Merge pull request #90 from fjtrujy/master

Adding support for `std:thread`

Dockerfile

@@ -13,9 +13,10 @@ RUN cd /src && \
     make -j $(getconf _NPROCESSORS_ONLN) && \
     make -j $(getconf _NPROCESSORS_ONLN) install
 
-## gcc needs to include libcglue libpsputility libpsprtc libpspnet_inet libpspnet_resolver libpspmodinfo libpspuser libpspkernel
+## gcc needs to include libcglue libpthreadglue libpsputility libpsprtc libpspnet_inet libpspnet_resolver libpspmodinfo libpspuser libpspkernel
 ## from pspsdk to be able to build executables, because they are part of the standard libraries
 RUN ln -sf "$PSPDEV/psp/sdk/lib/libcglue.a" "$PSPDEV/psp/lib/libcglue.a" || { exit 1; }
+RUN ln -sf "$PSPDEV/psp/sdk/lib/libpthreadglue.a" "$PSPDEV/psp/lib/libpthreadglue.a" || { exit 1; }
 RUN ln -sf "$PSPDEV/psp/sdk/lib/libpsputility.a" "$PSPDEV/psp/lib/libpsputility.a" || { exit 1; }
 RUN ln -sf "$PSPDEV/psp/sdk/lib/libpsprtc.a" "$PSPDEV/psp/lib/libpsprtc.a" || { exit 1; }
 RUN ln -sf "$PSPDEV/psp/sdk/lib/libpspnet_inet.a" "$PSPDEV/psp/lib/libpspnet_inet.a" || { exit 1; }

configure.ac

@@ -99,6 +99,7 @@ AC_CONFIG_FILES([Makefile
                  src/hprm/Makefile
                  src/kernel/Makefile
                  src/libcglue/Makefile
+                 src/libpthreadglue/Makefile
                  src/modinfo/Makefile
                  src/mp3/Makefile
                  src/mpeg/Makefile

src/Makefile.am

@@ -13,6 +13,7 @@ SUBDIRS = \
 	hprm \
 	kernel \
 	libcglue \
+	libpthreadglue \
 	modinfo \
 	mp3 \
 	mpeg \

src/base/build.mak

@@ -65,7 +65,7 @@ LDFLAGS +=  -Wl,-zmax-page-size=128
 ifeq ($(USE_KERNEL_LIBS),1)
 LIBS := -nostdlib $(LIBS) -lpspdebug -lpspdisplay_driver -lpspctrl_driver -lpspmodinfo -lpspsdk -lpspkernel
 else
-LIBS := $(LIBS) -lpspdebug -lpspdisplay -lpspge -lpspctrl -lpspsdk \
+LIBS := $(LIBS) -lpspdebug -lpspdisplay -lpspge -lpspctrl \
 		-lpspnet -lpspnet_apctl
 endif
 

src/base/build_prx.mak

@@ -33,7 +33,7 @@ LDFLAGS  := $(addprefix -L,$(LIBDIR)) -Wl,-q,-T$(PSPSDK)/lib/linkfile.prx -nosta
 ifeq ($(USE_KERNEL_LIBS),1)
 LIBS := -nostdlib $(LIBS) -lpspdebug -lpspdisplay_driver -lpspctrl_driver -lpspmodinfo -lpspsdk -lpspkernel
 else 
-LIBS := $(LIBS) -lpspdebug -lpspdisplay -lpspge -lpspctrl -lpspsdk
+LIBS := $(LIBS) -lpspdebug -lpspdisplay -lpspge -lpspctrl
 endif
 
 ifeq ($(PSP_FW_VERSION),)

src/libcglue/Makefile.am

@@ -31,7 +31,7 @@ GLUE_OBJS = __dummy_passwd.o __fill_stat.o __psp_heap_blockid.o __psp_free_heap.
 	fsync.o getpwnam.o getuid.o geteuid.o
 
 
-INIT_OBJS = __libcglue_init.o __libcglue_deinit.o _exit.o abort.o exit.o
+INIT_OBJS = __libpthreadglue_init.o __libcglue_init.o __libcglue_deinit.o _exit.o abort.o exit.o
 
 MUTEXMAN_OBJS = __malloc_mutex.o __sbrk_mutex.o __fdman_mutex.o __init_mutex.o __deinit_mutex.o
 

src/libcglue/init.c

@@ -23,11 +23,25 @@ void __init_cwd(char *argv_0);
 void __timezone_update();
 void __fdman_init();
 void __init_mutex();
+void pthread_init();
 void __psp_free_heap();
 void __deinit_mutex();
 
 extern int sce_newlib_nocreate_thread_in_start __attribute__((weak));
 
+#ifdef F___libpthreadglue_init
+/* Note: This function is being called from __libcglue_init.
+* It is a weak function because can be override by user program
+*/
+__attribute__((weak))
+void __libpthreadglue_init()
+{
+    pthread_init();
+}
+#else
+void __libpthreadglue_init();
+#endif
+
 #ifdef F___libcglue_init
 /* Note: This function is being called from crt0.c/crt0_prx.c.
 * It is a weak function because can be override by user program, 
@@ -45,6 +59,9 @@ void __libcglue_init(int argc, char *argv[])
 	/* Initialize filedescriptor management */
 	__fdman_init();
 
