0001-several-improvement-on-pthread-but-cause-performance.patch

From e295ffb109e34e24bbf97c179457ad560c26a31b Mon Sep 17 00:00:00 2001
From: MJChku <mjc97cooper@gmail.com>
Date: Tue, 27 Oct 2020 07:23:54 -0500
Subject: [PATCH] several improvement on pthread but cause performance
 degredation

---
 src/arch/x64/asm/boot.S                  | 22 +-------
 src/compat/pthread/Makefile              |  5 +-
 src/compat/pthread/create.c              |  8 +++
 src/compat/pthread/implement.h           |  2 +
 src/compat/pthread/nk/nk_compatibility.h | 72 ++++++++++++++++++++++++
 src/compat/pthread/nk/nk_osal.c          | 28 ++++++---
 src/compat/pthread/nk/pte_types.h        |  1 +
 src/compat/pthread/nk_get_pte_thread_t.c | 10 ++++
 src/compat/pthread/pte_detach.c          |  8 ++-
 src/compat/pthread/pte_threadStart.c     |  1 +
 src/compat/pthread/pte_throw.c           |  5 +-
 src/compat/pthread/pthread_exit.c        |  4 +-
 src/compat/pthread/pthread_self.c        | 22 +++++++-
 src/compat/pthread/pthread_setspecific.c |  5 +-
 14 files changed, 156 insertions(+), 37 deletions(-)
 create mode 100644 src/compat/pthread/nk/nk_compatibility.h
 create mode 100644 src/compat/pthread/nk_get_pte_thread_t.c

diff --git a/src/arch/x64/asm/boot.S b/src/arch/x64/asm/boot.S
index 260680f..8cb99f2 100644
--- a/src/arch/x64/asm/boot.S
+++ b/src/arch/x64/asm/boot.S
@@ -39,7 +39,6 @@ multiboot_hdr:
     /* BEGIN TAGS */
 
     /* sections */
- 
     .word 2, 0
     .long 24
     .long multiboot_hdr
@@ -123,31 +122,16 @@ paging_longmode_setup:
     orl $0x3, %eax
     movl $pdpt, %ebx
     movl %eax, (%ebx)
-   
-    //set first 4 entry of PDPT 4GB total
-    movl $4, %ecx
-    movl $pd, %edx
-    movl $0x83, %eax // set PS bit also (PDE -> 2MB page)
-
-.write_pdpe:
-    push %ecx
 
     // Identity map the first GB
     movl $512, %ecx
+    movl $pd, %edx
+    movl $0x83, %eax // set PS bit also (PDE -> 2MB page)
 .write_pde:
     movl %eax, (%edx)
     addl $0x200000, %eax
     addl $0x8, %edx
-    loop .write_pde 
-    
-    push %edx
-    orl $0x3, %edx
-    addl $0x8, %ebx  //next entry of pdpt
-    movl %edx, (%ebx) 
-    pop %edx
-    pop %ecx
-    loop .write_pdpe
-
+    loop .write_pde
 
     /* put pml4 address in cr3 */
     movl $pml4, %eax
diff --git a/src/compat/pthread/Makefile b/src/compat/pthread/Makefile
index 20ea250..11729f9 100644
--- a/src/compat/pthread/Makefile
+++ b/src/compat/pthread/Makefile
@@ -1,6 +1,9 @@
 NK_LIB =\
 	nk/
 
+NK_OBJS = \
+	  nk_get_pte_thread_t.o
+
 MUTEX_OBJS = \
   pthread_mutex_unlock.o \
   pthread_mutex_init.o \
@@ -152,7 +155,7 @@ CANCEL_OBJS = \
 
 
 
