- 12 Jun, 2019 19 commits
-
-
Eric Dumazet authored
Adding delays to TCP flows is crucial for studying behavior of TCP stacks, including congestion control modules. Linux offers netem module, but it has unpractical constraints : - Need root access to change qdisc - Hard to setup on egress if combined with non trivial qdisc like FQ - Single delay for all flows. EDT (Earliest Departure Time) adoption in TCP stack allows us to enable a per socket delay at a very small cost. Networking tools can now establish thousands of flows, each of them with a different delay, simulating real world conditions. This requires FQ packet scheduler or a EDT-enabled NIC. This patchs adds TCP_TX_DELAY socket option, to set a delay in usec units. unsigned int tx_delay = 10000; /* 10 msec */ setsockopt(fd, SOL_TCP, TCP_TX_DELAY, &tx_delay, sizeof(tx_delay)); Note that FQ packet scheduler limits might need some tweaking : man tc-fq PARAMETERS limit Hard limit on the real queue size. When this limit is reached, new packets are dropped. If the value is lowered, packets are dropped so that the new limit is met. Default is 10000 packets. flow_limit Hard limit on the maximum number of packets queued per flow. Default value is 100. Use of TCP_TX_DELAY option will increase number of skbs in FQ qdisc, so packets would be dropped if any of the previous limit is hit. Use of a jump label makes this support runtime-free, for hosts never using the option. Also note that TSQ (TCP Small Queues) limits are slightly changed with this patch : we need to account that skbs artificially delayed wont stop us providind more skbs to feed the pipe (netem uses skb_orphan_partial() for this purpose, but FQ can not use this trick) Because of that, using big delays might very well trigger old bugs in TSO auto defer logic and/or sndbuf limited detection. Signed-off-by:Eric Dumazet <edumazet@google.com> Signed-off-by:
David S. Miller <davem@davemloft.net>
-
David S. Miller authored
Sameeh Jubran says: ==================== Support for dynamic queue size changes This patchset introduces the following: * add new admin command for supporting different queue size for Tx/Rx * add support for Tx/Rx queues size modification through ethtool * allow queues allocation backoff when low on memory * update driver version Difference from v2: * Dropped superfluous range checks which are already done in ethtool. [patch 5/7] * Dropped inline keyword from function. [patch 4/7] * Added a new patch which drops inline keyword all *.c files. [patch 6/7] Difference from v1: * Changed ena_update_queue_sizes() signature to use u32 instead of int type for the size arguments. [patch 5/7] ==================== Signed-off-by: -
Sameeh Jubran authored
Update driver version to match device specification. Signed-off-by:
Sameeh Jubran <sameehj@amazon.com> Signed-off-by:
-
Sameeh Jubran authored
Let the compiler decide if the function should be inline in *.c files Signed-off-by:
-
Sameeh Jubran authored
Implement the set_ringparam() function of the ethtool interface to enable the changing of io queue sizes. Signed-off-by:
Arthur Kiyanovski <akiyano@amazon.com> Signed-off-by:
-
Sameeh Jubran authored
If there is not enough memory to allocate io queues the driver will try to allocate smaller queues. The backoff algorithm is as follows: 1. Try to allocate TX and RX and if successful. 1.1. return success 2. Divide by 2 the size of the larger of RX and TX queues (or both if their size is the same). 3. If TX or RX is smaller than 256 3.1. return failure. 4. else 4.1. go back to 1. Also change the tx_queue_size, rx_queue_size field names in struct adapter to requested_tx_queue_size and requested_rx_queue_size, and use RX and TX queue 0 for actual queue sizes. Explanation: The original fields were useless as they were simply used to assign values once from them to each of the queues in the adapter in ena_probe(). They could simply be deleted. However now that we have a backoff feature, we have use for them. In case of backoff there is a difference between the requested queue sizes and the actual sizes. Therefore there is a need to save the requested queue size for future retries of queue allocation (for example if allocation failed and then ifdown + ifup was called we want to start the allocation from the original requested size of the queues). Signed-off-by:
-
Sameeh Jubran authored
Currently ethtool -g shows the same size for current and max queue sizes. Signed-off-by:
-
Sameeh Jubran authored
Use MAX_QUEUES_EXT get feature capability to query the device. Signed-off-by:
Netanel Belgazal <netanel@amazon.com> Signed-off-by:
-
Arthur Kiyanovski authored
Add a new admin command to support different queue size for Tx/Rx queues (the change also support different SQ/CQ sizes) Signed-off-by:
-
David S. Miller authored
Ioana Radulescu says: ==================== dpaa2-eth: Add support for MQPRIO offloading Add support for adding multiple TX traffic classes with mqprio. We can have up to one netdev queue and hardware frame queue per TC per core. ==================== Signed-off-by: -
Ioana Radulescu authored
