seen during boot:
BUG: spinlock bad magic on CPU#2, swapper/0/1
lock: nf_connlabels_lock+0x0/0x60, .magic: 00000000, .owner: <none>/-1, .owner_cpu: 0
Call Trace:
do_raw_spin_lock+0x14e/0x1b0
nf_connlabels_get+0x15/0x40
ct_init_net+0xc4/0x270
ops_init+0x56/0x1c0
register_pernet_operations+0x1c8/0x350
register_pernet_subsys+0x1f/0x40
tcf_register_action+0x7c/0x1a0
do_one_initcall+0x13d/0x2d9
Problem is that ct action init function can run before
connlabels_init(). Lock has not been initialised yet.
Fix it by using a static initialiser.
Fixes: b57dc7c13e ("net/sched: Introduce action ct")
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Sparse warns about two tables not being declared.
CHECK net/netfilter/nf_nat_proto.c
net/netfilter/nf_nat_proto.c:725:26: warning: symbol 'nf_nat_ipv4_ops' was not declared. Should it be static?
net/netfilter/nf_nat_proto.c:964:26: warning: symbol 'nf_nat_ipv6_ops' was not declared. Should it be static?
And in fact they can indeed be static.
Signed-off-by: Valdis Kletnieks <valdis.kletnieks@vt.edu>
Acked-by: Florian Westphal <fw@strlen.de>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Sparse rightly complains about undeclared symbols.
CHECK net/netfilter/nft_set_hash.c
net/netfilter/nft_set_hash.c:647:21: warning: symbol 'nft_set_rhash_type' was not declared. Should it be static?
net/netfilter/nft_set_hash.c:670:21: warning: symbol 'nft_set_hash_type' was not declared. Should it be static?
net/netfilter/nft_set_hash.c:690:21: warning: symbol 'nft_set_hash_fast_type' was not declared. Should it be static?
CHECK net/netfilter/nft_set_bitmap.c
net/netfilter/nft_set_bitmap.c:296:21: warning: symbol 'nft_set_bitmap_type' was not declared. Should it be static?
CHECK net/netfilter/nft_set_rbtree.c
net/netfilter/nft_set_rbtree.c:470:21: warning: symbol 'nft_set_rbtree_type' was not declared. Should it be static?
Include nf_tables_core.h rather than nf_tables.h to pick up the additional definitions.
Signed-off-by: Valdis Kletnieks <valdis.kletnieks@vt.edu>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
linux/netfilter/ipset/ip_set.h included four other header files:
include/linux/netfilter/ipset/ip_set_comment.h
include/linux/netfilter/ipset/ip_set_counter.h
include/linux/netfilter/ipset/ip_set_skbinfo.h
include/linux/netfilter/ipset/ip_set_timeout.h
Of these the first three were not included anywhere else. The last,
ip_set_timeout.h, was included in a couple of other places, but defined
inline functions which call other inline functions defined in ip_set.h,
so ip_set.h had to be included before it.
Inlined all four into ip_set.h, and updated the other files that
included ip_set_timeout.h.
Signed-off-by: Jeremy Sowden <jeremy@azazel.net>
Acked-by: Jozsef Kadlecsik <kadlec@netfilter.org>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Store immediate data into offload context register. This allows follow
up instructions to take it from the corresponding source register.
This patch is required to support for payload mangling, although other
instructions that take data from source register will benefit from this
too.
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Extract mask from bitwise operation and store it into the corresponding
context register so the cmp instruction can set the mask accordingly.
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Let hidp_send_message return the number of successfully queued bytes
instead of an unconditional 0.
With the return value fixed to 0, other drivers relying on hidp, such as
hidraw, can not return meaningful values from their respective
implementations of write(). In particular, with the current behavior, a
hidraw device's write() will have different return values depending on
whether the device is connected via USB or Bluetooth, which makes it
harder to abstract away the transport layer.
Signed-off-by: Fabian Henneke <fabian.henneke@gmail.com>
Signed-off-by: Marcel Holtmann <marcel@holtmann.org>
We set the field 'addr_trial_end' to 'jiffies', instead of the current
value 0, at the moment the node address is initialized. This guarantees
we don't inadvertently enter an address trial period when the node
address is explicitly set by the user.
Signed-off-by: Chris Packham <chris.packham@alliedtelesis.co.nz>
Acked-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fix rxrpc_unuse_local() to handle a NULL local pointer as it can be called
on an unbound socket on which rx->local is not yet set.
