Re: First draft list of 2.3.x "Things to fix"

[Jeff Dike]

I'm carrying a few core kernel fixes which I need in order for the user-mode 
port to work.  This works for now since I'm distributing my stuff separately. 
I'm planning on getting it into 2.5, so it would be nice to have these fixes 
in 2.4.

- kswapd/bdflush race
	If wakeup_bdflush/try_to_free_pages manage to run before bdflush/kswapd
	have managed to be scheduled, bdlush_tsk/kswapd_process will be NULL,
	resulting in wake_up_process(NULL).  This can happen if there is always
	a runnable, normal-priority process to be scheduled.  The scheduler 
	seems to prefer recently-scheduled processes over not-recently-scheduled
	processes, so this can result in kswapd/bdflush being starved for 
	arbitrarily long times under the right conditions.

	I've brought this up before, proposing a test for NULL in
	wakeup_bdflush/try_to_free_pages, which Andrea rejected as adding code
	to a common path which would almost never be used.

	This time, I'd like to propose having kswapd_init/bdflush_init sleep on
	a wait_queue and have kswapd/bdflush wake it up after setting
	bdlush_tsk/kswapd_process.  This would add two lines of code in each
	case, which would run once per boot.

- move in_egroup_p() to sys.c
	This is a configuration bug, not specific to the user-mode port.  As 
	things stand, in_egroup_p() is controlled by CONFIG_SYSCTL but a caller
	of it is under CONFIG_QUOTA.  This makes a kernel with CONFIG_SYSCTL 
	off and CONFIG_QUOTA on fail to link.

	I pointed this out
	(http://kernelnotes.org/lnxlists/linux-kernel/lk_9912_01/msg01420.html)
	and Jan Kara responded with
	http://kernelnotes.org/lnxlists/linux-kernel/lk_9912_02/msg00039.html,
	which implied a	fix was on its way, but I haven't seen it come through
	yet.

				Jeff