linux/drivers/net/dsa/sja1105/sja1105_tas.c

896 lines
29 KiB
C
Raw Normal View History

net: dsa: sja1105: Configure the Time-Aware Scheduler via tc-taprio offload This qdisc offload is the closest thing to what the SJA1105 supports in hardware for time-based egress shaping. The switch core really is built around SAE AS6802/TTEthernet (a TTTech standard) but can be made to operate similarly to IEEE 802.1Qbv with some constraints: - The gate control list is a global list for all ports. There are 8 execution threads that iterate through this global list in parallel. I don't know why 8, there are only 4 front-panel ports. - Care must be taken by the user to make sure that two execution threads never get to execute a GCL entry simultaneously. I created a O(n^4) checker for this hardware limitation, prior to accepting a taprio offload configuration as valid. - The spec says that if a GCL entry's interval is shorter than the frame length, you shouldn't send it (and end up in head-of-line blocking). Well, this switch does anyway. - The switch has no concept of ADMIN and OPER configurations. Because it's so simple, the TAS settings are loaded through the static config tables interface, so there isn't even place for any discussion about 'graceful switchover between ADMIN and OPER'. You just reset the switch and upload a new OPER config. - The switch accepts multiple time sources for the gate events. Right now I am using the standalone clock source as opposed to PTP. So the base time parameter doesn't really do much. Support for the PTP clock source will be added in a future series. Signed-off-by: Vladimir Oltean <olteanv@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-09-15 10:00:02 +08:00
// SPDX-License-Identifier: GPL-2.0
/* Copyright (c) 2019, Vladimir Oltean <olteanv@gmail.com>
*/
#include "sja1105.h"
#define SJA1105_TAS_CLKSRC_DISABLED 0
#define SJA1105_TAS_CLKSRC_STANDALONE 1
#define SJA1105_TAS_CLKSRC_AS6802 2
#define SJA1105_TAS_CLKSRC_PTP 3
#define SJA1105_GATE_MASK GENMASK_ULL(SJA1105_NUM_TC - 1, 0)
net: dsa: sja1105: Implement state machine for TAS with PTP clock source Tested using the following bash script and the tc from iproute2-next: #!/bin/bash set -e -u -o pipefail NSEC_PER_SEC="1000000000" gatemask() { local tc_list="$1" local mask=0 for tc in ${tc_list}; do mask=$((${mask} | (1 << ${tc}))) done printf "%02x" ${mask} } if ! systemctl is-active --quiet ptp4l; then echo "Please start the ptp4l service" exit fi now=$(phc_ctl /dev/ptp1 get | gawk '/clock time is/ { print $5; }') # Phase-align the base time to the start of the next second. sec=$(echo "${now}" | gawk -F. '{ print $1; }') base_time="$(((${sec} + 1) * ${NSEC_PER_SEC}))" tc qdisc add dev swp5 parent root handle 100 taprio \ num_tc 8 \ map 0 1 2 3 5 6 7 \ queues 1@0 1@1 1@2 1@3 1@4 1@5 1@6 1@7 \ base-time ${base_time} \ sched-entry S $(gatemask 7) 100000 \ sched-entry S $(gatemask "0 1 2 3 4 5 6") 400000 \ clockid CLOCK_TAI flags 2 The "state machine" is a workqueue invoked after each manipulation command on the PTP clock (reset, adjust time, set time, adjust frequency) which checks over the state of the time-aware scheduler. So it is not monitored periodically, only in reaction to a PTP command typically triggered from a userspace daemon (linuxptp). Otherwise there is no reason for things to go wrong. Now that the timecounter/cyclecounter has been replaced with hardware operations on the PTP clock, the TAS Kconfig now depends upon PTP and the standalone clocksource operating mode has been removed. Signed-off-by: Vladimir Oltean <olteanv@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-11-12 08:11:54 +08:00
#define work_to_sja1105_tas(d) \
container_of((d), struct sja1105_tas_data, tas_work)
#define tas_to_sja1105(d) \
container_of((d), struct sja1105_private, tas_data)
static int sja1105_tas_set_runtime_params(struct sja1105_private *priv)
{
struct sja1105_tas_data *tas_data = &priv->tas_data;
net: dsa: sja1105: implement tc-gate using time-triggered virtual links Restrict the TTEthernet hardware support on this switch to operate as closely as possible to IEEE 802.1Qci as possible. This means that it can perform PTP-time-based ingress admission control on streams identified by {DMAC, VID, PCP}, which is useful when trying to ensure the determinism of traffic scheduled via IEEE 802.1Qbv. The oddity comes from the fact that in hardware (and in TTEthernet at large), virtual links always need a full-blown action, including not only the type of policing, but also the list of destination ports. So in practice, a single tc-gate action will result in all packets getting dropped. Additional actions (either "trap" or "redirect") need to be specified in the same filter rule such that the conforming packets are actually forwarded somewhere. Apart from the VL Lookup, Policing and Forwarding tables which need to be programmed for each flow (virtual link), the Schedule engine also needs to be told to open/close the admission gates for each individual virtual link. A fairly accurate (and detailed) description of how that works is already present in sja1105_tas.c, since it is already used to trigger the egress gates for the tc-taprio offload (IEEE 802.1Qbv). Key point here, we remember that the schedule engine supports 8 "subschedules" (execution threads that iterate through the global schedule in parallel, and that no 2 hardware threads must execute a schedule entry at the same time). For tc-taprio, each egress port used one of these 8 subschedules, leaving a total of 4 subschedules unused. In principle we could have allocated 1 subschedule for the tc-gate offload of each ingress port, but actually the schedules of all virtual links installed on each ingress port would have needed to be merged together, before they could have been programmed to hardware. So simplify our life and just merge the entire tc-gate configuration, for all virtual links on all ingress ports, into a single subschedule. Be sure to check that against the usual hardware scheduling conflicts, and program it to hardware alongside any tc-taprio subschedule that may be present. The following scenarios were tested: 1. Quantitative testing: tc qdisc add dev swp2 clsact tc filter add dev swp2 ingress flower skip_sw \ dst_mac 42:be:24:9b:76:20 \ action gate index 1 base-time 0 \ sched-entry OPEN 1200 -1 -1 \ sched-entry CLOSE 1200 -1 -1 \ action trap ping 192.168.1.2 -f PING 192.168.1.2 (192.168.1.2) 56(84) bytes of data. ............................. --- 192.168.1.2 ping statistics --- 948 packets transmitted, 467 received, 50.7384% packet loss, time 9671ms 2. Qualitative testing (with a phase-aligned schedule - the clocks are synchronized by ptp4l, not shown here): Receiver (sja1105): tc qdisc add dev swp2 clsact now=$(phc_ctl /dev/ptp1 get | awk '/clock time is/ {print $5}') && \ sec=$(echo $now | awk -F. '{print $1}') && \ base_time="$(((sec + 2) * 1000000000))" && \ echo "base time ${base_time}" tc filter add dev swp2 ingress flower skip_sw \ dst_mac 42:be:24:9b:76:20 \ action gate base-time ${base_time} \ sched-entry OPEN 60000 -1 -1 \ sched-entry CLOSE 40000 -1 -1 \ action trap Sender (enetc): now=$(phc_ctl /dev/ptp0 get | awk '/clock time is/ {print $5}') && \ sec=$(echo $now | awk -F. '{print $1}') && \ base_time="$(((sec + 2) * 1000000000))" && \ echo "base time ${base_time}" tc qdisc add dev eno0 parent root 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 ${base_time} \ sched-entry S 01 50000 \ sched-entry S 00 50000 \ flags 2 ping -A 192.168.1.1 PING 192.168.1.1 (192.168.1.1): 56 data bytes ... ^C --- 192.168.1.1 ping statistics --- 1425 packets transmitted, 1424 packets received, 0% packet loss round-trip min/avg/max = 0.322/0.361/0.990 ms And just for comparison, with the tc-taprio schedule deleted: ping -A 192.168.1.1 PING 192.168.1.1 (192.168.1.1): 56 data bytes ... ^C --- 192.168.1.1 ping statistics --- 33 packets transmitted, 19 packets received, 42% packet loss round-trip min/avg/max = 0.336/0.464/0.597 ms Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2020-05-06 03:20:56 +08:00
struct sja1105_gating_config *gating_cfg = &tas_data->gating_cfg;
net: dsa: sja1105: Implement state machine for TAS with PTP clock source Tested using the following bash script and the tc from iproute2-next: #!/bin/bash set -e -u -o pipefail NSEC_PER_SEC="1000000000" gatemask() { local tc_list="$1" local mask=0 for tc in ${tc_list}; do mask=$((${mask} | (1 << ${tc}))) done printf "%02x" ${mask} } if ! systemctl is-active --quiet ptp4l; then echo "Please start the ptp4l service" exit fi now=$(phc_ctl /dev/ptp1 get | gawk '/clock time is/ { print $5; }') # Phase-align the base time to the start of the next second. sec=$(echo "${now}" | gawk -F. '{ print $1; }') base_time="$(((${sec} + 1) * ${NSEC_PER_SEC}))" tc qdisc add dev swp5 parent root handle 100 taprio \ num_tc 8 \ map 0 1 2 3 5 6 7 \ queues 1@0 1@1 1@2 1@3 1@4 1@5 1@6 1@7 \ base-time ${base_time} \ sched-entry S $(gatemask 7) 100000 \ sched-entry S $(gatemask "0 1 2 3 4 5 6") 400000 \ clockid CLOCK_TAI flags 2 The "state machine" is a workqueue invoked after each manipulation command on the PTP clock (reset, adjust time, set time, adjust frequency) which checks over the state of the time-aware scheduler. So it is not monitored periodically, only in reaction to a PTP command typically triggered from a userspace daemon (linuxptp). Otherwise there is no reason for things to go wrong. Now that the timecounter/cyclecounter has been replaced with hardware operations on the PTP clock, the TAS Kconfig now depends upon PTP and the standalone clocksource operating mode has been removed. Signed-off-by: Vladimir Oltean <olteanv@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-11-12 08:11:54 +08:00
struct dsa_switch *ds = priv->ds;
s64 earliest_base_time = S64_MAX;
s64 latest_base_time = 0;
s64 its_cycle_time = 0;
s64 max_cycle_time = 0;
int port;
tas_data->enabled = false;
for (port = 0; port < SJA1105_NUM_PORTS; port++) {
const struct tc_taprio_qopt_offload *offload;
offload = tas_data->offload[port];
if (!offload)
continue;
tas_data->enabled = true;
if (max_cycle_time < offload->cycle_time)
max_cycle_time = offload->cycle_time;
if (latest_base_time < offload->base_time)
latest_base_time = offload->base_time;
if (earliest_base_time > offload->base_time) {
earliest_base_time = offload->base_time;
its_cycle_time = offload->cycle_time;
}
}
net: dsa: sja1105: implement tc-gate using time-triggered virtual links Restrict the TTEthernet hardware support on this switch to operate as closely as possible to IEEE 802.1Qci as possible. This means that it can perform PTP-time-based ingress admission control on streams identified by {DMAC, VID, PCP}, which is useful when trying to ensure the determinism of traffic scheduled via IEEE 802.1Qbv. The oddity comes from the fact that in hardware (and in TTEthernet at large), virtual links always need a full-blown action, including not only the type of policing, but also the list of destination ports. So in practice, a single tc-gate action will result in all packets getting dropped. Additional actions (either "trap" or "redirect") need to be specified in the same filter rule such that the conforming packets are actually forwarded somewhere. Apart from the VL Lookup, Policing and Forwarding tables which need to be programmed for each flow (virtual link), the Schedule engine also needs to be told to open/close the admission gates for each individual virtual link. A fairly accurate (and detailed) description of how that works is already present in sja1105_tas.c, since it is already used to trigger the egress gates for the tc-taprio offload (IEEE 802.1Qbv). Key point here, we remember that the schedule engine supports 8 "subschedules" (execution threads that iterate through the global schedule in parallel, and that no 2 hardware threads must execute a schedule entry at the same time). For tc-taprio, each egress port used one of these 8 subschedules, leaving a total of 4 subschedules unused. In principle we could have allocated 1 subschedule for the tc-gate offload of each ingress port, but actually the schedules of all virtual links installed on each ingress port would have needed to be merged together, before they could have been programmed to hardware. So simplify our life and just merge the entire tc-gate configuration, for all virtual links on all ingress ports, into a single subschedule. Be sure to check that against the usual hardware scheduling conflicts, and program it to hardware alongside any tc-taprio subschedule that may be present. The following scenarios were tested: 1. Quantitative testing: tc qdisc add dev swp2 clsact tc filter add dev swp2 ingress flower skip_sw \ dst_mac 42:be:24:9b:76:20 \ action gate index 1 base-time 0 \ sched-entry OPEN 1200 -1 -1 \ sched-entry CLOSE 1200 -1 -1 \ action trap ping 192.168.1.2 -f PING 192.168.1.2 (192.168.1.2) 56(84) bytes of data. ............................. --- 192.168.1.2 ping statistics --- 948 packets transmitted, 467 received, 50.7384% packet loss, time 9671ms 2. Qualitative testing (with a phase-aligned schedule - the clocks are synchronized by ptp4l, not shown here): Receiver (sja1105): tc qdisc add dev swp2 clsact now=$(phc_ctl /dev/ptp1 get | awk '/clock time is/ {print $5}') && \ sec=$(echo $now | awk -F. '{print $1}') && \ base_time="$(((sec + 2) * 1000000000))" && \ echo "base time ${base_time}" tc filter add dev swp2 ingress flower skip_sw \ dst_mac 42:be:24:9b:76:20 \ action gate base-time ${base_time} \ sched-entry OPEN 60000 -1 -1 \ sched-entry CLOSE 40000 -1 -1 \ action trap Sender (enetc): now=$(phc_ctl /dev/ptp0 get | awk '/clock time is/ {print $5}') && \ sec=$(echo $now | awk -F. '{print $1}') && \ base_time="$(((sec + 2) * 1000000000))" && \ echo "base time ${base_time}" tc qdisc add dev eno0 parent root 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 ${base_time} \ sched-entry S 01 50000 \ sched-entry S 00 50000 \ flags 2 ping -A 192.168.1.1 PING 192.168.1.1 (192.168.1.1): 56 data bytes ... ^C --- 192.168.1.1 ping statistics --- 1425 packets transmitted, 1424 packets received, 0% packet loss round-trip min/avg/max = 0.322/0.361/0.990 ms And just for comparison, with the tc-taprio schedule deleted: ping -A 192.168.1.1 PING 192.168.1.1 (192.168.1.1): 56 data bytes ... ^C --- 192.168.1.1 ping statistics --- 33 packets transmitted, 19 packets received, 42% packet loss round-trip min/avg/max = 0.336/0.464/0.597 ms Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2020-05-06 03:20:56 +08:00
if (!list_empty(&gating_cfg->entries)) {
tas_data->enabled = true;
if (max_cycle_time < gating_cfg->cycle_time)
max_cycle_time = gating_cfg->cycle_time;
if (latest_base_time < gating_cfg->base_time)
latest_base_time = gating_cfg->base_time;
if (earliest_base_time > gating_cfg->base_time) {
earliest_base_time = gating_cfg->base_time;
its_cycle_time = gating_cfg->cycle_time;
}
}
net: dsa: sja1105: Implement state machine for TAS with PTP clock source Tested using the following bash script and the tc from iproute2-next: #!/bin/bash set -e -u -o pipefail NSEC_PER_SEC="1000000000" gatemask() { local tc_list="$1" local mask=0 for tc in ${tc_list}; do mask=$((${mask} | (1 << ${tc}))) done printf "%02x" ${mask} } if ! systemctl is-active --quiet ptp4l; then echo "Please start the ptp4l service" exit fi now=$(phc_ctl /dev/ptp1 get | gawk '/clock time is/ { print $5; }') # Phase-align the base time to the start of the next second. sec=$(echo "${now}" | gawk -F. '{ print $1; }') base_time="$(((${sec} + 1) * ${NSEC_PER_SEC}))" tc qdisc add dev swp5 parent root handle 100 taprio \ num_tc 8 \ map 0 1 2 3 5 6 7 \ queues 1@0 1@1 1@2 1@3 1@4 1@5 1@6 1@7 \ base-time ${base_time} \ sched-entry S $(gatemask 7) 100000 \ sched-entry S $(gatemask "0 1 2 3 4 5 6") 400000 \ clockid CLOCK_TAI flags 2 The "state machine" is a workqueue invoked after each manipulation command on the PTP clock (reset, adjust time, set time, adjust frequency) which checks over the state of the time-aware scheduler. So it is not monitored periodically, only in reaction to a PTP command typically triggered from a userspace daemon (linuxptp). Otherwise there is no reason for things to go wrong. Now that the timecounter/cyclecounter has been replaced with hardware operations on the PTP clock, the TAS Kconfig now depends upon PTP and the standalone clocksource operating mode has been removed. Signed-off-by: Vladimir Oltean <olteanv@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-11-12 08:11:54 +08:00
if (!tas_data->enabled)
return 0;
/* Roll the earliest base time over until it is in a comparable
* time base with the latest, then compare their deltas.
