rooty/open-gpu-kernel-modules
1
0
Fork 0
mirror of https://github.com/NVIDIA/open-gpu-kernel-modules.git synced 2024-12-02 00:24:22 +01:00
Code Issues Projects Releases Wiki Activity
337e28efda
open-gpu-kernel-modules/kernel-open/nvidia-uvm/uvm_ioctl.h
Andy Ritger eb5c7665a1
535.43.02
2023-05-30 10:11:36 -07:00

1109 lines
38 KiB
C
Raw Blame History

/*******************************************************************************
Copyright (c) 2013-2019 NVidia Corporation
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to
deal in the Software without restriction, including without limitation the
rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
sell copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be
included in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
DEALINGS IN THE SOFTWARE.
*******************************************************************************/
#ifndef _UVM_IOCTL_H
#define _UVM_IOCTL_H
#include "uvm_types.h"
#ifdef __cplusplus
extern "C" {
#endif
//
// Please see the header file (uvm.h) for detailed documentation on each of the
// associated API calls.
//
#if defined(WIN32) || defined(WIN64)
# define UVM_IOCTL_BASE(i) CTL_CODE(FILE_DEVICE_UNKNOWN, 0x800+i, METHOD_BUFFERED, FILE_READ_DATA | FILE_WRITE_DATA)
#else
# define UVM_IOCTL_BASE(i) i
#endif
//
// UvmReserveVa
//
#define UVM_RESERVE_VA UVM_IOCTL_BASE(1)
typedef struct
{
NvU64 requestedBase NV_ALIGN_BYTES(8); // IN
NvU64 length NV_ALIGN_BYTES(8); // IN
NV_STATUS rmStatus; // OUT
} UVM_RESERVE_VA_PARAMS;
//
// UvmReleaseVa
//
#define UVM_RELEASE_VA UVM_IOCTL_BASE(2)
typedef struct
{
NvU64 requestedBase NV_ALIGN_BYTES(8); // IN
NvU64 length NV_ALIGN_BYTES(8); // IN
NV_STATUS rmStatus; // OUT
} UVM_RELEASE_VA_PARAMS;
//
// UvmRegionCommit
//
#define UVM_REGION_COMMIT UVM_IOCTL_BASE(3)
typedef struct
{
NvU64 requestedBase NV_ALIGN_BYTES(8); // IN
NvU64 length NV_ALIGN_BYTES(8); // IN
UvmStream streamId NV_ALIGN_BYTES(8); // IN
NvProcessorUuid gpuUuid; // IN
NV_STATUS rmStatus; // OUT
} UVM_REGION_COMMIT_PARAMS;
//
// UvmRegionDecommit
//
#define UVM_REGION_DECOMMIT UVM_IOCTL_BASE(4)
typedef struct
{
NvU64 requestedBase NV_ALIGN_BYTES(8); // IN
NvU64 length NV_ALIGN_BYTES(8); // IN
NV_STATUS rmStatus; // OUT
} UVM_REGION_DECOMMIT_PARAMS;
//
// UvmRegionSetStream
//
#define UVM_REGION_SET_STREAM UVM_IOCTL_BASE(5)
typedef struct
{
NvU64 requestedBase NV_ALIGN_BYTES(8); // IN
NvU64 length NV_ALIGN_BYTES(8); // IN
UvmStream newStreamId NV_ALIGN_BYTES(8); // IN
NvProcessorUuid gpuUuid; // IN
NV_STATUS rmStatus; // OUT
} UVM_REGION_SET_STREAM_PARAMS;
//
// UvmSetStreamRunning
//
#define UVM_SET_STREAM_RUNNING UVM_IOCTL_BASE(6)
typedef struct
{
UvmStream streamId NV_ALIGN_BYTES(8); // IN
NV_STATUS rmStatus; // OUT
} UVM_SET_STREAM_RUNNING_PARAMS;
//
// Due to limitations in how much we want to send per ioctl call, the nStreams
// member must be less than or equal to about 250. That's an upper limit.
//
// However, from a typical user-space driver's point of view (for example, the
// CUDA driver), a vast majority of the time, we expect there to be only one
// stream passed in. The second most common case is something like atmost 32
// streams being passed in. The cases where there are more than 32 streams are
// the most rare. So we might want to optimize the ioctls accordingly so that we
// don't always copy a 250 * sizeof(streamID) sized array when there's only one
// or a few streams.
//
// For that reason, UVM_MAX_STREAMS_PER_IOCTL_CALL is set to 32.
//
// If the higher-level (uvm.h) call requires more streams to be stopped than
// this value, then multiple ioctl calls should be made.
//
#define UVM_MAX_STREAMS_PER_IOCTL_CALL 32
//
// UvmSetStreamStopped
//
#define UVM_SET_STREAM_STOPPED UVM_IOCTL_BASE(7)
typedef struct
{
UvmStream streamIdArray[UVM_MAX_STREAMS_PER_IOCTL_CALL] NV_ALIGN_BYTES(8); // IN
NvU64 nStreams NV_ALIGN_BYTES(8); // IN
NV_STATUS rmStatus; // OUT
} UVM_SET_STREAM_STOPPED_PARAMS;
//
// UvmCallTestFunction
//
#define UVM_RUN_TEST UVM_IOCTL_BASE(9)
typedef struct
{
NvProcessorUuid gpuUuid; // IN
NvU32 test; // IN
struct
{
NvProcessorUuid peerGpuUuid; // IN
NvU32 peerId; // IN
} multiGpu;
NV_STATUS rmStatus; // OUT
} UVM_RUN_TEST_PARAMS;
//
// This is a magic offset for mmap. Any mapping of an offset above this
// threshold will be treated as a counters mapping, not as an allocation
// mapping. Since allocation offsets must be identical to the virtual address
// of the mapping, this threshold has to be an offset that cannot be
// a valid virtual address.
