KernelLand/Kernel/arch/x86_64/proc.c

   1 /*
   2  * Acess2 x86_64 port
   3  * proc.c
   4  */
   5 #include <acess.h>
   6 #include <proc.h>
   7 #include <threads.h>
   8 #include <threads_int.h>
   9 #include <desctab.h>
  10 #include <mm_virt.h>
  11 #include <errno.h>
  12 #if USE_MP
  13 # include <mp.h>
  14 #endif
  15 #include <arch_config.h>
  16 #include <hal_proc.h>
  17
  18 // === FLAGS ===
  19 #define DEBUG_TRACE_SWITCH      0
  20 #define BREAK_ON_SWITCH         0       // Break into bochs debugger on a task switch
  21
  22 // === CONSTANTS ===
  23
  24 // === TYPES ===
  25 typedef struct sCPU
  26 {
  27         Uint8   APICID;
  28         Uint8   State;  // 0: Unavaliable, 1: Idle, 2: Active
  29         Uint16  Resvd;
  30         tThread *Current;
  31         tThread *IdleThread;
  32 }       tCPU;
  33
  34 // === IMPORTS ===
  35 extern tGDT     gGDT[];
  36 extern void     APStartup(void);        // 16-bit AP startup code
  37
  38 extern Uint     GetRIP(void);   // start.asm
  39 extern Uint     SaveState(Uint *RSP, Uint *Regs);
  40 extern Uint     Proc_CloneInt(Uint *RSP, Uint *CR3);
  41 extern void     NewTaskHeader(void);    // Actually takes cdecl args
  42 extern void     Proc_InitialiseSSE(void);
  43 extern void     Proc_SaveSSE(Uint DestPtr);
  44 extern void     Proc_DisableSSE(void);
  45
  46 extern Uint64   gInitialPML4[512];      // start.asm
  47 extern int      giNumCPUs;
  48 extern int      giNextTID;
  49 extern int      giTotalTickets;
  50 extern int      giNumActiveThreads;
  51 extern tThread  gThreadZero;
  52 extern tProcess gProcessZero;
  53 extern void     Threads_Dump(void);
  54 extern void     Proc_ReturnToUser(tVAddr Handler, tVAddr KStackTop, int Argument);
  55 extern void     Time_UpdateTimestamp(void);
  56 extern void     SwitchTasks(Uint NewSP, Uint *OldSP, Uint NewIP, Uint *OldIO, Uint CR3);
  57
  58 // === PROTOTYPES ===
  59 //void  ArchThreads_Init(void);
  60 #if USE_MP
  61 void    MP_StartAP(int CPU);
  62 void    MP_SendIPI(Uint8 APICID, int Vector, int DeliveryMode);
  63 #endif
  64 void    Proc_IdleTask(void *unused);
  65 //void  Proc_Start(void);
  66 //tThread       *Proc_GetCurThread(void);
  67 // int  Proc_NewKThread(void (*Fcn)(void*), void *Data);
  68 // int  Proc_Clone(Uint *Err, Uint Flags);
  69 // int  Proc_SpawnWorker(void);
  70 Uint    Proc_MakeUserStack(void);
  71 //void  Proc_StartUser(Uint Entrypoint, Uint *Bases, int ArgC, char **ArgV, char **EnvP, int DataSize);
  72 void    Proc_StartProcess(Uint16 SS, Uint Stack, Uint Flags, Uint16 CS, Uint IP) NORETURN;
  73  int    Proc_Demote(Uint *Err, int Dest, tRegs *Regs);
  74 //void  Proc_CallFaultHandler(tThread *Thread);
  75 //void  Proc_DumpThreadCPUState(tThread *Thread);
  76 //void  Proc_Reschedule(void);
  77 void    Proc_Scheduler(int CPU, Uint RSP, Uint RIP);
  78
  79 // === GLOBALS ===
  80 //!\brief Used by desctab.asm in SyscallStub
  81 const int ci_offsetof_tThread_KernelStack = offsetof(tThread, KernelStack);
  82 // --- Multiprocessing ---
  83 #if USE_MP
  84 volatile int    giNumInitingCPUs = 0;
  85 tMPInfo *gMPFloatPtr = NULL;
  86 tAPIC   *gpMP_LocalAPIC = NULL;
  87 Uint8   gaAPIC_to_CPU[256] = {0};
  88 #endif
  89 tCPU    gaCPUs[MAX_CPUS] = {
  90         {.Current = &gThreadZero}
  91         };
  92 tTSS    *gTSSs = NULL;
  93 tTSS    gTSS0 = {0};
  94 // --- Error Recovery ---
  95 Uint32  gaDoubleFaultStack[1024];
  96
  97 // === CODE ===
  98 /**
  99  * \fn void ArchThreads_Init(void)
 100  * \brief Starts the process scheduler
 101  */
 102 void ArchThreads_Init(void)
 103 {
 104         Uint    pos = 0;
 105
 106         #if USE_MP
 107         tMPTable        *mptable;
 108
 109         // Mark BSP as active
 110         gaCPUs[0].State = 2;
 111
 112         // -- Initialise Multiprocessing
 113         // Find MP Floating Table
 114         // - EBDA/Last 1Kib (640KiB)
 115         for(pos = KERNEL_BASE|0x9F000; pos < (KERNEL_BASE|0xA0000); pos += 16) {
 116                 if( *(Uint*)(pos) == MPPTR_IDENT ) {
 117                         Log("Possible %p", pos);
