sdk/python/arvados/arvfile.py

   1 # Copyright (C) The Arvados Authors. All rights reserved.
   2 #
   3 # SPDX-License-Identifier: Apache-2.0
   4
   5 from __future__ import absolute_import
   6 from __future__ import division
   7 from future import standard_library
   8 from future.utils import listitems, listvalues
   9 standard_library.install_aliases()
  10 from builtins import range
  11 from builtins import object
  12 import bz2
  13 import collections
  14 import copy
  15 import errno
  16 import functools
  17 import hashlib
  18 import logging
  19 import os
  20 import queue
  21 import re
  22 import sys
  23 import threading
  24 import uuid
  25 import zlib
  26
  27 from . import config
  28 from .errors import KeepWriteError, AssertionError, ArgumentError
  29 from .keep import KeepLocator
  30 from ._normalize_stream import normalize_stream
  31 from ._ranges import locators_and_ranges, replace_range, Range, LocatorAndRange
  32 from .retry import retry_method
  33
  34 MOD = "mod"
  35 WRITE = "write"
  36
  37 _logger = logging.getLogger('arvados.arvfile')
  38
  39 def split(path):
  40     """split(path) -> streamname, filename
  41
  42     Separate the stream name and file name in a /-separated stream path and
  43     return a tuple (stream_name, file_name).  If no stream name is available,
  44     assume '.'.
  45
  46     """
  47     try:
  48         stream_name, file_name = path.rsplit('/', 1)
  49     except ValueError:  # No / in string
  50         stream_name, file_name = '.', path
  51     return stream_name, file_name
  52
  53
  54 class UnownedBlockError(Exception):
  55     """Raised when there's an writable block without an owner on the BlockManager."""
  56     pass
  57
  58
  59 class _FileLikeObjectBase(object):
  60     def __init__(self, name, mode):
  61         self.name = name
  62         self.mode = mode
  63         self.closed = False
  64
  65     @staticmethod
  66     def _before_close(orig_func):
  67         @functools.wraps(orig_func)
  68         def before_close_wrapper(self, *args, **kwargs):
  69             if self.closed:
  70                 raise ValueError("I/O operation on closed stream file")
  71             return orig_func(self, *args, **kwargs)
  72         return before_close_wrapper
  73
  74     def __enter__(self):
  75         return self
  76
  77     def __exit__(self, exc_type, exc_value, traceback):
  78         try:
  79             self.close()
  80         except Exception:
  81             if exc_type is None:
  82                 raise
  83
  84     def close(self):
  85         self.closed = True
  86
  87
  88 class ArvadosFileReaderBase(_FileLikeObjectBase):
  89     def __init__(self, name, mode, num_retries=None):
  90         super(ArvadosFileReaderBase, self).__init__(name, mode)
  91         self._filepos = 0
  92         self.num_retries = num_retries
  93         self._readline_cache = (None, None)
  94
  95     def __iter__(self):
  96         while True:
  97             data = self.readline()
  98             if not data:
  99                 break
 100             yield data
 101
 102     def decompressed_name(self):
 103         return re.sub('\.(bz2|gz)$', '', self.name)
 104
 105     @_FileLikeObjectBase._before_close
 106     def seek(self, pos, whence=os.SEEK_SET):
 107         if whence == os.SEEK_CUR:
 108             pos += self._filepos
 109         elif whence == os.SEEK_END:
 110             pos += self.size()
 111         if pos < 0:
 112             raise IOError(errno.EINVAL, "Tried to seek to negative file offset.")
 113         self._filepos = pos
 114         return self._filepos
 115
 116     def tell(self):
 117         return self._filepos
 118
 119     def readable(self):
 120         return True
 121
 122     def writable(self):
 123         return False
 124
 125     def seekable(self):
 126         return True
 127
 128     @_FileLikeObjectBase._before_close
 129     @retry_method
 130     def readall(self, size=2**20, num_retries=None):
 131         while True:
 132             data = self.read(size, num_retries=num_retries)
 133             if len(data) == 0:
 134                 break
 135             yield data
 136
 137     @_FileLikeObjectBase._before_close
 138     @retry_method
 139     def readline(self, size=float('inf'), num_retries=None):
 140         cache_pos, cache_data = self._readline_cache
 141         if self.tell() == cache_pos:
 142             data = [cache_data]
 143             self._filepos += len(cache_data)
 144         else:
 145             data = [b'']
 146         data_size = len(data[-1])
 147         while (data_size < size) and (b'\n' not in data[-1]):
 148             next_read = self.read(2 ** 20, num_retries=num_retries)
 149             if not next_read:
 150                 break
 151             data.append(next_read)
 152             data_size += len(next_read)
 153         data = b''.join(data)
 154         try:
 155             nextline_index = data.index(b'\n') + 1
 156         except ValueError:
 157             nextline_index = len(data)
 158         nextline_index = min(nextline_index, size)
 159         self._filepos -= len(data) - nextline_index
 160         self._readline_cache = (self.tell(), data[nextline_index:])
 161         return data[:nextline_index].decode()
 162
 163     @_FileLikeObjectBase._before_close
 164     @retry_method
 165     def decompress(self, decompress, size, num_retries=None):
 166         for segment in self.readall(size, num_retries=num_retries):
 167             data = decompress(segment)
 168             if data:
 169                 yield data
 170
 171     @_FileLikeObjectBase._before_close
 172     @retry_method
 173     def readall_decompressed(self, size=2**20, num_retries=None):
 174         self.seek(0)
 175         if self.name.endswith('.bz2'):
 176             dc = bz2.BZ2Decompressor()
 177             return self.decompress(dc.decompress, size,
 178                                    num_retries=num_retries)
 179         elif self.name.endswith('.gz'):
 180             dc = zlib.decompressobj(16+zlib.MAX_WBITS)
 181             return self.decompress(lambda segment: dc.decompress(dc.unconsumed_tail + segment),
 182                                    size, num_retries=num_retries)
 183         else:
 184             return self.readall(size, num_retries=num_retries)
 185
 186     @_FileLikeObjectBase._before_close
 187     @retry_method
 188     def readlines(self, sizehint=float('inf'), num_retries=None):
 189         data = []
 190         data_size = 0
 191         for s in self.readall(num_retries=num_retries):
 192             data.append(s)
 193             data_size += len(s)
 194             if data_size >= sizehint:
 195                 break
 196         return b''.join(data).decode().splitlines(True)
 197
 198     def size(self):
 199         raise IOError(errno.ENOSYS, "Not implemented")
 200
 201     def read(self, size, num_retries=None):
 202         raise IOError(errno.ENOSYS, "Not implemented")
 203
 204     def readfrom(self, start, size, num_retries=None):
 205         raise IOError(errno.ENOSYS, "Not implemented")
 206
 207
 208 class StreamFileReader(ArvadosFileReaderBase):
 209     class _NameAttribute(str):
 210         # The Python file API provides a plain .name attribute.
