X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/0d63ac0c2486a43198eb1015ba8d1028239139ee..0988acb472849dc08d576ee40493e70bde2132ca:/sdk/python/arvados/arvfile.py?ds=sidebyside diff --git a/sdk/python/arvados/arvfile.py b/sdk/python/arvados/arvfile.py index dd48b99d39..ce0e5e3564 100644 --- a/sdk/python/arvados/arvfile.py +++ b/sdk/python/arvados/arvfile.py @@ -2,23 +2,32 @@ import functools import os import zlib import bz2 -from .ranges import * -from arvados.retry import retry_method import config import hashlib -import hashlib import threading import Queue import copy import errno -from .errors import KeepWriteError, AssertionError +import re +import logging + +from .errors import KeepWriteError, AssertionError, ArgumentError from .keep import KeepLocator +from ._normalize_stream import normalize_stream +from ._ranges import locators_and_ranges, replace_range, Range +from .retry import retry_method + +MOD = "mod" +WRITE = "write" + +_logger = logging.getLogger('arvados.arvfile') def split(path): - """Separate the stream name and file name in a /-separated stream path and - return a tuple (stream_name, file_name). + """split(path) -> streamname, filename - If no stream name is available, assume '.'. + Separate the stream name and file name in a /-separated stream path and + return a tuple (stream_name, file_name). If no stream name is available, + assume '.'. """ try: @@ -27,7 +36,7 @@ def split(path): stream_name, file_name = '.', path return stream_name, file_name -class FileLikeObjectBase(object): +class _FileLikeObjectBase(object): def __init__(self, name, mode): self.name = name self.mode = mode @@ -56,16 +65,9 @@ class FileLikeObjectBase(object): self.closed = True -class ArvadosFileReaderBase(FileLikeObjectBase): - class _NameAttribute(str): - # The Python file API provides a plain .name attribute. - # Older SDK provided a name() method. - # This class provides both, for maximum compatibility. - def __call__(self): - return self - +class ArvadosFileReaderBase(_FileLikeObjectBase): def __init__(self, name, mode, num_retries=None): - super(ArvadosFileReaderBase, self).__init__(self._NameAttribute(name), mode) + super(ArvadosFileReaderBase, self).__init__(name, mode) self._filepos = 0L self.num_retries = num_retries self._readline_cache = (None, None) @@ -80,8 +82,8 @@ class ArvadosFileReaderBase(FileLikeObjectBase): def decompressed_name(self): return re.sub('\.(bz2|gz)$', '', self.name) - @FileLikeObjectBase._before_close - def seek(self, pos, whence=os.SEEK_CUR): + @_FileLikeObjectBase._before_close + def seek(self, pos, whence=os.SEEK_SET): if whence == os.SEEK_CUR: pos += self._filepos elif whence == os.SEEK_END: @@ -91,7 +93,7 @@ class ArvadosFileReaderBase(FileLikeObjectBase): def tell(self): return self._filepos - @FileLikeObjectBase._before_close + @_FileLikeObjectBase._before_close @retry_method def readall(self, size=2**20, num_retries=None): while True: @@ -100,7 +102,7 @@ class ArvadosFileReaderBase(FileLikeObjectBase): break yield data - @FileLikeObjectBase._before_close + @_FileLikeObjectBase._before_close @retry_method def readline(self, size=float('inf'), num_retries=None): cache_pos, cache_data = self._readline_cache @@ -124,15 +126,15 @@ class ArvadosFileReaderBase(FileLikeObjectBase): self._readline_cache = (self.tell(), data[nextline_index:]) return data[:nextline_index] - @FileLikeObjectBase._before_close + @_FileLikeObjectBase._before_close @retry_method def decompress(self, decompress, size, num_retries=None): - for segment in self.readall(size, num_retries): + for segment in self.readall(size, num_retries=num_retries): data = decompress(segment) if data: yield data - @FileLikeObjectBase._before_close + @_FileLikeObjectBase._before_close @retry_method def readall_decompressed(self, size=2**20, num_retries=None): self.seek(0) @@ -147,7 +149,7 @@ class ArvadosFileReaderBase(FileLikeObjectBase): else: return self.readall(size, num_retries=num_retries) - @FileLikeObjectBase._before_close + @_FileLikeObjectBase._before_close @retry_method def readlines(self, sizehint=float('inf'), num_retries=None): data = [] @@ -170,8 +172,15 @@ class ArvadosFileReaderBase(FileLikeObjectBase): class StreamFileReader(ArvadosFileReaderBase): + class _NameAttribute(str): + # The Python file API provides a plain .name attribute. + # Older SDK provided a name() method. + # This class provides both, for maximum compatibility. + def __call__(self): + return self + def __init__(self, stream, segments, name): - super(StreamFileReader, self).