Currently, Repository.getAllRefs() and Repository.getTags() silently
ignores an IOException and instead returns an empty map. Repository
is a public API and as such cannot be changed until the next major
revision change. Where possible, update the internal jgit APIs to
use the RefDatabase directly, since it propagates the error.

Change-Id: I4e4537d8bd0fa772f388262684c5c4ca1929dc4c

Enable streaming compression for any response that is bigger than
the 32 KiB buffer used by SmartOutputStream. This is useful on the
info/refs file which can have many branches and tags listed, and
is often bigger than 32 KiB, but also compresses by at least 50%.

Disable streaming compression on large git-upload-pack responses,
as these are usually highly compressed Git pack data. Trying to
compress these with gzip will only waste CPU time and additional
transfer space with the gzip wrapper. Small git-upload-pack data
is usually text based negotiation responses and can be squeezed
smaller with a little bit of CPU usage.

Change-Id: Ia13e63ed334f594d5e1ab53a97240eb2e8f550e2

It isn't strictly necessary to validate every reference's target
object is reachable in the repository before advertising it to a
client. This is an expensive operation when there are thousands of
references, and its very unlikely that a reference uses a missing
object, because garbage collection proceeds from the references and
walks down through the graph. So trying to hide a dangling reference
from clients is relatively pointless.

Even if we are trying to avoid giving a client a corrupt repository,
this simple check isn't sufficient.  It is possible for a reference to
point to a valid commit, but that commit to have a missing blob in its
root tree.  This can be caused by staging a file into the index,
waiting several weeks, then committing that file while also racing
against a prune.  The prune may delete the blob, since its
modification time is more than 2 weeks ago, but retain the commit,
since its modification time is right now.

Such graph corruption is already caught during PackWriter as it
enumerates the graph from the client's want list and digs back
to the roots or common base.  Leave the reference validation also
for that same phase, where we know we have to parse the object to
support the enumeration.

Change-Id: Iee70ead0d3ed2d2fcc980417d09d7a69b05f5c2f
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>

Update a number of calling sites of RevWalk to ensure the walker's
internal ObjectReader is released after the walk is no longer used.
Because the ObjectReader is likely to hold onto a native resource
like an Inflater, we don't want to leak them outside of their
useful scope.

Where possible we also try to share ObjectReaders across several
walk pools, or between a walker and a PackWriter.  This permits
the ObjectReader to actually do some caching if it felt inclined
to do so.

Not everything was updated, we'll probably need to come back and
update even more call sites, but these are some of the biggest
offenders.  Test cases in particular aren't updated.  My plan is to
move most storage-agnostic tests onto some purely in-memory storage
solution that doesn't do compression.

Change-Id: I04087ec79faeea208b19848939898ad7172b6672
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>

The HTTP server side code now uses the same approach that the smart
HTTP client code uses when preparing a request body.  The payload
is streamed into a TemporaryBuffer of limited size.  If the entire
data fits, its compressed with gzip if the user agent supports that,
and a Content-Length header is used to transmit the fixed length
body to the peer.  If however the data overflows the limited memory
segment, its streamed uncompressed to the peer.

One might initially think that larger contents which overflow
the buffer should also be compressed, rather than sent raw, since
they were deemed "large".  But usually these larger contents are
actually a pack file which has been already heavily compressed by
Git specific routines.  Trying to deflate that with gzip is probably
going to take up more space, not less, so the compression overhead
isn't worthwhile.

This buffer and compress optimization helps repositories with a
large number of references, as their text based advertisements
compress well. For example jgit's own native repository currently
requires 32,628 bytes for its full advertisement of 489 references.
Most repositories have fewer references, and thus could compress
their entire response in one buffer.

Change-Id: I790609c9f763339e0a1db9172aa570e29af96f42
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>

Don't copy and sort the set of references if they are passed through
in a RefMap or a SortedMap using the key's natural sort ordering.
Either map is already in the order we want to present the items
to the client in, so copying and sorting is a waste of local CPU
and memory.

Change-Id: I49ada7c1220e0fc2a163b9752c2b77525d9c82c1
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>

This commit actually does three major changes to the way references
are handled within JGit.  Unfortunately they were easier to do as
a single massive commit than to break them up into smaller units.

Disambiguate symbolic references:
---------------------------------

  Reporting a symbolic reference such as HEAD as though it were
  any other normal reference like refs/heads/master causes subtle
  programming errors.  We have been bitten by this error on several
  occasions, as have some downstream applications written by myself.

  Instead of reporting HEAD as a reference whose name differs from
  its "original name", report it as an actual SymbolicRef object
  that the application can test the type and examine the target of.

  With this change, Ref is now an abstract type with different
  subclasses for the different types.

