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 void     NewTaskHeader(void);    // Actually takes cdecl args
  41
  42 extern Uint64   gInitialPML4[512];      // start.asm
  43 extern tShortSpinlock   glThreadListLock;
  44 extern int      giNumCPUs;
  45 extern int      giNextTID;
  46 extern int      giTotalTickets;
  47 extern int      giNumActiveThreads;
  48 extern tThread  gThreadZero;
  49 extern void     Threads_Dump(void);
  50 extern void     Proc_ReturnToUser(void);
  51 extern void     Time_UpdateTimestamp(void);
  52
  53 // === PROTOTYPES ===
  54 //void  ArchThreads_Init(void);
  55 #if USE_MP
  56 void    MP_StartAP(int CPU);
  57 void    MP_SendIPI(Uint8 APICID, int Vector, int DeliveryMode);
  58 #endif
  59 void    Proc_IdleTask(void *unused);
  60 //void  Proc_Start(void);
  61 //tThread       *Proc_GetCurThread(void);
  62  int    Proc_NewKThread(void (*Fcn)(void*), void *Data);
  63 // int  Proc_Clone(Uint *Err, Uint Flags);
  64 // int  Proc_SpawnWorker(void);
  65 Uint    Proc_MakeUserStack(void);
  66 //void  Proc_StartUser(Uint Entrypoint, Uint *Bases, int ArgC, char **ArgV, char **EnvP, int DataSize);
  67 void    Proc_StartProcess(Uint16 SS, Uint Stack, Uint Flags, Uint16 CS, Uint IP) NORETURN;
  68  int    Proc_Demote(Uint *Err, int Dest, tRegs *Regs);
  69 //void  Proc_CallFaultHandler(tThread *Thread);
  70 //void  Proc_DumpThreadCPUState(tThread *Thread);
  71 void    Proc_Scheduler(int CPU, Uint RSP, Uint RIP);
  72
  73 // === GLOBALS ===
  74 //!\brief Used by desctab.asm in SyscallStub
  75 const int ci_offsetof_tThread_KernelStack = offsetof(tThread, KernelStack);
  76 // --- Multiprocessing ---
  77 #if USE_MP
  78 volatile int    giNumInitingCPUs = 0;
  79 tMPInfo *gMPFloatPtr = NULL;
  80 tAPIC   *gpMP_LocalAPIC = NULL;
  81 Uint8   gaAPIC_to_CPU[256] = {0};
  82 #endif
  83 tCPU    gaCPUs[MAX_CPUS];
  84 tTSS    *gTSSs = NULL;
  85 tTSS    gTSS0 = {0};
  86 // --- Error Recovery ---
  87 Uint32  gaDoubleFaultStack[1024];
  88
  89 // === CODE ===
  90 /**
  91  * \fn void ArchThreads_Init(void)
  92  * \brief Starts the process scheduler
  93  */
  94 void ArchThreads_Init(void)
  95 {
  96         Uint    pos = 0;
  97
  98         #if USE_MP
  99         tMPTable        *mptable;
 100
 101         // Mark BSP as active
 102         gaCPUs[0].State = 2;
 103
 104         // -- Initialise Multiprocessing
 105         // Find MP Floating Table
 106         // - EBDA/Last 1Kib (640KiB)
 107         for(pos = KERNEL_BASE|0x9F000; pos < (KERNEL_BASE|0xA0000); pos += 16) {
 108                 if( *(Uint*)(pos) == MPPTR_IDENT ) {
 109                         Log("Possible %p", pos);
 110                         if( ByteSum((void*)pos, sizeof(tMPInfo)) != 0 ) continue;
 111                         gMPFloatPtr = (void*)pos;
 112                         break;
 113                 }
 114         }
 115         // - Last KiB (512KiB base mem)
 116         if(!gMPFloatPtr) {
 117                 for(pos = KERNEL_BASE|0x7F000; pos < (KERNEL_BASE|0x80000); pos += 16) {
 118                         if( *(Uint*)(pos) == MPPTR_IDENT ) {
 119                                 Log("Possible %p", pos);
 120                                 if( ByteSum((void*)pos, sizeof(tMPInfo)) != 0 ) continue;
 121                                 gMPFloatPtr = (void*)pos;
 122                                 break;
 123                         }
 124                 }
 125         }
 126         // - BIOS ROM
 127         if(!gMPFloatPtr) {
 128                 for(pos = KERNEL_BASE|0xE0000; pos < (KERNEL_BASE|0x100000); pos += 16) {
 129                         if( *(Uint*)(pos) == MPPTR_IDENT ) {
 130                                 Log("Possible %p", pos);