+	/* Initialize pthread library */
+	__libpthreadglue_init();
+
 	/* Initialize cwd from this program's path */
 	__init_cwd(argv[0]);
 

src/libcglue/sleep.c

@@ -19,6 +19,9 @@
 /* Fuctions from errno.c */
 int __set_errno(int code);
 
+/* Fuctions from errno.c */
+int __set_errno(int code);
+
 #ifdef F_nanosleep
 /* note: we don't use rem as we have no signals */
 int nanosleep(const struct timespec *req, struct timespec *rem)

src/libpthreadglue/Makefile.am

@@ -0,0 +1,42 @@
+
+libdir = @PSPSDK_LIBDIR@
+
+CC = @PSP_CC@
+CCAS = $(CC)
+AR = @PSP_AR@
+RANLIB = @PSP_RANLIB@
+
+CPPFLAGS = -I$(top_srcdir)/src/base \
+	-I$(top_srcdir)/src/debug \
+	-I$(top_srcdir)/src/kernel \
+	-I$(top_srcdir)/src/sdk \
+	-I$(top_srcdir)/src/user
+
+CFLAGS = @PSPSDK_CFLAGS@
+CCASFLAGS = $(CFLAGS)
+
+OSAL_OBJS = __threadDataKey.o __getThreadData.o __pspStubThreadEntry.o \
+	pte_osInit.o pte_osTerminate.o pte_osThreadCreate.o pte_osThreadStart.o pte_osThreadDelete.o \
+	pte_osThreadExitAndDelete.o pte_osThreadExit.o pte_osThreadWaitForEnd.o pte_osThreadGetHandle.o pte_osThreadGetPriority.o pthread_num_processors_np.o \
+	pte_osThreadSetPriority.o pte_osThreadCancel.o pte_osThreadCheckCancel.o pte_osThreadSleep.o pte_osThreadGetMinPriority.o pte_osThreadGetMaxPriority.o \
+	pte_osThreadGetDefaultPriority.o pte_osMutexCreate.o pte_osMutexDelete.o pte_osMutexLock.o pte_osMutexTimedLock.o pte_osMutexUnlock.o \
+	pte_osSemaphoreCreate.o pte_osSemaphoreDelete.o pte_osSemaphorePost.o pte_osSemaphorePend.o pte_osSemaphoreCancellablePend.o \
+	pte_osAtomicExchange.o pte_osAtomicCompareExchange.o pte_osAtomicExchangeAdd.o pte_osAtomicDecrement.o pte_osAtomicIncrement.o
+
+TLS_HELPER_OBJS = __keysUsed.o __maxTlsValues.o __globalTlsLock.o __globalTls.o pteTlsGlobalInit.o pteTlsThreadInit.o \
+	__pteTlsAlloc.o pteTlsGetValue.o __pteTlsSetValue.o __getTlsStructFromThread.o getTlsStructFromThread.o pteTlsFree.o \
+	pteTlsThreadDestroy.o pteTlsGlobalDestroy.o pte_osTlsSetValue.o pte_osTlsGetValue.o pte_osTlsAlloc.o pte_osTlsFree.o
+
+lib_LIBRARIES = libpthreadglue.a
+
+libpthreadglue_a_SOURCES = \
+	osal.c \
+	tls-helper.c
+
+libpthreadglue_a_LIBADD  = $(OSAL_OBJS) $(TLS_HELPER_OBJS)
+
+$(OSAL_OBJS): osal.c
+	$(AM_V_CC)$(COMPILE) -DF_$* $< -c -o $@
+
+$(TLS_HELPER_OBJS): tls-helper.c
+	$(AM_V_CC)$(COMPILE) -DF_$* $< -c -o $@

