- 19 Sep, 2017 40 commits
-
-
David S. Miller authored
Eric Dumazet says: ==================== net: speedup netns create/delete time When rate of netns creation/deletion is high enough, we observe softlockups in cleanup_net() caused by huge list of netns and way too many rcu_barrier() calls. This patch series does some optimizations in kobject, and add batching to tunnels so that netns dismantles are less costly. IPv6 addrlabels also get a per netns list, and tcp_metrics also benefit from batch flushing. This gives me one order of magnitude gain. (~50 ms -> ~5 ms for one netns create/delete pair) Tested: for i in `seq 1 40` do (for j in `seq 1 100` ; do unshare -n /bin/true >/dev/null ; done) & done wait ; grep net_namespace /proc/slabinfo Before patch series : $ time ./add_del_unshare.sh net_namespace 116 258 5504 1 2 : tunables 8 4 0 : slabdata 116 258 0 real 3m24.910s user 0m0.747s sys 0m43.162s After : $ time ./add_del_unshare.sh net_namespace 135 291 5504 1 2 : tunables 8 4 0 : slabdata 135 291 0 real 0m22.117s user 0m0.728s sys 0m35.328s ==================== Signed-off-by:David S. Miller <davem@davemloft.net>
-
Eric Dumazet authored
Implement exit_batch() method to dismantle more devices per round. (rtnl_lock() ... unregister_netdevice_many() ... rtnl_unlock()) Tested: $ cat add_del_unshare.sh for i in `seq 1 40` do (for j in `seq 1 100` ; do unshare -n /bin/true >/dev/null ; done) & done wait ; grep net_namespace /proc/slabinfo Before patch : $ time ./add_del_unshare.sh net_namespace 126 282 5504 1 2 : tunables 8 4 0 : slabdata 126 282 0 real 1m38.965s user 0m0.688s sys 0m37.017s After patch: $ time ./add_del_unshare.sh net_namespace 135 291 5504 1 2 : tunables 8 4 0 : slabdata 135 291 0 real 0m22.117s user 0m0.728s sys 0m35.328s Signed-off-by:
Eric Dumazet <edumazet@google.com> Signed-off-by:
-
Eric Dumazet authored
Implement exit_batch() method to dismantle more devices per round. (rtnl_lock() ... unregister_netdevice_many() ... rtnl_unlock()) Tested: $ cat add_del_unshare.sh for i in `seq 1 40` do (for j in `seq 1 100` ; do unshare -n /bin/true >/dev/null ; done) & done wait ; grep net_namespace /proc/slabinfo Before patch : $ time ./add_del_unshare.sh net_namespace 110 267 5504 1 2 : tunables 8 4 0 : slabdata 110 267 0 real 3m25.292s user 0m0.644s sys 0m40.153s After patch: $ time ./add_del_unshare.sh net_namespace 126 282 5504 1 2 : tunables 8 4 0 : slabdata 126 282 0 real 1m38.965s user 0m0.688s sys 0m37.017s Signed-off-by:
-
Eric Dumazet authored
When dealing with a list of dismantling netns, we can scan tcp_metrics once, saving cpu cycles. Signed-off-by:
-
Eric Dumazet authored
Having a global list of labels do not scale to thousands of netns in the cloud era. This causes quadratic behavior on netns creation and deletion. This is time having a per netns list of ~10 labels. Tested: $ time perf record (for f in `seq 1 3000` ; do ip netns add tast$f; done) [ perf record: Woken up 1 times to write data ] [ perf record: Captured and wrote 3.637 MB perf.data (~158898 samples) ] real 0m20.837s # instead of 0m24.227s user 0m0.328s sys 0m20.338s # instead of 0m23.753s 16.17% ip [kernel.kallsyms] [k] netlink_broadcast_filtered 12.30% ip [kernel.kallsyms] [k] netlink_has_listeners 6.76% ip [kernel.kallsyms] [k] _raw_spin_lock_irqsave 5.78% ip [kernel.kallsyms] [k] memset_erms 5.77% ip [kernel.kallsyms] [k] kobject_uevent_env 5.18% ip [kernel.kallsyms] [k] refcount_sub_and_test 4.96% ip [kernel.kallsyms] [k] _raw_read_lock 3.82% ip [kernel.kallsyms] [k] refcount_inc_not_zero 3.33% ip [kernel.kallsyms] [k] _raw_spin_unlock_irqrestore 2.11% ip [kernel.kallsyms] [k] unmap_page_range 1.77% ip [kernel.kallsyms] [k] __wake_up 1.69% ip [kernel.kallsyms] [k] strlen 1.17% ip [kernel.kallsyms] [k] __wake_up_common 1.09% ip [kernel.kallsyms] [k] insert_header 1.04% ip [kernel.kallsyms] [k] page_remove_rmap 1.01% ip [kernel.kallsyms] [k] consume_skb 0.98% ip [kernel.kallsyms] [k] netlink_trim 0.51% ip [kernel.kallsyms] [k] kernfs_link_sibling 0.51% ip [kernel.kallsyms] [k] filemap_map_pages 0.46% ip [kernel.kallsyms] [k] memcpy_erms Signed-off-by: -
Eric Dumazet authored
