[PATCH v5 0/5] Fix opendir regression on some FS

[PATCH v5 0/5] Fix opendir regression on some FS

[Adhemerval Zanella]

Some filesystem might return a non-representable d_off on getdents
call, even if there are few entries in the directories (for instance
ext4, which returns a hash).  This trigger issues with non-LFS readdir
when it finds the first non representable entry, and also for LFS
interface when used along with telldir (which return a 'long int',
while d_off is potentially off64_t).

This patch changes non-LFS readdir so it ignore non representable
entries.  The opendir now uses getdents64, even for non-LFS and uses
a translation buffer to return the getdents entry from the getdents64
buffer.

The second part fixes the broken telldir on non LP64 ABIs, where
returning d_off as stream position might overflow.  For these ABIs,
readdir updates an internal list that maps the DIR object off64_t
offsets to the returned long int (the function return value).  The
seekdir will then set the correct offset from the internal list using
the telldir as the list key.

I have checked on x86_64-linux-gnu, i686-linux-gnu, and
arm-linux-gnueabihf.

Changes from v4:
  * Allocate the telldir map an readdir, thus avoid telldir failure.
  * The translation buffer now uses a fixed size struct.
  * Fixed the condition to use the long to off64_t map.

Changes from v3:
  * Rebase against master.
  * Fixed a missing lock unlock on telldir.
  * Renamed tst-seekdir2 to tst-opendir-nolfs and check the opendir
    result against getdents64.


Adhemerval Zanella (5):
  linux: Use getdents64 on non-LFS readdir
  linux: Set internal DIR filepos as off64_t (BZ #23960, BZ #24050)
  linux: Add __readdir64_unlocked
  linux: Add __old_readdir64_unlocked
  linux: Use getdents64 on readdir64 compat implementation

 dirent/tst-scandir.c                        |   6 +-
 dirent/tst-seekdir.c                        |   8 +
 include/dirent.h                            |   3 +-
 sysdeps/unix/sysv/linux/Makefile            |   1 +
 sysdeps/unix/sysv/linux/alpha/bits/dirent.h |   3 +
 sysdeps/unix/sysv/linux/bits/dirent.h       |   4 +
 sysdeps/unix/sysv/linux/closedir.c          |   4 +
 sysdeps/unix/sysv/linux/dirstream.h         |  22 ++-
 sysdeps/unix/sysv/linux/getdents64.c        |  93 ------------
 sysdeps/unix/sysv/linux/olddirent.h         |   4 +-
 sysdeps/unix/sysv/linux/opendir.c           |   3 +
 sysdeps/unix/sysv/linux/readdir.c           |  88 ++++++++---
 sysdeps/unix/sysv/linux/readdir64.c         |  92 ++++++++---
 sysdeps/unix/sysv/linux/readdir64_r.c       | 159 +++++---------------
 sysdeps/unix/sysv/linux/rewinddir.c         |   5 +
 sysdeps/unix/sysv/linux/seekdir.c           |  35 ++++-
 sysdeps/unix/sysv/linux/telldir.c           |  35 +++++
 sysdeps/unix/sysv/linux/telldir.h           |  65 ++++++++
 sysdeps/unix/sysv/linux/tst-opendir-nolfs.c | 146 ++++++++++++++++++
 19 files changed, 499 insertions(+), 277 deletions(-)
 create mode 100644 sysdeps/unix/sysv/linux/telldir.h
 create mode 100644 sysdeps/unix/sysv/linux/tst-opendir-nolfs.c

-- 
2.34.1

[PATCH v5 1/5] linux: Use getdents64 on non-LFS readdir

[Adhemerval Zanella]

The non-LFS opendir reserves a translation entry to be used to return
the entry and the dirent64 struct is translated to the temporary buffer
on each readdir call.

Entries that overflow d_off/d_ino and the buffer reallocation failure
(in case of large d_name) are ignored.

Checked on x86_64-linux-gnu and i686-linux-gnu.
---
 dirent/tst-scandir.c                |  6 ++-
 include/dirent.h                    |  2 +-
 sysdeps/unix/sysv/linux/dirstream.h |  5 ++
 sysdeps/unix/sysv/linux/readdir.c   | 83 +++++++++++++++++++----------
 4 files changed, 67 insertions(+), 29 deletions(-)

diff --git a/dirent/tst-scandir.c b/dirent/tst-scandir.c
index 8d87d4dd74..7bc666449e 100644
--- a/dirent/tst-scandir.c
+++ b/dirent/tst-scandir.c
@@ -155,8 +155,12 @@ do_test (void)
     }
   if (n != 6)
     {
+      /* Non-lfs opendir skips entries that can not be represented (for
+	 instance if d_off is not an offset but rather an internal filesystem
+	 representation.  For this case there is no point in continue the
+	 testcase.  */
       printf ("scandir returned %d entries instead of 6\n", n);
-      return 1;
+      return EXIT_UNSUPPORTED;
     }
 
   struct
diff --git a/include/dirent.h b/include/dirent.h
index d7567f5e86..17827176ba 100644
--- a/include/dirent.h
+++ b/include/dirent.h
@@ -1,8 +1,8 @@
 #ifndef _DIRENT_H
+# include <dirent/dirent.h>
 # ifndef _ISOMAC
 #  include <dirstream.h>
 # endif
-# include <dirent/dirent.h>
 # ifndef _ISOMAC
 # include <sys/stat.h>
 # include <stdbool.h>
diff --git a/sysdeps/unix/sysv/linux/dirstream.h b/sysdeps/unix/sysv/linux/dirstream.h
index 3cb313b410..adcf8234f1 100644
--- a/sysdeps/unix/sysv/linux/dirstream.h
+++ b/sysdeps/unix/sysv/linux/dirstream.h
@@ -18,6 +18,7 @@
 #ifndef	_DIRSTREAM_H
 #define	_DIRSTREAM_H	1
 
+#include <dirent.h>
 #include <sys/types.h>
 
 #include <libc-lock.h>
@@ -41,6 +42,10 @@ struct __dirstream
 
     int errcode;		/* Delayed error code.  */
 
+#if !defined __OFF_T_MATCHES_OFF64_T || !defined __INO_T_MATCHES_INO64_T
+    struct dirent tdp;
+#endif
+
     /* Directory block.  We must make sure that this block starts
        at an address that is aligned adequately enough to store
        dirent entries.  Using the alignment of "void *" is not
diff --git a/sysdeps/unix/sysv/linux/readdir.c b/sysdeps/unix/sysv/linux/readdir.c
index 4a4c00ea07..cd0ccaf33a 100644
--- a/sysdeps/unix/sysv/linux/readdir.c
+++ b/sysdeps/unix/sysv/linux/readdir.c
@@ -21,42 +21,71 @@
 #if !_DIRENT_MATCHES_DIRENT64
 #include <dirstream.h>
 
+/* Translate the DP64 entry to the non-LFS one in the translation entry
+   at dirstream DS.  Return true is the translation was possible or
+   false if either an internal field can not be represented in the non-LFS
+   entry or if the name is too long.  */
+static bool
+dirstream_entry (struct __dirstream *ds, const struct dirent64 *dp64)
+{
+  /* Check for overflow.  */
+  if (!in_off_t_range (dp64->d_off) || !in_ino_t_range (dp64->d_ino))
+    return false;
+
+  /* And if name is too large.  */
+  if (dp64->d_reclen - offsetof (struct dirent64, d_name) > NAME_MAX)
+    return false;
+
+  ds->filepos = dp64->d_off;
+
+  ds->tdp.d_off = dp64->d_off;
+  ds->tdp.d_ino = dp64->d_ino;
+  ds->tdp.d_reclen = sizeof (struct dirent)
+    + dp64->d_reclen - offsetof (struct dirent64, d_name);
+  ds->tdp.d_type = dp64->d_type;
+  memcpy (ds->tdp.d_name, dp64->d_name,
+	  dp64->d_reclen - offsetof (struct dirent64, d_name));
+
+  return true;
+}
+
 /* Read a directory entry from DIRP.  */
 struct dirent *
 __readdir_unlocked (DIR *dirp)
 {
-  struct dirent *dp;
   int saved_errno = errno;
 
-  if (dirp->offset >= dirp->size)
+  while (1)
     {
-      /* We've emptied out our buffer.  Refill it.  */
-
-      size_t maxread = dirp->allocation;
-      ssize_t bytes;
-
-      bytes = __getdents (dirp->fd, dirp->data, maxread);
-      if (bytes <= 0)
+      if (dirp->offset >= dirp->size)
 	{
-	  /* Linux may fail with ENOENT on some file systems if the
-	     directory inode is marked as dead (deleted).  POSIX
-	     treats this as a regular end-of-directory condition, so
-	     do not set errno in that case, to indicate success.  */
-	  if (bytes == 0 || errno == ENOENT)
-	    __set_errno (saved_errno);
-	  return NULL;
-	}
-      dirp->size = (size_t) bytes;
-
-      /* Reset the offset into the buffer.  */
-      dirp->offset = 0;
+	  /* We've emptied out our buffer.  Refill it.  */
+	  ssize_t bytes = __getdents64 (dirp->fd, dirp->data,
+					dirp->allocation);
+	  if (bytes <= 0)
+	    {
+	      /* Linux may fail with ENOENT on some file systems if the
+		 directory inode is marked as dead (deleted).  POSIX
+		 treats this as a regular end-of-directory condition, so
+		 do not set errno in that case, to indicate success.  */
+	      if (bytes < 0 && errno == ENOENT)
+		__set_errno (saved_errno);
+	      return NULL;
+	    }
+	  dirp->size = bytes;
+
+ 	  /* Reset the offset into the buffer.  */
+	  dirp->offset = 0;
+ 	}
+
+      struct dirent64 *dp64 = (struct dirent64 *) &dirp->data[dirp->offset];
+      dirp->offset += dp64->d_reclen;
+
+      /* Skip entries which might overflow d_off/d_ino or if the translation
+	 buffer can not be resized.  */
+      if (dirstream_entry (dirp, dp64))
+	return &dirp->tdp;
     }
-
-  dp = (struct dirent *) &dirp->data[dirp->offset];
-  dirp->offset += dp->d_reclen;
-  dirp->filepos = dp->d_off;
-
-  return dp;
 }
 
 struct dirent *
-- 
2.34.1

[PATCH v5 2/5] linux: Set internal DIR filepos as off64_t (BZ #23960, BZ #24050)

[Adhemerval Zanella]

It allows to obtain the expected entry offset on telldir and set
it correctly on seekdir on platforms where long int is smaller
than off64_t.