* We want to enforce that all ports' base times are within
* SJA1105_TAS_MAX_DELTA 200ns cycles of one another.
*/
earliest_base_time = future_base_time(earliest_base_time,
its_cycle_time,
latest_base_time);
while (earliest_base_time > latest_base_time)
earliest_base_time -= its_cycle_time;
if (latest_base_time - earliest_base_time >
sja1105_delta_to_ns(SJA1105_TAS_MAX_DELTA)) {
dev_err(ds->dev,
"Base times too far apart: min %llu max %llu\n",
earliest_base_time, latest_base_time);
return -ERANGE;
}
tas_data->earliest_base_time = earliest_base_time;
tas_data->max_cycle_time = max_cycle_time;
dev_dbg(ds->dev, "earliest base time %lld ns\n", earliest_base_time);
dev_dbg(ds->dev, "latest base time %lld ns\n", latest_base_time);
dev_dbg(ds->dev, "longest cycle time %lld ns\n", max_cycle_time);
return 0;
}
net: dsa: sja1105: Configure the Time-Aware Scheduler via tc-taprio offload This qdisc offload is the closest thing to what the SJA1105 supports in hardware for time-based egress shaping. The switch core really is built around SAE AS6802/TTEthernet (a TTTech standard) but can be made to operate similarly to IEEE 802.1Qbv with some constraints: - The gate control list is a global list for all ports. There are 8 execution threads that iterate through this global list in parallel. I don't know why 8, there are only 4 front-panel ports. - Care must be taken by the user to make sure that two execution threads never get to execute a GCL entry simultaneously. I created a O(n^4) checker for this hardware limitation, prior to accepting a taprio offload configuration as valid. - The spec says that if a GCL entry's interval is shorter than the frame length, you shouldn't send it (and end up in head-of-line blocking). Well, this switch does anyway. - The switch has no concept of ADMIN and OPER configurations. Because it's so simple, the TAS settings are loaded through the static config tables interface, so there isn't even place for any discussion about 'graceful switchover between ADMIN and OPER'. You just reset the switch and upload a new OPER config. - The switch accepts multiple time sources for the gate events. Right now I am using the standalone clock source as opposed to PTP. So the base time parameter doesn't really do much. Support for the PTP clock source will be added in a future series. Signed-off-by: Vladimir Oltean <olteanv@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-09-15 10:00:02 +08:00
/* Lo and behold: the egress scheduler from hell.
*
* At the hardware level, the Time-Aware Shaper holds a global linear arrray of
* all schedule entries for all ports. These are the Gate Control List (GCL)
* entries, let's call them "timeslots" for short. This linear array of
* timeslots is held in BLK_IDX_SCHEDULE.
*
* Then there are a maximum of 8 "execution threads" inside the switch, which
* iterate cyclically through the "schedule". Each "cycle" has an entry point
* and an exit point, both being timeslot indices in the schedule table. The
* hardware calls each cycle a "subschedule".
*
* Subschedule (cycle) i starts when
* ptpclkval >= ptpschtm + BLK_IDX_SCHEDULE_ENTRY_POINTS[i].delta.
*
* The hardware scheduler iterates BLK_IDX_SCHEDULE with a k ranging from
* k = BLK_IDX_SCHEDULE_ENTRY_POINTS[i].address to
* k = BLK_IDX_SCHEDULE_PARAMS.subscheind[i]
*
* For each schedule entry (timeslot) k, the engine executes the gate control
* list entry for the duration of BLK_IDX_SCHEDULE[k].delta.
*
* +---------+
* | | BLK_IDX_SCHEDULE_ENTRY_POINTS_PARAMS
* +---------+
* |
* +-----------------+
* | .actsubsch
* BLK_IDX_SCHEDULE_ENTRY_POINTS v
* +-------+-------+
* |cycle 0|cycle 1|
* +-------+-------+
* | | | |
* +----------------+ | | +-------------------------------------+
* | .subschindx | | .subschindx |
* | | +---------------+ |
* | .address | .address | |
* | | | |
* | | | |
* | BLK_IDX_SCHEDULE v v |
* | +-------+-------+-------+-------+-------+------+ |
* | |entry 0|entry 1|entry 2|entry 3|entry 4|entry5| |
* | +-------+-------+-------+-------+-------+------+ |
* | ^ ^ ^ ^ |
* | | | | | |
* | +-------------------------+ | | | |
* | | +-------------------------------+ | | |
* | | | +-------------------+ | |
* | | | | | |
* | +---------------------------------------------------------------+ |
* | |subscheind[0]<=subscheind[1]<=subscheind[2]<=...<=subscheind[7]| |
* | +---------------------------------------------------------------+ |
* | ^ ^ BLK_IDX_SCHEDULE_PARAMS |
* | | | |
* +--------+ +-------------------------------------------+
*
* In the above picture there are two subschedules (cycles):
*
* - cycle 0: iterates the schedule table from 0 to 2 (and back)
* - cycle 1: iterates the schedule table from 3 to 5 (and back)
*
* All other possible execution threads must be marked as unused by making
* their "subschedule end index" (subscheind) equal to the last valid
* subschedule's end index (in this case 5).
*/
net: dsa: sja1105: implement tc-gate using time-triggered virtual links Restrict the TTEthernet hardware support on this switch to operate as closely as possible to IEEE 802.1Qci as possible. This means that it can perform PTP-time-based ingress admission control on streams identified by {DMAC, VID, PCP}, which is useful when trying to ensure the determinism of traffic scheduled via IEEE 802.1Qbv. The oddity comes from the fact that in hardware (and in TTEthernet at large), virtual links always need a full-blown action, including not only the type of policing, but also the list of destination ports. So in practice, a single tc-gate action will result in all packets getting dropped. Additional actions (either "trap" or "redirect") need to be specified in the same filter rule such that the conforming packets are actually forwarded somewhere. Apart from the VL Lookup, Policing and Forwarding tables which need to be programmed for each flow (virtual link), the Schedule engine also needs to be told to open/close the admission gates for each individual virtual link. A fairly accurate (and detailed) description of how that works is already present in sja1105_tas.c, since it is already used to trigger the egress gates for the tc-taprio offload (IEEE 802.1Qbv). Key point here, we remember that the schedule engine supports 8 "subschedules" (execution threads that iterate through the global schedule in parallel, and that no 2 hardware threads must execute a schedule entry at the same time). For tc-taprio, each egress port used one of these 8 subschedules, leaving a total of 4 subschedules unused. In principle we could have allocated 1 subschedule for the tc-gate offload of each ingress port, but actually the schedules of all virtual links installed on each ingress port would have needed to be merged together, before they could have been programmed to hardware. So simplify our life and just merge the entire tc-gate configuration, for all virtual links on all ingress ports, into a single subschedule. Be sure to check that against the usual hardware scheduling conflicts, and program it to hardware alongside any tc-taprio subschedule that may be present. The following scenarios were tested: 1. Quantitative testing: tc qdisc add dev swp2 clsact tc filter add dev swp2 ingress flower skip_sw \ dst_mac 42:be:24:9b:76:20 \ action gate index 1 base-time 0 \ sched-entry OPEN 1200 -1 -1 \ sched-entry CLOSE 1200 -1 -1 \ action trap ping 192.168.1.2 -f PING 192.168.1.2 (192.168.1.2) 56(84) bytes of data. ............................. --- 192.168.1.2 ping statistics --- 948 packets transmitted, 467 received, 50.7384% packet loss, time 9671ms 2. Qualitative testing (with a phase-aligned schedule - the clocks are synchronized by ptp4l, not shown here): Receiver (sja1105): tc qdisc add dev swp2 clsact now=$(phc_ctl /dev/ptp1 get | awk '/clock time is/ {print $5}') && \ sec=$(echo $now | awk -F. '{print $1}') && \ base_time="$(((sec + 2) * 1000000000))" && \ echo "base time ${base_time}" tc filter add dev swp2 ingress flower skip_sw \ dst_mac 42:be:24:9b:76:20 \ action gate base-time ${base_time} \ sched-entry OPEN 60000 -1 -1 \ sched-entry CLOSE 40000 -1 -1 \ action trap Sender (enetc): now=$(phc_ctl /dev/ptp0 get | awk '/clock time is/ {print $5}') && \ sec=$(echo $now | awk -F. '{print $1}') && \ base_time="$(((sec + 2) * 1000000000))" && \ echo "base time ${base_time}" tc qdisc add dev eno0 parent root 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 ${base_time} \ sched-entry S 01 50000 \ sched-entry S 00 50000 \ flags 2 ping -A 192.168.1.1 PING 192.168.1.1 (192.168.1.1): 56 data bytes ... ^C --- 192.168.1.1 ping statistics --- 1425 packets transmitted, 1424 packets received, 0% packet loss round-trip min/avg/max = 0.322/0.361/0.990 ms And just for comparison, with the tc-taprio schedule deleted: ping -A 192.168.1.1 PING 192.168.1.1 (192.168.1.1): 56 data bytes ... ^C --- 192.168.1.1 ping statistics --- 33 packets transmitted, 19 packets received, 42% packet loss round-trip min/avg/max = 0.336/0.464/0.597 ms Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2020-05-06 03:20:56 +08:00
int sja1105_init_scheduling(struct sja1105_private *priv)
net: dsa: sja1105: Configure the Time-Aware Scheduler via tc-taprio offload This qdisc offload is the closest thing to what the SJA1105 supports in hardware for time-based egress shaping. The switch core really is built around SAE AS6802/TTEthernet (a TTTech standard) but can be made to operate similarly to IEEE 802.1Qbv with some constraints: - The gate control list is a global list for all ports. There are 8 execution threads that iterate through this global list in parallel. I don't know why 8, there are only 4 front-panel ports. - Care must be taken by the user to make sure that two execution threads never get to execute a GCL entry simultaneously. I created a O(n^4) checker for this hardware limitation, prior to accepting a taprio offload configuration as valid. - The spec says that if a GCL entry's interval is shorter than the frame length, you shouldn't send it (and end up in head-of-line blocking). Well, this switch does anyway. - The switch has no concept of ADMIN and OPER configurations. Because it's so simple, the TAS settings are loaded through the static config tables interface, so there isn't even place for any discussion about 'graceful switchover between ADMIN and OPER'. You just reset the switch and upload a new OPER config. - The switch accepts multiple time sources for the gate events. Right now I am using the standalone clock source as opposed to PTP. So the base time parameter doesn't really do much. Support for the PTP clock source will be added in a future series. Signed-off-by: Vladimir Oltean <olteanv@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-09-15 10:00:02 +08:00
{
struct sja1105_schedule_entry_points_entry *schedule_entry_points;
struct sja1105_schedule_entry_points_params_entry
*schedule_entry_points_params;
struct sja1105_schedule_params_entry *schedule_params;
struct sja1105_tas_data *tas_data = &priv->tas_data;
net: dsa: sja1105: implement tc-gate using time-triggered virtual links Restrict the TTEthernet hardware support on this switch to operate as closely as possible to IEEE 802.1Qci as possible. This means that it can perform PTP-time-based ingress admission control on streams identified by {DMAC, VID, PCP}, which is useful when trying to ensure the determinism of traffic scheduled via IEEE 802.1Qbv. The oddity comes from the fact that in hardware (and in TTEthernet at large), virtual links always need a full-blown action, including not only the type of policing, but also the list of destination ports. So in practice, a single tc-gate action will result in all packets getting dropped. Additional actions (either "trap" or "redirect") need to be specified in the same filter rule such that the conforming packets are actually forwarded somewhere. Apart from the VL Lookup, Policing and Forwarding tables which need to be programmed for each flow (virtual link), the Schedule engine also needs to be told to open/close the admission gates for each individual virtual link. A fairly accurate (and detailed) description of how that works is already present in sja1105_tas.c, since it is already used to trigger the egress gates for the tc-taprio offload (IEEE 802.1Qbv). Key point here, we remember that the schedule engine supports 8 "subschedules" (execution threads that iterate through the global schedule in parallel, and that no 2 hardware threads must execute a schedule entry at the same time). For tc-taprio, each egress port used one of these 8 subschedules, leaving a total of 4 subschedules unused. In principle we could have allocated 1 subschedule for the tc-gate offload of each ingress port, but actually the schedules of all virtual links installed on each ingress port would have needed to be merged together, before they could have been programmed to hardware. So simplify our life and just merge the entire tc-gate configuration, for all virtual links on all ingress ports, into a single subschedule. Be sure to check that against the usual hardware scheduling conflicts, and program it to hardware alongside any tc-taprio subschedule that may be present. The following scenarios were tested: 1. Quantitative testing: tc qdisc add dev swp2 clsact tc filter add dev swp2 ingress flower skip_sw \ dst_mac 42:be:24:9b:76:20 \ action gate index 1 base-time 0 \ sched-entry OPEN 1200 -1 -1 \ sched-entry CLOSE 1200 -1 -1 \ action trap ping 192.168.1.2 -f PING 192.168.1.2 (192.168.1.2) 56(84) bytes of data. ............................. --- 192.168.1.2 ping statistics --- 948 packets transmitted, 467 received, 50.7384% packet loss, time 9671ms 2. Qualitative testing (with a phase-aligned schedule - the clocks are synchronized by ptp4l, not shown here): Receiver (sja1105): tc qdisc add dev swp2 clsact now=$(phc_ctl /dev/ptp1 get | awk '/clock time is/ {print $5}') && \ sec=$(echo $now | awk -F. '{print $1}') && \ base_time="$(((sec + 2) * 1000000000))" && \ echo "base time ${base_time}" tc filter add dev swp2 ingress flower skip_sw \ dst_mac 42:be:24:9b:76:20 \ action gate base-time ${base_time} \ sched-entry OPEN 60000 -1 -1 \ sched-entry CLOSE 40000 -1 -1 \ action trap Sender (enetc): now=$(phc_ctl /dev/ptp0 get | awk '/clock time is/ {print $5}') && \ sec=$(echo $now | awk -F. '{print $1}') && \ base_time="$(((sec + 2) * 1000000000))" && \ echo "base time ${base_time}" tc qdisc add dev eno0 parent root 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 ${base_time} \ sched-entry S 01 50000 \ sched-entry S 00 50000 \ flags 2 ping -A 192.168.1.1 PING 192.168.1.1 (192.168.1.1): 56 data bytes ... ^C --- 192.168.1.1 ping statistics --- 1425 packets transmitted, 1424 packets received, 0% packet loss round-trip min/avg/max = 0.322/0.361/0.990 ms And just for comparison, with the tc-taprio schedule deleted: ping -A 192.168.1.1 PING 192.168.1.1 (192.168.1.1): 56 data bytes ... ^C --- 192.168.1.1 ping statistics --- 33 packets transmitted, 19 packets received, 42% packet loss round-trip min/avg/max = 0.336/0.464/0.597 ms Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2020-05-06 03:20:56 +08:00