__init__(name, 'rb', num_retries=stream.num_retries) + super(StreamFileReader, self).__init__(self._NameAttribute(name), 'rb', num_retries=stream.num_retries) self._stream = stream self.segments = segments @@ -182,7 +191,7 @@ class StreamFileReader(ArvadosFileReaderBase): n = self.segments[-1] return n.range_start + n.range_size - @FileLikeObjectBase._before_close + @_FileLikeObjectBase._before_close @retry_method def read(self, size, num_retries=None): """Read up to 'size' bytes from the stream, starting at the current file position""" @@ -193,14 +202,14 @@ class StreamFileReader(ArvadosFileReaderBase): available_chunks = locators_and_ranges(self.segments, self._filepos, size) if available_chunks: lr = available_chunks[0] - data = self._stream._readfrom(lr.locator+lr.segment_offset, + data = self._stream.readfrom(lr.locator+lr.segment_offset, lr.segment_size, num_retries=num_retries) self._filepos += len(data) return data - @FileLikeObjectBase._before_close + @_FileLikeObjectBase._before_close @retry_method def readfrom(self, start, size, num_retries=None): """Read up to 'size' bytes from the stream, starting at 'start'""" @@ -209,12 +218,11 @@ class StreamFileReader(ArvadosFileReaderBase): data = [] for lr in locators_and_ranges(self.segments, start, size): - data.append(self._stream._readfrom(lr.locator+lr.segment_offset, lr.segment_size, + data.append(self._stream.readfrom(lr.locator+lr.segment_offset, lr.segment_size, num_retries=num_retries)) return ''.join(data) def as_manifest(self): - from stream import normalize_stream segs = [] for r in self.segments: segs.extend(self._stream.locators_and_ranges(r.locator, r.range_size)) @@ -228,9 +236,15 @@ def synchronized(orig_func): return orig_func(self, *args, **kwargs) return synchronized_wrapper -class BufferBlock(object): - """A BufferBlock is a stand-in for a Keep block that is in the process of being - written. + +class StateChangeError(Exception): + def __init__(self, message, state, nextstate): + super(StateChangeError, self).__init__(message) + self.state = state + self.nextstate = nextstate + +class _BufferBlock(object): + """A stand-in for a Keep block that is in the process of being written. Writers can append to it, get the size, and compute the Keep locator. There are three valid states: @@ -252,6 +266,7 @@ class BufferBlock(object): WRITABLE = 0 PENDING = 1 COMMITTED = 2 + ERROR = 3 def __init__(self, blockid, starting_capacity, owner): """ @@ -269,10 +284,12 @@ class BufferBlock(object): self.buffer_block = bytearray(starting_capacity) self.buffer_view = memoryview(self.buffer_block) self.write_pointer = 0 - self._state = BufferBlock.WRITABLE + self._state = _BufferBlock.WRITABLE self._locator = None self.owner = owner self.lock = threading.Lock() + self.wait_for_commit = threading.Event() + self.error = None @synchronized def append(self, data): @@ -282,7 +299,7 @@ class BufferBlock(object): buffer, doubling capacity as needed to accomdate all the data. """ - if self._state == BufferBlock.WRITABLE: + if self._state == _BufferBlock.WRITABLE: while (self.write_pointer+len(data)) > len(self.buffer_block): new_buffer_block = bytearray(len(self.buffer_block) * 2) new_buffer_block[0:self.write_pointer] = self.buffer_block[0:self.write_pointer] @@ -294,12 +311,30 @@ class BufferBlock(object): else: raise AssertionError("Buffer block is not writable") - def set_state(self, nextstate): - if ((self._state == BufferBlock.WRITABLE and nextstate == BufferBlock.PENDING) or - (self._state == BufferBlock.PENDING and nextstate == BufferBlock.COMMITTED)): - self._state = nextstate - else: - raise AssertionError("Invalid state change from %s to %s" % (self.state, state)) + STATE_TRANSITIONS = frozenset([ + (WRITABLE, PENDING), + (PENDING, COMMITTED), + (PENDING, ERROR), + (ERROR, PENDING)]) + + @synchronized + def set_state(self, nextstate, val=None): + if (self._state, nextstate) not in self.STATE_TRANSITIONS: + raise StateChangeError("Invalid state change from %s to %s" % (self.state, nextstate), self.state, nextstate) + self._state = nextstate + + if