svn commit: r231814 - in head/sys: kern sys
brde at optusnet.com.au
Thu Feb 16 19:25:46 UTC 2012
On Thu, 16 Feb 2012, Kenneth D. Merry wrote:
> On Thu, Feb 16, 2012 at 11:13:09 +0200, Andriy Gapon wrote:
>> on 16/02/2012 11:00 Andriy Gapon said the following:
>>> on 16/02/2012 10:53 Julian Elischer said the following:
>>>> Bruce, this is a good example of a legitimate gripe going un-noticed because
>>>> you didn't shout loud enough at the right time, at the right people.
>>>> It's been about 20 years since we started working on this but I've finally
>>>> come to the point of saying that we need you to do more when you see problems.
>>>> object officially if you think things should be backed out!
>>>> your reasons here seem sound, so it's hard to see why you haven't been more
>>>> public about it.
>>> Just for the record: Bruce and I voiced opinions against the commit when it went
>>> in after a rather short notice for such an important thing. The opinions
>>> can be found in the archives of these lists.
>> To be more precise: s/opinions against/concerns about behavior in the edge cases/.
>> For me personally the immediate benefits in the common situations outweighed the
>> problems in the edge cases, although I still believe that we can get the former
>> without sacrifices in the latter.
> I will happily back the message buffer locking changes out if you and Bruce
> agree that that is the best thing to do.
Thanks. I'm not sure what backing it out would take us back to.
> I can also review any proposed fixes to the message buffer locking code,
> but I really don't have the bandwidth right now to come up with the "real"
> Sorry I didn't get around to dealing with it, I should have said something
> months ago.
> Now that you know a fix won't be coming from me, you two can let me know
> whether you'd like the changes backed out, or you are certainly free to
> commit a solution yourselves.
I seem to remember that I asked you to test my old simple changes for
printf serialization on your larger systems that apparently generates
lots of contending printfs.
Kernel printfs should be a very rare operation, and kernel syslog messages
not much more so, so concurrent printfs should be rare-squared. They
are a non-problem for me, so my old changes haven't been tested much.
But I needed serialization the other day for my printfs in trap() to
debug a deadlock in ddb, since the deadlock is related to to trap()
racing itself. I actually used PRINTF_BUFR_SIZE in a ~2008 kernel for
this. This is actually a good test of printf() under heavy load. Any
line buffering for printf() would have either broken the output format
(I wanted to show many checkpoint passes per line) or changed the timing
enough to make the race go away.
My changes do the following to serialize per printf (not per line):
- try to acquire a nonstandard spinlock implemented using atomic_cmpset.
- spin for up to ~1 second (or more like 1 ms if you want) waiting for it
- ignore the lock after the timeout and proceeed (with complications to
not let multiple CPUs proceed at the same time, I hope)
- when multiple CPUs proceed, use atomic atomic ops as before to avoid
races in msgbuf accesses, and depend on console drivers being correct
to avoid deadlocks and races in them. Most aren't correct.
The worst that is supposed to happen is for deadlock on the nonstandard
spinlock to occur except for the timeout; then the timeout reduces to
interleaved output with bugs limited to console drivers by correct
msgbuf code. The deadlock on the nonstanded spinlock occurs in buggy
trap handlers for traps in code holding the lock (mainly ddb traps).
Then it mainly gives the amusing behaviour that the buggy code is
punished by printing its output very slowly.
More information about the svn-src-head