diff --git a/bin/os b/bin/os
index e2fa090..7f55763 100755
Binary files a/bin/os and b/bin/os differ
diff --git a/check.img b/check.img
new file mode 100644
index 0000000..6dcc0ac
Binary files /dev/null and b/check.img differ
diff --git a/fortune b/fortune
new file mode 100644
index 0000000..ddad7f5
--- /dev/null
+++ b/fortune
@@ -0,0 +1,2 @@
+Many hands make light work.
+		-- John Heywood
diff --git a/image.hdd b/image.hdd
index 4851914..83dddff 100644
Binary files a/image.hdd and b/image.hdd differ
diff --git a/limine.sh b/limine.sh
index 0e26ae1..ca98929 100644
--- a/limine.sh
+++ b/limine.sh
@@ -16,13 +16,14 @@ make -C limine
 ./limine/limine bios-install image.hdd
 
 # Format the image as fat32.
-mformat -i image.hdd@@1M
+mformat -i image.hdd@@1M -F
 
 # Make relevant subdirectories.
-mmd -i image.hdd@@1M ::/EFI ::/EFI/BOOT ::/boot ::/boot/limine
+mmd -i image.hdd@@1M ::/EFI ::/EFI/BOOT ::/boot ::/boot/limine ::/fort
 
 # Copy over the relevant files.
 mcopy -i image.hdd@@1M bin/os ::/boot
 mcopy -i image.hdd@@1M limine.conf limine/limine-bios.sys ::/boot/limine
 mcopy -i image.hdd@@1M limine/BOOTX64.EFI ::/EFI/BOOT
 mcopy -i image.hdd@@1M limine/BOOTIA32.EFI ::/EFI/BOOT
+mcopy -i image.hdd@@1M fortune ::/fort/fortu.ne
diff --git a/mnt/EFI/BOOT/BOOTIA32.EFI b/mnt/EFI/BOOT/BOOTIA32.EFI
new file mode 100644
index 0000000..3246181
Binary files /dev/null and b/mnt/EFI/BOOT/BOOTIA32.EFI differ
diff --git a/mnt/EFI/BOOT/BOOTX64.EFI b/mnt/EFI/BOOT/BOOTX64.EFI
new file mode 100644
index 0000000..671fef4
Binary files /dev/null and b/mnt/EFI/BOOT/BOOTX64.EFI differ
diff --git a/mnt/boot/limine/limine-bios.sys b/mnt/boot/limine/limine-bios.sys
new file mode 100644
index 0000000..bb3c952
Binary files /dev/null and b/mnt/boot/limine/limine-bios.sys differ
diff --git a/mnt/boot/limine/limine.conf b/mnt/boot/limine/limine.conf
new file mode 100644
index 0000000..ec26b22
--- /dev/null
+++ b/mnt/boot/limine/limine.conf
@@ -0,0 +1,5 @@
+timeout: 5
+
+/LispOS
+	protocol: limine
+	path: boot():/boot/os
diff --git a/mnt/boot/os b/mnt/boot/os
new file mode 100755
index 0000000..e51b8cd
Binary files /dev/null and b/mnt/boot/os differ
diff --git a/obj/src/int.c.o b/obj/src/int.c.o
index 1906873..cb5e576 100644
Binary files a/obj/src/int.c.o and b/obj/src/int.c.o differ
diff --git a/obj/src/main.c.d b/obj/src/main.c.d
index 1efc91c..5b08ec0 100644
--- a/obj/src/main.c.d
+++ b/obj/src/main.c.d
@@ -1,19 +1,19 @@
 obj/src/main.c.o: src/main.c src/../include/limine.h src/prompt/lex.h \
-  src/prompt/file.h src/acpi.h src/common.h src/alloc.h src/apic.h \
-  src/kbd.h src/map.h src/print.h src/lib/stdio.h src/lib/malloc.h \
-  src/lib/../common.h src/lib/../kbd.h src/lib/../print.h
+  src/prompt/file.h src/acpi.h src/common.h src/ahci.h src/alloc.h \
+  src/apic.h src/fat32.h src/lib/stdio.h src/lib/malloc.h \
+  src/lib/../common.h src/lib/../kbd.h src/lib/../print.h src/print.h
 src/../include/limine.h:
 src/prompt/lex.h:
 src/prompt/file.h:
 src/acpi.h:
 src/common.h:
+src/ahci.h:
 src/alloc.h:
 src/apic.h:
-src/kbd.h:
-src/map.h:
-src/print.h:
+src/fat32.h:
 src/lib/stdio.h:
 src/lib/malloc.h:
 src/lib/../common.h:
 src/lib/../kbd.h:
 src/lib/../print.h:
+src/print.h:
diff --git a/obj/src/main.c.o b/obj/src/main.c.o
index 1711cd5..5767244 100644
Binary files a/obj/src/main.c.o and b/obj/src/main.c.o differ
diff --git a/run.sh b/run.sh
index 4feec7f..d71d485 100644
--- a/run.sh
+++ b/run.sh
@@ -1,2 +1,6 @@
-qemu-system-x86_64 image.hdd
+qemu-system-x86_64 \
+  -drive id=disk0,file=image.hdd,format=raw,if=none \
+  -device ich9-ahci,id=ahci0 \
+  -device ide-hd,drive=disk0,bus=ahci0.0
+
 #/mnt/c/'Program Files'/qemu/qemu-system-x86_64.exe image.hdd
diff --git a/src/ahci.c b/src/ahci.c
new file mode 100644
index 0000000..3568651
--- /dev/null
+++ b/src/ahci.c
@@ -0,0 +1,492 @@
+#include "alloc.h"
+#include "common.h"
+#include "map.h"
+#include "print.h"
+#include "lib/stdio.h"
+#include "map.h"
+#include "lib/malloc.h"
+
+#define PCI_CONFIG_ADDR 0xCF8
+#define PCI_CONFIG_DATA 0xCFC
+
+static u32 pci_read(u8 bus, u8 slot, u8 func, u8 offset) {
+    u32 addr = (1u << 31) | ((u32)bus << 16) | ((u32)slot << 11) |
+               ((u32)func << 8) | (offset & 0xFC);
+    outd(PCI_CONFIG_ADDR, addr);
+    return ind(PCI_CONFIG_DATA);
+}
+
+static void pci_write(u8 bus, u8 slot, u8 func, u8 offset, u32 value) {
+    u32 addr = (1u << 31) | ((u32)bus << 16) | ((u32)slot << 11) |
+               ((u32)func << 8) | (offset & 0xFC);
+    outd(PCI_CONFIG_ADDR, addr);
+    outd(PCI_CONFIG_DATA, value);
