Implement scheduling of IOP threads. (#1689)

* Use sleepthread in RPC loop

* Keep a pointer to current IOP thread

* Implement IOP thread scheduling based on priority

And implement DelayThread as an actual delay.

* Run IOP flat out

* Use information from scheduler in wait_run_iop

* Lock sif mutex in set_rpc_queue

* always use kernel dispatch with wait_run

* Loop in dispatch until no thread is ready

* Use timestamp for next wakeup

instead of duration

* Wrap IOP thread entrypoints for safety

Libco threads are not supposed to return from their entrypoint

* Use a queue for IOP thread wakeups from EE thread
This commit is contained in:
Ziemas
2022-07-27 03:15:37 +02:00
committed by GitHub
parent 7d5901aa95
commit f8bc883d48
8 changed files with 215 additions and 85 deletions
+4 -5
View File
@@ -244,7 +244,7 @@ void iop_runner(SystemThreadInterface& iface) {
bool complete = false;
start_overlord_wrapper(iop.overlord_argc, iop.overlord_argv, &complete); // todo!
while (complete == false) {
iop.kernel.dispatchAll();
iop.wait_run_iop(iop.kernel.dispatch());
}
// unblock the EE, the overlord is set up!
@@ -252,10 +252,9 @@ void iop_runner(SystemThreadInterface& iface) {
// IOP Kernel loop
while (!iface.get_want_exit() && !iop.want_exit) {
// the IOP kernel just runs at full blast, so we only run the IOP when the EE is waiting on the
// IOP. Each time the EE is waiting on the IOP, it will run an iteration of the IOP kernel.
iop.wait_run_iop();
iop.kernel.dispatchAll();
// The IOP scheduler informs us of how many microseconds are left until it has something to do.
// So we can wait for that long or until something else needs it to wake up.
iop.wait_run_iop(iop.kernel.dispatch());
}
}
} // namespace
+9 -3
View File
@@ -86,7 +86,14 @@ void* AllocSysMemory(int type, unsigned long size, void* addr) {
* Create a new thread
*/
s32 CreateThread(ThreadParam* param) {
return iop->kernel.CreateThread(param->name, (void (*)())param->entry);
return iop->kernel.CreateThread(param->name, (void (*)())param->entry, param->initPriority);
}
/*!
* Exit current thread
*/
s32 ExitThread() {
return iop->kernel.ExitThread();
}
/*!
@@ -158,8 +165,7 @@ int sceCdMmode(int media) {
}
void DelayThread(u32 usec) {
iop->kernel.SuspendThread();
(void)usec;
iop->kernel.DelayThread(usec);
}
int sceCdBreak() {
+1
View File
@@ -96,6 +96,7 @@ void CpuEnableIntr();
void SleepThread();
void DelayThread(u32 usec);
s32 CreateThread(ThreadParam* param);
s32 ExitThread();
s32 StartThread(s32 thid, u32 arg);
s32 WakeupThread(s32 thid);
+2 -2
View File
@@ -84,12 +84,12 @@ s32 sceSifCallRpc(sceSifClientData* bd,
ASSERT(!end_para);
ASSERT(mode == 1); // async
iop->kernel.sif_rpc(bd->rpcd.id, fno, mode, send, ssize, recv, rsize);
iop->signal_run_iop(false);
iop->signal_run_iop();
return 0;
}
s32 sceSifCheckStatRpc(sceSifRpcData* bd) {
iop->signal_run_iop(false);
iop->signal_run_iop();
return iop->kernel.sif_busy(bd->id);
}
+157 -40
View File
@@ -7,15 +7,46 @@
#include "game/sce/iop.h"
using namespace std::chrono;
/*
** wrap thread entry points to ensure they don't return into libco
*/
static void (*thread_entry)() = nullptr;
static cothread_t wrap_return;
void IopThread::functionWrapper() {
void (*f)() = thread_entry;
co_switch(wrap_return);
if (f != nullptr) {
f();
}
// libco threads must not return
while (true) {
iop::ExitThread();
}
}
/*
