On 05/06/2013 08:41 PM, Jeff King wrote:
> On Mon, May 06, 2013 at 02:03:40PM +0200, Michael Haggerty wrote:
>> The loose refs cache is only used by the for_each_ref() functions, not
>> for looking up individual references. Another approach would be to
>> change the top-level for_each_ref() functions to re-stat() all of the
>> loose references within the namespace that interests it, *then* verify
>> that the packed-ref cache is not stale, *then* start the iteration.
>> Then there would be no need to re-stat() files during the iteration.
>> (This would mean that we have to prohibit a second reference iteration
>> from being started while one is already in progress.)
> Hmm. Thinking on this more, I'm not sure that we need to stat the loose
> references at all. We do not need to care if the loose refs are up to
> date or not. Well, we might care, but the point here is not to pretend
> that we have an up-to-date atomic view of the loose refs; it is only to
> make sure that the fallback-to-packed behavior does not lie to us about
> the existence or value of a ref.
> IOW, it is OK to come up with a value for ref X that was true at the
> beginning of the program, even if it has been simultaneously updated.
> Our program can operate as if it happened in the instant it started,
> even though in real life it takes longer. But it is _not_ OK to miss the
> existence of a ref, or to come up with a value that it did not hold at
> some point during the program (e.g., it is not OK to return some cruft
> we wrote into the packed-refs file when we packed it three weeks ago).
This all sounds correct to me.
Another potential problem caused by the non-atomicity of loose reference
reading could be to confuse reachability tests if process 1 is reading
loose references while process 2 is renaming a reference:
1. Process 1 looks for refs/heads/aaaaa and finds it missing.
2. Process 2 renames zzzzz -> aaaaa
3. Process 1 looks for refs/heads/zzzzz and finds it missing.
Process 2 would think that any objects pointed to by aaaaa (formerly
zzzzz) are unreachable. This would be unfortunate if it is doing an
upload-pack and very bad if it is doing a gc. I wonder if this could be
a problem in practice? (Gee, wouldn't it be nice to keep reflogs for
deleted refs? :-) )
> That is a weaker guarantee, and I think we can provide it with:
> 1. Load all loose refs into cache for a particular enumeration.
> 2. Make sure the packed-refs cache is up-to-date (by checking its
> stat() information and reloading if necessary).
> 3. Run the usual iteration over the loose/packed ref caches.
This sounds reasonable, too, though I'll need some more time to digest
your suggestion in detail.
>> Of course, clearing (part of) the loose reference cache invalidates any
>> pointers that other callers might have retained to refnames in the old
>> version of the cache. I've never really investigated what callers might
>> hold onto such pointers under the assumption that they will live to the
>> end of the process.
> My proposal above gets rid of the need to invalidate the loose refs
> cache, so we can ignore that complexity.
The same concern applies to invalidating the packed-ref cache, which you
still want to do.
>> Given all of this trouble, there is an obvious question: why do we have
>> a loose reference cache in the first place? I think there are a few
>> 1. In case one git process has to iterate through the same part of the
>> reference namespace more than once. (Does this frequently happen?)
> I'm not sure how often it happens. There are a few obvious candidates if
> you "git grep 'for_each[a-z_]*ref'", but I'm not sure if the actual
> performance impact is measurable (the cache should be warm after the
> first run-through).
>> 2. Reading a bunch of loose references at the same time is more
>> efficient than reading them one by one interleaved with other file
>> reads. (I think this is a significant win.)
> Makes some sense, though I don't recall whether it was ever measured.
I think I measured it once and found it a significant benefit, though I
can't remember whether this was with a warm cache or only with a cold cache.
In fact, I experimented with some other strategies for loose reference
loading for performance reasons. Currently loose references are read
into the cache lazily, one directory at a time. I experimented with
changes in both directions:
* Preloading the whole tree of loose references before starting an
iteration. As I recall, this was a performance *win*. It was on my
to-do list of things to pursue when I have some free time (ha, ha). I
mostly wanted to check first that there are not callers who abort the
iteration soon after starting it. For example, imagine a caller who
tries to determine "are there any tags at all" by iterating over
"refs/tags" with a callback that just returns 1; such a caller would
suffer the cost of reading all of the loose references in "refs/tags".
* Lazy loading loose references one reference at a time. The ideas was
that this would allow the reference cache to be used for
single-reference lookups. This change alone caused a significant
performance loss, so it would have had to be combined with code for
preloading directories or subtrees before a for_each_ref() iteration.
To unsubscribe from this list: send the line "unsubscribe git" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html