+# Copyright (C) The Arvados Authors. All rights reserved.
+#
+# SPDX-License-Identifier: Apache-2.0
+
from __future__ import absolute_import
from __future__ import division
from future import standard_library
+from future.utils import listitems, listvalues
standard_library.install_aliases()
from builtins import range
from builtins import object
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 ._ranges import locators_and_ranges, replace_range, Range, LocatorAndRange
from .retry import retry_method
MOD = "mod"
pos += self._filepos
elif whence == os.SEEK_END:
pos += self.size()
- self._filepos = min(max(pos, 0), self.size())
+ if pos < 0:
+ raise IOError(errno.EINVAL, "Tried to seek to negative file offset.")
+ self._filepos = pos
+ return self._filepos
def tell(self):
return self._filepos
+ def readable(self):
+ return True
+
+ def writable(self):
+ return False
+
+ def seekable(self):
+ return True
+
@_FileLikeObjectBase._before_close
@retry_method
def readall(self, size=2**20, num_retries=None):
return b''.join(data).decode().splitlines(True)
def size(self):
- raise NotImplementedError()
+ raise IOError(errno.ENOSYS, "Not implemented")
def read(self, size, num_retries=None):
- raise NotImplementedError()
+ raise IOError(errno.ENOSYS, "Not implemented")
def readfrom(self, start, size, num_retries=None):
- raise NotImplementedError()
+ raise IOError(errno.ENOSYS, "Not implemented")
class StreamFileReader(ArvadosFileReaderBase):
PENDING = 1
COMMITTED = 2
ERROR = 3
+ DELETED = 4
def __init__(self, blockid, starting_capacity, owner):
"""
@synchronized
def clear(self):
+ self._state = _BufferBlock.DELETED
self.owner = None
self.buffer_block = None
self.buffer_view = None
+ @synchronized
+ def repack_writes(self):
+ """Optimize buffer block by repacking segments in file sequence.
+
+ When the client makes random writes, they appear in the buffer block in
+ the sequence they were written rather than the sequence they appear in
+ the file. This makes for inefficient, fragmented manifests. Attempt
+ to optimize by repacking writes in file sequence.
+
+ """
+ if self._state != _BufferBlock.WRITABLE:
+ raise AssertionError("Cannot repack non-writable block")
+
+ segs = self.owner.segments()
+
+ # Collect the segments that reference the buffer block.
+ bufferblock_segs = [s for s in segs if s.locator == self.blockid]
+
+ # Collect total data referenced by segments (could be smaller than
+ # bufferblock size if a portion of the file was written and
+ # then overwritten).
+ write_total = sum([s.range_size for s in bufferblock_segs])
+
+ if write_total < self.size() or len(bufferblock_segs) > 1:
+ # If there's more than one segment referencing this block, it is
+ # due to out-of-order writes and will produce a fragmented
+ # manifest, so try to optimize by re-packing into a new buffer.
