Black Lives Matter. Support the Equal Justice Initiative.

Source file src/cmd/go/internal/cache/cache.go

Documentation: cmd/go/internal/cache

     1  // Copyright 2017 The Go Authors. All rights reserved.
     2  // Use of this source code is governed by a BSD-style
     3  // license that can be found in the LICENSE file.
     4  
     5  // Package cache implements a build artifact cache.
     6  package cache
     7  
     8  import (
     9  	"bytes"
    10  	"crypto/sha256"
    11  	"encoding/hex"
    12  	"errors"
    13  	"fmt"
    14  	"io"
    15  	"io/fs"
    16  	"os"
    17  	"path/filepath"
    18  	"strconv"
    19  	"strings"
    20  	"time"
    21  
    22  	"cmd/go/internal/lockedfile"
    23  )
    24  
    25  // An ActionID is a cache action key, the hash of a complete description of a
    26  // repeatable computation (command line, environment variables,
    27  // input file contents, executable contents).
    28  type ActionID [HashSize]byte
    29  
    30  // An OutputID is a cache output key, the hash of an output of a computation.
    31  type OutputID [HashSize]byte
    32  
    33  // A Cache is a package cache, backed by a file system directory tree.
    34  type Cache struct {
    35  	dir string
    36  	now func() time.Time
    37  }
    38  
    39  // Open opens and returns the cache in the given directory.
    40  //
    41  // It is safe for multiple processes on a single machine to use the
    42  // same cache directory in a local file system simultaneously.
    43  // They will coordinate using operating system file locks and may
    44  // duplicate effort but will not corrupt the cache.
    45  //
    46  // However, it is NOT safe for multiple processes on different machines
    47  // to share a cache directory (for example, if the directory were stored
    48  // in a network file system). File locking is notoriously unreliable in
    49  // network file systems and may not suffice to protect the cache.
    50  //
    51  func Open(dir string) (*Cache, error) {
    52  	info, err := os.Stat(dir)
    53  	if err != nil {
    54  		return nil, err
    55  	}
    56  	if !info.IsDir() {
    57  		return nil, &fs.PathError{Op: "open", Path: dir, Err: fmt.Errorf("not a directory")}
    58  	}
    59  	for i := 0; i < 256; i++ {
    60  		name := filepath.Join(dir, fmt.Sprintf("%02x", i))
    61  		if err := os.MkdirAll(name, 0777); err != nil {
    62  			return nil, err
    63  		}
    64  	}
    65  	c := &Cache{
    66  		dir: dir,
    67  		now: time.Now,
    68  	}
    69  	return c, nil
    70  }
    71  
    72  // fileName returns the name of the file corresponding to the given id.
    73  func (c *Cache) fileName(id [HashSize]byte, key string) string {
    74  	return filepath.Join(c.dir, fmt.Sprintf("%02x", id[0]), fmt.Sprintf("%x", id)+"-"+key)
    75  }
    76  
    77  // An entryNotFoundError indicates that a cache entry was not found, with an
    78  // optional underlying reason.
    79  type entryNotFoundError struct {
    80  	Err error
    81  }
    82  
    83  func (e *entryNotFoundError) Error() string {
    84  	if e.Err == nil {
    85  		return "cache entry not found"
    86  	}
    87  	return fmt.Sprintf("cache entry not found: %v", e.Err)
    88  }
    89  
    90  func (e *entryNotFoundError) Unwrap() error {
    91  	return e.Err
    92  }
    93  
    94  const (
    95  	// action entry file is "v1 <hex id> <hex out> <decimal size space-padded to 20 bytes> <unixnano space-padded to 20 bytes>\n"
    96  	hexSize   = HashSize * 2
    97  	entrySize = 2 + 1 + hexSize + 1 + hexSize + 1 + 20 + 1 + 20 + 1
    98  )
    99  
   100  // verify controls whether to run the cache in verify mode.
   101  // In verify mode, the cache always returns errMissing from Get
   102  // but then double-checks in Put that the data being written
   103  // exactly matches any existing entry. This provides an easy
   104  // way to detect program behavior that would have been different
   105  // had the cache entry been returned from Get.
   106  //
   107  // verify is enabled by setting the environment variable
   108  // GODEBUG=gocacheverify=1.
   109  var verify = false
   110  
   111  var errVerifyMode = errors.New("gocacheverify=1")
   112  
   113  // DebugTest is set when GODEBUG=gocachetest=1 is in the environment.