On such cases opendir creates a map entry between the DIR d_off
offset and the returned long int (the telldir return value).
seekdir will then set the correct offset from the internal list
using the telldir as the list key.

It also removes the overflow check on readdir and the returned value
will be truncated by the non-LFS off_t size.  As Joseph has noted
in BZ #23960 comment #22, d_off is an opaque value and since
telldir/seekdir works regardless of the returned dirent d_off value.

Finally it removes the requirement to check for overflow values on
telldir (BZ #24050).

Checked on x86_64-linux-gnu, i686-linux-gnu, powerpc-linux-gnu,
and arm-linux-gnueabihf.
---
 dirent/tst-seekdir.c                        |   8 ++
 sysdeps/unix/sysv/linux/Makefile            |   1 +
 sysdeps/unix/sysv/linux/alpha/bits/dirent.h |   3 +
 sysdeps/unix/sysv/linux/bits/dirent.h       |   4 +
 sysdeps/unix/sysv/linux/closedir.c          |   4 +
 sysdeps/unix/sysv/linux/dirstream.h         |   6 +-
 sysdeps/unix/sysv/linux/opendir.c           |   3 +
 sysdeps/unix/sysv/linux/readdir.c           |  11 +-
 sysdeps/unix/sysv/linux/rewinddir.c         |   5 +
 sysdeps/unix/sysv/linux/seekdir.c           |  35 ++++-
 sysdeps/unix/sysv/linux/telldir.c           |  35 +++++
 sysdeps/unix/sysv/linux/telldir.h           |  65 +++++++++
 sysdeps/unix/sysv/linux/tst-opendir-nolfs.c | 146 ++++++++++++++++++++
 13 files changed, 319 insertions(+), 7 deletions(-)
 create mode 100644 sysdeps/unix/sysv/linux/telldir.h
 create mode 100644 sysdeps/unix/sysv/linux/tst-opendir-nolfs.c

diff --git a/dirent/tst-seekdir.c b/dirent/tst-seekdir.c
index dcdd699b09..187eda7584 100644
--- a/dirent/tst-seekdir.c
+++ b/dirent/tst-seekdir.c
@@ -41,6 +41,14 @@ do_test (void)
       if (i == 400)
 	break;
     }
+  if (i < 3)
+    {
+      /* Non-lfs opendir skips entries that can not be represented (for
+	 instance if d_off is not an offset but rather an internal filesystem
+	 representation.  For this case there is no point in continue the
+	 testcase.  */
+      return 77;
+    }
 
   printf ("going back past 4-th entry...\n");
 
diff --git a/sysdeps/unix/sysv/linux/Makefile b/sysdeps/unix/sysv/linux/Makefile
index f298878e8f..33708f910b 100644
--- a/sysdeps/unix/sysv/linux/Makefile
+++ b/sysdeps/unix/sysv/linux/Makefile
@@ -465,6 +465,7 @@ ifeq ($(subdir),dirent)
 sysdep_routines += getdirentries getdirentries64
 tests += \
   tst-getdents64 \
+  tst-opendir-nolfs \
   tst-readdir64-compat \
   # tests
 endif # $(subdir) == dirent
diff --git a/sysdeps/unix/sysv/linux/alpha/bits/dirent.h b/sysdeps/unix/sysv/linux/alpha/bits/dirent.h
index c8a0cfe93f..586d75586a 100644
--- a/sysdeps/unix/sysv/linux/alpha/bits/dirent.h
+++ b/sysdeps/unix/sysv/linux/alpha/bits/dirent.h
@@ -54,4 +54,7 @@ struct dirent64
 /* Inform libc code that these two types are effectively identical.  */
 #define _DIRENT_MATCHES_DIRENT64	1
 
+/* alpha 'long int' is enough to handle off64_t.  */
+#define _DIRENT_OFFSET_TRANSLATION	0
+
 #endif /* bits/dirent.h */
diff --git a/sysdeps/unix/sysv/linux/bits/dirent.h b/sysdeps/unix/sysv/linux/bits/dirent.h
index ab34d986ff..bb02dcb70a 100644
--- a/sysdeps/unix/sysv/linux/bits/dirent.h
+++ b/sysdeps/unix/sysv/linux/bits/dirent.h
@@ -57,3 +57,7 @@ struct dirent64
 #else
 # define _DIRENT_MATCHES_DIRENT64	0
 #endif
+
+/* The telldir function returns long int, which may not be large enough to
+   store off64_t values.  In this case, translation is required.  */
+#define _DIRENT_OFFSET_TRANSLATION (LONG_WIDTH < 64)
diff --git a/sysdeps/unix/sysv/linux/closedir.c b/sysdeps/unix/sysv/linux/closedir.c
index f1c2608642..9585a6ca3a 100644
--- a/sysdeps/unix/sysv/linux/closedir.c
+++ b/sysdeps/unix/sysv/linux/closedir.c
@@ -47,6 +47,10 @@ __closedir (DIR *dirp)
   __libc_lock_fini (dirp->lock);
 #endif
 
+#if _DIRENT_OFFSET_TRANSLATION
+  dirstream_loc_clear (&dirp->locs);
+#endif
+
   free ((void *) dirp);
 
   return __close_nocancel (fd);
diff --git a/sysdeps/unix/sysv/linux/dirstream.h b/sysdeps/unix/sysv/linux/dirstream.h
index adcf8234f1..8f58a1c3a6 100644
--- a/sysdeps/unix/sysv/linux/dirstream.h
+++ b/sysdeps/unix/sysv/linux/dirstream.h
@@ -22,6 +22,7 @@
 #include <sys/types.h>
 
 #include <libc-lock.h>
+#include <telldir.h>
 
 /* Directory stream type.
 
@@ -38,13 +39,16 @@ struct __dirstream
     size_t size;		/* Total valid data in the block.  */
     size_t offset;		/* Current offset into the block.  */
 
-    off_t filepos;		/* Position of next entry to read.  */
+    off64_t filepos;		/* Position of next entry to read.  */
 
     int errcode;		/* Delayed error code.  */
 
 #if !defined __OFF_T_MATCHES_OFF64_T || !defined __INO_T_MATCHES_INO64_T
     struct dirent tdp;
 #endif
+#if _DIRENT_OFFSET_TRANSLATION
+    struct dirstream_loc_t locs; /* off64_t to long int map for telldir.  */
+#endif
 
     /* Directory block.  We must make sure that this block starts
        at an address that is aligned adequately enough to store
diff --git a/sysdeps/unix/sysv/linux/opendir.c b/sysdeps/unix/sysv/linux/opendir.c
index 4336196a4d..3e2caabb9d 100644
--- a/sysdeps/unix/sysv/linux/opendir.c
+++ b/sysdeps/unix/sysv/linux/opendir.c
@@ -129,6 +129,9 @@ __alloc_dir (int fd, bool close_fd, int flags,
   dirp->offset = 0;
   dirp->filepos = 0;
   dirp->errcode = 0;
+#if _DIRENT_OFFSET_TRANSLATION
+  dirstream_loc_init (&dirp->locs);
+#endif
 
   return dirp;
 }
diff --git a/sysdeps/unix/sysv/linux/readdir.c b/sysdeps/unix/sysv/linux/readdir.c
index cd0ccaf33a..7a7f484c36 100644
--- a/sysdeps/unix/sysv/linux/readdir.c
+++ b/sysdeps/unix/sysv/linux/readdir.c
@@ -36,6 +36,15 @@ dirstream_entry (struct __dirstream *ds, const struct dirent64 *dp64)
   if (dp64->d_reclen - offsetof (struct dirent64, d_name) > NAME_MAX)
     return false;
 
+  /* telldir can not return an error, so preallocate the map if the entry can
+     not be packed directly.  */
+  if (telldir_need_dirstream (dp64->d_off))
+    {
+      dirstream_loc_add (&ds->locs, dp64->d_off);
+      if (dirstream_loc_has_failed (&ds->locs))
+	return false;
+    }
+
   ds->filepos = dp64->d_off;
 
   ds->tdp.d_off = dp64->d_off;
@@ -76,7 +85,7 @@ __readdir_unlocked (DIR *dirp)
 
  	  /* Reset the offset into the buffer.  */
 	  dirp->offset = 0;
- 	}
+	}
 
       struct dirent64 *dp64 = (struct dirent64 *) &dirp->data[dirp->offset];
       dirp->offset += dp64->d_reclen;
diff --git a/sysdeps/unix/sysv/linux/rewinddir.c b/sysdeps/unix/sysv/linux/rewinddir.c
index c0fb7aa765..1b158a584f 100644
--- a/sysdeps/unix/sysv/linux/rewinddir.c
+++ b/sysdeps/unix/sysv/linux/rewinddir.c
@@ -33,6 +33,11 @@ __rewinddir (DIR *dirp)
   dirp->offset = 0;
   dirp->size = 0;
   dirp->errcode = 0;
+
+#ifndef __LP64__
+  dirstream_loc_clear (&dirp->locs);
+#endif
+
 #if IS_IN (libc)
   __libc_lock_unlock (dirp->lock);
 #endif
diff --git a/sysdeps/unix/sysv/linux/seekdir.c b/sysdeps/unix/sysv/linux/seekdir.c
index 939ccc4447..30cce691a4 100644
--- a/sysdeps/unix/sysv/linux/seekdir.c
+++ b/sysdeps/unix/sysv/linux/seekdir.c
@@ -22,14 +22,39 @@
 #include <dirstream.h>
 
