Kernel: Track previous mode when entering/exiting traps

This allows us to determine what the previous mode (user or kernel)
was, e.g. in the timer interrupt. This is used e.g. to determine
whether a signal handler should be set up.

Fixes #5096

Kernel/Arch/i386/CPU.cpp

@@ -1463,6 +1463,7 @@ u32 Processor::init_context(Thread& thread, bool leave_crit)
     TrapFrame& trap = *reinterpret_cast<TrapFrame*>(stack_top);
     trap.regs = &iretframe;
     trap.prev_irq_level = 0;
+    trap.next_trap = nullptr;
 
     stack_top -= sizeof(u32); // pointer to TrapFrame
     *reinterpret_cast<u32*>(stack_top) = stack_top + 4;
@@ -1612,6 +1613,15 @@ void Processor::enter_trap(TrapFrame& trap, bool raise_irq)
     trap.prev_irq_level = m_in_irq;
     if (raise_irq)
         m_in_irq++;
+    if (m_current_thread) {
+        auto& current_trap = m_current_thread->current_trap();
+        trap.next_trap = current_trap;
+        current_trap = &trap;
+        // The cs register of this trap tells us where we will return back to
+        m_current_thread->set_previous_mode(((trap.regs->cs & 3) != 0) ? Thread::PreviousMode::UserMode : Thread::PreviousMode::KernelMode);
+    } else {
+        trap.next_trap = nullptr;
+    }
 }
 
 void Processor::exit_trap(TrapFrame& trap)
@@ -1624,6 +1634,21 @@ void Processor::exit_trap(TrapFrame& trap)
 
     if (!m_in_irq && !m_in_critical)
         check_invoke_scheduler();
+
+    if (m_current_thread) {
+        auto& current_trap = m_current_thread->current_trap();
+        current_trap = trap.next_trap;
+        if (current_trap) {
+            // If we have another higher level trap then we probably returned
+            // from an interrupt or irq handler. The cs register of the
+            // new/higher level trap tells us what the mode prior to it was
+            m_current_thread->set_previous_mode(((current_trap->regs->cs & 3) != 0) ? Thread::PreviousMode::UserMode : Thread::PreviousMode::KernelMode);
+        } else {
+            // If we don't have a higher level trap then we're back in user mode.
+            // Unless we're a kernel process, in which case we're always in kernel mode
+            m_current_thread->set_previous_mode(m_current_thread->process().is_kernel_process() ? Thread::PreviousMode::KernelMode : Thread::PreviousMode::UserMode);
+        }
+    }
 }
 
 void Processor::check_invoke_scheduler()

Kernel/Arch/i386/CPU.h

@@ -1078,6 +1078,7 @@ private:
 
 struct TrapFrame {
     u32 prev_irq_level;
+    TrapFrame* next_trap;
     RegisterState* regs; // must be last
 
     TrapFrame() = delete;
@@ -1087,7 +1088,7 @@ struct TrapFrame {
     TrapFrame& operator=(TrapFrame&&) = delete;
 };
 
-#define TRAP_FRAME_SIZE (2 * 4)
+#define TRAP_FRAME_SIZE (3 * 4)
 static_assert(TRAP_FRAME_SIZE == sizeof(TrapFrame));
 
 extern "C" void enter_trap_no_irq(TrapFrame*);

Kernel/Scheduler.cpp

@@ -369,11 +369,13 @@ bool Scheduler::context_switch(Thread* thread)
     ASSERT(thread == Thread::current());
 
 #if ARCH(I386)
+    if (thread->process().is_user_process()) {
         auto iopl = get_iopl_from_eflags(Thread::current()->get_register_dump_from_stack().eflags);
-    if (thread->process().is_user_process() && iopl != 0) {
+        if (iopl != 0) {
             dbgln("PANIC: Switched to thread {} with non-zero IOPL={}", Thread::current()->tid().value(), iopl);
             Processor::halt();
         }
+    }
 #endif
 
     return true;
@@ -392,7 +394,7 @@ void Scheduler::enter_current(Thread& prev_thread, bool is_first)
         // Check if we have any signals we should deliver (even if we don't
         // end up switching to another thread).
         auto current_thread = Thread::current();
-        if (!current_thread->is_in_block()) {
+        if (!current_thread->is_in_block() && current_thread->previous_mode() != Thread::PreviousMode::KernelMode) {
             ScopedSpinLock lock(current_thread->get_lock());
             if (current_thread->state() == Thread::Running && current_thread->pending_signals_for_state()) {
                 current_thread->dispatch_one_pending_signal();
@@ -485,6 +487,10 @@ void Scheduler::timer_tick(const RegisterState& regs)
     if (!current_thread)
         return;
 
+    // Sanity checks
+    ASSERT(current_thread->current_trap());
+    ASSERT(current_thread->current_trap()->regs == &regs);
+
     bool is_bsp = Processor::current().id() == 0;
     if (!is_bsp)
         return; // TODO: This prevents scheduling on other CPUs!

