1// SPDX-License-Identifier: GPL-2.0-or-later 2/* 3 * Cell Broadband Engine OProfile Support 4 * 5 * (C) Copyright IBM Corporation 2006 6 * 7 * Author: Maynard Johnson <maynardj@us.ibm.com> 8 */ 9 10/* The purpose of this file is to handle SPU event task switching 11 * and to record SPU context information into the OProfile 12 * event buffer. 13 * 14 * Additionally, the spu_sync_buffer function is provided as a helper 15 * for recoding actual SPU program counter samples to the event buffer. 16 */ 17#include <linux/dcookies.h> 18#include <linux/kref.h> 19#include <linux/mm.h> 20#include <linux/fs.h> 21#include <linux/file.h> 22#include <linux/module.h> 23#include <linux/notifier.h> 24#include <linux/numa.h> 25#include <linux/oprofile.h> 26#include <linux/slab.h> 27#include <linux/spinlock.h> 28#include "pr_util.h" 29 30#define RELEASE_ALL 9999 31 32static DEFINE_SPINLOCK(buffer_lock); 33static DEFINE_SPINLOCK(cache_lock); 34static int num_spu_nodes; 35static int spu_prof_num_nodes; 36 37struct spu_buffer spu_buff[MAX_NUMNODES * SPUS_PER_NODE]; 38struct delayed_work spu_work; 39static unsigned max_spu_buff; 40 41static void spu_buff_add(unsigned long int value, int spu) 42{ 43 /* spu buff is a circular buffer. Add entries to the 44 * head. Head is the index to store the next value. 45 * The buffer is full when there is one available entry 46 * in the queue, i.e. head and tail can't be equal. 47 * That way we can tell the difference between the 48 * buffer being full versus empty. 49 * 50 * ASSUMPTION: the buffer_lock is held when this function 51 * is called to lock the buffer, head and tail. 52 */ 53 int full = 1; 54 55 if (spu_buff[spu].head >= spu_buff[spu].tail) { 56 if ((spu_buff[spu].head - spu_buff[spu].tail) 57 < (max_spu_buff - 1)) 58 full = 0; 59 60 } else if (spu_buff[spu].tail > spu_buff[spu].head) { 61 if ((spu_buff[spu].tail - spu_buff[spu].head) 62 > 1) 63 full = 0; 64 } 65 66 if (!full) { 67 spu_buff[spu].buff[spu_buff[spu].head] = value; 68 spu_buff[spu].head++; 69 70 if (spu_buff[spu].head >= max_spu_buff) 71 spu_buff[spu].head = 0; 72 } else { 73 /* From the user's perspective make the SPU buffer 74 * size management/overflow look like we are using 75 * per cpu buffers. The user uses the same 76 * per cpu parameter to adjust the SPU buffer size. 77 * Increment the sample_lost_overflow to inform 78 * the user the buffer size needs to be increased. 79 */ 80 oprofile_cpu_buffer_inc_smpl_lost(); 81 } 82} 83 84/* This function copies the per SPU buffers to the 85 * OProfile kernel buffer. 86 */ 87static void sync_spu_buff(void) 88{ 89 int spu; 90 unsigned long flags; 91 int curr_head; 92 93 for (spu = 0; spu < num_spu_nodes; spu++) { 94 /* In case there was an issue and the buffer didn't 95 * get created skip it. 96 */ 97 if (spu_buff[spu].buff == NULL) 98 continue; 99 100 /* Hold the lock to make sure the head/tail 101 * doesn't change while spu_buff_add() is 102 * deciding if the buffer is full or not. 103 * Being a little paranoid. 104 */ 105 spin_lock_irqsave(&buffer_lock, flags); 106 curr_head = spu_buff[spu].head; 107 spin_unlock_irqrestore(&buffer_lock, flags); 108 109 /* Transfer the current contents to the kernel buffer. 110 * data can still be added to the head of the buffer. 