Bigarrays as memory buffers
We consider 1-dimensional bigarrays of chars as memory buffers. They have the useful property that the garbage collector cannot relocate them, i.e. the address is fixed. Also, one can mmap a file, and connect the bigarray with shared memory.
blit_memory_to_string src srcoff dst dstoff len copies len characters
from buffer src, starting at character number srcoff, to
string dst, starting at character number dstoff
Raise Invalid_argument if srcoff and len do not
designate a valid subbuffer of src, or if dstoff and len
do not designate a valid substring of dst.
Unsafe version
blit_string_to_memory src srcoff dst dstoff len copies len characters
from string src, starting at character number srcoff, to
buffer dst, starting at character number dstoff
Raise Invalid_argument if srcoff and len do not
designate a valid substring of src, or if dstoff and len
do not designate a valid subbuffer of dst.
Unsafe version
Returns the size of a page as reported by sysconf.
On many systems, a page has 4096 bytes, but this cannot be relied upon.
This function is only available if the system has sysconf.
The best guess at the page size
Allocates memory in units of pages. The memory buffer will start on a page boundary.
The passed int is the requested number of bytes. The size of the buffer is rounded up so a whole number of pages is allocated.
Optionally, one can request a certain address addr (which must
be a multiple of the page size). There is, however, no guarantee
that this wish can be fulfilled. In any way, one should check with
memory_address what the start address really is.
If exec, the memory region is marked as executable.
This function is only available if the system has sysconf, mmap,
and allows to allocate anonymous memory with mmap (outside POSIX
but common).
alloc_aligned_memory alignment size: Allocates a buffer of size
whose start address is a multiple of alignment. The alignment
must be a power of two, and at least Sys.word_size/8.
Aligned memory can be useful for ensuring that the whole memory
block is in the same cache line. A cache line typically has
64 or 128 bytes - but this is very platform-specific. (Linux:
look at /proc/cpuinfo.)
This function is only available if the system has posix_memalign.
memory_map_file fd shared size: Maps size bytes of the file
fd into memory, and returns the memory buffer like
Bigarray.Array1.map_file. pos and shared have the same
meaning as there. In addr one can suggest a start address.
There is, however, no guarantee that this wish can be fulfilled.
Unmaps the file. The memory block must have been allocated
with memory_map_file or with Bigarray.Array1.map_file.
Note that the data pointer of the bigarray is set to NULL, and that any further access of the array will trigger a segmentation violation! The intention of this function is to control when the file mapping is removed. Normally, this is done first when the GC finalizer is run.
It is required that there are no subarrays at the time of calling this function. (If so, the function does nothing.)
zero_pages m pos len: If possible, the memory pages in the
range pos to pos+len-1 of m are allocated again, so that
they replace the previous pages.
It is required that the start address of the range is a
multiple of the page size, and the len is a multiple of
the page size. Fails with Invalid_argument if the requirements
are not met, or the function is otherwise unavailable.
Calling zero_pages is sometimes an optimization when old
memory pages can be dropped, and when the alternative of
overwriting these pages would imply a copy-on-write operation.
grab addr len: Interprets the address range from addr to
addr+len-1 as memory bigarray.
This function does not allocate! It assumes that the given address range points to valid memory.
as_value mem offset: Returns a pointer to mem+offset. There
must be a valid boxed value at this address (i.e. at
the word preceding mem+offset there must be a valid block
header, followed by a valid value of the right type). However,
this is not checked:
This is an unsafe function that may crash the program if used in the wrong way!
It is possible that the memory block is deallocated while the returned value still exists. Any attempt to access the value will result into undefined behavior (anything from funny results to crashes may happen).
Some Ocaml primitives might not work on the returned values (polymorphic equality, marshalling, hashing) unless Netsys_mem.value_area is called for the memory block.
Marks the memory block as value area. This enables that the value primitives (polymorphic equality, marshalling, hashing) return meaningful results. The memory area is automatically unmarked when the finaliser for the memory block is run.
Be careful when marking sub arrays.
This function is first available since O'Caml 3.11.
Compares two strings like String.compare. This also works
when the strings reside outside the O'Caml heap, e.g. in a
memory block.
init_header mem offset tag size: Initializes the word at
mem+offset as an Ocaml value header with the given tag
and the given size (in words). The GC color is always set
to "white".
let voffset, bytelen = init_string mem offset len:
Initializes the memory at offset
and following bytes as Ocaml string with length len.
Returns in voffset the offset where the value starts
(i.e. offset plus one word), and in bytelen the number
of bytes used in mem.
offset must be a multiple of the word size in bytes.
The string can be accessed with
let s = (as_value mem voffset : string)The function is useful for initializing shared memory as string so that several processes can directly access the string.
The string has the GC color White.
Raises Out_of_space if the memory block is too small.
Returns bytelen if init_string was called with the passed
len.
let voffset, bytelen = init_array mem offset size:
Initializes the memory at offset
and following bytes as Ocaml array with size elements.
