Re: More information on scsi_cmd_cache leak... (bisect)

[Jens Axboe <axboe@suse.de>]

On Fri, Jan 27 2006, Mike Christie wrote:
> Jens Axboe wrote:
> >On Fri, Jan 27 2006, Jens Axboe wrote:
> >
> >>On Fri, Jan 27 2006, Neil Brown wrote:
> >>
> >>>On Friday January 27, chase.venters@clientec.com wrote:
> >>>
> >>>>Greetings,
> >>>>	Just a quick recap - there are at least 4 reports of 2.6.15 users 
> >>>>experiencing severe slab leaks with scsi_cmd_cache. It seems that a few 
> >>>>of us have a board (Asus P5GDC-V Deluxe) in common. We seem to have 
> >>>>raid in common. After dealing with this leak for a while, I decided 
> >>>>	to do some dancing around with git bisect. I've landed on a possible 
> >>>>point of regression:
> >>>>
> >>>>commit: a9701a30470856408d08657eb1bd7ae29a146190
> >>>>[PATCH] md: support BIO_RW_BARRIER for md/raid1
> >>>>
> >>>>	I spent about an hour and a half reading through the patch, trying 
> >>>>	to see if I could make sense of what might be wrong. The result (after 
> >>>>I dug into the code to make a change I foolishly thought made sense) 
> >>>>was a hung kernel.
> >>>>	This is important because when I rebooted into the kernel that had 
> >>>>	been giving me trouble, it started an md resync and I'm now watching 
> >>>>(at least during this resync) the slab usage for scsi_cmd_cache stay 
> >>>>sane:
> >>>>
> >>>>turbotaz ~ # cat /proc/slabinfo | grep scsi_cmd_cache
> >>>>scsi_cmd_cache        30     30    384   10    1 : tunables   54   27   
> >>>>8 : slabdata      3      3      0
> >>>>
> >>>
> >>>This suggests that the problem happens when a BIO_RW_BARRIER write is
> >>>sent to the device.  With this patch, md flags all superblock writes
> >>>as BIO_RW_BARRIER However md is not so likely to update the superblock 
> >>>often
> >>>during a resync.
> >>>
> >>>There is a (rough) count of the number of superblock writes in the
> >>>"Events" counter which "mdadm -D" will display.
> >>>You could try collecting 'Events' counter together with the
> >>>'active_objs' count from /proc/slabinfo and graph the pairs - see if
> >>>they are linear.
> >>>
> >>>I believe a BIO_RW_BARRIER is likely to send some sort of 'flush'
> >>>command to the device, and the driver for your particular device may
> >>>well be losing scsi_cmd_cache allocation when doing that, but I leave
> >>>that to someone how knows more about that code.
> >>
> >>I already checked up on that since I suspected barriers initially. The
> >>path there for scsi is sd.c:sd_issue_flush() which looks pretty straight
> >>forward. In the end it goes through the block layer and gets back to the
> >>SCSI layer as a regular REQ_BLOCK_PC request.
> >
> >
> >Sorry, that was for the ->issue_flush() that md also does but did before
> >the barrier addition as well. Most of the barrier handling is done in
> >the block layer, but it could show leaks in SCSI of course. FWIW, I
> >tested barriers with and without md on SCSI here a few days ago and
> >didn't see any leaks at all.
> >
> 
> It does not have anything to do with this in scsi_io_completion does it?
> 
>         if (blk_complete_barrier_rq(q, req, good_bytes >> 9))
>                 return;
> 
> For that case the scsi_cmnd does not get freed. Does it come back around 
> again and get released from a different path?

Certainly smells fishy. Unfortunately I cannot take a look at this until
monday :/

But adding some tracing there might be really interesting. Since we are
not seeing bio and/or req leaks, this does look very promising.

-- 
Jens Axboe