reddb_file/
file_format.rs

1//! Core persisted file-format constants.
2//!
3//! Runtime crates can own page management and columnar execution, but these
4//! durable magic/version values are file compatibility contracts.
5
6use std::fmt;
7
8/// Magic bytes for database file identification: `RDDB`.
9pub const PAGE_FILE_MAGIC: [u8; 4] = [0x52, 0x44, 0x44, 0x42];
10
11/// Database file version 1.0.0.
12pub const PAGE_FILE_VERSION: u32 = 0x0001_0000;
13
14/// Paged database page size.
15pub const PAGED_PAGE_SIZE: usize = 16_384;
16
17/// Paged database page-header size before the page-0 database header.
18pub const PAGED_PAGE_HEADER_SIZE: usize = 32;
19
20pub const PAGED_CELL_POINTER_SIZE: usize = 2;
21pub const PAGED_CELL_HEADER_SIZE: usize = 6;
22
23/// Raw persisted page header encoded at the start of every paged-store page.
24#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
25pub struct PagedPageHeader {
26    pub page_type: u8,
27    pub flags: u8,
28    pub cell_count: u16,
29    pub free_start: u16,
30    pub free_end: u16,
31    pub page_id: u32,
32    pub parent_id: u32,
33    pub right_child: u32,
34    pub lsn: u64,
35    pub checksum: u32,
36}
37
38pub fn encode_paged_page_header(header: &PagedPageHeader) -> [u8; PAGED_PAGE_HEADER_SIZE] {
39    let mut buf = [0u8; PAGED_PAGE_HEADER_SIZE];
40    buf[0] = header.page_type;
41    buf[1] = header.flags;
42    buf[2..4].copy_from_slice(&header.cell_count.to_le_bytes());
43    buf[4..6].copy_from_slice(&header.free_start.to_le_bytes());
44    buf[6..8].copy_from_slice(&header.free_end.to_le_bytes());
45    buf[8..12].copy_from_slice(&header.page_id.to_le_bytes());
46    buf[12..16].copy_from_slice(&header.parent_id.to_le_bytes());
47    buf[16..20].copy_from_slice(&header.right_child.to_le_bytes());
48    buf[20..28].copy_from_slice(&header.lsn.to_le_bytes());
49    buf[28..32].copy_from_slice(&header.checksum.to_le_bytes());
50    buf
51}
52
53pub fn decode_paged_page_header(buf: &[u8; PAGED_PAGE_HEADER_SIZE]) -> PagedPageHeader {
54    PagedPageHeader {
55        page_type: buf[0],
56        flags: buf[1],
57        cell_count: u16::from_le_bytes([buf[2], buf[3]]),
58        free_start: u16::from_le_bytes([buf[4], buf[5]]),
59        free_end: u16::from_le_bytes([buf[6], buf[7]]),
60        page_id: u32::from_le_bytes([buf[8], buf[9], buf[10], buf[11]]),
61        parent_id: u32::from_le_bytes([buf[12], buf[13], buf[14], buf[15]]),
62        right_child: u32::from_le_bytes([buf[16], buf[17], buf[18], buf[19]]),
63        lsn: u64::from_le_bytes([
64            buf[20], buf[21], buf[22], buf[23], buf[24], buf[25], buf[26], buf[27],
65        ]),
66        checksum: u32::from_le_bytes([buf[28], buf[29], buf[30], buf[31]]),
67    }
68}
69
70pub fn paged_page_type(page: &[u8; PAGED_PAGE_SIZE]) -> u8 {
71    page[0]
72}
73
74pub fn paged_page_id(page: &[u8; PAGED_PAGE_SIZE]) -> u32 {
75    read_u32(page, 8).expect("paged page has header")
76}
77
78pub fn paged_page_lsn(page: &[u8; PAGED_PAGE_SIZE]) -> u64 {
79    read_u64(page, 20).expect("paged page has header")
80}
81
82pub fn set_paged_page_lsn(page: &mut [u8; PAGED_PAGE_SIZE], lsn: u64) {
83    write_u64(page, 20, lsn).expect("paged page has header");
84}
85
86pub fn paged_page_cell_count(page: &[u8; PAGED_PAGE_SIZE]) -> u16 {
87    read_u16(page, 2).expect("paged page has header")
88}
89
90pub fn set_paged_page_cell_count(page: &mut [u8; PAGED_PAGE_SIZE], count: u16) {
91    write_u16(page, 2, count).expect("paged page has header");
92}
93
94pub fn paged_page_parent_id(page: &[u8; PAGED_PAGE_SIZE]) -> u32 {
95    read_u32(page, 12).expect("paged page has header")
96}
97
98pub fn set_paged_page_parent_id(page: &mut [u8; PAGED_PAGE_SIZE], parent_id: u32) {
99    write_u32(page, 12, parent_id).expect("paged page has header");
100}
101
102pub fn paged_page_right_child(page: &[u8; PAGED_PAGE_SIZE]) -> u32 {
103    read_u32(page, 16).expect("paged page has header")
104}
105
106pub fn set_paged_page_right_child(page: &mut [u8; PAGED_PAGE_SIZE], child_id: u32) {
107    write_u32(page, 16, child_id).expect("paged page has header");
108}
109
110pub fn paged_page_free_start(page: &[u8; PAGED_PAGE_SIZE]) -> u16 {
111    read_u16(page, 4).expect("paged page has header")
112}
113
114pub fn set_paged_page_free_start(page: &mut [u8; PAGED_PAGE_SIZE], offset: u16) {
115    write_u16(page, 4, offset).expect("paged page has header");
116}
117
118pub fn paged_page_free_end(page: &[u8; PAGED_PAGE_SIZE]) -> u16 {
119    read_u16(page, 6).expect("paged page has header")
120}
121
122pub fn set_paged_page_free_end(page: &mut [u8; PAGED_PAGE_SIZE], offset: u16) {
123    write_u16(page, 6, offset).expect("paged page has header");
124}
125
126pub fn paged_page_checksum(page: &[u8; PAGED_PAGE_SIZE]) -> u32 {
127    read_u32(page, 28).expect("paged page has header")
128}
129
130pub fn clear_paged_page_checksum(page: &mut [u8; PAGED_PAGE_SIZE]) {
131    write_u32(page, 28, 0).expect("paged page has header");
132}
133
134pub fn set_paged_page_checksum(page: &mut [u8; PAGED_PAGE_SIZE], checksum: u32) {
135    write_u32(page, 28, checksum).expect("paged page has header");
136}
137
138pub fn paged_cell_pointer_offset(index: usize) -> Option<usize> {
139    let offset = PAGED_PAGE_HEADER_SIZE.checked_add(index.checked_mul(PAGED_CELL_POINTER_SIZE)?)?;
140    (offset + PAGED_CELL_POINTER_SIZE <= PAGED_PAGE_SIZE).then_some(offset)
141}
142
143pub fn paged_cell_pointer(page: &[u8; PAGED_PAGE_SIZE], index: usize) -> Option<u16> {
144    read_u16(page, paged_cell_pointer_offset(index)?).ok()
145}
146
147pub fn set_paged_cell_pointer(
148    page: &mut [u8; PAGED_PAGE_SIZE],
149    index: usize,
150    pointer: u16,
151) -> bool {
152    let Some(offset) = paged_cell_pointer_offset(index) else {
153        return false;
154    };
155    if !paged_cell_pointer_is_valid(pointer) {
156        return false;
157    }
158    write_u16(page, offset, pointer).is_ok()
159}
160
161pub fn paged_cell_pointer_is_valid(pointer: u16) -> bool {
162    let pointer = pointer as usize;
163    (PAGED_PAGE_HEADER_SIZE..PAGED_PAGE_SIZE).contains(&pointer)
164}
165
166pub fn paged_cell_len(key_len: usize, value_len: usize) -> Option<usize> {
167    PAGED_CELL_HEADER_SIZE
168        .checked_add(key_len)?
