amd-drm-next-6.12-2024-09-13: amdgpu: - GPUVM sync fixes - kdoc fixes - Misc spelling mistakes - Add some raven GFXOFF quirks - Use clamp helper - DC fixes - JPEG fixes - Process isolation fix - Queue reset fix - W=1 cleanup - SMU14 fixes - JPEG fixes amdkfd: - Fetch cacheline info from IP discovery - Queue reset fix - RAS fix - Document SVM events - CRIU fixes - Race fix in dma-buf handling drm: - dma-buf fd race fixes Signed-off-by: Dave Airlie <airlied@redhat.com>
Showing the changes to include/uapi (1 of 50 files). Show all
Unified · Split
| @@ -540,26 +540,29 @@ enum kfd_smi_event { · 539 unmodified lines | ||
| 540 | 540 | KFD_SMI_EVENT_ALL_PROCESS = 64 |
| 541 | 541 | }; |
| 542 | 542 | |
| 543 | +/* The reason of the page migration event */ | |
| 543 | 544 | enum KFD_MIGRATE_TRIGGERS { |
| 544 | − KFD_MIGRATE_TRIGGER_PREFETCH, | |
| 545 | − KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU, | |
| 546 | − KFD_MIGRATE_TRIGGER_PAGEFAULT_CPU, | |
| 547 | − KFD_MIGRATE_TRIGGER_TTM_EVICTION | |
| 545 | + KFD_MIGRATE_TRIGGER_PREFETCH, /* Prefetch to GPU VRAM or system memory */ | |
| 546 | + KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU, /* GPU page fault recover */ | |
| 547 | + KFD_MIGRATE_TRIGGER_PAGEFAULT_CPU, /* CPU page fault recover */ | |
| 548 | + KFD_MIGRATE_TRIGGER_TTM_EVICTION /* TTM eviction */ | |
| 548 | 549 | }; |
| 549 | 550 | |
| 551 | +/* The reason of user queue evition event */ | |
| 550 | 552 | enum KFD_QUEUE_EVICTION_TRIGGERS { |
| 551 | − KFD_QUEUE_EVICTION_TRIGGER_SVM, | |
| 552 | − KFD_QUEUE_EVICTION_TRIGGER_USERPTR, | |
| 553 | − KFD_QUEUE_EVICTION_TRIGGER_TTM, | |
| 554 | − KFD_QUEUE_EVICTION_TRIGGER_SUSPEND, | |
| 555 | − KFD_QUEUE_EVICTION_CRIU_CHECKPOINT, | |
| 556 | − KFD_QUEUE_EVICTION_CRIU_RESTORE | |
| 553 | + KFD_QUEUE_EVICTION_TRIGGER_SVM, /* SVM buffer migration */ | |
| 554 | + KFD_QUEUE_EVICTION_TRIGGER_USERPTR, /* userptr movement */ | |
| 555 | + KFD_QUEUE_EVICTION_TRIGGER_TTM, /* TTM move buffer */ | |
| 556 | + KFD_QUEUE_EVICTION_TRIGGER_SUSPEND, /* GPU suspend */ | |
| 557 | + KFD_QUEUE_EVICTION_CRIU_CHECKPOINT, /* CRIU checkpoint */ | |
| 558 | + KFD_QUEUE_EVICTION_CRIU_RESTORE /* CRIU restore */ | |
| 557 | 559 | }; |
| 558 | 560 | |
| 561 | +/* The reason of unmap buffer from GPU event */ | |
| 559 | 562 | enum KFD_SVM_UNMAP_TRIGGERS { |
| 560 | − KFD_SVM_UNMAP_TRIGGER_MMU_NOTIFY, | |
| 561 | − KFD_SVM_UNMAP_TRIGGER_MMU_NOTIFY_MIGRATE, | |
| 562 | − KFD_SVM_UNMAP_TRIGGER_UNMAP_FROM_CPU | |
| 563 | + KFD_SVM_UNMAP_TRIGGER_MMU_NOTIFY, /* MMU notifier CPU buffer movement */ | |
| 564 | + KFD_SVM_UNMAP_TRIGGER_MMU_NOTIFY_MIGRATE,/* MMU notifier page migration */ | |
| 565 | + KFD_SVM_UNMAP_TRIGGER_UNMAP_FROM_CPU /* Unmap to free the buffer */ | |
| 563 | 566 | }; |
| 564 | 567 | |
| 565 | 568 | #define KFD_SMI_EVENT_MASK_FROM_INDEX(i) (1ULL << ((i) - 1)) |
| @@ -570,6 +573,77 @@ struct kfd_ioctl_smi_events_args { · 4 unmodified lines | ||
| 570 | 573 | __u32 anon_fd; /* from KFD */ |
| 571 | 574 | }; |
| 572 | 575 | |
| 576 | +/* | |
| 577 | + * SVM event tracing via SMI system management interface | |
| 578 | + * | |
| 579 | + * Open event file descriptor | |
| 580 | + * use ioctl AMDKFD_IOC_SMI_EVENTS, pass in gpuid and return a anonymous file | |
| 581 | + * descriptor to receive SMI events. | |
| 582 | + * If calling with sudo permission, then file descriptor can be used to receive | |
| 583 | + * SVM events from all processes, otherwise, to only receive SVM events of same | |
| 584 | + * process. | |
| 585 | + * | |
| 586 | + * To enable the SVM event | |
| 587 | + * Write event file descriptor with KFD_SMI_EVENT_MASK_FROM_INDEX(event) bitmap | |
