[dpdk-dev] [PATCH] net/nfp: add multiprocess support

Ferruh Yigit ferruh.yigit at intel.com
Wed Nov 28 18:28:48 CET 2018


On 11/28/2018 11:32 AM, Alejandro Lucero wrote:
> This patch introduces changes for supporting multiprocess support.
> This is trivial for VFs but comes with some limitations for the PF.
> 
> Due to restrictions when using NFP CPP interface, just one secondary
> process is supported by now for the PF.

Hi Alejandro,

I think we needs to clarify first what is the multiprocess support required in
PMD level, there are a few different implementations. cc'ed a few more folks for
helping clarification.

According my understanding,
Only *one* DPDK application (primary or secondary) should control a device.
There is no restriction in DPDK for it, this responsibility is pushed to
application, application should manage it.

Device initialization (probe()) done only by primary application.

Because of above two items, we only need RTE_PROC_PRIMARY check in driver
probe() but not in other eth_dev_ops.
If we need eth_dev_ops should be protected, I believe that should be done in
ethdev layer for all PMDs instead of each (some) PMD does it itself.


Following should be possible:
1)
- Primary starts with 5 ports and configures them
- Start 5 secondary apps in a way each control a specific port data path

2)
- Primary start with N ports
- A secondary per port configure the port and control data path