169        .checked_add(value_len)
170}
171
172pub fn paged_cell_total_len(page: &[u8; PAGED_PAGE_SIZE], pointer: u16) -> Option<usize> {
173    let pointer = pointer as usize;
174    if pointer + PAGED_CELL_HEADER_SIZE > PAGED_PAGE_SIZE {
175        return None;
176    }
177    let key_len = read_u16(page, pointer).ok()? as usize;
178    let value_len = read_u32(page, pointer + 2).ok()? as usize;
179    let total_len = paged_cell_len(key_len, value_len)?;
180    (pointer + total_len <= PAGED_PAGE_SIZE).then_some(total_len)
181}
182
183pub fn paged_cell_bytes(page: &[u8; PAGED_PAGE_SIZE], pointer: u16) -> Option<&[u8]> {
184    let total_len = paged_cell_total_len(page, pointer)?;
185    let pointer = pointer as usize;
186    Some(&page[pointer..pointer + total_len])
187}
188
189pub fn paged_cell_key_value(cell: &[u8]) -> Option<(&[u8], &[u8])> {
190    if cell.len() < PAGED_CELL_HEADER_SIZE {
191        return None;
192    }
193    let key_len = u16::from_le_bytes(cell[0..2].try_into().ok()?) as usize;
194    let value_len = u32::from_le_bytes(cell[2..6].try_into().ok()?) as usize;
195    let total_len = paged_cell_len(key_len, value_len)?;
196    if total_len > cell.len() {
197        return None;
198    }
199    let key_start = PAGED_CELL_HEADER_SIZE;
200    let value_start = key_start + key_len;
201    Some((
202        &cell[key_start..value_start],
203        &cell[value_start..value_start + value_len],
204    ))
205}
206
207pub fn write_paged_cell(
208    page: &mut [u8; PAGED_PAGE_SIZE],
209    offset: u16,
210    key: &[u8],
211    value: &[u8],
212) -> bool {
213    let Ok(key_len) = u16::try_from(key.len()) else {
214        return false;
215    };
216    let Ok(value_len) = u32::try_from(value.len()) else {
217        return false;
218    };
219    let Some(total_len) = paged_cell_len(key.len(), value.len()) else {
220        return false;
221    };
222    let offset = offset as usize;
223    if offset + total_len > PAGED_PAGE_SIZE {
224        return false;
225    }
226
227    page[offset..offset + 2].copy_from_slice(&key_len.to_le_bytes());
228    page[offset + 2..offset + 6].copy_from_slice(&value_len.to_le_bytes());
229    page[offset + PAGED_CELL_HEADER_SIZE..offset + PAGED_CELL_HEADER_SIZE + key.len()]
230        .copy_from_slice(key);
231    page[offset + PAGED_CELL_HEADER_SIZE + key.len()..offset + total_len].copy_from_slice(value);
232    true
233}
234
235/// Encryption marker embedded in page 0 when page encryption is enabled.
236pub const PAGED_ENCRYPTION_MARKER: [u8; 4] = *b"RDBE";
237
238/// Current page-0 encryption marker offset.
239///
240/// This preserves the existing on-disk placement. It overlaps the historical
241/// physical-header region, so callers must preserve page-0 bytes around normal
242/// header writes.
