On usual kernel, if a file is mmaped, and then memory read, and
underlying file implementation return -EIO, the kernel sends SIGBUS to
client thread, and if that SIGBUS is not handled the whole client process
is terminated with coredump.

In bigfile/virtmem.c until now we were doing similar thing - if in
vma_on_pagefault() a read request to loadblk() fails - we abort the
whole process.

This is however not very convenient, because if there is a multithreaded
server with each request mapped to thread, and a handling thread for
only 1 request fails this way, we kill the whole server process.

What could be convenient is to somehow propagate the error to calling
thread, e.g. unwinding the stack in a C++-style exceptions way and
turning that back to python exception at some point. And in the future
we maybe could try to do it.

For now we take a small step forward - we can terminate only the thread
which caused failed loadblk() - i.e. we still kill the code, without
providing it a way to recover, but we kill only the working thread, not
the whole process.

To test the functionality, we leverage our tfault framework which is now
extended to verify not only at which function a testcase dies, but more
generally it now examines the traceback (so that we can track coredump
to which thread terminated), and also it now verifies exit status code
and terminating signal of dying process.

NOTE on Linux it is not easy to terminate only 1 thread and produce a
coredump for debugging and have right process exit status if e.g. main
thread is terminated this way.

The reason is Linux hardcodes termination-with-coredump to kill all
threads from a process, and even separate processes which happen to
share virtual memory layout with the killing thread.

So to do such termination, we use hacks and workarounds - see comments
in newly introduced __abort_thread().

NOTE2 for getting separate coredump files for several faulting threads

    /proc/sys/kernel/core_pattern   or
    /proc/sys/kernel/core_uses_pid

are your friends.

/cc @Tyagov
/cc @klaus

Because numpy.ndarray does not accept it as buffer= argument

    https://github.com/numpy/numpy/issues/5935

and our memcpy crashes.

NOTE if we'll need to use memoryview, we can adapt our memcpy to use
array() directly which works with memoryview, as outlined in the above
numpy issue.

BigArrays can be big - up to 2^64 bytes, and thus in general it is not
possible to represent whole BigArray as ndarray view, because address
space is usually smaller on 64bit architectures.

However users often try to pass BigArrays to numpy functions as-is, and
numpy finds a way to convert, or start converting, BigArray to ndarray -
via detecting it as a sequence, and extracting elements one-by-one.
Which is slooooow.

Because of the above, we provide users a well-defined service:
- if virtual address space is available - we succeed at creating ndarray
  view for whole BigArray, without delay and copying.
- if not - we report properly the error and give hint how BigArrays have
  to be processed in chunks.

Verifying that big BigArrays cannot be converted to ndarray also tests
for behaviour and issues fixed in last 5 patches.

/cc @Tyagov
/cc @klaus

e.g.

    In [1]: multiply.reduce((1<<30, 1<<30, 1<<30))
    Out[1]: 0

instead of

    In [2]: (1<<30) * (1<<30) * (1<<30)
    Out[2]: 1237940039285380274899124224

    In [3]: 1<<90
    Out[3]: 1237940039285380274899124224

also multiply.reduce returns int64, instead of python int:

    In [4]: type( multiply.reduce([1,2,3]) )
    Out[4]: numpy.int64

which also leads to overflow-related problems if we further compute with
this value and other integers and results exceeds int64 - it becomes
float:

    In [5]: idx0_stop = 18446744073709551615

    In [6]: stride0   = numpy.int64(1)

    In [7]: byte0_stop = idx0_stop * stride0

    In [8]: byte0_stop
    Out[8]: 1.8446744073709552e+19

and then it becomes a real problem for BigArray.__getitem__()

    wendelin.core/bigarray/__init__.py:326: RuntimeWarning: overflow encountered in long_scalars
      page0_min  = min(byte0_start, byte0_stop+byte0_stride) // pagesize # TODO -> fileh.pagesize

