NFSv4.1 callbacks can arrive over an existing connection. This patch adds
the logic to read the RPC call direction (call or reply). It does this by
updating the state machine to look for the call direction invoking
xs_tcp_read_calldir(...) after reading the XID.
[nfs41: Keep track of RPC call/reply direction with a flag]
As per 11/14/08 review of RFC 53/85.
Add a new flag to track whether the incoming message is an RPC call or an
RPC reply. TCP_RPC_REPLY is set in the 'struct sock_xprt' tcp_flags in
xs_tcp_read_calldir() if the message is an RPC reply sent on the forechannel.
It is cleared if the message is an RPC request sent on the back channel.
Signed-off-by: Ricardo Labiaga <Ricardo.Labiaga@netapp.com>
Signed-off-by: Benny Halevy <bhalevy@panasas.com>
* 'for-2.6.30' of git://linux-nfs.org/~bfields/linux:
svcrdma: dma unmap the correct length for the RPCRDMA header page.
nfsd: Revert "svcrpc: take advantage of tcp autotuning"
nfsd: fix hung up of nfs client while sync write data to nfs server
The svcrdma module was incorrectly unmapping the RPCRDMA header page.
On IBM pserver systems this causes a resource leak that results in
running out of bus address space (10 cthon iterations will reproduce it).
The code was mapping the full page but only unmapping the actual header
length. The fix is to only map the header length.
I also cleaned up the use of ib_dma_map_page() calls since the unmap
logic always uses ib_dma_unmap_single(). I made these symmetrical.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Tom Tucker <tom@opengridcomputing.com>
Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
This reverts commit 47a14ef1af "svcrpc:
take advantage of tcp autotuning", which uncovered some further problems
in the server rpc code, causing significant performance regressions in
common cases.
We will likely reinstate this patch after releasing 2.6.30 and applying
some work on the underlying fixes to the problem (developed by Trond).
Reported-by: Jeff Moyer <jmoyer@redhat.com>
Cc: Olga Kornievskaia <aglo@citi.umich.edu>
Cc: Jim Rees <rees@umich.edu>
Cc: Trond Myklebust <trond.myklebust@fys.uio.no>
Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
mlx4/connectX FRMR requires local write enable together with remote
rdma write enable. This fixes NFS/RDMA operation over the ConnectX
Infiniband HCA in the default memreg mode.
Signed-off-by: Vu Pham <vu@mellanox.com>
Signed-off-by: Tom Talpey <tmtalpey@gmail.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
* 'for-2.6.30' of git://linux-nfs.org/~bfields/linux:
nfsd: silence lockdep warning
lockd: fix list corruption on lockd restart
nfsd4: check for negative dentry before use in nfsv4 readdir
nfsd41: slots are freed with session
svcrdma: clean up error paths.
svcrdma: Fix dma map direction for rdma read targets
These fixes resolved crashes due to resource leak BUG_ON checks. The
resource leaks were detected by introducing asynchronous transport errors.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Tom Tucker <tom@opengridcomputing.com>
Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
See http://bugzilla.kernel.org/show_bug.cgi?id=13034
If the port gets into a TIME_WAIT state, then we cannot reconnect without
binding to a new port.
