Only cpu seeing dst refcount going to 0 can safely
dereference dst->flags.
Otherwise an other cpu might already have freed the dst.
Fixes: 27b75c95f1 ("net: avoid RCU for NOCACHE dst")
Reported-by: Greg Thelen <gthelen@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The network namespace is already passed into dst_output pass it into
dst->output lwt->output and friends.
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
opts_size is only written and never read. Following patch
removes this unused variable.
Signed-off-by: Pravin B Shelar <pshelar@nicira.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently, the lwtunnel state resides in per-protocol data. This is
a problem if we encapsulate ipv6 traffic in an ipv4 tunnel (or vice versa).
The xmit function of the tunnel does not know whether the packet has been
routed to it by ipv4 or ipv6, yet it needs the lwtstate data. Moving the
lwtstate data to dst_entry makes such inter-protocol tunneling possible.
As a bonus, this brings a nice diffstat.
Signed-off-by: Jiri Benc <jbenc@redhat.com>
Acked-by: Roopa Prabhu <roopa@cumulusnetworks.com>
Acked-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>
Introduce helpers to let eBPF programs attached to TC manipulate tunnel metadata:
bpf_skb_[gs]et_tunnel_key(skb, key, size, flags)
skb: pointer to skb
key: pointer to 'struct bpf_tunnel_key'
size: size of 'struct bpf_tunnel_key'
flags: room for future extensions
First eBPF program that uses these helpers will allocate per_cpu
metadata_dst structures that will be used on TX.
On RX metadata_dst is allocated by tunnel driver.
Typical usage for TX:
struct bpf_tunnel_key tkey;
... populate tkey ...
bpf_skb_set_tunnel_key(skb, &tkey, sizeof(tkey), 0);
bpf_clone_redirect(skb, vxlan_dev_ifindex, 0);
RX:
struct bpf_tunnel_key tkey = {};
bpf_skb_get_tunnel_key(skb, &tkey, sizeof(tkey), 0);
... lookup or redirect based on tkey ...
'struct bpf_tunnel_key' will be extended in the future by adding
elements to the end and the 'size' argument will indicate which fields
are populated, thereby keeping backwards compatibility.
The 'flags' argument may be used as well when the 'size' is not enough or
to indicate completely different layout of bpf_tunnel_key.
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Acked-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>
Conflicts:
net/bridge/br_mdb.c
br_mdb.c conflict was a function call being removed to fix a bug in
'net' but whose signature was changed in 'net-next'.
Signed-off-by: David S. Miller <davem@davemloft.net>
Introduces a new dst_metadata which enables to carry per packet metadata
between forwarding and processing elements via the skb->dst pointer.
The structure is set up to be a union. Thus, each separate type of
metadata requires its own dst instance. If demand arises to carry
multiple types of metadata concurrently, metadata dst entries can be
made stackable.
The metadata dst entry is refcnt'ed as expected for now but a non
reference counted use is possible if the reference is forced before
queueing the skb.
In order to allow allocating dsts with variable length, the existing
dst_alloc() is split into a dst_alloc() and dst_init() function. The
existing dst_init() function to initialize the subsystem is being
renamed to dst_subsys_init() to make it clear what is what.
The check before ip_route_input() is changed to ignore metadata dsts
and drop the dst inside the routing function thus allowing to interpret
metadata in a later commit.
Signed-off-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>
Kernel generates a lot of warnings when dst entry reference counter
overflows and becomes negative. That bug was seen several times at
machines with outdated 3.10.y kernels. Most like it's already fixed
in upstream. Anyway that flood completely kills machine and makes
further debugging impossible.
Signed-off-by: Konstantin Khlebnikov <khlebnikov@yandex-team.ru>
Signed-off-by: David S. Miller <davem@davemloft.net>
Since commit f886497212 ("ipv4: fix dst race in sk_dst_get()")
DST_NOCACHE dst_entries get freed by RCU. So there is no need to get a
reference on them when we are in rcu protected sections.
Cc: Eric Dumazet <edumazet@google.com>
Cc: Julian Anastasov <ja@ssi.bg>
Signed-off-by: Hannes Frederic Sowa <hannes@stressinduktion.org>
Reviewed-by: Julian Anastasov <ja@ssi.bg>
Signed-off-by: David S. Miller <davem@davemloft.net>
When IP route cache had been removed in linux-3.6, we broke assumption
that dst entries were all freed after rcu grace period. DST_NOCACHE
dst were supposed to be freed from dst_release(). But it appears
we want to keep such dst around, either in UDP sockets or tunnels.
