- 16 Oct, 2015 37 commits
-
-
Jesse Brandeburg authored
As per Eric Dumazet's previous patches: (see commit (24d2e4a5) - tg3: use napi_complete_done()) Quoting verbatim: Using napi_complete_done() instead of napi_complete() allows us to use /sys/class/net/ethX/gro_flush_timeout GRO layer can aggregate more packets if the flush is delayed a bit, without having to set too big coalescing parameters that impact latencies. </end quote> Tested configuration: low latency via ethtool -C ethx adaptive-rx off rx-usecs 10 adaptive-tx off tx-usecs 15 workload: streaming rx using netperf TCP_MAERTS igb: MIGRATED TCP MAERTS TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to 10.0.0.1 () port 0 AF_INET : demo ... Interim result: 941.48 10^6bits/s over 1.000 seconds ending at 1440193171.589 Alignment Offset Bytes Bytes Recvs Bytes Sends Local Remote Local Remote Xfered Per Per Recv Send Recv Send Recv (avg) Send (avg) 8 8 0 0 1176930056 1475.36 797726 16384.00 71905 MIGRATED TCP MAERTS TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to 10.0.0.1 () port 0 AF_INET : demo ... Interim result: 941.49 10^6bits/s over 0.997 seconds ending at 1440193142.763 Alignment Offset Bytes Bytes Recvs Bytes Sends Local Remote Local Remote Xfered Per Per Recv Send Recv Send Recv (avg) Send (avg) 8 8 0 0 1175182320 50476.00 23282 16384.00 71816 i40e: Hard to test because the traffic is incoming so fast (24Gb/s) that GRO always receives 87kB, even at the highest interrupt rate. Other drivers were only compile tested. Signed-off-by:
Jesse Brandeburg <jesse.brandeburg@intel.com> Tested-by:
Andrew Bowers <andrewx.bowers@intel.com> Signed-off-by:
Jeff Kirsher <jeffrey.t.kirsher@intel.com>
-
Alexander Duyck authored
Signed-off-by:
Alexander Duyck <aduyck@mirantis.com> Tested-by:
-
Alexander Duyck authored
The code in i40e and i40evf is using an "IN_NETPOLL" flag that has never added any value due to the fact that the Rx clean-up is handled in NAPI. As such the flag was set, the queue was scheduled via NAPI, and then polled from the netpoll controller and if any Rx packets were processed the were processed in the wrong context. In addition the flag itself just added an unneeded conditional to the hot-path so it can safely be dropped and save us a few instructions. Signed-off-by:
-
Alexander Duyck authored
The polling routine for i40e was rounding up the budget for Rx cleanup to 1. This is incorrect as the netpoll poll call is expecting no Rx to be processed as the budget passed was 0. Signed-off-by:
-
David Ahern authored
This command: ip route add 192.168.1.0/24 nexthop via 10.2.1.5 dev eth1 nexthop via 10.2.2.5 dev eth2 generated this suspicious RCU usage message: [ 63.249262] [ 63.249939] =============================== [ 63.251571] [ INFO: suspicious RCU usage. ] [ 63.253250] 4.3.0-rc3+ #298 Not tainted [ 63.254724] ------------------------------- [ 63.256401] ../include/linux/inetdevice.h:205 suspicious rcu_dereference_check() usage! [ 63.259450] [ 63.259450] other info that might help us debug this: [ 63.259450] [ 63.262297] [ 63.262297] rcu_scheduler_active = 1, debug_locks = 1 [ 63.264647] 1 lock held by ip/2870: [ 63.265896] #0: (rtnl_mutex){+.+.+.}, at: [<ffffffff813ebfb7>] rtnl_lock+0x12/0x14 [ 63.268858] [ 63.268858] stack backtrace: [ 63.270409] CPU: 4 PID: 2870 Comm: ip Not tainted 4.3.0-rc3+ #298 [ 63.272478] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.7.5-20140531_083030-gandalf 04/01/2014 [ 63.275745] 0000000000000001 ffff8800b8c9f8b8 ffffffff8125f73c ffff88013afcf301 [ 63.278185] ffff8800bab7a380 ffff8800b8c9f8e8 ffffffff8107bf30 ffff8800bb728000 [ 63.280634] ffff880139fe9a60 0000000000000000 ffff880139fe9a00 ffff8800b8c9f908 [ 63.283177] Call Trace: [ 63.283959] [<ffffffff8125f73c>] dump_stack+0x4c/0x68 [ 63.285593] [<ffffffff8107bf30>] lockdep_rcu_suspicious+0xfa/0x103 [ 63.287500] [<ffffffff8144d752>] __in_dev_get_rcu+0x48/0x4f [ 63.289169] [<ffffffff8144d797>] fib_rebalance+0x3e/0x127 [ 63.290753] [<ffffffff8144d986>] ? rcu_read_unlock+0x3e/0x5f [ 63.292442] [<ffffffff8144ea45>] fib_create_info+0xaf9/0xdcc [ 63.294093] [<ffffffff8106c12f>] ? sched_clock_local+0x12/0x75 [ 63.295791] [<ffffffff8145236a>] fib_table_insert+0x8c/0x451 [ 63.297493] [<ffffffff8144bf9c>] ? fib_get_table+0x36/0x43 [ 63.299109] [<ffffffff8144c3ca>] inet_rtm_newroute+0x43/0x51 [ 