patches for DPDK stable branches
 help / color / mirror / Atom feed
From: Quentin Armitage <quentin@armitage.org.uk>
To: Stephen Hemminger <stephen@networkplumber.org>
Cc: Reshma Pattan <reshma.pattan@intel.com>,
	Ray Kinsella <mdr@ashroe.eu>,
	 dev@dpdk.org, stable@dpdk.org
Subject: Re: [PATCH] libpcapng: fix timestamp wrapping in output files
Date: Sat, 14 May 2022 18:14:48 +0100
Message-ID: <45bdbfad3857848faa2edc185db1ac639a901b23.camel@armitage.org.uk> (raw)
In-Reply-To: <20220511094655.4f885c84@hermes.local>

On Wed, 2022-05-11 at 09:46 -0700, Stephen Hemminger wrote:
> On Sat,  7 May 2022 17:12:36 +0100
> Quentin Armitage <quentin@armitage.org.uk> wrote:
> 
> > In pcap_tsc_to_ns(), delta * NSEC_PER_SEC will overflow approx 8
> > seconds after pcap_init is called when using a TSC with a frequency
> > of 2.5GHz.
> > 
> > To avoid the overflow, reread the time and TSC once
> > delta * NSEC_PER_SEC > (1 << 63). In order to ensure that there
> > is no overflow if there is a several second gap between calls to
> > pcapng_tsc_to_ns() the actual check to reread the clock is:
> >   delta > ((1ULL << 63) / NSEC_PER_SEC)
> > 
> > Fixes: 8d23ce8f5ee ("pcapng: add new library for writing pcapng files")
> > Cc: stable@dpdk.org
> > 
> > Signed-off-by: Quentin Armitage <quentin@armitage.org.uk>
> 
> What about something like this instead.
> 
> diff --git a/lib/pcapng/rte_pcapng.c b/lib/pcapng/rte_pcapng.c
> index 90b2f5bc6905..c5534301bf2c 100644
> --- a/lib/pcapng/rte_pcapng.c
> +++ b/lib/pcapng/rte_pcapng.c
> @@ -19,6 +19,7 @@
>  #include <rte_ether.h>
>  #include <rte_mbuf.h>
>  #include <rte_pcapng.h>
> +#include <rte_reciprocal.h>
>  #include <rte_time.h>
>  
>  #include "pcapng_proto.h"
> @@ -34,27 +35,39 @@ struct rte_pcapng {
>  };
>  
>  /* For converting TSC cycles to PCAPNG ns format */
> -struct pcapng_time {
> +#define TICK_SCALE 16u
> +static struct {
>         uint64_t ns;
>         uint64_t cycles;
> +       struct rte_reciprocal_u64 inverse;
>  } pcapng_time;
>  
>  RTE_INIT(pcapng_init)
>  {
>         struct timespec ts;
> +       uint64_t scale_tick_per_ns;
>  
>         pcapng_time.cycles = rte_get_tsc_cycles();
>         clock_gettime(CLOCK_REALTIME, &ts);
>         pcapng_time.ns = rte_timespec_to_ns(&ts);
> +
> +       scale_tick_per_ns = (rte_get_tsc_hz() * TICK_SCALE) / NSEC_PER_SEC;
> +       pcapng_time.inverse = rte_reciprocal_value_u64(scale_tick_per_ns);
>  }
>  
>  /* PCAPNG timestamps are in nanoseconds */
>  static uint64_t pcapng_tsc_to_ns(uint64_t cycles)
>  {
> -       uint64_t delta;
> +       uint64_t delta, elapsed;
>  
>         delta = cycles - pcapng_time.cycles;
> -       return pcapng_time.ns + (delta * NSEC_PER_SEC) / rte_get_tsc_hz();
> +
> +       /* Compute elapsed time in nanoseconds scaled by TICK_SCALE
> +        * since the start of the capture.
> +        * With scale of 4 this will roll over in 36 years.
> +        */
> +       elapsed = rte_reciprocal_divide_u64(delta, &pcapng_time.inverse);
> +       return pcapng_time.ns + elapsed / TICK_SCALE;
>  }
>  
>  /* length of option including padding */
> 

The final statement of pcapng_tsc_to_ns() should be:
	return pcapng_time.ns + elapsed * TICK_SCALE;

There is also a problem that rte_get_tsc_hz() returns eal_tsc_resolution_hz, but
this is not initialized until rte_eal_init() is called, so rte_get_tsc_hz()
cannot be called from a constructor function.

