Documentation / technical / hash-function-transition.txton commit am: avoid directory rename detection when calling recursive merge machinery (6aba117)
   1Git hash function transition
   2============================
   3
   4Objective
   5---------
   6Migrate Git from SHA-1 to a stronger hash function.
   7
   8Background
   9----------
  10At its core, the Git version control system is a content addressable
  11filesystem. It uses the SHA-1 hash function to name content. For
  12example, files, directories, and revisions are referred to by hash
  13values unlike in other traditional version control systems where files
  14or versions are referred to via sequential numbers. The use of a hash
  15function to address its content delivers a few advantages:
  16
  17* Integrity checking is easy. Bit flips, for example, are easily
  18  detected, as the hash of corrupted content does not match its name.
  19* Lookup of objects is fast.
  20
  21Using a cryptographically secure hash function brings additional
  22advantages:
  23
  24* Object names can be signed and third parties can trust the hash to
  25  address the signed object and all objects it references.
  26* Communication using Git protocol and out of band communication
  27  methods have a short reliable string that can be used to reliably
  28  address stored content.
  29
  30Over time some flaws in SHA-1 have been discovered by security
  31researchers. On 23 February 2017 the SHAttered attack
  32(https://shattered.io) demonstrated a practical SHA-1 hash collision.
  33
  34Git v2.13.0 and later subsequently moved to a hardened SHA-1
  35implementation by default, which isn't vulnerable to the SHAttered
  36attack.
  37
  38Thus Git has in effect already migrated to a new hash that isn't SHA-1
  39and doesn't share its vulnerabilities, its new hash function just
  40happens to produce exactly the same output for all known inputs,
  41except two PDFs published by the SHAttered researchers, and the new
  42implementation (written by those researchers) claims to detect future
  43cryptanalytic collision attacks.
  44
  45Regardless, it's considered prudent to move past any variant of SHA-1
  46to a new hash. There's no guarantee that future attacks on SHA-1 won't
  47be published in the future, and those attacks may not have viable
  48mitigations.
  49
  50If SHA-1 and its variants were to be truly broken, Git's hash function
  51could not be considered cryptographically secure any more. This would
  52impact the communication of hash values because we could not trust
  53that a given hash value represented the known good version of content
  54that the speaker intended.
  55
  56SHA-1 still possesses the other properties such as fast object lookup
  57and safe error checking, but other hash functions are equally suitable
  58that are believed to be cryptographically secure.
  59
  60Goals
  61-----
  62Where NewHash is a strong 256-bit hash function to replace SHA-1 (see
  63"Selection of a New Hash", below):
  64
  651. The transition to NewHash can be done one local repository at a time.
  66   a. Requiring no action by any other party.
  67   b. A NewHash repository can communicate with SHA-1 Git servers
  68      (push/fetch).
  69   c. Users can use SHA-1 and NewHash identifiers for objects
  70      interchangeably (see "Object names on the command line", below).
  71   d. New signed objects make use of a stronger hash function than
  72      SHA-1 for their security guarantees.
  732. Allow a complete transition away from SHA-1.
  74   a. Local metadata for SHA-1 compatibility can be removed from a
  75      repository if compatibility with SHA-1 is no longer needed.
  763. Maintainability throughout the process.
  77   a. The object format is kept simple and consistent.
  78   b. Creation of a generalized repository conversion tool.
  79
  80Non-Goals
  81---------
  821. Add NewHash support to Git protocol. This is valuable and the
  83   logical next step but it is out of scope for this initial design.
  842. Transparently improving the security of existing SHA-1 signed
  85   objects.
  863. Intermixing objects using multiple hash functions in a single
  87   repository.
  884. Taking the opportunity to fix other bugs in Git's formats and
  89   protocols.
  905. Shallow clones and fetches into a NewHash repository. (This will
  91   change when we add NewHash support to Git protocol.)
  926. Skip fetching some submodules of a project into a NewHash
  93   repository. (This also depends on NewHash support in Git
  94   protocol.)
  95
  96Overview
  97--------
  98We introduce a new repository format extension. Repositories with this
  99extension enabled use NewHash instead of SHA-1 to name their objects.
 100This affects both object names and object content --- both the names
 101of objects and all references to other objects within an object are
 102switched to the new hash function.
 103
 104NewHash repositories cannot be read by older versions of Git.
 105
 106Alongside the packfile, a NewHash repository stores a bidirectional
 107mapping between NewHash and SHA-1 object names. The mapping is generated
 108locally and can be verified using "git fsck". Object lookups use this
 109mapping to allow naming objects using either their SHA-1 and NewHash names
 110interchangeably.
