/* * SPDX-FileCopyrightText: Copyright (c) 2013-2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved. * SPDX-License-Identifier: MIT * * 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. */ // // This file provides the interface that RM exposes to UVM. // #ifndef _NV_UVM_INTERFACE_H_ #define _NV_UVM_INTERFACE_H_ // Forward references, to break circular header file dependencies: struct UvmOpsUvmEvents; #if defined(NVIDIA_UVM_ENABLED) // We are in the UVM build system, for a Linux target. #include "uvm_linux.h" #else // We are in the RM build system, for a Linux target: #include "nv-linux.h" #endif // NVIDIA_UVM_ENABLED #include "nvgputypes.h" #include "nvstatus.h" #include "nv_uvm_types.h" // Define the type here as it's Linux specific, used only by the Linux specific // nvUvmInterfaceRegisterGpu() API. typedef struct { struct pci_dev *pci_dev; // DMA addressable range of the device, mirrors fields in nv_state_t. NvU64 dma_addressable_start; NvU64 dma_addressable_limit; } UvmGpuPlatformInfo; /******************************************************************************* nvUvmInterfaceRegisterGpu Registers the GPU with the provided UUID for use. A GPU must be registered before its UUID can be used with any other API. This call is ref-counted so every nvUvmInterfaceRegisterGpu must be paired with a corresponding nvUvmInterfaceUnregisterGpu. You don't need to call nvUvmInterfaceSessionCreate before calling this. Error codes: NV_ERR_GPU_UUID_NOT_FOUND NV_ERR_NO_MEMORY NV_ERR_GENERIC */ NV_STATUS nvUvmInterfaceRegisterGpu(const NvProcessorUuid *gpuUuid, UvmGpuPlatformInfo *gpuInfo); /******************************************************************************* nvUvmInterfaceUnregisterGpu Unregisters the GPU with the provided UUID. This drops the ref count from nvUvmInterfaceRegisterGpu. Once the reference count goes to 0 the device may no longer be accessible until the next nvUvmInterfaceRegisterGpu call. No automatic resource freeing is performed, so only make the last unregister call after destroying all your allocations associated with that UUID (such as those from nvUvmInterfaceAddressSpaceCreate). If the UUID is not found, no operation is performed. */ void nvUvmInterfaceUnregisterGpu(const NvProcessorUuid *gpuUuid); /******************************************************************************* nvUvmInterfaceSessionCreate TODO: Creates session object. All allocations are tied to the session. The platformInfo parameter is filled by the callee with miscellaneous system information. Refer to the UvmPlatformInfo struct for details. Error codes: NV_ERR_GENERIC NV_ERR_NO_MEMORY */ NV_STATUS nvUvmInterfaceSessionCreate(uvmGpuSessionHandle *session, UvmPlatformInfo *platformInfo); /******************************************************************************* nvUvmInterfaceSessionDestroy Destroys a session object. All allocations are tied to the session will be destroyed. Error codes: NV_ERR_GENERIC NV_ERR_NO_MEMORY */ NV_STATUS nvUvmInterfaceSessionDestroy(uvmGpuSessionHandle session); /******************************************************************************* nvUvmInterfaceDeviceCreate Creates a device object under the given session for the GPU with the given UUID. Also creates a partition object for the device iff bCreateSmcPartition is true and pGpuInfo->smcEnabled is true. pGpuInfo->smcUserClientInfo will be used to determine the SMC partition in this case. A device handle is returned in the device output parameter. Error codes: NV_ERR_GENERIC NV_ERR_NO_MEMORY NV_ERR_INVALID_ARGUMENT NV_ERR_INSUFFICIENT_RESOURCES NV_ERR_OBJECT_NOT_FOUND */ NV_STATUS nvUvmInterfaceDeviceCreate(uvmGpuSessionHandle session, const UvmGpuInfo *pGpuInfo, const NvProcessorUuid *gpuUuid, uvmGpuDeviceHandle *device, NvBool bCreateSmcPartition); /******************************************************************************* nvUvmInterfaceDeviceDestroy Destroys the device object for the given handle. The handle must have been obtained in a prior call to nvUvmInterfaceDeviceCreate. */ void nvUvmInterfaceDeviceDestroy(uvmGpuDeviceHandle device); /******************************************************************************* nvUvmInterfaceAddressSpaceCreate This function creates an address space. This virtual address space is created on the GPU specified by device. Error codes: NV_ERR_GENERIC NV_ERR_NO_MEMORY */ NV_STATUS nvUvmInterfaceAddressSpaceCreate(uvmGpuDeviceHandle device, unsigned long long vaBase, unsigned long long vaSize, uvmGpuAddressSpaceHandle *vaSpace, UvmGpuAddressSpaceInfo *vaSpaceInfo); /******************************************************************************* nvUvmInterfaceDupAddressSpace This function will dup the given vaspace from the users client to the kernel client was created as an ops session. By duping the vaspace it is guaranteed that RM will refcount the vaspace object. Error codes: NV_ERR_GENERIC */ NV_STATUS nvUvmInterfaceDupAddressSpace(uvmGpuDeviceHandle device, NvHandle hUserClient, NvHandle hUserVASpace, uvmGpuAddressSpaceHandle *vaSpace, UvmGpuAddressSpaceInfo *vaSpaceInfo); /******************************************************************************* nvUvmInterfaceAddressSpaceDestroy Destroys an address space that was previously created via nvUvmInterfaceAddressSpaceCreate. */ void nvUvmInterfaceAddressSpaceDestroy(uvmGpuAddressSpaceHandle vaSpace); /******************************************************************************* nvUvmInterfaceMemoryAllocFB This function will allocate video memory and provide a mapped Gpu virtual address to this allocation. It also returns the Gpu physical offset if contiguous allocations are requested. This function will allocate a minimum page size if the length provided is 0 and will return a unique GPU virtual address. The default page size will be the small page size (as returned by query caps). The physical alignment will also be enforced to small page size(64K/128K). Arguments: vaSpace[IN] - Pointer to vaSpace object length [IN] - Length of the allocation gpuPointer[OUT] - GPU VA mapping allocInfo[IN/OUT] - Pointer to allocation info structure which contains below given fields allocInfo Members: gpuPhysOffset[OUT] - Physical offset of allocation returned only if contiguous allocation is requested. pageSize[IN] - Override the default page size (see above). alignment[IN] - gpuPointer GPU VA alignment. 