29.1 KiB, 1 run
created by r2848102244:738, which is this file's identity for as long as the history lasts, whatever it is later renamed to
download · who wrote it · its history
| 1 | //! What the working copy looked like when it last matched the history. |
| 2 | //! |
| 3 | //! Every verb but `init` begins by capturing, and capturing means rendering the |
| 4 | //! whole history and then reading every tracked file off the disk to compare |
| 5 | //! against it. On a 44,628 operation repository the render alone is 100 MB of |
| 6 | //! the 169 MB an `ore log` costs, and `ore log` never looks at a file. Git does |
| 7 | //! not do this: it keeps each file's size and modification time beside its name |
| 8 | //! and skips anything those two say is untouched, which is why `git status` |
| 9 | //! answers in milliseconds. |
| 10 | //! |
| 11 | //! This is that index. When it says the working copy is exactly where it was, |
| 12 | //! the capture reads no file and renders nothing at all. |
| 13 | //! |
| 14 | //! # It is a guess about content, and that is the whole difficulty |
| 15 | //! |
| 16 | //! The verdict cache next door could gamble safely because a lost gamble was |
| 17 | //! caught afterwards by a digest of the bytes. **Nothing catches this one.** If |
| 18 | //! the index says unchanged and the file changed, the edit is not refused and |
| 19 | //! not reported -- it is gone, and the history says it never happened. So every |
| 20 | //! rule here is written to fail towards reading, and an index that is |
| 21 | //! occasionally slow is right where one that is occasionally wrong is not: |
| 22 | //! |
| 23 | //! - **A path the walk finds and the index does not know, or the index knows |
| 24 | //! and the walk cannot find, is a change.** No attempt is made to work out |
| 25 | //! what happened to it. |
| 26 | //! - **Four things must all agree**, not two: the size, the modification time, |
| 27 | //! the change time and the inode. `touch -r` sets the modification time and |
| 28 | //! cannot set the change time; a restore from a backup, a `cp -p` or a `tar |
| 29 | //! -x` writes a new inode. Git checks these too and turns them off on |
| 30 | //! platforms that cannot supply them; here they are simply required, because |
| 31 | //! the tool is built for one that can. |
| 32 | //! - **The history must not have moved.** The index names the frontier it was |
| 33 | //! taken at, because a working copy that has not changed still differs from a |
| 34 | //! render that has. |
| 35 | //! - **Anything read too recently to be sure about is read again.** See below. |
| 36 | //! - **It is only written after a capture that appended nothing**, which is the |
| 37 | //! one moment the working copy is known to be what the history says. A |
| 38 | //! capture that recorded something leaves no index, and the command after it |
| 39 | //! pays one full capture to establish a fresh one. |
| 40 | //! |
| 41 | //! # The racily clean file, and the number this rests on |
| 42 | //! |
| 43 | //! A file written in the same clock tick that its size and time were recorded |
| 44 | //! in cannot be told apart from one that was not, if the write left the size |
| 45 | //! alone. Git meets this by comparing each file's time against the index file's |
| 46 | //! own and re-reading anything not strictly older. |
| 47 | //! |
| 48 | //! The rule here is stricter and needs no comparison with a file: the clock is |
| 49 | //! read **before** any file is looked at, that reading is what the index |
| 50 | //! records as `taken`, and an entry is believed only where its modification |
| 51 | //! time is at least [`RACY`] older than it. Any write from the moment the |
| 52 | //! command began stamps a time at or after `taken`, so it cannot leave a |
| 53 | //! recorded time that satisfies that; and a write in the recorded time's own |
| 54 | //! tick is excluded because that tick is required to have ended before the |
| 55 | //! command started. |
| 56 | //! |
| 57 | //! [`RACY`] is two seconds, and it is **not** a measurement of the filesystem |
| 58 | //! this was written on. Measured there, on ext4: twenty thousand consecutive |
| 59 | //! writes to one file produced twenty thousand distinct modification times, the |
| 60 | //! smallest gap being 16.8 microseconds, and three thousand files created back |
| 61 | //! to back shared no time with each other. So the real figure here is under |
| 62 | //! twenty microseconds. Two seconds is chosen to be coarser than any filesystem |
