Hi Tomasz,
> static int
> rte_eth_dev_tx_queue_config(struct rte_eth_dev *dev, uint16_t nb_queues)
> {
> uint16_t old_nb_queues = dev->data->nb_tx_queues;
> void **txq;
> + struct rte_eth_dev_tx_buffer *new_bufs;
> unsigned i;
>
> if (dev->data->tx_queues == NULL) { /* first time configuration */
> @@ -841,17 +872,40 @@ rte_eth_dev_tx_queue_config(struct rte_eth_dev *dev,
> uint16_t nb_queues)
> dev->data->nb_tx_queues = 0;
> return -(ENOMEM);
> }
> +
> + dev->data->txq_bufs = rte_zmalloc("ethdev->txq_bufs",
> + sizeof(*dev->data->txq_bufs) * nb_queues, 0);
> + if (dev->data->txq_bufs == NULL) {
> + dev->data->nb_tx_queues = 0;
> + rte_free(dev->data->tx_queues);
> + return -(ENOMEM);
> + }
> +
> } else { /* re-configure */
> +
> + /* flush the packets queued for all queues*/
> + for (i = 0; i < old_nb_queues; i++)
> + rte_eth_tx_buffer_flush(dev->data->port_id, i);
> +
I don't think it is safe to call tx_burst() at queue config stage.
Instead you need to flush (or just empty) your txq)bufs at tx_queue_stop stage.
> RTE_FUNC_PTR_OR_ERR_RET(*dev->dev_ops->tx_queue_release,
> -ENOTSUP);
>
> + /* get new buffer space first, but keep old space around */
> + new_bufs = rte_zmalloc("ethdev->txq_bufs",
> + sizeof(*dev->data->txq_bufs) * nb_queues, 0);
> + if (new_bufs == NULL)
> + return -(ENOMEM);
> +
Why not to allocate space for txq_bufs together with tx_queues (as one chunk
for both)?
As I understand there is always one to one mapping between them anyway.
Would simplify things a bit.
Or even introduce a new struct to group with all related tx queue info togetehr
struct rte_eth_txq_data {
void *queue; /*actual pmd queue*/
struct rte_eth_dev_tx_buffer buf;
uint8_t state;
}
And use it inside struct rte_eth_dev_data?
Would probably give a better data locality.
> txq = dev->data->tx_queues;
>
> for (i = nb_queues; i < old_nb_queues; i++)
> (*dev->dev_ops->tx_queue_release)(txq[i]);
> txq = rte_realloc(txq, sizeof(txq[0]) * nb_queues,
> RTE_CACHE_LINE_SIZE);
> - if (txq == NULL)
> - return -ENOMEM;
> + if (txq == NULL) {
> + rte_free(new_bufs);
> + return -(ENOMEM);
> + }
> +
> if (nb_queues > old_nb_queues) {
> uint16_t new_qs = nb_queues - old_nb_queues;
>
> @@ -861,6 +915,9 @@ rte_eth_dev_tx_queue_config(struct rte_eth_dev *dev,
> uint16_t nb_queues)
>
> dev->data->tx_queues = txq;
>
> + /* now replace old buffers with new */
> + rte_free(dev->data->txq_bufs);
> + dev->data->txq_bufs = new_bufs;
> }
> dev->data->nb_tx_queues = nb_queues;
> return 0;
> diff --git a/lib/librte_ether/rte_ethdev.h b/lib/librte_ether/rte_ethdev.h
> index bada8ad..23faa6a 100644
> --- a/lib/librte_ether/rte_ethdev.h
> +++ b/lib/librte_ether/rte_ethdev.h
> @@ -1,7 +1,7 @@
> /*-
> * BSD LICENSE
> *
> - * Copyright(c) 2010-2015 Intel Corporation. All rights reserved.
> + * Copyright(c) 2010-2016 Intel Corporation. All rights reserved.
> * All rights reserved.
