Kernel/arch/x86/proc.c

   1 /*
   2  * AcessOS Microkernel Version
   3  * proc.c
   4  */
   5 #include <acess.h>
   6 #include <threads.h>
   7 #include <proc.h>
   8 #include <desctab.h>
   9 #include <mm_virt.h>
  10 #include <errno.h>
  11 #if USE_MP
  12 # include <mp.h>
  13 #endif
  14 #include <hal_proc.h>
  15
  16 // === FLAGS ===
  17 #define DEBUG_TRACE_SWITCH      0
  18 #define DEBUG_DISABLE_DOUBLEFAULT       1
  19 #define DEBUG_VERY_SLOW_SWITCH  0
  20
  21 // === CONSTANTS ===
  22 #define SWITCH_MAGIC    0xFF5317C8      // FF SWITCH - There is no code in this area
  23 // Base is 1193182
  24 #define TIMER_BASE      1193182
  25 #if DEBUG_VERY_SLOW_PERIOD
  26 # define TIMER_DIVISOR  1193    //~10Hz switch, with 10 quantum = 1s per thread
  27 #else
  28 # define TIMER_DIVISOR  11932   //~100Hz
  29 #endif
  30
  31 // === TYPES ===
  32 #if USE_MP
  33 typedef struct sCPU
  34 {
  35         Uint8   APICID;
  36         Uint8   State;  // 0: Unavaliable, 1: Idle, 2: Active
  37         Uint16  Resvd;
  38         tThread *Current;
  39         tThread *IdleThread;
  40 }       tCPU;
  41 #endif
  42
  43 // === IMPORTS ===
  44 extern tGDT     gGDT[];
  45 extern tIDT     gIDT[];
  46 extern void APWait(void);       // 16-bit AP pause code
  47 extern void APStartup(void);    // 16-bit AP startup code
  48 extern Uint     GetEIP(void);   // start.asm
  49 extern Uint     GetEIP_Sched(void);     // proc.asm
  50 extern Uint32   gaInitPageDir[1024];    // start.asm
  51 extern char     Kernel_Stack_Top[];
  52 extern tShortSpinlock   glThreadListLock;
  53 extern int      giNumCPUs;
  54 extern int      giNextTID;
  55 extern tThread  gThreadZero;
  56 extern void     Isr8(void);     // Double Fault
  57 extern void     Proc_ReturnToUser(tVAddr Handler, Uint Argument, tVAddr KernelStack);
  58 extern void     scheduler_return;       // Return address in SchedulerBase
  59 extern void     IRQCommon;      // Common IRQ handler code
  60 extern void     IRQCommon_handled;      // IRQCommon call return location
  61 extern void     GetEIP_Sched_ret;       // GetEIP call return location
  62
  63 // === PROTOTYPES ===
  64 //void  ArchThreads_Init(void);
  65 #if USE_MP
  66 void    MP_StartAP(int CPU);
  67 void    MP_SendIPI(Uint8 APICID, int Vector, int DeliveryMode);
  68 #endif
  69 //void  Proc_Start(void);
  70 //tThread       *Proc_GetCurThread(void);
  71 void    Proc_ChangeStack(void);
  72 // int  Proc_Clone(Uint *Err, Uint Flags);
  73 Uint    Proc_MakeUserStack(void);
  74 //void  Proc_StartUser(Uint Entrypoint, Uint *Bases, int ArgC, char **ArgV, char **EnvP, int DataSize);
  75 void    Proc_StartProcess(Uint16 SS, Uint Stack, Uint Flags, Uint16 CS, Uint IP);
  76  int    Proc_Demote(Uint *Err, int Dest, tRegs *Regs);
  77 //void  Proc_CallFaultHandler(tThread *Thread);
  78 //void  Proc_DumpThreadCPUState(tThread *Thread);
  79 void    Proc_Scheduler(int CPU);
  80
  81 // === GLOBALS ===
  82 // --- Multiprocessing ---
  83 #if USE_MP
  84 volatile int    giNumInitingCPUs = 0;
  85 tMPInfo *gMPFloatPtr = NULL;
  86 volatile Uint32 giMP_TimerCount;        // Start Count for Local APIC Timer
  87 tAPIC   *gpMP_LocalAPIC = NULL;
  88 Uint8   gaAPIC_to_CPU[256] = {0};
  89 tCPU    gaCPUs[MAX_CPUS];
  90 tTSS    gaTSSs[MAX_CPUS];       // TSS Array
  91  int    giProc_BootProcessorID = 0;
  92 #else
  93 tThread *gCurrentThread = NULL;
  94 tThread *gpIdleThread = NULL;
  95 #endif
  96 tTSS    *gTSSs = NULL;  // Pointer to TSS array
  97 tTSS    gTSS0 = {0};
  98 // --- Error Recovery ---
  99 char    gaDoubleFaultStack[1024] __attribute__ ((section(".padata")));
 100 tTSS    gDoubleFault_TSS = {
 101         .ESP0 = (Uint)&gaDoubleFaultStack[1024],
 102         .SS0 = 0x10,
 103         .CR3 = (Uint)gaInitPageDir - KERNEL_BASE,
 104         .EIP = (Uint)Isr8,
 105         .ESP = (Uint)&gaDoubleFaultStack[1024],
 106         .CS = 0x08,     .SS = 0x10,
 107         .DS = 0x10,     .ES = 0x10,
 108         .FS = 0x10,     .GS = 0x10,
 109 };
 110
 111 // === CODE ===
 112 /**
 113  * \fn void ArchThreads_Init(void)
 114  * \brief Starts the process scheduler
 115  */
 116 void ArchThreads_Init(void)
 117 {
 118         Uint    pos = 0;
 119
 120         #if USE_MP
 121         tMPTable        *mptable;
 122
 123         // Mark BSP as active
 124         gaCPUs[0].State = 2;
 125
 126         // -- Initialise Multiprocessing
 127         // Find MP Floating Table
 128         // - EBDA/Last 1Kib (640KiB)
 129         for(pos = KERNEL_BASE|0x9F000; pos < (KERNEL_BASE|0xA0000); pos += 16) {
 130                 if( *(Uint*)(pos) == MPPTR_IDENT ) {
 131                         Log("Possible %p", pos);
 132                         if( ByteSum((void*)pos, sizeof(tMPInfo)) != 0 ) continue;
 133                         gMPFloatPtr = (void*)pos;
 134                         break;
 135                 }
 136         }
 137         // - Last KiB (512KiB base mem)
 138         if(!gMPFloatPtr) {
 139                 for(pos = KERNEL_BASE|0x7F000; pos < (KERNEL_BASE|0x80000); pos += 16) {