struct sja1105_gating_config *gating_cfg = &tas_data->gating_cfg;
net: dsa: sja1105: Configure the Time-Aware Scheduler via tc-taprio offload This qdisc offload is the closest thing to what the SJA1105 supports in hardware for time-based egress shaping. The switch core really is built around SAE AS6802/TTEthernet (a TTTech standard) but can be made to operate similarly to IEEE 802.1Qbv with some constraints: - The gate control list is a global list for all ports. There are 8 execution threads that iterate through this global list in parallel. I don't know why 8, there are only 4 front-panel ports. - Care must be taken by the user to make sure that two execution threads never get to execute a GCL entry simultaneously. I created a O(n^4) checker for this hardware limitation, prior to accepting a taprio offload configuration as valid. - The spec says that if a GCL entry's interval is shorter than the frame length, you shouldn't send it (and end up in head-of-line blocking). Well, this switch does anyway. - The switch has no concept of ADMIN and OPER configurations. Because it's so simple, the TAS settings are loaded through the static config tables interface, so there isn't even place for any discussion about 'graceful switchover between ADMIN and OPER'. You just reset the switch and upload a new OPER config. - The switch accepts multiple time sources for the gate events. Right now I am using the standalone clock source as opposed to PTP. So the base time parameter doesn't really do much. Support for the PTP clock source will be added in a future series. Signed-off-by: Vladimir Oltean <olteanv@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-09-15 10:00:02 +08:00
struct sja1105_schedule_entry *schedule;
struct sja1105_table *table;
int schedule_start_idx;
s64 entry_point_delta;
int schedule_end_idx;
int num_entries = 0;
int num_cycles = 0;
int cycle = 0;
int i, k = 0;
net: dsa: sja1105: Implement state machine for TAS with PTP clock source Tested using the following bash script and the tc from iproute2-next: #!/bin/bash set -e -u -o pipefail NSEC_PER_SEC="1000000000" gatemask() { local tc_list="$1" local mask=0 for tc in ${tc_list}; do mask=$((${mask} | (1 << ${tc}))) done printf "%02x" ${mask} } if ! systemctl is-active --quiet ptp4l; then echo "Please start the ptp4l service" exit fi now=$(phc_ctl /dev/ptp1 get | gawk '/clock time is/ { print $5; }') # Phase-align the base time to the start of the next second. sec=$(echo "${now}" | gawk -F. '{ print $1; }') base_time="$(((${sec} + 1) * ${NSEC_PER_SEC}))" tc qdisc add dev swp5 parent root handle 100 taprio \ num_tc 8 \ map 0 1 2 3 5 6 7 \ queues 1@0 1@1 1@2 1@3 1@4 1@5 1@6 1@7 \ base-time ${base_time} \ sched-entry S $(gatemask 7) 100000 \ sched-entry S $(gatemask "0 1 2 3 4 5 6") 400000 \ clockid CLOCK_TAI flags 2 The "state machine" is a workqueue invoked after each manipulation command on the PTP clock (reset, adjust time, set time, adjust frequency) which checks over the state of the time-aware scheduler. So it is not monitored periodically, only in reaction to a PTP command typically triggered from a userspace daemon (linuxptp). Otherwise there is no reason for things to go wrong. Now that the timecounter/cyclecounter has been replaced with hardware operations on the PTP clock, the TAS Kconfig now depends upon PTP and the standalone clocksource operating mode has been removed. Signed-off-by: Vladimir Oltean <olteanv@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-11-12 08:11:54 +08:00
int port, rc;
rc = sja1105_tas_set_runtime_params(priv);
if (rc < 0)
return rc;
net: dsa: sja1105: Configure the Time-Aware Scheduler via tc-taprio offload This qdisc offload is the closest thing to what the SJA1105 supports in hardware for time-based egress shaping. The switch core really is built around SAE AS6802/TTEthernet (a TTTech standard) but can be made to operate similarly to IEEE 802.1Qbv with some constraints: - The gate control list is a global list for all ports. There are 8 execution threads that iterate through this global list in parallel. I don't know why 8, there are only 4 front-panel ports. - Care must be taken by the user to make sure that two execution threads never get to execute a GCL entry simultaneously. I created a O(n^4) checker for this hardware limitation, prior to accepting a taprio offload configuration as valid. - The spec says that if a GCL entry's interval is shorter than the frame length, you shouldn't send it (and end up in head-of-line blocking). Well, this switch does anyway. - The switch has no concept of ADMIN and OPER configurations. Because it's so simple, the TAS settings are loaded through the static config tables interface, so there isn't even place for any discussion about 'graceful switchover between ADMIN and OPER'. You just reset the switch and upload a new OPER config. - The switch accepts multiple time sources for the gate events. Right now I am using the standalone clock source as opposed to PTP. So the base time parameter doesn't really do much. Support for the PTP clock source will be added in a future series. Signed-off-by: Vladimir Oltean <olteanv@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-09-15 10:00:02 +08:00
/* Discard previous Schedule Table */
table = &priv->static_config.tables[BLK_IDX_SCHEDULE];
if (table->entry_count) {
kfree(table->entries);
table->entry_count = 0;
}
/* Discard previous Schedule Entry Points Parameters Table */
table = &priv->static_config.tables[BLK_IDX_SCHEDULE_ENTRY_POINTS_PARAMS];
if (table->entry_count) {
kfree(table->entries);
table->entry_count = 0;
}
/* Discard previous Schedule Parameters Table */
table = &priv->static_config.tables[BLK_IDX_SCHEDULE_PARAMS];
if (table->entry_count) {
kfree(table->entries);
table->entry_count = 0;
}
/* Discard previous Schedule Entry Points Table */
table = &priv->static_config.tables[BLK_IDX_SCHEDULE_ENTRY_POINTS];
if (table->entry_count) {
kfree(table->entries);
table->entry_count = 0;
}
/* Figure out the dimensioning of the problem */
for (port = 0; port < SJA1105_NUM_PORTS; port++) {
if (tas_data->offload[port]) {
num_entries += tas_data->offload[port]->num_entries;
num_cycles++;
}
}
net: dsa: sja1105: implement tc-gate using time-triggered virtual links Restrict the TTEthernet hardware support on this switch to operate as closely as possible to IEEE 802.1Qci as possible. This means that it can perform PTP-time-based ingress admission control on streams identified by {DMAC, VID, PCP}, which is useful when trying to ensure the determinism of traffic scheduled via IEEE 802.1Qbv. The oddity comes from the fact that in hardware (and in TTEthernet at large), virtual links always need a full-blown action, including not only the type of policing, but also the list of destination ports. So in practice, a single tc-gate action will result in all packets getting dropped. Additional actions (either "trap" or "redirect") need to be specified in the same filter rule such that the conforming packets are actually forwarded somewhere. Apart from the VL Lookup, Policing and Forwarding tables which need to be programmed for each flow (virtual link), the Schedule engine also needs to be told to open/close the admission gates for each individual virtual link. A fairly accurate (and detailed) description of how that works is already present in sja1105_tas.c, since it is already used to trigger the egress gates for the tc-taprio offload (IEEE 802.1Qbv). Key point here, we remember that the schedule engine supports 8 "subschedules" (execution threads that iterate through the global schedule in parallel, and that no 2 hardware threads must execute a schedule entry at the same time). For tc-taprio, each egress port used one of these 8 subschedules, leaving a total of 4 subschedules unused. In principle we could have allocated 1 subschedule for the tc-gate offload of each ingress port, but actually the schedules of all virtual links installed on each ingress port would have needed to be merged together, before they could have been programmed to hardware. So simplify our life and just merge the entire tc-gate configuration, for all virtual links on all ingress ports, into a single subschedule. Be sure to check that against the usual hardware scheduling conflicts, and program it to hardware alongside any tc-taprio subschedule that may be present. The following scenarios were tested: 1. Quantitative testing: tc qdisc add dev swp2 clsact tc filter add dev swp2 ingress flower skip_sw \ dst_mac 42:be:24:9b:76:20 \ action gate index 1 base-time 0 \ sched-entry OPEN 1200 -1 -1 \ sched-entry CLOSE 1200 -1 -1 \ action trap ping 192.168.1.2 -f PING 192.168.1.2 (192.168.1.2) 56(84) bytes of data. ............................. --- 192.168.1.2 ping statistics --- 948 packets transmitted, 467 received, 50.7384% packet loss, time 9671ms 2. Qualitative testing (with a phase-aligned schedule - the clocks are synchronized by ptp4l, not shown here): Receiver (sja1105): tc qdisc add dev swp2 clsact now=$(phc_ctl /dev/ptp1 get | awk '/clock time is/ {print $5}') && \ sec=$(echo $now | awk -F. '{print $1}') && \ base_time="$(((sec + 2) * 1000000000))" && \ echo "base time ${base_time}" tc filter add dev swp2 ingress flower skip_sw \ dst_mac 42:be:24:9b:76:20 \ action gate base-time ${base_time} \ sched-entry OPEN 60000 -1 -1 \ sched-entry CLOSE 40000 -1 -1 \ action trap Sender (enetc): now=$(phc_ctl /dev/ptp0 get | awk '/clock time is/ {print $5}') && \ sec=$(echo $now | awk -F. '{print $1}') && \ base_time="$(((sec + 2) * 1000000000))" && \ echo "base time ${base_time}" tc qdisc add dev eno0 parent root 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 ${base_time} \ sched-entry S 01 50000 \ sched-entry S 00 50000 \ flags 2 ping -A 192.168.1.1 PING 192.168.1.1 (192.168.1.1): 56 data bytes ... ^C --- 192.168.1.1 ping statistics --- 1425 packets transmitted, 1424 packets received, 0% packet loss round-trip min/avg/max = 0.322/0.361/0.990 ms And just for comparison, with the tc-taprio schedule deleted: ping -A 192.168.1.1 PING 192.168.1.1 (192.168.1.1): 56 data bytes ... ^C --- 192.168.1.1 ping statistics --- 33 packets transmitted, 19 packets received, 42% packet loss round-trip min/avg/max = 0.336/0.464/0.597 ms Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2020-05-06 03:20:56 +08:00
if (!list_empty(&gating_cfg->entries)) {
num_entries += gating_cfg->num_entries;
num_cycles++;
}
net: dsa: sja1105: Configure the Time-Aware Scheduler via tc-taprio offload This qdisc offload is the closest thing to what the SJA1105 supports in hardware for time-based egress shaping. The switch core really is built around SAE AS6802/TTEthernet (a TTTech standard) but can be made to operate similarly to IEEE 802.1Qbv with some constraints: - The gate control list is a global list for all ports. There are 8 execution threads that iterate through this global list in parallel. I don't know why 8, there are only 4 front-panel ports. - Care must be taken by the user to make sure that two execution threads never get to execute a GCL entry simultaneously. I created a O(n^4) checker for this hardware limitation, prior to accepting a taprio offload configuration as valid. - The spec says that if a GCL entry's interval is shorter than the frame length, you shouldn't send it (and end up in head-of-line blocking). Well, this switch does anyway. - The switch has no concept of ADMIN and OPER configurations. Because it's so simple, the TAS settings are loaded through the static config tables interface, so there isn't even place for any discussion about 'graceful switchover between ADMIN and OPER'. You just reset the switch and upload a new OPER config. - The switch accepts multiple time sources for the gate events. Right now I am using the standalone clock source as opposed to PTP. So the base time parameter doesn't really do much. Support for the PTP clock source will be added in a future series. Signed-off-by: Vladimir Oltean <olteanv@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-09-15 10:00:02 +08:00
/* Nothing to do */
if (!num_cycles)
return 0;
/* Pre-allocate space in the static config tables */
/* Schedule Table */
table = &priv->static_config.tables[BLK_IDX_SCHEDULE];
table->entries = kcalloc(num_entries, table->ops->unpacked_entry_size,
GFP_KERNEL);
if (!table->entries)
return -ENOMEM;
table->entry_count = num_entries;
schedule = table->entries;
/* Schedule Points Parameters Table */
table = &priv->static_config.tables[BLK_IDX_SCHEDULE_ENTRY_POINTS_PARAMS];
table->entries = kcalloc(SJA1105_MAX_SCHEDULE_ENTRY_POINTS_PARAMS_COUNT,
table->ops->unpacked_entry_size, GFP_KERNEL);
if (!table->entries)
/* Previously allocated memory will be freed automatically in
* sja1105_static_config_free. This is true for all early
* returns below.