 111
 112"git cat-file" and "git hash-object" gain options to display an object
 113in its sha1 form and write an object given its sha1 form. This
 114requires all objects referenced by that object to be present in the
 115object database so that they can be named using the appropriate name
 116(using the bidirectional hash mapping).
 117
 118Fetches from a SHA-1 based server convert the fetched objects into
 119NewHash form and record the mapping in the bidirectional mapping table
 120(see below for details). Pushes to a SHA-1 based server convert the
 121objects being pushed into sha1 form so the server does not have to be
 122aware of the hash function the client is using.
 123
 124Detailed Design
 125---------------
 126Repository format extension
 127~~~~~~~~~~~~~~~~~~~~~~~~~~~
 128A NewHash repository uses repository format version `1` (see
 129Documentation/technical/repository-version.txt) with extensions
 130`objectFormat` and `compatObjectFormat`:
 131
 132        [core]
 133                repositoryFormatVersion = 1
 134        [extensions]
 135                objectFormat = newhash
 136                compatObjectFormat = sha1
 137
 138The combination of setting `core.repositoryFormatVersion=1` and
 139populating `extensions.*` ensures that all versions of Git later than
 140`v0.99.9l` will die instead of trying to operate on the NewHash
 141repository, instead producing an error message.
 142
 143        # Between v0.99.9l and v2.7.0
 144        $ git status
 145        fatal: Expected git repo version <= 0, found 1
 146        # After v2.7.0
 147        $ git status
 148        fatal: unknown repository extensions found:
 149                objectformat
 150                compatobjectformat
 151
 152See the "Transition plan" section below for more details on these
 153repository extensions.
 154
 155Object names
 156~~~~~~~~~~~~
 157Objects can be named by their 40 hexadecimal digit sha1-name or 64
 158hexadecimal digit newhash-name, plus names derived from those (see
 159gitrevisions(7)).
 160
 161The sha1-name of an object is the SHA-1 of the concatenation of its
 162type, length, a nul byte, and the object's sha1-content. This is the
 163traditional <sha1> used in Git to name objects.
 164
 165The newhash-name of an object is the NewHash of the concatenation of its
 166type, length, a nul byte, and the object's newhash-content.
 167
 168Object format
 169~~~~~~~~~~~~~
 170The content as a byte sequence of a tag, commit, or tree object named
 171by sha1 and newhash differ because an object named by newhash-name refers to
 172other objects by their newhash-names and an object named by sha1-name
 173refers to other objects by their sha1-names.
 174
 175The newhash-content of an object is the same as its sha1-content, except
 176that objects referenced by the object are named using their newhash-names
 177instead of sha1-names. Because a blob object does not refer to any
 178other object, its sha1-content and newhash-content are the same.
 179
 180The format allows round-trip conversion between newhash-content and
 181sha1-content.
 182
 183Object storage
 184~~~~~~~~~~~~~~
 185Loose objects use zlib compression and packed objects use the packed
 186format described in Documentation/technical/pack-format.txt, just like
 187today. The content that is compressed and stored uses newhash-content
 188instead of sha1-content.
 189
 190Pack index
 191~~~~~~~~~~
 192Pack index (.idx) files use a new v3 format that supports multiple
 193hash functions. They have the following format (all integers are in
 194network byte order):
 195
 196- A header appears at the beginning and consists of the following:
 197  - The 4-byte pack index signature: '\377t0c'
 198  - 4-byte version number: 3
 199  - 4-byte length of the header section, including the signature and
 200    version number
 201  - 4-byte number of objects contained in the pack
 202  - 4-byte number of object formats in this pack index: 2
 203  - For each object format:
 204    - 4-byte format identifier (e.g., 'sha1' for SHA-1)
 205    - 4-byte length in bytes of shortened object names. This is the
 206      shortest possible length needed to make names in the shortened
 207      object name table unambiguous.
 208    - 4-byte integer, recording where tables relating to this format
 209      are stored in this index file, as an offset from the beginning.
 210  - 4-byte offset to the trailer from the beginning of this file.
 211  - Zero or more additional key/value pairs (4-byte key, 4-byte
 212    value). Only one key is supported: 'PSRC'. See the "Loose objects
 213    and unreachable objects" section for supported values and how this
 214    is used.  All other keys are reserved. Readers must ignore
 215    unrecognized keys.
 216- Zero or more NUL bytes. This can optionally be used to improve the
 217  alignment of the full object name table below.