The following reproduced (includes omitted):
int main(void)
{
socket(AF_RXRPC, SOCK_DGRAM, AF_INET);
return 0;
}
causes the following oops to occur:
BUG: kernel NULL pointer dereference, address: 0000000000000010
...
RIP: 0010:rxrpc_unuse_local+0x8/0x1b
...
Call Trace:
rxrpc_release+0x2b5/0x338
__sock_release+0x37/0xa1
sock_close+0x14/0x17
__fput+0x115/0x1e9
task_work_run+0x72/0x98
do_exit+0x51b/0xa7a
? __context_tracking_exit+0x4e/0x10e
do_group_exit+0xab/0xab
__x64_sys_exit_group+0x14/0x17
do_syscall_64+0x89/0x1d4
entry_SYSCALL_64_after_hwframe+0x49/0xbe
Reported-by: syzbot+20dee719a2e090427b5f@syzkaller.appspotmail.com
Fixes: 730c5fd42c ("rxrpc: Fix local endpoint refcounting")
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Jeffrey Altman <jaltman@auristor.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Previous patch made the length of the per-CPU skb drop list
configurable. Expose a counter that shows how many packets could not be
enqueued to this list.
This allows users determine the desired queue length.
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In packet alert mode, each CPU holds a list of dropped skbs that need to
be processed in process context and sent to user space. To avoid
exhausting the system's memory the maximum length of this queue is
currently set to 1000.
Allow users to tune the length of this queue according to their needs.
The configured length is reported to user space when drop monitor
configuration is queried.
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Users should be able to query the current configuration of drop monitor
before they start using it. Add a command to query the existing
configuration which currently consists of alert mode and packet
truncation length.
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When sending dropped packets to user space it is not always necessary to
copy the entire packet as usually only the headers are of interest.
Allow user to specify the truncation length and add the original length
of the packet as additional metadata to the netlink message.
By default no truncation is performed.
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
So far drop monitor supported only one alert mode in which a summary of
locations in which packets were recently dropped was sent to user space.
This alert mode is sufficient in order to understand that packets were
dropped, but lacks information to perform a more detailed analysis.
Add a new alert mode in which the dropped packet itself is passed to
user space along with metadata: The drop location (as program counter
and resolved symbol), ingress netdevice and drop timestamp. More
metadata can be added in the future.
To avoid performing expensive operations in the context in which
kfree_skb() is invoked (can be hard IRQ), the dropped skb is cloned and
queued on per-CPU skb drop list. Then, in process context the netlink
message is allocated, prepared and finally sent to user space.
The per-CPU skb drop list is limited to 1000 skbs to prevent exhausting
the system's memory. Subsequent patches will make this limit
configurable and also add a counter that indicates how many skbs were
tail dropped.
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The next patch is going to add another alert mode in which the dropped
packet is notified to user space, instead of only a summary of recent
drops.
Abstract the differences between the modes by adding alert mode
operations. The operations are selected based on the currently
configured mode and associated with the probes and the work item just
before tracing starts.
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently, the configure command does not do anything but return an
error. Subsequent patches will enable the command to change various
configuration options such as alert mode and packet truncation.
Similar to other netlink-based configuration channels, make sure only
users with the CAP_NET_ADMIN capability set can execute this command.
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The function reset_per_cpu_data() allocates and prepares a new skb for
the summary netlink alert message ('NET_DM_CMD_ALERT'). The new skb is
stored in the per-CPU 'data' variable and the old is returned.
The function is invoked during module initialization and from the
workqueue, before an alert is sent. This means that it is possible to
receive an alert with stale data, if we stopped tracing when the
hysteresis timer ('data->send_timer') was pending.
Instead of invoking the function during module initialization, invoke it
just before we start tracing and ensure we get a fresh skb.
This also allows us to remove the calls to initialize the timer and the
work item from the module initialization path, since both could have
been triggered by the error paths of reset_per_cpu_data().
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The timer and work item are currently initialized once during module
init, but subsequent patches will need to associate different functions
with the work item, based on the configured alert mode.
Allow subsequent patches to make that change by initializing and
de-initializing these objects during tracing enable and disable.
This also guarantees that once the request to disable tracing returns,
no more netlink notifications will be generated.