In sk_dst_get() we need to make sure dst refcount is not 0
before incrementing it, or else we might end up freeing a dst
twice.
DST_NOCACHE set on a dst does not mean this dst can not be attached
to a socket or a tunnel.
Then, before actual freeing, we need to observe a rcu grace period
to make sure all other cpus can catch the fact the dst is no longer
usable.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Dormando <dormando@rydia.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
In the dst->output() path for ipv4, the code assumes the skb it has to
transmit is attached to an inet socket, specifically via
ip_mc_output() : The sk_mc_loop() test triggers a WARN_ON() when the
provider of the packet is an AF_PACKET socket.
The dst->output() method gets an additional 'struct sock *sk'
parameter. This needs a cascade of changes so that this parameter can
be propagated from vxlan to final consumer.
Fixes: 8f646c922d ("vxlan: keep original skb ownership")
Reported-by: lucien xin <lucien.xin@gmail.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
So far, only net_device * could be passed along with netdevice notifier
event. This patch provides a possibility to pass custom structure
able to provide info that event listener needs to know.
Signed-off-by: Jiri Pirko <jiri@resnulli.us>
v2->v3: fix typo on simeth
shortened dev_getter
shortened notifier_info struct name
v1->v2: fix notifier_call parameter in call_netdevice_notifier()
Signed-off-by: David S. Miller <davem@davemloft.net>
Rename skb_dst_set_noref to __skb_dst_set_noref and
add force flag as suggested by David Miller. The new wrapper
skb_dst_set_noref_force will force dst entries that are not
cached to be attached as skb dst without taking reference
as long as provided dst is reclaimed after RCU grace period.
Signed-off-by: Julian Anastasov <ja@ssi.bg>
Signed-off by: Hans Schillstrom <hans@schillstrom.com>
Acked-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Simon Horman <horms@verge.net.au>
Eric Dumazet wrote:
| Some strange crashes happen in rt6_check_expired(), with access
| to random addresses.
|
| At first glance, it looks like the RTF_EXPIRES and
| stuff added in commit 1716a96101
| (ipv6: fix problem with expired dst cache)
| are racy : same dst could be manipulated at the same time
| on different cpus.
|
| At some point, our stack believes rt->dst.from contains a dst pointer,
| while its really a jiffie value (as rt->dst.expires shares the same area
| of memory)
|
| rt6_update_expires() should be fixed, or am I missing something ?
|
| CC Neil because of https://bugzilla.redhat.com/show_bug.cgi?id=892060
Because we do not have any locks for dst_entry, we cannot change
essential structure in the entry; e.g., we cannot change reference
to other entity.
To fix this issue, split 'from' and 'expires' field in dst_entry
out of union. Once it is 'from' is assigned in the constructor,
keep the reference until the very last stage of the life time of
the object.
Of course, it is unsafe to change 'from', so make rt6_set_from simple
just for fresh entries.
Reported-by: Eric Dumazet <eric.dumazet@gmail.com>
Reported-by: Neil Horman <nhorman@tuxdriver.com>
CC: Gao Feng <gaofeng@cn.fujitsu.com>
Signed-off-by: YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reported-by: Steinar H. Gunderson <sesse@google.com>
Reviewed-by: Neil Horman <nhorman@tuxdriver.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Pull networking changes from David Miller:
1) GRE now works over ipv6, from Dmitry Kozlov.
2) Make SCTP more network namespace aware, from Eric Biederman.
3) TEAM driver now works with non-ethernet devices, from Jiri Pirko.
4) Make openvswitch network namespace aware, from Pravin B Shelar.
5) IPV6 NAT implementation, from Patrick McHardy.
6) Server side support for TCP Fast Open, from Jerry Chu and others.
7) Packet BPF filter supports MOD and XOR, from Eric Dumazet and Daniel
Borkmann.
8) Increate the loopback default MTU to 64K, from Eric Dumazet.
9) Use a per-task rather than per-socket page fragment allocator for
outgoing networking traffic. This benefits processes that have very
many mostly idle sockets, which is quite common.
From Eric Dumazet.
10) Use up to 32K for page fragment allocations, with fallbacks to
smaller sizes when higher order page allocations fail. Benefits are
a) less segments for driver to process b) less calls to page
allocator c) less waste of space.
From Eric Dumazet.