Thanks,
ferruh



> 
> Signed-off-by: Alejandro Lucero <alejandro.lucero at netronome.com>
> ---
>  doc/guides/nics/features/nfp.ini           |   1 +
>  doc/guides/nics/features/nfp_vf.ini        |   1 +
>  drivers/net/nfp/nfp_net.c                  | 133 +++++++++++++--------
>  drivers/net/nfp/nfpcore/nfp_cpp_pcie_ops.c | 103 ++++++++++++++--
>  4 files changed, 177 insertions(+), 61 deletions(-)
> 
> diff --git a/doc/guides/nics/features/nfp.ini b/doc/guides/nics/features/nfp.ini
> index d2899e7fb..70297b090 100644
> --- a/doc/guides/nics/features/nfp.ini
> +++ b/doc/guides/nics/features/nfp.ini
> @@ -24,5 +24,6 @@ Basic stats          = Y
>  Stats per queue      = Y
>  Linux UIO            = Y
>  Linux VFIO           = Y
> +Multiprocess aware   = Y
>  x86-64               = Y
>  Usage doc            = Y
> diff --git a/doc/guides/nics/features/nfp_vf.ini b/doc/guides/nics/features/nfp_vf.ini
> index d2899e7fb..70297b090 100644
> --- a/doc/guides/nics/features/nfp_vf.ini
> +++ b/doc/guides/nics/features/nfp_vf.ini
> @@ -24,5 +24,6 @@ Basic stats          = Y
>  Stats per queue      = Y
>  Linux UIO            = Y
>  Linux VFIO           = Y
> +Multiprocess aware   = Y
>  x86-64               = Y
>  Usage doc            = Y
> diff --git a/drivers/net/nfp/nfp_net.c b/drivers/net/nfp/nfp_net.c
> index 54c6da924..ffef97d80 100644
> --- a/drivers/net/nfp/nfp_net.c
> +++ b/drivers/net/nfp/nfp_net.c
> @@ -766,9 +766,12 @@ nfp_net_start(struct rte_eth_dev *dev)
>  		goto error;
>  	}
>  
> -	if (hw->is_pf)
> +	if (hw->is_pf && rte_eal_process_type() == RTE_PROC_PRIMARY)
>  		/* Configure the physical port up */
>  		nfp_eth_set_configured(hw->cpp, hw->pf_port_idx, 1);
> +	else
> +		nfp_eth_set_configured(dev->process_private,
> +				       hw->pf_port_idx, 1);
>  
>  	hw->ctrl = new_ctrl;
>  
> @@ -817,9 +820,12 @@ nfp_net_stop(struct rte_eth_dev *dev)
>  			(struct nfp_net_rxq *)dev->data->rx_queues[i]);
>  	}
>  
> -	if (hw->is_pf)
> +	if (hw->is_pf && rte_eal_process_type() == RTE_PROC_PRIMARY)
>  		/* Configure the physical port down */
>  		nfp_eth_set_configured(hw->cpp, hw->pf_port_idx, 0);
> +	else
> +		nfp_eth_set_configured(dev->process_private,
> +				       hw->pf_port_idx, 0);
>  }
>  
>  /* Reset and stop device. The device can not be restarted. */
> @@ -2918,16 +2924,16 @@ nfp_net_init(struct rte_eth_dev *eth_dev)
>  		     hw->mac_addr[0], hw->mac_addr[1], hw->mac_addr[2],
>  		     hw->mac_addr[3], hw->mac_addr[4], hw->mac_addr[5]);
>  
> -	/* Registering LSC interrupt handler */
> -	rte_intr_callback_register(&pci_dev->intr_handle,
> -				   nfp_net_dev_interrupt_handler,
> -				   (void *)eth_dev);
> -
> -	/* Telling the firmware about the LSC interrupt entry */
> -	nn_cfg_writeb(hw, NFP_NET_CFG_LSC, NFP_NET_IRQ_LSC_IDX);
> -
> -	/* Recording current stats counters values */
> -	nfp_net_stats_reset(eth_dev);
> +	if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
> +		/* Registering LSC interrupt handler */
> +		rte_intr_callback_register(&pci_dev->intr_handle,
> +					   nfp_net_dev_interrupt_handler,
> +					   (void *)eth_dev);
> +		/* Telling the firmware about the LSC interrupt entry */
> +		nn_cfg_writeb(hw, NFP_NET_CFG_LSC, NFP_NET_IRQ_LSC_IDX);
> +		/* Recording current stats counters values */
> +		nfp_net_stats_reset(eth_dev);
> +	}
>  
>  	return 0;
>  
> @@ -2947,7 +2953,7 @@ nfp_pf_create_dev(struct rte_pci_device *dev, int port, int ports,
>  	struct rte_eth_dev *eth_dev;
>  	struct nfp_net_hw *hw;
>  	char *port_name;
> -	int ret;
> +	int retval;
>  
>  	port_name = rte_zmalloc("nfp_pf_port_name", 100, 0);
>  	if (!port_name)
> @@ -2958,51 +2964,76 @@ nfp_pf_create_dev(struct rte_pci_device *dev, int port, int ports,
>  	else
>  		sprintf(port_name, "%s", dev->device.name);
>  
> -	eth_dev = rte_eth_dev_allocate(port_name);
> -	if (!eth_dev)
> -		return -ENOMEM;
>  
> -	if (port == 0) {
> -		*priv = rte_zmalloc(port_name,
> -				    sizeof(struct nfp_net_adapter) * ports,
> -				    RTE_CACHE_LINE_SIZE);
> -		if (!*priv) {
> -			rte_eth_dev_release_port(eth_dev);
> -			return -ENOMEM;
> +	if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
> +		eth_dev = rte_eth_dev_allocate(port_name);
> +		if (!eth_dev) {
> +			rte_free(port_name);