 118                         if( ByteSum((void*)pos, sizeof(tMPInfo)) != 0 ) continue;
 119                         gMPFloatPtr = (void*)pos;
 120                         break;
 121                 }
 122         }
 123         // - Last KiB (512KiB base mem)
 124         if(!gMPFloatPtr) {
 125                 for(pos = KERNEL_BASE|0x7F000; pos < (KERNEL_BASE|0x80000); pos += 16) {
 126                         if( *(Uint*)(pos) == MPPTR_IDENT ) {
 127                                 Log("Possible %p", pos);
 128                                 if( ByteSum((void*)pos, sizeof(tMPInfo)) != 0 ) continue;
 129                                 gMPFloatPtr = (void*)pos;
 130                                 break;
 131                         }
 132                 }
 133         }
 134         // - BIOS ROM
 135         if(!gMPFloatPtr) {
 136                 for(pos = KERNEL_BASE|0xE0000; pos < (KERNEL_BASE|0x100000); pos += 16) {
 137                         if( *(Uint*)(pos) == MPPTR_IDENT ) {
 138                                 Log("Possible %p", pos);
 139                                 if( ByteSum((void*)pos, sizeof(tMPInfo)) != 0 ) continue;
 140                                 gMPFloatPtr = (void*)pos;
 141                                 break;
 142                         }
 143                 }
 144         }
 145
 146         // If the MP Table Exists, parse it
 147         if(gMPFloatPtr)
 148         {
 149                  int    i;
 150                 tMPTable_Ent    *ents;
 151                 Log("gMPFloatPtr = %p", gMPFloatPtr);
 152                 Log("*gMPFloatPtr = {");
 153                 Log("\t.Sig = 0x%08x", gMPFloatPtr->Sig);
 154                 Log("\t.MPConfig = 0x%08x", gMPFloatPtr->MPConfig);
 155                 Log("\t.Length = 0x%02x", gMPFloatPtr->Length);
 156                 Log("\t.Version = 0x%02x", gMPFloatPtr->Version);
 157                 Log("\t.Checksum = 0x%02x", gMPFloatPtr->Checksum);
 158                 Log("\t.Features = [0x%02x,0x%02x,0x%02x,0x%02x,0x%02x]",
 159                         gMPFloatPtr->Features[0],       gMPFloatPtr->Features[1],
 160                         gMPFloatPtr->Features[2],       gMPFloatPtr->Features[3],
 161                         gMPFloatPtr->Features[4]
 162                         );
 163                 Log("}");
 164
 165                 mptable = (void*)( KERNEL_BASE|gMPFloatPtr->MPConfig );
 166                 Log("mptable = %p", mptable);
 167                 Log("*mptable = {");
 168                 Log("\t.Sig = 0x%08x", mptable->Sig);
 169                 Log("\t.BaseTableLength = 0x%04x", mptable->BaseTableLength);
 170                 Log("\t.SpecRev = 0x%02x", mptable->SpecRev);
 171                 Log("\t.Checksum = 0x%02x", mptable->Checksum);
 172                 Log("\t.OEMID = '%8c'", mptable->OemID);
 173                 Log("\t.ProductID = '%8c'", mptable->ProductID);
 174                 Log("\t.OEMTablePtr = %p'", mptable->OEMTablePtr);
 175                 Log("\t.OEMTableSize = 0x%04x", mptable->OEMTableSize);
 176                 Log("\t.EntryCount = 0x%04x", mptable->EntryCount);
 177                 Log("\t.LocalAPICMemMap = 0x%08x", mptable->LocalAPICMemMap);
 178                 Log("\t.ExtendedTableLen = 0x%04x", mptable->ExtendedTableLen);
 179                 Log("\t.ExtendedTableChecksum = 0x%02x", mptable->ExtendedTableChecksum);
 180                 Log("}");
 181
 182                 gpMP_LocalAPIC = (void*)MM_MapHWPage(mptable->LocalAPICMemMap, 1);
 183
 184                 ents = mptable->Entries;
 185                 giNumCPUs = 0;
 186
 187                 for( i = 0; i < mptable->EntryCount; i ++ )
 188                 {
 189                          int    entSize = 0;
 190                         switch( ents->Type )
 191                         {
 192                         case 0: // Processor
 193                                 entSize = 20;
 194                                 Log("%i: Processor", i);
 195                                 Log("\t.APICID = %i", ents->Proc.APICID);