0 means 4KB alignment. bContiguousPhysAlloc[IN] - Flag to request contiguous allocation. Default will follow the vidHeapControl default policy. bMemGrowsDown[IN] bPersistentVidmem[IN] - Allocate persistent vidmem. hPhysHandle[IN/OUT] - The handle will be used in allocation if provided. If not provided; allocator will return the handle it used eventually. Error codes: NV_ERR_INVALID_ARGUMENT NV_ERR_NO_MEMORY - Not enough physical memory to service allocation request with provided constraints NV_ERR_INSUFFICIENT_RESOURCES - Not enough available resources to satisfy allocation request NV_ERR_INVALID_OWNER - Target memory not accessible by specified owner NV_ERR_NOT_SUPPORTED - Operation not supported on broken FB */ NV_STATUS nvUvmInterfaceMemoryAllocFB(uvmGpuAddressSpaceHandle vaSpace, NvLength length, UvmGpuPointer * gpuPointer, UvmGpuAllocInfo * allocInfo); /******************************************************************************* nvUvmInterfaceMemoryAllocSys This function will allocate system memory and provide a mapped Gpu virtual address to this allocation. This function will allocate a minimum page size if the length provided is 0 and will return a unique GPU virtual address. The default page size will be the small page size (as returned by query caps) Arguments: vaSpace[IN] - Pointer to vaSpace object length [IN] - Length of the allocation gpuPointer[OUT] - GPU VA mapping allocInfo[IN/OUT] - Pointer to allocation info structure which contains below given fields allocInfo Members: gpuPhysOffset[OUT] - Physical offset of allocation returned only if contiguous allocation is requested. pageSize[IN] - Override the default page size (see above). alignment[IN] - gpuPointer GPU VA alignment. 0 means 4KB alignment. bContiguousPhysAlloc[IN] - Flag to request contiguous allocation. Default will follow the vidHeapControl default policy. bMemGrowsDown[IN] bPersistentVidmem[IN] - Allocate persistent vidmem. hPhysHandle[IN/OUT] - The handle will be used in allocation if provided. If not provided; allocator will return the handle it used eventually. Error codes: NV_ERR_INVALID_ARGUMENT NV_ERR_NO_MEMORY - Not enough physical memory to service allocation request with provided constraints NV_ERR_INSUFFICIENT_RESOURCES - Not enough available resources to satisfy allocation request NV_ERR_INVALID_OWNER - Target memory not accessible by specified owner NV_ERR_NOT_SUPPORTED - Operation not supported */ NV_STATUS nvUvmInterfaceMemoryAllocSys(uvmGpuAddressSpaceHandle vaSpace, NvLength length, UvmGpuPointer * gpuPointer, UvmGpuAllocInfo * allocInfo); /******************************************************************************* nvUvmInterfaceGetP2PCaps Obtain the P2P capabilities between two devices. Arguments: device1[IN] - Device handle of the first GPU (required) device2[IN] - Device handle of the second GPU (required) p2pCapsParams [OUT] - P2P capabilities between the two GPUs Error codes: NV_ERR_INVALID_ARGUMENT NV_ERR_GENERIC: Unexpected error. We try hard to avoid returning this error code,because it is not very informative. */ NV_STATUS nvUvmInterfaceGetP2PCaps(uvmGpuDeviceHandle device1, uvmGpuDeviceHandle device2, UvmGpuP2PCapsParams * p2pCapsParams); /******************************************************************************* nvUvmInterfaceGetPmaObject This function will return pointer to PMA object for the given GPU. This PMA object handle is required for page allocation. Arguments: device [IN] - Device handle allocated in nvUvmInterfaceDeviceCreate pPma [OUT] - Pointer to PMA object pPmaPubStats [OUT] - Pointer to UvmPmaStatistics object Error codes: NV_ERR_NOT_SUPPORTED - Operation not supported on broken FB NV_ERR_GENERIC: Unexpected error. We try hard to avoid returning this error code,because it is not very informative. */ NV_STATUS nvUvmInterfaceGetPmaObject(uvmGpuDeviceHandle device, void **pPma, const UvmPmaStatistics **pPmaPubStats); // Mirrors pmaEvictPagesCb_t, see its documentation in pma.h. typedef NV_STATUS (*uvmPmaEvictPagesCallback)(void *callbackData, NvU32 pageSize, NvU64 *pPages, NvU32 count, NvU64 physBegin, NvU64 physEnd); // Mirrors pmaEvictRangeCb_t, see its documentation in pma.h. typedef NV_STATUS (*uvmPmaEvictRangeCallback)(void *callbackData, NvU64 physBegin, NvU64 physEnd); /******************************************************************************* nvUvmInterfacePmaRegisterEvictionCallbacks Simple wrapper for pmaRegisterEvictionCb(), see its documentation in pma.h. */ NV_STATUS nvUvmInterfacePmaRegisterEvictionCallbacks(void *pPma, uvmPmaEvictPagesCallback evictPages, uvmPmaEvictRangeCallback evictRange, void *callbackData); /****************************************************************************** nvUvmInterfacePmaUnregisterEvictionCallbacks Simple wrapper for pmaUnregisterEvictionCb(), see its documentation in pma.h. */ void nvUvmInterfacePmaUnregisterEvictionCallbacks(void *pPma); /******************************************************************************* nvUvmInterfacePmaAllocPages @brief Synchronous API for allocating pages from the PMA. PMA will decide which pma regions to allocate from based on the provided flags. PMA will also initiate UVM evictions to make room for this allocation unless prohibited by PMA_FLAGS_DONT_EVICT. UVM callers must pass this flag to avoid deadlock. Only UVM may allocated unpinned memory from this API. For broadcast methods, PMA will guarantee the same physical frames are allocated on multiple GPUs, specified by the PMA objects passed in. If allocation is contiguous, only one page in pPages will be filled. Also, contiguous flag must be passed later to