[dpdk-dev] [PATCH v2] Allow -ve frame_overhead values

2018-01-16 Thread alanrobertsonatt
From: Alan Robertson 

When forwarding traffic across a TDM network the ethernet header will
be replaced with a ML-PPP one thereby reducing the size of the packet.

Signed-off-by: Alan Robertson 
---
 lib/librte_sched/rte_sched.c | 14 +-
 lib/librte_sched/rte_sched.h |  5 +++--
 2 files changed, 16 insertions(+), 3 deletions(-)

diff --git a/lib/librte_sched/rte_sched.c b/lib/librte_sched/rte_sched.c
index ad2f7c6d5..c971fd0d4 100644
--- a/lib/librte_sched/rte_sched.c
+++ b/lib/librte_sched/rte_sched.c
@@ -187,7 +187,7 @@ struct rte_sched_port {
uint32_t n_pipes_per_subport;
uint32_t rate;
uint32_t mtu;
-   uint32_t frame_overhead;
+   int32_t frame_overhead;
uint16_t qsize[RTE_SCHED_TRAFFIC_CLASSES_PER_PIPE];
uint32_t n_pipe_profiles;
uint32_t pipe_tc3_rate_max;
@@ -1591,6 +1591,10 @@ grinder_credits_check(struct rte_sched_port *port, 
uint32_t pos)
uint32_t pipe_tc_credits = pipe->tc_credits[tc_index];
int enough_credits;
 
+#ifdef RTE_SCHED_DEBUG
+   assert((int)(pkt->pkt_len + port->frame_overhead) > 0);
+#endif /* RTE_SCHED_DEBUG */
+
/* Check queue credits */
enough_credits = (pkt_len <= subport_tb_credits) &&
(pkt_len <= subport_tc_credits) &&
@@ -1629,6 +1633,10 @@ grinder_credits_check(struct rte_sched_port *port, 
uint32_t pos)
uint32_t pipe_tc_ov_credits = pipe_tc_ov_mask1[tc_index];
int enough_credits;
 
+#ifdef RTE_SCHED_DEBUG
+   assert((int)(pkt->pkt_len + port->frame_overhead) > 0);
+#endif /* RTE_SCHED_DEBUG */
+
/* Check pipe and subport credits */
enough_credits = (pkt_len <= subport_tb_credits) &&
(pkt_len <= subport_tc_credits) &&
@@ -1663,6 +1671,10 @@ grinder_schedule(struct rte_sched_port *port, uint32_t 
pos)
if (!grinder_credits_check(port, pos))
return 0;
 
+#ifdef RTE_SCHED_DEBUG
+   assert((int)(pkt->pkt_len + port->frame_overhead) > 0);
+#endif /* RTE_SCHED_DEBUG */
+
/* Advance port time */
port->time += pkt_len;
 
diff --git a/lib/librte_sched/rte_sched.h b/lib/librte_sched/rte_sched.h
index 5d2a688dc..3e135c1e5 100644
--- a/lib/librte_sched/rte_sched.h
+++ b/lib/librte_sched/rte_sched.h
@@ -190,8 +190,9 @@ struct rte_sched_port_params {
uint32_t mtu;/**< Maximum Ethernet frame size
  * (measured in bytes).
  * Should not include the framing 
overhead. */
-   uint32_t frame_overhead; /**< Framing overhead per packet
- * (measured in bytes) */
+   int32_t frame_overhead;
+   /**< Framing overhead per packet (measured in bytes).
+* Can have negative value. */
uint32_t n_subports_per_port;/**< Number of subports */
uint32_t n_pipes_per_subport;/**< Number of pipes per subport */
uint16_t qsize[RTE_SCHED_TRAFFIC_CLASSES_PER_PIPE];
-- 
2.11.0



