> > Copy public function pointers (rx_pkt_burst(), etc.) and related
> > pointers to internal data from rte_eth_dev structure into a
> > separate flat array. That array will remain in a public header.
> > The intention here is to make rte_eth_dev and related structures internal.
> > That should allow future possible changes to core eth_dev structures
> > to be transparent to the user and help to avoid ABI/API breakages.
> > The plan is to keep minimal part of data from rte_eth_dev public,
> > so we still can use inline functions for fast-path calls
> > (like rte_eth_rx_burst(), etc.) to avoid/minimize slowdown.
> > The whole idea beyond this new schema:
> > 1. PMDs keep to setup fast-path function pointers and related data
> > inside rte_eth_dev struct in the same way they did it before.
> > 2. Inside rte_eth_dev_start() and inside rte_eth_dev_probing_finish()
> > (for secondary process) we call eth_dev_fp_ops_setup, which
> > copies these function and data pointers into rte_eth_fp_ops[port_id].
> > 3. Inside rte_eth_dev_stop() and inside rte_eth_dev_release_port()
> > we call eth_dev_fp_ops_reset(), which resets rte_eth_fp_ops[port_id]
> > into some dummy values.
> > 4. fast-path ethdev API (rte_eth_rx_burst(), etc.) will use that new
> > flat array to call PMD specific functions.
> > That approach should allow us to make rte_eth_devices[] private
> > without introducing regression and help to avoid changes in drivers code.
> >
> > Signed-off-by: Konstantin Ananyev <konstantin.anan...@intel.com>
> > ---
> > lib/ethdev/ethdev_private.c | 52 ++++++++++++++++++++++++++++++++++
> > lib/ethdev/ethdev_private.h | 7 +++++
> > lib/ethdev/rte_ethdev.c | 27 ++++++++++++++++++
> > lib/ethdev/rte_ethdev_core.h | 55 ++++++++++++++++++++++++++++++++++++
> > 4 files changed, 141 insertions(+)
> >
> > diff --git a/lib/ethdev/ethdev_private.c b/lib/ethdev/ethdev_private.c
> > index 012cf73ca2..3eeda6e9f9 100644
> > --- a/lib/ethdev/ethdev_private.c
> > +++ b/lib/ethdev/ethdev_private.c
> > @@ -174,3 +174,55 @@ rte_eth_devargs_parse_representor_ports(char *str,
> > void *data)
> > RTE_LOG(ERR, EAL, "wrong representor format: %s\n", str);
> > return str == NULL ? -1 : 0;
> > }
> > +
> > +static uint16_t
> > +dummy_eth_rx_burst(__rte_unused void *rxq,
> > + __rte_unused struct rte_mbuf **rx_pkts,
> > + __rte_unused uint16_t nb_pkts)
> > +{
> > + RTE_ETHDEV_LOG(ERR, "rx_pkt_burst for unconfigured port\n");
> > + rte_errno = ENOTSUP;
> > + return 0;
> > +}
> > +
> > +static uint16_t
> > +dummy_eth_tx_burst(__rte_unused void *txq,
> > + __rte_unused struct rte_mbuf **tx_pkts,
> > + __rte_unused uint16_t nb_pkts)
> > +{
> > + RTE_ETHDEV_LOG(ERR, "tx_pkt_burst for unconfigured port\n");
> > + rte_errno = ENOTSUP;
> > + return 0;
> > +}
> > +
> > +void
> > +eth_dev_fp_ops_reset(struct rte_eth_fp_ops *fpo)
>
> The port_id parameter is preferable, this will hide rte_eth_fp_ops as much as
> possible.
Why do we need to hide it here?
rte_eth_fp_ops is a public structure, and it is a helper function that
just resets fields of this structure to some predefined dummy values.
Nice and simple, so I prefer to keep it like that.
>
> > +{
> > + static void *dummy_data[RTE_MAX_QUEUES_PER_PORT];
> > + static const struct rte_eth_fp_ops dummy_ops = {
> > + .rx_pkt_burst = dummy_eth_rx_burst,
> > + .tx_pkt_burst = dummy_eth_tx_burst,
> > + .rxq = {.data = dummy_data, .clbk = dummy_data,},
> > + .txq = {.data = dummy_data, .clbk = dummy_data,},
> > + };
> > +
> > + *fpo = dummy_ops;
> > +}
> > +
> > +void
> > +eth_dev_fp_ops_setup(struct rte_eth_fp_ops *fpo,
> > + const struct rte_eth_dev *dev)
>
> Because fp_ops and eth_dev is a one-to-one correspondence. It's better only
> use
> port_id parameter.
Same as above:
All this internal helper function does - copies some fields from one structure
to another.
Both structures are visible by ethdev layer.
No point to add extra assumptions and complexity here.
