Hackage Security

This is a library for Hackage security based on TUF, The Update Framework. In addition to the TUF website, the blog post that announced the framework is a good source of information.

Background Information

The Hackage security process is an implementation of The Update Framework (TUF), intended to stop software supply-chain attacks. TUF provides both index signing, which prevents mirrors or other middlemen from tampering with the contents of a software repository, and author signing, which prevents repositories from tampering with the contents of hosted packages. Thus far, however, Hackage implements only index signing. Additionally, Hackage differs somewhat from TUF's assumptions, and thus does things a little differently. Rather than attempting to describe the diff against TUF, this document simply describes how Hackage's security features work.

As an instance of TUF, the Hackage security process uses much of its jargon. In particular, a role refers to a manner of use of a particular key. A given key might, in principle, be used in multiple roles - for instance, the same key could be used to sign timestamps and mirrors. In this context, "the timestamp key" and "the mirror key" would refer to the same key used in two different ways. Each role is assigned a threshold that states how many signatures are expected for a given role. For instance, in Hackage, the timestamp role requires just one signature, while the root role requires three signatures. Similarly, keys are distinguished from key IDs, which are hashes of the key content. With the ed25519 keys used by this process, they are the same length, so be careful.

Hackage provides the following to build tools:

• An index, which contains the metadata for every package
• Packages, which contain the actual code
• A timestamp, with a frequently-updated signature that expires regularly

Malicious mirrors could attempt to interpose incorrect information into either. Hackage cryptographically signs the index, providing evidence of its authenticity, and the index itself contains hashes of each package file. Thus, mirrors cannot interpose new metadata or new packages, because both are secured by the signature. Additionally, these features prevent man-in-the-middle attacks against both Hackage and its mirrors, domain hijacking, and rollback attacks. The signed timestamp ensures that clients can detect replay attacks that are denying them new packages. They do not prevent malicious or compromised package authors from uploading malware to the index, nor do they protect against the Hackage server itself being compromised.

Brief overview of Hackage and cabal-install

Hackage makes all packages available from a single directory /package; for example, version 1.0 of package Foo is available at /package/Foo-1.0.tar.gz (see Footnote: Paths).

Additionally, Hackage offers a tarball, variously known as “the index” or “the index tarball”, located at /00-index.tar.gz. The index tarball contains the .cabal for all packages available on the server; cabal-install downloads this file whenever you call cabal update to figure out which packages are available, and it uses this file when you cabal install a package to figure out which dependencies to install. The .cabal file for Foo-1.0 is located at Foo/1.0/Foo.cabal in the index. (One side goal of the Hackage Security project is to make downloading the index incremental.)

Note that although Hackage additionally offers the (latest version of) the .cabal file at /package/Foo-1.0/Foo.cabal, this is never used by cabal-install.

In order to distinguish between paths on the server and paths in the index we will qualify them as <repo>/package/Foo-1.0.tar.gz and <index>/Foo/1.0/Foo.cabal respectively, both informally in this text and in formal delegation rules.

Formats and Tools

The files that are to be signed contain JSON objects that have two fields: signatures and signed. The Hackage signing tools will not sign any other format. The signatures field is expected to be an array of signatures, while the signed field may consist of arbitrary JSON. When signing, the signature is actually applied to the canonical JSON rendering of the contents of the signed field. This allows multiple signatures to be independently created and added, because new signatures do not sign the prior signatures.

There are two tools that are relevant: hackage-root-tool and hackage-repo-tool. hackage-root-tool is a minimal implementation of the cryptography, intended to be as small as possible so that it can be audited and run on an offline machine. hackage-repo-tool, on the other hand, has a number of features for managing file-based Hackage repositories in addition to signing.

Keys and Participants

The Root of Trust

The file root.json describes all of the keys used in Hackage Security. It lists the public parts of the keys, enumerates the roles used, and assigns keys and thresholds to each role. It is signed by keys in the root role.