243pub const PAGED_ENCRYPTION_MARKER_OFFSET: usize = PAGED_PAGE_HEADER_SIZE + 32;
244
245pub const PAGED_ENCRYPTION_SALT_SIZE: usize = 32;
246pub const PAGED_ENCRYPTION_KEY_CHECK_PLAINTEXT_SIZE: usize = 32;
247pub const PAGED_ENCRYPTION_KEY_CHECK_BLOB_SIZE: usize = 60;
248pub const PAGED_ENCRYPTION_HEADER_SIZE: usize =
249    PAGED_ENCRYPTION_SALT_SIZE + PAGED_ENCRYPTION_KEY_CHECK_BLOB_SIZE;
250
251#[derive(Debug, Clone, PartialEq, Eq)]
252pub struct PagedEncryptionHeader {
253    pub salt: [u8; PAGED_ENCRYPTION_SALT_SIZE],
254    pub key_check: Vec<u8>,
255}
256
257pub fn encode_paged_encryption_header(header: &PagedEncryptionHeader) -> Vec<u8> {
258    let mut out = Vec::with_capacity(PAGED_ENCRYPTION_HEADER_SIZE);
259    out.extend_from_slice(&header.salt);
260    out.extend_from_slice(&header.key_check);
261    out
262}
263
264pub fn decode_paged_encryption_header(
265    data: &[u8],
266) -> Result<PagedEncryptionHeader, DatabaseHeaderError> {
267    ensure_len(data, PAGED_ENCRYPTION_HEADER_SIZE)?;
268    let mut salt = [0u8; PAGED_ENCRYPTION_SALT_SIZE];
269    salt.copy_from_slice(&data[..PAGED_ENCRYPTION_SALT_SIZE]);
270    let key_check = data[PAGED_ENCRYPTION_SALT_SIZE..PAGED_ENCRYPTION_HEADER_SIZE].to_vec();
271    Ok(PagedEncryptionHeader { salt, key_check })
272}
273
274pub fn paged_encryption_marker_present(page: &[u8]) -> bool {
275    page.get(PAGED_ENCRYPTION_MARKER_OFFSET..PAGED_ENCRYPTION_MARKER_OFFSET + 4)
276        == Some(&PAGED_ENCRYPTION_MARKER)
277}
278
279pub fn paged_encryption_header_bytes(page: &[u8]) -> Option<&[u8]> {
280    let start = PAGED_ENCRYPTION_MARKER_OFFSET + PAGED_ENCRYPTION_MARKER.len();
281    page.get(start..start + PAGED_ENCRYPTION_HEADER_SIZE)
282}
283
284pub fn write_paged_encryption_marker_and_header(
285    page: &mut [u8],
286    header_bytes: &[u8],
287) -> Result<(), DatabaseHeaderError> {
288    let marker_end = PAGED_ENCRYPTION_MARKER_OFFSET + PAGED_ENCRYPTION_MARKER.len();
289    let header_end = marker_end + header_bytes.len();
290    ensure_len(page, header_end)?;
291    page[PAGED_ENCRYPTION_MARKER_OFFSET..marker_end].copy_from_slice(&PAGED_ENCRYPTION_MARKER);
292    page[marker_end..header_end].copy_from_slice(header_bytes);
293    Ok(())
294}
295
296const DB_MAGIC_OFFSET: usize = PAGED_PAGE_HEADER_SIZE;
297const DB_VERSION_OFFSET: usize = PAGED_PAGE_HEADER_SIZE + 4;
298const DB_PAGE_SIZE_OFFSET: usize = PAGED_PAGE_HEADER_SIZE + 8;
299const DB_PAGE_COUNT_OFFSET: usize = PAGED_PAGE_HEADER_SIZE + 12;
300const DB_FREELIST_HEAD_OFFSET: usize = PAGED_PAGE_HEADER_SIZE + 16;
301const DB_SCHEMA_VERSION_OFFSET: usize = PAGED_PAGE_HEADER_SIZE + 20;
302const DB_CHECKPOINT_LSN_OFFSET: usize = PAGED_PAGE_HEADER_SIZE + 24;
303const DB_PHYSICAL_FORMAT_VERSION_OFFSET: usize = PAGED_PAGE_HEADER_SIZE + 32;
304const DB_PHYSICAL_SEQUENCE_OFFSET: usize = PAGED_PAGE_HEADER_SIZE + 36;
305const DB_MANIFEST_ROOT_OFFSET: usize = PAGED_PAGE_HEADER_SIZE + 44;
306const DB_MANIFEST_OLDEST_ROOT_OFFSET: usize = PAGED_PAGE_HEADER_SIZE + 52;
307const DB_FREE_SET_ROOT_OFFSET: usize = PAGED_PAGE_HEADER_SIZE + 60;
308const DB_MANIFEST_PAGE_OFFSET: usize = PAGED_PAGE_HEADER_SIZE + 68;
309const DB_MANIFEST_CHECKSUM_OFFSET: usize = PAGED_PAGE_HEADER_SIZE + 72;
310const DB_COLLECTION_ROOTS_PAGE_OFFSET: usize = PAGED_PAGE_HEADER_SIZE + 80;
311const DB_COLLECTION_ROOTS_CHECKSUM_OFFSET: usize = PAGED_PAGE_HEADER_SIZE + 84;
312const DB_COLLECTION_ROOT_COUNT_OFFSET: usize = PAGED_PAGE_HEADER_SIZE + 92;
313const DB_SNAPSHOT_COUNT_OFFSET: usize = PAGED_PAGE_HEADER_SIZE + 96;
314const DB_INDEX_COUNT_OFFSET: usize = PAGED_PAGE_HEADER_SIZE + 100;
315const DB_CATALOG_COLLECTION_COUNT_OFFSET: usize = PAGED_PAGE_HEADER_SIZE + 104;
316const DB_CATALOG_TOTAL_ENTITIES_OFFSET: usize = PAGED_PAGE_HEADER_SIZE + 108;
317const DB_EXPORT_COUNT_OFFSET: usize = PAGED_PAGE_HEADER_SIZE + 116;
318const DB_GRAPH_PROJECTION_COUNT_OFFSET: usize = PAGED_PAGE_HEADER_SIZE + 120;
319const DB_ANALYTICS_JOB_COUNT_OFFSET: usize = PAGED_PAGE_HEADER_SIZE + 124;
320const DB_MANIFEST_EVENT_COUNT_OFFSET: usize = PAGED_PAGE_HEADER_SIZE + 128;
321const DB_REGISTRY_PAGE_OFFSET: usize = PAGED_PAGE_HEADER_SIZE + 132;
322const DB_REGISTRY_CHECKSUM_OFFSET: usize = PAGED_PAGE_HEADER_SIZE + 136;
323const DB_RECOVERY_PAGE_OFFSET: usize = PAGED_PAGE_HEADER_SIZE + 144;
324const DB_RECOVERY_CHECKSUM_OFFSET: usize = PAGED_PAGE_HEADER_SIZE + 148;
325const DB_CATALOG_PAGE_OFFSET: usize = PAGED_PAGE_HEADER_SIZE + 156;
326const DB_CATALOG_CHECKSUM_OFFSET: usize = PAGED_PAGE_HEADER_SIZE + 160;
327const DB_METADATA_STATE_PAGE_OFFSET: usize = PAGED_PAGE_HEADER_SIZE + 168;
328const DB_METADATA_STATE_CHECKSUM_OFFSET: usize = PAGED_PAGE_HEADER_SIZE + 172;
329const DB_VECTOR_ARTIFACT_PAGE_OFFSET: usize = PAGED_PAGE_HEADER_SIZE + 180;
330const DB_VECTOR_ARTIFACT_CHECKSUM_OFFSET: usize = PAGED_PAGE_HEADER_SIZE + 184;
331const DB_CHECKPOINT_IN_PROGRESS_OFFSET: usize = PAGED_PAGE_HEADER_SIZE + 192;
332const DB_CHECKPOINT_TARGET_LSN_OFFSET: usize = PAGED_PAGE_HEADER_SIZE + 193;
333const DB_HEADER_MIN_LEN: usize = DB_CHECKPOINT_TARGET_LSN_OFFSET + 8;
334
335/// Double-write-buffer file magic: `RDDW`.