self._state == _BufferBlock.PENDING: + self.wait_for_commit.clear() + + if self._state == _BufferBlock.COMMITTED: + self._locator = val + self.buffer_view = None + self.buffer_block = None + self.wait_for_commit.set() + + if self._state == _BufferBlock.ERROR: + self.error = val + self.wait_for_commit.set() @synchronized def state(self): @@ -316,6 +351,20 @@ class BufferBlock(object): self._locator = "%s+%i" % (hashlib.md5(self.buffer_view[0:self.write_pointer]).hexdigest(), self.size()) return self._locator + @synchronized + def clone(self, new_blockid, owner): + if self._state == _BufferBlock.COMMITTED: + raise AssertionError("Cannot duplicate committed buffer block") + bufferblock = _BufferBlock(new_blockid, self.size(), owner) + bufferblock.append(self.buffer_view[0:self.size()]) + return bufferblock + + @synchronized + def clear(self): + self.owner = None + self.buffer_block = None + self.buffer_view = None + class NoopLock(object): def __enter__(self): @@ -330,37 +379,39 @@ class NoopLock(object): def release(self): pass -SYNC_READONLY = 1 -SYNC_EXPLICIT = 2 -SYNC_LIVE = 3 def must_be_writable(orig_func): @functools.wraps(orig_func) def must_be_writable_wrapper(self, *args, **kwargs): - if self.sync_mode() == SYNC_READONLY: - raise IOError((errno.EROFS, "Collection is read only")) + if not self.writable(): + raise IOError(errno.EROFS, "Collection is read-only.") return orig_func(self, *args, **kwargs) return must_be_writable_wrapper -class BlockManager(object): - """BlockManager handles buffer blocks, background block uploads, and background - block prefetch for a Collection of ArvadosFiles. +class _BlockManager(object): + """BlockManager handles buffer blocks. + + Also handles background block uploads, and background block prefetch for a + Collection of ArvadosFiles. """ + + DEFAULT_PUT_THREADS = 2 + DEFAULT_GET_THREADS = 2 + def __init__(self, keep): """keep: KeepClient object to use""" self._keep = keep self._bufferblocks = {} self._put_queue = None - self._put_errors = None self._put_threads = None self._prefetch_queue = None self._prefetch_threads = None self.lock = threading.Lock() self.prefetch_enabled = True - self.num_put_threads = 2 - self.num_get_threads = 2 + self.num_put_threads = _BlockManager.DEFAULT_PUT_THREADS + self.num_get_threads = _BlockManager.DEFAULT_GET_THREADS @synchronized def alloc_bufferblock(self, blockid=None, starting_capacity=2**14, owner=None): @@ -378,14 +429,13 @@ class BlockManager(object): """ if blockid is None: blockid = "bufferblock%i" % len(self._bufferblocks) - bufferblock = BufferBlock(blockid, starting_capacity=starting_capacity, owner=owner) + bufferblock = _BufferBlock(blockid, starting_capacity=starting_capacity, owner=owner) self._bufferblocks[bufferblock.blockid] = bufferblock return bufferblock @synchronized def dup_block(self, block, owner): - """Create a new bufferblock in WRITABLE state, initialized with the content of - an existing bufferblock. + """Create a new bufferblock initialized with the content of an existing bufferblock. :block: the buffer block to copy. @@ -395,12 +445,7 @@ class BlockManager(object): """ new_blockid = "bufferblock%i" % len(self._bufferblocks) - with block.lock: - if block._state == BufferBlock.COMMITTED: - raise AssertionError("Can only duplicate a writable or pending buffer block") - - bufferblock = BufferBlock(new_blockid, block.size(), owner) - bufferblock.append(block.buffer_view[0:block.size()]) + bufferblock = block.clone(new_blockid, owner) self._bufferblocks[bufferblock.blockid] = bufferblock return bufferblock @@ -408,6 +453,70 @@ class BlockManager(object): def is_bufferblock(self, locator): return locator in self._bufferblocks + def _commit_bufferblock_worker(self): + """Background uploader thread.""" + + while True: + try: + bufferblock = self._put_queue.get() + if bufferblock is None: + return + + loc = self._keep.put(bufferblock.buffer_view[0:bufferblock.write_pointer].tobytes()) + bufferblock.set_state(_BufferBlock.COMMITTED, loc) + + except Exception as e: + bufferblock.set_state(_BufferBlock.ERROR, e) + finally: + if self._put_queue is not None: + self._put_queue.task_done() + + @synchronized + def start_put_threads(self): + if self._put_threads is None: + # Start uploader threads. + + # If we don't limit the Queue