This patch is suitable for just about any 2.6 kernel. It should go in
2.6.19 and 2.6.18.2 and possible even the .17 and .16 stable series.
This is a long standing bug that seems to have only recently become
apparent, presumably due to increasing use of NFS over TCP - many
distros seem to be making it the default.
The SK_CONN bit gets set when a listening socket may be ready
for an accept, just as SK_DATA is set when data may be available.
It is entirely possible for svc_tcp_accept to be called with neither
of these set. It doesn't happen often but there is a small race in
svc_sock_enqueue as SK_CONN and SK_DATA are tested outside the
spin_lock. They could be cleared immediately after the test and
before the lock is gained.
This normally shouldn't be a problem. The sockets are non-blocking so
trying to read() or accept() when ther is nothing to do is not a problem.
However: svc_tcp_recvfrom makes the decision "Should I accept() or
should I read()" based on whether SK_CONN is set or not. This usually
works but is not safe. The decision should be based on whether it is
a TCP_LISTEN socket or a TCP_CONNECTED socket.
Signed-off-by: Neil Brown <neilb@suse.de>
Cc: Adrian Bunk <bunk@stusta.de>
Cc: <stable@kernel.org>
Cc: Trond Myklebust <trond.myklebust@fys.uio.no>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
There is some confusion about the meaning of 'bufsz' for a sunrpc server.
In some cases it is the largest message that can be sent or received. In
other cases it is the largest 'payload' that can be included in a NFS
message.
In either case, it is not possible for both the request and the reply to be
this large. One of the request or reply may only be one page long, which
fits nicely with NFS.
So we remove 'bufsz' and replace it with two numbers: 'max_payload' and
'max_mesg'. Max_payload is the size that the server requests. It is used
by the server to check the max size allowed on a particular connection:
depending on the protocol a lower limit might be used.
max_mesg is the largest single message that can be sent or received. It is
calculated as the max_payload, rounded up to a multiple of PAGE_SIZE, and
with PAGE_SIZE added to overhead. Only one of the request and reply may be
this size. The other must be at most one page.
Cc: Greg Banks <gnb@sgi.com>
Cc: "J. Bruce Fields" <bfields@fieldses.org>
Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Speed up high call-rate workloads by caching the struct ip_map for the peer on
the connected struct svc_sock instead of looking it up in the ip_map cache
hashtable on every call. This helps workloads using AUTH_SYS authentication
over TCP.
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.
Profiling showed strcmp(), called from ip_map_match(), was taking 4.8% of each
CPU, and ip_map_lookup() was taking 2.9%. This patch drops both contribution
into the profile noise.
Note that the above result overstates this value of this patch for most
workloads. The synthetic clients are all using separate IP addresses, so
there are 64 entries in the ip_map cache hash. Because the kernel measured
contained the bug fixed in commit
commit 1f1e030bf7
and was running on 64bit little-endian machine, probably all of those 64
entries were on a single chain, thus increasing the cost of ip_map_lookup().
With a modern kernel you would need more clients to see the same amount of
performance improvement. This patch has helped to scale knfsd to handle a
deployment with 2000 NFS clients.
Signed-off-by: Greg Banks <gnb@melbourne.sgi.com>
Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
.. by allocating the array of 'kvec' in 'struct svc_rqst'.
As we plan to increase RPCSVC_MAXPAGES from 8 upto 256, we can no longer
allocate an array of this size on the stack. So we allocate it in 'struct
svc_rqst'.
However svc_rqst contains (indirectly) an array of the same type and size
(actually several, but they are in a union). So rather than waste space, we
move those arrays out of the separately allocated union and into svc_rqst to
share with the kvec moved out of svc_tcp_recvfrom (various arrays are used at
different times, so there is no conflict).
Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
We are planning to increase RPCSVC_MAXPAGES from about 8 to about 256. This
means we need to be a bit careful about arrays of size RPCSVC_MAXPAGES.
struct svc_rqst contains two such arrays. However the there are never more
that RPCSVC_MAXPAGES pages in the two arrays together, so only one array is
needed.
