Once the chip has booted, ROM accesses are requested over the system TL-UL bus. These come in through the TL-UL SRAM adapter (top-left of block diagram). In normal operation, the green multiplexer will give access to these TL reads. The address is scrambled at the first substitution-permutation network (marked S&P in the diagram).
In parallel with the ROM access, a reduced prim_prince
primitive (5 rounds with latency 1; equivalent to the cipher used for SRAM) computes a 39-bit truncated keystream for the block. On the following cycle, the scrambled data from ROM goes through a substitution-permutation network and is then XOR‘d with the keystream. This scheme is the same as that used by the SRAM controller, but is much simplified because the ROM doesn’t have to deal with writes, byte accesses or key changes.
The output from the XOR is the unscrambled 32-bit data, plus seven ECC bits. This data is passed straight through the TL-UL SRAM adapter; the ECC bits are used as a signal integrity check by the system bus.
The following diagram shows the timing of the different signals. The time from the req
output from the tlul_adapter_sram
to the response that appears on its rvalid
input is one cycle. The “scrambling scheme” for addresses in the diagram is to reverse their digits. The word stored at address 21 in the ROM is denoted w21
. The keystream value for address 12 is denoted k12
. The unscrambled ROM data for (logical) address 12 is denoted d12
.
{signal: [ {name: 'clk', wave: 'p....', period: 2}, {name: 'req', wave: '0.1...0...'}, {name: 'addr', wave: 'x.3.4.x...', data: ['12', '34']}, {name: 'scrambled addr', wave: 'x.3.4.x...', data: ['21', '43']}, {name: 'scrambled rdata + ecc', wave: 'x...3.4.x.', data: ['w21', 'w43']}, {name: 'keystream', wave: 'x...3.4.x.', data: ['k12', 'k34']}, {name: 'rdata + ecc', wave: 'x...3.4.x.', data: ['d12', 'd34']}, {name: 'rvalid', wave: '0...1...0.'}, ]}
The prim_prince
primitive and the two substitution-permutation networks are all parameterised by “keys”. For rom_ctrl
, these keys are global randomised netlist constants: they are assumed to be difficult to recover, but aren't considered secret data.
The ROM checker runs immediately after reset. Until it is done, it controls ROM address requests (through the green multiplexer). The select signal for this multiplexer has a redundant encoding to protect it against fault injection attacks. If the select signal has an invalid value, this will trigger a fatal alert. Before starting to read data, it starts a cSHAKE operation on the KMAC module using one of its application interfaces. We expect to use the cSHAKE256
algorithm, with prefix “ROM_CTRL”. The Application Interface section of the KMAC documentation details the parameters used.
The checker reads the ROM contents in address order, resulting in a scattered access pattern on the ROM itself because of the address scrambling. Each read produces 39 bits of data, which are padded with zeros to 64 bits to match the interface expected by the KMAC block. The checker FSM loops through almost all the words in ROM (from bottom to top), passing each to the KMAC block with the ready/valid interface and setting the kmac_data_o.last
bit for the last word that is sent. Once the last word has been sent, the FSM releases the multiplexer; this now switches over permanently to allow access through the TL-UL SRAM adapter.
The top eight words in ROM (by logical address) are interpreted as a 256-bit expected hash. Unlike the rest of ROM, their data is not stored scrambled, so the expected hash can be read directly. This is taken by the checker FSM (ignoring ECC bits) and will be compared with the digest that is read back from the KMAC block.
Once it comes back, the digest is forwarded directly to the Key Manager. It is also compared with the hash that was read from the top eight words of ROM. On a match, pwrmgr_data_o.good
is signalled as Mubi4True
. In either case, pwrmgr_data_o.done
goes high when the calculation is complete.
The diagram below shows the operation of the simple FSM.
One of the possible physical attacks on a system like OpenTitan is to subvert the ROM. The regular structure of a ROM is useful because it makes metal fixes easy, but (for the same reasons) it makes the ROM quite an easy target for an attacker. See [SKO-05][^SKO-05], section 2.1.1, for a description of ROMs and attacks on them.
[^SKO-05]: SKO-05: Skorobogatov, Semi-Invasive Attacks - A New Approach to Hardware Security Analysis, University of Cambridge Computer Laboratory Technical Report 630, 2005
Since the code in ROM is the first thing to execute, an attacker that modifies it undetected can completely subvert the chain of trust. As such, OpenTitan needs some form of ROM integrity checking and the ROM checker is the module in charge of providing it.
After bringing the ROM controller module out of reset, the power manager must wait until pwrgr_data_o.done
is asserted before starting the host processor. The ROM controller also passes the pwrmgr_data_o.good
signal. The power manager can use this to decide whether to boot (taking into account life cycle state). This provides an extra safety check, but the real security comes from key manager integration described below.
The simple KMAC interface assumes that KMAC is pre-configured to run the cSHAKE algorithm with a prefix specific to the ROM checker. The ROM checker will not assert kmac_data_o.valid
after finishing the one and only digest computation. The KMAC module may choose to add a check for this, to detect reset glitches affecting the rom_ctrl
block.
The integration with the key manager is based on forwarding the digest data in kmac_data_i
as keymgr_data_o.data
. This 256-bit digest will be incorporated into the CreatorRootKey
. The key manager should only allow one transaction (of 256 bits / 32 bits = 8 beats) after reset to pass this information across. On future messages, it should raise an alert, defeating an attacker that tries to trigger extra transactions before or after the real one.
CreatorRootKey
forms the first key in the chain described in Identities and Root Keys. An attacker who modifies the ROM will perturb CreatorRootKey
(to avoid doing so would require a preimage attack on the ROM checksum calculation or the KM_DERIVE
function). The result is that, while the chip will function, it will have the “wrong” root key and the chain of trust used for attestation will be broken.
The core integrity check, flowing from the ROM data to CreatorRootKey
, should be infeasible to subvert. However, rom_ctrl
also controls bus access to ROM data and interacts with other blocks. To avoid attacks propagating into the rest of the system, we take the following extra hardening steps:
mubi4_t
).mubi4_t
).Parameter | Default (Max) | Top Earlgrey | Description |
---|---|---|---|
RndCnstRomKey | (see RTL) | (see RTL) | Compile-time random default constant for scrambling key (used in prim_prince block). |
RndCnstRomNonce | (see RTL) | (see RTL) | Compile-time random default constant for scrambling nonce (used in prim_prince block and the two S&P blocks). |
The table below lists other ROM controller inter-module signals.
Software will mostly interact with the ROM controller by fetching code or loading data from ROM. For this, the block looks like a block of memory, accessible through a TL-UL window. However, there are a few registers that are accessible. Other than the standard ALERT_TEST
register, all are read-only.
The FATAL_ALERT_CAUSE
register might change value during operations (if an alert is signalled), but the other registers will all have fixed values by the time any software runs.
To get the computed ROM digest, software can read DIGEST_0
through DIGEST_7
. The ROM also contains an expected ROM digest. Unlike the rest of the contents of ROM, this isn‘t scrambled. As such, software can’t read it through the standard ROM interface (which would try to unscramble it again, resulting in rubbish data that would cause a failed ECC check). In case software needs access to this value, it can be read at EXP_DIGEST_0
through EXP_DIGEST_7
.