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Subsequent patches will need to enable / disable tracing based on the
configured alerting mode.
Reduce the nesting level and prepare for the introduction of this
functionality by splitting the tracing enable / disable operations into
two different functions.
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When calling debugfs functions, there is no need to ever check the
return value. The function can work or not, but the code logic should
never do something different based on this.
This cleans up a lot of unneeded code and logic around the debugfs wimax
files, making all of this much simpler and easier to understand.
Cc: Inaky Perez-Gonzalez <inaky.perez-gonzalez@intel.com>
Cc: linux-wimax@intel.com
Cc: netdev@vger.kernel.org
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Now that we swap the original proto and clear the ULP pointer
on close we have to make sure no callback will try to access
the freed state. sk_write_space is not part of sk_prot, remember
to swap it.
Reported-by: syzbot+dcdc9deefaec44785f32@syzkaller.appspotmail.com
Fixes: 95fa145479 ("bpf: sockmap/tls, close can race with map free")
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/sched/sch_taprio.c:680:32: warning:
entry_list_policy defined but not used [-Wunused-const-variable=]
One of the points of commit a3d43c0d56 ("taprio: Add support adding
an admin schedule") is that it removes support (it now returns "not
supported") for schedules using the TCA_TAPRIO_ATTR_SCHED_SINGLE_ENTRY
attribute (which were never used), the parsing of those types of schedules
was the only user of this policy. So removing this policy should be fine.
Reported-by: Hulk Robot <hulkci@huawei.com>
Suggested-by: Vinicius Costa Gomes <vinicius.gomes@intel.com>
Signed-off-by: YueHaibing <yuehaibing@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Generating and retrieving socket cookies are a useful feature that is
exposed to BPF for various program types through bpf_get_socket_cookie()
helper.
The fact that the cookie counter is per netns is quite a limitation
for BPF in practice in particular for programs in host namespace that
use socket cookies as part of a map lookup key since they will be
causing socket cookie collisions e.g. when attached to BPF cgroup hooks
or cls_bpf on tc egress in host namespace handling container traffic
from veth or ipvlan devices with peer in different netns. Change the
counter to be global instead.
Socket cookie consumers must assume the value as opqaue in any case.
Not every socket must have a cookie generated and knowledge of the
counter value itself does not provide much value either way hence
conversion to global is fine.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Alexei Starovoitov <ast@kernel.org>
Cc: Willem de Bruijn <willemb@google.com>
Cc: Martynas Pumputis <m@lambda.lt>
Signed-off-by: David S. Miller <davem@davemloft.net>
The current implementation of TCP MTU probing can considerably
underestimate the MTU on lossy connections allowing the MSS to get down to
48. We have found that in almost all of these cases on our networks these
paths can handle much larger MTUs meaning the connections are being
artificially limited. Even though TCP MTU probing can raise the MSS back up
we have seen this not to be the case causing connections to be "stuck" with
an MSS of 48 when heavy loss is present.
Prior to pushing out this change we could not keep TCP MTU probing enabled
b/c of the above reasons. Now with a reasonble floor set we've had it
enabled for the past 6 months.
The new sysctl will still default to TCP_MIN_SND_MSS (48), but gives
administrators the ability to control the floor of MSS probing.
Signed-off-by: Josh Hunt <johunt@akamai.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The size of the snapshot has to be the same as the size of the region,
therefore no need to pass it again during snapshot creation. Remove the
arg and use region->size instead.
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Starting from commit d41a69f1d3 ("tcp: make tcp_sendmsg() aware of socket backlog")
loopback flows got hurt, because for each skb sent, the socket receives an
immediate ACK and sk_flush_backlog() causes extra work.
Intent was to not let the backlog grow too much, but we went a bit too far.
We can check the backlog every 16 skbs (about 1MB chunks)
to increase TCP over loopback performance by about 15 %
Note that the call to sk_flush_backlog() handles a single ACK,
thanks to coalescing done on backlog, but cleans the 16 skbs
found in rtx rb-tree.
Reported-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This commit applies the consolidated list_for_each_entry_rcu() support
for lockdep conditions.
Acked-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
Don't bother generating maxSkew in the ACK packet as it has been obsolete
since AFS 3.1.
