os/bluestore/BlueFS: do not op_file_update deleted files

[liewegas]

Make truncate and preallocate no-op if the file has been
deleted. Otherwise we may get an op_file_update after
the op_file_remove in the log, confusing replay.

Signed-off-by: Sage Weil sage@redhat.com

[markhpc]

Checks are good, but it feels like we might be wallpapering over the issue. :)

[liewegas]

It's perfectly legal to operate on an unlinked file in unix/linux, so it's not surprising to me that rocksdb is doing this.

[varadakari]

Just a dumb question, If we unlink a file, don't file descriptor should be invalidated? Technically file no longer exist. Should we report the user, that file is already deleted, by returning an error?

[tchaikov]

If we unlink a file, don't file descriptor should be invalidated?

@varadakari no, it is still valid.

Technically file no longer exist. Should we report the user, that file is already deleted, by returning an error?

no, it's fine. like sage put, it's a common programming practice.

[liewegas]

The 'user' is rocksdb. It is probably doing something like

unlink(name);
delete handle;
flush(this);

or some other cleanup on file close that happens even when the file is
unlinked. It's probably harmless (as it is with a normal posix
unix-like file system).. I don't think we need to worry about.