Paul Eggert wrote:
> How about if we change usable_st_size to return false for these proc files

Attached is a better idea, I hope.  I audited the coreutils code to look for 
problematic uses of SEEK_END or st_size when reading files (I didn't look at 
writing; one can of worms at a time).

The attached patch still needs a changelog entry and test cases.  The basic idea 
is to not trust st_size when it's <= ST_BLKSIZE.  This fixes bugs in 'head', 
'od', 'split', 'tac', 'tail', and 'wc' when applied to input files in proc or 
sysfs file systems.

Here's an example bug fixed by this patch:

$ cat /sys/kernel/profiling
0
$ tail -2c /sys/kernel/profiling
$

This should be:

$ cat /sys/kernel/profiling
0
$ tail -2c /sys/kernel/profiling
0
$