2000-11-18 04:24:59

by Andrew Morton

[permalink] [raw]
Subject: [patch] Remove tq_scheduler


This patch removes tq_scheduler from the kernel. All uses of
tq_scheduler are migrated over to use schedule_task().

Notes:

- In two places: drivers/block/paride/pseudo.h and
drivers/net/wan/sdlamain.c we are re-adding tasks to tq_scheduler
within the callback. That means that these functions are being
called _every_ time the machine calls schedule().

It may be a limited case for paride, but sdlamain.c appears to be
doing this wicked thing all the time.

Now, if you do this with the current schedule_task() your machine
will hang up. So schedule_task() has been changed to support this
practice. If we see that the task queue has waiters on it we still
call schedule(), but we do it in state TASK_RUNNING.

- The patch adds to context.c a new API function
run_schedule_tasks() which immediately runs the queued tasks. Must
only be called from process context. serial.c needs this in its
shutdown routines.

- If anyone sleeps in a callback, then all other users of
schedule_task() also sleep. But there's nothing new here. Kinda
makes one wonder why schedule_task exists. But what-hey, it's neat.

- Note the careful massaging of module reference counts.

Yes my friends, much usage of task queues in modules is racy wrt
module removal. This patch fixes some of them.

The approach taken here is to increment the module refcount when we
enqueue a task and to decrement it in the handler:

mainline()
{
...
MOD_INC_USE_COUNT;
if (schedule_task(some_wq) == 0)
MOD_DEC_USE_COUNT;
}

handler()
{
...
MOD_DEC_USE_COUNT;
/* Wheee! Tiny race here */
return;
}

Note that queue_task and schedule_task have been enhanced to return
a success indicator. If this is non-zero you know that your task
will be run. If it returns zero then your tq_struct was already
queued and you lose.

The patch against test11-pre7 (1043 lines) is at

http://www.uow.edu.au/~andrewm/linux/tq_scheduler.patch

It affects the following files:

include/linux/tqueue.h
include/linux/sched.h
include/linux/compatmac.h
arch/ppc/8xx_io/uart.c
arch/ppc/8xx_io/fec.c
arch/ppc/8260_io/uart.c
drivers/net/wan/sdlamain.c
drivers/block/paride/pseudo.h
drivers/char/tty_io.c
drivers/char/esp.c
drivers/char/istallion.c
drivers/char/riscom8.c
drivers/char/serial.c
drivers/char/README.epca
drivers/char/specialix.c
drivers/char/epca.c
drivers/char/isicom.c
drivers/char/moxa.c
drivers/char/mxser.c
drivers/char/stallion.c
drivers/scsi/megaraid.c
drivers/scsi/qla1280.c
drivers/isdn/avmb1/b1capi.c
drivers/isdn/avmb1/kcapi.c
drivers/sbus/char/sab82532.c
drivers/sbus/char/aurora.c
drivers/ide/ide.c
drivers/usb/serial/keyspan_pda.c
drivers/usb/serial/digi_acceleport.c
drivers/i2o/i2o_lan.c
fs/smbfs/sock.c
kernel/sched.c
kernel/ksyms.c
kernel/timer.c
kernel/context.c


2000-11-19 23:41:56

by David Woodhouse

[permalink] [raw]
Subject: Re: [patch] Remove tq_scheduler

On Sat, 18 Nov 2000, Andrew Morton wrote:

>
> This patch removes tq_scheduler from the kernel. All uses of
> tq_scheduler are migrated over to use schedule_task().
>
> Notes:
> - If anyone sleeps in a callback, then all other users of
> schedule_task() also sleep. But there's nothing new here. Kinda
> makes one wonder why schedule_task exists. But what-hey, it's neat.

Because you're only supposed to use it if the task that is scheduled:
A) Doesn't care about a reasonable delay
B) Doesn't sleep for an unreasonable amount of time.

As long as there's some value of 'reasonable' to match the set of tasks
which you are using schedule_task() for at any given moment, you should be
fine.

If it's really necessary in 2.5, we can consider using multiple queues to
get round this problem - either a task per subsystem or a pool of worker
threads. Hopefully it won't be necessary though. We'll see.

> - Note the careful massaging of module reference counts.
>
> Yes my friends, much usage of task queues in modules is racy wrt
> module removal. This patch fixes some of them.

Cool. I was going to look into that. I had figured we should fix it
completely or not at all, though, which is why I didn't do the trick with
use counts. I probably should have done, though.

While you're in maintenance mode, do you feel like fixing up stuff to use
up_and_exit() for killing kernel threads? I started on net/sunrpc/sched.c
but it made my head hurt so I gave up and started hacking PCMCIA
instead :)

Also, drivers/usb/hub.c can probably use schedule_task() now instead of
its own kernel thread.

--
dwmw2


2000-11-20 17:37:02

by Alan Cox

[permalink] [raw]
Subject: Re: [patch] Remove tq_scheduler

> The patch against test11-pre7 (1043 lines) is at
>
> http://www.uow.edu.au/~andrewm/linux/tq_scheduler.patch
>
> It affects the following files:

Andrew, can you put your patches on a properly configured site. The site
admins have all ICMP packets blocked on http://www.uow.edu.au so you are a PMTU
blackhole and unreachable via my tunnel.

Alan

2000-11-20 18:28:27

by bert hubert

[permalink] [raw]
Subject: Re: [patch] Remove tq_scheduler

On Mon, Nov 20, 2000 at 05:07:14PM +0000, Alan Cox wrote:
> > The patch against test11-pre7 (1043 lines) is at
> >
> > http://www.uow.edu.au/~andrewm/linux/tq_scheduler.patch
> >
> > It affects the following files:
>
> Andrew, can you put your patches on a properly configured site. The site
> admins have all ICMP packets blocked on http://www.uow.edu.au so you are a PMTU
> blackhole and unreachable via my tunnel.

Or configure your host to do MSS Clamping. I have a similar situation over
here and no problems.

Regards,

bert hubert

--
PowerDNS Versatile DNS Services
Trilab The Technology People
'SYN! .. SYN|ACK! .. ACK!' - the mating call of the internet