//
#if defined(__linux__)
#if defined(NV_64_BITS)
#define UVM_EVENTS_OFFSET_BASE (1UL << 63)
#define UVM_COUNTERS_OFFSET_BASE (1UL << 62)
#else
#define UVM_EVENTS_OFFSET_BASE (1UL << 31)
#define UVM_COUNTERS_OFFSET_BASE (1UL << 30)
#endif
#endif // defined(__linux___)
//
// UvmAddSession
//
#define UVM_ADD_SESSION UVM_IOCTL_BASE(10)
typedef struct
{
NvU32 pidTarget; // IN
#ifdef __linux__
NvP64 countersBaseAddress NV_ALIGN_BYTES(8); // IN
NvS32 sessionIndex; // OUT (session index that got added)
#endif
NV_STATUS rmStatus; // OUT
} UVM_ADD_SESSION_PARAMS;
//
// UvmRemoveSession
//
#define UVM_REMOVE_SESSION UVM_IOCTL_BASE(11)
typedef struct
{
#ifdef __linux__
NvS32 sessionIndex; // IN (session index to be removed)
#endif
NV_STATUS rmStatus; // OUT
} UVM_REMOVE_SESSION_PARAMS;
#define UVM_MAX_COUNTERS_PER_IOCTL_CALL 32
//
// UvmEnableCounters
//
#define UVM_ENABLE_COUNTERS UVM_IOCTL_BASE(12)
typedef struct
{
#ifdef __linux__
NvS32 sessionIndex; // IN
#endif
UvmCounterConfig config[UVM_MAX_COUNTERS_PER_IOCTL_CALL]; // IN
NvU32 count; // IN
NV_STATUS rmStatus; // OUT
} UVM_ENABLE_COUNTERS_PARAMS;
//
// UvmMapCounter
//
#define UVM_MAP_COUNTER UVM_IOCTL_BASE(13)
typedef struct
{
#ifdef __linux__
NvS32 sessionIndex; // IN
#endif
NvU32 scope; // IN (UvmCounterScope)
NvU32 counterName; // IN (UvmCounterName)
NvProcessorUuid gpuUuid; // IN
NvP64 addr NV_ALIGN_BYTES(8); // OUT
NV_STATUS rmStatus; // OUT
} UVM_MAP_COUNTER_PARAMS;
//
// UvmCreateEventQueue
//
#define UVM_CREATE_EVENT_QUEUE UVM_IOCTL_BASE(14)
typedef struct
{
#ifdef __linux__
NvS32 sessionIndex; // IN
#endif
NvU32 eventQueueIndex; // OUT
NvU64 queueSize NV_ALIGN_BYTES(8); // IN
NvU64 notificationCount NV_ALIGN_BYTES(8); // IN
#if defined(WIN32) || defined(WIN64)
NvU64 notificationHandle NV_ALIGN_BYTES(8); // IN
#endif
NvU32 timeStampType; // IN (UvmEventTimeStampType)
NV_STATUS rmStatus; // OUT
} UVM_CREATE_EVENT_QUEUE_PARAMS;
//
// UvmRemoveEventQueue
//
#define UVM_REMOVE_EVENT_QUEUE UVM_IOCTL_BASE(15)
typedef struct
{
#ifdef __linux__
NvS32 sessionIndex; // IN
#endif
NvU32 eventQueueIndex; // IN
NV_STATUS rmStatus; // OUT
} UVM_REMOVE_EVENT_QUEUE_PARAMS;
//
// UvmMapEventQueue
//
#define UVM_MAP_EVENT_QUEUE UVM_IOCTL_BASE(16)
typedef struct
{
#ifdef __linux__
NvS32 sessionIndex; // IN
#endif
NvU32 eventQueueIndex; // IN
NvP64 userRODataAddr NV_ALIGN_BYTES(8); // IN
NvP64 userRWDataAddr NV_ALIGN_BYTES(8); // IN
NvP64 readIndexAddr NV_ALIGN_BYTES(8); // OUT
NvP64 writeIndexAddr NV_ALIGN_BYTES(8); // OUT
NvP64 queueBufferAddr NV_ALIGN_BYTES(8); // OUT
NV_STATUS rmStatus; // OUT
} UVM_MAP_EVENT_QUEUE_PARAMS;
//
// UvmEnableEvent
//
#define UVM_EVENT_CTRL UVM_IOCTL_BASE(17)
typedef struct
{
#ifdef __linux__
NvS32 sessionIndex; // IN
#endif
NvU32 eventQueueIndex; // IN
NvS32 eventType; // IN
NvU32 enable; // IN
NV_STATUS rmStatus; // OUT
} UVM_EVENT_CTRL_PARAMS;
//
// UvmRegisterMpsServer
//
#define UVM_REGISTER_MPS_SERVER UVM_IOCTL_BASE(18)
typedef struct
{
NvProcessorUuid gpuUuidArray[UVM_MAX_GPUS]; // IN
