import (
"bytes"
+ "context"
"crypto/md5"
- "crypto/sha1"
"errors"
"fmt"
"io"
"os"
"regexp"
"runtime"
+ "runtime/pprof"
"strings"
"sync"
"sync/atomic"
"testing"
"time"
- "git.curoverse.com/arvados.git/sdk/go/arvadostest"
check "gopkg.in/check.v1"
)
type keepClientStub struct {
blocks map[string][]byte
refreshable map[string]bool
- onPut func(bufcopy []byte) // called from PutB, before acquiring lock
+ onWrite func(bufcopy []byte) // called from WriteBlock, before acquiring lock
+ authToken string // client's auth token (used for signing locators)
+ sigkey string // blob signing key
+ sigttl time.Duration // blob signing ttl
sync.RWMutex
}
return copy(p, buf[off:]), nil
}
-func (kcs *keepClientStub) PutB(p []byte) (string, int, error) {
- locator := fmt.Sprintf("%x+%d+A12345@abcde", md5.Sum(p), len(p))
- buf := make([]byte, len(p))
- copy(buf, p)
- if kcs.onPut != nil {
- kcs.onPut(buf)
+func (kcs *keepClientStub) BlockWrite(_ context.Context, opts BlockWriteOptions) (BlockWriteResponse, error) {
+ if opts.Data == nil {
+ panic("oops, stub is not made for this")
+ }
+ locator := SignLocator(fmt.Sprintf("%x+%d", md5.Sum(opts.Data), len(opts.Data)), kcs.authToken, time.Now().Add(kcs.sigttl), kcs.sigttl, []byte(kcs.sigkey))
+ buf := make([]byte, len(opts.Data))
+ copy(buf, opts.Data)
+ if kcs.onWrite != nil {
+ kcs.onWrite(buf)
+ }
+ for _, sc := range opts.StorageClasses {
+ if sc != "default" {
+ return BlockWriteResponse{}, fmt.Errorf("stub does not write storage class %q", sc)
+ }
}
kcs.Lock()
defer kcs.Unlock()
kcs.blocks[locator[:32]] = buf
- return locator, 1, nil
+ return BlockWriteResponse{Locator: locator, Replicas: 1}, nil
}
-var localOrRemoteSignature = regexp.MustCompile(`\+[AR][^+]*`)
+var reRemoteSignature = regexp.MustCompile(`\+[AR][^+]*`)
func (kcs *keepClientStub) LocalLocator(locator string) (string, error) {
+ if strings.Contains(locator, "+A") {
+ return locator, nil
+ }
kcs.Lock()
defer kcs.Unlock()
if strings.Contains(locator, "+R") {
return "", fmt.Errorf("kcs.refreshable[%q]==false", locator)
}
}
- fakeSig := fmt.Sprintf("+A%x@%x", sha1.Sum(nil), time.Now().Add(time.Hour*24*14).Unix())
- return localOrRemoteSignature.ReplaceAllLiteralString(locator, fakeSig), nil
+ locator = reRemoteSignature.ReplaceAllLiteralString(locator, "")
+ locator = SignLocator(locator, kcs.authToken, time.Now().Add(kcs.sigttl), kcs.sigttl, []byte(kcs.sigkey))
+ return locator, nil
}
type CollectionFSSuite struct {
func (s *CollectionFSSuite) SetUpTest(c *check.C) {
s.client = NewClientFromEnv()
- err := s.client.RequestAndDecode(&s.coll, "GET", "arvados/v1/collections/"+arvadostest.FooAndBarFilesInDirUUID, nil, nil)
+ err := s.client.RequestAndDecode(&s.coll, "GET", "arvados/v1/collections/"+fixtureFooAndBarFilesInDirUUID, nil, nil)
c.Assert(err, check.IsNil)
s.kc = &keepClientStub{
blocks: map[string][]byte{
"3858f62230ac3c915f300c664312c63f": []byte("foobar"),
- }}
+ },
+ sigkey: fixtureBlobSigningKey,
+ sigttl: fixtureBlobSigningTTL,
+ authToken: fixtureActiveToken,
+ }
s.fs, err = s.coll.FileSystem(s.client, s.kc)
c.Assert(err, check.IsNil)
}
c.Check(ok, check.Equals, true)
}
+func (s *CollectionFSSuite) TestUnattainableStorageClasses(c *check.C) {
+ fs, err := (&Collection{
+ StorageClassesDesired: []string{"unobtainium"},
+ }).FileSystem(s.client, s.kc)
+ c.Assert(err, check.IsNil)
+
+ f, err := fs.OpenFile("/foo", os.O_CREATE|os.O_WRONLY, 0777)
+ c.Assert(err, check.IsNil)
+ _, err = f.Write([]byte("food"))
+ c.Assert(err, check.IsNil)
+ err = f.Close()
+ c.Assert(err, check.IsNil)
+ _, err = fs.MarshalManifest(".")