nvUvmInterfacePmaFreePages. Arguments: pPma[IN] - Pointer to PMA object pageCount [IN] - Number of pages required to be allocated. pageSize [IN] - 64kb, 128kb or 2mb. No other values are permissible. pPmaAllocOptions[IN] - Pointer to PMA allocation info structure. pPages[OUT] - Array of pointers, containing the PA base address of each page. Error codes: NV_ERR_NO_MEMORY: Internal memory allocation failed. NV_ERR_GENERIC: Unexpected error. We try hard to avoid returning this error code,because it is not very informative. */ NV_STATUS nvUvmInterfacePmaAllocPages(void *pPma, NvLength pageCount, NvU32 pageSize, UvmPmaAllocationOptions *pPmaAllocOptions, NvU64 *pPages); /******************************************************************************* nvUvmInterfacePmaPinPages This function will pin the physical memory allocated using PMA. The pages passed as input must be unpinned else this function will return an error and rollback any change if any page is not previously marked "unpinned". Arguments: pPma[IN] - Pointer to PMA object. pPages[IN] - Array of pointers, containing the PA base address of each page to be pinned. pageCount [IN] - Number of pages required to be pinned. pageSize [IN] - Page size of each page to be pinned. flags [IN] - UVM_PMA_CALLED_FROM_PMA_EVICTION if called from PMA eviction, 0 otherwise. Error codes: NV_ERR_INVALID_ARGUMENT - Invalid input arguments. NV_ERR_GENERIC - Unexpected error. We try hard to avoid returning this error code as is not very informative. NV_ERR_NOT_SUPPORTED - Operation not supported on broken FB */ NV_STATUS nvUvmInterfacePmaPinPages(void *pPma, NvU64 *pPages, NvLength pageCount, NvU32 pageSize, NvU32 flags); /******************************************************************************* nvUvmInterfacePmaUnpinPages This function will unpin the physical memory allocated using PMA. The pages passed as input must be already pinned, else this function will return an error and rollback any change if any page is not previously marked "pinned". Behaviour is undefined if any blacklisted pages are unpinned. Arguments: pPma[IN] - Pointer to PMA object. pPages[IN] - Array of pointers, containing the PA base address of each page to be unpinned. pageCount [IN] - Number of pages required to be unpinned. pageSize [IN] - Page size of each page to be unpinned. Error codes: NV_ERR_INVALID_ARGUMENT - Invalid input arguments. NV_ERR_GENERIC - Unexpected error. We try hard to avoid returning this error code as is not very informative. NV_ERR_NOT_SUPPORTED - Operation not supported on broken FB */ NV_STATUS nvUvmInterfacePmaUnpinPages(void *pPma, NvU64 *pPages, NvLength pageCount, NvU32 pageSize); /******************************************************************************* nvUvmInterfaceMemoryFree Free up a GPU allocation */ void nvUvmInterfaceMemoryFree(uvmGpuAddressSpaceHandle vaSpace, UvmGpuPointer gpuPointer); /******************************************************************************* nvUvmInterfacePmaFreePages This function will free physical memory allocated using PMA. It marks a list of pages as free. This operation is also used by RM to mark pages as "scrubbed" for the initial ECC sweep. This function does not fail. When allocation was contiguous, an appropriate flag needs to be passed. Arguments: pPma[IN] - Pointer to PMA object pPages[IN] - Array of pointers, containing the PA base address of each page. pageCount [IN] - Number of pages required to be allocated. pageSize [IN] - Page size of each page flags [IN] - Flags with information about allocation type with the same meaning as flags in options for nvUvmInterfacePmaAllocPages. When called from PMA eviction, UVM_PMA_CALLED_FROM_PMA_EVICTION needs to be added to flags. Error codes: NV_ERR_INVALID_ARGUMENT NV_ERR_NO_MEMORY - Not enough physical memory to service allocation request with provided constraints NV_ERR_INSUFFICIENT_RESOURCES - Not enough available resources to satisfy allocation request NV_ERR_INVALID_OWNER - Target memory not accessible by specified owner NV_ERR_NOT_SUPPORTED - Operation not supported on broken FB */ void nvUvmInterfacePmaFreePages(void *pPma, NvU64 *pPages, NvLength pageCount, NvU32 pageSize, NvU32 flags); /******************************************************************************* nvUvmInterfaceMemoryCpuMap This function creates a CPU mapping to the provided GPU address. If the address is not the same as what is returned by the Alloc function, then the function will map it from the address provided. This offset will be relative to the gpu offset obtained from the memory alloc functions. Error codes: NV_ERR_GENERIC NV_ERR_NO_MEMORY */ NV_STATUS nvUvmInterfaceMemoryCpuMap(uvmGpuAddressSpaceHandle vaSpace, UvmGpuPointer gpuPointer, NvLength length, void **cpuPtr, NvU32 pageSize); /******************************************************************************* uvmGpuMemoryCpuUnmap Unmaps the cpuPtr provided from the process virtual address space. */ void nvUvmInterfaceMemoryCpuUnMap(uvmGpuAddressSpaceHandle vaSpace, void *cpuPtr); /******************************************************************************* nvUvmInterfaceChannelAllocate This function will allocate a channel bound to a copy engine allocParams must contain an engineIndex as channels need to be bound to an engine type at allocation time. The possible values are [0, UVM_COPY_ENGINE_COUNT_MAX), but notably only the copy engines that have UvmGpuCopyEngineCaps::supported set to true can be allocated. This struct also contains information relative to GPFIFO and GPPut. channel is filled with the address of the corresponding channel handle. channelInfo is filled out with channel get/put. The errorNotifier is filled out when the channel hits an RC error. On Volta+ devices, it also computes the work submission token and the work submission offset to be used in the Host channel submission doorbell. Arguments: vaSpace[IN] - VA space linked to a client and a device under which the channel will be allocated allocParams[IN] - structure