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_PERIOD  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 }       tCPU;
  38
  39 // === IMPORTS ===
  40 extern tGDT     gGDT[];
  41 extern tIDT     gIDT[];
  42 extern void     APWait(void);   // 16-bit AP pause code
  43 extern void     APStartup(void);        // 16-bit AP startup code
  44 extern Uint     GetEIP(void);   // start.asm
  45 extern Uint     GetEIP_Sched(void);     // proc.asm
  46 extern void     NewTaskHeader(tThread *Thread, void *Fcn, int nArgs, ...);      // Actually takes cdecl args
  47 extern Uint     Proc_CloneInt(Uint *ESP, Uint32 *CR3);
  48 extern Uint32   gaInitPageDir[1024];    // start.asm
  49 extern char     Kernel_Stack_Top[];
  50 extern int      giNumCPUs;
  51 extern int      giNextTID;
  52 extern tThread  gThreadZero;
  53 extern void     Isr8(void);     // Double Fault
  54 extern void     Proc_ReturnToUser(tVAddr Handler, Uint Argument, tVAddr KernelStack);
  55 extern char     scheduler_return[];     // Return address in SchedulerBase
  56 extern char     IRQCommon[];    // Common IRQ handler code
  57 extern char     IRQCommon_handled[];    // IRQCommon call return location
  58 extern char     GetEIP_Sched_ret[];     // GetEIP call return location
  59 extern void     SwitchTasks(Uint NewSP, Uint *OldSP, Uint NewIP, Uint *OldIO, Uint CR3);
  60 extern void     Proc_InitialiseSSE(void);
  61 extern void     Proc_SaveSSE(Uint DestPtr);
  62 extern void     Proc_DisableSSE(void);
  63
  64 // === PROTOTYPES ===
  65 //void  ArchThreads_Init(void);
  66 #if USE_MP
  67 void    MP_StartAP(int CPU);
  68 void    MP_SendIPIVector(int CPU, Uint8 Vector);
  69 void    MP_SendIPI(Uint8 APICID, int Vector, int DeliveryMode);
  70 #endif
  71 void    Proc_IdleThread(void *Ptr);
  72 //void  Proc_Start(void);
  73 //tThread       *Proc_GetCurThread(void);
  74 void    Proc_ChangeStack(void);
  75 // int  Proc_NewKThread(void (*Fcn)(void*), void *Data);
  76 // int  Proc_Clone(Uint *Err, Uint Flags);
  77 Uint    Proc_MakeUserStack(void);
  78 //void  Proc_StartUser(Uint Entrypoint, Uint *Bases, int ArgC, char **ArgV, char **EnvP, int DataSize);
  79 void    Proc_StartProcess(Uint16 SS, Uint Stack, Uint Flags, Uint16 CS, Uint IP) NORETURN;
  80  int    Proc_Demote(Uint *Err, int Dest, tRegs *Regs);
  81 //void  Proc_CallFaultHandler(tThread *Thread);
  82 //void  Proc_DumpThreadCPUState(tThread *Thread);
  83 void    Proc_Scheduler(int CPU);
  84
  85 // === GLOBALS ===
  86 // --- Multiprocessing ---
  87 #if USE_MP
  88 volatile int    giNumInitingCPUs = 0;
  89 tMPInfo *gMPFloatPtr = NULL;
  90 volatile Uint32 giMP_TimerCount;        // Start Count for Local APIC Timer
  91 tAPIC   *gpMP_LocalAPIC = NULL;
  92 Uint8   gaAPIC_to_CPU[256] = {0};
  93  int    giProc_BootProcessorID = 0;
  94 tTSS    gaTSSs[MAX_CPUS];       // TSS Array
  95 #endif
  96 tCPU    gaCPUs[MAX_CPUS] = {
  97         {.Current = &gThreadZero}
  98         };
  99 tTSS    *gTSSs = NULL;  // Pointer to TSS array
 100 tTSS    gTSS0 = {0};
 101 // --- Error Recovery ---
 102 char    gaDoubleFaultStack[1024] __attribute__ ((section(".padata")));
 103 tTSS    gDoubleFault_TSS = {
 104         .ESP0 = (Uint)&gaDoubleFaultStack[1024],
 105         .SS0 = 0x10,
 106         .CR3 = (Uint)gaInitPageDir - KERNEL_BASE,
 107         .EIP = (Uint)Isr8,
 108         .ESP = (Uint)&gaDoubleFaultStack[1024],
 109         .CS = 0x08,     .SS = 0x10,
 110         .DS = 0x10,     .ES = 0x10,
 111         .FS = 0x10,     .GS = 0x10,
 112 };
 113
 114 // === CODE ===
 115 /**
 116  * \fn void ArchThreads_Init(void)
 117  * \brief Starts the process scheduler
 118  */
 119 void ArchThreads_Init(void)
 120 {
 121         Uint    pos = 0;
 122
 123         #if USE_MP
 124         tMPTable        *mptable;
 125
 126         // Mark BSP as active
 127         gaCPUs[0].State = 2;
 128
 129         // -- Initialise Multiprocessing
 130         // Find MP Floating Table
 131         // - EBDA/Last 1Kib (640KiB)
 132         for(pos = KERNEL_BASE|0x9F000; pos < (KERNEL_BASE|0xA0000); pos += 16) {
 133                 if( *(Uint*)(pos) == MPPTR_IDENT ) {
 134                         Log("Possible %p", pos);
 135                         if( ByteSum((void*)pos, sizeof(tMPInfo)) != 0 ) continue;
 136                         gMPFloatPtr = (void*)pos;
 137                         break;
 138                 }
 139         }
 140         // - Last KiB (512KiB base mem)
 141         if(!gMPFloatPtr) {
 142                 for(pos = KERNEL_BASE|0x7F000; pos < (KERNEL_BASE|0x80000); pos += 16) {
 143                         if( *(Uint*)(pos) == MPPTR_IDENT ) {
 144                                 Log("Possible %p", pos);
 145                                 if( ByteSum((void*)pos, sizeof(tMPInfo)) != 0 ) continue;
 146                                 gMPFloatPtr = (void*)pos;
 147                                 break;
 148                         }
 149                 }
 150         }
 151         // - BIOS ROM
