HDFS-4140. fuse-dfs handles open(O_TRUNC) poorly. Contributed by Colin Patrick McCabe.

git-svn-id: https://svn.apache.org/repos/asf/hadoop/common/trunk@1423257 13f79535-47bb-0310-9956-ffa450edef68

hadoop-hdfs-project/hadoop-hdfs/CHANGES.txt

@@ -622,6 +622,9 @@ Release 2.0.3-alpha - Unreleased
     HDFS-4315. DNs with multiple BPs can have BPOfferServices fail to start
     due to unsynchronized map access. (atm)
 
+    HDFS-4140. fuse-dfs handles open(O_TRUNC) poorly. (Colin Patrick McCabe
+    via atm)
+
   BREAKDOWN OF HDFS-3077 SUBTASKS
 
     HDFS-3077. Quorum-based protocol for reading and writing edit logs.

hadoop-hdfs-project/hadoop-hdfs/src/main/native/fuse-dfs/fuse_connect.c

@@ -131,7 +131,6 @@ static enum authConf discoverAuthConf(void)
 
 int fuseConnectInit(const char *nnUri, int port)
 {
-  const char *timerPeriod;
   int ret;
 
   gTimerPeriod = FUSE_CONN_DEFAULT_TIMER_PERIOD;

hadoop-hdfs-project/hadoop-hdfs/src/main/native/fuse-dfs/fuse_impls_open.c

@@ -24,12 +24,77 @@
 #include <stdio.h>
 #include <stdlib.h>
 
+static int get_hdfs_open_flags_from_info(hdfsFS fs, const char *path,
+                  int flags, int *outflags, const hdfsFileInfo *info);
+
+/**
+ * Given a set of FUSE flags, determine the libhdfs flags we need.
+ *
+ * This is complicated by two things:
+ * 1. libhdfs doesn't support O_RDWR at all;
+ * 2. when given O_WRONLY, libhdfs will truncate the file unless O_APPEND is
+ * also given.  In other words, there is an implicit O_TRUNC.
+ *
+ * Probably the next iteration of the libhdfs interface should not use the POSIX
+ * flags at all, since, as you can see, they don't really match up very closely
+ * to the POSIX meaning.  However, for the time being, this is the API.
+ *
+ * @param fs               The libhdfs object
+ * @param path             The path we're opening
+ * @param flags            The FUSE flags
+ *
+ * @return                 negative error code on failure; flags otherwise.
+ */
+static int64_t get_hdfs_open_flags(hdfsFS fs, const char *path, int flags)
+{
+  int hasContent;
+  int64_t ret;
+  hdfsFileInfo *info;
+
+  if ((flags & O_ACCMODE) == O_RDONLY) {
+    return O_RDONLY;
+  }
+  if (flags & O_TRUNC) {
+    /* If we're opening for write or read/write, O_TRUNC means we should blow
+     * away the file which is there and create our own file.
+     * */
+    return O_WRONLY;
+  }
+  info = hdfsGetPathInfo(fs, path);
+  if (info) {
+    if (info->mSize == 0) {
+      // If the file has zero length, we shouldn't feel bad about blowing it
+      // away.
+      ret = O_WRONLY;
+    } else if ((flags & O_ACCMODE) == O_RDWR) {
+      // HACK: translate O_RDWR requests into O_RDONLY if the file already
+      // exists and has non-zero length.
+      ret = O_RDONLY;
+    } else { // O_WRONLY
+      // HACK: translate O_WRONLY requests into append if the file already
+      // exists.
+      ret = O_WRONLY | O_APPEND;
+    }
+  } else { // !info
+    if (flags & O_CREAT) {
+      ret = O_WRONLY;
+    } else {
+      ret = -ENOENT;
+    }
+  }
+  if (info) {
+    hdfsFreeFileInfo(info, 1);
+  }
+  return ret;
+}
+
 int dfs_open(const char *path, struct fuse_file_info *fi)
 {
   hdfsFS fs = NULL;
   dfs_context *dfs = (dfs_context*)fuse_get_context()->private_data;
   dfs_fh *fh = NULL;
-  int mutexInit = 0, ret;
+  int mutexInit = 0, ret, flags = 0;
+  int64_t flagRet;
 
