archive/rpcs3
1
0
Fork 0
mirror of https://github.com/RPCS3/rpcs3.git synced 2025-07-08 16:01:42 +12:00
Code Issues Projects Releases Packages Wiki Activity Actions

SPURS: Update kernel to use lock line reservations

Browse source
This commit is contained in:
S Gopal Rajagopal 2015-01-29 20:20:34 +05:30
parent a7728c9067
commit 62e2d8d9a7
7 changed files with 414 additions and 297 deletions
Download patch file Download diff file Expand all files Collapse all files

8
rpcs3/Emu/Cell/MFC.h
View file

@ -35,6 +35,14 @@ enum
MFC_GETLLAR_SUCCESS = 4, MFC_GETLLAR_SUCCESS = 4,
}; };
// MFC Write Tag Status Update Request Channel (ch23) operations
enum
{
MFC_TAG_UPDATE_IMMEDIATE = 0,
MFC_TAG_UPDATE_ANY = 1,
MFC_TAG_UPDATE_ALL = 2,
};
enum enum
{ {
MFC_SPU_TO_PPU_MAILBOX_STATUS_MASK = 0x000000FF, MFC_SPU_TO_PPU_MAILBOX_STATUS_MASK = 0x000000FF,

9
rpcs3/Emu/Cell/SPUThread.cpp
View file

@ -1060,7 +1060,14 @@ void SPUThread::StopAndSignal(u32 code)
case 0x003: case 0x003:
{ {
m_code3_func(*this); auto iter = m_addr_to_hle_function_map.find(PC);
assert(iter != m_addr_to_hle_function_map.end());
auto return_to_caller = iter->second(*this);
if (return_to_caller)
{
SetBranch(GPR[0]._u32[3] & 0x3fffc);
}
break; break;
} }

31
rpcs3/Emu/Cell/SPUThread.h
View file

@ -290,6 +290,8 @@ public:
u32 m_event_mask; u32 m_event_mask;
u32 m_events; u32 m_events;
std::unordered_map<u32, std::function<bool(SPUThread& SPU)>> m_addr_to_hle_function_map;
struct IntrTag struct IntrTag
{ {
u32 enabled; // 1 == true u32 enabled; // 1 == true
@ -509,8 +511,35 @@ public:
void WriteLS64 (const u32 lsa, const u64& data) const { vm::write64 (lsa + m_offset, data); } void WriteLS64 (const u32 lsa, const u64& data) const { vm::write64 (lsa + m_offset, data); }
void WriteLS128(const u32 lsa, const u128& data) const { vm::write128(lsa + m_offset, data); } void WriteLS128(const u32 lsa, const u128& data) const { vm::write128(lsa + m_offset, data); }
void RegisterHleFuncion(u32 addr, std::function<bool(SPUThread & SPU)> function)
{
m_addr_to_hle_function_map[addr] = function;
WriteLS32(addr, 0x00000003); // STOP 3
}
void UnregisterHleFunction(u32 addr)
{
WriteLS32(addr, 0x00200000); // NOP
m_addr_to_hle_function_map.erase(addr);
}
void UnregisterHleFunctions(u32 start_addr, u32 end_addr)
{
for (auto iter = m_addr_to_hle_function_map.begin(); iter != m_addr_to_hle_function_map.end();)
{
if (iter->first >= start_addr && iter->first <= end_addr)
{
WriteLS32(iter->first, 0x00200000); // NOP
m_addr_to_hle_function_map.erase(iter++);
}
else
{
iter++;
}
}
}
std::function<void(SPUThread& SPU)> m_custom_task; std::function<void(SPUThread& SPU)> m_custom_task;
std::function<void(SPUThread& SPU)> m_code3_func;
public: public:
SPUThread(CPUThreadType type = CPU_THREAD_SPU); SPUThread(CPUThreadType type = CPU_THREAD_SPU);

17
rpcs3/Emu/SysCalls/Modules/cellSpurs.cpp
View file

@ -26,7 +26,7 @@ extern u32 libsre;
extern u32 libsre_rtoc; extern u32 libsre_rtoc;
#endif #endif
void spursKernelMain(SPUThread & spu); bool spursKernelMain(SPUThread & spu);
s64 cellSpursLookUpTasksetAddress(vm::ptr<CellSpurs> spurs, vm::ptr<CellSpursTaskset> taskset, u32 id); s64 cellSpursLookUpTasksetAddress(vm::ptr<CellSpurs> spurs, vm::ptr<CellSpursTaskset> taskset, u32 id);
s64 _cellSpursSendSignal(vm::ptr<CellSpursTaskset> taskset, u32 taskID); s64 _cellSpursSendSignal(vm::ptr<CellSpursTaskset> taskset, u32 taskID);
@ -156,6 +156,7 @@ s64 spursInit(
} }
#else #else
spurs->m.spuImg.addr = (u32)Memory.Alloc(0x40000, 4096); spurs->m.spuImg.addr = (u32)Memory.Alloc(0x40000, 4096);
spurs->m.spuImg.entry_point = isSecond ? CELL_SPURS_KERNEL2_ENTRY_ADDR : CELL_SPURS_KERNEL1_ENTRY_ADDR;
#endif #endif
s32 tgt = SYS_SPU_THREAD_GROUP_TYPE_NORMAL; s32 tgt = SYS_SPU_THREAD_GROUP_TYPE_NORMAL;
@ -179,17 +180,11 @@ s64 spursInit(
name += "CellSpursKernel0"; name += "CellSpursKernel0";
for (s32 num = 0; num < nSpus; num++, name[name.size() - 1]++) for (s32 num = 0; num < nSpus; num++, name[name.size() - 1]++)
{ {
spurs->m.spus[num] = spu_thread_initialize(tg, num, spurs->m.spuImg, name, SYS_SPU_THREAD_OPTION_DEC_SYNC_TB_ENABLE, 0, 0, 0, 0, [spurs, num](SPUThread& SPU) auto spu = spu_thread_initialize(tg, num, spurs->m.spuImg, name, SYS_SPU_THREAD_OPTION_DEC_SYNC_TB_ENABLE, num, spurs.addr(), 0, 0);
{ #ifndef PRX_DEBUG_XXX
SPU.GPR[3]._u32[3] = num; spu->RegisterHleFuncion(spurs->m.spuImg.entry_point, spursKernelMain);
SPU.GPR[4]._u64[1] = spurs.addr();
#ifdef PRX_DEBUG_XXX
return SPU.FastCall(SPU.PC);
#endif #endif
spurs->m.spus[num] = spu->GetId();
spursKernelMain(SPU);
})->GetId();
} }
if (flags & SAF_SPU_PRINTF_ENABLED) if (flags & SAF_SPU_PRINTF_ENABLED)