NvU32 numGpus; // IN
NvU64 serverId NV_ALIGN_BYTES(8); // OUT
NV_STATUS rmStatus; // OUT
} UVM_REGISTER_MPS_SERVER_PARAMS;
//
// UvmRegisterMpsClient
//
#define UVM_REGISTER_MPS_CLIENT UVM_IOCTL_BASE(19)
typedef struct
{
NvU64 serverId NV_ALIGN_BYTES(8); // IN
NV_STATUS rmStatus; // OUT
} UVM_REGISTER_MPS_CLIENT_PARAMS;
//
// UvmEventGetGpuUuidTable
//
#define UVM_GET_GPU_UUID_TABLE UVM_IOCTL_BASE(20)
typedef struct
{
NvProcessorUuid gpuUuidArray[UVM_MAX_GPUS]; // OUT
NvU32 validCount; // OUT
NV_STATUS rmStatus; // OUT
} UVM_GET_GPU_UUID_TABLE_PARAMS;
#if defined(WIN32) || defined(WIN64)
//
// UvmRegionSetBacking
//
#define UVM_REGION_SET_BACKING UVM_IOCTL_BASE(21)
typedef struct
{
NvProcessorUuid gpuUuid; // IN
NvU32 hAllocation; // IN
NvP64 vaAddr NV_ALIGN_BYTES(8); // IN
NvU64 regionLength NV_ALIGN_BYTES(8); // IN
NV_STATUS rmStatus; // OUT
} UVM_REGION_SET_BACKING_PARAMS;
//
// UvmRegionUnsetBacking
//
#define UVM_REGION_UNSET_BACKING UVM_IOCTL_BASE(22)
typedef struct
{
NvP64 vaAddr NV_ALIGN_BYTES(8); // IN
NvU64 regionLength NV_ALIGN_BYTES(8); // IN
NV_STATUS rmStatus; // OUT
} UVM_REGION_UNSET_BACKING_PARAMS;
#endif
#define UVM_CREATE_RANGE_GROUP UVM_IOCTL_BASE(23)
typedef struct
{
NvU64 rangeGroupId NV_ALIGN_BYTES(8); // OUT
NV_STATUS rmStatus; // OUT
} UVM_CREATE_RANGE_GROUP_PARAMS;
#define UVM_DESTROY_RANGE_GROUP UVM_IOCTL_BASE(24)
typedef struct
{
NvU64 rangeGroupId NV_ALIGN_BYTES(8); // IN
NV_STATUS rmStatus; // OUT
} UVM_DESTROY_RANGE_GROUP_PARAMS;
//
// UvmRegisterGpuVaSpace
//
#define UVM_REGISTER_GPU_VASPACE UVM_IOCTL_BASE(25)
typedef struct
{
NvProcessorUuid gpuUuid; // IN
NvS32 rmCtrlFd; // IN
NvHandle hClient; // IN
NvHandle hVaSpace; // IN
NV_STATUS rmStatus; // OUT
} UVM_REGISTER_GPU_VASPACE_PARAMS;
//
// UvmUnregisterGpuVaSpace
//
#define UVM_UNREGISTER_GPU_VASPACE UVM_IOCTL_BASE(26)
typedef struct
{
NvProcessorUuid gpuUuid; // IN
NV_STATUS rmStatus; // OUT
} UVM_UNREGISTER_GPU_VASPACE_PARAMS;
//
// UvmRegisterChannel
//
#define UVM_REGISTER_CHANNEL UVM_IOCTL_BASE(27)
typedef struct
{
NvProcessorUuid gpuUuid; // IN
NvS32 rmCtrlFd; // IN
NvHandle hClient; // IN
NvHandle hChannel; // IN
NvU64 base NV_ALIGN_BYTES(8); // IN
NvU64 length NV_ALIGN_BYTES(8); // IN
NV_STATUS rmStatus; // OUT
} UVM_REGISTER_CHANNEL_PARAMS;
//
// UvmUnregisterChannel
//
#define UVM_UNREGISTER_CHANNEL UVM_IOCTL_BASE(28)
typedef struct
{
NvProcessorUuid gpuUuid; // IN
NvHandle hClient; // IN
NvHandle hChannel; // IN
NV_STATUS rmStatus; // OUT
} UVM_UNREGISTER_CHANNEL_PARAMS;
//
// UvmEnablePeerAccess
//
#define UVM_ENABLE_PEER_ACCESS UVM_IOCTL_BASE(29)
typedef struct
{
NvProcessorUuid gpuUuidA; // IN
NvProcessorUuid gpuUuidB; // IN
NV_STATUS rmStatus; // OUT
} UVM_ENABLE_PEER_ACCESS_PARAMS;
//
// UvmDisablePeerAccess
//
#define UVM_DISABLE_PEER_ACCESS UVM_IOCTL_BASE(30)
typedef struct
{
NvProcessorUuid gpuUuidA; // IN
NvProcessorUuid gpuUuidB; // IN
NV_STATUS rmStatus; // OUT
} UVM_DISABLE_PEER_ACCESS_PARAMS;
//
// UvmSetRangeGroup
//
#define UVM_SET_RANGE_GROUP UVM_IOCTL_BASE(31)