 131                                 if( ByteSum((void*)pos, sizeof(tMPInfo)) != 0 ) continue;
 132                                 gMPFloatPtr = (void*)pos;
 133                                 break;
 134                         }
 135                 }
 136         }
 137
 138         // If the MP Table Exists, parse it
 139         if(gMPFloatPtr)
 140         {
 141                  int    i;
 142                 tMPTable_Ent    *ents;
 143                 Log("gMPFloatPtr = %p", gMPFloatPtr);
 144                 Log("*gMPFloatPtr = {");
 145                 Log("\t.Sig = 0x%08x", gMPFloatPtr->Sig);
 146                 Log("\t.MPConfig = 0x%08x", gMPFloatPtr->MPConfig);
 147                 Log("\t.Length = 0x%02x", gMPFloatPtr->Length);
 148                 Log("\t.Version = 0x%02x", gMPFloatPtr->Version);
 149                 Log("\t.Checksum = 0x%02x", gMPFloatPtr->Checksum);
 150                 Log("\t.Features = [0x%02x,0x%02x,0x%02x,0x%02x,0x%02x]",
 151                         gMPFloatPtr->Features[0],       gMPFloatPtr->Features[1],
 152                         gMPFloatPtr->Features[2],       gMPFloatPtr->Features[3],
 153                         gMPFloatPtr->Features[4]
 154                         );
 155                 Log("}");
 156
 157                 mptable = (void*)( KERNEL_BASE|gMPFloatPtr->MPConfig );
 158                 Log("mptable = %p", mptable);
 159                 Log("*mptable = {");
 160                 Log("\t.Sig = 0x%08x", mptable->Sig);
 161                 Log("\t.BaseTableLength = 0x%04x", mptable->BaseTableLength);
 162                 Log("\t.SpecRev = 0x%02x", mptable->SpecRev);
 163                 Log("\t.Checksum = 0x%02x", mptable->Checksum);
 164                 Log("\t.OEMID = '%8c'", mptable->OemID);
 165                 Log("\t.ProductID = '%8c'", mptable->ProductID);
 166                 Log("\t.OEMTablePtr = %p'", mptable->OEMTablePtr);
 167                 Log("\t.OEMTableSize = 0x%04x", mptable->OEMTableSize);
 168                 Log("\t.EntryCount = 0x%04x", mptable->EntryCount);
 169                 Log("\t.LocalAPICMemMap = 0x%08x", mptable->LocalAPICMemMap);
 170                 Log("\t.ExtendedTableLen = 0x%04x", mptable->ExtendedTableLen);
 171                 Log("\t.ExtendedTableChecksum = 0x%02x", mptable->ExtendedTableChecksum);
 172                 Log("}");
 173
 174                 gpMP_LocalAPIC = (void*)MM_MapHWPage(mptable->LocalAPICMemMap, 1);
 175
 176                 ents = mptable->Entries;
 177                 giNumCPUs = 0;
 178
 179                 for( i = 0; i < mptable->EntryCount; i ++ )
 180                 {
 181                          int    entSize = 0;
 182                         switch( ents->Type )
 183                         {
 184                         case 0: // Processor
 185                                 entSize = 20;
 186                                 Log("%i: Processor", i);
 187                                 Log("\t.APICID = %i", ents->Proc.APICID);
 188                                 Log("\t.APICVer = 0x%02x", ents->Proc.APICVer);
 189                                 Log("\t.CPUFlags = 0x%02x", ents->Proc.CPUFlags);
 190                                 Log("\t.CPUSignature = 0x%08x", ents->Proc.CPUSignature);
 191                                 Log("\t.FeatureFlags = 0x%08x", ents->Proc.FeatureFlags);
 192
 193
 194                                 if( !(ents->Proc.CPUFlags & 1) ) {
 195                                         Log("DISABLED");
 196                                         break;
 197                                 }
 198
 199                                 // Check if there is too many processors
 200                                 if(giNumCPUs >= MAX_CPUS) {