Tested-by: Petr Vandrovec <petr@vandrovec.name>
Tested-by: Jean Delvare <khali@linux-fr.org>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The nfs server rdma transport was mapping rdma read target pages for
TO_DEVICE instead of FROM_DEVICE. This causes data corruption on non
cache-coherent systems if frmrs are used.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
* 'for-2.6.30' of git://linux-nfs.org/~bfields/linux: (81 commits)
nfsd41: define nfsd4_set_statp as noop for !CONFIG_NFSD_V4
nfsd41: define NFSD_DRC_SIZE_SHIFT in set_max_drc
nfsd41: Documentation/filesystems/nfs41-server.txt
nfsd41: CREATE_EXCLUSIVE4_1
nfsd41: SUPPATTR_EXCLCREAT attribute
nfsd41: support for 3-word long attribute bitmask
nfsd: dynamically skip encoded fattr bitmap in _nfsd4_verify
nfsd41: pass writable attrs mask to nfsd4_decode_fattr
nfsd41: provide support for minor version 1 at rpc level
nfsd41: control nfsv4.1 svc via /proc/fs/nfsd/versions
nfsd41: add OPEN4_SHARE_ACCESS_WANT nfs4_stateid bmap
nfsd41: access_valid
nfsd41: clientid handling
nfsd41: check encode size for sessions maxresponse cached
nfsd41: stateid handling
nfsd: pass nfsd4_compound_state* to nfs4_preprocess_{state,seq}id_op
nfsd41: destroy_session operation
nfsd41: non-page DRC for solo sequence responses
nfsd41: Add a create session replay cache
nfsd41: create_session operation
...
* git://git.kernel.org/pub/scm/linux/kernel/git/rusty/linux-2.6-cpumask: (36 commits)
cpumask: remove cpumask allocation from idle_balance, fix
numa, cpumask: move numa_node_id default implementation to topology.h, fix
cpumask: remove cpumask allocation from idle_balance
x86: cpumask: x86 mmio-mod.c use cpumask_var_t for downed_cpus
x86: cpumask: update 32-bit APM not to mug current->cpus_allowed
x86: microcode: cleanup
x86: cpumask: use work_on_cpu in arch/x86/kernel/microcode_core.c
cpumask: fix CONFIG_CPUMASK_OFFSTACK=y cpu hotunplug crash
numa, cpumask: move numa_node_id default implementation to topology.h
cpumask: convert node_to_cpumask_map[] to cpumask_var_t
cpumask: remove x86 cpumask_t uses.
cpumask: use cpumask_var_t in uv_flush_tlb_others.
cpumask: remove cpumask_t assignment from vector_allocation_domain()
cpumask: make Xen use the new operators.
cpumask: clean up summit's send_IPI functions
cpumask: use new cpumask functions throughout x86
x86: unify cpu_callin_mask/cpu_callout_mask/cpu_initialized_mask/cpu_sibling_setup_mask
cpumask: convert struct cpuinfo_x86's llc_shared_map to cpumask_var_t
cpumask: convert node_to_cpumask_map[] to cpumask_var_t
x86: unify 32 and 64-bit node_to_cpumask_map
...
On an NFSv4.1 server cache miss that causes an upcall, NFS4ERR_DELAY will be
returned. It is up to the NFSv4.1 client to resend only the operations that
have not been processed.
Initialize rq_usedeferral to 1 in svc_process(). It sill be turned off in
nfsd4_proc_compound() only when NFSv4.1 Sessions are used.
Note: this isn't an adequate solution on its own. It's acceptable as a way
to get some minimal 4.1 up and working, but we're going to have to find a
way to avoid returning DELAY in all common cases before 4.1 can really be
considered ready.
Signed-off-by: Andy Adamson <andros@netapp.com>
Signed-off-by: Benny Halevy <bhalevy@panasas.com>
[nfsd41: reverse rq_nodeferral negative logic]
Signed-off-by: Benny Halevy <bhalevy@panasas.com>
[sunrpc: initialize rq_usedeferral]
Signed-off-by: Andy Adamson <andros@netapp.com>
Signed-off-by: Benny Halevy <bhalevy@panasas.com>
Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/jikos/trivial: (28 commits)
trivial: Update my email address
trivial: NULL noise: drivers/mtd/tests/mtd_*test.c
trivial: NULL noise: drivers/media/dvb/frontends/drx397xD_fw.h
trivial: Fix misspelling of "Celsius".
