Kernel/arch/x86_64/proc.c

   1 /*
   2  * Acess2 x86_64 port
   3  * proc.c
   4  */
   5 #include <acess.h>
   6 #include <proc.h>
   7 #include <threads.h>
   8 #include <threads_int.h>
   9 #include <desctab.h>
  10 #include <mm_virt.h>
  11 #include <errno.h>
  12 #if USE_MP
  13 # include <mp.h>
  14 #endif
  15 #include <arch_config.h>
  16 #include <hal_proc.h>
  17
  18 // === FLAGS ===
  19 #define DEBUG_TRACE_SWITCH      0
  20 #define BREAK_ON_SWITCH         0       // Break into bochs debugger on a task switch
  21
  22 // === CONSTANTS ===
  23
  24 // === TYPES ===
  25 typedef struct sCPU
  26 {
  27         Uint8   APICID;
  28         Uint8   State;  // 0: Unavaliable, 1: Idle, 2: Active
  29         Uint16  Resvd;
  30         tThread *Current;
  31         tThread *IdleThread;
  32 }       tCPU;
  33
  34 // === IMPORTS ===
  35 extern tGDT     gGDT[];
  36 extern void     APStartup(void);        // 16-bit AP startup code
  37
  38 extern Uint     GetRIP(void);   // start.asm
  39 extern Uint     SaveState(Uint *RSP, Uint *Regs);
  40 extern Uint     Proc_CloneInt(Uint *RSP, Uint *CR3);
  41 extern void     NewTaskHeader(void);    // Actually takes cdecl args
  42
  43 extern Uint64   gInitialPML4[512];      // start.asm
  44 extern tShortSpinlock   glThreadListLock;
  45 extern int      giNumCPUs;
  46 extern int      giNextTID;
  47 extern int      giTotalTickets;
  48 extern int      giNumActiveThreads;
  49 extern tThread  gThreadZero;
  50 extern void     Threads_Dump(void);
  51 extern void     Proc_ReturnToUser(tVAddr Handler, tVAddr KStackTop, int Argument);
  52 extern void     Time_UpdateTimestamp(void);
  53 extern void     SwitchTasks(Uint NewSP, Uint *OldSP, Uint NewIP, Uint *OldIO, Uint CR3);
  54
  55 // === PROTOTYPES ===
  56 //void  ArchThreads_Init(void);
  57 #if USE_MP
  58 void    MP_StartAP(int CPU);
  59 void    MP_SendIPI(Uint8 APICID, int Vector, int DeliveryMode);
  60 #endif
  61 void    Proc_IdleTask(void *unused);
  62 //void  Proc_Start(void);
  63 //tThread       *Proc_GetCurThread(void);
  64  int    Proc_NewKThread(void (*Fcn)(void*), void *Data);
  65 // int  Proc_Clone(Uint *Err, Uint Flags);
  66 // int  Proc_SpawnWorker(void);
  67 Uint    Proc_MakeUserStack(void);
  68 //void  Proc_StartUser(Uint Entrypoint, Uint *Bases, int ArgC, char **ArgV, char **EnvP, int DataSize);
  69 void    Proc_StartProcess(Uint16 SS, Uint Stack, Uint Flags, Uint16 CS, Uint IP) NORETURN;
  70  int    Proc_Demote(Uint *Err, int Dest, tRegs *Regs);
  71 //void  Proc_CallFaultHandler(tThread *Thread);
  72 //void  Proc_DumpThreadCPUState(tThread *Thread);
  73 //void  Proc_Reschedule(void);
  74 void    Proc_Scheduler(int CPU, Uint RSP, Uint RIP);
  75
  76 // === GLOBALS ===
  77 //!\brief Used by desctab.asm in SyscallStub
  78 const int ci_offsetof_tThread_KernelStack = offsetof(tThread, KernelStack);
  79 // --- Multiprocessing ---
  80 #if USE_MP
  81 volatile int    giNumInitingCPUs = 0;
  82 tMPInfo *gMPFloatPtr = NULL;
  83 tAPIC   *gpMP_LocalAPIC = NULL;
  84 Uint8   gaAPIC_to_CPU[256] = {0};
  85 #endif
  86 tCPU    gaCPUs[MAX_CPUS];
  87 tTSS    *gTSSs = NULL;
  88 tTSS    gTSS0 = {0};
  89 // --- Error Recovery ---
  90 Uint32  gaDoubleFaultStack[1024];
  91
  92 // === CODE ===
  93 /**
  94  * \fn void ArchThreads_Init(void)
  95  * \brief Starts the process scheduler
  96  */
  97 void ArchThreads_Init(void)
  98 {
  99         Uint    pos = 0;
 100
 101         #if USE_MP
 102         tMPTable        *mptable;
 103
 104         // Mark BSP as active
 105         gaCPUs[0].State = 2;
 106
 107         // -- Initialise Multiprocessing
 108         // Find MP Floating Table
 109         // - EBDA/Last 1Kib (640KiB)
 110         for(pos = KERNEL_BASE|0x9F000; pos < (KERNEL_BASE|0xA0000); pos += 16) {