with allocation settings channel[OUT] - pointer to the new channel handle channelInfo[OUT] - structure filled with channel information Error codes: NV_ERR_GENERIC NV_ERR_NO_MEMORY */ NV_STATUS nvUvmInterfaceChannelAllocate(uvmGpuAddressSpaceHandle vaSpace, const UvmGpuChannelAllocParams *allocParams, uvmGpuChannelHandle *channel, UvmGpuChannelInfo *channelInfo); /******************************************************************************* nvUvmInterfaceChannelDestroy This function destroys a given channel Arguments: channel[IN] - channel handle */ void nvUvmInterfaceChannelDestroy(uvmGpuChannelHandle channel); /******************************************************************************* nvUvmInterfaceQueryCaps Return capabilities for the provided GPU. If GPU does not exist, an error will be returned. If the client is only interested in the capabilities of the Copy Engines of the given GPU, use nvUvmInterfaceQueryCopyEnginesCaps instead. Error codes: NV_ERR_GENERIC NV_ERR_NO_MEMORY */ NV_STATUS nvUvmInterfaceQueryCaps(uvmGpuDeviceHandle device, UvmGpuCaps * caps); /******************************************************************************* nvUvmInterfaceQueryCopyEnginesCaps Return the capabilities of all the Copy Engines for the provided GPU. If the GPU does not exist, an error will be returned. Error codes: NV_ERR_GENERIC NV_ERR_NO_MEMORY */ NV_STATUS nvUvmInterfaceQueryCopyEnginesCaps(uvmGpuDeviceHandle device, UvmGpuCopyEnginesCaps *caps); /******************************************************************************* nvUvmInterfaceGetGpuInfo Return various gpu info, refer to the UvmGpuInfo struct for details. If no gpu matching the uuid is found, an error will be returned. On Ampere+ GPUs, pGpuClientInfo contains SMC information provided by the client regarding the partition targeted in this operation. Error codes: NV_ERR_GENERIC NV_ERR_INSUFFICIENT_RESOURCES */ NV_STATUS nvUvmInterfaceGetGpuInfo(const NvProcessorUuid *gpuUuid, const UvmGpuClientInfo *pGpuClientInfo, UvmGpuInfo *pGpuInfo); /******************************************************************************* nvUvmInterfaceServiceDeviceInterruptsRM Tells RM to service all pending interrupts. This is helpful in ECC error conditions when ECC error interrupt is set & error can be determined only after ECC notifier will be set or reset. Error codes: NV_ERR_GENERIC UVM_INVALID_ARGUMENTS */ NV_STATUS nvUvmInterfaceServiceDeviceInterruptsRM(uvmGpuDeviceHandle device); /******************************************************************************* nvUvmInterfaceSetPageDirectory Sets pageDirectory in the provided location. Also moves the existing PDE to the provided pageDirectory. RM will propagate the update to all channels using the provided VA space. All channels must be idle when this call is made. Arguments: vaSpace[IN} - VASpace Object physAddress[IN] - Physical address of new page directory numEntries[IN] - Number of entries including previous PDE which will be copied bVidMemAperture[IN] - If set pageDirectory will reside in VidMem aperture else sysmem pasid[IN] - PASID (Process Address Space IDentifier) of the process corresponding to the VA space. Ignored unless the VA space object has ATS enabled. Error codes: NV_ERR_GENERIC NV_ERR_INVALID_ARGUMENT */ NV_STATUS nvUvmInterfaceSetPageDirectory(uvmGpuAddressSpaceHandle vaSpace, NvU64 physAddress, unsigned numEntries, NvBool bVidMemAperture, NvU32 pasid); /******************************************************************************* nvUvmInterfaceUnsetPageDirectory Unsets/Restores pageDirectory to RM's defined location. Arguments: vaSpace[IN} - VASpace Object Error codes: NV_ERR_GENERIC NV_ERR_INVALID_ARGUMENT */ NV_STATUS nvUvmInterfaceUnsetPageDirectory(uvmGpuAddressSpaceHandle vaSpace); /******************************************************************************* nvUvmInterfaceDupAllocation Duplicate the given allocation in a different VA space. The physical handle backing the source allocation is duplicated in the GPU device associated with the destination VA space, and a new mapping is created in that VA space. The input allocation can be located in sysmem (i.e. allocated using nvUvmInterfaceMemoryAllocSys) or vidmem (i.e. allocated using nvUvmInterfaceMemoryAllocFB). If located in vidmem, duplication across GPUs is not supported. For duplication of physical memory use nvUvmInterfaceDupMemory. Arguments: srcVaSpace[IN] - Source VA space. srcAddress[IN] - GPU VA in the source VA space. The provided address should match one previously returned by nvUvmInterfaceMemoryAllocFB or nvUvmInterfaceMemoryAllocSys. dstVaSpace[IN] - Destination VA space where the new mapping will be created. dstVaAlignment[IN] - Alignment of the GPU VA in the destination VA space. 0 means 4KB alignment. dstAddress[OUT] - Pointer to the GPU VA in the destination VA space. Error codes: NV_ERR_INVALID_ARGUMENT - If any of the inputs is invalid, or the source and destination VA spaces are identical. NV_ERR_OBJECT_NOT_FOUND - If the input allocation is not found in under the provided VA space. NV_ERR_NO_MEMORY - If there is no memory to back the duplicate, or the associated metadata. NV_ERR_NOT_SUPPORTED - If trying to duplicate vidmem across GPUs. */ NV_STATUS nvUvmInterfaceDupAllocation(uvmGpuAddressSpaceHandle srcVaSpace, NvU64 srcAddress, uvmGpuAddressSpaceHandle dstVaSpace, NvU64 dstVaAlignment, NvU64 *dstAddress); /******************************************************************************* nvUvmInterfaceDupMemory Duplicates a physical memory allocation. If requested, provides information about the allocation. Arguments: device[IN] - Device linked to a client under which the phys memory needs to be duped. hClient[IN] - Client owning the memory. hPhysMemory[IN] - Phys memory which is to be duped. hDupedHandle[OUT] - Handle of the duped memory object. pGpuMemoryInfo[OUT] - see nv_uvm_types.h for more information. This parameter can be NULL. (optional) Error codes: NV_ERR_INVALID_ARGUMENT - If the parameter/s