   TRACE1("open", path)
 
@@ -38,10 +103,6 @@ int dfs_open(const char *path, struct fuse_file_info *fi)
   assert('/' == *path);
   assert(dfs);
 
-  // 0x8000 is always passed in and hadoop doesn't like it, so killing it here
-  // bugbug figure out what this flag is and report problem to Hadoop JIRA
-  int flags = (fi->flags & 0x7FFF);
-
   // retrieve dfs specific data
   fh = (dfs_fh*)calloc(1, sizeof (dfs_fh));
   if (!fh) {
@@ -57,22 +118,12 @@ int dfs_open(const char *path, struct fuse_file_info *fi)
     goto error;
   }
   fs = hdfsConnGetFs(fh->conn);
-
+  flagRet = get_hdfs_open_flags(fs, path, fi->flags);
-  if (flags & O_RDWR) {
+  if (flagRet < 0) {
-    hdfsFileInfo *info = hdfsGetPathInfo(fs, path);
+    ret = -flagRet;
-    if (info == NULL) {
+    goto error;
-      // File does not exist (maybe?); interpret it as a O_WRONLY
-      // If the actual error was something else, we'll get it again when
-      // we try to open the file.
-      flags ^= O_RDWR;
-      flags |= O_WRONLY;
-    } else {
-      // File exists; open this as read only.
-      flags ^= O_RDWR;
-      flags |= O_RDONLY;
   }
-  }
+  flags = flagRet;
-
   if ((fh->hdfsFH = hdfsOpenFile(fs, path, flags,  0, 0, 0)) == NULL) {
     ERROR("Could not open file %s (errno=%d)", path, errno);
     if (errno == 0 || errno == EINTERNAL) {
@@ -91,7 +142,7 @@ int dfs_open(const char *path, struct fuse_file_info *fi)
   }
   mutexInit = 1;
 
-  if (fi->flags & O_WRONLY || fi->flags & O_CREAT) {
+  if ((flags & O_ACCMODE) == O_WRONLY) {
     fh->buf = NULL;
   } else  {
     assert(dfs->rdbuffer_size > 0);

hadoop-hdfs-project/hadoop-hdfs/src/main/native/fuse-dfs/fuse_init.c

@@ -98,7 +98,7 @@ static void dfsPrintOptions(FILE *fp, const struct options *o)
           o->attribute_timeout, o->rdbuffer_size, o->direct_io);
 }
 
-void *dfs_init(void)
+void *dfs_init(struct fuse_conn_info *conn)
 {
   int ret;
 
@@ -143,6 +143,45 @@ void *dfs_init(void)
       exit(EXIT_FAILURE);
     }
   }
+
+#ifdef FUSE_CAP_ATOMIC_O_TRUNC
+  // If FUSE_CAP_ATOMIC_O_TRUNC is set, open("foo", O_CREAT | O_TRUNC) will
+  // result in dfs_open being called with O_TRUNC.
+  //
+  // If this capability is not present, fuse will try to use multiple
+  // operation to "simulate" open(O_TRUNC).  This doesn't work very well with
+  // HDFS.
+  // Unfortunately, this capability is only implemented on Linux 2.6.29 or so.
+  // See HDFS-4140 for details.
+  if (conn->capable & FUSE_CAP_ATOMIC_O_TRUNC) {
+    conn->want |= FUSE_CAP_ATOMIC_O_TRUNC;
+  }
+#endif
+
+#ifdef FUSE_CAP_ASYNC_READ
+  // We're OK with doing reads at the same time as writes.
+  if (conn->capable & FUSE_CAP_ASYNC_READ) {
+    conn->want |= FUSE_CAP_ASYNC_READ;
+  }
+#endif
+  
+#ifdef FUSE_CAP_BIG_WRITES
+  // Yes, we can read more than 4kb at a time.  In fact, please do!
+  if (conn->capable & FUSE_CAP_BIG_WRITES) {
+    conn->want |= FUSE_CAP_BIG_WRITES;
+  }
+#endif
+
+#ifdef FUSE_CAP_DONT_MASK
+  if ((options.no_permissions) && (conn->capable & FUSE_CAP_DONT_MASK)) {
+    // If we're handing permissions ourselves, we don't want the kernel
+    // applying its own umask.  HDFS already implements its own per-user
+    // umasks!  Sadly, this only actually does something on kernels 2.6.31 and
+    // later.
+    conn->want |= FUSE_CAP_DONT_MASK;
+  }
+#endif
+
   return (void*)dfs;
 }
 

