Skip to main content

mentedb_storage/
page.rs

1//! Page Manager: file-backed page storage with free list management.
2//!
3//! Pages are 16KB fixed-size blocks used as the fundamental I/O unit.
4//! The page file layout:
5//! - Page 0: file header (magic, version, page count, free list head)
6//! - Page 1..N: data pages
7
8use std::fs::{File, OpenOptions};
9use std::io::{Read, Seek, SeekFrom, Write};
10use std::path::Path;
11
12use mentedb_core::error::{MenteError, MenteResult};
13use tracing::{debug, info, trace};
14
15/// Page size: 64KB to accommodate large embedding vectors and long content.
16pub const PAGE_SIZE: usize = 64 * 1024;
17
18/// Magic number identifying a MenteDB page file ("MENTEDB1").
19const MAGIC: u64 = 0x4D454E_5445444231;
20
21/// File format version.
22const VERSION: u32 = 1;
23
24/// A page identifier.
25#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
26pub struct PageId(pub u64);
27
28/// Page type classification.
29#[derive(Debug, Clone, Copy, PartialEq, Eq)]
30#[repr(u8)]
31pub enum PageType {
32    Free = 0,
33    Data = 1,
34    Index = 2,
35    Overflow = 3,
36}
37
38impl From<u8> for PageType {
39    fn from(v: u8) -> Self {
40        match v {
41            1 => PageType::Data,
42            2 => PageType::Index,
43            3 => PageType::Overflow,
44            _ => PageType::Free,
45        }
46    }
47}
48
49/// Fixed-layout page header stored at the start of every page.
50#[repr(C)]
51#[derive(Debug, Clone, Copy)]
52pub struct PageHeader {
53    /// Which page this is.
54    pub page_id: u64,
55    /// Log Sequence Number: links to the WAL.
56    pub lsn: u64,
57    /// CRC-32 checksum of the page content.
58    pub checksum: u32,
59    /// Remaining free bytes in the data portion.
60    pub free_space: u16,
61    /// Number of occupied slots.
62    pub num_slots: u16,
63    /// Page type tag.
64    pub page_type: u8,
65    /// Padding for stable layout.
66    pub _padding: [u8; 7],
67}
68
69/// Size of the page header in bytes.
70pub const HEADER_SIZE: usize = std::mem::size_of::<PageHeader>();
71
72/// Usable data bytes per page (total page size minus header).
73pub const PAGE_DATA_SIZE: usize = PAGE_SIZE - HEADER_SIZE;
74
75/// A fixed-size, 4KB-aligned page.
76#[repr(C, align(4096))]
77pub struct Page {
78    /// The page header containing metadata (page ID, type, checksum, etc.).
79    pub header: PageHeader,
80    /// Raw page data payload.
81    pub data: [u8; PAGE_DATA_SIZE],
82}
83
84impl std::fmt::Debug for Page {
85    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
86        f.debug_struct("Page")
87            .field("header", &self.header)
88            .field("data_len", &self.data.len())
89            .finish()
90    }
91}
92
93impl Clone for Page {
94    fn clone(&self) -> Self {
95        let mut new_page = Page::zeroed();
96        new_page.header = self.header;
97        new_page.data.copy_from_slice(&self.data);
98        new_page
99    }
100}
101
102impl Page {
103    /// Create a zero-initialized page.
104    pub fn zeroed() -> Self {
105        // Safety: Page is #[repr(C)] with only primitive types; all-zeros is valid.
106        unsafe { std::mem::zeroed() }
107    }
108
109    /// View the raw bytes of this page.
110    fn as_bytes(&self) -> &[u8; PAGE_SIZE] {
111        // Safety: Page is #[repr(C)] with size == PAGE_SIZE.
112        unsafe { &*(self as *const Page as *const [u8; PAGE_SIZE]) }
113    }
114
115    /// Reconstruct a Page from a raw byte buffer.
116    fn from_bytes(bytes: &[u8; PAGE_SIZE]) -> Self {
117        // Safety: Page is #[repr(C)]; any bit pattern in data is valid.
118        unsafe { std::ptr::read(bytes.as_ptr() as *const Page) }
119    }
120
121    /// Compute CRC-32 checksum over header fields (excluding `checksum`) and data.
122    pub fn compute_checksum(&self) -> u32 {
123        let mut h = crc32fast::Hasher::new();
124        h.update(&self.header.page_id.to_le_bytes());
125        h.update(&self.header.lsn.to_le_bytes());
126        h.update(&self.header.free_space.to_le_bytes());
127        h.update(&self.header.num_slots.to_le_bytes());
128        h.update(&[self.header.page_type]);
129        h.update(&self.data);
130        h.finalize()
131    }
132}
133
134/// On-disk file header occupying the first bytes of page 0.