Signed-off-by: David Howells <dhowells@redhat.com>
Reviewed-by: Jeffrey Altman <jaltman@auristor.com>
The object lifetime management on the rxrpc_local struct is broken in that
the rxrpc_local_processor() function is expected to clean up and remove an
object - but it may get requeued by packets coming in on the backing UDP
socket once it starts running.
This may result in the assertion in rxrpc_local_rcu() firing because the
memory has been scheduled for RCU destruction whilst still queued:
rxrpc: Assertion failed
------------[ cut here ]------------
kernel BUG at net/rxrpc/local_object.c:468!
Note that if the processor comes around before the RCU free function, it
will just do nothing because ->dead is true.
Fix this by adding a separate refcount to count active users of the
endpoint that causes the endpoint to be destroyed when it reaches 0.
The original refcount can then be used to refcount objects through the work
processor and cause the memory to be rcu freed when that reaches 0.
Fixes: 4f95dd78a7 ("rxrpc: Rework local endpoint management")
Reported-by: syzbot+1e0edc4b8b7494c28450@syzkaller.appspotmail.com
Signed-off-by: David Howells <dhowells@redhat.com>
Flows that are in teardown state (due to RST / FIN TCP packet) still
have their offload flag set on. Hence, the conntrack garbage collector
may race to undo the timeout adjustment that the fixup routine performs,
leaving the conntrack entry in place with the internal offload timeout
(one day).
Update teardown flow state to ESTABLISHED and set tracking to liberal,
then once the offload bit is cleared, adjust timeout if it is more than
the default fixup timeout (conntrack might already have set a lower
timeout from the packet path).
Fixes: da5984e510 ("netfilter: nf_flow_table: add support for sending flows back to the slow path")
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Update conntrack entry to pick up expired flows, otherwise the conntrack
entry gets stuck with the internal offload timeout (one day). The TCP
state also needs to be adjusted to ESTABLISHED state and tracking is set
to liberal mode in order to give conntrack a chance to pick up the
expired flow.
Fixes: ac2a66665e ("netfilter: add generic flow table infrastructure")
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
If a rule that has already a bound anonymous set fails to be added, the
preparation phase releases the rule and the bound set. However, the
transaction object from the abort path still has a reference to the set
object that is stale, leading to a use-after-free when checking for the
set->bound field. Add a new field to the transaction that specifies if
the set is bound, so the abort path can skip releasing it since the rule
command owns it and it takes care of releasing it. After this update,
the set->bound field is removed.
[ 24.649883] Unable to handle kernel paging request at virtual address 0000000000040434
[ 24.657858] Mem abort info:
[ 24.660686] ESR = 0x96000004
[ 24.663769] Exception class = DABT (current EL), IL = 32 bits
[ 24.669725] SET = 0, FnV = 0
[ 24.672804] EA = 0, S1PTW = 0
[ 24.675975] Data abort info:
[ 24.678880] ISV = 0, ISS = 0x00000004
[ 24.682743] CM = 0, WnR = 0
[ 24.685723] user pgtable: 4k pages, 48-bit VAs, pgdp=0000000428952000
[ 24.692207] [0000000000040434] pgd=0000000000000000
[ 24.697119] Internal error: Oops: 96000004 [#1] SMP
[...]
[ 24.889414] Call trace:
[ 24.891870] __nf_tables_abort+0x3f0/0x7a0
[ 24.895984] nf_tables_abort+0x20/0x40
[ 24.899750] nfnetlink_rcv_batch+0x17c/0x588
[ 24.904037] nfnetlink_rcv+0x13c/0x190
[ 24.907803] netlink_unicast+0x18c/0x208
[ 24.911742] netlink_sendmsg+0x1b0/0x350
[ 24.915682] sock_sendmsg+0x4c/0x68
[ 24.919185] ___sys_sendmsg+0x288/0x2c8
[ 24.923037] __sys_sendmsg+0x7c/0xd0
[ 24.926628] __arm64_sys_sendmsg+0x2c/0x38
[ 24.930744] el0_svc_common.constprop.0+0x94/0x158
[ 24.935556] el0_svc_handler+0x34/0x90
[ 24.939322] el0_svc+0x8/0xc
[ 24.942216] Code: 37280300 f9404023 91014262 aa1703e0 (f9401863)
[ 24.948336] ---[ end trace cebbb9dcbed3b56f ]---
Fixes: f6ac858589 ("netfilter: nf_tables: unbind set in rule from commit path")
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
sk_validate_xmit_skb() and drivers depend on the sk member of
struct sk_buff to identify segments requiring encryption.