 140                         if( *(Uint*)(pos) == MPPTR_IDENT ) {
 141                                 Log("Possible %p", pos);
 142                                 if( ByteSum((void*)pos, sizeof(tMPInfo)) != 0 ) continue;
 143                                 gMPFloatPtr = (void*)pos;
 144                                 break;
 145                         }
 146                 }
 147         }
 148         // - BIOS ROM
 149         if(!gMPFloatPtr) {
 150                 for(pos = KERNEL_BASE|0xE0000; pos < (KERNEL_BASE|0x100000); pos += 16) {
 151                         if( *(Uint*)(pos) == MPPTR_IDENT ) {
 152                                 Log("Possible %p", pos);
 153                                 if( ByteSum((void*)pos, sizeof(tMPInfo)) != 0 ) continue;
 154                                 gMPFloatPtr = (void*)pos;
 155                                 break;
 156                         }
 157                 }
 158         }
 159
 160         // If the MP Table Exists, parse it
 161         if(gMPFloatPtr)
 162         {
 163                  int    i;
 164                 tMPTable_Ent    *ents;
 165                 Log("gMPFloatPtr = %p", gMPFloatPtr);
 166                 Log("*gMPFloatPtr = {");
 167                 Log("\t.Sig = 0x%08x", gMPFloatPtr->Sig);
 168                 Log("\t.MPConfig = 0x%08x", gMPFloatPtr->MPConfig);
 169                 Log("\t.Length = 0x%02x", gMPFloatPtr->Length);
 170                 Log("\t.Version = 0x%02x", gMPFloatPtr->Version);
 171                 Log("\t.Checksum = 0x%02x", gMPFloatPtr->Checksum);
 172                 Log("\t.Features = [0x%02x,0x%02x,0x%02x,0x%02x,0x%02x]",
 173                         gMPFloatPtr->Features[0],       gMPFloatPtr->Features[1],
 174                         gMPFloatPtr->Features[2],       gMPFloatPtr->Features[3],
 175                         gMPFloatPtr->Features[4]
 176                         );
 177                 Log("}");
 178
 179                 mptable = (void*)( KERNEL_BASE|gMPFloatPtr->MPConfig );
 180                 Log("mptable = %p", mptable);
 181                 Log("*mptable = {");
 182                 Log("\t.Sig = 0x%08x", mptable->Sig);
 183                 Log("\t.BaseTableLength = 0x%04x", mptable->BaseTableLength);
 184                 Log("\t.SpecRev = 0x%02x", mptable->SpecRev);
 185                 Log("\t.Checksum = 0x%02x", mptable->Checksum);
 186                 Log("\t.OEMID = '%8c'", mptable->OemID);
 187                 Log("\t.ProductID = '%8c'", mptable->ProductID);
 188                 Log("\t.OEMTablePtr = %p'", mptable->OEMTablePtr);
 189                 Log("\t.OEMTableSize = 0x%04x", mptable->OEMTableSize);
 190                 Log("\t.EntryCount = 0x%04x", mptable->EntryCount);
 191                 Log("\t.LocalAPICMemMap = 0x%08x", mptable->LocalAPICMemMap);
 192                 Log("\t.ExtendedTableLen = 0x%04x", mptable->ExtendedTableLen);
 193                 Log("\t.ExtendedTableChecksum = 0x%02x", mptable->ExtendedTableChecksum);
 194                 Log("}");
 195
 196                 gpMP_LocalAPIC = (void*)MM_MapHWPages(mptable->LocalAPICMemMap, 1);
 197
 198                 ents = mptable->Entries;
 199                 giNumCPUs = 0;
 200
 201                 for( i = 0; i < mptable->EntryCount; i ++ )
 202                 {
 203                          int    entSize = 0;
 204                         switch( ents->Type )
 205                         {
 206                         case 0: // Processor
 207                                 entSize = 20;
 208                                 Log("%i: Processor", i);
 209                                 Log("\t.APICID = %i", ents->Proc.APICID);
 210                                 Log("\t.APICVer = 0x%02x", ents->Proc.APICVer);
 211                                 Log("\t.CPUFlags = 0x%02x", ents->Proc.CPUFlags);
 212                                 Log("\t.CPUSignature = 0x%08x", ents->Proc.CPUSignature);
 213                                 Log("\t.FeatureFlags = 0x%08x", ents->Proc.FeatureFlags);
 214
 215
 216                                 if( !(ents->Proc.CPUFlags & 1) ) {
 217                                         Log("DISABLED");
 218                                         break;
 219                                 }
 220
 221                                 // Check if there is too many processors
 222                                 if(giNumCPUs >= MAX_CPUS) {
 223                                         giNumCPUs ++;   // If `giNumCPUs` > MAX_CPUS later, it will be clipped
 224                                         break;
 225                                 }
 226
 227                                 // Initialise CPU Info
 228                                 gaAPIC_to_CPU[ents->Proc.APICID] = giNumCPUs;
 229                                 gaCPUs[giNumCPUs].APICID = ents->Proc.APICID;
 230                                 gaCPUs[giNumCPUs].State = 0;
 231                                 giNumCPUs ++;
 232
 233                                 // Set BSP Variable
 234                                 if( ents->Proc.CPUFlags & 2 ) {
