diff --git a/Cargo.toml b/Cargo.toml
index 6e78296fc8..38d63919a7 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -50,6 +50,7 @@ doc = []
 graphical_debug = []
 lpss_debug = []
 multi_core = ["acpi"]
+profiling = []
 #TODO: remove when threading issues are fixed
 pti = []
 qemu_debug = []
diff --git a/src/arch/x86_64/gdt.rs b/src/arch/x86_64/gdt.rs
index 86fa4f1beb..80e26c15dc 100644
--- a/src/arch/x86_64/gdt.rs
+++ b/src/arch/x86_64/gdt.rs
@@ -201,6 +201,8 @@ pub unsafe fn init_paging(stack_offset: usize, cpu_id: LogicalCpuId) {
     pcr.percpu = PercpuBlock {
         cpu_id,
         switch_internals: Default::default(),
+
+        #[cfg(feature = "profiling")]
         profiling: None,
     };
 }
diff --git a/src/arch/x86_64/idt.rs b/src/arch/x86_64/idt.rs
index 35c1410fdb..a5ea9e045a 100644
--- a/src/arch/x86_64/idt.rs
+++ b/src/arch/x86_64/idt.rs
@@ -8,7 +8,6 @@ use hashbrown::HashMap;
 use x86::segmentation::Descriptor as X86IdtEntry;
 use x86::dtables::{self, DescriptorTablePointer};
 
-use crate::profiling::maybe_setup_timer;
 use crate::{interrupt::*, LogicalCpuId};
 use crate::interrupt::irq::{__generic_interrupts_end, __generic_interrupts_start};
 use crate::ipi::IpiKind;
@@ -255,7 +254,8 @@ pub unsafe fn init_generic(cpu_id: LogicalCpuId, idt: &mut Idt) {
     current_idt[0x80].set_flags(IdtFlags::PRESENT | IdtFlags::RING_3 | IdtFlags::INTERRUPT);
     idt.set_reserved_mut(0x80, true);
 
-    maybe_setup_timer(idt, cpu_id);
+    #[cfg(feature = "profiling")]
+    crate::profiling::maybe_setup_timer(idt, cpu_id);
 
     dtables::lidt(&idtr);
 }
diff --git a/src/arch/x86_64/interrupt/exception.rs b/src/arch/x86_64/interrupt/exception.rs
index 3094309f37..6a73d59079 100644
--- a/src/arch/x86_64/interrupt/exception.rs
+++ b/src/arch/x86_64/interrupt/exception.rs
@@ -46,7 +46,14 @@ interrupt_stack!(debug, @paranoid, |stack| {
 });
 
 interrupt_stack!(non_maskable, @paranoid, |stack| {
+    #[cfg(feature = "profiling")]
     crate::profiling::nmi_handler(stack);
+
+    #[cfg(not(feature = "profiling"))]
+    {
+        println!("Non-maskable interrupt");
+        stack.dump();
+    }
 });
 
 interrupt_stack!(breakpoint, |stack| {
diff --git a/src/arch/x86_64/interrupt/irq.rs b/src/arch/x86_64/interrupt/irq.rs
index c40babf4a9..36191ee762 100644
--- a/src/arch/x86_64/interrupt/irq.rs
+++ b/src/arch/x86_64/interrupt/irq.rs
@@ -252,6 +252,7 @@ interrupt!(lapic_timer, || {
     println!("Local apic timer interrupt");
     lapic_eoi();
 });
+#[cfg(feature = "profiling")]
 interrupt!(aux_timer, || {
     lapic_eoi();
     crate::ipi::ipi(IpiKind::Profile, IpiTarget::Other);
diff --git a/src/arch/x86_64/ipi.rs b/src/arch/x86_64/ipi.rs
index 00c2a0ec0f..d5b11df906 100644
--- a/src/arch/x86_64/ipi.rs
+++ b/src/arch/x86_64/ipi.rs
@@ -5,6 +5,8 @@ pub enum IpiKind {
     Tlb = 0x41,
     Switch = 0x42,
     Pit = 0x43,
+
+    #[cfg(feature = "profiling")]
     Profile = 0x44,
 }
 
@@ -25,6 +27,7 @@ pub fn ipi(_kind: IpiKind, _target: IpiTarget) {}
 pub fn ipi(kind: IpiKind, target: IpiTarget) {
     use crate::device::local_apic::LOCAL_APIC;
 
+    #[cfg(feature = "profiling")]
     if matches!(kind, IpiKind::Profile) {
         let icr = (target as u64) << 18 | 1 << 14 | 0b100 << 8;
         unsafe { LOCAL_APIC.set_icr(icr) };
diff --git a/src/arch/x86_64/start.rs b/src/arch/x86_64/start.rs
index a3bd054387..b7ac9fd919 100644
--- a/src/arch/x86_64/start.rs
+++ b/src/arch/x86_64/start.rs
@@ -6,7 +6,7 @@
 use core::slice;
 use core::sync::atomic::{AtomicBool, AtomicUsize, Ordering, AtomicU32};
 
-use crate::{allocator, memory, LogicalCpuId, profiling};
+use crate::{allocator, memory, LogicalCpuId};
 #[cfg(feature = "acpi")]
 use crate::acpi;
 use crate::arch::pti;
@@ -150,7 +150,8 @@ pub unsafe extern fn kstart(args_ptr: *const KernelArgs) -> ! {
         // Setup kernel heap
         allocator::init();
 