 196                                 Log("\t.APICVer = 0x%02x", ents->Proc.APICVer);
 197                                 Log("\t.CPUFlags = 0x%02x", ents->Proc.CPUFlags);
 198                                 Log("\t.CPUSignature = 0x%08x", ents->Proc.CPUSignature);
 199                                 Log("\t.FeatureFlags = 0x%08x", ents->Proc.FeatureFlags);
 200
 201
 202                                 if( !(ents->Proc.CPUFlags & 1) ) {
 203                                         Log("DISABLED");
 204                                         break;
 205                                 }
 206
 207                                 // Check if there is too many processors
 208                                 if(giNumCPUs >= MAX_CPUS) {
 209                                         giNumCPUs ++;   // If `giNumCPUs` > MAX_CPUS later, it will be clipped
 210                                         break;
 211                                 }
 212
 213                                 // Initialise CPU Info
 214                                 gaAPIC_to_CPU[ents->Proc.APICID] = giNumCPUs;
 215                                 gaCPUs[giNumCPUs].APICID = ents->Proc.APICID;
 216                                 gaCPUs[giNumCPUs].State = 0;
 217                                 giNumCPUs ++;
 218
 219                                 // Send IPI
 220                                 if( !(ents->Proc.CPUFlags & 2) )
 221                                 {
 222                                         MP_StartAP( giNumCPUs-1 );
 223                                 }
 224
 225                                 break;
 226                         case 1: // Bus
 227                                 entSize = 8;
 228                                 Log("%i: Bus", i);
 229                                 Log("\t.ID = %i", ents->Bus.ID);
 230                                 Log("\t.TypeString = '%6c'", ents->Bus.TypeString);
 231                                 break;
 232                         case 2: // I/O APIC
 233                                 entSize = 8;
 234                                 Log("%i: I/O APIC", i);
 235                                 Log("\t.ID = %i", ents->IOAPIC.ID);
 236                                 Log("\t.Version = 0x%02x", ents->IOAPIC.Version);
 237                                 Log("\t.Flags = 0x%02x", ents->IOAPIC.Flags);
 238                                 Log("\t.Addr = 0x%08x", ents->IOAPIC.Addr);
 239                                 break;
 240                         case 3: // I/O Interrupt Assignment
 241                                 entSize = 8;
 242                                 Log("%i: I/O Interrupt Assignment", i);
 243                                 Log("\t.IntType = %i", ents->IOInt.IntType);
 244                                 Log("\t.Flags = 0x%04x", ents->IOInt.Flags);
 245                                 Log("\t.SourceBusID = 0x%02x", ents->IOInt.SourceBusID);
 246                                 Log("\t.SourceBusIRQ = 0x%02x", ents->IOInt.SourceBusIRQ);
 247                                 Log("\t.DestAPICID = 0x%02x", ents->IOInt.DestAPICID);
 248                                 Log("\t.DestAPICIRQ = 0x%02x", ents->IOInt.DestAPICIRQ);
 249                                 break;
 250                         case 4: // Local Interrupt Assignment
 251                                 entSize = 8;
 252                                 Log("%i: Local Interrupt Assignment", i);
 253                                 Log("\t.IntType = %i", ents->LocalInt.IntType);
 254                                 Log("\t.Flags = 0x%04x", ents->LocalInt.Flags);
 255                                 Log("\t.SourceBusID = 0x%02x", ents->LocalInt.SourceBusID);
 256                                 Log("\t.SourceBusIRQ = 0x%02x", ents->LocalInt.SourceBusIRQ);
 257                                 Log("\t.DestLocalAPICID = 0x%02x", ents->LocalInt.DestLocalAPICID);
 258                                 Log("\t.DestLocalAPICIRQ = 0x%02x", ents->LocalInt.DestLocalAPICIRQ);
 259                                 break;
 260                         default:
 261                                 Log("%i: Unknown (%i)", i, ents->Type);
 262                                 break;
 263                         }
 264                         ents = (void*)( (Uint)ents + entSize );
 265                 }
 266
 267                 if( giNumCPUs > MAX_CPUS ) {
 268                         Warning("Too many CPUs detected (%i), only using %i of them", giNumCPUs, MAX_CPUS);
 269                         giNumCPUs = MAX_CPUS;
 270                 }
 271
 272                 while( giNumInitingCPUs )
 273                         MM_FinishVirtualInit();
 274
 275                 Panic("Uh oh... MP Table Parsing is unimplemented\n");
 276         }
 277         else {
 278                 Log("No MP Table was found, assuming uniprocessor\n");
 279                 giNumCPUs = 1;
 280                 gTSSs = &gTSS0;
 281         }
 282         #else
 283         giNumCPUs = 1;
 284         gTSSs = &gTSS0;
 285         MM_FinishVirtualInit();
 286         #endif
 287
 288         #if USE_MP
 289         // Initialise Normal TSS(s)
 290         for(pos=0;pos<giNumCPUs;pos++)
 291         {
 292         #else
 293         pos = 0;
 294         #endif
 295                 gTSSs[pos].RSP0 = 0;    // Set properly by scheduler
 296                 gGDT[7+pos*2].LimitLow = sizeof(tTSS) & 0xFFFF;
 297                 gGDT[7+pos*2].BaseLow = ((Uint)(&gTSSs[pos])) & 0xFFFF;
 298                 gGDT[7+pos*2].BaseMid = ((Uint)(&gTSSs[pos])) >> 16;
 299                 gGDT[7+pos*2].BaseHi = ((Uint)(&gTSSs[pos])) >> 24;
 300                 gGDT[7+pos*2+1].DWord[0] = ((Uint)(&gTSSs[pos])) >> 32;
 301         #if USE_MP
 302         }
 303         for(pos=0;pos<giNumCPUs;pos++) {
 304                 __asm__ __volatile__ ("ltr %%ax"::"a"(0x38+pos*16));
 305         }
 306         #else
 307         __asm__ __volatile__ ("ltr %%ax"::"a"(0x38));
 308         #endif
 309
 310         // Set Debug registers
 311         __asm__ __volatile__ ("mov %0, %%db0" : : "r"(&gThreadZero));
 312         __asm__ __volatile__ ("mov %%rax, %%db1" : : "a"(0));
 313
 314         gaCPUs[0].Current = &gThreadZero;
 315
 316         gProcessZero.MemState.CR3 = (Uint)gInitialPML4 - KERNEL_BASE;
 317         gThreadZero.CurCPU = 0;
 318         gThreadZero.KernelStack = 0xFFFFA00000000000 + KERNEL_STACK_SIZE;
 319
 320         // Set timer frequency
 321         outb(0x43, 0x34);       // Set Channel 0, Low/High, Rate Generator
 322         outb(0x40, PIT_TIMER_DIVISOR&0xFF);     // Low Byte of Divisor
 323         outb(0x40, (PIT_TIMER_DIVISOR>>8)&0xFF);        // High Byte
 324
 325         // Create Per-Process Data Block
 326         if( !MM_Allocate(MM_PPD_CFG) )
 327         {
 328                 Warning("Oh, hell, Unable to allocate PPD for Thread#0");
 329         }
 330
 331         Proc_InitialiseSSE();
 332
 333         Log_Log("Proc", "Multithreading initialised");
 334 }
 335
 336 #if USE_MP
 337 void MP_StartAP(int CPU)
 338 {
 339         Log("Starting AP %i (APIC %i)", CPU, gaCPUs[CPU].APICID);
 340         // Set location of AP startup code and mark for a warm restart
 341         *(Uint16*)(KERNEL_BASE|0x467) = (Uint)&APStartup - (KERNEL_BASE|0xFFFF0);
 342         *(Uint16*)(KERNEL_BASE|0x469) = 0xFFFF;
 343         outb(0x70, 0x0F);       outb(0x71, 0x0A);       // Warm Reset
 344         MP_SendIPI(gaCPUs[CPU].APICID, 0, 5);
 345         giNumInitingCPUs ++;
 346 }
 347
 348 void MP_SendIPI(Uint8 APICID, int Vector, int DeliveryMode)
 349 {
 350         Uint32  addr = (Uint)gpMP_LocalAPIC + 0x300;
 351         Uint32  val;
 352
 353         // High
 354         val = (Uint)APICID << 24;
 355         Log("*%p = 0x%08x", addr+0x10, val);
 356         *(Uint32*)(addr+0x10) = val;
 357         // Low (and send)
 358         val = ((DeliveryMode & 7) << 8) | (Vector & 0xFF);
 359         Log("*%p = 0x%08x", addr, val);
 360         *(Uint32*)addr = val;
 361 }
 362 #endif
 363
 364 /**
 365  * \brief Idle task
 366  */
 367 void Proc_IdleTask(void *ptr)
 368 {
 369         tCPU    *cpu = ptr;
