| Message ID | 20250313190309.2545711-1-ameryhung@gmail.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | bpf qdisc | expand |
Amery Hung <ameryhung@gmail.com> writes: > Hi all, > > This patchset aims to support implementing qdisc using bpf struct_ops. > This version takes a step back and only implements the minimum support > for bpf qdisc. 1) support of adding skb to bpf_list and bpf_rbtree > directly and 2) classful qdisc are deferred to future patchsets. In > addition, we only allow attaching bpf qdisc to root or mq for now. > This is to prevent accidentally breaking exisiting classful qdiscs > that rely on data in a child qdisc. This limit may be lifted in the > future after careful inspection. > > * Overview * > > This series supports implementing qdisc using bpf struct_ops. bpf qdisc > aims to be a flexible and easy-to-use infrastructure that allows users to > quickly experiment with different scheduling algorithms/policies. It only > requires users to implement core qdisc logic using bpf and implements the > mundane part for them. In addition, the ability to easily communicate > between qdisc and other components will also bring new opportunities for > new applications and optimizations. > > * struct_ops changes * > > To make struct_ops works better with bpf qdisc, two new changes are > introduced to bpf specifically for struct_ops programs. Frist, we > introduce "__ref" postfix for arguments in stub functions in patch 1-2. > It allows Qdisc_ops->enqueue to acquire an unique referenced kptr to the > skb argument. Through the reference object tracking mechanism in > the verifier, we can make sure that the acquired skb will be either > enqueued or dropped. Besides, no duplicate references can be acquired. > Then, we allow a referenced kptr to be returned from struct_ops programs > so that we can return an skb naturally. This is done and tested in patch 3 > and 4. > > * Performance of bpf qdisc * > > This patchset includes two qdisc examples, bpf_fifo and bpf_fq, for > __testing__ purposes. For performance test, we compare selftests and their > kernel counterparts to give you a sense of the performance of qdisc > implemented in bpf. > > The implementation of bpf_fq is fairly complex and slightly different from > fq so later we only compare the two fifo qdiscs. bpf_fq implements a > scheduling algorithm similar to fq before commit 29f834aa326e ("net_sched: > sch_fq: add 3 bands and WRR scheduling") was introduced. bpf_fifo uses a > single bpf_list as a queue instead of three queues for different > priorities in pfifo_fast. The time complexity of fifo however should be > similar since the queue selection time is negligible. > > Test setup: > > client -> qdisc -------------> server > ~~~~~~~~~~~~~~~ ~~~~~~ > nested VM1 @ DC1 VM2 @ DC2 > > Throghput: iperf3 -t 600, 5 times > > Qdisc Average (GBits/sec) > ---------- ------------------- > pfifo_fast 12.52 ± 0.26 > bpf_fifo 11.72 ± 0.32 > fq 10.24 ± 0.13 > bpf_fq 11.92 ± 0.64 > > Latency: sockperf pp --tcp -t 600, 5 times > > Qdisc Average (usec) > ---------- -------------- > pfifo_fast 244.58 ± 7.93 > bpf_fifo 244.92 ± 15.22 > fq 234.30 ± 19.25 > bpf_fq 221.34 ± 10.76 > > Looking at the two fifo qdiscs, the 6.4% drop in throughput in the bpf > implementatioin is consistent with previous observation (v8 throughput > test on a loopback device). This should be able to be mitigated by > supporting adding skb to bpf_list or bpf_rbtree directly in the future. > > * Clean up skb in bpf qdisc during reset * > > The current implementation relies on bpf qdisc implementors to correctly > release skbs in queues (bpf graphs or maps) in .reset, which might not be > a safe thing to do. The solution as Martin has suggested would be > supporting private data in struct_ops. This can also help simplifying > implementation of qdisc that works with mq. For examples, qdiscs in the > selftest mostly use global data. Therefore, even if user add multiple > qdisc instances under mq, they would still share the same queue. Very cool to see this progress! Are you aware that the series has a mix of commit author email addresses (mixing your bytedance.com and gmail addresses)? Otherwise, for the series: Acked-by: Toke Høiland-Jørgensen <toke@redhat.com>