19
rpcs3/Emu/SysCalls/Modules/cellSpurs.h
View file

@ -102,6 +102,12 @@ enum SPURSKernelInterfaces
CELL_SPURS_INTERRUPT_VECTOR = 0x0, CELL_SPURS_INTERRUPT_VECTOR = 0x0,
CELL_SPURS_LOCK_LINE = 0x80, CELL_SPURS_LOCK_LINE = 0x80,
CELL_SPURS_KERNEL_DMA_TAG_ID = 31, CELL_SPURS_KERNEL_DMA_TAG_ID = 31,
CELL_SPURS_KERNEL1_ENTRY_ADDR = 0x818,
CELL_SPURS_KERNEL2_ENTRY_ADDR = 0x848,
CELL_SPURS_KERNEL1_YIELD_ADDR = 0x808,
CELL_SPURS_KERNEL2_YIELD_ADDR = 0x838,
CELL_SPURS_KERNEL1_SELECT_WORKLOAD_ADDR = 0x290,
CELL_SPURS_KERNEL2_SELECT_WORKLOAD_ADDR = 0x290,
}; };
enum RangeofEventQueuePortNumbers enum RangeofEventQueuePortNumbers
@ -885,14 +891,23 @@ struct SpursKernelMgmtData
u8 spuIdling; // 0x1EB u8 spuIdling; // 0x1EB
be_t<u16> wklRunnable1; // 0x1EC be_t<u16> wklRunnable1; // 0x1EC
be_t<u16> wklRunnable2; // 0x1EE be_t<u16> wklRunnable2; // 0x1EE
u8 x1F0[0x210 - 0x1F0]; // 0x1F0 be_t<u32> x1F0; // 0x1F0
be_t<u32> x1F4; // 0x1F4
be_t<u32> x1F8; // 0x1F8
be_t<u32> x1FC; // 0x1FC
be_t<u32> x200; // 0x200
be_t<u32> x204; // 0x204
be_t<u32> x208; // 0x208
be_t<u32> x20C; // 0x20C
be_t<u64> traceBuffer; // 0x210 be_t<u64> traceBuffer; // 0x210
be_t<u32> traceMsgCount; // 0x218 be_t<u32> traceMsgCount; // 0x218
be_t<u32> traceMaxCount; // 0x21C be_t<u32> traceMaxCount; // 0x21C
u8 wklUniqueId[0x10]; // 0x220 u8 wklUniqueId[0x10]; // 0x220
u8 x230[0x280 - 0x230]; // 0x230
be_t<u32> guid[4]; // 0x280
}; };
static_assert(sizeof(SpursKernelMgmtData) == 0x130, "Incorrect size for SpursKernelMgmtData"); static_assert(sizeof(SpursKernelMgmtData) == 0x190, "Incorrect size for SpursKernelMgmtData");
// The SPURS taskset policy module data store. This resides at 0x2700 of the LS. // The SPURS taskset policy module data store. This resides at 0x2700 of the LS.
struct SpursTasksetPmMgmtData struct SpursTasksetPmMgmtData

