The blamed commit broke tc-taprio schedules such as this one:
tc qdisc replace dev $swp1 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 0 \
sched-entry S 0x7f 990000 \
sched-entry S 0x80 10000 \
flags 0x2
because the gate entry for TC 7 (S 0x80 10000 ns) now has a static guard
band added earlier than its 'gate close' event, such that packet
overruns won't occur in the worst case of the largest packet possible.
Since guard bands are statically determined based on the per-tc
QSYS_QMAXSDU_CFG_* with a fallback on the port-based QSYS_PORT_MAX_SDU,
we need to discuss what happens with TC 7 depending on kernel version,
since the driver, prior to commit
55a515b1f5a9 ("net: dsa: felix: drop
oversized frames with tc-taprio instead of hanging the port"), did not
touch QSYS_QMAXSDU_CFG_*, and therefore relied on QSYS_PORT_MAX_SDU.
1 (before vsc9959_tas_guard_bands_update): QSYS_PORT_MAX_SDU defaults to
1518, and at gigabit this introduces a static guard band (independent
of packet sizes) of 12144 ns, plus QSYS::HSCH_MISC_CFG.FRM_ADJ (bit
time of 20 octets => 160 ns). But this is larger than the time window
itself, of 10000 ns. So, the queue system never considers a frame with
TC 7 as eligible for transmission, since the gate practically never
opens, and these frames are forever stuck in the TX queues and hang
the port.
2 (after vsc9959_tas_guard_bands_update): Under the sole goal of
enabling oversized frame dropping, we make an effort to set
QSYS_QMAXSDU_CFG_7 to 1230 bytes. But QSYS_QMAXSDU_CFG_7 plays
one more role, which we did not take into account: per-tc static guard
band, expressed in L2 byte time (auto-adjusted for FCS and L1 overhead).
There is a discrepancy between what the driver thinks (that there is
no guard band, and 100% of min_gate_len[tc] is available for egress
scheduling) and what the hardware actually does (crops the equivalent
of QSYS_QMAXSDU_CFG_7 ns out of min_gate_len[tc]). In practice, this
means that the hardware thinks it has exactly 0 ns for scheduling tc 7.
In both cases, even minimum sized Ethernet frames are stuck on egress
rather than being considered for scheduling on TC 7, even if they would
fit given a proper configuration. Considering the current situation,
with vsc9959_tas_guard_bands_update(), frames between 60 octets and 1230
octets in size are not eligible for oversized dropping (because they are
smaller than QSYS_QMAXSDU_CFG_7), but won't be considered as eligible
for scheduling either, because the min_gate_len[7] (10000 ns) minus the
guard band determined by QSYS_QMAXSDU_CFG_7 (1230 octets * 8 ns per
octet == 9840 ns) minus the guard band auto-added for L1 overhead by
QSYS::HSCH_MISC_CFG.FRM_ADJ (20 octets * 8 ns per octet == 160 octets)
leaves 0 ns for scheduling in the queue system proper.
Investigating the hardware behavior, it becomes apparent that the queue
system needs precisely 33 ns of 'gate open' time in order to consider a
frame as eligible for scheduling to a tc. So the solution to this
problem is to amend vsc9959_tas_guard_bands_update(), by giving the
per-tc guard bands less space by exactly 33 ns, just enough for one
frame to be scheduled in that interval. This allows the queue system to
make forward progress for that port-tc, and prevents it from hanging.
Fixes: 297c4de6f780 ("net: dsa: felix: re-enable TAS guard band mode")
Reported-by: Xiaoliang Yang <xiaoliang.yang_1@nxp.com>
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
#define VSC9959_NUM_PORTS 6
#define VSC9959_TAS_GCL_ENTRY_MAX 63
+#define VSC9959_TAS_MIN_GATE_LEN_NS 33
#define VSC9959_VCAP_POLICER_BASE 63
#define VSC9959_VCAP_POLICER_MAX 383
#define VSC9959_SWITCH_PCI_BAR 4
mdiobus_free(felix->imdio);
}
+/* The switch considers any frame (regardless of size) as eligible for
+ * transmission if the traffic class gate is open for at least 33 ns.
+ * Overruns are prevented by cropping an interval at the end of the gate time
+ * slot for which egress scheduling is blocked, but we need to still keep 33 ns
+ * available for one packet to be transmitted, otherwise the port tc will hang.
+ * This function returns the size of a gate interval that remains available for
+ * setting the guard band, after reserving the space for one egress frame.
+ */
+static u64 vsc9959_tas_remaining_gate_len_ps(u64 gate_len_ns)
+{
+ /* Gate always open */
+ if (gate_len_ns == U64_MAX)
+ return U64_MAX;
+
+ return (gate_len_ns - VSC9959_TAS_MIN_GATE_LEN_NS) * PSEC_PER_NSEC;
+}
+
/* Extract shortest continuous gate open intervals in ns for each traffic class
* of a cyclic tc-taprio schedule. If a gate is always open, the duration is
* considered U64_MAX. If the gate is always closed, it is considered 0.
vsc9959_tas_min_gate_lengths(ocelot_port->taprio, min_gate_len);
for (tc = 0; tc < OCELOT_NUM_TC; tc++) {
+ u64 remaining_gate_len_ps;
u32 max_sdu;
- if (min_gate_len[tc] == U64_MAX /* Gate always open */ ||
- min_gate_len[tc] * PSEC_PER_NSEC > needed_bit_time_ps) {
+ remaining_gate_len_ps =
+ vsc9959_tas_remaining_gate_len_ps(min_gate_len[tc]);
+
+ if (remaining_gate_len_ps > needed_bit_time_ps) {
/* Setting QMAXSDU_CFG to 0 disables oversized frame
* dropping.
*/
/* If traffic class doesn't support a full MTU sized
* frame, make sure to enable oversize frame dropping
* for frames larger than the smallest that would fit.
+ *
+ * However, the exact same register, QSYS_QMAXSDU_CFG_*,
+ * controls not only oversized frame dropping, but also
+ * per-tc static guard band lengths, so it reduces the
+ * useful gate interval length. Therefore, be careful
+ * to calculate a guard band (and therefore max_sdu)
+ * that still leaves 33 ns available in the time slot.
*/
- max_sdu = div_u64(min_gate_len[tc] * PSEC_PER_NSEC,
- picos_per_byte);
+ max_sdu = div_u64(remaining_gate_len_ps, picos_per_byte);
/* A TC gate may be completely closed, which is a
* special case where all packets are oversized.
* Any limit smaller than 64 octets accomplishes this