| 63 | //! the tool is likely to meet -- FAT stamps to two seconds, HFS+ to one, and an |
| 64 | //! NFS server's clock may simply be behind this one's -- and what it costs is |
| 65 | //! that a file touched in the last two seconds is read rather than believed, |
| 66 | //! which is the file the reader wanted to read anyway. |
| 67 | //! |
| 68 | //! # It is replica-local, derived, and safe to delete |
| 69 | //! |
| 70 | //! It sits beside the log, it is never put in a segment or a sync message, and |
| 71 | //! losing it costs one full capture. Anything about it that does not parse is |
| 72 | //! treated as an absent index rather than as an error, which is the same |
| 73 | //! failure the absence itself is: read everything. |
| 74 | |
| 75 | use ore_store::keys::{ |
| 76 | bytes_of, |
| 77 | text_of, |
| 78 | }; |
| 79 | |
| 80 | use oxedyne_fe2o3_core::prelude::*; |
| 81 | use oxedyne_fe2o3_ore::id::OpId; |
| 82 | use oxedyne_fe2o3_ore::op::Op; |
| 83 | |
| 84 | use std::collections::BTreeMap; |
| 85 | use std::fs; |
| 86 | use std::os::unix::fs::MetadataExt; |
| 87 | use std::path::{ |
| 88 | Path, |
| 89 | PathBuf, |
| 90 | }; |
| 91 | use std::str::FromStr; |
| 92 | use std::time::{ |
| 93 | SystemTime, |
| 94 | UNIX_EPOCH, |
| 95 | }; |
| 96 | |
| 97 | |
| 98 | // The file, and what it says about itself |
| 99 | pub const STAT_FILE: &str = "stat"; |
| 100 | pub const STAT_FORMAT: &str = "ORESTAT 1"; |
| 101 | |
| 102 | /// How much older than the reading of the clock a file's modification time must |
| 103 | /// be before it is believed. See the module note: two seconds, deliberately far |
| 104 | /// coarser than the filesystem this was written on. |
| 105 | pub const RACY: i128 = 2_000_000_000; |
| 106 | |
| 107 | |
| 108 | /// What was seen at one path when it was last read. |
| 109 | #[derive(Clone, Copy, Debug, Eq, PartialEq)] |
| 110 | pub struct Seen { |
| 111 | len: u64, // bytes |
| 112 | mtime: i128, // nanoseconds since the epoch |
| 113 | ctime: i128, // which a write bumps and which `touch` cannot set |
| 114 | ino: u64, // which a restore, a copy or a checkout replaces |
| 115 | } |
| 116 | |
| 117 | impl Seen { |
| 118 | /// Reads what the filesystem says about a path, without following a link. |
| 119 | /// |
| 120 | /// `None` is anything that is not a regular file, a link included: the walk |
| 121 | /// records only what it would have read, and nothing else is content. |
| 122 | pub fn of(path: &Path) |
| 123 | -> Option<Self> |
| 124 | { |
| 125 | let meta = match fs::symlink_metadata(path) { |
| 126 | Ok(m) => m, |
| 127 | Err(_) => return None, |
| 128 | }; |
| 129 | if !meta.is_file() { |
| 130 | return None; |
| 131 | } |
| 132 | Some(Self { |
| 133 | len: meta.len(), |
| 134 | mtime: meta.mtime() as i128 * 1_000_000_000 + meta.mtime_nsec() as i128, |
| 135 | ctime: meta.ctime() as i128 * 1_000_000_000 + meta.ctime_nsec() as i128, |
| 136 | ino: meta.ino(), |
| 137 | }) |
| 138 | } |
| 139 | |
| 140 | /// Is this old enough to be believed, for an index taken at `taken`? |
| 141 | pub fn settled(&self, taken: i128) -> bool { |
| 142 | match self.mtime.checked_add(RACY) { |
| 143 | Some(after) => after <= taken, |
| 144 | None => false, |
| 145 | } |
| 146 | } |
| 147 | } |
| 148 | |
| 149 | |
| 150 | /// The working copy as it stood when it last matched the history. |
| 151 | #[derive(Clone, Debug, Default)] |
| 152 | pub struct Stats { |
| 153 | taken: i128, // the clock, read before a single file was looked at |
| 154 | at: Vec<OpId>, // the frontier this describes |
| 155 | seen: BTreeMap<Vec<u8>, Seen>, |
| 156 | } |
| 157 | |
| 158 | impl Stats { |
| 159 | |
| 160 | /// Begins an index, reading the clock now, before anything is looked at. |
| 161 | /// |
| 162 | /// The order is the whole of the racy rule: a reading taken afterwards would |
| 163 | /// be later than a write that raced the scan, and would believe it. |
| 164 | pub fn beginning() |
| 165 | -> Outcome<Self> |
| 166 | { |
| 167 | let now = res!(SystemTime::now().duration_since(UNIX_EPOCH)); |
| 168 | Ok(Self { |
| 169 | taken: now.as_nanos() as i128, |
| 170 | at: Vec::new(), |
| 171 | seen: BTreeMap::new(), |
| 172 | }) |
| 173 | } |
| 174 | |
| 175 | pub fn put(&mut self, path: &[u8], seen: Seen) { |
| 176 | self.seen.insert(path.to_vec(), seen); |
| 177 | } |
| 178 | |
| 179 | pub fn describes(&mut self, frontier: &[OpId]) { |
| 180 | self.at = frontier.to_vec(); |
| 181 | } |
| 182 | |
| 183 | pub fn path_of(dir: &Path) -> PathBuf { |
| 184 | dir.join(STAT_FILE) |