hadoop-hdfs-project/hadoop-hdfs/src/main/native/fuse-dfs/fuse_init.h

@@ -19,13 +19,15 @@
 #ifndef __FUSE_INIT_H__
 #define __FUSE_INIT_H__
 
+struct fuse_conn_info;
+
 /**
  * These are responsible for initializing connections to dfs and internal
  * data structures and then freeing them.
  * i.e., what happens on mount and unmount.
  *
  */
-void *dfs_init();
+void *dfs_init(struct fuse_conn_info *conn);
 void dfs_destroy (void *ptr);
 
 #endif

hadoop-hdfs-project/hadoop-hdfs/src/main/native/fuse-dfs/test/fuse_workload.c

@@ -16,6 +16,8 @@
  * limitations under the License.
  */
 
+#define FUSE_USE_VERSION 26
+
 #include "fuse-dfs/test/fuse_workload.h"
 #include "libhdfs/expect.h"
 #include "util/posix_util.h"
@@ -23,6 +25,7 @@
 #include <dirent.h>
 #include <errno.h>
 #include <fcntl.h>
+#include <fuse.h>
 #include <pthread.h>
 #include <stdio.h>
 #include <stdlib.h>
@@ -138,13 +141,89 @@ static int safeRead(int fd, void *buf, int c)
   return amt;
 }
 
+/* Bug: HDFS-2551.
+ * When a program writes a file, closes it, and immediately re-opens it,
+ * it might not appear to have the correct length.  This is because FUSE
+ * invokes the release() callback asynchronously.
+ *
+ * To work around this, we keep retrying until the file length is what we
+ * expect.
+ */
+static int closeWorkaroundHdfs2551(int fd, const char *path, off_t expectedSize)
+{
+  int ret, try;
+  struct stat stBuf;
+
+  RETRY_ON_EINTR_GET_ERRNO(ret, close(fd));
+  EXPECT_ZERO(ret);
+  for (try = 0; try < MAX_TRIES; try++) {
+    EXPECT_ZERO(stat(path, &stBuf));
+    EXPECT_NONZERO(S_ISREG(stBuf.st_mode));
+    if (stBuf.st_size == expectedSize) {
+      return 0;
+    }
+    sleepNoSig(1);
+  }
+  fprintf(stderr, "FUSE_WORKLOAD: error: expected file %s to have length "
+          "%lld; instead, it had length %lld\n",
+          path, (long long)expectedSize, (long long)stBuf.st_size);
+  return -EIO;
+}
+
+#ifdef FUSE_CAP_ATOMIC_O_TRUNC
+
+/**
+ * Test that we can create a file, write some contents to it, close that file,
+ * and then successfully re-open with O_TRUNC.
+ */
+static int testOpenTrunc(const char *base)
+{
+  int fd, err;
+  char path[PATH_MAX];
+  const char * const SAMPLE1 = "this is the first file that we wrote.";
+  const char * const SAMPLE2 = "this is the second file that we wrote.  "
+    "It's #2!";
+
+  snprintf(path, sizeof(path), "%s/trunc.txt", base);
+  fd = open(path, O_CREAT | O_TRUNC | O_WRONLY, 0644);
+  if (fd < 0) {
+    err = errno;
+    fprintf(stderr, "TEST_ERROR: testOpenTrunc(%s): first open "
+            "failed with error %d\n", path, err);
+    return -err;
+  }
+  EXPECT_ZERO(safeWrite(fd, SAMPLE1, strlen(SAMPLE1)));
+  EXPECT_ZERO(closeWorkaroundHdfs2551(fd, path, strlen(SAMPLE1)));
+  fd = open(path, O_CREAT | O_TRUNC | O_WRONLY, 0644);
+  if (fd < 0) {
+    err = errno;
+    fprintf(stderr, "TEST_ERROR: testOpenTrunc(%s): second open "
+            "failed with error %d\n", path, err);
+    return -err;
+  }
+  EXPECT_ZERO(safeWrite(fd, SAMPLE2, strlen(SAMPLE2)));
+  EXPECT_ZERO(closeWorkaroundHdfs2551(fd, path, strlen(SAMPLE2)));
+  return 0;
+}
+
+#else
+
+static int testOpenTrunc(const char *base)
+{
+  fprintf(stderr, "FUSE_WORKLOAD: We lack FUSE_CAP_ATOMIC_O_TRUNC support.  "
+          "Not testing open(O_TRUNC).\n");
+  return 0;
+}
+
+#endif
+
 int runFuseWorkloadImpl(const char *root, const char *pcomp,
     struct fileCtx *ctx)
 {
   char base[PATH_MAX], tmp[PATH_MAX], *tmpBuf;
   char src[PATH_MAX], dst[PATH_MAX];
   struct stat stBuf;
-  int ret, i, try;
+  int ret, i;
   struct utimbuf tbuf;
   struct statvfs stvBuf;
 