and then

    >           vma0 = self._fileh.mmap(page0_min, page0_max-page0_min+1)
    E           TypeError: integer argument expected, got float

~~~~

So just avoid multiple.reduce() and do our own mul() properly the same
way sum() is builtin into python, and we avoid overflow-related
problems.

We need this commit:

    http://git.ozlabs.org/?p=ccan;a=commitdiff;h=c38e11b508e52fb2921e67d1123b05d9bef90fd2

or else we segfault on really big arrays allocation instead of getting
ENOMEM and reporting it as MemoryError to python.

Consider e.g. this for pyvma:

    1. in pyfileh_mmap() pyvma is created

    2. next fileh_mmap(pyvma, pyfileh, ...) fails

    3. we need to deallocate pyvma which was not mapped

    4. in pyvma_dealloc() we unmap pyvma unconditionally -> boom.

The same story goes for pyfileh dealloc vs not fully constructing it in
pyfileh_open().

OverflowError when computing slice indices practically means we'll
cannot allocate so much address space at next step:

    In [1]: s = slice(None)

    In [2]: s.indices(1<<62)
    Out[2]: (0, 4611686018427387904, 1)

    In [3]: s.indices(1<<63)
    ---------------------------------------------------------------------------
    OverflowError                             Traceback (most recent call last)
    <ipython-input-4-5aa549641bc6> in <module>()
    ----> 1 s.indices(1<<63)

    OverflowError: cannot fit 'long' into an index-sized integer

So translate this OverflowError into MemoryError (preserving message
details), because we'll need such "no so much address space" cases to
show up as MemoryError in a sooner patch.

Currently we always raise RuntimeError for problems, which is
more-or-less ok for humans, but soon we'll need to distinguish "no
memory" errors from other error conditions in upper layers in code.

Introduce helper function for choosing appropriate exception type for an
error - MemoryError for when errno=ENOMEM and RuntimeError otherwise,
and use it where appropriate.

( Unfortunately Python does not provide such a helper... )

As discussion in https://bitbucket.org/pypa/setuptools/issue/313
unveiled, setuptools do not have to carry setuptools.file_finders
entrypoint with it in order to support _other_ projects to use
.file_finders entry point.

In our case it means that is is normal that we have to make sure that
the group we are going to register entry-point into, exists.

Remove erroneous comments introduced in 11d130d1 (setup: Ensure
setuptools.file_finders entry-point group is registered).

py2-only syntax introduced in acf7e91d (setup/runcmd: Properly report
errors, if running command is missing):

      File "setup.py", line 168
        except Exception, e:
                        ^
    SyntaxError: invalid syntax

The reason is some people are blocked to access git:// protocol/port
because of firewall setup in their environment, and git.ozlabs.org does
not provide a way to clone over http(s).

We can provide CCAN mirror and we support cloning it over http(s). Also,
it makes sense to not depend on thirdparty sites for hosting
strongly-related components.

We currently use subprocess.check_output() for running external command
(git), and check_output() checks running command status code, as its
name promises:

    https://github.com/python/cpython/blob/2.7/Lib/subprocess.py#L569

and reports it appropriately to user:

    https://github.com/python/cpython/blob/2.7/Lib/subprocess.py#L411

but if the command is not found, it is just single

    error: [Errno 2] No such file or directory

which gets output to the log and it is not clear what command failed, or
evern if it is from runcmd() call.

Redo our runcmd() manually, so that it always report erros with context
- what command it was trying to run. The error now becomes:

   RuntimeError: (['missing-command'],): [Errno 2] No such file or directory

/cc @Tyagov

If setuptools.file_finders group is not registered,
dist.get_entry_map('setuptools.file_finders') returns just {} not
connected to entry map, and further modifications of this dict go
nowhere (and thus, our git_lsfiles() is not hooked -> sdist fails to
produce correct source archive).

This is a workaround for setuptools 9.0 dropping
`setuptools.file_finders` entrypoint group registration:

    https://bitbucket.org/pypa/setuptools/commits/f191c8a1225bd58a5fb5aa9abb31b06dc710f0b9#Lsetup.pyF175
    https://bitbucket.org/pypa/setuptools/issue/313

issue reported back:

    https://bitbucket.org/pypa/setuptools/issue/313#comment-18430008

It was hanging with NumPy-1.9 before 425dc5d1 (bigarray: Raise
