A hang was encountered, apparently triggered by commit 63b503916d,
changing tls_pathbufs from malloc'ed to HeapAlloc'ed memory. After
lengthy debugging it transpired that adding the heap handle to the
tls_pathbuf struct added 8 bytes to the cygtls area, thus moving
the "context" member by 8 bytes, too, so it was suddently unaligned.
Fix this for now by changing the alignment.
Fix this once and for all, by adding code to the gentls_offsets script
to check if the alignment of the "context" member is 16 bytes. If not,
print a matching error message, remove the just generated file, and exit
with error.
FIXME: It would be really nice to find a way to auomate the correct
alignment of the "context" member, but I don't see any way to use
alignment attributes to get what we need here.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
Rather than using malloc/free for the buffers, we're now using
HeapAlloc/HeapFree on a HEAP_NO_SERIALIZE heap created for this
thread.
Advantages:
- Less contention. Our malloc/free doesn't scale well in
multithreaded scenarios
- Even faster heap allocation by using a non serialized heap.
- Internal, local, temporary data not cluttering the user heap.
- Internal, local, temporary data not copied over to child process
at fork().
Disadvantage:
- A forked process has to start allocating temporary buffers from
scratch. However, this should be alleviated by the fact that
buffer allocation usually reaches its peak very early in process
runtime, so the longer the proceess runs, the less buffers have
to allocated, and, only few processes don't exec after fork
anyway.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>