+}
+
+typedef struct {
+    u32 dba;
+    u32 dbau;
+    u32 reserved;
+    u32 dbc;        // byte count - 1, bit 31 = interrupt on completion
+} hba_prdt_t;
+
+#define AHCI_PRDT_PER_TABLE 8
+
+typedef struct {
+    u8         cfis[64];
+    u8         acmd[16];
+    u8         reserved[48];
+    hba_prdt_t prdt[AHCI_PRDT_PER_TABLE];
+} hba_cmd_table_t;
+
+// hba_cmd_table_t size = 64 + 16 + 48 + (8 * 16) = 256 bytes
+// two fit per 4KiB page
+typedef struct {
+    u16 flags;      // [4:0] FIS len in DWORDs, [6] write
+    u16 prdtl;
+    u32 prdbc;
+    u32 ctba;
+    u32 ctbau;
+    u32 reserved[4];
+} hba_cmd_header_t;
+
+typedef struct {
+    u8 dsfis[28];
+    u8 pad0[4];
+    u8 psfis[20];
+    u8 pad1[12];
+    u8 rfis[20];
+    u8 pad2[4];
+    u8 sdbfis[8];
+    u8 ufis[64];
+    u8 reserved[96];
+} hba_fis_t;
+
+// sizeof(hba_fis_t) = 256 bytes
+
+typedef volatile struct {
+    u32 clb, clbu;
+    u32 fb, fbu;
+    u32 is;
+    u32 ie;
+    u32 cmd;
+    u32 reserved0;
+    u32 tfd;
+    u32 sig;
+    u32 ssts;
+    u32 sctl;
+    u32 serr;
+    u32 sact;
+    u32 ci;
+    u32 sntf;
+    u32 fbs;
+    u32 reserved1[11];
+    u32 vendor[4];
+} hba_port_t;
+
+typedef volatile struct {
+    u32 cap;
+    u32 ghc;
+    u32 is;
+    u32 pi;
+    u32 vs;
+    u32 ccc_ctl;
+    u32 ccc_pts;
+    u32 em_loc;
+    u32 em_ctl;
+    u32 cap2;
+    u32 bohc;
+    u8  reserved[0xA0 - 0x2C];
+    u8  vendor[0x100 - 0xA0];
+    hba_port_t ports[32];
+} hba_mem_t;
+
+typedef struct {
+    u8  fis_type;   // 0x27
+    u8  c;          // bit 7 = command
+    u8  command;
+    u8  featurel;
+    u8  lba0, lba1, lba2;
+    u8  device;
+    u8  lba3, lba4, lba5;
+    u8  featureh;
+    u8  countl, counth;
+    u8  icc;
+    u8  control;
+    u8  reserved[4];
+} fis_reg_h2d_t;
+
+#define ATA_CMD_READ_DMA_EXT  0x25
+#define ATA_CMD_WRITE_DMA_EXT 0x35
+#define ATA_CMD_IDENTIFY      0xEC
+#define ATA_DEV_BUSY          0x80
+#define ATA_DEV_DRQ           0x08
+
+
+#define HBA_GHC_AE    (1u << 31)
+#define HBA_CMD_ST    (1u << 0)
+#define HBA_CMD_FRE   (1u << 4)
+#define HBA_CMD_FR    (1u << 14)
+#define HBA_CMD_CR    (1u << 15)
+#define HBA_IS_TFES   (1u << 30)   // task file error status
+
+
+#define AHCI_MAX_PORTS   32
+#define AHCI_CMD_SLOTS   32
+
+typedef struct {
+    hba_cmd_header_t *cmd_list;                     // virtual, 1 page (1KB used)
+    hba_fis_t        *fis;                          // virtual, fits after cmd_list
+    hba_cmd_table_t  *cmd_tables[AHCI_CMD_SLOTS];   // virtual pointers
+    int               active;
+    u64               sector_count;
+    u64               sector_size;
+} ahci_port_t;
+
+static hba_mem_t  *hba;
+static ahci_port_t ports[AHCI_MAX_PORTS];
+
+static void port_stop(hba_port_t *port) {
+    port->cmd &= ~HBA_CMD_ST;
+    port->cmd &= ~HBA_CMD_FRE;
+    while (port->cmd & (HBA_CMD_CR | HBA_CMD_FR))
+        ;
+}
+
+static void port_start(hba_port_t *port) {
+    while (port->cmd & HBA_CMD_CR)
+        ;
+    port->cmd |= HBA_CMD_FRE;
+    port->cmd |= HBA_CMD_ST;
+}
+
+static int find_free_slot(hba_port_t *port) {
+    u32 slots = port->sact | port->ci;
+    for (int i = 0; i < AHCI_CMD_SLOTS; i++)
+        if (!(slots & (1u << i)))
+            return i;
+    return -1;
+}
+
+static void port_init(int idx) {
+    hba_port_t  *port = &hba->ports[idx];
+    ahci_port_t *p    = &ports[idx];
+
+    port_stop(port);
+
+    //
+    // Page 0: command list (1024 bytes) + FIS buffer (256 bytes)
+    //
+    //   [0   .. 1023] : hba_cmd_header_t[32]   (command list)
+    //   [1024.. 1279] : hba_fis_t               (received FIS)
+    //
+    // Both fit in one 4KiB page. FIS is 256-byte aligned because 1024
+    // is a multiple of 256.
+    //
+    u8 *page0 = (u8 *)alloc(4096);
+    memset(page0, 0, 4096);
+
+    p->cmd_list = (hba_cmd_header_t *)page0;
+    p->fis      = (hba_fis_t *)(page0 + 1024);
+
+    u64 phys0        = virt_to_phys(page0);
+    u64 phys_fis     = phys0 + 1024;
+
+    port->clb  = (u32)(phys0);
+    port->clbu = (u32)(phys0 >> 32);
+    port->fb   = (u32)(phys_fis);
+    port->fbu  = (u32)(phys_fis >> 32);
+
+    //
+    // Command tables: 256 bytes each, two per 4KiB page.
+    // 32 slots → 16 pages.
+    //
+    for (int i = 0; i < AHCI_CMD_SLOTS; i += 2) {
+        u8 *page = (u8 *)alloc(4096);
+        memset(page, 0, 4096);
+
+        u64 phys = virt_to_phys(page);
+
+        for (int j = 0; j < 2 && (i + j) < AHCI_CMD_SLOTS; j++) {
+            int slot = i + j;