 218- Tables for the first object format:
 219  - A sorted table of shortened object names.  These are prefixes of
 220    the names of all objects in this pack file, packed together
 221    without offset values to reduce the cache footprint of the binary
 222    search for a specific object name.
 223
 224  - A table of full object names in pack order. This allows resolving
 225    a reference to "the nth object in the pack file" (from a
 226    reachability bitmap or from the next table of another object
 227    format) to its object name.
 228
 229  - A table of 4-byte values mapping object name order to pack order.
 230    For an object in the table of sorted shortened object names, the
 231    value at the corresponding index in this table is the index in the
 232    previous table for that same object.
 233
 234    This can be used to look up the object in reachability bitmaps or
 235    to look up its name in another object format.
 236
 237  - A table of 4-byte CRC32 values of the packed object data, in the
 238    order that the objects appear in the pack file. This is to allow
 239    compressed data to be copied directly from pack to pack during
 240    repacking without undetected data corruption.
 241
 242  - A table of 4-byte offset values. For an object in the table of
 243    sorted shortened object names, the value at the corresponding
 244    index in this table indicates where that object can be found in
 245    the pack file. These are usually 31-bit pack file offsets, but
 246    large offsets are encoded as an index into the next table with the
 247    most significant bit set.
 248
 249  - A table of 8-byte offset entries (empty for pack files less than
 250    2 GiB). Pack files are organized with heavily used objects toward
 251    the front, so most object references should not need to refer to
 252    this table.
 253- Zero or more NUL bytes.
 254- Tables for the second object format, with the same layout as above,
 255  up to and not including the table of CRC32 values.
 256- Zero or more NUL bytes.
 257- The trailer consists of the following:
 258  - A copy of the 20-byte NewHash checksum at the end of the
 259    corresponding packfile.
 260
 261  - 20-byte NewHash checksum of all of the above.
 262
 263Loose object index
 264~~~~~~~~~~~~~~~~~~
 265A new file $GIT_OBJECT_DIR/loose-object-idx contains information about
 266all loose objects. Its format is
 267
 268  # loose-object-idx
 269  (newhash-name SP sha1-name LF)*
 270
 271where the object names are in hexadecimal format. The file is not
 272sorted.
 273
 274The loose object index is protected against concurrent writes by a
 275lock file $GIT_OBJECT_DIR/loose-object-idx.lock. To add a new loose
 276object:
 277
 2781. Write the loose object to a temporary file, like today.
 2792. Open loose-object-idx.lock with O_CREAT | O_EXCL to acquire the lock.
 2803. Rename the loose object into place.
 2814. Open loose-object-idx with O_APPEND and write the new object
 2825. Unlink loose-object-idx.lock to release the lock.
 283
 284To remove entries (e.g. in "git pack-refs" or "git-prune"):
 285
 2861. Open loose-object-idx.lock with O_CREAT | O_EXCL to acquire the
 287   lock.
 2882. Write the new content to loose-object-idx.lock.
 2893. Unlink any loose objects being removed.
 2904. Rename to replace loose-object-idx, releasing the lock.
 291
 292Translation table
 293~~~~~~~~~~~~~~~~~
 294The index files support a bidirectional mapping between sha1-names
 295and newhash-names. The lookup proceeds similarly to ordinary object
 296lookups. For example, to convert a sha1-name to a newhash-name:
 297
 298 1. Look for the object in idx files. If a match is present in the
 299    idx's sorted list of truncated sha1-names, then:
 300    a. Read the corresponding entry in the sha1-name order to pack
 301       name order mapping.
 302    b. Read the corresponding entry in the full sha1-name table to
 303       verify we found the right object. If it is, then
 304    c. Read the corresponding entry in the full newhash-name table.
 305       That is the object's newhash-name.
 306 2. Check for a loose object. Read lines from loose-object-idx until
 307    we find a match.
 308
 309Step (1) takes the same amount of time as an ordinary object lookup:
 310O(number of packs * log(objects per pack)). Step (2) takes O(number of
 311loose objects) time. To maintain good performance it will be necessary
 312to keep the number of loose objects low. See the "Loose objects and
 313unreachable objects" section below for more details.
 314
 315Since all operations that make new objects (e.g., "git commit") add
 316the new objects to the corresponding index, this mapping is possible
 317for all objects in the object store.
 318
 319Reading an object's sha1-content
 320~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 321The sha1-content of an object can be read by converting all newhash-names
 322its newhash-content references to sha1-names using the translation table.
