When the NFSv4 server tells us the lease period, we usually want
to adjust down the timeout parameters on the TCP connection to
ensure that we don't miss lease renewals due to a faulty connection.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
A new argument is added to __skb_recv_datagram to provide
an explicit skb destructor, invoked under the receive queue
lock.
The UDP protocol uses such argument to perform memory
reclaiming on dequeue, so that the UDP protocol does not
set anymore skb->desctructor.
Instead explicit memory reclaiming is performed at close() time and
when skbs are removed from the receive queue.
The in kernel UDP protocol users now need to call a
skb_recv_udp() variant instead of skb_recv_datagram() to
properly perform memory accounting on dequeue.
Overall, this allows acquiring only once the receive queue
lock on dequeue.
Tested using pktgen with random src port, 64 bytes packet,
wire-speed on a 10G link as sender and udp_sink as the receiver,
using an l4 tuple rxhash to stress the contention, and one or more
udp_sink instances with reuseport.
nr sinks vanilla patched
1 440 560
3 2150 2300
6 3650 3800
9 4450 4600
12 6250 6450
v1 -> v2:
- do rmem and allocated memory scheduling under the receive lock
- do bulk scheduling in first_packet_length() and in udp_destruct_sock()
- avoid the typdef for the dequeue callback
Suggested-by: Eric Dumazet <edumazet@google.com>
Acked-by: Hannes Frederic Sowa <hannes@stressinduktion.org>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Completely avoid default sock memory accounting and replace it
with udp-specific accounting.
Since the new memory accounting model encapsulates completely
the required locking, remove the socket lock on both enqueue and
dequeue, and avoid using the backlog on enqueue.
Be sure to clean-up rx queue memory on socket destruction, using
udp its own sk_destruct.
Tested using pktgen with random src port, 64 bytes packet,
wire-speed on a 10G link as sender and udp_sink as the receiver,
using an l4 tuple rxhash to stress the contention, and one or more
udp_sink instances with reuseport.
nr readers Kpps (vanilla) Kpps (patched)
1 170 440
3 1250 2150
6 3000 3650
9 4200 4450
12 5700 6250
v4 -> v5:
- avoid unneeded test in first_packet_length
v3 -> v4:
- remove useless sk_rcvqueues_full() call
v2 -> v3:
- do not set the now unsed backlog_rcv callback
v1 -> v2:
- add memory pressure support
- fixed dropwatch accounting for ipv6
Acked-by: Hannes Frederic Sowa <hannes@stressinduktion.org>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Write space becoming available may race with putting the task to sleep
in xprt_wait_for_buffer_space(). The existing mechanism to avoid the
race does not work.
This (edited) partial trace illustrates the problem:
[1] rpc_task_run_action: task:43546@5 ... action=call_transmit
[2] xs_write_space <-xs_tcp_write_space
[3] xprt_write_space <-xs_write_space
[4] rpc_task_sleep: task:43546@5 ...
[5] xs_write_space <-xs_tcp_write_space
[1] Task 43546 runs but is out of write space.
[2] Space becomes available, xs_write_space() clears the
SOCKWQ_ASYNC_NOSPACE bit.
[3] xprt_write_space() attemts to wake xprt->snd_task (== 43546), but
this has not yet been queued and the wake up is lost.
[4] xs_nospace() is called which calls xprt_wait_for_buffer_space()
which queues task 43546.
[5] The call to sk->sk_write_space() at the end of xs_nospace() (which
is supposed to handle the above race) does not call
xprt_write_space() as the SOCKWQ_ASYNC_NOSPACE bit is clear and
thus the task is not woken.
Fix the race by resetting the SOCKWQ_ASYNC_NOSPACE bit in xs_nospace()
so the second call to sk->sk_write_space() calls xprt_write_space().
Suggested-by: Trond Myklebust <trondmy@primarydata.com>
Signed-off-by: David Vrabel <david.vrabel@citrix.com>
cc: stable@vger.kernel.org # 4.4
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
xprtrdma needs to allocate the Call and Reply buffers separately.
TBH, the reliance on using a single buffer for the pair of XDR
buffers is transport implementation-specific.
Instead of passing just the rq_buffer into the buf_free method, pass
the task structure and let buf_free take care of freeing both
XDR buffers at once.