11) Allow GRO to be used on GRE tunnels, from Eric Dumazet.
12) VXLAN device driver, one way to handle VLAN issues such as the
limitation of 4096 VLAN IDs yet still have some level of isolation.
From Stephen Hemminger.
13) As usual there is a large boatload of driver changes, with the scale
perhaps tilted towards the wireless side this time around.
Fix up various fairly trivial conflicts, mostly caused by the user
namespace changes.
* git://git.kernel.org/pub/scm/linux/kernel/git/davem/net-next: (1012 commits)
hyperv: Add buffer for extended info after the RNDIS response message.
hyperv: Report actual status in receive completion packet
hyperv: Remove extra allocated space for recv_pkt_list elements
hyperv: Fix page buffer handling in rndis_filter_send_request()
hyperv: Fix the missing return value in rndis_filter_set_packet_filter()
hyperv: Fix the max_xfer_size in RNDIS initialization
vxlan: put UDP socket in correct namespace
vxlan: Depend on CONFIG_INET
sfc: Fix the reported priorities of different filter types
sfc: Remove EFX_FILTER_FLAG_RX_OVERRIDE_IP
sfc: Fix loopback self-test with separate_tx_channels=1
sfc: Fix MCDI structure field lookup
sfc: Add parentheses around use of bitfield macro arguments
sfc: Fix null function pointer in efx_sriov_channel_type
vxlan: virtual extensible lan
igmp: export symbol ip_mc_leave_group
netlink: add attributes to fdb interface
tg3: unconditionally select HWMON support when tg3 is enabled.
Revert "net: ti cpsw ethernet: allow reading phy interface mode from DT"
gre: fix sparse warning
...
Pull workqueue changes from Tejun Heo:
"This is workqueue updates for v3.7-rc1. A lot of activities this
round including considerable API and behavior cleanups.
* delayed_work combines a timer and a work item. The handling of the
timer part has always been a bit clunky leading to confusing
cancelation API with weird corner-case behaviors. delayed_work is
updated to use new IRQ safe timer and cancelation now works as
expected.
* Another deficiency of delayed_work was lack of the counterpart of
mod_timer() which led to cancel+queue combinations or open-coded
timer+work usages. mod_delayed_work[_on]() are added.
These two delayed_work changes make delayed_work provide interface
and behave like timer which is executed with process context.
* A work item could be executed concurrently on multiple CPUs, which
is rather unintuitive and made flush_work() behavior confusing and
half-broken under certain circumstances. This problem doesn't
exist for non-reentrant workqueues. While non-reentrancy check
isn't free, the overhead is incurred only when a work item bounces
across different CPUs and even in simulated pathological scenario
the overhead isn't too high.
All workqueues are made non-reentrant. This removes the
distinction between flush_[delayed_]work() and
flush_[delayed_]_work_sync(). The former is now as strong as the
latter and the specified work item is guaranteed to have finished
execution of any previous queueing on return.
* In addition to the various bug fixes, Lai redid and simplified CPU
hotplug handling significantly.
* Joonsoo introduced system_highpri_wq and used it during CPU
hotplug.
There are two merge commits - one to pull in IRQ safe timer from
tip/timers/core and the other to pull in CPU hotplug fixes from
wq/for-3.6-fixes as Lai's hotplug restructuring depended on them."
Fixed a number of trivial conflicts, but the more interesting conflicts
were silent ones where the deprecated interfaces had been used by new
code in the merge window, and thus didn't cause any real data conflicts.
Tejun pointed out a few of them, I fixed a couple more.
* 'for-3.7' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/wq: (46 commits)
workqueue: remove spurious WARN_ON_ONCE(in_irq()) from try_to_grab_pending()
workqueue: use cwq_set_max_active() helper for workqueue_set_max_active()
workqueue: introduce cwq_set_max_active() helper for thaw_workqueues()
workqueue: remove @delayed from cwq_dec_nr_in_flight()
workqueue: fix possible stall on try_to_grab_pending() of a delayed work item
workqueue: use hotcpu_notifier() for workqueue_cpu_down_callback()
workqueue: use __cpuinit instead of __devinit for cpu callbacks
workqueue: rename manager_mutex to assoc_mutex
workqueue: WORKER_REBIND is no longer necessary for idle rebinding
workqueue: WORKER_REBIND is no longer necessary for busy rebinding
workqueue: reimplement idle worker rebinding
workqueue: deprecate __cancel_delayed_work()
workqueue: reimplement cancel_delayed_work() using try_to_grab_pending()
workqueue: use mod_delayed_work() instead of __cancel + queue
workqueue: use irqsafe timer for delayed_work
workqueue: clean up delayed_work initializers and add missing one
workqueue: make deferrable delayed_work initializer names consistent
workqueue: cosmetic whitespace updates for macro definitions
workqueue: deprecate system_nrt[_freezable]_wq
workqueue: deprecate flush[_delayed]_work_sync()
...