 323
 324Fetch
 325~~~~~
 326Fetching from a SHA-1 based server requires translating between SHA-1
 327and NewHash based representations on the fly.
 328
 329SHA-1s named in the ref advertisement that are present on the client
 330can be translated to NewHash and looked up as local objects using the
 331translation table.
 332
 333Negotiation proceeds as today. Any "have"s generated locally are
 334converted to SHA-1 before being sent to the server, and SHA-1s
 335mentioned by the server are converted to NewHash when looking them up
 336locally.
 337
 338After negotiation, the server sends a packfile containing the
 339requested objects. We convert the packfile to NewHash format using
 340the following steps:
 341
 3421. index-pack: inflate each object in the packfile and compute its
 343   SHA-1. Objects can contain deltas in OBJ_REF_DELTA format against
 344   objects the client has locally. These objects can be looked up
 345   using the translation table and their sha1-content read as
 346   described above to resolve the deltas.
 3472. topological sort: starting at the "want"s from the negotiation
 348   phase, walk through objects in the pack and emit a list of them,
 349   excluding blobs, in reverse topologically sorted order, with each
 350   object coming later in the list than all objects it references.
 351   (This list only contains objects reachable from the "wants". If the
 352   pack from the server contained additional extraneous objects, then
 353   they will be discarded.)
 3543. convert to newhash: open a new (newhash) packfile. Read the topologically
 355   sorted list just generated. For each object, inflate its
 356   sha1-content, convert to newhash-content, and write it to the newhash
 357   pack. Record the new sha1<->newhash mapping entry for use in the idx.
 3584. sort: reorder entries in the new pack to match the order of objects
 359   in the pack the server generated and include blobs. Write a newhash idx
 360   file
 3615. clean up: remove the SHA-1 based pack file, index, and
 362   topologically sorted list obtained from the server in steps 1
 363   and 2.
 364
 365Step 3 requires every object referenced by the new object to be in the
 366translation table. This is why the topological sort step is necessary.
 367
 368As an optimization, step 1 could write a file describing what non-blob
 369objects each object it has inflated from the packfile references. This
 370makes the topological sort in step 2 possible without inflating the
 371objects in the packfile for a second time. The objects need to be
 372inflated again in step 3, for a total of two inflations.
 373
 374Step 4 is probably necessary for good read-time performance. "git
 375pack-objects" on the server optimizes the pack file for good data
 376locality (see Documentation/technical/pack-heuristics.txt).
 377
 378Details of this process are likely to change. It will take some
 379experimenting to get this to perform well.
 380
 381Push
 382~~~~
 383Push is simpler than fetch because the objects referenced by the
 384pushed objects are already in the translation table. The sha1-content
 385of each object being pushed can be read as described in the "Reading
 386an object's sha1-content" section to generate the pack written by git
 387send-pack.
 388
 389Signed Commits
 390~~~~~~~~~~~~~~
 391We add a new field "gpgsig-newhash" to the commit object format to allow
 392signing commits without relying on SHA-1. It is similar to the
 393existing "gpgsig" field. Its signed payload is the newhash-content of the
 394commit object with any "gpgsig" and "gpgsig-newhash" fields removed.
 395
 396This means commits can be signed
 3971. using SHA-1 only, as in existing signed commit objects
 3982. using both SHA-1 and NewHash, by using both gpgsig-newhash and gpgsig
 399   fields.
 4003. using only NewHash, by only using the gpgsig-newhash field.
 401
 402Old versions of "git verify-commit" can verify the gpgsig signature in
 403cases (1) and (2) without modifications and view case (3) as an
 404ordinary unsigned commit.
 405
 406Signed Tags
 407~~~~~~~~~~~
 408We add a new field "gpgsig-newhash" to the tag object format to allow
 409signing tags without relying on SHA-1. Its signed payload is the
 410newhash-content of the tag with its gpgsig-newhash field and "-----BEGIN PGP
 411SIGNATURE-----" delimited in-body signature removed.
 412
 413This means tags can be signed
 4141. using SHA-1 only, as in existing signed tag objects
 4152. using both SHA-1 and NewHash, by using gpgsig-newhash and an in-body
 416   signature.
 4173. using only NewHash, by only using the gpgsig-newhash field.
 418
 419Mergetag embedding
 420~~~~~~~~~~~~~~~~~~
 421The mergetag field in the sha1-content of a commit contains the
 422sha1-content of a tag that was merged by that commit.
 423
 424The mergetag field in the newhash-content of the same commit contains the
 425newhash-content of the same tag.