+ c.Assert(err, check.ErrorMatches, `.*stub does not write storage class \"unobtainium\"`)
+}
+
func (s *CollectionFSSuite) TestColonInFilename(c *check.C) {
fs, err := (&Collection{
ManifestText: "./foo:foo 3858f62230ac3c915f300c664312c63f+3 0:3:bar:bar\n",
}
maxBlockSize = 8
- defer func() { maxBlockSize = 2 << 26 }()
+ defer func() { maxBlockSize = 1 << 26 }()
var wg sync.WaitGroup
for n := 0; n < 128; n++ {
c.Check(err, check.ErrorMatches, `invalid flag.*`)
}
-func (s *CollectionFSSuite) TestFlushFullBlocks(c *check.C) {
- defer func(wab, mbs int) {
- writeAheadBlocks = wab
+func (s *CollectionFSSuite) TestFlushFullBlocksWritingLongFile(c *check.C) {
+ defer func(cw, mbs int) {
+ concurrentWriters = cw
maxBlockSize = mbs
- }(writeAheadBlocks, maxBlockSize)
- writeAheadBlocks = 2
+ }(concurrentWriters, maxBlockSize)
+ concurrentWriters = 2
maxBlockSize = 1024
proceed := make(chan struct{})
var started, concurrent int32
blk2done := false
- s.kc.onPut = func([]byte) {
+ s.kc.onWrite = func([]byte) {
atomic.AddInt32(&concurrent, 1)
switch atomic.AddInt32(&started, 1) {
case 1:
default:
time.Sleep(time.Millisecond)
}
- c.Check(atomic.AddInt32(&concurrent, -1) < int32(writeAheadBlocks), check.Equals, true)
+ c.Check(atomic.AddInt32(&concurrent, -1) < int32(concurrentWriters), check.Equals, true)
}
fs, err := (&Collection{}).FileSystem(s.client, s.kc)
c.Check(currentMemExtents(), check.HasLen, 0)
}
+// Ensure blocks get flushed to disk if a lot of data is written to
+// small files/directories without calling sync().
+//
+// Write four 512KiB files into each of 256 top-level dirs (total
+// 512MiB), calling Flush() every 8 dirs. Ensure memory usage never
+// exceeds 24MiB (4 concurrentWriters * 2MiB + 8 unflushed dirs *
+// 2MiB).
+func (s *CollectionFSSuite) TestFlushAll(c *check.C) {
+ fs, err := (&Collection{}).FileSystem(s.client, s.kc)
+ c.Assert(err, check.IsNil)
+
+ s.kc.onWrite = func([]byte) {
+ // discard flushed data -- otherwise the stub will use
+ // unlimited memory
+ time.Sleep(time.Millisecond)
+ s.kc.Lock()
+ defer s.kc.Unlock()
+ s.kc.blocks = map[string][]byte{}
+ }
+ for i := 0; i < 256; i++ {
+ buf := bytes.NewBuffer(make([]byte, 524288))
+ fmt.Fprintf(buf, "test file in dir%d", i)
+
+ dir := fmt.Sprintf("dir%d", i)
+ fs.Mkdir(dir, 0755)
+ for j := 0; j < 2; j++ {
+ f, err := fs.OpenFile(fmt.Sprintf("%s/file%d", dir, j), os.O_WRONLY|os.O_CREATE, 0)
+ c.Assert(err, check.IsNil)
+ defer f.Close()
+ _, err = io.Copy(f, buf)
+ c.Assert(err, check.IsNil)
+ }
+
+ if i%8 == 0 {
+ fs.Flush("", true)
+ }
+
+ size := fs.MemorySize()
+ if !c.Check(size <= 1<<24, check.Equals, true) {
+ c.Logf("at dir%d fs.MemorySize()=%d", i, size)
+ return
+ }
+ }
+}
+
+// Ensure short blocks at the end of a stream don't get flushed by
+// Flush(false).
+//
+// Write 67x 1MiB files to each of 8 dirs, and check that 8 full 64MiB
+// blocks have been flushed while 8x 3MiB is still buffered in memory.