+ contents = self.buffer_view[0:self.write_pointer].tobytes()
+ new_bb = _BufferBlock(None, write_total, None)
+ for t in bufferblock_segs:
+ new_bb.append(contents[t.segment_offset:t.segment_offset+t.range_size])
+ t.segment_offset = new_bb.size() - t.range_size
+
+ self.buffer_block = new_bb.buffer_block
+ self.buffer_view = new_bb.buffer_view
+ self.write_pointer = new_bb.write_pointer
+ self._locator = None
+ new_bb.clear()
+ self.owner.set_segments(segs)
+
+ def __repr__(self):
+ return "<BufferBlock %s>" % (self.blockid)
+
class NoopLock(object):
def __enter__(self):
self.copies = copies
self._pending_write_size = 0
self.threads_lock = threading.Lock()
+ self.padding_block = None
@synchronized
def alloc_bufferblock(self, blockid=None, starting_capacity=2**14, owner=None):
def _alloc_bufferblock(self, blockid=None, starting_capacity=2**14, owner=None):
if blockid is None:
- blockid = "%s" % uuid.uuid4()
+ blockid = str(uuid.uuid4())
bufferblock = _BufferBlock(blockid, starting_capacity=starting_capacity, owner=owner)
self._bufferblocks[bufferblock.blockid] = bufferblock
return bufferblock
ArvadosFile that owns the new block
"""
- new_blockid = "bufferblock%i" % len(self._bufferblocks)
+ new_blockid = str(uuid.uuid4())
bufferblock = block.clone(new_blockid, owner)
self._bufferblocks[bufferblock.blockid] = bufferblock
return bufferblock
else:
loc = self._keep.put(bufferblock.buffer_view[0:bufferblock.write_pointer].tobytes(), copies=self.copies)
bufferblock.set_state(_BufferBlock.COMMITTED, loc)
-
except Exception as e:
bufferblock.set_state(_BufferBlock.ERROR, e)
finally:
@synchronized
def repack_small_blocks(self, force=False, sync=False, closed_file_size=0):
"""Packs small blocks together before uploading"""
+
self._pending_write_size += closed_file_size
# Check if there are enough small blocks for filling up one in full
- if force or (self._pending_write_size >= config.KEEP_BLOCK_SIZE):
+ if not (force or (self._pending_write_size >= config.KEEP_BLOCK_SIZE)):
+ return
- # Search blocks ready for getting packed together before being committed to Keep.
- # A WRITABLE block always has an owner.
- # A WRITABLE block with its owner.closed() implies that it's
- # size is <= KEEP_BLOCK_SIZE/2.
- try:
- small_blocks = [b for b in list(self._bufferblocks.values()) if b.state() == _BufferBlock.WRITABLE and b.owner.closed()]
- except AttributeError:
- # Writable blocks without owner shouldn't exist.
- raise UnownedBlockError()
+ # Search blocks ready for getting packed together before being
+ # committed to Keep.
+ # A WRITABLE block always has an owner.
+ # A WRITABLE block with its owner.closed() implies that its
+ # size is <= KEEP_BLOCK_SIZE/2.
+ try:
+ small_blocks = [b for b in listvalues(self._bufferblocks)
+ if b.state() == _BufferBlock.WRITABLE and b.owner.closed()]
+ except AttributeError:
+ # Writable blocks without owner shouldn't exist.
+ raise UnownedBlockError()
+
+ if len(small_blocks) <= 1:
+ # Not enough small blocks for repacking
+ return
- if len(small_blocks) <= 1:
- # Not enough small blocks for repacking
- return
+ for bb in small_blocks:
+ bb.repack_writes()
- # Update the pending write size count with its true value, just in case
- # some small file was opened, written and closed several times.
- self._pending_write_size = sum([b.size() for b in small_blocks])
- if self._pending_write_size < config.KEEP_BLOCK_SIZE and not force:
- return
+ # Update the pending write size count with its true value, just in case
+ # some small file was opened, written and closed several times.
+ self._pending_write_size = sum([b.size() for b in small_blocks])
+
+ if self._pending_write_size < config.KEEP_BLOCK_SIZE and not force:
+ return
- new_bb = self._alloc_bufferblock()
- while len(small_blocks) > 0 and (new_bb.write_pointer + small_blocks[0].size()) <= config.KEEP_BLOCK_SIZE:
- bb = small_blocks.pop(0)
- arvfile = bb.owner
- self._pending_write_size -= bb.size()
- new_bb.append(bb.buffer_view[0:bb.write_pointer].tobytes())
- arvfile.set_segments([Range(new_bb.blockid,
- 0,
- bb.size(),
- new_bb.write_pointer - bb.size())])
- self._delete_bufferblock(bb.blockid)
- self.commit_bufferblock(new_bb, sync=sync)
+ new_bb = self._alloc_bufferblock()
+ new_bb.owner = []
+ files = []
+ while len(small_blocks) > 0 and (new_bb.write_pointer + small_blocks[0].size()) <= config.KEEP_BLOCK_SIZE:
+ bb = small_blocks.pop(0)
+ new_bb.owner.append(bb.owner)
+ self._pending_write_size -= bb.size()
+ new_bb.append(bb.buffer_view[0:bb.write_pointer].tobytes())
+ files.append((bb, new_bb.write_pointer - bb.size()))
+
+ self.commit_bufferblock(new_bb, sync=sync)
+
+ for bb, new_bb_segment_offset in files:
+ newsegs = bb.owner.segments()
+ for s in newsegs:
+ if s.locator == bb.blockid:
+ s.locator = new_bb.blockid
+ s.segment_offset = new_bb_segment_offset+s.segment_offset
+ bb.owner.set_segments(newsegs)
+ self._delete_bufferblock(bb.blockid)
def commit_bufferblock(self, block, sync):
"""Initiate a background upload of a bufferblock.