 /* Seek to position POS in DIRP.  */
-/* XXX should be __seekdir ? */
 void
 seekdir (DIR *dirp, long int pos)
 {
+  off64_t filepos;
+
   __libc_lock_lock (dirp->lock);
-  (void) __lseek (dirp->fd, pos, SEEK_SET);
-  dirp->size = 0;
-  dirp->offset = 0;
-  dirp->filepos = pos;
+
+#if _DIRENT_OFFSET_TRANSLATION
+  union dirstream_packed dsp = { .l = pos };
+  if (dsp.p.is_packed == 1)
+    filepos = dsp.p.info;
+  else
+    {
+      size_t index = dsp.p.info;
+
+      if (index >= dirstream_loc_size (&dirp->locs))
+	{
+	  __libc_lock_unlock (dirp->lock);
+	  return;
+	}
+      filepos = *dirstream_loc_at (&dirp->locs, index);
+    }
+#else
+  filepos = pos;
+#endif
+
+  if (dirp->filepos != filepos)
+    {
+      __lseek64 (dirp->fd, filepos, SEEK_SET);
+      dirp->filepos = filepos;
+      dirp->offset = 0;
+      dirp->size = 0;
+    }
+
   __libc_lock_unlock (dirp->lock);
 }
diff --git a/sysdeps/unix/sysv/linux/telldir.c b/sysdeps/unix/sysv/linux/telldir.c
index 1e5c129e9f..c3ef14f3da 100644
--- a/sysdeps/unix/sysv/linux/telldir.c
+++ b/sysdeps/unix/sysv/linux/telldir.c
@@ -15,9 +15,11 @@
    License along with the GNU C Library; if not, see
    <https://www.gnu.org/licenses/>.  */
 
+#include <assert.h>
 #include <dirent.h>
 
 #include <dirstream.h>
+#include <telldir.h>
 
 /* Return the current position of DIRP.  */
 long int
@@ -26,7 +28,40 @@ telldir (DIR *dirp)
   long int ret;
 
   __libc_lock_lock (dirp->lock);
+
+#if _DIRENT_OFFSET_TRANSLATION
+  /* If the directory position fits in the packet structure, returns it.
+     Otherwise, check if the position is already been recorded in the
+     dynamic array.  If not, add the new record.  */
+
+  union dirstream_packed dsp;
+
+  if (!telldir_need_dirstream (dirp->filepos))
+    {
+      dsp.p.is_packed = 1;
+      dsp.p.info = dirp->filepos;
+    }
+  else
+    {
+      dsp.l = -1;
+
+      size_t i;
+      for (i = 0; i < dirstream_loc_size (&dirp->locs); i++)
+	if (*dirstream_loc_at (&dirp->locs, i) == dirp->filepos)
+	  break;
+      /* It should be pre-allocated on readdir.  */
+      assert (i != dirstream_loc_size (&dirp->locs));
+
+      dsp.p.is_packed = 0;
+      /* This assignment might overflow, however most likely ENOME would
+	 happen long before.  */
+      dsp.p.info = i;
+    }
+
+  ret = dsp.l;
+#else
   ret = dirp->filepos;
+#endif
   __libc_lock_unlock (dirp->lock);
 
   return ret;
diff --git a/sysdeps/unix/sysv/linux/telldir.h b/sysdeps/unix/sysv/linux/telldir.h
new file mode 100644
index 0000000000..758bcb0eb3
--- /dev/null
+++ b/sysdeps/unix/sysv/linux/telldir.h
@@ -0,0 +1,65 @@
+/* Linux internal telldir definitions.
+   Copyright (C) 2023 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <https://www.gnu.org/licenses/>.  */
+
+#ifndef _TELLDIR_H
+#define _TELLDIR_H 1
+
+#if _DIRENT_OFFSET_TRANSLATION
+/* On platforms where 'long int' is smaller than 'off64_t' this is how the
+   returned value is encoded and returned by 'telldir'.  If the directory
+   offset can be enconded in 31 bits it is returned in the 'info' member
+   with 'is_packed' set to 1.
+
+   Otherwise, the 'info' member describes an index in a dynamic array at
+   'DIR' structure.  */
+
+union dirstream_packed
+{
+  long int l;
+  struct
+  {
+    unsigned long int is_packed:1;
+    unsigned long int info:31;
+  } p;
+};
+
+_Static_assert (sizeof (long int) == sizeof (union dirstream_packed),
+		"sizeof (long int) != sizeof (union dirstream_packed)");
+
+/* telldir maintains a list of offsets that describe the obtained diretory
+   position if it can fit this information in the returned 'dirstream_packed'
+   struct.  */
+
+# define DYNARRAY_STRUCT  dirstream_loc_t
+# define DYNARRAY_ELEMENT off64_t
+# define DYNARRAY_PREFIX  dirstream_loc_
+# include <malloc/dynarray-skeleton.c>
+
+static __always_inline bool
+telldir_need_dirstream (__off64_t d_off)
+{
+  return d_off >= 1UL << 31;
+}
+#else
+
+_Static_assert (sizeof (long int) == sizeof (off64_t),
+		"sizeof (long int) != sizeof (off64_t)");
+
+#endif /* __LP64__  */
+
+#endif /* _TELLDIR_H  */
diff --git a/sysdeps/unix/sysv/linux/tst-opendir-nolfs.c b/sysdeps/unix/sysv/linux/tst-opendir-nolfs.c
new file mode 100644
index 0000000000..52e18171a7
--- /dev/null
+++ b/sysdeps/unix/sysv/linux/tst-opendir-nolfs.c
@@ -0,0 +1,146 @@
+/* Check multiple telldir and seekdir.
+   Copyright (C) 2023 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <https://www.gnu.org/licenses/>.  */
+
+#include <dirent.h>
+#include <fcntl.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <unistd.h>
+
+#include <support/check.h>
+#include <support/support.h>
+#include <support/temp_file.h>
+#include <support/xunistd.h>
+
+/* Some filesystems returns an arbitrary value for d_off direnty entry (ext4
+   for instance, where the value is an internal hash key).  The idea of create
+   a large number of file is to try trigger a overflow d_off value in a entry
+   to check if telldir/seekdir does work corretly in such case.  */
+static const char *dirname;
+/* The 2 extra files are '.' and '..'.  */
+static const size_t nfiles = (1<<14) + 2;
+
+static inline bool
+in_ino_t_range (ino64_t v)
+{
+  ino_t s = v;
+  return s == v;
+}
+
+static inline bool
+in_off_t_range (off64_t v)
+{
+  off_t s = v;
+  return s == v;
+}
+
+static void
+do_prepare (int argc, char *argv[])
+{
+  dirname = support_create_temp_directory ("tst-opendir-nolfs-");
+
+  for (size_t i = 0; i < nfiles - 2; i++)
+    {
+      int fd = create_temp_file_in_dir ("tempfile.", dirname, NULL);
+      TEST_VERIFY_EXIT (fd > 0);
+      close (fd);
+    }
+}
+#define PREPARE do_prepare
+
+static int
+do_test (void)
+{
+  DIR *dirp = opendir (dirname);
+  TEST_VERIFY_EXIT (dirp != NULL);
+
+  long int *tdirp = xmalloc (nfiles * sizeof (long int));
+  struct dirent **ddirp = xmalloc (nfiles * sizeof (struct dirent *));
+
+  /* For non-LFS, the entry is skipped if it can not be converted.  */
+  int count = 0;
+  for (; count < nfiles; count++)
+    {
+      tdirp[count] = telldir (dirp);
+      struct dirent *dp = readdir (dirp);
+      if (dp == NULL)
+	break;
+      ddirp[count] = xmalloc (dp->d_reclen);
+      memcpy (ddirp[count], dp, dp->d_reclen);
+    }
+
+  closedir (dirp);
+
+  /* Check against the getdents64 syscall.  */
+  int fd = xopen (dirname, O_RDONLY | O_DIRECTORY, 0);
+  int i = 0;
+  while (true)
+    {
+      struct
+      {
+	char buffer[1024];
+	struct dirent64 pad;
+      } data;
+
+      ssize_t ret = getdents64 (fd, &data.buffer, sizeof (data.buffer));
+      if (ret < 0)
+	FAIL_EXIT1 ("getdents64: %m");
+      if (ret == 0)
+	break;
+
+      char *current = data.buffer;
+      char *end = data.buffer + ret;
+      while (current != end)
+	{
+	  struct dirent64 entry;
+          memcpy (&entry, current, sizeof (entry));
+          /* Truncate overlong strings.  */
+          entry.d_name[sizeof (entry.d_name) - 1] = '\0';
+          TEST_VERIFY (strlen (entry.d_name) < sizeof (entry.d_name) - 1);
+
+	  if (in_ino_t_range (entry.d_ino) && in_off_t_range (entry.d_off))
+	    {
+	      TEST_COMPARE_STRING (entry.d_name, ddirp[i]->d_name);
+	      TEST_COMPARE (entry.d_ino, ddirp[i]->d_ino);
+	      TEST_COMPARE (entry.d_off, ddirp[i]->d_off);
+	      TEST_COMPARE (entry.d_type, ddirp[i]->d_type);
+
+	      /* Offset zero is reserved for the first entry.  */
+	      TEST_VERIFY (entry.d_off != 0);
+
+	      TEST_VERIFY_EXIT (entry.d_reclen <= end - current);
+	      i++;
+	    }
+
+	  current += entry.d_reclen;
+	}
+    }
+
+  /* direntries_read has been called more than once.  */
+  TEST_COMPARE (count, i);
+
+  free (tdirp);
+  for (int i = 0; i < count; i++)
+    free (ddirp[i]);
+  free (ddirp);
+
+  return 0;
+}
+
+#include <support/test-driver.c>
-- 
2.34.1

[PATCH v5 3/5] linux: Add __readdir64_unlocked

[Adhemerval Zanella]

And use it on readdir_r implementation.