   114  var DebugTest = false
   115  
   116  func init() { initEnv() }
   117  
   118  func initEnv() {
   119  	verify = false
   120  	debugHash = false
   121  	debug := strings.Split(os.Getenv("GODEBUG"), ",")
   122  	for _, f := range debug {
   123  		if f == "gocacheverify=1" {
   124  			verify = true
   125  		}
   126  		if f == "gocachehash=1" {
   127  			debugHash = true
   128  		}
   129  		if f == "gocachetest=1" {
   130  			DebugTest = true
   131  		}
   132  	}
   133  }
   134  
   135  // Get looks up the action ID in the cache,
   136  // returning the corresponding output ID and file size, if any.
   137  // Note that finding an output ID does not guarantee that the
   138  // saved file for that output ID is still available.
   139  func (c *Cache) Get(id ActionID) (Entry, error) {
   140  	if verify {
   141  		return Entry{}, &entryNotFoundError{Err: errVerifyMode}
   142  	}
   143  	return c.get(id)
   144  }
   145  
   146  type Entry struct {
   147  	OutputID OutputID
   148  	Size     int64
   149  	Time     time.Time
   150  }
   151  
   152  // get is Get but does not respect verify mode, so that Put can use it.
   153  func (c *Cache) get(id ActionID) (Entry, error) {
   154  	missing := func(reason error) (Entry, error) {
   155  		return Entry{}, &entryNotFoundError{Err: reason}
   156  	}
   157  	f, err := os.Open(c.fileName(id, "a"))
   158  	if err != nil {
   159  		return missing(err)
   160  	}
   161  	defer f.Close()
   162  	entry := make([]byte, entrySize+1) // +1 to detect whether f is too long
   163  	if n, err := io.ReadFull(f, entry); n > entrySize {
   164  		return missing(errors.New("too long"))
   165  	} else if err != io.ErrUnexpectedEOF {
   166  		if err == io.EOF {
   167  			return missing(errors.New("file is empty"))
   168  		}
   169  		return missing(err)
   170  	} else if n < entrySize {
   171  		return missing(errors.New("entry file incomplete"))
   172  	}
   173  	if entry[0] != 'v' || entry[1] != '1' || entry[2] != ' ' || entry[3+hexSize] != ' ' || entry[3+hexSize+1+hexSize] != ' ' || entry[3+hexSize+1+hexSize+1+20] != ' ' || entry[entrySize-1] != '\n' {
   174  		return missing(errors.New("invalid header"))
   175  	}
   176  	eid, entry := entry[3:3+hexSize], entry[3+hexSize:]
   177  	eout, entry := entry[1:1+hexSize], entry[1+hexSize:]
   178  	esize, entry := entry[1:1+20], entry[1+20:]
   179  	etime, entry := entry[1:1+20], entry[1+20:]
   180  	var buf [HashSize]byte
   181  	if _, err := hex.Decode(buf[:], eid); err != nil {
   182  		return missing(fmt.Errorf("decoding ID: %v", err))
   183  	} else if buf != id {
   184  		return missing(errors.New("mismatched ID"))
   185  	}
   186  	if _, err := hex.Decode(buf[:], eout); err != nil {
   187  		return missing(fmt.Errorf("decoding output ID: %v", err))
   188  	}
   189  	i := 0
   190  	for i < len(esize) && esize[i] == ' ' {
   191  		i++
   192  	}
   193  	size, err := strconv.ParseInt(string(esize[i:]), 10, 64)
   194  	if err != nil {
   195  		return missing(fmt.Errorf("parsing size: %v", err))
   196  	} else if size < 0 {
   197  		return missing(errors.New("negative size"))
   198  	}
   199  	i = 0
   200  	for i < len(etime) && etime[i] == ' ' {
   201  		i++
   202  	}
   203  	tm, err := strconv.ParseInt(string(etime[i:]), 10, 64)
   204  	if err != nil {
   205  		return missing(fmt.Errorf("parsing timestamp: %v", err))
   206  	} else if tm < 0 {
   207  		return missing(errors.New("negative timestamp"))
   208  	}
   209  
   210  	c.used(c.fileName(id, "a"))
   211  
   212  	return Entry{buf, size, time.Unix(0, tm)}, nil
   213  }
   214  
   215  // GetFile looks up the action ID in the cache and returns
   216  // the name of the corresponding data file.
   217  func (c *Cache) GetFile(id ActionID) (file string, entry Entry, err error) {
   218  	entry, err = c.Get(id)
   219  	if err != nil {
   220  		return "", Entry{}, err
   221  	}
   222  	file = c.OutputFile(entry.OutputID)
   223  	info, err := os.Stat(file)
   224  	if err != nil {
   225  		return "", Entry{}, &entryNotFoundError{Err: err}
   226  	}
   227  	if info.Size() != entry.Size {
   228  		return "", Entry{}, &entryNotFoundError{Err: errors.New("file incomplete")}
   229  	}
   230  	return file, entry, nil
   231  }
   232  
   233  // GetBytes looks up the action ID in the cache and returns
   234  // the corresponding output bytes.