While both of the above problems can easily be solved, I think there is a
problem with accuracy with this approach. With a 3GHz clock, scale_tick_per_ns
would be 48. For other clock speeds there can be a truncation in the
calculation. With a 3.3GHz clock, scale_tick_per_ns will be truncated from 52.8
to 52, resulting in a 1.5% or so error in the time returned by
pcapng_tsc_to_ns() (a 2.3GHz clock results in a 2.2% error). Increasing
TICK_SCALE reduces the %age error, but also reduces the time before overflow
occurs.

If the approach in the following patch is considered to be acceptable, I would
be very happy to submit an updated patch. The one concern I have about the patch
is introducing a new constructor priority, RTE_PRIORITY_TIMER, which may be
considered to be inappropriate. If it is inappropriate, then the simplest
alternative would be to introduce a new function rte_tsc_get_hz_init() which
calls set_tsc_freq() if eal_tsc_resolution_hz has not been initialized
(alternatively rte_get_tsc_hz() could be modified to make the check, but that
then produces an overhead every time the function is called).

diff --git a/lib/eal/common/eal_common_timer.c
b/lib/eal/common/eal_common_timer.c
index 5686a5102b..cb3fa1e240 100644
--- a/lib/eal/common/eal_common_timer.c
+++ b/lib/eal/common/eal_common_timer.c
@@ -54,6 +54,9 @@ set_tsc_freq(void)
 	struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
 	uint64_t freq;
 
+	if (eal_tsc_resolution_hz)
+		return;
+
 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
 		/*
 		 * Just use the primary process calculated TSC rate in any
@@ -86,3 +89,8 @@ RTE_INIT(rte_timer_init)
 	/* set rte_delay_us_block as a delay function */
 	rte_delay_us_callback_register(rte_delay_us_block);
 }
+
+RTE_INIT_PRIO(rte_tsc_init, TIMER)
+{
+	set_tsc_freq();
+}
diff --git a/lib/eal/include/rte_common.h b/lib/eal/include/rte_common.h
index 67587025ab..a0d64ff4f2 100644
--- a/lib/eal/include/rte_common.h
+++ b/lib/eal/include/rte_common.h
@@ -161,6 +161,7 @@ typedef uint16_t unaligned_uint16_t;
 
 #define RTE_PRIORITY_LOG 101
 #define RTE_PRIORITY_BUS 110
+#define RTE_PRIORITY_TIMER 115
 #define RTE_PRIORITY_CLASS 120
 #define RTE_PRIORITY_LAST 65535
 
diff --git a/lib/pcapng/rte_pcapng.c b/lib/pcapng/rte_pcapng.c
index 90b2f5bc69..09d42bbc9a 100644
--- a/lib/pcapng/rte_pcapng.c
+++ b/lib/pcapng/rte_pcapng.c
@@ -19,6 +19,7 @@
 #include <rte_ether.h>
 #include <rte_mbuf.h>
 #include <rte_pcapng.h>
+#include <rte_reciprocal.h>
 #include <rte_time.h>
 
 #include "pcapng_proto.h"
@@ -34,9 +35,11 @@ struct rte_pcapng {
 };
 
 /* For converting TSC cycles to PCAPNG ns format */
-struct pcapng_time {
+static struct {
 	uint64_t ns;
 	uint64_t cycles;
+	uint64_t tsc_hz;
+	struct rte_reciprocal_u64 tsc_hz_inverse;
 } pcapng_time;
 