 235                                         giProc_BootProcessorID = giNumCPUs-1;
 236                                 }
 237
 238                                 break;
 239
 240                         #if DUMP_MP_TABLES
 241                         case 1: // Bus
 242                                 entSize = 8;
 243                                 Log("%i: Bus", i);
 244                                 Log("\t.ID = %i", ents->Bus.ID);
 245                                 Log("\t.TypeString = '%6C'", ents->Bus.TypeString);
 246                                 break;
 247                         case 2: // I/O APIC
 248                                 entSize = 8;
 249                                 Log("%i: I/O APIC", i);
 250                                 Log("\t.ID = %i", ents->IOAPIC.ID);
 251                                 Log("\t.Version = 0x%02x", ents->IOAPIC.Version);
 252                                 Log("\t.Flags = 0x%02x", ents->IOAPIC.Flags);
 253                                 Log("\t.Addr = 0x%08x", ents->IOAPIC.Addr);
 254                                 break;
 255                         case 3: // I/O Interrupt Assignment
 256                                 entSize = 8;
 257                                 Log("%i: I/O Interrupt Assignment", i);
 258                                 Log("\t.IntType = %i", ents->IOInt.IntType);
 259                                 Log("\t.Flags = 0x%04x", ents->IOInt.Flags);
 260                                 Log("\t.SourceBusID = 0x%02x", ents->IOInt.SourceBusID);
 261                                 Log("\t.SourceBusIRQ = 0x%02x", ents->IOInt.SourceBusIRQ);
 262                                 Log("\t.DestAPICID = 0x%02x", ents->IOInt.DestAPICID);
 263                                 Log("\t.DestAPICIRQ = 0x%02x", ents->IOInt.DestAPICIRQ);
 264                                 break;
 265                         case 4: // Local Interrupt Assignment
 266                                 entSize = 8;
 267                                 Log("%i: Local Interrupt Assignment", i);
 268                                 Log("\t.IntType = %i", ents->LocalInt.IntType);
 269                                 Log("\t.Flags = 0x%04x", ents->LocalInt.Flags);
 270                                 Log("\t.SourceBusID = 0x%02x", ents->LocalInt.SourceBusID);
 271                                 Log("\t.SourceBusIRQ = 0x%02x", ents->LocalInt.SourceBusIRQ);
 272                                 Log("\t.DestLocalAPICID = 0x%02x", ents->LocalInt.DestLocalAPICID);
 273                                 Log("\t.DestLocalAPICIRQ = 0x%02x", ents->LocalInt.DestLocalAPICIRQ);
 274                                 break;
 275                         default:
 276                                 Log("%i: Unknown (%i)", i, ents->Type);
 277                                 break;
 278                         #endif
 279                         }
 280                         ents = (void*)( (Uint)ents + entSize );
 281                 }
 282
 283                 if( giNumCPUs > MAX_CPUS ) {
 284                         Warning("Too many CPUs detected (%i), only using %i of them", giNumCPUs, MAX_CPUS);
 285                         giNumCPUs = MAX_CPUS;
 286                 }
 287                 gTSSs = gaTSSs;
 288         }
 289         else {
 290                 Log("No MP Table was found, assuming uniprocessor\n");
 291                 giNumCPUs = 1;
 292                 gTSSs = &gTSS0;
 293         }
 294         #else
 295         giNumCPUs = 1;
 296         gTSSs = &gTSS0;
 297         #endif
 298
 299         #if !DEBUG_DISABLE_DOUBLEFAULT
 300         // Initialise Double Fault TSS
 301         gGDT[5].BaseLow = (Uint)&gDoubleFault_TSS & 0xFFFF;
 302         gGDT[5].BaseMid = (Uint)&gDoubleFault_TSS >> 16;
 303         gGDT[5].BaseHi = (Uint)&gDoubleFault_TSS >> 24;
 304
 305         // Set double fault IDT to use the new TSS
 306         gIDT[8].OffsetLo = 0;
 307         gIDT[8].CS = 5<<3;
 308         gIDT[8].Flags = 0x8500;
 309         gIDT[8].OffsetHi = 0;
 310         #endif
 311
 312         // Set timer frequency
 313         outb(0x43, 0x34);       // Set Channel 0, Low/High, Rate Generator
 314         outb(0x40, TIMER_DIVISOR&0xFF); // Low Byte of Divisor
 315         outb(0x40, (TIMER_DIVISOR>>8)&0xFF);    // High Byte
 316
 317         Log("Timer Frequency %i.%03i Hz",
 318                 TIMER_BASE/TIMER_DIVISOR,
 319                 ((Uint64)TIMER_BASE*1000/TIMER_DIVISOR)%1000
 320                 );
 321
 322         #if USE_MP
 323         // Get the count setting for APIC timer
 324         Log("Determining APIC Count");
 325         __asm__ __volatile__ ("sti");
 326         while( giMP_TimerCount == 0 )   __asm__ __volatile__ ("hlt");
 327         __asm__ __volatile__ ("cli");
 328         Log("APIC Count %i", giMP_TimerCount);
 329         {
 330                 Uint64  freq = giMP_TimerCount;
 331                 freq /= TIMER_DIVISOR;
 332                 freq *= TIMER_BASE;
 333                 if( (freq /= 1000) < 2*1000)