[dpdk-dev] [PATCH] Improve the shaper accuracy for large packets

2018-02-09 Thread alanrobertsonatt
From: Alan Robertson 

There were 2 issues, the first was time could be lost whilst updating
the traffic-class period, the second was a frame could be delayed if
not enough tokens were available for the full frame.  By allowing the
shaper to borrow credit from the next period the throughput is improved.
---
 lib/librte_sched/rte_sched.c | 58 +++-
 1 file changed, 41 insertions(+), 17 deletions(-)

diff --git a/lib/librte_sched/rte_sched.c b/lib/librte_sched/rte_sched.c
index 634486c..e53a424 100644
--- a/lib/librte_sched/rte_sched.c
+++ b/lib/librte_sched/rte_sched.c
@@ -57,7 +57,7 @@ struct rte_sched_subport {
/* Traffic classes (TCs) */
uint64_t tc_time; /* time of next update */
uint32_t tc_credits_per_period[RTE_SCHED_TRAFFIC_CLASSES_PER_PIPE];
-   uint32_t tc_credits[RTE_SCHED_TRAFFIC_CLASSES_PER_PIPE];
+   int32_t tc_credits[RTE_SCHED_TRAFFIC_CLASSES_PER_PIPE];
uint32_t tc_period;
 
/* TC oversubscription */
@@ -98,7 +98,7 @@ struct rte_sched_pipe {
 
/* Traffic classes (TCs) */
uint64_t tc_time; /* time of next update */
-   uint32_t tc_credits[RTE_SCHED_TRAFFIC_CLASSES_PER_PIPE];
+   int32_t tc_credits[RTE_SCHED_TRAFFIC_CLASSES_PER_PIPE];
 
/* Weighted Round Robin (WRR) */
uint8_t wrr_tokens[RTE_SCHED_QUEUES_PER_PIPE];
@@ -1451,6 +1451,8 @@ grinder_credits_update(struct rte_sched_port *port, 
uint32_t pos)
struct rte_sched_pipe *pipe = grinder->pipe;
struct rte_sched_pipe_profile *params = grinder->pipe_params;
uint64_t n_periods;
+   uint32_t tc;
+   uint64_t lapsed;
 
/* Subport TB */
n_periods = (port->time - subport->tb_time) / subport->tb_period;
@@ -1466,20 +1468,42 @@ grinder_credits_update(struct rte_sched_port *port, 
uint32_t pos)
 
/* Subport TCs */
if (unlikely(port->time >= subport->tc_time)) {
-   subport->tc_credits[0] = subport->tc_credits_per_period[0];
-   subport->tc_credits[1] = subport->tc_credits_per_period[1];
-   subport->tc_credits[2] = subport->tc_credits_per_period[2];
-   subport->tc_credits[3] = subport->tc_credits_per_period[3];
-   subport->tc_time = port->time + subport->tc_period;
+   for (tc = 0; tc < RTE_SCHED_TRAFFIC_CLASSES_PER_PIPE; tc++) {
+   if (subport->tc_credits[tc] < 0)
+   subport->tc_credits[tc] +=
+   subport->tc_credits_per_period[tc];
+   else
+   subport->tc_credits[tc] =
+   subport->tc_credits_per_period[tc];
+   }
+   /* If we've run into the next period only update the clock to
+* the time + tc_period so we'll replenish the tc tokens early
+* in the next tc_period to compensate. */
+   lapsed = port->time - subport->tc_time;
+   if (lapsed < subport->tc_period)
+   subport->tc_time += subport->tc_period;
+   else
+   subport->tc_time = port->time + subport->tc_period;
}
 
/* Pipe TCs */
if (unlikely(port->time >= pipe->tc_time)) {
-   pipe->tc_credits[0] = params->tc_credits_per_period[0];
-   pipe->tc_credits[1] = params->tc_credits_per_period[1];
-   pipe->tc_credits[2] = params->tc_credits_per_period[2];
-   pipe->tc_credits[3] = params->tc_credits_per_period[3];
-   pipe->tc_time = port->time + params->tc_period;
+   for (tc = 0; tc < RTE_SCHED_TRAFFIC_CLASSES_PER_PIPE; tc++) {
+   if (pipe->tc_credits[tc] < 0)
+   pipe->tc_credits[tc] +=
+   params->tc_credits_per_period[tc];
+   else
+   pipe->tc_credits[tc] =
+   params->tc_credits_per_period[tc];
+   }
+   /* If we've run into the next period only update the clock to
+* the time + tc_period so we'll replenish the tc tokens early
+* in the next tc_period to compensate. */
+   lapsed = port->time - pipe->tc_time;
+   if (lapsed < params->tc_period)
+   pipe->tc_time += params->tc_period;
+   else
+   pipe->tc_time = port->time + params->tc_period;
}
 }
 
