Uwe Hermann has described how to resize a root filesystem after booting from a live-cd or recovery disk . He makes some good points about resizing an LVM PV (which I hadn’t even realised was possible).
The following paragraph is outdated, see the update at the end:
Incidentally it should be noted that if your root filesystem is an LVM logical volume then it can’t be resized without booting from a different device because the way LVM appears to work is that the LV in question is locked, then files under /etc/lvm/ are written, and then the LV is unlocked. If the LV in question contains /etc/lvm then you deadlock your root filesystem and need to press reset. Of course if your root filesystem is on an LV which has been encrypted via cryptsetup (the LV is encrypted not the PV) then a live resize of the root filesystem can work as locking the LV merely means that write-backs from the encryption layer don’t get committed. I’m not sure if this means that data written to an encrypted device is less stable (testing this is on my todo list).
If your root filesystem is on a partition of a hard drive (such as /dev/hda2) then it is possible to extend it without booting from different media. There is nothing stopping you from running fdisk and deleting the partition of your root filesystem and recreating it. When you exit fdisk it will call an ioctl() to re-read the partition table, the kernel code counts the number of open file handles related to the device and if the number is greater than 1 (fdisk has one open handle) then it refuses to re-read the table. So you can use fdisk to change the root partition and then reboot to have the change be noticed. After that ext2online can be used to take advantage of the extra space (if the filesystem has a recent enough version of the ext3 disk format).
One thing he didn’t mention is that if you do need to boot from another device to manipulate your root filesystem (which should be quite rare if you are bold and know the tricks) then you can always use your swap space. To do this you simply run swapoff and then run mkfs on the device that had been used for swap (incidentally there is nothing really special about the swap space, but it does tend to often be used for such recovery operations simply because it has no data that persists across a reboot).
The minimum set of files that need to be copied to a temporary filesystem is usually /bin, /sbin, /dev, /lib (excluding all directories under /lib/modules apart from the one related to the kernel you are using), and /etc. Also you should make directories /proc, /sys, and /selinux (if the machine in question runs SE Linux). The aim is not to copy enough files for the machine to run in a regular manner, merely enough to allow manipulating all filesystems and logical volumes. Often for such recovery I boot with init=/bin/bash as a kernel parameter to skip the regular init and just start with a shell. Note that when you use init=/bin/bash you end up with a shell that has no job control and ^C is not enabled, if you want to run a command that might not terminate of it’s own accord then the command “openvt /bin/bash” can be used to start another session with reasonable terminal settings.
I recommend that anyone who wants to work as a sys-admin experiment with such procedures on a test machine. There are lots of interesting things that you can learn and interesting ways that you can break your system when performing such operations.