Any operation which removes or does not preserve the original TLS
socket such as skb_orphan() or skb_clone() will cause clear text
leaks.
Make the TCP socket underlying an offloaded TLS connection
mark all skbs as decrypted, if TLS TX is in offload mode.
Then in sk_validate_xmit_skb() catch skbs which have no socket
(or a socket with no validation) and decrypted flag set.
Note that CONFIG_SOCK_VALIDATE_XMIT, CONFIG_TLS_DEVICE and
sk->sk_validate_xmit_skb are slightly interchangeable right now,
they all imply TLS offload. The new checks are guarded by
CONFIG_TLS_DEVICE because that's the option guarding the
sk_buff->decrypted member.
Second, smaller issue with orphaning is that it breaks
the guarantee that packets will be delivered to device
queues in-order. All TLS offload drivers depend on that
scheduling property. This means skb_orphan_partial()'s
trick of preserving partial socket references will cause
issues in the drivers. We need a full orphan, and as a
result netem delay/throttling will cause all TLS offload
skbs to be dropped.
Reusing the sk_buff->decrypted flag also protects from
leaking clear text when incoming, decrypted skb is redirected
(e.g. by TC).
See commit 0608c69c9a ("bpf: sk_msg, sock{map|hash} redirect
through ULP") for justification why the internal flag is safe.
The only location which could leak the flag in is tcp_bpf_sendmsg(),
which is taken care of by clearing the previously unused bit.
v2:
- remove superfluous decrypted mark copy (Willem);
- remove the stale doc entry (Boris);
- rely entirely on EOR marking to prevent coalescing (Boris);
- use an internal sendpages flag instead of marking the socket
(Boris).
v3 (Willem):
- reorganize the can_skb_orphan_partial() condition;
- fix the flag leak-in through tcp_bpf_sendmsg.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Acked-by: Willem de Bruijn <willemb@google.com>
Reviewed-by: Boris Pismenny <borisp@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add get_fill_size() routine used to calculate the action size
when building a batch of events.
Fixes: ca9b0e27e ("pkt_action: add new action skbedit")
Signed-off-by: Roman Mashak <mrv@mojatatu.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fixes gcc '-Wunused-but-set-variable' warning:
net/sched/sch_fq_codel.c: In function fq_codel_dequeue:
net/sched/sch_fq_codel.c:288:23: warning: variable prev_ecn_mark set but not used [-Wunused-but-set-variable]
net/sched/sch_fq_codel.c:288:6: warning: variable prev_drop_count set but not used [-Wunused-but-set-variable]
They are not used since commit 77ddaff218 ("fq_codel: Kill
useless per-flow dropped statistic")
Reported-by: Hulk Robot <hulkci@huawei.com>
Signed-off-by: YueHaibing <yuehaibing@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Since node internal messages are passed directly to the socket, it is not
possible to observe those messages via tcpdump or wireshark.
We now remedy this by making it possible to clone such messages and send
the clones to the loopback interface. The clones are dropped at reception
and have no functional role except making the traffic visible.
The feature is enabled if network taps are active for the loopback device.
pcap filtering restrictions require the messages to be presented to the
receiving side of the loopback device.
v3 - Function dev_nit_active used to check for network taps.
- Procedure netif_rx_ni used to send cloned messages to loopback device.
Signed-off-by: John Rutherford <john.rutherford@dektech.com.au>
Acked-by: Jon Maloy <jon.maloy@ericsson.com>
Acked-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In error case, all entries should be freed from the sched list
before deleting it. For simplicity use rcu way.
Fixes: 5a781ccbd1 ("tc: Add support for configuring the taprio scheduler")
Acked-by: Vinicius Costa Gomes <vinicius.gomes@intel.com>
Signed-off-by: Ivan Khoronzhuk <ivan.khoronzhuk@linaro.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
move tc indirect block to flow_offload and rename
it to flow indirect block.The nf_tables can use the
indr block architecture.
Signed-off-by: wenxu <wenxu@ucloud.cn>
Acked-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When GRO decides not to coalesce a packet, in napi_frags_finish(), instead
of passing it to the stack immediately, place it on a list in the napi
struct. Then, at flush time (napi_complete_done(), napi_poll(), or
napi_busy_loop()), call netif_receive_skb_list_internal() on the list.