@@ -1586,16 +1610,16 @@ grinder_credits_check(struct rte_sched_port *port, 
uint32_t pos)
uint32_t tc_index = grinder->tc_index;
uint32_t pkt_len = pkt->pkt_len + port->frame_overhead;
uint32_t subport_tb_credits = subport->tb_credits;
-   uint32_t subport_tc_credits = subport->tc_credits[tc_index];
+   int32_t subport_tc_credits = subport->tc_credits[tc_ind

[dpdk-dev] [PATCH] Improve the shaper accuracy for large packets

2018-02-09 Thread alanrobertsonatt
From: Alan Robertson 

There were 2 issues, the first was time could be lost whilst updating
the traffic-class period, the second was a frame could be delayed if
not enough tokens were available for the full frame.  By allowing the
shaper to borrow credit from the next period the throughput is improved.

Signed-off-by: Alan Robertson 
---
 lib/librte_sched/rte_sched.c | 60 +++-
 1 file changed, 43 insertions(+), 17 deletions(-)

diff --git a/lib/librte_sched/rte_sched.c b/lib/librte_sched/rte_sched.c
index 634486c..7b06b0b 100644
--- a/lib/librte_sched/rte_sched.c
+++ b/lib/librte_sched/rte_sched.c
@@ -57,7 +57,7 @@ struct rte_sched_subport {
/* Traffic classes (TCs) */
uint64_t tc_time; /* time of next update */
uint32_t tc_credits_per_period[RTE_SCHED_TRAFFIC_CLASSES_PER_PIPE];
-   uint32_t tc_credits[RTE_SCHED_TRAFFIC_CLASSES_PER_PIPE];
+   int32_t tc_credits[RTE_SCHED_TRAFFIC_CLASSES_PER_PIPE];
uint32_t tc_period;
 
/* TC oversubscription */
@@ -98,7 +98,7 @@ struct rte_sched_pipe {
 
/* Traffic classes (TCs) */
uint64_t tc_time; /* time of next update */
-   uint32_t tc_credits[RTE_SCHED_TRAFFIC_CLASSES_PER_PIPE];
+   int32_t tc_credits[RTE_SCHED_TRAFFIC_CLASSES_PER_PIPE];
 
/* Weighted Round Robin (WRR) */
uint8_t wrr_tokens[RTE_SCHED_QUEUES_PER_PIPE];
@@ -1451,6 +1451,8 @@ grinder_credits_update(struct rte_sched_port *port, 
uint32_t pos)
struct rte_sched_pipe *pipe = grinder->pipe;
struct rte_sched_pipe_profile *params = grinder->pipe_params;
uint64_t n_periods;
+   uint32_t tc;
+   uint64_t lapsed;
 
/* Subport TB */
n_periods = (port->time - subport->tb_time) / subport->tb_period;
@@ -1466,20 +1468,44 @@ grinder_credits_update(struct rte_sched_port *port, 
uint32_t pos)
 
/* Subport TCs */
if (unlikely(port->time >= subport->tc_time)) {
-   subport->tc_credits[0] = subport->tc_credits_per_period[0];
-   subport->tc_credits[1] = subport->tc_credits_per_period[1];
-   subport->tc_credits[2] = subport->tc_credits_per_period[2];
-   subport->tc_credits[3] = subport->tc_credits_per_period[3];
-   subport->tc_time = port->time + subport->tc_period;
+   for (tc = 0; tc < RTE_SCHED_TRAFFIC_CLASSES_PER_PIPE; tc++) {
+   if (subport->tc_credits[tc] < 0)
+   subport->tc_credits[tc] +=
+   subport->tc_credits_per_period[tc];
+   else
+   subport->tc_credits[tc] =
+   subport->tc_credits_per_period[tc];
+   }
+   /* If we've run into the next period only update the clock to
+* the time + tc_period so we'll replenish the tc tokens early
+* in the next tc_period to compensate.
+*/
+   lapsed = port->time - subport->tc_time;
+   if (lapsed < subport->tc_period)
+   subport->tc_time += subport->tc_period;
+   else
+   subport->tc_time = port->time + subport->tc_period;
}
 