> +			return -ENODEV;
>  		}
> -	}
> -
> -	eth_dev->data->dev_private = *priv;
> -
> -	/*
> -	 * dev_private pointing to port0 dev_private because we need
> -	 * to configure vNIC bars based on port0 at nfp_net_init.
> -	 * Then dev_private is adjusted per port.
> -	 */
> -	hw = (struct nfp_net_hw *)(eth_dev->data->dev_private) + port;
> -	hw->cpp = cpp;
> -	hw->hwinfo = hwinfo;
> -	hw->sym_tbl = sym_tbl;
> -	hw->pf_port_idx = phys_port;
> -	hw->is_pf = 1;
> -	if (ports > 1)
> -		hw->pf_multiport_enabled = 1;
> +		if (port == 0) {
> +			*priv = rte_zmalloc(port_name,
> +					    sizeof(struct nfp_net_adapter) *
> +					    ports, RTE_CACHE_LINE_SIZE);
> +			if (!*priv) {
> +				rte_free(port_name);
> +				rte_eth_dev_release_port(eth_dev);
> +				return -ENOMEM;
> +			}
> +		}
> +		eth_dev->data->dev_private = *priv;
>  
> -	hw->total_ports = ports;
> +		/*
> +		 * dev_private pointing to port0 dev_private because we need
> +		 * to configure vNIC bars based on port0 at nfp_net_init.
> +		 * Then dev_private is adjusted per port.
> +		 */
> +		hw = (struct nfp_net_hw *)(eth_dev->data->dev_private) + port;
> +		hw->cpp = cpp;
> +		hw->hwinfo = hwinfo;
> +		hw->sym_tbl = sym_tbl;
> +		hw->pf_port_idx = phys_port;
> +		hw->is_pf = 1;
> +		if (ports > 1)
> +			hw->pf_multiport_enabled = 1;
> +
> +		hw->total_ports = ports;
> +	} else {
> +		eth_dev = rte_eth_dev_attach_secondary(port_name);
> +		if (!eth_dev) {
> +			RTE_LOG(ERR, EAL, "secondary process attach failed, "
> +				"ethdev doesn't exist");
> +			rte_free(port_name);
> +			return -ENODEV;
> +		}
> +		eth_dev->process_private = cpp;
> +	}
>  
>  	eth_dev->device = &dev->device;
>  	rte_eth_copy_pci_info(eth_dev, dev);
>  
> -	ret = nfp_net_init(eth_dev);
> +	retval = nfp_net_init(eth_dev);
>  
> -	if (ret)
> -		rte_eth_dev_release_port(eth_dev);
> -	else
> +	if (retval) {
> +		retval = -ENODEV;
> +		goto probe_failed;
> +	} else {
>  		rte_eth_dev_probing_finish(eth_dev);
> +	}
>  
>  	rte_free(port_name);
>  
> -	return ret;
> +	return retval;
> +
> +probe_failed:
> +	rte_free(port_name);
> +	/* free ports private data if primary process */
> +	if (rte_eal_process_type() == RTE_PROC_PRIMARY)
> +		rte_free(eth_dev->data->dev_private);
> +
> +	rte_eth_dev_release_port(eth_dev);
> +
> +	return retval;
>  }
>  
>  #define DEFAULT_FW_PATH       "/lib/firmware/netronome"
> @@ -3180,10 +3211,12 @@ static int nfp_pf_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
>  		return -EIO;
>  	}
>  
> -	if (nfp_fw_setup(dev, cpp, nfp_eth_table, hwinfo)) {
> -		PMD_DRV_LOG(INFO, "Error when uploading firmware");
> -		ret = -EIO;
> -		goto error;
> +	if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
> +		if (nfp_fw_setup(dev, cpp, nfp_eth_table, hwinfo)) {
> +			PMD_DRV_LOG(INFO, "Error when uploading firmware");
> +			ret = -EIO;
> +			goto error;
> +		}
>  	}
>  
>  	/* Now the symbol table should be there */
> diff --git a/drivers/net/nfp/nfpcore/nfp_cpp_pcie_ops.c b/drivers/net/nfp/nfpcore/nfp_cpp_pcie_ops.c
> index c68d9400f..39bd48a83 100644
> --- a/drivers/net/nfp/nfpcore/nfp_cpp_pcie_ops.c
> +++ b/drivers/net/nfp/nfpcore/nfp_cpp_pcie_ops.c
> @@ -112,6 +112,7 @@ struct nfp_pcie_user {
>  
>  	int device;
>  	int lock;
> +	int secondary_lock;
>  	char busdev[BUSDEV_SZ];
>  	int barsz;
>  	char *cfg;
> @@ -290,14 +291,32 @@ nfp_reconfigure_bar(struct nfp_pcie_user *nfp, struct nfp_bar *bar, int tgt,
>   * already mapped.
>   *
>   * BAR0.0: Reserved for General Mapping (for MSI-X access to PCIe SRAM)
> + *
> + *         Halving PCItoCPPBars for primary and secondary processes.
> + *         NFP PMD just requires two fixed slots, one for configuration BAR,
> + *         and another for accessing the hw queues. Another slot is needed
> + *         for setting the link up or down. Secondary processes do not need
> + *         to map the first two slots again, but it requires one slot for
> + *         accessing the link, even if it is not likely the secondary process
> + *         starting the port. This implies a limit of secondary processes
> + *         supported. Due to this requirement and future extensions requiring
> + *         new slots per process, only one secondary process is supported by
> + *         now.
