RE: [EXT] Re: [v2,net-next, 1/2] enetc: Configure the Time-Aware Scheduler via tc-taprio offload

[Po Liu <po.liu@nxp.com>]

Hi Ivan,

> -----Original Message-----
> From: Ivan Khoronzhuk <ivan.khoronzhuk@linaro.org>
> Sent: 2019年11月13日 5:10
> To: Po Liu <po.liu@nxp.com>
> Cc: Claudiu Manoil <claudiu.manoil@nxp.com>; davem@davemloft.net; linux-
> kernel@vger.kernel.org; netdev@vger.kernel.org; vinicius.gomes@intel.com;
> Vladimir Oltean <vladimir.oltean@nxp.com>; Alexandru Marginean
> <alexandru.marginean@nxp.com>; Xiaoliang Yang
> <xiaoliang.yang_1@nxp.com>; Roy Zang <roy.zang@nxp.com>; Mingkai Hu
> <mingkai.hu@nxp.com>; Jerry Huang <jerry.huang@nxp.com>; Leo Li
> <leoyang.li@nxp.com>
> Subject: [EXT] Re: [v2,net-next, 1/2] enetc: Configure the Time-Aware Scheduler
> via tc-taprio offload
> 
> Caution: EXT Email
> 
> Hello,
> 
> On Tue, Nov 12, 2019 at 08:42:49AM +0000, Po Liu wrote:
> >ENETC supports in hardware for time-based egress shaping according to
> >IEEE 802.1Qbv. This patch implement the Qbv enablement by the hardware
> >offload method qdisc tc-taprio method.
> >Also update cbdr writeback to up level since control bd ring may
> >writeback data to control bd ring.
> >
> >Signed-off-by: Po Liu <Po.Liu@nxp.com>
> >Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
> >Signed-off-by: Claudiu Manoil <claudiu.manoil@nxp.com>
> >---
> >changes:
> >- introduce a local define CONFIG_FSL_ENETC_QOS to fix the various
> >  configurations will result in link errors.
> >  Since the CONFIG_NET_SCH_TAPRIO depends on many Qos configs. Not
> >  to use it directly in driver. Add it to CONFIG_FSL_ENETC_QOS depends
> >  on list, so only CONFIG_NET_SCH_TAPRIO enabled, user can enable this
> >  tsn feature, or else, return not support.
> >
> > drivers/net/ethernet/freescale/enetc/Kconfig  |  10 ++
> > drivers/net/ethernet/freescale/enetc/Makefile |   1 +
> > drivers/net/ethernet/freescale/enetc/enetc.c  |  19 ++-
> > drivers/net/ethernet/freescale/enetc/enetc.h  |   7 +
> > .../net/ethernet/freescale/enetc/enetc_cbdr.c |   5 +-
> > .../net/ethernet/freescale/enetc/enetc_hw.h   | 150 ++++++++++++++++--
> > .../net/ethernet/freescale/enetc/enetc_qos.c  | 130 +++++++++++++++
> > 7 files changed, 300 insertions(+), 22 deletions(-) create mode 100644
> > drivers/net/ethernet/freescale/enetc/enetc_qos.c
> >
> 
> [...]
> 
> >
> >@@ -1483,6 +1479,19 @@ int enetc_setup_tc(struct net_device *ndev, enum
> tc_setup_type type,
> >       return 0;
> > }
> >
> >+int enetc_setup_tc(struct net_device *ndev, enum tc_setup_type type,
> >+                 void *type_data)
> >+{
> >+      switch (type) {
> >+      case TC_SETUP_QDISC_MQPRIO:
> >+              return enetc_setup_tc_mqprio(ndev, type_data);
> Sorry didn't see v2, so i duplicate my question here:
> 
> This patch is for taprio offload, I see that mqprio is related and is part of taprio
> offload configuration. But taprio offload has own mqprio settings.
> The taprio mqprio part is not offloaded with TC_SETUP_QDISC_MQPRIO.
> 
> So, a combination of mqprio and tario qdiscs used.
> Could you please share the commands were used for your setup?
> 

Example command: 
tc qdisc replace dev eth0 parent root handle 100  taprio num_tc 8 map 0 1 2 3 4 5 6 7 queues 1@0 1@1 1@2 1@3 1@4 1@5 1@6 1@7 base-time 01 sched-entry S 02 300000 flags 0x2

> And couple interesting questions about all of this:
> - The taprio qdisc has to have mqprio settings, but if it's done with mqprio then
> it just skipped (by reading tc class num).
> - If no separate mqprio qdisc configuration then mqprio conf from taprio is set,
> who should restore tc mappings when taprio qdisc is unloaded?
> Maybe there is reason to implement TC_SETUP_QDISC_MQPRIO offload in
> taprio since it's required feature?

