git/list[1] front-page[2] threads[3] people[4] search[5] about
 

Re: [PATCH 4/4] Fix delta offset overflow

From
Martin Koegler <martin.koegler@chello.at>
Date
Aug 11, 2017, 06:57 UTC
Message-ID
<20170811065732.GA15128@mail.zuhause>
In-Reply-To
<xmqqbmnn2lnv.fsf@gitster.mtv.corp.google.com>
On Thu, Aug 10, 2017 at 01:49:24PM -0700, Junio C Hamano wrote:
> The lower 4-byte of moff (before incrementing it with msize) were
> already encoded to the output stream before this hunk.  Shouldn't
> we be checking if moff would fit in uint32_t _before_ that happens?

moff is otherwise only decremented or assigned with an offset generated by create_delta_index. These offsets are limited by 4GB.

Any larger offets would be a programming bug - so qualify for just a "assert".
 
Show 6 quoted lines
> IOW, in a much earlier part of that "while (data < top)" loop, there
> is a code that tries to find a match that gives us a large msize by
> iterating over index->hash[], and it sets msize and moff as a potential
> location that we may want to use.  If moff is already big there, then
> we shouldn't consider it a match because we cannot encode its location
> using 4-byte anyway, no?
Recovering from an incorrect moff would add more complex code for a case, that
should not happen. A bailout would be sufficient.
 
Show 5 quoted lines
> Cutting it off at here by resetting msize to 0 might help the next
> iteration (I didn't check, but is the effect of it is to corrupt the
> "val" rolling checksum and make it unlikely that the hash
> computation would not find a correct match?) but it somehow feels
> like closing the barn door after the horse has already bolted...
The current code produces incorrect deltas - its not just a checksum issue.
By the way: 
Somebody interested in JGIT should also look at these two bugs:

https://github.com/eclipse/jgit/blob/005e5feb4ecd08c4e4d141a38b9e7942accb3212/org.eclipse.jgit/src/org/eclipse/jgit/internal/storage/pack/DeltaEncoder.java copy would also encode beyond 4GB - producing truncated delta offset.

https://github.com/eclipse/jgit/blob/005e5feb4ecd08c4e4d141a38b9e7942accb3212/org.eclipse.jgit/src/org/eclipse/jgit/internal/storage/pack/BinaryDelta.java apply uses int for decoding length values.

Regards, Martin

Previous: Junio C HamanoNext: Junio C Hamano
Message 4 of 6 in “Convert zlib.c to size_t”
  1. 3/4 Convert zlib.c to size_tMartin Koegler, Aug 10, 2017
  2. 4/4 Fix delta offset overflowMartin Koegler, Aug 10, 2017
  3. Junio C HamanoAug 10, 2017
  4. Martin KoeglerAug 11, 2017
  5. Junio C HamanoAug 11, 2017
  6. Junio C HamanoAug 10, 2017

Read the whole thread, see it on lore, or plain text.

$ cat FOOTERMessages come from the public archive at lore.kernel.org/git, fetched every hour. The front page is chosen and written each morning by an AI editor and can be wrong; the threads themselves are the record. About and API. For agents: an MCP server at https://gitlist.dev/mcp, and any thread, story or person page as Markdown by adding .md to its URL (or sending Accept: text/markdown). Details in /llms.txt.