 152         if(!gMPFloatPtr) {
 153                 for(pos = KERNEL_BASE|0xE0000; pos < (KERNEL_BASE|0x100000); pos += 16) {
 154                         if( *(Uint*)(pos) == MPPTR_IDENT ) {
 155                                 Log("Possible %p", pos);
 156                                 if( ByteSum((void*)pos, sizeof(tMPInfo)) != 0 ) continue;
 157                                 gMPFloatPtr = (void*)pos;
 158                                 break;
 159                         }
 160                 }
 161         }
 162
 163         // If the MP Table Exists, parse it
 164         if(gMPFloatPtr)
 165         {
 166                  int    i;
 167                 tMPTable_Ent    *ents;
 168                 #if DUMP_MP_TABLE
 169                 Log("gMPFloatPtr = %p", gMPFloatPtr);
 170                 Log("*gMPFloatPtr = {");
 171                 Log("\t.Sig = 0x%08x", gMPFloatPtr->Sig);
 172                 Log("\t.MPConfig = 0x%08x", gMPFloatPtr->MPConfig);
 173                 Log("\t.Length = 0x%02x", gMPFloatPtr->Length);
 174                 Log("\t.Version = 0x%02x", gMPFloatPtr->Version);
 175                 Log("\t.Checksum = 0x%02x", gMPFloatPtr->Checksum);
 176                 Log("\t.Features = [0x%02x,0x%02x,0x%02x,0x%02x,0x%02x]",
 177                         gMPFloatPtr->Features[0],       gMPFloatPtr->Features[1],
 178                         gMPFloatPtr->Features[2],       gMPFloatPtr->Features[3],
 179                         gMPFloatPtr->Features[4]
 180                         );
 181                 Log("}");
 182                 #endif
 183
 184                 mptable = (void*)( KERNEL_BASE|gMPFloatPtr->MPConfig );
 185                 #if DUMP_MP_TABLE
 186                 Log("mptable = %p", mptable);
 187                 Log("*mptable = {");
 188                 Log("\t.Sig = 0x%08x", mptable->Sig);
 189                 Log("\t.BaseTableLength = 0x%04x", mptable->BaseTableLength);
 190                 Log("\t.SpecRev = 0x%02x", mptable->SpecRev);
 191                 Log("\t.Checksum = 0x%02x", mptable->Checksum);
 192                 Log("\t.OEMID = '%8c'", mptable->OemID);
 193                 Log("\t.ProductID = '%8c'", mptable->ProductID);
 194                 Log("\t.OEMTablePtr = %p'", mptable->OEMTablePtr);
 195                 Log("\t.OEMTableSize = 0x%04x", mptable->OEMTableSize);
 196                 Log("\t.EntryCount = 0x%04x", mptable->EntryCount);
 197                 Log("\t.LocalAPICMemMap = 0x%08x", mptable->LocalAPICMemMap);
 198                 Log("\t.ExtendedTableLen = 0x%04x", mptable->ExtendedTableLen);
 199                 Log("\t.ExtendedTableChecksum = 0x%02x", mptable->ExtendedTableChecksum);
 200                 Log("}");
 201                 #endif
 202
 203                 gpMP_LocalAPIC = (void*)MM_MapHWPages(mptable->LocalAPICMemMap, 1);
 204
 205                 ents = mptable->Entries;
 206                 giNumCPUs = 0;
 207
 208                 for( i = 0; i < mptable->EntryCount; i ++ )
 209                 {
 210                          int    entSize = 0;
 211                         switch( ents->Type )
 212                         {
 213                         case 0: // Processor
 214                                 entSize = 20;
 215                                 #if DUMP_MP_TABLE
 216                                 Log("%i: Processor", i);
 217                                 Log("\t.APICID = %i", ents->Proc.APICID);
 218                                 Log("\t.APICVer = 0x%02x", ents->Proc.APICVer);
 219                                 Log("\t.CPUFlags = 0x%02x", ents->Proc.CPUFlags);
 220                                 Log("\t.CPUSignature = 0x%08x", ents->Proc.CPUSignature);
 221                                 Log("\t.FeatureFlags = 0x%08x", ents->Proc.FeatureFlags);
 222                                 #endif
 223
 224                                 if( !(ents->Proc.CPUFlags & 1) ) {
 225                                         Log("DISABLED");
 226                                         break;
 227                                 }
 228
 229                                 // Check if there is too many processors
 230                                 if(giNumCPUs >= MAX_CPUS) {
 231                                         giNumCPUs ++;   // If `giNumCPUs` > MAX_CPUS later, it will be clipped
 232                                         break;
 233                                 }
 234
 235                                 // Initialise CPU Info
 236                                 gaAPIC_to_CPU[ents->Proc.APICID] = giNumCPUs;
 237                                 gaCPUs[giNumCPUs].APICID = ents->Proc.APICID;
 238                                 gaCPUs[giNumCPUs].State = 0;
 239                                 giNumCPUs ++;
 240
 241                                 // Set BSP Variable
 242                                 if( ents->Proc.CPUFlags & 2 ) {
 243                                         giProc_BootProcessorID = giNumCPUs-1;