Root Keys

The Hackage root keys are held by trusted members of the Haskell community. A signature is valid when three keyholders have signed. This means that the overall system is not vulnerable to a single key being compromised; nor can service be denied by a single key being lost. Keyholders are strongly encouraged to keep their keys very secure. The current collection of keyholders, plus signatures that demonstrate that they have the keys, is available at https://github.com/haskell-infra/hackage-root-keys .

Root keys present a bootstrapping problem: how is a build tool to know whether it should trust a newly-downloaded root.json? To work around this, the public part of the root keys and the threshold policy for the root role are shipped with the build tools that need to verify Hackage downloads. Once this root.json has been verified, its policies override the built-in bootstrapping keys.

Because these root keys are so difficult to replace, they are not used for operations. The root keys are used to delegate roles to a set of operational keys, and these operational keys are used for the daily signing of indices by Hackage.

Operational Keys

The operational keys are signed by the root keys. Build tools have no in-built knowledge of them, but can instead discover them through root.json. The operational private keys are kept secure by the Hackage administrators, but because they are on an online machine, they are more vulnerable than the root keys.

Operational keys fulfill two roles:

• Snapshot keys are used to sign the Hackage index.
• Timestamp keys are used to sign the frequently-updated timestamp file.

Mirrors

A list of authorized mirrors of Hackage is provided in a file called mirrors.json. This list is signed by the mirror key. The mirrors list expires annually and must be re-signed by the mirror key. Clients check that the mirror key is signed by the root key, and that the mirror list is signed by the mirror key, before accepting a mirror list. According to the TUF spec,

The importance of using signed mirror lists depends on the application and the users of that application. There is minimal risk to the application’s security from being tricked into contacting the wrong mirrors. This is because the framework has very little trust in repositories.

The mirror list being signed is mostly for the sake of completeness, rather than out of concern for a particular threat.

Comparison with TUF

In this section we highlight some of the differences in the specifics of the implementation; generally we try to follow the TUF model as closely as possible.

This section is not (currently) intended to be complete.

Targets

In the current proposal we do not yet implement author signing, but when we do implement author signing we want to have a smooth transition, and moreover we want to be able to have a mixture of packages, some of which are author signed and some of which are not. That is, package authors must be able to opt-in to author signing (or not).

Unsigned packages

Package specific metadata

The package metadata files (“target files”) will be stored in the index. The metadata for Foo 1.0 will be stored in <index>/Foo/1.0/package.json, and will contain the hash and size of the package tarball:

{ "signed" : {
     "_type"   : "Targets"
   , "version" : VERSION
   , "expires" : never
   , "targets" : { "<repo>/package/Foo-1.0.tar.gz" : FILEINFO }
   }
, "signatures" : []
}

Note that expiry dates are relevant only for information that we expect to change over time (such as the snapshot). Since packages are immutable, they cannot expire. (Additionally, there is no point adding an expiry date to files that are protected only the snapshot key, as the snapshot itself will expire).

It is not necessary to list the file info of the .cabal files here: .cabal files are listed by value in the index tarball, and are therefore already protected by the snapshot key (but see author signing).

Delegation

Conceptually speaking we then need a top-level target file <index>/targets.json that contains the required delegation information:

{ "signed" : {
      "_type"       : Targets
    , "version"     : VERSION
    , "expires"     : never
    , "targets"     : []
    , "delegations" : {
          "keys"  : []
        , "roles" : [
               { "name"      : "<index>/$NAME/$VERSION/package.json"
               , "keyids"    : []
               , "threshold" : 0
               , "path"      : "<repo>/package/$NAME-$VERSION.tar.gz"
               }
             ]
       }
    }
, "signatures" : /* target keys */
}

This file itself is signed by the target keys (kept offline by the Hackage admins).