I noticed extra one second delay in device dismantle, tracked down to
a call to dst_dev_event() while some call_rcu() are still in RCU queues.
These call_rcu() were posted by rt_free(struct rtable *rt) calls.
We then wait a little (but one second) in netdev_wait_allrefs() before
kicking again NETDEV_UNREGISTER.
As the call_rcu() are now completed, dst_dev_event() can do the needed
device swap on busy dst.
To solve this problem, add a new NETDEV_UNREGISTER_FINAL, called
after a rcu_barrier(), but outside of RTNL lock.
Use NETDEV_UNREGISTER_FINAL with care !
Change dst_dev_event() handler to react to NETDEV_UNREGISTER_FINAL
Also remove NETDEV_UNREGISTER_BATCH, as its not used anymore after
IP cache removal.
With help from Gao feng
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Tom Herbert <therbert@google.com>
Cc: Mahesh Bandewar <maheshb@google.com>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Cc: Gao feng <gaofeng@cn.fujitsu.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Convert delayed_work users doing cancel_delayed_work() followed by
queue_delayed_work() to mod_delayed_work().
Most conversions are straight-forward. Ones worth mentioning are,
* drivers/edac/edac_mc.c: edac_mc_workq_setup() converted to always
use mod_delayed_work() and cancel loop in
edac_mc_reset_delay_period() is dropped.
* drivers/platform/x86/thinkpad_acpi.c: No need to remember whether
watchdog is active or not. @fan_watchdog_active and related code
dropped.
* drivers/power/charger-manager.c: Seemingly a lot of
delayed_work_pending() abuse going on here.
[delayed_]work_pending() are unsynchronized and racy when used like
this. I converted one instance in fullbatt_handler(). Please
conver the rest so that it invokes workqueue APIs for the intended
target state rather than trying to game work item pending state
transitions. e.g. if timer should be modified - call
mod_delayed_work(), canceled - call cancel_delayed_work[_sync]().
* drivers/thermal/thermal_sys.c: thermal_zone_device_set_polling()
simplified. Note that round_jiffies() calls in this function are
meaningless. round_jiffies() work on absolute jiffies not delta
delay used by delayed_work.
v2: Tomi pointed out that __cancel_delayed_work() users can't be
safely converted to mod_delayed_work(). They could be calling it
from irq context and if that happens while delayed_work_timer_fn()
is running, it could deadlock. __cancel_delayed_work() users are
dropped.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Henrique de Moraes Holschuh <hmh@hmh.eng.br>
Acked-by: Dmitry Torokhov <dmitry.torokhov@gmail.com>
Acked-by: Anton Vorontsov <cbouatmailru@gmail.com>
Acked-by: David Howells <dhowells@redhat.com>
Cc: Tomi Valkeinen <tomi.valkeinen@ti.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Jiri Kosina <jkosina@suse.cz>
Cc: Doug Thompson <dougthompson@xmission.com>
Cc: David Airlie <airlied@linux.ie>
Cc: Roland Dreier <roland@kernel.org>
Cc: "John W. Linville" <linville@tuxdriver.com>
Cc: Zhang Rui <rui.zhang@intel.com>
Cc: Len Brown <len.brown@intel.com>
Cc: "J. Bruce Fields" <bfields@fieldses.org>
Cc: Johannes Berg <johannes@sipsolutions.net>
While investigating on network performance problems, I found this little
gem :
$ nm -v vmlinux | grep -1 dst_default_metrics
ffffffff82736540 b busy.46605
ffffffff82736560 B dst_default_metrics
ffffffff82736598 b dst_busy_list
Apparently, declaring a const array without initializer put it in
(writeable) bss section, in middle of possibly often dirtied cache
lines.
Since we really want dst_default_metrics be const to avoid any possible
false sharing and catch any buggy writes, I force a null initializer.
ffffffff818a4c20 R dst_default_metrics
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Ben Hutchings <bhutchings@solarflare.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add a big comment explaining how the field works, and use defines
instead of magic constants for the values assigned to it.