size, the upload queue can quickly + # grow to take up gigabytes of RAM if the writing process is + # generating data more quickly than it can be send to the Keep + # servers. + # + # With two upload threads and a queue size of 2, this means up to 4 + # blocks pending. If they are full 64 MiB blocks, that means up to + # 256 MiB of internal buffering, which is the same size as the + # default download block cache in KeepClient. + self._put_queue = Queue.Queue(maxsize=2) + + self._put_threads = [] + for i in xrange(0, self.num_put_threads): + thread = threading.Thread(target=self._commit_bufferblock_worker) + self._put_threads.append(thread) + thread.daemon = False + thread.start() + + def _block_prefetch_worker(self): + """The background downloader thread.""" + while True: + try: + b = self._prefetch_queue.get() + if b is None: + return + self._keep.get(b) + except Exception: + pass + + @synchronized + def start_get_threads(self): + if self._prefetch_threads is None: + self._prefetch_queue = Queue.Queue() + self._prefetch_threads = [] + for i in xrange(0, self.num_get_threads): + thread = threading.Thread(target=self._block_prefetch_worker) + self._prefetch_threads.append(thread) + thread.daemon = True + thread.start() + + @synchronized def stop_threads(self): """Shut down and wait for background upload and download threads to finish.""" @@ -419,7 +528,6 @@ class BlockManager(object): t.join() self._put_threads = None self._put_queue = None - self._put_errors = None if self._prefetch_threads is not None: for t in self._prefetch_threads: @@ -429,68 +537,60 @@ class BlockManager(object): self._prefetch_threads = None self._prefetch_queue = None - def commit_bufferblock(self, block): + def __enter__(self): + return self + + def __exit__(self, exc_type, exc_value, traceback): + self.stop_threads() + + def __del__(self): + self.stop_threads() + + def commit_bufferblock(self, block, sync): """Initiate a background upload of a bufferblock. - This will block if the upload queue is at capacity, otherwise it will - return immediately. + :block: + The block object to upload + + :sync: + If `sync` is True, upload the block synchronously. + If `sync` is False, upload the block asynchronously. This will + return immediately unless the upload queue is at capacity, in + which case it will wait on an upload queue slot. """ - def commit_bufferblock_worker(self): - """Background uploader thread.""" - - while True: - try: - bufferblock = self._put_queue.get() - if bufferblock is None: - return - loc = self._keep.put(bufferblock.buffer_view[0:bufferblock.write_pointer].tobytes()) - with bufferblock.lock: - bufferblock._locator = loc - bufferblock.buffer_view = None - bufferblock.buffer_block = None - bufferblock.set_state(BufferBlock.COMMITTED) - - except Exception as e: - print e - self._put_errors.put((bufferblock.locator(), e)) - finally: - if self._put_queue is not None: - self._put_queue.task_done() + try: + # Mark the block as PENDING so to disallow any more appends. + block.set_state(_BufferBlock.PENDING) + except StateChangeError as e: + if e.state == _BufferBlock.PENDING and sync: + block.wait_for_commit.wait() + if block.state() == _BufferBlock.ERROR: + raise block.error + return - with self.lock: - if self._put_threads is None: - # Start uploader threads. - - # If we don't limit the Queue size, the upload queue can quickly - # grow to take up gigabytes of RAM if the writing process is - # generating data more quickly than it can be send to the Keep - # servers. - # - # With two upload threads and a queue size of 2, this means up to 4 - # blocks pending. If they are full 64 MiB blocks, that means up to - # 256 MiB of internal buffering, which is the same size as the - # default download block cache in KeepClient. - self._put_queue = Queue.Queue(maxsize=2) - self._put_errors = Queue.Queue() - - self._put_threads = [] - for i in xrange(0, self.num_put_threads): - thread = threading.Thread(target=commit_bufferblock_worker, args=(self,)) - self._put_threads.append(thread) - thread.daemon = True - thread.start() - - # Mark the block as PENDING so to disallow any more appends. - with block.lock: - block.set_state(BufferBlock.PENDING) - self._put_queue.put(block) + if sync: + try: + loc = self._keep.put(block.buffer_view[0:block.write_pointer].tobytes()) + block.set_state(_BufferBlock.COMMITTED, loc) + except Exception as e: + block.set_state(_BufferBlock.ERROR, e) + raise + else: + self.start_put_threads() + self._put_queue.put(block) @synchronized def get_bufferblock(self, locator): return self._bufferblocks.get(locator) + @synchronized + def delete_bufferblock(self, locator): + bb = self._bufferblocks[locator] + bb.clear() + del self._bufferblocks[locator] + def get_block_contents(self, locator, num_retries, cache_only=False): """Fetch a block. @@ -501,7 +601,7 @@ class BlockManager(object): with self.lock: if locator in self._bufferblocks: bufferblock = self._bufferblocks[locator] - if bufferblock.state() != BufferBlock.COMMITTED: + if bufferblock.state() != _BufferBlock.COMMITTED: return bufferblock.buffer_view[0:bufferblock.write_pointer].tobytes() else: locator = bufferblock._locator @@ -513,30 +613,33 @@ class BlockManager(object): def commit_all(self): """Commit all outstanding buffer blocks. - Unlike commit_bufferblock(), this is a synchronous call, and will not - return until all buffer blocks are uploaded. Raises - KeepWriteError() if any blocks failed to upload. + This is a synchronous call, and will not return until all buffer blocks + are uploaded. Raises KeepWriteError() if any blocks failed to upload. """ with self.lock: items = self._bufferblocks.items() for k,v in items: - if v.state() == BufferBlock.WRITABLE: - self.commit_bufferblock(v) + if v.state() != _BufferBlock.COMMITTED: + v.owner.flush(sync=False) with self.lock: if self._put_queue is not None: self._put_queue.join() - if not self._put_errors.empty(): - err = [] - try: - while True: - err.append(self._put_errors.get(False)) - except Queue.Empty: - pass - raise KeepWriteError("Error writing some blocks", err) + err = [] + for k,v in items: + if v.state() == _BufferBlock.ERROR: + err.append((v.locator(), v.error)) + if err: + raise KeepWriteError("Error writing some blocks", err, label="block") + + for k,v in items: + # flush again with sync=True to remove committed bufferblocks from + # the segments. + if v.owner: + v.owner.flush(sync=True) def block_prefetch(self, locator): """Initiate a background download of a block. @@ -551,33 +654,21 @@ class BlockManager(object): if not self.prefetch_enabled: return - def block_prefetch_worker(self): - """The background downloader thread.""" - while True: - try: - b = self._prefetch_queue.get() - if b is None: - return - self._keep.get(b) - except: - pass + if self._keep.get_from_cache(locator) is not None: + return with self.lock: if locator in self._bufferblocks: return - if self._prefetch_threads is None: - self._prefetch_queue = Queue.Queue() - self._prefetch_threads = [] - for i in xrange(0, self.num_get_threads): - thread = threading.Thread(target=block_prefetch_worker, args=(self,)) - self._prefetch_threads.append(thread) - thread.daemon = True - thread.start() + + self.start_get_threads() self._prefetch_queue.put(locator) class ArvadosFile(object): - """ArvadosFile manages the underlying representation of a file in Keep as a + """Represent a file in a Collection. + + ArvadosFile manages the underlying representation of a file in Keep as a sequence of segments spanning a set of blocks, and implements random read/write access. @@ -585,7 +676,7 @@ class ArvadosFile(object): """ - def __init__(self, parent, stream=[], segments=[]): + def __init__(self, parent, name, stream=[], segments=[]): """ ArvadosFile constructor. @@ -596,24 +687,25 @@ class ArvadosFile(object): a list of Range objects representing segments """ self.parent = parent - self._modified = True + self.name = name + self._committed = False self._segments = [] self.lock = parent.root_collection().lock for s in segments: self._add_segment(stream, s.locator, s.range_size) self._current_bblock = None - def sync_mode(self): - return self.parent.sync_mode() + def writable(self): + return self.parent.writable() @synchronized def segments(self): return copy.copy(self._segments) @synchronized - def clone(self, new_parent): + def clone(self, new_parent, new_name): """Make a copy of this file.""" - cp = ArvadosFile(new_parent) + cp = ArvadosFile(new_parent, new_name) cp.replace_contents(self) return cp @@ -624,20 +716,20 @@ class ArvadosFile(object): map_loc = {} self._segments = [] - for