src/libpthreadglue/osal.c

@@ -0,0 +1,788 @@
+/*
+ * PSP Software Development Kit - https://github.com/pspdev
+ * -----------------------------------------------------------------------
+ * Licensed under the BSD license, see LICENSE in PSPSDK root for details.
+ *
+ * osal.c - Pthread compatible system calls.
+ *
+ * Copyright (c) 2021 Francisco J Trujillo <fjtrujy@gmail.com>
+ *
+ */
+
+#include <pthread.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <pspkerror.h>
+#include <pspthreadman.h>
+#include <pspsdk.h>
+
+
+typedef int pte_osThreadHandle;
+typedef int pte_osSemaphoreHandle;
+typedef int pte_osMutexHandle;
+#include <sys/pte_generic_osal.h>
+
+#define MAX_PSP_UID 2048 // SWAG
+#define DEFAULT_STACK_SIZE_BYTES 4096
+#define PSP_MAX_TLS 32
+
+pte_osResult pteTlsGlobalInit(int maxEntries);
+void * pteTlsThreadInit(void);
+
+pte_osResult __pteTlsAlloc(unsigned int *pKey);
+void * pteTlsGetValue(void *pTlsThreadStruct, unsigned int index);
+pte_osResult __pteTlsSetValue(void *pTlsThreadStruct, unsigned int index, void * value);
+void *__getTlsStructFromThread(SceUID thid);
+pte_osResult pteTlsFree(unsigned int index);
+
+void pteTlsThreadDestroy(void * pTlsThreadStruct);
+void pteTlsGlobalDestroy(void);
+
+#if 0
+#define PSP_DEBUG(x) printf(x)
+#else
+#define PSP_DEBUG(x)
+#endif
+
+#define POLLING_DELAY_IN_us	100
+
+/* TLS key used to access pspThreadData struct for reach thread. */
+#ifdef F___threadDataKey
+unsigned int __threadDataKey;
+#else
+extern unsigned int __threadDataKey;
+#endif
+
+extern void *__globalTls;
+
+/*
+ * Data stored on a per-thread basis - allocated in pte_osThreadCreate
+ * and freed in pte_osThreadDelete.
+ */
+typedef struct pspThreadData
+  {
+    /* Entry point and parameters to thread's main function */
+    pte_osThreadEntryPoint entryPoint;
+    void * argv;
+
+    /* Semaphore used for cancellation.  Posted to by pte_osThreadCancel, 
+       polled in pte_osSemaphoreCancellablePend */
+    SceUID cancelSem;
+
+  } pspThreadData;
+
+
+/****************************************************************************
+ *
+ * Helper functions
+ *
+ ***************************************************************************/
+#ifdef F___getThreadData
+pspThreadData *__getThreadData(SceUID threadHandle)
+{
+  pspThreadData *pThreadData;
+  void *pTls;
+
+  pTls = __getTlsStructFromThread(threadHandle);
+  pThreadData = (pspThreadData *) pteTlsGetValue(pTls, __threadDataKey);
+
+  return pThreadData;
+}
+#else
+pspThreadData *__getThreadData(SceUID threadHandle);
+#endif
+
+/* A new thread's stub entry point.  It retrieves the real entry point from the per thread control
+ * data as well as any parameters to this function, and then calls the entry point.
+ */
+#ifdef F___pspStubThreadEntry
+int __pspStubThreadEntry(unsigned int argc, void *argv)
+{
+  int result;
+  pspThreadData *pThreadData;
+
+  pThreadData = __getThreadData(sceKernelGetThreadId());
+  result = (*(pThreadData->entryPoint))(pThreadData->argv);
+
+  return result;
+}
+#else 
+extern int __pspStubThreadEntry(unsigned int argc, void *argv);
+#endif
+
+/****************************************************************************
+ *
+ * Initialization
+ *
+ ***************************************************************************/
+#ifdef F_pte_osInit
+pte_osResult pte_osInit(void)
+{
+  pte_osResult result;
+  pspThreadData *pThreadData;
+  char cancelSemName[64];
+
+  /* Allocate and initialize TLS support */
+  result = pteTlsGlobalInit(PSP_MAX_TLS);
+
+  if (result == PTE_OS_OK) {
+    /* Allocate a key that we use to store control information (e.g. cancellation semaphore) per thread */
+    result = __pteTlsAlloc(&__threadDataKey);
+
+    if (result == PTE_OS_OK) {
+      /* Initialize the structure used to emulate TLS for 
+        * non-POSIX threads 
+        */
+      __globalTls = pteTlsThreadInit();
+
+      /* Also create a "thread data" structure for a single non-POSIX thread. */
+
+      /* Allocate some memory for our per-thread control data.  We use this for:
+        * 1. Entry point and parameters for the user thread's main function.
+        * 2. Semaphore used for thread cancellation.