 211         # Older SDK provided a name() method.
 212         # This class provides both, for maximum compatibility.
 213         def __call__(self):
 214             return self
 215
 216     def __init__(self, stream, segments, name):
 217         super(StreamFileReader, self).__init__(self._NameAttribute(name), 'rb', num_retries=stream.num_retries)
 218         self._stream = stream
 219         self.segments = segments
 220
 221     def stream_name(self):
 222         return self._stream.name()
 223
 224     def size(self):
 225         n = self.segments[-1]
 226         return n.range_start + n.range_size
 227
 228     @_FileLikeObjectBase._before_close
 229     @retry_method
 230     def read(self, size, num_retries=None):
 231         """Read up to 'size' bytes from the stream, starting at the current file position"""
 232         if size == 0:
 233             return b''
 234
 235         data = b''
 236         available_chunks = locators_and_ranges(self.segments, self._filepos, size)
 237         if available_chunks:
 238             lr = available_chunks[0]
 239             data = self._stream.readfrom(lr.locator+lr.segment_offset,
 240                                          lr.segment_size,
 241                                          num_retries=num_retries)
 242
 243         self._filepos += len(data)
 244         return data
 245
 246     @_FileLikeObjectBase._before_close
 247     @retry_method
 248     def readfrom(self, start, size, num_retries=None):
 249         """Read up to 'size' bytes from the stream, starting at 'start'"""
 250         if size == 0:
 251             return b''
 252
 253         data = []
 254         for lr in locators_and_ranges(self.segments, start, size):
 255             data.append(self._stream.readfrom(lr.locator+lr.segment_offset, lr.segment_size,
 256                                               num_retries=num_retries))
 257         return b''.join(data)
 258
 259     def as_manifest(self):
 260         segs = []
 261         for r in self.segments:
 262             segs.extend(self._stream.locators_and_ranges(r.locator, r.range_size))
 263         return " ".join(normalize_stream(".", {self.name: segs})) + "\n"
 264
 265
 266 def synchronized(orig_func):
 267     @functools.wraps(orig_func)
 268     def synchronized_wrapper(self, *args, **kwargs):
 269         with self.lock:
 270             return orig_func(self, *args, **kwargs)
 271     return synchronized_wrapper
 272
 273
 274 class StateChangeError(Exception):
 275     def __init__(self, message, state, nextstate):
 276         super(StateChangeError, self).__init__(message)
 277         self.state = state
 278         self.nextstate = nextstate
 279
 280 class _BufferBlock(object):
 281     """A stand-in for a Keep block that is in the process of being written.
 282
 283     Writers can append to it, get the size, and compute the Keep locator.
 284     There are three valid states:
 285
 286     WRITABLE
 287       Can append to block.
 288
 289     PENDING
 290       Block is in the process of being uploaded to Keep, append is an error.
 291
 292     COMMITTED
 293       The block has been written to Keep, its internal buffer has been
 294       released, fetching the block will fetch it via keep client (since we
 295       discarded the internal copy), and identifiers referring to the BufferBlock
 296       can be replaced with the block locator.
 297
 298     """
 299
 300     WRITABLE = 0
 301     PENDING = 1
 302     COMMITTED = 2
 303     ERROR = 3
 304     DELETED = 4
 305
 306     def __init__(self, blockid, starting_capacity, owner):
 307         """
 308         :blockid:
 309           the identifier for this block
 310
 311         :starting_capacity:
 312           the initial buffer capacity
 313
 314         :owner:
 315           ArvadosFile that owns this block
 316
 317         """
 318         self.blockid = blockid
 319         self.buffer_block = bytearray(starting_capacity)
 320         self.buffer_view = memoryview(self.buffer_block)
 321         self.write_pointer = 0
 322         self._state = _BufferBlock.WRITABLE
 323         self._locator = None
 324         self.owner = owner
 325         self.lock = threading.Lock()
 326         self.wait_for_commit = threading.Event()
 327         self.error = None
 328
 329     @synchronized
 330     def append(self, data):
 331         """Append some data to the buffer.
 332
 333         Only valid if the block is in WRITABLE state.  Implements an expanding
 334         buffer, doubling capacity as needed to accomdate all the data.
 335
 336         """
 337         if self._state == _BufferBlock.WRITABLE:
 338             if not isinstance(data, bytes) and not isinstance(data, memoryview):
 339                 data = data.encode()
 340             while (self.write_pointer+len(data)) > len(self.buffer_block):
 341                 new_buffer_block = bytearray(len(self.buffer_block) * 2)
 342                 new_buffer_block[0:self.write_pointer] = self.buffer_block[0:self.write_pointer]
 343                 self.buffer_block = new_buffer_block
 344                 self.buffer_view = memoryview(self.buffer_block)
 345             self.buffer_view[self.write_pointer:self.write_pointer+len(data)] = data
 346             self.write_pointer += len(data)
 347             self._locator = None
 348         else:
 349             raise AssertionError("Buffer block is not writable")
 350
 351     STATE_TRANSITIONS = frozenset([
 352             (WRITABLE, PENDING),
 353             (PENDING, COMMITTED),
 354             (PENDING, ERROR),
 355             (ERROR, PENDING)])
 356
 357     @synchronized
 358     def set_state(self, nextstate, val=None):
 359         if (self._state, nextstate) not in self.STATE_TRANSITIONS:
 360             raise StateChangeError("Invalid state change from %s to %s" % (self._state, nextstate), self._state, nextstate)
 361         self._state = nextstate
 362
 363         if self._state == _BufferBlock.PENDING:
 364             self.wait_for_commit.clear()
 365
 366         if self._state == _BufferBlock.COMMITTED:
 367             self._locator = val
 368             self.buffer_view = None
 369             self.buffer_block = None
 370             self.wait_for_commit.set()
 371
 372         if self._state == _BufferBlock.ERROR:
 373             self.error = val
 374             self.wait_for_commit.set()
 375
 376     @synchronized
 377     def state(self):
 378         return self._state
 379
 380     def size(self):
 381         """The amount of data written to the buffer."""
