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
  15 // === FLAGS ===
  16 #define DEBUG_TRACE_SWITCH      0
  17
  18 // === CONSTANTS ===
  19 #define SWITCH_MAGIC    0xFFFACE55      // There is no code in this area
  20 // Base is 1193182
  21 #define TIMER_BASE      1193182
  22 #define TIMER_DIVISOR   11931   //~100Hz
  23
  24 // === TYPES ===
  25 #if USE_MP
  26 typedef struct sCPU
  27 {
  28         Uint8   APICID;
  29         Uint8   State;  // 0: Unavaliable, 1: Idle, 2: Active
  30         Uint16  Resvd;
  31         tThread *Current;
  32         tThread *IdleThread;
  33 }       tCPU;
  34 #endif
  35
  36 // === IMPORTS ===
  37 extern tGDT     gGDT[];
  38 extern tIDT     gIDT[];
  39 extern void APWait(void);       // 16-bit AP pause code
  40 extern void APStartup(void);    // 16-bit AP startup code
  41 extern Uint     GetEIP(void);   // start.asm
  42 extern int      GetCPUNum(void);        // start.asm
  43 extern Uint32   gaInitPageDir[1024];    // start.asm
  44 extern void     Kernel_Stack_Top;
  45 extern tSpinlock        glThreadListLock;
  46 extern int      giNumCPUs;
  47 extern int      giNextTID;
  48 extern int      giTotalTickets;
  49 extern int      giNumActiveThreads;
  50 extern tThread  gThreadZero;
  51 extern tThread  *gActiveThreads;
  52 extern tThread  *gSleepingThreads;
  53 extern tThread  *gDeleteThreads;
  54 extern void     Threads_Dump(void);
  55 extern tThread  *Threads_CloneTCB(Uint *Err, Uint Flags);
  56 extern void     Isr8(void);     // Double Fault
  57 extern void     Proc_ReturnToUser(void);
  58
  59 // === PROTOTYPES ===
  60 void    ArchThreads_Init(void);
  61 #if USE_MP
  62 void    MP_StartAP(int CPU);
  63 void    MP_SendIPI(Uint8 APICID, int Vector, int DeliveryMode);
  64 #endif
  65 void    Proc_Start(void);
  66 tThread *Proc_GetCurThread(void);
  67 void    Proc_ChangeStack(void);
  68  int    Proc_Clone(Uint *Err, Uint Flags);
  69 void    Proc_StartProcess(Uint16 SS, Uint Stack, Uint Flags, Uint16 CS, Uint IP);
  70 void    Proc_CallFaultHandler(tThread *Thread);
  71 void    Proc_Scheduler(int CPU);
  72
  73 // === GLOBALS ===
  74 // --- Multiprocessing ---
  75 #if USE_MP
  76 volatile int    giNumInitingCPUs = 0;
  77 tMPInfo *gMPFloatPtr = NULL;
  78 volatile Uint32 giMP_TimerCount;        // Start Count for Local APIC Timer
  79 tAPIC   *gpMP_LocalAPIC = NULL;
  80 Uint8   gaAPIC_to_CPU[256] = {0};
  81 tCPU    gaCPUs[MAX_CPUS];
  82 tTSS    gaTSSs[MAX_CPUS];       // TSS Array
  83  int    giProc_BootProcessorID = 0;
  84 #else
  85 tThread *gCurrentThread = NULL;
  86 tThread *gpIdleThread = NULL;
  87 #endif
  88 #if USE_PAE
  89 Uint32  *gPML4s[4] = NULL;
  90 #endif
  91 tTSS    *gTSSs = NULL;  // Pointer to TSS array
  92 tTSS    gTSS0 = {0};
  93 // --- Error Recovery ---
  94 char    gaDoubleFaultStack[1024] __attribute__ ((section(".padata")));
  95 tTSS    gDoubleFault_TSS = {
  96         .ESP0 = (Uint)&gaDoubleFaultStack[1024],
  97         .SS0 = 0x10,
  98         .CR3 = (Uint)gaInitPageDir - KERNEL_BASE,
  99         .EIP = (Uint)Isr8,
 100         .ESP = (Uint)&gaDoubleFaultStack[1024],
 101         .CS = 0x08,     .SS = 0x10,
 102         .DS = 0x10,     .ES = 0x10,
 103         .FS = 0x10,     .GS = 0x10,
 104 };
 105
 106 // === CODE ===
 107 /**
 108  * \fn void ArchThreads_Init(void)
 109  * \brief Starts the process scheduler
 110  */
 111 void ArchThreads_Init(void)
 112 {
 113         Uint    pos = 0;
 114
 115         #if USE_MP
 116         tMPTable        *mptable;
 117
 118         // Mark BSP as active
 119         gaCPUs[0].State = 2;
 120
 121         // -- Initialise Multiprocessing
 122         // Find MP Floating Table
 123         // - EBDA/Last 1Kib (640KiB)
 124         for(pos = KERNEL_BASE|0x9F000; pos < (KERNEL_BASE|0xA0000); pos += 16) {
 125                 if( *(Uint*)(pos) == MPPTR_IDENT ) {
 126                         Log("Possible %p", pos);
 127                         if( ByteSum((void*)pos, sizeof(tMPInfo)) != 0 ) continue;
 128                         gMPFloatPtr = (void*)pos;
 129                         break;
 130                 }
 131         }
 132         // - Last KiB (512KiB base mem)
 133         if(!gMPFloatPtr) {
 134                 for(pos = KERNEL_BASE|0x7F000; pos < (KERNEL_BASE|0x80000); pos += 16) {
 135                         if( *(Uint*)(pos) == MPPTR_IDENT ) {