Checked on i686-linux-gnu.
---
 include/dirent.h                      |  1 +
 sysdeps/unix/sysv/linux/readdir64.c   | 20 +++++--
 sysdeps/unix/sysv/linux/readdir64_r.c | 80 ++++++---------------------
 3 files changed, 33 insertions(+), 68 deletions(-)

diff --git a/include/dirent.h b/include/dirent.h
index 17827176ba..f391476298 100644
--- a/include/dirent.h
+++ b/include/dirent.h
@@ -21,6 +21,7 @@ extern DIR *__fdopendir (int __fd) attribute_hidden;
 extern int __closedir (DIR *__dirp) attribute_hidden;
 extern struct dirent *__readdir (DIR *__dirp) attribute_hidden;
 extern struct dirent *__readdir_unlocked (DIR *__dirp) attribute_hidden;
+extern struct dirent64 *__readdir64_unlocked (DIR *__dirp) attribute_hidden;
 extern struct dirent64 *__readdir64 (DIR *__dirp);
 libc_hidden_proto (__readdir64)
 extern int __readdir_r (DIR *__dirp, struct dirent *__entry,
diff --git a/sysdeps/unix/sysv/linux/readdir64.c b/sysdeps/unix/sysv/linux/readdir64.c
index db1c6214d8..2327511736 100644
--- a/sysdeps/unix/sysv/linux/readdir64.c
+++ b/sysdeps/unix/sysv/linux/readdir64.c
@@ -28,15 +28,11 @@
 
 /* Read a directory entry from DIRP.  */
 struct dirent64 *
-__readdir64 (DIR *dirp)
+__readdir64_unlocked (DIR *dirp)
 {
   struct dirent64 *dp;
   int saved_errno = errno;
 
-#if IS_IN (libc)
-  __libc_lock_lock (dirp->lock);
-#endif
-
   if (dirp->offset >= dirp->size)
     {
       /* We've emptied out our buffer.  Refill it.  */
@@ -68,6 +64,20 @@ __readdir64 (DIR *dirp)
   dirp->offset += dp->d_reclen;
   dirp->filepos = dp->d_off;
 
+  return dp;
+}
+
+struct dirent64 *
+__readdir64 (DIR *dirp)
+{
+  struct dirent64 *dp;
+
+#if IS_IN (libc)
+  __libc_lock_lock (dirp->lock);
+#endif
+
+  dp = __readdir64_unlocked (dirp);
+
 #if IS_IN (libc)
   __libc_lock_unlock (dirp->lock);
 #endif
diff --git a/sysdeps/unix/sysv/linux/readdir64_r.c b/sysdeps/unix/sysv/linux/readdir64_r.c
index 285dc99509..5ae099bde7 100644
--- a/sysdeps/unix/sysv/linux/readdir64_r.c
+++ b/sysdeps/unix/sysv/linux/readdir64_r.c
@@ -32,89 +32,43 @@ __readdir64_r (DIR *dirp, struct dirent64 *entry, struct dirent64 **result)
 {
   struct dirent64 *dp;
   size_t reclen;
-  const int saved_errno = errno;
-  int ret;
 
   __libc_lock_lock (dirp->lock);
-
-  do
+  while (1)
     {
-      if (dirp->offset >= dirp->size)
-	{
-	  /* We've emptied out our buffer.  Refill it.  */
-
-	  size_t maxread = dirp->allocation;
-	  ssize_t bytes;
-
-	  maxread = dirp->allocation;
-
-	  bytes = __getdents64 (dirp->fd, dirp->data, maxread);
-	  if (bytes <= 0)
-	    {
-	      /* On some systems getdents fails with ENOENT when the
-		 open directory has been rmdir'd already.  POSIX.1
-		 requires that we treat this condition like normal EOF.  */
-	      if (bytes < 0 && errno == ENOENT)
-		{
-		  bytes = 0;
-		  __set_errno (saved_errno);
-		}
-	      if (bytes < 0)
-		dirp->errcode = errno;
-
-	      dp = NULL;
-	      break;
-	    }
-	  dirp->size = (size_t) bytes;
-
-	  /* Reset the offset into the buffer.  */
-	  dirp->offset = 0;
-	}
-
-      dp = (struct dirent64 *) &dirp->data[dirp->offset];
+      dp = __readdir64_unlocked (dirp);
+      if (dp == NULL)
+	break;
 
       reclen = dp->d_reclen;
+      if (reclen <= offsetof (struct dirent64, d_name) + NAME_MAX + 1)
+	break;
 
-      dirp->offset += reclen;
-
-      dirp->filepos = dp->d_off;
-
-      if (reclen > offsetof (struct dirent64, d_name) + NAME_MAX + 1)
+      /* The record is very long.  It could still fit into the caller-supplied
+	 buffer if we can skip padding at the end.  */
+      size_t namelen = _D_EXACT_NAMLEN (dp);
+      if (namelen <= NAME_MAX)
 	{
-	  /* The record is very long.  It could still fit into the
-	     caller-supplied buffer if we can skip padding at the
-	     end.  */
-	  size_t namelen = _D_EXACT_NAMLEN (dp);
-	  if (namelen <= NAME_MAX)
-	    reclen = offsetof (struct dirent64, d_name) + namelen + 1;
-	  else
-	    {
-	      /* The name is too long.  Ignore this file.  */
-	      dirp->errcode = ENAMETOOLONG;
-	      dp->d_ino = 0;
-	      continue;
-	    }
+	  reclen = offsetof (struct dirent64, d_name) + namelen + 1;
+	  break;
 	}
 
-      /* Skip deleted and ignored files.  */
+      /* The name is too long.  Ignore this file.  */
+      dirp->errcode = ENAMETOOLONG;
+      dp->d_ino = 0;
     }
-  while (dp->d_ino == 0);
 
   if (dp != NULL)
     {
       *result = memcpy (entry, dp, reclen);
       entry->d_reclen = reclen;
-      ret = 0;
     }
   else
-    {
-      *result = NULL;
-      ret = dirp->errcode;
-    }
+    *result = NULL;
 
   __libc_lock_unlock (dirp->lock);
 
-  return ret;
+  return dp != NULL ? 0 : dirp->errcode;
 }
 
 
-- 
2.34.1

[PATCH v5 4/5] linux: Add __old_readdir64_unlocked

[Adhemerval Zanella]

And use it __old_readdir64_r.

Checked on i686-linux-gnu.
---
 sysdeps/unix/sysv/linux/olddirent.h   |  2 +
 sysdeps/unix/sysv/linux/readdir64.c   | 24 +++++---
 sysdeps/unix/sysv/linux/readdir64_r.c | 79 ++++++---------------------
 3 files changed, 35 insertions(+), 70 deletions(-)

diff --git a/sysdeps/unix/sysv/linux/olddirent.h b/sysdeps/unix/sysv/linux/olddirent.h
index 9789ffae07..cde95e192e 100644
--- a/sysdeps/unix/sysv/linux/olddirent.h
+++ b/sysdeps/unix/sysv/linux/olddirent.h
@@ -32,6 +32,8 @@ struct __old_dirent64
 /* Now define the internal interfaces.  */
 extern struct __old_dirent64 *__old_readdir64 (DIR *__dirp);
 libc_hidden_proto (__old_readdir64);
+extern struct __old_dirent64 *__old_readdir64_unlocked (DIR *__dirp)
+        attribute_hidden;
 extern int __old_readdir64_r (DIR *__dirp, struct __old_dirent64 *__entry,
 			  struct __old_dirent64 **__result);
 extern __ssize_t __old_getdents64 (int __fd, char *__buf, size_t __nbytes)
diff --git a/sysdeps/unix/sysv/linux/readdir64.c b/sysdeps/unix/sysv/linux/readdir64.c
index 2327511736..b901071aa7 100644
--- a/sysdeps/unix/sysv/linux/readdir64.c
+++ b/sysdeps/unix/sysv/linux/readdir64.c
@@ -104,15 +104,11 @@ versioned_symbol (libc, __readdir64, readdir64, GLIBC_2_2);
 
 attribute_compat_text_section
 struct __old_dirent64 *
-__old_readdir64 (DIR *dirp)
+__old_readdir64_unlocked (DIR *dirp)
 {
   struct __old_dirent64 *dp;
   int saved_errno = errno;
 
-#if IS_IN (libc)
-  __libc_lock_lock (dirp->lock);
-#endif
-
   if (dirp->offset >= dirp->size)
     {
       /* We've emptied out our buffer.  Refill it.  */
@@ -129,9 +125,6 @@ __old_readdir64 (DIR *dirp)
 	     do not set errno in that case, to indicate success.  */
 	  if (bytes == 0 || errno == ENOENT)
 	    __set_errno (saved_errno);
-#if IS_IN (libc)
-	  __libc_lock_unlock (dirp->lock);
-#endif
 	  return NULL;
 	}
       dirp->size = (size_t) bytes;
@@ -144,6 +137,21 @@ __old_readdir64 (DIR *dirp)
   dirp->offset += dp->d_reclen;
   dirp->filepos = dp->d_off;
 
+  return dp;
+}
+
+attribute_compat_text_section
+struct __old_dirent64 *
+__old_readdir64 (DIR *dirp)
+{
+  struct __old_dirent64 *dp;
+
+#if IS_IN (libc)
+  __libc_lock_lock (dirp->lock);
+#endif
+
+  dp = __old_readdir64_unlocked (dirp);
+
 #if IS_IN (libc)
   __libc_lock_unlock (dirp->lock);
 #endif
diff --git a/sysdeps/unix/sysv/linux/readdir64_r.c b/sysdeps/unix/sysv/linux/readdir64_r.c
index 5ae099bde7..b499388de7 100644
--- a/sysdeps/unix/sysv/linux/readdir64_r.c
+++ b/sysdeps/unix/sysv/linux/readdir64_r.c
@@ -91,89 +91,44 @@ __old_readdir64_r (DIR *dirp, struct __old_dirent64 *entry,
 {
   struct __old_dirent64 *dp;
   size_t reclen;
-  const int saved_errno = errno;
-  int ret;
 
   __libc_lock_lock (dirp->lock);
 
-  do
+  while (1)
     {
-      if (dirp->offset >= dirp->size)
-	{
-	  /* We've emptied out our buffer.  Refill it.  */
-
-	  size_t maxread = dirp->allocation;
-	  ssize_t bytes;
-
-	  maxread = dirp->allocation;
-
-	  bytes = __old_getdents64 (dirp->fd, dirp->data, maxread);
-	  if (bytes <= 0)
-	    {
-	      /* On some systems getdents fails with ENOENT when the
-		 open directory has been rmdir'd already.  POSIX.1
-		 requires that we treat this condition like normal EOF.  */
-	      if (bytes < 0 && errno == ENOENT)
-		{
-		  bytes = 0;
-		  __set_errno (saved_errno);
-		}
-	      if (bytes < 0)
-		dirp->errcode = errno;
-
-	      dp = NULL;
-	      break;
-	    }
-	  dirp->size = (size_t) bytes;
-
-	  /* Reset the offset into the buffer.  */
-	  dirp->offset = 0;
-	}
-
-      dp = (struct __old_dirent64 *) &dirp->data[dirp->offset];
+      dp = __old_readdir64_unlocked (dirp);
+      if (dp == NULL)
+	break;
 
       reclen = dp->d_reclen;
+      if (reclen <= offsetof (struct __old_dirent64, d_name) + NAME_MAX + 1)
+	break;
 
