- 20 Aug, 2021 40 commits
-
-
Alex Elder authored
Replace the ipa_clock_get() call in ipa_uc_clock() when taking the "proxy" clock reference for the microcontroller with a call to pm_runtime_get_sync(). Replace calls of ipa_clock_put() for the microcontroller with pm_runtime_put() calls instead. There is a chance we get an error when taking the microcontroller power reference. This is an unlikely scenario, where system suspend is initiated just before we learn the modem is booting. For now we'll just accept that this could occur, and report it if it does. Signed-off-by:
Alex Elder <elder@linaro.org> Signed-off-by:
David S. Miller <davem@davemloft.net>
-
Alex Elder authored
If the "modem-init" Device Tree property is present for a platform, the modem performs early IPA hardware initialization, and signals this is complete with an "ipa-setup-ready" SMP2P interrupt. This triggers a call to ipa_setup(), which requires the hardware to be powered. Replace the call to ipa_clock_get() in this case with a call to pm_runtime_get_sync(). And replace the corresponding calls to ipa_clock_put() with calls to pm_runtime_put() instead. There is a chance we get an error when taking this power reference. This is an unlikely scenario, where system suspend is initiated just before the modem signals it has finished initializing the IPA hardware. For now we'll just accept that this could occur, and report it if it does. Signed-off-by:
-
Alex Elder authored
We need the hardware to be powered starting at the config stage of initialization when the IPA driver probes. And we need it powered when the driver is removed, at least until the deconfig stage has completed. Replace callers of ipa_clock_get() in ipa_probe() and ipa_exit(), calling pm_runtime_get_sync() instead. Replace the corresponding callers of ipa_clock_put(), calling pm_runtime_put() instead. The only error we expect when getting power would occur when the system is suspended. The ->probe and ->remove driver callbacks won't be called when suspended, so issue a WARN() call if an error is seen getting power. Signed-off-by:
-
Alex Elder authored
Jakub Kicinski pointed out a race condition in ipa_start_xmit() in a recently-accepted series of patches: https://lore.kernel.org/netdev/20210812195035.2816276-1-elder@linaro.org/ We are stopping the modem TX queue in that function if the power state is not active. We restart the TX queue again once hardware resume is complete. TX path Power Management ------- ---------------- pm_runtime_get(); no power Start resume Stop TX queue ... pm_runtime_put() Resume complete return NETDEV_TX_BUSY Start TX queue pm_runtime_get() Power present, transmit pm_runtime_put() (auto-suspend) The issue is that the power management (resume) activity and the network transmit activity can occur concurrently, and there's a chance the queue will be stopped *after* it has been started again. TX path Power Management ------- ---------------- Resume underway pm_runtime_get(); no power ... Resume complete Start TX queue Stop TX queue <-- No more transmits after this pm_runtime_put() return NETDEV_TX_BUSY We address this using a STARTED flag to indicate when the TX queue has been started from the resume path, and a spinlock to make the flag and queue updates happen atomically. TX path Power Management ------- ---------------- Resume underway pm_runtime_get(); no power Resume complete start TX queue \ If STARTED flag is *not* set: > atomic Stop TX queue set STARTED flag / pm_runtime_put() return NETDEV_TX_BUSY A second flag is used to address a different race that involves another path requesting power. TX path Other path Power Management ------- ---------- ---------------- pm_runtime_get_sync() Resume Start TX queue \ atomic Set STARTED flag / (do its thing) pm_runtime_put() (auto-suspend) pm_runtime_get() Mark delayed resume STARTED *is* set, so do *not* stop TX queue <-- Queue should be stopped here pm_runtime_put() return NETDEV_TX_BUSY Suspend done, resume Resume complete pm_runtime_get() Stop TX queue (STARTED is *not* set) Start TX queue \ atomic pm_runtime_put() Set STARTED flag / return NETDEV_TX_BUSY So a STOPPED flag is set in the transmit path when it has stopped the TX queue, and this pair of operations is also protected by the spinlock. The resume path only restarts the TX queue if the STOPPED flag is set. This case isn't a major problem, but it avoids the "non-trivial amount of useless work" done by the networking stack when NETDEV_TX_BUSY is returned. Fixes: 6b51f802 ("net: ipa: ensure hardware has power in ipa_start_xmit()") Signed-off-by:
-
David S. Miller authored
Vladimir Oltean says: ==================== Small ocelot VLAN improvements This small series propagates some VLAN restrictions via netlink extack and creates some helper functions instead of open-coding VLAN table manipulations from multiple places. This is split from the larger "DSA FDB isolation" series, hence the v2 tag: https://patchwork.kernel.org/project/netdevbpf/cover/20210818120150.892647-1-vladimir.oltean@nxp.com/ ==================== Signed-off-by:
-
Vladimir Oltean authored
This is a mostly cosmetic patch that creates some helpers for accessing the VLAN table. These helpers are also a bit more careful in that they do not modify the ocelot->vlan_mask unless the hardware operation succeeded. Not all callers check the return value (the init code doesn't), but anyway. Signed-off-by:
Vladimir Oltean <vladimir.oltean@nxp.com> Signed-off-by:
-
Vladimir Oltean authored
We need to transmit more restrictions in future patches, convert this one to netlink extack. Signed-off-by:
-
Vladimir Oltean authored
We need to reject some more configurations in future patches, convert the existing one to netlink extack. Signed-off-by:
-
David S. Miller authored
Vladimir Oltean says: ==================== Ocelot phylink fixes This series addresses a regression reported by Horatiu which introduced by the ocelot conversion to phylink: there are broken device trees in the wild, and the driver fails to probe the entire switch when a port fails to probe, which it previously did not do. Continue probing even when some ports fail to initialize properly. ==================== Signed-off-by: -
Vladimir Oltean authored
The existing ocelot device trees, like ocelot_pcb123.dts for example, have SERDES ports (ports 4 and higher) that do not have status = "disabled"; but on the other hand do not have a phy-handle or a fixed-link either. So from the perspective of phylink, they have broken DT bindings. Since the blamed commit, probing for the entire switch will fail when such a device tree binding is encountered on a port. There used to be this piece of code which skipped ports without a phy-handle: phy_node = of_parse_phandle(portnp, "phy-handle", 0); if (!phy_node) continue; but now it is gone. Anyway, fixed-link setups are a thing which should work out of the box with phylink, so it would not be in the best interest of the driver to add that check back. Instead, let's look at what other drivers do. Since commit 86f8b1c0 ("net: dsa: Do not make user port errors fatal"), DSA continues after a switch port fails to register, and works only with the ports that succeeded. We can achieve the same behavior in ocelot by unregistering the devlink port for ports where ocelot_port_phylink_create() failed (called via ocelot_probe_port), and clear the bit in devlink_ports_registered for that port. This will make the next iteration reconsider the port that failed to probe as an unused port, and re-register a devlink port of type UNUSED for it. No other cleanup should need to be performed, since ocelot_probe_port() should be self-contained when it fails. Fixes: e6e12df6 ("net: mscc: ocelot: convert to phylink") Reported-and-tested-by:
Horatiu Vultur <horatiu.vultur@microchip.com> Signed-off-by:
-
Horatiu Vultur authored
There are cases where we would like to continue probing the switch even if one port has failed to probe. When that happens, we need to unregister a devlink_port of type DEVLINK_PORT_FLAVOUR_PHYSICAL and re-register it of type DEVLINK_PORT_FLAVOUR_UNUSED. This is fine, except when calling devlink_port_attrs_set on a structure on which devlink_port_register has been previously called, there is a WARN_ON in devlink_port_attrs_set that devlink_port->devlink must be NULL. So don't assume that the memory behind dlp is clean when calling ocelot_port_devlink_init, just zero-initialize it. Signed-off-by:
-
David S. Miller authored
Vladimir Oltean says: ==================== dpaa2-switch phylink fixes This is fixing two regressions introduced by the recent conversion of the dpaa2-switch driver to phylink. ==================== Signed-off-by: -
Vladimir Oltean authored
Currently when probing returns an error, the netdev is freed but phylink_disconnect is not called. Create a common function between the unbind path and the error path, call it the opposite of dpaa2_switch_probe_port: dpaa2_switch_remove_port, and call it from both the unbind and the error path. Fixes: 84cba729 ("dpaa2-switch: integrate the MAC endpoint support") Signed-off-by:
Ioana Ciornei <ioana.ciornei@nxp.com> Signed-off-by:
-
Vladimir Oltean authored
There is an ASSERT_RTNL in phylink_disconnect_phy which triggers whenever dpaa2_switch_port_disconnect_mac is called. To follow the pattern established by dpaa2_eth_disconnect_mac, take the rtnl_mutex every time we call dpaa2_switch_port_disconnect_mac. Fixes: 84cba729 ("dpaa2-switch: integrate the MAC endpoint support") Signed-off-by:
-
David S. Miller authored
Gerhard Engleder says: ==================== Add Xilinx GMII2RGMII loopback support The Xilinx GMII2RGMII driver overrides PHY driver functions in order to configure the device according to the link speed of the PHY attached to it. This is implemented for a normal link but not for loopback. Andrew told me to use phy_loopback and this changes make phy_loopback work in combination with Xilinx GMII2RGMII. ==================== Signed-off-by: -
Gerhard Engleder authored
Configure speed if loopback is used. read_status is not called for loopback. Signed-off-by:
Gerhard Engleder <gerhard@engleder-embedded.com> Signed-off-by:
-
Gerhard Engleder authored
struct phy_device contains a pointer to the PHY driver and nearly everywhere this pointer is used to access the PHY driver. Only mdio_bus_phy_may_suspend() is still using to_phy_driver() instead of the PHY driver pointer. Uniform PHY driver access by eliminating to_phy_driver() use in mdio_bus_phy_may_suspend(). Only phy_bus_match() and phy_probe() are still using to_phy_driver(), because PHY driver pointer is not available there. Signed-off-by:
-
Gerhard Engleder authored
phy_read_status and various other PHY functions support PHY specific overriding of driver functions by using a PHY specific pointer to the PHY driver. Add support of PHY specific override to phy_loopback too. Signed-off-by:
-
David S. Miller authored
Steen Hegelund says: ==================== Adding Frame DMA functionality to Sparx5 v2: Removed an unused variable (proc_ctrl) from sparx5_fdma_start. This add frame DMA functionality to the Sparx5 platform. Until now the Sparx5 SwitchDev driver has been using register based injection and extraction when sending frames to/from the host CPU. With this series the Frame DMA functionality now added. The Frame DMA is only used if the Frame DMA interrupt is configured in the device tree; otherwise the existing register based injection and extraction is used. The Sparx5 has two ports that can be used for sending and receiving frames, but there are 8 channels that can be configured: 6 for injection and 2 for extraction. The additional channels can be used for more advanced scenarios e.g. where virtual cores are used, but currently the driver only uses port 0 and channel 0 and 6 respectively. DCB (data control block) structures are passed to the Frame DMA with suitable information about frame start/end etc, as well as pointers to DB (data blocks) buffers. The Frame DMA engine can use interrupts to signal back when the frames have been injected or extracted. There is a limitation on the DB alignment also for injection: Block must start on 16byte boundaries, and this is why the driver currently copies the data to into separate buffers. The Sparx5 switch core needs a IFH (Internal Frame Header) to pass information from the port to the switch core, and this header is added before injection and stripped after extraction. ==================== Signed-off-by: -
Steen Hegelund authored
This adds the interrupt for the Sparx5 Frame DMA. If this configuration is present the Sparx5 SwitchDev driver will use the Frame DMA feature, and if not it will use register based injection and extraction for sending and receiving frames to the CPU. Signed-off-by:
Steen Hegelund <steen.hegelund@microchip.com> Signed-off-by:
-
Steen Hegelund authored
This add frame DMA functionality to the Sparx5 platform. Ethernet frames can be extracted or injected autonomously to or from the device’s DDR3/DDR3L memory and/or PCIe memory space. Linked list data structures in memory are used for injecting or extracting Ethernet frames. The FDMA generates interrupts when frame extraction or injection is done and when the linked lists need updating. The FDMA implements two extraction channels, one per switch core port towards the VCore CPU system and a total of six injection channels. Extraction channels are mapped one-to-one to the CPU ports, while injection channels can be individually assigned to any CPU port. - FDMA channel 0 through 5 corresponds to CPU port 0 injection direction FDMA_CH_CFG[channel].CH_INJ_PORT is set to 0. - FDMA channel 0 through 5 corresponds