/*************************************************************************/ /*! @File @Title Device Memory Management internal utility functions @Copyright Copyright (c) Imagination Technologies Ltd. All Rights Reserved @Description Utility functions used internally by device memory management code. @License Dual MIT/GPLv2 The contents of this file are subject to the MIT license as set out below. 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. Alternatively, the contents of this file may be used under the terms of the GNU General Public License Version 2 ("GPL") in which case the provisions of GPL are applicable instead of those above. If you wish to allow use of your version of this file only under the terms of GPL, and not to allow others to use your version of this file under the terms of the MIT license, indicate your decision by deleting the provisions above and replace them with the notice and other provisions required by GPL as set out in the file called "GPL-COPYING" included in this distribution. If you do not delete the provisions above, a recipient may use your version of this file under the terms of either the MIT license or GPL. This License is also included in this distribution in the file called "MIT-COPYING". EXCEPT AS OTHERWISE STATED IN A NEGOTIATED AGREEMENT: (A) 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; AND (B) 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. */ /**************************************************************************/ #include "allocmem.h" #include "img_types.h" #include "img_defs.h" #include "pvrsrv_error.h" #include "ra.h" #include "devicemem_utils.h" #include "client_mm_bridge.h" #include "client_cache_bridge.h" #if defined(PVRSRV_ENABLE_GPU_MEMORY_INFO) #include "client_ri_bridge.h" #if defined(__KERNEL__) #include "pvrsrv.h" #else #include "pvr_bridge_client.h" #endif #endif #if defined(PVRSRV_ENABLE_PROCESS_STATS) #include "proc_stats.h" #endif #if defined(__KERNEL__) #include "srvcore.h" #else #include "srvcore_intern.h" #endif /* SVM heap management support functions for CPU (un)mapping */ #define DEVMEM_MAP_SVM_USER_MANAGED_RETRY 10 static inline PVRSRV_ERROR DevmemCPUMapSVMKernelManaged(DEVMEM_HEAP *psHeap, DEVMEM_IMPORT *psImport, IMG_UINT64 *ui64MapAddress) { PVRSRV_ERROR eError; IMG_UINT64 ui64SvmMapAddr; IMG_UINT64 ui64SvmMapAddrEnd; IMG_UINT64 ui64SvmHeapAddrEnd; /* SVM heap management always has XXX_MANAGER_KERNEL unless we have triggered the fall back code-path in which case we should not be calling into this code-path */ PVR_ASSERT(psHeap->ui32HeapManagerFlags == DEVMEM_HEAP_MANAGER_KERNEL); /* By acquiring the CPU virtual address here, it essentially means we lock-down the virtual address for the duration of the life-cycle of the allocation until a de-allocation request comes in. Thus the allocation is guaranteed not to change its virtual address on the CPU during its life-time. NOTE: Import might have already been CPU Mapped before now, normally this is not a problem, see fall back */ eError = DevmemImportStructCPUMap(psImport); if (eError != PVRSRV_OK) { PVR_LOG_ERROR(eError, "DevmemImportStructCPUMap"); eError = PVRSRV_ERROR_DEVICEMEM_MAP_FAILED; goto failSVM; } /* Supplied kernel mmap virtual address is also device virtual address; calculate the heap & kernel supplied mmap virtual address limits */ ui64SvmMapAddr = (IMG_UINT64)(uintptr_t)psImport->sCPUImport.pvCPUVAddr; ui64SvmHeapAddrEnd = psHeap->sBaseAddress.uiAddr + psHeap->uiSize; ui64SvmMapAddrEnd = ui64SvmMapAddr + psImport->uiSize; PVR_ASSERT(ui64SvmMapAddr != (IMG_UINT64)0); /* SVM limit test may fail if processor has more virtual address bits than device */ if ((ui64SvmMapAddr >= ui64SvmHeapAddrEnd || ui64SvmMapAddrEnd > ui64SvmHeapAddrEnd) || (ui64SvmMapAddr & ~(ui64SvmHeapAddrEnd - 1))) { /* Unmap incompatible SVM virtual address, this may not release address if it was elsewhere CPU Mapped before call into this function */ DevmemImportStructCPUUnmap(psImport); /* Flag incompatible SVM mapping */ eError = PVRSRV_ERROR_BAD_MAPPING; goto failSVM; } *ui64MapAddress = ui64SvmMapAddr; failSVM: /* either OK, MAP_FAILED or BAD_MAPPING */ return eError; } static inline void DevmemCPUUnmapSVMKernelManaged(DEVMEM_HEAP *psHeap, DEVMEM_IMPORT *psImport) { PVR_UNREFERENCED_PARAMETER(psHeap); DevmemImportStructCPUUnmap(psImport); } static inline PVRSRV_ERROR DevmemCPUMapSVMUserManaged(DEVMEM_HEAP *psHeap, DEVMEM_IMPORT *psImport, IMG_UINT uiAlign, IMG_UINT64 *ui64MapAddress) { RA_LENGTH_T uiAllocatedSize; RA_BASE_T uiAllocatedAddr; IMG_UINT64 ui64SvmMapAddr; IMG_UINT uiRetry = 0; IMG_UINT32 ui32TmpPaddingCount = 0; DEVMEM_SIZE_T uiTmpPaddingSize = 0; PVRSRV_ERROR eError; /* If SVM heap management has transitioned to XXX_MANAGER_USER, this is essentially a fall back approach that ensures we continue to satisfy SVM alloc. This approach is not without hazards in that we may specify a virtual address that is already in use by the user process */ PVR_ASSERT(psHeap->ui32HeapManagerFlags == DEVMEM_HEAP_MANAGER_USER); /* Normally, for SVM heap allocations, CPUMap _must_ be done before DevMap; ideally the initial CPUMap should be done by SVM functions though this is not a hard requirement as long as the prior elsewhere obtained CPUMap virtual address meets SVM address requirements. This is a fall-back code-pathway so we have to test that this assumption holds before we progress any further */ OSLockAcquire(psImport->sCPUImport.hLock); if (psImport->sCPUImport.ui32RefCount) { /* Already CPU Mapped SVM heap allocation, this prior elsewhere obtained virtual address is responsible for the above XXX_MANAGER_KERNEL failure. As we are not responsible for this, we cannot progress any further so need to fail */ PVR_DPF((PVR_DBG_ERROR, "%s: Previously obtained CPU map address not SVM compatible" , __func__)); /* Revert SVM heap to DEVMEM_HEAP_MANAGER_KERNEL */ psHeap->ui32HeapManagerFlags = DEVMEM_HEAP_MANAGER_KERNEL; PVR_DPF((PVR_DBG_MESSAGE, "%s: Reverting SVM heap back to kernel managed", __func__)); OSLockRelease(psImport->sCPUImport.hLock); /* Do we need a more specific error code here */ eError = PVRSRV_ERROR_DEVICEMEM_ALREADY_MAPPED; goto failSVM; } OSLockRelease(psImport->sCPUImport.hLock); do { /* Next we proceed to instruct the kernel to use the RA_Alloc supplied virtual address to map-in this SVM import suballocation; there is no guarantee that this RA_Alloc virtual address may not collide with an already in-use VMA range in the process */ eError = RA_Alloc(psHeap->psQuantizedVMRA, psImport->uiSize, RA_NO_IMPORT_MULTIPLIER, 0, /* flags: this RA doesn't use flags*/ uiAlign, "SVM_Virtual_Alloc", &uiAllocatedAddr, &uiAllocatedSize, NULL /* don't care about per-import priv data */); if (eError != PVRSRV_OK) { PVR_LOG_ERROR(eError, "RA_Alloc"); #if defined(PVRSRV_ENABLE_PROCESS_STATS) if (eError == PVRSRV_ERROR_RA_REQUEST_ALLOC_FAIL) { PVRSRV_ERROR eErr; eErr = BridgePVRSRVStatsUpdateOOMStat(GetBridgeHandle(psHeap->psCtx->hDevConnection), PVRSRV_DEVICE_STAT_TYPE_OOM_VIRTMEM_COUNT, OSGetCurrentProcessID()); PVR_LOG_IF_ERROR(eErr, "BridgePVRSRVStatsUpdateOOMStat"); } #endif goto failSVM; } /* No reason for allocated virtual size to be different from the PMR's size */ psImport->sCPUImport.pvCPUVAddr = (void*)(uintptr_t)uiAllocatedAddr; PVR_ASSERT(uiAllocatedSize == psImport->uiSize); /* Map the import or allocation using the RA_Alloc virtual address; the kernel may fail the request if the supplied virtual address is already in-use in which case we re-try using another virtual address obtained from the RA_Alloc */ eError = DevmemImportStructCPUMap(psImport); if (eError != PVRSRV_OK) { if (psHeap->ui32SVMBasePaddingCount == 0) { ui32TmpPaddingCount++; uiTmpPaddingSize = uiAllocatedSize; } /* For now we simply discard failed RA_Alloc() obtained virtual address (i.e. plenty of virtual space), this prevents us from re-using these and furthermore essentially blacklists these addresses from future SVM consideration; We exit fall-back attempt if retry exceeds the fall-back