** -----------------------------------------------------------------------------
** Functions callable by threads
** -----------------------------------------------------------------------------
*/
/*!
* Create a new thread. Will not run the thread.
*/
s32 IOP_Kernel::CreateThread(std::string name, void (*func)()) {
s32 IOP_Kernel::CreateThread(std::string name, void (*func)(), u32 priority) {
u32 ID = (u32)_nextThID++;
ASSERT(ID == threads.size());
// add entry
threads.emplace_back(name, func, ID);
threads.emplace_back(name, func, ID, priority);
// enter the function wrapper so it can put the actual thread enry on its stack
// to call it when the thread is eventually started
thread_entry = func;
wrap_return = co_active();
co_switch(threads.at(ID).thread);
return ID;
}
@@ -27,37 +58,23 @@ void IOP_Kernel::StartThread(s32 id) {
threads.at(id).state = IopThread::State::Ready;
}
/*!
* Run a thread (call from kernel)
*/
void IOP_Kernel::runThread(s32 id) {
ASSERT(_currentThread == -1); // should run in the kernel thread
_currentThread = id;
threads.at(id).state = IopThread::State::Run;
co_switch(threads.at(id).thread);
_currentThread = -1;
s32 IOP_Kernel::ExitThread() {
ASSERT(_currentThread);
_currentThread->state = IopThread::State::Dormant;
return 0;
}
/*!
* Return to kernel from a thread, not to be called from the kernel thread.
* Put a thread in Wait state for desired amount of usecs.
*/
void IOP_Kernel::exitThread() {
s32 oldThread = getCurrentThread();
co_switch(kernel_thread);
void IOP_Kernel::DelayThread(u32 usec) {
ASSERT(_currentThread);
// check kernel resumed us correctly
ASSERT(_currentThread == oldThread);
}
/*!
* Suspend a thread (call from user thread). Will simply allow other threads to run.
* Like yield
* This does not match the behaviour of any real IOP function.
*/
void IOP_Kernel::SuspendThread() {
ASSERT(getCurrentThread() >= 0);
threads.at(getCurrentThread()).state = IopThread::State::Ready;
_currentThread->state = IopThread::State::Wait;
_currentThread->waitType = IopThread::Wait::Delay;
_currentThread->resumeTime =
time_point_cast<microseconds>(steady_clock::now()) + microseconds(usec);
exitThread();
}
@@ -65,9 +82,9 @@ void IOP_Kernel::SuspendThread() {
* Sleep a thread. Must be explicitly woken up.
*/
void IOP_Kernel::SleepThread() {
ASSERT(getCurrentThread() >= 0);
ASSERT(_currentThread);
threads.at(getCurrentThread()).state = IopThread::State::Suspend;
_currentThread->state = IopThread::State::Suspend;
exitThread();
}
@@ -80,21 +97,114 @@ void IOP_Kernel::WakeupThread(s32 id) {
}
/*!
* Dispatch all IOP threads.
* Currently does no scheduling, on the real IOP the highest priority therad that is Ready
* will always be scheduled.
* Return to kernel from a thread, not to be called from the kernel thread.
*/
void IOP_Kernel::dispatchAll() {
for (s64 i = 0; i < threads.size(); i++) {
if (threads[i].state == IopThread::State::Ready) {
// printf("[IOP Kernel] Dispatch %s (%ld)\n", threads[i].name.c_str(), i);
runThread(i);
// printf("[IOP Kernel] back to kernel!\n");
void IOP_Kernel::exitThread() {
IopThread* oldThread = _currentThread;
co_switch(kernel_thread);
// check kernel resumed us correctly
ASSERT(_currentThread == oldThread);
}
/*
** -----------------------------------------------------------------------------
** Kernel functions.
** -----------------------------------------------------------------------------
*/
/*!
* Run a thread (call from kernel)