> *
> * Redistribution and use in source and binary forms, with or without
> @@ -182,6 +182,7 @@ extern "C" {
> #include <rte_pci.h>
> #include <rte_dev.h>
> #include <rte_devargs.h>
> +#include <rte_branch_prediction.h>
> #include "rte_ether.h"
> #include "rte_eth_ctrl.h"
> #include "rte_dev_info.h"
> @@ -1519,6 +1520,34 @@ enum rte_eth_dev_type {
> RTE_ETH_DEV_MAX /**< max value of this enum */
> };
>
> +typedef void (*buffer_tx_error_fn)(struct rte_mbuf **unsent, uint16_t count,
> + void *userdata);
> +
> +/**
> + * @internal
> + * Structure used to buffer packets for future TX
> + * Used by APIs rte_eth_tx_buffer and rte_eth_tx_buffer_flush
> + */
> +struct rte_eth_dev_tx_buffer {
> + struct rte_mbuf *pkts[RTE_ETHDEV_TX_BUFSIZE];
I think it is better to make size of pkts[] configurable at runtime.
There are a lot of different usage scenarios - hard to predict what would be an
optimal buffer size for all cases.
> + unsigned nb_pkts;
> + uint64_t errors;
> + /**< Total number of queue packets to sent that are dropped. */
> +};
> +
> +/**
> + * @internal
> + * Structure to hold a callback to be used on error when a tx_buffer_flush
> + * call fails to send all packets.
> + * This needs to be a separate structure, as it must go in the ethdev
> structure
> + * rather than ethdev_data, due to the use of a function pointer, which is
> not
> + * multi-process safe.
> + */
> +struct rte_eth_dev_tx_buffer_err_cb {
> + buffer_tx_error_fn flush_cb; /* callback for when tx_burst fails */
> + void *userdata; /* userdata for callback */
> +};
> +
> /**
> * @internal
> * The generic data structure associated with each ethernet device.
> @@ -1550,6 +1579,9 @@ struct rte_eth_dev {
> struct rte_eth_rxtx_callback *pre_tx_burst_cbs[RTE_MAX_QUEUES_PER_PORT];
> uint8_t attached; /**< Flag indicating the port is attached */
> enum rte_eth_dev_type dev_type; /**< Flag indicating the device type */
> +
> + /** Callbacks to be used on a tx_buffer_flush error */
> + struct rte_eth_dev_tx_buffer_err_cb
> tx_buf_err_cb[RTE_MAX_QUEUES_PER_PORT];
> };
>
> struct rte_eth_dev_sriov {
> @@ -1610,6 +1642,8 @@ struct rte_eth_dev_data {
> enum rte_kernel_driver kdrv; /**< Kernel driver passthrough */
> int numa_node; /**< NUMA node connection */
> const char *drv_name; /**< Driver name */
> + struct rte_eth_dev_tx_buffer *txq_bufs;
> + /**< space to allow buffered transmits */
> };
>
> /** Device supports hotplug detach */
> @@ -2661,8 +2695,181 @@ rte_eth_tx_burst(uint8_t port_id, uint16_t queue_id,
> }
>
> /**
> - * The eth device event type for interrupt, and maybe others in the future.
> + * Buffer a single packet for future transmission on a port and queue
> + *
> + * This function takes a single mbuf/packet and buffers it for later
> + * transmission on the particular port and queue specified. Once the buffer
> is
> + * full of packets, an attempt will be made to transmit all the buffered
> + * packets. In case of error, where not all packets can be transmitted, a
> + * callback is called with the unsent packets as a parameter. If no callback
> + * is explicitly set up, the unsent packets are just freed back to the owning
> + * mempool. The function returns the number of packets actually sent i.e.
> + * 0 if no buffer flush occurred, otherwise the number of packets
> successfully
> + * flushed
> + *
> + * @param port_id
> + * The port identifier of the Ethernet device.
> + * @param queue_id
> + * The index of the transmit queue through which output packets must be
> + * sent.
> + * The value must be in the range [0, nb_tx_queue - 1] previously supplied
> + * to rte_eth_dev_configure().
> + * @param tx_pkt
> + * Pointer to the packet mbuf to be sent.
> + * @return
> + * 0 = packet has been buffered for later transmission
> + * N > 0 = packet has been buffered, and the buffer was subsequently
> flushed,
> + * causing N packets to be sent, and the error callback to be called for
> + * the rest.