Note that this file uses various extensions to TUF spec:

• We can use wildcards in names as well as in paths. This means that we list a single path with a single replacement name. (Alternatively, we could have a list of pairs of paths and names.)
• Paths contain namespaces (<repo> versus <index>)
• Wildcards have more structure than TUF provides for.

The first one of these is the most important, as it has some security implications; see comments below.

New unsigned packages, as well as new versions of existing unsigned packages, can be uploaded to Hackage without any intervention from the Hackage admins (the offline target keys are not required).

Security

This setup is sufficient to allow for untrusted mirrors: since they do not have access to the snapshot key, they (or a man-in-the-middle) cannot change which packages are visible or change the packages themselves.

However, since the snapshot key is stored on the server, if the server itself is compromised almost all security guarantees are void.

Signed packages

We sketch the design here only, we do not actually intend to implement this yet in phase 1 of the project.

Package specific metadata

As for unsigned packages, we keep metadata specific for each package version. Unlike for unsigned packages, however, we store two files: one that can be signed by the package author, and one that can be signed by the Hackage trustees, who can upload new .cabal file revisions but not change the package contents.

As before we still have <index>/Foo/1.0/package.json containing

{ "signed" : {
     "_type"   : "Targets"
   , "version" : VERSION
   , "expires" : never
   , "targets" : { "<repo>/package/Foo-1.0.tar.gz" : FILEINFO }
   }
, "signatures" : /* signatures from package authors */
}

It is not necessary to separately sign the .cabal file that is listed inside the package .tar.gz file. However, this .cabal file may not match the one in the index, either because a Hackage trustee uploaded a revision, or because of an malicious attempt to fool the solver in installing different dependencies than intended.

Therefore, unlike for unsigned packages, listing the file info for the .cabal file in the index is useful for signed packages: although the .cabal files are listed by value in the index tarball, the index is only signed by the snapshot key. We may want to additionally check that the .cabal are properly author signed too. We record this in a different file <index>/Foo/1.0/revisions.json, which can be signed by either the package authors or the Hackage trustees.

{ "signed" : {
     "_type"   : "Targets"
   , "version" : VERSION
   , "expires" : never
   , "targets" : { "<index>/Foo/1.0/Foo.cabal" : FILEINFO }
   }
, "signatures" : /* signatures from package authors or Hackage trustees */
}

Delegation

Delegation for signed packages is a bit more complicated. We extend the top-level targets file to

{ "signed" : {
      "_type"       : Targets
    , "version"     : VERSION
    , "expires"     : EXPIRES
    , "targets"     : []
    , "delegations" : {
          "keys"  : /* Hackage trustee keys */
        , "roles" : [
               // Delegation for unsigned packages
               { "name"      : "<index>/$NAME/$VERSION/package.json"
               , "keyids"    : []
               , "threshold" : 0
               , "path"      : "<repo>/package/$NAME-$VERSION.tar.gz"
               }
             // Delegation for package Bar
             , { "name"      : "<index>/Bar/authors.json"
               , "keyids"    : /* top-level target keys */
               , "threshold" : THRESHOLD
               , "path"      : "<repo>/package/Bar-$VERSION.tar.gz"
               }
             , { "name"      : "<index>/Bar/authors.json"
               , "keyids"    : /* top-level target keys */
               , "threshold" : THRESHOLD
               , "path"      : "<index>/Bar/$VERSION/Bar.cabal"
               }
             , { "name"      : "<index>/Bar/$VERSION/revisions.json"
               , "keyids"    : /* Hackage trustee key IDs  */
               , "threshold" : THRESHOLD
               , "path"      : "<index>/Bar/$VERSION/Bar.cabal"
               }
             // .. delegation for other signed packages ..
             ]
        }
    }
, "signatures" : /* target keys */
}

Since this lists all signed packages, we must list an expiry date here so that attackers cannot mount freeze attacks (although this is somewhat less of an issue here as freezing this list would make an entire new package, rather than a new package version, invisible).