 244                                 }
 245
 246                                 break;
 247
 248                         #if DUMP_MP_TABLE >= 2
 249                         case 1: // Bus
 250                                 entSize = 8;
 251                                 Log("%i: Bus", i);
 252                                 Log("\t.ID = %i", ents->Bus.ID);
 253                                 Log("\t.TypeString = '%6C'", ents->Bus.TypeString);
 254                                 break;
 255                         case 2: // I/O APIC
 256                                 entSize = 8;
 257                                 Log("%i: I/O APIC", i);
 258                                 Log("\t.ID = %i", ents->IOAPIC.ID);
 259                                 Log("\t.Version = 0x%02x", ents->IOAPIC.Version);
 260                                 Log("\t.Flags = 0x%02x", ents->IOAPIC.Flags);
 261                                 Log("\t.Addr = 0x%08x", ents->IOAPIC.Addr);
 262                                 break;
 263                         case 3: // I/O Interrupt Assignment
 264                                 entSize = 8;
 265                                 Log("%i: I/O Interrupt Assignment", i);
 266                                 Log("\t.IntType = %i", ents->IOInt.IntType);
 267                                 Log("\t.Flags = 0x%04x", ents->IOInt.Flags);
 268                                 Log("\t.SourceBusID = 0x%02x", ents->IOInt.SourceBusID);
 269                                 Log("\t.SourceBusIRQ = 0x%02x", ents->IOInt.SourceBusIRQ);
 270                                 Log("\t.DestAPICID = 0x%02x", ents->IOInt.DestAPICID);
 271                                 Log("\t.DestAPICIRQ = 0x%02x", ents->IOInt.DestAPICIRQ);
 272                                 break;
 273                         case 4: // Local Interrupt Assignment
 274                                 entSize = 8;
 275                                 Log("%i: Local Interrupt Assignment", i);
 276                                 Log("\t.IntType = %i", ents->LocalInt.IntType);
 277                                 Log("\t.Flags = 0x%04x", ents->LocalInt.Flags);
 278                                 Log("\t.SourceBusID = 0x%02x", ents->LocalInt.SourceBusID);
 279                                 Log("\t.SourceBusIRQ = 0x%02x", ents->LocalInt.SourceBusIRQ);
 280                                 Log("\t.DestLocalAPICID = 0x%02x", ents->LocalInt.DestLocalAPICID);
 281                                 Log("\t.DestLocalAPICIRQ = 0x%02x", ents->LocalInt.DestLocalAPICIRQ);
 282                                 break;
 283                         default:
 284                                 Log("%i: Unknown (%i)", i, ents->Type);
 285                                 break;
 286                         #endif
 287                         }
 288                         ents = (void*)( (Uint)ents + entSize );
 289                 }
 290
 291                 if( giNumCPUs > MAX_CPUS ) {
 292                         Warning("Too many CPUs detected (%i), only using %i of them", giNumCPUs, MAX_CPUS);
 293                         giNumCPUs = MAX_CPUS;
 294                 }
 295                 gTSSs = gaTSSs;
 296         }
 297         else {
 298                 Log("No MP Table was found, assuming uniprocessor\n");
 299                 giNumCPUs = 1;
 300                 gTSSs = &gTSS0;
 301         }
 302         #else
 303         giNumCPUs = 1;
 304         gTSSs = &gTSS0;
 305         #endif
 306
 307         #if !DEBUG_DISABLE_DOUBLEFAULT
 308         // Initialise Double Fault TSS
 309         gGDT[5].BaseLow = (Uint)&gDoubleFault_TSS & 0xFFFF;
 310         gGDT[5].BaseMid = (Uint)&gDoubleFault_TSS >> 16;
 311         gGDT[5].BaseHi = (Uint)&gDoubleFault_TSS >> 24;
 312
 313         // Set double fault IDT to use the new TSS
 314         gIDT[8].OffsetLo = 0;
 315         gIDT[8].CS = 5<<3;
 316         gIDT[8].Flags = 0x8500;
 317         gIDT[8].OffsetHi = 0;
 318         #endif
 319
 320         // Set timer frequency
 321         outb(0x43, 0x34);       // Set Channel 0, Low/High, Rate Generator
 322         outb(0x40, TIMER_DIVISOR&0xFF); // Low Byte of Divisor
 323         outb(0x40, (TIMER_DIVISOR>>8)&0xFF);    // High Byte
 324
 325         Log_Debug("Proc", "PIT Frequency %i.%03i Hz",
 326                 TIMER_BASE/TIMER_DIVISOR,
 327                 ((Uint64)TIMER_BASE*1000/TIMER_DIVISOR)%1000
 328                 );
 329
 330         #if USE_MP
 331         // Get the count setting for APIC timer
 332         Log("Determining APIC Count");
 333         __asm__ __volatile__ ("sti");
 334         while( giMP_TimerCount == 0 )   __asm__ __volatile__ ("hlt");
 335         __asm__ __volatile__ ("cli");
 336         Log("APIC Count %i", giMP_TimerCount);
 337         {
 338                 Uint64  freq = giMP_TimerCount;
 339                 freq *= TIMER_BASE;
 340                 freq /= TIMER_DIVISOR;
 341                 if( (freq /= 1000) < 2*1000)
