libc - Fix bogus pthread_getspecific() return value due to bug in nmalloc

* nmalloc was calling pthread_set_specific() prior to calling
pthread_key_create(), causing it to use key 0 which might al

libc - Fix bogus pthread_getspecific() return value due to bug in nmalloc

* nmalloc was calling pthread_set_specific() prior to calling
pthread_key_create(), causing it to use key 0 which might already
have been allocated for other purposes.

* Reorder initializations in _nmalloc_thr_init() to solve the problem.

* This also solves certain application crashes (mail/milter-greylist).

Reported-by: Francois Tigeot <ftigeot@wolfpond.org>

show more ...

libc: Fix warnings fallout from the change to CSTD=gnu99.

Add __printflike's where possible and fix all related bugs & issues.

nmalloc - Remove startup locking

* A spinlock is no longer needed for malloc_init, it is run as a constructor
now, as such it is run before any other threads are started.

nmalloc - Optimize initialization function out of hot path

* Add a __constructor macro to cdefs.h

* Turn malloc_init into a gcc constructor

* Modify ifconfig to use the cdefs macro instead of its

nmalloc - Optimize initialization function out of hot path

* Add a __constructor macro to cdefs.h

* Turn malloc_init into a gcc constructor

* Modify ifconfig to use the cdefs macro instead of its own

show more ...

nmalloc - Fix depot race - this should fix remaining issues with firefox.

* nmalloc was accidently calling the libc _pthread_spin_lock() stub
even when threaded, causing all depot operations to ra

nmalloc - Fix depot race - this should fix remaining issues with firefox.

* nmalloc was accidently calling the libc _pthread_spin_lock() stub
even when threaded, causing all depot operations to race. No actual
lock was being acquired.

Change it over to the internal _SPINLOCK() API.

* Expand libthread_xu's base spinlocks from 128 to 256 and use those
for the depot lock. Note that the depot[] array can be upwards of
80 elements so lets not take any chances.

Reported-by: Pierre Abbat <phma@phma.optus.nu>, others

show more ...

libc - Move the newmag prestage code

* Avoid any possible assumptions due to reentrnacy by moving
the pre-stage code.

libc - Do some reorganization of nmalloc() and add asserts.

* Reorganize mtmagazine_alloc() and mtmagazine_free() to make the code
more clear. Add assertions on 'continue' loops so we abort on

libc - Do some reorganization of nmalloc() and add asserts.

* Reorganize mtmagazine_alloc() and mtmagazine_free() to make the code
more clear. Add assertions on 'continue' loops so we abort on
magazine memory corruption instead of looping forever.

* Pre-stage the new magazine that mtmagazine_free() might need to avoid
having to relock the depot. This also allows us to stage the new
magazine directory into the per-thread loaded field instead of running
it through the depot.

* Fix a deep or infinite recursion which can occur when mtmagazine_free()
calls _slaballoc(), which it does sometimes when it needs a new
magazine. Avoid reentrancy by temporarily disabling the per-thread
cache around the call.

show more ...

libc - Fix livelock in nmalloc

* free() -> _slabfree() was turning around and calling _slaballoc()
with the depot lock held, which could lead to a livelock. Fix by
unlocking the depot lock arou

libc - Fix livelock in nmalloc

* free() -> _slabfree() was turning around and calling _slaballoc()
with the depot lock held, which could lead to a livelock. Fix by
unlocking the depot lock around the call.

* Clean up and document other unrelated bits of nmalloc(). In particular,
clean up magazine_alloc().

Reported-by: Siju George <sgeorge.ml@gmail.com>, and others

show more ...

libc - Fix some recursion issues during thread teardown

* thread destructors called by sophisticated pthreaded programs such
as firefox can wind up allocating/freeing space multiple times
AFTER

libc - Fix some recursion issues during thread teardown

* thread destructors called by sophisticated pthreaded programs such
as firefox can wind up allocating/freeing space multiple times
AFTER nmalloc's destructor is run.

This can leave the mtmagazine in a weird state because it's destructor
fails to NULL-out tp->mags[i].loaded and tp->mags[i].prev.

* Properly NULL out tp->mags[i].{loaded,prev} in the destructor.

* Lockout mtmagazine use while the mtmagazine is being initialized or
destroyed.

* Permanently lockout mtmagazine use after its destructor has been run.

* Instead of trying to initialize the mtmagazine on the first free, which
might not occur until the destrutor is run (causing pthreads to complain
about destructors being left hanging), have libthread_xu call a new
function _nmalloc_thr_init() during thread creation and do all the slab
initialization there.

show more ...