to CPU port 1 injection direction when FDMA_CH_CFG[channel].CH_INJ_PORT is set to 1. - FDMA channel 6 corresponds to CPU port 0 extraction direction. - FDMA channel 7 corresponds to CPU port 1 extraction direction. The FDMA implements a strict priority scheme among channels. Extraction channels are prioritized over injection channels and secondarily channels with higher channel number are prioritized over channels with lower number. On the other hand, ports are being served on an equal-bandwidth principle both on injection and extraction directions. The equal-bandwidth principle will not force an equal bandwidth. Instead, it ensures that the ports perform at their best considering the operating conditions. When more than one injection channel is enabled for injection on the same CPU port, priority determines which channel can inject data. Ownership is re-arbitrated on frame boundaries. The FDMA processes linked lists of DMA Control Block Structures (DCBs). The DCBs have the same basic structure for both injection and extraction. A DCB must be placed on a 64-bit word-aligned address in memory. Each DCB has a per-channel configurable amount of associated data blocks in memory, where the frame data is stored. The data blocks that are used by extraction channels must be placed on 64-bit word aligned addresses in memory, and their length must be a multiple of 128 bytes. A DCB carries the pointer to the next DCB of the linked list, the INFO word which holds information for the DCB, and a pair of status word and memory pointer for every data block that it is associated with. Signed-off-by:
-
git://git.open-mesh.org/linux-mergeDavid S. Miller authored
Simon Wunderlich says: ==================== This (updated) cleanup patchset includes the following patches: - bump version strings, by Simon Wunderlich - update docs about move IRC channel away from freenode, by Sven Eckelmann (updated, added missing sign-off) - Switch to kstrtox.h for kstrtou64, by Sven Eckelmann - Update NULL checks, by Sven Eckelmann (2 patches) - remove remaining skb-copy calls for broadcast packets, by Linus Lüssing ==================== Signed-off-by: -
git://git.kernel.org/pub/scm/linux/kernel/git/saeed/linuxDavid S. Miller authored
Saeed Mahameed says: ==================== mlx5-updates-2021-08-19 This series introduces the support for two new mlx5 features: 1) Sample offload for tunneled traffic 2) devlink rate objects support 1) From Chris Mi: Sample offload for tunneled traffic ===================================================== Background and solution ----------------------- Currently the sample offload actions send the encapsulated packet to software. This series de-capsulates the packet before performing the sampling and set the tunnel properties on the skb metadata fields to make the behavior consistent with OVS sFlow. If de-capsulating first, we can't use the same match like before in default table. So instantiate a post action instance to continue processing the action list. If HW can preserve reg_c, also use the post action instance. Post action infrastructure -------------------------- Some tc actions are modeled in hardware using multiple tables causing a tc action list split. For example, CT action is modeled by jumping to a ct table which is controlled by nf flow table. sFlow jumps in hardware to a sample table, which continues to a "default table" where it should continue processing the action list. Multi table actions are modeled in hardware using a unique fte_id. The fte_id is set before jumping to a table. Split actions continue to a post-action table where the matched fte_id value continues the execution the tc action list. This series also introduces post action infrastructure. Both ct and sample use it. Sample for tunnel in TC SW -------------------------- tc filter add dev vxlan1 protocol ip parent ffff: prio 3 \ flower src_mac 24:25:d0:e1:00:00 dst_mac 02:25:d0:13:01:02 \ enc_src_ip 192.168.1.14 enc_dst_ip 192.168.1.13 \ enc_dst_port 4789 enc_key_id 4 \ action sample rate 1 group 6 \ action tunnel_key unset \ action mirred egress redirect dev enp4s0f0_1 MLX5 sample HW offload ---------------------- For the following typical flow table: +-------------------------------+ + original flow table + +-------------------------------+ + original match + +-------------------------------+ + sample action + other actions + +-------------------------------+ We translate the tc filter with sample action to the following HW model: +---------------------+ + original flow table + +---------------------+ + original match + +---------------------+ | set fte_id (if reg_c preserve cap) | do decap v +------------------------------------------------+ + Flow Sampler Object + +------------------------------------------------+ + sample ratio + +------------------------------------------------+ + sample table id | default table id + +------------------------------------------------+ | | v v +-----------------------------+ +-------------------+ + sample table + + default table + +-----------------------------+ +-------------------+ + forward to management vport + | +-----------------------------+ | +-------+------+ | |reg_c preserve cap | |or decap action v v +-----------------+ +-------------+ + per vport table + + post action + +-----------------+ +-------------+ + original match + +-----------------+ + other actions + +-----------------+ 2) From Dmytro Linkin: devlink rate object support for mlx5_core driver ======================================================================= HIGH-LEVEL OVERVIEW Devlink leaf rate objects created per vport (VF/SF, and PF on BlueField) in switchdev mode on devlink port registration. Implement devlink ops callbacks to create/destroy rate groups, set TX rate values of the vport/group, assign vport to the group. Driver accepts TX rate values as fraction of 1Mbps. Refactor existing eswitch QoS infrastructure to be accessible by legacy NDO rate API and new devlink rate API. NDO rate API is not removed/disabled in switchdev mode to not break existing users. Rate values configured with NDO rate API are not visible for devlink infrastructure, therefore APIs should not be used simultaneously. IMPLEMENTATION DETAILS Driver provide two level rate hierarchy to manage bandwidth - group level and vport level. Initially each vport added to internal unlimited group created by default. Each rate element (vport or group) receive bandwidth relative to its parent element (for groups the parent is a physical link itself) in a Round Robin manner, where element get bandwidth value according to its weight. Example: Created four rate groups with tx_share limits: $ devlink port function rate add \ pci/0000:06:00.0/group_1 tx_share 30gbit $ devlink port function rate add \ pci/0000:06:00.0/group_2 tx_share 20gbit $ devlink port function rate add \ pci/0000:06:00.0/group_3 tx_share 20gbit $ devlink port function rate add \ pci/0000:06:00.0/group_4 tx_share 10gbit Weights created in HW for each group are relative to the bigest tx_share value, which is 30gbit: <group_1> 1.0 <group_2> 0.67 <group_3> 0.67 <group_4> 0.33 Assuming link speed is 50 Gbit/sec and each group can sustain such amount of traffic, maximum bandwidth is 50 / (1.0 + 0.67 + 0.67 + 0.33) = ~18.75 Gbit/sec. Normilized bandwidth values for groups: <group_1> 18.75 * 1.0 = 18.75 Gbit/sec <group_2> 18.75 * 0.67 = 12.5 Gbit/sec <group_3> 18.75 * 0.67 = 12.5 Gbit/sec <group_4> 18.75 * 0.33 = 6.25 Gbit/sec If in example above group_1 doesn't produce any traffic, then maximum bandwidth becomes 50 / (0.67 + 0.67 + 0.33) = ~30.0 Gbit/sec. Normalized values: <group_2> 30.0 * 0.67 = 20.0 Gbit/sec <group_3> 30.0 * 0.67 = 20.0 Gbit/sec <group_4> 30.0 * 0.33 = 10.0 Gbit/sec Same normalization applied to each vport in the group. Normalized values are internal, therefore driver provides QoS tracepoints for next events: * vport rate element creation/deletion: * vport rate element configuration; * group rate element creation/deletion; * group rate element configuration. PATCHES OVERVIEW 1 - Moving and isolation of eswitch QoS logic in separate file; 2 - Implement devlink leaf rate object support for vports; 3 - Implement rate groups creation/deletion; 4 - Implement TX rate management for the groups; 5 - Implement parent set for vports; 6 - Eswitch QoS tracepoints. ==================== Signed-off-by: -
David S. Miller authored
Merge tag 'for-net-next-2021-08-19' of git://git.kernel.org/pub/scm/linux/kernel/git/bluetooth/bluetooth-next Luiz Augusto von Dentz says: ==================== bluetooth-next pull request for net-next: - Add support for Foxconn Mediatek Chip - Add support for LG LGSBWAC92/TWCM-K505D - hci_h5 flow control fixes and suspend support - Switch to use lock_sock for SCO and RFCOMM - Various fixes for extended advertising - Reword Intel's setup on btusb unifying the supported generations ==================== 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 - update docs about move IRC channel away from freenode, by Sven Eckelmann - Switch to kstrtox.h for kstrtou64, by Sven Eckelmann - Update NULL checks, by Sven Eckelmann (2 patches) - remove remaining skb-copy calls for broadcast packets, by Linus Lüssing ==================== Signed-off-by: -