 342                         Log("Bus Frequency %i KHz", freq);
 343                 else if( (freq /= 1000) < 2*1000)
 344                         Log("Bus Frequency %i MHz", freq);
 345                 else if( (freq /= 1000) < 2*1000)
 346                         Log("Bus Frequency %i GHz", freq);
 347                 else
 348                         Log("Bus Frequency %i THz", freq);
 349         }
 350
 351         // Initialise Normal TSS(s)
 352         for(pos=0;pos<giNumCPUs;pos++)
 353         {
 354         #else
 355         pos = 0;
 356         #endif
 357                 gTSSs[pos].SS0 = 0x10;
 358                 gTSSs[pos].ESP0 = 0;    // Set properly by scheduler
 359                 gGDT[6+pos].BaseLow = ((Uint)(&gTSSs[pos])) & 0xFFFF;
 360                 gGDT[6+pos].BaseMid = ((Uint)(&gTSSs[pos]) >> 16) & 0xFFFF;
 361                 gGDT[6+pos].BaseHi = ((Uint)(&gTSSs[pos])) >> 24;
 362         #if USE_MP
 363         }
 364         #endif
 365
 366         // Load the BSP's TSS
 367         __asm__ __volatile__ ("ltr %%ax"::"a"(0x30));
 368         // Set Current Thread and CPU Number in DR0 and DR1
 369         __asm__ __volatile__ ("mov %0, %%db0"::"r"(&gThreadZero));
 370         __asm__ __volatile__ ("mov %0, %%db1"::"r"(0));
 371
 372         gaCPUs[0].Current = &gThreadZero;
 373         gThreadZero.CurCPU = 0;
 374
 375         gThreadZero.MemState.CR3 = (Uint)gaInitPageDir - KERNEL_BASE;
 376
 377         // Create Per-Process Data Block
 378         if( !MM_Allocate(MM_PPD_CFG) )
 379         {
 380                 Panic("OOM - No space for initial Per-Process Config");
 381         }
 382
 383         // Initialise SSE support
 384         Proc_InitialiseSSE();
 385
 386         // Change Stacks
 387         Proc_ChangeStack();
 388 }
 389
 390 #if USE_MP
 391 /**
 392  * \brief Start an AP
 393  */
 394 void MP_StartAP(int CPU)
 395 {
 396         Log_Log("Proc", "Starting AP %i (APIC %i)", CPU, gaCPUs[CPU].APICID);
 397
 398         // Set location of AP startup code and mark for a warm restart
 399         *(Uint16*)(KERNEL_BASE|0x467) = (Uint)&APWait - (KERNEL_BASE|0xFFFF0);
 400         *(Uint16*)(KERNEL_BASE|0x469) = 0xFFFF;
 401         outb(0x70, 0x0F);       outb(0x71, 0x0A);       // Set warm reset flag
 402         MP_SendIPI(gaCPUs[CPU].APICID, 0, 5);   // Init IPI
 403
 404         // Delay
 405         inb(0x80); inb(0x80); inb(0x80); inb(0x80);
 406
 407         // TODO: Use a better address, preferably registered with the MM
 408         // - MM_AllocDMA mabye?
 409         // Create a far jump
 410         *(Uint8*)(KERNEL_BASE|0x11000) = 0xEA;  // Far JMP
 411         *(Uint16*)(KERNEL_BASE|0x11001) = (Uint)&APStartup - (KERNEL_BASE|0xFFFF0);     // IP
 412         *(Uint16*)(KERNEL_BASE|0x11003) = 0xFFFF;       // CS
 413         // Send a Startup-IPI to make the CPU execute at 0x11000 (which we
 414         // just filled)
 415         MP_SendIPI(gaCPUs[CPU].APICID, 0x11, 6);        // StartupIPI
 416
 417         giNumInitingCPUs ++;
 418 }
 419
 420 void MP_SendIPIVector(int CPU, Uint8 Vector)
 421 {
 422         MP_SendIPI(gaCPUs[CPU].APICID, Vector, 0);
 423 }
 424
 425 /**
 426  * \brief Send an Inter-Processor Interrupt
 427  * \param APICID        Processor's Local APIC ID
 428  * \param Vector        Argument of some kind
 429  * \param DeliveryMode  Type of signal
 430  * \note 3A 10.5 "APIC/Handling Local Interrupts"
 431  */
 432 void MP_SendIPI(Uint8 APICID, int Vector, int DeliveryMode)
 433 {
 434         Uint32  val;
 435
 436         // Hi
 437         val = (Uint)APICID << 24;
 438 //      Log("%p = 0x%08x", &gpMP_LocalAPIC->ICR[1], val);
 439         gpMP_LocalAPIC->ICR[1].Val = val;
 440         // Low (and send)
 441         val = ((DeliveryMode & 7) << 8) | (Vector & 0xFF);
 442 //      Log("%p = 0x%08x", &gpMP_LocalAPIC->ICR[0], val);
 443         gpMP_LocalAPIC->ICR[0].Val = val;
 444 }
 445 #endif
 446
 447 void Proc_IdleThread(void *Ptr)
 448 {
 449         tCPU    *cpu = &gaCPUs[GetCPUNum()];
 450         cpu->Current->ThreadName = strdup("Idle Thread");
 451         Threads_SetPriority( cpu->Current, -1 );        // Never called randomly
 452         cpu->Current->Quantum = 1;      // 1 slice quantum
 453         for(;;) {
 454                 __asm__ __volatile__ ("sti");   // Make sure interrupts are enabled
 455                 __asm__ __volatile__ ("hlt");   // Make sure interrupts are enabled
 456                 Proc_Reschedule();
 457         }
 458 }
 459
 460 /**
 461  * \fn void Proc_Start(void)
 462  * \brief Start process scheduler
 463  */
 464 void Proc_Start(void)
 465 {
 466          int    tid;
 467         #if USE_MP
 468          int    i;
 469         #endif
 470
 471         #if USE_MP
 472         // Start APs
 473         for( i = 0; i < giNumCPUs; i ++ )
 474         {
 475                 if(i)   gaCPUs[i].Current = NULL;
 476
 477                 // Create Idle Task
 478                 tid = Proc_NewKThread(Proc_IdleThread, &gaCPUs[i]);
 479