retry limit */ if (uiRetry++ > DEVMEM_MAP_SVM_USER_MANAGED_RETRY) { PVR_DPF((PVR_DBG_ERROR, "%s: Cannot find SVM compatible address, bad mapping", __func__)); eError = PVRSRV_ERROR_BAD_MAPPING; goto failSVM; } } else { /* Found compatible SVM virtual address, set as device virtual address */ ui64SvmMapAddr = (IMG_UINT64)(uintptr_t)psImport->sCPUImport.pvCPUVAddr; } } while (eError != PVRSRV_OK); if (psHeap->ui32SVMBasePaddingCount == 0) { psHeap->ui32SVMBasePaddingCount = ui32TmpPaddingCount; psHeap->uiSVMBasePaddingSize = uiTmpPaddingSize; } *ui64MapAddress = ui64SvmMapAddr; failSVM: return eError; } static inline void DevmemCPUUnmapSVMUserManaged(DEVMEM_HEAP *psHeap, DEVMEM_IMPORT *psImport) { RA_BASE_T uiAllocatedAddr; /* We only free SVM compatible addresses, all addresses in the blacklist are essentially excluded from future RA_Alloc */ uiAllocatedAddr = psImport->sDeviceImport.sDevVAddr.uiAddr; RA_Free(psHeap->psQuantizedVMRA, uiAllocatedAddr); DevmemImportStructCPUUnmap(psImport); } static inline PVRSRV_ERROR DevmemImportStructDevMapSVM(DEVMEM_HEAP *psHeap, DEVMEM_IMPORT *psImport, IMG_UINT uiAlign, IMG_UINT64 *ui64MapAddress) { PVRSRV_ERROR eError; switch (psHeap->ui32HeapManagerFlags) { case DEVMEM_HEAP_MANAGER_KERNEL: eError = DevmemCPUMapSVMKernelManaged(psHeap, psImport, ui64MapAddress); if (eError == PVRSRV_ERROR_BAD_MAPPING) { /* If the SVM map address is outside of SVM heap limits, change heap type to DEVMEM_HEAP_MANAGER_USER */ psHeap->ui32HeapManagerFlags = DEVMEM_HEAP_MANAGER_USER; PVR_DPF((PVR_DBG_WARNING, "%s: Kernel managed SVM heap is now user managed", __func__)); /* Retry using user managed fall-back approach */ eError = DevmemCPUMapSVMUserManaged(psHeap, psImport, uiAlign, ui64MapAddress); } break; case DEVMEM_HEAP_MANAGER_USER: eError = DevmemCPUMapSVMUserManaged(psHeap, psImport, uiAlign, ui64MapAddress); break; default: eError = PVRSRV_ERROR_INVALID_PARAMS; break; } return eError; } static inline void DevmemImportStructDevUnmapSVM(DEVMEM_HEAP *psHeap, DEVMEM_IMPORT *psImport) { switch (psHeap->ui32HeapManagerFlags) { case DEVMEM_HEAP_MANAGER_KERNEL: DevmemCPUUnmapSVMKernelManaged(psHeap, psImport); break; case DEVMEM_HEAP_MANAGER_USER: DevmemCPUUnmapSVMUserManaged(psHeap, psImport); break; default: break; } } /* The Devmem import structure is the structure we use to manage memory that is "imported" (which is page granular) from the server into our process, this includes allocations. This allows memory to be imported without requiring any CPU or device mapping. Memory can then be mapped into the device or CPU on demand, but neither is required. */ IMG_INTERNAL void DevmemImportStructAcquire(DEVMEM_IMPORT *psImport) { IMG_INT iRefCount = OSAtomicIncrement(&psImport->hRefCount); PVR_UNREFERENCED_PARAMETER(iRefCount); PVR_ASSERT(iRefCount != 1); DEVMEM_REFCOUNT_PRINT("%s (%p) %d->%d", __func__, psImport, iRefCount-1, iRefCount); } IMG_INTERNAL IMG_BOOL DevmemImportStructRelease(DEVMEM_IMPORT *psImport) { IMG_INT iRefCount = OSAtomicDecrement(&psImport->hRefCount); PVR_ASSERT(iRefCount >= 0); DEVMEM_REFCOUNT_PRINT("%s (%p) %d->%d", __func__, psImport, iRefCount+1, iRefCount); if (iRefCount == 0) { PVRSRV_ERROR eError = DestroyServerResource(psImport->hDevConnection, NULL, BridgePMRUnrefPMR, psImport->hPMR); PVR_ASSERT(eError == PVRSRV_OK); OSLockDestroy(psImport->sCPUImport.hLock); OSLockDestroy(psImport->sDeviceImport.hLock); OSLockDestroy(psImport->hLock); OSFreeMem(psImport); return IMG_TRUE; } return IMG_FALSE; } IMG_INTERNAL void DevmemImportDiscard(DEVMEM_IMPORT *psImport) { PVR_ASSERT(OSAtomicRead(&psImport->hRefCount) == 0); OSLockDestroy(psImport->sCPUImport.hLock); OSLockDestroy(psImport->sDeviceImport.hLock); OSLockDestroy(psImport->hLock); OSFreeMem(psImport); } IMG_INTERNAL PVRSRV_ERROR DevmemMemDescAlloc(DEVMEM_MEMDESC **ppsMemDesc) { DEVMEM_MEMDESC *psMemDesc; PVRSRV_ERROR eError; /* Must be zeroed in case it needs to be freed before it is initialised */ psMemDesc = OSAllocZMem(sizeof(DEVMEM_MEMDESC)); PVR_GOTO_IF_NOMEM(psMemDesc, eError, failAlloc); eError = OSLockCreate(&psMemDesc->hLock); PVR_GOTO_IF_ERROR(eError, failMDLock); eError = OSLockCreate(&psMemDesc->sDeviceMemDesc.hLock); PVR_GOTO_IF_ERROR(eError, failDMDLock); eError = OSLockCreate(&psMemDesc->sCPUMemDesc.hLock); PVR_GOTO_IF_ERROR(eError, failCMDLock); OSAtomicWrite(&psMemDesc->hRefCount, 0); *ppsMemDesc = psMemDesc; return PVRSRV_OK; failCMDLock: OSLockDestroy(psMemDesc->sDeviceMemDesc.hLock); failDMDLock: OSLockDestroy(psMemDesc->hLock); failMDLock: OSFreeMem(psMemDesc); failAlloc: PVR_ASSERT(eError != PVRSRV_OK); return eError; } /* Init the MemDesc structure */ IMG_INTERNAL void DevmemMemDescInit(DEVMEM_MEMDESC *psMemDesc, IMG_DEVMEM_OFFSET_T uiOffset, DEVMEM_IMPORT *psImport, IMG_DEVMEM_SIZE_T uiSize) { DEVMEM_REFCOUNT_PRINT("%s (%p) %d->%d", __func__, psMemDesc, 0, 1); psMemDesc->psImport = psImport; psMemDesc->uiOffset = uiOffset; psMemDesc->sDeviceMemDesc.ui32RefCount = 0; psMemDesc->sCPUMemDesc.ui32RefCount = 0; psMemDesc->uiAllocSize = uiSize; psMemDesc->hPrivData = NULL; psMemDesc->ui32AllocationIndex = DEVICEMEM_HISTORY_ALLOC_INDEX_NONE; #if defined(DEBUG) psMemDesc->bPoisonOnFree = IMG_FALSE; #endif OSAtomicWrite(&psMemDesc->hRefCount, 1); } #if defined(DEBUG) IMG_INTERNAL void DevmemMemDescSetPoF(DEVMEM_MEMDESC *psMemDesc, PVRSRV_MEMALLOCFLAGS_T uiFlags) { if (PVRSRV_CHECK_POISON_ON_FREE(uiFlags)) { psMemDesc->bPoisonOnFree = IMG_TRUE; } } #endif IMG_INTERNAL void DevmemMemDescAcquire(DEVMEM_MEMDESC *psMemDesc) { IMG_INT iRefCount = 0; iRefCount = OSAtomicIncrement(&psMemDesc->hRefCount); DEVMEM_REFCOUNT_PRINT("%s (%p) %d->%d", __func__, psMemDesc, iRefCount-1, iRefCount); PVR_UNREFERENCED_PARAMETER(iRefCount); } #if defined(DEBUG) static void _DevmemPoisonOnFree(DEVMEM_MEMDESC *psMemDesc) { void *pvAddr = NULL; IMG_UINT8 *pui8CPUVAddr; PVRSRV_ERROR eError; eError = DevmemCPUMapCheckImportProperties(psMemDesc); PVR_LOG_RETURN_VOID_IF_ERROR(eError, "DevmemCPUMapCheckImportProperties"); OSLockAcquire(psMemDesc->sCPUMemDesc.hLock); eError = DevmemImportStructCPUMap(psMemDesc->psImport); OSLockRelease(psMemDesc->sCPUMemDesc.hLock); PVR_LOG_RETURN_VOID_IF_ERROR(eError, "DevmemImportStructCPUMap"); pui8CPUVAddr = psMemDesc->psImport->sCPUImport.pvCPUVAddr; pui8CPUVAddr += psMemDesc->uiOffset; pvAddr = pui8CPUVAddr; DevmemCPUMemSet(pvAddr, PVRSRV_POISON_ON_FREE_VALUE, psMemDesc->uiAllocSize, psMemDesc->psImport->uiFlags); if (PVRSRV_CHECK_CPU_CACHE_COHERENT(psMemDesc->psImport->uiFlags) || PVRSRV_CHECK_CPU_CACHE_INCOHERENT(psMemDesc->psImport->uiFlags)) { eError = BridgeCacheOpExec(GetBridgeHandle(psMemDesc->psImport->hDevConnection), psMemDesc->psImport->hPMR, (IMG_UINT64) (uintptr_t) pvAddr - psMemDesc->uiOffset, psMemDesc->uiOffset, psMemDesc->uiAllocSize, PVRSRV_CACHE_OP_FLUSH); PVR_LOG_IF_ERROR(eError, "BridgeCacheOpExec"); } DevmemImportStructCPUUnmap(psMemDesc->psImport); pvAddr = NULL; } #endif IMG_INTERNAL IMG_BOOL DevmemMemDescRelease(DEVMEM_MEMDESC *psMemDesc) { IMG_INT iRefCount; PVR_ASSERT(psMemDesc != NULL); iRefCount = OSAtomicDecrement(&psMemDesc->hRefCount); PVR_ASSERT(iRefCount >= 0); DEVMEM_REFCOUNT_PRINT("%s (%p) %d->%d", __func__, psMemDesc, iRefCount+1, iRefCount); if (iRefCount == 0) { #if defined(PVRSRV_ENABLE_GPU_MEMORY_INFO) if (PVRSRVIsBridgeEnabled(GetBridgeHandle(psMemDesc->psImport->hDevConnection), PVRSRV_BRIDGE_RI) && (psMemDesc->hRIHandle)) { PVRSRV_ERROR eError; eError = DestroyServerResource(psMemDesc->psImport->hDevConnection, NULL, BridgeRIDeleteMEMDESCEntry, psMemDesc->hRIHandle); PVR_LOG_IF_ERROR(eError, "BridgeRIDeleteMEMDESCEntry"); } #endif OSLockAcquire(psMemDesc->psImport->hLock); if (psMemDesc->psImport->uiProperties & DEVMEM_PROPERTIES_SUBALLOCATABLE) { /* As soon as the first sub-allocation on the psImport is freed * we might get dirty memory when reusing it. * We have to delete the ZEROED, CLEAN & POISONED flag */ psMemDesc->psImport->uiProperties &= ~(DEVMEM_PROPERTIES_IMPORT_IS_ZEROED | DEVMEM_PROPERTIES_IMPORT_IS_CLEAN | DEVMEM_PROPERTIES_IMPORT_IS_POISONED); OSLockRelease(psMemDesc->psImport->hLock); #if defined(DEBUG) if (psMemDesc->bPoisonOnFree) { _DevmemPoisonOnFree(psMemDesc); } #endif RA_Free(psMemDesc->psImport->sDeviceImport.psHeap->psSubAllocRA, psMemDesc->psImport->sDeviceImport.sDevVAddr.uiAddr + psMemDesc->uiOffset); } else { OSLockRelease(psMemDesc->psImport->hLock); DevmemImportStructRelease(psMemDesc->psImport); } OSLockDestroy(psMemDesc->sCPUMemDesc.hLock); OSLockDestroy(psMemDesc->sDeviceMemDesc.hLock); OSLockDestroy(psMemDesc->hLock); OSFreeMem(psMemDesc); return IMG_TRUE; } return IMG_FALSE; } IMG_INTERNAL void DevmemMemDescDiscard(DEVMEM_MEMDESC *psMemDesc) { PVR_ASSERT(OSAtomicRead(&psMemDesc->hRefCount) == 0); OSLockDestroy(psMemDesc->sCPUMemDesc.hLock); OSLockDestroy(psMemDesc->sDeviceMemDesc.hLock); OSLockDestroy(psMemDesc->hLock); OSFreeMem(psMemDesc); } IMG_INTERNAL PVRSRV_ERROR DevmemValidateParams(IMG_DEVMEM_SIZE_T uiSize, IMG_DEVMEM_ALIGN_T uiAlign, PVRSRV_MEMALLOCFLAGS_T *puiFlags) { if ((*puiFlags & PVRSRV_MEMALLOCFLAG_ZERO_ON_ALLOC) && (*puiFlags & PVRSRV_MEMALLOCFLAG_POISON_ON_ALLOC)) { PVR_DPF((PVR_DBG_ERROR, "%s: Zero on Alloc and Poison on Alloc are mutually exclusive.", __func__)); return PVRSRV_ERROR_INVALID_PARAMS; } if (uiAlign & (uiAlign-1)) { PVR_DPF((PVR_DBG_ERROR, "%s: The requested alignment is not a power of two.", __func__)); return PVRSRV_ERROR_INVALID_PARAMS; } if (uiSize == 0) { PVR_DPF((PVR_DBG_ERROR, "%s: Please request a non-zero size value.", __func__)); return PVRSRV_ERROR_INVALID_PARAMS; } /* If zero or poison flags are set we have to have write access to the page. */ if (PVRSRV_CHECK_ZERO_ON_ALLOC(*puiFlags) || PVRSRV_CHECK_POISON_ON_ALLOC(*puiFlags) || #if defined(DEBUG) PVRSRV_CHECK_POISON_ON_FREE(*puiFlags) || #endif PVRSRV_CHECK_CPU_WRITEABLE(*puiFlags)) { (*puiFlags) |= PVRSRV_MEMALLOCFLAG_CPU_WRITEABLE | PVRSRV_MEMALLOCFLAG_CPU_READABLE; } return PVRSRV_OK; } /* Allocate and init an import structure */ IMG_INTERNAL PVRSRV_ERROR DevmemImportStructAlloc(SHARED_DEV_CONNECTION hDevConnection, DEVMEM_IMPORT **ppsImport) { DEVMEM_IMPORT *psImport; PVRSRV_ERROR eError; psImport = OSAllocMem(sizeof(*psImport)); PVR_RETURN_IF_FALSE(psImport != NULL, PVRSRV_ERROR_OUT_OF_MEMORY); /* Setup some known bad values for things we don't have yet */ psImport->sDeviceImport.hReservation = LACK_OF_RESERVATION_POISON; psImport->sDeviceImport.psHeap = NULL; psImport->sDeviceImport.bMapped = IMG_FALSE; eError = OSLockCreate(&psImport->sDeviceImport.hLock); PVR_GOTO_IF_ERROR(eError, failDIOSLockCreate); psImport->sCPUImport.hOSMMapData = NULL; psImport->sCPUImport.pvCPUVAddr = NULL; eError = OSLockCreate(&psImport->sCPUImport.hLock); PVR_GOTO_IF_ERROR(eError, failCIOSLockCreate); /* Set up common elements */ psImport->hDevConnection = hDevConnection; /* Setup properties */ psImport->uiProperties = 0; /* Setup refcounts */ psImport->sDeviceImport.ui32RefCount = 0; psImport->sCPUImport.ui32RefCount = 0; OSAtomicWrite(&psImport->hRefCount, 0); /* Create the lock */ eError = OSLockCreate(&psImport->hLock); PVR_GOTO_IF_ERROR(eError, failILockAlloc); *ppsImport = psImport; return PVRSRV_OK; failILockAlloc: OSLockDestroy(psImport->sCPUImport.hLock); failCIOSLockCreate: OSLockDestroy(psImport->sDeviceImport.hLock); failDIOSLockCreate: OSFreeMem(psImport); PVR_ASSERT(eError != PVRSRV_OK); return eError; } /* Initialise the import structure */ IMG_INTERNAL void DevmemImportStructInit(DEVMEM_IMPORT *psImport, IMG_DEVMEM_SIZE_T uiSize, IMG_DEVMEM_ALIGN_T uiAlign, PVRSRV_MEMALLOCFLAGS_T uiFlags, IMG_HANDLE hPMR, DEVMEM_PROPERTIES_T uiProperties) { DEVMEM_REFCOUNT_PRINT("%s (%p) %d->%d", __func__, psImport, 0, 1); psImport->uiSize = uiSize; psImport->uiAlign = uiAlign; psImport->uiFlags = uiFlags; psImport->hPMR = hPMR; psImport->uiProperties = uiProperties; OSAtomicWrite(&psImport->hRefCount, 1); } /* Allocate the requested device virtual address region * from the heap */ static PVRSRV_ERROR DevmemReserveVARange(DEVMEM_HEAP *psHeap, DEVMEM_SIZE_T