 426
 427Submodules
 428~~~~~~~~~~
 429To convert recorded submodule pointers, you need to have the converted
 430submodule repository in place. The translation table of the submodule
 431can be used to look up the new hash.
 432
 433Loose objects and unreachable objects
 434~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 435Fast lookups in the loose-object-idx require that the number of loose
 436objects not grow too high.
 437
 438"git gc --auto" currently waits for there to be 6700 loose objects
 439present before consolidating them into a packfile. We will need to
 440measure to find a more appropriate threshold for it to use.
 441
 442"git gc --auto" currently waits for there to be 50 packs present
 443before combining packfiles. Packing loose objects more aggressively
 444may cause the number of pack files to grow too quickly. This can be
 445mitigated by using a strategy similar to Martin Fick's exponential
 446rolling garbage collection script:
 447https://gerrit-review.googlesource.com/c/gerrit/+/35215
 448
 449"git gc" currently expels any unreachable objects it encounters in
 450pack files to loose objects in an attempt to prevent a race when
 451pruning them (in case another process is simultaneously writing a new
 452object that refers to the about-to-be-deleted object). This leads to
 453an explosion in the number of loose objects present and disk space
 454usage due to the objects in delta form being replaced with independent
 455loose objects.  Worse, the race is still present for loose objects.
 456
 457Instead, "git gc" will need to move unreachable objects to a new
 458packfile marked as UNREACHABLE_GARBAGE (using the PSRC field; see
 459below). To avoid the race when writing new objects referring to an
 460about-to-be-deleted object, code paths that write new objects will
 461need to copy any objects from UNREACHABLE_GARBAGE packs that they
 462refer to to new, non-UNREACHABLE_GARBAGE packs (or loose objects).
 463UNREACHABLE_GARBAGE are then safe to delete if their creation time (as
 464indicated by the file's mtime) is long enough ago.
 465
 466To avoid a proliferation of UNREACHABLE_GARBAGE packs, they can be
 467combined under certain circumstances. If "gc.garbageTtl" is set to
 468greater than one day, then packs created within a single calendar day,
 469UTC, can be coalesced together. The resulting packfile would have an
 470mtime before midnight on that day, so this makes the effective maximum
 471ttl the garbageTtl + 1 day. If "gc.garbageTtl" is less than one day,
 472then we divide the calendar day into intervals one-third of that ttl
 473in duration. Packs created within the same interval can be coalesced
 474together. The resulting packfile would have an mtime before the end of
 475the interval, so this makes the effective maximum ttl equal to the
 476garbageTtl * 4/3.
 477
 478This rule comes from Thirumala Reddy Mutchukota's JGit change
 479https://git.eclipse.org/r/90465.
 480
 481The UNREACHABLE_GARBAGE setting goes in the PSRC field of the pack
 482index. More generally, that field indicates where a pack came from:
 483
 484 - 1 (PACK_SOURCE_RECEIVE) for a pack received over the network
 485 - 2 (PACK_SOURCE_AUTO) for a pack created by a lightweight
 486   "gc --auto" operation
 487 - 3 (PACK_SOURCE_GC) for a pack created by a full gc
 488 - 4 (PACK_SOURCE_UNREACHABLE_GARBAGE) for potential garbage
 489   discovered by gc
 490 - 5 (PACK_SOURCE_INSERT) for locally created objects that were
 491   written directly to a pack file, e.g. from "git add ."
 492
 493This information can be useful for debugging and for "gc --auto" to
 494make appropriate choices about which packs to coalesce.
 495
 496Caveats
 497-------
 498Invalid objects
 499~~~~~~~~~~~~~~~
 500The conversion from sha1-content to newhash-content retains any
 501brokenness in the original object (e.g., tree entry modes encoded with
 502leading 0, tree objects whose paths are not sorted correctly, and
 503commit objects without an author or committer). This is a deliberate
 504feature of the design to allow the conversion to round-trip.
 505
 506More profoundly broken objects (e.g., a commit with a truncated "tree"
 507header line) cannot be converted but were not usable by current Git
 508anyway.
 509
 510Shallow clone and submodules
 511~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 512Because it requires all referenced objects to be available in the
 513locally generated translation table, this design does not support
 514shallow clone or unfetched submodules. Protocol improvements might
 515allow lifting this restriction.
 516
 517Alternates
 518~~~~~~~~~~
 519For the same reason, a newhash repository cannot borrow objects from a
 520sha1 repository using objects/info/alternates or
 521$GIT_ALTERNATE_OBJECT_REPOSITORIES.
