Wait for event with timeout (WFET) puts the CPU in a low power
mode and stays there until an event is signalled (SEV), loss of
an exclusive monitor or a timeout.
WFET is enabled selectively by checking FEAT_WFxT in Linux
auxiliary vector. If FEAT_WFxT is not available power management
will fallback to WFE.
WFE is available on all the Arm platforms supported by DPDK.
Therefore, the RTE_ARM_USE_WFE macro is not required to enable
the WFE feature for PMD power monitoring.
RTE_ARM_USE_WFE is used at the build time to use the WFE instruction
where applicable in the code at the developer's discretion rather
than as an indicator of the instruction's availability.
Signed-off-by: Wathsala Vithanage <wathsala.vithanage@arm.com>
Reviewed-by: Dhruv Tripathi <dhruv.tripathi@arm.com>
Reviewed-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
Reviewed-by: Jack Bond-Preston <jack.bond-preston@foss.arm.com>
Reviewed-by: Nick Connolly <nick.connolly@arm.com>
Reviewed-by: Vinod Krishna <vinod.krishna@arm.com>
---
.mailmap | 1 +
app/test/test_cpuflags.c | 3 +++
lib/eal/arm/include/rte_cpuflags_64.h | 3 +++
lib/eal/arm/include/rte_pause_64.h | 16 +++++++++--
lib/eal/arm/rte_cpuflags.c | 1 +
lib/eal/arm/rte_power_intrinsics.c | 39 ++++++++++++++++++---------
6 files changed, 49 insertions(+), 14 deletions(-)
@@ -1540,6 +1540,7 @@ Vincent Li <vincent.mc.li@gmail.com>
Vincent S. Cojot <vcojot@redhat.com>
Vinh Tran <vinh.t.tran10@gmail.com>
Vipin Padmam Ramesh <vipinp@vmware.com>
+Vinod Krishna <vinod.krishna@arm.com>
Vipin Varghese <vipin.varghese@amd.com> <vipin.varghese@intel.com>
Vipul Ashri <vipul.ashri@oracle.com>
Visa Hankala <visa@hankala.org>
@@ -156,6 +156,9 @@ test_cpuflags(void)
printf("Check for SVEBF16:\t");
CHECK_FOR_FLAG(RTE_CPUFLAG_SVEBF16);
+
+ printf("Check for WFXT:\t");
+ CHECK_FOR_FLAG(RTE_CPUFLAG_WFXT);
#endif
#if defined(RTE_ARCH_X86_64) || defined(RTE_ARCH_I686)
@@ -36,6 +36,9 @@ enum rte_cpu_flag_t {
RTE_CPUFLAG_SVEF64MM,
RTE_CPUFLAG_SVEBF16,
RTE_CPUFLAG_AARCH64,
+
+ /* WFET and WFIT instructions */
+ RTE_CPUFLAG_WFXT,
};
#include "generic/rte_cpuflags.h"
@@ -24,15 +24,27 @@ static inline void rte_pause(void)
asm volatile("yield" ::: "memory");
}
-/* Send a local event to quit WFE. */
+/* Send a local event to quit WFE/WFxT. */
#define __RTE_ARM_SEVL() { asm volatile("sevl" : : : "memory"); }
-/* Send a global event to quit WFE for all cores. */
+/* Send a global event to quit WFE/WFxT for all cores. */
#define __RTE_ARM_SEV() { asm volatile("sev" : : : "memory"); }
/* Put processor into low power WFE(Wait For Event) state. */
#define __RTE_ARM_WFE() { asm volatile("wfe" : : : "memory"); }
+/* Put processor into low power WFET (WFE with Timeout) state. */
+#ifdef RTE_ARM_FEATURE_WFXT
+#define __RTE_ARM_WFET(t) { \
+ asm volatile("wfet %x[to]" \
+ : \
+ : [to] "r" (t) \
+ : "memory"); \
+ }
+#else
+#define __RTE_ARM_WFET(t) { RTE_SET_USED(t); }
+#endif
+
/*
* Atomic exclusive load from addr, it returns the 8-bit content of
* *addr while making it 'monitored', when it is written by someone
@@ -115,6 +115,7 @@ const struct feature_entry rte_cpu_feature_table[] = {
FEAT_DEF(SVEF32MM, REG_HWCAP2, 10)
FEAT_DEF(SVEF64MM, REG_HWCAP2, 11)
FEAT_DEF(SVEBF16, REG_HWCAP2, 12)
+ FEAT_DEF(WFXT, REG_HWCAP2, 31)
FEAT_DEF(AARCH64, REG_PLATFORM, 0)
};
#endif /* RTE_ARCH */
@@ -4,19 +4,32 @@
#include <errno.h>
+#include "rte_cpuflags.h"
#include "rte_power_intrinsics.h"
/**
- * This function uses WFE instruction to make lcore suspend
+ * Set wfet_en if WFET is supported
+ */
+#ifdef RTE_ARCH_64
+static uint8_t wfet_en;
+#endif /* RTE_ARCH_64 */
+
+RTE_INIT(rte_power_intrinsics_init)
+{
+#ifdef RTE_ARCH_64
+ if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_WFXT))
+ wfet_en = 1;
+#endif /* RTE_ARCH_64 */
+}
+
+/**
+ * This function uses WFE/WFET instruction to make lcore suspend
* execution on ARM.
- * Note that timestamp based timeout is not supported yet.
*/
int
rte_power_monitor(const struct rte_power_monitor_cond *pmc,
const uint64_t tsc_timestamp)
{
- RTE_SET_USED(tsc_timestamp);
-
#ifdef RTE_ARCH_64
const unsigned int lcore_id = rte_lcore_id();
uint64_t cur_value;
@@ -33,28 +46,30 @@ rte_power_monitor(const struct rte_power_monitor_cond *pmc,
switch (pmc->size) {
case sizeof(uint8_t):
- __RTE_ARM_LOAD_EXC_8(pmc->addr, cur_value, rte_memory_order_relaxed)
- __RTE_ARM_WFE()
+ __RTE_ARM_LOAD_EXC_8(pmc->addr, cur_value, rte_memory_order_relaxed);
break;
case sizeof(uint16_t):
- __RTE_ARM_LOAD_EXC_16(pmc->addr, cur_value, rte_memory_order_relaxed)
- __RTE_ARM_WFE()
+ __RTE_ARM_LOAD_EXC_16(pmc->addr, cur_value, rte_memory_order_relaxed);
break;
case sizeof(uint32_t):
- __RTE_ARM_LOAD_EXC_32(pmc->addr, cur_value, rte_memory_order_relaxed)
- __RTE_ARM_WFE()
+ __RTE_ARM_LOAD_EXC_32(pmc->addr, cur_value, rte_memory_order_relaxed);
break;
case sizeof(uint64_t):
- __RTE_ARM_LOAD_EXC_64(pmc->addr, cur_value, rte_memory_order_relaxed)
- __RTE_ARM_WFE()
+ __RTE_ARM_LOAD_EXC_64(pmc->addr, cur_value, rte_memory_order_relaxed);
break;
default:
return -EINVAL; /* unexpected size */
}
+ if (wfet_en)
+ __RTE_ARM_WFET(tsc_timestamp)
+ else
+ __RTE_ARM_WFE()
+
return 0;
#else
RTE_SET_USED(pmc);
+ RTE_SET_USED(tsc_timestamp);
return -ENOTSUP;
#endif /* RTE_ARCH_64 */