typedef struct
{
NvU64 rangeGroupId NV_ALIGN_BYTES(8); // IN
NvU64 requestedBase NV_ALIGN_BYTES(8); // IN
NvU64 length NV_ALIGN_BYTES(8); // IN
NV_STATUS rmStatus; // OUT
} UVM_SET_RANGE_GROUP_PARAMS;
//
// UvmMapExternalAllocation
//
#define UVM_MAP_EXTERNAL_ALLOCATION UVM_IOCTL_BASE(33)
typedef struct
{
NvU64 base NV_ALIGN_BYTES(8); // IN
NvU64 length NV_ALIGN_BYTES(8); // IN
NvU64 offset NV_ALIGN_BYTES(8); // IN
UvmGpuMappingAttributes perGpuAttributes[UVM_MAX_GPUS]; // IN
NvU64 gpuAttributesCount NV_ALIGN_BYTES(8); // IN
NvS32 rmCtrlFd; // IN
NvU32 hClient; // IN
NvU32 hMemory; // IN
NV_STATUS rmStatus; // OUT
} UVM_MAP_EXTERNAL_ALLOCATION_PARAMS;
//
// UvmFree
//
#define UVM_FREE UVM_IOCTL_BASE(34)
typedef struct
{
NvU64 base NV_ALIGN_BYTES(8); // IN
NvU64 length NV_ALIGN_BYTES(8); // IN
NV_STATUS rmStatus; // OUT
} UVM_FREE_PARAMS;
//
// UvmMemMap
//
#define UVM_MEM_MAP UVM_IOCTL_BASE(35)
typedef struct
{
NvP64 regionBase NV_ALIGN_BYTES(8); // IN
NvU64 regionLength NV_ALIGN_BYTES(8); // IN
NV_STATUS rmStatus; // OUT
} UVM_MEM_MAP_PARAMS;
//
// UvmDebugAccessMemory
//
#define UVM_DEBUG_ACCESS_MEMORY UVM_IOCTL_BASE(36)
typedef struct
{
#ifdef __linux__
NvS32 sessionIndex; // IN
#endif
NvU64 baseAddress NV_ALIGN_BYTES(8); // IN
NvU64 sizeInBytes NV_ALIGN_BYTES(8); // IN
NvU32 accessType; // IN (UvmDebugAccessType)
NvU64 buffer NV_ALIGN_BYTES(8); // IN/OUT
NvBool isBitmaskSet; // OUT
NvU64 bitmask NV_ALIGN_BYTES(8); // IN/OUT
NV_STATUS rmStatus; // OUT
} UVM_DEBUG_ACCESS_MEMORY_PARAMS;
//
// UvmRegisterGpu
//
#define UVM_REGISTER_GPU UVM_IOCTL_BASE(37)
typedef struct
{
NvProcessorUuid gpu_uuid; // IN
NvBool numaEnabled; // OUT
NvS32 numaNodeId; // OUT
NvS32 rmCtrlFd; // IN
NvHandle hClient; // IN
NvHandle hSmcPartRef; // IN
NV_STATUS rmStatus; // OUT
} UVM_REGISTER_GPU_PARAMS;
//
// UvmUnregisterGpu
//
#define UVM_UNREGISTER_GPU UVM_IOCTL_BASE(38)
typedef struct
{
NvProcessorUuid gpu_uuid; // IN
NV_STATUS rmStatus; // OUT
} UVM_UNREGISTER_GPU_PARAMS;
#define UVM_PAGEABLE_MEM_ACCESS UVM_IOCTL_BASE(39)
typedef struct
{
NvBool pageableMemAccess; // OUT
NV_STATUS rmStatus; // OUT
} UVM_PAGEABLE_MEM_ACCESS_PARAMS;
//
// Due to limitations in how much we want to send per ioctl call, the numGroupIds
// member must be less than or equal to about 250. That's an upper limit.
//
// However, from a typical user-space driver's point of view (for example, the
// CUDA driver), a vast majority of the time, we expect there to be only one
// range group passed in. The second most common case is something like atmost 32
// range groups being passed in. The cases where there are more than 32 range
// groups are the most rare. So we might want to optimize the ioctls accordingly
// so that we don't always copy a 250 * sizeof(NvU64) sized array when there's
// only one or a few range groups.
//
// For that reason, UVM_MAX_RANGE_GROUPS_PER_IOCTL_CALL is set to 32.
//
// If the higher-level (uvm.h) call requires more range groups than
// this value, then multiple ioctl calls should be made.