 201                                         giNumCPUs ++;   // If `giNumCPUs` > MAX_CPUS later, it will be clipped
 202                                         break;
 203                                 }
 204
 205                                 // Initialise CPU Info
 206                                 gaAPIC_to_CPU[ents->Proc.APICID] = giNumCPUs;
 207                                 gaCPUs[giNumCPUs].APICID = ents->Proc.APICID;
 208                                 gaCPUs[giNumCPUs].State = 0;
 209                                 giNumCPUs ++;
 210
 211                                 // Send IPI
 212                                 if( !(ents->Proc.CPUFlags & 2) )
 213                                 {
 214                                         MP_StartAP( giNumCPUs-1 );
 215                                 }
 216
 217                                 break;
 218                         case 1: // Bus
 219                                 entSize = 8;
 220                                 Log("%i: Bus", i);
 221                                 Log("\t.ID = %i", ents->Bus.ID);
 222                                 Log("\t.TypeString = '%6c'", ents->Bus.TypeString);
 223                                 break;
 224                         case 2: // I/O APIC
 225                                 entSize = 8;
 226                                 Log("%i: I/O APIC", i);
 227                                 Log("\t.ID = %i", ents->IOAPIC.ID);
 228                                 Log("\t.Version = 0x%02x", ents->IOAPIC.Version);
 229                                 Log("\t.Flags = 0x%02x", ents->IOAPIC.Flags);
 230                                 Log("\t.Addr = 0x%08x", ents->IOAPIC.Addr);
 231                                 break;
 232                         case 3: // I/O Interrupt Assignment
 233                                 entSize = 8;
 234                                 Log("%i: I/O Interrupt Assignment", i);
 235                                 Log("\t.IntType = %i", ents->IOInt.IntType);
 236                                 Log("\t.Flags = 0x%04x", ents->IOInt.Flags);
 237                                 Log("\t.SourceBusID = 0x%02x", ents->IOInt.SourceBusID);
 238                                 Log("\t.SourceBusIRQ = 0x%02x", ents->IOInt.SourceBusIRQ);
 239                                 Log("\t.DestAPICID = 0x%02x", ents->IOInt.DestAPICID);
 240                                 Log("\t.DestAPICIRQ = 0x%02x", ents->IOInt.DestAPICIRQ);
 241                                 break;
 242                         case 4: // Local Interrupt Assignment
 243                                 entSize = 8;
 244                                 Log("%i: Local Interrupt Assignment", i);
 245                                 Log("\t.IntType = %i", ents->LocalInt.IntType);
 246                                 Log("\t.Flags = 0x%04x", ents->LocalInt.Flags);
 247                                 Log("\t.SourceBusID = 0x%02x", ents->LocalInt.SourceBusID);
 248                                 Log("\t.SourceBusIRQ = 0x%02x", ents->LocalInt.SourceBusIRQ);
 249                                 Log("\t.DestLocalAPICID = 0x%02x", ents->LocalInt.DestLocalAPICID);
 250                                 Log("\t.DestLocalAPICIRQ = 0x%02x", ents->LocalInt.DestLocalAPICIRQ);
 251                                 break;
 252                         default:
 253                                 Log("%i: Unknown (%i)", i, ents->Type);
 254                                 break;
 255                         }
 256                         ents = (void*)( (Uint)ents + entSize );
 257                 }
 258
 259                 if( giNumCPUs > MAX_CPUS ) {
 260                         Warning("Too many CPUs detected (%i), only using %i of them", giNumCPUs, MAX_CPUS);