63.300709] [<ffffffff813ef684>] rtnetlink_rcv_msg+0x182/0x195 [ 63.302334] [<ffffffff8107d04c>] ? trace_hardirqs_on+0xd/0xf [ 63.303888] [<ffffffff813ebfb7>] ? rtnl_lock+0x12/0x14 [ 63.305346] [<ffffffff813ef502>] ? __rtnl_unlock+0x12/0x12 [ 63.306878] [<ffffffff81407c4c>] netlink_rcv_skb+0x3d/0x90 [ 63.308437] [<ffffffff813ec00e>] rtnetlink_rcv+0x21/0x28 [ 63.309916] [<ffffffff81407742>] netlink_unicast+0xfa/0x17f [ 63.311447] [<ffffffff81407a5e>] netlink_sendmsg+0x297/0x2dc [ 63.313029] [<ffffffff813c6cd4>] sock_sendmsg_nosec+0x12/0x1d [ 63.314597] [<ffffffff813c835b>] ___sys_sendmsg+0x196/0x21b [ 63.316125] [<ffffffff8100bf9f>] ? native_sched_clock+0x1f/0x3c [ 63.317671] [<ffffffff8106c12f>] ? sched_clock_local+0x12/0x75 [ 63.319185] [<ffffffff8106c397>] ? sched_clock_cpu+0x9d/0xb6 [ 63.320693] [<ffffffff8107e2d7>] ? __lock_is_held+0x32/0x54 [ 63.322145] [<ffffffff81159fcb>] ? __fget_light+0x4b/0x77 [ 63.323541] [<ffffffff813c8726>] __sys_sendmsg+0x3d/0x5b [ 63.324947] [<ffffffff813c8751>] SyS_sendmsg+0xd/0x19 [ 63.326274] [<ffffffff814c8f57>] entry_SYSCALL_64_fastpath+0x12/0x6f It looks like all of the code paths to fib_rebalance are under rtnl. Fixes: 0e884c78 ("ipv4: L3 hash-based multipath") Cc: Peter Nørlund <pch@ordbogen.com> Signed-off-by:David Ahern <dsa@cumulusnetworks.com> Signed-off-by:
David S. Miller <davem@davemloft.net>
-
Tom Herbert authored
Currently, is only called from __prog_put_rcu in the bpf_prog_release path. Need this to call this from bpf_prog_put also to get correct accounting. Fixes: aaac3ba9 ("bpf: charge user for creation of BPF maps and programs") Signed-off-by:
Tom Herbert <tom@herbertland.com> Acked-by:
Alexei Starovoitov <ast@kernel.org> Acked-by:
Daniel Borkmann <daniel@iogearbox.net> Signed-off-by:
-
David S. Miller authored
Eric Dumazet says: ==================== tcp/dccp: make our listener code more robust This patch series addresses request sockets leaks and listener dismantle phase. This survives a stress test with listeners being added/removed quite randomly. ==================== Signed-off-by: -
Eric Dumazet authored
Under stress, a close() on a listener can trigger the WARN_ON(sk->sk_ack_backlog) in inet_csk_listen_stop() We need to test if listener is still active before queueing a child in inet_csk_reqsk_queue_add() Create a common inet_child_forget() helper, and use it from inet_csk_reqsk_queue_add() and inet_csk_listen_stop() Signed-off-by:
Eric Dumazet <edumazet@google.com> Signed-off-by:
-
Eric Dumazet authored
Let's reduce the confusion about inet_csk_reqsk_queue_drop() : In many cases we also need to release reference on request socket, so add a helper to do this, reducing code size and complexity. Fixes: 4bdc3d66 ("tcp/dccp: fix behavior of stale SYN_RECV request sockets") Signed-off-by:
-
Eric Dumazet authored
This reverts commit c6973669. At the time of above commit, tcp_req_err() and dccp_req_err() were dead code, as SYN_RECV request sockets were not yet in ehash table. Real bug was fixed later in a different commit. We need to revert to not leak a refcount on request socket. inet_csk_reqsk_queue_drop_and_put() will be added in following commit to make clean inet_csk_reqsk_queue_drop() does not release the reference owned by caller. Signed-off-by:
-
Ivan Vecera authored
Many drivers initialize uselessly n_priv_flags, n_stats, testinfo_len, eedump_len & regdump_len fields in their .get_drvinfo() ethtool op. It's not necessary as these fields is filled in ethtool_get_drvinfo(). v2: removed unused variable v3: removed another unused variable Signed-off-by:
Ivan Vecera <ivecera@redhat.com> Signed-off-by:
-
David S. Miller authored
Jon Maloy says: ==================== tipc: some link level code improvements Extensive testing has revealed some weaknesses and non-optimal solutions in the link level code. This commit series addresses those issues. ==================== Signed-off-by: -
Jon Paul Maloy authored