 111                 if( *(Uint*)(pos) == MPPTR_IDENT ) {
 112                         Log("Possible %p", pos);
 113                         if( ByteSum((void*)pos, sizeof(tMPInfo)) != 0 ) continue;
 114                         gMPFloatPtr = (void*)pos;
 115                         break;
 116                 }
 117         }
 118         // - Last KiB (512KiB base mem)
 119         if(!gMPFloatPtr) {
 120                 for(pos = KERNEL_BASE|0x7F000; pos < (KERNEL_BASE|0x80000); pos += 16) {
 121                         if( *(Uint*)(pos) == MPPTR_IDENT ) {
 122                                 Log("Possible %p", pos);
 123                                 if( ByteSum((void*)pos, sizeof(tMPInfo)) != 0 ) continue;
 124                                 gMPFloatPtr = (void*)pos;
 125                                 break;
 126                         }
 127                 }
 128         }
 129         // - BIOS ROM
 130         if(!gMPFloatPtr) {
 131                 for(pos = KERNEL_BASE|0xE0000; pos < (KERNEL_BASE|0x100000); pos += 16) {
 132                         if( *(Uint*)(pos) == MPPTR_IDENT ) {
 133                                 Log("Possible %p", pos);
 134                                 if( ByteSum((void*)pos, sizeof(tMPInfo)) != 0 ) continue;
 135                                 gMPFloatPtr = (void*)pos;
 136                                 break;
 137                         }
 138                 }
 139         }
 140
 141         // If the MP Table Exists, parse it
 142         if(gMPFloatPtr)
 143         {
 144                  int    i;
 145                 tMPTable_Ent    *ents;
 146                 Log("gMPFloatPtr = %p", gMPFloatPtr);
 147                 Log("*gMPFloatPtr = {");
 148                 Log("\t.Sig = 0x%08x", gMPFloatPtr->Sig);
 149                 Log("\t.MPConfig = 0x%08x", gMPFloatPtr->MPConfig);
 150                 Log("\t.Length = 0x%02x", gMPFloatPtr->Length);
 151                 Log("\t.Version = 0x%02x", gMPFloatPtr->Version);
 152                 Log("\t.Checksum = 0x%02x", gMPFloatPtr->Checksum);
 153                 Log("\t.Features = [0x%02x,0x%02x,0x%02x,0x%02x,0x%02x]",
 154                         gMPFloatPtr->Features[0],       gMPFloatPtr->Features[1],
 155                         gMPFloatPtr->Features[2],       gMPFloatPtr->Features[3],
 156                         gMPFloatPtr->Features[4]
 157                         );
 158                 Log("}");
 159
 160                 mptable = (void*)( KERNEL_BASE|gMPFloatPtr->MPConfig );
 161                 Log("mptable = %p", mptable);
 162                 Log("*mptable = {");
 163                 Log("\t.Sig = 0x%08x", mptable->Sig);
 164                 Log("\t.BaseTableLength = 0x%04x", mptable->BaseTableLength);
 165                 Log("\t.SpecRev = 0x%02x", mptable->SpecRev);
 166                 Log("\t.Checksum = 0x%02x", mptable->Checksum);
 167                 Log("\t.OEMID = '%8c'", mptable->OemID);
 168                 Log("\t.ProductID = '%8c'", mptable->ProductID);
 169                 Log("\t.OEMTablePtr = %p'", mptable->OEMTablePtr);
 170                 Log("\t.OEMTableSize = 0x%04x", mptable->OEMTableSize);
 171                 Log("\t.EntryCount = 0x%04x", mptable->EntryCount);
 172                 Log("\t.LocalAPICMemMap = 0x%08x", mptable->LocalAPICMemMap);
 173                 Log("\t.ExtendedTableLen = 0x%04x", mptable->ExtendedTableLen);
 174                 Log("\t.ExtendedTableChecksum = 0x%02x", mptable->ExtendedTableChecksum);
 175                 Log("}");
 176
 177                 gpMP_LocalAPIC = (void*)MM_MapHWPage(mptable->LocalAPICMemMap, 1);
 178
 179                 ents = mptable->Entries;
 180                 giNumCPUs = 0;
 181
 182                 for( i = 0; i < mptable->EntryCount; i ++ )
 183                 {