 334                         Log("Bus Frequency %i KHz", freq);
 335                 else if( (freq /= 1000) < 2*1000)
 336                         Log("Bus Frequency %i MHz", freq);
 337                 else if( (freq /= 1000) < 2*1000)
 338                         Log("Bus Frequency %i GHz", freq);
 339                 else
 340                         Log("Bus Frequency %i THz", freq);
 341         }
 342
 343         // Initialise Normal TSS(s)
 344         for(pos=0;pos<giNumCPUs;pos++)
 345         {
 346         #else
 347         pos = 0;
 348         #endif
 349                 gTSSs[pos].SS0 = 0x10;
 350                 gTSSs[pos].ESP0 = 0;    // Set properly by scheduler
 351                 gGDT[6+pos].BaseLow = ((Uint)(&gTSSs[pos])) & 0xFFFF;
 352                 gGDT[6+pos].BaseMid = ((Uint)(&gTSSs[pos]) >> 16) & 0xFFFF;
 353                 gGDT[6+pos].BaseHi = ((Uint)(&gTSSs[pos])) >> 24;
 354         #if USE_MP
 355         }
 356         #endif
 357
 358         // Load the BSP's TSS
 359         __asm__ __volatile__ ("ltr %%ax"::"a"(0x30));
 360         // Set Current Thread and CPU Number in DR0 and DR1
 361         __asm__ __volatile__ ("mov %0, %%db0"::"r"(&gThreadZero));
 362         __asm__ __volatile__ ("mov %0, %%db1"::"r"(0));
 363
 364         #if USE_MP
 365         gaCPUs[0].Current = &gThreadZero;
 366         #else
 367         gCurrentThread = &gThreadZero;
 368         #endif
 369         gThreadZero.CurCPU = 0;
 370
 371         gThreadZero.MemState.CR3 = (Uint)gaInitPageDir - KERNEL_BASE;
 372
 373         // Create Per-Process Data Block
 374         if( !MM_Allocate(MM_PPD_CFG) )
 375         {
 376                 Panic("OOM - No space for initial Per-Process Config");
 377         }
 378
 379         // Change Stacks
 380         Proc_ChangeStack();
 381 }
 382
 383 #if USE_MP
 384 void MP_StartAP(int CPU)
 385 {
 386         Log("Starting AP %i (APIC %i)", CPU, gaCPUs[CPU].APICID);
 387
 388         // Set location of AP startup code and mark for a warm restart
 389         *(Uint16*)(KERNEL_BASE|0x467) = (Uint)&APWait - (KERNEL_BASE|0xFFFF0);
 390         *(Uint16*)(KERNEL_BASE|0x469) = 0xFFFF;
 391         outb(0x70, 0x0F);       outb(0x71, 0x0A);       // Set warm reset flag
 392         MP_SendIPI(gaCPUs[CPU].APICID, 0, 5);   // Init IPI
 393
 394         // Delay
 395         inb(0x80); inb(0x80); inb(0x80); inb(0x80);
 396
 397         // TODO: Use a better address, preferably registered with the MM
 398         // - MM_AllocDMA mabye?
 399         // Create a far jump
 400         *(Uint8*)(KERNEL_BASE|0x11000) = 0xEA;  // Far JMP
 401         *(Uint16*)(KERNEL_BASE|0x11001) = (Uint)&APStartup - (KERNEL_BASE|0xFFFF0);     // IP
 402         *(Uint16*)(KERNEL_BASE|0x11003) = 0xFFFF;       // CS
 403         // Send a Startup-IPI to make the CPU execute at 0x11000 (which we
 404         // just filled)
 405         MP_SendIPI(gaCPUs[CPU].APICID, 0x11, 6);        // StartupIPI
 406
 407         giNumInitingCPUs ++;
 408 }
 409
 410 /**
 411  * \brief Send an Inter-Processor Interrupt
 412  * \param APICID        Processor's Local APIC ID
 413  * \param Vector        Argument of some kind
 414  * \param DeliveryMode  Type of signal?
 415  */
 416 void MP_SendIPI(Uint8 APICID, int Vector, int DeliveryMode)
 417 {
 418         Uint32  val;
 419
 420         // Hi
 421         val = (Uint)APICID << 24;
 422         Log("*%p = 0x%08x", &gpMP_LocalAPIC->ICR[1], val);
 423         gpMP_LocalAPIC->ICR[1].Val = val;
 424         // Low (and send)
 425         val = ((DeliveryMode & 7) << 8) | (Vector & 0xFF);
 426         Log("*%p = 0x%08x", &gpMP_LocalAPIC->ICR[0], val);
 427         gpMP_LocalAPIC->ICR[0].Val = val;
 428 }
 429 #endif
 430
 431 /**
 432  * \fn void Proc_Start(void)
 433  * \brief Start process scheduler
 434  */
 435 void Proc_Start(void)
 436 {
 437         #if USE_MP
 438          int    i;
 439         #endif
 440
 441         #if USE_MP
 442         // Start APs
 443         for( i = 0; i < giNumCPUs; i ++ )
 444         {
 445                  int    tid;
 446                 if(i)   gaCPUs[i].Current = NULL;
 447
 448                 // Create Idle Task
 449                 if( (tid = Proc_Clone(0, 0)) == 0)
 450                 {
 451                         for(;;) HALT(); // Just yeilds
 452                 }
 453                 gaCPUs[i].IdleThread = Threads_GetThread(tid);
 454                 gaCPUs[i].IdleThread->ThreadName = (char*)"Idle Thread";
 455                 Threads_SetPriority( gaCPUs[i].IdleThread, -1 );        // Never called randomly
 456                 gaCPUs[i].IdleThread->Quantum = 1;      // 1 slice quantum
 457
 458
 459                 // Start the AP
 460                 if( i != giProc_BootProcessorID ) {
 461                         MP_StartAP( i );
 462                 }
 463         }
 464
 465         // BSP still should run the current task
 466         gaCPUs[0].Current = &gThreadZero;
 467
 468         // Start interrupts and wait for APs to come up
 469         Log("Waiting for APs to come up\n");
