cp: Add FIEMAP support for efficient sparse file copy

* src/fiemap.h: Add fiemap.h for fiemap ioctl(2) support.
Copied from linux's include/linux/fiemap.h, with minor formatting changes.
* src/copy.c (copy_reg): Now, when `cp' invoked with --sparse=[WHEN] option, we
will try to do FIEMAP-copy if the underlaying file system support it, fall back
to a normal copy if it fails.

2 files changed, 261 insertions, 0 deletions

diff --git a/src/copy.c b/src/copy.c
index 171499c47..200338060 100644
--- a/src/copy.c
+++ b/src/copy.c
@@ -63,6 +63,10 @@
 
 #include <sys/ioctl.h>
 
+#ifndef HAVE_FIEMAP
+# include "fiemap.h"
+#endif
+
 #ifndef HAVE_FCHOWN
 # define HAVE_FCHOWN false
 # define fchown(fd, uid, gid) (-1)
@@ -149,6 +153,141 @@ clone_file (int dest_fd, int src_fd)
 #endif
 }
 
+#ifdef __linux__
+# ifndef FS_IOC_FIEMAP
+#  define FS_IOC_FIEMAP _IOWR ('f', 11, struct fiemap)
+# endif
+/* Perform FIEMAP(available in mainline 2.6.27) copy if possible.
+   Call ioctl(2) with FS_IOC_FIEMAP to efficiently map file allocation
+   excepts holes.  So the overhead to deal with holes with lseek(2) in
+   normal copy could be saved.  This would result in much faster backups
+   for any kind of sparse file.  */
+static bool
+fiemap_copy_ok (int src_fd, int dest_fd, size_t buf_size,
+                off_t src_total_size, char const *src_name,
+                char const *dst_name, bool *normal_copy_required)
+{
+  bool fail = false;
+  bool last = false;
+  char fiemap_buf[4096];
+  struct fiemap *fiemap = (struct fiemap *)fiemap_buf;
+  struct fiemap_extent *fm_ext = &fiemap->fm_extents[0];
+  uint32_t count = (sizeof (fiemap_buf) - sizeof (*fiemap)) /
+                    sizeof (struct fiemap_extent);
+  off_t last_ext_logical = 0;
+  uint64_t last_ext_len = 0;
+  uint64_t last_read_size = 0;
+  unsigned int i = 0;
+
+  /* This is required at least to initialize fiemap->fm_start,
+     but also serves (in May 2010) to appease valgrind, which
+     appears not to know the semantics of the FIEMAP ioctl. */
+  memset (fiemap_buf, 0, sizeof fiemap_buf);
+
+  do
+    {
+      fiemap->fm_length = FIEMAP_MAX_OFFSET;
+      fiemap->fm_extent_count = count;
+
+      /* When ioctl(2) fails, fall back to the normal copy only if it
+         is the first time we met.  */
+      if (ioctl (src_fd, FS_IOC_FIEMAP, fiemap) < 0)
+        {
+          /* If `i > 0', then at least one ioctl(2) has been performed before.  */
+          if (i == 0)
+            *normal_copy_required = true;
+          return false;
+        }
+
+      /* If 0 extents are returned, then more ioctls are not needed.  */
+      if (fiemap->fm_mapped_extents == 0)
+        break;
+
+      for (i = 0; i < fiemap->fm_mapped_extents; i++)
+        {
+          assert (fm_ext[i].fe_logical <= OFF_T_MAX);
+
+          off_t ext_logical = fm_ext[i].fe_logical;
+          uint64_t ext_len = fm_ext[i].fe_length;
+
+          if (lseek (src_fd, ext_logical, SEEK_SET) < 0LL)
+            {
+              error (0, errno, _("cannot lseek %s"), quote (src_name));
+              return fail;
+            }
+
+          if (lseek (dest_fd, ext_logical, SEEK_SET) < 0LL)
+            {
+              error (0, errno, _("cannot lseek %s"), quote (dst_name));
+              return fail;
+            }
+
+          if (fm_ext[i].fe_flags & FIEMAP_EXTENT_LAST)
+            {
+              last_ext_logical = ext_logical;
+              last_ext_len = ext_len;
+              last = true;
+            }
+
+          while (ext_len)
+            {
+              char buf[buf_size];
+
+              /* Avoid reading into the holes if the left extent
+                 length is shorter than the buffer size.  */
+              if (ext_len < buf_size)
+                buf_size = ext_len;
+
+              ssize_t n_read = read (src_fd, buf, buf_size);
+              if (n_read < 0)
+                {
+#ifdef EINTR
+                  if (errno == EINTR)
+                    continue;
+#endif
+                  error (0, errno, _("reading %s"), quote (src_name));
+                  return fail;
+                }
+
+              if (n_read == 0)
+                {
+                  /* Figure out how many bytes read from the last extent.  */
+                  last_read_size = last_ext_len - ext_len;
+                  break;
+                }
+
+              if (full_write (dest_fd, buf, n_read) != n_read)
+                {
+                  error (0, errno, _("writing %s"), quote (dst_name));
+                  return fail;
+                }
+
+              ext_len -= n_read;
+            }
+        }
+
+      fiemap->fm_start = fm_ext[i - 1].fe_logical + fm_ext[i - 1].fe_length;
+
+    } while (! last);
+
+  /* If a file ends up with holes, the sum of the last extent logical offset
+     and the read-returned size will be shorter than the actual size of the
+     file.  Use ftruncate to extend the length of the destination file.  */
+  if (last_ext_logical + last_read_size < src_total_size)
+    {
+      if (ftruncate (dest_fd, src_total_size) < 0)
+        {
+          error (0, errno, _("extending %s"), quote (dst_name));
+          return fail;
+        }
+    }
+
+  return ! fail;
+}
+#else
+static bool fiemap_copy_ok (ignored) { errno == ENOTSUP; return false; }
+#endif
+
 /* FIXME: describe */
 /* FIXME: rewrite this to use a hash table so we avoid the quadratic
    performance hit that's probably noticeable only on trees deeper
@@ -679,6 +818,25 @@ copy_reg (char const *src_name, char const *dst_name,
 #endif
         }
 