 480                 // Start the AP
 481                 if( i != giProc_BootProcessorID ) {
 482                         MP_StartAP( i );
 483                 }
 484         }
 485
 486         // BSP still should run the current task
 487         gaCPUs[0].Current = &gThreadZero;
 488
 489         // Start interrupts and wait for APs to come up
 490         Log_Debug("Proc", "Waiting for APs to come up");
 491         __asm__ __volatile__ ("sti");
 492         while( giNumInitingCPUs )       __asm__ __volatile__ ("hlt");
 493         #else
 494         // Create Idle Task
 495         tid = Proc_NewKThread(Proc_IdleThread, &gaCPUs[0]);
 496 //      gaCPUs[0].IdleThread = Threads_GetThread(tid);
 497
 498         // Set current task
 499         gaCPUs[0].Current = &gThreadZero;
 500
 501         // Start Interrupts (and hence scheduler)
 502         __asm__ __volatile__("sti");
 503         #endif
 504         MM_FinishVirtualInit();
 505 }
 506
 507 /**
 508  * \fn tThread *Proc_GetCurThread(void)
 509  * \brief Gets the current thread
 510  */
 511 tThread *Proc_GetCurThread(void)
 512 {
 513         #if USE_MP
 514         return gaCPUs[ GetCPUNum() ].Current;
 515         #else
 516         return gaCPUs[ 0 ].Current;
 517         #endif
 518 }
 519
 520 /**
 521  * \fn void Proc_ChangeStack(void)
 522  * \brief Swaps the current stack for a new one (in the proper stack reigon)
 523  */
 524 void Proc_ChangeStack(void)
 525 {
 526         Uint    esp, ebp;
 527         Uint    tmpEbp, oldEsp;
 528         Uint    curBase, newBase;
 529
 530         __asm__ __volatile__ ("mov %%esp, %0":"=r"(esp));
 531         __asm__ __volatile__ ("mov %%ebp, %0":"=r"(ebp));
 532
 533         oldEsp = esp;
 534
 535         // Create new KStack
 536         newBase = MM_NewKStack();
 537         // Check for errors
 538         if(newBase == 0) {
 539                 Panic("What the?? Unable to allocate space for initial kernel stack");
 540                 return;
 541         }
 542
 543         curBase = (Uint)&Kernel_Stack_Top;
 544
 545         LOG("curBase = 0x%x, newBase = 0x%x", curBase, newBase);
 546
 547         // Get ESP as a used size
 548         esp = curBase - esp;
 549         LOG("memcpy( %p, %p, 0x%x )", (void*)(newBase - esp), (void*)(curBase - esp), esp );
 550         // Copy used stack
 551         memcpy( (void*)(newBase - esp), (void*)(curBase - esp), esp );
 552         // Get ESP as an offset in the new stack
 553         esp = newBase - esp;
 554         // Adjust EBP
 555         ebp = newBase - (curBase - ebp);
 556
 557         // Repair EBPs & Stack Addresses
 558         // Catches arguments also, but may trash stack-address-like values
 559         for(tmpEbp = esp; tmpEbp < newBase; tmpEbp += 4)
 560         {
 561                 if(oldEsp < *(Uint*)tmpEbp && *(Uint*)tmpEbp < curBase)
 562                         *(Uint*)tmpEbp += newBase - curBase;
 563         }
 564
 565         Proc_GetCurThread()->KernelStack = newBase;
 566
 567         __asm__ __volatile__ ("mov %0, %%esp"::"r"(esp));
 568         __asm__ __volatile__ ("mov %0, %%ebp"::"r"(ebp));
 569 }
 570
 571 void Proc_ClearThread(tThread *Thread)
 572 {
 573         Log_Warning("Proc", "TODO: Nuke address space etc");
 574         if(Thread->SavedState.SSE) {
 575                 free(Thread->SavedState.SSE);
 576                 Thread->SavedState.SSE = NULL;
 577         }
 578 }
 579
 580 int Proc_NewKThread(void (*Fcn)(void*), void *Data)
 581 {
 582         Uint    esp;
 583         tThread *newThread, *cur;
 584
 585         cur = Proc_GetCurThread();
 586         newThread = Threads_CloneTCB(0);
 587         if(!newThread)  return -1;
 588
 589         // Set CR3
 590         newThread->MemState.CR3 = cur->MemState.CR3;
 591
 592         // Create new KStack
 593         newThread->KernelStack = MM_NewKStack();
 594         // Check for errors
 595         if(newThread->KernelStack == 0) {
 596                 free(newThread);
 597                 return -1;
 598         }
 599
 600         esp = newThread->KernelStack;
 601         *(Uint*)(esp-=4) = (Uint)Data;  // Data (shadowed)
 602         *(Uint*)(esp-=4) = 1;   // Number of params
 603         *(Uint*)(esp-=4) = (Uint)Fcn;   // Function to call
 604         *(Uint*)(esp-=4) = (Uint)newThread;     // Thread ID
 605
 606         newThread->SavedState.ESP = esp;
 607         newThread->SavedState.EIP = (Uint)&NewTaskHeader;
 608         newThread->SavedState.SSE = NULL;
 609 //      Log("New (KThread) %p, esp = %p", newThread->SavedState.EIP, newThread->SavedState.ESP);
 610
 611 //      MAGIC_BREAK();
 612         Threads_AddActive(newThread);
 613
 614         return newThread->TID;
 615 }
 616
 617 /**
 618  * \fn int Proc_Clone(Uint *Err, Uint Flags)
 619  * \brief Clone the current process
 620  */
 621 int Proc_Clone(Uint Flags)
 622 {
 623         tThread *newThread;
 624         tThread *cur = Proc_GetCurThread();
 625         Uint    eip;
 626
 627         // Sanity, please