*/
return -ENOMEM;
table->entry_count = SJA1105_MAX_SCHEDULE_ENTRY_POINTS_PARAMS_COUNT;
schedule_entry_points_params = table->entries;
/* Schedule Parameters Table */
table = &priv->static_config.tables[BLK_IDX_SCHEDULE_PARAMS];
table->entries = kcalloc(SJA1105_MAX_SCHEDULE_PARAMS_COUNT,
table->ops->unpacked_entry_size, GFP_KERNEL);
if (!table->entries)
return -ENOMEM;
table->entry_count = SJA1105_MAX_SCHEDULE_PARAMS_COUNT;
schedule_params = table->entries;
/* Schedule Entry Points Table */
table = &priv->static_config.tables[BLK_IDX_SCHEDULE_ENTRY_POINTS];
table->entries = kcalloc(num_cycles, table->ops->unpacked_entry_size,
GFP_KERNEL);
if (!table->entries)
return -ENOMEM;
table->entry_count = num_cycles;
schedule_entry_points = table->entries;
/* Finally start populating the static config tables */
net: dsa: sja1105: Implement state machine for TAS with PTP clock source Tested using the following bash script and the tc from iproute2-next: #!/bin/bash set -e -u -o pipefail NSEC_PER_SEC="1000000000" gatemask() { local tc_list="$1" local mask=0 for tc in ${tc_list}; do mask=$((${mask} | (1 << ${tc}))) done printf "%02x" ${mask} } if ! systemctl is-active --quiet ptp4l; then echo "Please start the ptp4l service" exit fi now=$(phc_ctl /dev/ptp1 get | gawk '/clock time is/ { print $5; }') # Phase-align the base time to the start of the next second. sec=$(echo "${now}" | gawk -F. '{ print $1; }') base_time="$(((${sec} + 1) * ${NSEC_PER_SEC}))" tc qdisc add dev swp5 parent root handle 100 taprio \ num_tc 8 \ map 0 1 2 3 5 6 7 \ queues 1@0 1@1 1@2 1@3 1@4 1@5 1@6 1@7 \ base-time ${base_time} \ sched-entry S $(gatemask 7) 100000 \ sched-entry S $(gatemask "0 1 2 3 4 5 6") 400000 \ clockid CLOCK_TAI flags 2 The "state machine" is a workqueue invoked after each manipulation command on the PTP clock (reset, adjust time, set time, adjust frequency) which checks over the state of the time-aware scheduler. So it is not monitored periodically, only in reaction to a PTP command typically triggered from a userspace daemon (linuxptp). Otherwise there is no reason for things to go wrong. Now that the timecounter/cyclecounter has been replaced with hardware operations on the PTP clock, the TAS Kconfig now depends upon PTP and the standalone clocksource operating mode has been removed. Signed-off-by: Vladimir Oltean <olteanv@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-11-12 08:11:54 +08:00
schedule_entry_points_params->clksrc = SJA1105_TAS_CLKSRC_PTP;
net: dsa: sja1105: Configure the Time-Aware Scheduler via tc-taprio offload This qdisc offload is the closest thing to what the SJA1105 supports in hardware for time-based egress shaping. The switch core really is built around SAE AS6802/TTEthernet (a TTTech standard) but can be made to operate similarly to IEEE 802.1Qbv with some constraints: - The gate control list is a global list for all ports. There are 8 execution threads that iterate through this global list in parallel. I don't know why 8, there are only 4 front-panel ports. - Care must be taken by the user to make sure that two execution threads never get to execute a GCL entry simultaneously. I created a O(n^4) checker for this hardware limitation, prior to accepting a taprio offload configuration as valid. - The spec says that if a GCL entry's interval is shorter than the frame length, you shouldn't send it (and end up in head-of-line blocking). Well, this switch does anyway. - The switch has no concept of ADMIN and OPER configurations. Because it's so simple, the TAS settings are loaded through the static config tables interface, so there isn't even place for any discussion about 'graceful switchover between ADMIN and OPER'. You just reset the switch and upload a new OPER config. - The switch accepts multiple time sources for the gate events. Right now I am using the standalone clock source as opposed to PTP. So the base time parameter doesn't really do much. Support for the PTP clock source will be added in a future series. Signed-off-by: Vladimir Oltean <olteanv@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-09-15 10:00:02 +08:00
schedule_entry_points_params->actsubsch = num_cycles - 1;
for (port = 0; port < SJA1105_NUM_PORTS; port++) {
const struct tc_taprio_qopt_offload *offload;
net: dsa: sja1105: Implement state machine for TAS with PTP clock source Tested using the following bash script and the tc from iproute2-next: #!/bin/bash set -e -u -o pipefail NSEC_PER_SEC="1000000000" gatemask() { local tc_list="$1" local mask=0 for tc in ${tc_list}; do mask=$((${mask} | (1 << ${tc}))) done printf "%02x" ${mask} } if ! systemctl is-active --quiet ptp4l; then echo "Please start the ptp4l service" exit fi now=$(phc_ctl /dev/ptp1 get | gawk '/clock time is/ { print $5; }') # Phase-align the base time to the start of the next second. sec=$(echo "${now}" | gawk -F. '{ print $1; }') base_time="$(((${sec} + 1) * ${NSEC_PER_SEC}))" tc qdisc add dev swp5 parent root handle 100 taprio \ num_tc 8 \ map 0 1 2 3 5 6 7 \ queues 1@0 1@1 1@2 1@3 1@4 1@5 1@6 1@7 \ base-time ${base_time} \ sched-entry S $(gatemask 7) 100000 \ sched-entry S $(gatemask "0 1 2 3 4 5 6") 400000 \ clockid CLOCK_TAI flags 2 The "state machine" is a workqueue invoked after each manipulation command on the PTP clock (reset, adjust time, set time, adjust frequency) which checks over the state of the time-aware scheduler. So it is not monitored periodically, only in reaction to a PTP command typically triggered from a userspace daemon (linuxptp). Otherwise there is no reason for things to go wrong. Now that the timecounter/cyclecounter has been replaced with hardware operations on the PTP clock, the TAS Kconfig now depends upon PTP and the standalone clocksource operating mode has been removed. Signed-off-by: Vladimir Oltean <olteanv@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-11-12 08:11:54 +08:00
/* Relative base time */
s64 rbt;
net: dsa: sja1105: Configure the Time-Aware Scheduler via tc-taprio offload This qdisc offload is the closest thing to what the SJA1105 supports in hardware for time-based egress shaping. The switch core really is built around SAE AS6802/TTEthernet (a TTTech standard) but can be made to operate similarly to IEEE 802.1Qbv with some constraints: - The gate control list is a global list for all ports. There are 8 execution threads that iterate through this global list in parallel. I don't know why 8, there are only 4 front-panel ports. - Care must be taken by the user to make sure that two execution threads never get to execute a GCL entry simultaneously. I created a O(n^4) checker for this hardware limitation, prior to accepting a taprio offload configuration as valid. - The spec says that if a GCL entry's interval is shorter than the frame length, you shouldn't send it (and end up in head-of-line blocking). Well, this switch does anyway. - The switch has no concept of ADMIN and OPER configurations. Because it's so simple, the TAS settings are loaded through the static config tables interface, so there isn't even place for any discussion about 'graceful switchover between ADMIN and OPER'. You just reset the switch and upload a new OPER config. - The switch accepts multiple time sources for the gate events. Right now I am using the standalone clock source as opposed to PTP. So the base time parameter doesn't really do much. Support for the PTP clock source will be added in a future series. Signed-off-by: Vladimir Oltean <olteanv@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-09-15 10:00:02 +08:00
offload = tas_data->offload[port];
if (!offload)
continue;
schedule_start_idx = k;
schedule_end_idx = k + offload->num_entries - 1;
net: dsa: sja1105: Implement state machine for TAS with PTP clock source Tested using the following bash script and the tc from iproute2-next: #!/bin/bash set -e -u -o pipefail NSEC_PER_SEC="1000000000" gatemask() { local tc_list="$1" local mask=0 for tc in ${tc_list}; do mask=$((${mask} | (1 << ${tc}))) done printf "%02x" ${mask} } if ! systemctl is-active --quiet ptp4l; then echo "Please start the ptp4l service" exit fi now=$(phc_ctl /dev/ptp1 get | gawk '/clock time is/ { print $5; }') # Phase-align the base time to the start of the next second. sec=$(echo "${now}" | gawk -F. '{ print $1; }') base_time="$(((${sec} + 1) * ${NSEC_PER_SEC}))" tc qdisc add dev swp5 parent root handle 100 taprio \ num_tc 8 \ map 0 1 2 3 5 6 7 \ queues 1@0 1@1 1@2 1@3 1@4 1@5 1@6 1@7 \ base-time ${base_time} \ sched-entry S $(gatemask 7) 100000 \ sched-entry S $(gatemask "0 1 2 3 4 5 6") 400000 \ clockid CLOCK_TAI flags 2 The "state machine" is a workqueue invoked after each manipulation command on the PTP clock (reset, adjust time, set time, adjust frequency) which checks over the state of the time-aware scheduler. So it is not monitored periodically, only in reaction to a PTP command typically triggered from a userspace daemon (linuxptp). Otherwise there is no reason for things to go wrong. Now that the timecounter/cyclecounter has been replaced with hardware operations on the PTP clock, the TAS Kconfig now depends upon PTP and the standalone clocksource operating mode has been removed. Signed-off-by: Vladimir Oltean <olteanv@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-11-12 08:11:54 +08:00
/* This is the base time expressed as a number of TAS ticks
* relative to PTPSCHTM, which we'll (perhaps improperly) call
* the operational base time.
*/
rbt = future_base_time(offload->base_time,
offload->cycle_time,
tas_data->earliest_base_time);
rbt -= tas_data->earliest_base_time;
/* UM10944.pdf 4.2.2. Schedule Entry Points table says that
* delta cannot be zero, which is shitty. Advance all relative
* base times by 1 TAS delta, so that even the earliest base
* time becomes 1 in relative terms. Then start the operational
* base time (PTPSCHTM) one TAS delta earlier than planned.
net: dsa: sja1105: Configure the Time-Aware Scheduler via tc-taprio offload This qdisc offload is the closest thing to what the SJA1105 supports in hardware for time-based egress shaping. The switch core really is built around SAE AS6802/TTEthernet (a TTTech standard) but can be made to operate similarly to IEEE 802.1Qbv with some constraints: - The gate control list is a global list for all ports. There are 8 execution threads that iterate through this global list in parallel. I don't know why 8, there are only 4 front-panel ports. - Care must be taken by the user to make sure that two execution threads never get to execute a GCL entry simultaneously. I created a O(n^4) checker for this hardware limitation, prior to accepting a taprio offload configuration as valid. - The spec says that if a GCL entry's interval is shorter than the frame length, you shouldn't send it (and end up in head-of-line blocking). Well, this switch does anyway. - The switch has no concept of ADMIN and OPER configurations. Because it's so simple, the TAS settings are loaded through the static config tables interface, so there isn't even place for any discussion about 'graceful switchover between ADMIN and OPER'. You just reset the switch and upload a new OPER config. - The switch accepts multiple time sources for the gate events. Right now I am using the standalone clock source as opposed to PTP. So the base time parameter doesn't really do much. Support for the PTP clock source will be added in a future series. Signed-off-by: Vladimir Oltean <olteanv@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-09-15 10:00:02 +08:00
*/
net: dsa: sja1105: Implement state machine for TAS with PTP clock source Tested using the following bash script and the tc from iproute2-next: #!/bin/bash set -e -u -o pipefail NSEC_PER_SEC="1000000000" gatemask() { local tc_list="$1" local mask=0 for tc in ${tc_list}; do mask=$((${mask} | (1 << ${tc}))) done printf "%02x" ${mask} } if ! systemctl is-active --quiet ptp4l; then echo "Please start the ptp4l service" exit fi now=$(phc_ctl /dev/ptp1 get | gawk '/clock time is/ { print $5; }') # Phase-align the base time to the start of the next second. sec=$(echo "${now}" | gawk -F. '{ print $1; }') base_time="$(((${sec} + 1) * ${NSEC_PER_SEC}))" tc qdisc add dev swp5 parent root handle 100 taprio \ num_tc 8 \ map 0 1 2 3 5 6 7 \ queues 1@0 1@1 1@2 1@3 1@4 1@5 1@6 1@7 \ base-time ${base_time} \ sched-entry S $(gatemask 7) 100000 \ sched-entry S $(gatemask "0 1 2 3 4 5 6") 400000 \ clockid CLOCK_TAI flags 2 The "state machine" is a workqueue invoked after each manipulation command on the PTP clock (reset, adjust time, set time, adjust frequency) which checks over the state of the time-aware scheduler. So it is not monitored periodically, only in reaction to a PTP command typically triggered from a userspace daemon (linuxptp). Otherwise there is no reason for things to go wrong. Now that the timecounter/cyclecounter has been replaced with hardware operations on the PTP clock, the TAS Kconfig now depends upon PTP and the standalone clocksource operating mode has been removed. Signed-off-by: Vladimir Oltean <olteanv@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-11-12 08:11:54 +08:00
entry_point_delta = ns_to_sja1105_delta(rbt) + 1;
net: dsa: sja1105: Configure the Time-Aware Scheduler via tc-taprio offload This qdisc offload is the closest thing to what the SJA1105 supports in hardware for time-based egress shaping. The switch core really is built around SAE AS6802/TTEthernet (a TTTech standard) but can be made to operate similarly to IEEE 802.1Qbv with some constraints: - The gate control list is a global list for all ports. There are 8 execution threads that iterate through this global list in parallel. I don't know why 8, there are only 4 front-panel ports. - Care must be taken by the user to make sure that two execution threads never get to execute a GCL entry simultaneously. I created a O(n^4) checker for this hardware limitation, prior to accepting a taprio offload configuration as valid. - The spec says that if a GCL entry's interval is shorter than the frame length, you shouldn't send it (and end up in head-of-line blocking). Well, this switch does anyway. - The switch has no concept of ADMIN and OPER configurations. Because it's so simple, the TAS settings are loaded through the static config tables interface, so there isn't even place for any discussion about 'graceful switchover between ADMIN and OPER'. You just reset the switch and upload a new OPER config. - The switch accepts multiple time sources for the gate events. Right now I am using the standalone clock source as opposed to PTP. So the base time parameter doesn't really do much. Support for the PTP clock source will be added in a future series. Signed-off-by: Vladimir Oltean <olteanv@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-09-15 10:00:02 +08:00
schedule_entry_points[cycle].subschindx = cycle;
schedule_entry_points[cycle].delta = entry_point_delta;
schedule_entry_points[cycle].address = schedule_start_idx;
/* The subschedule end indices need to be
* monotonically increasing.
