- 07 Oct, 2017 18 commits
-
-
David S. Miller authored
Wei Wang says: ==================== ipv6: replace rwlock with rcu and spinlock in fib6 table Currently, fib6 table is protected by rwlock. During route lookup, reader lock is taken and during route insertion, deletion or modification, writer lock is taken. This is a very inefficient implementation because the fastpath always has to do the operation to grab the reader lock. According to my latest syn flood test on an iota ivybridage machine with 2 10G mlx nics bonded together, each with 8 rx queues on 2 NUMA nodes, and with the upstream net-next kernel: ipv4 stack can handle around 4.2Mpps ipv6 stack can handle around 1.3Mpps In order to close the gap of the performance number between ipv4 and ipv6 stack, this patch series tries to get rid of the usage of the rwlock and replace it with rcu and spinlock protection. This will greatly speed up the fastpath performance as it only needs to hold rcu which is much less expensive than grabbing the reader lock. It also makes ipv6 fib implementation more consistent with ipv4. In order to be able to replace the current rwlock with rcu and spinlock, some preparation work is needed: Patch 1-8 introduces a per-route hash table (protected by rcu and a different spinlock) to store all cached routes created by pmtu and ip redirect under its main route. This makes the main fib6 tree only contain static routes. Patch 9-14 prepares all the reader path to be ready to tolerate concurrent writer. Patch 15 finally does the rwlock to rcu and spinlock conversion. Patch 16 takes care of rt6_stats. After this patch series, in the same syn flood test, ipv6 stack can now handle around 3.5Mpps compared to previous 1.3Mpps in my test setup. After this patch series, there are still some improvements that should be done in ipv6 stack: 1. During route lookup, dst_use() is called everytime on the selected route to update dst->__use and dst->lastuse. This dirties the cacheline and causes extra cacheline miss and should be avoided. 2. when no route is found in the current table, net->ip6.ipv6_null_entry is used and refcnt is taken on it. As there is no pcpu cache for this specific route, frequent change on the refcnt for this route causes quite some cacheline misses. And to make things worse, if CONFIG_IPV6_MULTIPLE_TABLES is defined, output path route lookup always starts with local table first and guarantees to hit net->ipv6.ip6_null_entry before continuing to do lookup in the main table. These operations on net->ipv6.ip6_null_entry could potentially be avoided. 3. ipv6 input path route lookup grabs refcnt on dst. This is different from ipv4. We could potentially change this behavior to let ipv6 input path route lookup not to grab refcnt on dst. However, it does not give us much performance boost as we currently have pcpu route cache for input path as well in ipv6. But this work probably is still worth doing to unify ipv6 and ipv4 route lookup behavior. The above issues will be addressed separately after this patch series has been accepted. This is a joint work with Martin KaFai Lau and Eric Dumazet. And many many thanks to them for their inspiring ideas and big big code review efforts. ==================== Reviewed-by:
Eric Dumazet <edumazet@google.com> Reviewed-by:
YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org> Signed-off-by:
David S. Miller <davem@davemloft.net>
-
Wei Wang authored
Currently, most of the rt6_stats are not hooked up correctly. As the last part of this patch series, hook up all existing rt6_stats and add one new stat fib_rt_uncache to indicate the number of routes in the uncached list. For details of the stats, please refer to the comments added in include/net/ip6_fib.h. Note: fib_rt_alloc and fib_rt_uncache are not guaranteed to be modified under a lock. So atomic_t is used for them. Signed-off-by:
Wei Wang <weiwan@google.com> Signed-off-by:
Martin KaFai Lau <kafai@fb.com> Signed-off-by:
-
Wei Wang authored
With all the preparation work before, we are now ready to replace rwlock with rcu and spinlock in fib6_table. That means now all fib6_node in fib6_table are protected by rcu. And when freeing fib6_node, call_rcu() is used to wait for the rcu grace period before releasing the memory. When accessing fib6_node, corresponding rcu APIs need to be used. And all previous sessions protected by the write lock will now be protected by the spin lock per table. All previous sessions protected by read lock will now be protected by rcu_read_lock(). A couple of things to note here: 1. As part of the work of replacing rwlock with rcu, the linked list of fn->leaf now has to be rcu protected as well. So both fn->leaf and rt->dst.rt6_next are now __rcu tagged and corresponding rcu APIs are used when manipulating them. 2. For fn->rr_ptr, first of all, it also needs to be rcu protected now and is tagged with __rcu and rcu APIs are used in corresponding places. Secondly, fn->rr_ptr is