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 #include <arch_int.h>
  16
  17 // === FLAGS ===
  18 #define DEBUG_TRACE_SWITCH      0
  19 #define DEBUG_DISABLE_DOUBLEFAULT       1
  20 #define DEBUG_VERY_SLOW_PERIOD  0
  21
  22 // === CONSTANTS ===
  23 // Base is 1193182
  24 #define TIMER_BASE      1193182
  25 #if DEBUG_VERY_SLOW_PERIOD
  26 # define TIMER_DIVISOR  1193    //~10Hz switch, with 10 quantum = 1s per thread
  27 #else
  28 # define TIMER_DIVISOR  11932   //~100Hz
  29 #endif
  30
  31 // === TYPES ===
  32 typedef struct sCPU
  33 {
  34         Uint8   APICID;
  35         Uint8   State;  // 0: Unavaliable, 1: Idle, 2: Active
  36         Uint16  Resvd;
  37         tThread *Current;
  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     SwitchTasks(Uint NewSP, Uint *OldSP, Uint NewIP, Uint *OldIO, Uint CR3);
  61 extern void     Proc_InitialiseSSE(void);
  62 extern void     Proc_SaveSSE(Uint DestPtr);
  63 extern void     Proc_DisableSSE(void);
  64
  65 // === PROTOTYPES ===
  66 //void  ArchThreads_Init(void);
  67 #if USE_MP
  68 void    MP_StartAP(int CPU);
  69 void    MP_SendIPIVector(int CPU, Uint8 Vector);
  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         {.Current = &gThreadZero}
  99         };
 100 tTSS    *gTSSs = NULL;  // Pointer to TSS array
 101 tTSS    gTSS0 = {0};
 102 // --- Error Recovery ---
 103 char    gaDoubleFaultStack[1024] __attribute__ ((section(".padata")));
 104 tTSS    gDoubleFault_TSS = {
 105         .ESP0 = (Uint)&gaDoubleFaultStack[1024],
 106         .SS0 = 0x10,
 107         .CR3 = (Uint)gaInitPageDir - KERNEL_BASE,
 108         .EIP = (Uint)Isr8,
 109         .ESP = (Uint)&gaDoubleFaultStack[1024],
 110         .CS = 0x08,     .SS = 0x10,
 111         .DS = 0x10,     .ES = 0x10,
 112         .FS = 0x10,     .GS = 0x10,
 113 };
 114
 115 // === CODE ===
 116 /**
 117  * \fn void ArchThreads_Init(void)
 118  * \brief Starts the process scheduler
 119  */
 120 void ArchThreads_Init(void)
 121 {
 122         Uint    pos = 0;
 123
 124         #if USE_MP
 125         tMPTable        *mptable;
 126
 127         // Mark BSP as active
 128         gaCPUs[0].State = 2;
 129
 130         // -- Initialise Multiprocessing
 131         // Find MP Floating Table
 132         // - EBDA/Last 1Kib (640KiB)
 133         for(pos = KERNEL_BASE|0x9F000; pos < (KERNEL_BASE|0xA0000); pos += 16) {
 134                 if( *(Uint*)(pos) == MPPTR_IDENT ) {
 135                         Log("Possible %p", pos);
 136                         if( ByteSum((void*)pos, sizeof(tMPInfo)) != 0 ) continue;
 137                         gMPFloatPtr = (void*)pos;
 138                         break;
 139                 }
 140         }
 141         // - Last KiB (512KiB base mem)
 142         if(!gMPFloatPtr) {
 143                 for(pos = KERNEL_BASE|0x7F000; pos < (KERNEL_BASE|0x80000); pos += 16) {
 144                         if( *(Uint*)(pos) == MPPTR_IDENT ) {
 145                                 Log("Possible %p", pos);
 146                                 if( ByteSum((void*)pos, sizeof(tMPInfo)) != 0 ) continue;
 147                                 gMPFloatPtr = (void*)pos;
 148                                 break;
 149                         }
 150                 }
 151         }
 152         // - BIOS ROM
 153         if(!gMPFloatPtr) {
 154                 for(pos = KERNEL_BASE|0xE0000; pos < (KERNEL_BASE|0x100000); pos += 16) {
 155                         if( *(Uint*)(pos) == MPPTR_IDENT ) {
 156                                 Log("Possible %p", pos);
 157                                 if( ByteSum((void*)pos, sizeof(tMPInfo)) != 0 ) continue;
 158                                 gMPFloatPtr = (void*)pos;
 159                                 break;
 160                         }
 161                 }
 162         }
 163
 164         // If the MP Table Exists, parse it
 165         if(gMPFloatPtr)
 166         {
 167                  int    i;
 168                 tMPTable_Ent    *ents;
 169                 #if DUMP_MP_TABLE
 170                 Log("gMPFloatPtr = %p", gMPFloatPtr);
 171                 Log("*gMPFloatPtr = {");
 172                 Log("\t.Sig = 0x%08x", gMPFloatPtr->Sig);
 173                 Log("\t.MPConfig = 0x%08x", gMPFloatPtr->MPConfig);
 174                 Log("\t.Length = 0x%02x", gMPFloatPtr->Length);
 175                 Log("\t.Version = 0x%02x", gMPFloatPtr->Version);
 176                 Log("\t.Checksum = 0x%02x", gMPFloatPtr->Checksum);
 177                 Log("\t.Features = [0x%02x,0x%02x,0x%02x,0x%02x,0x%02x]",
 178                         gMPFloatPtr->Features[0],       gMPFloatPtr->Features[1],
 179                         gMPFloatPtr->Features[2],       gMPFloatPtr->Features[3],