 261                         giNumCPUs = MAX_CPUS;
 262                 }
 263
 264                 while( giNumInitingCPUs )
 265                         MM_FinishVirtualInit();
 266
 267                 Panic("Uh oh... MP Table Parsing is unimplemented\n");
 268         }
 269         else {
 270                 Log("No MP Table was found, assuming uniprocessor\n");
 271                 giNumCPUs = 1;
 272                 gTSSs = &gTSS0;
 273         }
 274         #else
 275         giNumCPUs = 1;
 276         gTSSs = &gTSS0;
 277         MM_FinishVirtualInit();
 278         #endif
 279
 280         #if USE_MP
 281         // Initialise Normal TSS(s)
 282         for(pos=0;pos<giNumCPUs;pos++)
 283         {
 284         #else
 285         pos = 0;
 286         #endif
 287                 gTSSs[pos].RSP0 = 0;    // Set properly by scheduler
 288                 gGDT[7+pos*2].LimitLow = sizeof(tTSS) & 0xFFFF;
 289                 gGDT[7+pos*2].BaseLow = ((Uint)(&gTSSs[pos])) & 0xFFFF;
 290                 gGDT[7+pos*2].BaseMid = ((Uint)(&gTSSs[pos])) >> 16;
 291                 gGDT[7+pos*2].BaseHi = ((Uint)(&gTSSs[pos])) >> 24;
 292                 gGDT[7+pos*2+1].DWord[0] = ((Uint)(&gTSSs[pos])) >> 32;
 293         #if USE_MP
 294         }
 295         for(pos=0;pos<giNumCPUs;pos++) {
 296                 __asm__ __volatile__ ("ltr %%ax"::"a"(0x38+pos*16));
 297         }
 298         #else
 299         __asm__ __volatile__ ("ltr %%ax"::"a"(0x38));
 300         #endif
 301
 302         // Set Debug registers
 303         __asm__ __volatile__ ("mov %0, %%db0" : : "r"(&gThreadZero));
 304         __asm__ __volatile__ ("mov %%rax, %%db1" : : "a"(0));
 305
 306         gaCPUs[0].Current = &gThreadZero;
 307
 308         gThreadZero.MemState.CR3 = (Uint)gInitialPML4 - KERNEL_BASE;
 309         gThreadZero.CurCPU = 0;
 310         gThreadZero.KernelStack = 0xFFFFA00000000000 + KERNEL_STACK_SIZE;
 311
 312         // Set timer frequency
 313         outb(0x43, 0x34);       // Set Channel 0, Low/High, Rate Generator
 314         outb(0x40, PIT_TIMER_DIVISOR&0xFF);     // Low Byte of Divisor
 315         outb(0x40, (PIT_TIMER_DIVISOR>>8)&0xFF);        // High Byte
 316
 317         // Create Per-Process Data Block
 318         if( !MM_Allocate(MM_PPD_CFG) )
 319         {
 320                 Warning("Oh, hell, Unable to allocate PPD for Thread#0");
 321         }
 322
 323         Log_Log("Proc", "Multithreading initialised");
 324 }
 325
 326 #if USE_MP
 327 void MP_StartAP(int CPU)
 328 {
 329         Log("Starting AP %i (APIC %i)", CPU, gaCPUs[CPU].APICID);
 330         // Set location of AP startup code and mark for a warm restart
 331         *(Uint16*)(KERNEL_BASE|0x467) = (Uint)&APStartup - (KERNEL_BASE|0xFFFF0);
 332         *(Uint16*)(KERNEL_BASE|0x469) = 0xFFFF;
 333         outb(0x70, 0x0F);       outb(0x71, 0x0A);       // Warm Reset
 334         MP_SendIPI(gaCPUs[CPU].APICID, 0, 5);
 335         giNumInitingCPUs ++;
 336 }
 337
 338 void MP_SendIPI(Uint8 APICID, int Vector, int DeliveryMode)
 339 {
 340         Uint32  addr = (Uint)gpMP_LocalAPIC + 0x300;
 341         Uint32  val;
 342
 343         // High
 344         val = (Uint)APICID << 24;
 345         Log("*%p = 0x%08x", addr+0x10, val);
 346         *(Uint32*)(addr+0x10) = val;
 347         // Low (and send)
 348         val = ((DeliveryMode & 7) << 8) | (Vector & 0xFF);
 349         Log("*%p = 0x%08x", addr, val);
 350         *(Uint32*)addr = val;
 351 }
 352 #endif
 353
 354 /**
 355  * \brief Idle task
 356  */
 357 void Proc_IdleTask(void *ptr)
 358 {
 359         tCPU    *cpu = ptr;
 360         cpu->IdleThread = Proc_GetCurThread();
 361         cpu->IdleThread->ThreadName = (char*)"Idle Thread";