135#[repr(C)]
136struct FileHeader {
137    magic: u64,
138    version: u32,
139    _pad: u32,
140    page_count: u64,
141    free_list_head: u64,
142}
143
144/// Manages a file-backed page store with free-list allocation.
145pub struct PageManager {
146    file: File,
147    page_count: u64,
148    free_list_head: u64,
149}
150
151impl PageManager {
152    /// Open (or create) a page file at `dir_path/pages.db`.
153    pub fn open(dir_path: &Path) -> MenteResult<Self> {
154        let file_path = dir_path.join("pages.db");
155        let exists = file_path.exists()
156            && std::fs::metadata(&file_path)
157                .map(|m| m.len() > 0)
158                .unwrap_or(false);
159
160        let mut file = OpenOptions::new()
161            .read(true)
162            .write(true)
163            .create(true)
164            .truncate(false)
165            .open(&file_path)?;
166
167        if exists {
168            let mut buf = [0u8; std::mem::size_of::<FileHeader>()];
169            file.seek(SeekFrom::Start(0))?;
170            file.read_exact(&mut buf)?;
171            let hdr: FileHeader = unsafe { std::ptr::read(buf.as_ptr() as *const FileHeader) };
172
173            if hdr.magic != MAGIC {
174                return Err(MenteError::Storage("invalid page file magic number".into()));
175            }
176            if hdr.version != VERSION {
177                return Err(MenteError::Storage(format!(
178                    "unsupported page file version: {}",
179                    hdr.version
180                )));
181            }
182
183            info!(page_count = hdr.page_count, "opened existing page file");
184            Ok(Self {
185                file,
186                page_count: hdr.page_count,
187                free_list_head: hdr.free_list_head,
188            })
189        } else {
190            let mut pm = Self {
191                file,
192                page_count: 1,
193                free_list_head: 0,
194            };
195            // Write header page (page 0) — reserves first PAGE_SIZE bytes.
196            let mut header_page = Page::zeroed();
197            header_page.header.page_id = 0;
198            pm.write_page_raw(PageId(0), &header_page)?;
199            pm.write_file_header()?;
200            info!("created new page file");
201            Ok(pm)
202        }
203    }
204
205    /// Persist the file header into the beginning of page 0.
206    fn write_file_header(&mut self) -> MenteResult<()> {
207        let hdr = FileHeader {
208            magic: MAGIC,
209            version: VERSION,
210            _pad: 0,
211            page_count: self.page_count,
212            free_list_head: self.free_list_head,
213        };
214        let bytes = unsafe {
215            std::slice::from_raw_parts(
216                &hdr as *const FileHeader as *const u8,
217                std::mem::size_of::<FileHeader>(),
218            )
219        };
220        self.file.seek(SeekFrom::Start(0))?;
221        self.file.write_all(bytes)?;
222        self.file.flush()?;
223        Ok(())
224    }
225
226    /// Allocate a new page, reusing from the free list when possible.
227    pub fn allocate_page(&mut self) -> MenteResult<PageId> {
228        if self.free_list_head != 0 {
229            let page_id = PageId(self.free_list_head);
230            let page = self.read_page(page_id)?;
231            let next_free = u64::from_le_bytes(page.data[..8].try_into().unwrap());
232            self.free_list_head = next_free;
233            self.write_file_header()?;
234            debug!(page_id = page_id.0, "allocated page from free list");
235            return Ok(page_id);
236        }
237
238        let page_id = PageId(self.page_count);
239        self.page_count += 1;
240
241        let mut page = Page::zeroed();
242        page.header.page_id = page_id.0;
243        page.header.page_type = PageType::Data as u8;
244        page.header.free_space = PAGE_DATA_SIZE as u16;
245        self.write_page_raw(page_id, &page)?;
246        self.write_file_header()?;
247
248        debug!(page_id = page_id.0, "allocated new page");
249        Ok(page_id)
250    }
251
252    /// Read a page from disk.
253    pub fn read_page(&mut self, page_id: PageId) -> MenteResult<Box<Page>> {
254        if page_id.0 >= self.page_count {
255            return Err(MenteError::Storage(format!(
256                "page {} out of range (count={})",
257                page_id.0, self.page_count
258            )));
259        }
260
261        let offset = page_id.0 * PAGE_SIZE as u64;
262        let mut buf = [0u8; PAGE_SIZE];
263        self.file.seek(SeekFrom::Start(offset))?;
264        self.file.read_exact(&mut buf)?;
265
266        trace!(page_id = page_id.0, "read page from disk");
267        Ok(Box::new(Page::from_bytes(&buf)))
268    }
269
270    /// Write a page to disk.
271    pub fn write_page(&mut self, page_id: PageId, page: &Page) -> MenteResult<()> {
272        self.write_page_raw(page_id, page)
273    }
274
275    fn write_page_raw(&mut self, page_id: PageId, page: &Page) -> MenteResult<()> {
276        let offset = page_id.0 * PAGE_SIZE as u64;
277        self.file.seek(SeekFrom::Start(offset))?;
278        self.file.write_all(page.as_bytes())?;
279        trace!(page_id = page_id.0, "wrote page to disk");
280        Ok(())
281    }
282
283    /// Return a page to the free list.