 382         return self.write_pointer
 383
 384     @synchronized
 385     def locator(self):
 386         """The Keep locator for this buffer's contents."""
 387         if self._locator is None:
 388             self._locator = "%s+%i" % (hashlib.md5(self.buffer_view[0:self.write_pointer]).hexdigest(), self.size())
 389         return self._locator
 390
 391     @synchronized
 392     def clone(self, new_blockid, owner):
 393         if self._state == _BufferBlock.COMMITTED:
 394             raise AssertionError("Cannot duplicate committed buffer block")
 395         bufferblock = _BufferBlock(new_blockid, self.size(), owner)
 396         bufferblock.append(self.buffer_view[0:self.size()])
 397         return bufferblock
 398
 399     @synchronized
 400     def clear(self):
 401         self._state = _BufferBlock.DELETED
 402         self.owner = None
 403         self.buffer_block = None
 404         self.buffer_view = None
 405
 406     @synchronized
 407     def repack_writes(self):
 408         """Optimize buffer block by repacking segments in file sequence.
 409
 410         When the client makes random writes, they appear in the buffer block in
 411         the sequence they were written rather than the sequence they appear in
 412         the file.  This makes for inefficient, fragmented manifests.  Attempt
 413         to optimize by repacking writes in file sequence.
 414
 415         """
 416         if self._state != _BufferBlock.WRITABLE:
 417             raise AssertionError("Cannot repack non-writable block")
 418
 419         segs = self.owner.segments()
 420
 421         # Collect the segments that reference the buffer block.
 422         bufferblock_segs = [s for s in segs if s.locator == self.blockid]
 423
 424         # Collect total data referenced by segments (could be smaller than
 425         # bufferblock size if a portion of the file was written and
 426         # then overwritten).
 427         write_total = sum([s.range_size for s in bufferblock_segs])
 428
 429         if write_total < self.size() or len(bufferblock_segs) > 1:
 430             # If there's more than one segment referencing this block, it is
 431             # due to out-of-order writes and will produce a fragmented
 432             # manifest, so try to optimize by re-packing into a new buffer.
 433             contents = self.buffer_view[0:self.write_pointer].tobytes()
 434             new_bb = _BufferBlock(None, write_total, None)
 435             for t in bufferblock_segs:
 436                 new_bb.append(contents[t.segment_offset:t.segment_offset+t.range_size])
 437                 t.segment_offset = new_bb.size() - t.range_size
 438
 439             self.buffer_block = new_bb.buffer_block
 440             self.buffer_view = new_bb.buffer_view
 441             self.write_pointer = new_bb.write_pointer
 442             self._locator = None
 443             new_bb.clear()
 444             self.owner.set_segments(segs)
 445
 446     def __repr__(self):
 447         return "<BufferBlock %s>" % (self.blockid)
 448
 449
 450 class NoopLock(object):
 451     def __enter__(self):
 452         return self
 453
 454     def __exit__(self, exc_type, exc_value, traceback):
 455         pass
 456
 457     def acquire(self, blocking=False):
 458         pass
 459
 460     def release(self):
 461         pass
 462
 463
 464 def must_be_writable(orig_func):
 465     @functools.wraps(orig_func)
 466     def must_be_writable_wrapper(self, *args, **kwargs):
 467         if not self.writable():
 468             raise IOError(errno.EROFS, "Collection is read-only.")
 469         return orig_func(self, *args, **kwargs)
 470     return must_be_writable_wrapper
 471
 472
 473 class _BlockManager(object):
 474     """BlockManager handles buffer blocks.
 475
 476     Also handles background block uploads, and background block prefetch for a
 477     Collection of ArvadosFiles.
 478
 479     """
 480
 481     DEFAULT_PUT_THREADS = 2
 482     DEFAULT_GET_THREADS = 2
 483
 484     def __init__(self, keep, copies=None, put_threads=None, num_retries=None, storage_classes_func=None):
 485         """keep: KeepClient object to use"""
 486         self._keep = keep
 487         self._bufferblocks = collections.OrderedDict()
 488         self._put_queue = None
 489         self._put_threads = None
 490         self._prefetch_queue = None
 491         self._prefetch_threads = None
 492         self.lock = threading.Lock()
 493         self.prefetch_enabled = True
 494         self.num_put_threads = put_threads or _BlockManager.DEFAULT_PUT_THREADS
 495         self.num_get_threads = _BlockManager.DEFAULT_GET_THREADS
 496         self.copies = copies
 497         self.storage_classes = storage_classes_func or (lambda: [])
 498         self._pending_write_size = 0
 499         self.threads_lock = threading.Lock()
 500         self.padding_block = None
 501         self.num_retries = num_retries
 502
 503     @synchronized
 504     def alloc_bufferblock(self, blockid=None, starting_capacity=2**14, owner=None):
 505         """Allocate a new, empty bufferblock in WRITABLE state and return it.
 506
 507         :blockid:
 508           optional block identifier, otherwise one will be automatically assigned
 509
 510         :starting_capacity:
 511           optional capacity, otherwise will use default capacity
 512
 513         :owner:
 514           ArvadosFile that owns this block
 515
 516         """
 517         return self._alloc_bufferblock(blockid, starting_capacity, owner)
 518
 519     def _alloc_bufferblock(self, blockid=None, starting_capacity=2**14, owner=None):
 520         if blockid is None:
 521             blockid = str(uuid.uuid4())
 522         bufferblock = _BufferBlock(blockid, starting_capacity=starting_capacity, owner=owner)
 523         self._bufferblocks[bufferblock.blockid] = bufferblock
 524         return bufferblock
 525
 526     @synchronized
 527     def dup_block(self, block, owner):
 528         """Create a new bufferblock initialized with the content of an existing bufferblock.
 529
 530         :block:
 531           the buffer block to copy.
