More lockdep gifts, a 5-way lockup race: perf_event_create_kernel_counter() perf_event_alloc() perf_try_init_event() x86_pmu_event_init() __x86_pmu_event_init() x86_reserve_hardware() #0 mutex_lock(&pmc_reserve_mutex); reserve_ds_buffer() #1 get_online_cpus() perf_event_release_kernel() _free_event() hw_perf_event_destroy() x86_release_hardware() #0 mutex_lock(&pmc_reserve_mutex) release_ds_buffer() #1 get_online_cpus() #1 do_cpu_up() perf_event_init_cpu() #2 mutex_lock(&pmus_lock) #3 mutex_lock(&ctx->mutex) sys_perf_event_open() mutex_lock_double() #3 mutex_lock(ctx->mutex) #4 mutex_lock_nested(ctx->mutex, 1); perf_try_init_event() #4 mutex_lock_nested(ctx->mutex, 1) x86_pmu_event_init() intel_pmu_hw_config() x86_add_exclusive() #0 mutex_lock(&pmc_reserve_mutex) Fix it by using ordering constructs instead of locking.
Unified · Split
| @@ -372,10 +372,9 @@ static int alloc_pebs_buffer(int cpu) · 371 unmodified lines | |||
| 372 | static void release_pebs_buffer(int cpu) | 372 | static void release_pebs_buffer(int cpu) |
| 373 | { | 373 | { |
| 374 | struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu); | 374 | struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu); |
| 375 | struct debug_store *ds = hwev->ds; | ||
| 376 | void *cea; | 375 | void *cea; |
| 377 | 376 | ||
| 378 | if (!ds || !x86_pmu.pebs) | 377 | if (!x86_pmu.pebs) |
| 379 | return; | 378 | return; |
| 380 | 379 | ||
| 381 | kfree(per_cpu(insn_buffer, cpu)); | 380 | kfree(per_cpu(insn_buffer, cpu)); |
| @@ -384,7 +383,6 @@ static void release_pebs_buffer(int cpu) · 2 unmodified lines | |||
| 384 | /* Clear the fixmap */ | 383 | /* Clear the fixmap */ |
| 385 | cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.pebs_buffer; | 384 | cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.pebs_buffer; |
| 386 | ds_clear_cea(cea, x86_pmu.pebs_buffer_size); | 385 | ds_clear_cea(cea, x86_pmu.pebs_buffer_size); |
| 387 | ds->pebs_buffer_base = 0; | ||
| 388 | dsfree_pages(hwev->ds_pebs_vaddr, x86_pmu.pebs_buffer_size); | 386 | dsfree_pages(hwev->ds_pebs_vaddr, x86_pmu.pebs_buffer_size); |
| 389 | hwev->ds_pebs_vaddr = NULL; | 387 | hwev->ds_pebs_vaddr = NULL; |
| 390 | } | 388 | } |
| @@ -419,16 +417,14 @@ static int alloc_bts_buffer(int cpu) · 28 unmodified lines | |||
| 419 | static void release_bts_buffer(int cpu) | 417 | static void release_bts_buffer(int cpu) |
| 420 | { | 418 | { |
| 421 | struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu); | 419 | struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu); |
| 422 | struct debug_store *ds = hwev->ds; | ||
| 423 | void *cea; | 420 | void *cea; |
| 424 | 421 | ||
| 425 | if (!ds || !x86_pmu.bts) | 422 | if (!x86_pmu.bts) |
| 426 | return; | 423 | return; |
| 427 | 424 | ||
| 428 | /* Clear the fixmap */ | 425 | /* Clear the fixmap */ |
| 429 | cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.bts_buffer; | 426 | cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.bts_buffer; |
| 430 | ds_clear_cea(cea, BTS_BUFFER_SIZE); | 427 | ds_clear_cea(cea, BTS_BUFFER_SIZE); |
| 431 | ds->bts_buffer_base = 0; | ||
| 432 | dsfree_pages(hwev->ds_bts_vaddr, BTS_BUFFER_SIZE); | 428 | dsfree_pages(hwev->ds_bts_vaddr, BTS_BUFFER_SIZE); |
| 433 | hwev->ds_bts_vaddr = NULL; | 429 | hwev->ds_bts_vaddr = NULL; |
| 434 | } | 430 | } |
| @@ -454,16 +450,22 @@ void release_ds_buffers(void) · 19 unmodified lines | |||
| 454 | if (!x86_pmu.bts && !x86_pmu.pebs) | 450 | if (!x86_pmu.bts && !x86_pmu.pebs) |
| 455 | return; | 451 | return; |
| 456 | 452 | ||
| 457 | get_online_cpus(); | 453 | for_each_possible_cpu(cpu) |
| 458 | for_each_online_cpu(cpu) | 454 | release_ds_buffer(cpu); |
| 455 | |||
| 456 | for_each_possible_cpu(cpu) { | ||
| 457 | /* | ||
| 458 | * Again, ignore errors from offline CPUs, they will no longer | ||
| 459 | * observe cpu_hw_events.ds and not program the DS_AREA when | ||
| 460 | * they come up. | ||
| 461 | */ | ||
| 459 | fini_debug_store_on_cpu(cpu); | 462 | fini_debug_store_on_cpu(cpu); |
| 463 | } | ||
| 460 | 464 | ||
| 461 | for_each_possible_cpu(cpu) { | 465 | for_each_possible_cpu(cpu) { |
| 462 | release_pebs_buffer(cpu); | 466 | release_pebs_buffer(cpu); |
| 463 | release_bts_buffer(cpu); | 467 | release_bts_buffer(cpu); |
| 464 | release_ds_buffer(cpu); | ||
| 465 | } | 468 | } |
| 466 | put_online_cpus(); | ||
| 467 | } | 469 | } |
| 468 | 470 | ||
| 469 | void reserve_ds_buffers(void) | 471 | void reserve_ds_buffers(void) |
| @@ -483,8 +485,6 @@ void reserve_ds_buffers(void) · 13 unmodified lines | |||
| 483 | if (!x86_pmu.pebs) | 485 | if (!x86_pmu.pebs) |
| 484 | pebs_err = 1; | 486 | pebs_err = 1; |
| 485 | 487 | ||
| 486 | get_online_cpus(); | ||
| 487 | |||
| 488 | for_each_possible_cpu(cpu) { | 488 | for_each_possible_cpu(cpu) { |
| 489 | if (alloc_ds_buffer(cpu)) { | 489 | if (alloc_ds_buffer(cpu)) { |
| 490 | bts_err = 1; | 490 | bts_err = 1; |
| @@ -521,11 +521,14 @@ void reserve_ds_buffers(void) · 30 unmodified lines | |||
| 521 | if (x86_pmu.pebs && !pebs_err) | 521 | if (x86_pmu.pebs && !pebs_err) |
| 522 | x86_pmu.pebs_active = 1; | 522 | x86_pmu.pebs_active = 1; |
| 523 | 523 | ||
| 524 | for_each_online_cpu(cpu) | 524 | for_each_possible_cpu(cpu) { |
| 525 | /* | ||
| 526 | * Ignores wrmsr_on_cpu() errors for offline CPUs they | ||
| 527 | * will get this call through intel_pmu_cpu_starting(). | ||
| 528 | */ | ||
| 525 | init_debug_store_on_cpu(cpu); | 529 | init_debug_store_on_cpu(cpu); |
| 530 | } | ||
| 526 | } | 531 | } |
| 527 | |||
| 528 | put_online_cpus(); | ||
| 529 | } | 532 | } |
| 530 | 533 | ||
| 531 | /* | 534 | /* |