*/
void IOP_Kernel::runThread(IopThread* thread) {
ASSERT(_currentThread == nullptr); // should run in the kernel thread
_currentThread = thread;
thread->state = IopThread::State::Run;
co_switch(thread->thread);
_currentThread = nullptr;
}
/*!
** Update wait states for delayed threads
*/
void IOP_Kernel::updateDelay() {
for (auto& t : threads) {
if (t.waitType == IopThread::Wait::Delay) {
if (steady_clock::now() > t.resumeTime) {
t.waitType = IopThread::Wait::None;
t.state = IopThread::State::Ready;
}
}
}
}
time_stamp IOP_Kernel::nextWakeup() {
time_stamp lowest = time_point_cast<microseconds>(steady_clock::now()) + microseconds(1000);
for (auto& t : threads) {
if (t.waitType == IopThread::Wait::Delay) {
if (t.resumeTime < lowest) {
lowest = t.resumeTime;
}
}
}
return lowest;
}
/*!
** Get next thread to run.
** i.e. Highest prio in ready state.
*/
IopThread* IOP_Kernel::schedNext() {
IopThread* highest_prio = nullptr;
for (auto& t : threads) {
if (t.state == IopThread::State::Ready) {
if (highest_prio == nullptr) {
highest_prio = &t;
}
// Lower number = higher priority
if (t.priority < highest_prio->priority) {
highest_prio = &t;
}
}
}
return highest_prio;
};
void IOP_Kernel::processWakeups() {
std::scoped_lock lock(wakeup_mtx);
while (!wakeup_queue.empty()) {
WakeupThread(wakeup_queue.front());
wakeup_queue.pop();
}
}
/*!
* Run the next IOP thread.
*/
time_stamp IOP_Kernel::dispatch() {
updateDelay();
processWakeups();
IopThread* next = schedNext();
while (next != nullptr) {
// printf("[IOP Kernel] Dispatch %s (%d)\n", next->name.c_str(), next->thID);
runThread(next);
updateDelay();
next = schedNext();
// printf("[IOP Kernel] back to kernel!\n");
}
// printf("[IOP Kernel] No runnable threads\n");
return nextWakeup();
}
void IOP_Kernel::set_rpc_queue(iop::sceSifQueueData* qd, u32 thread) {
sif_mtx.lock();
for (const auto& r : sif_records) {
ASSERT(!(r.qd == qd || r.thread_to_wake == thread));
}
@@ -102,6 +212,7 @@ void IOP_Kernel::set_rpc_queue(iop::sceSifQueueData* qd, u32 thread) {
rec.thread_to_wake = thread;
rec.qd = qd;
sif_records.push_back(rec);
sif_mtx.unlock();
}
typedef void* (*sif_rpc_handler)(unsigned int, void*, int);
@@ -158,6 +269,11 @@ void IOP_Kernel::sif_rpc(s32 rpcChannel,
rec->cmd.started = false;
rec->cmd.finished = false;
{
std::scoped_lock lock(wakeup_mtx);
wakeup_queue.push(rec->thread_to_wake);
}
sif_mtx.unlock();
}
@@ -199,7 +315,8 @@ void IOP_Kernel::rpc_loop(iop::sceSifQueueData* qd) {
sif_mtx.unlock();
}
}
SuspendThread();
SleepThread();
}
}
+28 -10
View File
@@ -9,6 +9,7 @@
#include <queue>
#include <string>
#include <thread>
#include <utility>
#include <vector>
#include "common/common_types.h"
@@ -23,6 +24,8 @@ namespace iop {
struct sceSifQueueData;
}
using time_stamp = std::chrono::time_point<std::chrono::steady_clock, std::chrono::microseconds>;
struct SifRpcCommand {
bool started = true;
bool finished = true;
@@ -57,17 +60,21 @@ struct IopThread {
Delay,
};
IopThread(std::string n, void (*f)(), s32 ID) : name(n), function(f), thID(ID) {
thread = co_create(0x300000, f);
IopThread(std::string n, void (*f)(), s32 ID, u32 priority)
: name(std::move(n)), function(f), priority(priority), thID(ID) {
thread = co_create(0x300000, functionWrapper);
}
~IopThread() { co_delete(thread); }
static void functionWrapper();
std::string name;
void (*function)();
cothread_t thread;
State state = State::Dormant;
