Samuel Merritt 729430f349 Alternate DiskFile constructor for efficient auditing. ...

Before, to audit an object, the auditor:
 - calls listdir(object-hash-dir)
 - picks out the .data file from the listing
 - pulls out all N of its user.swift.metadata* xattrs
 - unpickles them
 - pulls out the value for 'name'
 - splits the name into a/c/o
 - then instantiates and opens a DiskFile(a, c, o),
which does the following
 - joins a/c/o back into a name
 - hashes the name
 - calls listdir(object-hash-dir) (AGAIN)
 - picks out the .data file (and maybe .meta) from the listing (AGAIN)
 - pulls out all N of its user.swift.metadata* xattrs (AGAIN)
 - unpickles them (AGAIN)
 - starts reading object's contents off disk

Now, the auditor simply locates the hash dir on the filesystem (saving
one listdir) and then hands it off to
DiskFileManager.get_diskfile_from_audit_location, which then
instantiates a DiskFile in a way that lazy-loads the name later
(saving one xattr reading).

As part of this, DiskFile.open() will now quarantine a hash
"directory" that's actually a file. Before, the audit location
generator would skip those, but now they make it clear into
DiskFile(). It's better to quarantine them anyway, as they're not
doing any good the way they are.

Also, removed the was_quarantined attribute on DiskFileReader. Now you
can pass in a quarantine_hook callable to DiskFile.reader() that gets
called if the file was quarantined. Default is to log quarantines, but
otherwise do nothing.

Change-Id: I04fc14569982a17fcc89e00832725ae71009335a

2013-11-06 11:16:42 -08:00

16 KiB

Raw Blame History

View Raw