 136                                 Log("Possible %p", pos);
 137                                 if( ByteSum((void*)pos, sizeof(tMPInfo)) != 0 ) continue;
 138                                 gMPFloatPtr = (void*)pos;
 139                                 break;
 140                         }
 141                 }
 142         }
 143         // - BIOS ROM
 144         if(!gMPFloatPtr) {
 145                 for(pos = KERNEL_BASE|0xE0000; pos < (KERNEL_BASE|0x100000); pos += 16) {
 146                         if( *(Uint*)(pos) == MPPTR_IDENT ) {
 147                                 Log("Possible %p", pos);
 148                                 if( ByteSum((void*)pos, sizeof(tMPInfo)) != 0 ) continue;
 149                                 gMPFloatPtr = (void*)pos;
 150                                 break;
 151                         }
 152                 }
 153         }
 154
 155         // If the MP Table Exists, parse it
 156         if(gMPFloatPtr)
 157         {
 158                  int    i;
 159                 tMPTable_Ent    *ents;
 160                 Log("gMPFloatPtr = %p", gMPFloatPtr);
 161                 Log("*gMPFloatPtr = {");
 162                 Log("\t.Sig = 0x%08x", gMPFloatPtr->Sig);
 163                 Log("\t.MPConfig = 0x%08x", gMPFloatPtr->MPConfig);
 164                 Log("\t.Length = 0x%02x", gMPFloatPtr->Length);
 165                 Log("\t.Version = 0x%02x", gMPFloatPtr->Version);
 166                 Log("\t.Checksum = 0x%02x", gMPFloatPtr->Checksum);
 167                 Log("\t.Features = [0x%02x,0x%02x,0x%02x,0x%02x,0x%02x]",
 168                         gMPFloatPtr->Features[0],       gMPFloatPtr->Features[1],
 169                         gMPFloatPtr->Features[2],       gMPFloatPtr->Features[3],
 170                         gMPFloatPtr->Features[4]
 171                         );
 172                 Log("}");
 173
 174                 mptable = (void*)( KERNEL_BASE|gMPFloatPtr->MPConfig );
 175                 Log("mptable = %p", mptable);
 176                 Log("*mptable = {");
 177                 Log("\t.Sig = 0x%08x", mptable->Sig);
 178                 Log("\t.BaseTableLength = 0x%04x", mptable->BaseTableLength);
 179                 Log("\t.SpecRev = 0x%02x", mptable->SpecRev);
 180                 Log("\t.Checksum = 0x%02x", mptable->Checksum);
 181                 Log("\t.OEMID = '%8c'", mptable->OemID);
 182                 Log("\t.ProductID = '%8c'", mptable->ProductID);
 183                 Log("\t.OEMTablePtr = %p'", mptable->OEMTablePtr);
 184                 Log("\t.OEMTableSize = 0x%04x", mptable->OEMTableSize);
 185                 Log("\t.EntryCount = 0x%04x", mptable->EntryCount);
 186                 Log("\t.LocalAPICMemMap = 0x%08x", mptable->LocalAPICMemMap);
 187                 Log("\t.ExtendedTableLen = 0x%04x", mptable->ExtendedTableLen);
 188                 Log("\t.ExtendedTableChecksum = 0x%02x", mptable->ExtendedTableChecksum);
 189                 Log("}");
 190
 191                 gpMP_LocalAPIC = (void*)MM_MapHWPages(mptable->LocalAPICMemMap, 1);
 192
 193                 ents = mptable->Entries;
 194                 giNumCPUs = 0;
 195
 196                 for( i = 0; i < mptable->EntryCount; i ++ )
 197                 {
 198                          int    entSize = 0;
 199                         switch( ents->Type )
 200                         {
 201                         case 0: // Processor
 202                                 entSize = 20;
 203                                 Log("%i: Processor", i);
 204                                 Log("\t.APICID = %i", ents->Proc.APICID);
 205                                 Log("\t.APICVer = 0x%02x", ents->Proc.APICVer);
 206                                 Log("\t.CPUFlags = 0x%02x", ents->Proc.CPUFlags);
 207                                 Log("\t.CPUSignature = 0x%08x", ents->Proc.CPUSignature);
 208                                 Log("\t.FeatureFlags = 0x%08x", ents->Proc.FeatureFlags);
 209
 210
 211                                 if( !(ents->Proc.CPUFlags & 1) ) {
 212                                         Log("DISABLED");
 213                                         break;
 214                                 }
 215
 216                                 // Check if there is too many processors
