Sunil Mushran wrote:
Jim Meyering wrote:
Hi Jeff,
I've included below the state of my local changes.
Unfortunately, with that 5-patch series, there is always a test failure
on F13/ext4. Maybe someone who knows more about extents can provide an
explanation?
Here's a small example to demonstrate:
Create a file with many extents:
perl -e 'BEGIN { $n = 19 * 1024; *F = *STDOUT }' \
-e 'for (1..100) { sysseek (*F, $n, 1)' \
-e '&& syswrite (*F, "."x$n) or die "$!"}'> j1
Using the patched "cp", repeat the following 10 or 20 times:
./cp --sparse=always j1 j2; sync
filefrag -v j1 | awk '/^ / {print $1,$2}'> ff1 || fail=1
filefrag -v j2 | awk '/^ / {print $1,$2}'> ff2 || fail=1
diff -u ff1 ff2 || fail=1
Usually there is no diff output, but occasionally it'll print this:
[hmm... today it consistently prints these differences every other time.]
The reason it does not work is because the sparse file created by cp
may not be sparse (or sparse enough). And that is because cp reads
is chunks of st_blocksize and skips the write only if the entire chunk
is zero. The perl script creates the file in 19K chunks (alternate writes
and holes).
Thanks for replying.
However, your description of how GNU cp works suggests that you're
looking at the pre-FIEMAP semantics. Please refer to the patches here
http://thread.gmane.org/gmane.comp.gnu.coreutils.bugs/20534/focus=20630
that make it use FIEMAP.
So on a 4K fs, the file created by the script will have 4 blocks as holes
(avg). But when cp makes it, it could fill out those holes because the read
granularity could be coarser. For example, ocfs2 fills out st_blocksize
with the fs cluster size (allocation size) which could be larger than the
block size.
My suggestion is to not use filefrag but to use md5sum to compare the two
files.
That would be pointless.
The goal of the test is to determine that the FIEMAP copy
did indeed preserve the extents. If I do as you suggest, even if
cp mistakenly filled in all holes, the test would still pass.