 180                         gMPFloatPtr->Features[4]
 181                         );
 182                 Log("}");
 183                 #endif
 184
 185                 mptable = (void*)( KERNEL_BASE|gMPFloatPtr->MPConfig );
 186                 #if DUMP_MP_TABLE
 187                 Log("mptable = %p", mptable);
 188                 Log("*mptable = {");
 189                 Log("\t.Sig = 0x%08x", mptable->Sig);
 190                 Log("\t.BaseTableLength = 0x%04x", mptable->BaseTableLength);
 191                 Log("\t.SpecRev = 0x%02x", mptable->SpecRev);
 192                 Log("\t.Checksum = 0x%02x", mptable->Checksum);
 193                 Log("\t.OEMID = '%8c'", mptable->OemID);
 194                 Log("\t.ProductID = '%8c'", mptable->ProductID);
 195                 Log("\t.OEMTablePtr = %p'", mptable->OEMTablePtr);
 196                 Log("\t.OEMTableSize = 0x%04x", mptable->OEMTableSize);
 197                 Log("\t.EntryCount = 0x%04x", mptable->EntryCount);
 198                 Log("\t.LocalAPICMemMap = 0x%08x", mptable->LocalAPICMemMap);
 199                 Log("\t.ExtendedTableLen = 0x%04x", mptable->ExtendedTableLen);
 200                 Log("\t.ExtendedTableChecksum = 0x%02x", mptable->ExtendedTableChecksum);
 201                 Log("}");
 202                 #endif
 203
 204                 gpMP_LocalAPIC = (void*)MM_MapHWPages(mptable->LocalAPICMemMap, 1);
 205
 206                 ents = mptable->Entries;
 207                 giNumCPUs = 0;
 208
 209                 for( i = 0; i < mptable->EntryCount; i ++ )
 210                 {
 211                          int    entSize = 0;
 212                         switch( ents->Type )
 213                         {
 214                         case 0: // Processor
 215                                 entSize = 20;
 216                                 #if DUMP_MP_TABLE
 217                                 Log("%i: Processor", i);
 218                                 Log("\t.APICID = %i", ents->Proc.APICID);
 219                                 Log("\t.APICVer = 0x%02x", ents->Proc.APICVer);
 220                                 Log("\t.CPUFlags = 0x%02x", ents->Proc.CPUFlags);
 221                                 Log("\t.CPUSignature = 0x%08x", ents->Proc.CPUSignature);
 222                                 Log("\t.FeatureFlags = 0x%08x", ents->Proc.FeatureFlags);
 223                                 #endif
 224
 225                                 if( !(ents->Proc.CPUFlags & 1) ) {
 226                                         Log("DISABLED");
 227                                         break;
 228                                 }
 229
 230                                 // Check if there is too many processors
 231                                 if(giNumCPUs >= MAX_CPUS) {
 232                                         giNumCPUs ++;   // If `giNumCPUs` > MAX_CPUS later, it will be clipped
 233                                         break;
 234                                 }
 235
 236                                 // Initialise CPU Info
 237                                 gaAPIC_to_CPU[ents->Proc.APICID] = giNumCPUs;
 238                                 gaCPUs[giNumCPUs].APICID = ents->Proc.APICID;
 239                                 gaCPUs[giNumCPUs].State = 0;
 240                                 giNumCPUs ++;
 241
 242                                 // Set BSP Variable
 243                                 if( ents->Proc.CPUFlags & 2 ) {
 244                                         giProc_BootProcessorID = giNumCPUs-1;
 245                                 }
 246
 247                                 break;
 248
 249                         #if DUMP_MP_TABLE >= 2
 250                         case 1: // Bus
 251                                 entSize = 8;
 252                                 Log("%i: Bus", i);
 253                                 Log("\t.ID = %i", ents->Bus.ID);
 254                                 Log("\t.TypeString = '%6C'", ents->Bus.TypeString);
 255                                 break;
 256                         case 2: // I/O APIC
 257                                 entSize = 8;
 258                                 Log("%i: I/O APIC", i);
 259                                 Log("\t.ID = %i", ents->IOAPIC.ID);
 260                                 Log("\t.Version = 0x%02x", ents->IOAPIC.Version);
 261                                 Log("\t.Flags = 0x%02x", ents->IOAPIC.Flags);
 262                                 Log("\t.Addr = 0x%08x", ents->IOAPIC.Addr);
 263                                 break;
 264                         case 3: // I/O Interrupt Assignment
 265                                 entSize = 8;
 266                                 Log("%i: I/O Interrupt Assignment", i);
