when throttling through proxy traffic 1000TPS within a few minutes we are seeing a memorry growth of some MB, continuing upwards until end of test.
When looking at the INK allocs and INK free's,
Our custom plugin transactions use a scheduleNext routine which schedules the next part of the transaction using 'TSHttpSchedule(m_contp, m_txn, 0);'.. this would seem the correct API as its a HTTP proxy we are running.
TSContCreates (INKContAllocator.alloc()) and TSContDestroys match up however INKContInternal::destroy does not simply delete the memory but checks a m_deletable var before doing so - this var is controlled by m_event_count - but in our case the count does not appear to be handled correctly - i.e. this counter is decremented and incremented depending on conditions.
It does not appear to be getting incremented in our case => memory never gets deleted as m_deletable var doesn't get set => we see a massive memory leak -
Looking at the code, the TSContSchedule functions correctly increment the event counter but TSHttpSchedule does not appear to -
i added in an incrementor:
INKContInternal *i = (INKContInternal *) contp;
if (ink_atomic_increment((int *) &i->m_event_count, 1) < 0)
ink_assert (!"not reached");
in TSHttpSchedule and reload tested - counter is correctly handled and memory leak is fixed -
IS this a bug in ATS or have i missed the correct way to run our http plugin using TSHttpSchedule (i can't see how we are mean't to use TSContSchedule as its TSHttpSchedule which requires the txn parameter so i think this is the correct API we are using)
using 3.0.2 and can supply above as a patch if required