This “middle level” targets file <index>/Bar/authors.json introduces the package author/maintainer keys and contains further delegation information:

{ "signed" : {
      "_type"       : "Targets"
    , "version"     : VERSION
    , "expires"     : never
    , "targets"     : {}
    , "delegations" : {
          "keys"  : /* package maintainer keys */
        , "roles" : [
              { "name"      : "<index>/Bar/$VERSION/package.json"
              , "keyids"    : /* package maintainer key IDs */
              , "threshold" : THRESHOLD
              , "path"      : "<repo>/Bar/$VERSION/Bar-$VERSION.tar.gz"
              }
            , { "name"      : "<index>/Bar/$VERSION/revisions.json"
              , "keyids"    : /* package maintainer key IDs */
              , "threshold" : THRESHOLD
              , "path"      : "<index>/Bar/$VERSION/Bar.cabal"
              }
            ]
    }
, "signatures" : /* signatures from top-level target keys */
}

Some notes:

1. When a new signed package is introduced, the Hackage admins need to create and sign a new targets.json that lists the package author keys and appropriate delegation information, as well as add corresponding entries to the top-level delegation file. However, once this is done, the Hackage admins do not need to be involved when package authors wish to upload new versions.

2. When package authors upload a new version, they need to sign only a single file that contains the information about that particular version.

3. Both package authors (through the package-specific “middle level” delegation information) and Hackage trustees (through the top-level delegation information) can sign .cabal file revisions, but only authors can sign the packages themselves.

4. Hackage trustees are listed only in the top-level delegation information, so when the set of trustees changes we only need to modify one file (as opposed to each middle-level package delegation information).

5. For signed packages that do not want to allow Hackage trustees to sign .cabal file revisions we can just omit the corresponding entry from the top-level delegations file.

6. There are two kinds of rule overlaps in these delegation rules: <repo>/package/Bar-1.0.tar.gz will match against the rule for unsigned packages (<repo>/package/$NAME-$VERSION.tar.gz) and against the rule for signed packages (<repo>/package/Bar-$VERSION.tar.gz). It is important here that the signed rule take precedence, because author signed packages must be author signed. The priority scheme can be simple: more specific rules should take precedence (the TUF specification leaves the priority scheme used open).

   The second kind of overlap occurs between the rule for the .cabal file in the the top-level targets.json and the corresponding rule in authors.json; in this case both rules match against precisely the same path (<index>/Bar/$VERSION/Bar.cabal), and indeed in this case there is no priority: as long as either rule matches (that is, the file is either signed by a package author or by a Hackage trustee) we're okay.

Transition packages from unsigned to signed

When a package that previously did not opt-in to author signing now wants author-signing, we just need to add the appropriate entries to the top-level delegation file and set up the appropriate middle-level delegation information.

Security

When the snapshot key is compromised, attackers still do not have access to package author keys, which are strictly kept offline. However, they can still mount freeze attacks on packages versions, because there is no file (which is signed with offline key) listing which versions are available.

We could increase security here by changing the middle-level authors.json to remove the wildcard rule, list all versions explicitly, and change the top-level delegation information to say that the middle-level file should be signed by the package authors instead.

Note that we do not use a wildcard for signed packages in the top-level targets.json for a similar reason: by listing all packages that we expect to be signed explicitly, we have a list of signed packages which is signed by offline keys (in this case, the target keys).

Snapshot

Interaction with the index tarball

According to the official specification we should have a file snapshot.json, signed with the snapshot key, which lists the hashes of all metadata files in the repo. In Hackage however we have the index tarball, which contains most of the metadata files in the repo (that is, it contains all the .cabal files, but it also contains all the various .json files). The only thing that is missing from the index tarball, compared to the snapshot.json file from the TUF spec, is the version, expiry time, and signatures. Therefore our snapshot.json looks like

{ "signed" : {
      "_type"   : "Snapshot"
    , "version" : VERSION
    , "expires" : EXPIRES
    , "meta"    : {
           "root.json"    : FILEINFO
         , "mirrors.json" : FILEINFO
         , "index.tar"    : FILEINFO
         , "index.tar.gz" : FILEINFO
        }
    }
, "signatures" : /* signatures from snapshot key */
}

Then the combination of snapshot.json together with the index tarball is a strict superset of the information in TUF's snapshot.json (instead of containing the hashes of the metadata in the repo, it contains the actual metadata themselves).