 370         cpu->IdleThread = Proc_GetCurThread();
 371         cpu->IdleThread->ThreadName = (char*)"Idle Thread";
 372         Threads_SetPriority( cpu->IdleThread, -1 );     // Never called randomly
 373         cpu->IdleThread->Quantum = 1;   // 1 slice quantum
 374         for(;;) {
 375                 HALT(); // Just yeilds
 376                 Threads_Yield();
 377         }
 378 }
 379
 380 /**
 381  * \fn void Proc_Start(void)
 382  * \brief Start process scheduler
 383  */
 384 void Proc_Start(void)
 385 {
 386         #if USE_MP
 387          int    i;
 388         #endif
 389
 390         #if USE_MP
 391         // Start APs
 392         for( i = 0; i < giNumCPUs; i ++ )
 393         {
 394                  int    tid;
 395                 if(i)   gaCPUs[i].Current = NULL;
 396
 397                 Proc_NewKThread(Proc_IdleTask, &gaCPUs[i]);
 398
 399                 // Create Idle Task
 400                 gaCPUs[i].IdleThread = Threads_GetThread(tid);
 401
 402
 403                 // Start the AP
 404                 if( i != giProc_BootProcessorID ) {
 405                         MP_StartAP( i );
 406                 }
 407         }
 408
 409         // BSP still should run the current task
 410         gaCPUs[0].Current = &gThreadZero;
 411         __asm__ __volatile__ ("mov %0, %%db0" : : "r"(&gThreadZero));
 412
 413         // Start interrupts and wait for APs to come up
 414         Log("Waiting for APs to come up\n");
 415         __asm__ __volatile__ ("sti");
 416         while( giNumInitingCPUs )       __asm__ __volatile__ ("hlt");
 417         #else
 418         Proc_NewKThread(Proc_IdleTask, &gaCPUs[0]);
 419
 420         // Start Interrupts (and hence scheduler)
 421         __asm__ __volatile__("sti");
 422         #endif
 423         MM_FinishVirtualInit();
 424         Log_Log("Proc", "Multithreading started");
 425 }
 426
 427 /**
 428  * \fn tThread *Proc_GetCurThread(void)
 429  * \brief Gets the current thread
 430  */
 431 tThread *Proc_GetCurThread(void)
 432 {
 433         #if USE_MP
 434         tThread *ret;
 435         __asm__ __volatile__ ("mov %%db0, %0" : "=r"(thread));
 436         return ret;     // gaCPUs[ GetCPUNum() ].Current;
 437         #else
 438         return gaCPUs[ 0 ].Current;
 439         #endif
 440 }
 441
 442 void Proc_ClearProcess(tProcess *Process)
 443 {
 444         Log_Warning("Proc", "TODO: Nuke address space etc");
 445 }
 446
 447 /*
 448  *
 449  */
 450 void Proc_ClearThread(tThread *Thread)
 451 {
 452 }
 453
 454 /**
 455  * \brief Create a new kernel thread
 456  */
 457 tTID Proc_NewKThread(void (*Fcn)(void*), void *Data)
 458 {
 459         Uint    rsp;
 460         tThread *newThread, *cur;
 461
 462         cur = Proc_GetCurThread();
 463         newThread = Threads_CloneTCB(0);
 464         if(!newThread)  return -1;
 465
 466         // Create new KStack
 467         newThread->KernelStack = MM_NewKStack();
 468         // Check for errors
 469         if(newThread->KernelStack == 0) {
 470                 free(newThread);
 471                 return -1;
 472         }
 473
 474         rsp = newThread->KernelStack;
 475         *(Uint*)(rsp-=8) = (Uint)Data;  // Data (shadowed)
 476         *(Uint*)(rsp-=8) = (Uint)Fcn;   // Function to call
 477         *(Uint*)(rsp-=8) = (Uint)newThread;     // Thread ID
 478
 479         newThread->SavedState.RSP = rsp;
 480         newThread->SavedState.RIP = (Uint)&NewTaskHeader;
 481         newThread->SavedState.SSE = NULL;
 482 //      Log("New (KThread) %p, rsp = %p\n", newThread->SavedState.RIP, newThread->SavedState.RSP);
 483
 484 //      MAGIC_BREAK();
 485         Threads_AddActive(newThread);
 486
 487         return newThread->TID;
 488 }
 489
 490 /**
 491  * \fn int Proc_Clone(Uint Flags)
 492  * \brief Clone the current process
 493  */
 494 tTID Proc_Clone(Uint Flags)
 495 {
 496         tThread *newThread, *cur = Proc_GetCurThread();
 497         Uint    rip;
 498
 499         // Sanity check
 500         if( !(Flags & CLONE_VM) ) {
 501                 Log_Error("Proc", "Proc_Clone: Don't leave CLONE_VM unset, use Proc_NewKThread instead");
