I just wrapped up busy three weeks worth of a family trip and two back to back conferences. I completely forgot how exhausting conferences are, and how little time they leave for anything else. There's a short backlog of posts, that I'm going to publish once I'm back home, but now that I had a chance to recover a bit, I'm going to release what I managed to work on during my travels.
I present for your consideration a proof of concept release of a vpatcher, that can press keccak patches. This implementation was modeled on vpatch parser/press that's used in btcbase to render patches and more importantly to produce an in-memory press. While the general architecture was chosen upfront, this implementation was authored incrementally, a development style that is surprisingly well supported by Ada1, so the bulk of functionality is contained in a single file. At this point I'm not convinced that some kind of split is required, though splitting it in the style of ascii's FFI might bring some clarity.
Vpatch is essentially standard unix patch tool, that supports strict application of a unified diff and verification of V hashes. It takes a patch in a standard input stream and attempts to press whatever content into the current directory. Existing files are patched, new files are created, old files are removed. Processing is done one file at a time, and the operation terminates with an error when an issue is encountered. Patcher keeps a temporary file around for the result output, which gets moved in place once the file's hash has been verified. This means that atomicity is preserved at a file level, but not at the patch level and failed press results in an inconsistent state of the directory. Individual files are always either in the previous or new state, which means that you get to inspect the offending file, but you have to fully redo the press on failure. This is a decision that I might have to reconsider, at the expense of increased complexity. Right now very little is kept in memory: information about the current file being patched, the current hunk and whatever simple data used to track the state.
To build the patcher use make vpatch, or call grpbuild on vpatch.gpr directly. To test the patcher, I reground trb stable release. Since the patcher doesn't verify sigs, I haven't signed the regrind, and it's provided for testing purposes only.
Press the genesis,
% vpatch < ps/1.vpatch creating bitcoin/.gitignore creating bitcoin/COPYING ... creating bitcoin/src/wallet.cpp creating bitcoin/src/wallet.h
Apply the next patch on top of it,
% vpatch < ps/2.vpatch patching bitcoin/src/bitcoinrpc.cpp patching bitcoin/src/db.cpp patching bitcoin/src/headers.h patching bitcoin/src/init.cpp deleting bitcoin/src/qtui.h patching bitcoin/src/util.h patching bitcoin/src/wallet.cpp
If we now try to rerun genesis in the same folder we get,
% vpatch < ps/1.vpatch creating bitcoin/.gitignore raised VPATCH.STATE : attempt to create a file, but file already exists
Likewise attempt to reapply second patch results in failure, since whatever files have invalid hash2
% vpatch < ps/2.vpatch patching bitcoin/src/bitcoinrpc.cpp raised VPATCH.STATE : from hash doesn't match
Same will happen if we attempt to apply a significantly later patch, since the necessary intermediate patches are missing,
% vpatch < ps/12.vpatch patching bitcoin/src/main.cpp raised VPATCH.STATE : from hash doesn't match
Finally applying the correct patch succeeds,
% vpatch < ps/3.vpatch patching bitcoin/src/db.cpp patching bitcoin/src/init.cpp patching bitcoin/src/main.cpp patching bitcoin/src/main.h patching bitcoin/src/makefile.linux-mingw patching bitcoin/src/makefile.unix patching bitcoin/src/net.cpp
Supporting pipe streaming means that we can start vpatch and incrementally feed it patches, moving the tree towards the press top. (In this case patches that we're pressing are named in order from 1 to 27.)
% cat ps/{1..27}.vpatch | vpatch
creating bitcoin/.gitignore
creating bitcoin/COPYING
...
creating bitcoin/deps/Makefile
creating bitcoin/deps/Manifest.sha512
patching bitcoin/src/db.cpp
patching bitcoin/src/init.cpp
patching bitcoin/src/main.cpp
creating bitcoin/verify.mk
The press can be sanity checked using e.g. checksum file from btcbase, but obviously the tool itself does both input and output hash verification as it goes.
There are some known issues, the biggest one is that \ No newline at end of file doesn't work yet, and the patcher fails when it encounters the dirrective. Half way through development I discovered that Text_IO is idiosyncratic: there's no machinery to produce a line without a newline at the end, or to figure out whether or not existing file has one.3 Ada always outputs a valid text file and there's no way to avoid it with Text_IO. Future developers beware! A solution to this problem is to use Sequential_IO specialized on Character, but that means writing own high level procedures like Get_Line. In the work in progress modification that uses Sequential_IO I was able to build a drop in replacement for Text_IO with minimum of changes to existing code, by gradually adding the missing functionality.
To understand the way this patcher works it's helpful to have some idea about the diff format, that we're using. There's three data structures that I use to keep track of patch data. The header,
type Header (From_L, To_L: Natural) Is record
From_Hash: Hash;
From_File: String(1..From_L);
To_Hash: Hash;
To_File: String(1..To_L);
end record;
which holds the source and the destination file information and corresponds to
diff -uNr a/bitcoin/.gitignore b/bitcoin/.gitignore --- a/bitcoin/.gitignore false +++ b/bitcoin/.gitignore 6654c7489c311585d7d3...
lines.
A hash is a variant record type that can either hold a specific value or when there's no hash, indicated by false label, it's explicitly marked empty, which happens when the file is either created or removed, with an empty from and to respectively.
type Hash_Type is (Empty, Value);
type Hash(The_Type: Hash_Type := Empty) is record
case The_Type is
when Value =>
Value: String(1..Hash_Length);
when Empty =>
null;
end case;
end record;
We distinguish between three possible file operations,
type Patch_Op is (Op_Create, Op_Delete, Op_Patch);
which depend on the presence of hashes. If we only have input hash, that means that the file has been deleted, likewise only output hash the file is being newly created. Both hashes indicate that the file is being patched.