IndexError for out-of-bound element access), because of the following
correct NumPy commit:

    https://github.com/numpy/numpy/commit/d36f8227

and in particular

    https://github.com/numpy/numpy/commit/d36f8227#diff-6d326badc0872de91e025cbfb0be1aafR522

That PySequence_Fast(obj)    (with obj being BigArray)

creates iterator on top of obj and before our previous IndexError fix in
425dc5d1, this was looping forever.

Test explicitly with both NumPy 1.8 and NumPy 1.9, that this construct
does not hang.

/cc @Tyagov

The way BigArray.__getitem__ works for element access is that for e.g.

    A[i]

it translates the request to

    A[i:i+1]

and remembers to lower the dimensionality at scalar index

    dim_adjust = (0,)

so, in full, A[i] is computed this way:

    A[i] -> A[i:i+1](0,)

( it is done this way to unify code for scalar / slice access in
  __getitem__ - see 0c826d5c "BigArray: An ndarray-like on top of
  BigFile memory mappings" )

The code for slice access also has a shortcut - if it sees that slice
results in empty array (e.g. for out-of-bound slice), we can avoid
spending time to create a file vma mapping only to create empty view on
top of it.

In 0c826d5c, that optimization, however forgot to apply the "lower the
dimensionality" step on top of resulting empty view, and that turned out
for not raising IndexError for out-of-bounds scalar access:

    A = BigArray((10,), uint8)
    In [1]: A[0]
    Out[1]: 0

    In [2]: A[1]
    Out[2]: 0

    In [3]: A[2]
    Out[3]: 0

    In [4]: A[9]
    Out[4]: 0

    In [5]: A[10]
    Out[5]: array([], dtype=uint8)

NOTE that A[10] returns empty array instead of raising IndexError.

So do not forget to apply the "reduce dimensionality" step for empty
views, and this way we get proper IndexError (because for empty view,
scalar access results in IndexError).

NOTE:

this bug was also preventing for e.g.

    list(A)

to work, because list(A) internally works this way:

    l = []
    i = iter(A)
    for _ in i:
        l.append(_)

but iterating would not stop after 10 elements - after array end, _ will
be always array([], dtype=uint8), and thus the loop never finished and
memory usage grow to infinity.

/cc @Tyagov

In NumPy speak advanced indexing is picking up arbitrarily requested
elemtnts, e.g.

    a = arange(10)
    a[[0,3,2]]  -> array([0, 3, 2])

The way this indexing schem works is - it creates a new array with
len = len(key), and picks up requested elements sequentially into new
area.

So it is very not the same as creating _view_ to original array data by
using basic indexing [1]

BigArray does not support advanced indexing, because its main job is to
organize an ndarray _view_ backed up by BigFile data and give that view
to clients, and then it is up to clients how to use that view with full
numpy api available with it.

So be explicit, and reject advanced indexing in __getitem__ right at the
beginning.

[1] http://docs.scipy.org/doc/numpy/reference/arrays.indexing.html

Else, in a clean repo the tests do not build from scratch:

    $ make test
    x86_64-linux-gnu-gcc -pthread -g -Wall -D_GNU_SOURCE -std=gnu99 -fplan9-extensions -Wno-declaration-after-statement -Wno-error=declaration-after-statement  -Iinclude -I3rdparty/ccan -I3rdparty/include   bigfile/tests/test_virtmem.c lib/bug.c lib/utils.c 3rdparty/ccan/ccan/tap/tap.c  -o bigfile/tests/test_virtmem.t
    In file included from include/wendelin/list.h:11:0,
                     from include/wendelin/bigfile/virtmem.h:33,
                     from bigfile/tests/../virtmem.c:23,
                     from bigfile/tests/test_virtmem.c:21:
    3rdparty/ccan/ccan/array_size/array_size.h:4:20: fatal error: config.h: No such file or directory
     #include "config.h"
                        ^
    compilation terminated.
    3rdparty/ccan/ccan/tap/tap.c:26:20: fatal error: config.h: No such file or directory
     #include "config.h"
                        ^
    compilation terminated.
    Makefile:99: recipe for target 'bigfile/tests/test_virtmem.t' failed
    make: *** [bigfile/tests/test_virtmem.t] Error 1

In NumPy, ndarray has .resize() but actually it does a whole array
copy into newly allocated larger segment which makes e.g. appending O(n).

For BigArray, we don't have that internal constraint NumPy has - to
keep the array itself contiguously _stored_ (compare to contiguously
_presented_ in memory). So we can have O(1) resize for big arrays.

NOTE having O(1) resize, here is how O(δ) append can be done:

    A                               # ZBigArray e.g. of shape   (N, 3)
    n = len(A)                      # lengh of A's major index  =N