+func (s *CollectionFSSuite) TestFlushFullBlocksOnly(c *check.C) {
+ fs, err := (&Collection{}).FileSystem(s.client, s.kc)
+ c.Assert(err, check.IsNil)
+
+ var flushed int64
+ s.kc.onWrite = func(p []byte) {
+ atomic.AddInt64(&flushed, int64(len(p)))
+ }
+
+ nDirs := int64(8)
+ megabyte := make([]byte, 1<<20)
+ for i := int64(0); i < nDirs; i++ {
+ dir := fmt.Sprintf("dir%d", i)
+ fs.Mkdir(dir, 0755)
+ for j := 0; j < 67; j++ {
+ f, err := fs.OpenFile(fmt.Sprintf("%s/file%d", dir, j), os.O_WRONLY|os.O_CREATE, 0)
+ c.Assert(err, check.IsNil)
+ defer f.Close()
+ _, err = f.Write(megabyte)
+ c.Assert(err, check.IsNil)
+ }
+ }
+ c.Check(fs.MemorySize(), check.Equals, int64(nDirs*67<<20))
+ c.Check(flushed, check.Equals, int64(0))
+
+ waitForFlush := func(expectUnflushed, expectFlushed int64) {
+ for deadline := time.Now().Add(5 * time.Second); fs.MemorySize() > expectUnflushed && time.Now().Before(deadline); time.Sleep(10 * time.Millisecond) {
+ }
+ c.Check(fs.MemorySize(), check.Equals, expectUnflushed)
+ c.Check(flushed, check.Equals, expectFlushed)
+ }
+
+ // Nothing flushed yet
+ waitForFlush((nDirs*67)<<20, 0)
+
+ // Flushing a non-empty dir "/" is non-recursive and there are
+ // no top-level files, so this has no effect
+ fs.Flush("/", false)
+ waitForFlush((nDirs*67)<<20, 0)
+
+ // Flush the full block in dir0
+ fs.Flush("dir0", false)
+ waitForFlush((nDirs*67-64)<<20, 64<<20)
+
+ err = fs.Flush("dir-does-not-exist", false)
+ c.Check(err, check.NotNil)
+
+ // Flush full blocks in all dirs
+ fs.Flush("", false)
+ waitForFlush(nDirs*3<<20, nDirs*64<<20)
+
+ // Flush non-full blocks, too
+ fs.Flush("", true)
+ waitForFlush(0, nDirs*67<<20)
+}
+
+// Even when writing lots of files/dirs from different goroutines, as
+// long as Flush(dir,false) is called after writing each file,
+// unflushed data should be limited to one full block per
+// concurrentWriter, plus one nearly-full block at the end of each
+// dir/stream.
+func (s *CollectionFSSuite) TestMaxUnflushed(c *check.C) {
+ nDirs := int64(8)
+ maxUnflushed := (int64(concurrentWriters) + nDirs) << 26
+
+ fs, err := (&Collection{}).FileSystem(s.client, s.kc)
+ c.Assert(err, check.IsNil)
+
+ release := make(chan struct{})
+ timeout := make(chan struct{})
+ time.AfterFunc(10*time.Second, func() { close(timeout) })
+ var putCount, concurrency int64
+ var unflushed int64
+ s.kc.onWrite = func(p []byte) {
+ defer atomic.AddInt64(&unflushed, -int64(len(p)))
+ cur := atomic.AddInt64(&concurrency, 1)
+ defer atomic.AddInt64(&concurrency, -1)
+ pc := atomic.AddInt64(&putCount, 1)
+ if pc < int64(concurrentWriters) {
+ // Block until we reach concurrentWriters, to
+ // make sure we're really accepting concurrent
+ // writes.
+ select {
+ case <-release:
+ case <-timeout:
+ c.Error("timeout")
+ }
+ } else if pc == int64(concurrentWriters) {
+ // Unblock the first N-1 PUT reqs.