 532
 533         :owner:
 534           ArvadosFile that owns the new block
 535
 536         """
 537         new_blockid = str(uuid.uuid4())
 538         bufferblock = block.clone(new_blockid, owner)
 539         self._bufferblocks[bufferblock.blockid] = bufferblock
 540         return bufferblock
 541
 542     @synchronized
 543     def is_bufferblock(self, locator):
 544         return locator in self._bufferblocks
 545
 546     def _commit_bufferblock_worker(self):
 547         """Background uploader thread."""
 548
 549         while True:
 550             try:
 551                 bufferblock = self._put_queue.get()
 552                 if bufferblock is None:
 553                     return
 554
 555                 if self.copies is None:
 556                     loc = self._keep.put(bufferblock.buffer_view[0:bufferblock.write_pointer].tobytes(), num_retries=self.num_retries, classes=self.storage_classes())
 557                 else:
 558                     loc = self._keep.put(bufferblock.buffer_view[0:bufferblock.write_pointer].tobytes(), num_retries=self.num_retries, copies=self.copies, classes=self.storage_classes())
 559                 bufferblock.set_state(_BufferBlock.COMMITTED, loc)
 560             except Exception as e:
 561                 bufferblock.set_state(_BufferBlock.ERROR, e)
 562             finally:
 563                 if self._put_queue is not None:
 564                     self._put_queue.task_done()
 565
 566     def start_put_threads(self):
 567         with self.threads_lock:
 568             if self._put_threads is None:
 569                 # Start uploader threads.
 570
 571                 # If we don't limit the Queue size, the upload queue can quickly
 572                 # grow to take up gigabytes of RAM if the writing process is
 573                 # generating data more quickly than it can be sent to the Keep
 574                 # servers.
 575                 #
 576                 # With two upload threads and a queue size of 2, this means up to 4
 577                 # blocks pending.  If they are full 64 MiB blocks, that means up to
 578                 # 256 MiB of internal buffering, which is the same size as the
 579                 # default download block cache in KeepClient.
 580                 self._put_queue = queue.Queue(maxsize=2)
 581
 582                 self._put_threads = []
 583                 for i in range(0, self.num_put_threads):
 584                     thread = threading.Thread(target=self._commit_bufferblock_worker)
 585                     self._put_threads.append(thread)
 586                     thread.daemon = True
 587                     thread.start()
 588
 589     def _block_prefetch_worker(self):
 590         """The background downloader thread."""
 591         while True:
 592             try:
 593                 b = self._prefetch_queue.get()
 594                 if b is None:
 595                     return
 596                 self._keep.get(b)
 597             except Exception:
 598                 _logger.exception("Exception doing block prefetch")
 599
 600     @synchronized
 601     def start_get_threads(self):
 602         if self._prefetch_threads is None:
 603             self._prefetch_queue = queue.Queue()
 604             self._prefetch_threads = []
 605             for i in range(0, self.num_get_threads):
 606                 thread = threading.Thread(target=self._block_prefetch_worker)
 607                 self._prefetch_threads.append(thread)
 608                 thread.daemon = True
 609                 thread.start()
 610
 611
 612     @synchronized
 613     def stop_threads(self):
 614         """Shut down and wait for background upload and download threads to finish."""
 615
 616         if self._put_threads is not None:
 617             for t in self._put_threads:
 618                 self._put_queue.put(None)
 619             for t in self._put_threads:
 620                 t.join()
 621         self._put_threads = None
 622         self._put_queue = None
 623
 624         if self._prefetch_threads is not None:
 625             for t in self._prefetch_threads:
 626                 self._prefetch_queue.put(None)
 627             for t in self._prefetch_threads:
 628                 t.join()
 629         self._prefetch_threads = None
 630         self._prefetch_queue = None
 631
 632     def __enter__(self):
 633         return self
 634
 635     def __exit__(self, exc_type, exc_value, traceback):
 636         self.stop_threads()
 637
 638     @synchronized
 639     def repack_small_blocks(self, force=False, sync=False, closed_file_size=0):
 640         """Packs small blocks together before uploading"""
 641
 642         self._pending_write_size += closed_file_size
 643
 644         # Check if there are enough small blocks for filling up one in full
 645         if not (force or (self._pending_write_size >= config.KEEP_BLOCK_SIZE)):
 646             return
 647
 648         # Search blocks ready for getting packed together before being
 649         # committed to Keep.
 650         # A WRITABLE block always has an owner.
 651         # A WRITABLE block with its owner.closed() implies that its
 652         # size is <= KEEP_BLOCK_SIZE/2.
 653         try:
 654             small_blocks = [b for b in listvalues(self._bufferblocks)
 655                             if b.state() == _BufferBlock.WRITABLE and b.owner.closed()]
 656         except AttributeError:
 657             # Writable blocks without owner shouldn't exist.
 658             raise UnownedBlockError()
 659
 660         if len(small_blocks) <= 1:
 661             # Not enough small blocks for repacking
 662             return
 663
 664         for bb in small_blocks:
 665             bb.repack_writes()
 666
 667         # Update the pending write size count with its true value, just in case
 668         # some small file was opened, written and closed several times.
 669         self._pending_write_size = sum([b.size() for b in small_blocks])
 670
 671         if self._pending_write_size < config.KEEP_BLOCK_SIZE and not force:
 672             return
 673
 674         new_bb = self._alloc_bufferblock()
 675         new_bb.owner = []
 676         files = []
 677         while len(small_blocks) > 0 and (new_bb.write_pointer + small_blocks[0].size()) <= config.KEEP_BLOCK_SIZE:
 678             bb = small_blocks.pop(0)
 679             new_bb.owner.append(bb.owner)
 680             self._pending_write_size -= bb.size()
 681             new_bb.append(bb.buffer_view[0:bb.write_pointer].tobytes())
 682             files.append((bb, new_bb.write_pointer - bb.size()))
 683
 684         self.commit_bufferblock(new_bb, sync=sync)
 685
 686         for bb, new_bb_segment_offset in files:
 687             newsegs = bb.owner.segments()
 688             for s in newsegs:
 689                 if s.locator == bb.blockid:
 690                     s.locator = new_bb.blockid
 691                     s.segment_offset = new_bb_segment_offset+s.segment_offset
 692             bb.owner.set_segments(newsegs)
 693             self._delete_bufferblock(bb.blockid)
 694
 695     def commit_bufferblock(self, block, sync):
 696         """Initiate a background upload of a bufferblock.