+            hba_cmd_table_t *tbl = (hba_cmd_table_t *)(page + j * 256);
+            p->cmd_tables[slot] = tbl;
+
+            u64 tbl_phys = phys + j * 256;
+            p->cmd_list[slot].ctba  = (u32)(tbl_phys);
+            p->cmd_list[slot].ctbau = (u32)(tbl_phys >> 32);
+            p->cmd_list[slot].prdtl = AHCI_PRDT_PER_TABLE;
+        }
+    }
+
+    // Clear pending errors and interrupts
+    port->serr = 0xFFFFFFFF;
+    port->is   = 0xFFFFFFFF;
+
+    // Enable D2H register FIS interrupt + task file error
+    port->ie = (1u << 0) | HBA_IS_TFES;
+
+    port_start(port);
+    p->active = 1;
+
+    
+    printf("AHCI: port %d initialized, sig=%x\n", idx, port->sig);
+}
+
+static void port_identify(int idx) {
+    ahci_port_t *p    = &ports[idx];
+    hba_port_t  *port = &hba->ports[idx];
+
+    // IDENTIFY needs a 512-byte buffer; alloc gives us a page
+    u16 *buf  = (u16 *)alloc(4096);
+    u64  phys = virt_to_phys(buf);
+
+    int slot = find_free_slot(port);
+    if (slot < 0) return;
+
+    hba_cmd_header_t *hdr = &p->cmd_list[slot];
+    hdr->flags = sizeof(fis_reg_h2d_t) / 4;  // FIS length, no write bit
+    hdr->prdtl = 1;
+    hdr->prdbc = 0;
+
+    hba_cmd_table_t *tbl = p->cmd_tables[slot];
+    memset(tbl, 0, sizeof(hba_cmd_table_t));
+
+    tbl->prdt[0].dba  = (u32)(phys);
+    tbl->prdt[0].dbau = (u32)(phys >> 32);
+    tbl->prdt[0].dbc  = 511; // 512 bytes - 1, no interrupt bit needed
+
+    fis_reg_h2d_t *fis = (fis_reg_h2d_t *)tbl->cfis;
+    fis->fis_type = 0x27;
+    fis->c        = (1 << 7);
+    fis->command  = ATA_CMD_IDENTIFY;
+    fis->device   = 0;
+
+    // Wait for idle
+    int spin = 0;
+    while ((port->tfd & (ATA_DEV_BUSY | ATA_DEV_DRQ)) && ++spin < 1000000)
+        ;
+
+    port->ci = (1u << slot);
+
+    while (port->ci & (1u << slot)) {
+        if (port->is & HBA_IS_TFES) {
+            printf("AHCI: identify error on port %d\n", idx);
+            return;
+        }
+    }
+
+    // Word 100-103: 48-bit LBA sector count
+    p->sector_count = ((u64)buf[103] << 48) | ((u64)buf[102] << 32) |
+                      ((u64)buf[101] << 16) | ((u64)buf[100]);
+
+
+    // word 106 contains sector size info
+    u16 w106 = buf[106];
+
+    if (w106 & (1 << 14) && !(w106 & (1 << 15))) {
+        // words 117-118: logical sector size in words (if bit 12 of w106 is set)
+        if (w106 & (1 << 12)) {
+            u32 logical_size = ((u32)buf[118] << 16) | buf[117];
+            logical_size *= 2; // words to bytes
+            p->sector_size = logical_size;
+        }
+    }
+
+    printf("AHCI: port %d sections = %d\n", idx, p->sector_count);
+}
+
+// buf must be physically contiguous. For >AHCI_PRDT_PER_TABLE*4MB you'd need
+// to split the request; at 8 PRDT entries of up to 4MB each that's 32MB max
+// per call which is plenty for now.
+static int ahci_rw(int idx, u64 lba, u32 sector_count, void *buf, int write) {
+    ahci_port_t *p    = &ports[idx];
+    hba_port_t  *port = &hba->ports[idx];
+
+    if (!p->active) {
+        print("AHCI: port not active\n");
+        return -1;
+    }
+
+    int slot = find_free_slot(port);
+    if (slot < 0) {
+        print("AHCI: no free command slots\n");
+        return -1;
+    }
+
+    hba_cmd_header_t *hdr = &p->cmd_list[slot];
+    hdr->flags = (u16)(sizeof(fis_reg_h2d_t) / 4);
+    if (write) hdr->flags |= (1 << 6);
+    hdr->prdbc = 0;
+
+    hba_cmd_table_t *tbl = p->cmd_tables[slot];
+    memset(tbl, 0, sizeof(hba_cmd_table_t));
+
+    // Fill PRDT — each entry covers up to 4MB (0x3FFFFF bytes + 1).
+    // We use 512KB chunks to stay conservative and keep entry count low.
+    #define CHUNK_SIZE (512 * 1024)
+
+    u32 remaining   = sector_count * 512;
+    u64 buff_phys    = virt_to_phys(buf);
+    u16 prdt_count  = 0;
+
+    while (remaining > 0 && prdt_count < AHCI_PRDT_PER_TABLE) {
+        u32 chunk = remaining < CHUNK_SIZE ? remaining : CHUNK_SIZE;
+
+        tbl->prdt[prdt_count].dba  = (u32)(buff_phys);
+        tbl->prdt[prdt_count].dbau = (u32)(buff_phys >> 32);
+        // dbc is bytes - 1; set interrupt on completion only on last entry
+        tbl->prdt[prdt_count].dbc  = (chunk - 1);
+
+        buff_phys  += chunk;
+        remaining -= chunk;
+        prdt_count++;
+    }
+
+    // Interrupt on last PRDT entry
+    tbl->prdt[prdt_count - 1].dbc |= (1u << 31);
+
+    hdr->prdtl = prdt_count;
+
+    // Build the command FIS
+    fis_reg_h2d_t *fis = (fis_reg_h2d_t *)tbl->cfis;
+    fis->fis_type = 0x27;
+    fis->c        = (1 << 7);
+    fis->command  = write ? ATA_CMD_WRITE_DMA_EXT : ATA_CMD_READ_DMA_EXT;