We list the file info of the root and mirrors metadata explicitly, rather than recording it in the index tarball, so that we can check them for updates during the update process (section 5.1, “The Client Application”, of the TUF spec) without downloading the entire index tarball.

Out-of-tarball targets

All versions of all packages are listed, along with their full .cabal file, in the Hackage index. This is useful because the constraint solver needs most of this information anyway. However, when we add additional kinds of targets we may not wish to add these to the index: people who are not interested in these new targets should not, or only minimally, be affected. In particular, new releases of these new kinds of targets should not result in a linear increase in what clients who are not interested in these targets need to download whenever they call cabal install.

To support these out-of-tarball targets we can use the regular TUF setup. Since the index does not serve as an exhaustive list of which targets (and which target versions) are available, it becomes important to have target metadata that list all targets exhaustively, to avoid freeze attacks. The file information (hash and filesize) of all these target metadata files must, by the TUF spec, be listed in the top-level snapshot; we should thus avoid introducing too many of them (in particular, we should avoid requiring a new metadata file for each new version of a particular kind of target).

It is important to store OOT targets under a different prefix than /package to avoid name clashes.

Collections

Package collections are a new Hackage feature that's currently in development. We want package collections to be signed, just like anything else.

Like packages, collections are versioned and immutable, so we have

collection/StackageNightly-2015.06.02.collection
collection/StackageNightly-2015.06.03.collection
collection/StackageNightly-...
collection/DebianJessie-...
collection/...

As for packages, collections should be able to opt-in for author signing (once we support author signing), but we should also support not-author-signed (“unsigned”) collections. Moreover, it should be possible for people to create new unsigned collections without the involvement of the Hackage admins. This rules out listing all collections explicitly in the top-level targets.json (which is signed with offline target keys).

Below we sketch a possible design (we may want to tweak this further).

Unsigned collections

For unsigned collections we add a single delegation rule to the top-level targets.json:

{ "name"      : "<repo>/collection/collections.json"
, "keyids"    : /* snapshot key */
, "threshold" : 1
, "path"      : "<repo>/collection/$NAME-$VERSION.collection"
}

The middle-level collections.json, signed with the snapshot role, lists delegation rules for all available collections:

[ { "name"      : "<repo>/collection/StackageNightly.json"
  , "keyids"    : /* snapshot key */
  , "threshold" : 1
  , "path"      : "<repo>/collection/StackageNightly-$VERSION.collection"
  }
, { "name"      : "<repo>/collection/DebianJessie.json"
  , "keyids"    : /* snapshot key */
  , "threshold" : 1
  , "path"      : "<repo>/collection/DebianJessie-$VERSION.collection"
  }
, ...
]

where StackageNightly and DebianJessie are two package collection (the Stackage Nightly collection and the set of Haskell package distributed with the Debian Jessie Linux distribution).

The final per-collection targets metadata finally lists all versions:

{ "signed" : {
     "_type"   : "Targets"
   , "version" : VERSION
   , "expires" : /* expiry */
   , "targets" : {
         "<repo>/collection/StackageNightly-2015.06.02.collection" : FILEINFO
       , "<repo>/collection/StackageNightly-2015.06.03.collection" : FILEINFO
       , ...
       }
   }
, "signatures" : /* signed with snapshot role */
}

Since we cannot rely on the index to have the list of all version we must list all versions explicitly here rather than using wildcards. Note that this means that the snapshot will get a new entry for each new collection introduced, but not for each new version of each collection.

