Internals of mamba#

Mamba comes with a C++ core (for speed and efficiency), and a clean Python API on top. The core of mamba uses:

  • libsolv to solve dependencies (the same library used in RedHat dnf and other Linux package managers)

  • curl for reliable and fast downloads

  • libarchive to extract the packages

libsolv#

pool#

The pool holds pointers of almost all the data manipulated in libsolv.

Some are presented here:

  • string pool Stringpool

  • relations of dependency Reldep

  • solvables/packages Solvable

  • repositories Repo

  • data about providers of a specific name or relation, as Id s and Offset s (resp. int and unsigned int) for maximum performance

Sizes of allocated memory or offset to the first free slot are also stored.

classDiagram class Pool{ +Stringpool ss +Reldep *rels +int nrels +Repo **repos +int nrepos +int urepos +Repo *installed +Solvable *solvables +int nsolvables +Offset *whatprovides +Offset *whatprovides_rel +Id *whatprovidesdata +Offset whatprovidesdataoff +int whatprovidesdataleft ... }

Ids#

Strings (package names, requirements, etc.) are converted to Id s using a hash table for efficient data manipulation (storage, usage).

  • Id pool_str2id(Pool *pool, const char *str, int create) is used to get an Id from a string

    • It eventually creates a fresh one if not existing

  • const char *pool_id2str(const Pool *pool, Id id) is used to get back the string from Id

Warning

Do not use solvable Id s with pool_id2str, use the equivalent const char *pool_solvid2str(Pool *pool, Id p)

Reldeps#

A Reldep describes a relation of dependency with:

  • a name Id

  • an epoque version evr Id

  • some flags: REL_CONDA in our case

classDiagram class Reldep{ +Id name +Id evr +int flags }

An example could be {1, 2, 30}, with:

  • 1 the Id for "xtensor"

  • 2 the Id for ">=0.20"

  • 30 the REL_CONDA flag

The pool holds a Reldep *rels pointer on the first memory location and the size int nrels of allocated memory.

Reldep are also handled with Id s that have bit 31 used as a flag to distinguish them from regular Id s.

There are multiple macros that help to convert those special Id s:

  • MAKERELDEP(id): turn a regular Id into a Reldep Id

  • ISRELDEP(id): test the bit 31

  • GETRELID(id): get the regular Id

  • GETRELDEP(pool, id): get the Reldep from its Id

Note

pool_id2str also works with Reldep Id s! But it will only returns the Reldep ‘s name

Offsets#

An Offset represents a positive or negative shift of a pointer on an array.

For example, a solvable does not contain all its data but rather holds multiple offsets on its repo->idarraydata storage.

  • idarraydata is an Id pointer

  • provides, requires, etc. are offsets in idarraydata

whatprovides#

A provider is a solvable fulfilling a specification. The following definitions are key to disambiguate how libsolv works:

  • a package is an identification of the resource handled: a name such as xtensor

  • a solvable is a specific version of the package. It can be assimilated to its tarball.

    • in Mamba, a solvable name MUST match the package name

    • libsolv handles cases where solvables are providing different entities than what identified in their names (example: pynum providing numpy)

      • this is not used, but important to know to understand the terminology

  • a specification: an expression to match solvables providing the same package

    • the package name can be used to select all providers/solvables

    • a more specific spec like s2="xtensor>=0.20" will only match a subset of solvables: the ones that have version >=0.20 (whatever the build string)

Let’s take a simple example to recap:

  • the package: xtensor

  • solvable(s):

    • xtensor=0.20.10=hc9558a2_0

    • xtensor=0.23.10=h4bd325d_0

    • etc.

  • specification(s):

    • xtensor

    • xtensor>=0.20

    • etc.

Note

It’s possible that a package is not provided by any solvable. It is then uninstallable.

%%{init: {'themeVariables':{'edgeLabelBackground':'white'}}}%% graph TD subgraph " " spec1((s1)) -.-> solvable1:::solvable spec1 -.-> solvable2:::solvable ...:::empty spec1 -.-> solvableN:::solvable spec2((s2)) -.-> solvableN:::solvable end classDef solvable fill:#f96; classDef empty fill:#ffffff00,stroke:#ffffff00;

Note

Specifications are stored as Id s (see the previous section)

Warning

While a solvable is a libsolv notion, specs and packages are not.

Storage#

The free function void pool_createwhatprovides(Pool *pool) is used to create hashes over pool of solvables to ease provide lookups.