+        */
+      pThreadData = (pspThreadData *) malloc(sizeof(pspThreadData));
+
+      if (pThreadData == NULL) {
+        result = PTE_OS_NO_RESOURCES;
+      } else {
+        /* Save a pointer to our per-thread control data as a TLS value */
+        __pteTlsSetValue(__globalTls, __threadDataKey, pThreadData);
+
+        /* Create a semaphore used to cancel threads */
+        snprintf(cancelSemName, sizeof(cancelSemName), "pthread_cancelSemGlobal");
+
+        pThreadData->cancelSem = sceKernelCreateSema(cancelSemName,
+                              0,          /* attributes (default) */
+                              0,          /* initial value        */
+                              255,        /* maximum value        */
+                              0);         /* options (default)    */
+        result = PTE_OS_OK;
+      }
+    }
+  }
+
+  return result;
+}
+#endif
+
+#ifdef F_pte_osTerminate
+pte_osResult pte_osTerminate(void) {
+  pteTlsGlobalDestroy();
+  return PTE_OS_OK;
+}
+#endif
+
+/****************************************************************************
+ *
+ * Threads
+ *
+ ***************************************************************************/
+#ifdef F_pte_osThreadCreate
+pte_osResult pte_osThreadCreate(pte_osThreadEntryPoint entryPoint,
+                                int stackSize,
+                                int initialPriority,
+                                void *argv,
+                                pte_osThreadHandle* ppte_osThreadHandle)
+{
+  char threadName[64];
+  char cancelSemName[64];
+  static int threadNum = 1;
+  int pspAttr;
+  void *pTls;
+  SceUID threadId;
+  pte_osResult result;
+  pspThreadData *pThreadData;
+
+  if (threadNum++ > MAX_PSP_UID) {
+    threadNum = 0;
+  }
+
+  /* Make sure that the stack we're going to allocate is big enough */
+  if (stackSize < DEFAULT_STACK_SIZE_BYTES) {
+    stackSize = DEFAULT_STACK_SIZE_BYTES;
+  }
+
+  /* Allocate TLS structure for this thread. */
+  pTls = pteTlsThreadInit();
+  if (pTls == NULL) {
+    PSP_DEBUG("pteTlsThreadInit: PTE_OS_NO_RESOURCES\n");
+    result = PTE_OS_NO_RESOURCES;
+    goto FAIL0;
+  }
+
+  /* Allocate some memory for our per-thread control data.  We use this for:
+   * 1. Entry point and parameters for the user thread's main function.
+   * 2. Semaphore used for thread cancellation.
+   */
+  pThreadData = (pspThreadData *) malloc(sizeof(pspThreadData));
+
+  if (pThreadData == NULL) {
+    pteTlsThreadDestroy(pTls);
+
+    PSP_DEBUG("malloc(pspThreadData): PTE_OS_NO_RESOURCES\n");
+    result = PTE_OS_NO_RESOURCES;
+    goto FAIL0;
+  }
+
+  /* Save a pointer to our per-thread control data as a TLS value */
+  __pteTlsSetValue(pTls, __threadDataKey, pThreadData);
+
+  pThreadData->entryPoint = entryPoint;
+  pThreadData->argv = argv;
+
+  /* Create a semaphore used to cancel threads */
+  snprintf(cancelSemName, sizeof(cancelSemName), "pthread_cancelSem%04d", threadNum);
+
+  pThreadData->cancelSem = sceKernelCreateSema(cancelSemName,
+                           0,          /* attributes (default) */
+                           0,          /* initial value        */
+                           255,        /* maximum value        */
+                           0);         /* options (default)    */
+
+
+  /* In order to emulate TLS functionality, we append the address of the TLS structure that we
+   * allocated above to the thread's name.  To set or get TLS values for this thread, the user
+   * needs to get the name of the thread from the OS and then parse the name to extract
+   * a pointer to the TLS structure.
+   */
+  snprintf(threadName, sizeof(threadName), "pthread%04d__%x", threadNum, (unsigned int) pTls);
+
+  pspAttr = 0;
+
+  //  printf("%s %p %d %d %d\n",threadName, __pspStubThreadEntry, initialPriority, stackSize, pspAttr);
+  threadId = sceKernelCreateThread(threadName,
+                                   __pspStubThreadEntry,
+                                   initialPriority,
+                                   stackSize,
+                                   pspAttr,
+                                   NULL);
+