is invalid. NV_ERR_NOT_SUPPORTED - If the allocation is not a physical allocation. NV_ERR_OBJECT_NOT_FOUND - If the allocation is not found in under the provided client. */ NV_STATUS nvUvmInterfaceDupMemory(uvmGpuDeviceHandle device, NvHandle hClient, NvHandle hPhysMemory, NvHandle *hDupMemory, UvmGpuMemoryInfo *pGpuMemoryInfo); /******************************************************************************* nvUvmInterfaceFreeDupedAllocation Free the allocation represented by the physical handle used to create the duped allocation. Arguments: device[IN] - Device handle used to dup the memory. hPhysHandle[IN] - Handle representing the phys allocation. Error codes: NV_ERROR NV_ERR_INVALID_ARGUMENT */ NV_STATUS nvUvmInterfaceFreeDupedHandle(uvmGpuDeviceHandle device, NvHandle hPhysHandle); /******************************************************************************* nvUvmInterfaceGetFbInfo Gets FB information from RM. Arguments: device[IN] - GPU device handle fbInfo [OUT] - Pointer to FbInfo structure which contains reservedHeapSize & heapSize Error codes: NV_ERROR NV_ERR_INVALID_ARGUMENT */ NV_STATUS nvUvmInterfaceGetFbInfo(uvmGpuDeviceHandle device, UvmGpuFbInfo * fbInfo); /******************************************************************************* nvUvmInterfaceGetEccInfo Gets ECC information from RM. Arguments: device[IN] - GPU device handle eccInfo [OUT] - Pointer to EccInfo structure Error codes: NV_ERROR NV_ERR_INVALID_ARGUMENT */ NV_STATUS nvUvmInterfaceGetEccInfo(uvmGpuDeviceHandle device, UvmGpuEccInfo * eccInfo); /******************************************************************************* nvUvmInterfaceOwnPageFaultIntr This function transfers ownership of the replayable page fault interrupt, between RM and UVM, for a particular GPU. bOwnInterrupts == NV_TRUE: UVM is taking ownership from the RM. This causes the following: RM will not service, enable or disable this interrupt and it is up to the UVM driver to handle this interrupt. In this case, replayable page fault interrupts are disabled by this function, before it returns. bOwnInterrupts == NV_FALSE: UVM is returning ownership to the RM: in this case, replayable page fault interrupts MUST BE DISABLED BEFORE CALLING this function. The cases above both result in transferring ownership of a GPU that has its replayable page fault interrupts disabled. Doing otherwise would make it very difficult to control which driver handles any interrupts that build up during the hand-off. The calling pattern should look like this: UVM setting up a new GPU for operation: UVM GPU LOCK nvUvmInterfaceOwnPageFaultIntr(..., NV_TRUE) UVM GPU UNLOCK Enable replayable page faults for that GPU UVM tearing down a GPU: Disable replayable page faults for that GPU UVM GPU GPU LOCK nvUvmInterfaceOwnPageFaultIntr(..., NV_FALSE) UVM GPU UNLOCK Arguments: gpuUuid[IN] - UUID of the GPU to operate on bOwnInterrupts - Set to NV_TRUE for UVM to take ownership of the replayable page fault interrupts. Set to NV_FALSE to return ownership of the page fault interrupts to RM. Error codes: NV_ERR_GENERIC NV_ERR_INVALID_ARGUMENT */ NV_STATUS nvUvmInterfaceOwnPageFaultIntr(uvmGpuDeviceHandle device, NvBool bOwnInterrupts); /******************************************************************************* nvUvmInterfaceInitFaultInfo This function obtains fault buffer address, size and a few register mappings for replayable faults, and creates a shadow buffer to store non-replayable faults if the GPU supports it. Arguments: device[IN] - Device handle associated with the gpu pFaultInfo[OUT] - information provided by RM for fault handling Error codes: NV_ERR_GENERIC NV_ERR_NO_MEMORY NV_ERR_INVALID_ARGUMENT */ NV_STATUS nvUvmInterfaceInitFaultInfo(uvmGpuDeviceHandle device, UvmGpuFaultInfo *pFaultInfo); /******************************************************************************* nvUvmInterfaceDestroyFaultInfo This function obtains destroys unmaps the fault buffer and clears faultInfo for replayable faults, and frees the shadow buffer for non-replayable faults. Arguments: device[IN] - Device handle associated with the gpu pFaultInfo[OUT] - information provided by RM for fault handling Error codes: NV_ERR_GENERIC NV_ERR_INVALID_ARGUMENT */ NV_STATUS nvUvmInterfaceDestroyFaultInfo(uvmGpuDeviceHandle device, UvmGpuFaultInfo *pFaultInfo); /******************************************************************************* nvUvmInterfaceHasPendingNonReplayableFaults This function tells whether there are pending non-replayable faults in the client shadow fault buffer ready to be consumed. NOTES: - This function uses a pre-allocated stack per GPU (stored in the UvmGpuFaultInfo object) for calls related to non-replayable faults from the top half. - Concurrent calls to this function using the same pFaultInfo are not thread-safe due to pre-allocated stack. Therefore, locking is the caller's responsibility. - This function DOES NOT acquire the RM API or GPU locks. That is because it is called during fault servicing, which could produce deadlocks. Arguments: pFaultInfo[IN] - information provided by RM for fault handling. Contains a pointer to the shadow fault buffer hasPendingFaults[OUT] - return value that tells if there are non-replayable faults ready to be consumed by the client Error codes: NV_ERR_INVALID_ARGUMENT */ NV_STATUS nvUvmInterfaceHasPendingNonReplayableFaults(UvmGpuFaultInfo *pFaultInfo, NvBool *hasPendingFaults); /******************************************************************************* nvUvmInterfaceGetNonReplayableFaults This function consumes all the non-replayable fault packets in the client shadow fault buffer and copies them to the given buffer. It also returns the number of faults that have been copied NOTES: - This function uses a pre-allocated stack per GPU (stored in the UvmGpuFaultInfo object) for calls from the bottom half that handles non-replayable faults. - See nvUvmInterfaceHasPendingNonReplayableFaults for the implications of using a shared stack. - This