+    fis->device   = (1 << 6); // LBA mode
+
+    fis->lba0 = (u8)(lba >>  0);
+    fis->lba1 = (u8)(lba >>  8);
+    fis->lba2 = (u8)(lba >> 16);
+    fis->lba3 = (u8)(lba >> 24);
+    fis->lba4 = (u8)(lba >> 32);
+    fis->lba5 = (u8)(lba >> 40);
+
+    fis->countl = (u8)(sector_count >> 0);
+    fis->counth = (u8)(sector_count >> 8);
+
+    // Wait until port is not busy
+    int spin = 0;
+    while ((port->tfd & (ATA_DEV_BUSY | ATA_DEV_DRQ)) && ++spin < 1000000)
+        ;
+    if (spin >= 1000000) {
+        print("AHCI: port hung before issue\n");
+        return -1;
+    }
+
+    port->is = 0xFFFFFFFF; // clear interrupts before issuing
+    port->ci = (1u << slot);
+
+    // Poll for completion
+    while (1) {
+        if (!(port->ci & (1u << slot)))
+            break;
+        if (port->is & HBA_IS_TFES) {
+            print("AHCI: task file error (port ");
+            print32(idx);
+            print(")\n");
+            return -1;
+        }
+    }
+
+    // Final error check after CI clears
+    if (port->is & HBA_IS_TFES) {
+        print("AHCI: task file error after completion\n");
+        return -1;
+    }
+
+    return 0;
+}
+
+int ahci_read(int port, u64 lba, u32 count, void *buf) {
+    int ret = ahci_rw(port, lba, count, buf, 0);
+    if(ret < 0) {
+        printf("Failed ahci read (%d, 0x", port);
+        print64(lba);
+        printf(", %u, %u)\n", count, buf);
+    }
+    return ret;
+}
+
+int ahci_write(int port, u64 lba, u32 count, void *buf) {
+    return ahci_rw(port, lba, count, buf, 1);
+}
+
+u64 ahci_sector_count(int port) {
+    return ports[port].sector_count;
+}
+
+int ahci_port_active(int port) {
+    return ports[port].active;
+}
+
+void ahci_init() {
+    // Scan PCI for AHCI controller (class=0x01, subclass=0x06)
+    u8  found_bus = 0, found_slot = 0, found_func = 0;
+    int found = 0;
+
+    for (u32 bus = 0; bus < 256 && !found; bus++) {
+        for (u8 slot = 0; slot < 32 && !found; slot++) {
+            for (u8 func = 0; func < 8 && !found; func++) {
+                u32 id = pci_read(bus, slot, func, 0x00);
+                if ((id & 0xFFFF) == 0xFFFF) continue;
+
+                u32 class = pci_read(bus, slot, func, 0x08);
+                if (((class >> 24) & 0xFF) == 0x01 &&
+                    ((class >> 16) & 0xFF) == 0x06) {
+                    found_bus  = bus;
+                    found_slot = slot;
+                    found_func = func;
+                    found = 1;
+                }
+            }
+        }
+    }
+
+    if (!found) {
+        print("AHCI: no controller found\n");
+        return;
+    }
+
+    // Enable bus mastering + memory space
+    u32 pcicmd = pci_read(found_bus, found_slot, found_func, 0x04);
+    pcicmd |= (1 << 1) | (1 << 2);
+    pci_write(found_bus, found_slot, found_func, 0x04, pcicmd);
+
+    // BAR5 = ABAR
+    u32 bar5      = pci_read(found_bus, found_slot, found_func, 0x24);
+    u64 abar_phys = (u64)(bar5 & 0xFFFFF000);
+
+    hba = (hba_mem_t *)map_mmio(abar_phys, 1<<12);
+    if (!hba) {
+        print("AHCI: failed to map HBA\n");
+        return;
+    }
+
+    printf("AHCI: HBA mapped, version=%d", hba->vs);
+
+    // Take AHCI ownership from BIOS if BOHC is supported (CAP2.BOH)
+    if (hba->cap2 & (1 << 0)) {
+        hba->bohc |= (1 << 1); // set OOS
+        while (hba->bohc & (1 << 0)) // wait for BOS to clear
+            ;
+    }
+
+    // Enable AHCI mode
+    hba->ghc |= HBA_GHC_AE;
+
+    // Init all ports that have a device attached
+    u32 pi = hba->pi;
+    for (int i = 0; i < AHCI_MAX_PORTS; i++) {
+        if (!(pi & (1u << i))) continue;
+
+        u32 ssts = hba->ports[i].ssts;
+        u8  det  = (ssts >> 0) & 0xF;  // device detection
+        u8  ipm  = (ssts >> 8) & 0xF;  // interface power management
+
+        if (det == 3 && ipm == 1) {
+            port_init(i);
+            port_identify(i);
+        }
+    }
+}
diff --git a/src/ahci.h b/src/ahci.h
new file mode 100644
index 0000000..93deac4
--- /dev/null
+++ b/src/ahci.h
@@ -0,0 +1,9 @@
+#pragma once
+
+#include "common.h"
+
+int ahci_read(int port, u64 lba, u32 count, void *buf);
+int ahci_write(int port, u64 lba, u32 count, void *buf);
+u64 ahci_sector_count(int port);
+int ahci_port_active(int port);
+void ahci_init();
diff --git a/src/alloc.c b/src/alloc.c
index 0deb914..af1a838 100644
--- a/src/alloc.c
+++ b/src/alloc.c
@@ -2,6 +2,7 @@
 #include "../include/limine.h"
 #include "common.h"
 #include "print.h"
+#include "lib/stdio.h"
 