Implement mqprio qdisc support by mapping traffic classes to different hardware enqueue priorities. The maximum number of supported traffic classes is an attribute of each DPNI object. The traffic classes map to hardware priorities from highest (0) to lowest (highest prio number). The skb priority information received from the stack is used to select the hardware Tx queue on which to enqueue the frame. Signed-off-by:
Ioana Radulescu <ruxandra.radulescu@nxp.com> Signed-off-by:
Bogdan Purcareata <bogdan.purcareata@nxp.com> Signed-off-by:
-
Ioana Radulescu authored
DPNI objects can have multiple traffic classes, as reflected by the num_tc attribute. Until now we ignored its value and only used traffic class 0. This patch adds support for multiple Tx traffic classes; we have num_queues x num_tcs hardware queues available for each interface. Signed-off-by:
-
Ioana Radulescu authored
Move the code configuring xps on the netdev TX queues to a separate function. A subsequent patch will need to call this in another context as well. Signed-off-by:
-
Grygorii Strashko authored
Add dependency to TI CPTS from Common CLK framework COMMON_CLK to fix allyesconfig build for Powerpc: drivers/net/ethernet/ti/cpts.c: In function 'cpts_of_mux_clk_setup': drivers/net/ethernet/ti/cpts.c:567:2: error: implicit declaration of function 'of_clk_parent_fill'; did you mean 'of_clk_get_parent_name'? [-Werror=implicit-function-declaration] of_clk_parent_fill(refclk_np, parent_names, num_parents); ^~~~~~~~~~~~~~~~~~ of_clk_get_parent_name Fixes: a3047a81 ("net: ethernet: ti: cpts: add support for ext rftclk selection") Reported-by:
Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by:
Grygorii Strashko <grygorii.strashko@ti.com> Signed-off-by:
-
Florian Fainelli authored
We need to specifically deal with phylink_of_phy_connect() returning -ENODEV, because this can happen when a CPU/DSA port does connect neither to a PHY, nor has a fixed-link property. This is a valid use case that is permitted by the binding and indicates to the switch: auto-configure port with maximum capabilities. Fixes: 0e279218 ("net: dsa: Use PHYLINK for the CPU/DSA ports") Signed-off-by:
Florian Fainelli <f.fainelli@gmail.com> Reviewed-by:
Andrew Lunn <andrew@lunn.ch> Signed-off-by:
-
Vivien Didelot authored
During a port FDB dump operation, the mutex protecting the concurrent access to the switch registers is currently held by the internal mv88e6xxx_port_db_dump and mv88e6xxx_port_db_dump_fid helpers. It must be held at the higher level in mv88e6xxx_port_fdb_dump which is called directly by DSA through ds->ops->port_fdb_dump. Fix this. Signed-off-by:
Vivien Didelot <vivien.didelot@gmail.com> Signed-off-by:
-
Nicolas Saenz Julienne authored
Add bindings for Wiznet's w5x00 series of SPI interfaced Ethernet chips. Based on the bindings for microchip,enc28j60. Signed-off-by:
Nicolas Saenz Julienne <nsaenzjulienne@suse.de> Signed-off-by:
-
Nicolas Saenz Julienne authored
The w5X00 chip provides an SPI to Ethernet inteface. This patch allows platform devices to be defined through the device tree. Signed-off-by:
-
Vlad Buslov authored
To remove rtnl lock dependency in tc filter update API when using ingress Qdisc, set QDISC_CLASS_OPS_DOIT_UNLOCKED flag in ingress Qdisc_class_ops. Ingress Qdisc ops don't require any modifications to be used without rtnl lock on tc filter update path. Ingress implementation never changes its q->block and only releases it when Qdisc is being destroyed. This means it is enough for RTM_{NEWTFILTER|DELTFILTER|GETTFILTER} message handlers to hold ingress Qdisc reference while using it without relying on rtnl lock protection. Unlocked Qdisc ops support is already implemented in filter update path by unlocked cls API patch set. Signed-off-by:Vlad Buslov <vladbu@mellanox.com> Signed-off-by:
-
- 11 Jun, 2019 17 commits
-
-
David S. Miller authored
Jakub Kicinski says: ==================== tls: add support for kernel-driven resync and nfp RX offload This series adds TLS RX offload for NFP and completes the offload by providing resync strategies. When TLS data stream looses segments or experiences reorder NIC can no longer perform in line offload. Resyncs provide information about placement of records in the stream so that offload can resume. Existing TLS resync mechanisms are not a great fit for the NFP. In particular the TX resync is hard to implement for packet-centric NICs. This patchset adds an ability to perform TX resync in a way similar to the way initial sync is done - by calling down to the driver when new record is created after driver indicated sync had been lost. Similarly on the RX side, we try to wait for a gap in the stream and send record information for the next record. This works very well for RPC workloads which are the primary focus at this time. ==================== Signed-off-by: -
Jakub Kicinski authored
When TCP stream gets out of sync (driver stops receiving skbs with expected TCP sequence numbers) request a TX resync from the kernel. We try to distinguish retransmissions from missed transmissions by comparing the sequence number to expected - if it's further than the expected one - we probably missed packets. Signed-off-by:
Jakub Kicinski <jakub.kicinski@netronome.com> Reviewed-by:
Dirk van der Merwe <dirk.vandermerwe@netronome.com> Signed-off-by:
-
Jakub Kicinski authored
TLS offload drivers keep track of TCP seq numbers to make sure the packets are fed into the HW in order. When packets get dropped on the way through the stack, the driver will get out of sync and have to use fallback encryption, but unless TCP seq number is resynced it will never match the packets correctly (or even worse - use incorrect record sequence number after TCP seq wraps). Existing drivers (mlx5) feed the entire record on every out-of-order event, allowing FW/HW to always be in sync. This patch adds an alternative, more akin to the RX resync. When driver sees a frame which is past its expected sequence number the stream must have gotten out of order (if the sequence number is smaller than expected its likely a retransmission which doesn't require resync). Driver will ask the stack to perform TX sync before it submits the next full record, and fall back to software crypto until stack has performed the sync. Signed-off-by:
-
Jakub Kicinski authored
Currently only RX direction is ever resynced, however, TX may also get out of sequence if packets get dropped on the way to the driver. Rename the resync callback and add a direction parameter. Signed-off-by:
-
Jakub Kicinski authored
Set ethtool TLS RX feature based on NIC capabilities, and enable TLS RX when connections are added for decryption. Signed-off-by:
-
Dirk van der Merwe authored
Enable kernel-controlled RX resync and propagate TLS connection RX resync from kernel TLS to firmware. Signed-off-by:
-
Jakub Kicinski authored
Some control messages must be sent from atomic context. The mailbox takes sleeping locks and uses a waitqueue so add a "posted" version of communication. Trylock the semaphore and if that's successful kick of the device communication. The device communication will be completed from a workqueue, which will also release the semaphore. If locks are taken queue the message and return. Schedule a different workqueue to take the semaphore and run the communication. Note that the there are currently no atomic users which would actually need the return value, so all replies to posted messages are just freed. Signed-off-by:
-
Jakub Kicinski authored
We need the name nfp_ccm_mbox_alloc() for allocating the mailbox communication channel itself. Signed-off-by:
-
Dirk van der Merwe authored
Firmware indicates when a packet has been decrypted by reusing the currently unused BPF flag. Transfer this information into the skb and provide a statistic of all decrypted segments. Signed-off-by:
-
Jakub Kicinski authored
TLS offload device may lose sync with the TCP stream if packets arrive out of order. Drivers can currently request a resync at a specific TCP sequence number. When a record is found starting at that sequence number kernel will inform the device of the corresponding record number. This requires the device to constantly scan the stream for a known pattern (constant bytes of the header) after sync is lost. This patch adds an alternative approach which is entirely under the control of the kernel. Kernel tracks records it had to fully decrypt, even though TLS socket is in TLS_HW mode. If multiple records did not have any decrypted parts - it's a pretty strong indication that the device is out of sync. We choose the min number of fully encrypted records to be 2, which should hopefully be more than will get retransmitted at a time. After kernel decides the device is out of sync it schedules a resync request. If the TCP socket is empty the resync gets performed immediately. If socket is not empty we leave the record parser to resync when next record comes. Before resync in message parser we peek at the TCP socket and don't attempt the sync if the socket already has some of the next record queued. On resync failure (encrypted data continues to flow in) we retry with exponential backoff, up to once every 128 records (with a 16k record thats at most once every 2M of data). Signed-off-by:
-
Jakub Kicinski authored
handle_device_resync() doesn't describe the function very well. The function checks if resync should be issued upon parsing of a new record. Signed-off-by:
-
Jakub Kicinski authored
TLS offload code casts record number to a u64. The buffer should be aligned to 8 bytes, but its actually a __be64, and the rest of the TLS code treats it as big int. Make the offload callbacks take a byte array, drivers can make the choice to do the ugly cast if they want to. Prepare for copying the record number onto the stack by defining a constant for max size of the byte array. Signed-off-by:
-
Jakub Kicinski authored
We subtract "TLS_HEADER_SIZE - 1" from req_seq, then if they match we add the same constant to seq. Just add it to seq, and we don't have to touch req_seq. Signed-off-by:
-
Mao Wenan authored