273
rpcs3/Emu/SysCalls/Modules/cellSpursSpu.cpp
View file

@ -14,11 +14,15 @@
void cellSpursModulePutTrace(CellSpursTracePacket * packet, unsigned tag); void cellSpursModulePutTrace(CellSpursTracePacket * packet, unsigned tag);
u32 cellSpursModulePollStatus(SPUThread & spu, u32 * status); u32 cellSpursModulePollStatus(SPUThread & spu, u32 * status);
bool spursDma(SPUThread & spu, u32 cmd, u64 ea, u32 lsa, u32 size, u32 tag);
u32 spursDmaGetCompletionStatus(SPUThread & spu, u32 tagMask);
u32 spursDmaWaitForCompletion(SPUThread & spu, u32 tagMask, bool waitForAll = true);
// //
// SPURS Kernel functions // SPURS Kernel functions
// //
void spursKernelSelectWorkload(SPUThread & spu); bool spursKernel1SelectWorkload(SPUThread & spu);
void spursKernelSelectWorkload2(SPUThread & spu); bool spursKernel2SelectWorkload(SPUThread & spu);
// //
// SPURS system service workload functions // SPURS system service workload functions
@ -31,7 +35,7 @@ void spursSysServiceUpdateWorkload(SPUThread & spu, SpursKernelMgmtData * mgmt);
void spursSysServiceProcessMessages(SPUThread & spu, SpursKernelMgmtData * mgmt); void spursSysServiceProcessMessages(SPUThread & spu, SpursKernelMgmtData * mgmt);
void spursSysServiceWaitOrExit(SPUThread & spu, SpursKernelMgmtData * mgmt); void spursSysServiceWaitOrExit(SPUThread & spu, SpursKernelMgmtData * mgmt);
void spursSysServiceWorkloadMain(SPUThread & spu, u32 pollStatus); void spursSysServiceWorkloadMain(SPUThread & spu, u32 pollStatus);
void spursSysServiceWorkloadEntry(SPUThread & spu); bool spursSysServiceWorkloadEntry(SPUThread & spu);
// //
// SPURS taskset polict module functions // SPURS taskset polict module functions
@ -54,9 +58,9 @@ u32 cellSpursModulePollStatus(SPUThread & spu, u32 * status) {
spu.GPR[3]._u32[3] = 1; spu.GPR[3]._u32[3] = 1;
if (mgmt->spurs->m.flags1 & SF1_32_WORKLOADS) { if (mgmt->spurs->m.flags1 & SF1_32_WORKLOADS) {
spursKernelSelectWorkload2(spu); spursKernel2SelectWorkload(spu);
} else { } else {
spursKernelSelectWorkload(spu); spursKernel1SelectWorkload(spu);
} }
auto result = spu.GPR[3]._u64[1]; auto result = spu.GPR[3]._u64[1];
@ -68,14 +72,51 @@ u32 cellSpursModulePollStatus(SPUThread & spu, u32 * status) {
return wklId == mgmt->wklCurrentId ? 0 : 1; return wklId == mgmt->wklCurrentId ? 0 : 1;
} }
/// Execute a DMA operation
bool spursDma(SPUThread & spu, u32 cmd, u64 ea, u32 lsa, u32 size, u32 tag) {
spu.WriteChannel(MFC_LSA, u128::from32r(lsa));
spu.WriteChannel(MFC_EAH, u128::from32r((u32)(ea >> 32)));
spu.WriteChannel(MFC_EAL, u128::from32r((u32)ea));
spu.WriteChannel(MFC_Size, u128::from32r(size));
spu.WriteChannel(MFC_TagID, u128::from32r(tag));
spu.WriteChannel(MFC_Cmd, u128::from32r(cmd));
if (cmd == MFC_GETLLAR_CMD || cmd == MFC_PUTLLC_CMD || cmd == MFC_PUTLLUC_CMD) {
u128 rv;
spu.ReadChannel(rv, MFC_RdAtomicStat);
return rv._u32[3] ? true : false;
}
return true;
}
/// Get the status of DMA operations
u32 spursDmaGetCompletionStatus(SPUThread & spu, u32 tagMask) {
u128 rv;
spu.WriteChannel(MFC_WrTagMask, u128::from32r(tagMask));
spu.WriteChannel(MFC_WrTagUpdate, u128::from32r(MFC_TAG_UPDATE_IMMEDIATE));
spu.ReadChannel(rv, MFC_RdTagStat);
return rv._u32[3];
}
/// Wait for DMA operations to complete
u32 spursDmaWaitForCompletion(SPUThread & spu, u32 tagMask, bool waitForAll) {
u128 rv;
spu.WriteChannel(MFC_WrTagMask, u128::from32r(tagMask));
spu.WriteChannel(MFC_WrTagUpdate, u128::from32r(waitForAll ? MFC_TAG_UPDATE_ALL : MFC_TAG_UPDATE_ANY));
spu.ReadChannel(rv, MFC_RdTagStat);
return rv._u32[3];
}
////////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////////
// SPURS kernel functions // SPURS kernel functions
////////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////////
/// Select a workload to run /// Select a workload to run
void spursKernelSelectWorkload(SPUThread & spu) { bool spursKernel1SelectWorkload(SPUThread & spu) {
LV2_LOCK(0); // TODO: lock-free implementation if possible
auto mgmt = vm::get_ptr<SpursKernelMgmtData>(spu.ls_offset + 0x100); auto mgmt = vm::get_ptr<SpursKernelMgmtData>(spu.ls_offset + 0x100);
// The first and only argument to this function is a boolean that is set to false if the function // The first and only argument to this function is a boolean that is set to false if the function
@ -83,42 +124,50 @@ void spursKernelSelectWorkload(SPUThread & spu) {
// If the first argument is true then the shared data is not updated with the result. // If the first argument is true then the shared data is not updated with the result.
const auto isPoll = spu.GPR[3]._u32[3]; const auto isPoll = spu.GPR[3]._u32[3];
u32 wklSelectedId;
u32 pollStatus;
do {
// DMA and lock the first 0x80 bytes of spurs
spursDma(spu, MFC_GETLLAR_CMD, mgmt->spurs.addr(), 0x100/*LSA*/, 0x80/*size*/, 0/*tag*/);