trivial: remove unused variable 'path' in alloc_file()
trivial: fix a pdlfush -> pdflush typo in comment
trivial: jbd header comment typo fix for JBD_PARANOID_IOFAIL
trivial: wusb: Storage class should be before const qualifier
trivial: drivers/char/bsr.c: Storage class should be before const qualifier
trivial: h8300: Storage class should be before const qualifier
trivial: fix where cgroup documentation is not correctly referred to
trivial: Give the right path in Documentation example
trivial: MTD: remove EOL from MODULE_DESCRIPTION
trivial: Fix typo in bio_split()'s documentation
trivial: PWM: fix of #endif comment
trivial: fix typos/grammar errors in Kconfig texts
trivial: Fix misspelling of firmware
trivial: cgroups: documentation typo and spelling corrections
trivial: Update contact info for Jochen Hein
trivial: fix typo "resgister" -> "register"
...
Also ensure that we use the protocol family instead of the address
family when calling sock_create_kern().
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Setting ->owner as done currently (pde->owner = THIS_MODULE) is racy
as correctly noted at bug #12454. Someone can lookup entry with NULL
->owner, thus not pinning enything, and release it later resulting
in module refcount underflow.
We can keep ->owner and supply it at registration time like ->proc_fops
and ->data.
But this leaves ->owner as easy-manipulative field (just one C assignment)
and somebody will forget to unpin previous/pin current module when
switching ->owner. ->proc_fops is declared as "const" which should give
some thoughts.
->read_proc/->write_proc were just fixed to not require ->owner for
protection.
rmmod'ed directories will be empty and return "." and ".." -- no harm.
And directories with tricky enough readdir and lookup shouldn't be modular.
We definitely don't want such modular code.
Removing ->owner will also make PDE smaller.
So, let's nuke it.
Kudos to Jeff Layton for reminding about this, let's say, oversight.
http://bugzilla.kernel.org/show_bug.cgi?id=12454
Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Impact: cleanup
Time to clean up remaining laggards using the old cpu_ functions.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Cc: Greg Kroah-Hartman <gregkh@suse.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Trond.Myklebust@netapp.com
We just augmented the kernel's RPC service registration code so that
it automatically adjusts to what is supported in user space. Thus we
no longer need the kernel configuration option to enable registering
RPC services with v4 -- it's all done automatically.
This patch is part of a series that addresses
http://bugzilla.kernel.org/show_bug.cgi?id=12256
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Move error reporting for RPC registration to rpcb_register's caller.
This way the caller can choose to recover silently from certain
errors, but report errors it does not recognize. Error reporting
for kernel RPC service registration is now handled in one place.
This patch is part of a series that addresses
http://bugzilla.kernel.org/show_bug.cgi?id=12256
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
The kernel registers RPC services with the local portmapper with an
rpcbind SET upcall to the local portmapper. Traditionally, this used
rpcbind v2 (PMAP), but registering RPC services that support IPv6
requires rpcbind v3 or v4.
Since we now want separate PF_INET and PF_INET6 listeners for each
kernel RPC service, svc_register() will do only one of those
registrations at a time.
For PF_INET, it tries an rpcb v4 SET upcall first; if that fails, it
does a legacy portmap SET. This makes it entirely backwards
compatible with legacy user space, but allows a proper v4 SET to be
used if rpcbind is available.
For PF_INET6, it does an rpcb v4 SET upcall. If that fails, it fails
the registration, and thus the transport creation. This let's the
kernel detect if user space is able to support IPv6 RPC services, and
thus whether it should maintain a PF_INET6 listener for each service
at all.
This provides complete backwards compatibilty with legacy user space
that only supports rpcbind v2. The only down-side is that registering
a new kernel RPC service may take an extra exchange with the local
portmapper on legacy systems, but this is an infrequent operation and
is done over UDP (no lingering sockets in TIMEWAIT), so it shouldn't
be consequential.
This patch is part of a series that addresses
http://bugzilla.kernel.org/show_bug.cgi?id=12256
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Our initial implementation of svc_unregister() assumed that PMAP_UNSET
cleared all rpcbind registrations for a [program, version] tuple.
