- 07 Mar, 2018 16 commits
-
-
Yossi Kuperman authored
Current hardware decrypts and authenticates incoming ESP packets. Subsequently, the software extracts the nexthdr field, truncates the trailer and adjusts csum accordingly. With this patch and a capable device, the trailer is being removed by the hardware and the nexthdr field is conveyed via PET. This way we avoid both the need to access the trailer (cache miss) and to compute its relative checksum, which significantly improve the performance. Experiment shows that trailer removal improves the performance by 2Gbps, (netperf). Both forwarding and host-to-host configurations. Signed-off-by:
Yossi Kuperman <yossiku@mellanox.com> Signed-off-by:
Aviad Yehezkel <aviadye@mellanox.com> Signed-off-by:
Saeed Mahameed <saeedm@mellanox.com>
-
Yossi Kuperman authored
The current code assume only SA QP commands. Refactor in order to pave the way for new QP commands: 1. Generic cmd response format. 2. SA cmd checks are in dedicated functions. 3. Aligned debug prints. Signed-off-by:
-
Saeed Mahameed authored
Fix build break of mlx5_accel_ipsec_device_caps is not defined when MLX5_ACCEL is not selected, use MLX5_IPSEC_DEV instead which handles such case. Signed-off-by:
Doug Ledford <dledford@redhat.com>
-
Aviad Yehezkel authored
Previously, deleting a flow steering entry only got the index. Since the FPGA implementation of FTE's deletion might need to dig inside the FTE itself, we would like to get the FTE's context. Changing the interface to pass the FTE context. Signed-off-by:
Matan Barak <matanb@mellanox.com> Signed-off-by:
-
Boris Pismenny authored
Add helper functions that check if a protocol is part of a flow steering match criteria. Signed-off-by:
Boris Pismenny <borisp@mellanox.com> Signed-off-by:
-
Matan Barak authored
fte objects contain the match value and action. Currently, extending the actions require in adding them both to the API and fs_fte. Signed-off-by:
-
Aviad Yehezkel authored
Currently, we don't support egress flow steering namespace in mlx5 flow steering core implementation. However, when we want to encrypt a packet, we model it as a flow steering rule in the egress path. To overcome this, we add an empty egress namespace to flow steering. This namespace is initialized only when ipsec support exists. In the future, this will grow to a full blown full steering implementation, resembling the ingress path. Signed-off-by:
-
Matan Barak authored
The shim layer allows each namespace to define possibly different functionality for add/delete/update commands. The shim layer introduced here, will be used to support flow steering with the FPGA. Signed-off-by:
-
Matan Barak authored
The has_tag member will indicate whether a tag action was specified in flow specification. A flow tag 0 = MLX5_FS_DEFAULT_FLOW_TAG is assumed a valid flow tag that is currently used by mlx5 RDMA driver, whereas in HW flow_tag = 0 means that the user doesn't care about flow_tag. HW always provide a flow_tag = 0 if all flow tags requested on a specific flow are 0. So we need a way (in the driver) to differentiate between a user really requesting flow_tag = 0 and a user who does not care, in order to be able to report conflicting flow tags on a specific flow. Signed-off-by:
-
Boris Pismenny authored
Group and pass all function arguments of parse_flow_attr call in one common struct mlx5_flow_act. This patch passes all the action arguments of parse_flow_attr in one common struct mlx5_flow_act. It allows us to scale the number of actions without adding new arguments to the function. Signed-off-by:
Jason Gunthorpe <jgg@mellanox.com>
-
Matan Barak authored
Some flow steering namespace initialization (i.e. egress namespace) might depend on FPGA capabilities. Changing the initialization order such that the FPGA will be initialized before flow steering. Flow steering fs cmds initialization might depend on IPSec capabilities. Changing the initialization order such that the IPSec will be initialized before flow steering as well. Signed-off-by:
-
Aviad Yehezkel authored
This is already done by xfrm layer between state_dev_del callback to state_dev_free callback. Signed-off-by:
-
Aviad Yehezkel authored
We can't allocate with GFP_KERNEL inside spinlock. Actually ida_simple doesn't require spinlock so remove it. Fixes: 547eede0 ("net/mlx5e: IPSec, Innova IPSec offload infrastructure") Signed-off-by:
-
Aviad Yehezkel authored
Generally, FPGA IPSec commands must always complete. We want to wait for one minute for them to complete gracefully also when killing a process. Signed-off-by:
-
Aviad Yehezkel authored
IPSec init and cleanup functions also depends on linux/mlx5/driver.h. Signed-off-by:
-
Aviad Yehezkel authored
Signed-off-by:
-
- 23 Feb, 2018 15 commits
-
-
Mark Bloch authored
When in switchdev mode, there is no need to do self loopback checks as we can't receive those packets, we insert steering rules to the eswitch that make sure packets can't be looped back. Signed-off-by:
Mark Bloch <markb@mellanox.com> Reviewed-by:
Or Gerlitz <ogerlitz@mellanox.com> Signed-off-by:
Leon Romanovsky <leon@kernel.org> Signed-off-by:
-
Mark Bloch authored
Up until this point it wasn't possible to activate IB representors when switching to switchdev mode, remove this limitation. We trigger reload of the PF IB interface in order to make sure that already allocated resources are invalid and new resources will be opened correctly with all the limitations of switchdev mode applied (only raw packet capabilities, without RoCE). We also move the remove/add to a place where the E-Switch mode is set/unset to better control when to trigger this action, this will allow the IB side to start in the correct mode. For better code reuse, create a function which reloads an interface and export it. Signed-off-by:
-
Mark Bloch authored
This commit adds full support for IB representor: 1) Representors profile, We add two new profiles: nic_rep_profile - This profile will be used to create an IB device that represents the PF/UPLINK. rep_profile - This profile will be used to create an IB device that represents VFs. Each VF will be its own representor. 2) Proper load/unload callbacks, Those are called by the E-Switch when moving to/from switchdev mode. 3) Different flow DB handling for when we in switchdev mode. Signed-off-by:
-
Mark Bloch authored
In order to forward traffic from representor's SQ to the right virtual function, every time an SQ is created also add the corresponding flow rule to the FDB. Signed-off-by:
-
Mark Bloch authored
When enabling many VFs and switching to switchdev mode, the total amount of mkeys we try to allocate when loading representors is very large and may cause timeouts on allocations, the same issues was observed on VFs and we employ the same fix that was done for them. We avoid allocating the full MR cache on load but still allow it to be manipulated once the IB device is loaded. Signed-off-by:
-
Mark Bloch authored
Currently in switchdev mode we allow only for raw packet QPs. Expose the right capabilities and set the gid table length to 0, also make sure we don't try to enable RoCE, so split the function to enable RoCE so representors can enable only the notifier needed for net device events. Signed-off-by:
-
Mark Bloch authored
Currently we listen to netdev register/unregister event based on PCI device. When in switchdev mode PF and representors share the same PCI device, so in order to pair ib device and netdev in switchdev mode compare the netdev that triggered the event to that of the representor. Expose a function that lets you receive the netdev associated what a given representor. Signed-off-by:
-
Mark Bloch authored
When we point to a representor, it means we are in switchdev mode. The flow db is shared between PF and virtual function representors so each rule created needs to have a match on its specific source port. Signed-off-by:
-
Mark Bloch authored
A flow DB is a shared resource between PF and representors, need to allocate it only when creating the PF IB device. Once we add IB representors, they will use the flow db which was created by the PF. Signed-off-by:
-
Mark Bloch authored
Create the basic infrastructure of registering and unregistering IB representors. The load/unload callbacks are left empty and proper implementation will be introduced in following patches. Signed-off-by:
-
Mark Bloch authored
Create a new representor type: REP_IB. which will be initialized by an IB device that is used as a logical representor of a eswitch vport (VF or uplink) just like we have a net device today in switchdev mode. Signed-off-by:
-
Mark Bloch authored
Under switchdev mode we insert an eswitch miss rule causing any unmatched traffic to be sent towards the PF vport. This miss rule can be optimized if we break it to two, one case is for multicast traffic and the other for unicast. Breaking the miss rule into two (unicast and multicast) allows the firmware to program the hardware in a more efficient way. Using ConncetX-5 Ex with IXIA and testpmd (which use IB representors): IXIA -> NIC -> PF -> IB representor -> NIC -> VF: - Without this optimization: 9.2 MPPS. - With this optimization: 18 MPPS. VF -> NIC -> IB representor-> PF -> NIC -> IXIA: - Without this optimization: 17 MPPS. - With this optimization: 23.4 MPPS. Signed-off-by: -
Mark Bloch authored
The max FTE number should be the max number of SQs that can be opened. Ethernet representors open one SQ each. Once we add IB representor this will increase (depends on the user). For now lets start with 31 per IB representor and if needed increase in the future. This increase only affects the number of FTEs in the slow path FDB, offloaded rules (done via TC on the fast path portion of the FDB) aren't affected. Signed-off-by:
-
Mark Bloch authored
In preparation for IB representors, move representors structs to a global scope, also expose functions needed for registration, unregistration, eswitch mode and creating a flow rule to direct traffic from SQs to the right VF. Signed-off-by:
-
Mark Bloch authored
