1
0
mirror of https://github.com/RPCS3/rpcs3.git synced 2024-11-25 12:12:50 +01:00

SPU: improve TSX usage

Reduce transaction failure amount
Remove vm::try_to_lock
This commit is contained in:
Nekotekina 2018-04-03 22:42:47 +03:00
parent d392379c7a
commit 2b5cf2455f
7 changed files with 361 additions and 221 deletions

View File

@ -100,6 +100,11 @@ public:
{
return (m_value.load() % c_one) != 0;
}
bool is_lockable() const
{
return m_value.load() >= c_min;
}
};
// Simplified shared (reader) lock implementation.

View File

@ -992,9 +992,10 @@ constexpr FORCE_INLINE error_code::not_an_error not_an_error(const T& value)
}
// Synchronization helper (cache-friendly busy waiting)
inline void busy_wait(std::size_t count = 100)
inline void busy_wait(std::size_t cycles = 3000)
{
while (count--) _mm_pause();
const u64 s = __rdtsc();
do _mm_pause(); while (__rdtsc() - s < cycles);
}
// Rotate helpers

View File

@ -964,7 +964,7 @@ extern bool ppu_stwcx(ppu_thread& ppu, u32 addr, u32 reg_value)
if (s_use_rtm && utils::transaction_enter())
{
if (!vm::reader_lock{vm::try_to_lock})
if (!vm::g_mutex.is_lockable())
{
_xabort(0);
}
@ -1008,7 +1008,7 @@ extern bool ppu_stdcx(ppu_thread& ppu, u32 addr, u64 reg_value)
if (s_use_rtm && utils::transaction_enter())
{
if (!vm::reader_lock{vm::try_to_lock})
if (!vm::g_mutex.is_lockable())
{
_xabort(0);
}