/* Pipe TCs */
if (unlikely(port->time >= pipe->tc_time)) {
-   pipe->tc_credits[0] = params->tc_credits_per_period[0];
-   pipe->tc_credits[1] = params->tc_credits_per_period[1];
-   pipe->tc_credits[2] = params->tc_credits_per_period[2];
-   pipe->tc_credits[3] = params->tc_credits_per_period[3];
-   pipe->tc_time = port->time + params->tc_period;
+   for (tc = 0; tc < RTE_SCHED_TRAFFIC_CLASSES_PER_PIPE; tc++) {
+   if (pipe->tc_credits[tc] < 0)
+   pipe->tc_credits[tc] +=
+   params->tc_credits_per_period[tc];
+   else
+   pipe->tc_credits[tc] =
+   params->tc_credits_per_period[tc];
+   }
+   /* If we've run into the next period only update the clock to
+* the time + tc_period so we'll replenish the tc tokens early
+* in the next tc_period to compensate.
+*/
+   lapsed = port->time - pipe->tc_time;
+   if (lapsed < params->tc_period)
+   pipe->tc_time += params->tc_period;
+   else
+   pipe->tc_time = port->time + params->tc_period;
}
 }
 
@@ -1586,16 +1612,16 @@ grinder_credits_check(struct rte_sched_port *port, 
uint32_t pos)
uint32_t tc_index = grinder->tc_index;
uint32_t pkt_len = pkt->pkt_len + port->frame_overhead;
uint32_t subport_tb_credits = subport->tb_credits;
-   uint32_t subport_tc_credits = subport->tc_credits[tc_index

[dpdk-dev] [PATCH] Allow -ve frame_overhead values

2018-01-08 Thread alanrobertsonatt
From: Alan Robertson 

When forwarding traffic across a TDM network the ethernet header will
be replaced with a ML-PPP one thereby reducing the size of the packet.

Signed-off-by: Alan Robertson 
---
 lib/librte_sched/rte_sched.c | 41 -
 lib/librte_sched/rte_sched.h |  2 +-
 2 files changed, 37 insertions(+), 6 deletions(-)

diff --git a/lib/librte_sched/rte_sched.c b/lib/librte_sched/rte_sched.c
index 7252f850d..5c88f1b62 100644
--- a/lib/librte_sched/rte_sched.c
+++ b/lib/librte_sched/rte_sched.c
@@ -216,7 +216,7 @@ struct rte_sched_port {
uint32_t n_pipes_per_subport;
uint32_t rate;
uint32_t mtu;
-   uint32_t frame_overhead;
+   int32_t frame_overhead;
uint16_t qsize[RTE_SCHED_TRAFFIC_CLASSES_PER_PIPE];
uint32_t n_pipe_profiles;
uint32_t pipe_tc3_rate_max;
@@ -643,7 +643,14 @@ rte_sched_port_config(struct rte_sched_port_params *params)
port->n_subports_per_port = params->n_subports_per_port;
port->n_pipes_per_subport = params->n_pipes_per_subport;
port->rate = params->rate;
-   port->mtu = params->mtu + params->frame_overhead;
+
+   /* Only add a +ve overhead.  A -ve overhead is to accommodate
+* for downstream TDM devices which won't have an ethernet header,
+* they obviously won't impact our local interface MTU size.
+*/
+   port->mtu = params->mtu;
+   if (params->frame_overhead > 0)
+   port->mtu += params->frame_overhead;
port->frame_overhead = params->frame_overhead;
memcpy(port->qsize, params->qsize, sizeof(params->qsize));
port->n_pipe_profiles = params->n_pipe_profiles;
@@ -1613,13 +1620,21 @@ grinder_credits_check(struct rte_sched_port *port, 
uint32_t pos)
struct rte_sched_pipe *pipe = grinder->pipe;
struct rte_mbuf *pkt = grinder->pkt;
uint32_t tc_index = grinder->tc_index;
-   uint32_t pkt_len = pkt->pkt_len + port->frame_overhead;
+   uint32_t pkt_len;
+   int32_t tpkt_len;
uint32_t subport_tb_credits = subport->tb_credits;
uint32_t subport_tc_credits = subport->tc_credits[tc_index];
uint32_t pipe_tb_credits = pipe->tb_credits;
uint32_t pipe_tc_credits = pipe->tc_credits[tc_index];
int enough_credits;
 
