exposing checksums to user-space programs

Hi Folks (please cc: any replies to me),

I have modified GNU Make to use checksums rather than timestamps. I would like the operating system and/or filesystem to compute and provide the checksums, because it can safely cache them.

Since btrfs checksums its files, I am trying to transfer its checksums to user space. I believe I saw a thread where Chris Mason said he didn't want to expose checksums, but then I found a patch that sounds like it does what I want:

    http://www.mail-archive.com/[email protected]/msg01182.html

In the thread, Chris said he would take the patch, but it evidently never made it to distribution. I applied the patch to kernel 2.6.30.8, which is old but available on one of my expendable laptops. I created a btrfs filesystem on a USB thumb drive and copied a file to it. Then, I cobbled together this example user-space program:

http://cs.boisestate.edu/~buff/files/trycsum.c

which produces this output:

spruce$ sudo mount /mnt/sdb1
spruce$ trycsum /mnt/sdb1/yum.list
File /mnt/sdb1/yum.list has 1 extents:
# Logical Physical Length Flags
0: 0000000000000000 0000000000c00000 000000000015b000 0001
sizeof(args)=420
size=0
csum[0]=0000007b
spruce$

The "size=0" tells me the file's one extent has no checksum. However:

spruce$ sudo btrfs-debug-tree /dev/sdb1 | grep -i csum
item 0 key (18446744073709551606 EXTENT_CSUM 12582912) itemoff 2607 itemsize 1388
extent csum item
spruce$

suggests it has a checksum.

I traced the function defined in the patch, with printk() calls. It seemed be accepting and processing its arguments correctly.

I'm not a kernel-hacking expert, but I was hoping someone could tell me what I'm doing wrong or give me some direction. Thanks!

-- Jim

PS: I think I also noticed that a small file has no checksum, and a big file has no checksum until after a umount/mount. Is this true?