diff --git a/kernel/src/arch/x86_64/disk/ATA.cpp b/kernel/src/arch/x86_64/disk/ATA.cpp
index dec6f2a5..6835ed98 100644
--- a/kernel/src/arch/x86_64/disk/ATA.cpp
+++ b/kernel/src/arch/x86_64/disk/ATA.cpp
@@ -3,6 +3,7 @@
 #include "arch/Serial.h"
 #include "arch/Timer.h"
 #include "arch/x86_64/IO.h"
+#include "memory/MemoryManager.h"
 #include <luna/Vector.h>
 
 SharedPtr<ATA::Controller> g_controller;
@@ -91,6 +92,26 @@ namespace ATA
         IO::outb(m_control_base + (u16)reg, value);
     }
 
+    u8 Channel::read_bm(BusmasterRegister reg)
+    {
+        return IO::inb(m_busmaster_base + (u16)reg);
+    }
+
+    void Channel::write_bm(BusmasterRegister reg, u8 value)
+    {
+        IO::outb(m_busmaster_base + (u16)reg, value);
+    }
+
+    u32 Channel::read_prdt_address()
+    {
+        return IO::inl(m_busmaster_base + (u16)BusmasterRegister::PRDTAddress);
+    }
+
+    void Channel::write_prdt_address(u32 value)
+    {
+        IO::outl(m_busmaster_base + (u16)BusmasterRegister::PRDTAddress, value);
+    }
+
     void Channel::delay_400ns()
     {
         // FIXME: We should use kernel_sleep(), but it doesn't support nanosecond granularity.
@@ -127,6 +148,15 @@ namespace ATA
         kernel_wait_for_event();
     }
 
+    bool Channel::wait_for_irq_or_timeout(u64 timeout)
+    {
+        m_thread = Scheduler::current();
+
+        kernel_sleep(timeout);
+
+        return !m_thread->sleep_ticks_left;
+    }
+
     bool Channel::wait_for_reg_set(Register reg, u8 value, u64 timeout)
     {
         u64 begin = Timer::ticks_ms();
@@ -320,6 +350,31 @@ namespace ATA
         kinfoln("ata: Drive IDENTIFY returned serial='%s', revision='%s' and model='%s'", m_serial.chars(),
                 m_revision.chars(), m_model.chars());
 
+        auto status = m_channel->read_bm(BusmasterRegister::Status);
+        if (status & BMS_SimplexOnly)
+        {
+            kwarnln("ata: Drive %d will not use DMA because of simplex shenanigans", m_drive_index);
+            m_uses_dma = false;
+        }
+
+        auto frame = MemoryManager::alloc_frame();
+        if (frame.has_error() || frame.value() > 0xffffffff)
+        {
+            kwarnln("ata: Failed to allocate memory below the 32-bit limit for the PRDT");
+            return false;
+        }
+        m_dma_prdt_phys = frame.release_value();
+        m_dma_prdt = (prdt_entry*)MMU::translate_physical_address(m_dma_prdt_phys);
+
+        frame = MemoryManager::alloc_frame();
+        if (frame.has_error() || frame.value() > 0xffffffff)
+        {
+            kwarnln("ata: Failed to allocate memory below the 32-bit limit for DMA memory");
+            return false;
+        }
+        m_dma_mem_phys = frame.release_value();
+        m_dma_mem = (void*)MMU::translate_physical_address(m_dma_mem_phys);
+
         return true;
     }
 
diff --git a/kernel/src/arch/x86_64/disk/ATA.h b/kernel/src/arch/x86_64/disk/ATA.h
index 45770de4..dc980529 100644
--- a/kernel/src/arch/x86_64/disk/ATA.h
+++ b/kernel/src/arch/x86_64/disk/ATA.h
@@ -32,6 +32,13 @@ namespace ATA
         DriveAddress = 1,
     };
 
+    enum class BusmasterRegister : u16
+    {
+        Command = 0,
+        Status = 2,
+        PRDTAddress = 4,
+    };
+
     enum StatusRegister
     {
         SR_Busy = 0x80,
@@ -50,9 +57,28 @@ namespace ATA
         CMD_Identify_Packet = 0xa1
     };
 
+    enum BusMasterStatus
+    {
+        BMS_SimplexOnly = 0x80,
+        BMS_SlaveInit = 0x40,
+        BMS_MasterInit = 0x20,
+        BMS_IRQPending = 0x4,
+        BMS_DMAFailure = 0x2,
+        BMS_DMAMode = 0x1
+    };
+
     class Controller;
     class Channel;
 
+    struct prdt_entry
+    {
+        u32 address;
+        u16 count;
+        u16 flags;
+    };
+
+    static constexpr u16 END_OF_PRDT = (1 << 15);
+
     class Drive
     {
       public:
@@ -74,6 +100,12 @@ namespace ATA
         };
 
         bool m_is_atapi { false };
+        bool m_uses_dma { true };
+
+        volatile prdt_entry* m_dma_prdt;
+        u64 m_dma_prdt_phys;
+        volatile void* m_dma_mem;
+        u64 m_dma_mem_phys;
 
         constexpr static usize SERIAL_LEN = 20;
         constexpr static usize REVISION_LEN = 8;
@@ -95,12 +127,18 @@ namespace ATA
         u8 read_control(ControlRegister reg);
         void write_control(ControlRegister reg, u8 value);
 
+        u8 read_bm(BusmasterRegister reg);
+        void write_bm(BusmasterRegister reg, u8 value);
+        u32 read_prdt_address();
+        void write_prdt_address(u32 value);
+
         bool wait_for_reg_set(Register reg, u8 value, u64 timeout);
         bool wait_for_reg_clear(Register reg, u8 value, u64 timeout);
 
         void delay_400ns();
 
         void wait_for_irq();
+        bool wait_for_irq_or_timeout(u64 timeout);
         void irq_handler(Registers*);
 
         void select(u8 drive);