   235  // GetBytes should only be used for data that can be expected to fit in memory.
   236  func (c *Cache) GetBytes(id ActionID) ([]byte, Entry, error) {
   237  	entry, err := c.Get(id)
   238  	if err != nil {
   239  		return nil, entry, err
   240  	}
   241  	data, _ := os.ReadFile(c.OutputFile(entry.OutputID))
   242  	if sha256.Sum256(data) != entry.OutputID {
   243  		return nil, entry, &entryNotFoundError{Err: errors.New("bad checksum")}
   244  	}
   245  	return data, entry, nil
   246  }
   247  
   248  // OutputFile returns the name of the cache file storing output with the given OutputID.
   249  func (c *Cache) OutputFile(out OutputID) string {
   250  	file := c.fileName(out, "d")
   251  	c.used(file)
   252  	return file
   253  }
   254  
   255  // Time constants for cache expiration.
   256  //
   257  // We set the mtime on a cache file on each use, but at most one per mtimeInterval (1 hour),
   258  // to avoid causing many unnecessary inode updates. The mtimes therefore
   259  // roughly reflect "time of last use" but may in fact be older by at most an hour.
   260  //
   261  // We scan the cache for entries to delete at most once per trimInterval (1 day).
   262  //
   263  // When we do scan the cache, we delete entries that have not been used for
   264  // at least trimLimit (5 days). Statistics gathered from a month of usage by
   265  // Go developers found that essentially all reuse of cached entries happened
   266  // within 5 days of the previous reuse. See golang.org/issue/22990.
   267  const (
   268  	mtimeInterval = 1 * time.Hour
   269  	trimInterval  = 24 * time.Hour
   270  	trimLimit     = 5 * 24 * time.Hour
   271  )
   272  
   273  // used makes a best-effort attempt to update mtime on file,
   274  // so that mtime reflects cache access time.
   275  //
   276  // Because the reflection only needs to be approximate,
   277  // and to reduce the amount of disk activity caused by using
   278  // cache entries, used only updates the mtime if the current
   279  // mtime is more than an hour old. This heuristic eliminates
   280  // nearly all of the mtime updates that would otherwise happen,
   281  // while still keeping the mtimes useful for cache trimming.
   282  func (c *Cache) used(file string) {
   283  	info, err := os.Stat(file)
   284  	if err == nil && c.now().Sub(info.ModTime()) < mtimeInterval {
   285  		return
   286  	}
   287  	os.Chtimes(file, c.now(), c.now())
   288  }
   289  
   290  // Trim removes old cache entries that are likely not to be reused.
   291  func (c *Cache) Trim() {
   292  	now := c.now()
   293  
   294  	// We maintain in dir/trim.txt the time of the last completed cache trim.
   295  	// If the cache has been trimmed recently enough, do nothing.
   296  	// This is the common case.
   297  	// If the trim file is corrupt, detected if the file can't be parsed, or the
   298  	// trim time is too far in the future, attempt the trim anyway. It's possible that
   299  	// the cache was full when the corruption happened. Attempting a trim on
   300  	// an empty cache is cheap, so there wouldn't be a big performance hit in that case.
   301  	if data, err := lockedfile.Read(filepath.Join(c.dir, "trim.txt")); err == nil {
   302  		if t, err := strconv.ParseInt(strings.TrimSpace(string(data)), 10, 64); err == nil {
   303  			lastTrim := time.Unix(t, 0)
   304  			if d := now.Sub(lastTrim); d < trimInterval && d > -mtimeInterval {
   305  				return
   306  			}
   307  		}
   308  	}
   309  
   310  	// Trim each of the 256 subdirectories.
   311  	// We subtract an additional mtimeInterval
   312  	// to account for the imprecision of our "last used" mtimes.
   313  	cutoff := now.Add(-trimLimit - mtimeInterval)
   314  	for i := 0; i < 256; i++ {
   315  		subdir := filepath.Join(c.dir, fmt.Sprintf("%02x", i))
   316  		c.trimSubdir(subdir, cutoff)
   317  	}
   318  
   319  	// Ignore errors from here: if we don't write the complete timestamp, the
   320  	// cache will appear older than it is, and we'll trim it again next time.
