Increase CPI Account Infos Limit
Summary
Increase the maximum number of account info's for cross-program invoked (CPI) instructions from 64 to 255 and consume compute units for serialized account info's and instruction account meta's.
Motivation
CPI's are restricted to a limit of 64 account info's passed to the syscall. This limit is burdensome for onchain programs which themselves were invoked with more than 64 accounts because it means they cannot simply pass through the same list of account info's that they were invoked with to another CPI syscall. They are faced with the burden of first deduplicating the account info's and constructing a new list before making the syscall.
This problem arises when onchain programs wrap another program (such as Jupiter) that composes many other programs and are invoked with over 64 accounts (including duplicates).
Along with the increasing the account info limit, start charging for both serialized account info's and instruction account meta's to avoid abuse of resources in CPI calls.
New Terminology
- Account Info: The serialized information for each account referenced in a CPI instruction used to read account info from the caller program.
Detailed Design
Maximum Account Info Length
Increase the maximum account info length imposed on CPI syscalls from 64 to 255. The maximum is inclusive, meaning that a list of account info's with a length of 255 is valid.
Account Info Cost
Consume 1 compute unit (CU) for every cpi_bytes_per_unit (currently 250) bytes of account info.
Fixed size of 80 bytes for each account_info broken down as
- 32 bytes for account address
- 32 bytes for owner address
- 8 bytes for lamport balance
- 8 bytes for data length
The total cost of account info's can be computed with (account_info_size * account_infos_len) / cpi_bytes_per_unit rounded down to the nearest CU.
Note that the SVM can skip duplicate or unused account info's but for simpler CU accounting, each account info has the same effect on cost regardless of whether it's actually used by the invoked instruction.
Instruction Account Metadata Cost
Consume 1 compute unit (CU) for every cpi_bytes_per_unit (currently 250) bytes of instruction account metadata.
Fixed size of 34 bytes for each ix_account_metadata broken down as
- 32 bytes for account address
- 1 byte for signer flag
- 1 byte for writable flag
The total cost of instruction account metadata can be computed with (ix_account_metadata_size * ix.accounts.len()) / cpi_bytes_per_unit rounded down to the nearest CU.
Reduce base CPI cost
The base cost of CPI syscalls is currently set to 1000 compute units. In order to avoid charging more CU's for existing onchain program behavior after the activation of this feature, the base cost of CPI syscalls will decreased by 54 compute units to 946 compute units to offset the new costs.
Before this SIMD, up to 255 instruction account metadata's could be passed to a CPI syscall. The base cost offset is therefore 255 * 34 / 250 = 34 CU's.
Before this SIMD, up to 64 account info's could be passed to a CPI syscall. The base cost offset is therefore 64 * 80 / 250 = 20 CU's.
Alternatives Considered
What alternative designs were considered and what pros/cons does this feature have relative to them?
Impact
Since the list of account info's passed to a CPI can now be ~4 times longer, there will be more overhead in the SVM to map each instruction account address to one of the account info's.
The number of allowed unique account info's remains unchanged so the maximum amount of accounts that need to be translated for a CPI syscall has not increased.
CPI's will consume compute units proportional to the number of account info's and instruction accounts. This gives an extra incentive to onchain programs to decrease the number of accounts passed to CPI's.
Security Considerations
NA
Drawbacks (Optional)
Why should we not do this?
Backwards Compatibility (Optional)
The max account info length increase for CPI's and new costs will be feature gated. Since the limit is being increased, existing behavior will not be restricted.