 470         __asm__ __volatile__ ("sti");
 471         while( giNumInitingCPUs )       __asm__ __volatile__ ("hlt");
 472         #else
 473         // Create Idle Task
 474         if(Proc_Clone(0) == 0)
 475         {
 476                 gpIdleThread = Proc_GetCurThread();
 477                 gpIdleThread->ThreadName = strdup("Idle Thread");
 478                 Threads_SetPriority( gpIdleThread, -1 );        // Never called randomly
 479                 gpIdleThread->Quantum = 1;      // 1 slice quantum
 480                 for(;;) HALT(); // Just yeilds
 481         }
 482
 483         // Set current task
 484         gCurrentThread = &gThreadZero;
 485
 486         // Start Interrupts (and hence scheduler)
 487         __asm__ __volatile__("sti");
 488         #endif
 489         MM_FinishVirtualInit();
 490 }
 491
 492 /**
 493  * \fn tThread *Proc_GetCurThread(void)
 494  * \brief Gets the current thread
 495  */
 496 tThread *Proc_GetCurThread(void)
 497 {
 498         #if USE_MP
 499         return gaCPUs[ GetCPUNum() ].Current;
 500         #else
 501         return gCurrentThread;
 502         #endif
 503 }
 504
 505 /**
 506  * \fn void Proc_ChangeStack(void)
 507  * \brief Swaps the current stack for a new one (in the proper stack reigon)
 508  */
 509 void Proc_ChangeStack(void)
 510 {
 511         Uint    esp, ebp;
 512         Uint    tmpEbp, oldEsp;
 513         Uint    curBase, newBase;
 514
 515         __asm__ __volatile__ ("mov %%esp, %0":"=r"(esp));
 516         __asm__ __volatile__ ("mov %%ebp, %0":"=r"(ebp));
 517
 518         oldEsp = esp;
 519
 520         // Create new KStack
 521         newBase = MM_NewKStack();
 522         // Check for errors
 523         if(newBase == 0) {
 524                 Panic("What the?? Unable to allocate space for initial kernel stack");
 525                 return;
 526         }
 527
 528         curBase = (Uint)&Kernel_Stack_Top;
 529
 530         LOG("curBase = 0x%x, newBase = 0x%x", curBase, newBase);
 531
 532         // Get ESP as a used size
 533         esp = curBase - esp;
 534         LOG("memcpy( %p, %p, 0x%x )", (void*)(newBase - esp), (void*)(curBase - esp), esp );
 535         // Copy used stack
 536         memcpy( (void*)(newBase - esp), (void*)(curBase - esp), esp );
 537         // Get ESP as an offset in the new stack
 538         esp = newBase - esp;
 539         // Adjust EBP
 540         ebp = newBase - (curBase - ebp);
 541
 542         // Repair EBPs & Stack Addresses
 543         // Catches arguments also, but may trash stack-address-like values
 544         for(tmpEbp = esp; tmpEbp < newBase; tmpEbp += 4)
 545         {
 546                 if(oldEsp < *(Uint*)tmpEbp && *(Uint*)tmpEbp < curBase)
 547                         *(Uint*)tmpEbp += newBase - curBase;
 548         }
 549
 550         Proc_GetCurThread()->KernelStack = newBase;
 551
 552         __asm__ __volatile__ ("mov %0, %%esp"::"r"(esp));
 553         __asm__ __volatile__ ("mov %0, %%ebp"::"r"(ebp));
 554 }
 555
 556 /**
 557  * \fn int Proc_Clone(Uint *Err, Uint Flags)
 558  * \brief Clone the current process
 559  */
 560 int Proc_Clone(Uint Flags)
 561 {
 562         tThread *newThread;
 563         tThread *cur = Proc_GetCurThread();
 564         Uint    eip, esp, ebp;
 565
 566         __asm__ __volatile__ ("mov %%esp, %0": "=r"(esp));
 567         __asm__ __volatile__ ("mov %%ebp, %0": "=r"(ebp));
 568
 569         newThread = Threads_CloneTCB(NULL, Flags);
 570         if(!newThread)  return -1;
 571
 572         // Initialise Memory Space (New Addr space or kernel stack)
 573         if(Flags & CLONE_VM) {
 574                 newThread->MemState.CR3 = MM_Clone();
 575                 // Check for errors
 576                 if(newThread->MemState.CR3 == 0) {
 577                         Threads_Kill(newThread, -2);
 578                         return -1;
 579                 }
 580                 newThread->KernelStack = cur->KernelStack;
 581         } else {
 582                 Uint    tmpEbp, oldEsp = esp;
 583
 584                 // Set CR3
 585                 newThread->MemState.CR3 = cur->MemState.CR3;
 586
 587                 // Create new KStack
 588                 newThread->KernelStack = MM_NewKStack();
 589                 // Check for errors
 590                 if(newThread->KernelStack == 0) {
 591                         Threads_Kill(newThread, -2);
 592                         return -1;
 593                 }
 594
 595                 // Get ESP as a used size
 596                 esp = cur->KernelStack - esp;
 597                 // Copy used stack
 598                 memcpy( (void*)(newThread->KernelStack - esp), (void*)(cur->KernelStack - esp), esp );
 599                 // Get ESP as an offset in the new stack
 600                 esp = newThread->KernelStack - esp;
 601                 // Adjust EBP
 602                 ebp = newThread->KernelStack - (cur->KernelStack - ebp);
 603
 604                 // Repair EBPs & Stack Addresses
 605                 // Catches arguments also, but may trash stack-address-like values