    A.resize((n+δ, A.shape[1:]))    # add δ new entries ; now len(A) =N+δ
    A[-δ:] = <new-data>             # set data for last new δ entries

/cc @klaus

test_bigarray_indexing_Nd() contains useful class to have a BigFile
connected to ndarray storage. Factor it out so that all tests could use
it.

BigFile_Data.storeblk() is newly introduced and is currently unused, but
will be convenient to have later.

In 5755a6b3 (Basic setup.py / Makefile to build/install/sdist stuff +
bigfile.so skeleton) we overlooked one thing:

    when tailing from setup.py to make level, we forgot to setup the
    PYTHON env variable properly, and that leads to changing building
    with std python.

Look e.g.

    $ python3 setup.py build_ext -i
    running build_ext
    python setup.py ll_build_ext --inplace
    ^^^

which is obviously not correct.

Fix it.

The tox setup we have setup for testing with multiple interpreters
(7af2b2d7 "Tox setup to test things with py27 & py34 and several
versions of ZODB") did not caught this, because for every python and
environment versions tox initializes separate virtualenv, in which
.../bin/python is a link to chosen python, and this way both python and
my-selected-for-build-python are the same.

Cc: @Tyagov

For example on Debian 7 Wheezy the kernel (linux 3.2) supports holepunching,
because holepunching was added in linux 2.6.38:

    http://git.kernel.org/linus/79124f18b335172e1916075c633745e12dae1dac

but glibc does not provide fallocate mode constants, and compilation fails:

    gcc -pthread -fno-strict-aliasing -g -O2 -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-prototypes -fPIC -D_GNU_SOURCE -I./include -I./3rdparty/ccan -I./3rdparty/include -I/opt/slapgrid/f2a5f59d0d2b521681b9333ee76a2859/parts/python2.7/include/python2.7 -c bigfile/ram_shmfs.c -o build/temp.linux-x86_64-2.7/bigfile/ram_shmfs.o -std=gnu99 -fplan9-extensions -fvisibility=hidden -Wno-declaration-after-statement -Wno-error=declaration-after-statement
    bigfile/ram_shmfs.c: In function ‘shmfs_drop_memory’:
    bigfile/ram_shmfs.c:135:31: error: ‘FALLOC_FL_PUNCH_HOLE’ undeclared (first use in this function)
    bigfile/ram_shmfs.c:135:31: note: each undeclared identifier is reported only once for each function it appears in
    bigfile/ram_shmfs.c:135:54: error: ‘FALLOC_FL_KEEP_SIZE’ undeclared (first use in this function)
    error: command 'gcc' failed with exit status 1

Try to fix it via fallbacking to provide the needed defines on older systems ourselves.

Cc: Klaus Wölfel <klaus@nexedi.com>
Cc: Ivan Tyagov <ivan@tyagov.com>

In C99 declaration after statement is ok, and we explicitly compile with -std=gnu99.

Python >= 3.4 however adds -Werror=declaration-after-statement even for extension
modules irregardless of their compilation flags:

  https://bugs.python.org/issue21121

and the build fails this way:

    building 'wendelin.bigfile._bigfile' extension
    creating build/temp.linux-x86_64-3.4
    creating build/temp.linux-x86_64-3.4/bigfile
    creating build/temp.linux-x86_64-3.4/lib
    gcc -pthread -Wno-unused-result -Werror=declaration-after-statement -DNDEBUG -fmessage-length=0 -grecord-gcc-switches -O2 -Wall
-D_FORTIFY_SOURCE=2 -fstack-protector -funwind-tables -fasynchronous-unwind-tables -g -DOPENSSL_LOAD_CONF -fPIC -D_GNU_SOURCE
-I./include -I./3rdparty/ccan -I./3rdparty/include -I/usr/include/python3.4m -c bigfile/_bigfile.c -o build/temp.linux-x86_64-3.4/bigfile/_bigfile.o
-std=gnu99 -fplan9-extensions -fvisibility=hidden
    bigfile/_bigfile.c: In function ‘pyfile_new’:
    bigfile/_bigfile.c:679:5: error: ISO C90 forbids mixed declarations and code [-Werror=declaration-after-statement]
         static char *kw_list[] = {"blksize", NULL};
         ^
    cc1: some warnings being treated as errors
    error: command 'gcc' failed with exit status 1

Ensure we turn off this warning and error because we already rely on compiling in C99 mode.
Reported-and-tested-by: Ivan Tyagov <ivan@tyagov.com>

This shows how to first generate such arrays (in steps, as every
transaction change should fit in memory), and then gather data from
whole array using C/Fortran/etc code.

It shows how to compute mean via NumPy's ndarray.mean()

It also shows that e.g. ndarray.var() wants to create temporaries in
size of original ndarray and that would fail, because it does not fit
into RAM.

ndarray.var() should not need to create such temporaries in principle -
all it has to do is to first compute mean, and then compute

    sum (Xi - <X>)^2

in a loop.

<X> is scalar, Xi - is just access to original array.