 184                          int    entSize = 0;
 185                         switch( ents->Type )
 186                         {
 187                         case 0: // Processor
 188                                 entSize = 20;
 189                                 Log("%i: Processor", i);
 190                                 Log("\t.APICID = %i", ents->Proc.APICID);
 191                                 Log("\t.APICVer = 0x%02x", ents->Proc.APICVer);
 192                                 Log("\t.CPUFlags = 0x%02x", ents->Proc.CPUFlags);
 193                                 Log("\t.CPUSignature = 0x%08x", ents->Proc.CPUSignature);
 194                                 Log("\t.FeatureFlags = 0x%08x", ents->Proc.FeatureFlags);
 195
 196
 197                                 if( !(ents->Proc.CPUFlags & 1) ) {
 198                                         Log("DISABLED");
 199                                         break;
 200                                 }
 201
 202                                 // Check if there is too many processors
 203                                 if(giNumCPUs >= MAX_CPUS) {
 204                                         giNumCPUs ++;   // If `giNumCPUs` > MAX_CPUS later, it will be clipped
 205                                         break;
 206                                 }
 207
 208                                 // Initialise CPU Info
 209                                 gaAPIC_to_CPU[ents->Proc.APICID] = giNumCPUs;
 210                                 gaCPUs[giNumCPUs].APICID = ents->Proc.APICID;
 211                                 gaCPUs[giNumCPUs].State = 0;
 212                                 giNumCPUs ++;
 213
 214                                 // Send IPI
 215                                 if( !(ents->Proc.CPUFlags & 2) )
 216                                 {
 217                                         MP_StartAP( giNumCPUs-1 );
 218                                 }
 219
 220                                 break;
 221                         case 1: // Bus
 222                                 entSize = 8;
 223                                 Log("%i: Bus", i);
 224                                 Log("\t.ID = %i", ents->Bus.ID);
 225                                 Log("\t.TypeString = '%6c'", ents->Bus.TypeString);
 226                                 break;
 227                         case 2: // I/O APIC
 228                                 entSize = 8;
 229                                 Log("%i: I/O APIC", i);
 230                                 Log("\t.ID = %i", ents->IOAPIC.ID);
 231                                 Log("\t.Version = 0x%02x", ents->IOAPIC.Version);
 232                                 Log("\t.Flags = 0x%02x", ents->IOAPIC.Flags);
 233                                 Log("\t.Addr = 0x%08x", ents->IOAPIC.Addr);
 234                                 break;
 235                         case 3: // I/O Interrupt Assignment
 236                                 entSize = 8;
 237                                 Log("%i: I/O Interrupt Assignment", i);
 238                                 Log("\t.IntType = %i", ents->IOInt.IntType);
 239                                 Log("\t.Flags = 0x%04x", ents->IOInt.Flags);
 240                                 Log("\t.SourceBusID = 0x%02x", ents->IOInt.SourceBusID);
 241                                 Log("\t.SourceBusIRQ = 0x%02x", ents->IOInt.SourceBusIRQ);
 242                                 Log("\t.DestAPICID = 0x%02x", ents->IOInt.DestAPICID);
 243                                 Log("\t.DestAPICIRQ = 0x%02x", ents->IOInt.DestAPICIRQ);
 244                                 break;
 245                         case 4: // Local Interrupt Assignment
 246                                 entSize = 8;
 247                                 Log("%i: Local Interrupt Assignment", i);