 RTE_INIT(pcapng_init)
@@ -45,16 +48,45 @@ RTE_INIT(pcapng_init)
 
 	pcapng_time.cycles = rte_get_tsc_cycles();
 	clock_gettime(CLOCK_REALTIME, &ts);
+	pcapng_time.cycles = (pcapng_time.cycles + rte_get_tsc_cycles()) / 2;
 	pcapng_time.ns = rte_timespec_to_ns(&ts);
+
+	pcapng_time.tsc_hz = rte_get_tsc_hz();
+	pcapng_time.tsc_hz_inverse =
rte_reciprocal_value_u64(pcapng_time.tsc_hz);
 }
 
 /* PCAPNG timestamps are in nanoseconds */
 static uint64_t pcapng_tsc_to_ns(uint64_t cycles)
 {
-	uint64_t delta;
-
+	uint64_t delta, secs;
+
+	/* In essence the calculation is:
+	 *   delta = (cycles - pcapng_time.cycles) * NSEC_PRE_SEC /
rte_get_tsc_hz()
+	 * but this overflows within 4 to 8 seconds depending on TSC frequency.
+	 * Instead, if delta >= pcapng_time.tsc_hz:
+	 *   Increase pcapng_time.ns and pcapng_time.cycles by the number of
+	 *   whole seconds in delta and reduce delta accordingly.
+	 * delta will therefore always lie in the interval [0,
pcapng_time.tsc_hz),
+	 * which will not overflow when multiplied by NSEC_PER_SEC provided the
+	 * TSC frequency < approx 18.4GHz.
+	 *
+	 * Currently all TSCs operate below 5GHz.
+	 */
 	delta = cycles - pcapng_time.cycles;
-	return pcapng_time.ns + (delta * NSEC_PER_SEC) / rte_get_tsc_hz();
+	if (unlikely(delta >= pcapng_time.tsc_hz)) {
+		if (likely(delta < pcapng_time.tsc_hz * 2)) {
+			delta -= pcapng_time.tsc_hz;
+			pcapng_time.cycles += pcapng_time.tsc_hz;
+			pcapng_time.ns += NSEC_PER_SEC;
+		} else {
+			secs = rte_reciprocal_divide_u64(delta,
&pcapng_time.tsc_hz_inverse);
+			delta -= secs * pcapng_time.tsc_hz;
+			pcapng_time.cycles += secs * pcapng_time.tsc_hz;
+			pcapng_time.ns += secs * NSEC_PER_SEC;
+		}
+	}
+
+	return pcapng_time.ns + rte_reciprocal_divide_u64(delta * NSEC_PER_SEC,
&pcapng_time.tsc_hz_inverse);
 }
 
 /* length of option including padding */



  reply	other threads:[~2022-05-14 17:15 UTC|newest]

Thread overview: 9+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2022-05-07 16:12 Quentin Armitage
2022-05-11 16:08 ` Stephen Hemminger
2022-05-11 16:46 ` Stephen Hemminger
2022-05-14 17:14   ` Quentin Armitage [this message]
2022-05-16 13:26     ` Stephen Hemminger
2022-05-17 10:01   ` [PATCH v2] " Quentin Armitage
2022-05-17 15:15     ` Stephen Hemminger
2022-06-01 14:38       ` Thomas Monjalon
2022-05-17 21:04     ` Stephen Hemminger

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=45bdbfad3857848faa2edc185db1ac639a901b23.camel@armitage.org.uk \
    --to=quentin@armitage.org.uk \
    --cc=dev@dpdk.org \
    --cc=mdr@ashroe.eu \
    --cc=reshma.pattan@intel.com \
    --cc=stable@dpdk.org \
    --cc=stephen@networkplumber.org \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link

patches for DPDK stable branches

This inbox may be cloned and mirrored by anyone:

	git clone --mirror http://inbox.dpdk.org/stable/0 stable/git/0.git

	# If you have public-inbox 1.1+ installed, you may
	# initialize and index your mirror using the following commands:
	public-inbox-init -V2 stable stable/ http://inbox.dpdk.org/stable \
		stable@dpdk.org
	public-inbox-index stable

Example config snippet for mirrors.
Newsgroup available over NNTP:
	nntp://inbox.dpdk.org/inbox.dpdk.stable


AGPL code for this site: git clone https://public-inbox.org/public-inbox.git