Rough CriticalSection work.

This commit is contained in:
Ben Vanik 2013-01-31 02:50:37 -08:00
parent fdbd30acb7
commit 56912aa2bd
1 changed files with 182 additions and 3 deletions

View File

@ -305,9 +305,182 @@ void RtlImageXexHeaderField_shim(
}
//RtlInitializeCriticalSection
//RtlEnterCriticalSection
//RtlLeaveCriticalSection
// Unfortunately the Windows RTL_CRITICAL_SECTION object is bigger than the one
// on the 360 (32b vs. 28b). This means that we can't do in-place splatting of
// the critical sections. Also, the 360 never calls RtlDeleteCriticalSection
// so we can't clean up the native handles.
//
// Because of this, we reimplement it poorly. Hooray.
// We have 28b to work with so we need to be careful. We map our struct directly
// into guest memory, as it should be opaque and so long as our size is right
// the user code will never know.
//
// This would be good to put in xethunk for inlining.
//
// Ref: http://msdn.microsoft.com/en-us/magazine/cc164040.aspx
// Ref: http://svn.reactos.org/svn/reactos/trunk/reactos/lib/rtl/critical.c?view=markup
// This structure tries to match the one on the 360 as best I can figure out.
// Unfortunately some games have the critical sections pre-initialized in
// their embedded data and InitializeCriticalSection will never be called.
namespace {
#pragma pack(push, 1)
typedef struct {
uint8_t unknown0;
uint8_t spin_count_div_256; // * 256
uint8_t unknown0b;
uint8_t unknown0c;
uint32_t unknown1; // maybe the handle to the event?
uint32_t unknown2; // head of queue, pointing to this offset
uint32_t unknown3; // tail of queue?
int32_t lock_count; // -1 -> 0 on first lock
uint16_t recursion_count; // 0 -> 1 on first lock
uint32_t owning_thread_id; // 0 unless locked
} X_RTL_CRITICAL_SECTION;
#pragma pack(pop)
}
void RtlInitializeCriticalSection_shim(
xe_ppc_state_t* ppc_state, KernelState* state) {
// VOID
// _Out_ LPCRITICAL_SECTION lpCriticalSection
uint32_t cs_ptr = SHIM_GET_ARG_32(0);
XELOGD(XT("RtlInitializeCriticalSection(%.8X)"), cs_ptr);
X_RTL_CRITICAL_SECTION* cs = (X_RTL_CRITICAL_SECTION*)SHIM_MEM_ADDR(cs_ptr);
cs->spin_count_div_256 = 0;
cs->lock_count = -1;
cs->recursion_count = 0;
cs->owning_thread_id = 0;
}
void RtlInitializeCriticalSectionAndSpinCount_shim(
xe_ppc_state_t* ppc_state, KernelState* state) {
// NTSTATUS
// _Out_ LPCRITICAL_SECTION lpCriticalSection,
// _In_ DWORD dwSpinCount
uint32_t cs_ptr = SHIM_GET_ARG_32(0);
uint32_t spin_count = SHIM_GET_ARG_32(1);
XELOGD(XT("RtlInitializeCriticalSectionAndSpinCount(%.8X, %d)"),
cs_ptr, spin_count);
// Spin count is rouned up to 256 intervals then packed in.
uint32_t spin_count_div_256 = (uint32_t)floor(spin_count / 256.0f + 0.5f);
X_RTL_CRITICAL_SECTION* cs = (X_RTL_CRITICAL_SECTION*)SHIM_MEM_ADDR(cs_ptr);
cs->spin_count_div_256 = spin_count_div_256;
cs->lock_count = -1;
cs->recursion_count = 0;
cs->owning_thread_id = 0;