changed in rt6_select() which is a reader thread. This makes the issue a bit complicated. We think a valid solution for it is to let rt6_select() grab the tb6_lock if it decides to change it. As it is not in the normal operation and only happens when there is no valid neighbor cache for the route, we think the performance impact should be low. 3. fib6_walk_continue() has to be called with tb6_lock held even in the route dumping related functions, e.g. inet6_dump_fib(), fib6_tables_dump() and ipv6_route_seq_ops. It is because fib6_walk_continue() makes modifications to the walker structure, and so are fib6_repair_tree() and fib6_del_route(). In order to do proper syncing between them, we need to let fib6_walk_continue() hold the lock. We may be able to do further improvement on the way we do the tree walk to get rid of the need for holding the spin lock. But not for now. 4. When fib6_del_route() removes a route from the tree, we no longer mark rt->dst.rt6_next to NULL to make simultaneous reader be able to further traverse the list with rcu. However, rt->dst.rt6_next is only valid within this same rcu period. No one should access it later. 5. All the operation of atomic_inc(rt->rt6i_ref) is changed to be performed before we publish this route (either by linking it to fn->leaf or insert it in the list pointed by fn->leaf) just to be safe because as soon as we publish the route, some read thread will be able to access it. Signed-off-by:
-
Wei Wang authored
After rwlock is replaced with rcu and spinlock, fib6_lookup() could potentially return an intermediate node if other thread is doing fib6_del() on a route which is the only route on the node so that fib6_repair_tree() will be called on this node and potentially assigns fn->leaf to the its child's fn->leaf. In order to detect this situation in rt6_select(), we have to check if fn->fn_bit is consistent with the key length stored in the route. And depending on if the fn is in the subtree or not, the key is either rt->rt6i_dst or rt->rt6i_src. If any inconsistency is found, that means the node no longer holds valid routes in it. So net->ipv6.ip6_null_entry is returned. Signed-off-by:
-
Wei Wang authored
If rwlock is replaced with rcu and spinlock, it is possible that the reader thread will see fn->leaf as NULL in the following scenarios: 1. fib6_add() is in progress and we have already inserted a new node but not yet inserted the route. 2. fib6_del_route() is in progress and we have already set fn->leaf to NULL but not yet freed the node because of rcu grace period. This patch makes sure all the reader threads check fn->leaf first before using it. And together with later patch to grab rcu_read_lock() and rcu_dereference() fn->leaf, it makes sure reader threads are safe when accessing fn->leaf. Signed-off-by:
-
Wei Wang authored
fib6_add() logic currently calls fib6_add_1() to figure out what node should be used for the newly added route and then call fib6_add_rt2node() to insert the route to the node. And during the call of fib6_add_1(), fn_sernum is updated for all nodes that share the same prefix as the new route. This does not have issue in the current code because reader thread will not be able to access the tree while writer thread is inserting new route to it. However, it is not the case once we transition to use RCU. Reader thread could potentially see the new fn_sernum before the new route is inserted. As a result, reader thread's route lookup will return a stale route with the new fn_sernum. In order to solve this issue, we remove all the update of fn_sernum in fib6_add_1(), and instead, introduce a new function that updates fn_sernum for all related nodes and call this functions once the route is successfully inserted to the tree. Also, smp_wmb() is used after a route is successfully inserted into the fib tree and right before the updated of fn->sernum. And smp_rmb() is used right after fn->sernum is accessed in rt6_get_cookie_safe(). This is to guarantee that when the reader thread sees the new fn->sernum, the new route is already inserted in the tree in memory. Signed-off-by:
-
Wei Wang authored
With rwlock, it is safe to call dst_hold() in the read thread because read thread is guaranteed to be separated from write thread. However, after we replace rwlock with rcu, it is no longer safe to use dst_hold(). A dst might already have been deleted but is waiting for the rcu grace period to pass before freeing the memory when a read thread is trying to do dst_hold(). This could potentially cause double free issue. So this commit replaces all dst_hold() with dst_hold_safe() in all read thread to avoid this double free issue. And in order to make the code more compact, a new function ip6_hold_safe() is introduced. It calls dst_hold_safe() first, and if that fails, it will either fall back to hold and return net->ipv6.ip6_null_entry or set rt to NULL according to the caller's need. Signed-off-by:
-
Wei Wang authored