*/
for (i = cycle; i < 8; i++)
schedule_params->subscheind[i] = schedule_end_idx;
for (i = 0; i < offload->num_entries; i++, k++) {
s64 delta_ns = offload->entries[i].interval;
schedule[k].delta = ns_to_sja1105_delta(delta_ns);
schedule[k].destports = BIT(port);
schedule[k].resmedia_en = true;
schedule[k].resmedia = SJA1105_GATE_MASK &
~offload->entries[i].gate_mask;
}
cycle++;
}
net: dsa: sja1105: implement tc-gate using time-triggered virtual links Restrict the TTEthernet hardware support on this switch to operate as closely as possible to IEEE 802.1Qci as possible. This means that it can perform PTP-time-based ingress admission control on streams identified by {DMAC, VID, PCP}, which is useful when trying to ensure the determinism of traffic scheduled via IEEE 802.1Qbv. The oddity comes from the fact that in hardware (and in TTEthernet at large), virtual links always need a full-blown action, including not only the type of policing, but also the list of destination ports. So in practice, a single tc-gate action will result in all packets getting dropped. Additional actions (either "trap" or "redirect") need to be specified in the same filter rule such that the conforming packets are actually forwarded somewhere. Apart from the VL Lookup, Policing and Forwarding tables which need to be programmed for each flow (virtual link), the Schedule engine also needs to be told to open/close the admission gates for each individual virtual link. A fairly accurate (and detailed) description of how that works is already present in sja1105_tas.c, since it is already used to trigger the egress gates for the tc-taprio offload (IEEE 802.1Qbv). Key point here, we remember that the schedule engine supports 8 "subschedules" (execution threads that iterate through the global schedule in parallel, and that no 2 hardware threads must execute a schedule entry at the same time). For tc-taprio, each egress port used one of these 8 subschedules, leaving a total of 4 subschedules unused. In principle we could have allocated 1 subschedule for the tc-gate offload of each ingress port, but actually the schedules of all virtual links installed on each ingress port would have needed to be merged together, before they could have been programmed to hardware. So simplify our life and just merge the entire tc-gate configuration, for all virtual links on all ingress ports, into a single subschedule. Be sure to check that against the usual hardware scheduling conflicts, and program it to hardware alongside any tc-taprio subschedule that may be present. The following scenarios were tested: 1. Quantitative testing: tc qdisc add dev swp2 clsact tc filter add dev swp2 ingress flower skip_sw \ dst_mac 42:be:24:9b:76:20 \ action gate index 1 base-time 0 \ sched-entry OPEN 1200 -1 -1 \ sched-entry CLOSE 1200 -1 -1 \ action trap ping 192.168.1.2 -f PING 192.168.1.2 (192.168.1.2) 56(84) bytes of data. ............................. --- 192.168.1.2 ping statistics --- 948 packets transmitted, 467 received, 50.7384% packet loss, time 9671ms 2. Qualitative testing (with a phase-aligned schedule - the clocks are synchronized by ptp4l, not shown here): Receiver (sja1105): tc qdisc add dev swp2 clsact now=$(phc_ctl /dev/ptp1 get | awk '/clock time is/ {print $5}') && \ sec=$(echo $now | awk -F. '{print $1}') && \ base_time="$(((sec + 2) * 1000000000))" && \ echo "base time ${base_time}" tc filter add dev swp2 ingress flower skip_sw \ dst_mac 42:be:24:9b:76:20 \ action gate base-time ${base_time} \ sched-entry OPEN 60000 -1 -1 \ sched-entry CLOSE 40000 -1 -1 \ action trap Sender (enetc): now=$(phc_ctl /dev/ptp0 get | awk '/clock time is/ {print $5}') && \ sec=$(echo $now | awk -F. '{print $1}') && \ base_time="$(((sec + 2) * 1000000000))" && \ echo "base time ${base_time}" tc qdisc add dev eno0 parent root 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 ${base_time} \ sched-entry S 01 50000 \ sched-entry S 00 50000 \ flags 2 ping -A 192.168.1.1 PING 192.168.1.1 (192.168.1.1): 56 data bytes ... ^C --- 192.168.1.1 ping statistics --- 1425 packets transmitted, 1424 packets received, 0% packet loss round-trip min/avg/max = 0.322/0.361/0.990 ms And just for comparison, with the tc-taprio schedule deleted: ping -A 192.168.1.1 PING 192.168.1.1 (192.168.1.1): 56 data bytes ... ^C --- 192.168.1.1 ping statistics --- 33 packets transmitted, 19 packets received, 42% packet loss round-trip min/avg/max = 0.336/0.464/0.597 ms Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2020-05-06 03:20:56 +08:00
if (!list_empty(&gating_cfg->entries)) {
struct sja1105_gate_entry *e;
/* Relative base time */
s64 rbt;
schedule_start_idx = k;
schedule_end_idx = k + gating_cfg->num_entries - 1;
rbt = future_base_time(gating_cfg->base_time,
gating_cfg->cycle_time,
tas_data->earliest_base_time);
rbt -= tas_data->earliest_base_time;
entry_point_delta = ns_to_sja1105_delta(rbt) + 1;
schedule_entry_points[cycle].subschindx = cycle;
schedule_entry_points[cycle].delta = entry_point_delta;
schedule_entry_points[cycle].address = schedule_start_idx;
for (i = cycle; i < 8; i++)
schedule_params->subscheind[i] = schedule_end_idx;
list_for_each_entry(e, &gating_cfg->entries, list) {
schedule[k].delta = ns_to_sja1105_delta(e->interval);
schedule[k].destports = e->rule->vl.destports;
schedule[k].setvalid = true;
schedule[k].txen = true;
schedule[k].vlindex = e->rule->vl.sharindx;
schedule[k].winstindex = e->rule->vl.sharindx;
if (e->gate_state) /* Gate open */
schedule[k].winst = true;
else /* Gate closed */
schedule[k].winend = true;
k++;
}
}
net: dsa: sja1105: Configure the Time-Aware Scheduler via tc-taprio offload This qdisc offload is the closest thing to what the SJA1105 supports in hardware for time-based egress shaping. The switch core really is built around SAE AS6802/TTEthernet (a TTTech standard) but can be made to operate similarly to IEEE 802.1Qbv with some constraints: - The gate control list is a global list for all ports. There are 8 execution threads that iterate through this global list in parallel. I don't know why 8, there are only 4 front-panel ports. - Care must be taken by the user to make sure that two execution threads never get to execute a GCL entry simultaneously. I created a O(n^4) checker for this hardware limitation, prior to accepting a taprio offload configuration as valid. - The spec says that if a GCL entry's interval is shorter than the frame length, you shouldn't send it (and end up in head-of-line blocking). Well, this switch does anyway. - The switch has no concept of ADMIN and OPER configurations. Because it's so simple, the TAS settings are loaded through the static config tables interface, so there isn't even place for any discussion about 'graceful switchover between ADMIN and OPER'. You just reset the switch and upload a new OPER config. - The switch accepts multiple time sources for the gate events. Right now I am using the standalone clock source as opposed to PTP. So the base time parameter doesn't really do much. Support for the PTP clock source will be added in a future series. Signed-off-by: Vladimir Oltean <olteanv@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-09-15 10:00:02 +08:00
return 0;
}
/* Be there 2 port subschedules, each executing an arbitrary number of gate
* open/close events cyclically.
* None of those gate events must ever occur at the exact same time, otherwise
* the switch is known to act in exotically strange ways.
* However the hardware doesn't bother performing these integrity checks.
* So here we are with the task of validating whether the new @admin offload
* has any conflict with the already established TAS configuration in
* tas_data->offload. We already know the other ports are in harmony with one
* another, otherwise we wouldn't have saved them.
* Each gate event executes periodically, with a period of @cycle_time and a
* phase given by its cycle's @base_time plus its offset within the cycle
* (which in turn is given by the length of the events prior to it).
* There are two aspects to possible collisions:
* - Collisions within one cycle's (actually the longest cycle's) time frame.
* For that, we need to compare the cartesian product of each possible
* occurrence of each event within one cycle time.
* - Collisions in the future. Events may not collide within one cycle time,
* but if two port schedules don't have the same periodicity (aka the cycle
* times aren't multiples of one another), they surely will some time in the
* future (actually they will collide an infinite amount of times).
*/
static bool
sja1105_tas_check_conflicts(struct sja1105_private *priv, int port,
const struct tc_taprio_qopt_offload *admin)
{
struct sja1105_tas_data *tas_data = &priv->tas_data;
const struct tc_taprio_qopt_offload *offload;
s64 max_cycle_time, min_cycle_time;
s64 delta1, delta2;
s64 rbt1, rbt2;
s64 stop_time;
s64 t1, t2;
int i, j;
s32 rem;
offload = tas_data->offload[port];
if (!offload)
return false;
/* Check if the two cycle times are multiples of one another.
* If they aren't, then they will surely collide.
*/
max_cycle_time = max(offload->cycle_time, admin->cycle_time);
min_cycle_time = min(offload->cycle_time, admin->cycle_time);
div_s64_rem(max_cycle_time, min_cycle_time, &rem);
if (rem)
return true;
/* Calculate the "reduced" base time of each of the two cycles
* (transposed back as close to 0 as possible) by dividing to
* the cycle time.
*/
div_s64_rem(offload->base_time, offload->cycle_time, &rem);
rbt1 = rem;
div_s64_rem(admin->base_time, admin->cycle_time, &rem);
rbt2 = rem;
stop_time = max_cycle_time + max(rbt1, rbt2);
/* delta1 is the relative base time of each GCL entry within
* the established ports' TAS config.
*/
for (i = 0, delta1 = 0;
i < offload->num_entries;
delta1 += offload->entries[i].interval, i++) {
/* delta2 is the relative base time of each GCL entry
* within the newly added TAS config.
*/
for (j = 0, delta2 = 0;
j < admin->num_entries;
delta2 += admin->entries[j].interval, j++) {
/* t1 follows all possible occurrences of the
* established ports' GCL entry i within the
* first cycle time.
*/
for (t1 = rbt1 + delta1;
t1 <= stop_time;
t1 += offload->cycle_time) {
/* t2 follows all possible occurrences
* of the newly added GCL entry j
* within the first cycle time.
*/
for (t2 = rbt2 + delta2;
t2 <= stop_time;
t2 += admin->cycle_time) {
if (t1 == t2) {
dev_warn(priv->ds->dev,
"GCL entry %d collides with entry %d of port %d\n",
j, i, port);
return true;
}
}
}
}
}
return false;
}
net: dsa: sja1105: implement tc-gate using time-triggered virtual links Restrict the TTEthernet hardware support on this switch to operate as closely as possible to IEEE 802.1Qci as possible. This means that it can perform PTP-time-based ingress admission control on streams identified by {DMAC, VID, PCP}, which is useful when trying to ensure the determinism of traffic scheduled via IEEE 802.1Qbv. The oddity comes from the fact that in hardware (and in TTEthernet at large), virtual links always need a full-blown action, including not only the type of policing, but also the list of destination ports. So in practice, a single tc-gate action will result in all packets getting dropped. Additional actions (either "trap" or "redirect") need to be specified in the same filter rule such that the conforming packets are actually forwarded somewhere. Apart from the VL Lookup, Policing and Forwarding tables which need to be programmed for each flow (virtual link), the Schedule engine also needs to be told to open/close the admission gates for each individual virtual link. A fairly accurate (and detailed) description of how that works is already present in sja1105_tas.c, since it is already used to trigger the egress gates for the tc-taprio offload (IEEE 802.1Qbv). Key point here, we remember that the schedule engine supports 8 "subschedules" (execution threads that iterate through the global schedule in parallel, and that no 2 hardware threads must execute a schedule entry at the same time). For tc-taprio, each egress port used one of these 8 subschedules, leaving a total of 4 subschedules unused. In principle we could have allocated 1 subschedule for the tc-gate offload of each ingress port, but actually the schedules of all virtual links installed on each ingress port would have needed to be merged together, before they could have been programmed to hardware. So simplify our life and just merge the entire tc-gate configuration, for all virtual links on all ingress ports, into a single subschedule. Be sure to check that against the usual hardware scheduling conflicts, and program it to hardware alongside any tc-taprio subschedule that may be present. The following scenarios were tested: 1. Quantitative testing: tc qdisc add dev swp2 clsact tc filter add dev swp2 ingress flower skip_sw \ dst_mac 42:be:24:9b:76:20 \ action gate index 1 base-time 0 \ sched-entry OPEN 1200 -1 -1 \ sched-entry CLOSE 1200 -1 -1 \ action trap ping 192.168.1.2 -f PING 192.168.1.2 (192.168.1.2) 56(84) bytes of data. ............................. --- 192.168.1.2 ping statistics --- 948 packets transmitted, 467 received, 50.7384% packet loss, time 9671ms 2. Qualitative testing (with a phase-aligned schedule - the clocks are synchronized by ptp4l, not shown here): Receiver (sja1105): tc qdisc add dev swp2 clsact now=$(phc_ctl /dev/ptp1 get | awk '/clock time is/ {print $5}') && \ sec=$(echo $now | awk -F. '{print $1}') && \ base_time="$(((sec + 2) * 1000000000))" && \ echo "base time ${base_time}" tc filter add dev swp2 ingress flower skip_sw \ dst_mac 42:be:24:9b:76:20 \ action gate base-time ${base_time} \ sched-entry OPEN 60000 -1 -1 \ sched-entry CLOSE 40000 -1 -1 \ action trap Sender (enetc): now=$(phc_ctl /dev/ptp0 get | awk '/clock time is/ {print $5}') && \ sec=$(echo $now | awk -F. '{print $1}') && \ base_time="$(((sec + 2) * 1000000000))" && \ echo "base time ${base_time}" tc qdisc add dev eno0 parent root 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 ${base_time} \ sched-entry S 01 50000 \ sched-entry S 00 50000 \ flags 2 ping -A 192.168.1.1 PING 192.168.1.1 (192.168.1.1): 56 data bytes ... ^C --- 192.168.1.1 ping statistics --- 1425 packets transmitted, 1424 packets received, 0% packet loss round-trip min/avg/max = 0.322/0.361/0.990 ms And just for comparison, with the tc-taprio schedule deleted: ping -A 192.168.1.1 PING 192.168.1.1 (192.168.1.1): 56 data bytes ... ^C --- 192.168.1.1 ping statistics --- 33 packets transmitted, 19 packets received, 42% packet loss round-trip min/avg/max = 0.336/0.464/0.597 ms Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2020-05-06 03:20:56 +08:00
/* Check the tc-taprio configuration on @port for conflicts with the tc-gate
* global subschedule. If @port is -1, check it against all ports.