| 185 | } |
| 186 | |
| 187 | /// The index of the `.ore` directory `dir`, or nothing. |
| 188 | /// |
| 189 | /// Absent, unreadable, a format this build does not know, one malformed line |
| 190 | /// -- all the same answer, which is that nothing is known and every file will |
| 191 | /// be read. A partial read is refused rather than half believed. |
| 192 | pub fn read(dir: &Path) |
| 193 | -> Option<Self> |
| 194 | { |
| 195 | let text = match fs::read_to_string(Self::path_of(dir)) { |
| 196 | Ok(t) => t, |
| 197 | Err(_) => return None, |
| 198 | }; |
| 199 | let mut lines = text.lines(); |
| 200 | match lines.next() { |
| 201 | Some(first) if first.trim() == STAT_FORMAT => (), |
| 202 | _ => return None, |
| 203 | } |
| 204 | let mut out = Self::default(); |
| 205 | let mut said_taken = false; |
| 206 | for line in lines { |
| 207 | if line.trim().is_empty() { |
| 208 | continue; |
| 209 | } |
| 210 | let field: Vec<&str> = line.split_whitespace().collect(); |
| 211 | match field.first() { |
| 212 | Some(&"taken") => { |
| 213 | let n = match field.get(1).map(|s| s.parse::<i128>()) { |
| 214 | Some(Ok(n)) => n, |
| 215 | _ => return None, |
| 216 | }; |
| 217 | out.taken = n; |
| 218 | said_taken = true; |
| 219 | }, |
| 220 | Some(&"at") => { |
| 221 | for said in &field[1..] { |
| 222 | match OpId::from_str(said) { |
| 223 | Ok(id) => out.at.push(id), |
| 224 | Err(_) => return None, |
| 225 | } |
| 226 | } |
| 227 | }, |
| 228 | _ => { |
| 229 | if field.len() != 5 { |
| 230 | return None; |
| 231 | } |
| 232 | let len = match field[0].parse::<u64>() { |
| 233 | Ok(n) => n, |
| 234 | Err(_) => return None, |
| 235 | }; |
| 236 | let mtime = match field[1].parse::<i128>() { |
| 237 | Ok(n) => n, |
| 238 | Err(_) => return None, |
| 239 | }; |
| 240 | let ctime = match field[2].parse::<i128>() { |
| 241 | Ok(n) => n, |
| 242 | Err(_) => return None, |
| 243 | }; |
| 244 | let ino = match field[3].parse::<u64>() { |
| 245 | Ok(n) => n, |
| 246 | Err(_) => return None, |
| 247 | }; |
| 248 | let path = match bytes_of(field[4]) { |
| 249 | Ok(b) => b, |
| 250 | Err(_) => return None, |
| 251 | }; |
| 252 | out.seen.insert(path, Seen { len, mtime, ctime, ino }); |
| 253 | }, |
| 254 | } |
| 255 | } |
| 256 | match said_taken { |
| 257 | true => Some(out), |
| 258 | false => None, |
| 259 | } |
| 260 | } |
| 261 | |
| 262 | /// Writes the index, replacing whatever was there. |
| 263 | /// |
| 264 | /// Aside and renamed over, so that a reader never meets half of one; a |
| 265 | /// half-written index would be discarded on the next read anyway, and this |
| 266 | /// costs nothing. |
| 267 | pub fn write(&self, dir: &Path) |
| 268 | -> Outcome<()> |
| 269 | { |
| 270 | let path = Self::path_of(dir); |
| 271 | let mut text = fmt!("{}\n", STAT_FORMAT); |
| 272 | text.push_str(&fmt!("taken {}\n", self.taken)); |
| 273 | let said: Vec<String> = self.at.iter().map(|id| fmt!("{}", id)).collect(); |
| 274 | text.push_str(&fmt!("at {}\n", said.join(" "))); |
| 275 | for (path, seen) in &self.seen { |
| 276 | text.push_str(&fmt!("{} {} {} {} {}\n", |
| 277 | seen.len, seen.mtime, seen.ctime, seen.ino, text_of(path))); |
| 278 | } |
| 279 | let aside = path.with_extension("new"); |
| 280 | match fs::write(&aside, text.as_bytes()) { |
| 281 | Ok(()) => (), |
| 282 | Err(e) => return Err(err!(e, |
| 283 | "The working copy index could not be written aside to {:?}.", aside; |
| 284 | IO, File, Write)), |
| 285 | } |
| 286 | match fs::rename(&aside, &path) { |
| 287 | Ok(()) => Ok(()), |
| 288 | Err(e) => { |
| 289 | let _ = fs::remove_file(&aside); |
| 290 | Err(err!(e, |
| 291 | "The working copy index at {:?} could not be moved over {:?}.", |
| 292 | aside, path; |
| 293 | IO, File, Write)) |
| 294 | }, |
| 295 | } |
| 296 | } |
| 297 | |
| 298 | /// Removes the index, because what it described is no longer what is there. |
| 299 | pub fn forget(dir: &Path) { |
| 300 | let _ = fs::remove_file(Self::path_of(dir)); |
| 301 | } |
| 302 | |
| 303 | /// Does this index describe the history as it stands? |
| 304 | pub fn taken_at(&self, frontier: &[OpId]) -> bool { |
| 305 | self.at == frontier |
| 306 | } |
| 307 | |
| 308 | /// The paths it holds, which is what the last scan read. |
| 309 | pub fn paths(&self) -> impl Iterator<Item = &Vec<u8>> { |
| 310 | self.seen.keys() |
| 311 | } |
| 312 | |
| 313 | /// Is every path still exactly what it was, and old enough to say so? |
| 314 | /// |
| 315 | /// `found` is what a stat-only walk of the working copy turned up, gathered |