+  if (threadId == (SceUID) SCE_KERNEL_ERROR_NO_MEMORY) {
+    free(pThreadData);
+    pteTlsThreadDestroy(pTls);
+
+    PSP_DEBUG("sceKernelCreateThread: PTE_OS_NO_RESOURCES\n");
+    result =  PTE_OS_NO_RESOURCES;
+  } else if (threadId < 0) {
+    free(pThreadData);
+    pteTlsThreadDestroy(pTls);
+
+    PSP_DEBUG("sceKernelCreateThread: PTE_OS_GENERAL_FAILURE\n");
+    result = PTE_OS_GENERAL_FAILURE;
+  } else {
+    *ppte_osThreadHandle = threadId;
+    result = PTE_OS_OK;
+  }
+
+FAIL0:
+  return result;
+}
+#endif
+
+#ifdef F_pte_osThreadStart
+pte_osResult pte_osThreadStart(pte_osThreadHandle osThreadHandle)
+{
+  sceKernelStartThread(osThreadHandle, 0, 0);
+
+  return PTE_OS_OK;
+}
+#endif
+
+#ifdef F_pte_osThreadDelete
+pte_osResult pte_osThreadDelete(pte_osThreadHandle handle)
+{
+  pspThreadData *pThreadData;
+  void *pTls;
+
+  pTls = __getTlsStructFromThread(handle);
+  pThreadData = __getThreadData(handle);
+  sceKernelDeleteSema(pThreadData->cancelSem);  
+  free(pThreadData);
+  pteTlsThreadDestroy(pTls);
+  sceKernelDeleteThread(handle);
+
+  return PTE_OS_OK;
+}
+#endif
+
+#ifdef F_pte_osThreadExitAndDelete
+pte_osResult pte_osThreadExitAndDelete(pte_osThreadHandle handle)
+{
+  pte_osThreadDelete(handle);
+  sceKernelExitDeleteThread(0);
+
+  return PTE_OS_OK;
+}
+#endif
+
+#ifdef F_pte_osThreadExit
+void pte_osThreadExit()
+{
+  sceKernelExitThread(0);
+}
+#endif
+
+/*
+ * This has to be cancellable, so we can't just call sceKernelWaitThreadEnd.
+ * Instead, poll on this in a loop, like we do for a cancellable semaphore.
+ */
+#ifdef F_pte_osThreadWaitForEnd
+pte_osResult pte_osThreadWaitForEnd(pte_osThreadHandle threadHandle)
+{
+  pte_osResult result;
+  pspThreadData *pThreadData;
+
+  pThreadData = __getThreadData(sceKernelGetThreadId());
+
+  while (1) {
+    SceKernelThreadRunStatus info;
+
+    /* Poll task to see if it has ended */
+    memset(&info,0,sizeof(info));
+    info.size = sizeof(info);
+    sceKernelReferThreadRunStatus(threadHandle, &info);      
+
+    if (info.status == PSP_THREAD_STOPPED) {
+      /* Thread has ended */
+      result = PTE_OS_OK;
+      break;
+    } else {
+      SceKernelSemaInfo semInfo;
+
+      if (pThreadData != NULL) {
+        SceUID osResult;
+
+        osResult = sceKernelReferSemaStatus(pThreadData->cancelSem, &semInfo);
+        if (osResult == SCE_KERNEL_ERROR_OK) {
+          if (semInfo.currentCount > 0) {
+            result = PTE_OS_INTERRUPTED;
+            break;
+          } else {
+            /* Nothing found and not timed out yet; let's yield so we're not
+              * in busy loop.
+              */
+            sceKernelDelayThread(POLLING_DELAY_IN_us);
+          }
+        } else {
+          result = PTE_OS_GENERAL_FAILURE;
+          break;
+        }
+      }
+    }
+  }
+
+  return result;
+}
+#endif
+
+#ifdef F_pte_osThreadGetHandle
+pte_osThreadHandle pte_osThreadGetHandle(void)
+{
+  return sceKernelGetThreadId();
+}
+#endif
+
+#ifdef F_pte_osThreadGetPriority
+int pte_osThreadGetPriority(pte_osThreadHandle threadHandle)
+{
+  SceKernelThreadInfo thinfo;
+
+  thinfo.size = sizeof(SceKernelThreadInfo);
+  sceKernelReferThreadStatus(threadHandle, &thinfo);
+
+  return thinfo.currentPriority;
+}
+#endif
+
+#ifdef F_pte_osThreadSetPriority
+pte_osResult pte_osThreadSetPriority(pte_osThreadHandle threadHandle, int newPriority)
+{
+  sceKernelChangeThreadPriority(threadHandle, newPriority);
+  return PTE_OS_OK;
+}
+#endif
+
+#ifdef F_pte_osThreadCancel
+pte_osResult pte_osThreadCancel(pte_osThreadHandle threadHandle)
+{
+  SceUID osResult;
+  pte_osResult result;
+  pspThreadData *pThreadData;
+
+  pThreadData = __getThreadData(threadHandle);
+  osResult = sceKernelSignalSema(pThreadData->cancelSem, 1);
+
+  if (osResult == SCE_KERNEL_ERROR_OK) {
+    result = PTE_OS_OK;
+  } else {
+    result = PTE_OS_GENERAL_FAILURE;
+  }
+
+  return result;
+}
+#endif
+
+#ifdef F_pte_osThreadCheckCancel
+pte_osResult pte_osThreadCheckCancel(pte_osThreadHandle threadHandle)
+{
+  pspThreadData *pThreadData;
+  SceKernelSemaInfo semInfo;
+  SceUID osResult;
+  pte_osResult result;
+
+  pThreadData = __getThreadData(threadHandle);