 248                                 Log("\t.IntType = %i", ents->LocalInt.IntType);
 249                                 Log("\t.Flags = 0x%04x", ents->LocalInt.Flags);
 250                                 Log("\t.SourceBusID = 0x%02x", ents->LocalInt.SourceBusID);
 251                                 Log("\t.SourceBusIRQ = 0x%02x", ents->LocalInt.SourceBusIRQ);
 252                                 Log("\t.DestLocalAPICID = 0x%02x", ents->LocalInt.DestLocalAPICID);
 253                                 Log("\t.DestLocalAPICIRQ = 0x%02x", ents->LocalInt.DestLocalAPICIRQ);
 254                                 break;
 255                         default:
 256                                 Log("%i: Unknown (%i)", i, ents->Type);
 257                                 break;
 258                         }
 259                         ents = (void*)( (Uint)ents + entSize );
 260                 }
 261
 262                 if( giNumCPUs > MAX_CPUS ) {
 263                         Warning("Too many CPUs detected (%i), only using %i of them", giNumCPUs, MAX_CPUS);
 264                         giNumCPUs = MAX_CPUS;
 265                 }
 266
 267                 while( giNumInitingCPUs )
 268                         MM_FinishVirtualInit();
 269
 270                 Panic("Uh oh... MP Table Parsing is unimplemented\n");
 271         }
 272         else {
 273                 Log("No MP Table was found, assuming uniprocessor\n");
 274                 giNumCPUs = 1;
 275                 gTSSs = &gTSS0;
 276         }
 277         #else
 278         giNumCPUs = 1;
 279         gTSSs = &gTSS0;
 280         MM_FinishVirtualInit();
 281         #endif
 282
 283         #if USE_MP
 284         // Initialise Normal TSS(s)
 285         for(pos=0;pos<giNumCPUs;pos++)
 286         {
 287         #else
 288         pos = 0;
 289         #endif
 290                 gTSSs[pos].RSP0 = 0;    // Set properly by scheduler
 291                 gGDT[7+pos*2].LimitLow = sizeof(tTSS) & 0xFFFF;
 292                 gGDT[7+pos*2].BaseLow = ((Uint)(&gTSSs[pos])) & 0xFFFF;
 293                 gGDT[7+pos*2].BaseMid = ((Uint)(&gTSSs[pos])) >> 16;
 294                 gGDT[7+pos*2].BaseHi = ((Uint)(&gTSSs[pos])) >> 24;
 295                 gGDT[7+pos*2+1].DWord[0] = ((Uint)(&gTSSs[pos])) >> 32;
 296         #if USE_MP
 297         }
 298         for(pos=0;pos<giNumCPUs;pos++) {
 299                 __asm__ __volatile__ ("ltr %%ax"::"a"(0x38+pos*16));
 300         }
 301         #else
 302         __asm__ __volatile__ ("ltr %%ax"::"a"(0x38));
 303         #endif
 304
 305         // Set Debug registers
 306         __asm__ __volatile__ ("mov %0, %%db0" : : "r"(&gThreadZero));
 307         __asm__ __volatile__ ("mov %%rax, %%db1" : : "a"(0));
 308
 309         gaCPUs[0].Current = &gThreadZero;
 310
 311         gThreadZero.MemState.CR3 = (Uint)gInitialPML4 - KERNEL_BASE;
 312         gThreadZero.CurCPU = 0;
 313         gThreadZero.KernelStack = 0xFFFFA00000000000 + KERNEL_STACK_SIZE;
 314
 315         // Set timer frequency
 316         outb(0x43, 0x34);       // Set Channel 0, Low/High, Rate Generator
 317         outb(0x40, PIT_TIMER_DIVISOR&0xFF);     // Low Byte of Divisor
 318         outb(0x40, (PIT_TIMER_DIVISOR>>8)&0xFF);        // High Byte
 319
 320         // Create Per-Process Data Block
 321         if( !MM_Allocate(MM_PPD_CFG) )
 322         {
 323                 Warning("Oh, hell, Unable to allocate PPD for Thread#0");
 324         }
 325
 326         Log_Log("Proc", "Multithreading initialised");
 327 }
 328
 329 #if USE_MP
 330 void MP_StartAP(int CPU)
 331 {
 332         Log("Starting AP %i (APIC %i)", CPU, gaCPUs[CPU].APICID);
 333         // Set location of AP startup code and mark for a warm restart
 334         *(Uint16*)(KERNEL_BASE|0x467) = (Uint)&APStartup - (KERNEL_BASE|0xFFFF0);