We can build one skb and let it be cloned in netlink. This is much faster, and use less memory (all clones will share the same skb->head) Tested: time perf record (for f in `seq 1 3000` ; do ip netns add tast$f; done) [ perf record: Woken up 1 times to write data ] [ perf record: Captured and wrote 4.110 MB perf.data (~179584 samples) ] real 0m24.227s # instead of 0m52.554s user 0m0.329s sys 0m23.753s # instead of 0m51.375s 14.77% ip [kernel.kallsyms] [k] __ip6addrlbl_add 14.56% ip [kernel.kallsyms] [k] netlink_broadcast_filtered 11.65% ip [kernel.kallsyms] [k] netlink_has_listeners 6.19% ip [kernel.kallsyms] [k] _raw_spin_lock_irqsave 5.66% ip [kernel.kallsyms] [k] kobject_uevent_env 4.97% ip [kernel.kallsyms] [k] memset_erms 4.67% ip [kernel.kallsyms] [k] refcount_sub_and_test 4.41% ip [kernel.kallsyms] [k] _raw_read_lock 3.59% ip [kernel.kallsyms] [k] refcount_inc_not_zero 3.13% ip [kernel.kallsyms] [k] _raw_spin_unlock_irqrestore 1.55% ip [kernel.kallsyms] [k] __wake_up 1.20% ip [kernel.kallsyms] [k] strlen 1.03% ip [kernel.kallsyms] [k] __wake_up_common 0.93% ip [kernel.kallsyms] [k] consume_skb 0.92% ip [kernel.kallsyms] [k] netlink_trim 0.87% ip [kernel.kallsyms] [k] insert_header 0.63% ip [kernel.kallsyms] [k] unmap_page_range Signed-off-by: -
Eric Dumazet authored
No need to iterate over strings, just copy in one efficient memcpy() call. Tested: time perf record "(for f in `seq 1 3000` ; do ip netns add tast$f; done)" [ perf record: Woken up 10 times to write data ] [ perf record: Captured and wrote 8.224 MB perf.data (~359301 samples) ] real 0m52.554s # instead of 1m7.492s user 0m0.309s sys 0m51.375s # instead of 1m6.875s 9.88% ip [kernel.kallsyms] [k] netlink_broadcast_filtered 8.86% ip [kernel.kallsyms] [k] string 7.37% ip [kernel.kallsyms] [k] __ip6addrlbl_add 5.68% ip [kernel.kallsyms] [k] netlink_has_listeners 5.52% ip [kernel.kallsyms] [k] memcpy_erms 4.76% ip [kernel.kallsyms] [k] __alloc_skb 4.54% ip [kernel.kallsyms] [k] vsnprintf 3.94% ip [kernel.kallsyms] [k] format_decode 3.80% ip [kernel.kallsyms] [k] kmem_cache_alloc_node_trace 3.71% ip [kernel.kallsyms] [k] kmem_cache_alloc_node 3.66% ip [kernel.kallsyms] [k] kobject_uevent_env 3.38% ip [kernel.kallsyms] [k] strlen 2.65% ip [kernel.kallsyms] [k] _raw_spin_lock_irqsave 2.20% ip [kernel.kallsyms] [k] kfree 2.09% ip [kernel.kallsyms] [k] memset_erms 2.07% ip [kernel.kallsyms] [k] ___cache_free 1.95% ip [kernel.kallsyms] [k] kmem_cache_free 1.91% ip [kernel.kallsyms] [k] _raw_read_lock 1.45% ip [kernel.kallsyms] [k] ksize 1.25% ip [kernel.kallsyms] [k] _raw_spin_unlock_irqrestore 1.00% ip [kernel.kallsyms] [k] widen_string Signed-off-by: -
Eric Dumazet authored
This removes some #ifdef pollution and will ease follow up patches. Signed-off-by:
-
Cong Wang authored
gen estimator has been rewritten in commit 1c0d32fd ("net_sched: gen_estimator: complete rewrite of rate estimators"), the caller no longer needs to wait for a grace period. So this patch gets rid of it. Cc: Jamal Hadi Salim <jhs@mojatatu.com> Cc: Eric Dumazet <edumazet@google.com> Signed-off-by:
Cong Wang <xiyou.wangcong@gmail.com> Acked-by:
-
Jim Hanko authored
If the hash to port mapping table does not have a valid port (i.e. when a port goes down), fall back to the simple hashing mechanism to avoid dropping packets. Signed-off-by:
Jim Hanko <hanko@drivescale.com> Acked-by:
Jiri Pirko <jiri@mellanox.com> Signed-off-by:
-
David S. Miller authored
Florian Westphal says: ==================== test_rhashtable: don't allocate huge static array Add a test case for the rhlist interface. While at it, cleanup current rhashtable test a bit and add a check for max_size support. No changes since v1, except in last patch. kbuild robot complained about large onstack allocation caused by struct rhltable when lockdep is enabled. ==================== Signed-off-by: -
Florian Westphal authored
also test rhltable. rhltable remove operations are slow as deletions require a list walk, thus test with 1/16th of the given entry count number to get a run duration similar to rhashtable one. Signed-off-by:
Florian Westphal <fw@strlen.de> Signed-off-by:
-
Florian Westphal authored