View File

@ -312,6 +312,8 @@ std::string SPUThread::dump() const
{
std::string ret = cpu_thread::dump();
// Print some transaction statistics
fmt::append(ret, "\nTX: %u; Fail: %u", tx_success, tx_failure);
fmt::append(ret, "\nTag Mask: 0x%08x", ch_tag_mask);
fmt::append(ret, "\nMFC Stall: 0x%08x", ch_stall_mask);
fmt::append(ret, "\nMFC Queue Size: %u", mfc_size);
@ -349,8 +351,8 @@ void SPUThread::cpu_init()
srr0 = 0;
mfc_size = 0;
mfc_barrier = -1;
mfc_fence = -1;
mfc_barrier = 0;
mfc_fence = 0;
ch_tag_upd = 0;
ch_tag_mask = 0;
mfc_prxy_mask = 0;
@ -487,14 +489,10 @@ void SPUThread::cpu_task()
void SPUThread::cpu_mem()
{
mfc_barrier = -1;
mfc_fence = -1;
}
void SPUThread::cpu_unmem()
{
mfc_barrier = -1;
mfc_fence = -1;
}
SPUThread::~SPUThread()
@ -749,13 +747,13 @@ bool SPUThread::do_dma_check(const spu_mfc_cmd& args)
for (u32 i = 0; i < mfc_size; i++)
{
if (mfc_queue[i].cmd == MFC_BARRIER_CMD)
if ((mfc_queue[i].cmd & ~0xc) == MFC_BARRIER_CMD)
{
mfc_barrier |= -1;
continue;
}
if (mfc_queue[i].cmd != MFC_EIEIO_CMD)
if (true)
{
const u32 _mask = 1u << mfc_queue[i].tag;
@ -772,12 +770,9 @@ bool SPUThread::do_dma_check(const spu_mfc_cmd& args)
if (mfc_barrier & mask || (args.cmd & MFC_FENCE_MASK && mfc_fence & mask))
{
mfc_barrier = -1;
mfc_fence = -1;
return false;
}
vm::passive_lock(*this);
return true;
}
@ -841,8 +836,14 @@ bool SPUThread::do_list_transfer(spu_mfc_cmd& args)
return true;
}
bool SPUThread::do_putlluc(const spu_mfc_cmd& args)
void SPUThread::do_putlluc(const spu_mfc_cmd& args)
{
if (raddr && args.eal == raddr)
{
ch_event_stat |= SPU_EVENT_LR;
raddr = 0;
}
const u32 addr = args.eal;
auto& data = vm::_ref<decltype(rdata)>(addr);
const auto to_write = _ref<decltype(rdata)>(args.lsa & 0x3ffff);
@ -852,7 +853,8 @@ bool SPUThread::do_putlluc(const spu_mfc_cmd& args)
// Store unconditionally
if (s_use_rtm && utils::transaction_enter())
{
if (!vm::reader_lock{vm::try_to_lock})
// First transaction attempt
if (!vm::g_mutex.is_lockable())
{
_xabort(0);
}
@ -861,34 +863,49 @@ bool SPUThread::do_putlluc(const spu_mfc_cmd& args)
vm::reservation_update(addr, 128);
vm::notify(addr, 128);
_xend();
tx_success++;
return;
}
else if (s_use_rtm)
{
vm::reader_lock lock;
if (utils::transaction_enter())
{
// Second transaction attempt
data = to_write;
vm::reservation_update(addr, 128);
_xend();
tx_success++;
vm::notify(addr, 128);
return;
}
else
{
vm::writer_lock lock(0);
data = to_write;
vm::reservation_update(addr, 128);
vm::notify(addr, 128);
tx_failure++;
}
}
return true;
vm::writer_lock lock(0);
vm::reservation_update(addr, 128, true);
_mm_sfence();
data = to_write;
_mm_sfence();
vm::reservation_update(addr, 128);
vm::notify(addr, 128);
}
void SPUThread::do_mfc()
void SPUThread::do_mfc(bool wait)
{
u32 removed = 0;
u32 barrier = 0;
u32 fence = 0;
// Check special value
if (UNLIKELY(mfc_barrier == -1 && mfc_fence == -1))