 335         *(Uint16*)(KERNEL_BASE|0x469) = 0xFFFF;
 336         outb(0x70, 0x0F);       outb(0x71, 0x0A);       // Warm Reset
 337         MP_SendIPI(gaCPUs[CPU].APICID, 0, 5);
 338         giNumInitingCPUs ++;
 339 }
 340
 341 void MP_SendIPI(Uint8 APICID, int Vector, int DeliveryMode)
 342 {
 343         Uint32  addr = (Uint)gpMP_LocalAPIC + 0x300;
 344         Uint32  val;
 345
 346         // High
 347         val = (Uint)APICID << 24;
 348         Log("*%p = 0x%08x", addr+0x10, val);
 349         *(Uint32*)(addr+0x10) = val;
 350         // Low (and send)
 351         val = ((DeliveryMode & 7) << 8) | (Vector & 0xFF);
 352         Log("*%p = 0x%08x", addr, val);
 353         *(Uint32*)addr = val;
 354 }
 355 #endif
 356
 357 /**
 358  * \brief Idle task
 359  */
 360 void Proc_IdleTask(void *ptr)
 361 {
 362         tCPU    *cpu = ptr;
 363         cpu->IdleThread = Proc_GetCurThread();
 364         cpu->IdleThread->ThreadName = (char*)"Idle Thread";
 365         Threads_SetPriority( cpu->IdleThread, -1 );     // Never called randomly
 366         cpu->IdleThread->Quantum = 1;   // 1 slice quantum
 367         for(;;) {
 368                 HALT(); // Just yeilds
 369                 Threads_Yield();
 370         }
 371 }
 372
 373 /**
 374  * \fn void Proc_Start(void)
 375  * \brief Start process scheduler
 376  */
 377 void Proc_Start(void)
 378 {
 379         #if USE_MP
 380          int    i;
 381         #endif
 382
 383         #if USE_MP
 384         // Start APs
 385         for( i = 0; i < giNumCPUs; i ++ )
 386         {
 387                  int    tid;
 388                 if(i)   gaCPUs[i].Current = NULL;
 389
 390                 Proc_NewKThread(Proc_IdleTask, &gaCPUs[i]);
 391
 392                 // Create Idle Task
 393                 gaCPUs[i].IdleThread = Threads_GetThread(tid);
 394
 395
 396                 // Start the AP
 397                 if( i != giProc_BootProcessorID ) {
 398                         MP_StartAP( i );
 399                 }
 400         }
 401
 402         // BSP still should run the current task
 403         gaCPUs[0].Current = &gThreadZero;
 404
 405         // Start interrupts and wait for APs to come up
 406         Log("Waiting for APs to come up\n");
 407         __asm__ __volatile__ ("sti");
 408         while( giNumInitingCPUs )       __asm__ __volatile__ ("hlt");
 409         #else
 410         Proc_NewKThread(Proc_IdleTask, &gaCPUs[0]);
 411
 412         // Start Interrupts (and hence scheduler)
 413         __asm__ __volatile__("sti");
 414         #endif
 415         MM_FinishVirtualInit();
 416         Log_Log("Proc", "Multithreading started");
 417 }
 418
 419 /**
 420  * \fn tThread *Proc_GetCurThread(void)
 421  * \brief Gets the current thread
 422  */
 423 tThread *Proc_GetCurThread(void)
 424 {
 425         #if USE_MP
 426         return gaCPUs[ GetCPUNum() ].Current;
 427         #else
 428         return gaCPUs[ 0 ].Current;
 429         #endif
 430 }
 431
 432 /**
 433  * \brief Create a new kernel thread
 434  */
 435 int Proc_NewKThread(void (*Fcn)(void*), void *Data)
 436 {
 437         Uint    rsp;
 438         tThread *newThread, *cur;
 439
 440         cur = Proc_GetCurThread();
 441         newThread = Threads_CloneTCB(0);
 442         if(!newThread)  return -1;
 443
 444         // Set CR3
 445         newThread->MemState.CR3 = cur->MemState.CR3;
 446
 447         // Create new KStack
 448         newThread->KernelStack = MM_NewKStack();
 449         // Check for errors
 450         if(newThread->KernelStack == 0) {
 451                 free(newThread);
 452                 return -1;
 453         }
 454
 455         rsp = newThread->KernelStack;
 456         *(Uint*)(rsp-=8) = (Uint)Data;  // Data (shadowed)
 457         *(Uint*)(rsp-=8) = 1;   // Number of params
