This page summarizes the key design principles that FIDL has adopted over time.
FIDL aims to respect the following priority of constituencies:
This list was adapted from that of the API Council Charter.
From RFC-0050: Syntax revamp:
FIDL is primarily concerned with defining Application Binary Interface (ABI) concerns, and second with Application Programming Interface (API) concerns.
From RFC-0050: Syntax revamp:
While many formats can represent FIDL messages, the FIDL Wire Format (or “FIDL Binary Wire Format”) is the one which has preferential treatment, and is catered to first ... we choose to over rotate towards the binary ABI format when making syntax choices.
From RFC-0131: Design principles of the FIDL wire format:
When faced with making a design tradeoff to support low level programming at the expense of high level programming (or the reverse), we typically opt for enabling low level programming.
From RFC-0050: Syntax revamp:
We strive to have the fewest features and rules, and aim to combine features to achieve use cases. In practice, when considering new features, we should first try to adapt or generalize other existing features rather than introduce new features.
From RFC-0027: You only pay for what you use:
When adding functionality to FIDL, we should evaluate the costs that adding that functionality imposes on people who use FIDL but do not use the new functionality. We should then have a very high bar for accepting functionality that imposes costs on people who do not use the functionality.
For example, RFC-0047: Tables followed this principle by adding tables to the language rather than replacing structs:
Tables are necessarily more complicated than structs, and so processing them will be slower and serializing them will use more space. As such, it's preferred to keep structs as is and introduce something new.
In contrast, RFC-0061: Extensible unions reached the opposite decision of replacing static unions with extensible unions, but only after a careful analysis of the tradeoffs. Unlike with tables, the extra cost imposed by extensible unions is marginal or nonexistent in most cases.
We design FIDL to solve real problems and address real needs, not imagined ones. We avoid designing a “solution in search of a problem”.
For example, FIDL initially did not support empty structs because it was unclear how to represent them in C/C++. In RFC-0056: Empty structs, we saw users were employing workarounds and recognized the need for an official solution. Only then did we add empty structs to the language.
Optimizing without data is useless at best and dangerous at worst. When designing optimizations (e.g. performance, binary size), we follow the data.
For example, RFC-0032: Efficient envelopes was initially accepted, but later rejected. In hindsight, at the time it should not have been accepted because there was no data to back it up. Later, it was re-proposed and accepted as RFC-0113: Efficient envelopes after there was data demonstrating a significant performance improvement.
Similarly, there was significant momentum behind using a sparser representation of table data. However, after investigation, there was an unfavorable tradeoff between design complexity and performance that led to a decision to not move forward (see RFC-0116: Sparser Tables). Without a significant prototyping and data-gathering phase, sparser tables would likely have been adopted and negatively impacted Fuchsia.
We strive to avoid breakage at a distance. Changes in one place should not cause surprising breakage in a faraway place. For example, it would be surprising if adding a struct named Foo
to a FIDL file broke compilation because another FIDL file in a completely different part of the codebase already had a type named Foo
. This is why FIDL, like most programming languages, uses namespaces to limit the scope of name collisions.
RFC-0029: Increasing method ordinals discusses this problem as it relates to protocol composition. RFC-0048: Explicit union ordinals revisits the topic, explaining why FIDL only uses hashing for protocols.
RFC-0057: Default no handles introduced a distinction between value and resource types. One motivation for this was providing the Clone
trait in Rust for types without handles without breakage at a distance:
Although FIDL bindings can conditionally enable code based on the presence of handles, doing so is undesirable because it breaks evolvability guarantees. For example, adding a field to a table is normally safe, but adding a handle field would become source-breaking — not only for that table, but for all types transitively containing it.
RFC-0149: FIDL encode validation not mandatory also touches upon this topic. It breaks down the classes of breakages that can occur and moves the decision of whether to perform encode-side validation to bindings. This allows there to be a more fine-grained discussion of the pros and cons of overhead from encode-side validation vs the risk of breakage at a distance.
We do not rigidly adhere to a “one way to do it” philosophy. When we are concerned that users will waste time deciding between trivial alternatives, we introduce restrictions in fidl-lint
or fidl-format
rather than in fidlc
.
For example, FIDL accepts modifier keywords in any order, but we intend to enforce a consistent ordering in the linter.
As another example, RFC-0040: Identifier uniqueness fixed the problem of identifiers colliding after case transformation by having fidlc
report an error if any two identifiers have the same canonical form. A simpler alternative would have been to enforce FIDL naming conventions in the compiler. However, this goes a step too far. There are valid reasons for using different naming styles, for example in describing the Kernel API, where snake_case
methods are strongly preferred.
From RFC-0131: Design principles of the FIDL wire format:
There must be a single unambiguous representation of a FIDL value, i.e. there is one and only one encoded representation of a FIDL value, and one and only one decoded representation of a FIDL value.
The FIDL wire format is canonical: there is exactly one encoding for a given message. As a corollary, every byte is accounted for: there is no byte that can be changed without altering the message's meaning.
For example, the specification requires that all padding bytes are zero. Similarly, RFC-0047: Tables disallows storing extraneous empty envelopes to ensure a canonical representation.
A canonical representation makes FIDL simpler and more secure. For example, allowing nonzero padding could result in FIDL messages leaking sensitive information that happened to occupy those bytes in memory. Allowing multiple representations for a given message also leads to rarely executed code paths that can hide bugs, e.g. the “extra empty envelopes” code path. A canonical representation also makes it easy to compare messages for equality without knowing the schema: for value types, it is a simple memcmp
.
From RFC-0131: Design principles of the FIDL wire format:
It must be possible to encode and decode in a single pass, without allocation beyond stack space (i.e. no dynamic heap allocation).
This requirement significantly influences the design of the wire format: you must be able to decode in place using only the stack. It is the reason the wire format uses presence indicators and a depth-first traversal order rather than, for example, and offset-based format that requires auxiliary data structures to keep track of information while decoding.
This principle is related to “You only pay for what you use”, in that it caters to very low-level uses of FIDL where malloc
may not yet exist, or is prohibitively expensive.
From RFC-0131: Design principles of the FIDL wire format:
Without explicit opt-in, a peer must not be allowed to perform reflection on a protocol, be it exposed methods, or exposed types.
While the binary wire format comes first, this does not mean FIDL should be tightly coupled to the Zircon channel transport. There are other important use cases to consider, such as describing the Kernel API, in-process messaging, and persistence.
RFC-0050: Syntax revamp describes the future direction for transport generalization.
RFC-0062: Method impossible was rejected in part because it coupled FIDL too closely to the Zircon channel transport.