1 # Copyright (C) The Arvados Authors. All rights reserved.
3 # SPDX-License-Identifier: Apache-2.0
16 _logger = logging.getLogger('arvados.keep')
18 cacheblock_suffix = ".keepcacheblock"
20 class DiskCacheSlot(object):
21 __slots__ = ("locator", "ready", "content", "cachedir", "filehandle")
23 def __init__(self, locator, cachedir):
24 self.locator = locator
25 self.ready = threading.Event()
27 self.cachedir = cachedir
28 self.filehandle = None
43 # Can't mmap a 0 length file
48 if self.content is not None:
49 # Has been set already
53 blockdir = os.path.join(self.cachedir, self.locator[0:3])
54 os.makedirs(blockdir, mode=0o700, exist_ok=True)
56 final = os.path.join(blockdir, self.locator) + cacheblock_suffix
58 self.filehandle = tempfile.NamedTemporaryFile(dir=blockdir, delete=False, prefix="tmp", suffix=cacheblock_suffix)
59 tmpfile = self.filehandle.name
60 os.chmod(tmpfile, stat.S_IRUSR | stat.S_IWUSR)
62 # aquire a shared lock, this tells other processes that
63 # we're using this block and to please not delete it.
64 fcntl.flock(self.filehandle, fcntl.LOCK_SH)
66 self.filehandle.write(value)
67 self.filehandle.flush()
68 os.rename(tmpfile, final)
71 self.content = mmap.mmap(self.filehandle.fileno(), 0, access=mmap.ACCESS_READ)
72 # only set the event when mmap is successful
75 if tmpfile is not None:
76 # If the tempfile hasn't been renamed on disk yet, try to delete it.
83 if self.content is None:
86 return len(self.content)
89 if self.content is not None and len(self.content) > 0:
90 # The mmap region might be in use when we decided to evict
91 # it. This can happen if the cache is too small.
93 # If we call close() now, it'll throw an error if
94 # something tries to access it.
96 # However, we don't need to explicitly call mmap.close()
98 # I confirmed in mmapmodule.c that that both close
99 # and deallocate do the same thing:
101 # a) close the file descriptor
102 # b) unmap the memory range
104 # So we can forget it in the cache and delete the file on
105 # disk, and it will tear it down after any other
106 # lingering Python references to the mapped memory are
109 blockdir = os.path.join(self.cachedir, self.locator[0:3])
110 final = os.path.join(blockdir, self.locator) + cacheblock_suffix
112 fcntl.flock(self.filehandle, fcntl.LOCK_UN)
114 # try to get an exclusive lock, this ensures other
115 # processes are not using the block. It is
116 # nonblocking and will throw an exception if we
117 # can't get it, which is fine because that means
118 # we just won't try to delete it.
120 # I should note here, the file locking is not
121 # strictly necessary, we could just remove it and
122 # the kernel would ensure that the underlying
123 # inode remains available as long as other
124 # processes still have the file open. However, if
125 # you have multiple processes sharing the cache
126 # and deleting each other's files, you'll end up
127 # with a bunch of ghost files that don't show up
128 # in the file system but are still taking up
129 # space, which isn't particularly user friendly.
130 # The locking strategy ensures that cache blocks
131 # in use remain visible.
133 fcntl.flock(self.filehandle, fcntl.LOCK_EX | fcntl.LOCK_NB)
140 self.filehandle = None
145 def get_from_disk(locator, cachedir):
146 blockdir = os.path.join(cachedir, locator[0:3])
147 final = os.path.join(blockdir, locator) + cacheblock_suffix
150 filehandle = open(final, "rb")
152 # aquire a shared lock, this tells other processes that
153 # we're using this block and to please not delete it.
154 fcntl.flock(filehandle, fcntl.LOCK_SH)
156 content = mmap.mmap(filehandle.fileno(), 0, access=mmap.ACCESS_READ)
157 dc = DiskCacheSlot(locator, cachedir)
158 dc.filehandle = filehandle
162 except FileNotFoundError:
164 except Exception as e:
165 traceback.print_exc()
170 def cache_usage(cachedir):
172 for root, dirs, files in os.walk(cachedir):
174 if not name.endswith(cacheblock_suffix):
177 blockpath = os.path.join(root, name)
178 res = os.stat(blockpath)
184 def init_cache(cachedir, maxslots):
186 # First check the disk cache works at all by creating a 1 byte cache entry
188 checkexists = DiskCacheSlot.get_from_disk('0cc175b9c0f1b6a831c399e269772661', cachedir)
189 ds = DiskCacheSlot('0cc175b9c0f1b6a831c399e269772661', cachedir)
191 if checkexists is None:
192 # Don't keep the test entry around unless it existed beforehand.
195 # map in all the files in the cache directory, up to max slots.
196 # after max slots, try to delete the excess blocks.
198 # this gives the calling process ownership of all the blocks
201 for root, dirs, files in os.walk(cachedir):
203 if not name.endswith(cacheblock_suffix):
206 blockpath = os.path.join(root, name)
207 res = os.stat(blockpath)
209 if len(name) == (32+len(cacheblock_suffix)) and not name.startswith("tmp"):
210 blocks.append((name[0:32], res.st_atime))
211 elif name.startswith("tmp") and ((time.time() - res.st_mtime) > 60):
212 # found a temporary file more than 1 minute old,
219 # sort by access time (atime), going from most recently
220 # accessed (highest timestamp) to least recently accessed
221 # (lowest timestamp).
222 blocks.sort(key=lambda x: x[1], reverse=True)
224 # Map in all the files we found, up to maxslots, if we exceed
225 # maxslots, start throwing things out.
228 got = DiskCacheSlot.get_from_disk(b[0], cachedir)
231 if len(cachelist) < maxslots:
232 cachelist.append(got)
234 # we found more blocks than maxslots, try to
235 # throw it out of the cache.