def get_bufferblock(self, locator):
return self._bufferblocks.get(locator)
+ @synchronized
+ def get_padding_block(self):
+ """Get a bufferblock 64 MB in size consisting of all zeros, used as padding
+ when using truncate() to extend the size of a file.
+
+ For reference (and possible future optimization), the md5sum of the
+ padding block is: 7f614da9329cd3aebf59b91aadc30bf0+67108864
+
+ """
+
+ if self.padding_block is None:
+ self.padding_block = self._alloc_bufferblock(starting_capacity=config.KEEP_BLOCK_SIZE)
+ self.padding_block.write_pointer = config.KEEP_BLOCK_SIZE
+ self.commit_bufferblock(self.padding_block, False)
+ return self.padding_block
+
@synchronized
def delete_bufferblock(self, locator):
self._delete_bufferblock(locator)
self.repack_small_blocks(force=True, sync=True)
with self.lock:
- items = list(self._bufferblocks.items())
+ items = listitems(self._bufferblocks)
for k,v in items:
if v.state() != _BufferBlock.COMMITTED and v.owner:
- v.owner.flush(sync=False)
+ # Ignore blocks with a list of owners, as if they're not in COMMITTED
+ # state, they're already being committed asynchronously.
+ if isinstance(v.owner, ArvadosFile):
+ v.owner.flush(sync=False)
with self.lock:
if self._put_queue is not None:
# flush again with sync=True to remove committed bufferblocks from
# the segments.
if v.owner:
- v.owner.flush(sync=True)
+ if isinstance(v.owner, ArvadosFile):
+ v.owner.flush(sync=True)
+ elif isinstance(v.owner, list) and len(v.owner) > 0:
+ # This bufferblock is referenced by many files as a result
+ # of repacking small blocks, so don't delete it when flushing
+ # its owners, just do it after flushing them all.
+ for owner in v.owner:
+ owner.flush(sync=True)
+ self.delete_bufferblock(k)
def block_prefetch(self, locator):
"""Initiate a background download of a block.
"""
+ __slots__ = ('parent', 'name', '_writers', '_committed',
+ '_segments', 'lock', '_current_bblock', 'fuse_entry')
+
def __init__(self, parent, name, stream=[], segments=[]):
"""
ArvadosFile constructor.
@must_be_writable
@synchronized
def truncate(self, size):
- """Shrink the size of the file.
+ """Shrink or expand the size of the file.
If `size` is less than the size of the file, the file contents after
`size` will be discarded. If `size` is greater than the current size
- of the file, an IOError will be raised.
+ of the file, it will be filled with zero bytes.
"""
if size < self.size():
self._segments = new_segs
self.set_committed(False)
elif size > self.size():
- raise IOError(errno.EINVAL, "truncate() does not support extending the file size")
+ padding = self.parent._my_block_manager().get_padding_block()
+ diff = size - self.size()
+ while diff > config.KEEP_BLOCK_SIZE:
+ self._segments.append(Range(padding.blockid, self.size(), config.KEEP_BLOCK_SIZE, 0))
+ diff -= config.KEEP_BLOCK_SIZE
+ if diff > 0:
+ self._segments.append(Range(padding.blockid, self.size(), diff, 0))
+ self.set_committed(False)
+ else:
+ # size == self.size()
+ pass
def readfrom(self, offset, size, num_retries, exact=False):
"""Read up to `size` bytes from the file starting at `offset`.
return b''.join(data)
- 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
- and to avoid leaking information.