+   /* Make sure we don't allow this to go -ve.  To accommodate
+* downstream TDM devices we may want to ignore the ethernet
+* header so allow -ve overhead values.
+*/
+   tpkt_len = pkt->pkt_len + port->frame_overhead;
+   pkt_len = (tpkt_len < 0) ? 1 : tpkt_len;
+
/* Check queue credits */
enough_credits = (pkt_len <= subport_tb_credits) &&
(pkt_len <= subport_tc_credits) &&
@@ -1648,7 +1663,8 @@ grinder_credits_check(struct rte_sched_port *port, 
uint32_t pos)
struct rte_sched_pipe *pipe = grinder->pipe;
struct rte_mbuf *pkt = grinder->pkt;
uint32_t tc_index = grinder->tc_index;
-   uint32_t pkt_len = pkt->pkt_len + port->frame_overhead;
+   uint32_t pkt_len;
+   int32_t tpkt_len;
uint32_t subport_tb_credits = subport->tb_credits;
uint32_t subport_tc_credits = subport->tc_credits[tc_index];
uint32_t pipe_tb_credits = pipe->tb_credits;
@@ -1658,6 +1674,13 @@ grinder_credits_check(struct rte_sched_port *port, 
uint32_t pos)
uint32_t pipe_tc_ov_credits = pipe_tc_ov_mask1[tc_index];
int enough_credits;
 
+   /* Make sure we don't allow this to go -ve.  To accommodate
+* downstream TDM devices we may want to ignore the ethernet
+* header so allow -ve overhead values.
+*/
+   tpkt_len = pkt->pkt_len + port->frame_overhead;
+   pkt_len = (tpkt_len < 0) ? 1 : tpkt_len;
+
/* Check pipe and subport credits */
enough_credits = (pkt_len <= subport_tb_credits) &&
(pkt_len <= subport_tc_credits) &&
@@ -1687,11 +1710,19 @@ grinder_schedule(struct rte_sched_port *port, uint32_t 
pos)
struct rte_sched_grinder *grinder = port->grinder + pos;
struct rte_sched_queue *queue = grinder->queue[grinder->qpos];
struct rte_mbuf *pkt = grinder->pkt;
-   uint32_t pkt_len = pkt->pkt_len + port->frame_overhead;
+   uint32_t pkt_len;
+   int32_t tpkt_len;
 
if (!grinder_credits_check(port, pos))
return 0;
 
+   /* Make sure we don't allow this to go -ve.  To accommodate
+* downstream TDM devices we may want to ignore the ethernet
+* header so allow -ve overhead values.
+*/
+   tpkt_len = pkt->pkt_len + port->frame_overhead;
+   pkt_len = (tpkt_len < 0) ? 1 : tpkt_len;
+
/* Advance port time */
port->time += pkt_len;
 
diff --git a/lib/librte_sched/rte_sched.h b/lib/librte_sched/rte_sched.h
index e9c281726..63677eefc 100644
--- a/lib/librte_sched/rte_sched.h
+++ b/lib/librte_sched/rte_sched.h
@@ -219,7 +219,7 @@ struct rte_sched_port_