 217                                 if(giNumCPUs >= MAX_CPUS) {
 218                                         giNumCPUs ++;   // If `giNumCPUs` > MAX_CPUS later, it will be clipped
 219                                         break;
 220                                 }
 221
 222                                 // Initialise CPU Info
 223                                 gaAPIC_to_CPU[ents->Proc.APICID] = giNumCPUs;
 224                                 gaCPUs[giNumCPUs].APICID = ents->Proc.APICID;
 225                                 gaCPUs[giNumCPUs].State = 0;
 226                                 giNumCPUs ++;
 227
 228                                 // Set BSP Variable
 229                                 if( ents->Proc.CPUFlags & 2 ) {
 230                                         giProc_BootProcessorID = giNumCPUs-1;
 231                                 }
 232
 233                                 break;
 234
 235                         #if DUMP_MP_TABLES
 236                         case 1: // Bus
 237                                 entSize = 8;
 238                                 Log("%i: Bus", i);
 239                                 Log("\t.ID = %i", ents->Bus.ID);
 240                                 Log("\t.TypeString = '%6C'", ents->Bus.TypeString);
 241                                 break;
 242                         case 2: // I/O APIC
 243                                 entSize = 8;
 244                                 Log("%i: I/O APIC", i);
 245                                 Log("\t.ID = %i", ents->IOAPIC.ID);
 246                                 Log("\t.Version = 0x%02x", ents->IOAPIC.Version);
 247                                 Log("\t.Flags = 0x%02x", ents->IOAPIC.Flags);
 248                                 Log("\t.Addr = 0x%08x", ents->IOAPIC.Addr);
 249                                 break;
 250                         case 3: // I/O Interrupt Assignment
 251                                 entSize = 8;
 252                                 Log("%i: I/O Interrupt Assignment", i);
 253                                 Log("\t.IntType = %i", ents->IOInt.IntType);
 254                                 Log("\t.Flags = 0x%04x", ents->IOInt.Flags);
 255                                 Log("\t.SourceBusID = 0x%02x", ents->IOInt.SourceBusID);
 256                                 Log("\t.SourceBusIRQ = 0x%02x", ents->IOInt.SourceBusIRQ);
 257                                 Log("\t.DestAPICID = 0x%02x", ents->IOInt.DestAPICID);
 258                                 Log("\t.DestAPICIRQ = 0x%02x", ents->IOInt.DestAPICIRQ);
 259                                 break;
 260                         case 4: // Local Interrupt Assignment
 261                                 entSize = 8;
 262                                 Log("%i: Local Interrupt Assignment", i);
 263                                 Log("\t.IntType = %i", ents->LocalInt.IntType);
 264                                 Log("\t.Flags = 0x%04x", ents->LocalInt.Flags);
 265                                 Log("\t.SourceBusID = 0x%02x", ents->LocalInt.SourceBusID);
 266                                 Log("\t.SourceBusIRQ = 0x%02x", ents->LocalInt.SourceBusIRQ);
 267                                 Log("\t.DestLocalAPICID = 0x%02x", ents->LocalInt.DestLocalAPICID);
 268                                 Log("\t.DestLocalAPICIRQ = 0x%02x", ents->LocalInt.DestLocalAPICIRQ);
 269                                 break;
 270                         default:
 271                                 Log("%i: Unknown (%i)", i, ents->Type);
 272                                 break;
 273                         #endif
 274                         }
 275                         ents = (void*)( (Uint)ents + entSize );
 276                 }
 277
 278                 if( giNumCPUs > MAX_CPUS ) {
 279                         Warning("Too many CPUs detected (%i), only using %i of them", giNumCPUs, MAX_CPUS);
 280                         giNumCPUs = MAX_CPUS;
 281                 }
 282                 gTSSs = gaTSSs;
 283         }
 284         else {
 285                 Log("No MP Table was found, assuming uniprocessor\n");
 286                 giNumCPUs = 1;
 287                 gTSSs = &gTSS0;
 288         }
 289         #else
 290         giNumCPUs = 1;
 291         gTSSs = &gTSS0;
 292         MM_FinishVirtualInit();
 293         #endif
 294