 628         if( !(Flags & CLONE_VM) ) {
 629                 Log_Error("Proc", "Proc_Clone: Don't leave CLONE_VM unset, use Proc_NewKThread instead");
 630                 return -1;
 631         }
 632
 633         // New thread
 634         newThread = Threads_CloneTCB(Flags);
 635         if(!newThread)  return -1;
 636
 637         newThread->KernelStack = cur->KernelStack;
 638
 639         // Clone state
 640         eip = Proc_CloneInt(&newThread->SavedState.ESP, &newThread->MemState.CR3);
 641         if( eip == 0 ) {
 642                 // ACK the interrupt
 643                 return 0;
 644         }
 645         newThread->SavedState.EIP = eip;
 646         newThread->SavedState.SSE = NULL;
 647         newThread->SavedState.bSSEModified = 0;
 648
 649         // Check for errors
 650         if( newThread->MemState.CR3 == 0 ) {
 651                 Log_Error("Proc", "Proc_Clone: MM_Clone failed");
 652                 Threads_Delete(newThread);
 653                 return -1;
 654         }
 655
 656         // Add the new thread to the run queue
 657         Threads_AddActive(newThread);
 658         return newThread->TID;
 659 }
 660
 661 /**
 662  * \fn int Proc_SpawnWorker(void)
 663  * \brief Spawns a new worker thread
 664  */
 665 int Proc_SpawnWorker(void (*Fcn)(void*), void *Data)
 666 {
 667         tThread *new;
 668         Uint    stack_contents[4];
 669
 670         // Create new thread
 671         new = Threads_CloneThreadZero();
 672         if(!new) {
 673                 Warning("Proc_SpawnWorker - Out of heap space!\n");
 674                 return -1;
 675         }
 676
 677         // Create the stack contents
 678         stack_contents[3] = (Uint)Data;
 679         stack_contents[2] = 1;
 680         stack_contents[1] = (Uint)Fcn;
 681         stack_contents[0] = (Uint)new;
 682
 683         // Create a new worker stack (in PID0's address space)
 684         new->KernelStack = MM_NewWorkerStack(stack_contents, sizeof(stack_contents));
 685
 686         // Save core machine state
 687         new->SavedState.ESP = new->KernelStack - sizeof(stack_contents);
 688         new->SavedState.EIP = (Uint)NewTaskHeader;
 689         new->SavedState.SSE = NULL;
 690         new->SavedState.bSSEModified = 0;
 691
 692         // Mark as active
 693         new->Status = THREAD_STAT_PREINIT;
 694         Threads_AddActive( new );
 695
 696         return new->TID;
 697 }
 698
 699 /**
 700  * \fn Uint Proc_MakeUserStack(void)
 701  * \brief Creates a new user stack
 702  */
 703 Uint Proc_MakeUserStack(void)
 704 {
 705          int    i;
 706         Uint    base = USER_STACK_TOP - USER_STACK_SZ;
 707
 708         // Check Prospective Space
 709         for( i = USER_STACK_SZ >> 12; i--; )
 710                 if( MM_GetPhysAddr( base + (i<<12) ) != 0 )
 711                         break;
 712
 713         if(i != -1)     return 0;
 714
 715         // Allocate Stack - Allocate incrementally to clean up MM_Dump output
 716         for( i = 0; i < USER_STACK_SZ/0x1000; i++ )
 717         {
 718                 if( !MM_Allocate( base + (i<<12) ) )
 719                 {
 720                         Warning("OOM: Proc_MakeUserStack");
 721                         return 0;
 722                 }
 723         }
 724
 725         return base + USER_STACK_SZ;
 726 }
 727
 728 void Proc_StartUser(Uint Entrypoint, Uint Base, int ArgC, char **ArgV, int DataSize)
 729 {
 730         Uint    *stack;
 731          int    i;
 732         char    **envp = NULL;
 733         Uint16  ss, cs;
 734
 735         // Copy data to the user stack and free original buffer
 736         stack = (void*)Proc_MakeUserStack();
 737         stack -= DataSize/sizeof(*stack);
 738         memcpy( stack, ArgV, DataSize );
 739         free(ArgV);
 740
 741         // Adjust Arguments and environment
 742         if( DataSize )
 743         {
 744                 Uint delta = (Uint)stack - (Uint)ArgV;
 745                 ArgV = (char**)stack;
 746                 for( i = 0; ArgV[i]; i++ )      ArgV[i] += delta;
 747                 envp = &ArgV[i+1];
 748                 for( i = 0; envp[i]; i++ )      envp[i] += delta;
 749         }
 750
 751         // User Mode Segments
 752         ss = 0x23;      cs = 0x1B;
 753
 754         // Arguments
 755         *--stack = (Uint)envp;
 756         *--stack = (Uint)ArgV;
 757         *--stack = (Uint)ArgC;
 758         *--stack = Base;
 759
 760         Proc_StartProcess(ss, (Uint)stack, 0x202, cs, Entrypoint);
 761 }
 762
 763 void Proc_StartProcess(Uint16 SS, Uint Stack, Uint Flags, Uint16 CS, Uint IP)
 764 {
 765         Uint    *stack = (void*)Stack;
 766         *--stack = SS;          //Stack Segment
 767         *--stack = Stack;       //Stack Pointer
 768         *--stack = Flags;       //EFLAGS (Resvd (0x2) and IF (0x20))
 769         *--stack = CS;          //Code Segment
 770         *--stack = IP;  //EIP
 771         //PUSHAD
 772         *--stack = 0xAAAAAAAA;  // eax
 773         *--stack = 0xCCCCCCCC;  // ecx
