- 04 Aug, 2017 9 commits
-
-
Willem de Bruijn authored
Introduce regression test for msg_zerocopy feature. Send traffic from one process to another with and without zerocopy. Evaluate tcp, udp, raw and packet sockets, including variants - udp: corking and corking with mixed copy/zerocopy calls - raw: with and without hdrincl - packet: at both raw and dgram level Test on both ipv4 and ipv6, optionally with ethtool changes to disable scatter-gather, tx checksum or tso offload. All of these can affect zerocopy behavior. The regression test can be run on a single machine if over a veth pair. Then skb_orphan_frags_rx must be modified to be identical to skb_orphan_frags to allow forwarding zerocopy locally. The msg_zerocopy.sh script will setup the veth pair in network namespaces and run all tests. Signed-off-by:
Willem de Bruijn <willemb@google.com> Signed-off-by:
David S. Miller <davem@davemloft.net>
-
Willem de Bruijn authored
Enable support for MSG_ZEROCOPY to the TCP stack. TSO and GSO are both supported. Only data sent to remote destinations is sent without copying. Packets looped onto a local destination have their payload copied to avoid unbounded latency. Tested: A 10x TCP_STREAM between two hosts showed a reduction in netserver process cycles by up to 70%, depending on packet size. Systemwide, savings are of course much less pronounced, at up to 20% best case. msg_zerocopy.sh 4 tcp: without zerocopy tx=121792 (7600 MB) txc=0 zc=n rx=60458 (7600 MB) with zerocopy tx=286257 (17863 MB) txc=286257 zc=y rx=140022 (17863 MB) This test opens a pair of sockets over veth, one one calls send with 64KB and optionally MSG_ZEROCOPY and on the other reads the initial bytes. The receiver truncates, so this is strictly an upper bound on what is achievable. It is more representative of sending data out of a physical NIC (when payload is not touched, either). Signed-off-by: -
Willem de Bruijn authored
Bound the number of pages that a user may pin. Follow the lead of perf tools to maintain a per-user bound on memory locked pages commit 789f90fc ("perf_counter: per user mlock gift") Signed-off-by:
-
Willem de Bruijn authored
In the simple case, each sendmsg() call generates data and eventually a zerocopy ready notification N, where N indicates the Nth successful invocation of sendmsg() with the MSG_ZEROCOPY flag on this socket. TCP and corked sockets can cause send() calls to append new data to an existing sk_buff and, thus, ubuf_info. In that case the notification must hold a range. odify ubuf_info to store a inclusive range [N..N+m] and add skb_zerocopy_realloc() to optionally extend an existing range. Also coalesce notifications in this common case: if a notification [1, 1] is about to be queued while [0, 0] is the queue tail, just modify the head of the queue to read [0, 1]. Coalescing is limited to a few TSO frames worth of data to bound notification latency. Signed-off-by:
-
Willem de Bruijn authored
Prepare the datapath for refcounted ubuf_info. Clone ubuf_info with skb_zerocopy_clone() wherever needed due to skb split, merge, resize or clone. Split skb_orphan_frags into two variants. The split, merge, .. paths support reference counted zerocopy buffers, so do not do a deep copy. Add skb_orphan_frags_rx for paths that may loop packets to receive sockets. That is not allowed, as it may cause unbounded latency. Deep copy all zerocopy copy buffers, ref-counted or not, in this path. The exact locations to modify were chosen by exhaustively searching through all code that might modify skb_frag references and/or the the SKBTX_DEV_ZEROCOPY tx_flags bit. The changes err on the safe side, in two ways. (1) legacy ubuf_info paths virtio and tap are not modified. They keep a 1:1 ubuf_info to sk_buff relationship. Calls to skb_orphan_frags still call skb_copy_ubufs and thus copy frags in this case. (2) not all copies deep in the stack are addressed yet. skb_shift, skb_split and skb_try_coalesce can be refined to avoid copying. These are not in the hot path and this patch is hairy enough as is, so that is left for future refinement. Signed-off-by: -
Willem de Bruijn authored
The send call ignores unknown flags. Legacy applications may already unwittingly pass MSG_ZEROCOPY. Continue to ignore this flag unless a socket opts in to zerocopy. Introduce socket option SO_ZEROCOPY to enable MSG_ZEROCOPY processing. Processes can also query this socket option to detect kernel support for the feature. Older kernels will return ENOPROTOOPT. Signed-off-by:
-
Willem de Bruijn authored