The two arrays are for the pages holding the request, and the pages holding
the reply. Instead of two arrays, we can simply keep an index into where the
first reply page is.
This patch also removes a number of small inline functions that probably
server to obscure what is going on rather than clarify it, and opencode the
needed functionality.
Also remove the 'rq_restailpage' variable as it is *always* 0. i.e. if the
response 'xdr' structure has a non-empty tail it is always in the same pages
as the head.
check counters are initilised and incr properly
check for consistant usage of ++ etc
maybe extra some inlines for common approach
general review
Signed-off-by: Neil Brown <neilb@suse.de>
Cc: Magnus Maatta <novell@kiruna.se>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Arrgg.. We cannot 'lockd_up' before 'svc_addsock' as we don't know the
protocol yet.... So switch it around again and save the name of the created
sockets so that it can be closed if lock_up fails.
Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
The refcount that nfsd holds on lockd is based on the number of open sockets.
So when we close a socket, we should decrement the ref (with lockd_down).
Currently when a socket is closed via writing to the portlist file, that
doesn't happen.
So: make sure we get an error return if the socket that was requested does is
not found, and call lockd_down if it was.
Cc: "J. Bruce Fields" <bfields@fieldses.org>
Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Actually implement multiple pools. On NUMA machines, allocate a svc_pool per
NUMA node; on SMP a svc_pool per CPU; otherwise a single global pool. Enqueue
sockets on the svc_pool corresponding to the CPU on which the socket bh is run
(i.e. the NIC interrupt CPU). Threads have their cpu mask set to limit them
to the CPUs in the svc_pool that owns them.
This is the patch that allows an Altix to scale NFS traffic linearly
beyond 4 CPUs and 4 NICs.
Incorporates changes and feedback from Neil Brown, Trond Myklebust, and
Christoph Hellwig.
Signed-off-by: Greg Banks <gnb@melbourne.sgi.com>
Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Split out the list of idle threads and pending sockets from svc_serv into a
new svc_pool structure, and allocate a fixed number (in this patch, 1) of
pools per svc_serv. The new structure contains a lock which takes over
several of the duties of svc_serv->sv_lock, which is now relegated to
protecting only sv_tempsocks, sv_permsocks, and sv_tmpcnt in svc_serv.
The point is to move the hottest fields out of svc_serv and into svc_pool,
allowing a following patch to arrange for a svc_pool per NUMA node or per CPU.
This is a major step towards making the NFS server NUMA-friendly.
Signed-off-by: Greg Banks <gnb@melbourne.sgi.com>
Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
The SK_BUSY bit in svc_sock->sk_flags ensures that we do not attempt to
enqueue a socket twice. Currently, setting and clearing the bit is protected
by svc_serv->sv_lock. As I intend to reduce the data that the lock protects
so it's not held when svc_sock_enqueue() tests and sets SK_BUSY, that test and
set needs to be atomic.
Signed-off-by: Greg Banks <gnb@melbourne.sgi.com>
Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Convert the svc_sock->sk_reserved variable from an int protected by
svc_serv->sv_lock, to an atomic. This reduces (by 1) the number of places we
need to take the (effectively global) svc_serv->sv_lock.
Signed-off-by: Greg Banks <gnb@melbourne.sgi.com>
Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Protect the svc_sock->sk_deferred list with a new lock svc_sock->sk_defer_lock
instead of svc_serv->sv_lock. Using the more fine-grained lock reduces the
number of places we need to take the svc_serv lock.
Signed-off-by: Greg Banks <gnb@melbourne.sgi.com>
Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Convert the svc_sock->sk_inuse counter from an int protected by
svc_serv->sv_lock, to an atomic. This reduces the number of places we need to
take the (effectively global) svc_serv->sv_lock.