284    pub fn free_page(&mut self, page_id: PageId) -> MenteResult<()> {
285        let mut page = Page::zeroed();
286        page.header.page_id = page_id.0;
287        page.header.page_type = PageType::Free as u8;
288        // Store the current free list head as a forward pointer.
289        page.data[..8].copy_from_slice(&self.free_list_head.to_le_bytes());
290
291        self.write_page(page_id, &page)?;
292        self.free_list_head = page_id.0;
293        self.write_file_header()?;
294
295        debug!(page_id = page_id.0, "freed page");
296        Ok(())
297    }
298
299    /// Total number of pages (including the header page).
300    pub fn page_count(&self) -> u64 {
301        self.page_count
302    }
303
304    /// Sync page file data to disk.
305    pub fn sync(&mut self) -> MenteResult<()> {
306        self.file.sync_data()?;
307        Ok(())
308    }
309}
310
311#[cfg(test)]
312mod tests {
313    use super::*;
314
315    fn setup() -> (tempfile::TempDir, PageManager) {
316        let dir = tempfile::tempdir().unwrap();
317        let pm = PageManager::open(dir.path()).unwrap();
318        (dir, pm)
319    }
320
321    #[test]
322    fn test_allocate_and_read_write() {
323        let (_dir, mut pm) = setup();
324
325        let pid = pm.allocate_page().unwrap();
326        assert_eq!(pid.0, 1); // page 0 is file header
327
328        let mut page = Page::zeroed();
329        page.header.page_id = pid.0;
330        page.header.page_type = PageType::Data as u8;
331        page.data[0..5].copy_from_slice(b"hello");
332        pm.write_page(pid, &page).unwrap();
333
334        let loaded = pm.read_page(pid).unwrap();
335        assert_eq!(&loaded.data[0..5], b"hello");
336    }
337
338    #[test]
339    fn test_free_and_reuse() {
340        let (_dir, mut pm) = setup();
341
342        let p1 = pm.allocate_page().unwrap();
343        let p2 = pm.allocate_page().unwrap();
344        assert_eq!(p1.0, 1);
345        assert_eq!(p2.0, 2);
346
347        // Free p1, then allocate — should get p1 back.
348        pm.free_page(p1).unwrap();
349        let p3 = pm.allocate_page().unwrap();
350        assert_eq!(p3.0, p1.0);
351
352        // Next allocation should be brand new.
353        let p4 = pm.allocate_page().unwrap();
354        assert_eq!(p4.0, 3);
355    }
356
357    #[test]
358    fn test_multiple_free_reuse() {
359        let (_dir, mut pm) = setup();
360
361        let p1 = pm.allocate_page().unwrap();
362        let p2 = pm.allocate_page().unwrap();
363        let _p3 = pm.allocate_page().unwrap();
364
365        // Free in order p1, p2 — free list becomes p2 -> p1.
366        pm.free_page(p1).unwrap();
367        pm.free_page(p2).unwrap();
368
369        // Allocations should come back in LIFO order: p2, p1.
370        let a1 = pm.allocate_page().unwrap();
371        let a2 = pm.allocate_page().unwrap();
372        assert_eq!(a1.0, p2.0);
373        assert_eq!(a2.0, p1.0);
374
375        // p3 is still allocated; next new page is 4.
376        let a3 = pm.allocate_page().unwrap();
377        assert_eq!(a3.0, 4);
378    }
379
380    #[test]
381    fn test_reopen() {
382        let dir = tempfile::tempdir().unwrap();
383        let pid;
384        {
385            let mut pm = PageManager::open(dir.path()).unwrap();
386            pid = pm.allocate_page().unwrap();
387            let mut page = Page::zeroed();
388            page.header.page_id = pid.0;
389            page.data[0..4].copy_from_slice(b"test");
390            pm.write_page(pid, &page).unwrap();
391            pm.sync().unwrap();
392        }
393        {
394            let mut pm = PageManager::open(dir.path()).unwrap();
395            let page = pm.read_page(pid).unwrap();
396            assert_eq!(&page.data[0..4], b"test");
397        }
398    }
399
400    #[test]
401    fn test_out_of_range() {
402        let (_dir, mut pm) = setup();
403        assert!(pm.read_page(PageId(999)).is_err());
404    }
405
406    #[test]
407    fn test_checksum() {
408        let mut page = Page::zeroed();
409        page.header.page_id = 42;
410        page.data[0] = 0xFF;
411        let c1 = page.compute_checksum();
412        page.data[0] = 0x00;
413        let c2 = page.compute_checksum();
414        assert_ne!(c1, c2);
415    }
416}