Re: [signal] BUG: unable to handle kernel NULL pointer dereference at 0000000000000001

[Al Viro <viro@ZenIV.linux.org.uk>]

On Mon, Sep 10, 2012 at 08:47:53AM +0800, Fengguang Wu wrote:
> Al,
> 
> Something bad happens since the below commit. Because commit [12/18] f411932
> was not tested, it might also be the first bad commit.

There's a damn good reason why the branch is called the way it's called ;-)
It's missing a lot of kernel_thread() stuff; I'm still not sure if it's
worth going that way, but there are some very attractive aspects.

Basically, if we had guaranteed that pt_regs instance on the bottom of the
stack will _not_ overlap anything for a kernel thread, we can move copying
pt_regs into kernel_execve() itself.  Which simplifies the hell out of
ret_from_kernel_execve() instances...

The price is that we need to leave that space aside for kernel threads.
It's not _that_ much and kernel threads are generally less stack-hungry
than the worst cases of syscalls (note that for userland process in the
middle of syscall that pt_regs instance *will* be there, no matter what).

FWIW, the plan for that branch is to do the following trick:
either split ret_from_kernel_thread away from ret_from_fork, setting the
right one to be used at copy_thread() time or make ret_from_fork itself
check if it's returning to kernel.  Either way, do *not* go through
return from syscall in return to kernel case; instead of that, have
ret_from_kernel_thread:
	schedule_tail();
	get the function to be called and its argument from regs and call it
	pass return value to do_exit() (BTW, I'm not at all sure we wouldn't
be better off if we took those do_exit() calls into 3--5 places in callbacks;
kernel_thread() is really very low-level and has few immediate callers).

Of course, such patches would have to go before the ret_from_kernel_execve()
ones.  Right now this branch is guaranteed to be broken; build testing does
make sense, but trying to boot it...  Not yet.