 502                 return -1;
 503         }
 504
 505         // Create new TCB
 506         newThread = Threads_CloneTCB(Flags);
 507         if(!newThread)  return -1;
 508
 509         // Save core machine state
 510         rip = Proc_CloneInt(&newThread->SavedState.RSP, &newThread->Process->MemState.CR3);
 511         if(rip == 0)    return 0;       // Child
 512         newThread->KernelStack = cur->KernelStack;
 513         newThread->SavedState.RIP = rip;
 514         newThread->SavedState.SSE = NULL;
 515
 516         // DEBUG
 517         #if 0
 518         Log("New (Clone) %p, rsp = %p, cr3 = %p", rip, newThread->SavedState.RSP, newThread->MemState.CR3);
 519         {
 520                 Uint cr3;
 521                 __asm__ __volatile__ ("mov %%cr3, %0" : "=r" (cr3));
 522                 Log("Current CR3 = 0x%x, PADDR(RSP) = 0x%x", cr3, MM_GetPhysAddr(newThread->SavedState.RSP));
 523         }
 524         #endif
 525         // /DEBUG
 526
 527         // Lock list and add to active
 528         Threads_AddActive(newThread);
 529
 530         return newThread->TID;
 531 }
 532
 533 /**
 534  * \fn int Proc_SpawnWorker(void)
 535  * \brief Spawns a new worker thread
 536  */
 537 int Proc_SpawnWorker(void (*Fcn)(void*), void *Data)
 538 {
 539         tThread *new, *cur;
 540         Uint    stack_contents[3];
 541
 542         cur = Proc_GetCurThread();
 543
 544         // Create new thread
 545         new = Threads_CloneThreadZero();
 546         if(!new) {
 547                 Warning("Proc_SpawnWorker - Out of heap space!\n");
 548                 return -1;
 549         }
 550
 551         // Create the stack contents
 552         stack_contents[2] = (Uint)Data;
 553         stack_contents[1] = (Uint)Fcn;
 554         stack_contents[0] = (Uint)new;
 555
 556         // Create a new worker stack (in PID0's address space)
 557         // - The stack is built by this code using stack_contents
 558         new->KernelStack = MM_NewWorkerStack(stack_contents, sizeof(stack_contents));
 559
 560         new->SavedState.RSP = new->KernelStack - sizeof(stack_contents);
 561         new->SavedState.RIP = (Uint)&NewTaskHeader;
 562         new->SavedState.SSE = NULL;
 563
 564 //      Log("New (Worker) %p, rsp = %p\n", new->SavedState.RIP, new->SavedState.RSP);
 565
 566         // Mark as active
 567         new->Status = THREAD_STAT_PREINIT;
 568         Threads_AddActive( new );
 569
 570         return new->TID;
 571 }
 572
 573 /**
 574  * \brief Creates a new user stack
 575  */
 576 Uint Proc_MakeUserStack(void)
 577 {
 578          int    i;
 579         Uint    base = USER_STACK_TOP - USER_STACK_SZ;
 580
 581         // Check Prospective Space
 582         for( i = USER_STACK_SZ >> 12; i--; )
 583         {
 584                 if( MM_GetPhysAddr( (void*)(base + (i<<12)) ) != 0 )
 585                         break;
 586         }
 587
 588         if(i != -1)     return 0;
 589
 590         // Allocate Stack - Allocate incrementally to clean up MM_Dump output
 591         // - Most of the user stack is the zero page
 592         for( i = 0; i < (USER_STACK_SZ-USER_STACK_PREALLOC)/0x1000; i++ )
 593         {
 594                 MM_AllocateZero( base + (i<<12) );
 595         }
 596         // - but the top USER_STACK_PREALLOC pages are actually allocated
 597         for( ; i < USER_STACK_SZ/0x1000; i++ )
 598         {
 599                 tPAddr  alloc = MM_Allocate( base + (i<<12) );
 600                 if( !alloc )
 601                 {
 602                         // Error
 603                         Log_Error("Proc", "Unable to allocate user stack (%i pages requested)", USER_STACK_SZ/0x1000);
 604                         while( i -- )
 605                                 MM_Deallocate( base + (i<<12) );
 606                         return 0;
 607                 }
 608         }
 609
 610         return base + USER_STACK_SZ;
 611 }
 612
 613 void Proc_StartUser(Uint Entrypoint, Uint Base, int ArgC, const char **ArgV, int DataSize)
 614 {
 615         Uint    *stack;
