Description of problem:
We get on the client side a memory leak on the size-8192 buckets when running
under a NFSV4 mount the bench iozone with -U option in a infinite loop.
We can see with "slabtop -s c" the number of size-8192 buckets constantly
increases without never decreases.
At the beginning :
33 33 100% 8.00K 33 1 264K size-8192
After half and hour:
89 89 100% 8.00K 89 1 712K size-8192
After one hour:
174 172 98% 8.00K 174 1 1392K size-8192
After two hours:
302 300 99% 8.00K 302 1 2416K size-8192
After four hours:
533 533 100% 8.00K 533 1 4264K size-8192
After six hours and half:
804 804 100% 8.00K 804 1 6432K size-8192
Version-Release number of selected component (if applicable):
Linux version 2.6.18-53.el5 (firstname.lastname@example.org) (gcc
version 4.1.2 20070626 (Red Hat 4.1.2-14)) #1 SMP Wed Oct 10 16:34:19 EDT 2007
Do the nfsv4 mount and run with the mounted directory:
It is possible to reproduce it on one machine with a NFSV4 mount in loopback.
For example on the machine nfs1gb machine:
mount -t nfs4 nfs1gb:/ /mnt/nosec
while true; do ./iozone -+q 30 -ace -r 64 -i 0 -i 1 -i 2 -f
/mnt/nosec/nfs1_nfs4_gb -U /mnt/nosec; date; sleep 30; done
The nfs1gb machine used is a Intel X86_64 64bits two-ways.
Steps to Reproduce:
Finally got a chance to look over this today. I see a similar memory leak when
just mounting and unmounting a NFS4 share in a loop:
# for i in `seq 1 100`; do mount /mnt/rhel4; umount /mnt/rhel4; done
...after this, the size-8192 slab gains ~100 more active objects. I'll have to
ponder how we can best track down what's actually doing these kmallocs. Maybe a
systemtap script that traps on kmalloc and does a dump_stack for any that are
>4096 and <8192 ?
Created attachment 294150 [details]
stap script for looking at size 8192 kmallocs and their corresponding kfrees
Systemtap script to try and track this down...
Created attachment 294151 [details]
output from stap script
Output from stap script. It looks like we have 6 kmallocs and 5 kfrees. The
lingering kmalloc seems to be the second one that returned 0xffff8800025b8000.
Stack trace from print_backtrace is:
size = 4120, addr = 0xffff8800025b8000
Returning from: 0xffffffff802c9899 : __kmalloc+0x0/0x9f 
Returning to : 0xffffffff802bd91a : __kzalloc+0x9/0x21 
0xffffffff802774b8 : kretprobe_trampoline_holder+0x0/0x2 
0xffffffff80404f07 : reqsk_queue_alloc+0x21/0x99 
0xffffffff8042bc7b : inet_csk_listen_start+0x1a/0x135 
0xffffffff8043c59b : inet_listen+0x42/0x68 
0xffffffff881e2464 : svc_makesock+0x127/0x183 [sunrpc]
0xffffffff881e18b7 : svc_create+0xee/0xf8 [sunrpc]
0xffffffff883711bf : nfs_callback_up+0x9c/0x14d [nfs]
0xffffffff8834fe2f : nfs_get_client+0xfd/0x3df [nfs]
0xffffffff88350158 : nfs4_set_client+0x47/0x173 [nfs]
0xffffffff88350909 : nfs4_create_server+0x7a/0x393 [nfs]
0xffffffff8025e823 : error_exit+0x0/0x6e 
0xffffffff883573b4 : nfs_copy_user_string+0x3c/0x89 [nfs]
0xffffffff88357cdc : nfs4_get_sb+0x1fc/0x323 [nfs]
0xffffffff8020adff : get_page_from_freelist+0x32e/0x3bc 
0xffffffff802cee21 : vfs_kern_mount+0x93/0x11a 
0xffffffff802ceeea : do_kern_mount+0x36/0x4d 
0xffffffff802d855b : do_mount+0x68c/0x6fc 
0xffffffff80418c8b : __qdisc_run+0x36/0x1bb 
0xffffffff8022bf6b : local_bh_enable+0x9/0xa5 
0xffffffff80230ebb : dev_queue_xmit+0x2f2/0x313 
0xffffffff80233001 : ip_output+0x29a/0x2dd 
0xffffffff802628b1 : _spin_lock_irqsave+0x9/0x14 
0xffffffff802229d4 : __up_read+0x19/0x7f 
0xffffffff802d72eb : copy_mount_options+0xce/0x127 
0xffffffff80297b70 : search_exception_tables+0x1d/0x2d 
0xffffffff802654f6 : do_page_fault+0x10e7/0x12cc 
0xffffffff80263786 : do_debug+0x70/0x151 
0xffffffff8020b663 : kfree+0x0/0xc5 
0xffffffff80263f9d : kprobe_handler+0x1ac/0x1c8 
0xffffffff80263ff4 : kprobe_exceptions_notify+0x3b/0x75 
0xffffffff802656fb : notifier_call_chain+0x20/0x32 
0xffffffff8020acc5 : get_page_from_freelist+0x1f4/0x3bc 
0xffffffff8020f2d0 : __alloc_pages+0x65/0x2ce 
0xffffffff8024c3cf : sys_mount+0x8a/0xcd 
0xffffffff8025e2f1 : tracesys+0xa7/0xb2 
I think there's a lot of garbage in there though, but that gives me some idea
of where to look...
