diff --git a/src/sfs.c b/src/sfs.c
new file mode 100644
index 0000000..cfd4dd0
--- /dev/null
+++ b/src/sfs.c
@@ -0,0 +1,158 @@
+#include "sfs.h"
+#include "ata.h"
+#include "heap.h"
+#include "string.h"
+#include "rtc.h"
+#include "fs.h"
+#include "sched.h"
+#include "vfs.h"
+
+#define sfsDirectBlocks 10
+#define sfsPtrsPerBlock (sfsBlockSize / sizeof(uint32)) /* 128 */
+
+/* 88 bytes: 512/88 = 5 inodes per block (72 bytes/block left unused --
+   inodeLocation and the format-time inodeTableBlocks math already derive
+   everything from sizeof(SfsInode) generically, so this shrinking density
+   is the honest cost of the extra fields, not a layout hack, same story as
+   84 bytes/6-per-block was before gid). Addressing is direct (10 blocks) +
+   single indirect (128 blocks) + double indirect (128 * 128 blocks) +
+   triple indirect (128 * 128 * 128 blocks): a max file size of roughly
+   (10 + 128 + 128*128 + 128*128*128) * 512 bytes ~= 1032 MiB, versus ~8 MiB
+   with direct/single/double alone. mtime/ctime (rtc.h's rtcNow(), seconds
+   since the Unix epoch, truncated to 32 bits -- valid through year 2106,
+   same honest scope as rtc.c's own 2000-2099 century limit), mode (fs.h's
+   fsModeOwner*, real chmod(2) bits), uid (fs.h's FbStat.uid, sched.h's
+   currentThreadUid() at creation) and now gid (fs.h's FbStat.gid, sched.h's
+   currentThreadGid() at creation) are the newest fields, genuinely growing
+   the struct like tripleIndirect did before them -- there is no padding
+   left to repurpose. */
+typedef struct SfsInode {
+    uint32 type;
+    uint32 size;
+    uint32 linkCount;
+    uint32 blockCount;
+    uint32 direct[sfsDirectBlocks];
+    uint32 indirect;
+    uint32 doubleIndirect;
+    uint32 tripleIndirect;
+    uint32 mtime;
+    uint32 ctime;
+    uint32 mode;
+    uint32 uid;
+    uint32 gid;
+} SfsInode;
+
+/* Exactly 64 bytes: 512/64 = 8 directory entries per block. name[56] leaves
+   room for sfsNameMax (55) chars plus the null terminator. */
+typedef struct SfsDirEntry {
+    uint32 inode;
+    uint32 type;
+    char name[56];
+} SfsDirEntry;
+
+typedef struct SfsSuperblock {
+    uint32 magic;
+    uint32 blockSize;
+    uint32 totalBlocks;
+    uint32 inodeCount;
+    uint32 inodeBitmapStart;
+    uint32 inodeBitmapBlocks;
+    uint32 blockBitmapStart;
+    uint32 blockBitmapBlocks;
+    uint32 inodeTableStart;
+    uint32 inodeTableBlocks;
+    uint32 dataStart;
+    uint32 rootInode;
+} SfsSuperblock;
+
+#define sfsMagic 0x53465331u /* "SFS1" */
+
+static SfsSuperblock sb;
+static uint8* inodeBitmap = nullptr;
+static uint8* blockBitmap = nullptr;
+static bool isMounted = false;
+
+/* See sfsGeneration()'s doc comment (sfs.h). Starts at 1 (not 0) so a caller
+   that zero-initializes its own cache can use generation 0 to mean "never
+   bound to anything yet" without that ever colliding with a real one. */
+static uint64 formatGeneration = 1;
+
+/* --- internal helpers (forward-declared so definition order can follow the
+   dependency graph without header noise) --- */
+static bool bitTest(const uint8* map, uint32 bit);
+static void bitSet(uint8* map, uint32 bit);
+static void bitClear(uint8* map, uint32 bit);
+static bool blockRead(uint32 blockAbs, void* buffer);
+static bool blockWrite(uint32 blockAbs, const void* buffer);
+static bool flushInodeBitmap(void);
+static bool flushBlockBitmap(void);
+static bool flushSuperblock(void);
+static void inodeLocation(int32 inodeNum, uint32* blockAbs, uint32* index);
+static bool readInode(int32 inodeNum, SfsInode* out);
+static bool writeInodeRaw(int32 inodeNum, const SfsInode* in);
+static int32 allocInodeNum(void);
+static void freeInodeNum(int32 inodeNum);
+static int32 allocDataBlockAbs(void);
+static void freeDataBlockAbs(uint32 blockAbs);
+static bool loadOrAllocPointerBlock(SfsInode* inode, uint32* holderAbs, uint8* buf);
+static bool getOrAllocBlockForWrite(SfsInode* inode, uint32 logicalIndex, uint32* outAbs);
+static bool getBlockForRead(const SfsInode* inode, uint32 logicalIndex, uint32* outAbs);
+static bool sfsWriteRaw(SfsInode* inode, int32 inodeNum, uint64 position, const void* data, uint64 length);
+static bool sfsTruncateInode(SfsInode* inode, int32 inodeNum, uint32 newSize);
+static bool splitPath(const char* path, char* parentOut, uint64 parentCap, char* leafOut, uint64 leafCap);
+static int32 dirFind(int32 dirInode, const char* name, uint64 nameLen);
+static bool dirAdd(int32 dirInode, const char* name, uint64 nameLen, int32 childInode, uint32 type);
+static bool dirRemove(int32 dirInode, const char* name, uint64 nameLen);
+static int32 createNode(const char* path, uint32 type);
+static bool readSymlinkTarget(int32 inodeNum, char* buf, uint64 cap);
+static int32 lookupFrom(const char* path, bool followFinal, int depth);
+
+static bool bitTest(const uint8* map, uint32 bit)
+{
+    return (map[bit / 8] >> (bit % 8)) & 1u;
+}
+
+static void bitSet(uint8* map, uint32 bit)
+{
+    map[bit / 8] = (uint8)(map[bit / 8] | (uint8)(1u << (bit % 8)));
+}
+
+static void bitClear(uint8* map, uint32 bit)
+{
+    map[bit / 8] = (uint8)(map[bit / 8] & (uint8)~(1u << (bit % 8)));
+}
+
+static bool blockRead(uint32 blockAbs, void* buffer)
+{
+    return ataReadSectors(sfsBaseLba + blockAbs, 1, buffer);
+}
+
+static bool blockWrite(uint32 blockAbs, const void* buffer)
+{
+    return ataWriteSectors(sfsBaseLba + blockAbs, 1, buffer);
+}
+
+static bool flushInodeBitmap(void)
+{
+    for (uint32 i = 0; i < sb.inodeBitmapBlocks; i++)
+        if (!blockWrite(sb.inodeBitmapStart + i, inodeBitmap + (uint64)i * sfsBlockSize))
+            return false;
+    return true;
+}
+
+static bool flushBlockBitmap(void)
+{
+    for (uint32 i = 0; i < sb.blockBitmapBlocks; i++)
+        if (!blockWrite(sb.blockBitmapStart + i, blockBitmap + (uint64)i * sfsBlockSize))
+            return false;
+    return true;
+}
+
+static bool flushSuperblock(void)
+{
+    uint8 buf[sfsBlockSize];
+    memset(buf, 0, sizeof(buf));
+    memcpy(buf, &sb, sizeof(sb));
+    return blockWrite(0, buf);
+}
+