//
#define UVM_MAX_RANGE_GROUPS_PER_IOCTL_CALL 32
//
// UvmPreventMigrationRangeGroups
//
#define UVM_PREVENT_MIGRATION_RANGE_GROUPS UVM_IOCTL_BASE(40)
typedef struct
{
NvU64 rangeGroupIds[UVM_MAX_RANGE_GROUPS_PER_IOCTL_CALL] NV_ALIGN_BYTES(8); // IN
NvU64 numGroupIds NV_ALIGN_BYTES(8); // IN
NV_STATUS rmStatus; // OUT
} UVM_PREVENT_MIGRATION_RANGE_GROUPS_PARAMS;
//
// UvmAllowMigrationRangeGroups
//
#define UVM_ALLOW_MIGRATION_RANGE_GROUPS UVM_IOCTL_BASE(41)
typedef struct
{
NvU64 rangeGroupIds[UVM_MAX_RANGE_GROUPS_PER_IOCTL_CALL] NV_ALIGN_BYTES(8); // IN
NvU64 numGroupIds NV_ALIGN_BYTES(8); // IN
NV_STATUS rmStatus; // OUT
} UVM_ALLOW_MIGRATION_RANGE_GROUPS_PARAMS;
//
// UvmSetPreferredLocation
//
#define UVM_SET_PREFERRED_LOCATION UVM_IOCTL_BASE(42)
typedef struct
{
NvU64 requestedBase NV_ALIGN_BYTES(8); // IN
NvU64 length NV_ALIGN_BYTES(8); // IN
NvProcessorUuid preferredLocation; // IN
NV_STATUS rmStatus; // OUT
} UVM_SET_PREFERRED_LOCATION_PARAMS;
//
// UvmUnsetPreferredLocation
//
#define UVM_UNSET_PREFERRED_LOCATION UVM_IOCTL_BASE(43)
typedef struct
{
NvU64 requestedBase NV_ALIGN_BYTES(8); // IN
NvU64 length NV_ALIGN_BYTES(8); // IN
NV_STATUS rmStatus; // OUT
} UVM_UNSET_PREFERRED_LOCATION_PARAMS;
//
// UvmEnableReadDuplication
//
#define UVM_ENABLE_READ_DUPLICATION UVM_IOCTL_BASE(44)
typedef struct
{
NvU64 requestedBase NV_ALIGN_BYTES(8); // IN
NvU64 length NV_ALIGN_BYTES(8); // IN
NV_STATUS rmStatus; // OUT
} UVM_ENABLE_READ_DUPLICATION_PARAMS;
//
// UvmDisableReadDuplication
//
#define UVM_DISABLE_READ_DUPLICATION UVM_IOCTL_BASE(45)
typedef struct
{
NvU64 requestedBase NV_ALIGN_BYTES(8); // IN
NvU64 length NV_ALIGN_BYTES(8); // IN
NV_STATUS rmStatus; // OUT
} UVM_DISABLE_READ_DUPLICATION_PARAMS;
//
// UvmSetAccessedBy
//
#define UVM_SET_ACCESSED_BY UVM_IOCTL_BASE(46)
typedef struct
{
NvU64 requestedBase NV_ALIGN_BYTES(8); // IN
NvU64 length NV_ALIGN_BYTES(8); // IN
NvProcessorUuid accessedByUuid; // IN
NV_STATUS rmStatus; // OUT
} UVM_SET_ACCESSED_BY_PARAMS;
//
// UvmUnsetAccessedBy
//
#define UVM_UNSET_ACCESSED_BY UVM_IOCTL_BASE(47)
typedef struct
{
NvU64 requestedBase NV_ALIGN_BYTES(8); // IN
NvU64 length NV_ALIGN_BYTES(8); // IN
NvProcessorUuid accessedByUuid; // IN
NV_STATUS rmStatus; // OUT
} UVM_UNSET_ACCESSED_BY_PARAMS;
// For managed allocations, UVM_MIGRATE implements the behavior described in
// UvmMigrate. If the input virtual range corresponds to system-allocated
// pageable memory, and the GPUs in the system support transparent access to
// pageable memory, the scheme is a bit more elaborate, potentially with
// several transitions betwen user and kernel spaces:
//
// 1) UVM_MIGRATE with the range base address and size. This will migrate
// anonymous vmas until:
// a) It finds a file-backed vma or no GPUs are registered in the VA space
// so no GPU can drive the copy. It will try to populate the vma using
// get_user_pages and return NV_WARN_NOTHING_TO_DO.
// b) It fails to allocate memory on the destination CPU node. It will return
// NV_ERR_MORE_PROCESSING_REQUIRED.
// c) It fails to populate pages directly on the destination GPU. It will try
// to populate the vma using get_user_pages and return.
// d) The full input range is migrated (or empty), this call will release
// the semaphore before returning.
// 2) The user-mode needs to handle the following error codes:
// a) NV_WARN_NOTHING_TO_DO: use move_pages to migrate pages for the VA
// range corresponding to the vma that couldn't be migrated in kernel
// mode. Then, it processes the remainder of the range, starting after
// that vma.
// b) NV_ERR_MORE_PROCESSING_REQUIRED: choose a different CPU NUMA node,
// trying to enforce the NUMA policies of the thread and retry the
// ioctl. If there are no more CPU NUMA nodes to try, try to populate
// the remainder of the range anywhere using the UVM_POPULATE_PAGEABLE
// ioctl.
// c) NV_OK: success. This only guarantees that pages were populated, not
// that they moved to the requested destination.
// 3) For cases 2.a) and 2.b) Goto 1
//
// If UVM_MIGRATE_FLAG_ASYNC is 0, the ioctl won't return until the migration is
// done and all mappings are updated, subject to the special rules for pageable
// memory described above. semaphoreAddress must be 0. semaphorePayload is
// ignored.