There's a micro-optimization here. In the common case, both
xprt_release and the transport's buf_free method were checking if
rq_buffer was NULL. Now the check is done only once per RPC.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
xprtrdma needs to allocate the Call and Reply buffers separately.
TBH, the reliance on using a single buffer for the pair of XDR
buffers is transport implementation-specific.
Transports that want to allocate separate Call and Reply buffers
will ignore the "size" argument anyway. Don't bother passing it.
The buf_alloc method can't return two pointers. Instead, make the
method's return value an error code, and set the rq_buffer pointer
in the method itself.
This gives call_allocate an opportunity to terminate an RPC instead
of looping forever when a permanent problem occurs. If a request is
just bogus, or the transport is in a state where it can't allocate
resources for any request, there needs to be a way to kill the RPC
right there and not loop.
This immediately fixes a rare problem in the backchannel send path,
which loops if the server happens to send a CB request whose
call+reply size is larger than a page (which it shouldn't do yet).
One more issue: looks like xprt_inject_disconnect was incorrectly
placed in the failure path in call_allocate. It needs to be in the
success path, as it is for other call-sites.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
The commit f9b2ee714c ("SUNRPC: Move UDP receive data path
into a workqueue context"), as a side effect, moved the
skb_free_datagram() call outside the scope of the related socket
lock, but UDP sockets require such lock to be held for proper
memory accounting.
Fix it by replacing skb_free_datagram() with
skb_free_datagram_locked().
Fixes: f9b2ee714c ("SUNRPC: Move UDP receive data path into a workqueue context")
Reported-and-tested-by: Jan Stancek <jstancek@redhat.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Cc: stable@vger.kernel.org # 4.4+
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
When the connect attempt fails and backs off, we should start the clock
at the last connection attempt, not time at which we queue up the
reconnect job.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
If the net.ipv6.conf.*.use_temp_addr sysctl is set to '2',
then TCP connections over IPv6 will prefer a 'private' source
address.
These eventually expire and become invalid, typically after a week,
but the time is configurable.
When the local address becomes invalid the client will not be able to
receive replies from the server. Eventually the connection will timeout
or break and a new connection will be established, but this can take
half an hour (typically TCP connection break time).
RFC 4941, which describes private IPv6 addresses, acknowledges that some
applications might not work well with them and that the application may
explicitly a request non-temporary (i.e. "public") address.
I believe this is correct for SUNRPC clients. Without this change, a
client will occasionally experience a long delay if private addresses
have been enabled.
The privacy offered by private addresses is of little value for an NFS
server which requires client authentication.
For NFSv3 this will often not be a problem because idle connections are
closed after 5 minutes. For NFSv4 connections never go idle due to the
period RENEW (or equivalent) request.
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
If the connect attempt immediately fails with an EADDRNOTAVAIL error, then
that means our choice of source port number was bad.
This error is expected when we set the SO_REUSEPORT socket option and we
have 2 sockets sharing the same source and destination address and port
combinations.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Fixes: 402e23b4ed ("SUNRPC: Fix stupid typo in xs_sock_set_reuseport")
Cc: stable@vger.kernel.org # v4.0+
The current min/max resvport settings are independently limited
by the entire range of allowed ports, so max_resvport can be
set to a port lower than min_resvport.
Prevent inversion of min/max values when set through sysfs and
module parameter by setting the limits dependent on each other.
Signed-off-by: Frank Sorenson <sorenson@redhat.com>
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
The current min/max resvport settings are independently limited
by the entire range of allowed ports, so max_resvport can be
set to a port lower than min_resvport.
Prevent inversion of min/max values when set through sysctl by
setting the limits dependent on each other.
Signed-off-by: Frank Sorenson <sorenson@redhat.com>
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
The range calculation for choosing the random reserved port will panic
with divide-by-zero when min_resvport == max_resvport, a range of one
port, not zero.
Fix the reserved port range calculation by adding one to the difference.
Signed-off-by: Frank Sorenson <sorenson@redhat.com>
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
The spec allows backchannels for multiple clients to share the same tcp
connection. When that happens, we need to use the same xprt for all of
them. Similarly, we need the same xps.
This fixes list corruption introduced by the multipath code.
Cc: stable@vger.kernel.org
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
Acked-by: Trond Myklebust <trondmy@primarydata.com>
rpciod can easily get congested due to the long list of queued rpc_tasks.