 295         #if 0
 296         // Initialise Double Fault TSS
 297         gGDT[5].BaseLow = (Uint)&gDoubleFault_TSS & 0xFFFF;
 298         gGDT[5].BaseMid = (Uint)&gDoubleFault_TSS >> 16;
 299         gGDT[5].BaseHi = (Uint)&gDoubleFault_TSS >> 24;
 300
 301         // Set double fault IDT to use the new TSS
 302         gIDT[8].OffsetLo = 0;
 303         gIDT[8].CS = 5<<3;
 304         gIDT[8].Flags = 0x8500;
 305         gIDT[8].OffsetHi = 0;
 306         #endif
 307
 308         // Set timer frequency
 309         outb(0x43, 0x34);       // Set Channel 0, Low/High, Rate Generator
 310         outb(0x40, TIMER_DIVISOR&0xFF); // Low Byte of Divisor
 311         outb(0x40, (TIMER_DIVISOR>>8)&0xFF);    // High Byte
 312
 313         #if USE_MP
 314         // Get the count setting for APIC timer
 315         Log("Determining APIC Count");
 316         __asm__ __volatile__ ("sti");
 317         while( giMP_TimerCount == 0 )   __asm__ __volatile__ ("hlt");
 318         __asm__ __volatile__ ("cli");
 319         Log("APIC Count %i", giMP_TimerCount);
 320         {
 321                 Uint64  freq = giMP_TimerCount;
 322                 freq /= TIMER_DIVISOR;
 323                 freq *= TIMER_BASE;
 324                 if( (freq /= 1000) < 2*1000)
 325                         Log("Bus Frequency %i KHz", freq);
 326                 else if( (freq /= 1000) < 2*1000)
 327                         Log("Bus Frequency %i MHz", freq);
 328                 else if( (freq /= 1000) < 2*1000)
 329                         Log("Bus Frequency %i GHz", freq);
 330                 else
 331                         Log("Bus Frequency %i THz", freq);
 332         }
 333
 334         // Initialise Normal TSS(s)
 335         for(pos=0;pos<giNumCPUs;pos++)
 336         {
 337         #else
 338         pos = 0;
 339         #endif
 340                 gTSSs[pos].SS0 = 0x10;
 341                 gTSSs[pos].ESP0 = 0;    // Set properly by scheduler
 342                 gGDT[6+pos].BaseLow = ((Uint)(&gTSSs[pos])) & 0xFFFF;
 343                 gGDT[6+pos].BaseMid = ((Uint)(&gTSSs[pos])) >> 16;
 344                 gGDT[6+pos].BaseHi = ((Uint)(&gTSSs[pos])) >> 24;
 345         #if USE_MP
 346         }
 347         #endif
 348
 349         // Load the BSP's TSS
 350         __asm__ __volatile__ ("ltr %%ax"::"a"(0x30));
 351
 352         #if USE_MP
 353         gaCPUs[0].Current = &gThreadZero;
 354         #else
 355         gCurrentThread = &gThreadZero;
 356         #endif
 357         gThreadZero.CurCPU = 0;
 358
 359         #if USE_PAE
 360         gThreadZero.MemState.PDP[0] = 0;
 361         gThreadZero.MemState.PDP[1] = 0;
 362         gThreadZero.MemState.PDP[2] = 0;
 363         #else
 364         gThreadZero.MemState.CR3 = (Uint)gaInitPageDir - KERNEL_BASE;
 365         #endif
 366
 367         // Create Per-Process Data Block
 368         MM_Allocate(MM_PPD_CFG);
 369
 370         // Change Stacks
 371         Proc_ChangeStack();
 372 }
 373
 374 #if USE_MP
 375 void MP_StartAP(int CPU)
 376 {
 377         Log("Starting AP %i (APIC %i)", CPU, gaCPUs[CPU].APICID);
 378
 379         // Set location of AP startup code and mark for a warm restart
 380         *(Uint16*)(KERNEL_BASE|0x467) = (Uint)&APWait - (KERNEL_BASE|0xFFFF0);
 381         *(Uint16*)(KERNEL_BASE|0x469) = 0xFFFF;
 382         outb(0x70, 0x0F);       outb(0x71, 0x0A);       // Warm Reset
 383         MP_SendIPI(gaCPUs[CPU].APICID, 0, 5);   // Init IPI
 384
 385         // Delay
 386         inb(0x80); inb(0x80); inb(0x80); inb(0x80);
 387
 388         // TODO: Use a better address, preferably registered with the MM
 389         // - MM_AllocDMA mabye?
 390         // Create a far jump
 391         *(Uint8*)(KERNEL_BASE|0x11000) = 0xEA;  // Far JMP
 392         *(Uint16*)(KERNEL_BASE|0x11001) = (Uint)&APStartup - (KERNEL_BASE|0xFFFF0);     // IP
 393         *(Uint16*)(KERNEL_BASE|0x11003) = 0xFFFF;       // CS
 394         // Send a Startup-IPI to make the CPU execute at 0x11000 (which we
 395         // just filled)
 396         MP_SendIPI(gaCPUs[CPU].APICID, 0x11, 6);        // StartupIPI
 397
 398         giNumInitingCPUs ++;
 399 }
 400
 401 /**
 402  * \brief Send an Inter-Processor Interrupt
 403  * \param APICID        Processor's Local APIC ID
 404  * \param Vector        Argument of some kind
 405  * \param DeliveryMode  Type of signal?