CellSpurs * spurs = (CellSpurs *)mgmt->tempArea;
// Calculate the contention (number of SPUs used) for each workload // Calculate the contention (number of SPUs used) for each workload
u8 contention[CELL_SPURS_MAX_WORKLOAD]; u8 contention[CELL_SPURS_MAX_WORKLOAD];
u8 pendingContention[CELL_SPURS_MAX_WORKLOAD]; u8 pendingContention[CELL_SPURS_MAX_WORKLOAD];
for (auto i = 0; i < CELL_SPURS_MAX_WORKLOAD; i++) { for (auto i = 0; i < CELL_SPURS_MAX_WORKLOAD; i++) {
contention[i] = mgmt->spurs->m.wklCurrentContention[i] - mgmt->wklLocContention[i]; contention[i] = spurs->m.wklCurrentContention[i] - mgmt->wklLocContention[i];
// If this is a poll request then the number of SPUs pending to context switch is also added to the contention presumably // If this is a poll request then the number of SPUs pending to context switch is also added to the contention presumably
// to prevent unnecessary jumps to the kernel // to prevent unnecessary jumps to the kernel
if (isPoll) { if (isPoll) {
pendingContention[i] = mgmt->spurs->m.wklPendingContention[i] - mgmt->wklLocPendingContention[i]; pendingContention[i] = spurs->m.wklPendingContention[i] - mgmt->wklLocPendingContention[i];
if (i != mgmt->wklCurrentId) { if (i != mgmt->wklCurrentId) {
contention[i] += pendingContention[i]; contention[i] += pendingContention[i];
} }
} }
} }
u32 wklSelectedId = CELL_SPURS_SYS_SERVICE_WORKLOAD_ID; wklSelectedId = CELL_SPURS_SYS_SERVICE_WORKLOAD_ID;
u32 pollStatus = 0; pollStatus = 0;
// The system service workload has the highest priority. Select the system service workload if // The system service workload has the highest priority. Select the system service workload if
// the system service message bit for this SPU is set. // the system service message bit for this SPU is set.
if (mgmt->spurs->m.sysSrvMessage.read_relaxed() & (1 << mgmt->spuNum)) { if (spurs->m.sysSrvMessage.read_relaxed() & (1 << mgmt->spuNum)) {
mgmt->spuIdling = 0; mgmt->spuIdling = 0;
if (!isPoll || mgmt->wklCurrentId == CELL_SPURS_SYS_SERVICE_WORKLOAD_ID) { if (!isPoll || mgmt->wklCurrentId == CELL_SPURS_SYS_SERVICE_WORKLOAD_ID) {
// Clear the message bit // Clear the message bit
mgmt->spurs->m.sysSrvMessage.write_relaxed(mgmt->spurs->m.sysSrvMessage.read_relaxed() & ~(1 << mgmt->spuNum)); spurs->m.sysSrvMessage.write_relaxed(spurs->m.sysSrvMessage.read_relaxed() & ~(1 << mgmt->spuNum));
} }
} else { } else {
// Caclulate the scheduling weight for each workload // Caclulate the scheduling weight for each workload
u16 maxWeight = 0; u16 maxWeight = 0;
for (auto i = 0; i < CELL_SPURS_MAX_WORKLOAD; i++) { for (auto i = 0; i < CELL_SPURS_MAX_WORKLOAD; i++) {
u16 runnable = mgmt->wklRunnable1 & (0x8000 >> i); u16 runnable = mgmt->wklRunnable1 & (0x8000 >> i);
u16 wklSignal = mgmt->spurs->m.wklSignal1.read_relaxed() & (0x8000 >> i); u16 wklSignal = spurs->m.wklSignal1.read_relaxed() & (0x8000 >> i);
u8 wklFlag = mgmt->spurs->m.wklFlag.flag.read_relaxed() == 0 ? mgmt->spurs->m.wklFlagReceiver.read_relaxed() == i ? 1 : 0 : 0; u8 wklFlag = spurs->m.wklFlag.flag.read_relaxed() == 0 ? spurs->m.wklFlagReceiver.read_relaxed() == i ? 1 : 0 : 0;
u8 readyCount = mgmt->spurs->m.wklReadyCount1[i].read_relaxed() > CELL_SPURS_MAX_SPU ? CELL_SPURS_MAX_SPU : mgmt->spurs->m.wklReadyCount1[i].read_relaxed(); u8 readyCount = spurs->m.wklReadyCount1[i].read_relaxed() > CELL_SPURS_MAX_SPU ? CELL_SPURS_MAX_SPU : spurs->m.wklReadyCount1[i].read_relaxed();
u8 idleSpuCount = mgmt->spurs->m.wklIdleSpuCountOrReadyCount2[i].read_relaxed() > CELL_SPURS_MAX_SPU ? CELL_SPURS_MAX_SPU : mgmt->spurs->m.wklIdleSpuCountOrReadyCount2[i].read_relaxed(); u8 idleSpuCount = spurs->m.wklIdleSpuCountOrReadyCount2[i].read_relaxed() > CELL_SPURS_MAX_SPU ? CELL_SPURS_MAX_SPU : spurs->m.wklIdleSpuCountOrReadyCount2[i].read_relaxed();
u8 requestCount = readyCount + idleSpuCount; u8 requestCount = readyCount + idleSpuCount;
// For a workload to be considered for scheduling: // For a workload to be considered for scheduling:
@ -128,7 +177,7 @@ void spursKernelSelectWorkload(SPUThread & spu) {
// 4. The number of SPUs allocated to it must be less than the number of SPUs requested (i.e. readyCount) // 4. The number of SPUs allocated to it must be less than the number of SPUs requested (i.e. readyCount)
// OR the workload must be signalled // OR the workload must be signalled
// OR the workload flag is 0 and the workload is configured as the wokload flag receiver // OR the workload flag is 0 and the workload is configured as the wokload flag receiver
if (runnable && mgmt->priority[i] != 0 && mgmt->spurs->m.wklMaxContention[i].read_relaxed() > contention[i]) { if (runnable && mgmt->priority[i] != 0 && spurs->m.wklMaxContention[i].read_relaxed() > contention[i]) {
