On 12/18/2010 09:13 PM, Mike Galbraith wrote:
On Sat, 2010-12-18 at 19:08 +0200, Avi Kivity wrote: > On 12/17/2010 09:15 AM, Mike Galbraith wrote: > > BTW, with this vruntime donation thingy, what prevents a task from > > forking off accomplices who do nothing but wait for a wakeup and > > yield_to(exploit)? > > > > What's the difference between that and forking off accomplices who > run(exploit) directly? The clock still advances. What I described is a clock stopper.
That's scheduler jargon, I'm not familiar with scheduler internals. Let's talk about this at a high level, since whenever I describe it in scheduler terms I'm likely to get it wrong.
If there are N tasks on the machine, each is entitled to 1/N of the machine's cpu resources (ignoring nice and cgroups for the moment). What I'd like is for one task to temporarily pass a portion of its entitlement to another. No other task's entitlement is affected; they still get their 1/N.
If a task donates its entitlement and immediately commits suicide, still we won't have fundamentally changed anything; the task could have kept itself alive and consumed that entitlement, so the scheduler was already prepared to give it this entitlement so nothing was gained by the forking task. The problem of fork() creating new entitlements out of thin air has nothing to do with directed yield, and is solved by cgroups.
We already do something similar with priority inheritance. This doesn't involve cpu entitlement, but we have the same situation where a task's ability to make progress depends on another task receiving cpu time.
-- I have a truly marvellous patch that fixes the bug which this signature is too narrow to contain. -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
