arch/x86/cpu_context.c

@@ -310,6 +310,43 @@ int cpu_kstate_init(struct cpu_state **ctxt, cpu_kstate_function_arg1_t *start_f
     return 0;
 }
 
+int cpu_ustate_init(struct cpu_state **ctx, uaddr_t startPC, uint32_t arg1, uint32_t arg2,
+                    uaddr_t startSP, vaddr_t kernelStackBottom, size_t kernelStackSize)
+{
+
+    // The user context is stacked above the usual cpu state by the CPU on context switch.
+    // So store it when the cpu expect it (See cpu_kstate_init for more details)
+    struct cpu_ustate *uctx =
+        (struct cpu_ustate *)(kernelStackBottom + kernelStackSize - sizeof(struct cpu_ustate));
+
+    /* If needed, poison the stack */
+#ifdef CPU_STATE_DETECT_UNINIT_KERNEL_VARS
+    memset((void *)kernelStackBottom, CPU_STATE_STACK_POISON, kernelStackSize);
+#elif defined(CPU_STATE_DETECT_KERNEL_STACK_OVERFLOW)
+    cpu_state_prepare_detect_kernel_stack_overflow(stack_bottom, stack_size);
+#endif
+
+    memset(uctx, 0, sizeof(struct cpu_ustate));
+
+    uctx->regs.eip = startPC;
+    uctx->regs.eax = arg1;
+    uctx->regs.ebx = arg2;
+
+    uctx->regs.cs      = BUILD_SEGMENT_REG_VALUE(3, FALSE, SEG_UCODE); // Code
+    uctx->regs.ds      = BUILD_SEGMENT_REG_VALUE(3, FALSE, SEG_UDATA); // Data
+    uctx->regs.es      = BUILD_SEGMENT_REG_VALUE(3, FALSE, SEG_UDATA); // Data
+    uctx->regs.cpl0_ss = BUILD_SEGMENT_REG_VALUE(3, FALSE, SEG_UDATA); // Kernel Stack
+    uctx->cpl3_ss      = BUILD_SEGMENT_REG_VALUE(3, FALSE, SEG_UDATA); // User Stack
+
+    uctx->cpl3_esp = startSP;
+
+    /* The newly created context is initially interruptible */
+    uctx->regs.eflags = (1 << 9); /* set IF bit */
+
+    *ctx = (struct cpu_state *)uctx;
+    return 0;
+}
+
 #if defined(CPU_STATE_DETECT_KERNEL_STACK_OVERFLOW)
 void cpu_state_prepare_detect_kernel_stack_overflow(const struct cpu_state *ctxt,
                                                     vaddr_t stack_bottom, size_t stack_size)

arch/x86/irq_pit.S

@@ -1,3 +1,5 @@
+#define ASM_SOURCE 1
+#include "segment.h"
 .file "irq_pit.S"
 
 .text
@@ -22,6 +24,13 @@ pit_handler:         // already got eflags, cs and eip on stack thanks to CPU
   pushw %fs          //              esp+2
   pushw %gs          //              esp
 
+  /* Set correct kernel segment descriptors' value */
+  movw $BUILD_SEGMENT_REG_VALUE(0, 0, SEG_KDATA), %di
+  pushw %di ; popw %ds
+  pushw %di ; popw %es
+  pushw %di ; popw %fs
+  pushw %di ; popw %gs
+
   /* Send EOI to PIC */
   movb  $0x20, %al
   outb  %al, $0x20

core/process.c

@@ -161,3 +161,8 @@ int processSetName(struct process *proc, char *name)
 
     return 0;
 }
+
+struct mmu_context *processGetMMUContext(struct process *proc)
+{
+    return proc->context;
+}

core/process.h

@@ -14,3 +14,4 @@ int processUnref(struct process *proc);
 int processSetName(struct process *proc, char *name);
 int processAddThread(struct process *proc, struct thread *th);
 int processRemoveThread(struct thread *th);
+struct mmu_context *processGetMMUContext(struct process *th);

core/thread.c

@@ -4,12 +4,15 @@
 #include "irq.h"
 #include "klibc.h"
 #include "list.h"
+#include "mmuContext.h"
 #include "time.h"
 #include "vga.h"
 
 static struct thread *currentThread;
 static struct thread *threadWithTimeout;
 
+static void threadPrepareContext(struct thread *th);
+
 void threadExit()
 {
     uint32_t flags;
@@ -92,13 +95,19 @@ void threadDelete(struct thread *thread)
     uint32_t flags;
     disable_IRQs(flags);
     list_delete(currentThread, thread);
+    restore_IRQs(flags);
+
+    if (thread->squattedContext) {
+        threadChangeCurrentContext(NULL);
+    }
+    if (thread->process)
+        processRemoveThread(thread);
 
 #ifdef DEBUG
     printf("Free stack at 0x%x struct at 0x%x\n", thread->stackAddr, thread);
 #endif
     free((void *)thread->stackAddr);
     free((void *)thread);
-    restore_IRQs(flags);
 }
 
 struct thread *threadSelectNext()
@@ -187,6 +196,7 @@ int threadWait(struct thread *current, struct thread *next, unsigned long msec)
 
     currentThread        = next;
     currentThread->state = RUNNING;
+    threadPrepareContext(next);
     cpu_context_switch(&current->cpuState, next->cpuState);
 
     return current->sleepHaveTimeouted;
@@ -213,6 +223,7 @@ int threadYield()
 
     currentThread        = next;
     currentThread->state = RUNNING;
+    threadPrepareContext(next);
     cpu_context_switch(&current->cpuState, next->cpuState);
     restore_IRQs(flags);
 