Kernel/Syscall.cpp

@@ -137,6 +137,7 @@ void syscall_handler(TrapFrame* trap)
 {
     auto& regs = *trap->regs;
     auto current_thread = Thread::current();
+    ASSERT(current_thread->previous_mode() == Thread::PreviousMode::UserMode);
     auto& process = current_thread->process();
 
     if (auto tracer = process.tracer(); tracer && tracer->is_tracing_syscalls()) {
@@ -206,6 +207,9 @@ void syscall_handler(TrapFrame* trap)
 
     current_thread->check_dispatch_pending_signal();
 
+    // If the previous mode somehow changed something is seriously messed up...
+    ASSERT(current_thread->previous_mode() == Thread::PreviousMode::UserMode);
+
     // Check if we're supposed to return to userspace or just die.
     current_thread->die_if_needed();
 

Kernel/Thread.cpp

@@ -699,6 +699,8 @@ DispatchSignalResult Thread::dispatch_signal(u8 signal)
         return DispatchSignalResult::Deferred;
     }
 
+    ASSERT(previous_mode() == PreviousMode::UserMode);
+
     auto& action = m_signal_action_data[signal];
     // FIXME: Implement SA_SIGINFO signal handlers.
     ASSERT(!(action.flags & SA_SIGINFO));
@@ -762,6 +764,9 @@ DispatchSignalResult Thread::dispatch_signal(u8 signal)
         return DispatchSignalResult::Continue;
     }
 
+    ASSERT(previous_mode() == PreviousMode::UserMode);
+    ASSERT(current_trap());
+
     ProcessPagingScope paging_scope(m_process);
 
     u32 old_signal_mask = m_signal_mask;
@@ -843,7 +848,19 @@ bool Thread::push_value_on_stack(FlatPtr value)
 
 RegisterState& Thread::get_register_dump_from_stack()
 {
-    return *(RegisterState*)(kernel_stack_top() - sizeof(RegisterState));
+    auto* trap = current_trap();
+
+    // We should *always* have a trap. If we don't we're probably a kernel
+    // thread that hasn't been pre-empted. If we want to support this, we
+    // need to capture the registers probably into m_tss and return it
+    ASSERT(trap);
+
+    while (trap) {
+        if (!trap->next_trap)
+            break;
+        trap = trap->next_trap;
+    }
+    return *trap->regs;
 }
 
 RefPtr<Thread> Thread::clone(Process& process)

Kernel/Thread.h

@@ -1106,6 +1106,14 @@ public:
     u32 ticks_in_user() const { return m_ticks_in_user; }
     u32 ticks_in_kernel() const { return m_ticks_in_kernel; }
 
+    enum class PreviousMode : u8 {
+        KernelMode = 0,
+        UserMode
+    };
+    PreviousMode previous_mode() const { return m_previous_mode; }
+    void set_previous_mode(PreviousMode mode) { m_previous_mode = mode; }
+    TrapFrame*& current_trap() { return m_current_trap; }
+
     RecursiveSpinLock& get_lock() const { return m_lock; }
 
 #if LOCK_DEBUG
@@ -1230,6 +1238,7 @@ private:
     NonnullRefPtr<Process> m_process;
     ThreadID m_tid { -1 };
     TSS32 m_tss;
+    TrapFrame* m_current_trap { nullptr };
     Atomic<u32> m_cpu { 0 };
     u32 m_cpu_affinity { THREAD_AFFINITY_DEFAULT };
     u32 m_ticks_left { 0 };
@@ -1261,6 +1270,7 @@ private:
     Atomic<bool> m_is_active { false };
     bool m_is_joinable { true };
     bool m_handling_page_fault { false };
+    PreviousMode m_previous_mode { PreviousMode::UserMode };
 
     unsigned m_syscall_count { 0 };
     unsigned m_inode_faults { 0 };