441 | printk(KERN_ERR "oprofile: failed to allocate note buffer of %u bytes\n", |
| 442 | notebufsize); |
| 443 | return -EFAULT; |
| 444 | } |
| 445 | note_pos = 0; |
| 446 | |
| 447 | // safe init |
| 448 | for (i = 0; i < smp_num_cpus; ++i) { |
| 449 | data = &oprof_data[i]; |
| 450 | data->buf_size = 0; |
| 451 | data->buffer = 0; |
| 452 | data->buf_watermark = 0; |
| 453 | data->nr_buffer_overflow = 0; |
| 454 | } |
| 455 | |
| 456 | buf_size = (sizeof(struct op_sample) * sysctl.buf_size); |
| 457 | |
| 458 | for (i = 0 ; i < smp_num_cpus ; ++i) { |
| 459 | data = &oprof_data[i]; |
| 460 | |
| 461 | data->buffer = vmalloc(buf_size); |
| 462 | if (!data->buffer) { |
| 463 | printk(KERN_ERR "oprofile: failed to allocate eviction buffer of %lu bytes\n", buf_size); |
| 464 | oprof_free_mem(i); |
| 465 | return -EFAULT; |
| 466 | } |
| 467 | |
| 468 | memset(data->buffer, 0, buf_size); |
| 469 | |
| 470 | data->buf_size = sysctl.buf_size; |
| 471 | data->buf_watermark = OP_PRE_WATERMARK(data->buf_size); |
| 472 | data->nextbuf = 0; |
| 473 | } |
| 474 | |
| 475 | return 0; |
| 476 | } |
| 477 | |
| 478 | static int parms_check(void) |
| 479 | { |
| 480 | int err; |
| 481 | |
| 482 | if ((err = check_range(sysctl.buf_size, OP_MIN_BUF_SIZE, OP_MAX_BUF_SIZE, |
| 483 | "sysctl.buf_size value %d not in range (%d %d)\n"))) |
| 484 | return err; |
| 485 | if ((err = check_range(sysctl.note_size, OP_MIN_NOTE_TABLE_SIZE, OP_MAX_NOTE_TABLE_SIZE, |
| 486 | "sysctl.note_size value %d not in range (%d %d)\n"))) |
| 487 | return err; |
| 488 | |
| 489 | if ((err = int_ops->check_params())) |
| 490 | return err; |
| 491 | |
| 492 | return 0; |
| 493 | } |
| 494 | |
| 495 | |
| 496 | static DECLARE_MUTEX(sysctlsem); |
| 497 | |
| 498 | |
| 499 | static int oprof_start(void) |
| 500 | { |
| 501 | int err = 0; |
| 502 | |
| 503 | down(&sysctlsem); |
| 504 | |
| 505 | /* save the sysctl settable things to protect against change through |
| 506 | * systcl the profiler params */ |
| 507 | sysctl_parms.cpu_type = sysctl.cpu_type; |
| 508 | sysctl = sysctl_parms; |
| 509 | |
| 510 | if ((err = oprof_init_data())) |
| 511 | goto out; |
| 512 | |
| 513 | if ((err = parms_check())) { |
| 514 | oprof_free_mem(smp_num_cpus); |
| 515 | goto out; |
| 516 | } |
| 517 | |
| 518 | if ((err = int_ops->setup())) { |
| 519 | oprof_free_mem(smp_num_cpus); |
| 520 | goto out; |
| 521 | } |
| 522 | |
| 523 | op_intercept_syscalls(); |
| 524 | |
| 525 | int_ops->start(); |
| 526 | |
| 527 | state = RUNNING; |
| 528 | |
| 529 | out: |
| 530 | up(&sysctlsem); |
| 531 | return err; |
| 532 | } |
| 533 | |
| 534 | /* |
| 535 | * stop interrupts being generated and notes arriving. |
| 536 | * This is idempotent. |
| 537 | */ |
| 538 | static void oprof_partial_stop(void) |
| 539 | { |
| 540 | BUG_ON(state == STOPPED); |
| 541 | |
| 542 | if (state == RUNNING) { |
| 543 | op_restore_syscalls(); |
| 544 | int_ops->stop(); |
| 545 | } |
| 546 | |
| 547 | state = STOPPING; |
| 548 | } |
| 549 | |
| 550 | static int oprof_stop(void) |
| 551 | { |
| 552 | uint i; |
| 553 | // FIXME: err not needed |
| 554 | int err = -EINVAL; |
| 555 | |
| 556 | down(&sysctlsem); |
| 557 | |