 267                                 Log("\t.IntType = %i", ents->IOInt.IntType);
 268                                 Log("\t.Flags = 0x%04x", ents->IOInt.Flags);
 269                                 Log("\t.SourceBusID = 0x%02x", ents->IOInt.SourceBusID);
 270                                 Log("\t.SourceBusIRQ = 0x%02x", ents->IOInt.SourceBusIRQ);
 271                                 Log("\t.DestAPICID = 0x%02x", ents->IOInt.DestAPICID);
 272                                 Log("\t.DestAPICIRQ = 0x%02x", ents->IOInt.DestAPICIRQ);
 273                                 break;
 274                         case 4: // Local Interrupt Assignment
 275                                 entSize = 8;
 276                                 Log("%i: Local Interrupt Assignment", i);
 277                                 Log("\t.IntType = %i", ents->LocalInt.IntType);
 278                                 Log("\t.Flags = 0x%04x", ents->LocalInt.Flags);
 279                                 Log("\t.SourceBusID = 0x%02x", ents->LocalInt.SourceBusID);
 280                                 Log("\t.SourceBusIRQ = 0x%02x", ents->LocalInt.SourceBusIRQ);
 281                                 Log("\t.DestLocalAPICID = 0x%02x", ents->LocalInt.DestLocalAPICID);
 282                                 Log("\t.DestLocalAPICIRQ = 0x%02x", ents->LocalInt.DestLocalAPICIRQ);
 283                                 break;
 284                         default:
 285                                 Log("%i: Unknown (%i)", i, ents->Type);
 286                                 break;
 287                         #endif
 288                         }
 289                         ents = (void*)( (Uint)ents + entSize );
 290                 }
 291
 292                 if( giNumCPUs > MAX_CPUS ) {
 293                         Warning("Too many CPUs detected (%i), only using %i of them", giNumCPUs, MAX_CPUS);
 294                         giNumCPUs = MAX_CPUS;
 295                 }
 296                 gTSSs = gaTSSs;
 297         }
 298         else {
 299                 Log("No MP Table was found, assuming uniprocessor\n");
 300                 giNumCPUs = 1;
 301                 gTSSs = &gTSS0;
 302         }
 303         #else
 304         giNumCPUs = 1;
 305         gTSSs = &gTSS0;
 306         #endif
 307
 308         #if !DEBUG_DISABLE_DOUBLEFAULT
 309         // Initialise Double Fault TSS
 310         gGDT[5].BaseLow = (Uint)&gDoubleFault_TSS & 0xFFFF;
 311         gGDT[5].BaseMid = (Uint)&gDoubleFault_TSS >> 16;
 312         gGDT[5].BaseHi = (Uint)&gDoubleFault_TSS >> 24;
 313
 314         // Set double fault IDT to use the new TSS
 315         gIDT[8].OffsetLo = 0;
 316         gIDT[8].CS = 5<<3;
 317         gIDT[8].Flags = 0x8500;
 318         gIDT[8].OffsetHi = 0;
 319         #endif
 320
 321         // Set timer frequency
 322         outb(0x43, 0x34);       // Set Channel 0, Low/High, Rate Generator
 323         outb(0x40, TIMER_DIVISOR&0xFF); // Low Byte of Divisor
 324         outb(0x40, (TIMER_DIVISOR>>8)&0xFF);    // High Byte
 325
 326         Log_Debug("Proc", "PIT Frequency %i.%03i Hz",
 327                 TIMER_BASE/TIMER_DIVISOR,
 328                 ((Uint64)TIMER_BASE*1000/TIMER_DIVISOR)%1000
 329                 );
 330
 331         #if USE_MP
 332         // Get the count setting for APIC timer
 333         Log("Determining APIC Count");
 334         __asm__ __volatile__ ("sti");
 335         while( giMP_TimerCount == 0 )   __asm__ __volatile__ ("hlt");
 336         __asm__ __volatile__ ("cli");
 337         Log("APIC Count %i", giMP_TimerCount);
 338         {
 339                 Uint64  freq = giMP_TimerCount;
 340                 freq *= TIMER_BASE;
 341                 freq /= TIMER_DIVISOR;
 342                 if( (freq /= 1000) < 2*1000)
 343                         Log("Bus Frequency %i KHz", freq);
 344                 else if( (freq /= 1000) < 2*1000)
 345                         Log("Bus Frequency %i MHz", freq);
 346                 else if( (freq /= 1000) < 2*1000)
 347                         Log("Bus Frequency %i GHz", freq);
 348                 else
 349                         Log("Bus Frequency %i THz", freq);
 350         }
 351
 352         // Initialise Normal TSS(s)
 353         for(pos=0;pos<giNumCPUs;pos++)
 354         {
 355         #else
 356         pos = 0;
 357         #endif
 358                 gTSSs[pos].SS0 = 0x10;
 359                 gTSSs[pos].ESP0 = 0;    // Set properly by scheduler
 360                 gGDT[6+pos].BaseLow = ((Uint)(&gTSSs[pos])) & 0xFFFF;
 361                 gGDT[6+pos].BaseMid = ((Uint)(&gTSSs[pos]) >> 16) & 0xFFFF;
 362                 gGDT[6+pos].BaseHi = ((Uint)(&gTSSs[pos])) >> 24;
 363         #if USE_MP
 364         }
 365         #endif
 366
 367         // Load the BSP's TSS