 522
 523git notes
 524~~~~~~~~~
 525The "git notes" tool annotates objects using their sha1-name as key.
 526This design does not describe a way to migrate notes trees to use
 527newhash-names. That migration is expected to happen separately (for
 528example using a file at the root of the notes tree to describe which
 529hash it uses).
 530
 531Server-side cost
 532~~~~~~~~~~~~~~~~
 533Until Git protocol gains NewHash support, using NewHash based storage
 534on public-facing Git servers is strongly discouraged. Once Git
 535protocol gains NewHash support, NewHash based servers are likely not
 536to support SHA-1 compatibility, to avoid what may be a very expensive
 537hash reencode during clone and to encourage peers to modernize.
 538
 539The design described here allows fetches by SHA-1 clients of a
 540personal NewHash repository because it's not much more difficult than
 541allowing pushes from that repository. This support needs to be guarded
 542by a configuration option --- servers like git.kernel.org that serve a
 543large number of clients would not be expected to bear that cost.
 544
 545Meaning of signatures
 546~~~~~~~~~~~~~~~~~~~~~
 547The signed payload for signed commits and tags does not explicitly
 548name the hash used to identify objects. If some day Git adopts a new
 549hash function with the same length as the current SHA-1 (40
 550hexadecimal digit) or NewHash (64 hexadecimal digit) objects then the
 551intent behind the PGP signed payload in an object signature is
 552unclear:
 553
 554        object e7e07d5a4fcc2a203d9873968ad3e6bd4d7419d7
 555        type commit
 556        tag v2.12.0
 557        tagger Junio C Hamano <gitster@pobox.com> 1487962205 -0800
 558
 559        Git 2.12
 560
 561Does this mean Git v2.12.0 is the commit with sha1-name
 562e7e07d5a4fcc2a203d9873968ad3e6bd4d7419d7 or the commit with
 563new-40-digit-hash-name e7e07d5a4fcc2a203d9873968ad3e6bd4d7419d7?
 564
 565Fortunately NewHash and SHA-1 have different lengths. If Git starts
 566using another hash with the same length to name objects, then it will
 567need to change the format of signed payloads using that hash to
 568address this issue.
 569
 570Object names on the command line
 571~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 572To support the transition (see Transition plan below), this design
 573supports four different modes of operation:
 574
 575 1. ("dark launch") Treat object names input by the user as SHA-1 and
 576    convert any object names written to output to SHA-1, but store
 577    objects using NewHash.  This allows users to test the code with no
 578    visible behavior change except for performance.  This allows
 579    allows running even tests that assume the SHA-1 hash function, to
 580    sanity-check the behavior of the new mode.
 581
 582 2. ("early transition") Allow both SHA-1 and NewHash object names in
 583    input. Any object names written to output use SHA-1. This allows
 584    users to continue to make use of SHA-1 to communicate with peers
 585    (e.g. by email) that have not migrated yet and prepares for mode 3.
 586
 587 3. ("late transition") Allow both SHA-1 and NewHash object names in
 588    input. Any object names written to output use NewHash. In this
 589    mode, users are using a more secure object naming method by
 590    default.  The disruption is minimal as long as most of their peers
 591    are in mode 2 or mode 3.
 592
 593 4. ("post-transition") Treat object names input by the user as
 594    NewHash and write output using NewHash. This is safer than mode 3
 595    because there is less risk that input is incorrectly interpreted
 596    using the wrong hash function.
 597
 598The mode is specified in configuration.
 599
 600The user can also explicitly specify which format to use for a
 601particular revision specifier and for output, overriding the mode. For
 602example:
 603
 604git --output-format=sha1 log abac87a^{sha1}..f787cac^{newhash}
 605
 606Selection of a New Hash
 607-----------------------
 608In early 2005, around the time that Git was written,  Xiaoyun Wang,
 609Yiqun Lisa Yin, and Hongbo Yu announced an attack finding SHA-1
 610collisions in 2^69 operations. In August they published details.
 611Luckily, no practical demonstrations of a collision in full SHA-1 were
 612published until 10 years later, in 2017.
 613
 614The hash function NewHash to replace SHA-1 should be stronger than
 615SHA-1 was: we would like it to be trustworthy and useful in practice
 616for at least 10 years.
 617
 618Some other relevant properties:
 619
 6201. A 256-bit hash (long enough to match common security practice; not
 621   excessively long to hurt performance and disk usage).