-OBJS = $(MUTEX_OBJS) $(MUTEXATTR_OBJS) $(THREAD_OBJS) $(SUPPORT_OBJS) $(TLS_OBJS) $(MISC_OBJS) $(SEM_OBJS) $(BARRIER_OBJS) $(SPIN_OBJS) $(CONDVAR_OBJS) $(RWLOCK_OBJS) $(CANCEL_OBJS)
+OBJS = $(MUTEX_OBJS) $(MUTEXATTR_OBJS) $(THREAD_OBJS) $(SUPPORT_OBJS) $(TLS_OBJS) $(MISC_OBJS) $(SEM_OBJS) $(BARRIER_OBJS) $(SPIN_OBJS) $(CONDVAR_OBJS) $(RWLOCK_OBJS) $(CANCEL_OBJS) $(NK_OBJS)
 obj-y +=$(OBJS) \
 	$(NK_LIB) \
 
diff --git a/src/compat/pthread/create.c b/src/compat/pthread/create.c
index 5a49458..4aded73 100644
--- a/src/compat/pthread/create.c
+++ b/src/compat/pthread/create.c
@@ -214,6 +214,14 @@ pthread_create (pthread_t * tid,
                                 parms,
                                 &(tp->threadId));
 
+  //mjc why this? ease pthread_self we don't need getspecific(pteself) anymore
+  nk_thread_t *threadptr = (nk_thread_t*) tp->threadId;
+  //plus 8 means add 8 bytes
+  pte_thread_t ** ptr = (pte_thread_t **) (threadptr->hwtls+8);
+  *ptr = tp; 
+  //mjc set this, seems to be a bug from pte repo
+  tp->ptHandle = thread;
+
   if (osResult == PTE_OS_OK)
     {
       pte_osThreadStart(tp->threadId);
diff --git a/src/compat/pthread/implement.h b/src/compat/pthread/implement.h
index b54e302..29a6476 100644
--- a/src/compat/pthread/implement.h
+++ b/src/compat/pthread/implement.h
@@ -479,6 +479,8 @@ extern "C"
     
     //mjc add out
     int pte_threadStart (void *vthreadParms, void ** out);
+    //mjc add
+    pte_thread_t * nk_get_pte_thread_t(void);
 
     void pte_callUserDestroyRoutines (pthread_t thread);
 