 406  */
 407 void MP_SendIPI(Uint8 APICID, int Vector, int DeliveryMode)
 408 {
 409         Uint32  val;
 410
 411         // Hi
 412         val = (Uint)APICID << 24;
 413         Log("*%p = 0x%08x", &gpMP_LocalAPIC->ICR[1], val);
 414         gpMP_LocalAPIC->ICR[1].Val = val;
 415         // Low (and send)
 416         val = ((DeliveryMode & 7) << 8) | (Vector & 0xFF);
 417         Log("*%p = 0x%08x", &gpMP_LocalAPIC->ICR[0], val);
 418         gpMP_LocalAPIC->ICR[0].Val = val;
 419 }
 420 #endif
 421
 422 /**
 423  * \fn void Proc_Start(void)
 424  * \brief Start process scheduler
 425  */
 426 void Proc_Start(void)
 427 {
 428         #if USE_MP
 429          int    i;
 430         #endif
 431
 432         #if USE_MP
 433         // Start APs
 434         for( i = 0; i < giNumCPUs; i ++ )
 435         {
 436                 // Create Idle Task
 437                 if(Proc_Clone(0, 0) == 0)
 438                 {
 439                         gaCPUs[i].IdleThread = Proc_GetCurThread();
 440                         gaCPUs[i].IdleThread->ThreadName = "Idle Thread";
 441                         gaCPUs[i].IdleThread->NumTickets = 0;   // Never called randomly
 442                         gaCPUs[i].IdleThread->Quantum = 1;      // 1 slice quantum
 443                         for(;;) HALT(); // Just yeilds
 444                 }
 445                 gaCPUs[i].Current = NULL;
 446
 447                 // Start the AP
 448                 if( i != giProc_BootProcessorID ) {
 449                         MP_StartAP( i );
 450                 }
 451         }
 452
 453         // BSP still should run the current task
 454         gaCPUs[0].Current = &gThreadZero;
 455
 456         // Start interrupts and wait for APs to come up
 457         Log("Waiting for APs to come up\n");
 458         __asm__ __volatile__ ("sti");
 459         while( giNumInitingCPUs )       __asm__ __volatile__ ("hlt");
 460         MM_FinishVirtualInit();
 461         #else
 462         // Create Idle Task
 463         if(Proc_Clone(0, 0) == 0)
 464         {
 465                 gpIdleThread = Proc_GetCurThread();
 466                 gpIdleThread->ThreadName = "Idle Thread";
 467                 gpIdleThread->NumTickets = 0;   // Never called randomly
 468                 gpIdleThread->Quantum = 1;      // 1 slice quantum
 469                 for(;;) HALT(); // Just yeilds
 470         }
 471
 472         // Set current task
 473         gCurrentThread = &gThreadZero;
 474
 475         // Start Interrupts (and hence scheduler)
 476         __asm__ __volatile__("sti");
 477         #endif
 478 }
 479
 480 /**
 481  * \fn tThread *Proc_GetCurThread(void)
 482  * \brief Gets the current thread
 483  */
 484 tThread *Proc_GetCurThread(void)
 485 {
 486         #if USE_MP
 487         //return gaCPUs[ gaAPIC_to_CPU[gpMP_LocalAPIC->ID.Val&0xFF] ].Current;
 488         return gaCPUs[ GetCPUNum() ].Current;
 489         #else
 490         return gCurrentThread;
 491         #endif
 492 }
 493
 494 /**
 495  * \fn void Proc_ChangeStack(void)
 496  * \brief Swaps the current stack for a new one (in the proper stack reigon)
 497  */
 498 void Proc_ChangeStack(void)
 499 {
 500         Uint    esp, ebp;
 501         Uint    tmpEbp, oldEsp;
 502         Uint    curBase, newBase;
 503
 504         __asm__ __volatile__ ("mov %%esp, %0":"=r"(esp));
 505         __asm__ __volatile__ ("mov %%ebp, %0":"=r"(ebp));
 506
 507         oldEsp = esp;
 508
 509         // Create new KStack
 510         newBase = MM_NewKStack();
 511         // Check for errors
 512         if(newBase == 0) {
 513                 Panic("What the?? Unable to allocate space for initial kernel stack");
 514                 return;
 515         }
 516
 517         curBase = (Uint)&Kernel_Stack_Top;
 518
 519         LOG("curBase = 0x%x, newBase = 0x%x", curBase, newBase);
 520
 521         // Get ESP as a used size
 522         esp = curBase - esp;
 523         LOG("memcpy( %p, %p, 0x%x )", (void*)(newBase - esp), (void*)(curBase - esp), esp );
 524         // Copy used stack
 525         memcpy( (void*)(newBase - esp), (void*)(curBase - esp), esp );
 526         // Get ESP as an offset in the new stack
 527         esp = newBase - esp;
 528         // Adjust EBP
 529         ebp = newBase - (curBase - ebp);
 530
 531         // Repair EBPs & Stack Addresses
 532         // Catches arguments also, but may trash stack-address-like values
 533         for(tmpEbp = esp; tmpEbp < newBase; tmpEbp += 4)
 534         {
 535                 if(oldEsp < *(Uint*)tmpEbp && *(Uint*)tmpEbp < curBase)
 536                         *(Uint*)tmpEbp += newBase - curBase;
 537         }
 538
 539         Proc_GetCurThread()->KernelStack = newBase;
 540
 541         __asm__ __volatile__ ("mov %0, %%esp"::"r"(esp));
 542         __asm__ __volatile__ ("mov %0, %%ebp"::"r"(ebp));
 543 }
 544
 545 /**
 546  * \fn int Proc_Clone(Uint *Err, Uint Flags)
 547  * \brief Clone the current process
 548  */
 549 int Proc_Clone(Uint *Err, Uint Flags)
 550 {
 551         tThread *newThread;
 552         tThread *cur = Proc_GetCurThread();
 553         Uint    eip, esp, ebp;
 554
 555         __asm__ __volatile__ ("mov %%esp, %0": "=r"(esp));
 556         __asm__ __volatile__ ("mov %%ebp, %0": "=r"(ebp));
 557
 558         newThread = Threads_CloneTCB(Err, Flags);
 559         if(!newThread)  return -1;
 560
 561         // Initialise Memory Space (New Addr space or kernel stack)
 562         if(Flags & CLONE_VM) {
 563                 newThread->MemState.CR3 = MM_Clone();