-      dirp->offset += reclen;
-
-      dirp->filepos = dp->d_off;
-
-      if (reclen > offsetof (struct __old_dirent64, d_name) + NAME_MAX + 1)
+      /* The record is very long.  It could still fit into the caller-supplied
+	 buffer if we can skip padding at the end.  */
+      size_t namelen = _D_EXACT_NAMLEN (dp);
+      if (namelen <= NAME_MAX)
 	{
-	  /* The record is very long.  It could still fit into the
-	     caller-supplied buffer if we can skip padding at the
-	     end.  */
-	  size_t namelen = _D_EXACT_NAMLEN (dp);
-	  if (namelen <= NAME_MAX)
-	    reclen = offsetof (struct __old_dirent64, d_name) + namelen + 1;
-	  else
-	    {
-	      /* The name is too long.  Ignore this file.  */
-	      dirp->errcode = ENAMETOOLONG;
-	      dp->d_ino = 0;
-	      continue;
-	    }
+	  reclen = offsetof (struct dirent64, d_name) + namelen + 1;
+	  break;
 	}
 
-      /* Skip deleted and ignored files.  */
+      /* The name is too long.  Ignore this file.  */
+      dirp->errcode = ENAMETOOLONG;
+      dp->d_ino = 0;
     }
-  while (dp->d_ino == 0);
 
   if (dp != NULL)
     {
       *result = memcpy (entry, dp, reclen);
       entry->d_reclen = reclen;
-      ret = 0;
     }
   else
-    {
-      *result = NULL;
-      ret = dirp->errcode;
-    }
+    *result = NULL;
 
   __libc_lock_unlock (dirp->lock);
 
-  return ret;
+  return dp != NULL ? 0 : dirp->errcode;
 }
 
 compat_symbol (libc, __old_readdir64_r, readdir64_r, GLIBC_2_1);
-- 
2.34.1

[PATCH v5 5/5] linux: Use getdents64 on readdir64 compat implementation

[Adhemerval Zanella]

It uses a similar strategy from the non-LFS readdir that also
uses getdents64 internally and uses a translation buffer to return
the compat readdir64 entry.

It allows to remove __old_getdents64.

Checked on i686-linux-gnu.
---
 sysdeps/unix/sysv/linux/dirstream.h  | 13 +++-
 sysdeps/unix/sysv/linux/getdents64.c | 93 ----------------------------
 sysdeps/unix/sysv/linux/olddirent.h  |  2 -
 sysdeps/unix/sysv/linux/readdir64.c  | 50 +++++++++++----
 4 files changed, 50 insertions(+), 108 deletions(-)

diff --git a/sysdeps/unix/sysv/linux/dirstream.h b/sysdeps/unix/sysv/linux/dirstream.h
index 8f58a1c3a6..b03ece4590 100644
--- a/sysdeps/unix/sysv/linux/dirstream.h
+++ b/sysdeps/unix/sysv/linux/dirstream.h
@@ -24,6 +24,11 @@
 #include <libc-lock.h>
 #include <telldir.h>
 
+#include <shlib-compat.h>
+#if SHLIB_COMPAT(libc, GLIBC_2_1, GLIBC_2_2)
+# include <olddirent.h>
+#endif
+
 /* Directory stream type.
 
    The miscellaneous Unix `readdir' implementations read directory data
@@ -44,7 +49,13 @@ struct __dirstream
     int errcode;		/* Delayed error code.  */
 
 #if !defined __OFF_T_MATCHES_OFF64_T || !defined __INO_T_MATCHES_INO64_T
-    struct dirent tdp;
+    union
+      {
+        struct dirent tdp;
+#if  SHLIB_COMPAT(libc, GLIBC_2_1, GLIBC_2_2)
+        struct __old_dirent64 tdp64;
+#    endif
+      };
 #endif
 #if _DIRENT_OFFSET_TRANSLATION
     struct dirstream_loc_t locs; /* off64_t to long int map for telldir.  */
diff --git a/sysdeps/unix/sysv/linux/getdents64.c b/sysdeps/unix/sysv/linux/getdents64.c
index 01c3517deb..db299864ed 100644
--- a/sysdeps/unix/sysv/linux/getdents64.c
+++ b/sysdeps/unix/sysv/linux/getdents64.c
@@ -36,97 +36,4 @@ weak_alias (__getdents64, getdents64)
 
 #if _DIRENT_MATCHES_DIRENT64
 strong_alias (__getdents64, __getdents)
-#else
-# include <shlib-compat.h>
-
-# if SHLIB_COMPAT(libc, GLIBC_2_1, GLIBC_2_2)
-#  include <olddirent.h>
-#  include <unistd.h>
-
-static ssize_t
-handle_overflow (int fd, __off64_t offset, ssize_t count)
-{
-  /* If this is the first entry in the buffer, we can report the
-     error.  */
-  if (offset == 0)
-    {
-      __set_errno (EOVERFLOW);
-      return -1;
-    }
-
-  /* Otherwise, seek to the overflowing entry, so that the next call
-     will report the error, and return the data read so far.  */
-  if (__lseek64 (fd, offset, SEEK_SET) != 0)
-    return -1;
-  return count;
-}
-
-ssize_t
-__old_getdents64 (int fd, char *buf, size_t nbytes)
-{
-  /* We do not move the individual directory entries.  This is only
-     possible if the target type (struct __old_dirent64) is smaller
-     than the source type.  */
-  _Static_assert (offsetof (struct __old_dirent64, d_name)
-		  <= offsetof (struct dirent64, d_name),
-		  "__old_dirent64 is larger than dirent64");
-  _Static_assert (__alignof__ (struct __old_dirent64)
-		  <= __alignof__ (struct dirent64),
-		  "alignment of __old_dirent64 is larger than dirent64");
-
-  ssize_t retval = INLINE_SYSCALL_CALL (getdents64, fd, buf, nbytes);
-  if (retval > 0)
-    {
-      /* This is the marker for the first entry.  Offset 0 is reserved
-	 for the first entry (see rewinddir).  Here, we use it as a
-	 marker for the first entry in the buffer.  We never actually
-	 seek to offset 0 because handle_overflow reports the error
-	 directly, so it does not matter that the offset is incorrect
-	 if entries have been read from the descriptor before (so that
-	 the descriptor is not actually at offset 0).  */
-      __off64_t previous_offset = 0;
-
-      char *p = buf;
-      char *end = buf + retval;
-      while (p < end)
-	{
-	  struct dirent64 *source = (struct dirent64 *) p;
-
-	  /* Copy out the fixed-size data.  */
-	  __ino_t ino = source->d_ino;
-	  __off64_t offset = source->d_off;
-	  unsigned int reclen = source->d_reclen;
-	  unsigned char type = source->d_type;
-
-	  /* Check for ino_t overflow.  */
-	  if (__glibc_unlikely (ino != source->d_ino))
-	    return handle_overflow (fd, previous_offset, p - buf);
-
-	  /* Convert to the target layout.  Use a separate struct and
-	     memcpy to side-step aliasing issues.  */
-	  struct __old_dirent64 result;
-	  result.d_ino = ino;
-	  result.d_off = offset;
-	  result.d_reclen = reclen;
-	  result.d_type = type;
-
-	  /* Write the fixed-sized part of the result to the
-	     buffer.  */
-	  size_t result_name_offset = offsetof (struct __old_dirent64, d_name);
-	  memcpy (p, &result, result_name_offset);
-
-	  /* Adjust the position of the name if necessary.  Copy
-	     everything until the end of the record, including the
-	     terminating NUL byte.  */
-	  if (result_name_offset != offsetof (struct dirent64, d_name))
-	    memmove (p + result_name_offset, source->d_name,
-		     reclen - offsetof (struct dirent64, d_name));
-
-	  p += reclen;
-	  previous_offset = offset;
-	}
-     }
-  return retval;
-}
-# endif /* SHLIB_COMPAT(libc, GLIBC_2_1, GLIBC_2_2)  */
 #endif /* _DIRENT_MATCHES_DIRENT64  */
diff --git a/sysdeps/unix/sysv/linux/olddirent.h b/sysdeps/unix/sysv/linux/olddirent.h
index cde95e192e..2d682a6919 100644
--- a/sysdeps/unix/sysv/linux/olddirent.h
+++ b/sysdeps/unix/sysv/linux/olddirent.h
@@ -36,8 +36,6 @@ extern struct __old_dirent64 *__old_readdir64_unlocked (DIR *__dirp)
         attribute_hidden;
 extern int __old_readdir64_r (DIR *__dirp, struct __old_dirent64 *__entry,
 			  struct __old_dirent64 **__result);
-extern __ssize_t __old_getdents64 (int __fd, char *__buf, size_t __nbytes)
-	attribute_hidden;
 int __old_scandir64 (const char * __dir,
 		     struct __old_dirent64 *** __namelist,
 		     int (*__selector) (const struct __old_dirent64 *),
diff --git a/sysdeps/unix/sysv/linux/readdir64.c b/sysdeps/unix/sysv/linux/readdir64.c
index b901071aa7..88e42c5e90 100644
--- a/sysdeps/unix/sysv/linux/readdir64.c
+++ b/sysdeps/unix/sysv/linux/readdir64.c
@@ -102,21 +102,43 @@ versioned_symbol (libc, __readdir64, readdir64, GLIBC_2_2);
 # if SHLIB_COMPAT(libc, GLIBC_2_1, GLIBC_2_2)
 #  include <olddirent.h>
 