| 588 | + * mask to start record the event to the kfifo, use bitmap mask combination | |
| 589 | + * for multiple events. New event mask will overwrite the previous event mask. | |
| 590 | + * KFD_SMI_EVENT_MASK_FROM_INDEX(KFD_SMI_EVENT_ALL_PROCESS) bit requires sudo | |
| 591 | + * permisson to receive SVM events from all process. | |
| 592 | + * | |
| 593 | + * To receive the event | |
| 594 | + * Application can poll file descriptor to wait for the events, then read event | |
| 595 | + * from the file into a buffer. Each event is one line string message, starting | |
| 596 | + * with the event id, then the event specific information. | |
| 597 | + * | |
| 598 | + * To decode event information | |
| 599 | + * The following event format string macro can be used with sscanf to decode | |
| 600 | + * the specific event information. | |
| 601 | + * event triggers: the reason to generate the event, defined as enum for unmap, | |
| 602 | + * eviction and migrate events. | |
| 603 | + * node, from, to, prefetch_loc, preferred_loc: GPU ID, or 0 for system memory. | |
| 604 | + * addr: user mode address, in pages | |
| 605 | + * size: in pages | |
| 606 | + * pid: the process ID to generate the event | |
| 607 | + * ns: timestamp in nanosecond-resolution, starts at system boot time but | |
| 608 | + * stops during suspend | |
| 609 | + * migrate_update: GPU page fault is recovered by 'M' for migrate, 'U' for update | |
| 610 | + * rw: 'W' for write page fault, 'R' for read page fault | |
| 611 | + * rescheduled: 'R' if the queue restore failed and rescheduled to try again | |
| 612 | + */ | |
| 613 | +#define KFD_EVENT_FMT_UPDATE_GPU_RESET(reset_seq_num, reset_cause)\ | |
| 614 | + "%x %s\n", (reset_seq_num), (reset_cause) | |
| 615 | + | |
| 616 | +#define KFD_EVENT_FMT_THERMAL_THROTTLING(bitmask, counter)\ | |
| 617 | + "%llx:%llx\n", (bitmask), (counter) | |
| 618 | + | |
| 619 | +#define KFD_EVENT_FMT_VMFAULT(pid, task_name)\ | |
| 620 | + "%x:%s\n", (pid), (task_name) | |
| 621 | + | |
| 622 | +#define KFD_EVENT_FMT_PAGEFAULT_START(ns, pid, addr, node, rw)\ | |
| 623 | + "%lld -%d @%lx(%x) %c\n", (ns), (pid), (addr), (node), (rw) | |
| 624 | + | |
| 625 | +#define KFD_EVENT_FMT_PAGEFAULT_END(ns, pid, addr, node, migrate_update)\ | |
| 626 | + "%lld -%d @%lx(%x) %c\n", (ns), (pid), (addr), (node), (migrate_update) | |
| 627 | + | |
| 628 | +#define KFD_EVENT_FMT_MIGRATE_START(ns, pid, start, size, from, to, prefetch_loc,\ | |
| 629 | + preferred_loc, migrate_trigger)\ | |
| 630 | + "%lld -%d @%lx(%lx) %x->%x %x:%x %d\n", (ns), (pid), (start), (size),\ | |
| 631 | + (from), (to), (prefetch_loc), (preferred_loc), (migrate_trigger) | |
| 632 | + | |
| 633 | +#define KFD_EVENT_FMT_MIGRATE_END(ns, pid, start, size, from, to, migrate_trigger)\ | |
| 634 | + "%lld -%d @%lx(%lx) %x->%x %d\n", (ns), (pid), (start), (size),\ | |
| 635 | + (from), (to), (migrate_trigger) | |
| 636 | + | |
| 637 | +#define KFD_EVENT_FMT_QUEUE_EVICTION(ns, pid, node, evict_trigger)\ | |
| 638 | + "%lld -%d %x %d\n", (ns), (pid), (node), (evict_trigger) | |
| 639 | + | |
| 640 | +#define KFD_EVENT_FMT_QUEUE_RESTORE(ns, pid, node, rescheduled)\ | |
| 641 | + "%lld -%d %x %c\n", (ns), (pid), (node), (rescheduled) | |
| 642 | + | |
| 643 | +#define KFD_EVENT_FMT_UNMAP_FROM_GPU(ns, pid, addr, size, node, unmap_trigger)\ | |
| 644 | + "%lld -%d @%lx(%lx) %x %d\n", (ns), (pid), (addr), (size),\ | |
| 645 | + (node), (unmap_trigger) | |
| 646 | + | |
| 573 | 647 | /************************************************************************************************** |
| 574 | 648 | * CRIU IOCTLs (Checkpoint Restore In Userspace) |
| 575 | 649 | * |