 564                 newThread->KernelStack = cur->KernelStack;
 565         } else {
 566                 Uint    tmpEbp, oldEsp = esp;
 567
 568                 // Set CR3
 569                 newThread->MemState.CR3 = cur->MemState.CR3;
 570
 571                 // Create new KStack
 572                 newThread->KernelStack = MM_NewKStack();
 573                 // Check for errors
 574                 if(newThread->KernelStack == 0) {
 575                         free(newThread);
 576                         return -1;
 577                 }
 578
 579                 // Get ESP as a used size
 580                 esp = cur->KernelStack - esp;
 581                 // Copy used stack
 582                 memcpy( (void*)(newThread->KernelStack - esp), (void*)(cur->KernelStack - esp), esp );
 583                 // Get ESP as an offset in the new stack
 584                 esp = newThread->KernelStack - esp;
 585                 // Adjust EBP
 586                 ebp = newThread->KernelStack - (cur->KernelStack - ebp);
 587
 588                 // Repair EBPs & Stack Addresses
 589                 // Catches arguments also, but may trash stack-address-like values
 590                 for(tmpEbp = esp; tmpEbp < newThread->KernelStack; tmpEbp += 4)
 591                 {
 592                         if(oldEsp < *(Uint*)tmpEbp && *(Uint*)tmpEbp < cur->KernelStack)
 593                                 *(Uint*)tmpEbp += newThread->KernelStack - cur->KernelStack;
 594                 }
 595         }
 596
 597         // Save core machine state
 598         newThread->SavedState.ESP = esp;
 599         newThread->SavedState.EBP = ebp;
 600         eip = GetEIP();
 601         if(eip == SWITCH_MAGIC) {
 602                 outb(0x20, 0x20);       // ACK Timer and return as child
 603                 __asm__ __volatile__ ("sti");   // Restart interrupts
 604                 return 0;
 605         }
 606
 607         // Set EIP as parent
 608         newThread->SavedState.EIP = eip;
 609
 610         // Lock list and add to active
 611         Threads_AddActive(newThread);
 612
 613         return newThread->TID;
 614 }
 615
 616 /**
 617  * \fn int Proc_SpawnWorker(void)
 618  * \brief Spawns a new worker thread
 619  */
 620 int Proc_SpawnWorker(void)
 621 {
 622         tThread *new, *cur;
 623         Uint    eip, esp, ebp;
 624
 625         cur = Proc_GetCurThread();
 626
 627         // Create new thread
 628         new = malloc( sizeof(tThread) );
 629         if(!new) {
 630                 Warning("Proc_SpawnWorker - Out of heap space!\n");
 631                 return -1;
 632         }
 633         memcpy(new, &gThreadZero, sizeof(tThread));
 634         // Set Thread ID
 635         new->TID = giNextTID++;
 636         // Create a new worker stack (in PID0's address space)
 637         // The stack is relocated by this code
 638         new->KernelStack = MM_NewWorkerStack();
 639
 640         // Get ESP and EBP based in the new stack
 641         __asm__ __volatile__ ("mov %%esp, %0": "=r"(esp));
 642         __asm__ __volatile__ ("mov %%ebp, %0": "=r"(ebp));
 643         esp = new->KernelStack - (cur->KernelStack - esp);
 644         ebp = new->KernelStack - (cur->KernelStack - ebp);
 645
 646         // Save core machine state
 647         new->SavedState.ESP = esp;
 648         new->SavedState.EBP = ebp;
 649         eip = GetEIP();
 650         if(eip == SWITCH_MAGIC) {
 651                 outb(0x20, 0x20);       // ACK Timer and return as child
 652                 return 0;
 653         }
 654
 655         // Set EIP as parent
 656         new->SavedState.EIP = eip;
 657         // Mark as active
 658         new->Status = THREAD_STAT_ACTIVE;
 659         Threads_AddActive( new );
 660
 661         return new->TID;
 662 }
 663
 664 /**
 665  * \fn Uint Proc_MakeUserStack(void)
 666  * \brief Creates a new user stack
 667  */
 668 Uint Proc_MakeUserStack(void)
 669 {
 670          int    i;
 671         Uint    base = USER_STACK_TOP - USER_STACK_SZ;
 672
 673         // Check Prospective Space
 674         for( i = USER_STACK_SZ >> 12; i--; )
 675                 if( MM_GetPhysAddr( base + (i<<12) ) != 0 )
 676                         break;
 677
 678         if(i != -1)     return 0;
 679
 680         // Allocate Stack - Allocate incrementally to clean up MM_Dump output
 681         for( i = 0; i < USER_STACK_SZ/0x1000; i++ )
 682                 MM_Allocate( base + (i<<12) );
 683
 684         return base + USER_STACK_SZ;
 685 }
 686
 687 /**
 688  * \fn void Proc_StartUser(Uint Entrypoint, Uint *Bases, int ArgC, char **ArgV, char **EnvP, int DataSize)
 689  * \brief Starts a user task
 690  */
 691 void Proc_StartUser(Uint Entrypoint, Uint *Bases, int ArgC, char **ArgV, char **EnvP, int DataSize)
 692 {
 693         Uint    *stack = (void*)Proc_MakeUserStack();
 694          int    i;
 695         Uint    delta;
 696         Uint16  ss, cs;
 697
 698         //Log("stack = %p", stack);
 699
 700         // Copy Arguments
 701         stack -= DataSize/sizeof(*stack);
 702         memcpy( stack, ArgV, DataSize );
 703
 704         //Log("stack = %p", stack);
 705
 706         if( DataSize )
 707         {
 708                 // Adjust Arguments and environment
 709                 delta = (Uint)stack - (Uint)ArgV;
 710                 ArgV = (char**)stack;
 711                 for( i = 0; ArgV[i]; i++ )
 712                         ArgV[i] += delta;
 713                 i ++;
 714
 715                 // Do we care about EnvP?