+/* Translate the DP64 entry to the old LFS one in the translation buffer
+   at dirstream DS.  Return true is the translation was possible or
+   false if either an internal fields can be represented in the non-LFS
+   entry or if the translation can not be resized.  */
+static bool
+dirstream_old_entry (struct __dirstream *ds, const struct dirent64 *dp64)
+{
+  /* Check for overflow.  */
+  if (!in_ino_t_range (dp64->d_ino))
+    return false;
+
+  /* And if name is too large.  */
+  if (dp64->d_reclen - offsetof (struct dirent64, d_name) > NAME_MAX)
+    return false;
+
+  ds->filepos = dp64->d_off;
+
+  ds->tdp64.d_off = dp64->d_off;
+  ds->tdp64.d_ino = dp64->d_ino;
+  ds->tdp64.d_reclen = dp64->d_reclen;
+  ds->tdp64.d_type = dp64->d_type;
+  memcpy (ds->tdp64.d_name, dp64->d_name,
+	  dp64->d_reclen - offsetof (struct dirent64, d_name));
+
+  return true;
+}
+
 attribute_compat_text_section
 struct __old_dirent64 *
 __old_readdir64_unlocked (DIR *dirp)
 {
-  struct __old_dirent64 *dp;
-  int saved_errno = errno;
+  const int saved_errno = errno;
 
   if (dirp->offset >= dirp->size)
     {
       /* We've emptied out our buffer.  Refill it.  */
-
-      size_t maxread = dirp->allocation;
-      ssize_t bytes;
-
-      bytes = __old_getdents64 (dirp->fd, dirp->data, maxread);
+      ssize_t bytes = __getdents64 (dirp->fd, dirp->data, dirp->allocation);
       if (bytes <= 0)
 	{
 	  /* Linux may fail with ENOENT on some file systems if the
@@ -127,17 +149,21 @@ __old_readdir64_unlocked (DIR *dirp)
 	    __set_errno (saved_errno);
 	  return NULL;
 	}
-      dirp->size = (size_t) bytes;
+      dirp->size = bytes;
 
       /* Reset the offset into the buffer.  */
       dirp->offset = 0;
     }
 
-  dp = (struct __old_dirent64 *) &dirp->data[dirp->offset];
-  dirp->offset += dp->d_reclen;
-  dirp->filepos = dp->d_off;
+  struct dirent64 *dp64 = (struct dirent64 *) &dirp->data[dirp->offset];
+  dirp->offset += dp64->d_reclen;
 
-  return dp;
+  /* Skip entries which might overflow d_ino or for memory allocation failure
+     in case of large file names.  */
+  if (dirstream_old_entry (dirp, dp64))
+    return &dirp->tdp64;
+
+  return NULL;
 }
 
 attribute_compat_text_section
-- 
2.34.1

Re: [PATCH v5 1/5] linux: Use getdents64 on non-LFS readdir

[Andreas Schwab]

On Jan 27 2023, Adhemerval Zanella via Libc-alpha wrote:

> +      /* Non-lfs opendir skips entries that can not be represented (for
> +	 instance if d_off is not an offset but rather an internal filesystem
> +	 representation.  For this case there is no point in continue the
> +	 testcase.  */

Missing close paren.  Either "in continuing" or "to continue".

-- 
Andreas Schwab, schwab@linux-m68k.org
GPG Key fingerprint = 7578 EB47 D4E5 4D69 2510  2552 DF73 E780 A9DA AEC1
"And now for something completely different."

Re: [PATCH v5 0/5] Fix opendir regression on some FS

[Andreas K. Huettel]

[-- Attachment #1: Type: text/plain, Size: 435 bytes --]

Am Freitag, 27. Januar 2023, 18:28:29 CET schrieben Sie:
> Some filesystem might return a non-representable d_off on getdents
> call, even if there are few entries in the directories (for instance
> ext4, which returns a hash).  This trigger issues with non-LFS readdir
[...]

Ping- anything still missing here? Thanks :)

-- 
Andreas K. Huettel
tel. +49 151 241 67748 (mobile)
e-mail mail@akhuettel.de
http://www.akhuettel.de/private/

[-- Attachment #2: This is a digitally signed message part. --]
[-- Type: application/pgp-signature, Size: 981 bytes --]

Re: [PATCH v5 0/5] Fix opendir regression on some FS

[Andreas K. Huettel]

[-- Attachment #1: Type: text/plain, Size: 453 bytes --]

Am Freitag, 27. Januar 2023, 18:28:29 CET schrieben Sie:
> Some filesystem might return a non-representable d_off on getdents
> call, even if there are few entries in the directories (for instance
> ext4, which returns a hash).  This trigger issues with non-LFS readdir
[...]

Ping- anything still missing here? Thanks :)

-- 
Andreas K. Hüttel
dilfridge@gentoo.org
Gentoo Linux developer
(council, toolchain, base-system, perl, libreoffice)

[-- Attachment #2: This is a digitally signed message part. --]
[-- Type: application/pgp-signature, Size: 981 bytes --]

Re: [PATCH v5 1/5] linux: Use getdents64 on non-LFS readdir

[Florian Weimer]

* Adhemerval Zanella:

> diff --git a/sysdeps/unix/sysv/linux/dirstream.h b/sysdeps/unix/sysv/linux/dirstream.h
> index 3cb313b410..adcf8234f1 100644
> --- a/sysdeps/unix/sysv/linux/dirstream.h
> +++ b/sysdeps/unix/sysv/linux/dirstream.h
> @@ -18,6 +18,7 @@
>  #ifndef	_DIRSTREAM_H
>  #define	_DIRSTREAM_H	1
>  
> +#include <dirent.h>
>  #include <sys/types.h>
>  
>  #include <libc-lock.h>
> @@ -41,6 +42,10 @@ struct __dirstream
>  
>      int errcode;		/* Delayed error code.  */
>  
> +#if !defined __OFF_T_MATCHES_OFF64_T || !defined __INO_T_MATCHES_INO64_T
> +    struct dirent tdp;
> +#endif

I don't quite see how this can work given that d_name in the struct may
not be large enough.  The new member needs a comment to explain its
purpose.

> diff --git a/sysdeps/unix/sysv/linux/readdir.c b/sysdeps/unix/sysv/linux/readdir.c
> index 4a4c00ea07..cd0ccaf33a 100644
> --- a/sysdeps/unix/sysv/linux/readdir.c
> +++ b/sysdeps/unix/sysv/linux/readdir.c
> @@ -21,42 +21,71 @@
>  #if !_DIRENT_MATCHES_DIRENT64
>  #include <dirstream.h>
>  
> +/* Translate the DP64 entry to the non-LFS one in the translation entry
> +   at dirstream DS.  Return true is the translation was possible or
> +   false if either an internal field can not be represented in the non-LFS
> +   entry or if the name is too long.  */
> +static bool
> +dirstream_entry (struct __dirstream *ds, const struct dirent64 *dp64)
> +{
> +  /* Check for overflow.  */
> +  if (!in_off_t_range (dp64->d_off) || !in_ino_t_range (dp64->d_ino))
> +    return false;
> +
> +  /* And if name is too large.  */
> +  if (dp64->d_reclen - offsetof (struct dirent64, d_name) > NAME_MAX)
> +    return false;

Sorry, I don't think this is the direction we should be going.  In
readdir_r, we at least delay the NAME_MAX error to the end of the
directory.  This just adds another rare case where 32-bit code fails and
64-bit code works.

struct dirent is always shorter than struct dirent64, right?  It should
be possible to do the translation in-place.  Or turn tdp into a pointer
and reallocate as needed.

However, I think we should fix only readdir64, not readdir.  It's simply
not possible to fix readdir fully because of d_ino, so applications
which use readdir instead of readdir64 will remain buggy even with this
change.

What do you think?

Thanks,
Florian

Re: [PATCH v5 1/5] linux: Use getdents64 on non-LFS readdir

[Florian Weimer]

* Florian Weimer:

>> diff --git a/sysdeps/unix/sysv/linux/readdir.c b/sysdeps/unix/sysv/linux/readdir.c
>> index 4a4c00ea07..cd0ccaf33a 100644
>> --- a/sysdeps/unix/sysv/linux/readdir.c
>> +++ b/sysdeps/unix/sysv/linux/readdir.c
>> @@ -21,42 +21,71 @@
>>  #if !_DIRENT_MATCHES_DIRENT64
>>  #include <dirstream.h>
>>  
>> +/* Translate the DP64 entry to the non-LFS one in the translation entry
>> +   at dirstream DS.  Return true is the translation was possible or
>> +   false if either an internal field can not be represented in the non-LFS
>> +   entry or if the name is too long.  */
>> +static bool
>> +dirstream_entry (struct __dirstream *ds, const struct dirent64 *dp64)
>> +{
>> +  /* Check for overflow.  */
>> +  if (!in_off_t_range (dp64->d_off) || !in_ino_t_range (dp64->d_ino))
>> +    return false;
>> +
>> +  /* And if name is too large.  */
>> +  if (dp64->d_reclen - offsetof (struct dirent64, d_name) > NAME_MAX)
>> +    return false;
>
> Sorry, I don't think this is the direction we should be going.  In
> readdir_r, we at least delay the NAME_MAX error to the end of the
> directory.  This just adds another rare case where 32-bit code fails and
> 64-bit code works.
>
> struct dirent is always shorter than struct dirent64, right?  It should
> be possible to do the translation in-place.  Or turn tdp into a pointer
> and reallocate as needed.
>
> However, I think we should fix only readdir64, not readdir.  It's simply
> not possible to fix readdir fully because of d_ino, so applications
> which use readdir instead of readdir64 will remain buggy even with this
> change.

Meh, during a walk it occurred tome that 64-bit d_ino is far less common
than 64-bit d_off.  So we really need all this rewriting. 8-(

Could you do it in-place at least?  To address the d_name sizing issue?

Thanks,
Florian

Re: [PATCH v5 2/5] linux: Set internal DIR filepos as off64_t (BZ #23960, BZ #24050)

[Florian Weimer]

* Adhemerval Zanella:

> diff --git a/dirent/tst-seekdir.c b/dirent/tst-seekdir.c
> index dcdd699b09..187eda7584 100644
> --- a/dirent/tst-seekdir.c
> +++ b/dirent/tst-seekdir.c
> @@ -41,6 +41,14 @@ do_test (void)
>        if (i == 400)
>  	break;
>      }
> +  if (i < 3)
> +    {
> +      /* Non-lfs opendir skips entries that can not be represented (for
> +	 instance if d_off is not an offset but rather an internal filesystem
> +	 representation.  For this case there is no point in continue the
> +	 testcase.  */
> +      return 77;
> +    }
>  
>    printf ("going back past 4-th entry...\n");

I think this needs to check errno properly (set to zero before readdir,
and if readdir fails, report a test error if errno is not zero).