  In the classical example of "HEAD" being a symbolic reference to
  branch "refs/heads/master", the Repository.getAllRefs() method
  will now return:

      Map<String, Ref> all = repository.getAllRefs();
      SymbolicRef HEAD = (SymbolicRef) all.get("HEAD");
      ObjectIdRef master = (ObjectIdRef) all.get("refs/heads/master");

      assertSame(master,               HEAD.getTarget());
      assertSame(master.getObjectId(), HEAD.getObjectId());

      assertEquals("HEAD",              HEAD.getName());
      assertEquals("refs/heads/master", master.getName());

  A nice side-effect of this change is the storage type of the
  symbolic reference is no longer ambiguous with the storge type
  of the underlying reference it targets.  In the above example,
  if master was only available in the packed-refs file, then the
  following is also true:

      assertSame(Ref.Storage.LOOSE,  HEAD.getStorage());
      assertSame(Ref.Storage.PACKED, master.getStorage());

  (Prior to this change we returned the ambiguous storage of
   LOOSE_PACKED for HEAD, which was confusing since it wasn't
   actually true on disk).

  Another nice side-effect of this change is all intermediate
  symbolic references are preserved, and are therefore visible
  to the application when they walk the target chain.  We can
  now correctly inspect chains of symbolic references.

  As a result of this change the Ref.getOrigName() method has been
  removed from the API.  Applications should identify a symbolic
  reference by testing for isSymbolic() and not by using an arcane
  string comparsion between properties.

Abstract the RefDatabase storage:
---------------------------------

  RefDatabase is now abstract, similar to ObjectDatabase, and a
  new concrete implementation called RefDirectory is used for the
  traditional on-disk storage layout.  In the future we plan to
  support additional implementations, such as a pure in-memory
  RefDatabase for unit testing purposes.

Optimize RefDirectory:
----------------------

  The implementation of the in-memory reference cache, reading, and
  update routines has been completely rewritten.  Much of the code
  was heavily borrowed or cribbed from the prior implementation,
  so copyright notices have been left intact as much as possible.

  The RefDirectory cache no longer confuses symbolic references
  with normal references.  This permits the cache to resolve the
  value of a symbolic reference as late as possible, ensuring it
  is always current, without needing to maintain reverse pointers.

  The cache is now 2 sorted RefLists, rather than 3 HashMaps.
  Using sorted lists allows the implementation to reduce the
  in-memory footprint when storing many refs.  Using specialized
  types for the elements allows the code to avoid additional map
  lookups for auxiliary stat information.

  To improve scan time during getRefs(), the lists are returned via
  a copy-on-write contract.  Most callers of getRefs() do not modify
  the returned collections, so the copy-on-write semantics improves
  access on repositories with a large number of packed references.

  Iterator traversals of the returned Map<String,Ref> are performed
  using a simple merge-join of the two cache lists, ensuring we can
  perform the entire traversal in linear time as a function of the
  number of references: O(PackedRefs + LooseRefs).

  Scans of the loose reference space to update the cache run in
  O(LooseRefs log LooseRefs) time, as the directory contents
  are sorted before being merged against the in-memory cache.
  Since the majority of stable references are kept packed, there
  typically are only a handful of reference names to be sorted,
  so the sorting cost should not be very high.

  Locking is reduced during getRefs() by taking advantage of the
  copy-on-write semantics of the improved cache data structure.
  This permits concurrent readers to pull back references without
  blocking each other.  If there is contention updating the cache
  during a scan, one or more updates are simply skipped and will
  get picked up again in a future scan.

  Writing to the $GIT_DIR/packed-refs during reference delete is
  now fully atomic.  The file is locked, reparsed fresh, and written
  back out if a change is necessary.  This avoids all race conditions
  with concurrent external updates of the packed-refs file.

  The RefLogWriter class has been fully folded into RefDirectory
  and is therefore deleted.  Maintaining the reference's log is
  the responsiblity of the database implementation, and not all
  implementations will use java.io for access.

  Future work still remains to be done to abstract the ReflogReader
  class away from local disk IO.

Change-Id: I26b9287c45a4b2d2be35ba2849daa316f5eec85d
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>

This is a simple HTTP server that provides the minimum server side
support required for dumb (non-git aware) transport clients.

We produce the info/refs and objects/info/packs file on the fly
from the local repository state, but otherwise serve data as raw
files from the on-disk structure.

In the future we could better optimize the FileSender class and the
servlets that use it to take advantage of direct file to network
APIs in more advanced servlet containers like Jetty.

Our glue package borrows the idea of a micro embedded DSL from
Google Guice and uses it to configure a collection of Filters
and HttpServlets, all of which are matched against requests using
regular expressions.  If a subgroup exists in the pattern, it is
extracted and used for the path info component of the request.

Change-Id: Ia0f1a425d07d035e344ae54faf8aeb04763e7487
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>