diff --git a/arch/amd64/hw/exc_s.S b/arch/amd64/hw/exc_s.S
index de22394..33e59b5 100644
--- a/arch/amd64/hw/exc_s.S
+++ b/arch/amd64/hw/exc_s.S
@@ -18,6 +18,15 @@ amd64_exc_isr_\n:
 
 .section .text
 amd64_exc_generic:
+    // 0x18: cs
+    // 0x10: rip
+    // 0x08: error code
+    // 0x00: error number
+    cmpq $0x08, 0x18(%rip)
+    jnz 1f
+    swapgs
+1:
+
     // Push full CPU context for dump
     pushq %rax
     pushq %rcx
@@ -37,9 +46,6 @@ amd64_exc_generic:
     pushq %r14
     pushq %r15
 
-    pushq %fs
-    pushq %gs
-
     movq $AMD64_STACK_CTX_CANARY, %rax
     pushq %rax
 
@@ -53,9 +59,6 @@ amd64_exc_generic:
 1:
     addq $8, %rsp
 
-    popq %gs
-    popq %fs
-
     popq %r15
     popq %r14
     popq %r13
diff --git a/arch/amd64/sys/spin_s.S b/arch/amd64/sys/spin_s.S
index bcbd2fe..9ed4b5c 100644
--- a/arch/amd64/sys/spin_s.S
+++ b/arch/amd64/sys/spin_s.S
@@ -1,3 +1,5 @@
+#include "arch/amd64/asm/asm_cpu.h"
+
 .section .text
 .global spin_lock
 .global spin_release
@@ -13,9 +15,12 @@ spin_lock_irqsave:
     cli
 
 spin_lock:
+    movq get_cpu(CPU_ID), %rax
+    shlq $1, %rax
     // %rdi - spinlock
-    lock bts $0, (%rdi)
+    lock btsq $0, (%rdi)
     jc 1f
+    orq %rax, (%rdi)
     retq
 
     // Failed to obtain lock immediately, wait
diff --git a/include/arch/amd64/asm/asm_cpu.h b/include/arch/amd64/asm/asm_cpu.h
index 531a403..761cf9e 100644
--- a/include/arch/amd64/asm/asm_cpu.h
+++ b/include/arch/amd64/asm/asm_cpu.h
@@ -24,6 +24,7 @@
 #define CPU_THREAD              0x08
 #define CPU_TSS                 0x10
 #define CPU_SYSCALL_RSP         0x18
+#define CPU_ID                  0x28
 
 #define TSS_RSP0                0x04
 
@@ -37,9 +38,9 @@ struct cpu {
     struct thread *thread;      // 0x08
     amd64_tss_t *tss;           // 0x10
     uint64_t syscall_rsp;       // 0x18
-    uint64_t ticks;
+    uint64_t ticks;             // 0x20
 
-    uint64_t processor_id;
+    uint64_t processor_id;      // 0x28
 
     // No need to define offsets for these: ther're not accessed
     // from assembly
diff --git a/include/arch/amd64/cpu.h b/include/arch/amd64/cpu.h
index 4a35f2f..b074706 100644
--- a/include/arch/amd64/cpu.h
+++ b/include/arch/amd64/cpu.h
@@ -34,7 +34,6 @@ extern struct cpu cpus[AMD64_MAX_SMP];
 
 static inline struct cpu *get_cpu(void) {
     struct cpu *cpu;
-    _assert(rdmsr(MSR_IA32_KERNEL_GS_BASE));
     asm volatile ("movq %%gs:0, %0":"=r"(cpu));
     return cpu;
 }
diff --git a/sys/char/ring.c b/sys/char/ring.c
index 61192eb..e4083d3 100644
--- a/sys/char/ring.c
+++ b/sys/char/ring.c
@@ -78,7 +78,11 @@ int ring_getc(struct thread *ctx, struct ring *ring, char *c, int err) {
                 ctx->wait_next = ctx;
                 ring->reader_head = ctx;
 
+                _assert(ctx->cpu >= 0);
                 sched_unqueue(ctx, THREAD_WAITING_IO);
+
+                thread_check_signal(ctx, 0);
+                _assert(ctx->cpu >= 0);
             } else {
                 break;
             }
diff --git a/sys/sched.c b/sys/sched.c
index 3e1f88a..05ff186 100644
--- a/sys/sched.c
+++ b/sys/sched.c
@@ -91,7 +91,7 @@ void sched_unqueue(struct thread *thr, enum thread_state new_state) {
 
     struct cpu *cpu = get_cpu();
 #if defined(AMD64_SMP)
-    _assert(thr->cpu >= 0);
+    assert(thr->cpu >= 0, "Tried to unqueue non-queued thread\n");
     if ((int) cpu->processor_id != thr->cpu) {
         // Need to ask another CPU to unqueue the task
         panic("TODO: implement cross-CPU unqueue\n");
diff --git a/sys/thread.c b/sys/thread.c
index d219878..4eba3ca 100644
--- a/sys/thread.c
+++ b/sys/thread.c
@@ -662,28 +662,28 @@ void thread_signal(struct thread *thr, int signum) {
     if (thr->cpu == (int) get_cpu()->processor_id) {
         if (thr == thread_self) {
             kdebug("Signal will be handled now\n");
-            thread_sigenter(signum);
+            //thread_sigenter(signum);
         } else {
             kdebug("Signal will be handled later\n");
             thread_signal_set(thr, signum);
 
-            if (thr->state == THREAD_WAITING) {
-                timer_remove_sleep(thr);
-            }
+            //if (thr->state == THREAD_WAITING) {
+            //    timer_remove_sleep(thr);
+            //}
 
-            sched_queue(thr);
+            //sched_queue(thr);
         }
     } else if (thr->cpu >= 0) {
         kdebug("Signal will be handled later (other cpu%d)\n", thr->cpu);
         thread_signal_set(thr, signum);
 
-        if (thr->state == THREAD_WAITING) {
-            timer_remove_sleep(thr);
-        }
+        //if (thr->state == THREAD_WAITING) {
+        //    timer_remove_sleep(thr);
+        //}
 
-        sched_queue(thr);
+        //sched_queue(thr);
     } else {
-        kdebug("Signal will be handled later\n");
+        kdebug("Signal will be handled later (not running)\n");
         thread_signal_set(thr, signum);
 
         if (thr->state == THREAD_WAITING) {