>
> > +{
> > + fpo->rx_pkt_burst = dev->rx_pkt_burst;
> > + fpo->tx_pkt_burst = dev->tx_pkt_burst;
> > + fpo->tx_pkt_prepare = dev->tx_pkt_prepare;
> > + fpo->rx_queue_count = dev->rx_queue_count;
> > + fpo->rx_descriptor_status = dev->rx_descriptor_status;
> > + fpo->tx_descriptor_status = dev->tx_descriptor_status;
> > +
> > + fpo->rxq.data = dev->data->rx_queues;
> > + fpo->rxq.clbk = (void **)(uintptr_t)dev->post_rx_burst_cbs;
> > +
> > + fpo->txq.data = dev->data->tx_queues;
> > + fpo->txq.clbk = (void **)(uintptr_t)dev->pre_tx_burst_cbs;
> > +}
> > diff --git a/lib/ethdev/ethdev_private.h b/lib/ethdev/ethdev_private.h
> > index 3724429577..5721be7bdc 100644
> > --- a/lib/ethdev/ethdev_private.h
> > +++ b/lib/ethdev/ethdev_private.h
> > @@ -26,4 +26,11 @@ eth_find_device(const struct rte_eth_dev *_start,
> > rte_eth_cmp_t cmp,
> > /* Parse devargs value for representor parameter. */
> > int rte_eth_devargs_parse_representor_ports(char *str, void *data);
> >
> > +/* reset eth fast-path API to dummy values */
> > +void eth_dev_fp_ops_reset(struct rte_eth_fp_ops *fpo);
> > +
> > +/* setup eth fast-path API to ethdev values */
> > +void eth_dev_fp_ops_setup(struct rte_eth_fp_ops *fpo,
> > + const struct rte_eth_dev *dev);
>
> Some drivers control the transmit/receive function during operation. E.g.
> for hns3 driver, when detect reset, primary process will set rx/tx burst to
> dummy, after
> process reset, primary process will set the correct rx/tx burst. During this
> process, the
> send and receive threads are still working, but the bursts they call are
> changed. So:
This text above is a bit too cryptic for me...
Are you saying that your driver changes rte_eth_dev.rx_pkt_burst(/
tx_pkt_burst) on the fly
(after dev_start() and before dev_stop())?
If so, then generally speaking, it is a bad idea.
While it might works for some limited scenarios, right now it is not supported
by ethdev framework,
and might introduce a lot of problems.
> 1. it is recommended that trace be deleted from the dummy function.
You are talking about:
RTE_ETHDEV_LOG(ERR, "rx_pkt_burst for unconfigured port\n");
right?
Dummy function is supposed to be set only when device is not able to do RX/TX
properly
(not attached, or attached but not configured, or attached and configured, but
not started).
Obviously if app calls rx/tx_burst for such port it is a major issue, that
should be flagged imemdiatelly.
So I believe having log here makes a perfect sense here.
> 2. public the eth_dev_fp_ops_reset/setup interface for driver usage.
You mean move their declarations into ethdev_driver.h?
I suppose that could be done, but still wonder why driver would need to
call these functions directly?
> > +
> > #endif /* _ETH_PRIVATE_H_ */
> > diff --git a/lib/ethdev/rte_ethdev.c b/lib/ethdev/rte_ethdev.c
> > index c8abda6dd7..9f7a0cbb8c 100644
> > --- a/lib/ethdev/rte_ethdev.c
> > +++ b/lib/ethdev/rte_ethdev.c
> > @@ -44,6 +44,9 @@
> > static const char *MZ_RTE_ETH_DEV_DATA = "rte_eth_dev_data";
> > struct rte_eth_dev rte_eth_devices[RTE_MAX_ETHPORTS];
> >
> > +/* public fast-path API */
> > +struct rte_eth_fp_ops rte_eth_fp_ops[RTE_MAX_ETHPORTS];
> > +
> > /* spinlock for eth device callbacks */
> > static rte_spinlock_t eth_dev_cb_lock = RTE_SPINLOCK_INITIALIZER;
> >
> > @@ -578,6 +581,8 @@ rte_eth_dev_release_port(struct rte_eth_dev *eth_dev)
> > rte_eth_dev_callback_process(eth_dev,
> > RTE_ETH_EVENT_DESTROY, NULL);
> >
> > + eth_dev_fp_ops_reset(rte_eth_fp_ops + eth_dev->data->port_id);
> > +
> > rte_spinlock_lock(ð_dev_shared_data->ownership_lock);
> >
> > eth_dev->state = RTE_ETH_DEV_UNUSED;
> > @@ -1787,6 +1792,9 @@ rte_eth_dev_start(uint16_t port_id)
> > (*dev->dev_ops->link_update)(dev, 0);
> > }
> >
> > + /* expose selection of PMD fast-path functions */
> > + eth_dev_fp_ops_setup(rte_eth_fp_ops + port_id, dev);
> > +
> > rte_ethdev_trace_start(port_id);
> > return 0;
> > }
> > @@ -1809,6 +1817,9 @@ rte_eth_dev_stop(uint16_t port_id)
> > return 0;
> > }
> >
> > + /* point fast-path functions to dummy ones */
> > + eth_dev_fp_ops_reset(rte_eth_fp_ops + port_id);
> > +
> > dev->data->dev_started = 0;
> > ret = (*dev->dev_ops->dev_stop)(dev);
> > rte_ethdev_trace_stop(port_id, ret);
> > @@ -4567,6 +4578,14 @@ rte_eth_mirror_rule_reset(uint16_t port_id, uint8_t
> > rule_id)
> > return eth_err(port_id, (*dev->dev_ops->mirror_rule_reset)(dev,
> > rule_id));
> > }
> >
> > +RTE_INIT(eth_dev_init_fp_ops)
> > +{
> > + uint32_t i;
> > +
> > + for (i = 0; i != RTE_DIM(rte_eth_fp_ops); i++)
> > + eth_dev_fp_ops_reset(rte_eth_fp_ops + i);
> > +}
> > +
> > RTE_INIT(eth_dev_init_cb_lists)
> > {
> > uint16_t i;
> > @@ -4735,6 +4754,14 @@ rte_eth_dev_probing_finish(struct rte_eth_dev *dev)
> > if (dev == NULL)
> > return;
> >
> > + /*
> > + * for secondary process, at that point we expect device
> > + * to be already 'usable', so shared data and all function pointers
> > + * for fast-path devops have to be setup properly inside rte_eth_dev.
