Re: Deadlocks / hangs in ZFS

On Mon, May 28, 2018 at 09:02:01AM +0200, Alexander Leidinger wrote:

> Quoting Slawa Olhovchenkov <slw_at_zxy.spb.ru> (from Mon, 28 May 2018  
> 01:06:12 +0300):
> 
> > On Sun, May 27, 2018 at 09:41:59PM +0200, Kirill Ponomarev wrote:
> >
> >> On 05/22, Slawa Olhovchenkov wrote:
> >> > > It has been a while since I tried Karl's patch the last time, and I
> >> > > stopped because it didn't apply to -current anymore at some point.
> >> > > Will what is provided right now in the patch work on -current?
> >> >
> >> > I am mean yes, after s/vm_cnt.v_free_count/vm_free_count()/g
> >> > I am don't know how to have two distinct patch (for stable and  
> >> current) in one review.
> >>
> >> I'm experiencing these issues sporadically as well, would you mind
> >> to publish this patch for fresh current?
> >
> > Week ago I am adopt and publish patch to fresh current and stable, is
> > adopt need again?
> 
> I applied the patch in the review yesterday to rev 333966, it applied  
> OK (with some fuzz). I will try to reproduce my issue with the patch.
> 
> Some thoughts I had after looking a little bit at the output of top...  
> half of the RAM of my machine is in use, the other half is listed as  
> free. Swap gets used while there is plenty of free RAM. I have NUMA in  
> my kernel (it's 2 socket Xeon system). I don't see any NUMA specific  
> code in the diff (and I don't expect something there), but could it be  
> that some NUMA related behavior comes into play here too? Does it make  
> sense to try without NUMA in the kernel?

Good question, NUMA in FreeBSD too new, nobody know it.
For Linux, some effectt exists: exhaust all memory in one NUMA domain
can cause memory deficit (swap/allocation failure/etc) simultaneous
with many free memory in other NUMA domain.

Yes, try w/o NUMA, this is may be interesting for NUMA developers.