111 */ 112 oprofile_put_buff(spu_buff[spu].buff, 113 spu_buff[spu].tail, 114 curr_head, max_spu_buff); 115 116 spin_lock_irqsave(&buffer_lock, flags); 117 spu_buff[spu].tail = curr_head; 118 spin_unlock_irqrestore(&buffer_lock, flags); 119 } 120 121} 122 123static void wq_sync_spu_buff(struct work_struct *work) 124{ 125 /* move data from spu buffers to kernel buffer */ 126 sync_spu_buff(); 127 128 /* only reschedule if profiling is not done */ 129 if (spu_prof_running) 130 schedule_delayed_work(&spu_work, DEFAULT_TIMER_EXPIRE); 131} 132 133/* Container for caching information about an active SPU task. */ 134struct cached_info { 135 struct vma_to_fileoffset_map *map; 136 struct spu *the_spu; /* needed to access pointer to local_store */ 137 struct kref cache_ref; 138}; 139 140static struct cached_info *spu_info[MAX_NUMNODES * 8]; 141 142static void destroy_cached_info(struct kref *kref) 143{ 144 struct cached_info *info; 145 146 info = container_of(kref, struct cached_info, cache_ref); 147 vma_map_free(info->map); 148 kfree(info); 149 module_put(THIS_MODULE); 150} 151 152/* Return the cached_info for the passed SPU number. 153 * ATTENTION: Callers are responsible for obtaining the 154 * cache_lock if needed prior to invoking this function. 155 */ 156static struct cached_info *get_cached_info(struct spu *the_spu, int spu_num) 157{ 158 struct kref *ref; 159 struct cached_info *ret_info; 160 161 if (spu_num >= num_spu_nodes) { 162 printk(KERN_ERR "SPU_PROF: " 163 "%s, line %d: Invalid index %d into spu info cache\n", 164 __func__, __LINE__, spu_num); 165 ret_info = NULL; 166 goto out; 167 } 168 if (!spu_info[spu_num] && the_spu) { 169 ref = spu_get_profile_private_kref(the_spu->ctx); 170 if (ref) { 171 spu_info[spu_num] = container_of(ref, struct cached_info, cache_ref); 172 kref_get(&spu_info[spu_num]->cache_ref); 173 } 174 } 175 176 ret_info = spu_info[spu_num]; 177 out: 178 return ret_info; 179} 180 181 182/* Looks for cached info for the passed spu. If not found, the 183 * cached info is created for the passed spu. 184 * Returns 0 for success; otherwise, -1 for error. 185 */ 186static int 187prepare_cached_spu_info(struct spu *spu, unsigned long objectId) 188{ 189 unsigned long flags; 190 struct vma_to_fileoffset_map *new_map; 191 int retval = 0; 192 struct cached_info *info; 193 194 /* We won't bother getting cache_lock here since 195 * don't do anything with the cached_info that's returned. 196 */ 197 info = get_cached_info(spu, spu->number); 198 199 if (info) { 200 pr_debug("Found cached SPU info.\n"); 201 goto out; 202 } 203 204 /* Create cached_info and set spu_info[spu->number] to point to it. 205 * spu->number is a system-wide value, not a per-node value. 206 */ 207 info = kzalloc(sizeof(*info), GFP_KERNEL); 208 if (!info) { 209 printk(KERN_ERR "SPU_PROF: " 210 "%s, line %d: create vma_map failed\n", 211 __func__, __LINE__); 212 retval = -ENOMEM; 213 goto err_alloc; 214 } 215 new_map = create_vma_map(spu, objectId); 216 if (!new_map) { 217 printk(KERN_ERR "SPU_PROF: " 218 "%s, line %d: create vma_map failed\n", 219 __func__, __LINE__); 220 retval = -ENOMEM; 221 goto err_alloc; 222 } 223 224 pr_debug("Created vma_map\n"); 225 info->map = new_map; 226 info->the_spu = spu; 227 kref_init(&info->cache_ref); 228 spin_lock_irqsave(&cache_lock, flags); 229 spu_info[spu->number] = info; 230 /* Increment count before passing off ref to SPUFS. */ 231 kref_get(&info->cache_ref); 232 233 /* We increment the module refcount here since SPUFS is 234 * responsible for the final destruction of the cached_info, 235 * and it must be able to access the destroy_cached_info() 236 * function defined in the OProfile module. We decrement 237 * the module refcount in destroy_cached_info. 238 */ 239 try_module_get(THIS_MODULE); 240 spu_set_profile_private_kref(spu->ctx, &info->cache_ref, 241 destroy_cached_info); 242 spin_unlock_irqrestore(&cache_lock, flags); 243 goto out; 244 245err_alloc: 246 kfree(info); 247out: 248 return retval; 249} 250 251/* 252 * NOTE: The caller is responsible for locking the 253 * cache_lock prior to calling this function. 