> > + */
> > + if (rte_eal_process_type() == RTE_PROC_SECONDARY)
> > + eth_dev_fp_ops_setup(rte_eth_fp_ops + dev->data->port_id, dev);
> > +
> > rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_NEW, NULL);
> >
> > dev->state = RTE_ETH_DEV_ATTACHED;
> > diff --git a/lib/ethdev/rte_ethdev_core.h b/lib/ethdev/rte_ethdev_core.h
> > index 51cd68de94..d5853dff86 100644
> > --- a/lib/ethdev/rte_ethdev_core.h
> > +++ b/lib/ethdev/rte_ethdev_core.h
> > @@ -50,6 +50,61 @@ typedef int (*eth_rx_descriptor_status_t)(void *rxq,
> > uint16_t offset);
> > typedef int (*eth_tx_descriptor_status_t)(void *txq, uint16_t offset);
> > /**< @internal Check the status of a Tx descriptor */
> >
> > +/**
> > + * @internal
> > + * Structure used to hold opaque pointers to internal ethdev Rx/Tx
> > + * queues data.
> > + * The main purpose to expose these pointers at all - allow compiler
> > + * to fetch this data for fast-path ethdev inline functions in advance.
> > + */
> > +struct rte_ethdev_qdata {
> > + void **data;
> > + /**< points to array of internal queue data pointers */
> > + void **clbk;
> > + /**< points to array of queue callback data pointers */
> > +};
> > +
> > +/**
> > + * @internal
> > + * fast-path ethdev functions and related data are hold in a flat array.
> > + * One entry per ethdev.
> > + * On 64-bit systems contents of this structure occupy exactly two 64B
> > lines.
> > + * On 32-bit systems contents of this structure fits into one 64B line.
> > + */
> > +struct rte_eth_fp_ops {
> > +
> > + /**
> > + * Rx fast-path functions and related data.
> > + * 64-bit systems: occupies first 64B line
> > + */
> > + eth_rx_burst_t rx_pkt_burst;
> > + /**< PMD receive function. */
> > + eth_rx_queue_count_t rx_queue_count;
> > + /**< Get the number of used RX descriptors. */
> > + eth_rx_descriptor_status_t rx_descriptor_status;
> > + /**< Check the status of a Rx descriptor. */
> > + struct rte_ethdev_qdata rxq;
> > + /**< Rx queues data. */
> > + uintptr_t reserved1[3];
> > +
> > + /**
> > + * Tx fast-path functions and related data.
> > + * 64-bit systems: occupies second 64B line
> > + */
> > + eth_tx_burst_t tx_pkt_burst;
>
> Why not place rx_pkt_burst/tx_pkt_burst/rxq /txq to the first cacheline ?
> Other function, e.g. rx_queue_count/descriptor_status are low frequency call
> functions.
I suppose you are talking about layout like that:
struct rte_eth_fp_ops {
/* first 64B line */
rx_pkt_burst;
tx_pkt_burst;
tx_pkt_prepare;
struct rte_ethdev_qdata rxq;
struct rte_ethdev_qdata txq;
reserved1[1];
/* second 64B line */
...
};
I thought about such ability, even tried it, but I didn't see any performance
gain.
From other side current layout seems better to me from structural point:
it is more uniform and easy to extend in future (both RX and TX data occupies
separate 64B line, each have equal rom for extension).
> > + /**< PMD transmit function. */
> > + eth_tx_prep_t tx_pkt_prepare;
> > + /**< PMD transmit prepare function. */
> > + eth_tx_descriptor_status_t tx_descriptor_status;
> > + /**< Check the status of a Tx descriptor. */
> > + struct rte_ethdev_qdata txq;
> > + /**< Tx queues data. */
> > + uintptr_t reserved2[3];
> > +
> > +} __rte_cache_aligned;
> > +
> > +extern struct rte_eth_fp_ops rte_eth_fp_ops[RTE_MAX_ETHPORTS];
> > +
> >
> > /**
> > * @internal
> >