It computes and store what solvables provide each spec, using a two-step indirect list:

  • from the spec Id, an offset is computed in Id *whatprovidesdata

    • Offset *whatprovides stores regular string specs

    • Offset *whatprovides_rel stores Reldep specs

  • whatprovidesdata at the given offset is a 0-terminated list of solvables Id s, providing the spec Id

../_images/whatprovides.svg

Lookups#

The pool_whatprovides(Pool *pool, Id d) function returns the offset of the first solvable Id in whatprovidesdata:

graph LR A{"ISRELDEP(d)?"} -->|Yes| B1["whatprovides[d]"]; A -->|No| B2["whatprovides_rel[GETRELID(d)]"]; B1 --> C{0?}; B2 --> C; C -->|Yes| D1["pool_addrelproviders[d]"]; C -->|No| D2((return)); D1 --> D2;

Rules#

A rule is all about solvables, it represents a logical disjunction OR between one or more literals.

Rules are created to translate in mathematical logic:

  • a specification: installation/removal/updates

    • (A) means A must be installed

    • (-A) means A must be removed or kept uninstalled

    • (A|B1|B2|...) means A could be updated with B1, B2, etc.

  • a dependency: A needs/requires b (a spec)

    • (-A|B1|B2|...) means A requires one of B1, B2, etc. (b providers)

  • an incompatibility: A1 can’t be installed alongside A2

    • (-A1|-A2), (-A1|-A3), ... means A1 is not compatible with A2, A3, etc.

    • this is a common case: multiple providers of the same package

  • etc.

Still for efficiency, rules are storing Id s of solvables and specs:

  • p is the package Id of A

  • d is the package Id offset into the list of providers (negative value means the rule is disabled)

  • w1 and w2 are watches

  • n1 and n2 are the next rules in linked-lists, corresponding to w1 and w2

classDiagram class Rule{ +Id p +Id d +Id w1 +Id w2 +Id n1 +Id n2 }

Dependencies#

Each dependency is turned into a rule to be satisfied during the solving stage.

Example:

  • p1 and p2 are 2 specs

  • p1 is provided by a single solvable s11

  • p2 is provided by s21 and s22

%%{init: {'themeVariables':{'edgeLabelBackground':'white'}}}%% graph LR subgraph p1 providers[ ] p1((p1)):::package -.-> s11:::solvable end s11:::solvable --> p2((p2)):::package subgraph p2 providers[ ] p2((p2)):::package -.-> s21:::solvable p2((p2)):::package -.-> s22:::solvable end classDef solvable fill:#f96;

The corresponding rule r is -s11|s21|s22.

That means that taking the decision to install s11:

  • -s11 is not satisfied

  • either s21 or s22 need to be satisfied

Note

Exclusive rules are used to avoid installation of multiple solvables providing the same package

Watches#

Watches are a way to link rules. They are triggered during a phase called propagation after decisions taken on previous rules for some solvable.

The possible decisions on solvables are installation or removal/conflict, this is stored as resp. positive and negative Ids.

Related rules are then evaluated during another level of decision: those are the one with an opposite first litteral.

Example:

  • the install spec is to install p1 provided by a single solvable s1: r1=(s1)

  • s1 depends on spec p2, provided by s21 or s22: r2=(-s1|s21|s22)

  • decision to install s1 triggers rule r2

Transaction#

Another important part of libsolv is the Transaction. A transaction governs what packages are installed or removed, and a transaction is the result of a successful solve process.

A transaction in libsolv is a single list (Queue) of Solvable Id's and is thus rather simple. The Queue contains either a positive or negative Id. Each negative Id is uninstalled from the environment, and each positive Id is to be installed. libsolv classifies the entire range of Id’s into different types of transaction operations. For example, if we have { -5, 5 } that would be a reinstall transaction for the Solvable with Id == 5. If the Id's are different then it can be a downgrade or upgrade operation (first, the previous package needs removal before the higher or lower version can be installed). The transaction_classify and transaction_classify_pkgs functions of libsolv take care of this classification to present a nice output to the user.

Another crucial libsolv function is transaction_order to order the transaction in a way that they are installed with the lowest dependency first (topological sort). This ensures that e.g. python is installed before any packages depending on python as they are sometimes needed during installation (for example for noarch packages with entry_points).

Lastly, we can force installation or explicitly install from URL’s by crafting transactions without using the solver – just by adding the correct Id's into the Transaction queue.