254 */ 255static int release_cached_info(int spu_index) 256{ 257 int index, end; 258 259 if (spu_index == RELEASE_ALL) { 260 end = num_spu_nodes; 261 index = 0; 262 } else { 263 if (spu_index >= num_spu_nodes) { 264 printk(KERN_ERR "SPU_PROF: " 265 "%s, line %d: " 266 "Invalid index %d into spu info cache\n", 267 __func__, __LINE__, spu_index); 268 goto out; 269 } 270 end = spu_index + 1; 271 index = spu_index; 272 } 273 for (; index < end; index++) { 274 if (spu_info[index]) { 275 kref_put(&spu_info[index]->cache_ref, 276 destroy_cached_info); 277 spu_info[index] = NULL; 278 } 279 } 280 281out: 282 return 0; 283} 284 285/* The source code for fast_get_dcookie was "borrowed" 286 * from drivers/oprofile/buffer_sync.c. 287 */ 288 289/* Optimisation. We can manage without taking the dcookie sem 290 * because we cannot reach this code without at least one 291 * dcookie user still being registered (namely, the reader 292 * of the event buffer). 293 */ 294static inline unsigned long fast_get_dcookie(const struct path *path) 295{ 296 unsigned long cookie; 297 298 if (path->dentry->d_flags & DCACHE_COOKIE) 299 return (unsigned long)path->dentry; 300 get_dcookie(path, &cookie); 301 return cookie; 302} 303 304/* Look up the dcookie for the task's mm->exe_file, 305 * which corresponds loosely to "application name". Also, determine 306 * the offset for the SPU ELF object. If computed offset is 307 * non-zero, it implies an embedded SPU object; otherwise, it's a 308 * separate SPU binary, in which case we retrieve it's dcookie. 309 * For the embedded case, we must determine if SPU ELF is embedded 310 * in the executable application or another file (i.e., shared lib). 311 * If embedded in a shared lib, we must get the dcookie and return 312 * that to the caller. 313 */ 314static unsigned long 315get_exec_dcookie_and_offset(struct spu *spu, unsigned int *offsetp, 316 unsigned long *spu_bin_dcookie, 317 unsigned long spu_ref) 318{ 319 unsigned long app_cookie = 0; 320 unsigned int my_offset = 0; 321 struct vm_area_struct *vma; 322 struct file *exe_file; 323 struct mm_struct *mm = spu->mm; 324 325 if (!mm) 326 goto out; 327 328 exe_file = get_mm_exe_file(mm); 329 if (exe_file) { 330 app_cookie = fast_get_dcookie(&exe_file->f_path); 331 pr_debug("got dcookie for %pD\n", exe_file); 332 fput(exe_file); 333 } 334 335 mmap_read_lock(mm); 336 for (vma = mm->mmap; vma; vma = vma->vm_next) { 337 if (vma->vm_start > spu_ref || vma->vm_end <= spu_ref) 338 continue; 339 my_offset = spu_ref - vma->vm_start; 340 if (!vma->vm_file) 341 goto fail_no_image_cookie; 342 343 pr_debug("Found spu ELF at %X(object-id:%lx) for file %pD\n", 344 my_offset, spu_ref, vma->vm_file); 345 *offsetp = my_offset; 346 break; 347 } 348 349 *spu_bin_dcookie = fast_get_dcookie(&vma->vm_file->f_path); 350 pr_debug("got dcookie for %pD\n", vma->vm_file); 351 352 mmap_read_unlock(mm); 353 354out: 355 return app_cookie; 356 357fail_no_image_cookie: 358 mmap_read_unlock(mm); 359 360 printk(KERN_ERR "SPU_PROF: " 361 "%s, line %d: Cannot find dcookie for SPU binary\n", 362 __func__, __LINE__); 363 goto out; 364} 365 366 367 368/* This function finds or creates cached context information for the 369 * passed SPU and records SPU context information into the OProfile 370 * event buffer. 371 */ 372static int process_context_switch(struct spu *spu, unsigned long objectId) 373{ 374 unsigned long flags; 375 int retval; 376 unsigned int offset = 0; 377 unsigned long spu_cookie = 0, app_dcookie; 378 379 retval = prepare_cached_spu_info(spu, objectId); 380 if (retval) 381 goto out; 382 383 /* Get dcookie first because a mutex_lock is taken in that 384 * code path, so interrupts must not be disabled. 