   321  	var b bytes.Buffer
   322  	fmt.Fprintf(&b, "%d", now.Unix())
   323  	if err := lockedfile.Write(filepath.Join(c.dir, "trim.txt"), &b, 0666); err != nil {
   324  		return
   325  	}
   326  }
   327  
   328  // trimSubdir trims a single cache subdirectory.
   329  func (c *Cache) trimSubdir(subdir string, cutoff time.Time) {
   330  	// Read all directory entries from subdir before removing
   331  	// any files, in case removing files invalidates the file offset
   332  	// in the directory scan. Also, ignore error from f.Readdirnames,
   333  	// because we don't care about reporting the error and we still
   334  	// want to process any entries found before the error.
   335  	f, err := os.Open(subdir)
   336  	if err != nil {
   337  		return
   338  	}
   339  	names, _ := f.Readdirnames(-1)
   340  	f.Close()
   341  
   342  	for _, name := range names {
   343  		// Remove only cache entries (xxxx-a and xxxx-d).
   344  		if !strings.HasSuffix(name, "-a") && !strings.HasSuffix(name, "-d") {
   345  			continue
   346  		}
   347  		entry := filepath.Join(subdir, name)
   348  		info, err := os.Stat(entry)
   349  		if err == nil && info.ModTime().Before(cutoff) {
   350  			os.Remove(entry)
   351  		}
   352  	}
   353  }
   354  
   355  // putIndexEntry adds an entry to the cache recording that executing the action
   356  // with the given id produces an output with the given output id (hash) and size.
   357  func (c *Cache) putIndexEntry(id ActionID, out OutputID, size int64, allowVerify bool) error {
   358  	// Note: We expect that for one reason or another it may happen
   359  	// that repeating an action produces a different output hash
   360  	// (for example, if the output contains a time stamp or temp dir name).
   361  	// While not ideal, this is also not a correctness problem, so we
   362  	// don't make a big deal about it. In particular, we leave the action
   363  	// cache entries writable specifically so that they can be overwritten.
   364  	//
   365  	// Setting GODEBUG=gocacheverify=1 does make a big deal:
   366  	// in verify mode we are double-checking that the cache entries
   367  	// are entirely reproducible. As just noted, this may be unrealistic
   368  	// in some cases but the check is also useful for shaking out real bugs.
   369  	entry := fmt.Sprintf("v1 %x %x %20d %20d\n", id, out, size, time.Now().UnixNano())
   370  	if verify && allowVerify {
   371  		old, err := c.get(id)
   372  		if err == nil && (old.OutputID != out || old.Size != size) {
   373  			// panic to show stack trace, so we can see what code is generating this cache entry.
   374  			msg := fmt.Sprintf("go: internal cache error: cache verify failed: id=%x changed:<<<\n%s\n>>>\nold: %x %d\nnew: %x %d", id, reverseHash(id), out, size, old.OutputID, old.Size)
   375  			panic(msg)
   376  		}
   377  	}
   378  	file := c.fileName(id, "a")
   379  
   380  	// Copy file to cache directory.
   381  	mode := os.O_WRONLY | os.O_CREATE
   382  	f, err := os.OpenFile(file, mode, 0666)
   383  	if err != nil {
   384  		return err
   385  	}
   386  	_, err = f.WriteString(entry)
   387  	if err == nil {
   388  		// Truncate the file only *after* writing it.
   389  		// (This should be a no-op, but truncate just in case of previous corruption.)
   390  		//
   391  		// This differs from os.WriteFile, which truncates to 0 *before* writing
   392  		// via os.O_TRUNC. Truncating only after writing ensures that a second write
   393  		// of the same content to the same file is idempotent, and does not — even
   394  		// temporarily! — undo the effect of the first write.
   395  		err = f.Truncate(int64(len(entry)))
   396  	}
   397  	if closeErr := f.Close(); err == nil {
   398  		err = closeErr
   399  	}
   400  	if err != nil {
   401  		// TODO(bcmills): This Remove potentially races with another go command writing to file.
   402  		// Can we eliminate it?
   403  		os.Remove(file)
   404  		return err
   405  	}
   406  	os.Chtimes(file, c.now(), c.now()) // mainly for tests
   407  
   408  	return nil
   409  }
   410  
   411  // Put stores the given output in the cache as the output for the action ID.
   412  // It may read file twice. The content of file must not change between the two passes.
   413  func (c *Cache) Put(id ActionID, file io.ReadSeeker) (OutputID, int64, error) {
   414  	return c.put(id, file, true)
   415  }
   416  
   417  // PutNoVerify is like Put but disables the verify check
   418  // when GODEBUG=goverifycache=1 is set.