uiSize, IMG_UINT uiAlign, RA_LENGTH_T *puiAllocatedSize, IMG_UINT64 ui64OptionalMapAddress) { PVRSRV_ERROR eError; /* Allocate space in the VM */ eError = RA_Alloc_Range(psHeap->psQuantizedVMRA, uiSize, 0, uiAlign, ui64OptionalMapAddress, puiAllocatedSize); if (PVRSRV_OK != eError) { #if defined(PVRSRV_ENABLE_PROCESS_STATS) if ((eError == PVRSRV_ERROR_RA_REQUEST_ALLOC_FAIL) || (eError == PVRSRV_ERROR_RA_REQUEST_VIRT_ADDR_FAIL)) { PVRSRV_ERROR eErr; eErr = BridgePVRSRVStatsUpdateOOMStat(GetBridgeHandle(psHeap->psCtx->hDevConnection), PVRSRV_DEVICE_STAT_TYPE_INVALID_VIRTMEM, OSGetCurrentProcessID()); PVR_LOG_IF_ERROR(eErr, "BridgePVRSRVStatsUpdateOOMStat"); } #endif return eError; } /* No reason for the allocated virtual size to be different from the PMR's size */ PVR_ASSERT(*puiAllocatedSize == uiSize); return PVRSRV_OK; } /* Map an import to the device */ IMG_INTERNAL PVRSRV_ERROR DevmemImportStructDevMap(DEVMEM_HEAP *psHeap, IMG_BOOL bMap, DEVMEM_IMPORT *psImport, IMG_UINT64 ui64OptionalMapAddress) { DEVMEM_DEVICE_IMPORT *psDeviceImport; RA_BASE_T uiAllocatedAddr; RA_LENGTH_T uiAllocatedSize; IMG_DEV_VIRTADDR sBase; PVRSRV_ERROR eError; IMG_UINT uiAlign; IMG_BOOL bDestroyed = IMG_FALSE; /* Round the provided import alignment to the configured heap alignment */ uiAlign = 1ULL << psHeap->uiLog2ImportAlignment; uiAlign = PVR_ALIGN(psImport->uiAlign, uiAlign); psDeviceImport = &psImport->sDeviceImport; OSLockAcquire(psDeviceImport->hLock); DEVMEM_REFCOUNT_PRINT("%s (%p) %d->%d", __func__, psImport, psDeviceImport->ui32RefCount, psDeviceImport->ui32RefCount+1); if (psDeviceImport->ui32RefCount++ == 0) { DevmemImportStructAcquire(psImport); OSAtomicIncrement(&psHeap->hImportCount); if (PVRSRV_CHECK_SVM_ALLOC(psImport->uiFlags)) { /* SVM (shared virtual memory) imports or allocations always need to acquire CPU virtual address first as address is used to map the allocation into the device virtual address space; i.e. the virtual address of the allocation for both the CPU/GPU must be identical. */ eError = DevmemImportStructDevMapSVM(psHeap, psImport, uiAlign, &ui64OptionalMapAddress); PVR_GOTO_IF_ERROR(eError, failVMRAAlloc); } if (ui64OptionalMapAddress == 0) { /* If heap is _completely_ managed by USER or KERNEL, we shouldn't * be here, as this is RA manager code-path */ if (psHeap->ui32HeapManagerFlags == DEVMEM_HEAP_MANAGER_USER || psHeap->ui32HeapManagerFlags == DEVMEM_HEAP_MANAGER_KERNEL) { PVR_DPF((PVR_DBG_ERROR, psHeap->ui32HeapManagerFlags == DEVMEM_HEAP_MANAGER_USER ? "%s: Heap is user managed, please use PVRSRVMapToDeviceAddress().": "%s: Heap is kernel managed, use right allocation flags (e.g. SVM).", __func__)); PVR_GOTO_WITH_ERROR(eError, PVRSRV_ERROR_INVALID_PARAMS, failVMRAAlloc); } if (psHeap->ui32HeapManagerFlags == DEVMEM_HEAP_MANAGER_UNKNOWN) { /* Only set the heap manager (to RA) at first map when heap manager * is unknown. It might be a dual heap (both, user and RA managed), * in which case heap manager is set at creation time */ psHeap->ui32HeapManagerFlags = DEVMEM_HEAP_MANAGER_RA; } /* Allocate space in the VM */ eError = RA_Alloc(psHeap->psQuantizedVMRA, psImport->uiSize, RA_NO_IMPORT_MULTIPLIER, 0, /* flags: this RA doesn't use flags*/ uiAlign, "Virtual_Alloc", &uiAllocatedAddr, &uiAllocatedSize, NULL /* don't care about per-import priv data */ ); if (PVRSRV_OK != eError) { #if defined(PVRSRV_ENABLE_PROCESS_STATS) if (eError == PVRSRV_ERROR_RA_REQUEST_ALLOC_FAIL) { PVRSRV_ERROR eErr; eErr = BridgePVRSRVStatsUpdateOOMStat(GetBridgeHandle(psHeap->psCtx->hDevConnection), PVRSRV_DEVICE_STAT_TYPE_OOM_VIRTMEM_COUNT, OSGetCurrentProcessID()); PVR_LOG_IF_ERROR(eErr, "BridgePVRSRVStatsUpdateOOMStat"); } #endif PVR_GOTO_WITH_ERROR(eError, PVRSRV_ERROR_DEVICEMEM_OUT_OF_DEVICE_VM, failVMRAAlloc); } /* No reason for the allocated virtual size to be different from the PMR's size */ PVR_ASSERT(uiAllocatedSize == psImport->uiSize); sBase.uiAddr = uiAllocatedAddr; } else { IMG_UINT64 ui64ValidEndAddr; /* Ensure supplied ui64OptionalMapAddress