| 558 | BUG_ON(state == STOPPED); |
| 559 | |
| 560 | /* here we need to : |
| 561 | * bring back the old system calls |
| 562 | * stop the perf counter |
| 563 | * bring back the old NMI handler |
| 564 | * reset the map buffer stuff and ready values |
| 565 | * |
| 566 | * Nothing will be able to write into the map buffer because |
| 567 | * we synchronise via the spinlocks |
| 568 | */ |
| 569 | |
| 570 | oprof_partial_stop(); |
| 571 | |
| 572 | spin_lock(¬e_lock); |
| 573 | |
| 574 | for (i = 0 ; i < smp_num_cpus; i++) { |
| 575 | struct _oprof_data * data = &oprof_data[i]; |
| 576 | oprof_ready[i] = 0; |
| 577 | data->nextbuf = 0; |
| 578 | } |
| 579 | |
| 580 | oprof_free_mem(smp_num_cpus); |
| 581 | |
| 582 | spin_unlock(¬e_lock); |
| 583 | err = 0; |
| 584 | |
| 585 | /* FIXME: can we really say this ? */ |
| 586 | state = STOPPED; |
| 587 | up(&sysctlsem); |
| 588 | return err; |
| 589 | } |
| 590 | |
| 591 | static struct file_operations oprof_fops = { |
| 592 | #ifdef HAVE_FILE_OPERATIONS_OWNER |
| 593 | owner: THIS_MODULE, |
| 594 | #endif |
| 595 | open: oprof_open, |
| 596 | release: oprof_release, |
| 597 | read: oprof_read, |
| 598 | mmap: oprof_mmap, |
| 599 | }; |
| 600 | |
| 601 | /* |
| 602 | * /proc/sys/dev/oprofile/ |
| 603 | * bufsize |
| 604 | * notesize |
| 605 | * dump |
| 606 | * dump_stop |
| 607 | * nr_interrupts |
| 608 | * #ctr/ |
| 609 | * event |
| 610 | * enabled |
| 611 | * count |
| 612 | * unit_mask |
| 613 | * kernel |
| 614 | * user |
| 615 | * |
| 616 | * #ctr is in [0-1] for PPro core, [0-3] for Athlon core |
| 617 | * |
| 618 | */ |
| 619 | |
| 620 | /* These access routines are basically not safe on SMP for module unload. |
| 621 | * And there is nothing we can do about it - the API is broken. We'll just |
| 622 | * make a best-efforts thing. Note the sem is needed to prevent parms_check |
| 623 | * bypassing during oprof_start(). |
| 624 | */ |
| 625 | |
| 626 | static void lock_sysctl(void) |
| 627 | { |
| 628 | MOD_INC_USE_COUNT; |
| 629 | down(&sysctlsem); |
| 630 | } |
| 631 | |
| 632 | static void unlock_sysctl(void) |
| 633 | { |
| 634 | up(&sysctlsem); |
| 635 | MOD_DEC_USE_COUNT; |
| 636 | } |
| 637 | |
| 638 | static int get_nr_interrupts(ctl_table * table, int write, struct file * filp, void * buffer, size_t * lenp) |
| 639 | { |
| 640 | uint cpu; |
| 641 | int ret = -EINVAL; |
| 642 | |
| 643 | lock_sysctl(); |
| 644 | |
| 645 | if (write) |
| 646 | goto out; |
| 647 | |
| 648 | sysctl.nr_interrupts = 0; |
| 649 | |
| 650 | for (cpu = 0 ; cpu < smp_num_cpus; cpu++) { |
| 651 | sysctl.nr_interrupts += oprof_data[cpu].nr_irq; |
| 652 | oprof_data[cpu].nr_irq = 0; |
| 653 | } |
| 654 | |
| 655 | ret = proc_dointvec(table, write, filp, buffer, lenp); |
| 656 | out: |
| 657 | unlock_sysctl(); |
| 658 | return ret; |
| 659 | } |
| 660 | |
| 661 | static int get_nr_buffer_overflow(ctl_table * table, int write, struct file * filp, void * buffer, size_t * lenp) |
| 662 | { |
| 663 | uint cpu; |
| 664 | int ret = -EINVAL; |
| 665 | |
| 666 | lock_sysctl(); |
| 667 | |
| 668 | if (write) |
| 669 | goto out; |
| 670 | |
| 671 | for (cpu = 0 ; cpu < smp_num_cpus; cpu++) { |