-
- """
- segs = self._segments
-
- # Sum up the segments to get the total bytes of the file referencing
- # into the buffer block.
- bufferblock_segs = [s for s in segs if s.locator == self._current_bblock.blockid]
- write_total = sum([s.range_size for s in bufferblock_segs])
-
- 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(contents[t.segment_offset:t.segment_offset+t.range_size])
- t.segment_offset = new_bb.size() - t.range_size
-
- self._current_bblock = new_bb
-
@must_be_writable
@synchronized
def writeto(self, offset, data, num_retries):
return
if offset > self.size():
- raise ArgumentError("Offset is past the end of the file")
+ self.truncate(offset)
if len(data) > config.KEEP_BLOCK_SIZE:
# Chunk it up into smaller writes
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(num_retries)
+ self._current_bblock.repack_writes()
if (self._current_bblock.size() + len(data)) > config.KEEP_BLOCK_SIZE:
self.parent._my_block_manager().commit_bufferblock(self._current_bblock, sync=False)
self._current_bblock = self.parent._my_block_manager().alloc_bufferblock(owner=self)
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)
+ self._current_bblock.repack_writes()
+ if self._current_bblock.state() != _BufferBlock.DELETED:
+ self.parent._my_block_manager().commit_bufferblock(self._current_bblock, sync=sync)
if sync:
to_delete = set()
to_delete.add(s.locator)
s.locator = bb.locator()
for s in to_delete:
- self.parent._my_block_manager().delete_bufferblock(s)
+ # Don't delete the bufferblock if it's owned by many files. It'll be
+ # deleted after all of its owners are flush()ed.
+ if self.parent._my_block_manager().get_bufferblock(s).owner is self:
+ self.parent._my_block_manager().delete_bufferblock(s)
self.parent.notify(MOD, self.parent, self.name, (self, self))
normalize=False, only_committed=False):
buf = ""
filestream = []
- for segment in self.segments:
+ for segment in self._segments:
loc = segment.locator
if self.parent._my_block_manager().is_bufferblock(loc):
if only_committed:
continue
- loc = self._bufferblocks[loc].calculate_locator()
+ loc = self.parent._my_block_manager().get_bufferblock(loc).locator()
if portable_locators:
loc = KeepLocator(loc).stripped()
- filestream.append(LocatorAndRange(loc, locator_block_size(loc),
+ filestream.append(LocatorAndRange(loc, KeepLocator(loc).size,
segment.segment_offset, segment.range_size))
- buf += ' '.join(normalize_stream(stream_name, {stream_name: filestream}))
+ buf += ' '.join(normalize_stream(stream_name, {self.name: filestream}))
buf += "\n"
return buf
super(ArvadosFileWriter, self).__init__(arvadosfile, mode=mode, num_retries=num_retries)
self.arvadosfile.add_writer(self)
+ def writable(self):
+ return True
+
@_FileLikeObjectBase._before_close
@retry_method
def write(self, data, num_retries=None):
if self.mode[0] == "a":
- self.arvadosfile.writeto(self.size(), data, num_retries)
- else:
- self.arvadosfile.writeto(self._filepos, data, num_retries)
- self._filepos += len(data)
+ self._filepos = self.size()
+ self.arvadosfile.writeto(self._filepos, data, num_retries)
+ self._filepos += len(data)
return len(data)
@_FileLikeObjectBase._before_close
if size is None:
size = self._filepos
self.arvadosfile.truncate(size)
- if self._filepos > self.size():
- self._filepos = self.size()
@_FileLikeObjectBase._before_close
def flush(self):