It looks like we're calling svc_setup_socket to create a socket for the nfs4
callback thread, but I don't see where that gets torn down. I suspect that's
where the problem is, but need to look a bit more closely.
The problem seems to be with the sv_nrthreads count for the callback thread.
It's at 2 when we do the umount:
RPC: svc_destroy(NFSv4 callback, 2)
...so it doesn't actually tear down the socket or the svc_serv. nfsd and lockd
also use those functions and when they go down their refcounts seem to be OK...
This appears to be an upstream bug too. On a rawhide machine after mounting and
svc: svc_destroy(NFSv4 callback, 2)
...and I don't see where the socket got torn down. I think I see the problem,
svc_create starts the service with sv_nrthreads==1. Then, svc_create_thread
increments that count.
nfs_callback_up() isn't handling this correctly. It should be calling
svc_destroy() on success and failure, but it isn't. As an example,
lockd_up_proto() handles this correctly. I'll post a patch here soon that I can
propose upstream to fix this.
Created attachment 294563 [details]
patch -- fix reference counting for NFS4 callback thread
This patch seems to fix the problem on rawhide. Backporting to RHEL5 and RHEL4
should be trivial. I'm assuming RHEL4 has this problem too, though I need to
check. I'll clone this BZ if so.
Patch posted upstream. Awaiting comment...
Looks like Trond applied the patch, so I'll plan to propose a similar one for
RHEL5. It's a bit too late for RHEL5.2, but I'll try to make sure we get
something for 5.3.
I'll also plan to take this patch into my test kernels for you to test. Once I
do, I'll post a note here so that you can test them.
Created attachment 295296 [details]
patch -- flush signals before taking down callback thread
Peter noticed that this seems to expose another problem with the callback
thread. It doesn't flush signals on shutdown and that makes the portmap
unregistration fail and throw an error. This patch seems to fix it, but he's
currently chasing an NFS related deadlock and I'd like to understand that
before I send this upstream.
Created attachment 302874 [details]
Patch that incorporates both patches that went upstream.
This request was evaluated by Red Hat Product Management for inclusion in a Red
Hat Enterprise Linux maintenance release. Product Management has requested
further review of this request by Red Hat Engineering, for potential
inclusion in a Red Hat Enterprise Linux Update release for currently deployed
products. This request is not yet committed for inclusion in an Update
You can download this test kernel from http://people.redhat.com/dzickus/el5
Partners, this bug should be fixed in the latest RHEL 5.3 Snapshot. We believe that you have some interest in its correct functionality, so we're making a friendly request to send us some testing feedback.
If you have a chance to test it, please share with us your findings. If you have successfully VERIFIED the fix, please add PartnerVerified to the Bugzilla keywords, along with a description of the results. Thanks!
~~ Snapshot 6 is out ~~ Partners, please test and let us know if this bug has been fixed. Add PartnerVerified keyword if everything works as expected. For any new issues encountered, CLONE this bug and report the issues in the new bug.
An advisory has been issued which should help the problem
described in this bug report. This report is therefore being
closed with a resolution of ERRATA. For more information
on therefore solution and/or where to find the updated files,
please follow the link below. You may reopen this bug report
if the solution does not work for you.