svn commit: r220755 - in head: . contrib/gcc/doc contrib/gcc/objc contrib/libobjc etc/mtree gnu/lib gnu/lib/libobjc gnu/usr.bin/cc gnu/usr.bin/cc/cc1obj gnu/usr.bin/cc/cc_tools gnu/usr.bin/cc/doc s...

[John Baldwin]

On Tuesday, April 19, 2011 2:04:53 pm Warner Losh wrote:
> 
> On Apr 19, 2011, at 6:40 AM, John Baldwin wrote:
> 
> > On Monday, April 18, 2011 3:59:45 pm Warner Losh wrote:
> >> 
> >> On Apr 18, 2011, at 12:18 PM, Doug Barton wrote:
> >> 
> >>> On 04/18/2011 11:14, Pawel Jakub Dawidek wrote:
> >>>> On Mon, Apr 18, 2011 at 11:06:42AM -0600, Warner Losh wrote:
> >>>>> 
> >>>>> On Apr 18, 2011, at 1:01 AM, Roman Divacky wrote:
> >>>>> 
> >>>>>> please mark this in src/UPDATING, maybe bump freebsd_version too?
> >>>>> 
> >>>>> Please do not bump freebsd_version just for this.  Ports wishing to know 
> > can go off the last bump, if there are any.
> >>>>> 
> >>>>> Every freebsd_version bump forces rebuilding all modules and such and is 
> > a pita.
> >>>> 
> >>>> I agree that this is a PITA, but there also should be a way to force
> >>>> module load even on version bump. This is PITA especially for
> >>>> developers.
> >>> 
> >>> .... who make up a tiny percentage of the FreeBSD user community. 
> > Seriously? We're going to whine because version bumps cause a little extra 
> > compile time?
> >> 
> >> The problem usually manifests itself when I got to debug a new problem, load 
> > a driver and find I have to rebuild everything else to use it, which forces an 
> > extra reboot on the machine in question.  Sometimes this can be quite 
> > disruptive to other things that machine is doing.
> > 
> > But that is only true if your source tree doesn't match your installed world.  
> > If the new driver is standalone and you build it as a module, it will use the 
> > headers from /sys and will work fine.
> 
> It is rarely the case that my source matches my install world.  My kernel and userland are asynchronously updated when I'm doing heavy kernel 
development.  My sources are updated all the time to pull in useful things that I need.  When I'm in this mode, I carefully track what's changed 
to make sure there were no recompile the world events.  When a userland thing changes that bumps freebsd_version, I have to recompile the kernel 
and reboot, even though there's no need to do so.  This is especially annoying when it happens a few times within one week, since it is a hit to 
my productivity each time it happens.  There's no great need to have userland features identified with the fidelity of a few dozen commits in -
current.  A few hundred or more would suffice for ports and those interested, so bumping more than once a week provides almost no benefit, but 
inflicts some pain.

Hmm, my development tree never matches my /usr/src either, but I never run
into this.  I always build a kernel in a p4 or whatever checkout and do
'make install KERNEL=foo' (or make installkernel INSTKERNNAME=foo) along
with 'nextboot -k foo' (or 'boot foo' at the loader prompt).  My current
testboxes actually tend to have 7.x worlds and I just boot 9 kernels on
them all the time.  But by building a new kernel each time I never hit
problems with __FreeBSD_version checks.

However, all that said, I would have no problem having something you can
set in make.conf or on the command line to disable the check for a given
module build.

Oh, one other thing, if you stick your module directory in a kernel tree
that matches whatever the running kernel is, then that also works fine.
I do this a lot where I've booted a 9.0 kernel on a 7.x machine, and then
have a sys/modules/foo in that source tree and I can build the module from
that directory and load/unload it without any problems.

> >> In this case, there was a new kernel thing just after, so it turned out OK.  
> > But let's not gratuitously bump the version since the granularity we have 
> > already allows the ports to make good choices on when to leave something in or 
> > out.
> > 
> > Except that that directly contradicts our previously established policy that 
> > these version bumps are cheap and that we should do more of them (this came up 
> > a few years ago when we changed the policy so that the new "stable" branch 
> > after a release starts at N + 500 (e.g. 802500) rather than N + 100 to give 
> > more room for version bumps on current).
> 
> We have never said that version bumps were totally cheap.  In -current, there's little benefit to bumping things more often than once a week.  
Since we generally develop stable for more than 2 years, 100 wasn't enough at the one a week rate.  But that doesn't mean that 5 per week is OK 
either, since now we suddenly have room for them.  I'm strongly suggesting that for userland-only changes that one check to see when the last bump 
was and if it is less than a few days (or a week), then to piggyback off the last one/next one to discriminate whether to use an old feature, or 
start using a new feature.  Especially for retroactive bumps that happen disconnected to the original feature going in.

I can't find the original thread, but we did in fact say that. :)

However, I'd be happy to only have one a week or what not.  Big interface
changes should generally have them though, especially if they affect KPIs used
by out of tree drivers or other software in ports.

-- 
John Baldwin