if (wklFlag || wklSignal || (readyCount != 0 && requestCount > contention[i])) { if (wklFlag || wklSignal || (readyCount != 0 && requestCount > contention[i])) {
// The scheduling weight of the workload is formed from the following parameters in decreasing order of priority: // The scheduling weight of the workload is formed from the following parameters in decreasing order of priority:
// 1. Wokload signal set or workload flag or ready count > contention // 1. Wokload signal set or workload flag or ready count > contention
@ -141,7 +190,7 @@ void spursKernelSelectWorkload(SPUThread & spu) {
u16 weight = (wklFlag || wklSignal || (readyCount > contention[i])) ? 0x8000 : 0; u16 weight = (wklFlag || wklSignal || (readyCount > contention[i])) ? 0x8000 : 0;
weight |= (u16)(mgmt->priority[i] & 0x7F) << 16; weight |= (u16)(mgmt->priority[i] & 0x7F) << 16;
weight |= i == mgmt->wklCurrentId ? 0x80 : 0x00; weight |= i == mgmt->wklCurrentId ? 0x80 : 0x00;
weight |= (contention[i] > 0 && mgmt->spurs->m.wklMinContention[i] > contention[i]) ? 0x40 : 0x00; weight |= (contention[i] > 0 && spurs->m.wklMinContention[i] > contention[i]) ? 0x40 : 0x00;
weight |= ((CELL_SPURS_MAX_SPU - contention[i]) & 0x0F) << 2; weight |= ((CELL_SPURS_MAX_SPU - contention[i]) & 0x0F) << 2;
weight |= mgmt->wklUniqueId[i] == mgmt->wklCurrentId ? 0x02 : 0x00; weight |= mgmt->wklUniqueId[i] == mgmt->wklCurrentId ? 0x02 : 0x00;
weight |= 0x01; weight |= 0x01;
@ -163,12 +212,12 @@ void spursKernelSelectWorkload(SPUThread & spu) {
if (!isPoll || wklSelectedId == mgmt->wklCurrentId) { if (!isPoll || wklSelectedId == mgmt->wklCurrentId) {
// Clear workload signal for the selected workload // Clear workload signal for the selected workload
mgmt->spurs->m.wklSignal1.write_relaxed(be_t<u16>::make(mgmt->spurs->m.wklSignal1.read_relaxed() & ~(0x8000 >> wklSelectedId))); spurs->m.wklSignal1.write_relaxed(be_t<u16>::make(spurs->m.wklSignal1.read_relaxed() & ~(0x8000 >> wklSelectedId)));
mgmt->spurs->m.wklSignal2.write_relaxed(be_t<u16>::make(mgmt->spurs->m.wklSignal1.read_relaxed() & ~(0x80000000u >> wklSelectedId))); spurs->m.wklSignal2.write_relaxed(be_t<u16>::make(spurs->m.wklSignal1.read_relaxed() & ~(0x80000000u >> wklSelectedId)));
// If the selected workload is the wklFlag workload then pull the wklFlag to all 1s // If the selected workload is the wklFlag workload then pull the wklFlag to all 1s
if (wklSelectedId == mgmt->spurs->m.wklFlagReceiver.read_relaxed()) { if (wklSelectedId == spurs->m.wklFlagReceiver.read_relaxed()) {
mgmt->spurs->m.wklFlag.flag.write_relaxed(be_t<u32>::make(0xFFFFFFFF)); spurs->m.wklFlag.flag.write_relaxed(be_t<u32>::make(0xFFFFFFFF));
} }
} }
} }
@ -181,7 +230,7 @@ void spursKernelSelectWorkload(SPUThread & spu) {
} }
for (auto i = 0; i < CELL_SPURS_MAX_WORKLOAD; i++) { for (auto i = 0; i < CELL_SPURS_MAX_WORKLOAD; i++) {
mgmt->spurs->m.wklCurrentContention[i] = contention[i]; spurs->m.wklCurrentContention[i] = contention[i];
mgmt->wklLocContention[i] = 0; mgmt->wklLocContention[i] = 0;
mgmt->wklLocPendingContention[i] = 0; mgmt->wklLocPendingContention[i] = 0;
} }
@ -199,7 +248,7 @@ void spursKernelSelectWorkload(SPUThread & spu) {
} }
for (auto i = 0; i < CELL_SPURS_MAX_WORKLOAD; i++) { for (auto i = 0; i < CELL_SPURS_MAX_WORKLOAD; i++) {
mgmt->spurs->m.wklPendingContention[i] = pendingContention[i]; spurs->m.wklPendingContention[i] = pendingContention[i];
mgmt->wklLocPendingContention[i] = 0; mgmt->wklLocPendingContention[i] = 0;
} }
@ -207,16 +256,16 @@ void spursKernelSelectWorkload(SPUThread & spu) {
mgmt->wklLocPendingContention[wklSelectedId] = 1; mgmt->wklLocPendingContention[wklSelectedId] = 1;
} }
} }
} while (spursDma(spu, MFC_PUTLLC_CMD, mgmt->spurs.addr(), 0x100/*LSA*/, 0x80/*size*/, 0/*tag*/) == false);
u64 result = (u64)wklSelectedId << 32; u64 result = (u64)wklSelectedId << 32;
result |= pollStatus; result |= pollStatus;
spu.GPR[3]._u64[1] = result; spu.GPR[3]._u64[1] = result;
return true;
} }
/// Select a workload to run /// Select a workload to run
void spursKernelSelectWorkload2(SPUThread & spu) { bool spursKernel2SelectWorkload(SPUThread & spu) {
LV2_LOCK(0); // TODO: lock-free implementation if possible
auto mgmt = vm::get_ptr<SpursKernelMgmtData>(spu.ls_offset + 0x100); auto mgmt = vm::get_ptr<SpursKernelMgmtData>(spu.ls_offset + 0x100);
// The first and only argument to this function is a boolean that is set to false if the function // The first and only argument to this function is a boolean that is set to false if the function
@ -224,17 +273,25 @@ void spursKernelSelectWorkload2(SPUThread & spu) {
// If the first argument is true then the shared data is not updated with the result. // If the first argument is true then the shared data is not updated with the result.