 362         Threads_SetPriority( cpu->IdleThread, -1 );     // Never called randomly
 363         cpu->IdleThread->Quantum = 1;   // 1 slice quantum
 364         for(;;) HALT(); // Just yeilds
 365 }
 366
 367 /**
 368  * \fn void Proc_Start(void)
 369  * \brief Start process scheduler
 370  */
 371 void Proc_Start(void)
 372 {
 373         #if USE_MP
 374          int    i;
 375         #endif
 376
 377         #if USE_MP
 378         // Start APs
 379         for( i = 0; i < giNumCPUs; i ++ )
 380         {
 381                  int    tid;
 382                 if(i)   gaCPUs[i].Current = NULL;
 383
 384                 Proc_NewKThread(Proc_IdleTask, &gaCPUs[i]);
 385
 386                 // Create Idle Task
 387                 gaCPUs[i].IdleThread = Threads_GetThread(tid);
 388
 389
 390                 // Start the AP
 391                 if( i != giProc_BootProcessorID ) {
 392                         MP_StartAP( i );
 393                 }
 394         }
 395
 396         // BSP still should run the current task
 397         gaCPUs[0].Current = &gThreadZero;
 398
 399         // Start interrupts and wait for APs to come up
 400         Log("Waiting for APs to come up\n");
 401         __asm__ __volatile__ ("sti");
 402         while( giNumInitingCPUs )       __asm__ __volatile__ ("hlt");
 403         #else
 404         Proc_NewKThread(Proc_IdleTask, &gaCPUs[0]);
 405
 406         // Start Interrupts (and hence scheduler)
 407         __asm__ __volatile__("sti");
 408         #endif
 409         MM_FinishVirtualInit();
 410         Log("Multithreading started");
 411 }
 412
 413 /**
 414  * \fn tThread *Proc_GetCurThread(void)
 415  * \brief Gets the current thread
 416  */
 417 tThread *Proc_GetCurThread(void)
 418 {
 419         #if USE_MP
 420         return gaCPUs[ GetCPUNum() ].Current;
 421         #else
 422         return gaCPUs[ 0 ].Current;
 423         #endif
 424 }
 425
 426 int Proc_NewKThread(void (*Fcn)(void*), void *Data)
 427 {
 428         Uint    rsp;
 429         tThread *newThread, *cur;
 430
 431         cur = Proc_GetCurThread();
 432         newThread = Threads_CloneTCB(NULL, 0);
 433         if(!newThread)  return -1;
 434
 435         // Set CR3
 436         newThread->MemState.CR3 = cur->MemState.CR3;
 437
 438         // Create new KStack
 439         newThread->KernelStack = MM_NewKStack();
 440         // Check for errors
 441         if(newThread->KernelStack == 0) {
 442                 free(newThread);
 443                 return -1;
 444         }
 445
 446         rsp = newThread->KernelStack;
 447         *(Uint*)(rsp-=8) = (Uint)Data;  // Data (shadowed)
 448         *(Uint*)(rsp-=8) = 1;   // Number of params
 449         *(Uint*)(rsp-=8) = (Uint)Fcn;   // Function to call
 450         *(Uint*)(rsp-=8) = (Uint)newThread;     // Thread ID
 451
 452         newThread->SavedState.RSP = rsp;
 453         newThread->SavedState.RIP = (Uint)&NewTaskHeader;
 454         Log("New (KThread) %p, rsp = %p\n", newThread->SavedState.RIP, newThread->SavedState.RSP);
 455
 456 //      MAGIC_BREAK();
 457         Threads_AddActive(newThread);
 458
 459         return newThread->TID;
 460 }
 461
 462 /**
 463  * \fn int Proc_Clone(Uint Flags)
 464  * \brief Clone the current process
 465  */
 466 int Proc_Clone(Uint Flags)
 467 {
 468         tThread *newThread, *cur = Proc_GetCurThread();
 469         Uint    rip;
 470         Uint    _savedregs[16+1];
 471
 472         newThread = Threads_CloneTCB(NULL, Flags);
 473         if(!newThread)  return -1;
 474
 475         // Save core machine state
 476         rip = SaveState(&newThread->SavedState.RSP, &_savedregs[0]);
 477         if(rip == 0) {
 478                 outb(0x20, 0x20);       // ACK Timer and return as child