diff --git a/src/compat/pthread/nk/nk_compatibility.h b/src/compat/pthread/nk/nk_compatibility.h
new file mode 100644
index 0000000..cdd4401
--- /dev/null
+++ b/src/compat/pthread/nk/nk_compatibility.h
@@ -0,0 +1,72 @@
+/* pte_types.h  */
+#ifndef NK_COMPAT_PTE_H
+#define NK_COMPAT_PTE_H
+
+#include <nautilus/scheduler.h>
+
+
+//this is duplicate but to be independent copy the struct here
+
+typedef enum { ARRIVED=0,          // no admission control done
+               ADMITTED,           // admitted
+               CHANGING,           // admitted for new constraints, now changing to them
+               YIELDING,           // explit yield of slice
+               SLEEPING,           // being removed from RT and non-RT run/arrival queues
+                                   // probably due to having been put into a wait queue
+               EXITING,            // being removed from RT and non-RT run/arrival queues
+                                   // will not return
+               DENIED,             // not admitted
+               REAPABLE,           // it's OK for the reaper to destroy the thread
+             } rt_status;
+
+typedef enum { RUNNABLE_QUEUE = 0,
+               PENDING_QUEUE = 1,
+               APERIODIC_QUEUE = 2} queue_type;
+
+typedef struct nk_sched_thread_state {
+    // how this thread is to be scheduled
+    struct nk_sched_constraints constraints;
+    // 
+    rt_status status;
+    // which queue the thread is currently on
+    queue_type q_type;
+    
+    int      is_intr;      // this is an interrupt thread
+    int      is_task;      // this is a task thread
+
+    uint64_t start_time;   // when last started
+    uint64_t cur_run_time; // how long it has run without being preempted
+    uint64_t run_time;     // how long it has run so far
+                           // full duration for aperiodic and sporadic
+                           // current slice for periodic
+    uint64_t deadline;     // current deadline / time of next arrival if pending
+                           // for an aperiodic task, this is its current dynamic
+                           // priority 
+    uint64_t exit_time;    // time of actual completion / arrival
+
+    // Statistics are reset when the constraints are changed
+    uint64_t arrival_count;   // how many times it has arrived (1 for aperiodic/sporadic)
+    uint64_t resched_count;   // how many times resched was invoked on this thread
+    uint64_t resched_long_count; // how many times the long path was taken for the thread
+    uint64_t switch_in_count; // number of times switched to
+    uint64_t miss_count;      // number of deadline misses
+    uint64_t miss_time_sum;   // sum of missed time
+    uint64_t miss_time_sum2;  // sum of squares of missed time
+
+    // the thread context itself
+    struct nk_thread *thread;
+
+    // the thread node in a thread list (the global thread list)
+    struct rt_node   *list; 
+
+} rt_thread ;
+
+typedef struct nk_sched_thread_state rt_thread;
+typedef struct rt_node {
+    rt_thread      *thread;
+    struct rt_node *next;
+    struct rt_node *prev;
+} rt_node;
+
+
+#endif
diff --git a/src/compat/pthread/nk/nk_osal.c b/src/compat/pthread/nk/nk_osal.c
index b80bb86..7574f22 100644
--- a/src/compat/pthread/nk/nk_osal.c
+++ b/src/compat/pthread/nk/nk_osal.c
@@ -2,6 +2,7 @@
 #include "../debug.h"
 #include "pte_types.h"
 #include "pte_osal.h"
+#include "nk_compatibility.h"
 #include <nautilus/libccompat.h>
 #include <nautilus/scheduler.h>
 #include <nautilus/semaphore.h>
@@ -149,7 +150,7 @@ pte_osResult pte_osThreadCreate(pte_osThreadEntryPoint entryPoint,
   //pte_osThreadHandle handleobj = malloc(sizeof(nk_thread_t));
   //pte_osThreadHandle handleobj = malloc(sizeof(struct thread_with_signal));
   //memset(handleobj,0,sizeof(nk_thread_t));
-
+  //
   //handleobj->priority = initialPriority;
  
   struct sys_info *sys = per_cpu_get(system);
@@ -159,10 +160,12 @@ pte_osResult pte_osThreadCreate(pte_osThreadEntryPoint entryPoint,
     ERROR("create error exit\n");
     return PTE_OS_NO_RESOURCES;
   }
-  
+   
   //*handle = handleobj;
   struct nk_thread* thread = (struct nk_thread*) (*handle);
 
+  //thread->sched_state->constraints.interrupt_priority_class = (uint8_t) initialPriority; 
+
   DEBUG("osThreadCreate %p, %lu ref %lu\n",  thread, thread->tid, thread->refcount);
  
   return PTE_OS_OK;
@@ -212,13 +215,20 @@ pte_osThreadHandle pte_osThreadGetHandle(void){
   //return (pte_osThreadHandle) pcontainer_of(thethread,struct thread_with_signal, tid);
 }
 
-int pte_osThreadGetPriority(pte_osThreadHandle threadHandle){
-  return 0;
+int pte_osThreadGetPriority(pte_osThreadHandle handle){
+
+    return 0;	
+    return  (int) ((nk_thread_t *) handle)->sched_state->constraints.interrupt_priority_class;
+    //return 0;
     //threadHandle->priority;
 }
 
-pte_osResult pte_osThreadSetPriority(pte_osThreadHandle threadHandle, int newPriority){
-  //threadHandle->priority = newPriority;
+pte_osResult pte_osThreadSetPriority(pte_osThreadHandle handle, int newPriority){
+
+ // struct nk_sched_constraints priority; 
+ // priority.interrupt_priority_class = newPriority;
+ // nk_sched_thread_change_constraints(&priority);
+	
   return PTE_OS_OK;
 
 }
@@ -291,7 +301,7 @@ void pte_osThreadSleep(unsigned int msecs){
  * Returns the maximum allowable priority
  *============================================*/
 int pte_osThreadGetMaxPriority(){
-  return 3;
+  return 0xe;
 