Linus Lüssing authored
We currently have two code paths for broadcast packets: A) self-generated, via batadv_interface_tx()-> batadv_send_bcast_packet(). B) received/forwarded, via batadv_recv_bcast_packet()-> batadv_forw_bcast_packet(). For A), self-generated broadcast packets: The only modifications to the skb data is the ethernet header which is added/pushed to the skb in batadv_send_broadcast_skb()->batadv_send_skb_packet(). However before doing so, batadv_skb_head_push() is called which calls skb_cow_head() to unshare the space for the to be pushed ethernet header. So for this case, it is safe to use skb clones. For B), received/forwarded packets: The same applies as in A) for the to be forwarded packets. Only the ethernet header is added. However after (queueing for) forwarding the packet in batadv_recv_bcast_packet()->batadv_forw_bcast_packet(), a packet is additionally decapsulated and is sent up the stack through batadv_recv_bcast_packet()->batadv_interface_rx(). Protocols higher up the stack are already required to check if the packet is shared and create a copy for further modifications. When the next (protocol) layer works correctly, it cannot happen that it tries to operate on the data behind the skb clone which is still queued up for forwarding. Co-authored-by:
Sven Eckelmann <sven@narfation.org> Signed-off-by:
Linus Lüssing <linus.luessing@c0d3.blue> Signed-off-by:
Simon Wunderlich <sw@simonwunderlich.de>
-
Sven Eckelmann authored
The check if a batman-adv related object is NULL or not is now directly in the batadv_*_put functions. It is not needed anymore to perform this check outside these function: The changes were generated using a coccinelle semantic patch: @@ expression E; @@ - if (likely(E != NULL)) ( batadv_backbone_gw_put | batadv_claim_put | batadv_dat_entry_put | batadv_gw_node_put | batadv_hardif_neigh_put | batadv_hardif_put | batadv_nc_node_put | batadv_nc_path_put | batadv_neigh_ifinfo_put | batadv_neigh_node_put | batadv_orig_ifinfo_put | batadv_orig_node_put | batadv_orig_node_vlan_put | batadv_softif_vlan_put | batadv_tp_vars_put | batadv_tt_global_entry_put | batadv_tt_local_entry_put | batadv_tt_orig_list_entry_put | batadv_tt_req_node_put | batadv_tvlv_container_put | batadv_tvlv_handler_put )(E); Signed-off-by:
-
Sven Eckelmann authored
The commit b37a4668 ("netdevice: add the case if dev is NULL") changed the way how the NULL check for net_devices have to be handled when trying to reduce its reference counter. Before this commit, it was the responsibility of the caller to check whether the object is NULL or not. But it was changed to behave more like kfree. Now the callee has to handle the NULL-case. The batman-adv code was scanned via cocinelle for similar places. These were changed to use the paradigm @@ identifier E, T, R, C; identifier put; @@ void put(struct T *E) { + if (!E) + return; kref_put(&E->C, R); } Functions which were used in other sources files were moved to the header to allow the compiler to inline the NULL check and the kref_put call. Signed-off-by:
-
Sven Eckelmann authored
The commit 4c527293 ("kernel.h: split out kstrtox() and simple_strtox() to a separate header") moved the kstrtou64 function to a new header called linux/kstrtox.h. Signed-off-by:
-
Sven Eckelmann authored
Due to recent developments around the Freenode.org IRC network, the opinions about the usage of this service shifted dramatically. The majority of the still active users of the #batman channel prefers a move to the hackint.org network. Signed-off-by:
-
Dmytro Linkin authored
Add tracepoints to log QoS enabling/disabling/configuration for vports and rate groups. Signed-off-by:
Dmytro Linkin <dlinkin@nvidia.com> Reviewed-by:
Huy Nguyen <huyn@nvidia.com> Reviewed-by:
Mark Bloch <mbloch@nvidia.com> Reviewed-by:
Saeed Mahameed <saeedm@nvidia.com>
-
Dmytro Linkin authored
Implement eswitch API that allows updating rate groups. If group pointer is NULL, then move the vport to internal unlimited group zero. Implement devlink_ops->rate_parent_node_set() callback in the terms of the new eswitch group update API. Enable QoS for all group's elements if a group has allocated BW share. Co-developed-by:
Vlad Buslov <vladbu@nvidia.com> Signed-off-by:
Parav Pandit <parav@nvidia.com> Reviewed-by:
-
Dmytro Linkin authored