Having the receive queue wait in turn for those tasks to complete can
therefore be a bottleneck.
Address the problem by separating the workqueues into:
- rpciod: manages rpc_tasks
- xprtiod: manages transport related work.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
The only difference between the two at this point is the reset of
the connection timeout, and since everyone expect tcp ignore that value,
we can just throw it into the generic function.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Highlights include:
Features:
- Add support for the NFS v4.2 COPY operation
- Add support for NFS/RDMA over IPv6
Bugfixes and cleanups:
- Avoid race that crashes nfs_init_commit()
- Fix oops in callback path
- Fix LOCK/OPEN race when unlinking an open file
- Choose correct stateids when using delegations in setattr, read and write
- Don't send empty SETATTR after OPEN_CREATE
- xprtrdma: Prevent server from writing a reply into memory client has released
- xprtrdma: Support using Read list and Reply chunk in one RPC call
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v2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=QMmY
-----END PGP SIGNATURE-----
Merge tag 'nfs-for-4.7-1' of git://git.linux-nfs.org/projects/anna/linux-nfs
Pull NFS client updates from Anna Schumaker:
"Highlights include:
Features:
- Add support for the NFS v4.2 COPY operation
- Add support for NFS/RDMA over IPv6
Bugfixes and cleanups:
- Avoid race that crashes nfs_init_commit()
- Fix oops in callback path
- Fix LOCK/OPEN race when unlinking an open file
- Choose correct stateids when using delegations in setattr, read and
write
- Don't send empty SETATTR after OPEN_CREATE
- xprtrdma: Prevent server from writing a reply into memory client
has released
- xprtrdma: Support using Read list and Reply chunk in one RPC call"
* tag 'nfs-for-4.7-1' of git://git.linux-nfs.org/projects/anna/linux-nfs: (61 commits)
pnfs: pnfs_update_layout needs to consider if strict iomode checking is on
nfs/flexfiles: Use the layout segment for reading unless it a IOMODE_RW and reading is disabled
nfs/flexfiles: Helper function to detect FF_FLAGS_NO_READ_IO
nfs: avoid race that crashes nfs_init_commit
NFS: checking for NULL instead of IS_ERR() in nfs_commit_file()
pnfs: make pnfs_layout_process more robust
pnfs: rework LAYOUTGET retry handling
pnfs: lift retry logic from send_layoutget to pnfs_update_layout
pnfs: fix bad error handling in send_layoutget
flexfiles: add kerneldoc header to nfs4_ff_layout_prepare_ds
flexfiles: remove pointless setting of NFS_LAYOUT_RETURN_REQUESTED
pnfs: only tear down lsegs that precede seqid in LAYOUTRETURN args
pnfs: keep track of the return sequence number in pnfs_layout_hdr
pnfs: record sequence in pnfs_layout_segment when it's created
pnfs: don't merge new ff lsegs with ones that have LAYOUTRETURN bit set
pNFS/flexfiles: When initing reads or writes, we might have to retry connecting to DSes
pNFS/flexfiles: When checking for available DSes, conditionally check for MDS io
pNFS/flexfile: Fix erroneous fall back to read/write through the MDS
NFS: Reclaim writes via writepage are opportunistic
NFSv4: Use the right stateid for delegations in setattr, read and write
...
RPC-over-RDMA transports have a limit on how large a backward
direction (backchannel) RPC message can be. Ensure that the NFSv4.x
CREATE_SESSION operation advertises this limit to servers.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Tested-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
sunrpc is using SOCKWQ_ASYNC_NOSPACE without setting SOCK_FASYNC,
so the recent optimizations done in sk_set_bit() and sk_clear_bit()
broke it.
There is still the risk that a subsequent sock_fasync() call
would clear SOCK_FASYNC, but sunrpc does not use this yet.
Fixes: 9317bb6982 ("net: SOCKWQ_ASYNC_NOSPACE optimizations")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Jiri Pirko <jiri@resnulli.us>
Reported-by: Huang, Ying <ying.huang@intel.com>
Tested-by: Jiri Pirko <jiri@resnulli.us>
Tested-by: Huang, Ying <ying.huang@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Rename UDP_INC_STATS_BH() to __UDP_INC_STATS(),
and UDP6_INC_STATS_BH() to __UDP6_INC_STATS()
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
sock_owned_by_user should not be used without socket lock held. It seems
to be a common practice to check .owned before lock reclassification, so
provide a little help to abstract this check away.