* To reuse the sja1105_tas_check_conflicts logic without refactoring it,
* convert the gating configuration to a dummy tc-taprio offload structure.
*/
bool sja1105_gating_check_conflicts(struct sja1105_private *priv, int port,
struct netlink_ext_ack *extack)
{
struct sja1105_gating_config *gating_cfg = &priv->tas_data.gating_cfg;
size_t num_entries = gating_cfg->num_entries;
struct tc_taprio_qopt_offload *dummy;
struct sja1105_gate_entry *e;
bool conflict;
int i = 0;
if (list_empty(&gating_cfg->entries))
return false;
dummy = kzalloc(struct_size(dummy, entries, num_entries), GFP_KERNEL);
net: dsa: sja1105: implement tc-gate using time-triggered virtual links Restrict the TTEthernet hardware support on this switch to operate as closely as possible to IEEE 802.1Qci as possible. This means that it can perform PTP-time-based ingress admission control on streams identified by {DMAC, VID, PCP}, which is useful when trying to ensure the determinism of traffic scheduled via IEEE 802.1Qbv. The oddity comes from the fact that in hardware (and in TTEthernet at large), virtual links always need a full-blown action, including not only the type of policing, but also the list of destination ports. So in practice, a single tc-gate action will result in all packets getting dropped. Additional actions (either "trap" or "redirect") need to be specified in the same filter rule such that the conforming packets are actually forwarded somewhere. Apart from the VL Lookup, Policing and Forwarding tables which need to be programmed for each flow (virtual link), the Schedule engine also needs to be told to open/close the admission gates for each individual virtual link. A fairly accurate (and detailed) description of how that works is already present in sja1105_tas.c, since it is already used to trigger the egress gates for the tc-taprio offload (IEEE 802.1Qbv). Key point here, we remember that the schedule engine supports 8 "subschedules" (execution threads that iterate through the global schedule in parallel, and that no 2 hardware threads must execute a schedule entry at the same time). For tc-taprio, each egress port used one of these 8 subschedules, leaving a total of 4 subschedules unused. In principle we could have allocated 1 subschedule for the tc-gate offload of each ingress port, but actually the schedules of all virtual links installed on each ingress port would have needed to be merged together, before they could have been programmed to hardware. So simplify our life and just merge the entire tc-gate configuration, for all virtual links on all ingress ports, into a single subschedule. Be sure to check that against the usual hardware scheduling conflicts, and program it to hardware alongside any tc-taprio subschedule that may be present. The following scenarios were tested: 1. Quantitative testing: tc qdisc add dev swp2 clsact tc filter add dev swp2 ingress flower skip_sw \ dst_mac 42:be:24:9b:76:20 \ action gate index 1 base-time 0 \ sched-entry OPEN 1200 -1 -1 \ sched-entry CLOSE 1200 -1 -1 \ action trap ping 192.168.1.2 -f PING 192.168.1.2 (192.168.1.2) 56(84) bytes of data. ............................. --- 192.168.1.2 ping statistics --- 948 packets transmitted, 467 received, 50.7384% packet loss, time 9671ms 2. Qualitative testing (with a phase-aligned schedule - the clocks are synchronized by ptp4l, not shown here): Receiver (sja1105): tc qdisc add dev swp2 clsact now=$(phc_ctl /dev/ptp1 get | awk '/clock time is/ {print $5}') && \ sec=$(echo $now | awk -F. '{print $1}') && \ base_time="$(((sec + 2) * 1000000000))" && \ echo "base time ${base_time}" tc filter add dev swp2 ingress flower skip_sw \ dst_mac 42:be:24:9b:76:20 \ action gate base-time ${base_time} \ sched-entry OPEN 60000 -1 -1 \ sched-entry CLOSE 40000 -1 -1 \ action trap Sender (enetc): now=$(phc_ctl /dev/ptp0 get | awk '/clock time is/ {print $5}') && \ sec=$(echo $now | awk -F. '{print $1}') && \ base_time="$(((sec + 2) * 1000000000))" && \ echo "base time ${base_time}" tc qdisc add dev eno0 parent root 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 ${base_time} \ sched-entry S 01 50000 \ sched-entry S 00 50000 \ flags 2 ping -A 192.168.1.1 PING 192.168.1.1 (192.168.1.1): 56 data bytes ... ^C --- 192.168.1.1 ping statistics --- 1425 packets transmitted, 1424 packets received, 0% packet loss round-trip min/avg/max = 0.322/0.361/0.990 ms And just for comparison, with the tc-taprio schedule deleted: ping -A 192.168.1.1 PING 192.168.1.1 (192.168.1.1): 56 data bytes ... ^C --- 192.168.1.1 ping statistics --- 33 packets transmitted, 19 packets received, 42% packet loss round-trip min/avg/max = 0.336/0.464/0.597 ms Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2020-05-06 03:20:56 +08:00
if (!dummy) {
NL_SET_ERR_MSG_MOD(extack, "Failed to allocate memory");
return true;
}
dummy->num_entries = num_entries;
dummy->base_time = gating_cfg->base_time;
dummy->cycle_time = gating_cfg->cycle_time;
list_for_each_entry(e, &gating_cfg->entries, list)
dummy->entries[i++].interval = e->interval;
if (port != -1) {
conflict = sja1105_tas_check_conflicts(priv, port, dummy);
} else {
for (port = 0; port < SJA1105_NUM_PORTS; port++) {
conflict = sja1105_tas_check_conflicts(priv, port,
dummy);
if (conflict)
break;
}
}
kfree(dummy);
return conflict;
}
net: dsa: sja1105: Configure the Time-Aware Scheduler via tc-taprio offload This qdisc offload is the closest thing to what the SJA1105 supports in hardware for time-based egress shaping. The switch core really is built around SAE AS6802/TTEthernet (a TTTech standard) but can be made to operate similarly to IEEE 802.1Qbv with some constraints: - The gate control list is a global list for all ports. There are 8 execution threads that iterate through this global list in parallel. I don't know why 8, there are only 4 front-panel ports. - Care must be taken by the user to make sure that two execution threads never get to execute a GCL entry simultaneously. I created a O(n^4) checker for this hardware limitation, prior to accepting a taprio offload configuration as valid. - The spec says that if a GCL entry's interval is shorter than the frame length, you shouldn't send it (and end up in head-of-line blocking). Well, this switch does anyway. - The switch has no concept of ADMIN and OPER configurations. Because it's so simple, the TAS settings are loaded through the static config tables interface, so there isn't even place for any discussion about 'graceful switchover between ADMIN and OPER'. You just reset the switch and upload a new OPER config. - The switch accepts multiple time sources for the gate events. Right now I am using the standalone clock source as opposed to PTP. So the base time parameter doesn't really do much. Support for the PTP clock source will be added in a future series. Signed-off-by: Vladimir Oltean <olteanv@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-09-15 10:00:02 +08:00
int sja1105_setup_tc_taprio(struct dsa_switch *ds, int port,
struct tc_taprio_qopt_offload *admin)
{
struct sja1105_private *priv = ds->priv;
struct sja1105_tas_data *tas_data = &priv->tas_data;
int other_port, rc, i;
/* Can't change an already configured port (must delete qdisc first).
* Can't delete the qdisc from an unconfigured port.
*/
if (!!tas_data->offload[port] == admin->enable)
return -EINVAL;
if (!admin->enable) {
taprio_offload_free(tas_data->offload[port]);
tas_data->offload[port] = NULL;
rc = sja1105_init_scheduling(priv);
if (rc < 0)
return rc;
return sja1105_static_config_reload(priv, SJA1105_SCHEDULING);
net: dsa: sja1105: Configure the Time-Aware Scheduler via tc-taprio offload This qdisc offload is the closest thing to what the SJA1105 supports in hardware for time-based egress shaping. The switch core really is built around SAE AS6802/TTEthernet (a TTTech standard) but can be made to operate similarly to IEEE 802.1Qbv with some constraints: - The gate control list is a global list for all ports. There are 8 execution threads that iterate through this global list in parallel. I don't know why 8, there are only 4 front-panel ports. - Care must be taken by the user to make sure that two execution threads never get to execute a GCL entry simultaneously. I created a O(n^4) checker for this hardware limitation, prior to accepting a taprio offload configuration as valid. - The spec says that if a GCL entry's interval is shorter than the frame length, you shouldn't send it (and end up in head-of-line blocking). Well, this switch does anyway. - The switch has no concept of ADMIN and OPER configurations. Because it's so simple, the TAS settings are loaded through the static config tables interface, so there isn't even place for any discussion about 'graceful switchover between ADMIN and OPER'. You just reset the switch and upload a new OPER config. - The switch accepts multiple time sources for the gate events. Right now I am using the standalone clock source as opposed to PTP. So the base time parameter doesn't really do much. Support for the PTP clock source will be added in a future series. Signed-off-by: Vladimir Oltean <olteanv@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-09-15 10:00:02 +08:00
}
/* The cycle time extension is the amount of time the last cycle from
* the old OPER needs to be extended in order to phase-align with the
* base time of the ADMIN when that becomes the new OPER.
* But of course our switch needs to be reset to switch-over between
* the ADMIN and the OPER configs - so much for a seamless transition.
* So don't add insult over injury and just say we don't support cycle
* time extension.