 }
 
@@ -299,7 +309,7 @@ int pte_osThreadGetMaxPriority(){
  * Returns the minimum allowable priority
  *=========================================*/
 int pte_osThreadGetMinPriority(){
-  return 0;
+  return 0x0;
 }
 
 /*==================================================================================*
@@ -307,7 +317,7 @@ int pte_osThreadGetMinPriority(){
  * explicitly set one.
  *===================================================================================*/
 int pte_osThreadGetDefaultPriority(){
-  return 0;
+  return 0x0;
 }
 
 /*=================================================================================*
diff --git a/src/compat/pthread/nk/pte_types.h b/src/compat/pthread/nk/pte_types.h
index 0e41578..caec682 100644
--- a/src/compat/pthread/nk/pte_types.h
+++ b/src/compat/pthread/nk/pte_types.h
@@ -51,4 +51,5 @@ struct timeb
   short dstflag;
 };
 
+
 #endif /* PTE_TYPES_H */
diff --git a/src/compat/pthread/nk_get_pte_thread_t.c b/src/compat/pthread/nk_get_pte_thread_t.c
new file mode 100644
index 0000000..5532c03
--- /dev/null
+++ b/src/compat/pthread/nk_get_pte_thread_t.c
@@ -0,0 +1,10 @@
+#include <nautilus/nautilus.h>
+#include "pthread.h"
+#include "implement.h"
+
+pte_thread_t * nk_get_pte_thread_t(){
+
+  nk_thread_t *cur = get_cur_thread();
+  return *((pte_thread_t **) cur->hwtls+8);
+
+}
diff --git a/src/compat/pthread/pte_detach.c b/src/compat/pthread/pte_detach.c
index f6195d7..da3a1b5 100644
--- a/src/compat/pthread/pte_detach.c
+++ b/src/compat/pthread/pte_detach.c
@@ -52,7 +52,9 @@ pte_thread_detach_common (unsigned char threadShouldExit)
        * Don't use pthread_self() - to avoid creating an implicit POSIX thread handle
        * unnecessarily.
        */
-      pte_thread_t * sp = (pte_thread_t *) pthread_getspecific (pte_selfThreadKey);
+      //mjc
+      pte_thread_t * sp = nk_get_pte_thread_t();
+      //pte_thread_t * sp = (pte_thread_t *) pthread_getspecific (pte_selfThreadKey);
 
       if (sp != NULL) // otherwise OS thread with no implicit POSIX handle.
         {
@@ -78,8 +80,8 @@ pte_thread_detach_common (unsigned char threadShouldExit)
                 {
                   pte_threadDestroy (sp->ptHandle);
                 }
-
-              // pte_osTlsSetValue (pte_selfThreadKey->key, NULL);
+                // mjc muted but no unmute
+                 pte_osTlsSetValue (pte_selfThreadKey->key, NULL);
             }
           else
             {
diff --git a/src/compat/pthread/pte_threadStart.c b/src/compat/pthread/pte_threadStart.c
index e7ef8f4..d9e9cf9 100644
--- a/src/compat/pthread/pte_threadStart.c
+++ b/src/compat/pthread/pte_threadStart.c
@@ -97,6 +97,7 @@ int pte_threadStart (void *vthreadParms, void ** out)
   arg = threadParms->arg;
 //  free (threadParms);
 
+  //mjc we don't need this
   pthread_setspecific (pte_selfThreadKey, sp);
 
   sp->state = PThreadStateRunning;
diff --git a/src/compat/pthread/pte_throw.c b/src/compat/pthread/pte_throw.c
index 7f1ee7e..5dfaa13 100644
--- a/src/compat/pthread/pte_throw.c
+++ b/src/compat/pthread/pte_throw.c
@@ -61,7 +61,10 @@ pte_throw (unsigned int exception)
    * Don't use pthread_self() to avoid creating an implicit POSIX thread handle
    * unnecessarily.
    */
-  pte_thread_t * sp = (pte_thread_t *) pthread_getspecific (pte_selfThreadKey);
+
+  //mjc
+  //pte_thread_t * sp = (pte_thread_t *) pthread_getspecific (pte_selfThreadKey);
+  pte_thread_t * sp = nk_get_pte_thread_t();
 