 #ifdef DBG
   #define dprint(x) print(x)
@@ -115,7 +116,7 @@ static void *addr(size_t page) { return alloc_start + page * PAGE_SIZE; }
 static void usepage(size_t page) {
   size_t idx = page / 64;
   size_t bit = page & 63;
-  alloc_bitmap[idx] |= (1 << bit);
+  alloc_bitmap[idx] |= (1ULL << bit);
 }
 
 static void usepagerange(size_t page, size_t len) {
@@ -127,7 +128,7 @@ static void usepagerange(size_t page, size_t len) {
 static void unusepage(size_t page) {
   size_t idx = page / 64;
   size_t bit = page & 63;
-  alloc_bitmap[idx] &= ~(1 << bit);
+  alloc_bitmap[idx] &= ~(1ULL << bit);
 }
 
 static void unusepagerange(size_t page, size_t len) {
@@ -190,6 +191,7 @@ void *alloc(size_t size) {
         if (++consecutive >= pages) {
           usepagerange(page, pages);  
           dprint("succeeded!\n");
+          
           return addr(page);
         }
       } else {
diff --git a/src/common.h b/src/common.h
index 829193e..738c498 100644
--- a/src/common.h
+++ b/src/common.h
@@ -34,12 +34,12 @@ static inline u8 inb(u16 port) {
 }
 static inline u16 inw(u16 port) {
   u16 value;
-  asm volatile("inw %w1, %b0" : "=a"(value) : "Nd"(port) : "memory");
+  asm volatile("inw %w1, %0" : "=a"(value) : "Nd"(port) : "memory");
   return value;
 }
 static inline u32 ind(u16 port) {
   u32 value;
-  asm volatile("ind %w1, %b0" : "=a"(value) : "Nd"(port) : "memory");
+  asm volatile("inl %w1, %0" : "=a"(value) : "Nd"(port) : "memory");
   return value;
 }
 