is within heap range */ ui64ValidEndAddr = psHeap->sBaseAddress.uiAddr + psHeap->uiSize; if ((ui64OptionalMapAddress + psImport->uiSize > ui64ValidEndAddr) || (ui64OptionalMapAddress < psHeap->sBaseAddress.uiAddr)) { PVR_DPF((PVR_DBG_ERROR, "%s: ui64OptionalMapAddress %p is outside of heap limits <%p:%p>." , __func__ , (void*)(uintptr_t)ui64OptionalMapAddress , (void*)(uintptr_t)psHeap->sBaseAddress.uiAddr , (void*)(uintptr_t)ui64ValidEndAddr)); PVR_GOTO_WITH_ERROR(eError, PVRSRV_ERROR_INVALID_PARAMS, failVMRAAlloc); } switch (psHeap->ui32HeapManagerFlags) { case DEVMEM_HEAP_MANAGER_UNKNOWN: /* DEVMEM_HEAP_MANAGER_USER can apply to _any_ heap and can only * be determined here. This heap type transitions from * DEVMEM_HEAP_MANAGER_UNKNOWN to DEVMEM_HEAP_MANAGER_USER on * 1st alloc. */ psHeap->ui32HeapManagerFlags = DEVMEM_HEAP_MANAGER_USER; break; case DEVMEM_HEAP_MANAGER_USER: case DEVMEM_HEAP_MANAGER_KERNEL: if (! psHeap->uiSize) { PVR_DPF((PVR_DBG_ERROR, psHeap->ui32HeapManagerFlags == DEVMEM_HEAP_MANAGER_USER ? "%s: Heap DEVMEM_HEAP_MANAGER_USER is disabled.": "%s: Heap DEVMEM_HEAP_MANAGER_KERNEL is disabled." , __func__)); PVR_GOTO_WITH_ERROR(eError, PVRSRV_ERROR_INVALID_HEAP, failVMRAAlloc); } break; case DEVMEM_HEAP_MANAGER_DUAL_USER_RA: /* When the heap is dual managed, ensure supplied ui64OptionalMapAddress * and import size are within heap address space range */ if (ui64OptionalMapAddress + psImport->uiSize <= psHeap->sBaseAddress.uiAddr + psHeap->uiReservedRegionSize) { break; } else { /* Allocate requested VM range */ eError = DevmemReserveVARange(psHeap, psImport->uiSize, uiAlign, &uiAllocatedSize, ui64OptionalMapAddress); if (eError != PVRSRV_OK) { PVR_GOTO_WITH_ERROR(eError, PVRSRV_ERROR_DEVICEMEM_VA_ALLOC_FAILED, failVMRAAlloc); } } break; case DEVMEM_HEAP_MANAGER_RA: /* Allocate requested VM range */ eError = DevmemReserveVARange(psHeap, psImport->uiSize, uiAlign, &uiAllocatedSize, ui64OptionalMapAddress); if (eError != PVRSRV_OK) { PVR_GOTO_WITH_ERROR(eError, PVRSRV_ERROR_DEVICEMEM_VA_ALLOC_FAILED, failVMRAAlloc); } break; default: break; } if (ui64OptionalMapAddress & ((1 << psHeap->uiLog2Quantum) - 1)) { PVR_DPF((PVR_DBG_ERROR, "%s: Invalid address to map to. Please provide an " "address aligned to a page multiple of the heap." , __func__)); PVR_GOTO_WITH_ERROR(eError, PVRSRV_ERROR_INVALID_PARAMS, failVMRAAlloc); } if (psImport->uiSize & ((1 << psHeap->uiLog2Quantum) - 1)) { PVR_DPF((PVR_DBG_ERROR, "%s: Invalid heap to map to. " "Please choose a heap that can handle smaller page sizes." , __func__)); PVR_GOTO_WITH_ERROR(eError, PVRSRV_ERROR_INVALID_PARAMS, failVMRAAlloc); } uiAllocatedAddr = ui64OptionalMapAddress; uiAllocatedSize = psImport->uiSize; sBase.uiAddr = uiAllocatedAddr; } if (psHeap->bPremapped) { /* No virtual address reservation and mapping are required for * memory that is already pre-mapped e.g. FW heaps in VZ configs */ psDeviceImport->hReservation = LACK_OF_RESERVATION_POISON; } else { PVRSRV_MEMALLOCFLAGS_T uiFlags; uiFlags = psImport->uiFlags & PVRSRV_MEMALLOCFLAGS_PERMAPPINGFLAGSMASK; if (bMap) { eError = BridgeDevmemIntReserveRangeAndMapPMR(GetBridgeHandle(psHeap->psCtx->hDevConnection), psHeap->hDevMemServerHeap, sBase, uiAllocatedSize, psImport->hPMR, uiFlags, &psDeviceImport->hReservation); PVR_GOTO_IF_ERROR(eError, failReserve); psDeviceImport->bMapped = IMG_TRUE; } else { /* Setup page tables for the allocated VM space */ eError = BridgeDevmemIntReserveRange(GetBridgeHandle(psHeap->psCtx->hDevConnection), psHeap->hDevMemServerHeap, sBase, uiAllocatedSize, uiFlags, &psDeviceImport->hReservation); PVR_GOTO_IF_ERROR(eError, failReserve); } } /* Setup device mapping specific parts of the mapping info */ psDeviceImport->sDevVAddr.uiAddr = uiAllocatedAddr; psDeviceImport->psHeap = psHeap; } else { /* Check that we've been asked to map it into the same heap 2nd time around */ if (psHeap != psDeviceImport->psHeap) { PVR_GOTO_WITH_ERROR(eError, PVRSRV_ERROR_INVALID_HEAP, failParams); } } OSLockRelease(psDeviceImport->hLock); return