Suggested by Joe Perches.
Signed-off-by: David S. Miller <davem@davemloft.net>
When a dst_confirm() happens, mark the confirmation as pending in the
dst. Then on the next packet out, when we have the neigh in-hand, do
the update.
This removes the dependency in dst_confirm() of dst's having an
attached neigh.
While we're here, remove the explicit 'dst' NULL check, all except 2
or 3 call sites ensure it's not NULL. So just fix those cases up.
Signed-off-by: David S. Miller <davem@davemloft.net>
To reflect the fact that a refrence is not obtained to the
resulting neighbour entry.
Signed-off-by: David S. Miller <davem@davemloft.net>
Acked-by: Roland Dreier <roland@purestorage.com>
Followup of commit f2c31e32b3 (fix NULL dereferences in
check_peer_redir()).
We need to make sure dst neighbour doesnt change in dst_ifdown().
Fix some sparse errors.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Now that there is a one-to-one correspondance between neighbour
and hh_cache entries, we no longer need:
1) dynamic allocation
2) attachment to dst->hh
3) refcounting
Initialization of the hh_cache entry is indicated by hh_len
being non-zero, and such initialization is always done with
the neighbour's lock held as a writer.
Signed-off-by: David S. Miller <davem@davemloft.net>
IPV6, unlike IPV4, doesn't have a routing cache.
Routing table entries, as well as clones made in response
to route lookup requests, all live in the same table. And
all of these things are together collected in the destination
cache table for ipv6.
This means that routing table entries count against the garbage
collection limits, even though such entries cannot ever be reclaimed
and are added explicitly by the administrator (rather than being
created in response to lookups).
Therefore it makes no sense to count ipv6 routing table entries
against the GC limits.
Add a DST_NOCOUNT destination cache entry flag, and skip the counting
if it is set. Use this flag bit in ipv6 when adding routing table
entries.
Signed-off-by: David S. Miller <davem@davemloft.net>
dst_default_metrics is readonly, we dont want to kfree() it later.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
* git://git.kernel.org/pub/scm/linux/kernel/git/davem/net-next-2.6: (1446 commits)
macvlan: fix panic if lowerdev in a bond
tg3: Add braces around 5906 workaround.
tg3: Fix NETIF_F_LOOPBACK error
macvlan: remove one synchronize_rcu() call
networking: NET_CLS_ROUTE4 depends on INET
irda: Fix error propagation in ircomm_lmp_connect_response()
irda: Kill set but unused variable 'bytes' in irlan_check_command_param()
irda: Kill set but unused variable 'clen' in ircomm_connect_indication()
rxrpc: Fix set but unused variable 'usage' in rxrpc_get_transport()
be2net: Kill set but unused variable 'req' in lancer_fw_download()
irda: Kill set but unused vars 'saddr' and 'daddr' in irlan_provider_connect_indication()
atl1c: atl1c_resume() is only used when CONFIG_PM_SLEEP is defined.
rxrpc: Fix set but unused variable 'usage' in rxrpc_get_peer().
rxrpc: Kill set but unused variable 'local' in rxrpc_UDP_error_handler()
rxrpc: Kill set but unused variable 'sp' in rxrpc_process_connection()
rxrpc: Kill set but unused variable 'sp' in rxrpc_rotate_tx_window()
pkt_sched: Kill set but unused variable 'protocol' in tc_classify()
isdn: capi: Use pr_debug() instead of ifdefs.
tg3: Update version to 3.119
tg3: Apply rx_discards fix to 5719/5720
...
Fix up trivial conflicts in arch/x86/Kconfig and net/mac80211/agg-tx.c
as per Davem.
Commit e66eed651f ("list: remove prefetching from regular list
iterators") removed the include of prefetch.h from list.h, which
uncovered several cases that had apparently relied on that rather
obscure header file dependency.
So this fixes things up a bit, using
grep -L linux/prefetch.h $(git grep -l '[^a-z_]prefetchw*(' -- '*.[ch]')
grep -L 'prefetchw*(' $(git grep -l 'linux/prefetch.h' -- '*.[ch]')
to guide us in finding files that either need <linux/prefetch.h>
inclusion, or have it despite not needing it.
There are more of them around (mostly network drivers), but this gets
many core ones.
Reported-by: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
It's way past it's usefulness. And this gets rid of a bunch
of stray ->rt_{dst,src} references.