However, we now have evidence that PMAP_UNSET clears only "inet"
entries, and not "inet6" entries, in the rpcbind database.
For backwards compatibility with the legacy portmapper, the
svc_unregister() function also must work if user space doesn't support
rpcbind version 4 at all.
Thus we'll send an rpcbind v4 UNSET, and if that fails, we'll send a
PMAP_UNSET.
This simplifies the code in svc_unregister() and provides better
backwards compatibility with legacy user space that does not support
rpcbind version 4. We can get rid of the conditional compilation in
here as well.
This patch is part of a series that addresses
http://bugzilla.kernel.org/show_bug.cgi?id=12256
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
The user space TI-RPC library uses an empty string for the universal
address when unregistering all target addresses for [program, version].
The kernel's rpcb client should behave the same way.
Here, we are switching between several registration methods based on
the protocol family of the incoming address. Rename the other rpcbind
v4 registration functions to make it clear that they, as well, are
switched on protocol family. In /etc/netconfig, this is either "inet"
or "inet6".
NB: The loopback protocol families are not supported in the kernel.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
RFC 1833 has little to say about the contents of r_owner; it only
specifies that it is a string, and states that it is used to control
who can UNSET an entry.
Our port of rpcbind (from Sun) assumes this string contains a numeric
UID value, not alphabetical or symbolic characters, but checks this
value only for AF_LOCAL RPCB_SET or RPCB_UNSET requests. In all other
cases, rpcbind ignores the contents of the r_owner string.
The reference user space implementation of rpcb_set(3) uses a numeric
UID for all SET/UNSET requests (even via the network) and an empty
string for all other requests. We emulate that behavior here to
maintain bug-for-bug compatibility.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Clean up: Simplify rpcb_v4_register() and its helpers by moving the
details of sockaddr type casting to rpcb_v4_register()'s helper
functions.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
The RPC client returns -EPROTONOSUPPORT if there is a protocol version
mismatch (ie the remote RPC server doesn't support the RPC protocol
version sent by the client).
Helpers for the svc_register() function return -EPROTONOSUPPORT if they
don't recognize the passed-in IPPROTO_ value.
These are two entirely different failure modes.
Have the helpers return -ENOPROTOOPT instead of -EPROTONOSUPPORT. This
will allow callers to determine more precisely what the underlying
problem is, and decide to report or recover appropriately.
This patch is part of a series that addresses
http://bugzilla.kernel.org/show_bug.cgi?id=12256
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
The kernel uses an IPv6 loopback address when registering its AF_INET6
RPC services so that it can tell whether the local portmapper is
actually IPv6-enabled.
Since the legacy portmapper doesn't listen on IPv6, however, this
causes a long timeout on older systems if the kernel happens to try
creating and registering an AF_INET6 RPC service. Originally I wanted
to use a connected transport (either TCP or connected UDP) so that the
upcall would fail immediately if the portmapper wasn't listening on
IPv6, but we never agreed on what transport to use.
In the end, it's of little consequence to the kernel whether the local
portmapper is listening on IPv6. It's only important whether the
portmapper supports rpcbind v4. And the kernel can't tell that at all
if it is sending requests via IPv6 -- the portmapper will just ignore
them.
So, send both rpcbind v2 and v4 SET/UNSET requests via IPv4 loopback
to maintain better backwards compatibility between new kernels and
legacy user space, and prevent multi-second hangs in some cases when
the kernel attempts to register RPC services.
This patch is part of a series that addresses
http://bugzilla.kernel.org/show_bug.cgi?id=12256
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
We are about to convert to using separate RPC listener sockets for
PF_INET and PF_INET6. This echoes the way IPv6 is handled in user
space by TI-RPC, and eliminates the need for ULPs to worry about
mapped IPv4 AF_INET6 addresses when doing address comparisons.