PVRSRV_OK; failReserve: if (ui64OptionalMapAddress == 0) { RA_Free(psHeap->psQuantizedVMRA, uiAllocatedAddr); } failVMRAAlloc: if ((ui64OptionalMapAddress) && PVRSRV_CHECK_SVM_ALLOC(psImport->uiFlags)) { DevmemImportStructDevUnmapSVM(psHeap, psImport); } bDestroyed = DevmemImportStructRelease(psImport); OSAtomicDecrement(&psHeap->hImportCount); failParams: if (!bDestroyed) { psDeviceImport->ui32RefCount--; OSLockRelease(psDeviceImport->hLock); } PVR_ASSERT(eError != PVRSRV_OK); return eError; } /* Unmap an import from the Device */ IMG_INTERNAL IMG_BOOL DevmemImportStructDevUnmap(DEVMEM_IMPORT *psImport) { PVRSRV_ERROR eError; DEVMEM_DEVICE_IMPORT *psDeviceImport; psDeviceImport = &psImport->sDeviceImport; OSLockAcquire(psDeviceImport->hLock); DEVMEM_REFCOUNT_PRINT("%s (%p) %d->%d", __func__, psImport, psDeviceImport->ui32RefCount, psDeviceImport->ui32RefCount-1); if (--psDeviceImport->ui32RefCount == 0) { DEVMEM_HEAP *psHeap = psDeviceImport->psHeap; if (!psHeap->bPremapped) { eError = DestroyServerResource(psImport->hDevConnection, NULL, BridgeDevmemIntUnreserveRange, psDeviceImport->hReservation); PVR_ASSERT(eError == PVRSRV_OK); } psDeviceImport->bMapped = IMG_FALSE; psDeviceImport->hReservation = LACK_OF_RESERVATION_POISON; /* DEVMEM_HEAP_MANAGER_RA can also come from a dual managed heap in which case, we need to check if the allocated VA falls within RA managed range */ if ((psHeap->ui32HeapManagerFlags & DEVMEM_HEAP_MANAGER_RA) && psDeviceImport->sDevVAddr.uiAddr >= (psHeap->sBaseAddress.uiAddr + psHeap->uiReservedRegionSize) && psDeviceImport->sDevVAddr.uiAddr < (psHeap->sBaseAddress.uiAddr + psHeap->uiSize)) { RA_Free(psHeap->psQuantizedVMRA, psDeviceImport->sDevVAddr.uiAddr); } if (PVRSRV_CHECK_SVM_ALLOC(psImport->uiFlags)) { DevmemImportStructDevUnmapSVM(psHeap, psImport); } OSLockRelease(psDeviceImport->hLock); DevmemImportStructRelease(psImport); OSAtomicDecrement(&psHeap->hImportCount); return IMG_TRUE; } else { OSLockRelease(psDeviceImport->hLock); return IMG_FALSE; } } /* Map an import into the CPU */ IMG_INTERNAL PVRSRV_ERROR DevmemImportStructCPUMap(DEVMEM_IMPORT *psImport) { PVRSRV_ERROR eError; DEVMEM_CPU_IMPORT *psCPUImport; size_t uiMappingLength; psCPUImport = &psImport->sCPUImport; OSLockAcquire(psCPUImport->hLock); DEVMEM_REFCOUNT_PRINT("%s (%p) %d->%d", __func__, psImport, psCPUImport->ui32RefCount, psCPUImport->ui32RefCount+1); if (psCPUImport->ui32RefCount++ == 0) { DevmemImportStructAcquire(psImport); eError = OSMMapPMR(GetBridgeHandle(psImport->hDevConnection), psImport->hPMR, psImport->uiSize, psImport->uiFlags, &psCPUImport->hOSMMapData, &psCPUImport->pvCPUVAddr, &uiMappingLength); PVR_GOTO_IF_ERROR(eError, failMap); /* MappingLength might be rounded up to page size */ PVR_ASSERT(uiMappingLength >= psImport->uiSize); } OSLockRelease(psCPUImport->hLock); return PVRSRV_OK; failMap: psCPUImport->ui32RefCount--; if (!DevmemImportStructRelease(psImport)) { OSLockRelease(psCPUImport->hLock); } PVR_ASSERT(eError != PVRSRV_OK); return eError; } /* Unmap an import from the CPU */ IMG_INTERNAL void DevmemImportStructCPUUnmap(DEVMEM_IMPORT *psImport) { DEVMEM_CPU_IMPORT *psCPUImport; psCPUImport = &psImport->sCPUImport; OSLockAcquire(psCPUImport->hLock); DEVMEM_REFCOUNT_PRINT("%s (%p) %d->%d", __func__, psImport, psCPUImport->ui32RefCount, psCPUImport->ui32RefCount-1); if (--psCPUImport->ui32RefCount == 0) { /* psImport->uiSize is a 64-bit quantity whereas the 5th * argument to OSUnmapPMR is a 32-bit quantity on 32-bit systems * hence a compiler warning of implicit cast and loss of data. * Added explicit cast and assert to remove warning. */ #if defined(__linux__) && defined(__i386__) PVR_ASSERT(psImport->uiSizehDevConnection), psImport->hPMR, psCPUImport->hOSMMapData, psCPUImport->pvCPUVAddr, (size_t)psImport->uiSize); psCPUImport->hOSMMapData = NULL; psCPUImport->pvCPUVAddr = NULL; OSLockRelease(psCPUImport->hLock); DevmemImportStructRelease(psImport); } else { OSLockRelease(psCPUImport->hLock); } }