   419  // It is meant for data that is OK to cache but that we expect to vary slightly from run to run,
   420  // like test output containing times and the like.
   421  func (c *Cache) PutNoVerify(id ActionID, file io.ReadSeeker) (OutputID, int64, error) {
   422  	return c.put(id, file, false)
   423  }
   424  
   425  func (c *Cache) put(id ActionID, file io.ReadSeeker, allowVerify bool) (OutputID, int64, error) {
   426  	// Compute output ID.
   427  	h := sha256.New()
   428  	if _, err := file.Seek(0, 0); err != nil {
   429  		return OutputID{}, 0, err
   430  	}
   431  	size, err := io.Copy(h, file)
   432  	if err != nil {
   433  		return OutputID{}, 0, err
   434  	}
   435  	var out OutputID
   436  	h.Sum(out[:0])
   437  
   438  	// Copy to cached output file (if not already present).
   439  	if err := c.copyFile(file, out, size); err != nil {
   440  		return out, size, err
   441  	}
   442  
   443  	// Add to cache index.
   444  	return out, size, c.putIndexEntry(id, out, size, allowVerify)
   445  }
   446  
   447  // PutBytes stores the given bytes in the cache as the output for the action ID.
   448  func (c *Cache) PutBytes(id ActionID, data []byte) error {
   449  	_, _, err := c.Put(id, bytes.NewReader(data))
   450  	return err
   451  }
   452  
   453  // copyFile copies file into the cache, expecting it to have the given
   454  // output ID and size, if that file is not present already.
   455  func (c *Cache) copyFile(file io.ReadSeeker, out OutputID, size int64) error {
   456  	name := c.fileName(out, "d")
   457  	info, err := os.Stat(name)
   458  	if err == nil && info.Size() == size {
   459  		// Check hash.
   460  		if f, err := os.Open(name); err == nil {
   461  			h := sha256.New()
   462  			io.Copy(h, f)
   463  			f.Close()
   464  			var out2 OutputID
   465  			h.Sum(out2[:0])
   466  			if out == out2 {
   467  				return nil
   468  			}
   469  		}
   470  		// Hash did not match. Fall through and rewrite file.
   471  	}
   472  
   473  	// Copy file to cache directory.
   474  	mode := os.O_RDWR | os.O_CREATE
   475  	if err == nil && info.Size() > size { // shouldn't happen but fix in case
   476  		mode |= os.O_TRUNC
   477  	}
   478  	f, err := os.OpenFile(name, mode, 0666)
   479  	if err != nil {
   480  		return err
   481  	}
   482  	defer f.Close()
   483  	if size == 0 {
   484  		// File now exists with correct size.
   485  		// Only one possible zero-length file, so contents are OK too.
   486  		// Early return here makes sure there's a "last byte" for code below.
   487  		return nil
   488  	}
   489  
   490  	// From here on, if any of the I/O writing the file fails,
   491  	// we make a best-effort attempt to truncate the file f
   492  	// before returning, to avoid leaving bad bytes in the file.
   493  
   494  	// Copy file to f, but also into h to double-check hash.
   495  	if _, err := file.Seek(0, 0); err != nil {
   496  		f.Truncate(0)
   497  		return err
   498  	}
   499  	h := sha256.New()
   500  	w := io.MultiWriter(f, h)
   501  	if _, err := io.CopyN(w, file, size-1); err != nil {
   502  		f.Truncate(0)
   503  		return err
   504  	}
   505  	// Check last byte before writing it; writing it will make the size match
   506  	// what other processes expect to find and might cause them to start
   507  	// using the file.
   508  	buf := make([]byte, 1)
   509  	if _, err := file.Read(buf); err != nil {
   510  		f.Truncate(0)
   511  		return err
   512  	}
   513  	h.Write(buf)
   514  	sum := h.Sum(nil)
   515  	if !bytes.Equal(sum, out[:]) {
   516  		f.Truncate(0)
   517  		return fmt.Errorf("file content changed underfoot")
   518  	}
   519  
   520  	// Commit cache file entry.
   521  	if _, err := f.Write(buf); err != nil {
   522  		f.Truncate(0)
   523  		return err
   524  	}
   525  	if err := f.Close(); err != nil {
   526  		// Data might not have been written,
   527  		// but file may look like it is the right size.
   528  		// To be extra careful, remove cached file.
   529  		os.Remove(name)
   530  		return err
   531  	}
   532  	os.Chtimes(name, c.now(), c.now()) // mainly for tests
   533  
   534  	return nil
   535  }
   536  

View as plain text