 697
 698         :block:
 699           The block object to upload
 700
 701         :sync:
 702           If `sync` is True, upload the block synchronously.
 703           If `sync` is False, upload the block asynchronously.  This will
 704           return immediately unless the upload queue is at capacity, in
 705           which case it will wait on an upload queue slot.
 706
 707         """
 708         try:
 709             # Mark the block as PENDING so to disallow any more appends.
 710             block.set_state(_BufferBlock.PENDING)
 711         except StateChangeError as e:
 712             if e.state == _BufferBlock.PENDING:
 713                 if sync:
 714                     block.wait_for_commit.wait()
 715                 else:
 716                     return
 717             if block.state() == _BufferBlock.COMMITTED:
 718                 return
 719             elif block.state() == _BufferBlock.ERROR:
 720                 raise block.error
 721             else:
 722                 raise
 723
 724         if sync:
 725             try:
 726                 if self.copies is None:
 727                     loc = self._keep.put(block.buffer_view[0:block.write_pointer].tobytes(), num_retries=self.num_retries, classes=self.storage_classes())
 728                 else:
 729                     loc = self._keep.put(block.buffer_view[0:block.write_pointer].tobytes(), num_retries=self.num_retries, copies=self.copies, classes=self.storage_classes())
 730                 block.set_state(_BufferBlock.COMMITTED, loc)
 731             except Exception as e:
 732                 block.set_state(_BufferBlock.ERROR, e)
 733                 raise
 734         else:
 735             self.start_put_threads()
 736             self._put_queue.put(block)
 737
 738     @synchronized
 739     def get_bufferblock(self, locator):
 740         return self._bufferblocks.get(locator)
 741
 742     @synchronized
 743     def get_padding_block(self):
 744         """Get a bufferblock 64 MB in size consisting of all zeros, used as padding
 745         when using truncate() to extend the size of a file.
 746
 747         For reference (and possible future optimization), the md5sum of the
 748         padding block is: 7f614da9329cd3aebf59b91aadc30bf0+67108864
 749
 750         """
 751
 752         if self.padding_block is None:
 753             self.padding_block = self._alloc_bufferblock(starting_capacity=config.KEEP_BLOCK_SIZE)
 754             self.padding_block.write_pointer = config.KEEP_BLOCK_SIZE
 755             self.commit_bufferblock(self.padding_block, False)
 756         return self.padding_block
 757
 758     @synchronized
 759     def delete_bufferblock(self, locator):
 760         self._delete_bufferblock(locator)
 761
 762     def _delete_bufferblock(self, locator):
 763         bb = self._bufferblocks[locator]
 764         bb.clear()
 765         del self._bufferblocks[locator]
 766
 767     def get_block_contents(self, locator, num_retries, cache_only=False):
 768         """Fetch a block.
 769
 770         First checks to see if the locator is a BufferBlock and return that, if
 771         not, passes the request through to KeepClient.get().
 772
 773         """
 774         with self.lock:
 775             if locator in self._bufferblocks:
 776                 bufferblock = self._bufferblocks[locator]
 777                 if bufferblock.state() != _BufferBlock.COMMITTED:
 778                     return bufferblock.buffer_view[0:bufferblock.write_pointer].tobytes()
 779                 else:
 780                     locator = bufferblock._locator
 781         if cache_only:
 782             return self._keep.get_from_cache(locator)
 783         else:
 784             return self._keep.get(locator, num_retries=num_retries)
 785
 786     def commit_all(self):
 787         """Commit all outstanding buffer blocks.
 788
 789         This is a synchronous call, and will not return until all buffer blocks
 790         are uploaded.  Raises KeepWriteError() if any blocks failed to upload.
 791
 792         """
 793         self.repack_small_blocks(force=True, sync=True)
 794
 795         with self.lock:
 796             items = listitems(self._bufferblocks)
 797
 798         for k,v in items:
 799             if v.state() != _BufferBlock.COMMITTED and v.owner:
 800                 # Ignore blocks with a list of owners, as if they're not in COMMITTED
 801                 # state, they're already being committed asynchronously.
 802                 if isinstance(v.owner, ArvadosFile):
 803                     v.owner.flush(sync=False)
 804
 805         with self.lock:
 806             if self._put_queue is not None:
 807                 self._put_queue.join()
 808
 809                 err = []
 810                 for k,v in items:
 811                     if v.state() == _BufferBlock.ERROR:
 812                         err.append((v.locator(), v.error))
 813                 if err:
 814                     raise KeepWriteError("Error writing some blocks", err, label="block")
 815
 816         for k,v in items:
 817             # flush again with sync=True to remove committed bufferblocks from
 818             # the segments.
 819             if v.owner:
 820                 if isinstance(v.owner, ArvadosFile):
 821                     v.owner.flush(sync=True)
 822                 elif isinstance(v.owner, list) and len(v.owner) > 0:
 823                     # This bufferblock is referenced by many files as a result
 824                     # of repacking small blocks, so don't delete it when flushing
 825                     # its owners, just do it after flushing them all.
 826                     for owner in v.owner:
 827                         owner.flush(sync=True)
 828                     self.delete_bufferblock(k)
 829
 830     def block_prefetch(self, locator):
 831         """Initiate a background download of a block.
 832
 833         This assumes that the underlying KeepClient implements a block cache,
 834         so repeated requests for the same block will not result in repeated
 835         downloads (unless the block is evicted from the cache.)  This method
 836         does not block.
 837
 838         """
 839
 840         if not self.prefetch_enabled:
 841             return
 842
 843         if self._keep.get_from_cache(locator) is not None:
 844             return
 845
 846         with self.lock:
 847             if locator in self._bufferblocks:
 848                 return
 849
 850         self.start_get_threads()
 851         self._prefetch_queue.put(locator)
 852
 853
 854 class ArvadosFile(object):
 855     """Represent a file in a Collection.
 856
 857     ArvadosFile manages the underlying representation of a file in Keep as a
 858     sequence of segments spanning a set of blocks, and implements random
 859     read/write access.