 616          int    i;
 617         const char      **envp = NULL;
 618         Uint16  ss, cs;
 619
 620
 621         // Copy Arguments
 622         stack = (void*)Proc_MakeUserStack();
 623         if(!stack) {
 624                 Log_Error("Proc", "Unable to create user stack!");
 625                 Threads_Exit(0, -1);
 626         }
 627         stack -= (DataSize+7)/8;
 628         memcpy( stack, ArgV, DataSize );
 629         free(ArgV);
 630
 631         // Adjust Arguments and environment
 632         if(DataSize)
 633         {
 634                 Uint    delta = (Uint)stack - (Uint)ArgV;
 635                 ArgV = (const char**)stack;
 636                 for( i = 0; ArgV[i]; i++ )      ArgV[i] += delta;
 637                 envp = &ArgV[i+1];
 638                 for( i = 0; envp[i]; i++ )      envp[i] += delta;
 639         }
 640
 641         // User Mode Segments
 642         // 0x2B = 64-bit
 643         ss = 0x23;      cs = 0x2B;
 644
 645         // Arguments
 646         *--stack = (Uint)envp;
 647         *--stack = (Uint)ArgV;
 648         *--stack = (Uint)ArgC;
 649         *--stack = Base;
 650
 651         Proc_StartProcess(ss, (Uint)stack, 0x202, cs, Entrypoint);
 652 }
 653
 654 void Proc_StartProcess(Uint16 SS, Uint Stack, Uint Flags, Uint16 CS, Uint IP)
 655 {
 656         if( !(CS == 0x1B || CS == 0x2B) || SS != 0x23 ) {
 657                 Log_Error("Proc", "Proc_StartProcess: CS / SS are not valid (%x, %x)",
 658                         CS, SS);
 659                 Threads_Exit(0, -1);
 660         }
 661 //      Log("Proc_StartProcess: (SS=%x, Stack=%p, Flags=%x, CS=%x, IP=%p)", SS, Stack, Flags, CS, IP);
 662 //      MM_DumpTables(0, USER_MAX);
 663         if(CS == 0x1B)
 664         {
 665                 // 32-bit return
 666                 __asm__ __volatile__ (
 667                         "mov %0, %%rsp;\n\t"    // Set stack pointer
 668                         "mov %2, %%r11;\n\t"    // Set RFLAGS
 669                         "sysret;\n\t"
 670                         : : "r" (Stack), "c" (IP), "r" (Flags)
 671                         );
 672         }
 673         else
 674         {
 675                 // 64-bit return
 676                 __asm__ __volatile__ (
 677                         "mov %0, %%rsp;\n\t"    // Set stack pointer
 678                         "mov %2, %%r11;\n\t"    // Set RFLAGS
 679                         "sysretq;\n\t"
 680                         : : "r" (Stack), "c" (IP), "r" (Flags)
 681                         );
 682         }
 683         for(;;);
 684 }
 685
 686 /**
 687  * \fn int Proc_Demote(Uint *Err, int Dest, tRegs *Regs)
 688  * \brief Demotes a process to a lower permission level
 689  * \param Err   Pointer to user's errno
 690  * \param Dest  New Permission Level
 691  * \param Regs  Pointer to user's register structure
 692  */
 693 int Proc_Demote(Uint *Err, int Dest, tRegs *Regs)
 694 {
 695          int    cpl = Regs->CS & 3;
 696         // Sanity Check
 697         if(Dest > 3 || Dest < 0) {
 698                 *Err = -EINVAL;
 699                 return -1;
 700         }
 701
 702         // Permission Check
 703         if(cpl > Dest) {
 704                 *Err = -EACCES;
 705                 return -1;
 706         }
 707
 708         // Change the Segment Registers
 709         Regs->CS = (((Dest+1)<<4) | Dest) - 8;
 710         Regs->SS = ((Dest+1)<<4) | Dest;
 711
 712         return 0;
 713 }
 714
 715 /**
 716  * \brief Calls a signal handler in user mode
 717  * \note Used for signals
 718  */
 719 void Proc_CallFaultHandler(tThread *Thread)
 720 {
 721         // Never returns
 722         Proc_ReturnToUser(Thread->FaultHandler, Thread->KernelStack, Thread->CurFaultNum);
 723         for(;;);
 724 }
 725
 726 void Proc_DumpThreadCPUState(tThread *Thread)
 727 {
 728         Log("  At %04x:%016llx", Thread->SavedState.UserCS, Thread->SavedState.UserRIP);
 729 }
 730
 731 void Proc_Reschedule(void)
 732 {
 733         tThread *nextthread, *curthread;
 734          int    cpu = GetCPUNum();
 735
 736         // TODO: Wait for it?
