On Mon, 9 Nov 2009 10:39:10 -0500
Herbert Xu <herbert@gondor.apana.org.au> wrote:

> However, I have some reservations as to whether this is the ideal
> situation.  Unless I'm mistaken, this patch may cause IPsec SAs
> to expire if the system clock was out of sync prior to IPsec startup
> and is subsequently resynced by ntpdate or similar.
> 
> For example, it's quite common for clocks to be out-of-sync by
> 10 hours in Australia due to time zone issues with BIOS clocks.
> So potentially ntpdate could move the clock forward by 10 hours
> or more on bootup thus causing IPsec SAs to expire prematurely
> with this patch.
> 
> This shouldn't really be a problem in itself except that there
> are some dodgy IPsec gateways out there that refuse to reestablish
> IPsec SAs if the interval between two successive connections is
> too small.  This could render the SA inoperable for hours.

But why would it be inoperable for hours? 

I think that the following will happen:
* racoon will recreate SAD entry in the larval state, wait 30s and drop
  it (since dodgy-gw filtered out all keyexchange packets)
* The next time there is a connect() with a match in the SPD, racoon
  will again try to recreate the SAD entry. If there dodgy-gw still
  filters out, the larval SAD entry dies after 30s.

So the inoperability will only last as long as dodgy-gw filters
keyexchanges. 

In any case, running ntpdate before racoon fixes the problem.

> 
> So the upshot of all this is that we definitely want the effect
> of this patch for suspend/resume, but it would be great if we can
> avoid it for settimeofday(2).

I think the natural solution is to have CLOCK_BOOTBASED hrtimers. I.e.
something in the spirit of monotonic_to_bootbased() and getboottime().
I understand that doing +=total_sleep_time is against the core idea of
hires timers, but perhaps there is a nicer way.

Best,
Y