//
// If UVM_MIGRATE_FLAG_ASYNC is 1, the ioctl may return before the migration is
// complete. If semaphoreAddress is 0, semaphorePayload is ignored and no
// notification will be given on completion. If semaphoreAddress is non-zero
// and the returned error code is NV_OK, semaphorePayload will be written to
// semaphoreAddress once the migration is complete.
#define UVM_MIGRATE_FLAG_ASYNC 0x00000001
// When the migration destination is the CPU, skip the step which creates new
// virtual mappings on the CPU. Creating CPU mappings must wait for the
// migration to complete, so skipping this step allows the migration to be
// fully asynchronous. This flag is ignored for pageable migrations if the GPUs
// in the system support transparent access to pageable memory.
//
// The UVM driver must have builtin tests enabled for the API to use this flag.
#define UVM_MIGRATE_FLAG_SKIP_CPU_MAP 0x00000002
// By default UVM_MIGRATE returns an error if the destination UUID is a GPU
// without a registered GPU VA space. Setting this flag skips that check, so the
// destination GPU only needs to have been registered.
//
// This can be used in tests to trigger migrations of physical memory without
// the overhead of GPU PTE mappings.
//
// The UVM driver must have builtin tests enabled for the API to use this flag.
#define UVM_MIGRATE_FLAG_NO_GPU_VA_SPACE 0x00000004
#define UVM_MIGRATE_FLAGS_TEST_ALL (UVM_MIGRATE_FLAG_SKIP_CPU_MAP | \
UVM_MIGRATE_FLAG_NO_GPU_VA_SPACE)
#define UVM_MIGRATE_FLAGS_ALL (UVM_MIGRATE_FLAG_ASYNC | \
UVM_MIGRATE_FLAGS_TEST_ALL)
// For pageable migrations, cpuNumaNode is used as the destination NUMA node if
// destinationUuid is the CPU.
//
// If NV_WARN_NOTHING_TO_DO is returned, user-space is responsible for
// completing the migration of the VA range described by userSpaceStart and
// userSpaceLength using move_pages.
//
// If NV_ERR_MORE_PROCESSING_REQUIRED is returned, user-space is responsible
// for re-trying with a different cpuNumaNode, starting at userSpaceStart.
#define UVM_MIGRATE UVM_IOCTL_BASE(51)
typedef struct
{
NvU64 base NV_ALIGN_BYTES(8); // IN
NvU64 length NV_ALIGN_BYTES(8); // IN
NvProcessorUuid destinationUuid; // IN
NvU32 flags; // IN
NvU64 semaphoreAddress NV_ALIGN_BYTES(8); // IN
NvU32 semaphorePayload; // IN
NvU32 cpuNumaNode; // IN
NvU64 userSpaceStart NV_ALIGN_BYTES(8); // OUT
NvU64 userSpaceLength NV_ALIGN_BYTES(8); // OUT
NV_STATUS rmStatus; // OUT
} UVM_MIGRATE_PARAMS;
#define UVM_MIGRATE_RANGE_GROUP UVM_IOCTL_BASE(53)
typedef struct
{
NvU64 rangeGroupId NV_ALIGN_BYTES(8); // IN
NvProcessorUuid destinationUuid; // IN
NV_STATUS rmStatus; // OUT
} UVM_MIGRATE_RANGE_GROUP_PARAMS;
//
// UvmEnableSystemWideAtomics
//
#define UVM_ENABLE_SYSTEM_WIDE_ATOMICS UVM_IOCTL_BASE(54)
typedef struct
{
NvProcessorUuid gpu_uuid; // IN
NV_STATUS rmStatus; // OUT
} UVM_ENABLE_SYSTEM_WIDE_ATOMICS_PARAMS;
//
// UvmDisableSystemWideAtomics
//
#define UVM_DISABLE_SYSTEM_WIDE_ATOMICS UVM_IOCTL_BASE(55)
typedef struct
{
NvProcessorUuid gpu_uuid; // IN
NV_STATUS rmStatus; // OUT
} UVM_DISABLE_SYSTEM_WIDE_ATOMICS_PARAMS;
//
// Initialize any tracker object such as a queue or counter
// UvmToolsCreateEventQueue, UvmToolsCreateProcessAggregateCounters, UvmToolsCreateProcessorCounters
//
#define UVM_TOOLS_INIT_EVENT_TRACKER UVM_IOCTL_BASE(56)
typedef struct
{
NvU64 queueBuffer NV_ALIGN_BYTES(8); // IN
NvU64 queueBufferSize NV_ALIGN_BYTES(8); // IN
NvU64 controlBuffer NV_ALIGN_BYTES(8); // IN
NvProcessorUuid processor; // IN
NvU32 allProcessors; // IN
NvU32 uvmFd; // IN
NV_STATUS rmStatus; // OUT
} UVM_TOOLS_INIT_EVENT_TRACKER_PARAMS;
//
// UvmToolsSetNotificationThreshold
