Memory checks for direct mapping planned for epoch 1092
On 20 November 2026 Anza wrote that direct mapping is planned for mainnet this cycle. The upgrade page names epoch 1092 and also leaves the date open.
On 20 November 2026 Anza wrote on X that feature-gate activations are back after Breakpoint. In the same post, direct mapping is a breaking change planned for mainnet this cycle. The post links the Agave v4.5 release schedule.
The Solana Foundation page Account Data Direct Mapping is dated October 2026. Its mainnet column for virtual_address_space_adjustments reads “Live in Epoch 1092”. The expected mainnet date on the same page is “To be determined”.
The page still lists both gates as pending on Anza’s feature-gate tracker. Epoch 1092 is a plan here, not a completed activation. The date field on the page stays open.
Two gates
Two gates carry the change. virtual_address_space_adjustments adds the memory checks and can break programs. Its address is 7VgiehxNxu53KdxgLspGQY8myE6f7UokaWa4jsGcaSz.
account_data_direct_mapping is supposed to remove the copies of account data. The path of the data to the program changes, and the allowed accesses stay the same. On mainnet the gate is labeled “Not activated”.
The second address is CR3dVN2Yoo95Y96kLSTaziWDAQT2MNEpiWh5cqVq2pNE. Anza and Firedancer split SIMD-0219 so each part could be tested on its own. The checks came first.
Memory checks
Until now the validator copies account data into the virtual machine and back. A call into another program repeats that in both directions. With direct mapping the virtual machine points at the account data itself.
The runtime then checks every read and write as it happens. A read past the end of an account is supposed to fail with an instruction error. The same applies to a write on a read-only account, even when the program puts the old value back.
An access that crosses two memory regions in one step is supposed to fail as well. According to the page, most programs are not affected. A violation often hits only some instructions, such as a close, a resize, or one CPI path.
On devnet and testnet both gates have been running since May 2026, the page says. It names epochs 1073 and 1074 for devnet. It names epochs 957 and 958 for testnet.
The copies go away only once account_data_direct_mapping applies, according to the page. Instructions that use CPI or touch large accounts save the most. A test on devnet or testnet covers both gates, because both are running there.
What developers check
Users have nothing to do, the page says. Validator operators have nothing to do either. App and client developers have something to do only when a program they call breaks the rules.
Programs get tested on devnet. A program is at risk if it reads account bytes directly instead of through the SDK account types. Rust unsafe, raw pointers, and a resize of an account are on that list.
Anchor before version 1.2.2 adds one case. A typed account is reassigned there through a CPI.
A read past the end of the account data needs a data_len() check first. A write to a read-only account, or to an account owned by someone else, has to go. That still holds when the old value is back in the account at the end.
A resize past the limits, or one that uses realloc, moves to resize. The limits are 10 KiB per instruction and 10 MiB per account, plus the transaction’s growth cap. AccountInfo::realloc was deprecated in solana-account-info 2.3.0 and removed in 3.0.0.
With zero_init set to false the length grows without writing the new bytes. The first read of those bytes fails with InvalidRealloc. AccountInfo::resize fills the new bytes with zeros.
A write inside a writable account the program owns stays as it is. A write past the limits ends as an access violation. The log shows that as ProgramFailedToComplete.
The page maps the other errors like this. A read past the end of a read-only or foreign-owned account returns AccountDataTooSmall. A read past the end of a writable account the program owns returns InvalidRealloc.
Without write permission the error is ReadonlyDataModified. On an account the program does not own, it is ExternalAccountDataModified. Past the maximum size or the growth limits, it is InvalidRealloc again.
Logs print the sentence, for example “account data too small for instruction”. Simulation and getTransaction name the variant, such as AccountDataTooSmall. Catching the error in the calling program still aborts the transaction.
The runtime checks the write itself. Putting the old bytes back does not save it.
Stack and testing
The gaps between stack frames go away, including for sBPFv0 programs. On sBPFv0, call or callx moves the stack pointer by 4 KiB instead of 8 KiB. The total stack size stays the same.
An access outside the new range ends in an access violation. An access inside it can read or write the neighboring frame. The transaction then keeps running, with different data and no error.
A passing transaction does not show that a stack-dependent path is sound. Absolute stack addresses, and accesses across a frame boundary, belong in a devnet test. According to the page, Anza replayed mainnet traffic and contacted the teams it found.
That replay only covers traffic Anza has seen. solana feature status, with the address of virtual_address_space_adjustments, shows where a gate stands. solana-test-validator turns on every gate it knows about by default.
For a run without the gates, the page names a start with --reset and --deactivate-feature for both addresses. If the message says the argument was ignored because the ledger already exists, the gates are still on. The next run needs --reset.
For a real mainnet transaction the page points to SVM Scope, a community tool covering the last 30 days. Replay the transaction, switch on the gate labeled direct mapping, and replay it again. If it fails only with the gate on, the program needs a fix.
The gate syscall_parameter_address_restrictions has been running on mainnet since 30 June 2026, the page says. It is not part of this upgrade. Where CPI already runs on mainnet, the page says those pointer rules are met.
SIMD-0460 is still in review, the page says. The gate address comes from Agave’s feature set. The proposal does not list a feature key of its own there.
What to watch
- Whether virtual_address_space_adjustments takes effect on mainnet with epoch 1092
- Whether account_data_direct_mapping gets a mainnet date
- Whether CPI depth goes from 4 to 8 under SIMD-0268. The page places that after SIMD-0219
Sources
- Anza — Post on X, 20 November 2026, 12:38 a.m. Berlin. Feature gates after Breakpoint, direct mapping planned for mainnet activation this cycle, link to the Agave v4.5 schedule
- Solana Foundation — Account Data Direct Mapping, dated October 2026, read 20 November 2026. Mainnet column “Live in Epoch 1092” for virtual_address_space_adjustments (7VgiehxNxu53KdxgLspGQY8myE6f7UokaWa4jsGcaSz), “Not activated” for account_data_direct_mapping (CR3dVN2Yoo95Y96kLSTaziWDAQT2MNEpiWh5cqVq2pNE), expected mainnet date “To be determined”, both gates pending on Anza’s tracker. Devnet epochs 1073 and 1074, testnet epochs 957 and 958, May 2026