 
   if (exception != PTE_EPS_CANCEL && exception != PTE_EPS_EXIT)
diff --git a/src/compat/pthread/pthread_exit.c b/src/compat/pthread/pthread_exit.c
index ef48e9b..ea0e280 100644
--- a/src/compat/pthread/pthread_exit.c
+++ b/src/compat/pthread/pthread_exit.c
@@ -74,7 +74,9 @@ pthread_exit (void *value_ptr)
    * Don't use pthread_self() to avoid creating an implicit POSIX thread handle
    * unnecessarily.
    */
-  sp = (pte_thread_t *) pthread_getspecific (pte_selfThreadKey);
+  //mjc
+  //  sp = (pte_thread_t *) pthread_getspecific (pte_selfThreadKey);
+  sp = nk_get_pte_thread_t();
 
   if (NULL == sp)
     {
diff --git a/src/compat/pthread/pthread_self.c b/src/compat/pthread/pthread_self.c
index e5f075d..d0aca89 100644
--- a/src/compat/pthread/pthread_self.c
+++ b/src/compat/pthread/pthread_self.c
@@ -46,6 +46,13 @@
 #include "pthread.h"
 #include "implement.h"
 
+#define offsetof(TYPE, MEMBER) ((size_t) &((TYPE *)0)->MEMBER)
+#define container_of(ptr, type, member) ({                      \
+        const typeof( ((type *)0)->member ) *__mptr = (ptr);    \
+        (type *)( (char *)__mptr - offsetof(type,member) );})
+
+
+
 pthread_t
 pthread_self (void)
 /*
@@ -71,7 +78,10 @@ pthread_self (void)
   pthread_t self;
   pte_thread_t * sp;
 
-  sp = (pte_thread_t *) pthread_getspecific (pte_selfThreadKey);
+  //mjc
+  sp = nk_get_pte_thread_t();
+
+  //sp = (pte_thread_t *) pthread_getspecific (pte_selfThreadKey);
 
   if (sp != NULL)
     {
@@ -107,7 +117,15 @@ pthread_self (void)
            * because the new handle is not yet public.
            */
           sp->sched_priority = 0;
-
+          
+	  nk_thread_t *threadptr = (nk_thread_t*) sp->threadId;
+          //plus 8 means add 8 bytes
+          pte_thread_t ** ptr = (pte_thread_t **) (threadptr->hwtls+8);
+          *ptr = sp;
+          //mjc set this, seems to be a bug from pte repo
+          sp->ptHandle = self;
+  
+	  //mjc we don't need this
           pthread_setspecific (pte_selfThreadKey, (void *) sp);
         }
     }
diff --git a/src/compat/pthread/pthread_setspecific.c b/src/compat/pthread/pthread_setspecific.c
index 62928a3..b8312f5 100644
--- a/src/compat/pthread/pthread_setspecific.c
+++ b/src/compat/pthread/pthread_setspecific.c
@@ -83,6 +83,7 @@ pthread_setspecific (pthread_key_t key, const void *value)
        * an instance of pthread_t for the current
        * thread if one wasn't explicitly created
        */
+      //mjc confirm this
       self = pthread_self ();
       if (self.p == NULL)
         {
@@ -95,7 +96,9 @@ pthread_setspecific (pthread_key_t key, const void *value)
        * Resolve catch-22 of registering thread with selfThread
        * key
        */
-      pte_thread_t * sp = (pte_thread_t *) pthread_getspecific (pte_selfThreadKey);
+      //mjc
+      //pte_thread_t * sp = (pte_thread_t *) pthread_getspecific (pte_selfThreadKey);
+      pte_thread_t * sp = nk_get_pte_thread_t();
 
       if (sp == NULL)
         {
-- 
2.25.1