Cc: linux-cifs@vger.kernel.org
Cc: linux-bluetooth@vger.kernel.org
Cc: linux-nfs@vger.kernel.org
Signed-off-by: Hannes Frederic Sowa <hannes@stressinduktion.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Commit e6afc8ace6 modified the udp receive path by pulling the udp
header before queuing an skbuff onto the receive queue.
Sunrpc also calls skb_recv_datagram to dequeue an skb from a udp
socket. Modify this receive path to also no longer expect udp
headers.
Fixes: e6afc8ace6 ("udp: remove headers from UDP packets before queueing")
Reported-by: Franklin S Cooper Jr. <fcooper@ti.com>
Signed-off-by: Willem de Bruijn <willemb@google.com>
Tested-by: Thierry Reding <treding@nvidia.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
* bugfixes:
SUNRPC: Fixup socket wait for memory
SUNRPC: Fix a missing break in rpc_anyaddr()
pNFS/flexfiles: Fix an Oopsable typo in ff_mirror_match_fh()
NFS: Fix attribute cache revalidation
NFS: Ensure we revalidate attributes before using execute_ok()
NFS: Flush reclaim writes using FLUSH_COND_STABLE
NFS: Background flush should not be low priority
NFSv4.1/pnfs: Fixup an lo->plh_block_lgets imbalance in layoutreturn
NFSv4: Don't perform cached access checks before we've OPENed the file
NFS: Allow the combination pNFS and labeled NFS
NFS42: handle layoutstats stateid error
nfs: Fix race in __update_open_stateid()
nfs: fix missing assignment in nfs4_sequence_done tracepoint
We're seeing hangs in the NFS client code, with loops of the form:
RPC: 30317 xmit incomplete (267368 left of 524448)
RPC: 30317 call_status (status -11)
RPC: 30317 call_transmit (status 0)
RPC: 30317 xprt_prepare_transmit
RPC: 30317 xprt_transmit(524448)
RPC: xs_tcp_send_request(267368) = -11
RPC: 30317 xmit incomplete (267368 left of 524448)
RPC: 30317 call_status (status -11)
RPC: 30317 call_transmit (status 0)
RPC: 30317 xprt_prepare_transmit
RPC: 30317 xprt_transmit(524448)
Turns out commit ceb5d58b21 ("net: fix sock_wake_async() rcu protection")
moved SOCKWQ_ASYNC_NOSPACE out of sock->flags and into sk->sk_wq->flags,
however it never tried to fix up the code in net/sunrpc.
The new idiom is to use the flags in the RCU protected struct socket_wq.
While we're at it, clear out the now redundant places where we set/clear
SOCKWQ_ASYNC_NOSPACE and SOCK_NOSPACE. In principle, sk_stream_wait_memory()
is supposed to set these for us, so we only need to clear them in the
particular case of our ->write_space() callback.
Fixes: ceb5d58b21 ("net: fix sock_wake_async() rcu protection")
Cc: Eric Dumazet <edumazet@google.com>
Cc: stable@vger.kernel.org # 4.4
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
xs_reclassify_socket4() and friends used to be called directly.
xs_reclassify_socket() is called instead nowadays.
The xs_reclassify_socketX() helper functions are empty when
CONFIG_DEBUG_LOCK_ALLOC is not defined. Drop them since they have no
callers.
Note that AF_LOCAL still calls xs_reclassify_socketu() directly but is
easily converted to generic xs_reclassify_socket().
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
This patch is a cleanup to make following patch easier to
review.
Goal is to move SOCK_ASYNC_NOSPACE and SOCK_ASYNC_WAITDATA
from (struct socket)->flags to a (struct socket_wq)->flags
to benefit from RCU protection in sock_wake_async()
To ease backports, we rename both constants.
Two new helpers, sk_set_bit(int nr, struct sock *sk)
and sk_clear_bit(int net, struct sock *sk) are added so that
following patch can change their implementation.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Due to incorrect len type bc_send_request returned always zero.
The problem has been detected using proposed semantic patch
scripts/coccinelle/tests/assign_signed_to_unsigned.cocci [1].