diff --git a/src/fat32.h b/src/fat32.h
new file mode 100644
index 0000000..ba75555
--- /dev/null
+++ b/src/fat32.h
@@ -0,0 +1,143 @@
+#pragma once
+
+#include "ahci.h"
+#include "alloc.h"
+#include "common.h"
+#include "lib/stdio.h"
+
+// FAT32 BPB (BIOS Parameter Block) sits at the start of sector 0
+typedef pstruct {
+    u8  jump[3];
+    u8  oem[8];
+    u16 bytes_per_sector;
+    u8  sectors_per_cluster;
+    u16 reserved_sectors;
+    u8  fat_count;
+    u16 root_entry_count;   // 0 for FAT32
+    u16 total_sectors_16;   // 0 for FAT32
+    u8  media_type;
+    u16 fat_size_16;        // 0 for FAT32
+    u16 sectors_per_track;
+    u16 head_count;
+    u32 hidden_sectors;
+    u32 total_sectors_32;
+
+    // FAT32 extended fields
+    u32 fat_size_32;
+    u16 ext_flags;
+    u16 fs_version;         // must be 0x0000
+    u32 root_cluster;       // usually 2
+    u16 fs_info_sector;
+    u16 backup_boot_sector;
+    u8  reserved[12];
+    u8  drive_number;
+    u8  reserved1;
+    u8  boot_signature;     // 0x29
+    u32 volume_id;
+    u8  volume_label[11];
+    u8  fs_type[8];         // "FAT32   "
+ } fat32_bpb_t;
+
+
+typedef pstruct {
+    u8 status; // 0x80 = bootable
+    u8 chs_first[3]; // legacy
+    u8 type; // partition type
+    u8 chs_last[3]; // legacy
+    u32 lba_start;  // first sector of the partition
+    u32 sector_count; // size in sectors
+} fat32_mbr_partition_t;
+
+typedef pstruct {
+     u8 bootcode[446];
+     fat32_mbr_partition_t partitions[4];
+     u16 signature;
+ } fat32_mbr_t;
+
+ typedef pstruct {
+     u32 fat_start;
+     u32 data_start;
+     u32 sectors_per_cluster;
+     u32 root_cluster;
+     u32 part_start;
+ } fat32_vol_t;
+
+ static fat32_vol_t vol_from_bpb(fat32_bpb_t* bpb, u32 part_start) {
+     fat32_vol_t vol;
+     vol.part_start = part_start;
+     vol.sectors_per_cluster = bpb->sectors_per_cluster;
+     vol.fat_start = part_start + bpb->reserved_sectors;
+     vol.data_start = vol.fat_start + bpb->fat_count * bpb->fat_size_32;
+     vol.root_cluster = bpb->root_cluster;
+
+     return vol;
+ }
+
+#define TODO_MAKE_ME_DYNAMIC_BLOCK_SIZE 512
+
+static u32 fat32_next(int port, fat32_vol_t* vol, u32 cluster) {
+    u32 fat_off = cluster*4;
+    u32 fat_sect = vol->fat_start + fat_off / TODO_MAKE_ME_DYNAMIC_BLOCK_SIZE;
+    u32 fat_idx = (fat_off & (TODO_MAKE_ME_DYNAMIC_BLOCK_SIZE-1));
+
+    printf("fat_next: cluster=%u fat_off=%u fat_sect=%u byte_off=%u\n",
+        cluster, fat_off, fat_sect, fat_off);
+
+    u8* buff = (u8*) alloc(4096);
+    if(ahci_read(port, fat_sect, 2, buff) < 0) {
+        printf("fat_next: read failed at lba=%u\n", fat_sect);
+        return 0x0FFFFFFF;
+    }
+
+    u32* tbl = (u32*) (buff + fat_idx);
+    // mask top 4
+    return *tbl & 0x0FFFFFFF;
+}
+
+static bool fat32_is_eof(u32 cluster) {
+    return cluster >= 0x0FFFFFF8;
+}
+
+static u64 fat32_cluster_to_lba(fat32_vol_t* vol, u32 cluster) {
+    // first 2 are resv.
+    
+    u64 lba = vol->data_start + (u64)(cluster-2) * vol->sectors_per_cluster;
+
+    return lba;
+}
+
+static int fat32_read_cluster(int port, fat32_vol_t* vol, u32 cluster, u8* buff) {
+    u64 lba = fat32_cluster_to_lba(vol, cluster);
+    return ahci_read(port, lba, vol->sectors_per_cluster, buff);
+}
+
+/* directory entries and files */
+typedef pstruct {
+    u8 name[8];
+    u8 ext[3];
+    u8 attr;
+    u8 resv;
+    u8 creation_time_tenth;
+    u16 creation_time;
+    u16 creation_date;
+    u16 access_date;
+    u16 cluster_high; // high 16 bits of first cluster
+    u16 modify_time;
+    u16 modify_date;
+    u16 cluster_low; //  low 16 bits of first cluster
+    u32 file_size;
+} fat32_dirent_t;
+
+enum fat32_attr {
+    ATTR_READ_ONLY = 1 << 0,
+    ATTR_HIDDEN = 1 << 1,
+    ATTR_SYSTEM = 1 << 2,
+    ATTR_VOL_ID = 1 << 3,
+    ATTR_DIR = 1 << 4,
+    ATTR_ARCH = 1 << 5,
+    ATTR_LFN = 0xF, // long file name entry
+};
+
+static u32 fat32_dirent_cluster(fat32_dirent_t* dirent) {
+    return ((u32)dirent->cluster_high << 16) | dirent->cluster_low;
+}
diff --git a/src/main.c b/src/main.c
index 9f7e76e..c7445a3 100644
--- a/src/main.c
+++ b/src/main.c
@@ -9,11 +9,11 @@
 void print_tok(struct token* tok);
 
