The global replace in changeset bd80903cf987 was a bit too sweeping: a comment
mentioning the OS function wait() (which exists) was inadvertently changed to
talk about an OS function suspend() (which does not).
For reasons not yet diagnosed, specifically in Mac Release builds, the tests
in test_httprequest.hpp consistently crash with a backtrace suggesting that
the worker thread is calling LLCore::HttpLibcurl::completeRequest() after the
foreground thread calls HttpRequest::destroyService().
Weirdly, even executing a tut::skip() call in every test<n>() function up to
the point of the crash does not eliminate the crash.
NickyD discovered that the substitute default allocator used for llcorehttp
tests was returning badly-aligned storage, which caused access violations on
alignment-sensitive data such as std::atomic. Thanks Nicky!!
Moreover, the llcorehttp test assertions regarding memory usage, well-
intentioned though they are, have been causing us trouble for years. Many have
already been disabled.
The problem is that use of test_allocator.h affected *everything* defined with
that header file's declarations visible. That inevitably included specific
functions in other subsystems. Those functions then (unintentionally) consumed
the special allocator, throwing off the memory tracking and making certain
memory-related assertions consistently fail.
This is a particular, observable bad effect of One Definition Rule violations.
Within a given program, C++ allows multiple definitions for the same entity,
but requires that all such definitions be the same. Partial visibility of the
global operator new() and operator delete() overrides meant that some
definitions of certain entities used the default global allocator, some used
llcorehttp's. There may have been other, more subtle bad effects of these ODR
violations.
If one wanted to reimplement verification of the memory consumption of
llcorehttp classes:
* Each llcorehttp class (for which memory tracking was desired) should declare
class-specific operator new() and operator delete() methods. Naturally,
these would all consume a central llcorehttp-specific allocator, but that
allocator should *not* be named global operator new().
* Presumably that would require runtime indirection to allow using the default
allocator in production while substituting the special allocator for tests.
* Recording and verifying the memory consumption in each test should be
performed in the test-object constructor and destructor, rather than being
sprinkled throughout the test<n>() methods.
* With that mechanism in place, the test object should provide methods to
adjust (or entirely disable) memory verification for a particular test.
* The test object should also provide a "yes, we're still consuming llcorehttp
memory" method to be used for spot checks in the middle of tests -- instead
of sprinkling in explicit comparisons as before.
* In fact, the llcorehttp test object in each test_*.hpp file should be
derived from a central llcorehttp test-object base class providing those
methods.
Using boost::fibers::unbuffered_channel can block the mainthread when calling mPendingCoprocs.push (LLCoprocedurePool::enqueueCoprocedure)
From the documentation:
- If a fiber attempts to send a value through an unbuffered channel and no fiber is waiting to receive the value, the channel will block the sending fiber.
This can happen if LLCoprocedurePool::coprocedureInvokerCoro is running a coroutine and this coroutine calls yield, resuming the viewers main loop. If inside
the main loop someone calls LLCoprocedurePool::enqueueCoprocedure now push will block, as there's no one waiting for a result right now.
The wait would be in LLCoprocedurePool::coprocedureInvokerCoro at the start of the while loop, but we have not reached that yet again as LLCoprocedurePool::coprocedureInvokerCoro
did yield before reaching pop_wait_for.
The result is a deadlock.
boost::fibers::buffered_channel will not block as long as there's space in the channel. A size of 4096 (DEFAULT_QUEUE_SIZE) should be plenty enough for this.
On login failure, LLLogin now tries to sync up with SLVersionChecker. It waits
for up to 10 seconds before shrugging and giving up. Since that coroutine can
now block for that long, make the llogin_test failure cases wait at least that
long too.
Nat Goodspeed