Even the comment documenting the macro was inaccurate (stated
default was 1 when it's 0).
If reintroduced, it should be done properly, with dynamic debug
facilities.
Signed-off-by: David S. Miller <davem@davemloft.net>
Make dst_alloc() and it's users explicitly initialize the entire
entry.
The zero'ing done by kmem_cache_zalloc() was almost entirely
redundant.
Signed-off-by: David S. Miller <davem@davemloft.net>
This allows avoiding multiple writes to the initial __refcnt.
The most simplest cases of wanting an initial reference of "1"
in ipv4 and ipv6 have been converted, the rest have been left
along and kept at the existing "0".
Signed-off-by: David S. Miller <davem@davemloft.net>
If there are no explicit metrics attached to a route, hook
fi->fib_info up to dst_default_metrics.
Signed-off-by: David S. Miller <davem@davemloft.net>
Routing metrics are now copy-on-write.
Initially a route entry points it's metrics at a read-only location.
If a routing table entry exists, it will point there. Else it will
point at the all zero metric place-holder called 'dst_default_metrics'.
The writeability state of the metrics is stored in the low bits of the
metrics pointer, we have two bits left to spare if we want to store
more states.
For the initial implementation, COW is implemented simply via kmalloc.
However future enhancements will change this to place the writable
metrics somewhere else, in order to increase sharing. Very likely
this "somewhere else" will be the inetpeer cache.
Note also that this means that metrics updates may transiently fail
if we cannot COW the metrics successfully.
But even by itself, this patch should decrease memory usage and
increase cache locality especially for routing workloads. In those
cases the read-only metric copies stay in place and never get written
to.
TCP workloads where metrics get updated, and those rare cases where
PMTU triggers occur, will take a very slight performance hit. But
that hit will be alleviated when the long-term writable metrics
move to a more sharable location.
Since the metrics storage went from a u32 array of RTAX_MAX entries to
what is essentially a pointer, some retooling of the dst_entry layout
was necessary.
Most importantly, we need to preserve the alignment of the reference
count so that it doesn't share cache lines with the read-mostly state,
as per Eric Dumazet's alignment assertion checks.
The only non-trivial bit here is the move of the 'flags' member into
the writeable cacheline. This is OK since we are always accessing the
flags around the same moment when we made a modification to the
reference count.
Signed-off-by: David S. Miller <davem@davemloft.net>
Followup of commit ef885afbf8 (net: use rcu_barrier() in
rollback_registered_many)
dst_dev_event() scans a garbage dst list that might be feeded by various
network notifiers at device dismantle time.
Its important to call dst_dev_event() after other notifiers, or we might
enter the infamous msleep(250) in netdev_wait_allrefs(), and wait one
second before calling again call_netdevice_notifiers(NETDEV_UNREGISTER,
dev) to properly remove last device references.
Use priority -10 to let dst_dev_notifier be called after other network
notifiers (they have the default 0 priority)
Reported-by: Ben Greear <greearb@candelatech.com>
Reported-by: Nicolas Dichtel <nicolas.dichtel@6wind.com>
Reported-by: Octavian Purdila <opurdila@ixiacom.com>
Reported-by: Benjamin LaHaise <bcrl@kvack.org>
Tested-by: Ben Greear <greearb@candelatech.com>
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
There is no point using RCU for dst we allocate for a very short time
(used once).
Change dst_release() to take DST_NOCACHE into account, but also change
skb_dst_set_noref() to force a refcount increment for such dst.
This is a _huge_ gain, because we dont waste memory to store xx thousand
of dsts. Instead of queueing them to RCU, we can free them instantly.
CPU caches can stay hot, re-using same memory blocks to hold temporary
dsts.
Note : remove unneeded smp_mb__before_atomic_dec(); in dst_release(),
since atomic_dec_return() implies a full memory barrier.
Stress test, 160.000.000 udp frames sent, IP route cache disabled
(DDOS).
Before:
real 0m38.091s
user 0m13.189s
sys 7m53.018s
After:
real 0m29.946s
user 0m12.157s
sys 7m40.605s
For reference, if IP route cache was enabled :
real 0m32.030s
user 0m10.521s
sys 8m15.243s
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
struct dst_ops tracks number of allocated dst in an atomic_t field,
subject to high cache line contention in stress workload.
Switch to a percpu_counter, to reduce number of time we need to dirty a
central location. Place it on a separate cache line to avoid dirtying
read only fields.