| 316 | /// under the same rules the reading walk uses. Everything is a reason to say |
| 317 | /// no: a path on one side and not the other, a field that differs, or an |
| 318 | /// entry recorded too recently for its time to mean anything. |
| 319 | pub fn agrees(&self, found: &BTreeMap<Vec<u8>, Seen>) -> bool { |
| 320 | if found.len() != self.seen.len() { |
| 321 | return false; |
| 322 | } |
| 323 | for (path, was) in &self.seen { |
| 324 | if !was.settled(self.taken) { |
| 325 | return false; |
| 326 | } |
| 327 | match found.get(path) { |
| 328 | Some(now) if now == was => (), |
| 329 | _ => return false, |
| 330 | } |
| 331 | } |
| 332 | true |
| 333 | } |
| 334 | } |
| 335 | |
| 336 | |
| 337 | |
| 338 | /// Does this operation leave every file exactly where it was? |
| 339 | /// |
| 340 | /// The question the index needs answered, and it is narrower than "renders |
| 341 | /// nothing": what the index claims is that the working copy equals the history, |
| 342 | /// and a working copy is paths, bytes and modes. An operation that changes none |
| 343 | /// of those leaves the claim true however much else it adds. |
| 344 | /// |
| 345 | /// [`Op::Mark`] and [`Op::Note`] both answer yes, and the reason is in what the |
| 346 | /// placement machinery asks of an operation. `Slots::place_without`, |
| 347 | /// `Claims::build_without` and `Dead::build` read `origins()`, `regions()` and |
| 348 | /// the inserted bytes; `Op::Mark` and `Op::Note` have no `insert`, and both |
| 349 | /// return `(None, None)` and `&[]` from the first two. Neither answers |
| 350 | /// `names_file()`. Neither is named anywhere in `seq/slot.rs`, `seq/claim.rs` |
| 351 | /// or `seq/atom.rs` at all. A note does name content -- through `note_on()`, |
| 352 | /// which is a separate accessor -- and what that buys it is a place in |
| 353 | /// `Rendered::notes`, which no working copy materialises. |
| 354 | /// |
| 355 | /// **The match is exhaustive on purpose.** A new operation in the vocabulary |
| 356 | /// must fail to compile here and be decided, because a catch-all answering yes |
| 357 | /// would be an operation that changed a file and an index that said it had not, |
| 358 | /// and what that costs is somebody's edit. |
| 359 | /// |
| 360 | /// Read as an assertion this would not be worth having; it is asserted on |
| 361 | /// running code by `a_mark_and_a_note_leave_every_file_exactly_where_it_was` in |
| 362 | /// `cli/tests/stat.rs`, which renders a repository before and after each and |
| 363 | /// compares the bytes and the modes of every file. |
| 364 | pub fn leaves_the_files_alone(op: &Op) -> bool { |
| 365 | match op { |
| 366 | Op::Mark { .. } => true, |
| 367 | Op::Note { .. } => true, |
| 368 | // Everything below writes, moves, names or removes content, or is a |
| 369 | // claim this lane has not measured. No is the safe answer and it costs |
| 370 | // one full capture. |
| 371 | Op::FileCreate { .. } => false, |
| 372 | Op::FileDelete { .. } => false, |
| 373 | Op::FileRename { .. } => false, |
| 374 | Op::FileMode { .. } => false, |
| 375 | Op::Splice { .. } => false, |
| 376 | Op::Move { .. } => false, |
| 377 | Op::Proposal { .. } => false, |
| 378 | Op::Said { .. } => false, |
| 379 | Op::Settled { .. } => false, |
| 380 | Op::Reverts { .. } => false, |
| 381 | } |
| 382 | } |
| 383 | |
| 384 | /// Moves an index's frontier on, where everything appended since it was taken |
| 385 | /// left every file where it was. |
| 386 | /// |
| 387 | /// Every command that appends anything ends by marking the point it reached, so |
| 388 | /// without this an `ore mark` over an unchanged working copy costs the command |
| 389 | /// after it a full capture: 0.65 s and 169 MB against 0.34 s and 69 MB, |
| 390 | /// measured. The mark moves the frontier and moves nothing else, so what is |
| 391 | /// stale about the index is one field. |
| 392 | /// |
| 393 | /// **Only that field is written.** Every file fact stays exactly as it was |
| 394 | /// measured, which is what keeps this safe rather than convenient: a verb that |
| 395 | /// wrote to the working copy is caught by the size, the times and the inode |
| 396 | /// afterwards exactly as it was before, and a verb that appended anything else |
| 397 | /// takes the index away instead. |
| 398 | pub fn restamp<'a, I>(dir: &Path, was: &[OpId], now: &[OpId], appended: I) |
| 399 | -> Outcome<()> |
| 400 | where |
| 401 | I: IntoIterator<Item = &'a Op>, |
| 402 | { |
| 403 | if was == now { |