| 672 | sysctl.nr_buffer_overflow += oprof_data[cpu].nr_buffer_overflow; |
| 673 | oprof_data[cpu].nr_buffer_overflow = 0; |
| 674 | } |
| 675 | |
| 676 | ret = proc_dointvec(table, write, filp, buffer, lenp); |
| 677 | out: |
| 678 | unlock_sysctl(); |
| 679 | return ret; |
| 680 | } |
| 681 | |
| 682 | int lproc_dointvec(ctl_table * table, int write, struct file * filp, void * buffer, size_t * lenp) |
| 683 | { |
| 684 | int err; |
| 685 | |
| 686 | lock_sysctl(); |
| 687 | err = proc_dointvec(table, write, filp, buffer, lenp); |
| 688 | unlock_sysctl(); |
| 689 | |
| 690 | return err; |
| 691 | } |
| 692 | |
| 693 | static void do_actual_dump(void) |
| 694 | { |
| 695 | uint cpu; |
| 696 | |
| 697 | for (cpu = 0 ; cpu < smp_num_cpus; cpu++) |
| 698 | oprof_ready[cpu] = 2; |
| 699 | oprof_wake_up(&oprof_wait); |
| 700 | } |
| 701 | |
| 702 | static int sysctl_do_dump(ctl_table * table, int write, struct file * filp, void * buffer, size_t * lenp) |
| 703 | { |
| 704 | int err = -EINVAL; |
| 705 | |
| 706 | lock_sysctl(); |
| 707 | |
| 708 | if (state != RUNNING) |
| 709 | goto out; |
| 710 | |
| 711 | if (!write) { |
| 712 | err = proc_dointvec(table, write, filp, buffer, lenp); |
| 713 | goto out; |
| 714 | } |
| 715 | |
| 716 | do_actual_dump(); |
| 717 | |
| 718 | err = 0; |
| 719 | out: |
| 720 | unlock_sysctl(); |
| 721 | return err; |
| 722 | } |
| 723 | |
| 724 | static int sysctl_do_dump_stop(ctl_table * table, int write, struct file * filp, void * buffer, size_t * lenp) |
| 725 | { |
| 726 | int err = -EINVAL; |
| 727 | |
| 728 | lock_sysctl(); |
| 729 | |
| 730 | if (state != RUNNING) |
| 731 | goto out; |
| 732 | |
| 733 | if (!write) { |
| 734 | err = proc_dointvec(table, write, filp, buffer, lenp); |
| 735 | goto out; |
| 736 | } |
| 737 | |
| 738 | oprof_partial_stop(); |
| 739 | |
| 740 | /* also wakes up daemon */ |
| 741 | do_actual_dump(); |
| 742 | |
| 743 | err = 0; |
| 744 | out: |
| 745 | unlock_sysctl(); |
| 746 | return err; |
| 747 | } |
| 748 | |
| 749 | static int const nr_oprof_static = 8; |
| 750 | |
| 751 | static ctl_table oprof_table[] = { |
| 752 | { 1, "bufsize", &sysctl_parms.buf_size, sizeof(int), 0644, NULL, &lproc_dointvec, NULL, }, |
| 753 | { 1, "dump", &sysctl_parms.dump, sizeof(int), 0666, NULL, &sysctl_do_dump, NULL, }, |
| 754 | { 1, "dump_stop", &sysctl_parms.dump_stop, sizeof(int), 0644, NULL, &sysctl_do_dump_stop, NULL, }, |
| 755 | { 1, "nr_interrupts", &sysctl.nr_interrupts, sizeof(int), 0444, NULL, &get_nr_interrupts, NULL, }, |
| 756 | { 1, "notesize", &sysctl_parms.note_size, sizeof(int), 0644, NULL, &lproc_dointvec, NULL, }, |
| 757 | { 1, "cpu_type", &sysctl.cpu_type, sizeof(int), 0444, NULL, &lproc_dointvec, NULL, }, |
| 758 | { 1, "note_buffer_overflow", &sysctl.nr_note_buffer_overflow, sizeof(int), 0444, NULL, &lproc_dointvec, NULL, }, |
| 759 | { 1, "buffer_overflow", &sysctl.nr_buffer_overflow, sizeof(int), 0444, NULL, &get_nr_buffer_overflow, NULL, }, |
| 760 | { 0, }, { 0, }, { 0, }, { 0, }, { 0, }, { 0, }, { 0, }, { 0, }, |
| 761 | { 0, }, |
| 762 | }; |
| 763 | |
| 764 | static ctl_table oprof_root[] = { |
| 765 | {1, "oprofile", NULL, 0, 0755, oprof_table}, |
| 766 | {0, }, |
| 767 | }; |
| 768 | |