 716                 if( EnvP ) {
 717                         EnvP = &ArgV[i];
 718                         for( i = 0; EnvP[i]; i++ )
 719                                 EnvP[i] += delta;
 720                 }
 721         }
 722
 723         // User Mode Segments
 724         ss = 0x23;      cs = 0x1B;
 725
 726         // Arguments
 727         *--stack = (Uint)EnvP;
 728         *--stack = (Uint)ArgV;
 729         *--stack = (Uint)ArgC;
 730         while(*Bases)
 731                 *--stack = *Bases++;
 732         *--stack = 0;   // Return Address
 733
 734         Proc_StartProcess(ss, (Uint)stack, 0x202, cs, Entrypoint);
 735 }
 736
 737 void Proc_StartProcess(Uint16 SS, Uint Stack, Uint Flags, Uint16 CS, Uint IP)
 738 {
 739         Uint    *stack = (void*)Stack;
 740         *--stack = SS;          //Stack Segment
 741         *--stack = Stack;       //Stack Pointer
 742         *--stack = Flags;       //EFLAGS (Resvd (0x2) and IF (0x20))
 743         *--stack = CS;          //Code Segment
 744         *--stack = IP;  //EIP
 745         //PUSHAD
 746         *--stack = 0xAAAAAAAA;  // eax
 747         *--stack = 0xCCCCCCCC;  // ecx
 748         *--stack = 0xDDDDDDDD;  // edx
 749         *--stack = 0xBBBBBBBB;  // ebx
 750         *--stack = 0xD1D1D1D1;  // edi
 751         *--stack = 0x54545454;  // esp - NOT POPED
 752         *--stack = 0x51515151;  // esi
 753         *--stack = 0xB4B4B4B4;  // ebp
 754         //Individual PUSHs
 755         *--stack = SS;  // ds
 756         *--stack = SS;  // es
 757         *--stack = SS;  // fs
 758         *--stack = SS;  // gs
 759
 760         __asm__ __volatile__ (
 761         "mov %%eax,%%esp;\n\t"  // Set stack pointer
 762         "pop %%gs;\n\t"
 763         "pop %%fs;\n\t"
 764         "pop %%es;\n\t"
 765         "pop %%ds;\n\t"
 766         "popa;\n\t"
 767         "iret;\n\t" : : "a" (stack));
 768         for(;;);
 769 }
 770
 771 /**
 772  * \fn int Proc_Demote(Uint *Err, int Dest, tRegs *Regs)
 773  * \brief Demotes a process to a lower permission level
 774  * \param Err   Pointer to user's errno
 775  * \param Dest  New Permission Level
 776  * \param Regs  Pointer to user's register structure
 777  */
 778 int Proc_Demote(Uint *Err, int Dest, tRegs *Regs)
 779 {
 780          int    cpl = Regs->cs & 3;
 781         // Sanity Check
 782         if(Dest > 3 || Dest < 0) {
 783                 *Err = -EINVAL;
 784                 return -1;
 785         }
 786
 787         // Permission Check
 788         if(cpl > Dest) {
 789                 *Err = -EACCES;
 790                 return -1;
 791         }
 792
 793         // Change the Segment Registers
 794         Regs->cs = (((Dest+1)<<4) | Dest) - 8;
 795         Regs->ss = ((Dest+1)<<4) | Dest;
 796         // Check if the GP Segs are GDT, then change them
 797         if(!(Regs->ds & 4))     Regs->ds = ((Dest+1)<<4) | Dest;
 798         if(!(Regs->es & 4))     Regs->es = ((Dest+1)<<4) | Dest;
 799         if(!(Regs->fs & 4))     Regs->fs = ((Dest+1)<<4) | Dest;
 800         if(!(Regs->gs & 4))     Regs->gs = ((Dest+1)<<4) | Dest;
 801
 802         return 0;
 803 }
 804
 805 /**
 806  * \brief Calls a signal handler in user mode
 807  * \note Used for signals
 808  */
 809 void Proc_CallFaultHandler(tThread *Thread)
 810 {
 811         // Rewinds the stack and calls the user function
 812         // Never returns
 813         __asm__ __volatile__ ("mov %0, %%ebp;\n\tcall Proc_ReturnToUser" :: "r"(Thread->FaultHandler));
 814         for(;;);
 815 }
 816
 817 /**
 818  * \fn void Proc_Scheduler(int CPU)
 819  * \brief Swap current thread and clears dead threads
 820  */
 821 void Proc_Scheduler(int CPU)
 822 {
 823         Uint    esp, ebp, eip;
 824         tThread *thread;
 825
 826         // If the spinlock is set, let it complete
 827         if(IS_LOCKED(&glThreadListLock))        return;
 828
 829         // Clear Delete Queue
 830         while(gDeleteThreads)
 831         {
 832                 thread = gDeleteThreads->Next;