+ close(release)
+ }
+ c.Assert(cur <= int64(concurrentWriters), check.Equals, true)
+ c.Assert(atomic.LoadInt64(&unflushed) <= maxUnflushed, check.Equals, true)
+ }
+
+ var owg sync.WaitGroup
+ megabyte := make([]byte, 1<<20)
+ for i := int64(0); i < nDirs; i++ {
+ dir := fmt.Sprintf("dir%d", i)
+ fs.Mkdir(dir, 0755)
+ owg.Add(1)
+ go func() {
+ defer owg.Done()
+ defer fs.Flush(dir, true)
+ var iwg sync.WaitGroup
+ defer iwg.Wait()
+ for j := 0; j < 67; j++ {
+ iwg.Add(1)
+ go func(j int) {
+ defer iwg.Done()
+ f, err := fs.OpenFile(fmt.Sprintf("%s/file%d", dir, j), os.O_WRONLY|os.O_CREATE, 0)
+ c.Assert(err, check.IsNil)
+ defer f.Close()
+ n, err := f.Write(megabyte)
+ c.Assert(err, check.IsNil)
+ atomic.AddInt64(&unflushed, int64(n))
+ fs.Flush(dir, false)
+ }(j)
+ }
+ }()
+ }
+ owg.Wait()
+ fs.Flush("", true)
+}
+
+func (s *CollectionFSSuite) TestFlushStress(c *check.C) {
+ done := false
+ defer func() { done = true }()
+ time.AfterFunc(10*time.Second, func() {
+ if !done {
+ pprof.Lookup("goroutine").WriteTo(os.Stderr, 1)
+ panic("timeout")
+ }
+ })
+
+ wrote := 0
+ s.kc.onWrite = func(p []byte) {
+ s.kc.Lock()
+ s.kc.blocks = map[string][]byte{}
+ wrote++
+ defer c.Logf("wrote block %d, %d bytes", wrote, len(p))
+ s.kc.Unlock()
+ time.Sleep(20 * time.Millisecond)
+ }
+
+ fs, err := (&Collection{}).FileSystem(s.client, s.kc)
+ c.Assert(err, check.IsNil)
+
+ data := make([]byte, 1<<20)
+ for i := 0; i < 3; i++ {
+ dir := fmt.Sprintf("dir%d", i)
+ fs.Mkdir(dir, 0755)
+ for j := 0; j < 200; j++ {
+ data[0] = byte(j)
+ f, err := fs.OpenFile(fmt.Sprintf("%s/file%d", dir, j), os.O_WRONLY|os.O_CREATE, 0)
+ c.Assert(err, check.IsNil)
+ _, err = f.Write(data)
+ c.Assert(err, check.IsNil)
+ f.Close()
+ fs.Flush(dir, false)
+ }
+ _, err := fs.MarshalManifest(".")
+ c.Check(err, check.IsNil)
+ }
+}
+
+func (s *CollectionFSSuite) TestFlushShort(c *check.C) {
+ s.kc.onWrite = func([]byte) {
+ s.kc.Lock()
+ s.kc.blocks = map[string][]byte{}
+ s.kc.Unlock()
+ }
+ fs, err := (&Collection{}).FileSystem(s.client, s.kc)
+ c.Assert(err, check.IsNil)
+ for _, blocksize := range []int{8, 1000000} {
+ dir := fmt.Sprintf("dir%d", blocksize)
+ err = fs.Mkdir(dir, 0755)
+ c.Assert(err, check.IsNil)
+ data := make([]byte, blocksize)
+ for i := 0; i < 100; i++ {
+ f, err := fs.OpenFile(fmt.Sprintf("%s/file%d", dir, i), os.O_WRONLY|os.O_CREATE, 0)
+ c.Assert(err, check.IsNil)
+ _, err = f.Write(data)
+ c.Assert(err, check.IsNil)
+ f.Close()
+ fs.Flush(dir, false)
+ }
+ fs.Flush(dir, true)
+ _, err := fs.MarshalManifest(".")
+ c.Check(err, check.IsNil)
+ }
+}
+
func (s *CollectionFSSuite) TestBrokenManifests(c *check.C) {
for _, txt := range []string{
"\n",
}
}
+var bigmanifest = func() string {
+ var buf bytes.Buffer
+ for i := 0; i < 2000; i++ {
+ fmt.Fprintf(&buf, "./dir%d", i)
+ for i := 0; i < 100; i++ {
+ fmt.Fprintf(&buf, " d41d8cd98f00b204e9800998ecf8427e+99999")
+ }
+ for i := 0; i < 2000; i++ {
+ fmt.Fprintf(&buf, " 1200000:300000:file%d", i)
+ }
+ fmt.Fprintf(&buf, "\n")
+ }
+ return buf.String()
+}()
+
+func (s *CollectionFSSuite) BenchmarkParseManifest(c *check.C) {
+ DebugLocksPanicMode = false
+ c.Logf("test manifest is %d bytes", len(bigmanifest))
+ for i := 0; i < c.N; i++ {
+ fs, err := (&Collection{ManifestText: bigmanifest}).FileSystem(s.client, s.kc)
+ c.Check(err, check.IsNil)
+ c.Check(fs, check.NotNil)
+ }
+}
+
func (s *CollectionFSSuite) checkMemSize(c *check.C, f File) {
fn := f.(*filehandle).inode.(*filenode)
var memsize int64
projects / arvados.git / blobdiff