Returns in voffset the offset where the value starts
(i.e. offset plus one word), and in bytelen the number
of bytes used in mem.
The array cannot be used as float array.
offset must be a multiple of the word size in bytes.
The array can be accessed with
let a = (as_value mem voffset : _ array)The elements of the array have a value but it might not be valid for the element type of the array. Because of this, it is unwise to access the elements before setting them for the first time.
The array has the GC color White.
Raises Out_of_space if the memory block is too small.
Returns bytelen if init_array was called with the passed
size.
Same for arrays of floats
let voffset, bytelen = init_value mem offset v flags:
Initializes the memory at offset and following bytes as
copy of the boxed value v.
Returns in voffset the offset where the value starts
(i.e. offset plus one word), and in bytelen the number
of bytes used in mem.
The copied value can then be accessed with
let v' = (as_value mem voffset : 'a)offset must be a multiple of the word size in bytes.
The input value v must be heap-allocated. Also, a number of
restrictions and caveats apply:
Copy_bigarray flag
is given. In this case, a copy of the bigarray is also made
and appended to the value copy (i.e. it is also placed into
the buffer mem).int32, int64, and nativeint the flag Copy_custom_int
enables copying, and for bigarrays the flag Copy_bigarray.
Generally, there is a function pointer in such data blocks which
might be invalid when the memory buffer is loaded into a
different executable. This specific problem can be fixed
by passing target_custom_ops with the right pointers.Copy_atom or Keep_atom flags are present,
otherwise the function fails. Keep_atom means here to keep
atoms as-is. This is correct, but also keeps references to
the atom definitions which live outside mem. Copy_atom means to
create a copy of the atom as a zero-sized block outside the
atom table. This way the value in mem is self-contained,
but this unfortunately breaks some assumptions of the
OCaml code generator. In particular, comparisons like
if array=[| |] then... may yield wrong results.The function raises Out_of_space if the memory block is too small.
Cyclic input values are supported, and value sharing is kept intact.
If the Copy_simulate flag is given, mem is not modified.
In simulation mode, it is pretended that mem is as large
as necessary to hold the value, no matter how large mem really
is. The returned values voffset and bytelen reflect how much
of mem would have been used.
If the targetaddr argument is passed, it is assumed that the
memory block is mapped at this address and not at the address it
is really mapped. This is useful for preparing memory that is going
to be mapped at a different address than it is right now.
The new value has the GC color White.
If bigarrays are copied, the copy also includes the data part. The data part is directly following the bigarray block, and is represented in a special implementation-defined way.
If the Copy_conditionally flag is set, the condition cc is
evaluated for every block, and only if cc returns true, the block
is copied. cc is a list of addresses (start,end), and a block
is not copied if its address lies in any of these address ranges.
Otherwise the block is copied. As an exception of the foregoing,
the first block (i.e. v) is always copied.
Returns custom ops for a sample value (or Invalid_argument)
copy_value flags v: Creates a deep copy of v and returns it.
The copy is allocated in the normal Ocaml heap.
Restrictions:
Copy_bigarray flag
is given. In this case, a copy of bigarrays are also made,
and placed into additional buffers obtained via stat_alloc.int32, int64, and nativeint the flag Copy_custom_int
enables copying, and for bigarrays the flag Copy_bigarray.Copy_atoms is ignored.Cyclic input values are supported. Copy_simulate is ignored.
memory as buffersA version of Unix.read that uses a memory buffer.
Some OS allow faster I/O when memory is page-aligned
(see alloc_memory_pages). Also, a copy in the stub function
can be avoided. Both effects can result in a considerable speedup.
A version of Unix.single_write that uses a memory buffer.
Versions of Unix.recv, and Unix.send
using memory buffers.
A pool of memory blocks that are all the same size and page-aligned
(if the OS supports this). The pool tries to bundle memory allocations
so that not for every block a system call is required. This reduces
the number of system calls, and the number of entries in the process
page table. Also, unused blocks are automatically returned to the
pool.
Create a new pool. The argument is the size of the memory blocks (must be a multiple of the page size)
let m = pool_alloc_memory p:
Gets a memory block m from the pool p. If required, new blocks are
automatically allocated and added to the pool. This function is
thread-safe.
The memory block is automatically garbage-collected.
let m, free = pool_alloc_memory2 p:
Gets a memory block m from the pool p like pool_alloc_memory.
This function also returns the function free marking the
block as free again. The block can then be immediately recycled
for another use.
If free is not called, the block m is first recycled when it
is not referenced any more (like in pool_alloc_memory).
The default block size, normally 64 K (or better, 16 times the page size)
The default pool with the default block size. This pool is used by Ocamlnet itself as much as possible
The block size of small_pool, normally 4K (or better, the
page size)
Another standard pool where the blocks are smaller than in
default_pool.
Returns a report describing the memory allocation in the pool