+  if (pThreadData != NULL) {
+    osResult = sceKernelReferSemaStatus(pThreadData->cancelSem, &semInfo);
+
+    if (osResult == SCE_KERNEL_ERROR_OK) {
+      if (semInfo.currentCount > 0) {
+        result = PTE_OS_INTERRUPTED;
+      } else {
+        result = PTE_OS_OK;
+      }
+    } else {
+      /* sceKernelReferSemaStatus returned an error */
+      result = PTE_OS_GENERAL_FAILURE;
+    }
+  } else {
+    /* For some reason, we couldn't get thread data */
+    result = PTE_OS_GENERAL_FAILURE;
+  }
+
+  return result;
+}
+#endif
+
+#ifdef F_pte_osThreadSleep
+void pte_osThreadSleep(unsigned int msecs)
+{
+  sceKernelDelayThread(msecs*1000);
+}
+#endif
+
+#ifdef  F_pte_osThreadGetMinPriority
+int pte_osThreadGetMinPriority()
+{
+  return 17;
+}
+#endif
+
+#ifdef F_pte_osThreadGetMaxPriority
+int pte_osThreadGetMaxPriority()
+{
+  return 30;
+}
+#endif
+
+#ifdef F_pte_osThreadGetDefaultPriority
+int pte_osThreadGetDefaultPriority()
+{
+  return 18;
+}
+#endif
+
+#ifdef F_pthread_num_processors_np
+int pthread_num_processors_np(void)
+{
+  return 1;
+}
+#endif
+
+/****************************************************************************
+ *
+ * Mutexes
+ *
+ ****************************************************************************/
+#ifdef F_pte_osMutexCreate
+pte_osResult pte_osMutexCreate(pte_osMutexHandle *pHandle)
+{
+  static int mutexCtr = 0;
+  char mutexName[32];
+  pte_osMutexHandle handle;
+
+  if (mutexCtr++ > MAX_PSP_UID) {
+    mutexCtr = 0;
+  }
+
+  snprintf(mutexName,sizeof(mutexName),"mutex%d",mutexCtr);
+  handle = sceKernelCreateSema(mutexName,
+                               0,          /* attributes (default) */
+                               1,          /* initial value        */
+                               1,          /* maximum value        */
+                               0);         /* options (default)    */
+
+  *pHandle = handle;
+  return PTE_OS_OK;
+}
+#endif
+
+#ifdef F_pte_osMutexDelete
+pte_osResult pte_osMutexDelete(pte_osMutexHandle handle)
+{
+  sceKernelDeleteSema(handle);
+
+  return PTE_OS_OK;
+}
+#endif
+
+#ifdef F_pte_osMutexLock
+pte_osResult pte_osMutexLock(pte_osMutexHandle handle)
+{
+  sceKernelWaitSema(handle, 1, NULL);
+
+  return PTE_OS_OK;
+}
+#endif
+
+#ifdef F_pte_osMutexTimedLock
+pte_osResult pte_osMutexTimedLock(pte_osMutexHandle handle, unsigned int timeoutMsecs)
+{
+  pte_osResult result;
+  SceUInt timeoutUsecs = timeoutMsecs*1000;
+
+  int status = sceKernelWaitSema(handle, 1, &timeoutUsecs);
+  if (status < 0) {
+    // Assume that any error from sceKernelWaitSema was due to a timeout
+    result = PTE_OS_TIMEOUT;
+  } else {
+    result = PTE_OS_OK;
+  }
+
+  return result;
+}
+#endif
+
+#ifdef F_pte_osMutexUnlock
+pte_osResult pte_osMutexUnlock(pte_osMutexHandle handle)
+{
+  sceKernelSignalSema(handle, 1);
+  return PTE_OS_OK;
+}
+#endif
+
+/****************************************************************************
+ *
+ * Semaphores
+ *
+ ***************************************************************************/
+#ifdef F_pte_osSemaphoreCreate
+pte_osResult pte_osSemaphoreCreate(int initialValue, pte_osSemaphoreHandle *pHandle)
+{
+  pte_osSemaphoreHandle handle;
+  static int semCtr = 0;
+  char semName[32];
+
+  if (semCtr++ > MAX_PSP_UID) {
+    semCtr = 0;
+  }
+
+  snprintf(semName,sizeof(semName),"pthread_sem%d",semCtr);
+  handle = sceKernelCreateSema(semName,
+                               0,              /* attributes (default) */
+                               initialValue,   /* initial value        */
+                               SEM_VALUE_MAX,  /* maximum value        */
+                               0);             /* options (default)    */
+
+  *pHandle = handle;
+  return PTE_OS_OK;
+}
+#endif
+
+#ifdef F_pte_osSemaphoreDelete
+pte_osResult pte_osSemaphoreDelete(pte_osSemaphoreHandle handle)
+{
+  sceKernelDeleteSema(handle);
+  return PTE_OS_OK;
+}
+#endif
+
+#ifdef F_pte_osSemaphorePost
+pte_osResult pte_osSemaphorePost(pte_osSemaphoreHandle handle, int count)
+{
+  sceKernelSignalSema(handle, count);
+  return PTE_OS_OK;
+}
+#endif
+
+#ifdef F_pte_osSemaphorePend