Start by setting the IPV6ONLY flag on PF_INET6 RPC listener sockets.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Since an RPC service listener's protocol family is specified now via
svc_create_xprt(), it no longer needs to be passed to svc_create() or
svc_create_pooled(). Remove that argument from the synopsis of those
functions, and remove the sv_family field from the svc_serv struct.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
The sv_family field is going away. Pass a protocol family argument to
svc_create_xprt() instead of extracting the family from the passed-in
svc_serv struct.
Again, as this is a listener socket and not an address, we make this
new argument an "int" protocol family, instead of an "sa_family_t."
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Since the sv_family field is going away, modify svc_setup_socket() to
extract the protocol family from the passed-in socket instead of from
the passed-in svc_serv struct.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
The sv_family field is going away. Instead of using sv_family, have
the svc_register() function take a protocol family argument.
Since this argument represents a protocol family, and not an address
family, this argument takes an int, as this is what is passed to
sock_create_kern(). Also make sure svc_register's helpers are
checking for PF_FOO instead of AF_FOO. The value of [AP]F_FOO are
equivalent; this is simply a symbolic change to reflect the semantics
of the value stored in that variable.
sock_create_kern() should return EPFNOSUPPORT if the passed-in
protocol family isn't supported, but it uses EAFNOSUPPORT for this
case. We will stick with that tradition here, as svc_register()
is called by the RPC server in the same path as sock_create_kern().
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Clean up: add documentating comment and use appropriate data types for
svc_find_xprt()'s arguments.
This also eliminates a mixed sign comparison: @port was an int, while
the return value of svc_xprt_local_port() is an unsigned short.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
In 2007, commit e65fe3976f added
additional sanity checking to rpcb_decode_getaddr() to make sure we
were getting a reply that was long enough to be an actual universal
address. If the uaddr string isn't long enough, the XDR decoder
returns EIO.
However, an empty string is a valid RPCB_GETADDR response if the
requested service isn't registered. Moreover, "::.n.m" is also a
valid RPCB_GETADDR response for IPv6 addresses that is shorter
than rpcb_decode_getaddr()'s lower limit of 11. So this sanity
check introduced a regression for rpcbind requests against IPv6
remotes.
So revert the lower bound check added by commit
e65fe3976f, and add an explicit check
for an empty uaddr string, similar to libtirpc's rpcb_getaddr(3).
Pointed-out-by: Jeff Layton <jlayton@redhat.com>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
There is no need to set rqstp->rq_server to serv, while serv is initialized as rqstp->rq_server at previous line. And between these two lines, there is no change to rqstp->rq_server.
Signed-off-by: ideawu <ideawu@163.com>
Reviewed-by: Tom Tucker <tom@opengridcomputing.com>
Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
printk formats in prior commit were reversed/incorrect.
Compiled without warning on x86 and x86_64, but detected on ppc.
Signed-off-by: Tom Talpey <tmtalpey@gmail.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
As long as one task is holding the socket lock, then calls to
xprt_force_disconnect(xprt) will not succeed in shutting down the socket.
In particular, this would mean that a server initiated shutdown will not
succeed until the lock is relinquished.
In order to avoid the deadlock, we should ensure that xs_tcp_send_request()
closes the socket on EPIPE errors too.
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
This fixes a regression against FreeBSD servers as reported by Tomas
Kasparek. Apparently when using RPC over a TCP socket, the FreeBSD servers
don't ever react to the client closing the socket, and so commit
e06799f958 (SUNRPC: Use shutdown() instead of
close() when disconnecting a TCP socket) causes the setup to hang forever
whenever the client attempts to close and then reconnect.
We break the deadlock by adding a 'linger2' style timeout to the socket,
after which, the client will abort the connection using a TCP 'RST'.
The default timeout is set to 15 seconds. A subsequent patch will put it
under user control by means of a systctl.