336pub const DWB_MAGIC: [u8; 4] = [0x52, 0x44, 0x44, 0x57];
337pub const PAGED_DWB_HEADER_SIZE: usize = 12;
338pub const PAGED_DWB_ENTRY_HEADER_SIZE: usize = 4;
339pub const PAGED_DWB_ENTRY_SIZE: usize = PAGED_DWB_ENTRY_HEADER_SIZE + PAGED_PAGE_SIZE;
340
341#[derive(Debug, Clone, PartialEq, Eq)]
342pub struct PagedDwbEntry {
343    pub page_id: u32,
344    pub page: [u8; PAGED_PAGE_SIZE],
345}
346
347#[derive(Debug, Clone, PartialEq, Eq)]
348pub enum PagedDwbFrameError {
349    ShortHeader { got: usize },
350    InvalidMagic,
351    IncompleteFrame { expected: usize, got: usize },
352    ChecksumMismatch { expected: u32, actual: u32 },
353}
354
355impl fmt::Display for PagedDwbFrameError {
356    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
357        match self {
358            Self::ShortHeader { got } => write!(f, "DWB frame too short: got {got} bytes"),
359            Self::InvalidMagic => write!(f, "invalid DWB frame magic"),
360            Self::IncompleteFrame { expected, got } => {
361                write!(
362                    f,
363                    "incomplete DWB frame: expected {expected} bytes, got {got}"
364                )
365            }
366            Self::ChecksumMismatch { expected, actual } => write!(
367                f,
368                "DWB frame checksum mismatch: expected {expected}, actual {actual}"
369            ),
370        }
371    }
372}
373
374impl std::error::Error for PagedDwbFrameError {}
375
376pub fn encode_paged_dwb_frame<'a>(
377    pages: impl IntoIterator<Item = (u32, &'a [u8; PAGED_PAGE_SIZE])>,
378) -> Vec<u8> {
379    let pages: Vec<_> = pages.into_iter().collect();
380    let total = PAGED_DWB_HEADER_SIZE + pages.len() * PAGED_DWB_ENTRY_SIZE;
381    let mut out = Vec::with_capacity(total);
382
383    out.extend_from_slice(&DWB_MAGIC);
384    out.extend_from_slice(&(pages.len() as u32).to_le_bytes());
385    out.extend_from_slice(&[0u8; 4]);
386
387    for (page_id, page) in pages {
388        out.extend_from_slice(&page_id.to_le_bytes());
389        out.extend_from_slice(page);
390    }
391
392    let checksum = crc32(&out[PAGED_DWB_HEADER_SIZE..]);
393    out[8..12].copy_from_slice(&checksum.to_le_bytes());
394    out
395}
396
397pub fn decode_paged_dwb_frame(bytes: &[u8]) -> Result<Vec<PagedDwbEntry>, PagedDwbFrameError> {
398    if bytes.len() < PAGED_DWB_HEADER_SIZE {
399        return Err(PagedDwbFrameError::ShortHeader { got: bytes.len() });
400    }
401    if bytes[0..4] != DWB_MAGIC {
402        return Err(PagedDwbFrameError::InvalidMagic);
403    }
404
405    let count = u32::from_le_bytes(bytes[4..8].try_into().expect("len checked")) as usize;
406    let stored_checksum = u32::from_le_bytes(bytes[8..12].try_into().expect("len checked"));
407    let expected_len = PAGED_DWB_HEADER_SIZE + count * PAGED_DWB_ENTRY_SIZE;
408    if bytes.len() < expected_len {
409        return Err(PagedDwbFrameError::IncompleteFrame {
410            expected: expected_len,
411            got: bytes.len(),
412        });
413    }
414
415    let actual_checksum = crc32(&bytes[PAGED_DWB_HEADER_SIZE..expected_len]);
416    if actual_checksum != stored_checksum {
417        return Err(PagedDwbFrameError::ChecksumMismatch {
418            expected: stored_checksum,
419            actual: actual_checksum,
420        });
421    }
422
423    let mut offset = PAGED_DWB_HEADER_SIZE;
424    let mut entries = Vec::with_capacity(count);
425    for _ in 0..count {
426        let page_id =
427            u32::from_le_bytes(bytes[offset..offset + 4].try_into().expect("len checked"));
428        offset += PAGED_DWB_ENTRY_HEADER_SIZE;
429        let mut page = [0u8; PAGED_PAGE_SIZE];
430        page.copy_from_slice(&bytes[offset..offset + PAGED_PAGE_SIZE]);
431        offset += PAGED_PAGE_SIZE;
432        entries.push(PagedDwbEntry { page_id, page });
433    }
434    Ok(entries)
435}
436
437fn crc32(data: &[u8]) -> u32 {
438    crc32fast::hash(data)
439}
440
441/// `b"RDCC"` — RedDB Columnar Chunk. Opens and closes every column block.
442pub const COLUMN_BLOCK_MAGIC: [u8; 4] = *b"RDCC";
443
444/// Column block on-disk format version.
445pub const COLUMN_BLOCK_VERSION_V1: u16 = 1;
446
447/// Reusable segment-level bloom header magic.
448pub const BLOOM_SEGMENT_MAGIC: u8 = 0xBF;
449
450/// Legacy vector B-tree on-disk page format.
451pub const VECTOR_BTREE_FORMAT_VERSION_V1: u16 = 1;
452
453/// Current vector B-tree on-disk page format.
454pub const VECTOR_BTREE_FORMAT_VERSION_V2: u16 = 2;
455
456/// Vector B-tree format stamped into freshly-written page headers.
457pub const VECTOR_BTREE_FORMAT_VERSION: u16 = VECTOR_BTREE_FORMAT_VERSION_V2;
458
459/// Database file header information persisted in page 0 after the page header.
