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