 #include "acpi.h"
+#include "ahci.h"
 #include "alloc.h"
 #include "apic.h"
 #include "common.h"
-#include "kbd.h"
-#include "map.h"
+#include "fat32.h"
 #include "print.h"
 #include "lib/stdio.h"
 
@@ -108,11 +108,11 @@ void hang() {
 
 struct limine_framebuffer *fb;
 
-void draw(u8 vect) {
+void draw() {
   for (size_t y = 0; y < fb->height; y++) {
     for (size_t x = 0; x < fb->width; x++) {
       volatile u32 *fb_ptr = fb->address;
-      int tone = ((x ^ y) & 0xFF) >> 6;
+      int tone = ((x ^ y) & 0xFF);
       fb_ptr[y * fb->width + x] = (tone << 16);
     }
   }
@@ -142,6 +142,64 @@ void enable_ps2_keyboard_interrupts(void) {
   outb(0x64, 0xAE);
 }
 
+
+void ls(int port, int indent, fat32_vol_t* vol, u32 cluster) {
+  u32 cluster_size = vol->sectors_per_cluster * TODO_MAKE_ME_DYNAMIC_BLOCK_SIZE;
+
+  u8* lsbuff = alloc(4096);
+  do {
+    fat32_read_cluster(port, vol, cluster, lsbuff);
+    fat32_dirent_t* dir = (fat32_dirent_t*) lsbuff;
+    u32 cnt = cluster_size / sizeof(fat32_dirent_t);
+    for(size_t i = 0; i < cnt; i++) {
+      if(dir[i].name[0] == 0x00) {
+        return;
+      }
+
+      for(int i = 0; i < indent; i++) {
+        printf("----");
+      }
+
+      printf("[%d] ", i);
+
+      if(dir[i].name[0] == 0xE5) {
+        printf("[DELETED]\n");
+        continue;
+      }
+
+      if((dir[i].attr & 0x3F) == ATTR_LFN) {
+        printf("Long File Name\n", i);
+        continue;
+      }
+      if(dir[i].attr & ATTR_VOL_ID) {
+        printf("Volume ID\n");
+        continue;
+      }
+
+
+      for(size_t c = 0; c < 8 && dir[i].name[c] != ' '; c++) {
+        put(dir[i].name[c]);
+      }
+      if(dir[i].ext[0] && dir[i].ext[0] != ' ') {
+        put('.');
+        for(size_t c = 0; c < 8 && dir[i].ext[c] != ' '; c++) {
+          put(dir[i].ext[c]);
+        }
+      }
+      put('\n');
+      
+      if(dir[i].attr & ATTR_DIR) {
+        if(dir[i].name[0] != '.')
+          ls(port, indent + 1, vol, dir[i].cluster_low + ((u32)(dir[i].cluster_high) << 16));
+      }
+    }
+
+    cluster = fat32_next(port, vol, cluster);
+  } while(!fat32_is_eof(cluster));
+
+  printf("All Done\n");
+}
+
 void kmain() {
   if (LIMINE_BASE_REVISION_SUPPORTED(limine_base_revision) == false) {
     hang();
@@ -153,7 +211,7 @@ void kmain() {
 
   fb = fb_req.response->framebuffers[0];
 
-  draw(1);
+  draw();
   print_init(fb);
   init_idt();
   fg = 0xFFFFFF;
@@ -203,42 +261,112 @@ void kmain() {
 
   asm volatile("sti");
 
-#define LINES 40
-  size_t max_len = MAX_LINE_LEN;
-  char* lines[LINES];
 