 737         if(IS_LOCKED(&glThreadListLock))        return;
 738
 739         curthread = gaCPUs[cpu].Current;
 740
 741         nextthread = Threads_GetNextToRun(cpu, curthread);
 742
 743         if(nextthread == curthread)     return ;
 744         if(!nextthread)
 745                 nextthread = gaCPUs[cpu].IdleThread;
 746         if(!nextthread)
 747                 return ;
 748
 749         #if DEBUG_TRACE_SWITCH
 750         LogF("\nSwitching to task CR3 = 0x%x, RIP = %p, RSP = %p - %i (%s)\n",
 751                 nextthread->Process->MemState.CR3,
 752                 nextthread->SavedState.RIP,
 753                 nextthread->SavedState.RSP,
 754                 nextthread->TID,
 755                 nextthread->ThreadName
 756                 );
 757         #endif
 758
 759         // Update CPU state
 760         gaCPUs[cpu].Current = nextthread;
 761         gTSSs[cpu].RSP0 = nextthread->KernelStack-4;
 762         __asm__ __volatile__ ("mov %0, %%db0" : : "r" (nextthread));
 763
 764         if( curthread )
 765         {
 766                 // Save FPU/MMX/XMM/SSE state
 767                 if( curthread->SavedState.SSE )
 768                 {
 769                         Proc_SaveSSE( ((Uint)curthread->SavedState.SSE + 0xF) & ~0xF );
 770                         curthread->SavedState.bSSEModified = 0;
 771                         Proc_DisableSSE();
 772                 }
 773                 SwitchTasks(
 774                         nextthread->SavedState.RSP, &curthread->SavedState.RSP,
 775                         nextthread->SavedState.RIP, &curthread->SavedState.RIP,
 776                         nextthread->Process->MemState.CR3
 777                         );
 778         }
 779         else
 780         {
 781                 Uint    tmp;
 782                 SwitchTasks(
 783                         nextthread->SavedState.RSP, &tmp,
 784                         nextthread->SavedState.RIP, &tmp,
 785                         nextthread->Process->MemState.CR3
 786                         );
 787         }
 788         return ;
 789 }
 790
 791 /**
 792  * \fn void Proc_Scheduler(int CPU)
 793  * \brief Swap current thread and clears dead threads
 794  */
 795 void Proc_Scheduler(int CPU, Uint RSP, Uint RIP)
 796 {
 797 #if 0
 798         tThread *thread;
 799
 800         // If the spinlock is set, let it complete
 801         if(IS_LOCKED(&glThreadListLock))        return;
 802
 803         // Get current thread
 804         thread = gaCPUs[CPU].Current;
 805
 806         if( thread )
 807         {
 808                 tRegs   *regs;
 809                 // Reduce remaining quantum and continue timeslice if non-zero
 810                 if(thread->Remaining--) return;
 811                 // Reset quantum for next call
 812                 thread->Remaining = thread->Quantum;
 813
 814                 // TODO: Make this more stable somehow
 815                 {
 816                         regs = (tRegs*)(RSP+(1)*8);     // CurThread
 817                         thread->SavedState.UserCS = regs->CS;
 818                         thread->SavedState.UserRIP = regs->RIP;
 819                 }
 820         }
 821
 822         // ACK Timer here?
 823
 824         Proc_Reschedule();
 825 #endif
 826 }
 827
 828 // === EXPORTS ===
 829 EXPORT(Proc_SpawnWorker);