const auto isPoll = spu.GPR[3]._u32[3]; const auto isPoll = spu.GPR[3]._u32[3];
u32 wklSelectedId;
u32 pollStatus;
do {
// DMA and lock the first 0x80 bytes of spurs
spursDma(spu, MFC_GETLLAR_CMD, mgmt->spurs.addr(), 0x100/*LSA*/, 0x80/*size*/, 0/*tag*/);
CellSpurs * spurs = (CellSpurs *)mgmt->tempArea;
// Calculate the contention (number of SPUs used) for each workload // Calculate the contention (number of SPUs used) for each workload
u8 contention[CELL_SPURS_MAX_WORKLOAD2]; u8 contention[CELL_SPURS_MAX_WORKLOAD2];
u8 pendingContention[CELL_SPURS_MAX_WORKLOAD2]; u8 pendingContention[CELL_SPURS_MAX_WORKLOAD2];
for (auto i = 0; i < CELL_SPURS_MAX_WORKLOAD2; i++) { for (auto i = 0; i < CELL_SPURS_MAX_WORKLOAD2; i++) {
contention[i] = mgmt->spurs->m.wklCurrentContention[i & 0x0F] - mgmt->wklLocContention[i & 0x0F]; contention[i] = spurs->m.wklCurrentContention[i & 0x0F] - mgmt->wklLocContention[i & 0x0F];
contention[i] = i < CELL_SPURS_MAX_WORKLOAD ? contention[i] & 0x0F : contention[i] >> 4; contention[i] = i < CELL_SPURS_MAX_WORKLOAD ? contention[i] & 0x0F : contention[i] >> 4;
// If this is a poll request then the number of SPUs pending to context switch is also added to the contention presumably // If this is a poll request then the number of SPUs pending to context switch is also added to the contention presumably
// to prevent unnecessary jumps to the kernel // to prevent unnecessary jumps to the kernel
if (isPoll) { if (isPoll) {
pendingContention[i] = mgmt->spurs->m.wklPendingContention[i & 0x0F] - mgmt->wklLocPendingContention[i & 0x0F]; pendingContention[i] = spurs->m.wklPendingContention[i & 0x0F] - mgmt->wklLocPendingContention[i & 0x0F];
pendingContention[i] = i < CELL_SPURS_MAX_WORKLOAD ? pendingContention[i] & 0x0F : pendingContention[i] >> 4; pendingContention[i] = i < CELL_SPURS_MAX_WORKLOAD ? pendingContention[i] & 0x0F : pendingContention[i] >> 4;
if (i != mgmt->wklCurrentId) { if (i != mgmt->wklCurrentId) {
contention[i] += pendingContention[i]; contention[i] += pendingContention[i];
@ -242,17 +299,17 @@ void spursKernelSelectWorkload2(SPUThread & spu) {
} }
} }
u32 wklSelectedId = CELL_SPURS_SYS_SERVICE_WORKLOAD_ID; wklSelectedId = CELL_SPURS_SYS_SERVICE_WORKLOAD_ID;
u32 pollStatus = 0; pollStatus = 0;
// The system service workload has the highest priority. Select the system service workload if // The system service workload has the highest priority. Select the system service workload if
// the system service message bit for this SPU is set. // the system service message bit for this SPU is set.
if (mgmt->spurs->m.sysSrvMessage.read_relaxed() & (1 << mgmt->spuNum)) { if (spurs->m.sysSrvMessage.read_relaxed() & (1 << mgmt->spuNum)) {
// Not sure what this does. Possibly Mark the SPU as in use. // Not sure what this does. Possibly Mark the SPU as in use.
mgmt->spuIdling = 0; mgmt->spuIdling = 0;
if (!isPoll || mgmt->wklCurrentId == CELL_SPURS_SYS_SERVICE_WORKLOAD_ID) { if (!isPoll || mgmt->wklCurrentId == CELL_SPURS_SYS_SERVICE_WORKLOAD_ID) {
// Clear the message bit // Clear the message bit
mgmt->spurs->m.sysSrvMessage.write_relaxed(mgmt->spurs->m.sysSrvMessage.read_relaxed() & ~(1 << mgmt->spuNum)); spurs->m.sysSrvMessage.write_relaxed(spurs->m.sysSrvMessage.read_relaxed() & ~(1 << mgmt->spuNum));
} }
} else { } else {
// Caclulate the scheduling weight for each workload // Caclulate the scheduling weight for each workload
@ -261,10 +318,10 @@ void spursKernelSelectWorkload2(SPUThread & spu) {
auto j = i & 0x0F; auto j = i & 0x0F;
u16 runnable = i < CELL_SPURS_MAX_WORKLOAD ? mgmt->wklRunnable1 & (0x8000 >> j) : mgmt->wklRunnable2 & (0x8000 >> j); u16 runnable = i < CELL_SPURS_MAX_WORKLOAD ? mgmt->wklRunnable1 & (0x8000 >> j) : mgmt->wklRunnable2 & (0x8000 >> j);
u8 priority = i < CELL_SPURS_MAX_WORKLOAD ? mgmt->priority[j] & 0x0F : mgmt->priority[j] >> 4; u8 priority = i < CELL_SPURS_MAX_WORKLOAD ? mgmt->priority[j] & 0x0F : mgmt->priority[j] >> 4;
u8 maxContention = i < CELL_SPURS_MAX_WORKLOAD ? mgmt->spurs->m.wklMaxContention[j].read_relaxed() & 0x0F : mgmt->spurs->m.wklMaxContention[j].read_relaxed() >> 4; u8 maxContention = i < CELL_SPURS_MAX_WORKLOAD ? spurs->m.wklMaxContention[j].read_relaxed() & 0x0F : spurs->m.wklMaxContention[j].read_relaxed() >> 4;
u16 wklSignal = i < CELL_SPURS_MAX_WORKLOAD ? mgmt->spurs->m.wklSignal1.read_relaxed() & (0x8000 >> j) : mgmt->spurs->m.wklSignal2.read_relaxed() & (0x8000 >> j); u16 wklSignal = i < CELL_SPURS_MAX_WORKLOAD ? spurs->m.wklSignal1.read_relaxed() & (0x8000 >> j) : spurs->m.wklSignal2.read_relaxed() & (0x8000 >> j);
u8 wklFlag = mgmt->spurs->m.wklFlag.flag.read_relaxed() == 0 ? mgmt->spurs->m.wklFlagReceiver.read_relaxed() == i ? 1 : 0 : 0; u8 wklFlag = spurs->m.wklFlag.flag.read_relaxed() == 0 ? spurs->m.wklFlagReceiver.read_relaxed() == i ? 1 : 0 : 0;