~~~~

So this also show NumPy can be incrementally improved to avoid creating
such temporaries, and then it will work.

    - for virtual memory subsytem
    - for ZBigFiles

They are not currently great, e.g. for virtmem we have in-kernel
overhead of page clearing - in perf profiles, for bigfile_mmap compared
to file_read kernel's clear_page_c raises significantly.

That is the worker for clearing page memory and we currently cannot
avoid that - any memory obtained from kernel (MAP_ANONYMOUS, mmap(file)
with hole, etc...) comes pre-initialized to zeros to userspace.

This can be seen in the benchmarks as well: file_readbig differs from
file_read in only that the latter uses 1 small buffer and the first
allocates large memory (cleared by kernel + python does the memset).

    bigfile/tests/bench_virtmem.py@125::bench_file_mmap_adler32     0.47  (0.86 0.49 0.47)
    bigfile/tests/bench_virtmem.py@126::bench_file_read_adler32     0.69  (1.11 0.71 0.69)
    bigfile/tests/bench_virtmem.py@127::bench_file_readbig_adler32  1.41  (1.70 1.42 1.41)
    bigfile/tests/bench_virtmem.py@128::bench_bigfile_mmap_adler32  1.42  (1.45 1.42 1.51)

    bigfile/tests/bench_virtmem.py@130::bench_file_mmap_md5         1.52  (1.91 1.54 1.52)
    bigfile/tests/bench_virtmem.py@131::bench_file_read_md5         1.73  (2.10 1.75 1.73)
    bigfile/tests/bench_virtmem.py@132::bench_file_readbig_md5      2.44  (2.73 2.46 2.44)
    bigfile/tests/bench_virtmem.py@133::bench_bigfile_mmap_md5      2.40  (2.48 2.40 2.53)

There is MAP_UNINITIALIZED which works only for non-mmu targets and only
if explicitly allowed when configuring kernel (off by default).

There were patches to disable that pages zeroing, as it gives
significant speedup for people's workloads, e.g. [1,2] but all of them
did not got merged for security reasons.

[1] http://marc.info/?t=132691315900001&r=1&w=2
[2] http://thread.gmane.org/gmane.linux.kernel/548926

~~~~

For ZBigFile - it is the storage who is dominating in profiles.

For this, a small wrapper over py.test is developed (to discover/collect
functions to benchmar, etc) and then it runs such functions several
times in a boxed enveronment.

Benchmarks should be named bench_*.py

This is like to BigArray, like ZBigFile is to BigFile (4174b84a
"bigfile: BigFile backend to store data in ZODB")

I.e. something like numpy.memmap for numpy.ndarray and OS files. The whole
bigarray cannot be used as a drop-in replacement for numpy arrays, but BigArray
_slices_ are real ndarrays and can be used everywhere ndarray can be used,
including in C/Fortran code. Slice size is limited by mapping-size (=
address-space size) limit, i.e. to ~ max 127TB on Linux/amd64.

Changes to bigarray memory are changes to bigfile memory mapping and as such
can be discarded or saved back to bigfile using mapping (= BigFileH) dirty
discard/writeout interface.

For the same reason the whole amount of changes to memory is limited by amount
of physical RAM.

This adds transactionality and with e.g. NEO[1] allows to distribute
objects to nodes into cluster.

We hook into ZODB two-phase commit process as a separate data manager,
and synchronize changes to memory, to changes to object only at that
time.

Alternative would be to get notified on every page change, and mark
appropriate object as dirty right at that moment.

But I wanted to stay close to filesystem design (we don't get
notification for every file change from kernel) - that's why it is done
the first way.

[1] http://www.neoppod.org/

In Python. To demonstrate how backends could be implemented and test
system on simple things.

Exposes BigFile - this way users can define BigFile backend in Python.

Also exposed are BigFile handles, and VMA objects which are results of
mmaping.