 860
 861     This object may be accessed from multiple threads.
 862
 863     """
 864
 865     __slots__ = ('parent', 'name', '_writers', '_committed',
 866                  '_segments', 'lock', '_current_bblock', 'fuse_entry')
 867
 868     def __init__(self, parent, name, stream=[], segments=[]):
 869         """
 870         ArvadosFile constructor.
 871
 872         :stream:
 873           a list of Range objects representing a block stream
 874
 875         :segments:
 876           a list of Range objects representing segments
 877         """
 878         self.parent = parent
 879         self.name = name
 880         self._writers = set()
 881         self._committed = False
 882         self._segments = []
 883         self.lock = parent.root_collection().lock
 884         for s in segments:
 885             self._add_segment(stream, s.locator, s.range_size)
 886         self._current_bblock = None
 887
 888     def writable(self):
 889         return self.parent.writable()
 890
 891     @synchronized
 892     def permission_expired(self, as_of_dt=None):
 893         """Returns True if any of the segment's locators is expired"""
 894         for r in self._segments:
 895             if KeepLocator(r.locator).permission_expired(as_of_dt):
 896                 return True
 897         return False
 898
 899     @synchronized
 900     def has_remote_blocks(self):
 901         """Returns True if any of the segment's locators has a +R signature"""
 902
 903         for s in self._segments:
 904             if '+R' in s.locator:
 905                 return True
 906         return False
 907
 908     @synchronized
 909     def _copy_remote_blocks(self, remote_blocks={}):
 910         """Ask Keep to copy remote blocks and point to their local copies.
 911
 912         This is called from the parent Collection.
 913
 914         :remote_blocks:
 915             Shared cache of remote to local block mappings. This is used to avoid
 916             doing extra work when blocks are shared by more than one file in
 917             different subdirectories.
 918         """
 919
 920         for s in self._segments:
 921             if '+R' in s.locator:
 922                 try:
 923                     loc = remote_blocks[s.locator]
 924                 except KeyError:
 925                     loc = self.parent._my_keep().refresh_signature(s.locator)
 926                     remote_blocks[s.locator] = loc
 927                 s.locator = loc
 928                 self.parent.set_committed(False)
 929         return remote_blocks
 930
 931     @synchronized
 932     def segments(self):
 933         return copy.copy(self._segments)
 934
 935     @synchronized
 936     def clone(self, new_parent, new_name):
 937         """Make a copy of this file."""
 938         cp = ArvadosFile(new_parent, new_name)
 939         cp.replace_contents(self)
 940         return cp
 941
 942     @must_be_writable
 943     @synchronized
 944     def replace_contents(self, other):
 945         """Replace segments of this file with segments from another `ArvadosFile` object."""
 946
 947         map_loc = {}
 948         self._segments = []
 949         for other_segment in other.segments():
 950             new_loc = other_segment.locator
 951             if other.parent._my_block_manager().is_bufferblock(other_segment.locator):
 952                 if other_segment.locator not in map_loc:
 953                     bufferblock = other.parent._my_block_manager().get_bufferblock(other_segment.locator)
 954                     if bufferblock.state() != _BufferBlock.WRITABLE:
 955                         map_loc[other_segment.locator] = bufferblock.locator()
 956                     else:
 957                         map_loc[other_segment.locator] = self.parent._my_block_manager().dup_block(bufferblock, self).blockid
 958                 new_loc = map_loc[other_segment.locator]
 959
 960             self._segments.append(Range(new_loc, other_segment.range_start, other_segment.range_size, other_segment.segment_offset))
 961
 962         self.set_committed(False)
 963
 964     def __eq__(self, other):
 965         if other is self:
 966             return True
 967         if not isinstance(other, ArvadosFile):
 968             return False
 969
 970         othersegs = other.segments()
 971         with self.lock:
 972             if len(self._segments) != len(othersegs):
 973                 return False
 974             for i in range(0, len(othersegs)):
 975                 seg1 = self._segments[i]
 976                 seg2 = othersegs[i]
 977                 loc1 = seg1.locator
 978                 loc2 = seg2.locator
 979
 980                 if self.parent._my_block_manager().is_bufferblock(loc1):
 981                     loc1 = self.parent._my_block_manager().get_bufferblock(loc1).locator()
 982
 983                 if other.parent._my_block_manager().is_bufferblock(loc2):
 984                     loc2 = other.parent._my_block_manager().get_bufferblock(loc2).locator()
 985
 986                 if (KeepLocator(loc1).stripped() != KeepLocator(loc2).stripped() or
 987                     seg1.range_start != seg2.range_start or
 988                     seg1.range_size != seg2.range_size or
 989                     seg1.segment_offset != seg2.segment_offset):
 990                     return False
 991
 992         return True
 993
 994     def __ne__(self, other):
 995         return not self.__eq__(other)
 996
 997     @synchronized
 998     def set_segments(self, segs):
 999         self._segments = segs
1000
1001     @synchronized
1002     def set_committed(self, value=True):
1003         """Set committed flag.
1004
1005         If value is True, set committed to be True.
1006
1007         If value is False, set committed to be False for this and all parents.
1008         """
1009         if value == self._committed:
1010             return
1011         self._committed = value
1012         if self._committed is False and self.parent is not None:
1013             self.parent.set_committed(False)
1014
1015     @synchronized
1016     def committed(self):
1017         """Get whether this is committed or not."""
1018         return self._committed
1019
1020     @synchronized
1021     def add_writer(self, writer):
1022         """Add an ArvadosFileWriter reference to the list of writers"""
1023         if isinstance(writer, ArvadosFileWriter):
1024             self._writers.add(writer)
1025
1026     @synchronized
1027     def remove_writer(self, writer, flush):
1028         """
1029         Called from ArvadosFileWriter.close(). Remove a writer reference from the list
1030         and do some block maintenance tasks.
1031         """
1032         self._writers.remove(writer)
1033
1034         if flush or self.size() > config.KEEP_BLOCK_SIZE // 2:
1035             # File writer closed, not small enough for repacking
1036             self.flush()
1037         elif self.closed():
1038             # All writers closed and size is adequate for repacking
1039             self.parent._my_block_manager().repack_small_blocks(closed_file_size=self.size())
1040
1041     def closed(self):
1042         """
1043         Get whether this is closed or not. When the writers list is empty, the file
1044         is supposed to be closed.
1045         """
1046         return len(self._writers) == 0
1047
1048     @must_be_writable
1049     @synchronized
1050     def truncate(self, size):
1051         """Shrink or expand the size of the file.
1052
1053         If `size` is less than the size of the file, the file contents after
1054         `size` will be discarded.  If `size` is greater than the current size
1055         of the file, it will be filled with zero bytes.