+      if (make_holes)
+        {
+          bool require_normal_copy = false;
+          /* Perform efficient FIEMAP copy for sparse files, fall back to the
+             standard copy only if the ioctl(2) fails.  */
+          if (fiemap_copy_ok (source_desc, dest_desc, buf_size,
+                              src_open_sb.st_size, src_name,
+                              dst_name, &require_normal_copy))
+            goto preserve_metadata;
+          else
+            {
+              if (! require_normal_copy)
+                {
+                  return_val = false;
+                  goto close_src_and_dst_desc;
+                }
+            }
+        }
+
       /* If not making a sparse file, try to use a more-efficient
          buffer size.  */
       if (! make_holes)
@@ -807,6 +965,7 @@ copy_reg (char const *src_name, char const *dst_name,
         }
     }
 
+preserve_metadata:
   if (x->preserve_timestamps)
     {
       struct timespec timespec[2];
diff --git a/src/fiemap.h b/src/fiemap.h
new file mode 100644
index 000000000..d33293b5c
--- /dev/null
+++ b/src/fiemap.h
@@ -0,0 +1,102 @@
+/* FS_IOC_FIEMAP ioctl infrastructure.
+   Some portions copyright (C) 2007 Cluster File Systems, Inc
+   Authors: Mark Fasheh <mfasheh@suse.com>
+            Kalpak Shah <kalpak.shah@sun.com>
+            Andreas Dilger <adilger@sun.com>.  */
+
+/* Copy from kernel, modified to respect GNU code style by Jie Liu.  */
+
+#ifndef _LINUX_FIEMAP_H
+# define _LINUX_FIEMAP_H
+
+# include <linux/types.h>
+
+struct fiemap_extent
+{
+  /* Logical offset in bytes for the start of the extent
+     from the beginning of the file.  */
+  uint64_t fe_logical;
+
+  /* Physical offset in bytes for the start of the extent
+     from the beginning of the disk.  */
+  uint64_t fe_physical;
+
+  /* Length in bytes for this extent.  */
+  uint64_t fe_length;
+
+  uint64_t fe_reserved64[2];
+
+  /* FIEMAP_EXTENT_* flags for this extent.  */
+  uint32_t fe_flags;
+
+  uint32_t fe_reserved[3];
+};
+
+struct fiemap
+{
+  /* Logical offset(inclusive) at which to start mapping(in).  */
+  uint64_t fm_start;
+
+  /* Logical length of mapping which userspace wants(in).  */
+  uint64_t fm_length;
+
+  /* FIEMAP_FLAG_* flags for request(in/out).  */
+  uint32_t fm_flags;
+
+  /* Number of extents that were mapped(out).  */
+  uint32_t fm_mapped_extents;
+
+  /* Size of fm_extents array(in).  */
+  uint32_t fm_extent_count;
+
+  uint32_t fm_reserved;
+
+  /* Array of mapped extents(out).  */
+  struct fiemap_extent fm_extents[0];
+};
+
+/* The maximum offset can be mapped for a file.  */
+# define FIEMAP_MAX_OFFSET       (~0ULL)
+
+/* Sync file data before map.  */
+# define FIEMAP_FLAG_SYNC        0x00000001
+
+/* Map extented attribute tree.  */
+# define FIEMAP_FLAG_XATTR       0x00000002
+
+# define FIEMAP_FLAGS_COMPAT     (FIEMAP_FLAG_SYNC | FIEMAP_FLAG_XATTR)
+
+/* Last extent in file.  */
+# define FIEMAP_EXTENT_LAST              0x00000001
+
+/* Data location unknown.  */
+# define FIEMAP_EXTENT_UNKNOWN           0x00000002
+
+/* Location still pending, Sets EXTENT_UNKNOWN.  */
+# define FIEMAP_EXTENT_DELALLOC          0x00000004
+
+/* Data can not be read while fs is unmounted.  */
+# define FIEMAP_EXTENT_ENCODED           0x00000008
+
+/* Data is encrypted by fs.  Sets EXTENT_NO_BYPASS.  */
+# define FIEMAP_EXTENT_DATA_ENCRYPTED    0x00000080
+
+/* Extent offsets may not be block aligned.  */
+# define FIEMAP_EXTENT_NOT_ALIGNED       0x00000100
+
+/* Data mixed with metadata.  Sets EXTENT_NOT_ALIGNED.  */
+# define FIEMAP_EXTENT_DATA_INLINE       0x00000200
+
+/* Multiple files in block.  Set EXTENT_NOT_ALIGNED.  */
+# define FIEMAP_EXTENT_DATA_TAIL         0x00000400
+
+/* Space allocated, but not data (i.e. zero).  */
+# define FIEMAP_EXTENT_UNWRITTEN         0x00000800
+
+/* File does not natively support extents.  Result merged for efficiency.  */
+# define FIEMAP_EXTENT_MERGED		0x00001000
+
+/* Space shared with other files.  */
+# define FIEMAP_EXTENT_SHARED            0x00002000
+
+#endif