The change made in the previous commit revealed a small flaw in the way the node FSM is updated. When the function tipc_node_link_down() is called for the last link to a node, we should check whether this was caused by a local reset or by a received RESET message from the peer. In the latter case, we can directly issue a PEER_LOST_CONTACT_EVT to the node FSM, so that it is ready to re-establish contact. If this is not done, the peer node will sometimes have to go through a second establish cycle before the link becomes stable. We fix this in this commit by conditionally issuing the mentioned event in the function tipc_node_link_down(). We also move LINK_RESET FSM even away from the link_reset() function and into the caller function, partially because it is easier to follow the code when state changes are gathered at a limited number of locations, partially because there will be cases in future commits where we don't want the link to go RESET mode when link_reset() is called. Signed-off-by:
Jon Maloy <jon.maloy@ericsson.com> Acked-by:
Ying Xue <ying.xue@windriver.com> Signed-off-by:
-
Jon Paul Maloy authored
When a link is taken down because of a node local event, such as disabling of a bearer or an interface, we currently leave it to the peer node to discover the broken communication. The default time for such failure discovery is 1.5-2 seconds. If we instead allow the terminating link endpoint to send out a RESET message at the moment it is reset, we can achieve the impression that both endpoints are going down instantly. Since this is a very common scenario, we find it worthwhile to make this small modification. Apart from letting the link produce the said message, we also have to ensure that the interface is able to transmit it before TIPC is detached. We do this by performing the disabling of a bearer in three steps: 1) Disable reception of TIPC packets from the interface in question. 2) Take down the links, while allowing them so send out a RESET message. 3) Disable transmission of TIPC packets on the interface. Apart from this, we now have to react on the NETDEV_GOING_DOWN event, instead of as currently the NEDEV_DOWN event, to ensure that such transmission is possible during the teardown phase. Signed-off-by:
-
Jon Paul Maloy authored
Link establishing, just like link teardown, is a non-atomic action, in the sense that discovering that conditions are right to establish a link, and the actual adding of the link to one of the node's send slots is done in two different lock contexts. The link FSM is designed to help bridging the gap between the two contexts in a safe manner. We have now discovered a weakness in the implementaton of this FSM. Because we directly let the link go from state LINK_ESTABLISHING to state LINK_ESTABLISHED already in the first lock context, we are unable to distinguish between a fully established link, i.e., a link that has been added to its slot, and a link that has not yet reached the second lock context. It may hence happen that a manual intervention, e.g., when disabling an interface, causes the function tipc_node_link_down() to try removing the link from the node slots, decrementing its active link counter etc, although the link was never added there in the first place. We solve this by delaying the actual state change until we reach the second lock context, inside the function tipc_node_link_up(). This makes it possible for potentail callers of __tipc_node_link_down() to know if they should proceed or not, and the problem is solved. Unforunately, the situation described above also has a second problem. Since there by necessity is a tipc_node_link_up() call pending once the node lock has been released, we must defuse that call by setting the link back from LINK_ESTABLISHING to LINK_RESET state. This forces us to make a slight modification to the link FSM, which will now look as follows. +------------------------------------+ |RESET_EVT | | | | +--------------+ | +-----------------| SYNCHING |-----------------+ | |FAILURE_EVT +--------------+ PEER_RESET_EVT| | | A | | | | | | | | | | | | | | |SYNCH_ |SYNCH_ | | | |BEGIN_EVT |END_EVT | | | | | | | V | V V | +-------------+ +--------------+ +------------+ | | RESETTING |<---------| ESTABLISHED |--------->| PEER_RESET | | +-------------+ FAILURE_ +--------------+ PEER_ +------------+ | | EVT | A RESET_EVT | | | | | | | | +----------------+ | | | RESET_EVT| |RESET_EVT | | | | | | | | | | |ESTABLISH_EVT | | | | +-------------+ | | | | | | RESET_EVT | | | | | | | | | | | V V V | | | | +-------------+ +--------------+ RESET_EVT| +--->| RESET |--------->| ESTABLISHING |<----------------+ +-------------+ PEER_ +--------------+ | A RESET_EVT | | | | | | | |FAILOVER_ |FAILOVER_ |FAILOVER_ |BEGIN_EVT |END_EVT |BEGIN_EVT | | | V | | +-------------+ | | FAILINGOVER |<----------------+ +-------------+ Signed-off-by: -