Stress test :
(Sending 160.000.000 UDP frames,
IP route cache disabled, dual E5540 @2.53GHz,
32bit kernel, FIB_TRIE, SLUB/NUMA)
Before:
real 0m51.179s
user 0m15.329s
sys 10m15.942s
After:
real 0m45.570s
user 0m15.525s
sys 9m56.669s
With a small reordering of struct neighbour fields, subject of a
following patch, (to separate refcnt from other read mostly fields)
real 0m41.841s
user 0m15.261s
sys 8m45.949s
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When a new dst is used to send a frame, neigh_resolve_output() tries to
associate an struct hh_cache to this dst, calling neigh_hh_init() with
the neigh rwlock write locked.
Most of the time, hh_cache is already known and linked into neighbour,
so we find it and increment its refcount.
This patch changes the logic so that we call neigh_hh_init() with
neighbour lock read locked only, so that fast path can be run in
parallel by concurrent cpus.
This brings part of the speedup we got with commit c7d4426a98
(introduce DST_NOCACHE flag) for non cached dsts, even for cached ones,
removing one of the contention point that routers hit on multiqueue
enabled machines.
Further improvements would need to use a seqlock instead of an rwlock to
protect neigh->ha[], to not dirty neigh too often and remove two atomic
ops.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The function dst_ifdown is called only two places but in a non-
performance critical code path, there is no reason to inline it.
Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fix coding style errors and warnings output while running checkpatch.pl
on the file net/core/dst.c.
Signed-off-by: chavey <chavey@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
percpu.h is included by sched.h and module.h and thus ends up being
included when building most .c files. percpu.h includes slab.h which
in turn includes gfp.h making everything defined by the two files
universally available and complicating inclusion dependencies.
percpu.h -> slab.h dependency is about to be removed. Prepare for
this change by updating users of gfp and slab facilities include those
headers directly instead of assuming availability. As this conversion
needs to touch large number of source files, the following script is
used as the basis of conversion.
http://userweb.kernel.org/~tj/misc/slabh-sweep.py
The script does the followings.
* Scan files for gfp and slab usages and update includes such that
only the necessary includes are there. ie. if only gfp is used,
gfp.h, if slab is used, slab.h.
* When the script inserts a new include, it looks at the include
blocks and try to put the new include such that its order conforms
to its surrounding. It's put in the include block which contains
core kernel includes, in the same order that the rest are ordered -
alphabetical, Christmas tree, rev-Xmas-tree or at the end if there
doesn't seem to be any matching order.
* If the script can't find a place to put a new include (mostly
because the file doesn't have fitting include block), it prints out
an error message indicating which .h file needs to be added to the
file.
The conversion was done in the following steps.
1. The initial automatic conversion of all .c files updated slightly
over 4000 files, deleting around 700 includes and adding ~480 gfp.h
and ~3000 slab.h inclusions. The script emitted errors for ~400
files.
2. Each error was manually checked. Some didn't need the inclusion,
some needed manual addition while adding it to implementation .h or
embedding .c file was more appropriate for others. This step added
inclusions to around 150 files.
3. The script was run again and the output was compared to the edits
from #2 to make sure no file was left behind.
4. Several build tests were done and a couple of problems were fixed.
e.g. lib/decompress_*.c used malloc/free() wrappers around slab
APIs requiring slab.h to be added manually.
5. The script was run on all .h files but without automatically
editing them as sprinkling gfp.h and slab.h inclusions around .h
files could easily lead to inclusion dependency hell. Most gfp.h
inclusion directives were ignored as stuff from gfp.h was usually
wildly available and often used in preprocessor macros. Each
slab.h inclusion directive was examined and added manually as
necessary.
6. percpu.h was updated not to include slab.h.
7. Build test were done on the following configurations and failures
were fixed. CONFIG_GCOV_KERNEL was turned off for all tests (as my
distributed build env didn't work with gcov compiles) and a few
more options had to be turned off depending on archs to make things
build (like ipr on powerpc/64 which failed due to missing writeq).
* x86 and x86_64 UP and SMP allmodconfig and a custom test config.
* powerpc and powerpc64 SMP allmodconfig
* sparc and sparc64 SMP allmodconfig
* ia64 SMP allmodconfig
* s390 SMP allmodconfig
* alpha SMP allmodconfig
* um on x86_64 SMP allmodconfig
8. percpu.h modifications were reverted so that it could be applied as
a separate patch and serve as bisection point.