>   */
>  static int
>  nfp_enable_bars(struct nfp_pcie_user *nfp)
>  {
>  	struct nfp_bar *bar;
> -	int x;
> +	int x, start, end;
>  
> -	for (x = ARRAY_SIZE(nfp->bar); x > 0; x--) {
> +	if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
> +		start = 4;
> +		end = 1;
> +	} else {
> +		start = 7;
> +		end = 4;
> +	}
> +	for (x = start; x > end; x--) {
>  		bar = &nfp->bar[x - 1];
>  		bar->barcfg = 0;
>  		bar->nfp = nfp;
> @@ -320,9 +339,16 @@ static struct nfp_bar *
>  nfp_alloc_bar(struct nfp_pcie_user *nfp)
>  {
>  	struct nfp_bar *bar;
> -	int x;
> +	int x, start, end;
>  
> -	for (x = ARRAY_SIZE(nfp->bar); x > 0; x--) {
> +	if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
> +		start = 4;
> +		end = 1;
> +	} else {
> +		start = 7;
> +		end = 4;
> +	}
> +	for (x = start; x > end; x--) {
>  		bar = &nfp->bar[x - 1];
>  		if (!bar->lock) {
>  			bar->lock = 1;
> @@ -336,9 +362,17 @@ static void
>  nfp_disable_bars(struct nfp_pcie_user *nfp)
>  {
>  	struct nfp_bar *bar;
> -	int x;
> +	int x, start, end;
> +
> +	if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
> +		start = 4;
> +		end = 1;
> +	} else {
> +		start = 7;
> +		end = 4;
> +	}
>  
> -	for (x = ARRAY_SIZE(nfp->bar); x > 0; x--) {
> +	for (x = start; x > end; x--) {
>  		bar = &nfp->bar[x - 1];
>  		if (bar->iomem) {
>  			bar->iomem = NULL;
> @@ -633,6 +667,47 @@ nfp_acquire_process_lock(struct nfp_pcie_user *desc)
>  	return 0;
>  }
>  
> +static int
> +nfp_acquire_secondary_process_lock(struct nfp_pcie_user *desc)
> +{
> +	int rc;
> +	struct flock lock;
> +	const char *lockname = "/.lock_nfp_secondary";
> +	char *home_path;
> +	char *lockfile;
> +
> +	memset(&lock, 0, sizeof(lock));
> +
> +	/*
> +	 * Using user's home directory. Note this can be called in a DPDK app
> +	 * being executed as non-root. This is not the case for the previous
> +	 * function nfp_acquire_process_lock which is invoked only when UIO
> +	 * driver is used because that implies root user.
> +	 */
> +	home_path = getenv("HOME");
> +	lockfile = calloc(strlen(home_path) + strlen(lockname) + 1,
> +			  sizeof(char));
> +
> +	strcat(lockfile, home_path);
> +	strcat(lockfile, "/.lock_nfp_secondary");
> +	desc->secondary_lock = open(lockfile, O_RDWR | O_CREAT | O_NONBLOCK,
> +				    0666);
> +	if (desc->secondary_lock < 0) {
> +		RTE_LOG(ERR, PMD, "NFP lock for secondary process failed\n");
> +		return desc->secondary_lock;
> +	}
> +
> +	lock.l_type = F_WRLCK;
> +	lock.l_whence = SEEK_SET;
> +	rc = fcntl(desc->secondary_lock, F_SETLK, &lock);
> +	if (rc < 0) {
> +		RTE_LOG(ERR, PMD, "NFP lock for secondary process failed\n");
> +		close(desc->secondary_lock);
> +	}
> +
> +	return rc;
> +}
> +
>  static int
>  nfp6000_set_model(struct rte_pci_device *dev, struct nfp_cpp *cpp)
>  {
> @@ -774,7 +849,6 @@ static int
>  nfp6000_init(struct nfp_cpp *cpp, struct rte_pci_device *dev)
>  {
>  	int ret = 0;
> -	uint32_t model;
>  	struct nfp_pcie_user *desc;
>  
>  	desc = malloc(sizeof(*desc));
> @@ -785,12 +859,20 @@ nfp6000_init(struct nfp_cpp *cpp, struct rte_pci_device *dev)
>  	memset(desc->busdev, 0, BUSDEV_SZ);
>  	strlcpy(desc->busdev, dev->device.name, sizeof(desc->busdev));
>  
> -	if (cpp->driver_lock_needed) {
> +	if (rte_eal_process_type() == RTE_PROC_PRIMARY &&
> +	    cpp->driver_lock_needed) {
>  		ret = nfp_acquire_process_lock(desc);
>  		if (ret)
>  			return -1;
>  	}
>  
> +	/* Just support for one secondary process */
> +	if (rte_eal_process_type() != RTE_PROC_PRIMARY) {
> +		ret = nfp_acquire_secondary_process_lock(desc);
> +		if (ret)
> +			return -1;
> +	}
> +
>  	if (nfp6000_set_model(dev, cpp) < 0)
>  		return -1;
>  	if (nfp6000_set_interface(dev, cpp) < 0)
> @@ -806,9 +888,6 @@ nfp6000_init(struct nfp_cpp *cpp, struct rte_pci_device *dev)
>  
>  	nfp_cpp_priv_set(cpp, desc);
>  
> -	model = __nfp_cpp_model_autodetect(cpp);
> -	nfp_cpp_model_set(cpp, model);
> -
>  	return ret;
>  }
>  
> @@ -820,6 +899,8 @@ nfp6000_free(struct nfp_cpp *cpp)
>  	nfp_disable_bars(desc);
>  	if (cpp->driver_lock_needed)
>  		close(desc->lock);
> +	if (rte_eal_process_type() != RTE_PROC_PRIMARY)
> +		close(desc->secondary_lock);
>  	close(desc->device);
>  	free(desc);
>  }
> 



More information about the dev mailing list