*/
if (admin->cycle_time_extension)
return -ENOTSUPP;
for (i = 0; i < admin->num_entries; i++) {
s64 delta_ns = admin->entries[i].interval;
s64 delta_cycles = ns_to_sja1105_delta(delta_ns);
bool too_long, too_short;
too_long = (delta_cycles >= SJA1105_TAS_MAX_DELTA);
too_short = (delta_cycles == 0);
if (too_long || too_short) {
dev_err(priv->ds->dev,
"Interval %llu too %s for GCL entry %d\n",
delta_ns, too_long ? "long" : "short", i);
return -ERANGE;
}
}
for (other_port = 0; other_port < SJA1105_NUM_PORTS; other_port++) {
if (other_port == port)
continue;
if (sja1105_tas_check_conflicts(priv, other_port, admin))
return -ERANGE;
}
net: dsa: sja1105: implement tc-gate using time-triggered virtual links Restrict the TTEthernet hardware support on this switch to operate as closely as possible to IEEE 802.1Qci as possible. This means that it can perform PTP-time-based ingress admission control on streams identified by {DMAC, VID, PCP}, which is useful when trying to ensure the determinism of traffic scheduled via IEEE 802.1Qbv. The oddity comes from the fact that in hardware (and in TTEthernet at large), virtual links always need a full-blown action, including not only the type of policing, but also the list of destination ports. So in practice, a single tc-gate action will result in all packets getting dropped. Additional actions (either "trap" or "redirect") need to be specified in the same filter rule such that the conforming packets are actually forwarded somewhere. Apart from the VL Lookup, Policing and Forwarding tables which need to be programmed for each flow (virtual link), the Schedule engine also needs to be told to open/close the admission gates for each individual virtual link. A fairly accurate (and detailed) description of how that works is already present in sja1105_tas.c, since it is already used to trigger the egress gates for the tc-taprio offload (IEEE 802.1Qbv). Key point here, we remember that the schedule engine supports 8 "subschedules" (execution threads that iterate through the global schedule in parallel, and that no 2 hardware threads must execute a schedule entry at the same time). For tc-taprio, each egress port used one of these 8 subschedules, leaving a total of 4 subschedules unused. In principle we could have allocated 1 subschedule for the tc-gate offload of each ingress port, but actually the schedules of all virtual links installed on each ingress port would have needed to be merged together, before they could have been programmed to hardware. So simplify our life and just merge the entire tc-gate configuration, for all virtual links on all ingress ports, into a single subschedule. Be sure to check that against the usual hardware scheduling conflicts, and program it to hardware alongside any tc-taprio subschedule that may be present. The following scenarios were tested: 1. Quantitative testing: tc qdisc add dev swp2 clsact tc filter add dev swp2 ingress flower skip_sw \ dst_mac 42:be:24:9b:76:20 \ action gate index 1 base-time 0 \ sched-entry OPEN 1200 -1 -1 \ sched-entry CLOSE 1200 -1 -1 \ action trap ping 192.168.1.2 -f PING 192.168.1.2 (192.168.1.2) 56(84) bytes of data. ............................. --- 192.168.1.2 ping statistics --- 948 packets transmitted, 467 received, 50.7384% packet loss, time 9671ms 2. Qualitative testing (with a phase-aligned schedule - the clocks are synchronized by ptp4l, not shown here): Receiver (sja1105): tc qdisc add dev swp2 clsact now=$(phc_ctl /dev/ptp1 get | awk '/clock time is/ {print $5}') && \ sec=$(echo $now | awk -F. '{print $1}') && \ base_time="$(((sec + 2) * 1000000000))" && \ echo "base time ${base_time}" tc filter add dev swp2 ingress flower skip_sw \ dst_mac 42:be:24:9b:76:20 \ action gate base-time ${base_time} \ sched-entry OPEN 60000 -1 -1 \ sched-entry CLOSE 40000 -1 -1 \ action trap Sender (enetc): now=$(phc_ctl /dev/ptp0 get | awk '/clock time is/ {print $5}') && \ sec=$(echo $now | awk -F. '{print $1}') && \ base_time="$(((sec + 2) * 1000000000))" && \ echo "base time ${base_time}" tc qdisc add dev eno0 parent root 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 ${base_time} \ sched-entry S 01 50000 \ sched-entry S 00 50000 \ flags 2 ping -A 192.168.1.1 PING 192.168.1.1 (192.168.1.1): 56 data bytes ... ^C --- 192.168.1.1 ping statistics --- 1425 packets transmitted, 1424 packets received, 0% packet loss round-trip min/avg/max = 0.322/0.361/0.990 ms And just for comparison, with the tc-taprio schedule deleted: ping -A 192.168.1.1 PING 192.168.1.1 (192.168.1.1): 56 data bytes ... ^C --- 192.168.1.1 ping statistics --- 33 packets transmitted, 19 packets received, 42% packet loss round-trip min/avg/max = 0.336/0.464/0.597 ms Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2020-05-06 03:20:56 +08:00
if (sja1105_gating_check_conflicts(priv, port, NULL)) {
dev_err(ds->dev, "Conflict with tc-gate schedule\n");
return -ERANGE;
}
net: dsa: sja1105: Configure the Time-Aware Scheduler via tc-taprio offload This qdisc offload is the closest thing to what the SJA1105 supports in hardware for time-based egress shaping. The switch core really is built around SAE AS6802/TTEthernet (a TTTech standard) but can be made to operate similarly to IEEE 802.1Qbv with some constraints: - The gate control list is a global list for all ports. There are 8 execution threads that iterate through this global list in parallel. I don't know why 8, there are only 4 front-panel ports. - Care must be taken by the user to make sure that two execution threads never get to execute a GCL entry simultaneously. I created a O(n^4) checker for this hardware limitation, prior to accepting a taprio offload configuration as valid. - The spec says that if a GCL entry's interval is shorter than the frame length, you shouldn't send it (and end up in head-of-line blocking). Well, this switch does anyway. - The switch has no concept of ADMIN and OPER configurations. Because it's so simple, the TAS settings are loaded through the static config tables interface, so there isn't even place for any discussion about 'graceful switchover between ADMIN and OPER'. You just reset the switch and upload a new OPER config. - The switch accepts multiple time sources for the gate events. Right now I am using the standalone clock source as opposed to PTP. So the base time parameter doesn't really do much. Support for the PTP clock source will be added in a future series. Signed-off-by: Vladimir Oltean <olteanv@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-09-15 10:00:02 +08:00
tas_data->offload[port] = taprio_offload_get(admin);
rc = sja1105_init_scheduling(priv);
if (rc < 0)
return rc;
return sja1105_static_config_reload(priv, SJA1105_SCHEDULING);
net: dsa: sja1105: Configure the Time-Aware Scheduler via tc-taprio offload This qdisc offload is the closest thing to what the SJA1105 supports in hardware for time-based egress shaping. The switch core really is built around SAE AS6802/TTEthernet (a TTTech standard) but can be made to operate similarly to IEEE 802.1Qbv with some constraints: - The gate control list is a global list for all ports. There are 8 execution threads that iterate through this global list in parallel. I don't know why 8, there are only 4 front-panel ports. - Care must be taken by the user to make sure that two execution threads never get to execute a GCL entry simultaneously. I created a O(n^4) checker for this hardware limitation, prior to accepting a taprio offload configuration as valid. - The spec says that if a GCL entry's interval is shorter than the frame length, you shouldn't send it (and end up in head-of-line blocking). Well, this switch does anyway. - The switch has no concept of ADMIN and OPER configurations. Because it's so simple, the TAS settings are loaded through the static config tables interface, so there isn't even place for any discussion about 'graceful switchover between ADMIN and OPER'. You just reset the switch and upload a new OPER config. - The switch accepts multiple time sources for the gate events. Right now I am using the standalone clock source as opposed to PTP. So the base time parameter doesn't really do much. Support for the PTP clock source will be added in a future series. Signed-off-by: Vladimir Oltean <olteanv@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-09-15 10:00:02 +08:00
}
net: dsa: sja1105: Implement state machine for TAS with PTP clock source Tested using the following bash script and the tc from iproute2-next: #!/bin/bash set -e -u -o pipefail NSEC_PER_SEC="1000000000" gatemask() { local tc_list="$1" local mask=0 for tc in ${tc_list}; do mask=$((${mask} | (1 << ${tc}))) done printf "%02x" ${mask} } if ! systemctl is-active --quiet ptp4l; then echo "Please start the ptp4l service" exit fi now=$(phc_ctl /dev/ptp1 get | gawk '/clock time is/ { print $5; }') # Phase-align the base time to the start of the next second. sec=$(echo "${now}" | gawk -F. '{ print $1; }') base_time="$(((${sec} + 1) * ${NSEC_PER_SEC}))" tc qdisc add dev swp5 parent root handle 100 taprio \ num_tc 8 \ map 0 1 2 3 5 6 7 \ queues 1@0 1@1 1@2 1@3 1@4 1@5 1@6 1@7 \ base-time ${base_time} \ sched-entry S $(gatemask 7) 100000 \ sched-entry S $(gatemask "0 1 2 3 4 5 6") 400000 \ clockid CLOCK_TAI flags 2 The "state machine" is a workqueue invoked after each manipulation command on the PTP clock (reset, adjust time, set time, adjust frequency) which checks over the state of the time-aware scheduler. So it is not monitored periodically, only in reaction to a PTP command typically triggered from a userspace daemon (linuxptp). Otherwise there is no reason for things to go wrong. Now that the timecounter/cyclecounter has been replaced with hardware operations on the PTP clock, the TAS Kconfig now depends upon PTP and the standalone clocksource operating mode has been removed. Signed-off-by: Vladimir Oltean <olteanv@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-11-12 08:11:54 +08:00
static int sja1105_tas_check_running(struct sja1105_private *priv)
{
struct sja1105_tas_data *tas_data = &priv->tas_data;
struct dsa_switch *ds = priv->ds;
struct sja1105_ptp_cmd cmd = {0};
int rc;
rc = sja1105_ptp_commit(ds, &cmd, SPI_READ);
if (rc < 0)
return rc;
if (cmd.ptpstrtsch == 1)
/* Schedule successfully started */
tas_data->state = SJA1105_TAS_STATE_RUNNING;
else if (cmd.ptpstopsch == 1)
/* Schedule is stopped */
tas_data->state = SJA1105_TAS_STATE_DISABLED;
else
/* Schedule is probably not configured with PTP clock source */
rc = -EINVAL;
return rc;
}
/* Write to PTPCLKCORP */
static int sja1105_tas_adjust_drift(struct sja1105_private *priv,
u64 correction)
{
const struct sja1105_regs *regs = priv->info->regs;
u32 ptpclkcorp = ns_to_sja1105_ticks(correction);
return sja1105_xfer_u32(priv, SPI_WRITE, regs->ptpclkcorp,
&ptpclkcorp, NULL);
}
/* Write to PTPSCHTM */
static int sja1105_tas_set_base_time(struct sja1105_private *priv,
u64 base_time)
{
const struct sja1105_regs *regs = priv->info->regs;
u64 ptpschtm = ns_to_sja1105_ticks(base_time);
return sja1105_xfer_u64(priv, SPI_WRITE, regs->ptpschtm,
&ptpschtm, NULL);
}
static int sja1105_tas_start(struct sja1105_private *priv)
{
struct sja1105_tas_data *tas_data = &priv->tas_data;
struct sja1105_ptp_cmd *cmd = &priv->ptp_data.cmd;
struct dsa_switch *ds = priv->ds;
int rc;
dev_dbg(ds->dev, "Starting the TAS\n");
if (tas_data->state == SJA1105_TAS_STATE_ENABLED_NOT_RUNNING ||
tas_data->state == SJA1105_TAS_STATE_RUNNING) {
dev_err(ds->dev, "TAS already started\n");
return -EINVAL;
}
cmd->ptpstrtsch = 1;
cmd->ptpstopsch = 0;
rc = sja1105_ptp_commit(ds, cmd, SPI_WRITE);
if (rc < 0)
return rc;
tas_data->state = SJA1105_TAS_STATE_ENABLED_NOT_RUNNING;
return 0;
}
static int sja1105_tas_stop(struct sja1105_private *priv)
{
struct sja1105_tas_data *tas_data = &priv->tas_data;
struct sja1105_ptp_cmd *cmd = &priv->ptp_data.cmd;
struct dsa_switch *ds = priv->ds;
int rc;
dev_dbg(ds->dev, "Stopping the TAS\n");
if (tas_data->state == SJA1105_TAS_STATE_DISABLED) {
dev_err(ds->dev, "TAS already disabled\n");
return -EINVAL;
}
cmd->ptpstopsch = 1;
cmd->ptpstrtsch = 0;
rc = sja1105_ptp_commit(ds, cmd, SPI_WRITE);
if (rc < 0)
return rc;
tas_data->state = SJA1105_TAS_STATE_DISABLED;
return 0;
}
/* The schedule engine and the PTP clock are driven by the same oscillator, and
* they run in parallel. But whilst the PTP clock can keep an absolute
* time-of-day, the schedule engine is only running in 'ticks' (25 ticks make
* up a delta, which is 200ns), and wrapping around at the end of each cycle.
* The schedule engine is started when the PTP clock reaches the PTPSCHTM time
* (in PTP domain).
* Because the PTP clock can be rate-corrected (accelerated or slowed down) by
* a software servo, and the schedule engine clock runs in parallel to the PTP
* clock, there is logic internal to the switch that periodically keeps the
* schedule engine from drifting away. The frequency with which this internal
* syntonization happens is the PTP clock correction period (PTPCLKCORP). It is
* a value also in the PTP clock domain, and is also rate-corrected.
* To be precise, during a correction period, there is logic to determine by
* how many scheduler clock ticks has the PTP clock drifted. At the end of each
* correction period/beginning of new one, the length of a delta is shrunk or
* expanded with an integer number of ticks, compared with the typical 25.
* So a delta lasts for 200ns (or 25 ticks) only on average.
* Sometimes it is longer, sometimes it is shorter. The internal syntonization
* logic can adjust for at most 5 ticks each 20 ticks.
*
* The first implication is that you should choose your schedule correction
* period to be an integer multiple of the schedule length. Preferably one.
* In case there are schedules of multiple ports active, then the correction
* period needs to be a multiple of them all. Given the restriction that the
* cycle times have to be multiples of one another anyway, this means the
* correction period can simply be the largest cycle time, hence the current
* choice. This way, the updates are always synchronous to the transmission
* cycle, and therefore predictable.
*
* The second implication is that at the beginning of a correction period, the
* first few deltas will be modulated in time, until the schedule engine is
* properly phase-aligned with the PTP clock. For this reason, you should place
* your best-effort traffic at the beginning of a cycle, and your
* time-triggered traffic afterwards.
*
* The third implication is that once the schedule engine is started, it can
* only adjust for so much drift within a correction period. In the servo you
* can only change the PTPCLKRATE, but not step the clock (PTPCLKADD). If you
* want to do the latter, you need to stop and restart the schedule engine,
* which is what the state machine handles.
*/
static void sja1105_tas_state_machine(struct work_struct *work)
{
struct sja1105_tas_data *tas_data = work_to_sja1105_tas(work);
struct sja1105_private *priv = tas_to_sja1105(tas_data);
struct sja1105_ptp_data *ptp_data = &priv->ptp_data;
struct timespec64 base_time_ts, now_ts;
struct dsa_switch *ds = priv->ds;
struct timespec64 diff;
s64 base_time, now;
int rc = 0;
mutex_lock(&ptp_data->lock);
switch (tas_data->state) {
case SJA1105_TAS_STATE_DISABLED:
/* Can't do anything at all if clock is still being stepped */
if (tas_data->last_op != SJA1105_PTP_ADJUSTFREQ)
break;
rc = sja1105_tas_adjust_drift(priv, tas_data->max_cycle_time);
if (rc < 0)
break;
rc = __sja1105_ptp_gettimex(ds, &now, NULL);
if (rc < 0)
break;
/* Plan to start the earliest schedule first. The others
* will be started in hardware, by way of their respective
* entry points delta.
* Try our best to avoid fringe cases (race condition between
* ptpschtm and ptpstrtsch) by pushing the oper_base_time at
* least one second in the future from now. This is not ideal,
* but this only needs to buy us time until the
* sja1105_tas_start command below gets executed.