 833                 if(gDeleteThreads->IsLocked) {  // Only free if structure is unused
 834                         gDeleteThreads->Status = THREAD_STAT_NULL;
 835                         free( gDeleteThreads );
 836                 }
 837                 gDeleteThreads = thread;
 838         }
 839
 840         // Check if there is any tasks running
 841         if(giNumActiveThreads == 0) {
 842                 #if 0
 843                 Log("No Active threads, sleeping");
 844                 #endif
 845                 #if USE_MP
 846                 if(CPU)
 847                         gpMP_LocalAPIC->EOI.Val = 0;
 848                 else
 849                 #endif
 850                         outb(0x20, 0x20);
 851                 __asm__ __volatile__ ("hlt");
 852                 return;
 853         }
 854
 855         // Get current thread
 856         #if USE_MP
 857         thread = gaCPUs[CPU].Current;
 858         #else
 859         thread = gCurrentThread;
 860         #endif
 861
 862         if( thread )
 863         {
 864                 // Reduce remaining quantum and continue timeslice if non-zero
 865                 if(thread->Remaining--) return;
 866                 // Reset quantum for next call
 867                 thread->Remaining = thread->Quantum;
 868
 869                 // Get machine state
 870                 __asm__ __volatile__ ("mov %%esp, %0":"=r"(esp));
 871                 __asm__ __volatile__ ("mov %%ebp, %0":"=r"(ebp));
 872                 eip = GetEIP();
 873                 if(eip == SWITCH_MAGIC) return; // Check if a switch happened
 874
 875                 // Save machine state
 876                 thread->SavedState.ESP = esp;
 877                 thread->SavedState.EBP = ebp;
 878                 thread->SavedState.EIP = eip;
 879         }
 880
 881         // Get next thread to run
 882         thread = Threads_GetNextToRun(CPU, thread);
 883
 884         // No avaliable tasks, just go into low power mode
 885         if(thread == NULL) {
 886                 //HALT();
 887                 //return;
 888                 #if USE_MP
 889                 thread = gaCPUs[CPU].IdleThread;
 890                 #else
 891                 thread = gpIdleThread;
 892                 #endif
 893         }
 894
 895         #if DEBUG_TRACE_SWITCH
 896         Log("Switching to task %i, CR3 = 0x%x, EIP = %p",
 897                 thread->TID,
 898                 thread->MemState.CR3,
 899                 thread->SavedState.EIP
 900                 );
 901         #endif
 902
 903         // Set current thread
 904         #if USE_MP
 905         gaCPUs[CPU].Current = thread;
 906         #else
 907         gCurrentThread = thread;
 908         #endif
 909
 910         //Log("CPU = %i", CPU);
 911
 912         // Update Kernel Stack pointer
 913         gTSSs[CPU].ESP0 = thread->KernelStack-4;
 914
 915         // Set address space
 916         #if USE_PAE
 917         # error "Todo: Implement PAE Address space switching"
 918         #else
 919                 __asm__ __volatile__ ("mov %0, %%cr3"::"a"(thread->MemState.CR3));
 920         #endif
 921
 922         #if 0
 923         if(thread->SavedState.ESP > 0xC0000000
 924         && thread->SavedState.ESP < thread->KernelStack-0x2000) {
 925                 Log_Warning("Proc", "Possible bad ESP %p (PID %i)", thread->SavedState.ESP);
 926         }
 927         #endif
 928
 929         // Switch threads
 930         __asm__ __volatile__ (
 931                 "mov %1, %%esp\n\t"     // Restore ESP
 932                 "mov %2, %%ebp\n\t"     // and EBP
 933                 "jmp *%3" : :   // And return to where we saved state (Proc_Clone or Proc_Scheduler)
 934                 "a"(SWITCH_MAGIC), "b"(thread->SavedState.ESP),
 935                 "d"(thread->SavedState.EBP), "c"(thread->SavedState.EIP)
 936                 );
 937         for(;;);        // Shouldn't reach here
 938 }
 939
 940 // === EXPORTS ===
 941 EXPORT(Proc_SpawnWorker);