+pte_osResult pte_osSemaphorePend(pte_osSemaphoreHandle handle, unsigned int *pTimeoutMsecs)
+{
+  unsigned int timeoutUsecs;
+  unsigned int *pTimeoutUsecs;
+  SceUInt result;
+  pte_osResult osResult;
+
+  if (pTimeoutMsecs == NULL) {
+    pTimeoutUsecs = NULL;
+  } else {
+    timeoutUsecs = *pTimeoutMsecs * 1000;
+    pTimeoutUsecs = &timeoutUsecs;
+  }
+
+  result = sceKernelWaitSema(handle, 1, pTimeoutUsecs);
+  if (result == SCE_KERNEL_ERROR_OK) {
+    osResult = PTE_OS_OK;
+  } else if (result == SCE_KERNEL_ERROR_WAIT_TIMEOUT) {
+    osResult = PTE_OS_TIMEOUT;
+  } else {
+    osResult = PTE_OS_GENERAL_FAILURE;
+  }
+
+  return osResult;
+}
+#endif
+
+
+/*
+ * Pend on a semaphore- and allow the pend to be cancelled.
+ *
+ * PSP OS provides no functionality to asynchronously interrupt a blocked call.  We simulte
+ * this by polling on the main semaphore and the cancellation semaphore and sleeping in a loop.
+ */
+#ifdef F_pte_osSemaphoreCancellablePend
+pte_osResult pte_osSemaphoreCancellablePend(pte_osSemaphoreHandle semHandle, unsigned int *pTimeout)
+{
+  pspThreadData *pThreadData;
+
+  pThreadData = __getThreadData(sceKernelGetThreadId());
+
+  clock_t start_time;
+  pte_osResult result =  PTE_OS_OK;
+  unsigned int timeout;
+  unsigned char timeoutEnabled;
+
+  start_time = clock();
+
+  // clock() is in microseconds, timeout as passed in was in milliseconds
+  if (pTimeout == NULL) {
+    timeout = 0;
+    timeoutEnabled = 0;
+  } else {
+    timeout = *pTimeout * 1000;
+    timeoutEnabled = 1;
+  }
+
+  while (1) {
+    SceUInt semTimeout;
+    int status;
+
+    /* Poll semaphore */
+    semTimeout = 0;
+    status = sceKernelWaitSema(semHandle, 1, &semTimeout);
+
+    if (status == SCE_KERNEL_ERROR_OK) {
+      /* User semaphore posted to */
+      result = PTE_OS_OK;
+      break;
+    } else if ((timeoutEnabled) && ((clock() - start_time) > timeout)) {
+      /* Timeout expired */
+      result = PTE_OS_TIMEOUT;
+      break;
+    } else {
+      SceKernelSemaInfo semInfo;
+
+      if (pThreadData != NULL) {
+        SceUID osResult;
+
+        osResult = sceKernelReferSemaStatus(pThreadData->cancelSem, &semInfo);
+        if (osResult == SCE_KERNEL_ERROR_OK) {
+          if (semInfo.currentCount > 0) {
+              result = PTE_OS_INTERRUPTED;
+              break;
+          } else {
+            /* Nothing found and not timed out yet; let's yield so we're not
+              * in busy loop.
+              */
+            sceKernelDelayThread(POLLING_DELAY_IN_us);
+          }
+        } else {
+          result = PTE_OS_GENERAL_FAILURE;
+          break;
+        }
+      }
+    }
+  }
+
+  return result;
+}
+#endif
+
+
+/****************************************************************************
+ *
+ * Atomic Operations
+ *
+ ***************************************************************************/
+#ifdef F_pte_osAtomicExchange
+int pte_osAtomicExchange(int *ptarg, int val)
+{
+  int intc = pspSdkDisableInterrupts();
+  int origVal;
+
+  origVal = *ptarg;
+  *ptarg = val;
+
+  pspSdkEnableInterrupts(intc);
+  return origVal;
+}
+#endif
+
+#ifdef F_pte_osAtomicCompareExchange
+int pte_osAtomicCompareExchange(int *pdest, int exchange, int comp)
+{
+  int intc = pspSdkDisableInterrupts();
+  int origVal;
+
+  origVal = *pdest;
+  if (*pdest == comp){
+    *pdest = exchange;
+  }
+
+  pspSdkEnableInterrupts(intc);
+  return origVal;
+}
+#endif
+
+#ifdef F_pte_osAtomicExchangeAdd
+int pte_osAtomicExchangeAdd(int volatile* pAddend, int value)
+{
+  int origVal;
+  int intc = pspSdkDisableInterrupts();
+
+  origVal = *pAddend;
+  *pAddend += value;
+
+  pspSdkEnableInterrupts(intc);
+  return origVal;
+}
+#endif
+
+#ifdef F_pte_osAtomicDecrement
+int pte_osAtomicDecrement(int *pdest)
+{
+  int val;
+  int intc = pspSdkDisableInterrupts();
+
+  (*pdest)--;
+  val = *pdest;
+
+  pspSdkEnableInterrupts(intc);
+  return val;
+}
+#endif
+
+#ifdef F_pte_osAtomicIncrement
+int pte_osAtomicIncrement(int *pdest)
+{
+  int val;
+  int intc = pspSdkDisableInterrupts();
+
+  (*pdest)++;
+  val = *pdest;
+
+  pspSdkEnableInterrupts(intc);
+  return val;
+}
+#endif