 606                 for(tmpEbp = esp; tmpEbp < newThread->KernelStack; tmpEbp += 4)
 607                 {
 608                         if(oldEsp < *(Uint*)tmpEbp && *(Uint*)tmpEbp < cur->KernelStack)
 609                                 *(Uint*)tmpEbp += newThread->KernelStack - cur->KernelStack;
 610                 }
 611         }
 612
 613         // Save core machine state
 614         newThread->SavedState.ESP = esp;
 615         newThread->SavedState.EBP = ebp;
 616         eip = GetEIP();
 617         if(eip == SWITCH_MAGIC) {
 618                 //__asm__ __volatile__ ("mov %0, %%db0" : : "r" (newThread) );
 619                 #if USE_MP
 620                 // ACK the interrupt
 621                 if( GetCPUNum() )
 622                         gpMP_LocalAPIC->EOI.Val = 0;
 623                 else
 624                 #endif
 625                         outb(0x20, 0x20);       // ACK Timer and return as child
 626                 __asm__ __volatile__ ("sti");   // Restart interrupts
 627                 return 0;
 628         }
 629
 630         // Set EIP as parent
 631         newThread->SavedState.EIP = eip;
 632
 633         // Lock list and add to active
 634         Threads_AddActive(newThread);
 635
 636         return newThread->TID;
 637 }
 638
 639 /**
 640  * \fn int Proc_SpawnWorker(void)
 641  * \brief Spawns a new worker thread
 642  */
 643 int Proc_SpawnWorker(void)
 644 {
 645         tThread *new, *cur;
 646         Uint    eip, esp, ebp;
 647
 648         cur = Proc_GetCurThread();
 649
 650         // Create new thread
 651         new = Threads_CloneThreadZero();
 652         if(!new) {
 653                 Warning("Proc_SpawnWorker - Out of heap space!\n");
 654                 return -1;
 655         }
 656         // Create a new worker stack (in PID0's address space)
 657         // - The stack is relocated by this function
 658         new->KernelStack = MM_NewWorkerStack();
 659
 660         // Get ESP and EBP based in the new stack
 661         __asm__ __volatile__ ("mov %%esp, %0": "=r"(esp));
 662         __asm__ __volatile__ ("mov %%ebp, %0": "=r"(ebp));
 663         esp = new->KernelStack - (cur->KernelStack - esp);
 664         ebp = new->KernelStack - (cur->KernelStack - ebp);
 665
 666         // Save core machine state
 667         new->SavedState.ESP = esp;
 668         new->SavedState.EBP = ebp;
 669         eip = GetEIP();
 670         if(eip == SWITCH_MAGIC) {
 671                 //__asm__ __volatile__ ("mov %0, %%db0" : : "r"(new));
 672                 #if USE_MP
 673                 // ACK the interrupt
 674                 if(GetCPUNum())
 675                         gpMP_LocalAPIC->EOI.Val = 0;
 676                 else
 677                 #endif
 678                         outb(0x20, 0x20);       // ACK Timer and return as child
 679                 __asm__ __volatile__ ("sti");   // Restart interrupts
 680                 return 0;
 681         }
 682
 683         // Set EIP as parent
 684         new->SavedState.EIP = eip;
 685         // Mark as active
 686         Threads_AddActive( new );
 687
 688         return new->TID;
 689 }
 690
 691 /**
 692  * \fn Uint Proc_MakeUserStack(void)
 693  * \brief Creates a new user stack
 694  */
 695 Uint Proc_MakeUserStack(void)
 696 {
 697          int    i;
 698         Uint    base = USER_STACK_TOP - USER_STACK_SZ;
 699
 700         // Check Prospective Space
 701         for( i = USER_STACK_SZ >> 12; i--; )
 702                 if( MM_GetPhysAddr( base + (i<<12) ) != 0 )
 703                         break;
 704
 705         if(i != -1)     return 0;
 706
 707         // Allocate Stack - Allocate incrementally to clean up MM_Dump output
 708         for( i = 0; i < USER_STACK_SZ/0x1000; i++ )
 709         {
 710                 if( !MM_Allocate( base + (i<<12) ) )
 711                 {
 712                         Warning("OOM: Proc_MakeUserStack");
 713                         return 0;
 714                 }
 715         }
 716
 717         return base + USER_STACK_SZ;
 718 }
 719
 720 /**
 721  * \fn void Proc_StartUser(Uint Entrypoint, Uint *Bases, int ArgC, char **ArgV, char **EnvP, int DataSize)
 722  * \brief Starts a user task
 723  */
 724 void Proc_StartUser(Uint Entrypoint, Uint *Bases, int ArgC, char **ArgV, char **EnvP, int DataSize)
 725 {
 726         Uint    *stack = (void*)Proc_MakeUserStack();
 727          int    i;
 728         Uint    delta;
 729         Uint16  ss, cs;
 730
 731         //Log("stack = %p", stack);
 732
 733         // Copy Arguments
 734         stack -= DataSize/sizeof(*stack);
 735         memcpy( stack, ArgV, DataSize );
 736
 737         //Log("stack = %p", stack);
 738
 739         if( DataSize )
 740         {
 741                 // Adjust Arguments and environment
 742                 delta = (Uint)stack - (Uint)ArgV;
 743                 ArgV = (char**)stack;
 744                 for( i = 0; ArgV[i]; i++ )
 745                         ArgV[i] += delta;
 746                 i ++;
 747
 748                 // Do we care about EnvP?