> + */
> +static inline uint16_t __attribute__((always_inline))
> +rte_eth_tx_buffer(uint8_t port_id, uint16_t queue_id, struct rte_mbuf
> *tx_pkt)
> +{
> + struct rte_eth_dev *dev = &rte_eth_devices[port_id];
> + struct rte_eth_dev_tx_buffer *qbuf = &dev->data->txq_bufs[queue_id];
> + uint16_t i;
> +
> + qbuf->pkts[qbuf->nb_pkts++] = tx_pkt;
> + if (qbuf->nb_pkts < RTE_ETHDEV_TX_BUFSIZE)
> + return 0;
> +
Probably just call rte_eth_tx_buffer_flush() here to avoid duplication.
> + const uint16_t sent = rte_eth_tx_burst(port_id, queue_id, qbuf->pkts,
> + RTE_ETHDEV_TX_BUFSIZE);
> +
> + qbuf->nb_pkts = 0;
> +
> + /* All packets sent, or to be dealt with by callback below */
> + if (unlikely(sent != RTE_ETHDEV_TX_BUFSIZE)) {
> + if (dev->tx_buf_err_cb[queue_id].flush_cb)
> + dev->tx_buf_err_cb[queue_id].flush_cb(&qbuf->pkts[sent],
> + RTE_ETHDEV_TX_BUFSIZE - sent,
> + dev->tx_buf_err_cb[queue_id].userdata);
> + else {
> + qbuf->errors += RTE_ETHDEV_TX_BUFSIZE - sent;
> + for (i = sent; i < RTE_ETHDEV_TX_BUFSIZE; i++)
> + rte_pktmbuf_free(qbuf->pkts[i]);
> + }
> + }
> +
> + return sent;
> +}
> +
> +/**
> + * Send any packets queued up for transmission on a port and HW queue
> + *
> + * This causes an explicit flush of packets previously buffered via the
> + * rte_eth_tx_buffer() function. It returns the number of packets
> successfully
> + * sent to the NIC, and calls the error callback for any unsent packets.
> Unless
> + * explicitly set up otherwise, the default callback simply frees the unsent
> + * packets back to the owning mempool.
> + *
> + * @param port_id
> + * The port identifier of the Ethernet device.
> + * @param queue_id
> + * The index of the transmit queue through which output packets must be
> + * sent.
> + * The value must be in the range [0, nb_tx_queue - 1] previously supplied
> + * to rte_eth_dev_configure().
> + * @return
> + * The number of packets successfully sent to the Ethernet device. The
> error
> + * callback is called for any packets which could not be sent.
> + */
> +static inline uint16_t
> +rte_eth_tx_buffer_flush(uint8_t port_id, uint16_t queue_id)
> +{
> + uint16_t i;
> + struct rte_eth_dev *dev = &rte_eth_devices[port_id];
> + struct rte_eth_dev_tx_buffer *qbuf = &dev->data->txq_bufs[queue_id];
> +
> + if (qbuf->nb_pkts == 0)
> + return 0;
> +
> + const uint16_t to_send = qbuf->nb_pkts;
> +
> + const uint16_t sent = rte_eth_tx_burst(port_id, queue_id, qbuf->pkts,
> + to_send);
Try to avoid defining variables in the middle of the code block.
Again no much value in having these 2 above variables as 'const'.
Konstantin
> +
> + qbuf->nb_pkts = 0;
> +
> + /* All packets sent, or to be dealt with by callback below */
> + if (unlikely(sent != to_send)) {
> + if (dev->tx_buf_err_cb[queue_id].flush_cb)
> + dev->tx_buf_err_cb[queue_id].flush_cb(&qbuf->pkts[sent],
> + to_send - sent,
> + dev->tx_buf_err_cb[queue_id].userdata);
> + else {
> + qbuf->errors += to_send - sent;
> + for (i = sent; i < to_send; i++)
> + rte_pktmbuf_free(qbuf->pkts[i]);
> + }
> + }
> +
> + return sent;
> +}
> +
