Re: dm-crypt using kthread

[Jeff Garzik <jgarzik@pobox.com>]

Christophe Saout wrote:
> Am Mo, den 16.02.2004 schrieb Jeff Garzik um 03:10:
> 
> 
>>>+		/*
>>>+		 * if additional pages cannot be allocated without waiting,
>>>+		 * return a partially allocated bio, the caller will then try
>>>+		 * to allocate additional bios while submitting this partial bio
>>>+		 */
>>>+		if ((i - bio->bi_idx) == (MIN_BIO_PAGES - 1))
>>>+			gfp_mask = (gfp_mask | __GFP_NOWARN) & ~__GFP_WAIT;
>>
>>If the caller said they can wait, why not wait?
> 
> 
> How can the caller say this?

There is a gfp_mask there :)


>>>+		set_task_state(current, TASK_INTERRUPTIBLE);
>>>+		while (!(bio = kcryptd_get_bios())) {
>>>+			schedule();
>>>+			if (signal_pending(current))
>>>+				return 0;
>>>+		}
>>>+		set_task_state(current, TASK_RUNNING);
>>
>>You just keep calling schedule() rapid-fire until you get a bio?  That's 
>>a bit sub-optimal.
> 
> 
> That's wrong anyway. I was just making sure I was calling
> kcryptd_get_bios after schedule. schedule() will sleep and woken after
> someone added a bio to the list.
> 
> I've changed it to an if now and call kcryptd_get_bios after schedule.
> 
> I'm calling it twice because it is likely that someone started a new
> list while the old list is being processed and I don't want to sleep in
> this case, just fall through.
> 
> The kcryptd_get_bios needs to be after state = TASK_INTERRUPTIBLE to
> avoid a race. If someone wakes the process after kcryptd_get_bios but
> before schedule it resets the state to TASK_RUNNING so that the schedule
> won't sleep.

This sounds like a lot of work, just to reimplement what a semaphore 
does for you :)

When you down(), you sleep, waiting for new work.  Each time new work 
occurs, on any cpu, you call up().  This is perfect for a kernel thread, 
which can sleep, just like a semaphore needs.  If you want to be 
interruptible by signals (such as sysadmin killing your thread, for some 
reason), then use down_interruptible().

There is typically one special case -- killing your thread on shutdown. 
  The typical solution is to set a flag thread_shutdown, and then up().

	Jeff