> index adcf8234f1..8f58a1c3a6 100644
> --- a/sysdeps/unix/sysv/linux/dirstream.h
> +++ b/sysdeps/unix/sysv/linux/dirstream.h
> @@ -22,6 +22,7 @@
>  #include <sys/types.h>
>  
>  #include <libc-lock.h>
> +#include <telldir.h>
>  
>  /* Directory stream type.
>  
> @@ -38,13 +39,16 @@ struct __dirstream
>      size_t size;		/* Total valid data in the block.  */
>      size_t offset;		/* Current offset into the block.  */
>  
> -    off_t filepos;		/* Position of next entry to read.  */
> +    off64_t filepos;		/* Position of next entry to read.  */
>  
>      int errcode;		/* Delayed error code.  */
>  
>  #if !defined __OFF_T_MATCHES_OFF64_T || !defined __INO_T_MATCHES_INO64_T
>      struct dirent tdp;
>  #endif
> +#if _DIRENT_OFFSET_TRANSLATION
> +    struct dirstream_loc_t locs; /* off64_t to long int map for telldir.  */
> +#endif

Please add a comment explaining the purpose of the new field.

> diff --git a/sysdeps/unix/sysv/linux/readdir.c b/sysdeps/unix/sysv/linux/readdir.c
> index cd0ccaf33a..7a7f484c36 100644
> --- a/sysdeps/unix/sysv/linux/readdir.c
> +++ b/sysdeps/unix/sysv/linux/readdir.c
> @@ -36,6 +36,15 @@ dirstream_entry (struct __dirstream *ds, const struct dirent64 *dp64)
>    if (dp64->d_reclen - offsetof (struct dirent64, d_name) > NAME_MAX)
>      return false;
>  
> +  /* telldir can not return an error, so preallocate the map if the entry can
> +     not be packed directly.  */
> +  if (telldir_need_dirstream (dp64->d_off))
> +    {
> +      dirstream_loc_add (&ds->locs, dp64->d_off);
> +      if (dirstream_loc_has_failed (&ds->locs))
> +	return false;
> +    }

I recommend to reserve the space here (but only for d_off values that
need it), and perform the assignment only if there is an actual telldir
call.  This avoids all the additional processing and allocations for the
common case that telldir is never called.  This will matter quite a bit
for large directories.

Reservation can be implemented by adding a 0 dummy value (offset 0 will
never be recorded), unless the last element is already 0.  I don't think
we need to change dynarray for this.


> diff --git a/sysdeps/unix/sysv/linux/seekdir.c b/sysdeps/unix/sysv/linux/seekdir.c
> index 939ccc4447..30cce691a4 100644
> --- a/sysdeps/unix/sysv/linux/seekdir.c
> +++ b/sysdeps/unix/sysv/linux/seekdir.c

> +
> +  if (dirp->filepos != filepos)
> +    {
> +      __lseek64 (dirp->fd, filepos, SEEK_SET);
> +      dirp->filepos = filepos;
> +      dirp->offset = 0;
> +      dirp->size = 0;
> +    }

I think we should seek unconditionally, just in case someone messed with
dirfd directly.

> diff --git a/sysdeps/unix/sysv/linux/telldir.c b/sysdeps/unix/sysv/linux/telldir.c
> index 1e5c129e9f..c3ef14f3da 100644
> --- a/sysdeps/unix/sysv/linux/telldir.c
> +++ b/sysdeps/unix/sysv/linux/telldir.c
> @@ -15,9 +15,11 @@
>     License along with the GNU C Library; if not, see
>     <https://www.gnu.org/licenses/>.  */

> +  else
> +    {
> +      dsp.l = -1;
> +
> +      size_t i;
> +      for (i = 0; i < dirstream_loc_size (&dirp->locs); i++)
> +	if (*dirstream_loc_at (&dirp->locs, i) == dirp->filepos)
> +	  break;
> +      /* It should be pre-allocated on readdir.  */
> +      assert (i != dirstream_loc_size (&dirp->locs));

This should overwrite the final 0 element if the offset could not be
found.

I think the quadratic behavior is acceptable in this case, given that we
do not really expect telldir to be called.

> diff --git a/sysdeps/unix/sysv/linux/telldir.h b/sysdeps/unix/sysv/linux/telldir.h
> new file mode 100644
> index 0000000000..758bcb0eb3
> --- /dev/null
> +++ b/sysdeps/unix/sysv/linux/telldir.h
> @@ -0,0 +1,65 @@
> +/* Linux internal telldir definitions.
> +   Copyright (C) 2023 Free Software Foundation, Inc.
> +   This file is part of the GNU C Library.
> +
> +   The GNU C Library is free software; you can redistribute it and/or
> +   modify it under the terms of the GNU Lesser General Public
> +   License as published by the Free Software Foundation; either
> +   version 2.1 of the License, or (at your option) any later version.
> +
> +   The GNU C Library is distributed in the hope that it will be useful,
> +   but WITHOUT ANY WARRANTY; without even the implied warranty of
> +   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
> +   Lesser General Public License for more details.
> +
> +   You should have received a copy of the GNU Lesser General Public
> +   License along with the GNU C Library; if not, see
> +   <https://www.gnu.org/licenses/>.  */
> +
> +#ifndef _TELLDIR_H
> +#define _TELLDIR_H 1
> +
> +#if _DIRENT_OFFSET_TRANSLATION
> +/* On platforms where 'long int' is smaller than 'off64_t' this is how the
> +   returned value is encoded and returned by 'telldir'.  If the directory
> +   offset can be enconded in 31 bits it is returned in the 'info' member
> +   with 'is_packed' set to 1.
> +
> +   Otherwise, the 'info' member describes an index in a dynamic array at
> +   'DIR' structure.  */
> +
> +union dirstream_packed
> +{
> +  long int l;
> +  struct
> +  {
> +    unsigned long int is_packed:1;
> +    unsigned long int info:31;
> +  } p;
> +};

Good explanation in the comment, but … can't you use negative values for
indices, and ~ for encoding/decoding?  I think that's a bit clearer.

> +static __always_inline bool
> +telldir_need_dirstream (__off64_t d_off)
> +{
> +  return d_off >= 1UL << 31;
> +}

Hmm, can d_off be negative?

> diff --git a/sysdeps/unix/sysv/linux/tst-opendir-nolfs.c b/sysdeps/unix/sysv/linux/tst-opendir-nolfs.c
> new file mode 100644
> index 0000000000..52e18171a7

> +static int
> +do_test (void)
> +{
> +  DIR *dirp = opendir (dirname);
> +  TEST_VERIFY_EXIT (dirp != NULL);
> +
> +  long int *tdirp = xmalloc (nfiles * sizeof (long int));
> +  struct dirent **ddirp = xmalloc (nfiles * sizeof (struct dirent *));

xreallocarray?

> +  /* For non-LFS, the entry is skipped if it can not be converted.  */
> +  int count = 0;
> +  for (; count < nfiles; count++)
> +    {
> +      tdirp[count] = telldir (dirp);
> +      struct dirent *dp = readdir (dirp);
> +      if (dp == NULL)
> +	break;

Should probably verify the count variable here (and check errno too?).

> +      ddirp[count] = xmalloc (dp->d_reclen);
> +      memcpy (ddirp[count], dp, dp->d_reclen);
> +    }
> +
> +  closedir (dirp);
> +
> +  /* Check against the getdents64 syscall.  */
> +  int fd = xopen (dirname, O_RDONLY | O_DIRECTORY, 0);
> +  int i = 0;
> +  while (true)

If we create predictable file names (maybe just numbers?), this kind of
checking becomes way easier.

Thanks,
Florian

Re: [PATCH v5 1/5] linux: Use getdents64 on non-LFS readdir

[Adhemerval Zanella Netto]

On 20/02/23 14:55, Florian Weimer wrote:
> * Florian Weimer:
> 
>>> diff --git a/sysdeps/unix/sysv/linux/readdir.c b/sysdeps/unix/sysv/linux/readdir.c
>>> index 4a4c00ea07..cd0ccaf33a 100644
>>> --- a/sysdeps/unix/sysv/linux/readdir.c
>>> +++ b/sysdeps/unix/sysv/linux/readdir.c
>>> @@ -21,42 +21,71 @@
>>>  #if !_DIRENT_MATCHES_DIRENT64
>>>  #include <dirstream.h>
>>>  
>>> +/* Translate the DP64 entry to the non-LFS one in the translation entry
>>> +   at dirstream DS.  Return true is the translation was possible or
>>> +   false if either an internal field can not be represented in the non-LFS
>>> +   entry or if the name is too long.  */
>>> +static bool
>>> +dirstream_entry (struct __dirstream *ds, const struct dirent64 *dp64)
>>> +{
>>> +  /* Check for overflow.  */
>>> +  if (!in_off_t_range (dp64->d_off) || !in_ino_t_range (dp64->d_ino))
>>> +    return false;
>>> +
>>> +  /* And if name is too large.  */
>>> +  if (dp64->d_reclen - offsetof (struct dirent64, d_name) > NAME_MAX)
>>> +    return false;
>>
>> Sorry, I don't think this is the direction we should be going.  In
>> readdir_r, we at least delay the NAME_MAX error to the end of the
>> directory.  This just adds another rare case where 32-bit code fails and
>> 64-bit code works.
>>
>> struct dirent is always shorter than struct dirent64, right?  It should
>> be possible to do the translation in-place.  Or turn tdp into a pointer
>> and reallocate as needed.
>>
>> However, I think we should fix only readdir64, not readdir.  It's simply
>> not possible to fix readdir fully because of d_ino, so applications
>> which use readdir instead of readdir64 will remain buggy even with this
>> change.
> 
> Meh, during a walk it occurred tome that 64-bit d_ino is far less common
> than 64-bit d_off.  So we really need all this rewriting. 8-(
> 
> Could you do it in-place at least?  To address the d_name sizing issue?