 368         __asm__ __volatile__ ("ltr %%ax"::"a"(0x30));
 369         // Set Current Thread and CPU Number in DR0 and DR1
 370         __asm__ __volatile__ ("mov %0, %%db0"::"r"(&gThreadZero));
 371         __asm__ __volatile__ ("mov %0, %%db1"::"r"(0));
 372
 373         gaCPUs[0].Current = &gThreadZero;
 374         gThreadZero.CurCPU = 0;
 375
 376         gThreadZero.MemState.CR3 = (Uint)gaInitPageDir - KERNEL_BASE;
 377
 378         // Create Per-Process Data Block
 379         if( !MM_Allocate(MM_PPD_CFG) )
 380         {
 381                 Panic("OOM - No space for initial Per-Process Config");
 382         }
 383
 384         // Initialise SSE support
 385         Proc_InitialiseSSE();
 386
 387         // Change Stacks
 388         Proc_ChangeStack();
 389 }
 390
 391 #if USE_MP
 392 /**
 393  * \brief Start an AP
 394  */
 395 void MP_StartAP(int CPU)
 396 {
 397         Log_Log("Proc", "Starting AP %i (APIC %i)", CPU, gaCPUs[CPU].APICID);
 398
 399         // Set location of AP startup code and mark for a warm restart
 400         *(Uint16*)(KERNEL_BASE|0x467) = (Uint)&APWait - (KERNEL_BASE|0xFFFF0);
 401         *(Uint16*)(KERNEL_BASE|0x469) = 0xFFFF;
 402         outb(0x70, 0x0F);       outb(0x71, 0x0A);       // Set warm reset flag
 403         MP_SendIPI(gaCPUs[CPU].APICID, 0, 5);   // Init IPI
 404
 405         // Delay
 406         inb(0x80); inb(0x80); inb(0x80); inb(0x80);
 407
 408         // TODO: Use a better address, preferably registered with the MM
 409         // - MM_AllocDMA mabye?
 410         // Create a far jump
 411         *(Uint8*)(KERNEL_BASE|0x11000) = 0xEA;  // Far JMP
 412         *(Uint16*)(KERNEL_BASE|0x11001) = (Uint)&APStartup - (KERNEL_BASE|0xFFFF0);     // IP
 413         *(Uint16*)(KERNEL_BASE|0x11003) = 0xFFFF;       // CS
 414         // Send a Startup-IPI to make the CPU execute at 0x11000 (which we
 415         // just filled)
 416         MP_SendIPI(gaCPUs[CPU].APICID, 0x11, 6);        // StartupIPI
 417
 418         giNumInitingCPUs ++;
 419 }
 420
 421 void MP_SendIPIVector(int CPU, Uint8 Vector)
 422 {
 423         MP_SendIPI(gaCPUs[CPU].APICID, Vector, 0);
 424 }
 425
 426 /**
 427  * \brief Send an Inter-Processor Interrupt
 428  * \param APICID        Processor's Local APIC ID
 429  * \param Vector        Argument of some kind
 430  * \param DeliveryMode  Type of signal
 431  * \note 3A 10.5 "APIC/Handling Local Interrupts"
 432  */
 433 void MP_SendIPI(Uint8 APICID, int Vector, int DeliveryMode)
 434 {
 435         Uint32  val;
 436
 437         // Hi
 438         val = (Uint)APICID << 24;
 439 //      Log("%p = 0x%08x", &gpMP_LocalAPIC->ICR[1], val);
 440         gpMP_LocalAPIC->ICR[1].Val = val;
 441         // Low (and send)
 442         val = ((DeliveryMode & 7) << 8) | (Vector & 0xFF);
 443 //      Log("%p = 0x%08x", &gpMP_LocalAPIC->ICR[0], val);
 444         gpMP_LocalAPIC->ICR[0].Val = val;
 445 }
 446 #endif
 447
 448 void Proc_IdleThread(void *Ptr)
 449 {
 450         tCPU    *cpu = &gaCPUs[GetCPUNum()];
 451         cpu->Current->ThreadName = strdup("Idle Thread");
 452         Threads_SetPriority( cpu->Current, -1 );        // Never called randomly
 453         cpu->Current->Quantum = 1;      // 1 slice quantum
 454         for(;;) {
 455                 __asm__ __volatile__ ("sti");   // Make sure interrupts are enabled
 456                 __asm__ __volatile__ ("hlt");   // Make sure interrupts are enabled
 457                 Proc_Reschedule();
 458         }
 459 }
 460
 461 /**
 462  * \fn void Proc_Start(void)
 463  * \brief Start process scheduler
 464  */
 465 void Proc_Start(void)
 466 {
 467          int    tid;
 468         #if USE_MP
 469          int    i;
 470         #endif
 471
 472         #if USE_MP
 473         // Start APs
 474         for( i = 0; i < giNumCPUs; i ++ )
 475         {
 476                 if(i)   gaCPUs[i].Current = NULL;
 477
 478                 // Create Idle Task
 479                 tid = Proc_NewKThread(Proc_IdleThread, &gaCPUs[i]);
 480
 481                 // Start the AP
 482                 if( i != giProc_BootProcessorID ) {
 483                         MP_StartAP( i );
 484                 }
 485         }
 486
 487         // BSP still should run the current task
 488         gaCPUs[0].Current = &gThreadZero;
 489
 490         // Start interrupts and wait for APs to come up
 491         Log_Debug("Proc", "Waiting for APs to come up");
 492         __asm__ __volatile__ ("sti");
 493         while( giNumInitingCPUs )       __asm__ __volatile__ ("hlt");
 494         #else
 495         // Create Idle Task
 496         tid = Proc_NewKThread(Proc_IdleThread, &gaCPUs[0]);
 497 //      gaCPUs[0].IdleThread = Threads_GetThread(tid);
 498
 499         // Set current task
