Bluetooth: Mesh: Fix regression in PB-ADV

[trond-snekvik]

After #35702, the provisioner is unable to mark a link as closed, as
it depends on the send_end callback to be called, so it can start its
timer. PB-Adv keeps a reference to the buffers of reliable messages,
which prevents this callback to be invoked, as the buffer destructor is
never called.

Move scheduling of the retransmit timer to the initial transmission, and
replace the timer based LINK_CLOSE message tx duration with a message
counting solution.

Signed-off-by: Trond Einar Snekvik [email protected]

[trond-snekvik]

I'm not 100 sure about this solution, as it's changing the retransmit logic to not take the time spent sending the messages into account, which can make retransmits come earlier than necessary.

I also believe I owe you a beer, @jhedberg, you pointed out that this might happen, and I dismissed it 🙁

[LingaoM]

I'm not 100 sure about this solution, as it's changing the retransmit logic to not take the time spent sending the messages into account, which can make retransmits come earlier than necessary.

I also believe I owe you a beer, @jhedberg, you pointed out that this might happen, and I dismissed it 🙁

Sorry for this, i thought bsim would automatically catch these problems, or it was a mistake. Is it necessary for us to add a check for the available buffer under bsim?

[trond-snekvik]

Sorry for this, i thought bsim would automatically catch these problems, or it was a mistake. Is it necessary for us to add a check for the available buffer under bsim?

That's how I discovered it, actually. #35446 fails because of this bug. I'll rebase it after this goes in, and we'll have coverage.