{
vm::passive_lock(*this);
}
// Process enqueued commands
std::remove_if(mfc_queue + 0, mfc_queue + mfc_size, [&](spu_mfc_cmd& args)
{
if (args.cmd == MFC_BARRIER_CMD || args.cmd == MFC_EIEIO_CMD)
if ((args.cmd & ~0xc) == MFC_BARRIER_CMD)
{
if (&args - mfc_queue <= removed)
{
@ -898,12 +915,8 @@ void SPUThread::do_mfc()
return true;
}
if (args.cmd == MFC_BARRIER_CMD)
{
// Block all tags
barrier |= -1;
}
return false;
}
@ -916,44 +929,70 @@ void SPUThread::do_mfc()
return false;
}
if (args.cmd & MFC_FENCE_MASK && fence & mask)
{
return false;
}
if (args.cmd & MFC_LIST_MASK)
{
if (!test(ch_stall_mask, mask) && do_list_transfer(args))
if (!test(ch_stall_mask, mask))
{
if (s_use_rtm)
{
if (do_list_transfer(args))
{
removed++;
return true;
}
}
else if (vm::passive_lock(*this, wait))
{
if (do_list_transfer(args))
{
vm::passive_unlock(*this);
removed++;
return true;
}
vm::passive_unlock(*this);
}
}
if (args.cmd & MFC_BARRIER_MASK)
{
barrier |= mask;
}
if (test(state, cpu_flag::stop))
fence |= mask;
return false;
}
if (args.size)
{
barrier |= -1;
if (s_use_rtm)
{
do_dma_transfer(args);
}
else if (vm::passive_lock(*this, wait))
{
do_dma_transfer(args);
vm::passive_unlock(*this);
}
else
{
if (args.cmd & MFC_BARRIER_MASK)
{
barrier |= mask;
}
fence |= mask;
return false;
}
if (args.cmd == MFC_PUTQLLUC_CMD)
{
if (do_putlluc(args))
{
removed++;
return true;
}
barrier |= -1;
return false;
}
// Also ignore MFC_SYNC_CMD
if (args.size)
else if (args.cmd == MFC_PUTQLLUC_CMD)
{
do_dma_transfer(args);
do_putlluc(args);
}
removed++;
@ -983,25 +1022,20 @@ void SPUThread::do_mfc()
u32 SPUThread::get_mfc_completed()
{
u32 completed = ch_tag_mask;
for (u32 i = 0; i < mfc_size; i++)
{
if (mfc_queue[i].cmd != MFC_BARRIER_CMD && mfc_queue[i].cmd != MFC_EIEIO_CMD)
{
completed &= ~(1u << mfc_queue[i].tag);
}
}
return completed;
return ch_tag_mask & ~mfc_fence;
}
bool SPUThread::process_mfc_cmd(spu_mfc_cmd args)
{
// Stall infinitely if MFC queue is full
while (mfc_size >= 16)
while (UNLIKELY(mfc_size >= 16))
{
vm::temporary_unlock(*this);
do_mfc();
if (mfc_size < 16)
{
break;
}
if (test(state, cpu_flag::stop))
{
@ -1020,19 +1054,14 @@ bool SPUThread::process_mfc_cmd(spu_mfc_cmd args)
{
auto& data = vm::_ref<decltype(rdata)>(args.eal);
const u32 _addr = args.eal;
const u64 _time = vm::reservation_acquire(raddr, 128);
if (raddr && raddr != args.eal)
{
ch_event_stat |= SPU_EVENT_LR;
}
const bool is_polling = false;// raddr == _addr && rtime == _time; // TODO
raddr = args.eal;
_mm_lfence();
raddr = _addr;
rtime = _time;
const bool is_polling = false; // TODO
if (is_polling)
{
@ -1056,31 +1085,48 @@ bool SPUThread::process_mfc_cmd(spu_mfc_cmd args)
thread_ctrl::wait_for(100);
}
}
else if (s_use_rtm && utils::transaction_enter())
{
if (!vm::reader_lock{vm::try_to_lock})
{
_xabort(0);
}
// Do several attemps
for (uint i = 0; i < 5; i++)
{