 774         *--stack = 0xDDDDDDDD;  // edx
 775         *--stack = 0xBBBBBBBB;  // ebx
 776         *--stack = 0xD1D1D1D1;  // edi
 777         *--stack = 0x54545454;  // esp - NOT POPED
 778         *--stack = 0x51515151;  // esi
 779         *--stack = 0xB4B4B4B4;  // ebp
 780         //Individual PUSHs
 781         *--stack = SS;  // ds
 782         *--stack = SS;  // es
 783         *--stack = SS;  // fs
 784         *--stack = SS;  // gs
 785
 786         __asm__ __volatile__ (
 787         "mov %%eax,%%esp;\n\t"  // Set stack pointer
 788         "pop %%gs;\n\t"
 789         "pop %%fs;\n\t"
 790         "pop %%es;\n\t"
 791         "pop %%ds;\n\t"
 792         "popa;\n\t"
 793         "iret;\n\t" : : "a" (stack));
 794         for(;;);
 795 }
 796
 797 /**
 798  * \fn int Proc_Demote(Uint *Err, int Dest, tRegs *Regs)
 799  * \brief Demotes a process to a lower permission level
 800  * \param Err   Pointer to user's errno
 801  * \param Dest  New Permission Level
 802  * \param Regs  Pointer to user's register structure
 803  */
 804 int Proc_Demote(Uint *Err, int Dest, tRegs *Regs)
 805 {
 806          int    cpl = Regs->cs & 3;
 807         // Sanity Check
 808         if(Dest > 3 || Dest < 0) {
 809                 *Err = -EINVAL;
 810                 return -1;
 811         }
 812
 813         // Permission Check
 814         if(cpl > Dest) {
 815                 *Err = -EACCES;
 816                 return -1;
 817         }
 818
 819         // Change the Segment Registers
 820         Regs->cs = (((Dest+1)<<4) | Dest) - 8;
 821         Regs->ss = ((Dest+1)<<4) | Dest;
 822         // Check if the GP Segs are GDT, then change them
 823         if(!(Regs->ds & 4))     Regs->ds = ((Dest+1)<<4) | Dest;
 824         if(!(Regs->es & 4))     Regs->es = ((Dest+1)<<4) | Dest;
 825         if(!(Regs->fs & 4))     Regs->fs = ((Dest+1)<<4) | Dest;
 826         if(!(Regs->gs & 4))     Regs->gs = ((Dest+1)<<4) | Dest;
 827
 828         return 0;
 829 }
 830
 831 /**
 832  * \brief Calls a signal handler in user mode
 833  * \note Used for signals
 834  */
 835 void Proc_CallFaultHandler(tThread *Thread)
 836 {
 837         // Rewinds the stack and calls the user function
 838         // Never returns
 839         Proc_ReturnToUser( Thread->FaultHandler, Thread->CurFaultNum, Thread->KernelStack );
 840         for(;;);
 841 }
 842
 843 void Proc_DumpThreadCPUState(tThread *Thread)
 844 {
 845         if( Thread->CurCPU > -1 )
 846         {
 847                  int    maxBacktraceDistance = 6;
 848                 tRegs   *regs = NULL;
 849                 Uint32  *stack;
 850
 851                 if( Thread->CurCPU != GetCPUNum() ) {
 852                         Log("  Currently running");
 853                         return ;
 854                 }
 855
 856                 // Backtrace to find the IRQ entrypoint
 857                 // - This will usually only be called by an IRQ, so this should
 858                 //   work
 859                 __asm__ __volatile__ ("mov %%ebp, %0" : "=r" (stack));
 860                 while( maxBacktraceDistance -- )
 861                 {
 862                         // [ebp] = oldEbp
 863                         // [ebp+4] = retaddr
 864
 865                         if( stack[1] == (tVAddr)&IRQCommon_handled ) {
 866                                 regs = (void*)stack[2];
 867                                 break;
 868                         }
 869
 870                         stack = (void*)stack[0];
 871                 }
 872
 873                 if( !regs ) {
 874                         Log("  Unable to find IRQ Entry");
 875                         return ;
 876                 }
 877
 878                 Log("  at %04x:%08x", regs->cs, regs->eip);
 879                 return ;
 880         }
 881
 882         tVAddr  diffFromScheduler = Thread->SavedState.EIP - (tVAddr)SwitchTasks;
 883         tVAddr  diffFromClone = Thread->SavedState.EIP - (tVAddr)Proc_CloneInt;
 884         tVAddr  diffFromSpawn = Thread->SavedState.EIP - (tVAddr)NewTaskHeader;
 885
 886         if( diffFromClone > 0 && diffFromClone < 40 )   // When I last checked, .newTask was at .+27
 887         {
 888                 Log("  Creating process");
 889                 return ;
 890         }
 891
 892         if( diffFromSpawn == 0 )
 893         {
 894                 Log("  Creating thread");
 895                 return ;
 896         }
 897
 898         if( diffFromScheduler > 0 && diffFromScheduler < 128 )  // When I last checked, GetEIP was at .+0x30
 899         {
 900                 // Scheduled out
 901                 Log("  At %04x:%08x", Thread->SavedState.UserCS, Thread->SavedState.UserEIP);
 902                 return ;
 903         }
 904
 905         Log("  Just created (unknown %p)", Thread->SavedState.EIP);
 906 }
 907
 908 void Proc_Reschedule(void)
 909 {
 910         tThread *nextthread, *curthread;
 911          int    cpu = GetCPUNum();
 912
 913         // TODO: Wait for the lock?