//
#define UVM_TOOLS_SET_NOTIFICATION_THRESHOLD UVM_IOCTL_BASE(57)
typedef struct
{
NvU32 notificationThreshold; // IN
NV_STATUS rmStatus; // OUT
} UVM_TOOLS_SET_NOTIFICATION_THRESHOLD_PARAMS;
//
// UvmToolsEventQueueEnableEvents
//
#define UVM_TOOLS_EVENT_QUEUE_ENABLE_EVENTS UVM_IOCTL_BASE(58)
typedef struct
{
NvU64 eventTypeFlags NV_ALIGN_BYTES(8); // IN
NV_STATUS rmStatus; // OUT
} UVM_TOOLS_EVENT_QUEUE_ENABLE_EVENTS_PARAMS;
//
// UvmToolsEventQueueDisableEvents
//
#define UVM_TOOLS_EVENT_QUEUE_DISABLE_EVENTS UVM_IOCTL_BASE(59)
typedef struct
{
NvU64 eventTypeFlags NV_ALIGN_BYTES(8); // IN
NV_STATUS rmStatus; // OUT
} UVM_TOOLS_EVENT_QUEUE_DISABLE_EVENTS_PARAMS;
//
// UvmToolsEnableCounters
//
#define UVM_TOOLS_ENABLE_COUNTERS UVM_IOCTL_BASE(60)
typedef struct
{
NvU64 counterTypeFlags NV_ALIGN_BYTES(8); // IN
NV_STATUS rmStatus; // OUT
} UVM_TOOLS_ENABLE_COUNTERS_PARAMS;
//
// UvmToolsDisableCounters
//
#define UVM_TOOLS_DISABLE_COUNTERS UVM_IOCTL_BASE(61)
typedef struct
{
NvU64 counterTypeFlags NV_ALIGN_BYTES(8); // IN
NV_STATUS rmStatus; // OUT
} UVM_TOOLS_DISABLE_COUNTERS_PARAMS;
//
// UvmToolsReadProcessMemory
//
#define UVM_TOOLS_READ_PROCESS_MEMORY UVM_IOCTL_BASE(62)
typedef struct
{
NvU64 buffer NV_ALIGN_BYTES(8); // IN
NvU64 size NV_ALIGN_BYTES(8); // IN
NvU64 targetVa NV_ALIGN_BYTES(8); // IN
NvU64 bytesRead NV_ALIGN_BYTES(8); // OUT
NV_STATUS rmStatus; // OUT
} UVM_TOOLS_READ_PROCESS_MEMORY_PARAMS;
//
// UvmToolsWriteProcessMemory
//
#define UVM_TOOLS_WRITE_PROCESS_MEMORY UVM_IOCTL_BASE(63)
typedef struct
{
NvU64 buffer NV_ALIGN_BYTES(8); // IN
NvU64 size NV_ALIGN_BYTES(8); // IN
NvU64 targetVa NV_ALIGN_BYTES(8); // IN
NvU64 bytesWritten NV_ALIGN_BYTES(8); // OUT
NV_STATUS rmStatus; // OUT
} UVM_TOOLS_WRITE_PROCESS_MEMORY_PARAMS;
//
// UvmToolsGetProcessorUuidTable
//
#define UVM_TOOLS_GET_PROCESSOR_UUID_TABLE UVM_IOCTL_BASE(64)
typedef struct
{
NvU64 tablePtr NV_ALIGN_BYTES(8); // IN
NvU32 count; // IN/OUT
NV_STATUS rmStatus; // OUT
} UVM_TOOLS_GET_PROCESSOR_UUID_TABLE_PARAMS;
//
// UvmMapDynamicParallelismRegion
//
#define UVM_MAP_DYNAMIC_PARALLELISM_REGION UVM_IOCTL_BASE(65)
typedef struct
{
NvU64 base NV_ALIGN_BYTES(8); // IN
NvU64 length NV_ALIGN_BYTES(8); // IN
NvProcessorUuid gpuUuid; // IN
NV_STATUS rmStatus; // OUT
} UVM_MAP_DYNAMIC_PARALLELISM_REGION_PARAMS;
//
// UvmUnmapExternal
//
#define UVM_UNMAP_EXTERNAL UVM_IOCTL_BASE(66)
typedef struct
{
NvU64 base NV_ALIGN_BYTES(8); // IN
NvU64 length NV_ALIGN_BYTES(8); // IN
NvProcessorUuid gpuUuid; // IN
NV_STATUS rmStatus; // OUT
} UVM_UNMAP_EXTERNAL_PARAMS;
//
// UvmToolsFlushEvents
//
#define UVM_TOOLS_FLUSH_EVENTS UVM_IOCTL_BASE(67)
typedef struct
{
NV_STATUS rmStatus; // OUT
} UVM_TOOLS_FLUSH_EVENTS_PARAMS;
//
// UvmAllocSemaphorePool
//
#define UVM_ALLOC_SEMAPHORE_POOL UVM_IOCTL_BASE(68)
typedef struct
{
NvU64 base NV_ALIGN_BYTES(8); // IN
NvU64 length NV_ALIGN_BYTES(8); // IN
UvmGpuMappingAttributes perGpuAttributes[UVM_MAX_GPUS]; // IN
NvU64 gpuAttributesCount NV_ALIGN_BYTES(8); // IN
NV_STATUS rmStatus; // OUT
} UVM_ALLOC_SEMAPHORE_POOL_PARAMS;
//
// UvmCleanUpZombieResources
//
#define UVM_CLEAN_UP_ZOMBIE_RESOURCES UVM_IOCTL_BASE(69)
typedef struct
{
NV_STATUS rmStatus; // OUT
} UVM_CLEAN_UP_ZOMBIE_RESOURCES_PARAMS;
//
// UvmIsPageableMemoryAccessSupportedOnGpu
//
#define UVM_PAGEABLE_MEM_ACCESS_ON_GPU UVM_IOCTL_BASE(70)
typedef struct
{
NvProcessorUuid gpu_uuid; // IN
NvBool pageableMemAccess; // OUT
NV_STATUS rmStatus; // OUT
} UVM_PAGEABLE_MEM_ACCESS_ON_GPU_PARAMS;
//
// UvmPopulatePageable
//
#define UVM_POPULATE_PAGEABLE UVM_IOCTL_BASE(71)
// Allow population of managed ranges.