The variable 'status' in __packet_lookup_frame_in_block() is never used since introduction in commit f6fb8f10 ("af-packet: TPACKET_V3 flexible buffer implementation."), we can remove it. Signed-off-by:
Mao Wenan <maowenan@huawei.com> Signed-off-by:
-
Taehee Yoo authored
ASSERT_OVSL() in ovs_vport_del() is unnecessary because ovs_vport_del() is only called by ovs_dp_detach_port() and ovs_dp_detach_port() calls ASSERT_OVSL() too. Signed-off-by:
Taehee Yoo <ap420073@gmail.com> Reviewed-by:
Greg Rose <gvrose8192@gmail.com> Signed-off-by:
-
Taehee Yoo authored
netlink_walk_start() needed to return an error code because of rhashtable_walk_init(). but that was converted to rhashtable_walk_enter() and it is a void type function. so now netlink_walk_start() doesn't need any return value. Signed-off-by:
-
Stefano Brivio authored
This test checks that route exceptions can be successfully listed and flushed using ip -6 route {list,flush} cache. Signed-off-by:Stefano Brivio <sbrivio@redhat.com> Signed-off-by:
-
- 10 Jun, 2019 4 commits
-
-
David S. Miller authored
David Ahern says: ==================== net: Enable nexthop objects with IPv4 and IPv6 routes This is the final set of the initial nexthop object work. When I started this idea almost 2 years ago, it took 18 seconds to inject 700k+ IPv4 routes with 1 hop and about 28 seconds for 4-paths. Some of that time was due to inefficiencies in 'ip', but most of it was kernel side with excessive synchronize_rcu calls in ipv4, and redundant processing validating a nexthop spec (device, gateway, encap). Worse, the time increased dramatically as the number of legs in the routes increased; for example, taking over 72 seconds for 16-path routes. After this set, with increased dirty memory limits (fib_sync_mem sysctl), an improved ip and nexthop objects a full internet fib (743,799 routes based on a pull in January 2019) can be pushed to the kernel in 4.3 seconds. Even better, the time to insert is "almost" constant with increasing number of paths. The 'almost constant' time is due to expanding the nexthop definitions when generating notifications. A follow on patch will be sent adding a sysctl that allows an admin to avoid the nexthop expansion and truly get constant route insert time regardless of the number of paths in a route! (Useful once all programs used for a deployment that care about routes understand nexthop objects). To be clear, 'ip' is used for benchmarking for no other reason than 'ip -batch' is a trivial to use for the tests. FRR, for example, better manages nexthops and route changes and the way those are pushed to the kernel and thus will have less userspace processing times than 'ip -batch'. Patches 1-10 iterate over fib6_nh with a nexthop invoke a processing function per fib6_nh. Prior to nexthop objects, a fib6_info referenced a single fib6_nh. Multipath routes were added as separate fib6_info for each leg of the route and linked as siblings: f6i -> sibling -> sibling ... -> sibling | | +--------- multipath route ---------+ With nexthop objects a single fib6_info references an external nexthop which may have a series of fib6_nh: f6i ---> nexthop ---> fib6_nh ... fib6_nh making IPv6 routes similar to IPv4. The side effect is that a single fib6_info now indirectly references a series of fib6_nh so the code needs to walk each entry and call the local, per-fib6_nh processing function. Patches 11 and 13 wire up use of nexthops with fib entries for IPv4 and IPv6. With these commits you can actually use nexthops with routes. Patch 12 is an optimization for IPv4 when using nexthops in the most predominant use case (no metrics). Patches 14 handles replace of a nexthop config. Patches 15-18 add update pmtu and redirect tests to use both old and new routing. Patches 19 and 20 add new tests for the nexthop infrastructure. The first is single nexthop is used by multiple prefixes to communicate with remote hosts. This is on top of the functional tests already committed. The second verifies multipath selection. v4 - changed return to 'goto out' in patch 9 since the rcu_read_lock is held (noticed by Wei) v3 - removed found arg in patch 7 and changed rt6_nh_remove_exception_rt to return 1 when a match is found for an exception v2 - changed ++i to i++ in patches 1 and 14 as noticed by DaveM - improved commit message for patch 14 (nexthop replace) - removed the skip_fib argument to remove_nexthop; vestige of an older design ==================== Reviewed-By:Wei Wang <weiwan@google.com> Signed-off-by:
-
David Ahern authored
Add a version of router_multipath.sh that uses nexthop objects for routes. Ido requested a version that does not cause regressions with mlxsw testing since it does not support nexthop objects yet. Signed-off-by:
David Ahern <dsahern@gmail.com> Signed-off-by:
-
David Ahern authored
Add tests where multiple FIB entries use the same nexthop object. Generate per-cpu cached routes for each by running ping on each cpu, and then generate exceptions unique to each prefix (remote host) with different mtus. Signed-off-by:
-
David Ahern authored
Add a second pass to icmp_redirect.sh to use nexthop objects for routes. Signed-off-by:
-