 500         gaCPUs[0].Current = &gThreadZero;
 501
 502         // Start Interrupts (and hence scheduler)
 503         __asm__ __volatile__("sti");
 504         #endif
 505         MM_FinishVirtualInit();
 506 }
 507
 508 /**
 509  * \fn tThread *Proc_GetCurThread(void)
 510  * \brief Gets the current thread
 511  */
 512 tThread *Proc_GetCurThread(void)
 513 {
 514         #if USE_MP
 515         return gaCPUs[ GetCPUNum() ].Current;
 516         #else
 517         return gaCPUs[ 0 ].Current;
 518         #endif
 519 }
 520
 521 /**
 522  * \fn void Proc_ChangeStack(void)
 523  * \brief Swaps the current stack for a new one (in the proper stack reigon)
 524  */
 525 void Proc_ChangeStack(void)
 526 {
 527         Uint    esp, ebp;
 528         Uint    tmpEbp, oldEsp;
 529         Uint    curBase, newBase;
 530
 531         __asm__ __volatile__ ("mov %%esp, %0":"=r"(esp));
 532         __asm__ __volatile__ ("mov %%ebp, %0":"=r"(ebp));
 533
 534         oldEsp = esp;
 535
 536         // Create new KStack
 537         newBase = MM_NewKStack();
 538         // Check for errors
 539         if(newBase == 0) {
 540                 Panic("What the?? Unable to allocate space for initial kernel stack");
 541                 return;
 542         }
 543
 544         curBase = (Uint)&Kernel_Stack_Top;
 545
 546         LOG("curBase = 0x%x, newBase = 0x%x", curBase, newBase);
 547
 548         // Get ESP as a used size
 549         esp = curBase - esp;
 550         LOG("memcpy( %p, %p, 0x%x )", (void*)(newBase - esp), (void*)(curBase - esp), esp );
 551         // Copy used stack
 552         memcpy( (void*)(newBase - esp), (void*)(curBase - esp), esp );
 553         // Get ESP as an offset in the new stack
 554         esp = newBase - esp;
 555         // Adjust EBP
 556         ebp = newBase - (curBase - ebp);
 557
 558         // Repair EBPs & Stack Addresses
 559         // Catches arguments also, but may trash stack-address-like values
 560         for(tmpEbp = esp; tmpEbp < newBase; tmpEbp += 4)
 561         {
 562                 if(oldEsp < *(Uint*)tmpEbp && *(Uint*)tmpEbp < curBase)
 563                         *(Uint*)tmpEbp += newBase - curBase;
 564         }
 565
 566         Proc_GetCurThread()->KernelStack = newBase;
 567
 568         __asm__ __volatile__ ("mov %0, %%esp"::"r"(esp));
 569         __asm__ __volatile__ ("mov %0, %%ebp"::"r"(ebp));
 570 }
 571
 572 void Proc_ClearThread(tThread *Thread)
 573 {
 574         MM_ClearSpace(Thread->MemState.CR3);
 575         if(Thread->SavedState.SSE) {
 576                 free(Thread->SavedState.SSE);
 577                 Thread->SavedState.SSE = NULL;
 578         }
 579 }
 580
 581 int Proc_NewKThread(void (*Fcn)(void*), void *Data)
 582 {
 583         Uint    esp;
 584         tThread *newThread, *cur;
 585
 586         cur = Proc_GetCurThread();
 587         newThread = Threads_CloneTCB(0);
 588         if(!newThread)  return -1;
 589
 590         // Set CR3
 591         MM_RefPhys( cur->MemState.CR3 );
 592         newThread->MemState.CR3 = cur->MemState.CR3;
 593
 594         // Create new KStack
 595         newThread->KernelStack = MM_NewKStack();
 596         // Check for errors
 597         if(newThread->KernelStack == 0) {
 598                 free(newThread);
 599                 return -1;
 600         }
 601
 602         esp = newThread->KernelStack;
 603         *(Uint*)(esp-=4) = (Uint)Data;  // Data (shadowed)
 604         *(Uint*)(esp-=4) = 1;   // Number of params
 605         *(Uint*)(esp-=4) = (Uint)Fcn;   // Function to call
 606         *(Uint*)(esp-=4) = (Uint)newThread;     // Thread ID
 607
 608         newThread->SavedState.ESP = esp;
 609         newThread->SavedState.EIP = (Uint)&NewTaskHeader;
 610         newThread->SavedState.SSE = NULL;
 611 //      Log("New (KThread) %p, esp = %p", newThread->SavedState.EIP, newThread->SavedState.ESP);
 612
 613 //      MAGIC_BREAK();
 614         Threads_AddActive(newThread);
 615
 616         return newThread->TID;
 617 }
 618
 619 /**
 620  * \fn int Proc_Clone(Uint *Err, Uint Flags)
 621  * \brief Clone the current process
 622  */
 623 int Proc_Clone(Uint Flags)
 624 {
 625         tThread *newThread;
 626         tThread *cur = Proc_GetCurThread();
 627         Uint    eip;
 628
 629         // Sanity, please
 630         if( !(Flags & CLONE_VM) ) {
 631                 Log_Error("Proc", "Proc_Clone: Don't leave CLONE_VM unset, use Proc_NewKThread instead");
 632                 return -1;
 633         }
 634
 635         // New thread
 636         newThread = Threads_CloneTCB(Flags);
 637         if(!newThread)  return -1;
 638
 639         newThread->KernelStack = cur->KernelStack;
 640
 641         // Clone state
 642         eip = Proc_CloneInt(&newThread->SavedState.ESP, &newThread->MemState.CR3);