The kernel supports zerocopy sendmsg in virtio and tap. Expand the infrastructure to support other socket types. Introduce a completion notification channel over the socket error queue. Notifications are returned with ee_origin SO_EE_ORIGIN_ZEROCOPY. ee_errno is 0 to avoid blocking the send/recv path on receiving notifications. Add reference counting, to support the skb split, merge, resize and clone operations possible with SOCK_STREAM and other socket types. The patch does not yet modify any datapaths. Signed-off-by:
-
Willem de Bruijn authored
Refine skb_copy_ubufs to support compound pages. With upcoming TCP zerocopy sendmsg, such fragments may appear. The existing code replaces each page one for one. Splitting each compound page into an independent number of regular pages can result in exceeding limit MAX_SKB_FRAGS if data is not exactly page aligned. Instead, fill all destination pages but the last to PAGE_SIZE. Split the existing alloc + copy loop into separate stages: 1. compute bytelength and minimum number of pages to store this. 2. allocate 3. copy, filling each page except the last to PAGE_SIZE bytes 4. update skb frag array Signed-off-by:
-
Willem de Bruijn authored
Add sock_omalloc and sock_ofree to be able to allocate control skbs, for instance for looping errors onto sk_error_queue. The transmit budget (sk_wmem_alloc) is involved in transmit skb shaping, most notably in TCP Small Queues. Using this budget for control packets would impact transmission. Signed-off-by:
-
- 03 Aug, 2017 31 commits
-
-
David S. Miller authored
Jiri Pirko says: ==================== mlxsw: Support for IPv6 UC router Ido says: This set adds support for IPv6 unicast routes offload. The first four patches make the FIB notification chain generic so that it could be used by address families other than IPv4. This is done by having each address family register its callbacks with the common code, so that its FIB tables and rules could be dumped upon registration to the chain, while ensuring the integrity of the dump. The exact mechanics are explained in detail in the first patch. The next six patches build upon this work and add the necessary callbacks in IPv6 code. This allows listeners of the chain to receive notifications about IPv6 routes addition, deletion and replacement as well as FIB rules notifications. Unlike user space notifications for IPv6 multipath routes, the FIB notification chain notifies these on a per-nexthop basis. This allows us to keep the common code lean and is also unnecessary, as notifications are serialized by each table's lock whereas applications maintaining netlink caches may suffer from concurrent dumps and deletions / additions of routes. The next five patches audit the different code paths reading the route's reference count (rt6i_ref) and remove assumptions regarding its meaning. This is needed since non-FIB users need to be able to hold a reference on the route and a non-zero reference count no longer means the route is in the FIB. The last six patches enable the mlxsw driver to offload IPv6 unicast routes to the Spectrum ASIC. Without resorting to ACLs, lookup is done solely based on the destination IP, so the abort mechanism is invoked upon the addition of source-specific routes. Follow-up patch sets will increase the scale of gatewayed routes by consolidating identical nexthop groups to one adjacency entry in the device's adjacency table (as in IPv4), as well as add support for NH_{ADD,DEL} events which enable support for the 'ignore_routes_with_linkdown' sysctl. Changes in v2: * Provide offload indication for individual nexthops (David Ahern). * Use existing route reference count instead of adding another one. This resulted in several new patches to remove assumptions regarding current semantics of the existing reference count (David Ahern). * Add helpers to allow non-FIB users to take a reference on route. * Remove use of tb6_lock in mlxsw (David Ahern). * Add IPv6 dependency to mlxsw. ==================== Signed-off-by: -
Ido Schimmel authored
We now have all the necessary IPv6 infrastructure in place, so stop ignoring these notifications. Signed-off-by:
Ido Schimmel <idosch@mellanox.com> Signed-off-by:
Jiri Pirko <jiri@mellanox.com> Signed-off-by:
-
Ido Schimmel authored
Without resorting to ACLs, the device performs route lookup solely based on the destination IP address. In case source-specific routing is needed, an error is returned and the abort mechanism is activated, thus allowing the kernel to take over forwarding decisions. Instead of aborting, we can trap specific destination prefixes where source-specific routes are present, but this will result in a lot more code that is unlikely to ever be used. Signed-off-by:
-
Ido Schimmel authored