Provide eswitch API to allow controlling group rate limits. Use it to implement devlink_ops->mlx5_devlink_rate_node_tx_{share|max}_set(). The share rate will create relative bandwidth share on the groups level while within the group the user can set shared rate on the member vports of that group and this rate will be relative to the group's share rate. The group with the highest shared rate will get a BW share of 100 and the rest of the groups will get a value that reflects the ratio between their share rate and the maximum share rate. Example: Created four rate groups with tx_share limits: $ devlink port function rate add \ pci/0000:06:00.0/group_1 tx_share 30gbit $ devlink port function rate add \ pci/0000:06:00.0/group_2 tx_share 20gbit $ devlink port function rate add \ pci/0000:06:00.0/group_3 tx_share 20gbit $ devlink port function rate add \ pci/0000:06:00.0/group_4 tx_share 10gbit Assuming link speed is 50 Gbit/sec ratio divider will be 50 / (30+20+20+10) = 0.625. Normalized rate values for the groups: <group_1> 30 * 0.625 = 18.75 Gbit/sec <group_2> 20 * 0.625 = 12.5 Gbit/sec <group_3> 20 * 0.625 = 12.5 Gbit/sec <group_4> 10 * 0.625 = 6.25 Gbit/sec Rate group with unlimited tx_share rate will receive minimum BW value (1Mbit/sec) if presented any group with tx_share rate limit. This allow to not drop all packets in case of heavy traffic. Co-developed-by: -
Dmytro Linkin authored
Extend eswitch API with rate limiting groups: - Define new struct mlx5_esw_rate_group that is used to hold all internal group data. - Implement functions that allow creation, destruction and cleanup of groups. - Assign all vports to internal unlimited zero group by default. This commit lays the groundwork for group rate limiting by implementing devlink_ops->rate_node_{new|del}() callbacks to support creating and deleting groups through devlink rate node objects. APIs that allows setting rates and adding/removing members are implemented in following patches. Co-developed-by: -
Dmytro Linkin authored
Register devlink rate leaf object for every eswitch vport. Implement devlink ops that enable setting shared and max tx rates through devlink API. Extract common eswitch code from existing tx rate set function that is accessed through NDO to be reused for the devlink. Values configured with NDO API are not visible for the devlink API, therefore shouldn't be used simultaneously. When normalizing the BW share value, dividing the desired minimum rate by the common divider results in losing information since the quotient is rounded down. This has a significant affect on configurations of low rate where the round down eliminates a large percentage of the total rate. To improve the formula, round up the division result to make sure that the BW share is at least the value it was supposed to be and won't lost a significant amount of the expected value. Co-developed-by:
-
Dmytro Linkin authored
Move eswitch QoS related code into dedicated file. Provide eswitch API to access this code meaning it is isolated and restricted to be used only by eswitch.c. Exception is legacy NDO vf set rate, which moved to esw/legacy.c. Signed-off-by:
-
Chris Mi authored
Currently the sample offload actions send the encapsulated packet to software. This commit decapsulates the packet before performing the sampling and set the tunnel properties on the skb metadata fields to make the behavior consistent with OVS sFlow. If decapsulating first, we can't use the same match like before in default table. So instantiate a post action instance to continue processing the action list. If HW can preserve reg_c, also use the post action instance. Signed-off-by:
Chris Mi <cmi@nvidia.com> Reviewed-by:
Oz Shlomo <ozsh@nvidia.com> Reviewed-by:
Roi Dayan <roid@nvidia.com> Reviewed-by:
-
Chris Mi authored
Currently the sample offload actions send the encapsulated packet to software. sFlow expects tunneled packets to be decapsulated while having the tunnel properties on the skb metadata fields. Reuse the functions used by connection tracking to map the outer header properties to a unique id. The next patch will use that id to restore the tunnel information of decapsulated packets onto the skb. Signed-off-by:
-
Chris Mi authored
CONFIG_NET_TC_SKB_EXT controls the SKB extension support for restoring chain ids. SKB extension is not required for tunnel restoration. Remove the CONFIG_NET_TC_SKB_EXT dependency as a pre-step for using the tunnel restore methods for sample offload use cases. Signed-off-by:
-
Chris Mi authored
Move post action table management to common library providing add/del/get API. Refactor the ct action offload to use the common API. Signed-off-by:
-