Jon Paul Maloy authored
After the previous commits, we are guaranteed that no packets of type LINK_PROTOCOL or with illegal sequence numbers will be attempted added to the link deferred queue. This makes it possible to make some simplifications to the sorting algorithm in the function tipc_skb_queue_sorted(). We also alter the function so that it will drop packets if one with the same seqeunce number is already present in the queue. This is necessary because we have identified weird packet sequences, involving duplicate packets, where a legitimate in-sequence packet may advance to the head of the queue without being detected and de-queued. Finally, we make this function outline, since it will now be called only in exceptional cases. Signed-off-by:
-
Jon Paul Maloy authored
The sequence number of an incoming packet is currently only checked for less than, equality to, or bigger than the next expected number, meaning that the receive window in practice becomes one half sequence number cycle, or U16_MAX/2. This does not make sense, and may not even be safe if there are extreme delays in the network. Any packet sent by the peer during the ongoing cycle must belong inside his current send window, or should otherwise be dropped if possible. Since a link endpoint cannot know its peer's current send window, it has to base this sanity check on a worst-case assumption, i.e., that the peer is using a maximum sized window of 8191 packets. Using this assumption, we now add a check that the sequence number is not bigger than next_expected + TIPC_MAX_LINK_WIN. We also re-order the checks done, so that the receive window test is performed before the gap test. This way, we are guaranteed that no packet with illegal sequence numbers are ever added to the deferred queue. Signed-off-by:
-
Jon Paul Maloy authored
Currently, all packets received in tipc_link_rcv() are unconditionally added to the packet deferred queue, whereafter that queue is walked and all its buffers evaluated for delivery. This is both non-optimal and and makes the queue sorting function unnecessary complex. This commit changes the loop so that an arrived packet is evaluated first, and added to the deferred queue only when a sequence number gap is discovered. A non-empty deferred queue is walked until it is empty or until its head's sequence number doesn't fit. Signed-off-by:
-
Jon Paul Maloy authored
During packet reception, the function tipc_link_rcv() adds its accepted packets to a temporary buffer queue, before finally splicing this queue into the lock protected input queue that will be delivered up to the socket layer. The purpose is to reduce potential contention on the input queue lock. However, since the vast majority of packets arrive in sequence, they will anyway be added one by one to the input queue, and the use of the temporary queue becomes a sub-optimization. The only case where this queue makes sense is when unpacking buffers from a bundle packet; here we want to avoid dozens of small buffers to be added individually to the lock-protected input queue in a tight loop. In this commit, we remove the general usage of the temporary queue, and keep it only for the packet unbundling case. Signed-off-by:
-
Insu Yun authored
When allocation fails, mlx4_alloc_cmd_mailbox returns -ENOMEM. Since there is no case that mlx4_alloc_cmd_mailbox returns NULL, it needs to be checked by IS_ERR, not IS_ERR_OR_NULL Signed-off-by:
Insu Yun <wuninsu@gmail.com> Signed-off-by:
-
Eric Dumazet authored
If using a sixtofour device on top of a bonding device, skb segmentation of TCP traffic is done right before calling bonding xmit, because bonding only enables TSO for IPv4. This patch improves single flow performance by about 120 % on my hosts, because segmentation is deferred right before calling slave xmit. Signed-off-by:
-
David S. Miller authored
Jiri Pirko says: ==================== mlxsw: Driver update, cleanups This patchset contains various cleanups and improvements in mlxsw driver. ==================== Signed-off-by: -
Ido Schimmel authored
The meaning of certain parameters in the profile passed to the device during initialization has changed, so update their documentation accordingly. Signed-off-by:
Ido Schimmel <idosch@mellanox.com> Signed-off-by:
Jiri Pirko <jiri@mellanox.com> Signed-off-by:
-
Ido Schimmel authored
Previously, we trapped flooded and control packets using the same trap group. This can cause flooded packets to overflow the PCI bus and prevent control packets (e.g. STP, LACP) from getting to the CPU. Solve this by splitting the RX trap group to RX and control, which allows us to configure a policer on the first, thereby preventing it from overflowing the PCI bus. Signed-off-by:
-
Ido Schimmel authored
The Host Trap Group Table (HTGT) register configures trap groups, which are populated with trap IDs using the Host PacKet Trap (HPKT) register. However, a trap ID can only be present inside one trap group (the last configured). Instead of passing both the trap group and ID for the function that packs HPKT, pass only the trap ID and derive from it the trap group. Signed-off-by:
-
Jiri Pirko authored
Introduce separate helper for packing SPMS VIDs, as it can be used for multiple VIDs and not only for one as previous SPMS pack function provided. Signed-off-by:
-
Ido Schimmel authored
Define max which would be needed later on. Signed-off-by:
-
Jiri Pirko authored
Signed-off-by:
-
Jiri Pirko authored
Signed-off-by:
-
Jiri Pirko authored
Signed-off-by:
-
Jiri Pirko authored
Be symmetric with mlxsw_emad_init and don't use EMADs in mlxsw_emad_fini cleanup function. Use command interface instead. Signed-off-by:
-
Jiri Pirko authored
Firmware accepts only limited number of mapping entries for MAP_FA command. In order to prevent overflow, introduce a limit and in case the number of entries is bigger, call MAP_FA multiple times. Signed-off-by:
-
Jiri Pirko authored
Remove strict number check of queues count as various ASICs have different counts. Signed-off-by:
-
Jiri Pirko authored
Use mlxsw_pci_sdq_count helper instead. Signed-off-by:
-
Jiri Pirko authored
The count of CQs can be different for various ASICs, so just define maximal value and check for that. Signed-off-by:
-
Jiri Pirko authored
Signed-off-by:
-
Ido Schimmel authored
With respect to a firmware change, the Switch Multicast ID (SMID) register is no longer needed, so the related configuration code can be removed. Signed-off-by:
-
- 15 Oct, 2015 3 commits
-
-
Lendacky, Thomas authored
A previous patch switched from using the system workqueue to the device workqueue for various operations. During a device restart the device workqueue is flushed so the restart cannot use this workqueue or else a deadlock results. Move the device restart back to using the system workqueue. Signed-off-by:
Tom Lendacky <thomas.lendacky@amd.com> Signed-off-by:
-
David S. Miller authored
Jiri Pirko says: ==================== switchdev: change locking This is something which I'm currently struggling with. Callers of attr_set and obj_add/del often hold not only RTNL, but also spinlock (bridge). So in that case, the driver implementing the op cannot sleep. The way rocker is dealing with this now is just to invoke driver operation and go out, without any checking or reporting of the operation status. Since it would be nice to at least put a warning in case the operation fails, it makes sense to do this in delayed work directly in switchdev core instead of implementing this in separate drivers. And that is what this patchset is introducing. So from now on, the locking of switchdev mod ops is consistent. Caller either holds rtnl mutex or in case it does not, caller sets defer flag, telling switchdev core to process the op later, in deferred queue. Function to force to process switchdev deferred ops can be called by op caller in appropriate location, for example after it releases spin lock, to force switchdev core to process pending ops. v1->v2: - rebased on current net-next head (including Scott's ageing patchset) v2->v3: - fixed comment s/of/or/ typo suggested by Nik v3->v4: - the actual patchset is sent instead of different branch I send in v3 :/ v4->v5: - added patch to "const" attr param - reworked deferred ops infrastructure (mainly patch number 1 and internal users (patch 3 and 5)) - resolves the issue pointed out by John ==================== Signed-off-by: -
Jiri Pirko authored
netdev_for_each_lower_dev has to be called with rtnl mutex held. So better enforce it in switchdev functions. Signed-off-by:
-