 622
 6232. High quality implementations should be widely available (e.g. in
 624   OpenSSL).
 625
 6263. The hash function's properties should match Git's needs (e.g. Git
 627   requires collision and 2nd preimage resistance and does not require
 628   length extension resistance).
 629
 6304. As a tiebreaker, the hash should be fast to compute (fortunately
 631   many contenders are faster than SHA-1).
 632
 633Some hashes under consideration are SHA-256, SHA-512/256, SHA-256x16,
 634K12, and BLAKE2bp-256.
 635
 636Transition plan
 637---------------
 638Some initial steps can be implemented independently of one another:
 639- adding a hash function API (vtable)
 640- teaching fsck to tolerate the gpgsig-newhash field
 641- excluding gpgsig-* from the fields copied by "git commit --amend"
 642- annotating tests that depend on SHA-1 values with a SHA1 test
 643  prerequisite
 644- using "struct object_id", GIT_MAX_RAWSZ, and GIT_MAX_HEXSZ
 645  consistently instead of "unsigned char *" and the hardcoded
 646  constants 20 and 40.
 647- introducing index v3
 648- adding support for the PSRC field and safer object pruning
 649
 650
 651The first user-visible change is the introduction of the objectFormat
 652extension (without compatObjectFormat). This requires:
 653- implementing the loose-object-idx
 654- teaching fsck about this mode of operation
 655- using the hash function API (vtable) when computing object names
 656- signing objects and verifying signatures
 657- rejecting attempts to fetch from or push to an incompatible
 658  repository
 659
 660Next comes introduction of compatObjectFormat:
 661- translating object names between object formats
 662- translating object content between object formats
 663- generating and verifying signatures in the compat format
 664- adding appropriate index entries when adding a new object to the
 665  object store
 666- --output-format option
 667- ^{sha1} and ^{newhash} revision notation
 668- configuration to specify default input and output format (see
 669  "Object names on the command line" above)
 670
 671The next step is supporting fetches and pushes to SHA-1 repositories:
 672- allow pushes to a repository using the compat format
 673- generate a topologically sorted list of the SHA-1 names of fetched
 674  objects
 675- convert the fetched packfile to newhash format and generate an idx
 676  file
 677- re-sort to match the order of objects in the fetched packfile
 678
 679The infrastructure supporting fetch also allows converting an existing
 680repository. In converted repositories and new clones, end users can
 681gain support for the new hash function without any visible change in
 682behavior (see "dark launch" in the "Object names on the command line"
 683section). In particular this allows users to verify NewHash signatures
 684on objects in the repository, and it should ensure the transition code
 685is stable in production in preparation for using it more widely.
 686
 687Over time projects would encourage their users to adopt the "early
 688transition" and then "late transition" modes to take advantage of the
 689new, more futureproof NewHash object names.
 690
 691When objectFormat and compatObjectFormat are both set, commands
 692generating signatures would generate both SHA-1 and NewHash signatures
 693by default to support both new and old users.
 694
 695In projects using NewHash heavily, users could be encouraged to adopt
 696the "post-transition" mode to avoid accidentally making implicit use
 697of SHA-1 object names.
 698
 699Once a critical mass of users have upgraded to a version of Git that
 700can verify NewHash signatures and have converted their existing
 701repositories to support verifying them, we can add support for a
 702setting to generate only NewHash signatures. This is expected to be at
 703least a year later.
 704
 705That is also a good moment to advertise the ability to convert
 706repositories to use NewHash only, stripping out all SHA-1 related
 707metadata. This improves performance by eliminating translation
 708overhead and security by avoiding the possibility of accidentally
 709relying on the safety of SHA-1.
 710
 711Updating Git's protocols to allow a server to specify which hash
 712functions it supports is also an important part of this transition. It
 713is not discussed in detail in this document but this transition plan
 714assumes it happens. :)
 715
 716Alternatives considered
 717-----------------------
 718Upgrading everyone working on a particular project on a flag day
 719~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 720Projects like the Linux kernel are large and complex enough that
 721flipping the switch for all projects based on the repository at once
 722is infeasible.
 723
 724Not only would all developers and server operators supporting
 725developers have to switch on the same flag day, but supporting tooling
 726(continuous integration, code review, bug trackers, etc) would have to
 727be adapted as well. This also makes it difficult to get early feedback
 728from some project participants testing before it is time for mass
 729adoption.
 730
 731Using hash functions in parallel
 732~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 733(e.g. https://public-inbox.org/git/22708.8913.864049.452252@chiark.greenend.org.uk/ )
 734Objects newly created would be addressed by the new hash, but inside
 735such an object (e.g. commit) it is still possible to address objects
 736using the old hash function.