Add a callback interface to get a protocol device (per representor type). The Ethernet representors will expose their netdev via this interface. This functionality can be later used by IB representor in order to find the corresponding net device representor. Signed-off-by:
-
- 15 Feb, 2018 7 commits
-
-
Yonatan Cohen authored
The current implementation of create CQ requires contiguous memory, such requirement is problematic once the memory is fragmented or the system is low in memory, it causes for failures in dma_zalloc_coherent(). This patch implements new scheme of fragmented CQ to overcome this issue by introducing new type: 'struct mlx5_frag_buf_ctrl' to allocate fragmented buffers, rather than contiguous ones. Base the Completion Queues (CQs) on this new fragmented buffer. It fixes following crashes: kworker/29:0: page allocation failure: order:6, mode:0x80d0 CPU: 29 PID: 8374 Comm: kworker/29:0 Tainted: G OE 3.10.0 Workqueue: ib_cm cm_work_handler [ib_cm] Call Trace: [<>] dump_stack+0x19/0x1b [<>] warn_alloc_failed+0x110/0x180 [<>] __alloc_pages_slowpath+0x6b7/0x725 [<>] __alloc_pages_nodemask+0x405/0x420 [<>] dma_generic_alloc_coherent+0x8f/0x140 [<>] x86_swiotlb_alloc_coherent+0x21/0x50 [<>] mlx5_dma_zalloc_coherent_node+0xad/0x110 [mlx5_core] [<>] ? mlx5_db_alloc_node+0x69/0x1b0 [mlx5_core] [<>] mlx5_buf_alloc_node+0x3e/0xa0 [mlx5_core] [<>] mlx5_buf_alloc+0x14/0x20 [mlx5_core] [<>] create_cq_kernel+0x90/0x1f0 [mlx5_ib] [<>] mlx5_ib_create_cq+0x3b0/0x4e0 [mlx5_ib] Signed-off-by:
Yonatan Cohen <yonatanc@mellanox.com> Reviewed-by:
Tariq Toukan <tariqt@mellanox.com> Signed-off-by:
-
Saeed Mahameed authored
EQ structure and API is private to mlx5_core driver only, external drivers should not have access or the means to manipulate EQ objects. Remove redundant exports and move API functions out of the linux/mlx5 include directory into the driver's mlx5_core.h private include file. Signed-off-by:
Gal Pressman <galp@mellanox.com>
-
Saeed Mahameed authored
Since CQ tree is now per EQ, CQ completion and event forwarding became specific implementation of EQ logic, this patch moves that logic to eq.c and makes those functions static. Signed-off-by:
-
Saeed Mahameed authored
Now as the CQ table is per EQ, add an API to hold/put CQ to be used from eq.c in downstream patch. Signed-off-by:
-
Saeed Mahameed authored
Add API to add/del CQ to/from EQs CQ table to be used in cq.c upon CQ creation/destruction, as CQ table is now private to eq.c. Signed-off-by:
-
Saeed Mahameed authored
If a hardware event is targeting a CQ, that CQ should exist. Add unlikely to error handling flows. Signed-off-by:
-
Saeed Mahameed authored
Before this patch the driver had one CQ database protected via one spinlock, this spinlock is meant to synchronize between CQ adding/removing and CQ IRQ interrupt handling. On a system with large number of CPUs and on a work load that requires lots of interrupts, this global spinlock becomes a very nasty hotspot and introduces a contention between the active cores, which will significantly hurt performance and becomes a bottleneck that prevents seamless cpu scaling. To solve this we simply move the CQ database and its spinlock to be per EQ (IRQ), thus per core. Tested with: system: 2 sockets, 14 cores per socket, hyperthreading, 2x14x2=56 cores netperf command: ./super_netperf 200 -P 0 -t TCP_RR -H <server> -l 30 -- -r 300,300 -o -s 1M,1M -S 1M,1M WITHOUT THIS PATCH: Average: CPU %usr %nice %sys %iowait %irq %soft %steal %guest %gnice %idle Average: all 4.32 0.00 36.15 0.09 0.00 34.02 0.00 0.00 0.00 25.41 Samples: 2M of event 'cycles:pp', Event count (approx.): 1554616897271 Overhead Command Shared Object Symbol + 14.28% swapper [kernel.vmlinux] [k] intel_idle + 12.25% swapper [kernel.vmlinux] [k] queued_spin_lock_slowpath + 10.29% netserver [kernel.vmlinux] [k] queued_spin_lock_slowpath + 1.32% netserver [kernel.vmlinux] [k] mlx5e_xmit WITH THIS PATCH: Average: CPU %usr %nice %sys %iowait %irq %soft %steal %guest %gnice %idle Average: all 4.27 0.00 34.31 0.01 0.00 18.71 0.00 0.00 0.00 42.69 Samples: 2M of event 'cycles:pp', Event count (approx.): 1498132937483 Overhead Command Shared Object Symbol + 23.33% swapper [kernel.vmlinux] [k] intel_idle + 1.69% netserver [kernel.vmlinux] [k] mlx5e_xmit Tested-by:
Song Liu <songliubraving@fb.com> Signed-off-by:
-
- 11 Feb, 2018 2 commits
-
-
Linus Torvalds authored
-
Al Viro authored
except, again, POLLFREE and POLL_BUSY_LOOP. With this, we finally get to the promised end result: - POLL{IN,OUT,...} are plain integers and *not* in __poll_t, so any stray instances of ->poll() still using those will be caught by sparse. - eventpoll.c and select.c warning-free wrt __poll_t - no more kernel-side definitions of POLL... - userland ones are visible through the entire kernel (and used pretty much only for mangle/demangle) - same behavior as after the first series (i.e. sparc et.al. epoll(2) working correctly). Signed-off-by:Al Viro <viro@zeniv.linux.org.uk> Signed-off-by:
Linus Torvalds <torvalds@linux-foundation.org>
-