u8 readyCount = i < CELL_SPURS_MAX_WORKLOAD ? mgmt->spurs->m.wklReadyCount1[j].read_relaxed() : mgmt->spurs->m.wklIdleSpuCountOrReadyCount2[j].read_relaxed(); u8 readyCount = i < CELL_SPURS_MAX_WORKLOAD ? spurs->m.wklReadyCount1[j].read_relaxed() : spurs->m.wklIdleSpuCountOrReadyCount2[j].read_relaxed();
// For a workload to be considered for scheduling: // For a workload to be considered for scheduling:
// 1. Its priority must be greater than 0 // 1. Its priority must be greater than 0
@ -299,12 +356,12 @@ void spursKernelSelectWorkload2(SPUThread & spu) {
if (!isPoll || wklSelectedId == mgmt->wklCurrentId) { if (!isPoll || wklSelectedId == mgmt->wklCurrentId) {
// Clear workload signal for the selected workload // Clear workload signal for the selected workload
mgmt->spurs->m.wklSignal1.write_relaxed(be_t<u16>::make(mgmt->spurs->m.wklSignal1.read_relaxed() & ~(0x8000 >> wklSelectedId))); spurs->m.wklSignal1.write_relaxed(be_t<u16>::make(spurs->m.wklSignal1.read_relaxed() & ~(0x8000 >> wklSelectedId)));
mgmt->spurs->m.wklSignal2.write_relaxed(be_t<u16>::make(mgmt->spurs->m.wklSignal1.read_relaxed() & ~(0x80000000u >> wklSelectedId))); spurs->m.wklSignal2.write_relaxed(be_t<u16>::make(spurs->m.wklSignal1.read_relaxed() & ~(0x80000000u >> wklSelectedId)));
// If the selected workload is the wklFlag workload then pull the wklFlag to all 1s // If the selected workload is the wklFlag workload then pull the wklFlag to all 1s
if (wklSelectedId == mgmt->spurs->m.wklFlagReceiver.read_relaxed()) { if (wklSelectedId == spurs->m.wklFlagReceiver.read_relaxed()) {
mgmt->spurs->m.wklFlag.flag.write_relaxed(be_t<u32>::make(0xFFFFFFFF)); spurs->m.wklFlag.flag.write_relaxed(be_t<u32>::make(0xFFFFFFFF));
} }
} }
} }
@ -317,7 +374,7 @@ void spursKernelSelectWorkload2(SPUThread & spu) {
} }
for (auto i = 0; i < (CELL_SPURS_MAX_WORKLOAD2 >> 1); i++) { for (auto i = 0; i < (CELL_SPURS_MAX_WORKLOAD2 >> 1); i++) {
mgmt->spurs->m.wklCurrentContention[i] = contention[i] | (contention[i + 0x10] << 4); spurs->m.wklCurrentContention[i] = contention[i] | (contention[i + 0x10] << 4);
mgmt->wklLocContention[i] = 0; mgmt->wklLocContention[i] = 0;
mgmt->wklLocPendingContention[i] = 0; mgmt->wklLocPendingContention[i] = 0;
} }
@ -332,94 +389,98 @@ void spursKernelSelectWorkload2(SPUThread & spu) {
} }
for (auto i = 0; i < (CELL_SPURS_MAX_WORKLOAD2 >> 1); i++) { for (auto i = 0; i < (CELL_SPURS_MAX_WORKLOAD2 >> 1); i++) {
mgmt->spurs->m.wklPendingContention[i] = pendingContention[i] | (pendingContention[i + 0x10] << 4); spurs->m.wklPendingContention[i] = pendingContention[i] | (pendingContention[i + 0x10] << 4);
mgmt->wklLocPendingContention[i] = 0; mgmt->wklLocPendingContention[i] = 0;
} }
mgmt->wklLocPendingContention[wklSelectedId & 0x0F] = wklSelectedId < CELL_SPURS_MAX_WORKLOAD ? 0x01 : wklSelectedId < CELL_SPURS_MAX_WORKLOAD2 ? 0x10 : 0; mgmt->wklLocPendingContention[wklSelectedId & 0x0F] = wklSelectedId < CELL_SPURS_MAX_WORKLOAD ? 0x01 : wklSelectedId < CELL_SPURS_MAX_WORKLOAD2 ? 0x10 : 0;
} }
} while (spursDma(spu, MFC_PUTLLC_CMD, mgmt->spurs.addr(), 0x100/*LSA*/, 0x80/*size*/, 0/*tag*/) == false);
u64 result = (u64)wklSelectedId << 32; u64 result = (u64)wklSelectedId << 32;
result |= pollStatus; result |= pollStatus;
spu.GPR[3]._u64[1] = result; spu.GPR[3]._u64[1] = result;
return true;
} }
/// Entry point of the SPURS kernel /// SPURS kernel main
void spursKernelMain(SPUThread & spu) { bool spursKernelMain(SPUThread & spu) {
SpursKernelMgmtData * mgmt = vm::get_ptr<SpursKernelMgmtData>(spu.ls_offset + 0x100); SpursKernelMgmtData * mgmt = vm::get_ptr<SpursKernelMgmtData>(spu.ls_offset + 0x100);
bool isKernel2;
u32 pollStatus;
const CellSpurs::WorkloadInfo * wklInfo;
if (spu.PC == CELL_SPURS_KERNEL1_ENTRY_ADDR || spu.PC == CELL_SPURS_KERNEL2_ENTRY_ADDR) {
// Entry point of SPURS kernel
// Save arguments
mgmt->spuNum = spu.GPR[3]._u32[3]; mgmt->spuNum = spu.GPR[3]._u32[3];
mgmt->dmaTagId = 0x1F;
mgmt->spurs.set(spu.GPR[4]._u64[1]); mgmt->spurs.set(spu.GPR[4]._u64[1]);
mgmt->wklCurrentId = CELL_SPURS_SYS_SERVICE_WORKLOAD_ID;
isKernel2 = mgmt->spurs->m.flags1 & SF1_32_WORKLOADS ? true : false;
memset(mgmt, 0, sizeof(SpursKernelMgmtData));
// Initialise the SPURS management area to its initial values
mgmt->dmaTagId = CELL_SPURS_KERNEL_DMA_TAG_ID;
mgmt->wklCurrentUniqueId = 0x20; mgmt->wklCurrentUniqueId = 0x20;
mgmt->wklCurrentId = CELL_SPURS_SYS_SERVICE_WORKLOAD_ID;
bool isSecond = mgmt->spurs->m.flags1 & SF1_32_WORKLOADS ? true : false; mgmt->yieldToKernelAddr = isKernel2 ? CELL_SPURS_KERNEL2_YIELD_ADDR : CELL_SPURS_KERNEL1_YIELD_ADDR;
mgmt->yieldToKernelAddr = isSecond ? 0x838 : 0x808; mgmt->selectWorkloadAddr = isKernel2 ? CELL_SPURS_KERNEL2_SELECT_WORKLOAD_ADDR : CELL_SPURS_KERNEL1_SELECT_WORKLOAD_ADDR;