rtime = vm::reservation_acquire(raddr, 128);
rdata = data;
_xend();
_mm_lfence();
// Check LSB: atomic store may be in progress
if (LIKELY((rtime & 1) == 0))
{
rdata = data;
_mm_lfence();
if (LIKELY(vm::reservation_acquire(raddr, 128) == rtime))
{
// Copy to LS
tx_success++;
_ref<decltype(rdata)>(args.lsa & 0x3ffff) = rdata;
ch_atomic_stat.set_value(MFC_GETLLAR_SUCCESS);
return true;
}
else
{
rdata = data;
_mm_lfence();
}
// Hack: ensure no other atomic updates have happened during reading the data
if (is_polling || UNLIKELY(vm::reservation_acquire(raddr, 128) != rtime))
busy_wait(300);
}
if (s_use_rtm && utils::transaction_enter())
{
// TODO: vm::check_addr
rtime = vm::reservation_acquire(raddr, 128);
if (rtime & 1)
{
_xabort(0);
}
rdata = data;
_xend();
tx_success++;
}
else
{
tx_failure++;
vm::reader_lock lock;
rtime = vm::reservation_acquire(raddr, 128);
rdata = data;
@ -1088,7 +1134,6 @@ bool SPUThread::process_mfc_cmd(spu_mfc_cmd args)
// Copy to LS
_ref<decltype(rdata)>(args.lsa & 0x3ffff) = rdata;
ch_atomic_stat.set_value(MFC_GETLLAR_SUCCESS);
return true;
}
@ -1101,12 +1146,13 @@ bool SPUThread::process_mfc_cmd(spu_mfc_cmd args)
bool result = false;
if (raddr == args.eal && rtime == vm::reservation_acquire(raddr, 128) && rdata == data)
if (raddr == args.eal && rtime == vm::reservation_acquire(raddr, 128))
{
// TODO: vm::check_addr
if (s_use_rtm && utils::transaction_enter())
{
if (!vm::reader_lock{vm::try_to_lock})
// First transaction attempt
if (!vm::g_mutex.is_lockable())
{
_xabort(0);
}
@ -1121,20 +1167,59 @@ bool SPUThread::process_mfc_cmd(spu_mfc_cmd args)
}
_xend();
tx_success++;
}
else
else if (s_use_rtm)
{
vm::writer_lock lock;
// Second transaction attempt
vm::reader_lock lock;
if (utils::transaction_enter())
{
if (rtime == vm::reservation_acquire(raddr, 128) && rdata == data)
{
data = to_write;
result = true;
vm::reservation_update(raddr, 128);
}
_xend();
tx_success++;
if (result)
{
// First transaction attempt usually fails on vm::notify
vm::notify(raddr, 128);
}
}
else
{
// Don't fallback to heavyweight lock, just give up
tx_failure++;
}
}
else if (rdata == data)
{
// Full lock (heavyweight)
vm::writer_lock lock(1);
if (rtime == vm::reservation_acquire(raddr, 128) && rdata == data)
{
vm::reservation_update(raddr, 128, true);
_mm_sfence();
data = to_write;
_mm_sfence();
result = true;
vm::reservation_update(raddr, 128);
vm::notify(raddr, 128);
tx_success++;
}
else
{
tx_failure++;
}
}
}
if (result)
@ -1156,49 +1241,21 @@ bool SPUThread::process_mfc_cmd(spu_mfc_cmd args)
}
case MFC_PUTLLUC_CMD:
{
if (raddr && args.eal == raddr)
{
ch_event_stat |= SPU_EVENT_LR;
raddr = 0;
}
auto& data = vm::_ref<decltype(rdata)>(args.eal);
const auto to_write = _ref<decltype(rdata)>(args.lsa & 0x3ffff);
vm::reservation_acquire(args.eal, 128);
// Store unconditionally
// TODO: vm::check_addr
if (s_use_rtm && utils::transaction_enter())
