cp: work around linux kernel bug: short-read != EOF on /proc

Remove the optimization that avoided up to 50% of cp's read syscalls.
Do not assume that a short read on a regular file indicates EOF.
When reading from a file in /proc on linux [at least 2.6.9 - 2.6.29]
into a 4k-byte buffer or larger, a short read does not
always indicate EOF.  For example, "cp /proc/slabinfo /tmp"
copies only 4068 of the total 7493 bytes.  This optimization
(25719a33154f0c62ea9881f0c79ae312dd4cec7a, Improve performance a bit
by optimizing away; 2005-11-24) appears to have been worth less than
a 2% speed-up (and usually much less), so the impact of removing it
is negligible.

* src/copy.c (copy_reg): Don't exit the loop early.
* tests/cp/proc-short-read: New test, lightly based on a suggestion
from Mike Frysinger, to exercise this fix.
* tests/Makefile.am (TESTS): Add cp/proc-short-read.
* NEWS (Improve robustness): Mention this change.

1 files changed, 12 insertions, 0 deletions

diff --git a/NEWS b/NEWS
index 5951bb5e8..18db634bc 100644
--- a/NEWS
+++ b/NEWS
@@ -16,6 +16,18 @@ GNU coreutils NEWS                                    -*- outline -*-
   default should proceed at the speed of the disk.  Previously /dev/urandom
   was used if available, which is relatively slow on GNU/Linux systems.
 
+** Improved robustness
+
+  cp would exit successfully after copying less than the full contents
+  of a file larger than ~4000 bytes from a linux-/proc file system to a
+  destination file system with a fundamental block size of 4KiB or greater.
+  Reading into a 4KiB-or-larger buffer, cp's "read" syscall would return
+  a value smaller than 4096, and cp would interpret that as EOF (POSIX
+  allows this).  This optimization, now removed, saved 50% of cp's read
+  syscalls when copying small files.  Affected linux kernels: at least
+  2.6.9 through 2.6.29.
+  [the optimization was introduced in coreutils-6.0]
+
 ** Portability
 
   `id -G $USER` now works correctly even on Darwin and NetBSD. Previously it