 643         if( eip == 0 ) {
 644                 // ACK the interrupt
 645                 return 0;
 646         }
 647         newThread->SavedState.EIP = eip;
 648         newThread->SavedState.SSE = NULL;
 649         newThread->SavedState.bSSEModified = 0;
 650
 651         // Check for errors
 652         if( newThread->MemState.CR3 == 0 ) {
 653                 Log_Error("Proc", "Proc_Clone: MM_Clone failed");
 654                 Threads_Delete(newThread);
 655                 return -1;
 656         }
 657
 658         // Add the new thread to the run queue
 659         Threads_AddActive(newThread);
 660         return newThread->TID;
 661 }
 662
 663 /**
 664  * \fn int Proc_SpawnWorker(void)
 665  * \brief Spawns a new worker thread
 666  */
 667 int Proc_SpawnWorker(void (*Fcn)(void*), void *Data)
 668 {
 669         tThread *new;
 670         Uint    stack_contents[4];
 671
 672         // Create new thread
 673         new = Threads_CloneThreadZero();
 674         if(!new) {
 675                 Warning("Proc_SpawnWorker - Out of heap space!\n");
 676                 return -1;
 677         }
 678
 679         // Create the stack contents
 680         stack_contents[3] = (Uint)Data;
 681         stack_contents[2] = 1;
 682         stack_contents[1] = (Uint)Fcn;
 683         stack_contents[0] = (Uint)new;
 684
 685         // Create a new worker stack (in PID0's address space)
 686         new->KernelStack = MM_NewWorkerStack(stack_contents, sizeof(stack_contents));
 687
 688         // Save core machine state
 689         new->SavedState.ESP = new->KernelStack - sizeof(stack_contents);
 690         new->SavedState.EIP = (Uint)NewTaskHeader;
 691         new->SavedState.SSE = NULL;
 692         new->SavedState.bSSEModified = 0;
 693
 694         // Mark as active
 695         new->Status = THREAD_STAT_PREINIT;
 696         Threads_AddActive( new );
 697
 698         return new->TID;
 699 }
 700
 701 /**
 702  * \fn Uint Proc_MakeUserStack(void)
 703  * \brief Creates a new user stack
 704  */
 705 Uint Proc_MakeUserStack(void)
 706 {
 707          int    i;
 708         Uint    base = USER_STACK_TOP - USER_STACK_SZ;
 709
 710         // Check Prospective Space
 711         for( i = USER_STACK_SZ >> 12; i--; )
 712                 if( MM_GetPhysAddr( base + (i<<12) ) != 0 )
 713                         break;
 714
 715         if(i != -1)     return 0;
 716
 717         // Allocate Stack - Allocate incrementally to clean up MM_Dump output
 718         for( i = 0; i < USER_STACK_SZ/0x1000; i++ )
 719         {
 720                 if( !MM_Allocate( base + (i<<12) ) )
 721                 {
 722                         Warning("OOM: Proc_MakeUserStack");
 723                         return 0;
 724                 }
 725         }
 726
 727         return base + USER_STACK_SZ;
 728 }
 729
 730 void Proc_StartUser(Uint Entrypoint, Uint Base, int ArgC, char **ArgV, int DataSize)
 731 {
 732         Uint    *stack;
 733          int    i;
 734         char    **envp = NULL;
 735         Uint16  ss, cs;
 736
 737         // Copy data to the user stack and free original buffer
 738         stack = (void*)Proc_MakeUserStack();
 739         stack -= DataSize/sizeof(*stack);
 740         memcpy( stack, ArgV, DataSize );
 741         free(ArgV);
 742
 743         // Adjust Arguments and environment
 744         if( DataSize )
 745         {
 746                 Uint delta = (Uint)stack - (Uint)ArgV;
 747                 ArgV = (char**)stack;
 748                 for( i = 0; ArgV[i]; i++ )      ArgV[i] += delta;
 749                 envp = &ArgV[i+1];
 750                 for( i = 0; envp[i]; i++ )      envp[i] += delta;
 751         }
 752
 753         // User Mode Segments
 754         ss = 0x23;      cs = 0x1B;
 755
 756         // Arguments
 757         *--stack = (Uint)envp;
 758         *--stack = (Uint)ArgV;
 759         *--stack = (Uint)ArgC;
 760         *--stack = Base;
 761
 762         Proc_StartProcess(ss, (Uint)stack, 0x202, cs, Entrypoint);
 763 }
 764
 765 void Proc_StartProcess(Uint16 SS, Uint Stack, Uint Flags, Uint16 CS, Uint IP)
 766 {
 767         Uint    *stack = (void*)Stack;
 768         *--stack = SS;          //Stack Segment
 769         *--stack = Stack;       //Stack Pointer
 770         *--stack = Flags;       //EFLAGS (Resvd (0x2) and IF (0x20))
 771         *--stack = CS;          //Code Segment
 772         *--stack = IP;  //EIP
 773         //PUSHAD
 774         *--stack = 0xAAAAAAAA;  // eax
 775         *--stack = 0xCCCCCCCC;  // ecx
 776         *--stack = 0xDDDDDDDD;  // edx
 777         *--stack = 0xBBBBBBBB;  // ebx
 778         *--stack = 0xD1D1D1D1;  // edi
 779         *--stack = 0x54545454;  // esp - NOT POPED
 780         *--stack = 0x51515151;  // esi