On Fri, Mar 14, 2025 at 3:52 AM Toke Høiland-Jørgensen <toke@redhat.com> wrote: > > Amery Hung <ameryhung@gmail.com> writes: > > > Hi all, > > > > This patchset aims to support implementing qdisc using bpf struct_ops. > > This version takes a step back and only implements the minimum support > > for bpf qdisc. 1) support of adding skb to bpf_list and bpf_rbtree > > directly and 2) classful qdisc are deferred to future patchsets. In > > addition, we only allow attaching bpf qdisc to root or mq for now. > > This is to prevent accidentally breaking exisiting classful qdiscs > > that rely on data in a child qdisc. This limit may be lifted in the > > future after careful inspection. > > Very cool to see this progress! > > Are you aware that the series has a mix of commit author email addresses > (mixing your bytedance.com and gmail addresses)? > Yes. I have changed my affiliation so some recently added patches use another email. Thanks for reviewing! Amery > Otherwise, for the series: > > Acked-by: Toke Høiland-Jørgensen <toke@redhat.com> >
Hi all, This patchset aims to support implementing qdisc using bpf struct_ops. This version takes a step back and only implements the minimum support for bpf qdisc. 1) support of adding skb to bpf_list and bpf_rbtree directly and 2) classful qdisc are deferred to future patchsets. In addition, we only allow attaching bpf qdisc to root or mq for now. This is to prevent accidentally breaking exisiting classful qdiscs that rely on data in a child qdisc. This limit may be lifted in the future after careful inspection. * Overview * This series supports implementing qdisc using bpf struct_ops. bpf qdisc aims to be a flexible and easy-to-use infrastructure that allows users to quickly experiment with different scheduling algorithms/policies. It only requires users to implement core qdisc logic using bpf and implements the mundane part for them. In addition, the ability to easily communicate between qdisc and other components will also bring new opportunities for new applications and optimizations. * struct_ops changes * To make struct_ops works better with bpf qdisc, two new changes are introduced to bpf specifically for struct_ops programs. Frist, we introduce "__ref" postfix for arguments in stub functions in patch 1-2. It allows Qdisc_ops->enqueue to acquire an unique referenced kptr to the skb argument. Through the reference object tracking mechanism in the verifier, we can make sure that the acquired skb will be either enqueued or dropped. Besides, no duplicate references can be acquired. Then, we allow a referenced kptr to be returned from struct_ops programs so that we can return an skb naturally. This is done and tested in patch 3 and 4. * Performance of bpf qdisc * This patchset includes two qdisc examples, bpf_fifo and bpf_fq, for __testing__ purposes. For performance test, we compare selftests and their kernel counterparts to give you a sense of the performance of qdisc implemented in bpf. The implementation of bpf_fq is fairly complex and slightly different from fq so later we only compare the two fifo qdiscs. bpf_fq implements a scheduling algorithm similar to fq before commit 29f834aa326e ("net_sched: sch_fq: add 3 bands and WRR scheduling") was introduced. bpf_fifo uses a single bpf_list as a queue instead of three queues for different priorities in pfifo_fast. The time complexity of fifo however should be similar since the queue selection time is negligible. Test setup: client -> qdisc -------------> server ~~~~~~~~~~~~~~~ ~~~~~~ nested VM1 @ DC1 VM2 @ DC2 Throghput: iperf3 -t 600, 5 times Qdisc Average (GBits/sec) ---------- ------------------- pfifo_fast 12.52 ± 0.26 bpf_fifo 11.72 ± 0.32 fq 10.24 ± 0.13 bpf_fq 11.92 ± 0.64 Latency: sockperf pp --tcp -t 600, 5 times Qdisc Average (usec) ---------- -------------- pfifo_fast 244.58 ± 7.93 bpf_fifo 244.92 ± 15.22 fq 234.30 ± 19.25 bpf_fq 221.34 ± 10.76 Looking at the two fifo qdiscs, the 6.4% drop in throughput in the bpf implementatioin is consistent with previous observation (v8 throughput test on a loopback device). This should be able to be mitigated by supporting adding skb to bpf_list or bpf_rbtree directly in the future. * Clean up skb in bpf qdisc during reset * The current implementation relies on bpf qdisc implementors to correctly release skbs in queues (bpf graphs or maps) in .reset, which might not be a safe thing to do. The solution as Martin has suggested would be supporting private data in struct_ops. This can also help simplifying implementation of qdisc that works with mq. For examples, qdiscs in the selftest mostly use global data. Therefore, even if user add multiple qdisc instances under mq, they would still share the same queue. --- v5: * Rebase to bpf-next/master * Remove prerequisite bpf patches that has landed * Add Acked-by from Cong v4: * Rebase to bpf-next/master * Move the root/mq attaching check to bpf_qdisc.c (now patch 15) * Replace netdevsim with veth for