{
if (!vm::reader_lock{vm::try_to_lock})
{
_xabort(0);
}
data = to_write;
vm::reservation_update(args.eal, 128);
vm::notify(args.eal, 128);
_xend();
ch_atomic_stat.set_value(MFC_PUTLLUC_SUCCESS);
return true;
}
vm::writer_lock lock(0);
data = to_write;
vm::reservation_update(args.eal, 128);
vm::notify(args.eal, 128);
do_putlluc(args);
ch_atomic_stat.set_value(MFC_PUTLLUC_SUCCESS);
return true;
}
case MFC_PUTQLLUC_CMD:
{
if (UNLIKELY(!do_dma_check(args) || !do_putlluc(args)))
if (UNLIKELY(!do_dma_check(args)))
{
args.size = 0;
mfc_queue[mfc_size++] = args;
mfc_fence |= 1u << args.tag;
}
else
{
do_putlluc(args);
}
return true;
@ -1222,15 +1279,34 @@ bool SPUThread::process_mfc_cmd(spu_mfc_cmd args)
{
if (LIKELY(args.size <= 0x4000))
{
if (UNLIKELY(!do_dma_check(args)))
if (LIKELY(do_dma_check(args)))
{
mfc_queue[mfc_size++] = args;
return true;
}
if (LIKELY(args.size))
{
if (s_use_rtm)
{
do_dma_transfer(args);
return true;
}
else if (vm::passive_lock(*this, false))
{
do_dma_transfer(args);
vm::passive_unlock(*this);
return true;
}
}
else
{
return true;
}
}
mfc_queue[mfc_size++] = args;
mfc_fence |= 1u << args.tag;
if (args.cmd & MFC_BARRIER_MASK)
{
mfc_barrier |= 1u << args.tag;
}
return true;
@ -1250,9 +1326,33 @@ bool SPUThread::process_mfc_cmd(spu_mfc_cmd args)
{
if (LIKELY(args.size <= 0x4000))
{
if (UNLIKELY(!do_dma_check(args) || test(ch_stall_mask, 1u << args.tag) || !do_list_transfer(args)))
if (LIKELY(do_dma_check(args) && !test(ch_stall_mask, 1u << args.tag)))
{
if (s_use_rtm)
{
if (LIKELY(do_list_transfer(args)))
{
return true;
}
}
else if (vm::passive_lock(*this, false))
{
if (LIKELY(do_list_transfer(args)))
{
vm::passive_unlock(*this);
return true;
}
vm::passive_unlock(*this);
}
}
mfc_queue[mfc_size++] = args;
mfc_fence |= 1u << args.tag;
if (args.cmd & MFC_BARRIER_MASK)
{
mfc_barrier |= 1u << args.tag;
}
return true;
@ -1262,6 +1362,7 @@ bool SPUThread::process_mfc_cmd(spu_mfc_cmd args)
}
case MFC_BARRIER_CMD:
case MFC_EIEIO_CMD:
case MFC_SYNC_CMD:
{
if (mfc_size == 0)
{
@ -1270,20 +1371,7 @@ bool SPUThread::process_mfc_cmd(spu_mfc_cmd args)
else
{
mfc_queue[mfc_size++] = args;
}
return true;
}
case MFC_SYNC_CMD:
{
if (LIKELY(do_dma_check(args)))
{
_mm_mfence();
}
else
{
args.size = 0;
mfc_queue[mfc_size++] = args;
mfc_barrier |= -1;
}
return true;
@ -1394,14 +1482,18 @@ bool SPUThread::get_ch_value(u32 ch, u32& out)
{
for (int i = 0; i < 10 && channel.get_count() == 0; i++)
{
vm::temporary_unlock(*this);
if (!s_use_rtm && mfc_size)
{
do_mfc();
}
else
{
busy_wait();
}
}
while (!channel.try_pop(out))
{
vm::temporary_unlock(*this);
if (test(state, cpu_flag::stop))
{
return false;
@ -1426,9 +1518,15 @@ bool SPUThread::get_ch_value(u32 ch, u32& out)
{
for (int i = 0; i < 10 && ch_in_mbox.get_count() == 0; i++)
{
vm::temporary_unlock(*this);
if (!s_use_rtm && mfc_size)
{
do_mfc();
}
else
{
busy_wait();