Each header is followed by one or more "hunks", a line count prelude followed by the line related commands,
type Line_Numbers is record
Start: Natural;
Count: Natural;
end record;
type Hunk is record
From_File_Line_Numbers: Line_Numbers;
To_File_Line_Numbers: Line_Numbers;
end record;
A hunk holds the actual change details that is a sequence of optional context lines which we use for sanity checking ("patch claims line foo does input file actually have foo in that line"), followed by some number of additions or deletions, followed by optional context lines. The line commands are not actually stored in memory, instead they are processed as they are encountered. A typical hunk looks like this,
@@ -435,7 +432,7 @@
{
BOOST_FOREACH(string strAddr, mapMultiArgs["-addnode"])
{
- CAddress addr(strAddr, fAllowDNS);
+ CAddress addr(strAddr);
addr.nTime = 0; // so it won't relay unless successfully connected
if (addr.IsValid())
AddAddress(addr);
Only the line in bold is kept in memory. Each record has a corresponding Get procedure which reads the record from input stream. This way you can say, e.g. Get(A_Hunk) and that'll read the @@ -435,7 +432,7 @@ line from input stream.
The parser is naive in that it looks at the stream one character at a time, and dispatches to various handlers, rarelly needing to read the whole of something, before making a decision. This is a traditional lisp parsing technique, which is also well supported in Ada. The bulk of work happens inside an unreasonably large Process_Hunks_For_Header procedure. I will eventually attempt to refactor it, but right now it does all the pre and post checks, parses the hunk body and performs relevant modifications. It relies on record Gets to parse the input patch. The are two loops in the body, Hunk_Loop, which handles each hunk under the header, and Hunk_Body_Loop, which actually handles individual line changes within a hunk. The core of hunk body loop is a dispatch on the first character in the line,
exit Hunk_Body_Loop when From_Count = 0 and To_Count = 0;
Look_Ahead(C, EOL);
if EOL then
raise Parse with "blank line in hunk";
end if;
case C is
when '+' => -- line added
...
when '-' => -- line deleted
...
when ' ' => -- line stays the same
...
when others =>
raise Parse with "unexpected character "
...
end case;
Attentive reader will note that we exit the loop based exclusively on the expected line count, which is the information communicated in the hunk @@ prelude.
There are some obvious improvements for future releases, the aformentioned newline at end of file issue. I'd also like to port this implementation to SHA 512 branch, to allow testing in the current workflow. The SHA port particularly will let me test Ada to C interoperability. Going back to the Wednesday schedule, I will address one of these in the next release.
The original plan to get vpatch released this week fell through, instead there's more bug fixes.
The exercise of trimming GNU diff left a bad taste in my mouth, the end result, while significantly reduced and thus easier to study, is still a significant chunk of "clever" C code clocking at 3383 lines. But the exercise was worthwhile1 in that it allowed me to explicitly preserve diff's quirks when it comes to hunk construction, in order to be able to replicate existing vpatches.
Same consideration doesn't apply to a patcher, since a patcher is entirely dumb machinery, a kind of player piano, executing instructions from pre-recorded tape. As I've been enjoying my brief foray into Ada programming, thrust as it was on me by the republic, I decided to stick to the environment and use it to implement the patcher also. There is some rational reasons for using Ada for patcher instead of C. Where differ can afford to be sloppy in operation-- an operator can identify issues by reading the patch-- a press absolutely must result in a tree of files claimed by the press chain, or fail explicitly.
Current version of ada patcher was modeled on btcbase's internal Lisp implementation, and at 490 lines, can successfully press trb's "stable" branch. Unfortunately in the process of testing the patcher I've discovered another bug in the current keccak differ, that ate up the rest of my allocated time.
The possibility of that bug was hinted at in the recent conversation with diana_coman on the subject of Ada and C interoperability. Vtools interfaces to SMG's Keccak has two functions that among other things transfer arrays of characters between diff's C code and Keccak's Ada, C_Hash and C_End. The first takes in the text of the files under the consideration, in chunks, and the later sends back the final hash value, encoded as an array of bits, and on the C side represented as a string. Last week's vdiff uses Interfaces.C. Char_Array type to point to a shared char buffer, and standard functions Interfaces.C. To_Ada / Interfaces.C. To_C to convert the data between languages.
Well, in our conversation, Diana mentioned that in her experiments with To_Ada it sometime stoped too soon and failed to copy the entire contents of buffer. My reaction at the time was essentially "well, works for me"2, except now it doesn't, and on the most recent pass the issue came back with vengeance, almost entirely failing to transfer the data3. The test environment ostensibly stayed the same, so I'm completely mystified by the behavior. I'll attempt a dive into Ada's code to at least understand what's going on, but meanwhile, I rewrote the offending bits using yet another C to Ada copy method.
This is not the approach that diana_coman took in her code, which still uses Interfaces.C. Char_Array for data type, but instead of using To_C/Ada her functions explicitly walk the buffer in a loop, copying each character one by one. I instead listened to the siren's call of Ada's documentation4 and went with Interfaces.C. Pointers, which even provides a helpful example of a roundtrip at the end of the section. The details of the implementation can be seen in the patch, but they follow almost directly the blueprint in the documentation. The approach is similar to what diana_coman does, in that characters are copied in an explicit loop, but instead of Char_Array I'm now using a pointer abstraction, which mimics C's behavior and requires explicit advancement.
The method that I've implemented turned out to have already been condemned by ave1. Apparently he wrote what looks like an extensive investigation into dealing with char array. All in all much back to the drawing board.
The patch also includes a backport of bounds error fix for smg_keccak, documented in extensive detail on diana_coman's blog.