 479                 __asm__ __volatile__ ("sti");
 480                 return 0;
 481         }
 482
 483         // Initialise Memory Space (New Addr space or kernel stack)
 484         if(Flags & CLONE_VM) {
 485                 newThread->MemState.CR3 = MM_Clone();
 486                 newThread->KernelStack = cur->KernelStack;
 487         }
 488         else {
 489                 Log_Error("Proc", "Proc_Clone: Don't leave CLONE_VM unset, use Proc_NewKThread instead");
 490                 return -1;
 491         }
 492
 493         Log("New (Clone) %p, rsp = %p", rip, newThread->SavedState.RSP);
 494         {
 495                 Uint cr3;
 496                 __asm__ __volatile__ ("mov %%cr3, %0" : "=r" (cr3));
 497                 Log(" CR3 = %x, PADDR(RSP) = %x", cr3, MM_GetPhysAddr(newThread->SavedState.RSP));
 498         }
 499
 500         // Set EIP as parent
 501         newThread->SavedState.RIP = rip;
 502
 503         // Lock list and add to active
 504         Threads_AddActive(newThread);
 505
 506         return newThread->TID;
 507 }
 508
 509 /**
 510  * \fn int Proc_SpawnWorker(void)
 511  * \brief Spawns a new worker thread
 512  */
 513 int Proc_SpawnWorker(void (*Fcn)(void*), void *Data)
 514 {
 515         tThread *new, *cur;
 516         Uint    stack_contents[4];
 517
 518         cur = Proc_GetCurThread();
 519
 520         // Create new thread
 521         new = malloc( sizeof(tThread) );
 522         if(!new) {
 523                 Warning("Proc_SpawnWorker - Out of heap space!\n");
 524                 return -1;
 525         }
 526         memcpy(new, &gThreadZero, sizeof(tThread));
 527         // Set Thread ID
 528         new->TID = giNextTID++;
 529
 530         // Create the stack contents
 531         stack_contents[3] = (Uint)Data;
 532         stack_contents[2] = 1;
 533         stack_contents[1] = (Uint)Fcn;
 534         stack_contents[0] = (Uint)new;
 535
 536         // Create a new worker stack (in PID0's address space)
 537         // The stack is relocated by this code
 538         new->KernelStack = MM_NewWorkerStack(stack_contents, sizeof(stack_contents));
 539
 540         new->SavedState.RSP = new->KernelStack - sizeof(stack_contents);
 541         new->SavedState.RIP = (Uint)&NewTaskHeader;
 542
 543         Log("New (Worker) %p, rsp = %p\n", new->SavedState.RIP, new->SavedState.RSP);
 544
 545         // Mark as active
 546         new->Status = THREAD_STAT_PREINIT;
 547         Threads_AddActive( new );
 548
 549         return new->TID;
 550 }
 551
 552 /**
 553  * \fn Uint Proc_MakeUserStack(void)
 554  * \brief Creates a new user stack
 555  */
 556 Uint Proc_MakeUserStack(void)
 557 {
 558          int    i;
 559         Uint    base = USER_STACK_TOP - USER_STACK_SZ;
 560
 561         // Check Prospective Space
 562         for( i = USER_STACK_SZ >> 12; i--; )
 563                 if( MM_GetPhysAddr( base + (i<<12) ) != 0 )
 564                         break;
 565
 566         if(i != -1)     return 0;
 567
 568         // Allocate Stack - Allocate incrementally to clean up MM_Dump output
 569         for( i = 0; i < USER_STACK_SZ/0x1000; i++ )
 570         {
 571                 if( !MM_Allocate( base + (i<<12) ) )
 572                 {
 573                         // Error
 574                         Log_Error("Proc", "Unable to allocate user stack (%i pages requested)", USER_STACK_SZ/0x1000);
 575                         while( i -- )
 576                                 MM_Deallocate( base + (i<<12) );
 577                         return 0;
 578                 }
 579         }
 580
 581         return base + USER_STACK_SZ;
 582 }
 583
 584
 585 /**
 586  * \fn void Proc_StartUser(Uint Entrypoint, Uint *Bases, int ArgC, char **ArgV, char **EnvP, int DataSize)
