[1/2] timer: use rte_mp_msg to pass TSC hz to secondary procs
Checks
Commit Message
rte_eal_init() is much faster in secondary processes since
hugepages don't need to be zeroed. But there's still
non-trivial delays in the timer subsystem initialization
due to the 100ms sleep used to calculate TSC hz. So use
the rte_mp_msg framework to allow secondary processes
to get the TSC hz from the primary process.
This cuts rte_eal_init() execution time in a secondary
process from 165ms to 66ms in my test program.
Signed-off-by: Jim Harris <james.r.harris@intel.com>
---
lib/librte_eal/common/eal_common_timer.c | 70 +++++++++++++++++++++++++++++-
1 file changed, 68 insertions(+), 2 deletions(-)
Comments
On 15-Aug-19 12:27 PM, Jim Harris wrote:
> rte_eal_init() is much faster in secondary processes since
> hugepages don't need to be zeroed. But there's still
> non-trivial delays in the timer subsystem initialization
> due to the 100ms sleep used to calculate TSC hz. So use
> the rte_mp_msg framework to allow secondary processes
> to get the TSC hz from the primary process.
>
> This cuts rte_eal_init() execution time in a secondary
> process from 165ms to 66ms in my test program.
>
> Signed-off-by: Jim Harris <james.r.harris@intel.com>
> ---
<snip>
> @@ -89,6 +96,65 @@ set_tsc_freq(void)
> eal_tsc_resolution_hz = freq;
> }
>
> +static void
> +set_tsc_freq_secondary(void)
> +{
> + struct rte_mp_msg mp_req;
> + struct rte_mp_reply mp_reply;
> + struct timer_mp_param *r;
> + struct timespec ts = {.tv_sec = 1, .tv_nsec = 0};
> +
> + memset(&mp_req, 0, sizeof(mp_req));
> + strcpy(mp_req.name, EAL_TIMER_MP);
> + if (rte_mp_request_sync(&mp_req, &mp_reply, &ts) || mp_reply.nb_received != 1) {
If rte_mp_request_sync returns 0 but mp_reply.nb_receieved isn't set to
1, you'll be leaking mp_reply.msgs.
On 8/16/19, 12:56 AM, "Burakov, Anatoly" <anatoly.burakov@intel.com> wrote:
<snip>
> @@ -89,6 +96,65 @@ set_tsc_freq(void)
> eal_tsc_resolution_hz = freq;
> }
>
> +static void
> +set_tsc_freq_secondary(void)
> +{
> + struct rte_mp_msg mp_req;
> + struct rte_mp_reply mp_reply;
> + struct timer_mp_param *r;
> + struct timespec ts = {.tv_sec = 1, .tv_nsec = 0};
> +
> + memset(&mp_req, 0, sizeof(mp_req));
> + strcpy(mp_req.name, EAL_TIMER_MP);
> + if (rte_mp_request_sync(&mp_req, &mp_reply, &ts) || mp_reply.nb_received != 1) {
If rte_mp_request_sync returns 0 but mp_reply.nb_receieved isn't set to
1, you'll be leaking mp_reply.msgs.
Ha - of course you're right. I didn't notice this when using the VFIO code as a reference for
how to code this. I'll respin this patch and then push patches to fix the VFIO code separately.
Thanks,
-Jim
On Thu, Aug 15, 2019 at 04:27:35AM -0700, Jim Harris wrote:
> rte_eal_init() is much faster in secondary processes since
> hugepages don't need to be zeroed. But there's still
> non-trivial delays in the timer subsystem initialization
> due to the 100ms sleep used to calculate TSC hz. So use
> the rte_mp_msg framework to allow secondary processes
> to get the TSC hz from the primary process.
>
> This cuts rte_eal_init() execution time in a secondary
> process from 165ms to 66ms in my test program.
>
> Signed-off-by: Jim Harris <james.r.harris@intel.com>
> ---
Rather than messaging, can we not just move the CPU frequency to being
stored in a shared memory location? It's not something where different
processes are going to need to be provided with different values.
@@ -15,9 +15,16 @@
#include <rte_log.h>
#include <rte_cycles.h>
#include <rte_pause.h>
+#include <rte_eal.h>
#include "eal_private.h"
+#define EAL_TIMER_MP "eal_timer_mp_sync"
+
+struct timer_mp_param {
+ uint64_t tsc;
+};
+
/* The frequency of the RDTSC timer resolution */
static uint64_t eal_tsc_resolution_hz;
@@ -74,8 +81,8 @@ estimate_tsc_freq(void)
return RTE_ALIGN_MUL_NEAR(rte_rdtsc() - start, CYC_PER_10MHZ);
}
-void
-set_tsc_freq(void)
+static void
+set_tsc_freq_primary(void)
{
uint64_t freq;
@@ -89,6 +96,65 @@ set_tsc_freq(void)
eal_tsc_resolution_hz = freq;
}
+static void
+set_tsc_freq_secondary(void)
+{
+ struct rte_mp_msg mp_req;
+ struct rte_mp_reply mp_reply;
+ struct timer_mp_param *r;
+ struct timespec ts = {.tv_sec = 1, .tv_nsec = 0};
+
+ memset(&mp_req, 0, sizeof(mp_req));
+ strcpy(mp_req.name, EAL_TIMER_MP);
+ if (rte_mp_request_sync(&mp_req, &mp_reply, &ts) || mp_reply.nb_received != 1) {
+ /* We weren't able to get the tsc hz from the primary process. So we will
+ * just calculate it here in the secondary process instead.
+ */
+ set_tsc_freq_primary();
+ return;
+ }
+
+ r = (struct timer_mp_param *)mp_reply.msgs[0].param;
+ eal_tsc_resolution_hz = r->tsc;
+ free(mp_reply.msgs);
+}
+
+static int
+timer_mp_primary(__attribute__((unused)) const struct rte_mp_msg *msg, const void *peer)
+{
+ struct rte_mp_msg reply;
+ struct timer_mp_param *r = (struct timer_mp_param *)reply.param;
+
+ memset(&reply, 0, sizeof(reply));
+ r->tsc = 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)
+{
+ int rc;
+
+ /* We use a 100ms timer to calculate the TSC hz. We can save this 100ms in
+ * secondary processes, by getting the TSC hz from the primary process.
+ * So register an mp_action callback in the primary process, which secondary
+ * processes will use to get the TSC hz.
+ */
+ if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
+ set_tsc_freq_primary();
+ rc = rte_mp_action_register(EAL_TIMER_MP, timer_mp_primary);
+ if (rc) {
+ RTE_LOG(WARNING, EAL, "Could not register mp_action - secondary "
+ " processes will calculate TSC independently.\n");
+ }
+ } else {
+ set_tsc_freq_secondary();
+ }
+}
+
void rte_delay_us_callback_register(void (*userfunc)(unsigned int))
{
rte_delay_us = userfunc;