In case we got a replace event, then the replaced route must exist. If the route isn't capable of multipath, then replace first matching non-multipath capable route. If the route is capable of multipath and matching multipath capable route is found, then replace it. Otherwise, replace first matching non-multipath capable route. The new route is inserted before the replaced one. In case the replaced route is currently offloaded, then it's overwritten in the device's table by the new route and later deleted, thus not impacting routed traffic. Signed-off-by:
-
Ido Schimmel authored
Allow directly connected and remote unicast IPv6 routes to be programmed to the device's tables. As with IPv4, identical routes - sharing the same destination prefix - are ordered in a FIB node according to their table ID and then the metric. While the kernel doesn't share the same trie for the local and main table, this does happen in the device, so ordering according to table ID is needed. Since individual nexthops can be added and deleted in IPv6, each FIB entry stores a linked list of the rt6_info structs it represents. Upon the addition or deletion of a nexthop, a new nexthop group is allocated according to the new configuration and the old one is destroyed. Identical groups aren't currently consolidated, but will be in a follow-up patchset. Signed-off-by:
-
Ido Schimmel authored
We only allow FIB offload in the presence of default rules or an l3mdev rule. In a similar fashion to IPv4 FIB rules, sanitize IPv6 rules. Signed-off-by:
-
Ido Schimmel authored
The FIB notification block currently only handles IPv4 events, but we want to start handling IPv6 events soon, so lay the groundwork now. Do that by preparing the work item and process it according to the notified address family. Signed-off-by:
-
Ido Schimmel authored
Similar to commit 1c677b3d ("ipv4: fib: Add fib_info_hold() helper") and commit b423cb10 ("ipv4: fib: Export free_fib_info()") add an helper to hold a reference on rt6_info and export rt6_release() to drop it and potentially release the route. This is needed so that drivers capable of FIB offload could hold a reference on the route before queueing it for offload and drop it after the route has been programmed to the device's tables. Signed-off-by:
-
Ido Schimmel authored
When an interface is brought back up, the kernel tries to restore the host routes tied to its permanent addresses. However, if the host route was removed from the FIB, then we need to reinsert it. This is done by releasing the current dst and allocating a new, so as to not reuse a dst with obsolete values. Since this function is called under RTNL and using the same explanation from the previous patch, we can test if the route is in the FIB by checking its node pointer instead of its reference count. Tested using the following script and Andrey's reproducer mentioned in commit 8048ced9 ("net: ipv6: regenerate host route if moved to gc list") and linked below: $ ip link set dev lo up $ ip link add dummy1 type dummy $ ip -6 address add cafe::1/64 dev dummy1 $ ip link set dev lo down # cafe::1/128 is removed $ ip link set dev dummy1 up $ ip link set dev lo up The host route is correctly regenerated. Signed-off-by:
-
Ido Schimmel authored
When the loopback device is brought back up we need to check if the host route attached to the address is still in the FIB and regenerate one in case it's not. Host routes using the loopback device are always inserted into and removed from the FIB under RTNL (under which this function is called), so we can test their node pointer instead of the reference count in order to check if the route is in the FIB or not. Tested using the following script from Nicolas mentioned in commit a220445f ("ipv6: correctly add local routes when lo goes up"): $ ip link add dummy1 type dummy $ ip link set dummy1 up $ ip link set lo down ; ip link set lo up The host route is correctly regenerated. Signed-off-by:
-
Ido Schimmel authored
When a route is deleted its node pointer is set to NULL to indicate it's no longer linked to its node. Do the same for routes that are replaced. This will later allow us to test if a route is still in the FIB by checking its node pointer instead of its reference count. Signed-off-by:
-
Ido Schimmel authored
The code currently assumes that only FIB nodes can hold a reference on routes. Therefore, after fib6_purge_rt() has run and the route is no longer present in any intermediate nodes, it's assumed that its reference count would be 1 - taken by the node where it's currently stored. However, we're going to allow users other than the FIB to take a reference on a route, so this assumption is no longer valid and the BUG_ON() needs to be removed. Note that purging only takes place if the initial reference count is different than 1. I've left that check intact, as in the majority of systems (where routes are only referenced by the FIB), it does actually mean the route is present in intermediate nodes. Signed-off-by:
-
Ido Schimmel authored
Allow user space applications to see which routes are offloaded and which aren't by setting the RTNH_F_OFFLOAD flag when dumping them. To be consistent with IPv4, offload indication is provided on a per-nexthop basis. Signed-off-by:
-
Ido Schimmel authored
Dump all the FIB tables in each net namespace upon registration to the FIB notification chain so that the callee will have a complete view of the tables. The integrity of the dump is ensured by a per-table sequence counter that is incremented (under write lock) whenever a route is added or deleted from the table. All the sequence counters are read (under each table's read lock) and summed, prior and after the dump. In case the counters differ, then the dump is either restarted or the registration fails. While it's possible for a table to be modified after its counter has been read, this isn't really a problem. In case it happened before it was read the second time, then the comparison at the end will fail. If it happened afterwards, then we're guaranteed to be notified about the change, as the notification block is registered prior to the second read. Signed-off-by:
-
Ido Schimmel authored
Allow users of the FIB notification chain to receive a complete view of the IPv6 FIB rules upon registration to the chain. The integrity of the dump is ensured by a per-family sequence counter that is incremented (under RTNL) whenever a rule is added or deleted. All the sequence counters are read (under RTNL) and summed, prior and after the dump. In case the counters differ, then the dump is either restarted or the registration fails. Signed-off-by:
-
Ido Schimmel authored
As with IPv4, allow listeners of the FIB notification chain to receive notifications whenever a route is added, replaced or deleted. This is done by placing calls to the FIB notification chain in the two lowest level functions that end up performing these operations - namely, fib6_add_rt2node() and fib6_del_route(). Unlike IPv4, APPEND notifications aren't sent as the kernel doesn't distinguish between "append" (NLM_F_CREATE|NLM_F_APPEND) and "prepend" (NLM_F_CREATE). If NLM_F_EXCL isn't set, duplicate routes are always added after the existing duplicate routes. Signed-off-by:
-
Ido Schimmel authored
We're about to add IPv6 FIB offload support, so implement the necessary callbacks in IPv6 code, which will later allow us to add routes and rules notifications. Signed-off-by:
-
Ido Schimmel authored
As explained in commit 3c71006d ("ipv4: fib_rules: Check if rule is a default rule"), drivers supporting IPv6 FIB offload need to be able to sanitize the rules they don't support and potentially flush their tables. Add an IPv6 helper to check if a FIB rule is a default rule. Signed-off-by:
-
Ido Schimmel authored
Unlike the routing tables, the FIB rules share a common core, so instead of replicating the same logic for each address family we can simply dump the rules and send notifications from the core itself. To protect the integrity of the dump, a rules-specific sequence counter is added for each address family and incremented whenever a rule is added or deleted (under RTNL). Signed-off-by:
-
Ido Schimmel authored
As in previous patch, ignore IPv6 notifications since the driver doesn't support these. Signed-off-by:
-
Ido Schimmel authored
We're about to add IPv6 notifications in the FIB notification chain, but the driver currently doesn't support these, so ignore them. Signed-off-by:
-
Ido Schimmel authored
The FIB notification chain is currently soley used by IPv4 code. However, we're going to introduce IPv6 FIB offload support, which requires these notification as well. As explained in commit c3852ef7 ("ipv4: fib: Replay events when registering FIB notifier"), upon registration to the chain, the callee receives a full dump of the FIB tables and rules by traversing all the net namespaces. The integrity of the dump is ensured by a per-namespace sequence counter that is incremented whenever a change to the tables or rules occurs. In order to allow more address families to use the chain, each family is expected to register its fib_notifier_ops in its pernet init. These operations allow the common code to read the family's sequence counter as well as dump its tables and rules in the given net namespace. Additionally, a 'family' parameter is added to sent notifications, so that listeners could distinguish between the different families. Implement the common code that allows listeners to register to the chain and for address families to register their fib_notifier_ops. Subsequent patches will implement these operations in IPv6. In the future, ipmr and ip6mr will be extended to provide these notifications as well. Signed-off-by:
-
David S. Miller authored
Thomas Petazzoni says: ==================== net: mvpp2: add TX interrupts support So far, the mvpp2 driver was using an hrtimer to handle TX completion. This patch series adds support for using TX interrupts (for each CPU) on PPv2.2, the variant of the IP used on Marvell Armada 7K/8K. Dave: this version can be applied right away, it no longer depends on Antoine's patch series. Antoine series had some comments, so he will have to respin later on. Therefore, let's merge this smaller patch series first. Changes since v1: - Rebased on top of net-next, instead of on top of Antoine's series. - Removed the Device Tree patch, as it shouldn't go through the net tree. ==================== Signed-off-by: -
Thomas Petazzoni authored
The PPv2.2 unit has several interrupts used for TX completion notification. This commit updates the Device Tree binding describing this HW block to mention such interrupts. While at it, we update the example to use a recent Device Tree example, that uses interrupts going through the ICU, and not to the GIC directly. Signed-off-by:
Thomas Petazzoni <thomas.petazzoni@free-electrons.com> Signed-off-by:
-
Thomas Petazzoni authored
This commit adds the support for two related features: - Support for TX interrupts, with one interrupt for each CPU - Support for different RX queue distribution modes MVPP2_QDIST_SINGLE_MODE where a single interrupt, shared by all CPUs, receives the RX events, and MVPP2_QDIST_MULTI_MODE, where the per-CPU interrupts used for TX events are also used for RX events. Since additional interrupts are needed, an update to the Device Tree binding is needed. However, backward compatibility is preserved with the old Device Tree binding, by gracefully degrading to the original behavior, with only one RX interrupt, and TX completion being handled by an hrtimer. Signed-off-by:
-
Thomas Petazzoni authored
In preparation to the introduction of TX interrupts and improved RX queue distribution, this commit introduces the concept of "queue vector". A queue vector represents a number of RX and/or TX queues, and an associated NAPI instance and interrupt. This commit currently only creates a single queue_vector, so there are no changes in behavior, but it paves the way for additional queue_vector in the next commits. Signed-off-by:
-
Thomas Petazzoni authored
The PPv2.2 IP has a concept of "software thread", with all registers of the PPv2.2 mapped 8 times, for concurrent accesses by 8 "software threads". In addition, interrupts on RX queues are associated to such "software thread". For most cases, we map a "software thread" to the more conventional concept of CPU, but we will soon have one exception: we will have a model where we have one TX interrupt per CPU (each using one software thread), and all RX events mapped to another software thread (associated to another interrupt). In preparation for this change, it makes sense to change the naming from MVPP2_MAX_CPUS to MVPP2_MAX_THREADS, and plan for 8 software threads instead of 4 currently. Signed-off-by:
-
Thomas Petazzoni authored
Currently, the global variables rxq_number and txq_number hold the number of per-port TXQs and RXQs. Until now, such numbers were constant regardless of the driver configuration. As we are going to introduce different modes for TX and RX queues, these numbers will depend on the configuration (PPv2.1 vs. PPv2.2, exact queue distribution logic). Therefore, as a preparation, we move the number of RXQs and TXQs in the 'struct mvpp2_port' structure, next to the RXQs and TXQs descriptor arrays. For now, they remain initialized to the same default values as rxq_number/txq_number used to be initialized, but this will change in future commits. The only non-mechanical change in this patch is that the check to verify hardware constraints on the number of RXQs and TXQs is moved from mvpp2_probe() to mvpp2_port_probe(), since it's now in mvpp2_port_probe() that we initialize the per-port count of RXQ and TXQ. Signed-off-by:
-
Thomas Petazzoni authored
The RX queue group allocation is anyway re-done later in mvpp2_port_init(), so resetting it in mvpp2_init() is not very useful, and will be annoying as we are going to rework the RX queue group allocation logic. Signed-off-by:
-
Thomas Petazzoni authored
The MVPP21_ISR_RXQ_GROUP_REG register is not indexed by rxq, but by port, so we fix the parameter name accordingly. There are no functional changes. Signed-off-by:
-
Romain Perier authored
This operation is required for handling ioctl commands like SIOCGMIIREG, when debugging MDIO registers from userspace. This commit adds support for this operation. Signed-off-by:
Romain Perier <romain.perier@collabora.com> Signed-off-by:
-