| 404 | return Ok(()); |
| 405 | } |
| 406 | for op in appended { |
| 407 | if !leaves_the_files_alone(op) { |
| 408 | // The index describes a state that has gone. The frontier check would |
| 409 | // refuse it anyway; taking it away says so rather than leaving it to |
| 410 | // be refused again on every command until something replaces it. |
| 411 | Stats::forget(dir); |
| 412 | return Ok(()); |
| 413 | } |
| 414 | } |
| 415 | let mut index = match Stats::read(dir) { |
| 416 | Some(index) => index, |
| 417 | None => return Ok(()), |
| 418 | }; |
| 419 | if !index.taken_at(was) { |
| 420 | Stats::forget(dir); |
| 421 | return Ok(()); |
| 422 | } |
| 423 | index.describes(now); |
| 424 | index.write(dir) |
| 425 | } |
| 426 | |
| 427 | |
| 428 | #[cfg(test)] |
| 429 | mod tests { |
| 430 | use super::*; |
| 431 | |
| 432 | fn seen_at(mtime: i128) -> Seen { |
| 433 | Seen { len: 10, mtime, ctime: mtime, ino: 7 } |
| 434 | } |
| 435 | |
| 436 | fn one(taken: i128, mtime: i128) -> Stats { |
| 437 | let mut out = Stats { taken, at: Vec::new(), seen: BTreeMap::new() }; |
| 438 | out.seen.insert(b"a.txt".to_vec(), seen_at(mtime)); |
| 439 | out |
| 440 | } |
| 441 | |
| 442 | /// A file whose time is not [`RACY`] older than the clock reading is read |
| 443 | /// again, however exactly the disk agrees with what was recorded. |
| 444 | /// |
| 445 | /// This is the case nothing else can catch: a write in the same tick that |
| 446 | /// leaves the size alone. The fixture is built so that the disk agrees |
| 447 | /// **perfectly** -- the same entry, byte for byte -- so the only thing that |
| 448 | /// can make the answer no is the age. |
| 449 | #[test] |
| 450 | fn a_time_too_recent_to_mean_anything_is_not_believed() -> Outcome<()> { |
| 451 | let taken = 10 * RACY; |
| 452 | let mut found = BTreeMap::new(); |
| 453 | |
| 454 | // One nanosecond too young, and the disk agrees exactly. |
| 455 | let index = one(taken, taken - RACY + 1); |
| 456 | found.insert(b"a.txt".to_vec(), seen_at(taken - RACY + 1)); |
| 457 | assert!(!index.agrees(&found), "a time inside the window means nothing"); |
| 458 | |
| 459 | // One nanosecond older, and the same comparison is believed. |
| 460 | let index = one(taken, taken - RACY); |
| 461 | let mut found = BTreeMap::new(); |
| 462 | found.insert(b"a.txt".to_vec(), seen_at(taken - RACY)); |
| 463 | assert!(index.agrees(&found), "and outside it, it does"); |
| 464 | |
| 465 | // A time in the future says the clock or the filesystem is not to be |
| 466 | // reasoned with, so it is not reasoned with. |
| 467 | let index = one(taken, taken + RACY); |
| 468 | let mut found = BTreeMap::new(); |
| 469 | found.insert(b"a.txt".to_vec(), seen_at(taken + RACY)); |
| 470 | assert!(!index.agrees(&found), "and a time after the reading is not believed"); |
| 471 | Ok(()) |
| 472 | } |
| 473 | |
| 474 | /// Each of the four fields is enough on its own to make an entry disagree. |
| 475 | #[test] |
| 476 | fn every_field_is_load_bearing() -> Outcome<()> { |
| 477 | let taken = 10 * RACY; |
| 478 | let mtime = taken - 2 * RACY; |
| 479 | let index = one(taken, mtime); |
| 480 | let same = Seen { len: 10, mtime, ctime: mtime, ino: 7 }; |
| 481 | |
| 482 | let mut found = BTreeMap::new(); |
| 483 | found.insert(b"a.txt".to_vec(), same); |
| 484 | assert!(index.agrees(&found), "the fixture agrees before it is disturbed"); |
| 485 | |
| 486 | for (what, changed) in [ |
| 487 | ("size", Seen { len: 11, ..same }), |
| 488 | ("time", Seen { mtime: mtime + 1, ..same }), |
| 489 | ("change time", Seen { ctime: mtime + 1, ..same }), |
| 490 | ("inode", Seen { ino: 8, ..same }), |
| 491 | ] { |
| 492 | let mut found = BTreeMap::new(); |
| 493 | found.insert(b"a.txt".to_vec(), changed); |
| 494 | assert!(!index.agrees(&found), "a different {} is a change", what); |
| 495 | } |
| 496 | |
| 497 | // And a path on one side and not the other. |
| 498 | let mut found = BTreeMap::new(); |
| 499 | found.insert(b"a.txt".to_vec(), same); |
| 500 | found.insert(b"b.txt".to_vec(), same); |
| 501 | assert!(!index.agrees(&found), "a path that arrived is a change"); |
| 502 | assert!(!index.agrees(&BTreeMap::new()), "and one that left is too"); |
| 503 | Ok(()) |
| 504 | } |
| 505 | |
| 506 | /// A working copy that has not moved still differs from a history that has. |
| 507 | #[test] |
| 508 | fn an_index_names_the_frontier_it_was_taken_at() -> Outcome<()> { |
| 509 | use oxedyne_fe2o3_ore::id::ReplicaId; |
| 510 | |
| 511 | let mut index = one(10 * RACY, 0); |