385 */ 386 app_dcookie = get_exec_dcookie_and_offset(spu, &offset, &spu_cookie, objectId); 387 if (!app_dcookie || !spu_cookie) { 388 retval = -ENOENT; 389 goto out; 390 } 391 392 /* Record context info in event buffer */ 393 spin_lock_irqsave(&buffer_lock, flags); 394 spu_buff_add(ESCAPE_CODE, spu->number); 395 spu_buff_add(SPU_CTX_SWITCH_CODE, spu->number); 396 spu_buff_add(spu->number, spu->number); 397 spu_buff_add(spu->pid, spu->number); 398 spu_buff_add(spu->tgid, spu->number); 399 spu_buff_add(app_dcookie, spu->number); 400 spu_buff_add(spu_cookie, spu->number); 401 spu_buff_add(offset, spu->number); 402 403 /* Set flag to indicate SPU PC data can now be written out. If 404 * the SPU program counter data is seen before an SPU context 405 * record is seen, the postprocessing will fail. 406 */ 407 spu_buff[spu->number].ctx_sw_seen = 1; 408 409 spin_unlock_irqrestore(&buffer_lock, flags); 410 smp_wmb(); /* insure spu event buffer updates are written */ 411 /* don't want entries intermingled... */ 412out: 413 return retval; 414} 415 416/* 417 * This function is invoked on either a bind_context or unbind_context. 418 * If called for an unbind_context, the val arg is 0; otherwise, 419 * it is the object-id value for the spu context. 420 * The data arg is of type 'struct spu *'. 421 */ 422static int spu_active_notify(struct notifier_block *self, unsigned long val, 423 void *data) 424{ 425 int retval; 426 unsigned long flags; 427 struct spu *the_spu = data; 428 429 pr_debug("SPU event notification arrived\n"); 430 if (!val) { 431 spin_lock_irqsave(&cache_lock, flags); 432 retval = release_cached_info(the_spu->number); 433 spin_unlock_irqrestore(&cache_lock, flags); 434 } else { 435 retval = process_context_switch(the_spu, val); 436 } 437 return retval; 438} 439 440static struct notifier_block spu_active = { 441 .notifier_call = spu_active_notify, 442}; 443 444static int number_of_online_nodes(void) 445{ 446 u32 cpu; u32 tmp; 447 int nodes = 0; 448 for_each_online_cpu(cpu) { 449 tmp = cbe_cpu_to_node(cpu) + 1; 450 if (tmp > nodes) 451 nodes++; 452 } 453 return nodes; 454} 455 456static int oprofile_spu_buff_create(void) 457{ 458 int spu; 459 460 max_spu_buff = oprofile_get_cpu_buffer_size(); 461 462 for (spu = 0; spu < num_spu_nodes; spu++) { 463 /* create circular buffers to store the data in. 464 * use locks to manage accessing the buffers 465 */ 466 spu_buff[spu].head = 0; 467 spu_buff[spu].tail = 0; 468 469 /* 470 * Create a buffer for each SPU. Can't reliably 471 * create a single buffer for all spus due to not 472 * enough contiguous kernel memory. 473 */ 474 475 spu_buff[spu].buff = kzalloc((max_spu_buff 476 * sizeof(unsigned long)), 477 GFP_KERNEL); 478 479 if (!spu_buff[spu].buff) { 480 printk(KERN_ERR "SPU_PROF: " 481 "%s, line %d: oprofile_spu_buff_create " 482 "failed to allocate spu buffer %d.\n", 483 __func__, __LINE__, spu); 484 485 /* release the spu buffers that have been allocated */ 486 while (spu >= 0) { 487 kfree(spu_buff[spu].buff); 488 spu_buff[spu].buff = 0; 489 spu--; 490 } 491 return -ENOMEM; 492 } 493 } 494 return 0; 495} 496 497/* The main purpose of this function is to synchronize 498 * OProfile with SPUFS by registering to be notified of 499 * SPU task switches. 500 * 501 * NOTE: When profiling SPUs, we must ensure that only 502 * spu_sync_start is invoked and not the generic sync_start 503 * in drivers/oprofile/oprof.c. A return value of 504 * SKIP_GENERIC_SYNC or SYNC_START_ERROR will 505 * accomplish this. 506 */ 507int spu_sync_start(void) 508{ 509 int spu; 510 int ret = SKIP_GENERIC_SYNC; 511 int register_ret; 512 unsigned long flags = 0; 513 514 spu_prof_num_nodes = number_of_online_nodes(); 515 num_spu_nodes = spu_prof_num_nodes * 8; 516 INIT_DELAYED_WORK(&spu_work, wq_sync_spu_buff); 517 518 /* create buffer for storing the SPU data to put in 519 * the kernel buffer. 