460#[derive(Debug, Clone, PartialEq, Eq)]
461pub struct DatabaseHeader {
462    pub version: u32,
463    pub page_size: u32,
464    pub page_count: u32,
465    pub freelist_head: u32,
466    pub schema_version: u32,
467    pub checkpoint_lsn: u64,
468    pub checkpoint_in_progress: bool,
469    pub checkpoint_target_lsn: u64,
470    pub physical: PhysicalFileHeader,
471}
472
473impl Default for DatabaseHeader {
474    fn default() -> Self {
475        Self {
476            version: PAGE_FILE_VERSION,
477            page_size: PAGED_PAGE_SIZE as u32,
478            page_count: 1,
479            freelist_head: 0,
480            schema_version: 0,
481            checkpoint_lsn: 0,
482            checkpoint_in_progress: false,
483            checkpoint_target_lsn: 0,
484            physical: PhysicalFileHeader::default(),
485        }
486    }
487}
488
489/// Minimal physical state mirrored into page 0 for paged databases.
490///
491/// The pager owns when this is read and written, but the field layout is a
492/// durable file compatibility contract.
493#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
494pub struct PhysicalFileHeader {
495    pub format_version: u32,
496    pub sequence: u64,
497    pub manifest_oldest_root: u64,
498    pub manifest_root: u64,
499    pub free_set_root: u64,
500    pub manifest_page: u32,
501    pub manifest_checksum: u64,
502    pub collection_roots_page: u32,
503    pub collection_roots_checksum: u64,
504    pub collection_root_count: u32,
505    pub snapshot_count: u32,
506    pub index_count: u32,
507    pub catalog_collection_count: u32,
508    pub catalog_total_entities: u64,
509    pub export_count: u32,
510    pub graph_projection_count: u32,
511    pub analytics_job_count: u32,
512    pub manifest_event_count: u32,
513    pub registry_page: u32,
514    pub registry_checksum: u64,
515    pub recovery_page: u32,
516    pub recovery_checksum: u64,
517    pub catalog_page: u32,
518    pub catalog_checksum: u64,
519    pub metadata_state_page: u32,
520    pub metadata_state_checksum: u64,
521    pub vector_artifact_page: u32,
522    pub vector_artifact_checksum: u64,
523}
524
525#[derive(Debug, Clone, PartialEq, Eq)]
526pub enum DatabaseHeaderError {
527    ShortPage { need: usize, got: usize },
528    InvalidMagic,
529    UnsupportedPageSize(u32),
530    UnsupportedDatabaseVersion { file_version: u32, supported: u32 },
531}
532
533impl fmt::Display for DatabaseHeaderError {
534    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
535        match self {
536            Self::ShortPage { need, got } => {
537                write!(f, "database header page too short: need {need} bytes, got {got}")
538            }
539            Self::InvalidMagic => write!(f, "invalid database header magic"),
540            Self::UnsupportedPageSize(size) => write!(f, "Unsupported page size: {size}"),
541            Self::UnsupportedDatabaseVersion {
542                file_version,
543                supported,
544            } => write!(
545                f,
546                "Unsupported database version: file version {file_version} is newer than supported {supported}"
547            ),
548        }
549    }
550}
551
552impl std::error::Error for DatabaseHeaderError {}
553
554pub fn database_header_magic_matches(page: &[u8]) -> bool {
555    page.get(DB_MAGIC_OFFSET..DB_MAGIC_OFFSET + PAGE_FILE_MAGIC.len()) == Some(&PAGE_FILE_MAGIC)
556}
557
558pub fn init_database_header_page(
559    page: &mut [u8],
560    page_count: u32,
561) -> Result<(), DatabaseHeaderError> {
562    encode_database_header(
563        page,
564        &DatabaseHeader {
565            page_count,
566            ..DatabaseHeader::default()
567        },
568    )
569}
570
571pub fn database_header_page_count(page: &[u8]) -> Result<u32, DatabaseHeaderError> {
572    read_u32(page, DB_PAGE_COUNT_OFFSET)
573}
574
575pub fn set_database_header_version(
576    page: &mut [u8],
577    version: u32,
578) -> Result<(), DatabaseHeaderError> {
579    write_u32(page, DB_VERSION_OFFSET, version)
580}
581
582pub fn set_database_header_page_count(
583    page: &mut [u8],
584    page_count: u32,
585) -> Result<(), DatabaseHeaderError> {
586    write_u32(page, DB_PAGE_COUNT_OFFSET, page_count)
587}
588
589pub fn database_header_freelist_head(page: &[u8]) -> Result<u32, DatabaseHeaderError> {
590    read_u32(page, DB_FREELIST_HEAD_OFFSET)
591}
592
593pub fn set_database_header_freelist_head(
594    page: &mut [u8],
595    page_id: u32,
596) -> Result<(), DatabaseHeaderError> {
597    write_u32(page, DB_FREELIST_HEAD_OFFSET, page_id)
598}
599
600pub fn database_header_page_size(page: &[u8]) -> Result<u32, DatabaseHeaderError> {
601    read_u32(page, DB_PAGE_SIZE_OFFSET)
602}
603
604pub fn decode_database_header(page: &[u8]) -> Result<DatabaseHeader, DatabaseHeaderError> {
605    ensure_len(page, DB_HEADER_MIN_LEN)?;
606    if !database_header_magic_matches(page) {
607        return Err(DatabaseHeaderError::InvalidMagic);
608    }
609
610    let version = read_u32(page, DB_VERSION_OFFSET)?;
611    let page_size = read_u32(page, DB_PAGE_SIZE_OFFSET)?;
612    if page_size != PAGED_PAGE_SIZE as u32 {
613        return Err(DatabaseHeaderError::UnsupportedPageSize(page_size));
614    }
615    if version > PAGE_FILE_VERSION {
616        return Err(DatabaseHeaderError::UnsupportedDatabaseVersion {
617            file_version: version,
618            supported: PAGE_FILE_VERSION,
619        });
620    }
621
622    Ok(DatabaseHeader {
