mirror/rpcs3
1
0
Fork 0
mirror of https://github.com/RPCS3/rpcs3.git synced 2024-11-17 08:11:51 +00:00
Code Issues Packages Projects Releases Wiki Activity

SPU: Cache reservation memory direct access handle (optimization)

Browse Source
This commit is contained in:
Eladash 2021-12-03 17:28:39 +02:00 committed by Ivan
parent 1458349cb3
commit 3dda72e47f
2 changed files with 64 additions and 31 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
Utilities/Thread.h
View File

@ -276,8 +276,8 @@ public:
}
// Wait for both thread sync var and provided atomic var
template <atomic_wait::op Op = atomic_wait::op::eq, typename T, typename U>
static inline void wait_on(T& wait, U old, u64 usec = -1)
template <uint Max, typename Func>
static inline void wait_on_custom(Func&& setter, u64 usec = -1)
{
auto _this = g_tls_this_thread;
@ -286,13 +286,19 @@ public:
return;
}
atomic_wait::list<3> list{};
list.set<0, Op>(wait, old);
list.set<1>(_this->m_sync, 0, 4 + 1);
list.set<2>(_this->m_taskq, nullptr);
atomic_wait::list<Max + 2> list{};
list.template set<Max>(_this->m_sync, 0, 4 + 1);
list.template set<Max + 1>(_this->m_taskq, nullptr);
setter(list);
list.wait(atomic_wait_timeout{usec <= 0xffff'ffff'ffff'ffff / 1000 ? usec * 1000 : 0xffff'ffff'ffff'ffff});
}
template <atomic_wait::op Op = atomic_wait::op::eq, typename T, typename U>
static inline void wait_on(T& wait, U old, u64 usec = -1)
{
wait_on_custom<1>([&](atomic_wait::list<3>& list){ list.set<0, Op>(wait, old); }, usec);
}
// Exit.
[[noreturn]] static void emergency_exit(std::string_view reason);

77
rpcs3/Emu/Cell/SPUThread.cpp
View File

@ -3942,40 +3942,46 @@ s64 spu_thread::get_ch_value(u32 ch)
spu_function_logger logger(*this, "MFC Events read");
if (mask1 & SPU_EVENT_LR && raddr)
state += cpu_flag::wait;
using resrv_ptr = std::add_pointer_t<decltype(rdata)>;
resrv_ptr resrv_mem{};
std::shared_ptr<utils::shm> rdata_shm;
if (raddr && mask1 & SPU_EVENT_LR)
{
if (mask1 != SPU_EVENT_LR && mask1 != SPU_EVENT_LR + SPU_EVENT_TM)
auto area = vm::get(vm::any, raddr);
if (area && (area->flags & vm::preallocated) && vm::check_addr(raddr))
{
// Combining LR with other flags needs another solution
fmt::throw_exception("Not supported: event mask 0x%x", mask1);
// Obtain pointer to pre-allocated storage
resrv_mem = vm::get_super_ptr<decltype(rdata)>(raddr);
}
else if (area)
{
// Ensure possesion over reservation memory so it won't be deallocated
auto [base_addr, shm_] = area->peek(raddr);
if (shm_)
{
const u32 data_offs = raddr - base_addr;
rdata_shm = std::move(shm_);
vm::writer_lock{}, resrv_mem = reinterpret_cast<resrv_ptr>(rdata_shm->map_self() + data_offs);
}
}
for (; !events.count; events = get_events(mask1, false, true))
if (!resrv_mem)
{
const auto old = state.add_fetch(cpu_flag::wait);
if (is_stopped(old))
{
return -1;
}
if (is_paused(old))
{
// Ensure reservation data won't change while paused for debugging purposes
check_state();
continue;
}
vm::reservation_notifier(raddr).wait(rtime, -128, atomic_wait_timeout{100'000});
spu_log.error("A dangling reservation address has been found while reading SPU_RdEventStat channel. (addr=0x%x, events_mask=0x%x)", raddr, mask1);
raddr = 0;
set_events(SPU_EVENT_LR);
}
check_state();
return events.events & mask1;
}
for (; !events.count; events = get_events(mask1, true, true))
for (; !events.count; events = get_events(mask1 & ~SPU_EVENT_LR, true, true))
{
const auto old = state.add_fetch(cpu_flag::wait);
const auto old = +state;
if (is_stopped(old))
{
@ -3984,7 +3990,28 @@ s64 spu_thread::get_ch_value(u32 ch)
if (is_paused(old))
{
// Ensure spu_thread::rdata's stagnancy while the thread is paused for debugging purposes
check_state();
state += cpu_flag::wait;
continue;
}
// Optimized check
if (raddr && (!vm::check_addr(raddr) || rtime != vm::reservation_acquire(raddr) || !cmp_rdata(rdata, *resrv_mem)))
{
raddr = 0;
set_events(SPU_EVENT_LR);
continue;
}
if (raddr)
{
thread_ctrl::wait_on_custom<2>([&](atomic_wait::list<4>& list)
{
list.set<0>(state, old);
list.set<1>(vm::reservation_notifier(raddr), rtime, -128);
}, 100);
continue;
}