We'd like to do that in napi_gro_flush(), but it's not called if
!napi->gro_bitmask, so we have to do it in the callers instead. (There are
a handful of drivers that call napi_gro_flush() themselves, but it's not
clear why, or whether this will affect them.)
Because a full 64 packets is an inefficiently large batch, also consume the
list whenever it exceeds gro_normal_batch, a new net/core sysctl that
defaults to 8.
Signed-off-by: Edward Cree <ecree@solarflare.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Before commit d4289fcc9b ("net: IP6 defrag: use rbtrees for IPv6
defrag"), a netperf UDP_STREAM test[0] using big IPv6 datagrams (thus
generating many fragments) and running over an IPsec tunnel, reported
more than 6Gbps throughput. After that patch, the same test gets only
9Mbps when receiving on a be2net nic (driver can make a big difference
here, for example, ixgbe doesn't seem to be affected).
By reusing the IPv4 defragmentation code, IPv6 lost fragment coalescing
(IPv4 fragment coalescing was dropped by commit 14fe22e334 ("Revert
"ipv4: use skb coalescing in defragmentation"")).
Without fragment coalescing, be2net runs out of Rx ring entries and
starts to drop frames (ethtool reports rx_drops_no_frags errors). Since
the netperf traffic is only composed of UDP fragments, any lost packet
prevents reassembly of the full datagram. Therefore, fragments which
have no possibility to ever get reassembled pile up in the reassembly
queue, until the memory accounting exeeds the threshold. At that point
no fragment is accepted anymore, which effectively discards all
netperf traffic.
When reassembly timeout expires, some stale fragments are removed from
the reassembly queue, so a few packets can be received, reassembled
and delivered to the netperf receiver. But the nic still drops frames
and soon the reassembly queue gets filled again with stale fragments.
These long time frames where no datagram can be received explain why
the performance drop is so significant.
Re-introducing fragment coalescing is enough to get the initial
performances again (6.6Gbps with be2net): driver doesn't drop frames
anymore (no more rx_drops_no_frags errors) and the reassembly engine
works at full speed.
This patch is quite conservative and only coalesces skbs for local
IPv4 and IPv6 delivery (in order to avoid changing skb geometry when
forwarding). Coalescing could be extended in the future if need be, as
more scenarios would probably benefit from it.
[0]: Test configuration
Sender:
ip xfrm policy flush
ip xfrm state flush
ip xfrm state add src fc00:1::1 dst fc00:2::1 proto esp spi 0x1000 aead 'rfc4106(gcm(aes))' 0x0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b 96 mode transport sel src fc00:1::1 dst fc00:2::1
ip xfrm policy add src fc00:1::1 dst fc00:2::1 dir in tmpl src fc00:1::1 dst fc00:2::1 proto esp mode transport action allow
ip xfrm state add src fc00:2::1 dst fc00:1::1 proto esp spi 0x1001 aead 'rfc4106(gcm(aes))' 0x0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b 96 mode transport sel src fc00:2::1 dst fc00:1::1
ip xfrm policy add src fc00:2::1 dst fc00:1::1 dir out tmpl src fc00:2::1 dst fc00:1::1 proto esp mode transport action allow
netserver -D -L fc00:2::1
Receiver:
ip xfrm policy flush
ip xfrm state flush
ip xfrm state add src fc00:2::1 dst fc00:1::1 proto esp spi 0x1001 aead 'rfc4106(gcm(aes))' 0x0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b 96 mode transport sel src fc00:2::1 dst fc00:1::1
ip xfrm policy add src fc00:2::1 dst fc00:1::1 dir in tmpl src fc00:2::1 dst fc00:1::1 proto esp mode transport action allow
ip xfrm state add src fc00:1::1 dst fc00:2::1 proto esp spi 0x1000 aead 'rfc4106(gcm(aes))' 0x0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b0b 96 mode transport sel src fc00:1::1 dst fc00:2::1
ip xfrm policy add src fc00:1::1 dst fc00:2::1 dir out tmpl src fc00:1::1 dst fc00:2::1 proto esp mode transport action allow
netperf -H fc00:2::1 -f k -P 0 -L fc00:1::1 -l 60 -t UDP_STREAM -I 99,5 -i 5,5 -T5,5 -6
Signed-off-by: Guillaume Nault <gnault@redhat.com>
Acked-by: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>