*/
base_time = future_base_time(tas_data->earliest_base_time,
tas_data->max_cycle_time,
now + 1ull * NSEC_PER_SEC);
base_time -= sja1105_delta_to_ns(1);
rc = sja1105_tas_set_base_time(priv, base_time);
if (rc < 0)
break;
tas_data->oper_base_time = base_time;
rc = sja1105_tas_start(priv);
if (rc < 0)
break;
base_time_ts = ns_to_timespec64(base_time);
now_ts = ns_to_timespec64(now);
dev_dbg(ds->dev, "OPER base time %lld.%09ld (now %lld.%09ld)\n",
base_time_ts.tv_sec, base_time_ts.tv_nsec,
now_ts.tv_sec, now_ts.tv_nsec);
break;
case SJA1105_TAS_STATE_ENABLED_NOT_RUNNING:
if (tas_data->last_op != SJA1105_PTP_ADJUSTFREQ) {
/* Clock was stepped.. bad news for TAS */
sja1105_tas_stop(priv);
break;
}
/* Check if TAS has actually started, by comparing the
* scheduled start time with the SJA1105 PTP clock
*/
rc = __sja1105_ptp_gettimex(ds, &now, NULL);
if (rc < 0)
break;
if (now < tas_data->oper_base_time) {
/* TAS has not started yet */
diff = ns_to_timespec64(tas_data->oper_base_time - now);
dev_dbg(ds->dev, "time to start: [%lld.%09ld]",
diff.tv_sec, diff.tv_nsec);
break;
}
/* Time elapsed, what happened? */
rc = sja1105_tas_check_running(priv);
if (rc < 0)
break;
if (tas_data->state != SJA1105_TAS_STATE_RUNNING)
/* TAS has started */
dev_err(ds->dev,
"TAS not started despite time elapsed\n");
break;
case SJA1105_TAS_STATE_RUNNING:
/* Clock was stepped.. bad news for TAS */
if (tas_data->last_op != SJA1105_PTP_ADJUSTFREQ) {
sja1105_tas_stop(priv);
break;
}
rc = sja1105_tas_check_running(priv);
if (rc < 0)
break;
if (tas_data->state != SJA1105_TAS_STATE_RUNNING)
dev_err(ds->dev, "TAS surprisingly stopped\n");
break;
default:
if (net_ratelimit())
dev_err(ds->dev, "TAS in an invalid state (incorrect use of API)!\n");
}
if (rc && net_ratelimit())
dev_err(ds->dev, "An operation returned %d\n", rc);
mutex_unlock(&ptp_data->lock);
}
void sja1105_tas_clockstep(struct dsa_switch *ds)
{
struct sja1105_private *priv = ds->priv;
struct sja1105_tas_data *tas_data = &priv->tas_data;
if (!tas_data->enabled)
return;
tas_data->last_op = SJA1105_PTP_CLOCKSTEP;
schedule_work(&tas_data->tas_work);
}
void sja1105_tas_adjfreq(struct dsa_switch *ds)
{
struct sja1105_private *priv = ds->priv;
struct sja1105_tas_data *tas_data = &priv->tas_data;
if (!tas_data->enabled)
return;
/* No reason to schedule the workqueue, nothing changed */
if (tas_data->state == SJA1105_TAS_STATE_RUNNING)
return;
tas_data->last_op = SJA1105_PTP_ADJUSTFREQ;
schedule_work(&tas_data->tas_work);
}
net: dsa: sja1105: Configure the Time-Aware Scheduler via tc-taprio offload This qdisc offload is the closest thing to what the SJA1105 supports in hardware for time-based egress shaping. The switch core really is built around SAE AS6802/TTEthernet (a TTTech standard) but can be made to operate similarly to IEEE 802.1Qbv with some constraints: - The gate control list is a global list for all ports. There are 8 execution threads that iterate through this global list in parallel. I don't know why 8, there are only 4 front-panel ports. - Care must be taken by the user to make sure that two execution threads never get to execute a GCL entry simultaneously. I created a O(n^4) checker for this hardware limitation, prior to accepting a taprio offload configuration as valid. - The spec says that if a GCL entry's interval is shorter than the frame length, you shouldn't send it (and end up in head-of-line blocking). Well, this switch does anyway. - The switch has no concept of ADMIN and OPER configurations. Because it's so simple, the TAS settings are loaded through the static config tables interface, so there isn't even place for any discussion about 'graceful switchover between ADMIN and OPER'. You just reset the switch and upload a new OPER config. - The switch accepts multiple time sources for the gate events. Right now I am using the standalone clock source as opposed to PTP. So the base time parameter doesn't really do much. Support for the PTP clock source will be added in a future series. Signed-off-by: Vladimir Oltean <olteanv@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-09-15 10:00:02 +08:00
void sja1105_tas_setup(struct dsa_switch *ds)
{
net: dsa: sja1105: Implement state machine for TAS with PTP clock source Tested using the following bash script and the tc from iproute2-next: #!/bin/bash set -e -u -o pipefail NSEC_PER_SEC="1000000000" gatemask() { local tc_list="$1" local mask=0 for tc in ${tc_list}; do mask=$((${mask} | (1 << ${tc}))) done printf "%02x" ${mask} } if ! systemctl is-active --quiet ptp4l; then echo "Please start the ptp4l service" exit fi now=$(phc_ctl /dev/ptp1 get | gawk '/clock time is/ { print $5; }') # Phase-align the base time to the start of the next second. sec=$(echo "${now}" | gawk -F. '{ print $1; }') base_time="$(((${sec} + 1) * ${NSEC_PER_SEC}))" tc qdisc add dev swp5 parent root handle 100 taprio \ num_tc 8 \ map 0 1 2 3 5 6 7 \ queues 1@0 1@1 1@2 1@3 1@4 1@5 1@6 1@7 \ base-time ${base_time} \ sched-entry S $(gatemask 7) 100000 \ sched-entry S $(gatemask "0 1 2 3 4 5 6") 400000 \ clockid CLOCK_TAI flags 2 The "state machine" is a workqueue invoked after each manipulation command on the PTP clock (reset, adjust time, set time, adjust frequency) which checks over the state of the time-aware scheduler. So it is not monitored periodically, only in reaction to a PTP command typically triggered from a userspace daemon (linuxptp). Otherwise there is no reason for things to go wrong. Now that the timecounter/cyclecounter has been replaced with hardware operations on the PTP clock, the TAS Kconfig now depends upon PTP and the standalone clocksource operating mode has been removed. Signed-off-by: Vladimir Oltean <olteanv@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-11-12 08:11:54 +08:00
struct sja1105_private *priv = ds->priv;
struct sja1105_tas_data *tas_data = &priv->tas_data;
INIT_WORK(&tas_data->tas_work, sja1105_tas_state_machine);
tas_data->state = SJA1105_TAS_STATE_DISABLED;
tas_data->last_op = SJA1105_PTP_NONE;
net: dsa: sja1105: implement tc-gate using time-triggered virtual links Restrict the TTEthernet hardware support on this switch to operate as closely as possible to IEEE 802.1Qci as possible. This means that it can perform PTP-time-based ingress admission control on streams identified by {DMAC, VID, PCP}, which is useful when trying to ensure the determinism of traffic scheduled via IEEE 802.1Qbv. The oddity comes from the fact that in hardware (and in TTEthernet at large), virtual links always need a full-blown action, including not only the type of policing, but also the list of destination ports. So in practice, a single tc-gate action will result in all packets getting dropped. Additional actions (either "trap" or "redirect") need to be specified in the same filter rule such that the conforming packets are actually forwarded somewhere. Apart from the VL Lookup, Policing and Forwarding tables which need to be programmed for each flow (virtual link), the Schedule engine also needs to be told to open/close the admission gates for each individual virtual link. A fairly accurate (and detailed) description of how that works is already present in sja1105_tas.c, since it is already used to trigger the egress gates for the tc-taprio offload (IEEE 802.1Qbv). Key point here, we remember that the schedule engine supports 8 "subschedules" (execution threads that iterate through the global schedule in parallel, and that no 2 hardware threads must execute a schedule entry at the same time). For tc-taprio, each egress port used one of these 8 subschedules, leaving a total of 4 subschedules unused. In principle we could have allocated 1 subschedule for the tc-gate offload of each ingress port, but actually the schedules of all virtual links installed on each ingress port would have needed to be merged together, before they could have been programmed to hardware. So simplify our life and just merge the entire tc-gate configuration, for all virtual links on all ingress ports, into a single subschedule. Be sure to check that against the usual hardware scheduling conflicts, and program it to hardware alongside any tc-taprio subschedule that may be present. The following scenarios were tested: 1. Quantitative testing: tc qdisc add dev swp2 clsact tc filter add dev swp2 ingress flower skip_sw \ dst_mac 42:be:24:9b:76:20 \ action gate index 1 base-time 0 \ sched-entry OPEN 1200 -1 -1 \ sched-entry CLOSE 1200 -1 -1 \ action trap ping 192.168.1.2 -f PING 192.168.1.2 (192.168.1.2) 56(84) bytes of data. ............................. --- 192.168.1.2 ping statistics --- 948 packets transmitted, 467 received, 50.7384% packet loss, time 9671ms 2. Qualitative testing (with a phase-aligned schedule - the clocks are synchronized by ptp4l, not shown here): Receiver (sja1105): tc qdisc add dev swp2 clsact now=$(phc_ctl /dev/ptp1 get | awk '/clock time is/ {print $5}') && \ sec=$(echo $now | awk -F. '{print $1}') && \ base_time="$(((sec + 2) * 1000000000))" && \ echo "base time ${base_time}" tc filter add dev swp2 ingress flower skip_sw \ dst_mac 42:be:24:9b:76:20 \ action gate base-time ${base_time} \ sched-entry OPEN 60000 -1 -1 \ sched-entry CLOSE 40000 -1 -1 \ action trap Sender (enetc): now=$(phc_ctl /dev/ptp0 get | awk '/clock time is/ {print $5}') && \ sec=$(echo $now | awk -F. '{print $1}') && \ base_time="$(((sec + 2) * 1000000000))" && \ echo "base time ${base_time}" tc qdisc add dev eno0 parent root 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 ${base_time} \ sched-entry S 01 50000 \ sched-entry S 00 50000 \ flags 2 ping -A 192.168.1.1 PING 192.168.1.1 (192.168.1.1): 56 data bytes ... ^C --- 192.168.1.1 ping statistics --- 1425 packets transmitted, 1424 packets received, 0% packet loss round-trip min/avg/max = 0.322/0.361/0.990 ms And just for comparison, with the tc-taprio schedule deleted: ping -A 192.168.1.1 PING 192.168.1.1 (192.168.1.1): 56 data bytes ... ^C --- 192.168.1.1 ping statistics --- 33 packets transmitted, 19 packets received, 42% packet loss round-trip min/avg/max = 0.336/0.464/0.597 ms Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2020-05-06 03:20:56 +08:00
INIT_LIST_HEAD(&tas_data->gating_cfg.entries);
net: dsa: sja1105: Configure the Time-Aware Scheduler via tc-taprio offload This qdisc offload is the closest thing to what the SJA1105 supports in hardware for time-based egress shaping. The switch core really is built around SAE AS6802/TTEthernet (a TTTech standard) but can be made to operate similarly to IEEE 802.1Qbv with some constraints: - The gate control list is a global list for all ports. There are 8 execution threads that iterate through this global list in parallel. I don't know why 8, there are only 4 front-panel ports. - Care must be taken by the user to make sure that two execution threads never get to execute a GCL entry simultaneously. I created a O(n^4) checker for this hardware limitation, prior to accepting a taprio offload configuration as valid. - The spec says that if a GCL entry's interval is shorter than the frame length, you shouldn't send it (and end up in head-of-line blocking). Well, this switch does anyway. - The switch has no concept of ADMIN and OPER configurations. Because it's so simple, the TAS settings are loaded through the static config tables interface, so there isn't even place for any discussion about 'graceful switchover between ADMIN and OPER'. You just reset the switch and upload a new OPER config. - The switch accepts multiple time sources for the gate events. Right now I am using the standalone clock source as opposed to PTP. So the base time parameter doesn't really do much. Support for the PTP clock source will be added in a future series. Signed-off-by: Vladimir Oltean <olteanv@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-09-15 10:00:02 +08:00
}
void sja1105_tas_teardown(struct dsa_switch *ds)
{
struct sja1105_private *priv = ds->priv;
struct tc_taprio_qopt_offload *offload;
int port;
net: dsa: sja1105: Implement state machine for TAS with PTP clock source Tested using the following bash script and the tc from iproute2-next: #!/bin/bash set -e -u -o pipefail NSEC_PER_SEC="1000000000" gatemask() { local tc_list="$1" local mask=0 for tc in ${tc_list}; do mask=$((${mask} | (1 << ${tc}))) done printf "%02x" ${mask} } if ! systemctl is-active --quiet ptp4l; then echo "Please start the ptp4l service" exit fi now=$(phc_ctl /dev/ptp1 get | gawk '/clock time is/ { print $5; }') # Phase-align the base time to the start of the next second. sec=$(echo "${now}" | gawk -F. '{ print $1; }') base_time="$(((${sec} + 1) * ${NSEC_PER_SEC}))" tc qdisc add dev swp5 parent root handle 100 taprio \ num_tc 8 \ map 0 1 2 3 5 6 7 \ queues 1@0 1@1 1@2 1@3 1@4 1@5 1@6 1@7 \ base-time ${base_time} \ sched-entry S $(gatemask 7) 100000 \ sched-entry S $(gatemask "0 1 2 3 4 5 6") 400000 \ clockid CLOCK_TAI flags 2 The "state machine" is a workqueue invoked after each manipulation command on the PTP clock (reset, adjust time, set time, adjust frequency) which checks over the state of the time-aware scheduler. So it is not monitored periodically, only in reaction to a PTP command typically triggered from a userspace daemon (linuxptp). Otherwise there is no reason for things to go wrong. Now that the timecounter/cyclecounter has been replaced with hardware operations on the PTP clock, the TAS Kconfig now depends upon PTP and the standalone clocksource operating mode has been removed. Signed-off-by: Vladimir Oltean <olteanv@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-11-12 08:11:54 +08:00
cancel_work_sync(&priv->tas_data.tas_work);
net: dsa: sja1105: Configure the Time-Aware Scheduler via tc-taprio offload This qdisc offload is the closest thing to what the SJA1105 supports in hardware for time-based egress shaping. The switch core really is built around SAE AS6802/TTEthernet (a TTTech standard) but can be made to operate similarly to IEEE 802.1Qbv with some constraints: - The gate control list is a global list for all ports. There are 8 execution threads that iterate through this global list in parallel. I don't know why 8, there are only 4 front-panel ports. - Care must be taken by the user to make sure that two execution threads never get to execute a GCL entry simultaneously. I created a O(n^4) checker for this hardware limitation, prior to accepting a taprio offload configuration as valid. - The spec says that if a GCL entry's interval is shorter than the frame length, you shouldn't send it (and end up in head-of-line blocking). Well, this switch does anyway. - The switch has no concept of ADMIN and OPER configurations. Because it's so simple, the TAS settings are loaded through the static config tables interface, so there isn't even place for any discussion about 'graceful switchover between ADMIN and OPER'. You just reset the switch and upload a new OPER config. - The switch accepts multiple time sources for the gate events. Right now I am using the standalone clock source as opposed to PTP. So the base time parameter doesn't really do much. Support for the PTP clock source will be added in a future series. Signed-off-by: Vladimir Oltean <olteanv@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-09-15 10:00:02 +08:00
for (port = 0; port < SJA1105_NUM_PORTS; port++) {
offload = priv->tas_data.offload[port];
if (!offload)
continue;
taprio_offload_free(offload);
}
}