}
}
if (const uint old_count = ch_in_mbox.try_pop(out))
{
@ -1440,8 +1538,6 @@ bool SPUThread::get_ch_value(u32 ch, u32& out)
return true;
}
vm::temporary_unlock(*this);
if (test(state & cpu_flag::stop))
{
return false;
@ -1453,6 +1549,11 @@ bool SPUThread::get_ch_value(u32 ch, u32& out)
case MFC_RdTagStat:
{
if (!s_use_rtm && mfc_size && ch_tag_upd)
{
do_mfc();
}
if (ch_tag_stat.get_count())
{
out = ch_tag_stat.get_value();
@ -1533,6 +1634,11 @@ bool SPUThread::get_ch_value(u32 ch, u32& out)
return true;
}
if (!s_use_rtm)
{
do_mfc();
}
vm::waiter waiter;
if (ch_event_mask & SPU_EVENT_LR)
@ -1547,8 +1653,6 @@ bool SPUThread::get_ch_value(u32 ch, u32& out)
while (!(res = get_events(true)))
{
vm::temporary_unlock(*this);
if (test(state & cpu_flag::stop))
{
return false;
@ -1587,12 +1691,15 @@ bool SPUThread::set_ch_value(u32 ch, u32 value)
case SPU_WrOutIntrMbox:
{
if (!s_use_rtm)
{
do_mfc(false);
}
if (offset >= RAW_SPU_BASE_ADDR)
{
while (!ch_out_intr_mbox.try_push(value))
{
vm::temporary_unlock(*this);
if (test(state & cpu_flag::stop))
{
return false;
@ -1737,10 +1844,13 @@ bool SPUThread::set_ch_value(u32 ch, u32 value)
case SPU_WrOutMbox:
{
if (!s_use_rtm)
{
do_mfc(false);
}
while (!ch_out_mbox.try_push(value))
{
vm::temporary_unlock(*this);
if (test(state & cpu_flag::stop))
{
return false;
@ -1782,6 +1892,11 @@ bool SPUThread::set_ch_value(u32 ch, u32 value)
break;
}
if (!s_use_rtm)
{
do_mfc(false);
}
const u32 completed = get_mfc_completed();
if (!value)
@ -1846,7 +1961,7 @@ bool SPUThread::set_ch_value(u32 ch, u32 value)
// Reset stall status for specified tag
if (::test_and_reset(ch_stall_mask, 1u << value))
{
do_mfc();
do_mfc(false);
}
return true;
@ -1901,6 +2016,11 @@ bool SPUThread::stop_and_signal(u32 code)
{
LOG_TRACE(SPU, "stop_and_signal(code=0x%x)", code);
if (!s_use_rtm)
{
do_mfc();
}
if (offset >= RAW_SPU_BASE_ADDR)
{
status.atomic_op([code](u32& status)
@ -1941,8 +2061,6 @@ bool SPUThread::stop_and_signal(u32 code)
// HACK: wait for executable code
while (!_ref<u32>(pc))
{
vm::temporary_unlock(*this);
if (test(state & cpu_flag::stop))
{
return false;
@ -1956,7 +2074,6 @@ bool SPUThread::stop_and_signal(u32 code)
case 0x001:
{
vm::temporary_unlock(*this);
thread_ctrl::wait_for(1000); // hack
return true;
}
@ -1971,8 +2088,6 @@ bool SPUThread::stop_and_signal(u32 code)
{
/* ===== sys_spu_thread_receive_event ===== */
vm::temporary_unlock(*this);
u32 spuq;
if (!ch_out_mbox.try_pop(spuq))
@ -2128,8 +2243,6 @@ bool SPUThread::stop_and_signal(u32 code)
{
/* ===== sys_spu_thread_group_exit ===== */
vm::temporary_unlock(*this);
u32 value;
if (!ch_out_mbox.try_pop(value))
@ -2163,8 +2276,6 @@ bool SPUThread::stop_and_signal(u32 code)
{
/* ===== sys_spu_thread_exit ===== */
vm::temporary_unlock(*this);
if (!ch_out_mbox.get_count())
{
fmt::throw_exception("sys_spu_thread_exit(): Out_MBox is empty" HERE);