 749                 if( EnvP ) {
 750                         EnvP = &ArgV[i];
 751                         for( i = 0; EnvP[i]; i++ )
 752                                 EnvP[i] += delta;
 753                 }
 754         }
 755
 756         // User Mode Segments
 757         ss = 0x23;      cs = 0x1B;
 758
 759         // Arguments
 760         *--stack = (Uint)EnvP;
 761         *--stack = (Uint)ArgV;
 762         *--stack = (Uint)ArgC;
 763         while(*Bases)
 764                 *--stack = *Bases++;
 765         *--stack = 0;   // Return Address
 766
 767         Proc_StartProcess(ss, (Uint)stack, 0x202, cs, Entrypoint);
 768 }
 769
 770 void Proc_StartProcess(Uint16 SS, Uint Stack, Uint Flags, Uint16 CS, Uint IP)
 771 {
 772         Uint    *stack = (void*)Stack;
 773         *--stack = SS;          //Stack Segment
 774         *--stack = Stack;       //Stack Pointer
 775         *--stack = Flags;       //EFLAGS (Resvd (0x2) and IF (0x20))
 776         *--stack = CS;          //Code Segment
 777         *--stack = IP;  //EIP
 778         //PUSHAD
 779         *--stack = 0xAAAAAAAA;  // eax
 780         *--stack = 0xCCCCCCCC;  // ecx
 781         *--stack = 0xDDDDDDDD;  // edx
 782         *--stack = 0xBBBBBBBB;  // ebx
 783         *--stack = 0xD1D1D1D1;  // edi
 784         *--stack = 0x54545454;  // esp - NOT POPED
 785         *--stack = 0x51515151;  // esi
 786         *--stack = 0xB4B4B4B4;  // ebp
 787         //Individual PUSHs
 788         *--stack = SS;  // ds
 789         *--stack = SS;  // es
 790         *--stack = SS;  // fs
 791         *--stack = SS;  // gs
 792
 793         __asm__ __volatile__ (
 794         "mov %%eax,%%esp;\n\t"  // Set stack pointer
 795         "pop %%gs;\n\t"
 796         "pop %%fs;\n\t"
 797         "pop %%es;\n\t"
 798         "pop %%ds;\n\t"
 799         "popa;\n\t"
 800         "iret;\n\t" : : "a" (stack));
 801         for(;;);
 802 }
 803
 804 /**
 805  * \fn int Proc_Demote(Uint *Err, int Dest, tRegs *Regs)
 806  * \brief Demotes a process to a lower permission level
 807  * \param Err   Pointer to user's errno
 808  * \param Dest  New Permission Level
 809  * \param Regs  Pointer to user's register structure
 810  */
 811 int Proc_Demote(Uint *Err, int Dest, tRegs *Regs)
 812 {
 813          int    cpl = Regs->cs & 3;
 814         // Sanity Check
 815         if(Dest > 3 || Dest < 0) {
 816                 *Err = -EINVAL;
 817                 return -1;
 818         }
 819
 820         // Permission Check
 821         if(cpl > Dest) {
 822                 *Err = -EACCES;
 823                 return -1;
 824         }
 825
 826         // Change the Segment Registers
 827         Regs->cs = (((Dest+1)<<4) | Dest) - 8;
 828         Regs->ss = ((Dest+1)<<4) | Dest;
 829         // Check if the GP Segs are GDT, then change them
 830         if(!(Regs->ds & 4))     Regs->ds = ((Dest+1)<<4) | Dest;
 831         if(!(Regs->es & 4))     Regs->es = ((Dest+1)<<4) | Dest;
 832         if(!(Regs->fs & 4))     Regs->fs = ((Dest+1)<<4) | Dest;
 833         if(!(Regs->gs & 4))     Regs->gs = ((Dest+1)<<4) | Dest;
 834
 835         return 0;
 836 }
 837
 838 /**
 839  * \brief Calls a signal handler in user mode
 840  * \note Used for signals
 841  */
 842 void Proc_CallFaultHandler(tThread *Thread)
 843 {
 844         // Rewinds the stack and calls the user function
 845         // Never returns
 846         Proc_ReturnToUser( Thread->FaultHandler, Thread->CurFaultNum, Thread->KernelStack );
 847         for(;;);
 848 }
 849
 850 void Proc_DumpThreadCPUState(tThread *Thread)
 851 {
 852         if( Thread->CurCPU > -1 )
 853         {
 854                  int    maxBacktraceDistance = 6;
 855                 tRegs   *regs = NULL;
 856                 Uint32  *stack;
 857
 858                 if( Thread->CurCPU != GetCPUNum() ) {
 859                         Log("  Currently running");
 860                         return ;
 861                 }
 862
 863                 // Backtrace to find the IRQ entrypoint
 864                 // - This will usually only be called by an IRQ, so this should
 865                 //   work
 866                 __asm__ __volatile__ ("mov %%ebp, %0" : "=r" (stack));
 867                 while( maxBacktraceDistance -- )
 868                 {
 869                         // [ebp] = oldEbp
 870                         // [ebp+4] = retaddr
 871
 872                         if( stack[1] == (tVAddr)&IRQCommon_handled ) {
 873                                 regs = (void*)stack[2];
 874                                 break;
 875                         }
 876
 877                         stack = (void*)stack[0];
 878                 }
 879
 880                 if( !regs ) {
 881                         Log("  Unable to find IRQ Entry");
 882                         return ;
 883                 }
 884
 885                 Log("  at %04x:%08x", regs->cs, regs->eip);
 886                 return ;
 887         }
 888
 889         #if 1
 890         tVAddr  diffFromScheduler = Thread->SavedState.EIP - (tVAddr)Proc_Scheduler;
 891         tVAddr  diffFromClone = Thread->SavedState.EIP - (tVAddr)Proc_Clone;
 892         tVAddr  diffFromSpawn = Thread->SavedState.EIP - (tVAddr)Proc_SpawnWorker;
 893
 894         if( diffFromClone > 0 && diffFromClone < 512 )  // When I last checked, GetEIP was at .+0x183
 895         {
 896                 Log("  Creating full thread");
 897                 return ;
 898         }
 899
 900         if( diffFromSpawn > 0 && diffFromSpawn < 512 )  // When I last checked, GetEIP was at .+0x99
 901         {
 902                 Log("  Creating worker thread");
 903                 return ;
 904         }
 905
 906         if( diffFromScheduler > 0 && diffFromScheduler < 256 )  // When I last checked, GetEIP was at .+0x60
 907         #else
 908         Uint32  data[3];
 909         MM_ReadFromAddrSpace(Thread->MemState.CR3, Thread->SavedState.EBP, data, 12);
 910         if( data[1] == (Uint32)&IRQCommon + 25 )
 911         {
 912                 tRegs   *regs = (void *) data[2];
 913                 Log("  oldebp = 0x%08x, ret = 0x%08x, regs = 0x%x",
 914                         data[0], data[1], data[2]
 915                         );
 916                 // [EBP] = old EBP
 917                 // [EBP+0x04] = Return Addr
 918                 // [EBP+0x08] = Arg 1 (CPU Number)
 919                 // [EBP+0x0C] = Arg 2 (Thread)
 920                 // [EBP+0x10] = GS (start of tRegs)
 921                 Log("  IRQ%i from %02x:%08x", regs->int_num regs->cs, regs->eip);
 922         }
 923         if( stack[1] == (Uint32)&scheduler_return )
 924         #endif
 925         {
 926                 // Scheduled out
 927                 Log("  At %04x:%08x", Thread->SavedState.UserCS, Thread->SavedState.UserEIP);
 928                 return ;
 929         }
 930
 931         Log("  Just created (unknow %p)", Thread->SavedState.EIP);
 932 }
 933
 934 /**
 935  * \fn void Proc_Scheduler(int CPU)
 936  * \brief Swap current thread and clears dead threads
 937  */
 938 void Proc_Scheduler(int CPU)
 939 {
 940         Uint    esp, ebp, eip;
 941         tThread *thread;
 942
 943         // If the spinlock is set, let it complete
 944         if(IS_LOCKED(&glThreadListLock))        return;
 945
 946         // Get current thread
 947         #if USE_MP
 948         thread = gaCPUs[CPU].Current;
 949         #else
 950         thread = gCurrentThread;
 951         #endif
 952
 953         // NOTE:
 954         // 2011-04-05
 955         // Bug may be caused by DR0 not being maintained somewhere, hence
 956         // login is getting loaded with the idle state.