 737* You cannot trust its history (needed for bisectability) in the
 738  future without further work
 739* Maintenance burden as the number of supported hash functions grows
 740  (they will never go away, so they accumulate). In this proposal, by
 741  comparison, converted objects lose all references to SHA-1.
 742
 743Signed objects with multiple hashes
 744~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 745Instead of introducing the gpgsig-newhash field in commit and tag objects
 746for newhash-content based signatures, an earlier version of this design
 747added "hash newhash <newhash-name>" fields to strengthen the existing
 748sha1-content based signatures.
 749
 750In other words, a single signature was used to attest to the object
 751content using both hash functions. This had some advantages:
 752* Using one signature instead of two speeds up the signing process.
 753* Having one signed payload with both hashes allows the signer to
 754  attest to the sha1-name and newhash-name referring to the same object.
 755* All users consume the same signature. Broken signatures are likely
 756  to be detected quickly using current versions of git.
 757
 758However, it also came with disadvantages:
 759* Verifying a signed object requires access to the sha1-names of all
 760  objects it references, even after the transition is complete and
 761  translation table is no longer needed for anything else. To support
 762  this, the design added fields such as "hash sha1 tree <sha1-name>"
 763  and "hash sha1 parent <sha1-name>" to the newhash-content of a signed
 764  commit, complicating the conversion process.
 765* Allowing signed objects without a sha1 (for after the transition is
 766  complete) complicated the design further, requiring a "nohash sha1"
 767  field to suppress including "hash sha1" fields in the newhash-content
 768  and signed payload.
 769
 770Lazily populated translation table
 771~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 772Some of the work of building the translation table could be deferred to
 773push time, but that would significantly complicate and slow down pushes.
 774Calculating the sha1-name at object creation time at the same time it is
 775being streamed to disk and having its newhash-name calculated should be
 776an acceptable cost.
 777
 778Document History
 779----------------
 780
 7812017-03-03
 782bmwill@google.com, jonathantanmy@google.com, jrnieder@gmail.com,
 783sbeller@google.com
 784
 785Initial version sent to
 786http://public-inbox.org/git/20170304011251.GA26789@aiede.mtv.corp.google.com
 787
 7882017-03-03 jrnieder@gmail.com
 789Incorporated suggestions from jonathantanmy and sbeller:
 790* describe purpose of signed objects with each hash type
 791* redefine signed object verification using object content under the
 792  first hash function
 793
 7942017-03-06 jrnieder@gmail.com
 795* Use SHA3-256 instead of SHA2 (thanks, Linus and brian m. carlson).[1][2]
 796* Make sha3-based signatures a separate field, avoiding the need for
 797  "hash" and "nohash" fields (thanks to peff[3]).
 798* Add a sorting phase to fetch (thanks to Junio for noticing the need
 799  for this).
 800* Omit blobs from the topological sort during fetch (thanks to peff).
 801* Discuss alternates, git notes, and git servers in the caveats
 802  section (thanks to Junio Hamano, brian m. carlson[4], and Shawn
 803  Pearce).
 804* Clarify language throughout (thanks to various commenters,
 805  especially Junio).
 806
 8072017-09-27 jrnieder@gmail.com, sbeller@google.com
 808* use placeholder NewHash instead of SHA3-256
 809* describe criteria for picking a hash function.
 810* include a transition plan (thanks especially to Brandon Williams
 811  for fleshing these ideas out)
 812* define the translation table (thanks, Shawn Pearce[5], Jonathan
 813  Tan, and Masaya Suzuki)
 814* avoid loose object overhead by packing more aggressively in
 815  "git gc --auto"
 816
 817[1] http://public-inbox.org/git/CA+55aFzJtejiCjV0e43+9oR3QuJK2PiFiLQemytoLpyJWe6P9w@mail.gmail.com/
 818[2] http://public-inbox.org/git/CA+55aFz+gkAsDZ24zmePQuEs1XPS9BP_s8O7Q4wQ7LV7X5-oDA@mail.gmail.com/
 819[3] http://public-inbox.org/git/20170306084353.nrns455dvkdsfgo5@sigill.intra.peff.net/
 820[4] http://public-inbox.org/git/20170304224936.rqqtkdvfjgyezsht@genre.crustytoothpaste.net
 821[5] https://public-inbox.org/git/CAJo=hJtoX9=AyLHHpUJS7fueV9ciZ_MNpnEPHUz8Whui6g9F0A@mail.gmail.com/