All changes to the Linux kernel are distributed in the form of patches, which are the output of the GNU diff(1) program in a form that is readable by the patch(1) program. The simplest way to generate a patch is to have two source trees, one that is the vanilla stock kernel and another that is the stock tree with your modifications. A common scheme is to name the stock tree linux-x.y.z (which is what the source tarball extracts to, initially) and to name your modified tree simply linux. Then, to generate a patch of the two trees, issue the command
diff -urN linux-x.y.z/ linux/ > my-patch
from one directory below your trees. This is typically done somewhere in your home, and not /usr/src/linux, so that you do not need to be root. The -u flag specifies that the unified diff format should be used. Without this, the patch is ugly and not very readable by humans. The -r flag specifies to recursively diff all directories, and the -N flag specifies that new files in the modified tree should be included in the diff. Alternatively, if you need to diff only a single file, you can do
diff -u linux-x.y.z/some/file linux/some/file > my-patch
Note that it is important to always diff the trees from one directory below your source trees. This creates a patch that is very usable by others, even if their directory names differ. To apply a patch made in this format, do
patch -p1 < ../my-patch
from the root of your source tree. In this example, the patch is named my_patch and is created one directory below the current. The -p1 flag specifies that the first directory name is to be stripped from the patch. This enables you to apply a patch regardless of the directory-naming convention used by the patch maker.
diffstat -p1 my-patch
It is often useful to include this output when you post a patch to lkml. Because the patch(1) program ignores all lines until a diff is detected, you can even include a short description at the top of the patch itself.