mgmt->selectWorkloadAddr = 0x290; if (!isKernel2) {
spu.WriteLS32(mgmt->yieldToKernelAddr, 2); // hack for cellSpursModuleExit mgmt->x1F0 = 0xF0020000;
spu.WriteLS32(mgmt->selectWorkloadAddr, 3); // hack for cellSpursModulePollStatus mgmt->x200 = 0x20000;
spu.WriteLS32(mgmt->selectWorkloadAddr + 4, 0x35000000); // bi $0 mgmt->guid[0] = 0x423A3A02;
spu.m_code3_func = isSecond ? spursKernelSelectWorkload2 : spursKernelSelectWorkload; mgmt->guid[1] = 0x43F43A82;
mgmt->guid[2] = 0x43F26502;
u32 wid = CELL_SPURS_SYS_SERVICE_WORKLOAD_ID; mgmt->guid[3] = 0x420EB382;
u32 pollStatus = 0; } else {
while (true) { mgmt->guid[0] = 0x43A08402;
if (Emu.IsStopped()) { mgmt->guid[1] = 0x43FB0A82;
cellSpurs->Warning("Spurs Kernel aborted"); mgmt->guid[2] = 0x435E9302;
return; mgmt->guid[3] = 0x43A3C982;
} }
// Get current workload info spu.UnregisterHleFunctions(0, 0x40000); // TODO: use a symbolic constant
auto & wkl = wid < CELL_SPURS_MAX_WORKLOAD ? mgmt->spurs->m.wklInfo1[wid] : (wid < CELL_SPURS_MAX_WORKLOAD2 && isSecond ? mgmt->spurs->m.wklInfo2[wid & 0xf] : mgmt->spurs->m.wklInfoSysSrv); spu.RegisterHleFuncion(isKernel2 ? CELL_SPURS_KERNEL2_ENTRY_ADDR : CELL_SPURS_KERNEL1_ENTRY_ADDR, spursKernelMain);
spu.RegisterHleFuncion(mgmt->yieldToKernelAddr, spursKernelMain);
spu.RegisterHleFuncion(mgmt->selectWorkloadAddr, isKernel2 ? spursKernel2SelectWorkload : spursKernel1SelectWorkload);
if (mgmt->wklCurrentAddr != wkl.addr) { // Start the system service workload
if (wkl.addr.addr() != SPURS_IMG_ADDR_SYS_SRV_WORKLOAD) { spu.RegisterHleFuncion(0xA00, spursSysServiceWorkloadEntry);
// Load executable code wklInfo = &mgmt->spurs->m.wklInfoSysSrv;
memcpy(vm::get_ptr<void>(spu.ls_offset + 0xA00), wkl.addr.get_ptr(), wkl.size); pollStatus = 0;
} } else if (spu.PC == mgmt->yieldToKernelAddr) {
mgmt->wklCurrentAddr = wkl.addr; isKernel2 = mgmt->spurs->m.flags1 & SF1_32_WORKLOADS ? true : false;
mgmt->wklCurrentUniqueId = wkl.uniqueId.read_relaxed();
// Select next workload to run
spu.GPR[3].clear();
if (isKernel2) {
spursKernel2SelectWorkload(spu);
} else {
spursKernel1SelectWorkload(spu);
} }
if (!isSecond) { pollStatus = (u32)(spu.GPR[3]._u64[1]);
auto wid = (u32)(spu.GPR[3]._u64[1] >> 32);
wklInfo = wid < CELL_SPURS_MAX_WORKLOAD ? &mgmt->spurs->m.wklInfo1[wid] :
(wid < CELL_SPURS_MAX_WORKLOAD2 && isKernel2 ? &mgmt->spurs->m.wklInfo2[wid & 0xf] : &mgmt->spurs->m.wklInfoSysSrv);
} else {
assert(0);
}
if (!isKernel2) {
mgmt->moduleId[0] = 0; mgmt->moduleId[0] = 0;
mgmt->moduleId[1] = 0; mgmt->moduleId[1] = 0;
} }
// Run workload // Run workload
spu.GPR[0]._u32[3] = mgmt->yieldToKernelAddr;
spu.GPR[1]._u32[3] = 0x3FFB0; spu.GPR[1]._u32[3] = 0x3FFB0;
spu.GPR[3]._u32[3] = 0x100; spu.GPR[3]._u32[3] = 0x100;
spu.GPR[4]._u64[1] = wkl.arg; spu.GPR[4]._u64[1] = wklInfo->arg;
spu.GPR[5]._u32[3] = pollStatus; spu.GPR[5]._u32[3] = pollStatus;
spu.SetPc(0xA00); spu.SetBranch(0xA00);
switch (mgmt->wklCurrentAddr.addr()) { return false;
case SPURS_IMG_ADDR_SYS_SRV_WORKLOAD:
spursSysServiceWorkloadEntry(spu);
break;
default:
spu.FastCall(0xA00);
break;
}
// Check status
auto status = spu.SPU.Status.GetValue();
if (status == SPU_STATUS_STOPPED_BY_STOP) {
return;
} else {
assert(status == SPU_STATUS_RUNNING);
}
// Select next workload to run
spu.GPR[3].clear();
if (isSecond) {
spursKernelSelectWorkload2(spu);
} else {
spursKernelSelectWorkload(spu);
}
u64 res = spu.GPR[3]._u64[1];
pollStatus = (u32)(res);
wid = (u32)(res >> 32);
}
} }
////////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////////
@ -783,7 +844,7 @@ poll:
} }
/// Entry point of the system service workload /// Entry point of the system service workload
void spursSysServiceWorkloadEntry(SPUThread & spu) { bool spursSysServiceWorkloadEntry(SPUThread & spu) {
auto mgmt = vm::get_ptr<SpursKernelMgmtData>(spu.ls_offset + spu.GPR[3]._u32[3]); auto mgmt = vm::get_ptr<SpursKernelMgmtData>(spu.ls_offset + spu.GPR[3]._u32[3]);
auto arg = spu.GPR[4]._u64[1]; auto arg = spu.GPR[4]._u64[1];
auto pollStatus = spu.GPR[5]._u32[3]; auto pollStatus = spu.GPR[5]._u32[3];
@ -800,7 +861,7 @@ void spursSysServiceWorkloadEntry(SPUThread & spu) {
} }
// TODO: Ensure that this function always returns to the SPURS kernel // TODO: Ensure that this function always returns to the SPURS kernel
return; return false;
} }
////////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////////

2
rpcs3/stdafx.h
View file

@ -34,6 +34,8 @@
#include <algorithm> #include <algorithm>
#include <random> #include <random>
#include <unordered_set> #include <unordered_set>
#include <map>
#include <unordered_map>
#include <sys/stat.h> #include <sys/stat.h>
#include "Utilities/GNU.h" #include "Utilities/GNU.h"