1056
1057         """
1058         if size < self.size():
1059             new_segs = []
1060             for r in self._segments:
1061                 range_end = r.range_start+r.range_size
1062                 if r.range_start >= size:
1063                     # segment is past the trucate size, all done
1064                     break
1065                 elif size < range_end:
1066                     nr = Range(r.locator, r.range_start, size - r.range_start, 0)
1067                     nr.segment_offset = r.segment_offset
1068                     new_segs.append(nr)
1069                     break
1070                 else:
1071                     new_segs.append(r)
1072
1073             self._segments = new_segs
1074             self.set_committed(False)
1075         elif size > self.size():
1076             padding = self.parent._my_block_manager().get_padding_block()
1077             diff = size - self.size()
1078             while diff > config.KEEP_BLOCK_SIZE:
1079                 self._segments.append(Range(padding.blockid, self.size(), config.KEEP_BLOCK_SIZE, 0))
1080                 diff -= config.KEEP_BLOCK_SIZE
1081             if diff > 0:
1082                 self._segments.append(Range(padding.blockid, self.size(), diff, 0))
1083             self.set_committed(False)
1084         else:
1085             # size == self.size()
1086             pass
1087
1088     def readfrom(self, offset, size, num_retries, exact=False):
1089         """Read up to `size` bytes from the file starting at `offset`.
1090
1091         :exact:
1092          If False (default), return less data than requested if the read
1093          crosses a block boundary and the next block isn't cached.  If True,
1094          only return less data than requested when hitting EOF.
1095         """
1096
1097         with self.lock:
1098             if size == 0 or offset >= self.size():
1099                 return b''
1100             readsegs = locators_and_ranges(self._segments, offset, size)
1101             prefetch = locators_and_ranges(self._segments, offset + size, config.KEEP_BLOCK_SIZE, limit=32)
1102
1103         locs = set()
1104         data = []
1105         for lr in readsegs:
1106             block = self.parent._my_block_manager().get_block_contents(lr.locator, num_retries=num_retries, cache_only=(bool(data) and not exact))
1107             if block:
1108                 blockview = memoryview(block)
1109                 data.append(blockview[lr.segment_offset:lr.segment_offset+lr.segment_size].tobytes())
1110                 locs.add(lr.locator)
1111             else:
1112                 break
1113
1114         for lr in prefetch:
1115             if lr.locator not in locs:
1116                 self.parent._my_block_manager().block_prefetch(lr.locator)
1117                 locs.add(lr.locator)
1118
1119         return b''.join(data)
1120
1121     @must_be_writable
1122     @synchronized
1123     def writeto(self, offset, data, num_retries):
1124         """Write `data` to the file starting at `offset`.
1125
1126         This will update existing bytes and/or extend the size of the file as
1127         necessary.
1128
1129         """
1130         if not isinstance(data, bytes) and not isinstance(data, memoryview):
1131             data = data.encode()
1132         if len(data) == 0:
1133             return
1134
1135         if offset > self.size():
1136             self.truncate(offset)
1137
1138         if len(data) > config.KEEP_BLOCK_SIZE:
1139             # Chunk it up into smaller writes
1140             n = 0
1141             dataview = memoryview(data)
1142             while n < len(data):
1143                 self.writeto(offset+n, dataview[n:n + config.KEEP_BLOCK_SIZE].tobytes(), num_retries)
1144                 n += config.KEEP_BLOCK_SIZE
1145             return
1146
1147         self.set_committed(False)
1148
1149         if self._current_bblock is None or self._current_bblock.state() != _BufferBlock.WRITABLE:
1150             self._current_bblock = self.parent._my_block_manager().alloc_bufferblock(owner=self)
1151
1152         if (self._current_bblock.size() + len(data)) > config.KEEP_BLOCK_SIZE:
1153             self._current_bblock.repack_writes()
1154             if (self._current_bblock.size() + len(data)) > config.KEEP_BLOCK_SIZE:
1155                 self.parent._my_block_manager().commit_bufferblock(self._current_bblock, sync=False)
1156                 self._current_bblock = self.parent._my_block_manager().alloc_bufferblock(owner=self)
1157
1158         self._current_bblock.append(data)
1159
1160         replace_range(self._segments, offset, len(data), self._current_bblock.blockid, self._current_bblock.write_pointer - len(data))
1161
1162         self.parent.notify(WRITE, self.parent, self.name, (self, self))
1163
1164         return len(data)
1165
1166     @synchronized
1167     def flush(self, sync=True, num_retries=0):
1168         """Flush the current bufferblock to Keep.
1169
1170         :sync:
1171           If True, commit block synchronously, wait until buffer block has been written.
1172           If False, commit block asynchronously, return immediately after putting block into
1173           the keep put queue.
1174         """
1175         if self.committed():
1176             return
1177
1178         if self._current_bblock and self._current_bblock.state() != _BufferBlock.COMMITTED:
1179             if self._current_bblock.state() == _BufferBlock.WRITABLE:
1180                 self._current_bblock.repack_writes()
1181             if self._current_bblock.state() != _BufferBlock.DELETED:
1182                 self.parent._my_block_manager().commit_bufferblock(self._current_bblock, sync=sync)
1183
1184         if sync:
1185             to_delete = set()
1186             for s in self._segments:
1187                 bb = self.parent._my_block_manager().get_bufferblock(s.locator)
1188                 if bb:
1189                     if bb.state() != _BufferBlock.COMMITTED:
1190                         self.parent._my_block_manager().commit_bufferblock(bb, sync=True)
1191                     to_delete.add(s.locator)
1192                     s.locator = bb.locator()
1193             for s in to_delete:
1194                 # Don't delete the bufferblock if it's owned by many files. It'll be
1195                 # deleted after all of its owners are flush()ed.
1196                 if self.parent._my_block_manager().get_bufferblock(s).owner is self:
1197                     self.parent._my_block_manager().delete_bufferblock(s)
1198
1199         self.parent.notify(MOD, self.parent, self.name, (self, self))
1200
1201     @must_be_writable
1202     @synchronized
1203     def add_segment(self, blocks, pos, size):
1204         """Add a segment to the end of the file.
1205
1206         `pos` and `offset` reference a section of the stream described by
1207         `blocks` (a list of Range objects)
1208
1209         """
1210         self._add_segment(blocks, pos, size)
1211
1212     def _add_segment(self, blocks, pos, size):
1213         """Internal implementation of add_segment."""