 781         *--stack = 0xB4B4B4B4;  // ebp
 782         //Individual PUSHs
 783         *--stack = SS;  // ds
 784         *--stack = SS;  // es
 785         *--stack = SS;  // fs
 786         *--stack = SS;  // gs
 787
 788         __asm__ __volatile__ (
 789         "mov %%eax,%%esp;\n\t"  // Set stack pointer
 790         "pop %%gs;\n\t"
 791         "pop %%fs;\n\t"
 792         "pop %%es;\n\t"
 793         "pop %%ds;\n\t"
 794         "popa;\n\t"
 795         "iret;\n\t" : : "a" (stack));
 796         for(;;);
 797 }
 798
 799 /**
 800  * \fn int Proc_Demote(Uint *Err, int Dest, tRegs *Regs)
 801  * \brief Demotes a process to a lower permission level
 802  * \param Err   Pointer to user's errno
 803  * \param Dest  New Permission Level
 804  * \param Regs  Pointer to user's register structure
 805  */
 806 int Proc_Demote(Uint *Err, int Dest, tRegs *Regs)
 807 {
 808          int    cpl = Regs->cs & 3;
 809         // Sanity Check
 810         if(Dest > 3 || Dest < 0) {
 811                 *Err = -EINVAL;
 812                 return -1;
 813         }
 814
 815         // Permission Check
 816         if(cpl > Dest) {
 817                 *Err = -EACCES;
 818                 return -1;
 819         }
 820
 821         // Change the Segment Registers
 822         Regs->cs = (((Dest+1)<<4) | Dest) - 8;
 823         Regs->ss = ((Dest+1)<<4) | Dest;
 824         // Check if the GP Segs are GDT, then change them
 825         if(!(Regs->ds & 4))     Regs->ds = ((Dest+1)<<4) | Dest;
 826         if(!(Regs->es & 4))     Regs->es = ((Dest+1)<<4) | Dest;
 827         if(!(Regs->fs & 4))     Regs->fs = ((Dest+1)<<4) | Dest;
 828         if(!(Regs->gs & 4))     Regs->gs = ((Dest+1)<<4) | Dest;
 829
 830         return 0;
 831 }
 832
 833 /**
 834  * \brief Calls a signal handler in user mode
 835  * \note Used for signals
 836  */
 837 void Proc_CallFaultHandler(tThread *Thread)
 838 {
 839         // Rewinds the stack and calls the user function
 840         // Never returns
 841         Proc_ReturnToUser( Thread->FaultHandler, Thread->CurFaultNum, Thread->KernelStack );
 842         for(;;);
 843 }
 844
 845 void Proc_DumpThreadCPUState(tThread *Thread)
 846 {
 847         if( Thread->CurCPU > -1 )
 848         {
 849                  int    maxBacktraceDistance = 6;
 850                 tRegs   *regs = NULL;
 851                 Uint32  *stack;
 852
 853                 if( Thread->CurCPU != GetCPUNum() ) {
 854                         Log("  Currently running");
 855                         return ;
 856                 }
 857
 858                 // Backtrace to find the IRQ entrypoint
 859                 // - This will usually only be called by an IRQ, so this should
 860                 //   work
 861                 __asm__ __volatile__ ("mov %%ebp, %0" : "=r" (stack));
 862                 while( maxBacktraceDistance -- )
 863                 {
 864                         // [ebp] = oldEbp
 865                         // [ebp+4] = retaddr
 866
 867                         if( stack[1] == (tVAddr)&IRQCommon_handled ) {
 868                                 regs = (void*)stack[2];
 869                                 break;
 870                         }
 871
 872                         stack = (void*)stack[0];
 873                 }
 874
 875                 if( !regs ) {
 876                         Log("  Unable to find IRQ Entry");
 877                         return ;
 878                 }
 879
 880                 Log("  at %04x:%08x", regs->cs, regs->eip);
 881                 return ;
 882         }
 883
 884         tVAddr  diffFromScheduler = Thread->SavedState.EIP - (tVAddr)SwitchTasks;
 885         tVAddr  diffFromClone = Thread->SavedState.EIP - (tVAddr)Proc_CloneInt;
 886         tVAddr  diffFromSpawn = Thread->SavedState.EIP - (tVAddr)NewTaskHeader;
 887
 888         if( diffFromClone > 0 && diffFromClone < 40 )   // When I last checked, .newTask was at .+27
 889         {
 890                 Log("  Creating process");
 891                 return ;
 892         }
 893
 894         if( diffFromSpawn == 0 )
 895         {
 896                 Log("  Creating thread");
 897                 return ;
 898         }
 899
 900         if( diffFromScheduler > 0 && diffFromScheduler < 128 )  // When I last checked, GetEIP was at .+0x30
 901         {
 902                 // Scheduled out
 903                 Log("  At %04x:%08x", Thread->SavedState.UserCS, Thread->SavedState.UserEIP);
 904                 return ;
 905         }
 906
 907         Log("  Just created (unknown %p)", Thread->SavedState.EIP);
 908 }
 909
 910 void Proc_Reschedule(void)
 911 {
 912         tThread *nextthread, *curthread;
 913          int    cpu = GetCPUNum();
 914
 915         // TODO: Wait for the lock?