 587  * \brief Starts a user task
 588  */
 589 void Proc_StartUser(Uint Entrypoint, Uint *Bases, int ArgC, char **ArgV, char **EnvP, int DataSize)
 590 {
 591         Uint    *stack = (void*)Proc_MakeUserStack();
 592          int    i;
 593         Uint    delta;
 594         Uint16  ss, cs;
 595
 596         LOG("stack = 0x%x", stack);
 597
 598         // Copy Arguments
 599         stack = (void*)( (Uint)stack - DataSize );
 600         memcpy( stack, ArgV, DataSize );
 601
 602         // Adjust Arguments and environment
 603         delta = (Uint)stack - (Uint)ArgV;
 604         ArgV = (char**)stack;
 605         for( i = 0; ArgV[i]; i++ )      ArgV[i] += delta;
 606         i ++;
 607         EnvP = &ArgV[i];
 608         for( i = 0; EnvP[i]; i++ )      EnvP[i] += delta;
 609
 610         // User Mode Segments
 611         // 0x2B = 64-bit
 612         ss = 0x23;      cs = 0x2B;
 613
 614         // Arguments
 615         *--stack = (Uint)EnvP;
 616         *--stack = (Uint)ArgV;
 617         *--stack = (Uint)ArgC;
 618         while(*Bases)
 619                 *--stack = *Bases++;
 620
 621         Proc_StartProcess(ss, (Uint)stack, 0x202, cs, Entrypoint);
 622 }
 623
 624 void Proc_StartProcess(Uint16 SS, Uint Stack, Uint Flags, Uint16 CS, Uint IP)
 625 {
 626         if( !(CS == 0x1B || CS == 0x2B) || SS != 0x23 ) {
 627                 Log_Error("Proc", "Proc_StartProcess: CS / SS are not valid (%x, %x)",
 628                         CS, SS);
 629                 Threads_Exit(0, -1);
 630         }
 631         Log("Proc_StartProcess: (SS=%x, Stack=%p, Flags=%x, CS=%x, IP=%p)",
 632                 SS, Stack, Flags, CS, IP);
 633         if(CS == 0x1B)
 634         {
 635                 // 32-bit return
 636                 __asm__ __volatile__ (
 637                         "mov %0, %%rsp;\n\t"    // Set stack pointer
 638                         "mov %2, %%r11;\n\t"    // Set RFLAGS
 639                         "sysret;\n\t"
 640                         : : "r" (Stack), "c" (IP), "r" (Flags)
 641                         );
 642         }
 643         else
 644         {
 645                 // 64-bit return
 646                 __asm__ __volatile__ (
 647                         "mov %0, %%rsp;\n\t"    // Set stack pointer
 648                         "mov %2, %%r11;\n\t"    // Set RFLAGS
 649                         "sysretq;\n\t"
 650                         : : "r" (Stack), "c" (IP), "r" (Flags)
 651                         );
 652         }
 653         for(;;);
 654 }
 655
 656 /**
 657  * \fn int Proc_Demote(Uint *Err, int Dest, tRegs *Regs)
 658  * \brief Demotes a process to a lower permission level
 659  * \param Err   Pointer to user's errno
 660  * \param Dest  New Permission Level
 661  * \param Regs  Pointer to user's register structure
 662  */
 663 int Proc_Demote(Uint *Err, int Dest, tRegs *Regs)
 664 {
 665          int    cpl = Regs->CS & 3;
 666         // Sanity Check
 667         if(Dest > 3 || Dest < 0) {
 668                 *Err = -EINVAL;
 669                 return -1;
 670         }
 671
 672         // Permission Check
 673         if(cpl > Dest) {
 674                 *Err = -EACCES;
 675                 return -1;
 676         }
 677
 678         // Change the Segment Registers
 679         Regs->CS = (((Dest+1)<<4) | Dest) - 8;
 680         Regs->SS = ((Dest+1)<<4) | Dest;
 681
 682         return 0;
 683 }
 684
 685 /**
 686  * \brief Calls a signal handler in user mode
 687  * \note Used for signals
 688  */
 689 void Proc_CallFaultHandler(tThread *Thread)
 690 {
 691         // Rewinds the stack and calls the user function
 692         // Never returns
 693         __asm__ __volatile__ ("mov %0, %%rbp;\n\tcall Proc_ReturnToUser" :: "r"(Thread->FaultHandler));