View File

@ -591,12 +591,15 @@ public:
std::shared_ptr<class spu_recompiler_base> spu_rec;
u32 recursion_level = 0;
u64 tx_success = 0;
u64 tx_failure = 0;
void push_snr(u32 number, u32 value);
void do_dma_transfer(const spu_mfc_cmd& args);
bool do_dma_check(const spu_mfc_cmd& args);
bool do_list_transfer(spu_mfc_cmd& args);
bool do_putlluc(const spu_mfc_cmd& args);
void do_mfc();
void do_putlluc(const spu_mfc_cmd& args);
void do_mfc(bool wait = true);
u32 get_mfc_completed();
bool process_mfc_cmd(spu_mfc_cmd args);

View File

@ -66,25 +66,56 @@ namespace vm
}
}
void passive_lock(cpu_thread& cpu)
bool passive_lock(cpu_thread& cpu, bool wait)
{
if (g_tls_locked && *g_tls_locked == &cpu)
if (UNLIKELY(g_tls_locked && *g_tls_locked == &cpu))
{
return;
return true;
}
if (LIKELY(g_mutex.is_lockable()))
{
// Optimistic path (hope that mutex is not exclusively locked)
_register_lock(&cpu);
if (UNLIKELY(!g_mutex.is_lockable()))
{
passive_unlock(cpu);
if (!wait)
{
return false;
}
::reader_lock lock(g_mutex);
_register_lock(&cpu);
}
}
else
{
if (!wait)
{
return false;
}
::reader_lock lock(g_mutex);
_register_lock(&cpu);
}
return true;
}
void passive_unlock(cpu_thread& cpu)
{
if (g_tls_locked)
if (auto& ptr = g_tls_locked)
{
g_tls_locked->compare_and_swap_test(&cpu, nullptr);
::reader_lock lock(g_mutex);
g_tls_locked = nullptr;
*ptr = nullptr;
ptr = nullptr;
if (test(cpu.state, cpu_flag::memory))
{
cpu.state -= cpu_flag::memory;
}
}
}
@ -140,11 +171,6 @@ namespace vm
}
}
reader_lock::reader_lock(const try_to_lock_t&)
: locked(g_mutex.try_lock_shared())
{
}
reader_lock::~reader_lock()
{
if (locked)
@ -196,11 +222,6 @@ namespace vm
}
}
writer_lock::writer_lock(const try_to_lock_t&)
: locked(g_mutex.try_lock())
{
}
writer_lock::~writer_lock()
{
if (locked)
@ -250,16 +271,16 @@ namespace vm
return g_pages[addr >> 12][addr].load(std::memory_order_acquire);
}
void reservation_update(u32 addr, u32 _size)
void reservation_update(u32 addr, u32 _size, bool lsb)
{
// Update reservation info with new timestamp (unsafe, assume allocated)
(*g_pages[addr >> 12].reservations)[(addr & 0xfff) >> 7].store(__rdtsc(), std::memory_order_release);
(*g_pages[addr >> 12].reservations)[(addr & 0xfff) >> 7].store((__rdtsc() & -2) | lsb, std::memory_order_release);
}
void waiter::init()
{
// Register waiter
writer_lock lock(0);
vm::writer_lock lock(0);
g_waiters.emplace_back(this);
}
@ -292,7 +313,7 @@ namespace vm
waiter::~waiter()
{
// Unregister waiter
writer_lock lock(0);
vm::writer_lock lock(0);
// Find waiter
const auto found = std::find(g_waiters.cbegin(), g_waiters.cend(), this);
@ -360,7 +381,7 @@ namespace vm
bool page_protect(u32 addr, u32 size, u8 flags_test, u8 flags_set, u8 flags_clear)
{
writer_lock lock;
vm::writer_lock lock(0);
if (!size || (size | addr) % 4096)
{
@ -546,7 +567,7 @@ namespace vm
block_t::~block_t()
{
writer_lock lock;
vm::writer_lock lock(0);
// Deallocate all memory
for (auto& entry : m_map)
@ -557,7 +578,7 @@ namespace vm
u32 block_t::alloc(const u32 orig_size, u32 align, const uchar* data, u32 sup)
{
writer_lock lock;
vm::writer_lock lock(0);
// Align to minimal page size
const u32 size = ::align(orig_size, 4096);
@ -604,7 +625,7 @@ namespace vm
u32 block_t::falloc(u32 addr, const u32 orig_size, const uchar* data, u32 sup)
{
writer_lock lock;
vm::writer_lock lock(0);
// align to minimal page size
const u32 size = ::align(orig_size, 4096);
@ -641,7 +662,7 @@ namespace vm
u32 block_t::dealloc(u32 addr, uchar* data_out, u32* sup_out)
{
writer_lock lock;
vm::writer_lock lock(0);
const auto found = m_map.find(addr);
@ -690,14 +711,14 @@ namespace vm
u32 block_t::used()
{
writer_lock lock(0);
vm::writer_lock lock(0);
return imp_used(lock);
}
std::shared_ptr<block_t> map(u32 addr, u32 size, u64 flags)
{
writer_lock lock(0);
vm::writer_lock lock(0);
if (!size || (size | addr) % 4096)
{
@ -734,7 +755,7 @@ namespace vm
std::shared_ptr<block_t> unmap(u32 addr, bool must_be_empty)
{
writer_lock lock(0);
vm::writer_lock lock(0);
for (auto it = g_locations.begin(); it != g_locations.end(); it++)
{
@ -756,7 +777,7 @@ namespace vm
std::shared_ptr<block_t> get(memory_location_t location, u32 addr)
{
reader_lock lock;
vm::reader_lock lock;
if (location != any)
{

View File

@ -4,6 +4,7 @@
#include <functional>
#include <memory>
class shared_mutex;
class named_thread;
class cpu_thread;
@ -59,10 +60,12 @@ namespace vm
// Address type
enum addr_t : u32 {};
extern shared_mutex g_mutex;
extern thread_local atomic_t<cpu_thread*>* g_tls_locked;
// Register reader
void passive_lock(cpu_thread& cpu);
bool passive_lock(cpu_thread& cpu, bool wait = true);
// Unregister reader
void passive_unlock(cpu_thread& cpu);
@ -74,15 +77,12 @@ namespace vm
void temporary_unlock(cpu_thread& cpu) noexcept;
void temporary_unlock() noexcept;
constexpr struct try_to_lock_t{} try_to_lock{};
struct reader_lock final
{
const bool locked;
reader_lock(const reader_lock&) = delete;
reader_lock();
reader_lock(const try_to_lock_t&);
~reader_lock();
explicit operator bool() const { return locked; }
@ -93,8 +93,7 @@ namespace vm
const bool locked;
writer_lock(const writer_lock&) = delete;
writer_lock(int full = 1);
writer_lock(const try_to_lock_t&);
writer_lock(int full);
~writer_lock();
explicit operator bool() const { return locked; }
@ -104,7 +103,7 @@ namespace vm
u64 reservation_acquire(u32 addr, u32 size);
// End atomic update
void reservation_update(u32 addr, u32 size);
void reservation_update(u32 addr, u32 size, bool lsb = false);
// Check and notify memory changes at address
void notify(u32 addr, u32 size);