 458         *(Uint*)(rsp-=8) = (Uint)Fcn;   // Function to call
 459         *(Uint*)(rsp-=8) = (Uint)newThread;     // Thread ID
 460
 461         newThread->SavedState.RSP = rsp;
 462         newThread->SavedState.RIP = (Uint)&NewTaskHeader;
 463 //      Log("New (KThread) %p, rsp = %p\n", newThread->SavedState.RIP, newThread->SavedState.RSP);
 464
 465 //      MAGIC_BREAK();
 466         Threads_AddActive(newThread);
 467
 468         return newThread->TID;
 469 }
 470
 471 /**
 472  * \fn int Proc_Clone(Uint Flags)
 473  * \brief Clone the current process
 474  */
 475 int Proc_Clone(Uint Flags)
 476 {
 477         tThread *newThread, *cur = Proc_GetCurThread();
 478         Uint    rip;
 479
 480         // Sanity check
 481         if( !(Flags & CLONE_VM) ) {
 482                 Log_Error("Proc", "Proc_Clone: Don't leave CLONE_VM unset, use Proc_NewKThread instead");
 483                 return -1;
 484         }
 485
 486         // Create new TCB
 487         newThread = Threads_CloneTCB(Flags);
 488         if(!newThread)  return -1;
 489
 490         // Save core machine state
 491         rip = Proc_CloneInt(&newThread->SavedState.RSP, &newThread->MemState.CR3);
 492         if(rip == 0)    return 0;       // Child
 493         newThread->KernelStack = cur->KernelStack;
 494         newThread->SavedState.RIP = rip;
 495
 496         // DEBUG
 497         #if 0
 498         Log("New (Clone) %p, rsp = %p, cr3 = %p", rip, newThread->SavedState.RSP, newThread->MemState.CR3);
 499         {
 500                 Uint cr3;
 501                 __asm__ __volatile__ ("mov %%cr3, %0" : "=r" (cr3));
 502                 Log("Current CR3 = 0x%x, PADDR(RSP) = 0x%x", cr3, MM_GetPhysAddr(newThread->SavedState.RSP));
 503         }
 504         #endif
 505         // /DEBUG
 506
 507         // Lock list and add to active
 508         Threads_AddActive(newThread);
 509
 510         return newThread->TID;
 511 }
 512
 513 /**
 514  * \fn int Proc_SpawnWorker(void)
 515  * \brief Spawns a new worker thread
 516  */
 517 int Proc_SpawnWorker(void (*Fcn)(void*), void *Data)
 518 {
 519         tThread *new, *cur;
 520         Uint    stack_contents[4];
 521
 522         cur = Proc_GetCurThread();
 523
 524         // Create new thread
 525         new = malloc( sizeof(tThread) );
 526         if(!new) {
 527                 Warning("Proc_SpawnWorker - Out of heap space!\n");
 528                 return -1;
 529         }
 530         memcpy(new, &gThreadZero, sizeof(tThread));
 531         // Set Thread ID
 532         new->TID = giNextTID++;
 533
 534         // Create the stack contents
 535         stack_contents[3] = (Uint)Data;
 536         stack_contents[2] = 1;
 537         stack_contents[1] = (Uint)Fcn;
 538         stack_contents[0] = (Uint)new;
 539
 540         // Create a new worker stack (in PID0's address space)
 541         // The stack is relocated by this code
 542         new->KernelStack = MM_NewWorkerStack(stack_contents, sizeof(stack_contents));
 543
 544         new->SavedState.RSP = new->KernelStack - sizeof(stack_contents);
 545         new->SavedState.RIP = (Uint)&NewTaskHeader;
 546
 547 //      Log("New (Worker) %p, rsp = %p\n", new->SavedState.RIP, new->SavedState.RSP);
 548
 549         // Mark as active
 550         new->Status = THREAD_STAT_PREINIT;
 551         Threads_AddActive( new );
 552
 553         return new->TID;
 554 }
 555
 556 /**
 557  * \brief Creates a new user stack
 558  */
 559 Uint Proc_MakeUserStack(void)
 560 {
 561          int    i;
 562         Uint    base = USER_STACK_TOP - USER_STACK_SZ;
 563
 564         // Check Prospective Space
 565         for( i = USER_STACK_SZ >> 12; i--; )
 566         {
 567                 if( MM_GetPhysAddr( base + (i<<12) ) != 0 )
 568                         break;
 569         }
 570
