NVIDIA/open-gpu-kernel-modules
1
0
Fork 0
mirror of https://github.com/NVIDIA/open-gpu-kernel-modules.git synced 2026-01-31 13:39:47 +00:00
Code Issues Packages Projects Releases Wiki Activity

550.40.07

Browse Source
This commit is contained in:
Bernhard Stoeckner
2024-01-24 17:51:53 +01:00
parent bb2dac1f20
commit 91676d6628
1411 changed files with 261367 additions and 145959 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
src/common/unix/common/utils/unix_rm_handle.c
View File

@@ -195,8 +195,8 @@ NvU32 nvGenerateUnixRmHandleInternal(NVUnixRmHandleAllocatorPtr pAllocator)
/* Find free handle */
handleId = 1;
while (USED(pAllocator->bitmap, handleId) &&
(handleId <= pAllocator->maxHandles)) {
while ((handleId <= pAllocator->maxHandles) &&
USED(pAllocator->bitmap, handleId)) {
handleId++;
}

12
src/common/unix/nvidia-push/interface/nvidia-push-methods.h
View File

@@ -72,6 +72,18 @@ static inline void nvPushSetMethodDataU64(NvPushChannelPtr p, const NvU64 data)
__nvPushSetMethodDataSegmentU64(&p->main, data);
}
/* Little-endian: least significant bits first. */
static inline void __nvPushSetMethodDataSegmentU64LE(NvPushChannelSegmentPtr s, const NvU64 data)
{
__nvPushSetMethodDataSegment(s, NvU64_LO32(data));
__nvPushSetMethodDataSegment(s, NvU64_HI32(data));
}
static inline void nvPushSetMethodDataU64LE(NvPushChannelPtr p, const NvU64 data)
{
__nvPushSetMethodDataSegmentU64LE(&p->main, data);
}
void __nvPushMoveDWORDS(NvU32* dst, const NvU32* src, int dwords);
static inline void

27
src/common/unix/nvidia-push/src/nvidia-push.c
View File

@@ -332,13 +332,7 @@ static void InsertProgressTracker(NvPushChannelPtr p, NvU32 putOffset,
DRF_DEF(C36F, _SEM_EXECUTE, _RELEASE_WFI, _DIS));
__nvPushStart(p, progressTracker, 0, NVC36F_SEM_ADDR_LO, 5, _INC_METHOD);
/*
* NVC36F_SEM_ADDR_{LO,HI} is backwards from most 64-bit
* addresses spread over two methods, so we cannot use
* __nvPushSetMethodDataSegmentU64().
*/
__nvPushSetMethodDataSegment(segment, NvU64_LO32(p->progressSemaphore.gpuVA));
__nvPushSetMethodDataSegment(segment, NvU64_HI32(p->progressSemaphore.gpuVA));
__nvPushSetMethodDataSegmentU64LE(segment, p->progressSemaphore.gpuVA);
__nvPushSetMethodDataSegment(segment, payload);
__nvPushSetMethodDataSegment(segment, 0);
__nvPushSetMethodDataSegment(segment, semaphoreOperation);
@@ -638,11 +632,16 @@ NvBool __nvPushTestPushBuffer(NvPushChannelPtr p)
/*
* Immediately after allocating the pushbuffer, push a channel NOP and
* babysit the channel until it's consumed as a quick sanity check.
* Note we use a full long timeout (10 seconds) when performing this
* sanity test. In normal operation, idling will happen very quickly.
* However, when the GPU is under heavy load in stress tests, it can
* take much longer to idle the channel.
*/
WriteGetOffset(p, 0);
ret = IdleChannel(p, FALSE /* progressTrackerWFI */,
NV_PUSH_NOTIFIER_SHORT_TIMEOUT);
NV_PUSH_NOTIFIER_LONG_TIMEOUT);
if (!ret) {
nvPushImportLogError(p->pDevice, "Failed to initialize DMA.");
@@ -1047,10 +1046,8 @@ static void VoltaReleaseTimelineSemaphore(
NvU64 val)
{
nvPushMethod(p, 0, NVC36F_SEM_ADDR_LO, 5);
nvPushSetMethodData(p, NvU64_LO32(gpuAddress));
nvPushSetMethodData(p, NvU64_HI32(gpuAddress)); // NVC36F_SEM_ADDR_HI
nvPushSetMethodData(p, NvU64_LO32(val)); // NVC36F_SEM_PAYLOAD_LO
nvPushSetMethodData(p, NvU64_HI32(val)); // NVC36F_SEM_PAYLOAD_HI
nvPushSetMethodDataU64LE(p, gpuAddress); // NVC36F_SEM_ADDR_LO/HI
nvPushSetMethodDataU64LE(p, val); // NVC36F_SEM_PAYLOAD_LO/HI
nvPushSetMethodData(p, // NVC36F_SEM_EXECUTE
DRF_DEF(C36F, _SEM_EXECUTE, _OPERATION, _RELEASE) |
DRF_DEF(C36F, _SEM_EXECUTE, _RELEASE_WFI, _EN) |
@@ -1067,10 +1064,8 @@ static void VoltaAcquireTimelineSemaphore(
NvU64 val)
{
nvPushMethod(p, 0, NVC36F_SEM_ADDR_LO, 5);
nvPushSetMethodData(p, NvU64_LO32(gpuAddress));
nvPushSetMethodData(p, NvU64_HI32(gpuAddress)); // NVC36F_SEM_ADDR_HI
nvPushSetMethodData(p, NvU64_LO32(val)); // NVC36F_SEM_PAYLOAD_LO
nvPushSetMethodData(p, NvU64_HI32(val)); // NVC36F_SEM_PAYLOAD_HI
nvPushSetMethodDataU64LE(p, gpuAddress); // NVC36F_SEM_ADDR_LO/HI
nvPushSetMethodDataU64LE(p, val); // NVC36F_SEM_PAYLOAD_LO/HI
nvPushSetMethodData(p, // NVC36F_SEM_EXECUTE
DRF_DEF(C36F, _SEM_EXECUTE, _OPERATION, _ACQ_STRICT_GEQ) |
DRF_DEF(C36F, _SEM_EXECUTE, _ACQUIRE_SWITCH_TSG, _EN) |