Given the fact that I had only a couple of failures from tests on step
6, I'm fairly confident about the coverage of this conversion patch.
If there is a breakage, it's likely to be something in one of the arch
headers which should be easily discoverable easily on most builds of
the specific arch.
Signed-off-by: Tejun Heo <tj@kernel.org>
Guess-its-ok-by: Christoph Lameter <cl@linux-foundation.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Kernel bugzilla #15239
On some workloads, it is quite possible to get a huge dst list to
process in dst_gc_task(), and trigger soft lockup detection.
Fix is to call cond_resched(), as we run in process context.
Reported-by: Pawel Staszewski <pstaszewski@itcare.pl>
Tested-by: Pawel Staszewski <pstaszewski@itcare.pl>
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
During tbench/oprofile sessions, I found that dst_release() was in third position.
CPU: Core 2, speed 2999.68 MHz (estimated)
Counted CPU_CLK_UNHALTED events (Clock cycles when not halted) with a unit mask of 0x00 (Unhalted core cycles) count 100000
samples % symbol name
483726 9.0185 __copy_user_zeroing_intel
191466 3.5697 __copy_user_intel
185475 3.4580 dst_release
175114 3.2648 ip_queue_xmit
153447 2.8608 tcp_sendmsg
108775 2.0280 tcp_recvmsg
102659 1.9140 sysenter_past_esp
101450 1.8914 tcp_current_mss
95067 1.7724 __copy_from_user_ll
86531 1.6133 tcp_transmit_skb
Of course, all CPUS fight on the dst_entry associated with 127.0.0.1
Instead of first checking the refcount value, then decrement it,
we use atomic_dec_return() to help CPU to make the right memory transaction
(ie getting the cache line in exclusive mode)
dst_release() is now at the fifth position, and tbench a litle bit faster ;)
CPU: Core 2, speed 3000.1 MHz (estimated)
Counted CPU_CLK_UNHALTED events (Clock cycles when not halted) with a unit mask of 0x00 (Unhalted core cycles) count 100000
samples % symbol name
647107 8.8072 __copy_user_zeroing_intel
258840 3.5229 ip_queue_xmit
258302 3.5155 __copy_user_intel
209629 2.8531 tcp_sendmsg
165632 2.2543 dst_release
149232 2.0311 tcp_current_mss
147821 2.0119 tcp_recvmsg
137893 1.8767 sysenter_past_esp
127473 1.7349 __copy_from_user_ll
121308 1.6510 ip_finish_output
118510 1.6129 tcp_transmit_skb
109295 1.4875 tcp_v4_rcv
Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The dst garbage collector dst_gc_task() may not be scheduled as we
expect it to be in __dst_free().
Indeed, when the dst_gc_timer was replaced by the delayed_work
dst_gc_work, the mod_timer() call used to schedule the garbage
collector at an earlier date was replaced by a schedule_delayed_work()
(see commit 86bba269d0).
But, the behaviour of mod_timer() and schedule_delayed_work() is
different in the way they handle the delay.
mod_timer() stops the timer and re-arm it with the new given delay,
whereas schedule_delayed_work() only check if the work is already
queued in the workqueue (and queue it (with delay) if it is not)
BUT it does NOT take into account the new delay (even if the new delay
is earlier in time).
schedule_delayed_work() returns 0 if it didn't queue the work,
but we don't check the return code in __dst_free().
If I understand the code in __dst_free() correctly, we want dst_gc_task
to be queued after DST_GC_INC jiffies if we pass the test (and not in
some undetermined time in the future), so I think we should add a call
to cancel_delayed_work() before schedule_delayed_work(). Patch below.
Or we should at least test the return code of schedule_delayed_work(),
and reset the values of dst_garbage.timer_inc and dst_garbage.timer_expires
back to their former values if schedule_delayed_work() failed.
Otherwise the subsequent calls to __dst_free will test the wrong values
and assume wrong thing about when the garbage collector is supposed to
be scheduled.
dst_gc_task() also calls schedule_delayed_work() without checking
its return code (or calling cancel_scheduled_work() first), but it
should fine there: dst_gc_task is the routine of the delayed_work, so
no dst_gc_work should be pending in the queue when it's running.
Signed-off-by: Benjamin Thery <benjamin.thery@bull.net>
Acked-by: Eric Dumazet <dada1@cosmosbay.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Eric Dumazet