-        profiling::init();
+        #[cfg(feature = "profiling")]
+        crate::profiling::init();
 
         // Set up double buffer for grpahical debug now that heap is available
         #[cfg(feature = "graphical_debug")]
@@ -236,7 +237,8 @@ pub unsafe extern fn kstart_ap(args_ptr: *const KernelArgsAp) -> ! {
         // Set up GDT with TLS
         gdt::init_paging(stack_end, cpu_id);
 
-        profiling::init();
+        #[cfg(feature = "profiling")]
+        crate::profiling::init();
 
         // Set up IDT for AP
         idt::init_paging_post_heap(cpu_id);
diff --git a/src/main.rs b/src/main.rs
index 37e09cf8f9..046a8d893f 100644
--- a/src/main.rs
+++ b/src/main.rs
@@ -122,6 +122,7 @@ mod percpu;
 mod ptrace;
 
 /// Performance profiling of the kernel
+#[cfg(feature = "profiling")]
 pub mod profiling;
 
 /// Schemes, filesystem handlers
@@ -190,6 +191,8 @@ fn kmain(cpu_count: u32, bootstrap: Bootstrap) -> ! {
     info!("Env: {:?}", ::core::str::from_utf8(bootstrap.env));
 
     BOOTSTRAP.call_once(|| bootstrap);
+
+    #[cfg(feature = "profiling")]
     profiling::ready_for_profiling();
 
     match context::contexts_mut().spawn(userspace_init) {
@@ -221,6 +224,7 @@ fn kmain(cpu_count: u32, bootstrap: Bootstrap) -> ! {
 /// This is the main kernel entry point for secondary CPUs
 #[allow(unreachable_code, unused_variables)]
 fn kmain_ap(cpu_id: LogicalCpuId) -> ! {
+    #[cfg(feature = "profiling")]
     profiling::maybe_run_profiling_helper_forever(cpu_id);
 
     if cfg!(feature = "multi_core") {
@@ -229,6 +233,7 @@ fn kmain_ap(cpu_id: LogicalCpuId) -> ! {
         let pid = syscall::getpid();
         info!("AP {}: {:?}", cpu_id, pid);
 
+        #[cfg(feature = "profiling")]
         profiling::ready_for_profiling();
 
         loop {
diff --git a/src/percpu.rs b/src/percpu.rs
index c7bd84b7c2..2107ec60d0 100644
--- a/src/percpu.rs
+++ b/src/percpu.rs
@@ -1,6 +1,5 @@
 use crate::LogicalCpuId;
 use crate::context::switch::ContextSwitchPercpu;
-use crate::profiling::RingBuffer;
 
 /// The percpu block, that stored all percpu variables.
 pub struct PercpuBlock {
@@ -13,7 +12,8 @@ pub struct PercpuBlock {
     // TODO: Put mailbox queues here, e.g. for TLB shootdown? Just be sure to 128-byte align it
     // first to avoid cache invalidation.
 
-    pub profiling: Option<&'static RingBuffer>,
+    #[cfg(feature = "profiling")]
+    pub profiling: Option<&'static crate::profiling::RingBuffer>,
 }
 
 // PercpuBlock::current() is implemented somewhere in the arch-specific modules
diff --git a/src/scheme/debug.rs b/src/scheme/debug.rs
index e2cbbd6c31..a00daaeb8e 100644
--- a/src/scheme/debug.rs
+++ b/src/scheme/debug.rs
@@ -1,7 +1,6 @@
 use core::sync::atomic::{AtomicUsize, Ordering};
 use spin::RwLock;
 
-use crate::LogicalCpuId;
 use crate::arch::debug::Writer;
 use crate::event;
 use crate::scheme::*;
@@ -45,6 +44,8 @@ impl KernelScheme for DebugScheme {
 
         let num = match path {
             "" => !0,
+
+            #[cfg(feature = "profiling")]
             p if p.starts_with("profiling-") => {
                 path[10..].parse().map_err(|_| Error::new(ENOENT))?
             }
@@ -110,12 +111,13 @@ impl KernelScheme for DebugScheme {
             *handles.get(&id).ok_or(Error::new(EBADF))?
         };
 
-        if handle.num == !0 {
-            INPUT
-                .receive_into_user(buf, handle.flags & O_NONBLOCK != O_NONBLOCK, "DebugScheme::read")
-        } else {
-            crate::profiling::drain_buffer(LogicalCpuId::new(handle.num as u32), buf)
+        #[cfg(feature = "profiling")]
+        if handle.num != !0 {
+            return crate::profiling::drain_buffer(crate::LogicalCpuId::new(handle.num as u32), buf);
         }
+
+        INPUT
+            .receive_into_user(buf, handle.flags & O_NONBLOCK != O_NONBLOCK, "DebugScheme::read")
     }
 
     fn kwrite(&self, id: usize, buf: UserSliceRo) -> Result<usize> {
diff --git a/src/scheme/serio.rs b/src/scheme/serio.rs
index 679c53a6f1..13979c5d0e 100644
--- a/src/scheme/serio.rs
+++ b/src/scheme/serio.rs
@@ -27,6 +27,7 @@ static HANDLES: RwLock<BTreeMap<usize, Handle>> = RwLock::new(BTreeMap::new());
 
 /// Add to the input queue
 pub fn serio_input(index: usize, data: u8) {
+    #[cfg(feature = "profiling")]
     crate::profiling::serio_command(index, data);
 
     INPUT[index].send(data);