 916         if(IS_LOCKED(&glThreadListLock))        return;
 917
 918         curthread = Proc_GetCurThread();
 919
 920         nextthread = Threads_GetNextToRun(cpu, curthread);
 921
 922         if(!nextthread || nextthread == curthread)
 923                 return ;
 924
 925         #if DEBUG_TRACE_SWITCH
 926         // HACK: Ignores switches to the idle threads
 927         if( nextthread->TID == 0 || nextthread->TID > giNumCPUs )
 928         {
 929                 LogF("\nSwitching CPU %i to %p (%i %s) - CR3 = 0x%x, EIP = %p, ESP = %p\n",
 930                         GetCPUNum(),
 931                         nextthread, nextthread->TID, nextthread->ThreadName,
 932                         nextthread->MemState.CR3,
 933                         nextthread->SavedState.EIP,
 934                         nextthread->SavedState.ESP
 935                         );
 936         }
 937         #endif
 938
 939         // Update CPU state
 940         gaCPUs[cpu].Current = nextthread;
 941         gTSSs[cpu].ESP0 = nextthread->KernelStack-4;
 942         __asm__ __volatile__("mov %0, %%db0\n\t" : : "r"(nextthread) );
 943
 944         // Save FPU/MMX/XMM/SSE state
 945         if( curthread && curthread->SavedState.SSE )
 946         {
 947                 Proc_SaveSSE( ((Uint)curthread->SavedState.SSE + 0xF) & ~0xF );
 948                 curthread->SavedState.bSSEModified = 0;
 949                 Proc_DisableSSE();
 950         }
 951
 952         if( curthread )
 953         {
 954                 SwitchTasks(
 955                         nextthread->SavedState.ESP, &curthread->SavedState.ESP,
 956                         nextthread->SavedState.EIP, &curthread->SavedState.EIP,
 957                         nextthread->MemState.CR3
 958                         );
 959         }
 960         else
 961         {
 962                 SwitchTasks(
 963                         nextthread->SavedState.ESP, 0,
 964                         nextthread->SavedState.EIP, 0,
 965                         nextthread->MemState.CR3
 966                         );
 967         }
 968
 969         return ;
 970 }
 971
 972 /**
 973  * \fn void Proc_Scheduler(int CPU)
 974  * \brief Swap current thread and clears dead threads
 975  */
 976 void Proc_Scheduler(int CPU)
 977 {
 978 #if 0
 979         tThread *thread;
 980
 981         // If the spinlock is set, let it complete
 982         if(IS_LOCKED(&glThreadListLock))        return;
 983
 984         // Get current thread
 985         thread = gaCPUs[CPU].Current;
 986
 987         if( thread )
 988         {
 989                 tRegs   *regs;
 990                 Uint    ebp;
 991                 // Reduce remaining quantum and continue timeslice if non-zero
 992                 if( thread->Remaining-- )
 993                         return;
 994                 // Reset quantum for next call
 995                 thread->Remaining = thread->Quantum;
 996
 997                 // TODO: Make this more stable somehow
 998                 __asm__ __volatile__("mov %%ebp, %0" : "=r" (ebp));
 999                 regs = (tRegs*)(ebp+(2+2)*4);   // EBP,Ret + CPU,CurThread
1000                 thread->SavedState.UserCS = regs->cs;
1001                 thread->SavedState.UserEIP = regs->eip;
1002
1003                 if(thread->bInstrTrace) {
1004                         regs->eflags |= 0x100;  // Set TF
1005                         Log("%p De-scheduled", thread);
1006                 }
1007                 else
1008                         regs->eflags &= ~0x100; // Clear TF
1009         }
1010
1011         // TODO: Ack timer?
1012         #if USE_MP
1013         if( GetCPUNum() )
1014                 gpMP_LocalAPIC->EOI.Val = 0;
1015         else
1016         #endif
1017                 outb(0x20, 0x20);
1018         __asm__ __volatile__ ("sti");
1019         Proc_Reschedule();
1020 #endif
1021 }
1022
1023 // === EXPORTS ===
1024 EXPORT(Proc_SpawnWorker);