Author-signed collections

For author-signed collections we only need to make a single change. Suppose that the DebianJessie collection is signed. Then we move the rule for DebianJessie from collection/collections.json and instead list it in the top-level targets.json (as for packages, introducing a signed collection necessarily requires the involvement of the Hackage admins):

{ "name"      : "<repo>/collection/DebianJessie.json"
, "keyids"    : /* DebianJessie maintainer keys */
, "threshold" : /* threshold */
, "path"      : "<repo>/collection/DebianJessie-$VERSION.collection"
}

No other changes are required (apart from of course that <repo>/collection/DebianJessie.json will now be signed with the DebianJessie maintainer keys rather than the snapshot key). As for packages, this requires a priority scheme for delegation rules.

(One difference between this scheme and the scheme we use for packages is that this means that the top-level targets.json file lists all keys for all collection authors. If we want to avoid that we need a further level of indirection.)

Note that in a sense author-signed collections are snapshots of the server. As such, it would be good if these collections listed the file info (hashes and filesizes) of the packages the collection.

Project phases and shortcuts

Phase 1 of the project will implement the basic TUF framework, but leave out author signing; support for author signed packages (and other targets) will added in phase 2.

Shortcuts taken in phase 1 (aka TODOs for phase 2)

This list is currenty not exhaustive.

Core library

• Although the infrastructure is in place for target metadata, including typed data types representing pattern matches and replacements, we have not yet actually implementing target delegation proper. We don't need to: we only support one kind of target (not-author-signed packages), and we know statically where the target information for packages can be found (in /package/version/targets.json). This is currently hardcoded.

  Once we have author signing we need a proper implementation of delegation, including a priority scheme between rules. This will also involve lazily downloading additional target metadata.

• Out-of-tarballs targets are not yet implemented. The main difficulty here is that they require a proper implementation of delegation; once that is done (required anyway for author signing) support for OOT targets should be straightforward.

Integration in cabal-install

• The cabal integration uses the hackage-security library to check for updates (that is, update the local copy of the index) and to download packages (verifying the downloaded file against the file info that was recorded in the package metadata, which itself is stored in the index). However, it does not use the library to get a list of available packages, nor to access .cabal files.

  Once we have author signing however we may want to do additional checks:

  • We should look at the top-level targets.json file (in addition to the index) to figure out which packages are available. (The top-level targets file, signed by offline keys, will enumerate all author-signed packages.)

  • If we do allow package authors to sign list of package versions (as detailed above) we should use these “middle level” target files to figure out which versions are available for these packages.

  • We might want to verify the .cabal files to make sure that they match the file info listed in the now author-signed metadata.

  Therefore cabal-install should be modified to go through the hackage-security library get the list of available packages, package versions, and to access the actual .cabal files.

Open questions / TODOs

• The set of maintainers of a package can change over time, and can even change so much that the old maintainers of a package are no longer maintainters. But we would still like to be able to install and verify old packages. How do we deal with this?

Ongoing Maintenance

Mirror Keys: Every Year

The mirror list requires annual resigning by a holder of a mirror key. To do this, use the following steps:

1. Install hackage-root-tool on the signing machine and ensure that the key is present.
2. Create a new mirrors.json file by incrementing the version field of the existing file and adding a year to the expiration date. Delete the signature(s), replacing them with an empty array. Place the file on the signing machine.
3. Sign the file using hackage-root-tool sign KEY mirrors.json, and place the resulting signature array into the signatures field of mirrors.json.
4. Commit the updated file to https://github.com/haskell-infra/hackage-root-keys and inform the Hackage admins so they can install it.

Root Data: Every Other Year

The holders of the root keys are, each year, signing the root information file root.json that directs clients to the operational and mirror keys. The keyholders are attesting that they believe that the root and operational keys are not compromised, that Hackage is still under the control of trusted administrators, and that everything is working about the way it usually does.