520 */ 521 ret = oprofile_spu_buff_create(); 522 if (ret) 523 goto out; 524 525 spin_lock_irqsave(&buffer_lock, flags); 526 for (spu = 0; spu < num_spu_nodes; spu++) { 527 spu_buff_add(ESCAPE_CODE, spu); 528 spu_buff_add(SPU_PROFILING_CODE, spu); 529 spu_buff_add(num_spu_nodes, spu); 530 } 531 spin_unlock_irqrestore(&buffer_lock, flags); 532 533 for (spu = 0; spu < num_spu_nodes; spu++) { 534 spu_buff[spu].ctx_sw_seen = 0; 535 spu_buff[spu].last_guard_val = 0; 536 } 537 538 /* Register for SPU events */ 539 register_ret = spu_switch_event_register(&spu_active); 540 if (register_ret) { 541 ret = SYNC_START_ERROR; 542 goto out; 543 } 544 545 pr_debug("spu_sync_start -- running.\n"); 546out: 547 return ret; 548} 549 550/* Record SPU program counter samples to the oprofile event buffer. */ 551void spu_sync_buffer(int spu_num, unsigned int *samples, 552 int num_samples) 553{ 554 unsigned long long file_offset; 555 unsigned long flags; 556 int i; 557 struct vma_to_fileoffset_map *map; 558 struct spu *the_spu; 559 unsigned long long spu_num_ll = spu_num; 560 unsigned long long spu_num_shifted = spu_num_ll << 32; 561 struct cached_info *c_info; 562 563 /* We need to obtain the cache_lock here because it's 564 * possible that after getting the cached_info, the SPU job 565 * corresponding to this cached_info may end, thus resulting 566 * in the destruction of the cached_info. 567 */ 568 spin_lock_irqsave(&cache_lock, flags); 569 c_info = get_cached_info(NULL, spu_num); 570 if (!c_info) { 571 /* This legitimately happens when the SPU task ends before all 572 * samples are recorded. 573 * No big deal -- so we just drop a few samples. 574 */ 575 pr_debug("SPU_PROF: No cached SPU context " 576 "for SPU #%d. Dropping samples.\n", spu_num); 577 goto out; 578 } 579 580 map = c_info->map; 581 the_spu = c_info->the_spu; 582 spin_lock(&buffer_lock); 583 for (i = 0; i < num_samples; i++) { 584 unsigned int sample = *(samples+i); 585 int grd_val = 0; 586 file_offset = 0; 587 if (sample == 0) 588 continue; 589 file_offset = vma_map_lookup( map, sample, the_spu, &grd_val); 590 591 /* If overlays are used by this SPU application, the guard 592 * value is non-zero, indicating which overlay section is in 593 * use. We need to discard samples taken during the time 594 * period which an overlay occurs (i.e., guard value changes). 595 */ 596 if (grd_val && grd_val != spu_buff[spu_num].last_guard_val) { 597 spu_buff[spu_num].last_guard_val = grd_val; 598 /* Drop the rest of the samples. */ 599 break; 600 } 601 602 /* We must ensure that the SPU context switch has been written 603 * out before samples for the SPU. Otherwise, the SPU context 604 * information is not available and the postprocessing of the 605 * SPU PC will fail with no available anonymous map information. 606 */ 607 if (spu_buff[spu_num].ctx_sw_seen) 608 spu_buff_add((file_offset | spu_num_shifted), 609 spu_num); 610 } 611 spin_unlock(&buffer_lock); 612out: 613 spin_unlock_irqrestore(&cache_lock, flags); 614} 615 616 617int spu_sync_stop(void) 618{ 619 unsigned long flags = 0; 620 int ret; 621 int k; 622 623 ret = spu_switch_event_unregister(&spu_active); 624 625 if (ret) 626 printk(KERN_ERR "SPU_PROF: " 627 "%s, line %d: spu_switch_event_unregister " \ 628 "returned %d\n", 629 __func__, __LINE__, ret); 630 631 /* flush any remaining data in the per SPU buffers */ 632 sync_spu_buff(); 633 634 spin_lock_irqsave(&cache_lock, flags); 635 ret = release_cached_info(RELEASE_ALL); 636 spin_unlock_irqrestore(&cache_lock, flags); 637 638 /* remove scheduled work queue item rather then waiting 639 * for every queued entry to execute. Then flush pending 640 * system wide buffer to event buffer. 641 */ 642 cancel_delayed_work(&spu_work); 643 644 for (k = 0; k < num_spu_nodes; k++) { 645 spu_buff[k].ctx_sw_seen = 0; 646 647 /* 648 * spu_sys_buff will be null if there was a problem 649 * allocating the buffer. Only delete if it exists. 650 */ 651 kfree(spu_buff[k].buff); 652 spu_buff[k].buff = 0; 653 } 654 pr_debug("spu_sync_stop -- done.\n"); 655 return ret; 656} 657 658