ENTROPY_SRC HWIP Technical Specification

Overview

This document specifies ENTROPY_SRC hardware IP functionality. This module conforms to the Comportable guideline for peripheral functionality.

Features

• This revision provides an interface to an external physical random noise generator (also referred to as a physical true random number generator. The PTRNG external source is a physical true random noise source. A noise source and its relation to an entropy source are defined by SP 800-90B.
• A set of registers is provided for firmware to obtain entropy bits.
• Interrupts are supported:
  • Entropy bits are available for firmware consumption.
  • The internal health tests have detected a test failure.
  • An internal FIFO error has occurred.
• Two health checks that are defined by SP 800-90B are performed by this revision: Repetition Count and Adaptive Proportion tests.
• Two additional hardware health checks are performed by this revision: Bucket and Markov tests.
• Firmware-defined (mailbox-based) and vendor-defined health checks are also supported.
• A health check failure alert is supported by this revision.

Description

This IP block provides an entropy source that is capable of using a PTRNG noise source to generate random values in a manner that is compliant both with FIPS (though NIST SP 800-90B) and CC (AIS31) recommendations.

The random values generated by this block serve as non-deterministic seeds for the CSRNG block. The outputs of the CSRNG are then used either directly by firmware or are distributed to other hardware blocks through the Entropy Distribution Network.

In the terms of AIS31 classes, this block is meant to satisfy the requirements for a PTG.2 class physical entropy source, with “internal” entropy (an AIS31 term, meaning the min-entropy as measured just before the output pins) exceeding 0.997 entropy-bits/output-bit. In NIST terms, this block satisfies the requirements for a “full-entropy” source, which requires even smaller deviations from ideal entropy, at the level of less than one part in 2⁶⁴. When operating in full-compliance mode, raw outputs from a PTRNG noise source (external to this block) are passed through a conditioning function based on a suitable secure hash function (SHA-2 or SHA-3) which has been vetted by NIST to meet these stringent requirements.

These raw noise bits are subjected to a sequence of health-checks to screen the raw signals for statistical defects which would cause any significant deviations from ideal entropy at the output of the conditioning block. These tests include:

• The Repetition Count test, which screens for stuck-bits, or a complete failure of the PTRNG noise source,
• The Adaptive Proportion test, which screens for statistical bias in the number of 1’s or 0’s output by the noise source,
• The “Bucket Test”, which looks for correlations between the individual noise channels that the external noise source concatenates together to produce the raw noise sequence,
• The “Markov Test”, which looks for unexpected first-order temporal correlations between bits output by the individual noise channels,
• The “Mailbox Test”, in which raw-noise data is transferred to firmware in contiguous 2kbit-blocks, so that firmware can perform custom tests, and signal a failure through the same path as the other tests, and
• Optional Vendor Specific tests, which allow silicon creators to extend the health checks by adding a top-level block external to this IP.

The Repetition Count and Adaptive Proportion test are specifically recommended by SP 800-90B, and are implemented in accordance with those recommendations. In FIPS/CC-compliance mode, all checks except the Repetition Count test are performed on fixed window of data, typically consisting of 2048 bits each. Per the definition in SP 800-90B, the Repetition Count test does not operate on a fixed window. The repetition count test fails if any sequence of bits continuously asserts the same value for too many samples, as determined by the programmable threshold, regardless of whether that sequence crosses any window boundaries. The thresholds for these tests should be chosen to achieve a low false-positive rate (α) given a conservative estimate of the manufacturing tolerances of the PTRNG noise source. The combined choice of threshold and window size then determine the false-negative rate (β), or the probability of missing statistical defects at any particular magnitude.

When the IP is disabled by clearing the ENABLE bit in CONF, all heath checks are disabled and all counters internal to the health checks are reset.