| 512 | let was = vec![OpId::new(ReplicaId::new(1), 4)]; |
| 513 | index.describes(&was); |
| 514 | assert!(index.taken_at(&was)); |
| 515 | assert!(!index.taken_at(&[OpId::new(ReplicaId::new(1), 5)])); |
| 516 | assert!(!index.taken_at(&[])); |
| 517 | Ok(()) |
| 518 | } |
| 519 | |
| 520 | /// **The claim `leaves_the_files_alone` makes, put to the render itself.** |
| 521 | /// |
| 522 | /// The integration test in `cli/tests/stat.rs` compares the working copy, |
| 523 | /// which a mark never writes to, so it could not tell a mark that changed the |
| 524 | /// render from one that did not. This compares what the render produces: |
| 525 | /// every path, every byte and every mode, before and after. |
| 526 | /// |
| 527 | /// **The fixture is built to be able to disagree**, and that is asserted |
| 528 | /// before anything else: two files, one filled and then spliced again so its |
| 529 | /// bytes come from more than one operation, one carrying bytes that are not |
| 530 | /// text, one with a mode that is not the default. An ordinary splice is |
| 531 | /// appended first and the comparison is required to notice it. |
| 532 | #[test] |
| 533 | fn a_mark_and_a_note_change_nothing_the_render_produces() -> Outcome<()> { |
| 534 | use oxedyne_fe2o3_ore::id::{ |
| 535 | Anchor, |
| 536 | ReplicaId, |
| 537 | }; |
| 538 | use oxedyne_fe2o3_ore::log::OpLog; |
| 539 | use oxedyne_fe2o3_ore::op::{ |
| 540 | Header, |
| 541 | Mode, |
| 542 | Record, |
| 543 | }; |
| 544 | use ore_store::tree; |
| 545 | |
| 546 | let me = ReplicaId::new(1); |
| 547 | let mut log = OpLog::new(); |
| 548 | let mut counter = 0u64; |
| 549 | let mut next = |counter: &mut u64| { *counter += 1; OpId::new(me, *counter) }; |
| 550 | let mut parents: Vec<OpId> = Vec::new(); |
| 551 | |
| 552 | let one = next(&mut counter); |
| 553 | res!(log.append(Record::new(res!(Header::new(one, parents.clone())), |
| 554 | Op::FileCreate { path: b"one.txt".to_vec() }))); |
| 555 | parents = vec![one]; |
| 556 | let id = next(&mut counter); |
| 557 | res!(log.append(Record::new(res!(Header::new(id, parents.clone())), Op::Splice { |
| 558 | left: Some(Anchor::origin(one)), |
| 559 | right: None, |
| 560 | remove: Vec::new(), |
| 561 | insert: b"alpha\nbeta\ngamma\n".to_vec().into(), |
| 562 | }))); |
| 563 | parents = vec![id]; |
| 564 | let two = next(&mut counter); |
| 565 | res!(log.append(Record::new(res!(Header::new(two, parents.clone())), |
| 566 | Op::FileCreate { path: b"deep/two.txt".to_vec() }))); |
| 567 | parents = vec![two]; |
| 568 | let id = next(&mut counter); |
| 569 | res!(log.append(Record::new(res!(Header::new(id, parents.clone())), Op::Splice { |
| 570 | left: Some(Anchor::origin(two)), |
| 571 | right: None, |
| 572 | remove: Vec::new(), |
| 573 | insert: vec![0u8, 1, 2, 0xff, 0xfe, b'\n', 0x80].into(), |
| 574 | }))); |
| 575 | parents = vec![id]; |
| 576 | let id = next(&mut counter); |
| 577 | res!(log.append(Record::new(res!(Header::new(id, parents.clone())), |
| 578 | Op::FileMode { file: two, mode: Mode::Executable }))); |
| 579 | parents = vec![id]; |
| 580 | // A second edit inside the first file, so its bytes come from two |
| 581 | // operations and the render has runs to order rather than one run a file. |
| 582 | let built = res!(tree::whole(&log)); |
| 583 | let edit = { |
| 584 | let view = res!(built.repo.file(one).ok_or_else(|| err!( |
| 585 | "the render lost a file it was just given"; Test, Missing))); |
| 586 | res!(view.splice(6, 4, b"BETA".to_vec())) |
| 587 | }; |
| 588 | let id = next(&mut counter); |
| 589 | res!(log.append(Record::new(res!(Header::new(id, parents.clone())), edit))); |
| 590 | parents = vec![id]; |
| 591 | |
| 592 | let built = res!(tree::whole(&log)); |
| 593 | let was = res!(built.contents()); |
| 594 | let modes = res!(built.modes()); |
| 595 | assert_eq!(was.len(), 2, "the fixture holds both files"); |
| 596 | assert!(was.values().any(|b| b.contains(&0xff)), "and bytes that are not text"); |
| 597 | assert!(modes.values().any(|m| *m == Mode::Executable), |
| 598 | "and a mode that is not the default"); |
| 599 | |
| 600 | // FIRST: the comparison must be able to fail. |
| 601 | let disturb = { |
| 602 | let view = res!(built.repo.file(one).ok_or_else(|| err!( |
| 603 | "missing file"; Test, Missing))); |
| 604 | res!(view.splice(0, 0, b"zero\n".to_vec())) |
| 605 | }; |
| 606 | let mut probe = log.clone(); |
| 607 | let id = next(&mut counter); |
| 608 | res!(probe.append(Record::new(res!(Header::new(id, parents.clone())), disturb))); |
