[PATCH 1/2] sched: fix byte clock computation for extreme port rates
Alexander.Chesnokov at kaspersky.com
Alexander.Chesnokov at kaspersky.com
Fri Nov 20 07:00:43 CET 2026
From: Alexander Chesnokov <Alexander.Chesnokov at kaspersky.com>
rte_sched_port_config() computes the port byte clock as
cycles_per_byte = (rte_get_tsc_hz() << RTE_SCHED_TIME_SHIFT)
/ params->rate;
and passes the result to rte_reciprocal_value(). The parameter
check rejects rate == 0 only. For rate above
rte_get_tsc_hz() << RTE_SCHED_TIME_SHIFT the division yields 0
and rte_reciprocal_value() divides by zero in turn: with d == 0,
rte_fls_u32(d - 1) returns 32 and the following m /= d faults,
crashing the application with SIGFPE. For rate below
rte_get_tsc_hz() >> (32 - RTE_SCHED_TIME_SHIFT) the quotient
exceeds UINT32_MAX and is silently truncated when stored in the
uint32_t cycles_per_byte, corrupting the byte clock.
The byte clock cannot represent rates outside that range, so
configurations within it behave exactly as before.
Move the computation to a helper shared by the parameter check
and the port configuration, and reject configurations whose byte
clock value is zero or does not fit in 32 bits, so that
rte_sched_port_config() and rte_sched_port_get_memory_footprint()
fail cleanly instead of crashing or mis-scheduling.
Fixes: 03d00293ca76 ("sched: eliminate floating point in calculating byte clock")
Cc: stable at dpdk.org
Signed-off-by: Alexander Chesnokov <Alexander.Chesnokov at kaspersky.com>
---
Cc: stephen at networkplumber.org
lib/sched/rte_sched.c | 19 +++++++++++++++++--
1 file changed, 17 insertions(+), 2 deletions(-)
diff --git a/lib/sched/rte_sched.c b/lib/sched/rte_sched.c
index daa31e8b99..ef88df124e 100644
--- a/lib/sched/rte_sched.c
+++ b/lib/sched/rte_sched.c
@@ -436,10 +436,17 @@ subport_profile_check(struct rte_sched_subport_profile_params *params,
return 0;
}
+static uint64_t
+rte_sched_port_cycles_per_byte(uint64_t rate)
+{
+ return (rte_get_tsc_hz() << RTE_SCHED_TIME_SHIFT) / rate;
+}
+
static int
rte_sched_port_check_params(struct rte_sched_port_params *params)
{
uint32_t i;
+ uint64_t cycles_per_byte;
if (params == NULL) {
SCHED_LOG(ERR,
@@ -461,6 +468,15 @@ rte_sched_port_check_params(struct rte_sched_port_params *params)
return -EINVAL;
}
+ cycles_per_byte = rte_sched_port_cycles_per_byte(params->rate);
+ /* rate: the byte clock value must be non-zero and fit in 32 bits */
+ if (cycles_per_byte == 0 ||
+ cycles_per_byte > UINT32_MAX) {
+ SCHED_LOG(ERR,
+ "%s: Incorrect value for rate", __func__);
+ return -EINVAL;
+ }
+
/* mtu */
if (params->mtu == 0) {
SCHED_LOG(ERR,
@@ -1006,8 +1022,7 @@ rte_sched_port_config(struct rte_sched_port_params *params)
/* Subport profile table */
rte_sched_port_config_subport_profile_table(port, params, port->rate);
- cycles_per_byte = (rte_get_tsc_hz() << RTE_SCHED_TIME_SHIFT)
- / params->rate;
+ cycles_per_byte = rte_sched_port_cycles_per_byte(params->rate);
port->inv_cycles_per_byte = rte_reciprocal_value(cycles_per_byte);
port->cycles_per_byte = cycles_per_byte;
--
2.43.0
More information about the dev
mailing list