function DOES NOT acquire the RM API or GPU locks. That is because it is called during fault servicing, which could produce deadlocks. Arguments: pFaultInfo[IN] - information provided by RM for fault handling. Contains a pointer to the shadow fault buffer pFaultBuffer[OUT] - buffer provided by the client where fault buffers are copied when they are popped out of the shadow fault buffer (which is a circular queue). numFaults[OUT] - return value that tells the number of faults copied to the client's buffer Error codes: NV_ERR_INVALID_ARGUMENT */ NV_STATUS nvUvmInterfaceGetNonReplayableFaults(UvmGpuFaultInfo *pFaultInfo, void *pFaultBuffer, NvU32 *numFaults); /******************************************************************************* nvUvmInterfaceInitAccessCntrInfo This function obtains access counter buffer address, size and a few register mappings Arguments: device[IN] - Device handle associated with the gpu pAccessCntrInfo[OUT] - Information provided by RM for access counter handling Error codes: NV_ERR_GENERIC NV_ERR_INVALID_ARGUMENT */ NV_STATUS nvUvmInterfaceInitAccessCntrInfo(uvmGpuDeviceHandle device, UvmGpuAccessCntrInfo *pAccessCntrInfo); /******************************************************************************* nvUvmInterfaceDestroyAccessCntrInfo This function obtains, destroys, unmaps the access counter buffer and clears accessCntrInfo Arguments: device[IN] - Device handle associated with the gpu pAccessCntrInfo[IN] - Information provided by RM for access counter handling Error codes: NV_ERR_GENERIC NV_ERR_INVALID_ARGUMENT */ NV_STATUS nvUvmInterfaceDestroyAccessCntrInfo(uvmGpuDeviceHandle device, UvmGpuAccessCntrInfo *pAccessCntrInfo); /******************************************************************************* nvUvmInterfaceEnableAccessCntr This function enables access counters using the given configuration UVM is also taking ownership from the RM. This causes the following: RM will not service, enable or disable this interrupt and it is up to the UVM driver to handle this interrupt. In this case, access counter notificaion interrupts are enabled by this function before it returns. Arguments: device[IN] - Device handle associated with the gpu pAccessCntrInfo[IN] - Pointer to structure filled out by nvUvmInterfaceInitAccessCntrInfo pAccessCntrConfig[IN] - Configuration for access counters Error codes: NV_ERR_GENERIC NV_ERR_INVALID_ARGUMENT */ NV_STATUS nvUvmInterfaceEnableAccessCntr(uvmGpuDeviceHandle device, UvmGpuAccessCntrInfo *pAccessCntrInfo, UvmGpuAccessCntrConfig *pAccessCntrConfig); /******************************************************************************* nvUvmInterfaceDisableAccessCntr This function disables acccess counters UVM is also returning ownership to the RM: RM can service, enable or disable this interrupt. In this case, access counter notificaion interrupts are disabled by this function before it returns. Arguments: device[IN] - Device handle associated with the gpu pAccessCntrInfo[IN] - Pointer to structure filled out by nvUvmInterfaceInitAccessCntrInfo Error codes: NV_ERR_GENERIC NV_ERR_INVALID_ARGUMENT */ NV_STATUS nvUvmInterfaceDisableAccessCntr(uvmGpuDeviceHandle device, UvmGpuAccessCntrInfo *pAccessCntrInfo); // // Called by the UVM driver to register operations with RM. Only one set of // callbacks can be registered by any driver at a time. If another set of // callbacks was already registered, NV_ERR_IN_USE is returned. // NV_STATUS nvUvmInterfaceRegisterUvmCallbacks(struct UvmOpsUvmEvents *importedUvmOps); // // Counterpart to nvUvmInterfaceRegisterUvmCallbacks. This must only be called // if nvUvmInterfaceRegisterUvmCallbacks returned NV_OK. // // Upon return, the caller is guaranteed that any outstanding callbacks are done // and no new ones will be invoked. // void nvUvmInterfaceDeRegisterUvmOps(void); /******************************************************************************* nvUvmInterfaceP2pObjectCreate This API creates an NV50_P2P object for the GPUs with the given device handles, and returns the handle to the object. Arguments: device1[IN] - first GPU device handle device2[IN] - second GPU device handle hP2pObject[OUT] - handle to the created P2p object. Error codes: NV_ERR_INVALID_ARGUMENT NV_ERR_OBJECT_NOT_FOUND : If device object associated with the uuids aren't found. */ NV_STATUS nvUvmInterfaceP2pObjectCreate(uvmGpuDeviceHandle device1, uvmGpuDeviceHandle device2, NvHandle *hP2pObject); /******************************************************************************* nvUvmInterfaceP2pObjectDestroy This API destroys the NV50_P2P associated with the passed handle. Arguments: session[IN] - Session handle. hP2pObject[IN] - handle to an P2p object. Error codes: NONE */ void nvUvmInterfaceP2pObjectDestroy(uvmGpuSessionHandle session, NvHandle hP2pObject); /******************************************************************************* nvUvmInterfaceGetExternalAllocPtes The interface builds the RM PTEs using the provided input parameters. Arguments: vaSpace[IN] - vaSpace handle. hMemory[IN] - Memory handle. offset [IN] - Offset from the beginning of the allocation where PTE mappings should begin. Should be aligned with pagesize associated with the allocation. size [IN] - Length of the allocation for which PTEs should be built. Should be aligned with pagesize associated with the allocation. size = 0 will be interpreted as the total size of the allocation. gpuExternalMappingInfo[IN/OUT] - See nv_uvm_types.h for more information. Error codes: NV_ERR_INVALID_ARGUMENT - Invalid parameter/s is passed. NV_ERR_INVALID_OBJECT_HANDLE - Invalid memory handle is passed. NV_ERR_NOT_SUPPORTED - Functionality is not supported (see comments in nv_gpu_ops.c) NV_ERR_INVALID_BASE - offset is beyond the allocation size NV_ERR_INVALID_LIMIT - (offset + size) is beyond the allocation size. NV_ERR_BUFFER_TOO_SMALL - gpuExternalMappingInfo.pteBufferSize is insufficient to store single PTE. NV_ERR_NOT_READY - Returned when querying the PTEs requires a deferred setup