//
// The UVM driver must have builtin tests enabled for the API to use the
// following two flags.
#define UVM_POPULATE_PAGEABLE_FLAG_ALLOW_MANAGED 0x00000001
// By default UVM_POPULATE_PAGEABLE returns an error if the destination vma
// does not have read permission. This flag skips that check.
#define UVM_POPULATE_PAGEABLE_FLAG_SKIP_PROT_CHECK 0x00000002
#define UVM_POPULATE_PAGEABLE_FLAGS_TEST_ALL (UVM_POPULATE_PAGEABLE_FLAG_ALLOW_MANAGED | \
UVM_POPULATE_PAGEABLE_FLAG_SKIP_PROT_CHECK)
#define UVM_POPULATE_PAGEABLE_FLAGS_ALL UVM_POPULATE_PAGEABLE_FLAGS_TEST_ALL
typedef struct
{
NvU64 base NV_ALIGN_BYTES(8); // IN
NvU64 length NV_ALIGN_BYTES(8); // IN
NvU32 flags; // IN
NV_STATUS rmStatus; // OUT
} UVM_POPULATE_PAGEABLE_PARAMS;
//
// UvmValidateVaRange
//
#define UVM_VALIDATE_VA_RANGE UVM_IOCTL_BASE(72)
typedef struct
{
NvU64 base NV_ALIGN_BYTES(8); // IN
NvU64 length NV_ALIGN_BYTES(8); // IN
NV_STATUS rmStatus; // OUT
} UVM_VALIDATE_VA_RANGE_PARAMS;
#define UVM_CREATE_EXTERNAL_RANGE UVM_IOCTL_BASE(73)
typedef struct
{
NvU64 base NV_ALIGN_BYTES(8); // IN
NvU64 length NV_ALIGN_BYTES(8); // IN
NV_STATUS rmStatus; // OUT
} UVM_CREATE_EXTERNAL_RANGE_PARAMS;
#define UVM_MAP_EXTERNAL_SPARSE UVM_IOCTL_BASE(74)
typedef struct
{
NvU64 base NV_ALIGN_BYTES(8); // IN
NvU64 length NV_ALIGN_BYTES(8); // IN
NvProcessorUuid gpuUuid; // IN
NV_STATUS rmStatus; // OUT
} UVM_MAP_EXTERNAL_SPARSE_PARAMS;
//
// Used to initialise a secondary UVM file-descriptor which holds a
// reference on the memory map to prevent it being torn down without
// first notifying UVM. This is achieved by preventing mmap() calls on
// the secondary file-descriptor so that on process exit
// uvm_mm_release() will be called while the memory map is present
// such that UVM can cleanly shutdown the GPU by handling faults
// instead of cancelling them.
//
// This ioctl must be called after the primary file-descriptor has
// been initialised with the UVM_INITIALIZE ioctl. The primary FD
// should be passed in the uvmFd field and the UVM_MM_INITIALIZE ioctl
// will hold a reference on the primary FD. Therefore uvm_release() is
// guaranteed to be called after uvm_mm_release().
//
// Once this file-descriptor has been closed the UVM context is
// effectively dead and subsequent operations requiring a memory map
// will fail. Calling UVM_MM_INITIALIZE on a context that has already
// been initialized via any FD will return NV_ERR_INVALID_STATE.
//
// Calling this with a non-UVM file-descriptor in uvmFd will return
// NV_ERR_INVALID_ARGUMENT. Calling this on the same file-descriptor
// as UVM_INITIALIZE or more than once on the same FD will return
// NV_ERR_IN_USE.
//
// Not all platforms require this secondary file-descriptor. On those
// platforms NV_WARN_NOTHING_TO_DO will be returned and users may
// close the file-descriptor at anytime.
#define UVM_MM_INITIALIZE UVM_IOCTL_BASE(75)
typedef struct
{
NvS32 uvmFd; // IN
NV_STATUS rmStatus; // OUT
} UVM_MM_INITIALIZE_PARAMS;
//
// Temporary ioctls which should be removed before UVM 8 release
// Number backwards from 2047 - highest custom ioctl function number
// windows can handle.
//
//
// UvmIs8Supported
//
#define UVM_IS_8_SUPPORTED UVM_IOCTL_BASE(2047)
typedef struct
{
NvU32 is8Supported; // OUT
NV_STATUS rmStatus; // OUT
} UVM_IS_8_SUPPORTED_PARAMS;
#ifdef __cplusplus
}
#endif
#endif // _UVM_IOCTL_H
Reference in New Issue View Git Blame Copy Permalink