SHIM_SET_RETURN(X_STATUS_SUCCESS);
}
void RtlEnterCriticalSection_shim(
xe_ppc_state_t* ppc_state, KernelState* state) {
// VOID
// _Inout_ LPCRITICAL_SECTION lpCriticalSection
uint32_t cs_ptr = SHIM_GET_ARG_32(0);
XELOGD(XT("RtlEnterCriticalSection(%.8X)"), cs_ptr);
X_RTL_CRITICAL_SECTION* cs = (X_RTL_CRITICAL_SECTION*)SHIM_MEM_ADDR(cs_ptr);
// TODO(benvanik): get current thread ID.
uint32_t thread_id = 0;
uint32_t spin_wait_remaining = cs->spin_count_div_256 * 256;
spin:
if (xe_atomic_inc_32(&cs->lock_count)) {
// If this thread already owns the CS increment the recursion count.
if (cs->owning_thread_id == thread_id) {
++cs->recursion_count;
return;
}
// Thread was locked - spin wait.
if (--spin_wait_remaining) {
goto spin;
}
// All out of spin waits, create a full waiter.
// TODO(benvanik): contention - do a real wait!
XELOGE(XT("RtlEnterCriticalSection tried to really lock!"));
}
// Now own the lock.
cs->owning_thread_id = thread_id;
cs->recursion_count = 1;
}
void RtlTryEnterCriticalSection_shim(
xe_ppc_state_t* ppc_state, KernelState* state) {
// DWORD
// _Inout_ LPCRITICAL_SECTION lpCriticalSection
uint32_t cs_ptr = SHIM_GET_ARG_32(0);
XELOGD(XT("RtlTryEnterCriticalSection(%.8X)"), cs_ptr);
X_RTL_CRITICAL_SECTION* cs = (X_RTL_CRITICAL_SECTION*)SHIM_MEM_ADDR(cs_ptr);
// TODO(benvanik): get current thread ID.
uint32_t thread_id = 0;
if (xe_atomic_cas_32(-1, 0, &cs->lock_count)) {
// Able to steal the lock right away.
cs->owning_thread_id = thread_id;
cs->recursion_count = 1;
SHIM_SET_RETURN(1);
return;
} else if (cs->owning_thread_id == thread_id) {
xe_atomic_inc_32(&cs->lock_count);
++cs->recursion_count;
SHIM_SET_RETURN(1);
return;
}
SHIM_SET_RETURN(0);
}
void RtlLeaveCriticalSection_shim(
xe_ppc_state_t* ppc_state, KernelState* state) {
// VOID
// _Inout_ LPCRITICAL_SECTION lpCriticalSection
uint32_t cs_ptr = SHIM_GET_ARG_32(0);
XELOGD(XT("RtlLeaveCriticalSection(%.8X)"), cs_ptr);
X_RTL_CRITICAL_SECTION* cs = (X_RTL_CRITICAL_SECTION*)SHIM_MEM_ADDR(cs_ptr);
// Drop recursion count - if we are still not zero'ed return.
uint32_t recursion_count = --cs->recursion_count;
if (recursion_count) {
xe_atomic_dec_32(&cs->lock_count);
return;
}
// Unlock!
cs->owning_thread_id = 0;
if (xe_atomic_dec_32(&cs->lock_count) != -1) {
// There were waiters - wake one of them.
// TODO(benvanik): wake a waiter.
XELOGE(XT("RtlLeaveCriticalSection would have woken a waiter"));
}
}
}
@ -330,5 +503,11 @@ void xe::kernel::xboxkrnl::RegisterRtlExports(
SHIM_SET_MAPPING(0x0000012B, RtlImageXexHeaderField_shim, NULL);
SHIM_SET_MAPPING(0x0000012E, RtlInitializeCriticalSection_shim, NULL);
SHIM_SET_MAPPING(0x0000012F, RtlInitializeCriticalSectionAndSpinCount_shim, NULL);
SHIM_SET_MAPPING(0x00000125, RtlEnterCriticalSection_shim, NULL);
SHIM_SET_MAPPING(0x00000141, RtlTryEnterCriticalSection_shim, NULL);
SHIM_SET_MAPPING(0x00000130, RtlLeaveCriticalSection_shim, NULL);
#undef SET_MAPPING
}