 914         if(IS_LOCKED(&glThreadListLock))        return;
 915
 916         curthread = Proc_GetCurThread();
 917
 918         nextthread = Threads_GetNextToRun(cpu, curthread);
 919
 920         if(!nextthread || nextthread == curthread)
 921                 return ;
 922
 923         #if DEBUG_TRACE_SWITCH
 924         // HACK: Ignores switches to the idle threads
 925         if( nextthread->TID == 0 || nextthread->TID > giNumCPUs )
 926         {
 927                 LogF("\nSwitching CPU %i to %p (%i %s) - CR3 = 0x%x, EIP = %p, ESP = %p\n",
 928                         GetCPUNum(),
 929                         nextthread, nextthread->TID, nextthread->ThreadName,
 930                         nextthread->MemState.CR3,
 931                         nextthread->SavedState.EIP,
 932                         nextthread->SavedState.ESP
 933                         );
 934         }
 935         #endif
 936
 937         // Update CPU state
 938         gaCPUs[cpu].Current = nextthread;
 939         gTSSs[cpu].ESP0 = nextthread->KernelStack-4;
 940         __asm__ __volatile__("mov %0, %%db0\n\t" : : "r"(nextthread) );
 941
 942         // Save FPU/MMX/XMM/SSE state
 943         if( curthread && curthread->SavedState.SSE )
 944         {
 945                 Proc_SaveSSE( ((Uint)curthread->SavedState.SSE + 0xF) & ~0xF );
 946                 curthread->SavedState.bSSEModified = 0;
 947                 Proc_DisableSSE();
 948         }
 949
 950         if( curthread )
 951         {
 952                 SwitchTasks(
 953                         nextthread->SavedState.ESP, &curthread->SavedState.ESP,
 954                         nextthread->SavedState.EIP, &curthread->SavedState.EIP,
 955                         nextthread->MemState.CR3
 956                         );
 957         }
 958         else
 959         {
 960                 SwitchTasks(
 961                         nextthread->SavedState.ESP, 0,
 962                         nextthread->SavedState.EIP, 0,
 963                         nextthread->MemState.CR3
 964                         );
 965         }
 966
 967         return ;
 968 }
 969
 970 /**
 971  * \fn void Proc_Scheduler(int CPU)
 972  * \brief Swap current thread and clears dead threads
 973  */
 974 void Proc_Scheduler(int CPU)
 975 {
 976         tThread *thread;
 977
 978         // If the spinlock is set, let it complete
 979         if(IS_LOCKED(&glThreadListLock))        return;
 980
 981         // Get current thread
 982         thread = gaCPUs[CPU].Current;
 983
 984         if( thread )
 985         {
 986                 tRegs   *regs;
 987                 Uint    ebp;
 988                 // Reduce remaining quantum and continue timeslice if non-zero
 989                 if( thread->Remaining-- )
 990                         return;
 991                 // Reset quantum for next call
 992                 thread->Remaining = thread->Quantum;
 993
 994                 // TODO: Make this more stable somehow
 995                 __asm__ __volatile__("mov %%ebp, %0" : "=r" (ebp));
 996                 regs = (tRegs*)(ebp+(2+2)*4);   // EBP,Ret + CPU,CurThread
 997                 thread->SavedState.UserCS = regs->cs;
 998                 thread->SavedState.UserEIP = regs->eip;
 999
1000                 if(thread->bInstrTrace) {
1001                         regs->eflags |= 0x100;  // Set TF
1002                         Log("%p De-scheduled", thread);
1003                 }
1004                 else
1005                         regs->eflags &= ~0x100; // Clear TF
1006         }
1007
1008 #if 0
1009         // TODO: Ack timer?
1010         #if USE_MP
1011         if( GetCPUNum() )
1012                 gpMP_LocalAPIC->EOI.Val = 0;
1013         else
1014         #endif
1015                 outb(0x20, 0x20);
1016         __asm__ __volatile__ ("sti");
1017         Proc_Reschedule();
1018 #endif
1019 }
1020
1021 // === EXPORTS ===
1022 EXPORT(Proc_SpawnWorker);