@@ -241,34 +320,9 @@ int runFuseWorkloadImpl(const char *root, const char *pcomp,
     EXPECT_ZERO(safeWrite(ctx[i].fd, ctx[i].str, ctx[i].strLen));
   }
   for (i = 0; i < NUM_FILE_CTX; i++) {
-    RETRY_ON_EINTR_GET_ERRNO(ret, close(ctx[i].fd));
+    EXPECT_ZERO(closeWorkaroundHdfs2551(ctx[i].fd, ctx[i].path, ctx[i].strLen));
-    EXPECT_ZERO(ret);
     ctx[i].fd = -1;
   }
-  for (i = 0; i < NUM_FILE_CTX; i++) {
-    /* Bug: HDFS-2551.
-     * When a program writes a file, closes it, and immediately re-opens it,
-     * it might not appear to have the correct length.  This is because FUSE
-     * invokes the release() callback asynchronously.
-     *
-     * To work around this, we keep retrying until the file length is what we
-     * expect.
-     */
-    for (try = 0; try < MAX_TRIES; try++) {
-      EXPECT_ZERO(stat(ctx[i].path, &stBuf));
-      EXPECT_NONZERO(S_ISREG(stBuf.st_mode));
-      if (ctx[i].strLen == stBuf.st_size) {
-        break;
-      }
-      sleepNoSig(1);
-    }
-    if (try == MAX_TRIES) {
-      fprintf(stderr, "FUSE_WORKLOAD: error: expected file %s to have length "
-              "%d; instead, it had length %lld\n",
-              ctx[i].path, ctx[i].strLen, (long long)stBuf.st_size);
-      return -EIO;
-    }
-  }
   for (i = 0; i < NUM_FILE_CTX; i++) {
     ctx[i].fd = open(ctx[i].path, O_RDONLY);
     if (ctx[i].fd < 0) {
@@ -308,6 +362,7 @@ int runFuseWorkloadImpl(const char *root, const char *pcomp,
   for (i = 0; i < NUM_FILE_CTX; i++) {
     free(ctx[i].path);
   }
+  EXPECT_ZERO(testOpenTrunc(base));
   EXPECT_ZERO(recursiveDelete(base));
   return 0;
 }