cat,cp,mv,install,split: Set the minimum IO block size used to 32KiB

This is following on from this change:
[02c3dc9d 2008-03-06 cat: use larger buffer sizes ...]
which increased the IO block size used by cat by 8 times,
but also capped it at 32KiB.
* NEWS: Mention the change in behavior.
* src/system.h: Add a new io_blksize() function that
returns the max of ST_BLKSIZE or 32KiB, as this was
seen as a good value for a minimum block size to use
to get good performance while minimizing system call overhead.
* src/cat.c: Use it.
* src/copy.c: ditto
* src/split.c: ditto

4 files changed, 53 insertions, 22 deletions

diff --git a/src/cat.c b/src/cat.c
index 04eb204d4..18fa1f150 100644
--- a/src/cat.c
+++ b/src/cat.c
@@ -78,12 +78,6 @@ static char *line_num_end = line_buf + LINE_COUNTER_BUF_LEN - 3;
 /* Preserves the `cat' function's local `newlines' between invocations.  */
 static int newlines2 = 0;
 
-static inline size_t
-compute_buffer_size (struct stat st)
-{
-  return MIN (8 * ST_BLKSIZE (st), 32 * 1024);
-}
-
 void
 usage (int status)
 {
@@ -642,7 +636,7 @@ main (int argc, char **argv)
   if (fstat (STDOUT_FILENO, &stat_buf) < 0)
     error (EXIT_FAILURE, errno, _("standard output"));
 
-  outsize = compute_buffer_size (stat_buf);
+  outsize = io_blksize (stat_buf);
   /* Input file can be output file for non-regular files.
      fstat on pipes returns S_IFSOCK on some systems, S_IFIFO
      on others, so the checking should not be done for those types,
@@ -706,7 +700,7 @@ main (int argc, char **argv)
 	  ok = false;
 	  goto contin;
 	}
-      insize = compute_buffer_size (stat_buf);
+      insize = io_blksize (stat_buf);
 
       /* Compare the device and i-node numbers of this input file with
 	 the corresponding values of the (output file associated with)
diff --git a/src/copy.c b/src/copy.c
index 191867126..e37fead52 100644
--- a/src/copy.c
+++ b/src/copy.c
@@ -568,7 +568,7 @@ copy_reg (char const *src_name, char const *dst_name,
     /* Choose a suitable buffer size; it may be adjusted later.  */
     size_t buf_alignment = lcm (getpagesize (), sizeof (word));
     size_t buf_alignment_slop = sizeof (word) + buf_alignment - 1;
-    size_t buf_size = ST_BLKSIZE (sb);
+    size_t buf_size = io_blksize (sb);
 
     /* Deal with sparse files.  */
     bool last_write_made_hole = false;
@@ -596,21 +596,12 @@ copy_reg (char const *src_name, char const *dst_name,
        buffer size.  */
     if (! make_holes)
       {
-	/* These days there's no point ever messing with buffers smaller
-	   than 8 KiB.  It would be nice to configure SMALL_BUF_SIZE
-	   dynamically for this host and pair of files, but there doesn't
-	   seem to be a good way to get readahead info portably.  */
-	enum { SMALL_BUF_SIZE = 8 * 1024 };
-
 	/* Compute the least common multiple of the input and output
 	   buffer sizes, adjusting for outlandish values.  */
 	size_t blcm_max = MIN (SIZE_MAX, SSIZE_MAX) - buf_alignment_slop;
-	size_t blcm = buffer_lcm (ST_BLKSIZE (src_open_sb), buf_size,
+	size_t blcm = buffer_lcm (io_blksize (src_open_sb), buf_size,
 				  blcm_max);
 
-	/* Do not use a block size that is too small.  */
-	buf_size = MAX (SMALL_BUF_SIZE, blcm);
-
 	/* Do not bother with a buffer larger than the input file, plus one
 	   byte to make sure the file has not grown while reading it.  */
 	if (S_ISREG (src_open_sb.st_mode) && src_open_sb.st_size < buf_size)
diff --git a/src/split.c b/src/split.c
index 1d8a94ca7..f8e26835b 100644
--- a/src/split.c
+++ b/src/split.c
@@ -554,7 +554,7 @@ main (int argc, char **argv)
 
   if (fstat (STDIN_FILENO, &stat_buf) != 0)
     error (EXIT_FAILURE, errno, "%s", infile);
-  in_blk_size = ST_BLKSIZE (stat_buf);
+  in_blk_size = io_blksize (stat_buf);
 
   buf = ptr_align (xmalloc (in_blk_size + 1 + page_size - 1), page_size);
 
diff --git a/src/system.h b/src/system.h
index ba74da448..dbb4da1ed 100644
--- a/src/system.h
+++ b/src/system.h
@@ -147,6 +147,9 @@ enum
 # define D_INO(dp) NOT_AN_INODE_NUMBER
 #endif
 
+/* include here for SIZE_MAX.  */
+#include <inttypes.h>
+
 /* Get or fake the disk device blocksize.
    Usually defined by sys/param.h (if at all).  */
 #if !defined DEV_BSIZE && defined BSIZE
@@ -218,6 +221,51 @@ enum
 # endif
 #endif
 
+/* As of Mar 2009, 32KiB is determined to be the minimium
+   blksize to best minimize system call overhead.
+   This can be tested with this script with the results
+   shown for a 1.7GHz pentium-m with 2GB of 400MHz DDR2 RAM:
+
+   for i in $(seq 0 10); do
+     size=$((8*1024**3)) #ensure this is big enough
+     bs=$((1024*2**$i))
+     printf "%7s=" $bs
+     dd bs=$bs if=/dev/zero of=/dev/null count=$(($size/$bs)) 2>&1 |
+     sed -n 's/.* \([0-9.]* [GM]B\/s\)/\1/p'
+   done
+
+      1024=734 MB/s
+      2048=1.3 GB/s
+      4096=2.4 GB/s
+      8192=3.5 GB/s
+     16384=3.9 GB/s
+     32768=5.2 GB/s
+     65536=5.3 GB/s
+    131072=5.5 GB/s
+    262144=5.7 GB/s
+    524288=5.7 GB/s
+   1048576=5.8 GB/s
+
+   Note that this is to minimize system call overhead.
+   Other values may be appropriate to minimize file system
+   or disk overhead.  For example on my current linux system
+   the readahead setting is 128KiB which was read using:
+
+   file="."
+   device=$(df -P --local "$file" | tail -n1 | cut -d' ' -f1)
+   echo $(( $(blockdev --getra $device) * 512 ))
+
+   However there isn't a portable way to get the above.
+   In the future we could use the above method if available
+   and default to io_blksize() if not.
+ */
+enum { IO_BUFSIZE = 32*1024 };
+static inline size_t
+io_blksize (struct stat sb)
+{
+  return MAX (IO_BUFSIZE, ST_BLKSIZE (sb));
+}
+
 /* Redirection and wildcarding when done by the utility itself.
    Generally a noop, but used in particular for native VMS. */
 #ifndef initialize_main
@@ -228,8 +276,6 @@ enum
 
 #include "timespec.h"
 
-#include <inttypes.h>
-
 #include <ctype.h>
 
 #if ! (defined isblank || HAVE_DECL_ISBLANK)