-  size_t n_lines = 0;
-  for(n_lines = 0; n_lines < LINES; n_lines++) {
-    char* line = getline_alloc(max_len);
-    if(!line)
-      break;
-    if(line[0] == 'F' && line[1] == 'I' && line[2] == '\n')
-      break;
-    lines[n_lines] = line;
+  ahci_init();
+
+  i8 port = -1;
+  for(u8 i = 0; i < 32; i++) {
+    if(ahci_port_active(i)) {
+      // assume we want the first port
+      if(port < 0)
+        port = i;
+      printf("Port %d is active!\n", i);
+    }
   }
 
-  const char* name = "stdin";
-  struct file input = {
-    .row = 0,
-    .col = 0,
-    .rows = n_lines,
-    .lines = (char**)lines,
-    .name = name,
-    .tokens = NULL,
-    .tail = &input.tokens,
-  };
-
-  int res = lex(&input);
-
-  print("Number of Errors: ");
-  print8(res);
-
-  printf("\n# of Tokens: %d\n", input.tokens->len);
-  struct token* token = input.tokens;
-
-  for(;token; token = token->next) {
-    print_tok(token);
+  if(port < 0) {
+    printf("Couldn't find a port to read the FS from :(\n");
+    hang();
   }
+  
+
+  fat32_mbr_t* mbr = alloc(1<<12);
+  ahci_read(port, 0, 1, mbr);
+
+  i8 bootpart = -1;
+  fat32_mbr_partition_t bpart;
+  printf("Fat32 Partitions:\n");
+  for(size_t i = 0; i < 4; i++) {
+    fat32_mbr_partition_t* part = &mbr->partitions[i];
+    u32 lba_start = part->lba_start;
+    u32 lba_end = lba_start + part->sector_count;
+
+    if(bootpart < 0 && part->status == 0x80) {
+      bootpart = i;
+      bpart = *part;
+    }
+    
+    printf("\t%s, %d Sectors [0x%x-0x%x]\n",
+      (part->status == 0x80)? "[BOOTABLE]" : "[UNBOOTABLE]",
+      part->sector_count,
+      lba_start, lba_end);
+  }
+
+  if(bootpart < 0) {
+    printf("Failed to find boot partition :(\n");
+  }
+
+
+  // reuse buffer
+  fat32_bpb_t* bpb = (fat32_bpb_t*) mbr;
+  mbr = NULL;
+
+  // hardcode for now :(
+  u32 part_lba = 2048;
+  ahci_read(port, part_lba, 1, bpb);
+  
+  fat32_vol_t vol = vol_from_bpb(bpb, part_lba);
+
+  u64 root_lba = fat32_cluster_to_lba(&vol, vol.root_cluster);
+  printf("root_lba=%u\n", (u32)root_lba);
+
+  u8 *raw = alloc(4096);
+  ahci_read(port, root_lba, vol.sectors_per_cluster, raw);
+  printf("raw: ");
+  for(int i = 0; i < 64; i++)
+      printf("%x ", raw[i]);
+  printf("\n");
+
+  ls(port, 0, &vol, vol.root_cluster);
+
+  //dealloc((void*) bpb, 1<<12);
+
+// #define LINES 40
+//   size_t max_len = MAX_LINE_LEN;
+//   char* lines[LINES];
+
+//   size_t n_lines = 0;
+//   for(n_lines = 0; n_lines < LINES; n_lines++) {
+//     char* line = getline_alloc(max_len);
+//     if(!line)
+//       break;
+//     if(line[0] == 'F' && line[1] == 'I' && line[2] == '\n')
+//       break;
+//     lines[n_lines] = line;
+//   }
+
+//   const char* name = "stdin";
+//   struct file input = {
+//     .row = 0,
+//     .col = 0,
+//     .rows = n_lines,
+//     .lines = (char**)lines,
+//     .name = name,
+//     .tokens = NULL,
+//     .tail = &input.tokens,
+//   };
+
+//   int res = lex(&input);
+
+//   print("Number of Errors: ");
+//   print8(res);
+
+//   printf("\n# of Tokens: %d\n", input.tokens->len);
+//   struct token* token = input.tokens;
+
+//   for(;token; token = token->next) {
+//     print_tok(token);
+//   }
 
   hang();
 }
diff --git a/src/print.c b/src/print.c
index 13dd1f4..1d53c92 100644
--- a/src/print.c
+++ b/src/print.c
@@ -75,7 +75,7 @@ void put(int ch) {
       size_t off = (py + y) * fb_width + (px + x);
       bool padding =
           (y == 0 || y == GLYPH_HEIGHT - 1 || x == 0 || x == GLYPH_WIDTH - 1);
-      fb_scr[off] = ((~glyph & 1) | padding) ? bg : fg;
+      fb_scr[off] = ((~glyph & 1) | padding) ? ((fb_scr[off] == fg) ? bg : fb_scr[off]) : fg;
       glyph >>= !padding;
     }
   }
@@ -132,4 +132,4 @@ void unput(void) {
     for (size_t y = 0; y < GLYPH_HEIGHT; y++)
         for (size_t x = 0; x < GLYPH_WIDTH; x++)
             fb_scr[(py + y) * fb_width + (px + x)] = bg;
-}
\ No newline at end of file
+}