By accident the size of the empty dict keys object matched the
size of values array.

Few bytes at the begin and at the end of the reallocated blocks, as well
as the header and the trailer, now are erased before calling realloc()
in debug build.  This will help to detect using or double freeing the
reallocated block.

bpo-31764: Prevent a crash in sqlite3.Cursor.close() in case the Cursor object is uninitialized (#3958)

bpo-31770: Prevent a crash and refleaks when calling sqlite3.Cursor.__init__() more than once (#3968)

blocksize was hardcoded to 8192, preventing efficient upload when using
file-like body. Add blocksize argument to __init__, so users can
configure the blocksize to fit their needs.

I tested this uploading data from /dev/zero to a web server dropping the
received data, to test the overhead of the HTTPConnection.send() with a
file-like object.

Here is an example 10g upload with the default buffer size (8192):

$ time ~/src/cpython/release/python upload-httplib.py 10 https://localhost:8000/
Uploaded 10.00g in 17.53 seconds (584.00m/s)

real	0m17.574s
user	0m8.887s
sys	0m5.971s

Same with 512k blocksize:

$ time ~/src/cpython/release/python upload-httplib.py 10 https://localhost:8000/
Uploaded 10.00g in 6.60 seconds (1551.15m/s)

real	0m6.641s
user	0m3.426s
sys	0m2.162s

In real world usage the difference will be smaller, depending on the
local and remote storage and the network.

See https://github.com/nirs/http-bench for more info.

MemoryError raised when normalizing a RecursionError raised during exception normalization now not always causes a fatal error.

While technically a purely internal change, bpo-31845 was
a fairly significant externally visible bug caused by
these changes (environment variable based configuration
was being ignored due to a change in the relative order
of reading the environment and reading command line settings,
and the test suite was only testing the command line options)

Hence this note to essentially say "If you see odd startup
problems in 3.7 that you've never seen in previous releases,
it's probably our fault, so let us know, and we'll fix it".

This speeds up it to 20-40%.

* bpo-21423: Add an initializer argument to {Process,Thread}PoolExecutor

* Fix docstring

* Using ifconfig on NetBSD and OpenBSD.
* Using arp on Linux, FreeBSD, NetBSD and OpenBSD.

Based on patch by Takayuki Shimizukawa.

bpo-31936

All Blake2 params have to be encoded in little-endian byte order. For
the two multi-byte integer params, leaf_length and node_offset, that
means that assigning a native-endian integer to them appears to work on
little-endian platforms, but gives the wrong result on big-endian. The
current libb2 API doesn't make that very clear, and @sneves is working
on new API functions in the GH issue above. In the meantime, we can work
around the problem by explicitly assigning little-endian values to the
parameter block.

See https://github.com/BLAKE2/libb2/issues/12.

Bug report and patch by Jeroen Demeyer.

* bpo-31310: multiprocessing's semaphore tracker should be launched again if crashed

* Avoid mucking with process state in test.
Add a warning if the semaphore process died, as semaphores may then be leaked.

* Add NEWS entry

* bpo-31308: If multiprocessing's forkserver dies, launch it again when necessary.

* Fix test on Windows

* Add NEWS entry

* Adopt a different approach: ignore SIGINT and SIGTERM, as in semaphore tracker.

* Fix comment

* Make sure the test doesn't muck with process state

* Also test previously-started processes

* Update 2017-08-30-17-59-36.bpo-31308.KbexyC.rst

* Avoid masking SIGTERM in forkserver.  It's not necessary and causes a race condition in test_many_processes.

When a single .c file contains several functions and/or methods with
the same name, a safety _METHODDEF #define statement is generated
only for one of them.

This fixes the bug by using the full name of the function to avoid
duplicates rather than just the name.

Support PYTHONPROFILEIMPORTTIME envvar equivalent to -X importtime

Add METH_STACKLESS to prevent future collisions.

* bpo-28643: Record profile-opt build progress with stamp files

The profile-opt makefile target is expensive to build. Since the
makefile does not contain complete dependency information for this
target, much extra work can get done if the build is interrupted and
re-started.  Even running "make" a second time will result in a huge
amount of redundant work.

As a minimal fix (rather than removing recursive "make" and adding a
proper dependency graph), split the profile-opt target into parts:

- ensure tree is clean (profile-clean-stamp)
- build with profile generation enabled (profile-gen-stamp)
- run task to generate profile information (profile-run-stamp)
- build optimized Python using above information (profile-opt)

We use "stamp" files to record completion of the steps.  Running
"make clean" will not remove the profile-run-stamp file.

Other minor changes:

- remove the "build_all_use_profile" target.  I don't expect callers
  of the makefile to use this target so that should be safe.

- remove execution of "profile-removal" at end of "profile-opt".  I
  don't see any reason to not to keep the profile information, given
  the cost to generate it.  Removing the "profile-run-stamp" file
  will force re-generation of it.

Add new time functions:

* time.clock_gettime_ns()
* time.clock_settime_ns()
* time.monotonic_ns()
* time.perf_counter_ns()
* time.process_time_ns()
* time.time_ns()

Add new _PyTime functions:

* _PyTime_FromTimespec()
* _PyTime_FromNanosecondsObject()
* _PyTime_FromTimeval()

Other changes:

* Add also os.times() tests to test_os.
* pytime_fromtimeval() and pytime_fromtimeval() now return
  _PyTime_MAX or _PyTime_MIN on overflow, rather than undefined
  behaviour
* _PyTime_FromNanoseconds() parameter type changes from long long to
  _PyTime_t