Wait waitType = Wait::None;
time_stamp resumeTime = {};
u32 priority = 0;
s32 thID = -1;
};
@@ -84,19 +91,20 @@ class IOP_Kernel {
IOP_Kernel() {
// this ugly hack
threads.reserve(16);
CreateThread("null-thread", nullptr);
CreateThread("null-thread", nullptr, 0);
CreateMbx();
kernel_thread = co_active();
}
~IOP_Kernel();
s32 CreateThread(std::string n, void (*f)());
s32 CreateThread(std::string n, void (*f)(), u32 priority);
s32 ExitThread();
void StartThread(s32 id);
void SuspendThread();
void DelayThread(u32 usec);
void SleepThread();
void WakeupThread(s32 id);
void dispatchAll();
time_stamp dispatch();
void set_rpc_queue(iop::sceSifQueueData* qd, u32 thread);
void rpc_loop(iop::sceSifQueueData* qd);
void shutdown();
@@ -104,7 +112,10 @@ class IOP_Kernel {
/*!
* Get current thread ID.
*/
s32 getCurrentThread() { return _currentThread; }
s32 getCurrentThread() {
ASSERT(_currentThread);
return _currentThread->thID;
}
/*!
* Create a message box
@@ -158,18 +169,25 @@ class IOP_Kernel {
s32 recvSize);
private:
void runThread(s32 id);
void runThread(IopThread* thread);
void exitThread();
void updateDelay();
void processWakeups();
IopThread* schedNext();
time_stamp nextWakeup();
cothread_t kernel_thread;
s32 _nextThID = 0;
s32 _currentThread = {-1};
IopThread* _currentThread = nullptr;
std::vector<IopThread> threads;
std::vector<std::queue<void*>> mbxs;
std::vector<SifRecord> sif_records;
std::vector<Semaphore> semas;
std::queue<int> wakeup_queue;
bool mainThreadSleep = false;
FILE* iso_disc_file = nullptr;
std::mutex sif_mtx;
std::mutex sif_mtx, wakeup_mtx;
};
#endif // JAK_IOP_KERNEL_H
+9 -20
View File
@@ -55,33 +55,22 @@ void* IOP::iop_alloc(int size) {
return mem;
}
void IOP::wait_run_iop() {
std::unique_lock<std::mutex> lk(iters_mutex);
if (iop_iters_des > iop_iters_act) {
iop_iters_act++;
return;
}
iop_run_cv.wait(lk, [&] { return iop_iters_des > iop_iters_act; });
iop_iters_act++;
void IOP::wait_run_iop(
std::chrono::time_point<std::chrono::steady_clock, std::chrono::microseconds> wakeup) {
std::unique_lock<std::mutex> lk(run_cv_mutex);
iop_run_cv.wait_until(lk, wakeup);
}
void IOP::kill_from_ee() {
want_exit = true;
signal_run_iop(true);
signal_run_iop();
}
void IOP::signal_run_iop(bool force) {
std::unique_lock<std::mutex> lk(iters_mutex);
if (iop_iters_act == iop_iters_des || force) {
iop_iters_des++; // todo, tune this
if (iop_iters_des - iop_iters_act > 500) {
iop_iters_des = iop_iters_act + 500;
}
iop_run_cv.notify_all();
}
void IOP::signal_run_iop() {
std::unique_lock<std::mutex> lk(run_cv_mutex);
iop_run_cv.notify_all();
}
IOP::~IOP() {
reset_allocator();
}
}
+5 -5
View File
@@ -19,8 +19,10 @@ class IOP {
void wait_for_overlord_start_cmd();
void wait_for_overlord_init_finish();
void signal_overlord_init_finish();
void signal_run_iop(bool force);
void wait_run_iop();
void signal_run_iop();
void wait_run_iop(
std::chrono::time_point<std::chrono::steady_clock, std::chrono::microseconds> duration);
void kill_from_ee();
void set_ee_main_mem(u8* mem) { ee_main_mem = mem; }
@@ -33,14 +35,12 @@ class IOP {
IOP_Kernel kernel;
u8* ee_main_mem = nullptr;
u64 iop_iters_des = 0;
u64 iop_iters_act = 0;
bool want_exit = false;
private:
std::vector<void*> allocations;
std::condition_variable cv;
std::mutex iop_mutex, iters_mutex;
std::mutex iop_mutex, run_cv_mutex;
bool overlord_init_done = false;
std::condition_variable iop_run_cv;
};