623        version,
624        page_size,
625        page_count: read_u32(page, DB_PAGE_COUNT_OFFSET)?,
626        freelist_head: read_u32(page, DB_FREELIST_HEAD_OFFSET)?,
627        schema_version: read_u32(page, DB_SCHEMA_VERSION_OFFSET)?,
628        checkpoint_lsn: read_u64(page, DB_CHECKPOINT_LSN_OFFSET)?,
629        checkpoint_in_progress: page[DB_CHECKPOINT_IN_PROGRESS_OFFSET] != 0,
630        checkpoint_target_lsn: read_u64(page, DB_CHECKPOINT_TARGET_LSN_OFFSET)?,
631        physical: PhysicalFileHeader {
632            format_version: read_u32(page, DB_PHYSICAL_FORMAT_VERSION_OFFSET)?,
633            sequence: read_u64(page, DB_PHYSICAL_SEQUENCE_OFFSET)?,
634            manifest_oldest_root: read_u64(page, DB_MANIFEST_OLDEST_ROOT_OFFSET)?,
635            manifest_root: read_u64(page, DB_MANIFEST_ROOT_OFFSET)?,
636            free_set_root: read_u64(page, DB_FREE_SET_ROOT_OFFSET)?,
637            manifest_page: read_u32(page, DB_MANIFEST_PAGE_OFFSET)?,
638            manifest_checksum: read_u64(page, DB_MANIFEST_CHECKSUM_OFFSET)?,
639            collection_roots_page: read_u32(page, DB_COLLECTION_ROOTS_PAGE_OFFSET)?,
640            collection_roots_checksum: read_u64(page, DB_COLLECTION_ROOTS_CHECKSUM_OFFSET)?,
641            collection_root_count: read_u32(page, DB_COLLECTION_ROOT_COUNT_OFFSET)?,
642            snapshot_count: read_u32(page, DB_SNAPSHOT_COUNT_OFFSET)?,
643            index_count: read_u32(page, DB_INDEX_COUNT_OFFSET)?,
644            catalog_collection_count: read_u32(page, DB_CATALOG_COLLECTION_COUNT_OFFSET)?,
645            catalog_total_entities: read_u64(page, DB_CATALOG_TOTAL_ENTITIES_OFFSET)?,
646            export_count: read_u32(page, DB_EXPORT_COUNT_OFFSET)?,
647            graph_projection_count: read_u32(page, DB_GRAPH_PROJECTION_COUNT_OFFSET)?,
648            analytics_job_count: read_u32(page, DB_ANALYTICS_JOB_COUNT_OFFSET)?,
649            manifest_event_count: read_u32(page, DB_MANIFEST_EVENT_COUNT_OFFSET)?,
650            registry_page: read_u32(page, DB_REGISTRY_PAGE_OFFSET)?,
651            registry_checksum: read_u64(page, DB_REGISTRY_CHECKSUM_OFFSET)?,
652            recovery_page: read_u32(page, DB_RECOVERY_PAGE_OFFSET)?,
653            recovery_checksum: read_u64(page, DB_RECOVERY_CHECKSUM_OFFSET)?,
654            catalog_page: read_u32(page, DB_CATALOG_PAGE_OFFSET)?,
655            catalog_checksum: read_u64(page, DB_CATALOG_CHECKSUM_OFFSET)?,
656            metadata_state_page: read_u32(page, DB_METADATA_STATE_PAGE_OFFSET)?,
657            metadata_state_checksum: read_u64(page, DB_METADATA_STATE_CHECKSUM_OFFSET)?,
658            vector_artifact_page: read_u32(page, DB_VECTOR_ARTIFACT_PAGE_OFFSET)?,
659            vector_artifact_checksum: read_u64(page, DB_VECTOR_ARTIFACT_CHECKSUM_OFFSET)?,
660        },
661    })
662}
663
664pub fn encode_database_header(
665    page: &mut [u8],
666    header: &DatabaseHeader,
667) -> Result<(), DatabaseHeaderError> {
668    ensure_len(page, DB_HEADER_MIN_LEN)?;
669    page[DB_MAGIC_OFFSET..DB_MAGIC_OFFSET + PAGE_FILE_MAGIC.len()]
670        .copy_from_slice(&PAGE_FILE_MAGIC);
671    write_u32(page, DB_VERSION_OFFSET, header.version)?;
672    write_u32(page, DB_PAGE_SIZE_OFFSET, header.page_size)?;
673    write_u32(page, DB_PAGE_COUNT_OFFSET, header.page_count)?;
674    write_u32(page, DB_FREELIST_HEAD_OFFSET, header.freelist_head)?;
675    write_u32(page, DB_SCHEMA_VERSION_OFFSET, header.schema_version)?;
676    write_u64(page, DB_CHECKPOINT_LSN_OFFSET, header.checkpoint_lsn)?;
677    write_u32(
678        page,
679        DB_PHYSICAL_FORMAT_VERSION_OFFSET,
680        header.physical.format_version,
681    )?;
682    write_u64(page, DB_PHYSICAL_SEQUENCE_OFFSET, header.physical.sequence)?;
683    write_u64(page, DB_MANIFEST_ROOT_OFFSET, header.physical.manifest_root)?;
684    write_u64(
685        page,
686        DB_MANIFEST_OLDEST_ROOT_OFFSET,
687        header.physical.manifest_oldest_root,
688    )?;
689    write_u64(page, DB_FREE_SET_ROOT_OFFSET, header.physical.free_set_root)?;
690    write_u32(page, DB_MANIFEST_PAGE_OFFSET, header.physical.manifest_page)?;
691    write_u64(
692        page,
693        DB_MANIFEST_CHECKSUM_OFFSET,
694        header.physical.manifest_checksum,
695    )?;
696    write_u32(
697        page,
698        DB_COLLECTION_ROOTS_PAGE_OFFSET,
699        header.physical.collection_roots_page,
700    )?;
701    write_u64(
702        page,
703        DB_COLLECTION_ROOTS_CHECKSUM_OFFSET,
704        header.physical.collection_roots_checksum,
705    )?;
706    write_u32(
707        page,
708        DB_COLLECTION_ROOT_COUNT_OFFSET,
709        header.physical.collection_root_count,
710    )?;
711    write_u32(
712        page,
713        DB_SNAPSHOT_COUNT_OFFSET,
714        header.physical.snapshot_count,
715    )?;
716    write_u32(page, DB_INDEX_COUNT_OFFSET, header.physical.index_count)?;
717    write_u32(
718        page,
719        DB_CATALOG_COLLECTION_COUNT_OFFSET,
720        header.physical.catalog_collection_count,
721    )?;
722    write_u64(
723        page,
724        DB_CATALOG_TOTAL_ENTITIES_OFFSET,
725        header.physical.catalog_total_entities,
726    )?;
727    write_u32(page, DB_EXPORT_COUNT_OFFSET, header.physical.export_count)?;
728    write_u32(
729        page,
730        DB_GRAPH_PROJECTION_COUNT_OFFSET,
731        header.physical.graph_projection_count,
732    )?;
733    write_u32(
734        page,
735        DB_ANALYTICS_JOB_COUNT_OFFSET,
736        header.physical.analytics_job_count,