r in other.segments(): - new_loc = r.locator - if other.parent._my_block_manager().is_bufferblock(r.locator): - if r.locator not in map_loc: - bufferblock = other.parent._my_block_manager().get_bufferblock(r.locator) - if bufferblock.state() != BufferBlock.WRITABLE: - map_loc[r.locator] = bufferblock.locator() + for other_segment in other.segments(): + new_loc = other_segment.locator + if other.parent._my_block_manager().is_bufferblock(other_segment.locator): + if other_segment.locator not in map_loc: + bufferblock = other.parent._my_block_manager().get_bufferblock(other_segment.locator) + if bufferblock.state() != _BufferBlock.WRITABLE: + map_loc[other_segment.locator] = bufferblock.locator() else: - map_loc[r.locator] = self.parent._my_block_manager().dup_block(bufferblock, self).blockid - new_loc = map_loc[r.locator] + map_loc[other_segment.locator] = self.parent._my_block_manager().dup_block(bufferblock, self).blockid + new_loc = map_loc[other_segment.locator] - self._segments.append(Range(new_loc, r.range_start, r.range_size, r.segment_offset)) + self._segments.append(Range(new_loc, other_segment.range_start, other_segment.range_size, other_segment.segment_offset)) - self._modified = True + self._committed = False def __eq__(self, other): if other is self: @@ -673,14 +765,14 @@ class ArvadosFile(object): return not self.__eq__(other) @synchronized - def set_unmodified(self): - """Clear the modified flag""" - self._modified = False + def set_committed(self): + """Set committed flag to False""" + self._committed = True @synchronized - def modified(self): - """Test the modified flag""" - return self._modified + def committed(self): + """Get whether this is committed or not.""" + return self._committed @must_be_writable @synchronized @@ -700,7 +792,7 @@ class ArvadosFile(object): # segment is past the trucate size, all done break elif size < range_end: - nr = Range(r.locator, r.range_start, size - r.range_start) + nr = Range(r.locator, r.range_start, size - r.range_start, 0) nr.segment_offset = r.segment_offset new_segs.append(nr) break @@ -708,34 +800,45 @@ class ArvadosFile(object): new_segs.append(r) self._segments = new_segs - self._modified = True + self._committed = False elif size > self.size(): - raise IOError("truncate() does not support extending the file size") + raise IOError(errno.EINVAL, "truncate() does not support extending the file size") + + def readfrom(self, offset, size, num_retries, exact=False): + """Read up to `size` bytes from the file starting at `offset`. - def readfrom(self, offset, size, num_retries): - """Read upto `size` bytes from the file starting at `offset`.""" + :exact: + If False (default), return less data than requested if the read + crosses a block boundary and the next block isn't cached. If True, + only return less data than requested when hitting EOF. + """ with self.lock: if size == 0 or offset >= self.size(): return '' - prefetch = locators_and_ranges(self._segments, offset, size + config.KEEP_BLOCK_SIZE) readsegs = locators_and_ranges(self._segments, offset, size) + prefetch = locators_and_ranges(self._segments, offset + size, config.KEEP_BLOCK_SIZE, limit=32) - for lr in prefetch: - self.parent._my_block_manager().block_prefetch(lr.locator) - + locs = set() data = [] for lr in readsegs: - block = self.parent._my_block_manager().get_block_contents(lr.locator, num_retries=num_retries, cache_only=bool(data)) + block = self.parent._my_block_manager().get_block_contents(lr.locator, num_retries=num_retries, cache_only=(bool(data) and not exact)) if block: - data.append(block[lr.segment_offset:lr.segment_offset+lr.segment_size]) + blockview = memoryview(block) + data.append(blockview[lr.segment_offset:lr.segment_offset+lr.segment_size].tobytes()) + locs.add(lr.locator) else: break + + for lr in prefetch: + if lr.locator not in locs: + self.parent._my_block_manager().block_prefetch(lr.locator) + locs.add(lr.locator) + return ''.join(data) - def _repack_writes(self): - """Test if the buffer block has more data than is referenced by actual - segments. + def _repack_writes(self, num_retries): + """Test if the buffer block has more data than actual segments. This happens when a buffered write over-writes a file range written in a previous buffered write. Re-pack the buffer block for