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.

We'll use py.test for unit-testing of Python part.

Does similar things to what kernel does - users can mmap file parts into
address space and access them read/write. The manager will be getting
invoked by hardware/OS kernel for cases when there is no page loaded for
read, or when a previousle read-only page is being written to.

Additionally to features provided in kernel, it support to be used to
store back changes in transactional way (see fileh_dirty_writeout()) and
potentially use huge pages for mappings (though this is currently TODO)

Users can inherit from BigFile and provide custom ->loadblk() and
->storeblk() to load/store file blocks from a database or some other
storage. The system then could use such files to memory map them into
user address space (see next patch).

Useful to debug (via either printk or gdb) the kernel - one can boot it,
and tweak compile test program on host and be ready to test it inside
qemu on tested kernel. Another use case is to add tracing printk to
kernel and to boot it on each iteration. Booting via qemu means the
workflow is not disrupted as would with rebooting the hardware.

With this tool it was straightforward to find out why mmap-alias through
mremap does not work for huge pages:

static struct vm_area_struct *vma_to_resize(unsigned long addr,
        unsigned long old_len, unsigned long new_len, unsigned long *p)
{
        struct mm_struct *mm = current->mm;
        struct vm_area_struct *vma = find_vma(mm, addr);

        if (!vma || vma->vm_start > addr)
                goto Efault;

        if (is_vm_hugetlb_page(vma))            <--
                goto Einval;

mremap_to(...)
{
        /* lots of checks, then */
        vma_to_resize(...)

This thing allows to get aliasable RAM from OS kernel and to manage it.
Currently we get memory from a tmpfs mount, and hugetlbfs should also
work, but is TODO because hugetlbfs in the kernel needs to be improved.

We need aliasing because we'll need to be able to memory map the same
page into several places in address space, e.g. for taking two slices
overlapping slice of the same array at different times.

Comes with test programs that show we aliasing does not work for
anonymous memory.

This will be the core of virtual memory subsystem. For now we just
define a structure to describe pages of memory and add utility to
allocate address space from OS.

We hook into SIGSEGV and handle read/write pagefaults this way.

In this patch there goes stub code that only detects faults and
determines (in arch specific way) whether fault was for read or write
and there is a TODO to pass that information to higher level.

It also comes with tests to detect we still crash if we access something
incorrectly, so people could have coredumps and investigate them.

For BigFiles we'll needs to maintain `{} offset-in-file -> void *` mapping. A
hash or a binary tree could be used there, but since we know files are
most of the time accessed sequentially and locally in pages-batches, we
can also organize the mapping in batches of keys.

Specifically offset bits are so divided into parts, that every part
addresses 1 entry in a table of hardware-page in size. To get to the actual
value, the system lookups first table by first part of offset, then from
first table and next part from address - second table, etc.

To clients this looks like a dictionary with get/set/del & clear methods,
but lookups are O(1) time always, and in contrast to hashes values are
stored with locality (= adjacent lookups almost always access the same tables).

We'll be testing C code in a similiar-to-python way - keep test_*.c
files nearby and compile/executing them as needed.

Tests are run several times in several ways - as plainly compiled and
also with additional error checkers:

 - AddressSanitizer
 - ThreadSanitizer
 - and similiar checkers from Valgrind

Like C bzero / memset & memcopy - but work on python buffers. We
leverage NumPy for doing actual work, and this way NumPy becomes a
depenency.

Having NumPy as a dependency is ok - we'll for sure need it later as we
are trying to build out-of-core ndarrays.

Like taking an exact integer log2, upcasting pointers for C-style
inheritance done in a Plan9 way, and wrappers to functions which should
never fail.

Modelled by ones used in Linux kernel.

It is an early project decision to use gnu99 & Plan9 C extensions, to
simplify C code.

So far we only build stub wendelin/bigfile/_bigfile.so .

Makefile is introduced because there will be targets which are easier to
handle at make level. For end users no make knowledge is required -
usual `python setup.py build|install|...` work, redirecting to make
where necessary.

As was promised (e870781d "Top-level in-tree import redirector")
setup.py contains install-time hooks to handle in-tree wendelin.py and
install it as a module namespace.

For sdist, we just use `git ls-files` info if we are in a checkout.

Prepare for building stuff and add commonly produced files to our
.gitignore .

This will be a module to allow users to program custom file-like
backends and latter memory-map content of files from the backend.

For now just start the module structure.

Unfortunately in CCAN they made separation of list_head into list_head &
list_node (for better type safety I suppose), but that leads e.g. to
inablility to insert element in the middle of the list.

We take the code from util-linux.git (v2.25-165-g9138d6f)[1] which claims
LGPL license for it, with which we are fine.

[1] https://git.kernel.org/cgit/utils/util-linux/util-linux.git/tree/include/list.h?id=v2.25-165-g9138d6f

CCAN[1] is a "Comprehensive C Archive Network" - a collection of useful
C libraries to use.

Wendelin.core will be a C project in big part, so having CCAN at hand is
useful.

We hook CCAN as git submodule, without copying it into git repo. Later
on making source tarball distribution, the code will be injected into
generated tarball, so there will be no additional dependencies to
download something from network for end users.

[1] http://ccodearchive.net/

For working copy checkout we would like to work on the source
right-away, without putting it under additional wendelin/ (and possibly
/core/) subdirectories.

Add in-tree modules redirector which tweaks plain-python and
pkg_resources-based imports so that e.g.

    import wendelin.bigarray

while in-tree resolves to

    bigarray/__init__.py

In installed state, there will be no such redirector, and code will be
installed with conventional structure.

As a first step we want to make sure the code can be used by open-source
projects without a problem.

So GPLv3+ with wide exception for all open-source licenses.

The project to implement Out-of-core NumPy arrays in backward-compatible
way.