1214         self.set_committed(False)
1215         for lr in locators_and_ranges(blocks, pos, size):
1216             last = self._segments[-1] if self._segments else Range(0, 0, 0, 0)
1217             r = Range(lr.locator, last.range_start+last.range_size, lr.segment_size, lr.segment_offset)
1218             self._segments.append(r)
1219
1220     @synchronized
1221     def size(self):
1222         """Get the file size."""
1223         if self._segments:
1224             n = self._segments[-1]
1225             return n.range_start + n.range_size
1226         else:
1227             return 0
1228
1229     @synchronized
1230     def manifest_text(self, stream_name=".", portable_locators=False,
1231                       normalize=False, only_committed=False):
1232         buf = ""
1233         filestream = []
1234         for segment in self._segments:
1235             loc = segment.locator
1236             if self.parent._my_block_manager().is_bufferblock(loc):
1237                 if only_committed:
1238                     continue
1239                 loc = self.parent._my_block_manager().get_bufferblock(loc).locator()
1240             if portable_locators:
1241                 loc = KeepLocator(loc).stripped()
1242             filestream.append(LocatorAndRange(loc, KeepLocator(loc).size,
1243                                  segment.segment_offset, segment.range_size))
1244         buf += ' '.join(normalize_stream(stream_name, {self.name: filestream}))
1245         buf += "\n"
1246         return buf
1247
1248     @must_be_writable
1249     @synchronized
1250     def _reparent(self, newparent, newname):
1251         self.set_committed(False)
1252         self.flush(sync=True)
1253         self.parent.remove(self.name)
1254         self.parent = newparent
1255         self.name = newname
1256         self.lock = self.parent.root_collection().lock
1257
1258
1259 class ArvadosFileReader(ArvadosFileReaderBase):
1260     """Wraps ArvadosFile in a file-like object supporting reading only.
1261
1262     Be aware that this class is NOT thread safe as there is no locking around
1263     updating file pointer.
1264
1265     """
1266
1267     def __init__(self, arvadosfile, mode="r", num_retries=None):
1268         super(ArvadosFileReader, self).__init__(arvadosfile.name, mode=mode, num_retries=num_retries)
1269         self.arvadosfile = arvadosfile
1270
1271     def size(self):
1272         return self.arvadosfile.size()
1273
1274     def stream_name(self):
1275         return self.arvadosfile.parent.stream_name()
1276
1277     def readinto(self, b):
1278         data = self.read(len(b))
1279         b[:len(data)] = data
1280         return len(data)
1281
1282     @_FileLikeObjectBase._before_close
1283     @retry_method
1284     def read(self, size=None, num_retries=None):
1285         """Read up to `size` bytes from the file and return the result.
1286
1287         Starts at the current file position.  If `size` is None, read the
1288         entire remainder of the file.
1289         """
1290         if size is None:
1291             data = []
1292             rd = self.arvadosfile.readfrom(self._filepos, config.KEEP_BLOCK_SIZE, num_retries)
1293             while rd:
1294                 data.append(rd)
1295                 self._filepos += len(rd)
1296                 rd = self.arvadosfile.readfrom(self._filepos, config.KEEP_BLOCK_SIZE, num_retries)
1297             return b''.join(data)
1298         else:
1299             data = self.arvadosfile.readfrom(self._filepos, size, num_retries, exact=True)
1300             self._filepos += len(data)
1301             return data
1302
1303     @_FileLikeObjectBase._before_close
1304     @retry_method
1305     def readfrom(self, offset, size, num_retries=None):
1306         """Read up to `size` bytes from the stream, starting at the specified file offset.
1307
1308         This method does not change the file position.
1309         """
1310         return self.arvadosfile.readfrom(offset, size, num_retries)
1311
1312     def flush(self):
1313         pass
1314
1315
1316 class ArvadosFileWriter(ArvadosFileReader):
1317     """Wraps ArvadosFile in a file-like object supporting both reading and writing.
1318
1319     Be aware that this class is NOT thread safe as there is no locking around
1320     updating file pointer.
1321
1322     """
1323
1324     def __init__(self, arvadosfile, mode, num_retries=None):
1325         super(ArvadosFileWriter, self).__init__(arvadosfile, mode=mode, num_retries=num_retries)
1326         self.arvadosfile.add_writer(self)
1327
1328     def writable(self):
1329         return True
1330
1331     @_FileLikeObjectBase._before_close
1332     @retry_method
1333     def write(self, data, num_retries=None):
1334         if self.mode[0] == "a":
1335             self._filepos = self.size()
1336         self.arvadosfile.writeto(self._filepos, data, num_retries)
1337         self._filepos += len(data)
1338         return len(data)
1339
1340     @_FileLikeObjectBase._before_close
1341     @retry_method
1342     def writelines(self, seq, num_retries=None):
1343         for s in seq:
1344             self.write(s, num_retries=num_retries)
1345
1346     @_FileLikeObjectBase._before_close
1347     def truncate(self, size=None):
1348         if size is None:
1349             size = self._filepos
1350         self.arvadosfile.truncate(size)
1351
1352     @_FileLikeObjectBase._before_close
1353     def flush(self):
1354         self.arvadosfile.flush()
1355
1356     def close(self, flush=True):
1357         if not self.closed:
1358             self.arvadosfile.remove_writer(self, flush)
1359             super(ArvadosFileWriter, self).close()
1360
1361
1362 class WrappableFile(object):
1363     """An interface to an Arvados file that's compatible with io wrappers.
1364
1365     """
1366     def __init__(self, f):
1367         self.f = f
1368         self.closed = False
1369     def close(self):
1370         self.closed = True
1371         return self.f.close()
1372     def flush(self):
1373         return self.f.flush()
1374     def read(self, *args, **kwargs):
1375         return self.f.read(*args, **kwargs)
1376     def readable(self):
1377         return self.f.readable()
1378     def readinto(self, *args, **kwargs):
1379         return self.f.readinto(*args, **kwargs)
1380     def seek(self, *args, **kwargs):
1381         return self.f.seek(*args, **kwargs)
1382     def seekable(self):
1383         return self.f.seekable()
1384     def tell(self):
1385         return self.f.tell()
1386     def writable(self):
1387         return self.f.writable()
1388     def write(self, *args, **kwargs):
1389         return self.f.write(*args, **kwargs)