Signed-off-by: Greg Banks <gnb@melbourne.sgi.com>
Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Following are 11 patches from Greg Banks which combine to make knfsd more
Numa-aware. They reduce hitting on 'global' data structures, and create some
data-structures that can be node-local.
knfsd threads are bound to a particular node, and the thread to handle a new
request is chosen from the threads that are attach to the node that received
the interrupt.
The distribution of threads across nodes can be controlled by a new file in
the 'nfsd' filesystem, though the default approach of an even spread is
probably fine for most sites.
Some (old) numbers that show the efficacy of these patches: N == number of
NICs == number of CPUs == nmber of clients. Number of NUMA nodes == N/2
N Throughput, MiB/s CPU usage, % (max=N*100)
Before After Before After
--- ------ ---- ----- -----
4 312 435 350 228
6 500 656 501 418
8 562 804 690 589
This patch:
Move the aging of RPC/TCP connection sockets from the main svc_recv() loop to
a timer which uses a mark-and-sweep algorithm every 6 minutes. This reduces
the amount of work that needs to be done in the main RPC loop and the length
of time we need to hold the (effectively global) svc_serv->sv_lock.
[akpm@osdl.org: cleanup]
Signed-off-by: Greg Banks <gnb@melbourne.sgi.com>
Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
It isn't needed as it is available in rqstp->rq_server, and dropping it allows
some local vars to be dropped.
[akpm@osdl.org: build fix]
Cc: "J. Bruce Fields" <bfields@fieldses.org>
Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Userspace should create and bind a socket (but not connectted) and write the
'fd' to portlist. This will cause the nfs server to listen on that socket.
To close a socket, the name of the socket - as read from 'portlist' can be
written to 'portlist' with a preceding '-'.
Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This file will list all ports that nfsd has open.
Default when TCP enabled will be
ipv4 udp 0.0.0.0 2049
ipv4 tcp 0.0.0.0 2049
Later, the list of ports will be settable.
'portlist' chosen rather than 'ports', to avoid unnecessary confusion with
non-mainline patches which created 'ports' with different semantics.
[akpm@osdl.org: cleanups, build fix]
Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
coverity spotted this one as possible dereference in the dprintk(),
but since there is only one caller of svc_create_socket(), which always
passes a valid sin, we dont need this check.
Signed-off-by: Eric Sesterhenn <snakebyte@gmx.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
pure s/u32/__be32/
[AV: large part based on Alexey's patches]
Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>
Semaphore to mutex conversion.
The conversion was generated via scripts, and the result was validated
automatically via a script as well.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
The server code currently keeps track of the destination address on every
request so that it can reply using the same address. However we forget to do
that in the case of a deferred request. Remedy this oversight. >From folks
at PolyServe.
Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
I submitted this one previously - svc_tcp_recvfrom currently returns
any errors to the caller, including ECONNRESET and the like.
This is something svc_recv isn't able to deal with:
len = svsk->sk_recvfrom(rqstp);
[...]
if (len == 0 || len == -EAGAIN) {
[...]
return -EAGAIN;
}
[...]
return len;
The nfsd main loop will exit when it sees an error code other than
EAGAIN.
The following patch fixes this problem
svc_recv is not equipped to deal with error codes other than EAGAIN,
and will propagate anything else (such as ECONNRESET) up to nfsd,
causing it to exit.
Signed-off-by: Olaf Kirch <okir@suse.de>
Cc: Trond Myklebust <trond.myklebust@fys.uio.no>
Cc: Neil Brown <neilb@cse.unsw.edu.au>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
I noticed that some of 'struct proto_ops' used in the kernel may share
a cache line used by locks or other heavily modified data. (default
linker alignement is 32 bytes, and L1_CACHE_LINE is 64 or 128 at
least)
This patch makes sure a 'struct proto_ops' can be declared as const,
so that all cpus can share all parts of it without false sharing.
This is not mandatory : a driver can still use a read/write structure
if it needs to (and eventually a __read_mostly)
I made a global stubstitute to change all existing occurences to make
them const.
