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 */
next prev parent 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