Computing Poseidon[0] hashes is too expensive to be done in a Solana
program in one transaction. Poseidon is a zero-knowlege proof friendly
hash function, used by the majority of ZK-based projects, including the
ones built on top of Solana.
This change introduces the `sol_poseidon` syscall which takes 2D byte
slice as an input and then calculates a Poseidon hash using a BN254
curve and the following Poseidon parameters:
* x^5 S-boxes
* width - 2 ≤ t ≤ 13
* inputs - 1 ≤ n ≤ 12
* 8 full rounds and partial rounds depending on t: [56, 57, 56, 60, 60,
63, 64, 63, 60, 66, 60, 65]
Computation of Poseidon hashes is done with the light-poseidon[1]
crate, which is audited[2] and compatible with Circom[3] (BN254 curve,
the same parameters and constants).
Proposed compute costs depend on number of inputs and are based on
light-poseidon benchmarks[4].
[0] https://www.poseidon-hash.info/
[1] https://crates.io/crates/light-poseidon
[2] https://github.com/Lightprotocol/light-poseidon/blob/main/assets/audit.pdf
[3] https://docs.circom.io/
[4] https://github.com/Lightprotocol/light-poseidon/tree/main#performance
- BankForks is not an optional argument, so remove dated comment
- Given that BankForks is always present, no need for special values to
initialize variables before the loop
- Root slot can be retrieved from root bank, no need to call
BankForks::root() which will load the underlying atomic a second time
- Use BankForks::highest_slot() instead of .slot() on .working_bank() to
avoid the extra clone that .working_bank() performs
- Move several operations outside of BankForks read lock scope to
minimize lock time
* transaction_context: update make_data_mut comment
* bpf_loader: cpi: pass SerializeAccountMetadata to CallerAccount::from*
We now have a way to provide CallerAccount with trusted values coming
from our internal serialization code and not from untrusted vm space
* bpf_loader: direct_mapping: enforce account info pointers to be immutable
When direct mapping is enabled, we might need to update account data
memory regions across CPI calls. Since the only way we have to retrieve
the regions is based on their vm addresses, we enforce vm addresses to
be stable. Accounts can still be mutated and resized of course, but it
must be done in place.
This also locks all other AccountInfo pointers, since there's no legitimate
reason to make them point to anything else.
* bpf_loader: cpi: access ref_to_len_in_vm through VmValue
Direct mapping needs to translate vm values at each access since
permissions of the underlying memory might have changed.
* direct mapping: improve memory permission tracking across CPI calls
Ensure that the data and realloc regions of an account always track the
account's permissions. In order to do this, we also need to split
realloc regions in their own self contained regions, where before we
had:
[account fields][account data][account realloc + more account fields + next account fields][next account data][...]
we now have:
[account fields][account data][account realloc][more account fields + next account fields][next account data][...]
Tested in TEST_[FORBID|ALLOW]_WRITE_AFTER_OWNERSHIP_CHANGE*
Additionally when direct mapping is on, we must update all perms at once before
doing account data updates. Otherwise, updating an account might write into
another account whose perms we haven't updated yet. Tested in
TEST_FORBID_LEN_UPDATE_AFTER_OWNERSHIP_CHANGE.
* bpf_loader: serialization: address review comment don't return vm_addr from push_account_region
* bpf_loader: rename push_account_region to push_account_data_region
* cpi: fix slow edge case zeroing extra account capacity after shrinking an account
When returning from CPI we need to zero all the account memory up to the
original length only if we know we're potentially dealing with uninitialized
memory.
When we know that the spare capacity has deterministic content, we only need to
zero new_len..prev_len.
This fixes a slow edge case that was triggerable by the following scenario:
- load a large account (say 10MB) into the vm
- shrink to 10 bytes - would memset 10..10MB
- shrink to 9 bytes - would memset 9..10MB
- shrink to 8 bytes - would memset 8..10MB
- ...
Now instead in the scenario above the following will happen:
- load a large account (say 10MB) into the vm
- shrink to 10 bytes - memsets 10..10MB
- shrink to 9 bytes - memsets 9..10
- shrink to 8 bytes - memset 8..9
- ...
* bpf_loader: add account_data_region_memory_state helper
Shared between serialization and CPI to figure out the MemoryState of an
account.
* cpi: direct_mapping: error out if ref_to_len_in_vm points to account memory
If ref_to_len_in_vm is allowed to be in account memory, calles could mutate it,
essentially letting callees directly mutate callers memory.
* bpf_loader: direct_mapping: map AccessViolation -> InstructionError
Return the proper ReadonlyDataModified / ExecutableDataModified /
ExternalAccountDataModified depending on where the violation occurs
* bpf_loader: cpi: remove unnecessary infallible slice::get call
* remove unnecessary hashes around raw string literals
* remove unncessary literal `unwrap()`s
* remove panicking `unwrap()`
* remove unnecessary `unwrap()`
* use `[]` instead of `vec![]` where applicable
* remove (more) unnecessary explicit `into_iter()` calls
* remove redundant pattern matching
* don't cast to same type and constness
* do not `cfg(any(...` a single item
* remove needless pass by `&mut`
* prefer `or_default()` to `or_insert_with(T::default())`
* `filter_map()` better written as `filter()`
* incorrect `PartialOrd` impl on `Ord` type
* replace "slow zero-filled `Vec` initializations"
* remove redundant local bindings
* add required lifetime to associated constant
* dereplicode address alignment check
* Uses `checked_div` and `checked_rem` in built-in loaders.
* Uses `checked_div` and `checked_rem`.
* sdk: replace sub() with saturating_sub()
* eliminate `String` "arithmetic"
* allow arithmetic side-effects in tests and benches and on types we don't control
---------
Co-authored-by: Trent Nelson <trent@solana.com>
* Rename blockheight -> block height
* Clean up comments
* Simplify tpu-client implementation
* client: Destructure more in match arm
* client: Avoid passing Arc where it isn't needed
* Refactor sending into a new function
* Chain work after send to make last part cleaner
* Do all sending at once in the same future
* Sleep before sending to avoid overwhelming
Bank::get_leader_schedule_epoch returns: 1 + EpochSchedule::get_epoch.
As a result, at the epoch boundaries when the propagated shred is from
the next epoch, we are looking for epoch stakes for 2 epochs ahead of
the root or working bank's epoch.
However, the bank structure only contains epoch stakes for one epoch
ahead. This results in shreds propagated at epoch boundary having
unknown epoch stakes.
Created an overload of get_leader_tpus: get_leader_tpus_with_slots to get the tpu address along with the slot. And put that information into the debug log. This information can be compared with where the the transaction confirmed to measure leader accuracy. This help us better understand the txn latency.