| 609 | assert_ne!(was, res!(res!(tree::whole(&probe)).contents()), |
| 610 | "an ordinary splice must show up in this comparison, or it proves nothing"); |
| 611 | |
| 612 | // A MARK. |
| 613 | let mut probe = log.clone(); |
| 614 | let id = next(&mut counter); |
| 615 | res!(probe.append(Record::new(res!(Header::new(id, parents.clone())), |
| 616 | Op::Mark { name: fmt!("a point"), body: None, time: Some(1) }))); |
| 617 | let after = res!(tree::whole(&probe)); |
| 618 | assert_eq!(was, res!(after.contents()), "a mark changes no path and no byte"); |
| 619 | assert_eq!(modes, res!(after.modes()), "and no mode"); |
| 620 | |
| 621 | // A NOTE, which is the one of the two that names content. |
| 622 | let on = { |
| 623 | let view = res!(built.repo.file(one).ok_or_else(|| err!( |
| 624 | "missing file"; Test, Missing))); |
| 625 | res!(view.span(0, 5)) |
| 626 | }; |
| 627 | assert!(!on.is_empty(), "the note names real content, not nothing"); |
| 628 | let mut probe = log.clone(); |
| 629 | let id = next(&mut counter); |
| 630 | res!(probe.append(Record::new(res!(Header::new(id, parents.clone())), |
| 631 | Op::Note { on, text: b"about the first line".to_vec() }))); |
| 632 | let after = res!(tree::whole(&probe)); |
| 633 | assert_eq!(was, res!(after.contents()), "a note changes no path and no byte"); |
| 634 | assert_eq!(modes, res!(after.modes()), "and no mode"); |
| 635 | assert!(!after.repo.notes().is_empty(), |
| 636 | "while the note really did reach the render, so this is not comparing a \ |
| 637 | repository it never got to"); |
| 638 | Ok(()) |
| 639 | } |
| 640 | |
| 641 | /// An operation that moves a byte takes the index away, and one that does not |
| 642 | /// moves its frontier on. |
| 643 | /// |
| 644 | /// The check `restamp` makes, asserted where it can be decided: driving the |
| 645 | /// binary cannot reach it, because every verb whose own operations change |
| 646 | /// content also either materialises the working copy or leaves it for the |
| 647 | /// next capture, and in both cases the FILE FACTS disagree and the index is |
| 648 | /// refused before this ever matters. That makes the check depth rather than |
| 649 | /// the only line -- worth having, because the case it guards is a render that |
| 650 | /// moved while the disk did not, which is a working copy walking away from |
| 651 | /// its history rather than an edit being lost -- and it means the assertion |
| 652 | /// belongs here. |
| 653 | #[test] |
| 654 | fn only_an_operation_that_moves_nothing_carries_the_index_forward() -> Outcome<()> { |
| 655 | use oxedyne_fe2o3_ore::id::{ |
| 656 | Anchor, |
| 657 | ReplicaId, |
| 658 | }; |
| 659 | |
| 660 | let dir = std::env::temp_dir().join(fmt!( |
| 661 | "ore_stat_restamp_{}_{}", std::process::id(), |
| 662 | res!(SystemTime::now().duration_since(UNIX_EPOCH)).as_nanos())); |
| 663 | let was = vec![OpId::new(ReplicaId::new(1), 4)]; |
| 664 | let now = vec![OpId::new(ReplicaId::new(1), 5)]; |
| 665 | let laid = |dir: &Path| -> Outcome<()> { |
| 666 | let _ = fs::remove_dir_all(dir); |
| 667 | res!(fs::create_dir_all(dir)); |
| 668 | let mut index = res!(Stats::beginning()); |
| 669 | index.describes(&was); |
| 670 | index.put(b"a.txt", seen_at(1)); |
| 671 | res!(index.write(dir)); |
| 672 | Ok(()) |
| 673 | }; |
| 674 | |
| 675 | // A mark carries it forward, and only the frontier moves. |
| 676 | res!(laid(&dir)); |
| 677 | let before = res!(fs::read_to_string(Stats::path_of(&dir))); |
| 678 | let mark = Op::Mark { name: fmt!("a point"), body: None, time: Some(1) }; |
| 679 | res!(restamp(&dir, &was, &now, [&mark])); |
| 680 | let after = res!(fs::read_to_string(Stats::path_of(&dir))); |
| 681 | let facts = |t: &str| -> Vec<String> { |
| 682 | t.lines().filter(|l| !l.starts_with("at ")).map(|l| fmt!("{}", l)).collect() |
| 683 | }; |
| 684 | assert_eq!(facts(&before), facts(&after), "every file fact is what it was"); |
| 685 | assert_ne!(before, after, "and the frontier is not"); |
| 686 | let held = res!(Stats::read(&dir).ok_or_else(|| err!("gone"; Test, Missing))); |
| 687 | assert!(held.taken_at(&now), "the index now names where the history is"); |
| 688 | |
| 689 | // A note likewise. |
| 690 | res!(laid(&dir)); |
| 691 | let note = Op::Note { on: Vec::new(), text: b"said".to_vec() }; |
| 692 | res!(restamp(&dir, &was, &now, [¬e])); |
| 693 | assert!(res!(Stats::read(&dir).ok_or_else(|| err!("gone"; Test, Missing))) |
| 694 | .taken_at(&now), "a note carries it forward too"); |
| 695 | |
| 696 | // Anything that moves a byte takes it away, even beside a mark. |
