Re: [PATCH v5 04/13] csum-file: add CSUM_KEEP_OPEN flag

Derrick Stolee <stolee@gmail.com> writes:

Let's follow the commit-graph writing codepath to see what happens:

	fd = hold_lock_file_for_update(&lk, graph_name, 0);
	...
	f = hashfd(lk.tempfile->fd, lk.tempfile->filename.buf);

The caller creates a lockfile, and then wraps its file descriptor in a hashfile.

	hashwrite_be32(f, GRAPH_SIGNATURE);
	...

Then it goes on writing to the hashfile, growing the lockfile.

        ...
	write_graph_chunk_large_edges(f, commits.list, commits.nr);

	close_commit_graph();

And after writing all data out (oh by the way, why aren't we passing commit_graph instance around and instead relying on a file-scope static global?)...

	hashclose(f, final_hash, CSUM_CLOSE | CSUM_FSYNC | CSUM_KEEP_OPEN);

We ask for the final hash value to be written to the file (and also returned to us---although you do not seem to use that at all). See a comment on this, though, at the end.

	commit_lock_file(&lk);

And then, we put the lockfile to its final place, while closing its file descriptor.

The overall API sounds sensible, from the above.

However.

The function whose name is hashclose() that takes a flag word whose possible bit value includes "Please close this thing" feels strange enough (does it mean the hashclose() function does not close it if CSUM_CLOSE is not given?), but adding another to the mix that lets us say "Please close this (with or without FSYNC), oh by the way please leave it open" feels a bit borderline to insanity.

I _think_ the word "close" in the name hashclose() is about closing the (virtual) stream for the hashing that is overlayed on top of the underlying file descriptor, and being able to choose between closing and not closing the underlying file descriptor when "closing" the hashing layer sort of makes sense. So I won't complain too much about hashclose() that takes optional CSUM_CLOSE flag.

But then what does it mean to give KEEP_OPEN and CLOSE together?

The new caller (which is the only one that wants the nominally nonsensical CLOSE|KEEP_OPEN combination, which is shown above) wants the final checksum of the data sent over the (virtual) stream computed and written, and the file descriptor fsync'ed, but the file descriptor kept open. As we _DO_ want to keep the verbs in flags CSUM_CLOSE and CSUM_FSYNC to be about the underlying file descriptor, I think your new code for KEEP_OPEN that is inside the if() block that is for CSUM_CLOSE is an ugly hack, and your asking for improvements is very much appreciated.

Let's step back and see what different behaviours the existing code wants to support before your patch:

    - hashclose() is always about finializing the hash computation
      over the data sent through the struct hashfile (i.e. the
      virtual stream opened by hashfd()).  The optional *result can
      be used to receive this hash value, even when the caller does
      not want to write that hash value to the output stream.

    - when CSUM_CLOSE is given, however, the hash value is written
      out as the trailing record to the output stream and the stream
      is closed.  CSUM_FSYNC can instead be used to ensure that the
      data hits the disk platter when the output stream is closed.

    - when CSUM_CLOSE nor CSUM_FSYNC is not given, hash value is not
      written to the output stream (the caller takes responsibility
      of using *result), and the output stream is left open.

I think the first mistake in the existing code is to associate "close the underlying stream" and "write the hash out to the underlying stream" more closely than it should. It should be possible to "close the underlying steam" without first writing the hash out to the underlying stream", and vice versa.

IOW, I think

        hashclose() {
                hashflush();
                the_hash_algo->final_fn();
                if (result)             
                        hashcpy(result, f->buffer);
        +       if (flags & CSUM_HASH_IN_STREAM)
        +               flush(f, f->buffer, the_hash_algo->rawsz);
        +       if (flags & CSUM_FSYNC)
        +               fsync_or_die();
                if (flags & (CSUM_CLOSE | CSUM_FSYNC)) {
        -               flush();                
        -               if (flags & CSUM_FSYNC)
        -                       fsync_or_die();
                        if (close(f->fd))
                                die_errno();
                        fd = 0;
                } else
                        fd = f->fd;
                if (0 <= f->check_fd) {
                        ...
                }
                free(f);
                return fd;
        }

with would be a good first "preliminary preparation" step.

Existing callers that pass CSUM_FSYNC or CSUM_CLOSE now need to also say "I want the resulting hash in the output stream", but that allows your later caller to omit CSUM_CLOSE and then ask for HASH_IN_STREAM alone.

Existing callers can expect that FSYNC alone means fsync and close, but your caller wants hashclose() to compute the hash, write the hash to the output stream, and fsync the output stream, and return without closing the output stream. For that, you'd make FSYNC not to imply CLOSE, and you'd need to vet all the existing callers that use FSYNC are OK with such a change. And then the above would become

        hashclose() {
                hashflush();
                the_hash_algo->final_fn();
                if (result)             
                        hashcpy(result, f->buffer);
                if (flags & CSUM_HASH_IN_STREAM)
                        flush(f, f->buffer, the_hash_algo->rawsz);
                if (flags & CSUM_FSYNC)
                        fsync_or_die();
                if (flags & CSUM_CLOSE) {
                        if (close(f->fd))
                                die_errno();
                        fd = 0;
                } else
                        fd = f->fd;
                if (0 <= f->check_fd) {
                        ...
                }
                free(f);
                return fd;
        }

Once we reach that state, the new caller in write_commit_graph() does not have to pass nonsensical CLOSE|KEEP_OPEN combination. Instead we can do

	hashclose(f, NULL, CSUM_HASH_IN_STREAM | CSUM_FSYNC);

or something like that, I would think, without having KEEP_OPEN.

I am actually wondering if it is worth making CSUM_FSYNC not imply CSUM_CLOSE. There aren't that many existing callers of hashclose() that uses FSYNC, so vetting all of them and replacing their FSYNC with (FSYNC|CLOSE) is not all that difficult, but if this new caller is an oddball then another strategy may be to do the fsync_or_die() on the caller side, something like:

                hashclose(f, NULL, CSUM_HASH_IN_STREAM);
        +       fsync_or_die(fd, get_lock_file_path(&lk));
                commit_lock_file(&lk);

And then we can keep the "FSYNC means fsync and then close" the current set of callers rely on. I dunno if that is a major issue, but I do think "close this, or no, keep it open" is far worse than "do we want the resulting hash in the stream?"

An alternative design of the above is without making CSUM_HASH_IN_STREAM a new flag bit. I highly suspect that the calling codepath _knows_ whether the resulting final hash will be written out at the end of the stream or not when it wraps an fd with a hashfile structure, so "struct hashfile" could gain a bit to tell hashclose() whether the resulting hash need to be written (or not). That would be a bit larger change than what I outlined above, and I do not know if it is worth doing, though.