On 08/21, Jim Harris wrote:
>Ideally, get_tsc_freq_arch() is able to provide the
>TSC rate using architecture-specific means. When that
>is not possible, DPDK reverts to calculating the
>TSC rate with a 100ms nanosleep or 1s sleep. The latter
>occurs more frequently in VMs which often do not have
>access to the data they need from arch-specific means
>(CPUID leaf 0x15 or MSR 0xCE on x86).
>
>In secondary processes, the extra 100ms is especially
>noticeable and consumes the bulk of rte_eal_init()
>execution time. So in secondary processes, if
>we cannot get the TSC rate using get_tsc_freq_arch(),
>try to get the TSC rate from the primary process
>instead using rte_mp_msg. This is much faster than
>100ms.
>
>Reduces rte_eal_init() execution time in a secondary
>process from 165ms to 66ms on my test system.
>
>Signed-off-by: Jim Harris <james.r.har...@intel.com>
>Change-Id: I584419ed1c7d6f47841e0a0eb23f34c9f1186d35
This Change-Id line is unnecessary.
Thanks,
Xiaolong
>---
> lib/librte_eal/common/eal_common_timer.c | 62 ++++++++++++++++++++++++++++++
> 1 file changed, 62 insertions(+)
>
>diff --git a/lib/librte_eal/common/eal_common_timer.c
>b/lib/librte_eal/common/eal_common_timer.c
>index 145543de7..ad965455d 100644
>--- a/lib/librte_eal/common/eal_common_timer.c
>+++ b/lib/librte_eal/common/eal_common_timer.c
>@@ -15,9 +15,17 @@
> #include <rte_log.h>
> #include <rte_cycles.h>
> #include <rte_pause.h>
>+#include <rte_eal.h>
>+#include <rte_errno.h>
>
> #include "eal_private.h"
>
>+#define EAL_TIMER_MP "eal_timer_mp_sync"
>+
>+struct timer_mp_param {
>+ uint64_t tsc_hz;
>+};
>+
> /* The frequency of the RDTSC timer resolution */
> static uint64_t eal_tsc_resolution_hz;
>
>@@ -74,12 +82,58 @@ estimate_tsc_freq(void)
> return RTE_ALIGN_MUL_NEAR(rte_rdtsc() - start, CYC_PER_10MHZ);
> }
>
>+static uint64_t
>+get_tsc_freq_from_primary(void)
>+{
>+ struct rte_mp_msg mp_req = {0};
>+ struct rte_mp_reply mp_reply = {0};
>+ struct timer_mp_param *r;
>+ struct timespec ts = {.tv_sec = 1, .tv_nsec = 0};
>+ uint64_t tsc_hz;
>+
>+ strcpy(mp_req.name, EAL_TIMER_MP);
>+ if (rte_mp_request_sync(&mp_req, &mp_reply, &ts) ||
>+ mp_reply.nb_received != 1) {
>+ tsc_hz = 0;
>+ } else {
>+ r = (struct timer_mp_param *)mp_reply.msgs[0].param;
>+ tsc_hz = r->tsc_hz;
>+ }
>+
>+ free(mp_reply.msgs);
>+ return tsc_hz;
>+}
>+
>+static int
>+timer_mp_primary(__attribute__((unused)) const struct rte_mp_msg *msg,
>+ const void *peer)
>+{
>+ struct rte_mp_msg reply = {0};
>+ struct timer_mp_param *r = (struct timer_mp_param *)reply.param;
>+
>+ r->tsc_hz = eal_tsc_resolution_hz;
>+ strcpy(reply.name, EAL_TIMER_MP);
>+ reply.len_param = sizeof(*r);
>+
>+ return rte_mp_reply(&reply, peer);
>+}
>+
> void
> set_tsc_freq(void)
> {
> uint64_t freq;
>+ int rc;
>
> freq = get_tsc_freq_arch();
>+ if (!freq && rte_eal_process_type() != RTE_PROC_PRIMARY) {
>+ /* We couldn't get the TSC frequency through arch-specific
>+ * means. If this is a secondary process, try to get the
>+ * TSC frequency from the primary process - this will
>+ * be much faster than get_tsc_freq() or estimate_tsc_freq()
>+ * below.
>+ */
>+ freq = get_tsc_freq_from_primary();
>+ }
> if (!freq)
> freq = get_tsc_freq();
> if (!freq)
>@@ -87,6 +141,14 @@ set_tsc_freq(void)
>
> RTE_LOG(DEBUG, EAL, "TSC frequency is ~%" PRIu64 " KHz\n", freq / 1000);
> eal_tsc_resolution_hz = freq;
>+ if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
>+ rc = rte_mp_action_register(EAL_TIMER_MP, timer_mp_primary);
>+ if (rc && rte_errno != ENOTSUP) {
>+ RTE_LOG(WARNING, EAL, "Could not register mp_action - "
>+ "secondary processes will calculate TSC rate "
>+ "independently.\n");
>+ }
>+ }
> }
>
> void rte_delay_us_callback_register(void (*userfunc)(unsigned int))
>