737    )?;
738    write_u32(
739        page,
740        DB_MANIFEST_EVENT_COUNT_OFFSET,
741        header.physical.manifest_event_count,
742    )?;
743    write_u32(page, DB_REGISTRY_PAGE_OFFSET, header.physical.registry_page)?;
744    write_u64(
745        page,
746        DB_REGISTRY_CHECKSUM_OFFSET,
747        header.physical.registry_checksum,
748    )?;
749    write_u32(page, DB_RECOVERY_PAGE_OFFSET, header.physical.recovery_page)?;
750    write_u64(
751        page,
752        DB_RECOVERY_CHECKSUM_OFFSET,
753        header.physical.recovery_checksum,
754    )?;
755    write_u32(page, DB_CATALOG_PAGE_OFFSET, header.physical.catalog_page)?;
756    write_u64(
757        page,
758        DB_CATALOG_CHECKSUM_OFFSET,
759        header.physical.catalog_checksum,
760    )?;
761    write_u32(
762        page,
763        DB_METADATA_STATE_PAGE_OFFSET,
764        header.physical.metadata_state_page,
765    )?;
766    write_u64(
767        page,
768        DB_METADATA_STATE_CHECKSUM_OFFSET,
769        header.physical.metadata_state_checksum,
770    )?;
771    write_u32(
772        page,
773        DB_VECTOR_ARTIFACT_PAGE_OFFSET,
774        header.physical.vector_artifact_page,
775    )?;
776    write_u64(
777        page,
778        DB_VECTOR_ARTIFACT_CHECKSUM_OFFSET,
779        header.physical.vector_artifact_checksum,
780    )?;
781    page[DB_CHECKPOINT_IN_PROGRESS_OFFSET] = if header.checkpoint_in_progress { 1 } else { 0 };
782    write_u64(
783        page,
784        DB_CHECKPOINT_TARGET_LSN_OFFSET,
785        header.checkpoint_target_lsn,
786    )
787}
788
789fn ensure_len(bytes: &[u8], need: usize) -> Result<(), DatabaseHeaderError> {
790    if bytes.len() < need {
791        return Err(DatabaseHeaderError::ShortPage {
792            need,
793            got: bytes.len(),
794        });
795    }
796    Ok(())
797}
798
799fn read_u32(bytes: &[u8], offset: usize) -> Result<u32, DatabaseHeaderError> {
800    ensure_len(bytes, offset + 4)?;
801    Ok(u32::from_le_bytes(
802        bytes[offset..offset + 4].try_into().expect("len checked"),
803    ))
804}
805
806fn read_u16(bytes: &[u8], offset: usize) -> Result<u16, DatabaseHeaderError> {
807    ensure_len(bytes, offset + 2)?;
808    Ok(u16::from_le_bytes(
809        bytes[offset..offset + 2].try_into().expect("len checked"),
810    ))
811}
812
813fn read_u64(bytes: &[u8], offset: usize) -> Result<u64, DatabaseHeaderError> {
814    ensure_len(bytes, offset + 8)?;
815    Ok(u64::from_le_bytes(
816        bytes[offset..offset + 8].try_into().expect("len checked"),
817    ))
818}
819
820fn write_u32(bytes: &mut [u8], offset: usize, value: u32) -> Result<(), DatabaseHeaderError> {
821    ensure_len(bytes, offset + 4)?;
822    bytes[offset..offset + 4].copy_from_slice(&value.to_le_bytes());
823    Ok(())
824}
825
826fn write_u16(bytes: &mut [u8], offset: usize, value: u16) -> Result<(), DatabaseHeaderError> {
827    ensure_len(bytes, offset + 2)?;
828    bytes[offset..offset + 2].copy_from_slice(&value.to_le_bytes());
829    Ok(())
830}
831
832fn write_u64(bytes: &mut [u8], offset: usize, value: u64) -> Result<(), DatabaseHeaderError> {
833    ensure_len(bytes, offset + 8)?;
834    bytes[offset..offset + 8].copy_from_slice(&value.to_le_bytes());
835    Ok(())
836}
837
838#[cfg(test)]
839mod tests {
840    use super::*;
841
842    #[test]
843    fn paged_page_header_round_trips_raw_layout() {
844        let header = PagedPageHeader {
845            page_type: 13,
846            flags: 0b1010_0001,
847            cell_count: 42,
848            free_start: 44,
849            free_end: 16_000,
850            page_id: 99,
851            parent_id: 88,
852            right_child: 77,
853            lsn: 66,
854            checksum: 55,
855        };
856
857        let encoded = encode_paged_page_header(&header);
858
859        assert_eq!(encoded.len(), PAGED_PAGE_HEADER_SIZE);
860        assert_eq!(decode_paged_page_header(&encoded), header);
861    }
862
863    #[test]
864    fn paged_page_header_accessors_update_expected_offsets() {
865        let mut page = [0u8; PAGED_PAGE_SIZE];
866        page[0] = 13;
867
868        set_paged_page_cell_count(&mut page, 2);
869        set_paged_page_free_start(&mut page, 34);
870        set_paged_page_free_end(&mut page, 4096);
871        set_paged_page_parent_id(&mut page, 7);
872        set_paged_page_right_child(&mut page, 8);
873        set_paged_page_lsn(&mut page, 9);
874        set_paged_page_checksum(&mut page, 10);
875
876        assert_eq!(paged_page_type(&page), 13);
877        assert_eq!(paged_page_cell_count(&page), 2);
878        assert_eq!(paged_page_free_start(&page), 34);
879        assert_eq!(paged_page_free_end(&page), 4096);
880        assert_eq!(paged_page_parent_id(&page), 7);
881        assert_eq!(paged_page_right_child(&page), 8);
882        assert_eq!(paged_page_lsn(&page), 9);
883        assert_eq!(paged_page_checksum(&page), 10);
884        clear_paged_page_checksum(&mut page);
885        assert_eq!(paged_page_checksum(&page), 0);
886    }
887
888    #[test]
889    fn paged_cell_helpers_round_trip_pointer_and_payload() {
890        let mut page = [0u8; PAGED_PAGE_SIZE];
891        let pointer = (PAGED_PAGE_SIZE - 14) as u16;
892
893        assert!(write_paged_cell(&mut page, pointer, b"key", b"value"));
894        assert!(set_paged_cell_pointer(&mut page, 0, pointer));
895
896        assert_eq!(paged_cell_pointer(&page, 0), Some(pointer));
897        let cell = paged_cell_bytes(&page, pointer).expect("cell bytes");
