Re: [PATCH v3 2.6.39-rc1-tip 12/26] 12: uprobes: slot allocationfor uprobes

[Peter Zijlstra]

(Dropped the systemtap list since its mis-behaving, please leave it out on future postings)

On Tue, 2011-04-19 at 11:56 +0530, Srikar Dronamraju wrote:
> > > TODO: On massively threaded processes (or if a huge number of processes
> > > share the same mm), there is a possiblilty of running out of slots.
> > > One alternative could be to extend the slots as when slots are required.
> > 
> > As long as you're single stepping things and not using boosted probes
> > you can fully serialize the slot usage. Claim a slot on trap and release
> > the slot on finish. Claiming can wait on a free slot since you already
> > have the whole SLEEPY thing.
> > 
> 
> Yes, thats certainly one approach but that approach makes every
> breakpoint hit contend for spinlock. (Infact we will have to change it
> to mutex lock (as you rightly pointed out) so that we allow threads to
> wait when slots are not free). Assuming a 4K page, we would be taxing
> applications that have less than 32 threads (which is probably the
> default case). If we continue with the current approach, then we
> could only add additional page(s) for apps which has more than 32
> threads and only when more than 32 __live__ threads have actually hit a
> breakpoint. 

That very much depends on what you do, some folks think its entirely
reasonable for processes to have thousands of threads. Now I completely
agree with you that that is not 'normal', but then I think using Java
isn't normal either ;-)

Anyway, avoiding that spinlock/mutex for each trap isn't hard, avoiding
a process wide cacheline bounce is slightly harder but still not
impossible. 

With 32 slots in 4k you have 128 bytes to play with, all we need is a
single bit per slot to mark it being in-use. If a task remembers what
slot it used last and tries to claim that using an atomic test and set
for that bit it will, in the 'normal' case, never contend on a process
wide cacheline.

In case it does find the slot taken, it'll have to go the slow route and
scan for a free slot and possibly wait for one to become free.

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/