 571         if(i != -1)     return 0;
 572
 573         // Allocate Stack - Allocate incrementally to clean up MM_Dump output
 574         for( i = 0; i < (USER_STACK_SZ-USER_STACK_PREALLOC)/0x1000; i++ )
 575         {
 576                 MM_AllocateZero( base + (i<<12) );
 577         }
 578         for( ; i < USER_STACK_SZ/0x1000; i++ )
 579         {
 580                 tPAddr  alloc = MM_Allocate( base + (i<<12) );
 581                 if( !alloc )
 582                 {
 583                         // Error
 584                         Log_Error("Proc", "Unable to allocate user stack (%i pages requested)", USER_STACK_SZ/0x1000);
 585                         while( i -- )
 586                                 MM_Deallocate( base + (i<<12) );
 587                         return 0;
 588                 }
 589         }
 590
 591         return base + USER_STACK_SZ;
 592 }
 593
 594
 595 /**
 596  * \fn void Proc_StartUser(Uint Entrypoint, Uint *Bases, int ArgC, char **ArgV, char **EnvP, int DataSize)
 597  * \brief Starts a user task
 598  */
 599 void Proc_StartUser(Uint Entrypoint, Uint *Bases, int ArgC, char **ArgV, char **EnvP, int DataSize)
 600 {
 601         Uint    *stack = (void*)Proc_MakeUserStack();
 602          int    i;
 603         Uint    delta;
 604         Uint16  ss, cs;
 605
 606         LOG("stack = 0x%x", stack);
 607
 608         // Copy Arguments
 609         stack = (void*)( (Uint)stack - DataSize );
 610         memcpy( stack, ArgV, DataSize );
 611
 612         // Adjust Arguments and environment
 613         delta = (Uint)stack - (Uint)ArgV;
 614         ArgV = (char**)stack;
 615         for( i = 0; ArgV[i]; i++ )      ArgV[i] += delta;
 616         i ++;
 617         EnvP = &ArgV[i];
 618         for( i = 0; EnvP[i]; i++ )      EnvP[i] += delta;
 619
 620         // User Mode Segments
 621         // 0x2B = 64-bit
 622         ss = 0x23;      cs = 0x2B;
 623
 624         // Arguments
 625         *--stack = (Uint)EnvP;
 626         *--stack = (Uint)ArgV;
 627         *--stack = (Uint)ArgC;
 628         while(*Bases)
 629                 *--stack = *Bases++;
 630
 631         Proc_StartProcess(ss, (Uint)stack, 0x202, cs, Entrypoint);
 632 }
 633
 634 void Proc_StartProcess(Uint16 SS, Uint Stack, Uint Flags, Uint16 CS, Uint IP)
 635 {
 636         if( !(CS == 0x1B || CS == 0x2B) || SS != 0x23 ) {
 637                 Log_Error("Proc", "Proc_StartProcess: CS / SS are not valid (%x, %x)",
 638                         CS, SS);
 639                 Threads_Exit(0, -1);
 640         }
 641 //      Log("Proc_StartProcess: (SS=%x, Stack=%p, Flags=%x, CS=%x, IP=%p)", SS, Stack, Flags, CS, IP);
 642 //      MM_DumpTables(0, USER_MAX);
 643         if(CS == 0x1B)
 644         {
 645                 // 32-bit return
 646                 __asm__ __volatile__ (
 647                         "mov %0, %%rsp;\n\t"    // Set stack pointer
 648                         "mov %2, %%r11;\n\t"    // Set RFLAGS
 649                         "sysret;\n\t"
 650                         : : "r" (Stack), "c" (IP), "r" (Flags)
 651                         );
 652         }
 653         else
 654         {
 655                 // 64-bit return
 656                 __asm__ __volatile__ (
 657                         "mov %0, %%rsp;\n\t"    // Set stack pointer
 658                         "mov %2, %%r11;\n\t"    // Set RFLAGS
 659                         "sysretq;\n\t"
 660                         : : "r" (Stack), "c" (IP), "r" (Flags)
 661                         );
 662         }
 663         for(;;);
 664 }
 665
 666 /**
 667  * \fn int Proc_Demote(Uint *Err, int Dest, tRegs *Regs)
 668  * \brief Demotes a process to a lower permission level
 669  * \param Err   Pointer to user's errno
 670  * \param Dest  New Permission Level
 671  * \param Regs  Pointer to user's register structure