Mqprio offload with TC_SETUP_QDISC_MQPRIO would be good or even the plus with num_tc would fix some hack. This has a discussion with Vinicius Gomes for a future patch fix. 
I know the problem is the mqprio will be set after the offload function. But offload function may use some for hardware set.

> Would be better to move changes for mqprio in separate patch with
> explanation.
> 
> >+      case TC_SETUP_QDISC_TAPRIO:
> >+              return enetc_setup_tc_taprio(ndev, type_data);
> >+      default:
> >+              return -EOPNOTSUPP;
> >+      }
> >+}
> >+
> 
> [...]
> 
> >diff --git a/drivers/net/ethernet/freescale/enetc/enetc_qos.c
> >b/drivers/net/ethernet/freescale/enetc/enetc_qos.c
> >new file mode 100644
> >index 000000000000..036bb39c7a0b
> >--- /dev/null
> >+++ b/drivers/net/ethernet/freescale/enetc/enetc_qos.c
> 
> [...]
> 
> >+static int enetc_setup_taprio(struct net_device *ndev,
> >+                            struct tc_taprio_qopt_offload *admin_conf)
> >+{
> >+      struct enetc_ndev_priv *priv = netdev_priv(ndev);
> >+      struct enetc_cbd cbd = {.cmd = 0};
> >+      struct tgs_gcl_conf *gcl_config;
> >+      struct tgs_gcl_data *gcl_data;
> >+      struct gce *gce;
> >+      dma_addr_t dma;
> >+      u16 data_size;
> >+      u16 gcl_len;
> >+      u32 temp;
> >+      int i;
> >+
> >+      gcl_len = admin_conf->num_entries;
> >+      if (gcl_len > enetc_get_max_gcl_len(&priv->si->hw))
> >+              return -EINVAL;
> >+
> >+      if (admin_conf->enable) {
> >+              enetc_wr(&priv->si->hw,
> >+                       ENETC_QBV_PTGCR_OFFSET,
> >+                       temp & (~ENETC_QBV_TGE));
> >+              usleep_range(10, 20);
> >+              enetc_wr(&priv->si->hw,
> >+                       ENETC_QBV_PTGCR_OFFSET,
> >+                       temp | ENETC_QBV_TGE);
> >+      } else {
> >+              enetc_wr(&priv->si->hw,
> >+                       ENETC_QBV_PTGCR_OFFSET,
> >+                       temp & (~ENETC_QBV_TGE));
> >+              return 0;
> >+      }
> 
> Better do the upper qbv enable/disable procedure closer to enetc_send_cmd()
> or at least after kzalloc that can fail, no need to restore then.

After saving the 'if' suggest by Simon Horman, the enable part could move close to the cmd().

> 
> >+
> >+      /* Configure the (administrative) gate control list using the
> >+       * control BD descriptor.
> >+       */
> >+      gcl_config = &cbd.gcl_conf;
> >+
> >+      data_size = sizeof(struct tgs_gcl_data) + gcl_len *
> >+ sizeof(struct gce);
> >+
> >+      gcl_data = kzalloc(data_size, __GFP_DMA | GFP_KERNEL);
> >+      if (!gcl_data)
> >+              return -ENOMEM;
> >+
> >+      gce = (struct gce *)(gcl_data + 1);
> >+
> >+      /* Since no initial state config in taprio, set gates open as default.
> >+       */
> tc-taprio and IEEE Qbv allows to change configuration in flight, so that oper
> state is active till new admin start time. So, here comment says it does initial
> state config, if in-flight feature is not supported then error has to be returned
> instead of silently rewriting configuration. But if it can be implemented then
> state should be remembered/verified in order to not brake oper configuration?

I think this is ok as per standard. Also see this comment in
net/sched/sch_taprio.c:

	/* Until the schedule starts, all the queues are open */
I would change the comment.