 694         for(;;);
 695 }
 696
 697 void Proc_DumpThreadCPUState(tThread *Thread)
 698 {
 699         Log("  At %04x:%016llx", Thread->SavedState.UserCS, Thread->SavedState.UserRIP);
 700 }
 701
 702 /**
 703  * \fn void Proc_Scheduler(int CPU)
 704  * \brief Swap current thread and clears dead threads
 705  */
 706 void Proc_Scheduler(int CPU, Uint RSP, Uint RIP)
 707 {
 708         tThread *thread;
 709
 710         if( CPU == 0 )
 711                 Time_UpdateTimestamp();
 712
 713         // If the spinlock is set, let it complete
 714         if(IS_LOCKED(&glThreadListLock))        return;
 715
 716         // Get current thread
 717         thread = gaCPUs[CPU].Current;
 718
 719         if( thread )
 720         {
 721                 tRegs   *regs;
 722                 // Reduce remaining quantum and continue timeslice if non-zero
 723                 if(thread->Remaining--) return;
 724                 // Reset quantum for next call
 725                 thread->Remaining = thread->Quantum;
 726
 727                 // Save machine state
 728                 thread->SavedState.RSP = RSP;
 729                 thread->SavedState.RIP = RIP;
 730
 731 //              LogF("\nSaved %i, RIP = %p, rsp = %p\n", thread->TID, thread->SavedState.RIP, thread->SavedState.RSP);
 732
 733                 // TODO: Make this more stable somehow
 734                 {
 735                         regs = (tRegs*)(RSP+(1)*8);     // CurThread
 736                         thread->SavedState.UserCS = regs->CS;
 737                         thread->SavedState.UserRIP = regs->RIP;
 738                 }
 739         }
 740
 741         #if BREAK_ON_SWITCH
 742         {
 743         tThread *oldthread = thread;
 744         #endif
 745
 746         // Get next thread
 747         thread = Threads_GetNextToRun(CPU, thread);
 748
 749         // Error Check
 750         if(thread == NULL) {
 751                 thread = gaCPUs[CPU].IdleThread;
 752                 //Warning("Hmm... Threads_GetNextToRun returned NULL, I don't think this should happen.\n");
 753 //              LogF("Zzzzz.\n");
 754                 //return;
 755         }
 756         if(thread == NULL ) {
 757                 return ;
 758         }
 759         #if BREAK_ON_SWITCH
 760         if( thread != oldthread ) {
 761                 MAGIC_BREAK();
 762         }
 763         }
 764         #endif
 765
 766         #if DEBUG_TRACE_SWITCH
 767         LogF("\nSwitching to task %i, CR3 = 0x%x, RIP = %p, RSP = %p\n",
 768                 thread->TID,
 769                 thread->MemState.CR3,
 770                 thread->SavedState.RIP,
 771                 thread->SavedState.RSP
 772                 );
 773         #endif
 774
 775
 776         if(CPU > MAX_CPUS)
 777                 LogF("CPU = %i", CPU);
 778         // Set current thread
 779         gaCPUs[CPU].Current = thread;
 780
 781         // Update Kernel Stack pointer
 782         gTSSs[CPU].RSP0 = thread->KernelStack-4;
 783
 784         // Switch threads
 785         __asm__ __volatile__ (
 786                 "mov %2, %%cr3\n\t"
 787                 "mov %0, %%rsp\n\t"     // Restore RSP
 788                 "invlpg (%%rsp)\n\t"
 789                 "invlpg 0x1000(%%rsp)\n\t"
 790                 "invlpg -0x1000(%%rsp)\n\t"
 791                 "xor %%eax, %%eax\n\t"
 792                 "jmp *%1" : :   // And return to where we saved state (Proc_Clone or Proc_Scheduler)
 793                 "r"(thread->SavedState.RSP), "r"(thread->SavedState.RIP),
 794                 "r"(thread->MemState.CR3)
 795                 );
 796         for(;;);        // Shouldn't reach here
 797 }
 798
 799 // === EXPORTS ===
 800 EXPORT(Proc_SpawnWorker);