898        let (key, value) = paged_cell_key_value(cell).expect("key value");
899
900        assert_eq!(key, b"key");
901        assert_eq!(value, b"value");
902        assert_eq!(paged_cell_total_len(&page, pointer), Some(14));
903    }
904
905    #[test]
906    fn database_header_field_helpers_preserve_page_zero_offsets() {
907        let mut page = vec![0u8; PAGED_PAGE_SIZE];
908
909        init_database_header_page(&mut page, 7).expect("init database header page");
910        set_database_header_version(&mut page, PAGE_FILE_VERSION + 1).expect("set version");
911        set_database_header_page_count(&mut page, 9).expect("set page count");
912        set_database_header_freelist_head(&mut page, 4).expect("set freelist head");
913
914        assert!(database_header_magic_matches(&page));
915        assert!(matches!(
916            decode_database_header(&page),
917            Err(DatabaseHeaderError::UnsupportedDatabaseVersion { .. })
918        ));
919        set_database_header_version(&mut page, PAGE_FILE_VERSION).expect("restore version");
920        assert_eq!(
921            database_header_page_size(&page).unwrap(),
922            PAGED_PAGE_SIZE as u32
923        );
924        assert_eq!(database_header_page_count(&page).unwrap(), 9);
925        assert_eq!(database_header_freelist_head(&page).unwrap(), 4);
926    }
927
928    #[test]
929    fn paged_dwb_frame_round_trips_entries_and_validates_checksum() {
930        let mut page = [0u8; PAGED_PAGE_SIZE];
931        page[0] = 7;
932        let frame = encode_paged_dwb_frame([(42, &page)]);
933
934        let entries = decode_paged_dwb_frame(&frame).expect("decode DWB frame");
935        assert_eq!(entries.len(), 1);
936        assert_eq!(entries[0].page_id, 42);
937        assert_eq!(entries[0].page, page);
938
939        let mut corrupted = frame;
940        let last = corrupted.len() - 1;
941        corrupted[last] ^= 0xFF;
942        assert!(matches!(
943            decode_paged_dwb_frame(&corrupted),
944            Err(PagedDwbFrameError::ChecksumMismatch { .. })
945        ));
946    }
947
948    #[test]
949    fn paged_encryption_header_round_trips_marker_and_raw_bytes() {
950        let header = PagedEncryptionHeader {
951            salt: [7u8; PAGED_ENCRYPTION_SALT_SIZE],
952            key_check: vec![9u8; PAGED_ENCRYPTION_KEY_CHECK_BLOB_SIZE],
953        };
954        let mut page = vec![0u8; PAGED_PAGE_SIZE];
955        let bytes = encode_paged_encryption_header(&header);
956
957        write_paged_encryption_marker_and_header(&mut page, &bytes)
958            .expect("write encryption marker");
959
960        assert!(paged_encryption_marker_present(&page));
961        let raw = paged_encryption_header_bytes(&page).expect("header bytes");
962        assert_eq!(decode_paged_encryption_header(raw).unwrap(), header);
963    }
964
965    #[test]
966    fn database_header_round_trips_page_zero_contract() {
967        let header = DatabaseHeader {
968            version: PAGE_FILE_VERSION,
969            page_size: PAGED_PAGE_SIZE as u32,
970            page_count: 42,
971            freelist_head: 7,
972            schema_version: 3,
973            checkpoint_lsn: 99,
974            checkpoint_in_progress: true,
975            checkpoint_target_lsn: 123,
976            physical: PhysicalFileHeader {
977                format_version: 11,
978                sequence: 12,
979                manifest_oldest_root: 13,
980                manifest_root: 14,
981                free_set_root: 15,
982                manifest_page: 16,
983                manifest_checksum: 17,
984                collection_roots_page: 18,
985                collection_roots_checksum: 19,
986                collection_root_count: 20,
987                snapshot_count: 21,
988                index_count: 22,
989                catalog_collection_count: 23,
990                catalog_total_entities: 24,
991                export_count: 25,
992                graph_projection_count: 26,
993                analytics_job_count: 27,
994                manifest_event_count: 28,
995                registry_page: 29,
996                registry_checksum: 30,
997                recovery_page: 31,
998                recovery_checksum: 32,
999                catalog_page: 33,
1000                catalog_checksum: 34,
1001                metadata_state_page: 35,
1002                metadata_state_checksum: 36,
1003                vector_artifact_page: 37,
1004                vector_artifact_checksum: 38,
1005            },
1006        };
1007        let mut page = vec![0u8; PAGED_PAGE_SIZE];
1008
1009        encode_database_header(&mut page, &header).expect("encode header");
1010
1011        assert!(database_header_magic_matches(&page));
1012        let decoded = decode_database_header(&page).expect("decode header");
1013        assert_eq!(decoded, header);
1014    }
1015
1016    #[test]
1017    fn database_header_rejects_newer_versions() {
1018        let mut page = vec![0u8; PAGED_PAGE_SIZE];
1019        let header = DatabaseHeader {
1020            version: PAGE_FILE_VERSION + 1,
1021            ..DatabaseHeader::default()
1022        };
1023        encode_database_header(&mut page, &header).expect("encode header");
1024
1025        assert!(matches!(
1026            decode_database_header(&page),
1027            Err(DatabaseHeaderError::UnsupportedDatabaseVersion { .. })
1028        ));
1029    }
1030}
reddb_file/file_format.rs

reddb_file/
file_format.rs