I added the translation buffer only to simplify the copy logic, but it should
be doable to do it in place (similar to what getdents64 for mips64 does).

Re: [PATCH v5 2/5] linux: Set internal DIR filepos as off64_t (BZ #23960, BZ #24050)

[Adhemerval Zanella Netto]

On 20/02/23 15:39, Florian Weimer wrote:
> * Adhemerval Zanella:
> 
>> diff --git a/dirent/tst-seekdir.c b/dirent/tst-seekdir.c
>> index dcdd699b09..187eda7584 100644
>> --- a/dirent/tst-seekdir.c
>> +++ b/dirent/tst-seekdir.c
>> @@ -41,6 +41,14 @@ do_test (void)
>>        if (i == 400)
>>  	break;
>>      }
>> +  if (i < 3)
>> +    {
>> +      /* Non-lfs opendir skips entries that can not be represented (for
>> +	 instance if d_off is not an offset but rather an internal filesystem
>> +	 representation.  For this case there is no point in continue the
>> +	 testcase.  */
>> +      return 77;
>> +    }
>>  
>>    printf ("going back past 4-th entry...\n");
> 
> I think this needs to check errno properly (set to zero before readdir,
> and if readdir fails, report a test error if errno is not zero).

It can't in fact, since the entries will be ignored nad not reported as
errors.  Yes, it is subpar; but the whole non-LFS interface has similar
corner cases and I think we can't paper over to much of its limitations.

> 
>> index adcf8234f1..8f58a1c3a6 100644
>> --- a/sysdeps/unix/sysv/linux/dirstream.h
>> +++ b/sysdeps/unix/sysv/linux/dirstream.h
>> @@ -22,6 +22,7 @@
>>  #include <sys/types.h>
>>  
>>  #include <libc-lock.h>
>> +#include <telldir.h>
>>  
>>  /* Directory stream type.
>>  
>> @@ -38,13 +39,16 @@ struct __dirstream
>>      size_t size;		/* Total valid data in the block.  */
>>      size_t offset;		/* Current offset into the block.  */
>>  
>> -    off_t filepos;		/* Position of next entry to read.  */
>> +    off64_t filepos;		/* Position of next entry to read.  */
>>  
>>      int errcode;		/* Delayed error code.  */
>>  
>>  #if !defined __OFF_T_MATCHES_OFF64_T || !defined __INO_T_MATCHES_INO64_T
>>      struct dirent tdp;
>>  #endif
>> +#if _DIRENT_OFFSET_TRANSLATION
>> +    struct dirstream_loc_t locs; /* off64_t to long int map for telldir.  */
>> +#endif
> 
> Please add a comment explaining the purpose of the new field.

Ack.

> 
>> diff --git a/sysdeps/unix/sysv/linux/readdir.c b/sysdeps/unix/sysv/linux/readdir.c
>> index cd0ccaf33a..7a7f484c36 100644
>> --- a/sysdeps/unix/sysv/linux/readdir.c
>> +++ b/sysdeps/unix/sysv/linux/readdir.c
>> @@ -36,6 +36,15 @@ dirstream_entry (struct __dirstream *ds, const struct dirent64 *dp64)
>>    if (dp64->d_reclen - offsetof (struct dirent64, d_name) > NAME_MAX)
>>      return false;
>>  
>> +  /* telldir can not return an error, so preallocate the map if the entry can
>> +     not be packed directly.  */
>> +  if (telldir_need_dirstream (dp64->d_off))
>> +    {
>> +      dirstream_loc_add (&ds->locs, dp64->d_off);
>> +      if (dirstream_loc_has_failed (&ds->locs))
>> +	return false;
>> +    }
> 
> I recommend to reserve the space here (but only for d_off values that
> need it), and perform the assignment only if there is an actual telldir
> call.  This avoids all the additional processing and allocations for the
> common case that telldir is never called.  This will matter quite a bit
> for large directories.
> 
> Reservation can be implemented by adding a 0 dummy value (offset 0 will
> never be recorded), unless the last element is already 0.  I don't think
> we need to change dynarray for this.

We need the d_off value, since it will be used by telldir to actually return
the correct dynarray position to return as dirstream_packed union. 

> 
> 
>> diff --git a/sysdeps/unix/sysv/linux/seekdir.c b/sysdeps/unix/sysv/linux/seekdir.c
>> index 939ccc4447..30cce691a4 100644
>> --- a/sysdeps/unix/sysv/linux/seekdir.c
>> +++ b/sysdeps/unix/sysv/linux/seekdir.c
> 
>> +
>> +  if (dirp->filepos != filepos)
>> +    {
>> +      __lseek64 (dirp->fd, filepos, SEEK_SET);
>> +      dirp->filepos = filepos;
>> +      dirp->offset = 0;
>> +      dirp->size = 0;
>> +    }
> 
> I think we should seek unconditionally, just in case someone messed with
> dirfd directly.

Alright.

> 
>> diff --git a/sysdeps/unix/sysv/linux/telldir.c b/sysdeps/unix/sysv/linux/telldir.c
>> index 1e5c129e9f..c3ef14f3da 100644
>> --- a/sysdeps/unix/sysv/linux/telldir.c
>> +++ b/sysdeps/unix/sysv/linux/telldir.c
>> @@ -15,9 +15,11 @@
>>     License along with the GNU C Library; if not, see
>>     <https://www.gnu.org/licenses/>.  */
> 
>> +  else
>> +    {
>> +      dsp.l = -1;
>> +
>> +      size_t i;
>> +      for (i = 0; i < dirstream_loc_size (&dirp->locs); i++)
>> +	if (*dirstream_loc_at (&dirp->locs, i) == dirp->filepos)
>> +	  break;
>> +      /* It should be pre-allocated on readdir.  */
>> +      assert (i != dirstream_loc_size (&dirp->locs));
> 
> This should overwrite the final 0 element if the offset could not be
> found.
> 
> I think the quadratic behavior is acceptable in this case, given that we
> do not really expect telldir to be called.

Ack.

> 
>> diff --git a/sysdeps/unix/sysv/linux/telldir.h b/sysdeps/unix/sysv/linux/telldir.h
>> new file mode 100644
>> index 0000000000..758bcb0eb3
>> --- /dev/null
>> +++ b/sysdeps/unix/sysv/linux/telldir.h
>> @@ -0,0 +1,65 @@
>> +/* Linux internal telldir definitions.
>> +   Copyright (C) 2023 Free Software Foundation, Inc.
>> +   This file is part of the GNU C Library.
>> +
>> +   The GNU C Library is free software; you can redistribute it and/or
>> +   modify it under the terms of the GNU Lesser General Public
>> +   License as published by the Free Software Foundation; either
>> +   version 2.1 of the License, or (at your option) any later version.
>> +
>> +   The GNU C Library is distributed in the hope that it will be useful,
>> +   but WITHOUT ANY WARRANTY; without even the implied warranty of
>> +   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
>> +   Lesser General Public License for more details.
>> +
>> +   You should have received a copy of the GNU Lesser General Public
>> +   License along with the GNU C Library; if not, see
>> +   <https://www.gnu.org/licenses/>.  */
>> +
>> +#ifndef _TELLDIR_H
>> +#define _TELLDIR_H 1
>> +
>> +#if _DIRENT_OFFSET_TRANSLATION
>> +/* On platforms where 'long int' is smaller than 'off64_t' this is how the
>> +   returned value is encoded and returned by 'telldir'.  If the directory
>> +   offset can be enconded in 31 bits it is returned in the 'info' member
>> +   with 'is_packed' set to 1.
>> +
>> +   Otherwise, the 'info' member describes an index in a dynamic array at
>> +   'DIR' structure.  */
>> +
>> +union dirstream_packed
>> +{
>> +  long int l;
>> +  struct
>> +  {
>> +    unsigned long int is_packed:1;
>> +    unsigned long int info:31;
>> +  } p;
>> +};
> 
> Good explanation in the comment, but … can't you use negative values for
> indices, and ~ for encoding/decoding?  I think that's a bit clearer.

Do you mean remove the bitfield?

> 
>> +static __always_inline bool
>> +telldir_need_dirstream (__off64_t d_off)
>> +{
>> +  return d_off >= 1UL << 31;
>> +}
> 
> Hmm, can d_off be negative?

But it is an unsigned comparion, so it should not matter afaiu.

> 
>> diff --git a/sysdeps/unix/sysv/linux/tst-opendir-nolfs.c b/sysdeps/unix/sysv/linux/tst-opendir-nolfs.c
>> new file mode 100644
>> index 0000000000..52e18171a7
> 
>> +static int
>> +do_test (void)
>> +{
>> +  DIR *dirp = opendir (dirname);
>> +  TEST_VERIFY_EXIT (dirp != NULL);
>> +
>> +  long int *tdirp = xmalloc (nfiles * sizeof (long int));
>> +  struct dirent **ddirp = xmalloc (nfiles * sizeof (struct dirent *));
> 
> xreallocarray?

Ack.

> 
>> +  /* For non-LFS, the entry is skipped if it can not be converted.  */
>> +  int count = 0;
>> +  for (; count < nfiles; count++)
>> +    {
>> +      tdirp[count] = telldir (dirp);
>> +      struct dirent *dp = readdir (dirp);
>> +      if (dp == NULL)
>> +	break;
> 
> Should probably verify the count variable here (and check errno too?).

The issue is we don't know if the entry will be skipped due overflow (the test
check for non-LFS interface).  And this lead me to add a test for LFS interface
and realize that we need to use dirstream_packed for LFS as well if the
ABI sets _DIRENT_OFFSET_TRANSLATION.

> 
>> +      ddirp[count] = xmalloc (dp->d_reclen);
>> +      memcpy (ddirp[count], dp, dp->d_reclen);
>> +    }
>> +
>> +  closedir (dirp);
>> +
>> +  /* Check against the getdents64 syscall.  */
>> +  int fd = xopen (dirname, O_RDONLY | O_DIRECTORY, 0);
>> +  int i = 0;
>> +  while (true)
> 
> If we create predictable file names (maybe just numbers?), this kind of
> checking becomes way easier.

I am not sure because the test can have a 'count' less than 'nfiles'.