| 769 | static ctl_table dev_root[] = { |
| 770 | {CTL_DEV, "dev", NULL, 0, 0555, oprof_root}, |
| 771 | {0, }, |
| 772 | }; |
| 773 | |
| 774 | static struct ctl_table_header * sysctl_header; |
| 775 | |
| 776 | /* NOTE: we do *not* support sysctl() syscall */ |
| 777 | |
| 778 | static int __init init_sysctl(void) |
| 779 | { |
| 780 | int err = 0; |
| 781 | ctl_table * next = &oprof_table[nr_oprof_static]; |
| 782 | |
| 783 | /* these sysctl parms need sensible value */ |
| 784 | sysctl_parms.buf_size = OP_DEFAULT_BUF_SIZE; |
| 785 | sysctl_parms.note_size = OP_DEFAULT_NOTE_SIZE; |
| 786 | |
| 787 | if ((err = int_ops->add_sysctls(next))) |
| 788 | return err; |
| 789 | |
| 790 | sysctl_header = register_sysctl_table(dev_root, 0); |
| 791 | return err; |
| 792 | } |
| 793 | |
| 794 | /* not safe to mark as __exit since used from __init code */ |
| 795 | static void cleanup_sysctl(void) |
| 796 | { |
| 797 | ctl_table * next = &oprof_table[nr_oprof_static]; |
| 798 | unregister_sysctl_table(sysctl_header); |
| 799 | |
| 800 | int_ops->remove_sysctls(next); |
| 801 | |
| 802 | return; |
| 803 | } |
| 804 | |
| 805 | static int can_unload(void) |
| 806 | { |
| 807 | int can = -EBUSY; |
| 808 | down(&sysctlsem); |
| 809 | |
| 810 | if (allow_unload && state == STOPPED && !GET_USE_COUNT(THIS_MODULE)) |
| 811 | can = 0; |
| 812 | up(&sysctlsem); |
| 813 | return can; |
| 814 | } |
| 815 | |
| 816 | int __init oprof_init(void) |
| 817 | { |
| 818 | int err = 0; |
| 819 | |
| 820 | if (sysctl.cpu_type != CPU_RTC) { |
| 821 | int_ops = op_int_interface(); |
| 822 | |
| 823 | // try to init, fall back to rtc if not |
| 824 | if ((err = int_ops->init())) { |
| 825 | int_ops = &op_rtc_ops; |
| 826 | if ((err = int_ops->init())) |
| 827 | return err; |
| 828 | sysctl.cpu_type = CPU_RTC; |
| 829 | } |
| 830 | } else { |
| 831 | int_ops = &op_rtc_ops; |
| 832 | if ((err = int_ops->init())) |
| 833 | return err; |
| 834 | } |
| 835 | |
| 836 | if ((err = init_sysctl())) |
| 837 | goto out_err; |
| 838 | |
| 839 | err = op_major = register_chrdev(0, "oprof", &oprof_fops); |
| 840 | if (err < 0) |
| 841 | goto out_err2; |
| 842 | |
| 843 | err = oprof_init_hashmap(); |
| 844 | if (err < 0) { |
| 845 | printk(KERN_ERR "oprofile: couldn't allocate hash map !\n"); |
| 846 | unregister_chrdev(op_major, "oprof"); |
| 847 | goto out_err2; |
| 848 | } |
| 849 | |
| 850 | /* module might not be unloadable */ |
| 851 | THIS_MODULE->can_unload = can_unload; |
| 852 | |
| 853 | /* do this now so we don't have to track save/restores later */ |
| 854 | op_save_syscalls(); |
| 855 | |
| 856 | printk(KERN_INFO "%s loaded, major %u\n", op_version, op_major); |
| 857 | return 0; |
| 858 | |
| 859 | out_err2: |
| 860 | cleanup_sysctl(); |
| 861 | out_err: |
| 862 | int_ops->deinit(); |
| 863 | return err; |
| 864 | } |
| 865 | |
| 866 | void __exit oprof_exit(void) |
| 867 | { |
| 868 | oprof_free_hashmap(); |
| 869 | |
| 870 | unregister_chrdev(op_major, "oprof"); |
| 871 | |
| 872 | cleanup_sysctl(); |
| 873 | |
| 874 | int_ops->deinit(); |
| 875 | } |
| 876 | |
| 877 | /* |
| 878 | * "The most valuable commodity I know of is information." |
| 879 | * - Gordon Gekko |
| 880 | */ |