On 08/03/16 11:40 PM, Bruce Evans wrote:
> On Wed, 3 Aug 2016, Jung-uk Kim wrote:
>
>> Log:
>> Support nanosecond time stamps for pcap_dispatch(3) and pcap_loop(3).
>>
>> Modified:
>> head/contrib/libpcap/pcap-bpf.c
>>
>> Modified: head/contrib/libpcap/pcap-bpf.c
>> ==============================================================================
>>
>> --- head/contrib/libpcap/pcap-bpf.c Wed Aug 3 19:23:22 2016
>> (r303732)
>> +++ head/contrib/libpcap/pcap-bpf.c Wed Aug 3 20:08:39 2016
>> (r303733)
>> @@ -1008,7 +1028,25 @@ pcap_read_bpf(pcap_t *p, int cnt, pcap_h
>> if (pb->filtering_in_kernel ||
>> bpf_filter(p->fcode.bf_insns, datap, bhp->bh_datalen,
>> caplen)) {
>> struct pcap_pkthdr pkthdr;
>> +#ifdef BIOCSTSTAMP
>> + struct bintime bt;
>> +
>> + bt.sec = bhp->bh_tstamp.bt_sec;
>> + bt.frac = bhp->bh_tstamp.bt_frac;
>
> The names are very confusing since bt_sec and bt_frac are only misnamed as
> sec and frac in struct bintime.
bt_* means BPF timestamp, not bintime.
>> + if (p->opt.tstamp_precision == PCAP_TSTAMP_PRECISION_NANO) {
>> + struct timespec ts;
>> +
>> + bintime2timespec(&bt, &ts);
>> + pkthdr.ts.tv_sec = ts.tv_sec;
>> + pkthdr.ts.tv_usec = ts.tv_nsec;
>
> And this abuses tv_usec to hold nanoseconds.
libpcap decided to "abuse" it few years ago, not to break ABI/API.
http://www.tcpdump.org/manpages/pcap-savefile.5.html
> Old code is even more confusing, and at least partly wrong.
>
> X contrib/libpcap/pcap-bpf.c: pkthdr.ts.tv_usec =
> bhp->bh_tstamp.tv_usec/1000;
>
> This is to convert for tv_usec actually being tv_nsec on AIX. If the above
> works with no conversion, then it might work for AIX too.
Of course. However, we don't maintain AIX. ;-)
> X sys/net/bpf.c: struct timeval32 bh_tstamp; /* time stamp */
>
> Banal comment. The complexities are from what sort of timestamp this is.
> It is obviously a timestamp.
>
> This bh_tstamp is in struct bpf_hdr32 for the !BURN_BRIDGES case. There
> is also struct timeval bh_timestamp in struct bpf_hdr. This header is
> bogusly marked Obsolete.
It is superceded by struct bpf_xhdr although we kept it for backward
compatibility. New applications should avoid it.
> X sys/net/bpf.c: hdr32_old.bh_tstamp.tv_usec = ts.bt_frac;
>
> This is in the !BURN_BRIDGES && COMPAT_FREEBSD32 case. Since struct
> timeval32
> always has a 32-bit tv_usec, this assignment discards the most significant
> bits in bt_frac but keeps the noise.
>
> X sys/net/bpf.c: hdr_old.bh_tstamp.tv_usec = ts.bt_frac;
>
> This is in the !BURN_BRIDGES && !COMPAT_FREEBSD32 case. Since tv_sec in a
> normal timetamp is bogusly long, this accidentally preserves all of the
> bits
> in bt_frac on 64-bit arches. On 32-bit arches, it loses the signal as for
> the COMPAT_FREEBSD32 case.
Note struct bpf_ts is not struct bintime.
Jung-uk Kim
signature.asc
Description: OpenPGP digital signature