> >+      gcl_config->atc = 0xff;
> >+      gcl_config->acl_len = cpu_to_le16(gcl_len);
> 
> Ok, this is maximum number of schedules.
> According to tc-taprio it's possible to set cycle period more then schedules
> actually can consume. If cycle time is more, then last gate's state can be kept
> till the end of cycle. But if last schedule has it's own interval set then gates
> should be closed till the end of cycle or no? if it has to be closed, then one more
> endl schedule should be present closing gates at the end of list for the rest cycle
> time. Can be implemented in h/w but just to be sure, how it's done in h/w?
> 
There is already check the list len in up code.
if (admin_conf->num_entries > enetc_get_max_gcl_len(&priv->si->hw))
	return -EINVAL;
gcl_len = admin_conf->num_entries; 

> >+
> >+      if (!admin_conf->base_time) {
> >+              gcl_data->btl =
> >+                      cpu_to_le32(enetc_rd(&priv->si->hw, ENETC_SICTR0));
> >+              gcl_data->bth =
> >+                      cpu_to_le32(enetc_rd(&priv->si->hw, ENETC_SICTR1));
> >+      } else {
> >+              gcl_data->btl =
> >+                      cpu_to_le32(lower_32_bits(admin_conf->base_time));
> >+              gcl_data->bth =
> >+                      cpu_to_le32(upper_32_bits(admin_conf->base_time));
> >+      }
> >+
> >+      gcl_data->ct = cpu_to_le32(admin_conf->cycle_time);
> >+      gcl_data->cte = cpu_to_le32(admin_conf->cycle_time_extension);
> 
> Not sure it's good idea to write values w/o verification, The cycle time and time
> extension is 64 values, so it's supposed them to be more then 4,3 seconds, it's
> probably not a case, but better return error if it's more.
> 

Can add a check for cycle time and extension since type not match.
 
> >+
> >+      for (i = 0; i < gcl_len; i++) {
> >+              struct tc_taprio_sched_entry *temp_entry;
> >+              struct gce *temp_gce = gce + i;
> >+
> >+              temp_entry = &admin_conf->entries[i];
> >+
> >+              temp_gce->gate = cpu_to_le32(temp_entry->gate_mask);
> 
> So, gate_mask can have up to 32 traffic classes? :-|.

Gate_mask is defined u32 type. Simon has suggested endian issue. Would change in next patch.

> 
> >+              temp_gce->period = cpu_to_le32(temp_entry->interval);
> 
> So, the interval can be up to 4.3 seconds for one schedule?
> That is, one cycle can be one schedule.
> great.
> 
> >+      }
> 
> There is no schedule cmd set, so only SetGateStates is supported?
> But anyway it's Ok.
> 
> >+
> >+      cbd.length = cpu_to_le16(data_size);
> >+      cbd.status_flags = 0;
> >+
> >+      dma = dma_map_single(&priv->si->pdev->dev, gcl_data,
> >+                           data_size, DMA_TO_DEVICE);
> >+      if (dma_mapping_error(&priv->si->pdev->dev, dma)) {
> >+              netdev_err(priv->si->ndev, "DMA mapping failed!\n");
> >+              kfree(gcl_data);
> >+              return -ENOMEM;
> >+      }
> >+
> >+      cbd.addr[0] = lower_32_bits(dma);
> >+      cbd.addr[1] = upper_32_bits(dma);
> >+      cbd.cls = BDCR_CMD_PORT_GCL;
> >+
> >+      /* Updated by ENETC on completion of the configuration
> >+       * command. A zero value indicates success.
> >+       */
> >+      cbd.status_flags = 0;
> 
> It's updated on completion by setting 0 on success, then why it's here set to 0
> but not 1 and not verified afterwards?
>

This byte is feedback by hardware after enetc_send_cmd. Hardware require the cbd space set status_flags 0 before send to hardware.
 No choice of larger than 0 before send to hardware.
But you mind me the enetc_send_cmd() need to check return value.

> >+
> >+      enetc_send_cmd(priv->si, &cbd);
> >+
> >+      dma_unmap_single(&priv->si->pdev->dev, dma, data_size,
> DMA_TO_DEVICE);
> >+      kfree(gcl_data);
> >+
> >+      return 0;
> >+}
> >+
> >+int enetc_setup_tc_taprio(struct net_device *ndev, void *type_data) {
> >+      struct tc_taprio_qopt_offload *taprio = type_data;
> >+      struct enetc_ndev_priv *priv = netdev_priv(ndev);
> >+      int i;
> >+
> >+      for (i = 0; i < priv->num_tx_rings; i++)
> >+              enetc_set_bdr_prio(&priv->si->hw,
> >+                                 priv->tx_ring[i]->index,
> >+                                 taprio->enable ? i : 0);
> 
> then why enable/disable at the beginning for whole qbv scheduler, maybe
> better do it together? Or better say, what if setup_taprio failed, who restore
> configuration?

You remind me the enetc_send_cmd() need to check return value.

> >+
> >+      return enetc_setup_taprio(ndev, taprio); }
> >--
> >2.17.1
> >
> 
> --
> Regards,
> Ivan Khoronzhuk