which has not yet completed. It is expected that the caller will reattempt the call until a different code is returned. */ NV_STATUS nvUvmInterfaceGetExternalAllocPtes(uvmGpuAddressSpaceHandle vaSpace, NvHandle hMemory, NvU64 offset, NvU64 size, UvmGpuExternalMappingInfo *gpuExternalMappingInfo); /******************************************************************************* nvUvmInterfaceRetainChannel Validates and returns information about the user's channel and its resources (local CTX buffers + global CTX buffers). The state is refcounted and must be released by calling nvUvmInterfaceReleaseChannel. Arguments: vaSpace[IN] - vaSpace handle. hClient[IN] - Client handle hChannel[IN] - Channel handle retainedChannel[OUT] - Opaque pointer to use to refer to this channel in other nvUvmInterface APIs. channelInstanceInfo[OUT] - Channel instance information to be filled out. See nv_uvm_types.h for details. Error codes: NV_ERR_INVALID_ARGUMENT : If the parameter/s are invalid. NV_ERR_OBJECT_NOT_FOUND : If the object associated with the handle isn't found. NV_ERR_INVALID_CHANNEL : If the channel verification fails. NV_ERR_INSUFFICIENT_RESOURCES : If no memory available to store the resource information. */ NV_STATUS nvUvmInterfaceRetainChannel(uvmGpuAddressSpaceHandle vaSpace, NvHandle hClient, NvHandle hChannel, void **retainedChannel, UvmGpuChannelInstanceInfo *channelInstanceInfo); /******************************************************************************* nvUvmInterfaceBindChannelResources Associates the mapping address of the channel resources (VAs) provided by the caller with the channel. Arguments: retainedChannel[IN] - Channel pointer returned by nvUvmInterfaceRetainChannel channelResourceBindParams[IN] - Buffer of initialized UvmGpuChannelInstanceInfo::resourceCount entries. See nv_uvm_types.h for details. Error codes: NV_ERR_INVALID_ARGUMENT : If the parameter/s are invalid. NV_ERR_OBJECT_NOT_FOUND : If the object associated with the handle aren't found. NV_ERR_INSUFFICIENT_RESOURCES : If no memory available to store the resource information. */ NV_STATUS nvUvmInterfaceBindChannelResources(void *retainedChannel, UvmGpuChannelResourceBindParams *channelResourceBindParams); /******************************************************************************* nvUvmInterfaceReleaseChannel Releases state retained by nvUvmInterfaceRetainChannel. */ void nvUvmInterfaceReleaseChannel(void *retainedChannel); /******************************************************************************* nvUvmInterfaceStopChannel Idles the channel and takes it off the runlist. Arguments: retainedChannel[IN] - Channel pointer returned by nvUvmInterfaceRetainChannel bImmediate[IN] - If true, kill the channel without attempting to wait for it to go idle. */ void nvUvmInterfaceStopChannel(void *retainedChannel, NvBool bImmediate); /******************************************************************************* nvUvmInterfaceGetChannelResourcePtes The interface builds the RM PTEs using the provided input parameters. Arguments: vaSpace[IN] - vaSpace handle. resourceDescriptor[IN] - The channel resource descriptor returned by returned by nvUvmInterfaceRetainChannelResources. offset[IN] - Offset from the beginning of the allocation where PTE mappings should begin. Should be aligned with pagesize associated with the allocation. size[IN] - Length of the allocation for which PTEs should be built. Should be aligned with pagesize associated with the allocation. size = 0 will be interpreted as the total size of the allocation. gpuExternalMappingInfo[IN/OUT] - See nv_uvm_types.h for more information. Error codes: NV_ERR_INVALID_ARGUMENT - Invalid parameter/s is passed. NV_ERR_INVALID_OBJECT_HANDLE - Invalid memory handle is passed. NV_ERR_NOT_SUPPORTED - Functionality is not supported. NV_ERR_INVALID_BASE - offset is beyond the allocation size NV_ERR_INVALID_LIMIT - (offset + size) is beyond the allocation size. NV_ERR_BUFFER_TOO_SMALL - gpuExternalMappingInfo.pteBufferSize is insufficient to store single PTE. */ NV_STATUS nvUvmInterfaceGetChannelResourcePtes(uvmGpuAddressSpaceHandle vaSpace, NvP64 resourceDescriptor, NvU64 offset, NvU64 size, UvmGpuExternalMappingInfo *externalMappingInfo); /******************************************************************************* nvUvmInterfaceReportNonReplayableFault The interface communicates a nonreplayable fault packet from UVM to RM, which will log the fault, notify the clients and then trigger RC on the channel. Arguments: device[IN] - The device where the fault happened. pFaultPacket[IN] - The opaque pointer from UVM that will be later converted to a MMU_FAULT_PACKET type. Error codes: NV_ERR_INVALID_ARGUMENT - Invalid parameter/s is passed. NV_ERR_NOT_SUPPORTED - Functionality is not supported. */ NV_STATUS nvUvmInterfaceReportNonReplayableFault(uvmGpuDeviceHandle device, const void *pFaultPacket); /******************************************************************************* nvUvmInterfacePagingChannelAllocate In SR-IOV heavy, this function requests the allocation of a paging channel (i.e. a privileged CE channel) bound to a specified copy engine. Unlike channels allocated via nvUvmInterfaceChannelAllocate, the caller cannot push methods to a paging channel directly, but instead relies on the nvUvmInterfacePagingChannelPushStream API to do so. SR-IOV heavy only. The implementation of this interface can acquire RM or GPU locks. Arguments: device[IN] - device under which the paging channel will be allocated allocParams[IN] - structure with allocation settings channel[OUT] - pointer to the allocated paging channel handle channelInfo[OUT] - structure filled with channel information Error codes: NV_ERR_INVALID_ARGUMENT - Invalid parameter/s is passed. NV_ERR_NO_MEMORY - Not enough memory to allocate paging channel/shadow notifier. NV_ERR_NOT_SUPPORTED - SR-IOV heavy mode is disabled. */ NV_STATUS nvUvmInterfacePagingChannelAllocate(uvmGpuDeviceHandle device, const UvmGpuPagingChannelAllocParams *allocParams, UvmGpuPagingChannelHandle *channel, UvmGpuPagingChannelInfo *channelInfo); /******************************************************************************* nvUvmInterfacePagingChannelDestroy This function destroys a given paging channel. SR-IOV heavy only. The implementation of this interface can acquire RM or GPU locks. Arguments: channel[IN] - paging channel handle. If the passed handle is the NULL pointer, the function returns immediately. */ void nvUvmInterfacePagingChannelDestroy(UvmGpuPagingChannelHandle channel); /******************************************************************************* nvUvmInterfacePagingChannelsMap Map a guest allocation in the address space associated with all the paging channels allocated under the given device. SR-IOV heavy only. The implementation of this interface can acquire RM or GPU locks. Arguments: srcVaSpace[IN] - VA space handle used to allocate the input pointer srcAddress. srcAddress[IN] - virtual address returned by nvUvmInterfaceMemoryAllocFB or nvUvmInterfaceMemoryAllocSys. The entire allocation backing this guest VA is mapped. device[IN] - device under which paging channels were allocated dstAddress[OUT] - a virtual address that is valid (i.e. is mapped) in all the paging channels allocated under the given vaSpace. Error codes: NV_ERR_INVALID_ARGUMENT - Invalid parameter/s is passed. NV_ERR_NOT_SUPPORTED - SR-IOV heavy mode is disabled. */ NV_STATUS nvUvmInterfacePagingChannelsMap(uvmGpuAddressSpaceHandle srcVaSpace, UvmGpuPointer srcAddress, uvmGpuDeviceHandle device, NvU64 *dstAddress); /******************************************************************************* nvUvmInterfacePagingChannelsUnmap Unmap a VA returned by nvUvmInterfacePagingChannelsMap. SR-IOV heavy only. The implementation of this interface can acquire RM or GPU locks. Arguments: srcVaSpace[IN] - VA space handle that was passed to prevous mapping. srcAddress[IN] - virtual address that was passed to prevous mapping. device[IN] - device under which paging channels were allocated. */ void nvUvmInterfacePagingChannelsUnmap(uvmGpuAddressSpaceHandle srcVaSpace, UvmGpuPointer srcAddress, uvmGpuDeviceHandle device); /******************************************************************************* nvUvmInterfacePagingChannelPushStream Used for remote execution of the passed methods; the UVM driver uses this interface to ask the vGPU plugin to execute certain HW methods on its behalf. The callee should push the methods in the specified order i.e. is not allowed to do any reordering. The API is asynchronous. The UVM driver can wait on the remote execution by inserting a semaphore release method at the end of the method stream, and then loop until the semaphore value reaches the completion value indicated in the release method. The valid HW methods that can be passed by the UVM driver follow; the source functions listed contain the exact formatting (encoding) of the HW method used by the UVM driver for Ampere. - TLB invalidation targeting a VA range. See uvm_hal_volta_host_tlb_invalidate_va. - TLB invalidation targeting certain levels in the page tree (including the possibility of invalidating everything). See uvm_hal_pascal_host_tlb_invalidate_all. - Replayable fault replay. See uvm_hal_volta_replay_faults. - Replayable fault cancellation targeting a guest virtual address. See uvm_hal_volta_cancel_faults_va - Membar, scoped to device or to the entire system. See uvm_hal_pascal_host_membar_gpu and uvm_hal_pascal_host_membar_sys - Host semaphore acquire, see uvm_hal_turing_host_semaphore_acquire. The virtual address specified in the semaphore operation must lie within a buffer previously mapped by nvUvmInterfacePagingChannelsMap. - CE semaphore release, see uvm_hal_pascal_ce_semaphore_release. The virtual address specified in the semaphore operation must lie within a buffer previously mapped by nvUvmInterfacePagingChannelsMap. - 64 bits-wide memset, see uvm_hal_kepler_ce_memset_8. The destination address is a physical address in vidmem. - No-op, see uvm_hal_kepler_host_noop. Used to store the source buffer of a memcopy method within the input stream itself. - Memcopy, see uvm_hal_kepler_ce_memcopy. The destination address is a physical address in vidmem. The source address is an offset within methodStream, in bytes, indicating the location of the (inlined) source buffer. The copy size does not exceed 4KB. - CE semaphore release with timestamp, see uvm_hal_kepler_ce_semaphore_timestamp. The virtual address specified in the semaphore operation must lie within a buffer previously mapped by nvUvmInterfacePagingChannelsMap. - CE semaphore reduction, see uvm_hal_kepler_ce_semaphore_reduction_inc. The virtual address specified in the semaphore operation must lie within a buffer previously mapped by nvUvmInterfacePagingChannelsMap. Only invoked in SR-IOV heavy mode. NOTES: - This function uses a pre-allocated stack per paging channel (stored in the UvmGpuPagingChannel object) - This function DOES NOT acquire the RM API or GPU locks. That is because it is called during fault servicing, which could produce deadlocks. - Concurrent calls to this function using channels under same device are not allowed due to: a. pre-allocated stack b. the fact that internal RPC infrastructure doesn't acquire GPU lock. Therefore, locking is the caller's responsibility. - This function DOES NOT sleep (does not allocate memory or acquire locks) so it can be invoked while holding a spinlock. Arguments: channel[IN] - paging channel handle obtained via nvUvmInterfacePagingChannelAllocate methodStream[IN] - HW methods to be pushed to the paging channel. methodStreamSize[IN] - Size of methodStream, in bytes. The maximum push size is 128KB. Error codes: NV_ERR_INVALID_ARGUMENT - Invalid parameter/s is passed. NV_ERR_NOT_SUPPORTED - SR-IOV heavy mode is disabled. */ NV_STATUS nvUvmInterfacePagingChannelPushStream(UvmGpuPagingChannelHandle channel, char *methodStream, NvU32 methodStreamSize); #endif // _NV_UVM_INTERFACE_H_