On 8/24/2021 3:33 AM, Jason Gunthorpe wrote:
On Wed, Aug 18, 2021 at 02:24:18PM +0300, Mark Zhang wrote:
Hi,
This series from Aharon and Neta provides an extension to the rdma
statistics tool that allows to add and remove optional counters
dynamically, using new netlink commands.
The idea of having optional counters is to provide to the users the
ability to get statistics of counters that hurts performance.
Once an optional counter was added, its statistics will be presented
along with all the counters, using the show command.
Binding objects to the optional counters is currently not supported,
neither in auto mode nor in manual mode.
To get the list of optional counters that are supported on this device,
use "rdma statistic mode supported". To see which counters are currently
enabled, use "rdma statistic mode".
$ rdma statistic mode supported
link rocep8s0f0/1
Optional-set: cc_rx_ce_pkts cc_rx_cnp_pkts cc_tx_cnp_pkts
link rocep8s0f1/1
Optional-set: cc_rx_ce_pkts cc_rx_cnp_pkts cc_tx_cnp_pkts
$ sudo rdma statistic add link rocep8s0f0/1 optional-set cc_rx_ce_pkts
$ rdma statistic mode
link rocep8s0f0/1
Optional-set: cc_rx_ce_pkts
$ sudo rdma statistic add link rocep8s0f0/1 optional-set cc_tx_cnp_pkts
$ rdma statistic mode
link rocep8s0f0/1
Optional-set: cc_rx_ce_pkts cc_tx_cnp_pkts
This doesn't look like the right output to iproute to me, the two
command should not be using the same tag and the output of iproute
should always be formed to be valid input to iproute
So it should be like this:
$ rdma statistic mode supported
link rocep8s0f0/1 optional-set cc_rx_ce_pkts cc_rx_cnp_pkts cc_tx_cnp_pkts
link rocep8s0f1/1 optional-set cc_rx_ce_pkts cc_rx_cnp_pkts cc_tx_cnp_pkts
$ sudo rdma statistic add link rocep8s0f0/1 optional-set cc_rx_ce_pkts
$ rdma statistic mode
link rocep8s0f0/1 optional-set cc_rx_ce_pkts
$ sudo rdma statistic add link rocep8s0f0/1 optional-set cc_tx_cnp_pkts
$ rdma statistic mode
link rocep8s0f0/1 optional-set cc_rx_ce_pkts cc_tx_cnp_pkts
$ rdma statistic show link rocep8s0f0/1
link rocep8s0f0/1 rx_write_requests 0 rx_read_requests 0 rx_atomic_requests 0 out_of_buffer 0
out_of_sequence 0 duplicate_request 0 rnr_nak_retry_err 0 packet_seq_err 0 implied_nak_seq_err 0
local_ack_timeout_err 0 resp_local_length_error 0 resp_cqe_error 0 req_cqe_error 0
req_remote_invalid_request 0 req_remote_access_errors 0 resp_remote_access_errors 0
resp_cqe_flush_error 0 req_cqe_flush_error 0 roce_adp_retrans 0 roce_adp_retrans_to 0
roce_slow_restart 0 roce_slow_restart_cnps 0 roce_slow_restart_trans 0 rp_cnp_ignored 0
rp_cnp_handled 0 np_ecn_marked_roce_packets 0 np_cnp_sent 0 rx_icrc_encapsulated 0
Optional-set: cc_rx_ce_pkts 0 cc_tx_cnp_pkts 0
Also looks bad, optional counters should not be marked specially at
this point.
Will put optional counters in the last, like this:
$ rdma statistic show link rocep8s0f0/1
link rocep8s0f0/1 rx_write_requests 0 rx_read_requests 0
rx_atomic_requests 0 out_of_buffer 0
out_of_sequence 0 duplicate_request 0 rnr_nak_retry_err 0 packet_seq_err
0 implied_nak_seq_err 0
local_ack_timeout_err 0 resp_local_length_error 0 resp_cqe_error 0
req_cqe_error 0
req_remote_invalid_request 0 req_remote_access_errors 0
resp_remote_access_errors 0
resp_cqe_flush_error 0 req_cqe_flush_error 0 roce_adp_retrans 0
roce_adp_retrans_to 0
roce_slow_restart 0 roce_slow_restart_cnps 0 roce_slow_restart_trans 0
rp_cnp_ignored 0
rp_cnp_handled 0 np_ecn_marked_roce_packets 0 np_cnp_sent 0
rx_icrc_encapsulated 0 cc_rx_ce_pkts 0 cc_tx_cnp_pkts 0
Aharon Landau (9):
net/mlx5: Add support in bth_opcode as a match criteria
net/mlx5: Add priorities for counters in RDMA namespaces
RDMA/counters: Support to allocate per-port optional counter
statistics
RDMA/mlx5: Add alloc_op_port_stats() support
RDMA/mlx5: Add steering support in optional flow counters
RDMA/nldev: Add support to add and remove optional counters
RDMA/mlx5: Add add_op_stat() and remove_op_stat() support
RDMA/mlx5: Add get_op_stats() support
RDMA/nldev: Add support to get current enabled optional counters
Neta Ostrovsky (1):
RDMA/nldev: Add support to get optional counters statistics
This series is in a poor order, all the core update should come first
and the commit messages should explain what is going on when building
out the new APIs.
The RDMA/mlx5 patches can go last
Will fix it, thanks Jason.
Jason