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Allow the NFSv4 server to make use of TCP autotuning behaviour, which
was previously disabled by setting the sk_userlocks variable.
Set the receive buffers to be big enough to receive the whole RPC
request, and set this for the listening socket, not the accept socket.
Remove the code that readjusts the receive/send buffer sizes for the
accepted socket. Previously this code was used to influence the TCP
window management behaviour, which is no longer needed when autotuning
is enabled.
This can improve IO bandwidth on networks with high bandwidth-delay
products, where a large tcp window is required. It also simplifies
performance tuning, since getting adequate tcp buffers previously
required increasing the number of nfsd threads.
Signed-off-by: Olga Kornievskaia <aglo@citi.umich.edu>
Cc: Jim Rees <rees@umich.edu>
Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
Add /proc/fs/nfsd/pool_stats to export to userspace various
statistics about the operation of rpc server thread pools.
This patch is based on a forward-ported version of
knfsd-add-pool-thread-stats which has been shipping in the SGI
"Enhanced NFS" product since 2006 and which was previously
posted:
http://article.gmane.org/gmane.linux.nfs/10375
It has also been updated thus:
* moved EXPORT_SYMBOL() to near the function it exports
* made the new struct struct seq_operations const
* used SEQ_START_TOKEN instead of ((void *)1)
* merged fix from SGI PV 990526 "sunrpc: use dprintk instead of
printk in svc_pool_stats_*()" by Harshula Jayasuriya.
* merged fix from SGI PV 964001 "Crash reading pool_stats before
nfsds are started".
Signed-off-by: Greg Banks <gnb@sgi.com>
Signed-off-by: Harshula Jayasuriya <harshula@sgi.com>
Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
Avoid overloading the CPU scheduler with enormous load averages
when handling high call-rate NFS loads. When the knfsd bottom half
is made aware of an incoming call by the socket layer, it tries to
choose an nfsd thread and wake it up. As long as there are idle
threads, one will be woken up.
If there are lot of nfsd threads (a sensible configuration when
the server is disk-bound or is running an HSM), there will be many
more nfsd threads than CPUs to run them. Under a high call-rate
low service-time workload, the result is that almost every nfsd is
runnable, but only a handful are actually able to run. This situation
causes two significant problems:
1. The CPU scheduler takes over 10% of each CPU, which is robbing
the nfsd threads of valuable CPU time.
2. At a high enough load, the nfsd threads starve userspace threads
of CPU time, to the point where daemons like portmap and rpc.mountd
do not schedule for tens of seconds at a time. Clients attempting
to mount an NFS filesystem timeout at the very first step (opening
a TCP connection to portmap) because portmap cannot wake up from
select() and call accept() in time.
Disclaimer: these effects were observed on a SLES9 kernel, modern
kernels' schedulers may behave more gracefully.
The solution is simple: keep in each svc_pool a counter of the number
of threads which have been woken but have not yet run, and do not wake
any more if that count reaches an arbitrary small threshold.
Testing was on a 4 CPU 4 NIC Altix using 4 IRIX clients, each with 16
synthetic client threads simulating an rsync (i.e. recursive directory
listing) workload reading from an i386 RH9 install image (161480
regular files in 10841 directories) on the server. That tree is small
enough to fill in the server's RAM so no disk traffic was involved.
This setup gives a sustained call rate in excess of 60000 calls/sec
before being CPU-bound on the server. The server was running 128 nfsds.
Profiling showed schedule() taking 6.7% of every CPU, and __wake_up()
taking 5.2%. This patch drops those contributions to 3.0% and 2.2%.
Load average was over 120 before the patch, and 20.9 after.
This patch is a forward-ported version of knfsd-avoid-nfsd-overload
which has been shipping in the SGI "Enhanced NFS" product since 2006.
It has been posted before:
http://article.gmane.org/gmane.linux.nfs/10374
Signed-off-by: Greg Banks <gnb@sgi.com>
Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