src/libpthreadglue/tls-helper.c

@@ -0,0 +1,254 @@
+/*
+ * PSP Software Development Kit - https://github.com/pspdev
+ * -----------------------------------------------------------------------
+ * Licensed under the BSD license, see LICENSE in PSPSDK root for details.
+ *
+ * tls-helper.c - Pthread compatible system calls.
+ *
+ * Copyright (c) 2021 Francisco J Trujillo <fjtrujy@gmail.com>
+ *
+ */
+
+#include <stdio.h>
+#include <stdlib.h>
+
+#include <pspkerneltypes.h>
+#include <pspthreadman.h>
+
+typedef int pte_osThreadHandle;
+typedef int pte_osSemaphoreHandle;
+typedef int pte_osMutexHandle;
+#include <sys/pte_generic_osal.h>
+
+#ifdef F___keysUsed
+int *__keysUsed;
+#else
+extern int *__keysUsed;
+#endif
+
+#ifdef F___maxTlsValues
+int __maxTlsValues;
+#else
+extern int __maxTlsValues;
+#endif
+
+#ifdef F___globalTlsLock
+pte_osMutexHandle __globalTlsLock;
+#else
+extern pte_osMutexHandle __globalTlsLock;
+#endif
+
+/* Structure used to emulate TLS on non-POSIX threads.  
+ * This limits us to one non-POSIX thread that can
+ * call pthread functions. */
+#ifdef F___globalTls
+void *__globalTls;
+#else
+extern void *__globalTls;
+#endif
+
+#ifdef F_pteTlsGlobalInit
+pte_osResult pteTlsGlobalInit(int maxEntries)
+{
+    int i;
+    pte_osResult result;
+
+    pte_osMutexCreate(&__globalTlsLock);
+    __keysUsed = (int *)malloc(maxEntries * sizeof(int));
+
+    if (__keysUsed != NULL) {
+        for (i = 0; i < maxEntries; i++) {
+            __keysUsed[i] = 0;
+        }
+
+        __maxTlsValues = maxEntries;
+        result = PTE_OS_OK;
+    } else {
+        result = PTE_OS_NO_RESOURCES;
+    }
+
+    return result;
+}
+#endif
+
+#ifdef F_pteTlsThreadInit
+void *pteTlsThreadInit(void)
+{
+    void **pTlsStruct;
+    int i;
+
+    pTlsStruct = (void **)malloc(__maxTlsValues * sizeof(void *));
+
+    // PTE library assumes that keys are initialized to zero
+    for (i = 0; i < __maxTlsValues; i++) {
+        pTlsStruct[i] = 0;
+    }
+
+    return (void *)pTlsStruct;
+}
+#endif
+
+#ifdef F___pteTlsAlloc
+pte_osResult __pteTlsAlloc(unsigned int *pKey)
+{
+    int i;
+    pte_osResult result = PTE_OS_NO_RESOURCES;
+
+    pte_osMutexLock(__globalTlsLock);
+
+    for (i = 0; i < __maxTlsValues; i++) {
+        if (__keysUsed[i] == 0) {
+            __keysUsed[i] = 1;
+
+            *pKey = i + 1;
+            result = PTE_OS_OK;
+            break;
+        }
+    }
+
+    pte_osMutexUnlock(__globalTlsLock);
+    return result;
+}
+#else
+pte_osResult __pteTlsAlloc(unsigned int *pKey);
+#endif
+
+#ifdef F_pteTlsGetValue
+void *pteTlsGetValue(void *pTlsThreadStruct, unsigned int index)
+{
+    void **pTls = (void **)pTlsThreadStruct;
+
+    if (__keysUsed[index - 1] && pTls != NULL) {
+        return pTls[index - 1];
+    } else {
+        return NULL;
+    }
+}
+#else
+void *pteTlsGetValue(void *pTlsThreadStruct, unsigned int index);
+#endif
+
+#ifdef F___pteTlsSetValue
+pte_osResult __pteTlsSetValue(void *pTlsThreadStruct, unsigned int index, void *value)
+{
+    pte_osResult result;
+    void **pTls = (void **)pTlsThreadStruct;
+
+    if (pTls != NULL) {
+        pTls[index - 1] = value;
+        result = PTE_OS_OK;
+    } else {
+        result = PTE_OS_INVALID_PARAM;
+    }
+
+    return result;
+}
+#else
+pte_osResult __pteTlsSetValue(void *pTlsThreadStruct, unsigned int index, void *value);
+#endif
+
+#ifdef F___getTlsStructFromThread
+void *__getTlsStructFromThread(SceUID thid)
+{
+  SceKernelThreadInfo thinfo;
+  unsigned int ptr;
+  unsigned int thrNum;
+  void *pTls;
+  int numMatches;
+
+
+  thinfo.size = sizeof(SceKernelThreadInfo);
+  sceKernelReferThreadStatus(thid, &thinfo);
+  numMatches = sscanf(thinfo.name,"pthread%04d__%x", &thrNum, &ptr);
+
+  /* If we were called from a pthread, use the TLS allocated when the thread
+   * was created.  Otherwise, we were called from a non-pthread, so use the
+   * "global".  This is a pretty bad hack, but necessary due to lack of TLS on PSP.
+   */
+  if (numMatches == 2) {
+    pTls = (void *) ptr;
+  } else {
+    pTls = __globalTls;
+  }
+
+  return pTls;
+}
+#else
+void *__getTlsStructFromThread(SceUID thid);
+#endif
+
+#ifdef F_pteTlsFree
+pte_osResult pteTlsFree(unsigned int index)
+{
+    pte_osResult result;
+
+    if (__keysUsed != NULL) {
+        pte_osMutexLock(__globalTlsLock);
+        __keysUsed[index - 1] = 0;
+        pte_osMutexUnlock(__globalTlsLock);
+
+        result = PTE_OS_OK;
+    } else {
+        result = PTE_OS_GENERAL_FAILURE;
+    }
+
+    return result;
+}
+#else
+pte_osResult pteTlsFree(unsigned int index);
+#endif
+
+#ifdef F_pteTlsThreadDestroy
+void pteTlsThreadDestroy(void *pTlsThreadStruct)
+{
+    free(pTlsThreadStruct);
+}
+#endif
+
+#ifdef F_pteTlsGlobalDestroy
+void pteTlsGlobalDestroy(void)
+{
+    pte_osMutexDelete(__globalTlsLock);
+    free(__keysUsed);
+}
+#endif
+
+#ifdef F_pte_osTlsSetValue
+pte_osResult pte_osTlsSetValue(unsigned int key, void * value)
+{
+  void *pTls;
+
+  pTls = __getTlsStructFromThread(sceKernelGetThreadId());
+
+  return __pteTlsSetValue(pTls, key, value);
+}
+#endif
+
+#ifdef F_pte_osTlsGetValue
+void * pte_osTlsGetValue(unsigned int index)
+{
+  void *pTls;
+
+  pTls = __getTlsStructFromThread(sceKernelGetThreadId());
+
+  return (void *) pteTlsGetValue(pTls, index);
+
+}
+#endif
+
+#ifdef F_pte_osTlsAlloc
+pte_osResult pte_osTlsAlloc(unsigned int *pKey)
+{
+  __getTlsStructFromThread(sceKernelGetThreadId());
+
+  return __pteTlsAlloc(pKey);
+
+}
+#endif
+
+#ifdef F_pte_osTlsFree
+pte_osResult pte_osTlsFree(unsigned int index)
+{
+  return pteTlsFree(index);
+}
+#endif

src/user/pspmoduleinfo.h

@@ -107,4 +107,8 @@ enum PspModuleInfoAttr
 	void __libcglue_init(int argc, char *argv[]) {} \
 	void __libcglue_deinit() {}
 
+/* Disable the auto start of pthread on init for reducing binary size if not used. */
+#define PSP_DISABLE_AUTOSTART_PTHREAD() \
+	void __libpthreadglue_init() {}
+
 #endif /* PSPMODULEINFO_H */