[1]: http://permalink.gmane.org/gmane.linux.kernel/2046107
Signed-off-by: Andrzej Hajda <a.hajda@samsung.com>
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
In addition to a variety of bugfixes, these patches are mostly geared at
enabling both swap and backchannel support to the NFS over RDMA client.
Signed-off-by: Anna Schumake <Anna.Schumaker@Netapp.com>
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v2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=Ush5
-----END PGP SIGNATURE-----
Merge tag 'nfs-rdma-4.4-2' of git://git.linux-nfs.org/projects/anna/nfs-rdma
NFS: NFSoRDMA Client Side Changes
In addition to a variety of bugfixes, these patches are mostly geared at
enabling both swap and backchannel support to the NFS over RDMA client.
Signed-off-by: Anna Schumake <Anna.Schumaker@Netapp.com>
Forechannel transports get their own "bc_up" method to create an
endpoint for the backchannel service.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
[Anna Schumaker: Add forward declaration of struct net to xprt.h]
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
xprt_{setup,destroy}_backchannel() won't be adequate for RPC/RMDA
bi-direction. In particular, receive buffers have to be pre-
registered and posted in order to receive incoming backchannel
requests.
Add a virtual function call to allow the insertion of appropriate
backchannel setup and destruction methods for each transport.
In addition, freeing a backchannel request is a little different
for RPC/RDMA. Introduce an rpc_xprt_op to handle the difference.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Reviewed-by: Sagi Grimberg <sagig@mellanox.com>
Tested-By: Devesh Sharma <devesh.sharma@avagotech.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
If we're sending more than one page via kernel_sendpage(), then set
MSG_SENDPAGE_NOTLAST between the pages so that we don't send suboptimal
frames (see commit 2f53384424 and commit 35f9c09fe9).
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Stream protocols such as TCP can often build up a backlog of data to be
read due to ordering. Combine this with the fact that some workloads such
as NFS read()-intensive workloads need to receive a lot of data per RPC
call, and it turns out that receiving the data from inside a softirq
context can cause starvation.
The following patch moves the TCP data receive into a workqueue context.
We still end up calling tcp_read_sock(), but we do so from a process
context, meaning that softirqs are enabled for most of the time.
With this patch, I see a doubling of read bandwidth when running a
multi-threaded iozone workload between a virtual client and server setup.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Move the TCP data receive loop out of xs_tcp_data_ready(). Doing so
will allow us to move the data receive out of the softirq context in
a set of followup patches.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Under all conditions, it should be quite sufficient just to mark
the socket as disconnected. It will then be closed by the
transport shutdown or reconnect code.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Commit 718ba5b873, moved the responsibility for unlocking the socket to
xs_tcp_setup_socket, meaning that the socket will be unlocked before we
know that it has finished trying to connect. The following patch is based on
an initial patch by Russell King to ensure that we delay clearing the
XPRT_CONNECTING flag until we either know that we failed to initiate
a connection attempt, or the connection attempt itself failed.
Fixes: 718ba5b873 ("SUNRPC: Add helpers to prevent socket create from racing")
Reported-by: Russell King <linux@arm.linux.org.uk>
Reported-by: Russell King <rmk+kernel@arm.linux.org.uk>
Tested-by: Russell King <rmk+kernel@arm.linux.org.uk>
Tested-by: Benjamin Coddington <bcodding@redhat.com>
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
When we're destroying the socket transport, we need to ensure that
we cancel any existing delayed connection attempts, and order them
w.r.t. the call to xs_close().
Reported-by:"Suzuki K. Poulose" <suzuki.poulose@arm.com>
Acked-by: Jeff Layton <jlayton@poochiereds.net>
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Add a shutdown() call before we release the socket in order to ensure the
reset is sent before we try to reconnect.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Follow up to commit c4a7ca7749 ("SUNRPC: Allow waiting on memory
allocation"). Allows the RPC socket code to do non-IO blocking.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
It is rather pointless to test the value of transport->inet after
calling xs_reset_transport(), since it will always be zero, and
so we will never see any exponential back off behaviour.
Also don't force early connections for SOFTCONN tasks. If the server
disconnects us, we should respect the exponential backoff.
Cc: stable@vger.kernel.org # 4.0+
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Trond Myklebust