After rwlock is replaced with rcu and spinlock, route lookup can happen simultanously with route deletion. This patch removes the call to free_percpu(rt->rt6i_pcpu) from rt6_release() to avoid the race condition between rt6_release() and rt6_get_pcpu_route(). And as free_percpu(rt->rt6i_pcpu) is already called in ip6_dst_destroy() after the rcu grace period, it is safe to do this change. Signed-off-by:
-
Wei Wang authored
After rwlock is replaced with rcu and spinlock, ip6_pol_route() will be called with only rcu held. That means rt6 route deletion could happen simultaneously with rt6_make_pcpu_rt(). This could potentially cause memory leak if rt6_release() is called right before rt6_make_pcpu_rt() on the same route. This patch grabs rt->rt6i_ref safely before calling rt6_make_pcpu_rt() to make sure rt6_release() will not get triggered while rt6_make_pcpu_rt() is in progress. And rt6_release() is called after rt6_make_pcpu_rt() is finished. Note: As we are incrementing rt->rt6i_ref in ip6_pol_route(), there is a very slim chance that fib6_purge_rt() will be triggered unnecessarily when deleting a route if ip6_pol_route() running on another thread picks this route as well and tries to make pcpu cache for it. Signed-off-by:
-
Wei Wang authored
This commit makes use of the exception hash table implementation to store dst caches created by pmtu discovery and ip redirect into the hash table under the rt_info and no longer inserts these routes into fib6 tree. This makes the fib6 tree only contain static configured routes and could now be protected by rcu instead of a rw lock. With this change, in the route lookup related functions, after finding the rt6_info with the longest prefix, we also need to search for the exception table before doing backtracking. In the route delete function, if the route being deleted is not a dst cache, deletion of this route also need to flush the whole hash table under it. If it is a dst cache, then only delete the cached dst in the hash table. Note: for fib6_walk_continue() function, w->root now is always pointing to a root node considering that fib6_prune_clones() is removed from the code. So we add a WARN_ON() msg to make sure w->root always points to a root node and also removed the update of w->root in fib6_repair_tree(). This is a prerequisite for later patch because we don't need to make w->root as rcu protected when replacing rwlock with RCU. Also, we remove all prune related variables as it is no longer used. Signed-off-by:
-
Wei Wang authored
fib6_locate() is used to find the fib6_node according to the passed in prefix address key. It currently tries to find the fib6_node with the exact match of the passed in key. However, when we move cached routes into the exception table, fib6_locate() will fail to find the fib6_node for it as the cached routes will be stored in the exception table under the fib6_node with the longest prefix match of the cache's dst addr key. This commit adds a new parameter to let the caller specify if it needs exact match or longest prefix match. Right now, all callers still does exact match when calling fib6_locate(). It will be changed in later commit where exception table is hooked up to store cached routes. Signed-off-by:
-
Wei Wang authored
If all dst cache entries are stored in the exception table under the main route, we have to go through them during fib6_age() when doing garbage collecting. Introduce a new function rt6_age_exception() which goes through all dst entries in the exception table and remove those entries that are expired. This function is called in fib6_age() so that all dst caches are also garbage collected. Signed-off-by:
-
Wei Wang authored
If we move all cached dst into the exception table under the main route, current rt6_clean_tohost() will no longer be able to access them. This commit makes fib6_clean_tohost() to also go through all cached routes in exception table and removes cached gateway routes to the passed in gateway. This is a preparation in order to move all cached routes into the exception table. Signed-off-by:
-
Wei Wang authored
If we move all cached dst into the exception table under the main route, current rt6_mtu_change() will no longer be able to access them. This commit makes rt6_mtu_change_route() function to also go through all cached routes in the exception table under the main route and do proper updates on the mtu. This is a preparation in order to move all cached routes into the exception table. Signed-off-by:
-
Wei Wang authored
After we move cached dst entries into the exception table under its parent route, current fib6_remove_prefsrc() no longer can access them. This commit makes fib6_remove_prefsrc() also go through all routes in the exception table to remove the pref src. This is a preparation patch in order to move all cached dst into the exception table. Signed-off-by:
-
Wei Wang authored