@@ -240,6 +251,7 @@ int threadMsleep(unsigned long msec)
 
     currentThread        = next;
     currentThread->state = RUNNING;
+    threadPrepareContext(next);
     cpu_context_switch(&current->cpuState, next->cpuState);
     restore_IRQs(flags);
     return current->sleepHaveTimeouted == 1;
@@ -261,17 +273,61 @@ int threadAddThread(struct thread *th)
     return 0;
 }
 
-void threadPrepareSyscallSwitchBack(struct cpu_state *cpu_state){
+static void threadPrepareContext(struct thread *th)
-    (void)cpu_state;
+{
-        return;
+    if (cpu_context_is_in_user_mode(th->cpuState)) {
+        assert(th->process != NULL);
+        assert(th->squattedContext == NULL);
+        mmuContextSwitch(processGetMMUContext(th->process));
+    } else if (th->squattedContext) {
+        mmuContextSwitch(th->squattedContext);
+    }
 }
 
-void threadPrepareExceptionSwitchBack(struct cpu_state *cpu_state){
+int threadChangeCurrentContext(struct mmu_context *ctx)
-    (void)cpu_state;
+{
-        return;
+    uint32_t flags;
+    struct mmu_context *prev = currentThread->squattedContext;
+
+    if (ctx != NULL) {
+        assert(prev == NULL);
+    } else {
+        assert(prev != NULL);
     }
 
-void threadPrepareIrqSwitchBack(struct cpu_state *cpu_state){
+    disable_IRQs(flags);
-    (void)cpu_state;
+    currentThread->squattedContext = ctx;
-        return;
+
+    if (ctx != NULL) {
+        mmuContextRef(ctx);
+        mmuContextSwitch(ctx);
+    } else {
+        mmuContextUnref(prev);
+    }
+    restore_IRQs(flags);
+
+    return 0;
+}
+
+void threadPrepareSyscallSwitchBack(struct cpu_state *cpuState)
+{
+    currentThread->cpuState = cpuState;
+    threadPrepareContext(currentThread);
+}
+
+void threadPrepareExceptionSwitchBack(struct cpu_state *cpuState)
+{
+    currentThread->cpuState = cpuState;
+    threadPrepareContext(currentThread);
+}
+
+void threadPrepareIrqServicing(struct cpu_state *cpuState)
+{
+    currentThread->cpuState = cpuState;
+}
+
+void threadPrepareIrqSwitchBack(struct cpu_state *cpuState)
+{
+    currentThread->cpuState = cpuState;
+    threadPrepareContext(currentThread);
 }

core/thread.h

@@ -32,6 +32,40 @@ struct thread {
     // For User thread only
     struct thread *nextInProcess, *prevInProcess;
     struct process *process;
+
+  /**
+   * Address space currently "squatted" by the thread, or used to be
+   * active when the thread was interrupted/preempted. This is the MMU
+   * configuration expected before the cpu_state of the thread is
+   * restored on CPU.
+   *   - For kernel threads: should normally be NULL, meaning that the
+   *     thread will squat the current mm_context currently set in the
+   *     MMU. Might be NON NULL when a kernel thread squats a given
+   *     process to manipulate its address space.
+   *   - For user threads: should normally be NULL. More precisely:
+   *       - in user mode: the thread->process.mm_context is ALWAYS
+   *         set on MMU. squatted_mm_context is ALWAYS NULL in this
+   *         situation, meaning that the thread in user mode uses its
+   *         process-space as expected
+   *       - in kernel mode: NULL means that we keep on using the
+   *         mm_context currently set on MMU, which might be the
+   *         mm_context of another process. This is natural since a
+   *         thread in kernel mode normally only uses data in kernel
+   *         space. BTW, this limits the number of TLB flushes. However,
+   *         there are exceptions where this squatted_mm_context will
+   *         NOT be NULL. One is the copy_from/to_user API, which can
+   *         force the effective mm_context so that the MMU will be
+   *         (re)configured upon every context to the thread to match
+   *         the squatted_mm_context. Another exception is when a parent
+   *         thread creates the address space of a child process, in
+   *         which case the parent thread might temporarilly decide to
+   *         switch to the child's process space.
+   *
+   * This is the SOS/matos implementation of the Linux "Lazy TLB" and
+   * address-space loaning.
+   */
+  struct mmu_context *squattedContext;
+
 };
 
 int threadSetup(vaddr_t mainStack, size_t mainStackSize);
@@ -50,3 +84,4 @@ int threadMsleep(unsigned long msec);
 int threadOnJieffiesTick();
 struct thread *getCurrentThread();
 int threadAddThread(struct thread *th);
+int threadChangeCurrentContext(struct mmu_context *ctx);

core/types.h

@@ -33,3 +33,6 @@ typedef unsigned long vaddr_t;
 
 // Physical address
 typedef unsigned long paddr_t;
+
+// Userspace vaddr
+typedef unsigned long uaddr_t;