> > This series introduces event-mode additions to ipsec-secgw. > > With this series, ipsec-secgw would be able to run in eventmode. The > worker thread (executing loop) would be receiving events and would be > submitting it back to the eventdev after the processing. This way, > multicore scaling and h/w assisted scheduling is achieved by making use > of the eventdev capabilities. > > Since the underlying event device would be having varying capabilities, > the worker thread could be drafted differently to maximize performance. > This series introduces usage of multiple worker threads, among which the > one to be used will be determined by the operating conditions and the > underlying device capabilities. > > For example, if an event device - eth device pair has Tx internal port, > then application can do tx_adapter_enqueue() instead of regular > event_enqueue(). So a thread making an assumption that the device pair > has internal port will not be the right solution for another pair. The > infrastructure added with these patches aims to help application to have > multiple worker threads, there by extracting maximum performance from > every device without affecting existing paths/use cases. > > The eventmode configuration is predefined. All packets reaching one eth > port will hit one event queue. All event queues will be mapped to all > event ports. So all cores will be able to receive traffic from all ports. > When schedule_type is set as RTE_SCHED_TYPE_ORDERED/ATOMIC, event device > will ensure the ordering. Ordering would be lost when tried in PARALLEL. > > Following command line options are introduced, > > --transfer-mode: to choose between poll mode & event mode > --event-schedule-type: to specify the scheduling type > (RTE_SCHED_TYPE_ORDERED/ > RTE_SCHED_TYPE_ATOMIC/ > RTE_SCHED_TYPE_PARALLEL) > > Additionally the event mode introduces two modes of processing packets: > > Driver-mode: This mode will have bare minimum changes in the application > to support ipsec. There woudn't be any lookup etc done in > the application. And for inline-protocol use case, the > thread would resemble l2fwd as the ipsec processing would be > done entirely in the h/w. This mode can be used to benchmark > the raw performance of the h/w. All the application side > steps (like lookup) can be redone based on the requirement > of the end user. Hence the need for a mode which would > report the raw performance. > > App-mode: This mode will have all the features currently implemented with > ipsec-secgw (non librte_ipsec mode). All the lookups etc > would follow the existing methods and would report numbers > that can be compared against regular ipsec-secgw benchmark > numbers. > > The driver mode is selected with existing --single-sa option > (used also by poll mode). When --single-sa option is used > in conjution with event mode then index passed to --single-sa > is ignored. > > Example commands to execute ipsec-secgw in various modes on OCTEON TX2 > platform, > > #Inbound and outbound app mode > ipsec-secgw -w 0002:02:00.0,ipsec_in_max_spi=128 -w > 0002:03:00.0,ipsec_in_max_spi=128 -w 0002:0e:00.0 -w 0002:10:00.1 > --log-level=8 > -c 0x1 -- -P -p 0x3 -u 0x1 -f aes-gcm.cfg --transfer-mode event > --event-schedule-type parallel > > #Inbound and outbound driver mode > ipsec-secgw -w 0002:02:00.0,ipsec_in_max_spi=128 -w > 0002:03:00.0,ipsec_in_max_spi=128 -w 0002:0e:00.0 -w 0002:10:00.1 > --log-level=8 > -c 0x1 -- -P -p 0x3 -u 0x1 -f aes-gcm.cfg --transfer-mode event > --event-schedule-type parallel --single-sa 0 > > This series adds non burst tx internal port workers only. It provides > infrastructure > for non internal port workers, however does not define any. Also, only inline > ipsec > protocol mode is supported by the worker threads added. > > Following are planned features, > 1. Add burst mode workers. > 2. Add non internal port workers. > 3. Verify support for Rx core (the support is added but lack of h/w to > verify). > 4. Add lookaside protocol support. > > Following are features that Marvell won't be attempting. > 1. Inline crypto support. > 2. Lookaside crypto support. > > For the features that Marvell won't be attempting, new workers can be > introduced by the respective stake holders. > > This series is tested on Marvell OCTEON TX2. > This series is targeted for 20.05 release. > > Changes in v5: > * Rename function check_params() to check_poll_mode_params() and > check_eh_conf() to check_event_mode_params() in order to make it clear > what is their purpose. > * Forbid usage of --config option in event mode. > * Replace magic numbers on return with enum values in > process_ipsec_ev_inbound() > and process_ipsec_ev_outbound() functions. > * Add session_priv_pool for both inbound and outbound configuration in > ipsec_wrkr_non_burst_int_port_app_mode worker. > * Add check of event type in ipsec_wrkr_non_burst_int_port_app_mode worker. > * Update description of --config option in both ipsec-secgw help and > documentation. > > Changes in v4: > * Update ipsec-secgw documentation to describe the new options as well as > event mode support. > * In event mode reserve number of crypto queues equal to number of eth ports > in order to meet inline protocol offload requirements. > * Add calculate_nb_mbufs() function to calculate number of mbufs in a pool > and include