Today, there are five active root keys, because three of the original eight never completed the setup process.

To prepare the updated roots.json for signing by the root keyholders, a coordinator should perform the following edits:

1. If any new keys are to be admitted, collect their key IDs and add them to the keys field.
2. Modify the expiration date.
3. Increment the version field.
4. Delete the existing signatures.

Each holder of root keys should do the following:

1. Verify that the contents of root.json are correct. Any changes to the set of keys and their roles or to the threshold policies should be trustworthy, and any new keyholders need to have their identity verified. Please be a stickler for details here.
2. Install hackage-root-tool on the signing machine and ensure that the key is present.
3. Place the updated roots.json file on the signing machine.
4. Sign the file using hackage-root-tool sign KEY roots.json, and send the resulting signature back to the person who is coordinating the signing.

Finally, the coordinator should insert the provided signatures and commit the updated file to https://github.com/haskell-infra/hackage-root-keys and inform the Hackage admins so they can install it.

Operational Keys

The operational keys do not presently require regeneration, unless the private keys have been lost or compromised.

Footnotes

Footnote: Paths

The situation with paths in cabal/hackage is a bit of a mess. In this footnote we describe the situation before the work on the Hackage Security library.

The index tarball

The index tarball contains paths of the form

<package-name>/<package-version>/<package-name>.cabal

For example:

mtl/1.0/mtl.cabal

as well as a single top-level preferred-versions file.

Package resources offered by Hackage

Hackage offers a number of resources for package tarballs: one “official” one and a few redirects.

1. The official location of package tarballs on a Hackage server is

   /package/<package-id>/<package-id>.tar.gz
   

   for example

   /package/mtl-2.2.1/mtl-2.2.1.tar.gz
   

   (This was the official location from very early on).

2. It provides a redirect for

   /package/<package-id>.tar.gz
   

   for example

   /package/mtl-2.2.1.tar.gz
   

   (that is, a request for /package/mtl-2.2.1.tar.gz will get a 301 Moved Permanently response, and is redirected to /package/mtl-2.2.1/mtl-2.2.1.tar.gz).

3. It provides a redirect for Hackage-1 style URLs of the form

   /packages/archive/<package-name>/<package-version>/<package-id>.tar.gz
   

   for example

   /packages/archive/mtl/2.2.1/mtl-2.2.1.tar.gz
   

Locations used by cabal-install to find packages

There are two kinds of repositories supported by cabal-install: local and remote.

1. For a local repository cabal-install looks for packages at

   <local-dir>/<package-name>/<package-version>/<package-id>.tar.gz
   

2. For remote repositories however cabal-install looks for packages in one of two locations.

   a. If the remote repository (<repo>) is http://hackage.haskell.org/packages/archive (this value is hardcoded) then it looks for the package at

   <repo>/<package-name>/<package-version>/<package-id>.tar.gz
   

   b. For any other repository it looks for the package at

   <repo>/package/<package-id>.tar.gz
   

Some notes:

1. Files downloaded from a remote repository are cached locally as

   <cache>/<package-name>/<package-version>/<package-id>.tar.gz
   

   I.e., the layout of the local cache matches the layout of a local repository (and matches the structure of the index tarball too).

2. Somewhat bizarrely, when cabal-install creates a new initial config file it uses http://hackage.haskell.org/packages/archive as the repo base URI (even in newer versions of cabal-install; this was changed only very recently).

3. However, notice that even when we give cabal a “new-style” URI the address used by cabal still causes a redirect (from /package/<package-id>.tar.gz to /package/<package-id>/<package-id>.tar.gz).

The most important observation however is the following: It is not possible to serve a local repository as a remote repository (by pointing a web server at a local repository) because the layouts are completely different. (Note that the location of packages on Hackage-1 did match the layout of local repositories, but that doesn't help because the only repository that cabal-install will regard as a Hackage-1 repository is one hosted on hackage.haskell.org).