attaching mq v3: * Rebase to bpf-next/master * Remove the requirement in the verifier that "__ref arguments must acquire ref_obj_id first" by making each prog keeping a copy of arg_ctx_info and saving ref_obj_id in it. * Generalize prog_ops_moff() to work with any struct_op (now called bpf_struct_ops_prog_moff()) * Use bpf_struct_ops_prog_moff(prog) instead of prog->aux->attach_func_name to infer the ops of a program * Limit attach to root and mq for now and add corresponding selftests * Simplify qdisc selftests with network_helper * Fix fq_remove_flow() not deleting the stashed flow v2: Rebase to bpf-next/master Patch 1-4 Remove the use of ctx_arg_info->ref_obj_id when acquiring referenced kptr from struct_ops arg Improve type comparison when checking kptr return from struct_ops Simplify selftests with test_loader and nomerge attribute Patch 5 Remove redundant checks in qdisc_init Disallow tail_call Patch 6 Improve kfunc ops availabilty filter by i) Checking struct_ops->type ii) Defining op-specific kfunc set Patch 7 Search and add bpf_kfunc_desc after gen_prologue/epilogue Patch 8 Use gen_prologue/epilogue to init/cancel watchdog timer Patch 12 Mark read-only func arg and struct member const in libbpf Link: https://lore.kernel.org/bpf/20241220195619.2022866-1-amery.hung@gmail.com/ v1: Fix struct_ops referenced kptr acquire/return mechanisms Allow creating dynptr from skb Add bpf qdisc kfunc filter Support updating bstats and qstats Update qdiscs in selftest to update stats Add gc, handle hash collision and fix bugs in fq_bpf Link: https://lore.kernel.org/bpf/20241213232958.2388301-1-amery.hung@bytedance.com/ past RFCs v9: Drop classful qdisc operations and kfuncs Drop support of enqueuing skb directly to bpf_rbtree/list Link: https://lore.kernel.org/bpf/20240714175130.4051012-1-amery.hung@bytedance.com/ v8: Implement support of bpf qdisc using struct_ops Allow struct_ops to acquire referenced kptr via argument Allow struct_ops to release and return referenced kptr Support enqueuing sk_buff to bpf_rbtree/list Move examples from samples to selftests Add a classful qdisc selftest Link: https://lore.kernel.org/netdev/20240510192412.3297104-15-amery.hung@bytedance.com/ v7: Reference skb using kptr to sk_buff instead of __sk_buff Use the new bpf rbtree/link to for skb queues Add reset and init programs Add a bpf fq qdisc sample Add a bpf netem qdisc sample Link: https://lore.kernel.org/netdev/cover.1705432850.git.amery.hung@bytedance.com/ v6: switch to kptr based approach v5: mv kernel/bpf/skb_map.c net/core/skb_map.c implement flow map as map-in-map rename bpf_skb_tc_classify() and move it to net/sched/cls_api.c clean up eBPF qdisc program context v4: get rid of PIFO, use rbtree directly v3: move priority queue from sch_bpf to skb map introduce skb map and its helpers introduce bpf_skb_classify() use netdevice notifier to reset skb's Rebase on latest bpf-next v2: Rebase on latest net-next Make the code more complete (but still incomplete) Amery Hung (13): bpf: Prepare to reuse get_ctx_arg_idx bpf: Generalize finding member offset of struct_ops prog bpf: net_sched: Support implementation of Qdisc_ops in bpf bpf: net_sched: Add basic bpf qdisc kfuncs bpf: net_sched: Add a qdisc watchdog timer bpf: net_sched: Support updating bstats bpf: net_sched: Support updating qstats bpf: net_sched: Allow writing to more Qdisc members bpf: net_sched: Disable attaching bpf qdisc to non root libbpf: Support creating and destroying qdisc selftests/bpf: Add a basic fifo qdisc test selftests/bpf: Add a bpf fq qdisc to selftest selftests/bpf: Test attaching bpf qdisc to mq and non root include/linux/bpf.h | 2 + include/linux/btf.h | 1 + kernel/bpf/bpf_struct_ops.c | 15 + kernel/bpf/btf.c | 6 +- net/ipv4/bpf_tcp_ca.c | 23 +- net/sched/Kconfig | 12 + net/sched/Makefile | 1 + net/sched/bpf_qdisc.c | 465 ++++++++++++ net/sched/sch_api.c | 7 +- net/sched/sch_generic.c | 3 +- tools/lib/bpf/libbpf.h | 5 +- tools/lib/bpf/netlink.c | 20 +- tools/testing/selftests/bpf/config | 2 + .../selftests/bpf/prog_tests/bpf_qdisc.c | 178 +++++ .../selftests/bpf/progs/bpf_qdisc_common.h | 27 + .../selftests/bpf/progs/bpf_qdisc_fifo.c | 117 +++ .../selftests/bpf/progs/bpf_qdisc_fq.c | 718 ++++++++++++++++++ 17 files changed, 1570 insertions(+), 32 deletions(-) create mode 100644 net/sched/bpf_qdisc.c create mode 100644 tools/testing/selftests/bpf/prog_tests/bpf_qdisc.c create mode 100644 tools/testing/selftests/bpf/progs/bpf_qdisc_common.h create mode 100644 tools/testing/selftests/bpf/progs/bpf_qdisc_fifo.c create mode 100644 tools/testing/selftests/bpf/progs/bpf_qdisc_fq.c