 957         if( thread )
 958         {
 959                 tRegs   *regs;
 960                 // Reduce remaining quantum and continue timeslice if non-zero
 961                 if( thread->Remaining-- )
 962                         return;
 963                 // Reset quantum for next call
 964                 thread->Remaining = thread->Quantum;
 965
 966                 // Get machine state
 967                 __asm__ __volatile__ ( "mov %%esp, %0" : "=r" (esp) );
 968                 __asm__ __volatile__ ( "mov %%ebp, %0" : "=r" (ebp) );
 969                 eip = GetEIP();
 970                 if(eip == SWITCH_MAGIC) return; // Check if a switch happened
 971
 972                 // Save machine state
 973                 thread->SavedState.ESP = esp;
 974                 thread->SavedState.EBP = ebp;
 975                 thread->SavedState.EIP = eip;
 976
 977                 // TODO: Make this more stable somehow
 978                 regs = (tRegs*)(ebp+(2+2)*4);   // EBP,Ret + CPU,CurThread
 979                 thread->SavedState.UserCS = regs->cs;
 980                 thread->SavedState.UserEIP = regs->eip;
 981
 982                 if(thread->bInstrTrace) {
 983                         regs->eflags |= 0x100;  // Set TF
 984                         Log("%p De-scheduled", thread);
 985                 }
 986                 else
 987                         regs->eflags &= ~0x100; // Clear TF
 988         }
 989
 990         // Get next thread to run
 991         thread = Threads_GetNextToRun(CPU, thread);
 992
 993
 994         // No avaliable tasks, just go into low power mode (idle thread)
 995         if(thread == NULL) {
 996                 #if USE_MP
 997                 thread = gaCPUs[CPU].IdleThread;
 998                 Log("CPU %i Running Idle Thread", CPU);
 999                 #else
1000                 thread = gpIdleThread;
1001                 #endif
1002         }
1003
1004         #if DEBUG_TRACE_SWITCH
1005         if(thread && thread != Proc_GetCurThread() ) {
1006                 Log("Switching to task %i(%s), CR3 = 0x%x, EIP = %p",
1007                         thread->TID,
1008                         thread->ThreadName,
1009                         thread->MemState.CR3,
1010                         thread->SavedState.EIP
1011                         );
1012         }
1013         #endif
1014
1015         // Set current thread
1016         #if USE_MP
1017         gaCPUs[CPU].Current = thread;
1018         #else
1019         gCurrentThread = thread;
1020         #endif
1021
1022         #if USE_MP      // MP Debug
1023 //      Log("CPU = %i, Thread %p", CPU, thread);
1024         #endif
1025
1026         // Update Kernel Stack pointer
1027         gTSSs[CPU].ESP0 = thread->KernelStack-4;
1028
1029         #if 0
1030         if(thread->SavedState.ESP > 0xC0000000
1031         && thread->SavedState.ESP < thread->KernelStack-0x2000) {
1032                 Log_Warning("Proc", "Possible bad ESP %p (PID %i)", thread->SavedState.ESP);
1033         }
1034         #endif
1035
1036         if( thread->bInstrTrace ) {
1037                 Log("%p Scheduled", thread);
1038         }
1039
1040         // Set thread pointer
1041         __asm__ __volatile__("mov %0, %%db0\n\t" : : "r"(thread) );
1042         // Switch threads
1043         __asm__ __volatile__ (
1044                 "mov %4, %%cr3\n\t"     // Set address space
1045                 "mov %1, %%esp\n\t"     // Restore ESP
1046                 "mov %2, %%ebp\n\t"     // and EBP
1047                 "or %5, 72(%%ebp)\n\t"  // or trace flag to eflags (2+2+4+8+2)*4
1048                 "jmp *%3" : :   // And return to where we saved state (Proc_Clone or Proc_Scheduler)
1049                 "a"(SWITCH_MAGIC), "b"(thread->SavedState.ESP),
1050                 "d"(thread->SavedState.EBP), "c"(thread->SavedState.EIP),
1051                 "r"(thread->MemState.CR3),
1052                 "r"(thread->bInstrTrace&&thread->SavedState.EIP==(Uint)&GetEIP_Sched_ret?0x100:0)
1053                 );
1054         for(;;);        // Shouldn't reach here
1055 }
1056
1057 // === EXPORTS ===
1058 EXPORT(Proc_SpawnWorker);