efficiency @@ -752,9 +855,10 @@ class ArvadosFile(object): if write_total < self._current_bblock.size(): # There is more data in the buffer block than is actually accounted for by segments, so # re-pack into a new buffer by copying over to a new buffer block. + contents = self.parent._my_block_manager().get_block_contents(self._current_bblock.blockid, num_retries) new_bb = self.parent._my_block_manager().alloc_bufferblock(self._current_bblock.blockid, starting_capacity=write_total, owner=self) for t in bufferblock_segs: - new_bb.append(self._current_bblock.buffer_view[t.segment_offset:t.segment_offset+t.range_size].tobytes()) + new_bb.append(contents[t.segment_offset:t.segment_offset+t.range_size]) t.segment_offset = new_bb.size() - t.range_size self._current_bblock = new_bb @@ -775,37 +879,80 @@ class ArvadosFile(object): raise ArgumentError("Offset is past the end of the file") if len(data) > config.KEEP_BLOCK_SIZE: - raise ArgumentError("Please append data in chunks smaller than %i bytes (config.KEEP_BLOCK_SIZE)" % (config.KEEP_BLOCK_SIZE)) + # Chunk it up into smaller writes + n = 0 + dataview = memoryview(data) + while n < len(data): + self.writeto(offset+n, dataview[n:n + config.KEEP_BLOCK_SIZE].tobytes(), num_retries) + n += config.KEEP_BLOCK_SIZE + return - self._modified = True + self._committed = False - if self._current_bblock is None or self._current_bblock.state() != BufferBlock.WRITABLE: + if self._current_bblock is None or self._current_bblock.state() != _BufferBlock.WRITABLE: self._current_bblock = self.parent._my_block_manager().alloc_bufferblock(owner=self) if (self._current_bblock.size() + len(data)) > config.KEEP_BLOCK_SIZE: - self._repack_writes() + self._repack_writes(num_retries) if (self._current_bblock.size() + len(data)) > config.KEEP_BLOCK_SIZE: - self.parent._my_block_manager().commit_bufferblock(self._current_bblock) + self.parent._my_block_manager().commit_bufferblock(self._current_bblock, sync=False) self._current_bblock = self.parent._my_block_manager().alloc_bufferblock(owner=self) self._current_bblock.append(data) replace_range(self._segments, offset, len(data), self._current_bblock.blockid, self._current_bblock.write_pointer - len(data)) + self.parent.notify(WRITE, self.parent, self.name, (self, self)) + + return len(data) + + @synchronized + def flush(self, sync=True, num_retries=0): + """Flush the current bufferblock to Keep. + + :sync: + If True, commit block synchronously, wait until buffer block has been written. + If False, commit block asynchronously, return immediately after putting block into + the keep put queue. + """ + if self.committed(): + return + + if self._current_bblock and self._current_bblock.state() != _BufferBlock.COMMITTED: + if self._current_bblock.state() == _BufferBlock.WRITABLE: + self._repack_writes(num_retries) + self.parent._my_block_manager().commit_bufferblock(self._current_bblock, sync=sync) + + if sync: + to_delete = set() + for s in self._segments: + bb = self.parent._my_block_manager().get_bufferblock(s.locator) + if bb: + if bb.state() != _BufferBlock.COMMITTED: + self.parent._my_block_manager().commit_bufferblock(self._current_bblock, sync=True) + to_delete.add(s.locator) + s.locator = bb.locator() + for s in to_delete: + self.parent._my_block_manager().delete_bufferblock(s) + + self.parent.notify(MOD, self.parent, self.name, (self, self)) + @must_be_writable @synchronized def add_segment(self, blocks, pos, size): - """Add a segment to the end of the file, with `pos` and `offset` referencing a - section of the stream described by `blocks` (a list of Range objects) + """Add a segment to the end of the file. + + `pos` and `offset` reference a section of the stream described by + `blocks` (a list of Range objects) """ self._add_segment(blocks, pos, size) def _add_segment(self, blocks, pos, size): """Internal implementation of add_segment.""" - self._modified = True + self._committed = False for lr in locators_and_ranges(blocks, pos, size): - last = self._segments[-1] if self._segments else Range(0, 0, 0) + last = self._segments[-1] if self._segments else Range(0, 0, 0, 0) r = Range(lr.locator, last.range_start+last.range_size, lr.segment_size, lr.segment_offset) self._segments.append(r) @@ -818,6 +965,33 @@ class ArvadosFile(object): else: return 0 + @synchronized + def manifest_text(self, stream_name=".", portable_locators=False, normalize=False): + buf = "" + filestream = [] + for segment in self.segments: + loc = segment.locator + if loc.startswith("bufferblock"): + loc = self._bufferblocks[loc].calculate_locator() + if portable_locators: + loc = KeepLocator(loc).stripped() + filestream.append(LocatorAndRange(loc, locator_block_size(loc), + segment.segment_offset, segment.range_size)) + buf += ' '.join(normalize_stream(stream_name, {stream_name: filestream})) + buf += "\n" + return buf + + @must_be_writable + @synchronized + def _reparent(self, newparent, newname): + self._committed = False + self.flush(sync=True) + self.parent.remove(self.name) + self.parent = newparent + self.name = newname + self.lock = self.parent.root_collection().lock + + class ArvadosFileReader(ArvadosFileReaderBase): """Wraps ArvadosFile in a file-like object supporting reading only. @@ -826,25 +1000,44 @@ class ArvadosFileReader(ArvadosFileReaderBase): """ - def __init__(self, arvadosfile, name, mode="r", num_retries=None): - super(ArvadosFileReader, self).__init__(name, mode, num_retries=num_retries) + def __init__(self, arvadosfile, num_retries=None): + super(ArvadosFileReader, self).__init__(arvadosfile.name, "r", num_retries=num_retries) self.arvadosfile = arvadosfile def size(self): return self.arvadosfile.size() - @FileLikeObjectBase._before_close + def stream_name(self): + return self.arvadosfile.parent.stream_name() + + @_FileLikeObjectBase._before_close @retry_method - def read(self, size, num_retries=None): - """Read up to `size` bytes from the stream, starting at the current file position.""" - data = self.arvadosfile.readfrom(self._filepos, size, num_retries) - self._filepos += len(data) - return data + def read(self, size=None, num_retries=None): + """Read up to `size` bytes from the file and return the result. - @FileLikeObjectBase._before_close + Starts at the current file position. If `size` is None, read the + entire remainder of the file. + """ + if size is None: + data = [] + rd = self.arvadosfile.readfrom(self._filepos, config.KEEP_BLOCK_SIZE, num_retries) + while rd: + data.append(rd) + self._filepos += len(rd) + rd = self.arvadosfile.readfrom(self._filepos, config.KEEP_BLOCK_SIZE, num_retries) + return ''.join(data) + else: + data = self.arvadosfile.readfrom(self._filepos, size, num_retries, exact=True) + self._filepos += len(data) + return data + + @_FileLikeObjectBase._before_close @retry_method def readfrom(self, offset, size, num_retries=None): - """Read up to `size` bytes from the stream, starting at the current file position.""" + """Read up to `size` bytes from the stream, starting at the specified file offset. + + This method does not change the file position. + """ return self.arvadosfile.readfrom(offset, size, num_retries) def flush(self): @@ -859,10 +1052,11 @@ class ArvadosFileWriter(ArvadosFileReader): """ - def __init__(self, arvadosfile, name, mode, num_retries=None): - super(ArvadosFileWriter, self).__init__(arvadosfile, name, mode, num_retries=num_retries) + def __init__(self, arvadosfile, mode, num_retries=None): + super(ArvadosFileWriter, self).__init__(arvadosfile, num_retries=num_retries) + self.mode = mode - @FileLikeObjectBase._before_close + @_FileLikeObjectBase._before_close @retry_method def write(self, data, num_retries=None): if self.mode[0] == "a": @@ -870,13 +1064,15 @@ class ArvadosFileWriter(ArvadosFileReader): else: self.arvadosfile.writeto(self._filepos, data, num_retries) self._filepos += len(data) + return len(data) - @FileLikeObjectBase._before_close + @_FileLikeObjectBase._before_close @retry_method def writelines(self, seq, num_retries=None): for s in seq: - self.write(s, num_retries) + self.write(s, num_retries=num_retries) + @_FileLikeObjectBase._before_close def truncate(self, size=None): if size is None: size = self._filepos @@ -884,6 +1080,11 @@ class ArvadosFileWriter(ArvadosFileReader): if self._filepos > self.size(): self._filepos = self.size() + @_FileLikeObjectBase._before_close + def flush(self): + self.arvadosfile.flush() + def close(self): - if self.arvadosfile.parent.sync_mode() == SYNC_LIVE: - self.arvadosfile.parent.root_collection().save() + if not self.closed: + self.flush() + super(ArvadosFileWriter, self).close()