Who needs this filesystem malarkey anyway? (20 Jul 2003)

You know, my home system didn't feel slow till I started using the dual Opteron system. Heck, even the dual Xeon-HT's don't feel as nippy and it's running 32-bit code.

I posted a note to python-dev today about finding the size of types at configure time. Almost nothing except glibc, gcc and binutils works cleanly when cross compiling. GNU autoconf should mean that setting --host and --build makes everything work magically. Does it hell.

(p.s. glibc 2.3.2 cannot be cross compiled, use 2.3.1. And both of these versions misdefine sscanf - you have to correct it in stdio-common/sscanf.c first.)

One of the dumbest things in configure scripts is when they don't try tests because they can't run the compiled code (because it's x86-64 code). The script knows it's a cross and just gives up. In the case of Python it assumes that the sizes of int etc are for 32-bit. (Except for fpos_t, for which it's correct for an 8-bit system). But there's no reason to run compiled code to get this information.

#include <asm/types.h>
#include <sys/types.h>

const   __u8    sizeof_int[sizeof(int)];

And so on. Then compile the code and objdump -t sizeof.o | grep 'sizeof_[^ ]+$' | awk '{ print $5 " " $1; } will give you all the information. Works perfectly for native and cross compilers.

DJB exchanged emails about his call for a disablenetwork() syscall. My point was basically that he was thinking about it the wrong way round. It shouldn't be a disablenetwork call, but a case of "I didn't explictly give you a network capability".

I also remarked that if you were going to go a capability you could also chroot() everything and give it a UNIX domain socket via which it could make its filesystem calls. This would make restricting programs pretty simple as you have one point of access for all filesystem control. (It would be a UNIX domain socket because they can have file descriptors passed between processes over them).

He suggested that few programs really need the filesystem (and would you look at the date on that?) and that it has more than security implications:

A small interface (for example,
a descriptor allowing read() or write()) supports many implementations
(disk files; network connections; and all sorts of interesting programs
via pipes), dramatically expanding the user's power to combine programs.
A big interface (for example, a file descriptor that allows directory
operations) naturally has far fewer implementations.

Which is actually really cool. Most programs could do without the full fledged filesystem and it would be useful to be able to redirect their access down a pipe or socket. There are a number of problems with being able to do this that would probably need a kernel patch; mmaping for one and the problem of no being able to pass fds between machines.