| 697 | let splice = Op::Splice { |
| 698 | left: Some(Anchor::origin(OpId::new(ReplicaId::new(1), 1))), |
| 699 | right: None, |
| 700 | remove: Vec::new(), |
| 701 | insert: b"more".to_vec().into(), |
| 702 | }; |
| 703 | for (what, ops) in [ |
| 704 | ("a splice", vec![&splice]), |
| 705 | ("a splice after a mark", vec![&mark, &splice]), |
| 706 | ("a splice before a mark", vec![&splice, &mark]), |
| 707 | ] { |
| 708 | res!(laid(&dir)); |
| 709 | res!(restamp(&dir, &was, &now, ops)); |
| 710 | assert!(Stats::read(&dir).is_none(), "{} takes the index away", what); |
| 711 | } |
| 712 | |
| 713 | // And an index taken somewhere else is not carried forward either. |
| 714 | res!(laid(&dir)); |
| 715 | let elsewhere = vec![OpId::new(ReplicaId::new(9), 9)]; |
| 716 | res!(restamp(&dir, &elsewhere, &now, [&mark])); |
| 717 | assert!(Stats::read(&dir).is_none(), |
| 718 | "an index describing another point is not re-stamped, it is discarded"); |
| 719 | let _ = fs::remove_dir_all(&dir); |
| 720 | Ok(()) |
| 721 | } |
| 722 | |
| 723 | /// A line this build cannot read discards the whole index. |
| 724 | /// |
| 725 | /// Half a cache is not worth the question of which half, and a format that |
| 726 | /// grows will be met by a build that does not know it. Nothing downstream |
| 727 | /// depends on this being strict -- an index that lost an entry disagrees on |
| 728 | /// the path count and is refused anyway -- so this is depth rather than the |
| 729 | /// only line, and it is asserted here because the integration test cannot |
| 730 | /// tell the two apart: every fixture it could use is caught by the count. |
| 731 | #[test] |
| 732 | fn a_line_this_build_cannot_read_discards_the_index() -> Outcome<()> { |
| 733 | let dir = std::env::temp_dir().join(fmt!( |
| 734 | "ore_stat_bad_{}_{}", std::process::id(), |
| 735 | res!(SystemTime::now().duration_since(UNIX_EPOCH)).as_nanos())); |
| 736 | res!(fs::create_dir_all(&dir)); |
| 737 | let good = fmt!("{}\ntaken 5000000000\nat r1:2\n10 1 1 7 {}\n", |
| 738 | STAT_FORMAT, text_of(b"a.txt")); |
| 739 | |
| 740 | assert!(Stats::read(&dir).is_none(), "an absent index is no index"); |
| 741 | |
| 742 | res!(fs::write(Stats::path_of(&dir), good.as_bytes())); |
| 743 | let read = res!(Stats::read(&dir).ok_or_else(|| err!( |
| 744 | "a sound index did not read"; Test, Missing))); |
| 745 | assert_eq!(read.seen.len(), 1, "the fixture reads before it is disturbed"); |
| 746 | |
| 747 | for bad in [ |
| 748 | fmt!("{}\ntaken 5000000000\nat r1:2\n10 1 1 7\n", STAT_FORMAT), |
| 749 | fmt!("{}\ntaken 5000000000\nat r1:2\n10 1 1 7 {} extra\n", |
| 750 | STAT_FORMAT, text_of(b"a.txt")), |
| 751 | fmt!("{}\ntaken 5000000000\nat r1:2\nten 1 1 7 {}\n", |
| 752 | STAT_FORMAT, text_of(b"a.txt")), |
| 753 | fmt!("{}\ntaken 5000000000\nat r1:2\n10 1 1 7 !!not-encodable!!\n", |
| 754 | STAT_FORMAT), |
| 755 | fmt!("{}\nat r1:2\n10 1 1 7 {}\n", STAT_FORMAT, text_of(b"a.txt")), |
| 756 | fmt!("ORESTAT 2\ntaken 5000000000\nat r1:2\n10 1 1 7 {}\n", |
| 757 | text_of(b"a.txt")), |
| 758 | fmt!("{}\ntaken 5000000000\nat not-an-op-id\n", STAT_FORMAT), |
| 759 | ] { |
| 760 | res!(fs::write(Stats::path_of(&dir), bad.as_bytes())); |
| 761 | assert!(Stats::read(&dir).is_none(), |
| 762 | "this must read as no index at all: {:?}", bad); |
| 763 | } |
| 764 | let _ = fs::remove_dir_all(&dir); |
| 765 | Ok(()) |
| 766 | } |
| 767 | |
| 768 | /// It round-trips, including a path that is not UTF-8. |
| 769 | #[test] |
| 770 | fn what_is_written_is_what_is_read() -> Outcome<()> { |
| 771 | use oxedyne_fe2o3_ore::id::ReplicaId; |
| 772 | |
| 773 | let dir = std::env::temp_dir().join(fmt!( |
| 774 | "ore_stat_{}_{}", std::process::id(), |
| 775 | res!(SystemTime::now().duration_since(UNIX_EPOCH)).as_nanos())); |
| 776 | res!(fs::create_dir_all(&dir)); |
| 777 | |
| 778 | let mut index = res!(Stats::beginning()); |
| 779 | index.describes(&[OpId::new(ReplicaId::new(9), 3)]); |
| 780 | index.put(b"plain.txt", seen_at(1)); |
| 781 | index.put(&[0xff, 0xfe, b'/', 0x80], seen_at(2)); |
| 782 | res!(index.write(&dir)); |
| 783 | |
| 784 | let back = res!(Stats::read(&dir).ok_or_else(|| err!( |
| 785 | "What was just written did not read back."; Test, Missing))); |
| 786 | assert_eq!(back.taken, index.taken); |
| 787 | assert_eq!(back.at, index.at); |
| 788 | assert_eq!(back.seen, index.seen, "a path that is not text survives it"); |
| 789 | |
| 790 | let _ = fs::remove_dir_all(&dir); |
| 791 | Ok(()) |
| 792 | } |
| 793 | } |