 672  */
 673 int Proc_Demote(Uint *Err, int Dest, tRegs *Regs)
 674 {
 675          int    cpl = Regs->CS & 3;
 676         // Sanity Check
 677         if(Dest > 3 || Dest < 0) {
 678                 *Err = -EINVAL;
 679                 return -1;
 680         }
 681
 682         // Permission Check
 683         if(cpl > Dest) {
 684                 *Err = -EACCES;
 685                 return -1;
 686         }
 687
 688         // Change the Segment Registers
 689         Regs->CS = (((Dest+1)<<4) | Dest) - 8;
 690         Regs->SS = ((Dest+1)<<4) | Dest;
 691
 692         return 0;
 693 }
 694
 695 /**
 696  * \brief Calls a signal handler in user mode
 697  * \note Used for signals
 698  */
 699 void Proc_CallFaultHandler(tThread *Thread)
 700 {
 701         // Never returns
 702         Proc_ReturnToUser(Thread->FaultHandler, Thread->KernelStack, Thread->CurFaultNum);
 703         for(;;);
 704 }
 705
 706 void Proc_DumpThreadCPUState(tThread *Thread)
 707 {
 708         Log("  At %04x:%016llx", Thread->SavedState.UserCS, Thread->SavedState.UserRIP);
 709 }
 710
 711 void Proc_Reschedule(void)
 712 {
 713         tThread *nextthread, *curthread;
 714          int    cpu = GetCPUNum();
 715
 716         // TODO: Wait for it?
 717         if(IS_LOCKED(&glThreadListLock))        return;
 718
 719         curthread = gaCPUs[cpu].Current;
 720
 721         nextthread = Threads_GetNextToRun(cpu, curthread);
 722
 723         if(nextthread == curthread)     return ;
 724         if(!nextthread)
 725                 nextthread = gaCPUs[cpu].IdleThread;
 726         if(!nextthread)
 727                 return ;
 728
 729         #if DEBUG_TRACE_SWITCH
 730         LogF("\nSwitching to task CR3 = 0x%x, RIP = %p, RSP = %p - %i (%s)\n",
 731                 nextthread->MemState.CR3,
 732                 nextthread->SavedState.RIP,
 733                 nextthread->SavedState.RSP,
 734                 nextthread->TID,
 735                 nextthread->ThreadName
 736                 );
 737         #endif
 738
 739         // Update CPU state
 740         gaCPUs[cpu].Current = nextthread;
 741         gTSSs[cpu].RSP0 = nextthread->KernelStack-4;
 742         __asm__ __volatile__ ("mov %0, %%db0" : : "r" (nextthread));
 743
 744         SwitchTasks(
 745                 nextthread->SavedState.RSP, &curthread->SavedState.RSP,
 746                 nextthread->SavedState.RIP, &curthread->SavedState.RIP,
 747                 nextthread->MemState.CR3
 748                 );
 749         return ;
 750 }
 751
 752 /**
 753  * \fn void Proc_Scheduler(int CPU)
 754  * \brief Swap current thread and clears dead threads
 755  */
 756 void Proc_Scheduler(int CPU, Uint RSP, Uint RIP)
 757 {
 758 #if 0
 759         {
 760         tThread *thread;
 761
 762         // If the spinlock is set, let it complete
 763         if(IS_LOCKED(&glThreadListLock))        return;
 764
 765         // Get current thread
 766         thread = gaCPUs[CPU].Current;
 767
 768         if( thread )
 769         {
 770                 tRegs   *regs;
 771                 // Reduce remaining quantum and continue timeslice if non-zero
 772                 if(thread->Remaining--) return;
 773                 // Reset quantum for next call
 774                 thread->Remaining = thread->Quantum;
 775
 776                 // TODO: Make this more stable somehow
 777                 {
 778                         regs = (tRegs*)(RSP+(1)*8);     // CurThread
 779                         thread->SavedState.UserCS = regs->CS;
 780                         thread->SavedState.UserRIP = regs->RIP;
 781                 }
 782         }
 783
 784         // ACK Timer here?
 785
 786         Proc_Reschedule();
 787         }
 788 #endif
 789 }
 790
 791 // === EXPORTS ===
 792 EXPORT(Proc_SpawnWorker);