add a test that tries to insert more than max_size elements. Signed-off-by:
-
Florian Westphal authored
pass the entries to test as an argument instead. Followup patch will add an rhlist test case; rhlist delete opererations are slow so we need to use a smaller number to test it. Signed-off-by:
-
Florian Westphal authored
Signed-off-by:
-
David S. Miller authored
Florian Fainelli says: ==================== net: dsa: b53/bcm_sf2 cleanups This patch series is a first pass set of clean-ups to reduce the number of LOCs between b53 and bcm_sf2 and sharing as many functions as possible. There is a number of additional cleanups queued up locally that require more thorough testing. Changes in v3: - remove one extra argument for the b53_build_io_op macro (David Laight) - added additional Reviewed-by tags from Vivien Changes in v2: - added Reviewed-by tags from Vivien - added a missing EXPORT_SYMBOL() in patch 8 - fixed a typo in patch 5 ==================== Signed-off-by: -
Florian Fainelli authored
Export b53_{enable,disable}_port and use these two functions in bcm_sf2_port_setup and bcm_sf2_port_disable. The generic functions cannot be used without wrapping because we need to manage additional switch integration details (PHY, Broadcom tag etc.). Reviewed-by:Vivien Didelot <vivien.didelot@savoirfairelinux.com> Signed-off-by:
Florian Fainelli <f.fainelli@gmail.com> Signed-off-by:
-
Florian Fainelli authored
The magic number 8 in 3 locations in bcm_sf2_cfp.c actually designates the number of switch port egress queues, so use that define instead of open-coding it. Reviewed-by:
-
Florian Fainelli authored
bcm_sf2 and b53 do exactly the same thing, so share that piece. Reviewed-by:
-
Florian Fainelli authored
Add support for enabling and disabling EEE, as well as re-negotiating it in .adjust_link() and in .port_enable(). Reviewed-by:
-
Florian Fainelli authored
Move the bcm_sf2 EEE-related functions to the b53 driver because this is shared code amongst Gigabit capable switch, only 5325 and 5365 are too old to support that. Reviewed-by:
-
Florian Fainelli authored
In preparation for migrating the EEE code from bcm_sf2 to b53, define the full EEE register page and offsets within that page. Reviewed-by:
-
Florian Fainelli authored
The code to enable Broadcom tags/headers is largely switch independent, and in preparation for enabling it for multiple devices with b53, move the code we have in bcm_sf2.c to b53_common.c Reviewed-by:
-
Florian Fainelli authored
Instead of repeating the same pattern: acquire mutex, read/write, release mutex, define a macro: b53_build_op() which takes the type (read|write), I/O size, and value (scalar or pointer). This helps with fixing bugs that could exist (e.g: missing barrier, lock etc.). Reviewed-by:
-
Florian Fainelli authored
There is no need to configure the enabled ports once in bcm_sf2_sw_setup() and then a second time around when dsa_switch_ops::port_enable is called, just do it when port_enable is called which is better in terms of power consumption and correctness. Reviewed-by:
-
Florian Fainelli authored
There is no need to configure the enabled ports once in b53_setup() and then a second time around when dsa_switch_ops::port_enable is called, just do it when port_enable is called which is better in terms of power consumption and correctness. Reviewed-by:
-
Florian Fainelli authored
In preparation for future changes allowing the configuring of multiple CPU ports, make b53_enable_cpu_port() take a port argument. Reviewed-by:
-
Florian Fainelli authored
This is not used anywhere, so remove it. Reviewed-by:
-
David S. Miller authored
Vivien Didelot says: ==================== net: dsa: move master ethtool code The DSA core overrides the master device's ethtool_ops structure so that it can inject statistics and such of its dedicated switch CPU port. This ethtool code is currently called on unnecessary conditions or before the master interface and its switch CPU port get wired up. This patchset fixes this. Similarly to slave.c where the DSA slave net_device is the entry point of the dsa_slave_* functions, this patchset also isolates the master's ethtool code in a new master.c file, where the DSA master net_device is the entry point of the dsa_master_* functions. This is a first step towards better control of the master device and support for multiple CPU ports. ==================== Tested-by:
-
Vivien Didelot authored
DSA overrides the master device ethtool ops, so that it can inject stats from its dedicated switch CPU port as well. The related code is currently split in dsa.c and slave.c, but it only scopes the master net device. Move it to a new master.c DSA core file. This file will be later extented with master net device specific code. Signed-off-by:
-
Vivien Didelot authored
DSA overrides the master's ethtool ops so that we can inject its CPU port's statistics. Because of that, we need to setup the ethtool ops after the master's dsa_ptr pointer has been assigned, not before. This patch setups the ethtool ops after dsa_ptr is assigned, and restores them before it gets cleared. Signed-off-by:
-
Vivien Didelot authored
When a DSA switch tree is meant to be applied, it already has a CPU port. Thus remove the condition of dst->cpu_dp. Moreover, the next lines access dst->cpu_dp unconditionally. Signed-off-by:
-
Vivien Didelot authored
There is no need to store a copy of the master ethtool ops, storing the original pointer in DSA and the new one in the master netdev itself is enough. In the meantime, set orig_ethtool_ops to NULL when restoring the master ethtool ops and check the presence of the master original ethtool ops as well as its needed functions before calling them. Signed-off-by:
-
Florian Westphal authored
Signed-off-by:
-
Eric Dumazet authored
skb->rbnode shares space with skb->next, skb->prev and skb->tstamp Current uses (TCP receive ofo queue and netem) need to save/restore tstamp, while skb->dev is either NULL (TCP) or a constant for a given queue (netem). Since we plan using an RB tree for TCP retransmit queue to speedup SACK processing with large BDP, this patch exchanges skb->dev and skb->tstamp. This saves some overhead in both TCP and netem. v2: removes the swtstamp field from struct tcp_skb_cb Signed-off-by:
Soheil Hassas Yeganeh <soheil@google.com> Signed-off-by:
-
David S. Miller authored
Jiri Pirko says: ==================== mlxsw: Prepare for multicast router offload Yotam says: This patch-set makes various preparations needed for the multicast router offloading, which include: - Add the needed registers. - Add needed ACL actions. - Add new traps and trap groups. - Exporting needed private structs and enums. ==================== Signed-off-by: -
Yotam Gigi authored
Add three new traps needed for multicast routing: - PIM: Trap for PIM protocol control packets. - RPF: Trap for packets that fail the RPF check on a specific hardware route entry. - MULTICAST: Generic trap for multicast. It is used for routes that trap the packets to the CPU. The RPF and MULTICAST traps have rate limiters as these traps may have line-rate of packets trapped. The PIM trap has a rate limiter similarly to other L3 control protocols. The rate limiters are implemented by adding three new trap groups for the newly introduced traps. Signed-off-by:
Yotam Gigi <yotamg@mellanox.com> Reviewed-by:
Ido Schimmel <idosch@mellanox.com> Signed-off-by:
-
Yotam Gigi authored
The mlxsw_sp_rif struct, defined as private struct in spectrum_router.c will be used in the multicast router source file. Due to the fact that the dev field will be needed by the multicast router logic, add an access function to it. Signed-off-by:
-
Yotam Gigi authored
Turn on two bits on the Spectrum RIF configuration: - IPv4 multicast: when a multicast packet arrives on a RIF, send it to go through multicast routes lookup. - IPv4 multicast forwarding enable: when multicast packet arrives on a RIF, allow it to be forwarded by multicast routes. If this bit is not set, multicast packets will go through multicast routing lookup but will be dropped at the egress of the ports. Signed-off-by:
-
Yotam Gigi authored
The RRCR register is used for copying and moving TCAM multicast routes from different offsets. It will be used to allow routes relocation for parman ops as part of the multicast router offloading logic. Signed-off-by:
-
