Reserved Bits in Vector/String/Array Count Fields
FIDL vectors (and strings) use a 64-bit count field to represent the number of encoded bytes. This FTP proposes a few changes:
Vector, string and array counts max out at 32-bits (max value 232-1).
The upper 8-bits of vector/string count fields are reserved for use by bindings for in-memory use. They must be filled with zero on the wire.
The next 24-bits of vector/string count fields are reserved but are unused and must be filled with zero on the wire. They cannot be used by bindings and may be reallocated in a future FTP.
A visualization of the breakdown of the existing 64-bit count field.
This would decrease the maximum vector, string and array encoding size from 18.45 exabytes to 4.29 gigabytes.
The LLCPP bindings enable in-place encoding of objects. As part of the LLCPP Builder effort, the bindings now keep track of memory ownership to simplify object creation for users. In particular, for vectors, the most significant bit (MSB) of the count field is used to store ownership information (see vector_view.h). This bit is zeroed during linearization and doesn’t affect the FIDL wire format, but it does prevent the LLCPP bindings from using the MSB for count values.
The goal of this is to formalize the reservation of the MSB of the vector’s count field for use by bindings and extend the reservation to the upper 8-bits.
Separately, there is some inconsistency in maximum counts between bindings. Various parts of the C++ language bindings and compiler assume a maximum size of 32-bits, but this size was never formalized and isn’t followed by other bindings. This FTP formalizes a maximum size of 32-bits for vector/string/array counts.
The upper 8-bits of the 64-bit vector (and string) count field are reserved for in-memory use by bindings. The upper 8-bits MUST be zero when sent on the wire.
In addition, the next uppermost 24 bits (bits 32-55) are reserved and must be unused by the bindings. They also MUST be zero on the wire.
The ABI is unchanged.
Encode and decode logic in each of the bindings will be updated to validate that the count is at most 2^32-1. This includes validating that the upper bits are zero. The channel will be closed if this constraint is violated during decoding.
Documentation and Examples
The wire format documentation will be updated to show the reserved bits.
Channels max out at a byte size of 65536 (16-bits) per message so parts of the system that FIDL-encode into channel messages won’t have any compatibility issues. Additionally, a vector/string/array max count of 232-1 is already assumed by some parts of the code.
There is negligible performance impact. The only implementation change is additional validation checks.
This should not introduce any security risks. The count field will have the same value over the wire and there are now additional validation checks.
Each binding implementation is responsible for testing its implementation of the feature, including testing validation checks.
After this FTP, bindings are able to use the upper 8 bits of the count field for any purpose. Therefore, it is difficult or impossible to reclaim those bits for a new purpose - it would require exhaustively validating that each binding leaves the bits unused and potentially migrating them away from using these bits.
Instead of storing vector ownership information in the count field, it could instead be stored inside of the pointer field of the vector.
Two possible approaches:
Assume >= 2 byte alignment and store an ownership boolean in the least-significant-bit. The problem with this approach is that the assumption of 2-byte alignment is hard to enforce and often not true. There are places in the codebase where bytes are read from arbitrary offsets in a buffer.
Use properties of the address space to assign unused bits to hold memory ownership information. The problem with this is that there are no guarantees that bits are unused in the memory space. Even if this is true today, tools like address sanitizer tend to store information in available bits in pointers and there could be a collision between those tools and LLCPP.
Separately, there was discussion around the number of bits to reserve for bindings and how large the vector count should be. An argument for a larger vector count is that it is hard to predict future needs and other systems have run into issues with arbitrary size limitations. That said, there are few realistic use cases for writing a FIDL object greater than 4.29 gb. Encoding or decoding an object with this many elements would take over a second under the outrageous assumption of 1/cycle per element on current CPUs. Large primitive arrays, which can be encoded more efficiently should generally be sent with alternative mechanisms such as VMOs rather than FIDL. Still, reserving too many bits for bindings might limit future use cases. As a compromise, 8 bits are reserved for bindings, 24 bits unused and 32 bits for the count. This allows for future changes as needs arise.
Prior Art and References