Add a hash table into struct rt6_info in order to store dst caches created by pmtu discovery and ip redirect in ipv6 routing code. APIs to add dst cache, delete dst cache, find dst cache and update dst cache in the hash table are implemented and will be used in later commits. This is a preparation work to move all cache routes into the exception table instead of getting inserted into the fib6 tree. Signed-off-by:
-
Wei Wang authored
This function takes a route as input and tries to update the sernum in the fib6_node this route is associated with. It will be used in later commit when adding a cached route into the exception table under that route. Signed-off-by:
-
Jonathan Toppins authored
Instead of zeroing out bnxt_tc.c with a #ifdef foo, instead don't compile the file when the option is not enabled. Now make and the preprocessor do not have to waste time compiling a no-op. Signed-off-by:
Jonathan Toppins <jtoppins@redhat.com> Acked-by:
Michael Chan <michael.chan@broadcom.com> Signed-off-by:
-
- 06 Oct, 2017 22 commits
-
-
David S. Miller authored
Eric Dumazet says: ==================== tcp: implement rb-tree based retransmit queue This patch series implement RB-tree based retransmit queue for TCP, to better match modern BDP. Tested: On receiver : netem on ingress : delay 150ms 200us loss 1 GRO disabled to force stress and SACK storms. for f in `seq 1 10` do ./netperf -H lpaa6 -l30 -- -K bbr -o THROUGHPUT|tail -1 done | awk '{print $0} {sum += $0} END {printf "%7u\n",sum}' Before patch : 323.87 351.48 339.59 338.62 306.72 204.07 304.93 291.88 202.47 176.88 -> 2840 After patch: 1700.83 2207.98 2070.17 1544.26 2114.76 2124.89 1693.14 1080.91 2216.82 1299.94 -> 18053 Average of 1805 Mbits istead of 284 Mbits. ==================== Signed-off-by: -
Eric Dumazet authored
Using a linear list to store all skbs in write queue has been okay for quite a while : O(N) is not too bad when N < 500. Things get messy when N is the order of 100,000 : Modern TCP stacks want 10Gbit+ of throughput even with 200 ms RTT flows. 40 ns per cache line miss means a full scan can use 4 ms, blowing away CPU caches. SACK processing often can use various hints to avoid parsing whole retransmit queue. But with high packet losses and/or high reordering, hints no longer work. Sender has to process thousands of unfriendly SACK, accumulating a huge socket backlog, burning a cpu and massively dropping packets. Using an rb-tree for retransmit queue has been avoided for years because it added complexity and overhead, but now is the time to be more resistant and say no to quadratic behavior. 1) RTX queue is no longer part of the write queue : already sent skbs are stored in one rb-tree. 2) Since reaching the head of write queue no longer needs sk->sk_send_head, we added an union of sk_send_head and tcp_rtx_queue Tested: On receiver : netem on ingress : delay 150ms 200us loss 1 GRO disabled to force stress and SACK storms. for f in `seq 1 10` do ./netperf -H lpaa6 -l30 -- -K bbr -o THROUGHPUT|tail -1 done | awk '{print $0} {sum += $0} END {printf "%7u\n",sum}' Before patch : 323.87 351.48 339.59 338.62 306.72 204.07 304.93 291.88 202.47 176.88 2840 After patch: 1700.83 2207.98 2070.17 1544.26 2114.76 2124.89 1693.14 1080.91 2216.82 1299.94 18053 Signed-off-by: -
Eric Dumazet authored
No need to recompute previous skb, as it will be a bit more expensive when rtx queue is converted to RB tree. Signed-off-by:
-
Eric Dumazet authored
With upcoming rb-tree implementation, the checks will trigger more often, and this is expected. Signed-off-by:
-
Eric Dumazet authored
It will be a bit more expensive to get the head of rtx queue once rtx queue is converted to an rb-tree. We can avoid this extra cost in case tp->lost_skb_hint is set. Signed-off-by:
-
Eric Dumazet authored
tcp_write_queue_tail() call can be factorized. Signed-off-by:
-
Eric Dumazet authored
Since the upcoming rtx rbtree will add some extra code, it is time to not inline this fat function anymore. Signed-off-by:
-
Eric Dumazet authored
Geeralize private netem_rb_to_skb() TCP rtx queue will soon be converted to rb-tree, so we will need skb_rbtree_walk() helpers. Signed-off-by:
-
David S. Miller authored
Jeff Kirsher says: ==================== 40GbE Intel Wired LAN Driver Updates 2017-10-06 This series contains updates to i40e and i40evf only. Rami fixes a typo in the code comments. Mitch adds an ethtool private flag to control source pruning to resolve an issue where our default behavior is to enable source pruning which breaks ARP monitoring in channel bonding. Fixes a couple of register definitions, which were incorrect. Jake fixes an issue with multiple logical CPUs per core (simultaneous multithreading - SMT) and how we set an affinity hint based on the v_idx of that q_vector, which is an incremental value and might lead to multiple offline CPUs being assigned to a q_vector. Instead, we should only assign hints for CPUs which are online, so look to use cpumask_local_spread(). Also fixed a VF VLAN tag