This should reduce the possibility of false sharing on SMP, and
speedup some socket system calls.
Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Being kernel-threads, nfsd servers don't get pre-empted (depending on
CONFIG). If there is a steady stream of NFS requests that can be served
from cache, an nfsd thread may hold on to a cpu indefinitely, which isn't
very friendly.
So it is good to have a cond_resched in there (just before looking for a
new request to serve), to make sure we play nice.
Signed-off-by: Neil Brown <neilb@suse.de>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Here is the patch that introduces the generic skb_checksum_complete
which also checks for hardware RX checksum faults. If that happens,
it'll call netdev_rx_csum_fault which currently prints out a stack
trace with the device name. In future it can turn off RX checksum.
I've converted every spot under net/ that does RX checksum checks to
use skb_checksum_complete or __skb_checksum_complete with the
exceptions of:
* Those places where checksums are done bit by bit. These will call
netdev_rx_csum_fault directly.
* The following have not been completely checked/converted:
ipmr
ip_vs
netfilter
dccp
This patch is based on patches and suggestions from Stephen Hemminger
and David S. Miller.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
Convert nanoseconds to microseconds correctly.
Spotted by Steve Dickson <SteveD@redhat.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Clean-up: Move some code that is common to both RPC client- and server-side
socket transports into its own source file, net/sunrpc/socklib.c.
Test-plan:
Compile kernel with CONFIG_NFS enabled. Millions of fsx operations over
UDP, client and server. Connectathon over UDP.
Version: Thu, 11 Aug 2005 16:03:09 -0400
Signed-off-by: Chuck Lever <cel@netapp.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Change a printk(KERN_WARNING to dprintk, and it is really only interesting
when trying to debug a problem, and can occur normally without error.
Remove various gratuitous gotos in surrounding code, and remove some
type-cast assignments from inside 'if' conditionals, as that is just
obscuring what it going on.
Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Use schedule_timeout_{,un}interruptible() instead of
set_current_state()/schedule_timeout() to reduce kernel size. Also use
human-time conversion functions instead of hard-coded division to avoid
rounding issues.
Signed-off-by: Nishanth Aravamudan <nacc@us.ibm.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Lots of places just needs the states, not even linux/tcp.h, where this
enum was, needs it.
This speeds up development of the refactorings as less sources are
rebuilt when things get moved from net/tcp.h.
Signed-off-by: Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
We need to divide, not multiply. While we're here,
use NSEC_PER_USEC instead of a magic constant.
Based upon a report from Josip Loncaric and a patch
by Andrew Morton.
Signed-off-by: David S. Miller <davem@davemloft.net>
1. Establish a simple API for process freezing defined in linux/include/sched.h:
frozen(process) Check for frozen process
freezing(process) Check if a process is being frozen
freeze(process) Tell a process to freeze (go to refrigerator)
thaw_process(process) Restart process
frozen_process(process) Process is frozen now
2. Remove all references to PF_FREEZE and PF_FROZEN from all
kernel sources except sched.h
3. Fix numerous locations where try_to_freeze is manually done by a driver
4. Remove the argument that is no longer necessary from two function calls.
5. Some whitespace cleanup
6. Clear potential race in refrigerator (provides an open window of PF_FREEZE
cleared before setting PF_FROZEN, recalc_sigpending does not check
PF_FROZEN).
This patch does not address the problem of freeze_processes() violating the rule
that a task may only modify its own flags by setting PF_FREEZE. This is not clean
in an SMP environment. freeze(process) is therefore not SMP safe!
Signed-off-by: Christoph Lameter <christoph@lameter.com>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Initial git repository build. I'm not bothering with the full history,
even though we have it. We can create a separate "historical" git
archive of that later if we want to, and in the meantime it's about
3.2GB when imported into git - space that would just make the early
git days unnecessarily complicated, when we don't have a lot of good
infrastructure for it.
Let it rip!