fragments table size into the calculation. > * Move structures ipsec_xf and ipsec_sad to ipsec.h and remove static keyword > from sa_out, nb_sa_out, sa_in and nb_sa_in in sa.c. > * Update process_ipsec_ev_inbound(), process_ipsec_ev_outbound(), check_sp() > and prepare_out_sessions_tbl() functions as a result of changes introduced > by SAD feature. > * Remove setting sa->cdev_id_qp to 0 in create_inline_session as sa_ctx > is created with rte_zmalloc. > * Minor cleanup enhancements: > - In eh_set_default_conf_eventdev() function in event_helper.c put definition > of int local vars in one line, remove invalid comment, put > "eventdev_config->ev_queue_mode = RTE_EVENT_QUEUE_CFG_ALL_TYPES" in one line > instead of two. > - Remove extern "C" from event_helper.h. > - Put local vars in reverse xmas tree order in eh_dev_has_rx_internal_port() > and > eh_dev_has_tx_internal_port() functions in event_helper.c. > - Put #include <rte_bitmap.h> in alphabetical order in ipsec-secgw.c. > - Move "extern volatile bool force_quit" and "#include <stdbool.h>" to > ipsec-secgw.h, > remove #include <rte_hash.h>. > - Remove not needed includes in ipsec_worker.c. > - Remove expired todo from ipsec_worker.h. > > Changes in v3: > * Move eh_conf_init() and eh_conf_uninit() functions to event_helper.c > including minor rework. > * Rename --schedule-type option to --event-schedule-type. > * Replace macro UNPROTECTED_PORT with static inline function > is_unprotected_port(). > * Move definitions of global variables used by multiple modules > to .c files and add externs in .h headers. > * Add eh_check_conf() which validates ipsec-secgw configuration > for event mode. > * Add dynamic calculation of number of buffers in a pool based > on number of cores, ports and crypto queues. > * Fix segmentation fault in event mode driver worker which happens > when there are no inline outbound sessions configured. > * Remove change related to updating number of crypto queues > in cryptodevs_init(). The update of crypto queues will be handled > in a separate patch. > * Fix compilation error on 32-bit platforms by using userdata instead > of udata64 from rte_mbuf. > > Changes in v2: > * Remove --process-dir option. Instead use existing unprotected port mask > option (-u) to decide wheter port handles inbound or outbound traffic. > * Remove --process-mode option. Instead use existing --single-sa option > to select between app and driver modes. > * Add handling of PKT_RX_SEC_OFFLOAD_FAIL result in app worker thread. > * Fix passing of req_rx_offload flags to create_default_ipsec_flow(). > * Move destruction of flows to a location where eth ports are stopped > and closed. > * Print error and exit when event mode --schedule-type option is used > in poll mode. > * Reduce number of goto statements replacing them with loop constructs. > * Remove sec_session_fixed table and replace it with locally build > table in driver worker thread. Table is indexed by port identifier > and holds first inline session pointer found for a given port. > * Print error and exit when sessions other than inline are configured > in event mode. > * When number of event queues is less than number of eth ports then > map all eth ports to one event queue. > * Cleanup and minor improvements in code as suggested by Konstantin > > Ankur Dwivedi (1): > examples/ipsec-secgw: add default rte flow for inline Rx > > Anoob Joseph (5): > examples/ipsec-secgw: add framework for eventmode helper > examples/ipsec-secgw: add eventdev port-lcore link > examples/ipsec-secgw: add Rx adapter support > examples/ipsec-secgw: add Tx adapter support > examples/ipsec-secgw: add routines to display config > > Lukasz Bartosik (9): > examples/ipsec-secgw: add routines to launch workers > examples/ipsec-secgw: add support for internal ports > examples/ipsec-secgw: add event helper config init/uninit > examples/ipsec-secgw: add eventmode to ipsec-secgw > examples/ipsec-secgw: add driver mode worker > examples/ipsec-secgw: add app mode worker > examples/ipsec-secgw: make number of buffers dynamic > doc: add event mode support to ipsec-secgw > examples/ipsec-secgw: reserve crypto queues in event mode > > doc/guides/sample_app_ug/ipsec_secgw.rst | 135 ++- > examples/ipsec-secgw/Makefile | 2 + > examples/ipsec-secgw/event_helper.c | 1812 > ++++++++++++++++++++++++++++++ > examples/ipsec-secgw/event_helper.h | 327 ++++++ > examples/ipsec-secgw/ipsec-secgw.c | 506 +++++++-- > examples/ipsec-secgw/ipsec-secgw.h | 88 ++ > examples/ipsec-secgw/ipsec.c | 5 +- > examples/ipsec-secgw/ipsec.h | 53 +- > examples/ipsec-secgw/ipsec_worker.c | 649 +++++++++++ > examples/ipsec-secgw/ipsec_worker.h | 41 + > examples/ipsec-secgw/meson.build | 6 +- > examples/ipsec-secgw/sa.c | 21 +- > examples/ipsec-secgw/sad.h | 5 - > 13 files changed, 3516 insertions(+), 134 deletions(-) > create mode 100644 examples/ipsec-secgw/event_helper.c > create mode 100644 examples/ipsec-secgw/event_helper.h > create mode 100644 examples/ipsec-secgw/ipsec-secgw.h > create mode 100644 examples/ipsec-secgw/ipsec_worker.c > create mode 100644 examples/ipsec-secgw/ipsec_worker.h > > --
Have to say I didn't look extensively on event mode. My primary concern was poll-mode and common code changes. >From that perspective - LGTM. Series Acked-by: Konstantin Ananyev <konstantin.anan...@intel.com> > 2.7.4