stripping issue, where the flag created to change this feature was seen as unchangeable. Lastly, organized and re-numbered the feature flags. Alan re-enables PTP L4 for XL710 devices with firmware version 6.0 or greater, now that the previous bug in the older firmware is fixed. Implements the PCI error handlers for reset_prepare() and reset_done() to allow us to handle function level resets. Alice cleans up code that was added to the incorrect function during a merge. Filip adds a change to display an error message when a module is inserted that does not meet the thermal requirements, Talking Heads "Burning Down the House" comes to mind. Also fixed a flow director filter issue where a variable was not being cleared which stores the filter number to be removed from the list when the firmware refused to add the requested filter. ==================== Signed-off-by: -
git://git.open-mesh.org/linux-mergeDavid S. Miller authored
Simon Wunderlich says: ==================== This cleanup patchset includes the following patches: - bump version strings, by Simon Wunderlich - Cleanup patches to make checkpatch happy, by Sven Eckelmann (3 patches) ==================== Signed-off-by: -
Bjorn Helgaas authored
Use pci_ari_enabled() from the PCI core instead of the identical local copy bnx2x_ari_enabled(). No functional change intended. Signed-off-by:
Bjorn Helgaas <bhelgaas@google.com> Signed-off-by:
-
David S. Miller authored
Jesper Dangaard Brouer says: ==================== Improve xdp_monitor samples/bpf Here are some improvements to the xdp_monitor tool currently located under samples/bpf/. Once the tools library libbpf become more feature complete, xdp_monitor should be converted to use it, and be moved into tools/bpf/xdp/ or tools/xdp/. ==================== Signed-off-by: -
Jesper Dangaard Brouer authored
Other concurrent running programs, like perf or the XDP program what needed to be monitored, might take up part of the max locked memory limit. Thus, the xdp_monitor tool have to set the RLIMIT_MEMLOCK to RLIM_INFINITY, as it cannot determine a more sane limit. Using the man exit(3) specified EXIT_FAILURE return exit code, and correct other users too. Signed-off-by:
Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by:
-
Jesper Dangaard Brouer authored
Also monitor the tracepoint xdp_exception. This tracepoint is usually invoked by the drivers. Programs themselves can activate this by returning XDP_ABORTED, which will drop the packet but also trigger the tracepoint. This is useful for distinguishing intentional (XDP_DROP) vs. ebpf-program error cases that cased a drop (XDP_ABORTED). Drivers also use this tracepoint for reporting on XDP actions that are unknown to the specific driver. This can help the user to detect if a driver e.g. doesn't implement XDP_REDIRECT yet. Signed-off-by:
-
Jesper Dangaard Brouer authored
The first 8 bytes of the tracepoint context struct are not accessible by the bpf code. This is a choice that dates back to the original inclusion of this code. See explaination in: commit 98b5c2c6 ("perf, bpf: allow bpf programs attach to tracepoints") Signed-off-by:
Alexei Starovoitov <ast@kernel.org> Signed-off-by:
-
David S. Miller authored
Simon Horman says: ==================== nfp: extend match and action for flower offload Pieter says: This series extends flower offload match and action capabilities. It specifically adds offload capabilities for matching on MPLS, TTL, TOS and flow label. Furthermore offload capabilities for action have been expanded to include set ethernet, ipv4, ipv6, tcp and udp headers. ==================== Signed-off-by: -
Pieter Jansen van Vuuren authored
Previously we did not have offloading support for set TCP/UDP actions. This patch enables TC flower offload of set TCP/UDP sport and dport actions. Signed-off-by:
Pieter Jansen van Vuuren <pieter.jansenvanvuuren@netronome.com> Reviewed-by:
Jakub Kicinski <jakub.kicinski@netronome.com> Signed-off-by:
Simon Horman <simon.horman@netronome.com> Signed-off-by:
-
Pieter Jansen van Vuuren authored
Previously we did not have offloading support for set IPv6 actions. This patch enables TC flower offload of set IPv6 src and dst address actions. Signed-off-by:
-
Pieter Jansen van Vuuren authored
Previously we did not have offloading support for set IPv4 actions. This patch enables TC flower offload of set IPv4 src and dst address actions. Signed-off-by:
-
Pieter Jansen van Vuuren authored
Previously we did not have offloading support for set ethernet actions. This patch enables TC flower offload of set ethernet actions. Signed-off-by:
-
Pieter Jansen van Vuuren authored
Previously matching on IPv6 ttl and tos fields were not offloaded. This patch enables offloading IPv6 ttl and tos as match fields. Signed-off-by:
-
Pieter Jansen van Vuuren authored
Previously matching on IPv4 ttl and tos fields were not offloaded. This patch enables offloading IPv4 ttl and tos as match fields. Signed-off-by:
-
