On Sat, May 03, 2025 at 02:21:31AM +0200, Samuel Thibault wrote:
> Zhaoming Luo, le sam. 03 mai 2025 08:17:52 +0800, a ecrit:
> > On Sat, May 03, 2025 at 02:10:12AM +0200, Samuel Thibault wrote:
> > > Zhaoming Luo, le ven. 02 mai 2025 22:45:44 +0800, a ecrit:
> > > > Get the time with higher resolution from kernel using host_get_time64()
> > > > when possible.
> > >
> > > It seems there's a misunderstanding. I explained that the mapped time
> > > does have 64bit precision fields, that's what I mean we want to use,
> > > rather than the costly RPC. When we discussed about it on IRC, I didn't
> > > know that libdiskfs already uses maptime, so advised to just use the RPC
> > > since it's simpler. But since libdiskfs already uses maptime, better
> > > just extend maptime with 64b capability (falling back to 32b if the 64b
> > > fields are zero).
> > >
> > But the issue is that maptime_read does not support 10ns resolution
>
> Honestly, I believe 10ns resolution is less useful for files timestamps
> than using the effective mapfile vs costly RPC. Files take time to
> write, so it's not unexpected to have coarse resolution for them.
>
I agree, but one of the reason I'm interested in it is that some
software tests require high-precision file timestamps. For example vim:
https://github.com/vim/vim/issues/16658.
Zhaoming
>
> > which is host_get_time64 can offer, and I haven't got an idea of how to
> > do that on maptime_read. The function call of hpclock is provided in
> > kernel.
> >
> > Zhaoming
> > >
> > > > - maptime_read (diskfs_mtime, &t);
> > > > + /* Get time from kernel. If host_get_time64 is not supported by the
> > > > + kernel, step back to use maptime_read. */
> > > > + if (host_get_time64 (mach_host_self (), &t) == MIG_BAD_ID)
> > > > + {
> > > > + maptime_read (diskfs_mtime, &mt);
> > > > + t.seconds = mt.tv_sec;
> > > > + t.nanoseconds = mt.tv_usec * 1000;
> > > > + }
> > > >
> > > > /* We are careful to test and reset each of these individually, so
> > > > there
> > > > is no race condition where a dn_set_?time flag setting gets lost.
> > > > It
> > > > @@ -85,22 +94,22 @@ diskfs_set_node_times (struct node *np)
> > > > the update will happen at the next call. */
> > > > if (np->dn_set_mtime)
> > > > {
> > > > - np->dn_stat.st_mtim.tv_sec = t.tv_sec;
> > > > - np->dn_stat.st_mtim.tv_nsec = t.tv_usec * 1000;
> > > > + np->dn_stat.st_mtim.tv_sec = t.seconds;
> > > > + np->dn_stat.st_mtim.tv_nsec = t.nanoseconds;
> > > > np->dn_stat_dirty = 1;
> > > > np->dn_set_mtime = 0;
> > > > }
> > > > if (np->dn_set_atime)
> > > > {
> > > > - np->dn_stat.st_atim.tv_sec = t.tv_sec;
> > > > - np->dn_stat.st_atim.tv_nsec = t.tv_usec * 1000;
> > > > + np->dn_stat.st_atim.tv_sec = t.seconds;
> > > > + np->dn_stat.st_atim.tv_nsec = t.nanoseconds;
> > > > np->dn_stat_dirty = 1;
> > > > np->dn_set_atime = 0;
> > > > }
> > > > if (np->dn_set_ctime)
> > > > {
> > > > - np->dn_stat.st_ctim.tv_sec = t.tv_sec;
> > > > - np->dn_stat.st_ctim.tv_nsec = t.tv_usec * 1000;
> > > > + np->dn_stat.st_ctim.tv_sec = t.seconds;
> > > > + np->dn_stat.st_ctim.tv_nsec = t.nanoseconds;
> > > > np->dn_stat_dirty = 1;
> > > > np->dn_set_ctime = 0;
> > > > }
> > > > --
> > > > 2.49.0
> > > >
> > > >
> > >
> > > --
> > > Samuel
> > > * D a decide de peter un cable dans son rapport de pfp
> > > <c> et il a bien raison ;-)
> > > <c> tu vas dire quoi ?
> > > <D> j'ai mis les paroles de "le coq est mort" en en-tete
> > > -+- #ens-mim et la peufeupeu -+-
> >
>
> --
> Samuel
> * D a decide de peter un cable dans son rapport de pfp
> <c> et il a bien raison ;-)
> <c> tu vas dire quoi ?
> <D> j'ai mis les paroles de "le coq est mort" en en-tete
> -+- #ens-mim et la peufeupeu -+-