In order to compensate for the fact our tests (like all realistic statistical tests) have finite resolution for detecting defects, we conservatively use 2048 bits of PTRNG noise source to construct each 384 bit conditioned entropy sample. When passed through the conditioning block, the resultant entropy stream will be full entropy unless the PTRNG noise source has encountered some statistical defect serious enough to reduce the raw min-entropy to a level below 0.375 bits of entropy per output bit. We choose this level as our definition of “non-tolerable statistical defects” for the purposes of evaluating this system under AIS31. Given this definition of “non-tolerable defects”, the health-checks as implemented for this block will almost certainly detect any of the previously mentioned defects in a single iteration of the health checks (i.e. such serious defects will be detected with very low β).

In addition to the brief, low-latency health checks, various long-term statistics are accumulated in registers for additional diagnostic purposes or for in-depth analysis. For the hardware interface, entropy bits will be transferred whenever the downstream consumer is ready to receive them.

At a high-level, this ENTROPY_SRC block, when enabled, will continuously collect entropy bits from the entropy source into a FIFO that can be read from the TL-UL bus, or sent out through a hardware interface. The PTRNG noise source will generate a relatively slow set of bit streams, with a total bit rate on the order of 50 kbps. The digital source can be programmed to be faster, or much slower if desired. A digital source rate of flexible speed might be useful in demonstrating how the downstream applications use the entropy.

Once a noise source has been selected through configuration, the raw noise data processing begins. This ENTROPY_SRC block will do health checks, collect and pack bits, and finally send them through a conditioning unit. There are two modes in which entropy bits are delivered, boot-time and FIPS/CC mode. Boot-time mode will deliver bits sooner for specific on-boot obfuscation applications, though the bits may not yet have been subjected to the same level of startup health checks required for FIPS or CC compliance.

In boot-time mode health checks only operate on a window of 384 bits. The boot-time health checks are the same as the FIPS/CC health-checks, though with different thresholds. They are sensitive to the same types of statistical defects, though at reduced statistical resolution. With suitable thresholds, the boot-time health checks can be operate both with low false-alarm rates (low α), while still confirming with low β that the total entropy of the first seed contains at least 80 bits of total entropy. During start up the initial 384 bits are held in a buffer until the boot-time start-up health checks are performed. Storing the seed in this buffer, allows this seed to released to the CSRNG immediately after the entropy has been confirmed.

Boot-time mode also has the feature that it bypasses the SHA conditioning function, as only 384 bits are used in the initial boot-time seed.

For maximal flexibility in normal operation, the conditioning function can also be implemented by firmware. When this firmware conditioning feature is activated, data read directly out of the noise source can be reinjected into the entropy pipeline via a TL-UL register after it has been processed by firmware. It should be noted that this firmware algorithm must be vetted by NIST to satisfy the requirements for a full-entropy source. This feature can also be disabled for security purposes, either by locking the feature via the REGEN register at boot, or by a write to one-time programmable (OTP) memory.

Compatibility

This IP block does not have any direct hardware compatibility requirements. However, the general design of this block follows the overall NIST recommendations, as described by SP 800-90B.

Theory of Operations

As already described, this IP block will collect bits of entropy for firmware or hardware consumption. This revision supports only an external interface for a PTRNG noise source implementation.

The first step is initialization and enabling. The PTRNG noise source mode is selected when the ENABLE field will be set. After the block is enabled and initialized, entropy bits will be collected up indefinitely until disabled.

After a reset, the ENTROPY_SRC block will start up in boot-time mode by default. This feature is designed to provide an initial seed’s worth of entropy with lower latency than the normal FIPS/CC compliant health check process. Health testing will still be performed on boot-time mode entropy, but the window of checking is, by default, 384 bits instead of 2048 bits. When entropy is delivered to the downstream hardware block, a signal will indicate what type of entropy it is - FIPS compliant or not. Boot-time mode can be completely disabled in the CONF register.

Once the initial boot-time mode phase has completed, the ENTROPY_SRC block will switch to FIPS compliant mode. In this mode, once the raw entropy has been health checked, it will be passed into a conditioner block. This block will compress the bits such that the entropy bits/physical bits, or min-entropy value, should be improved over the raw data source min-entropy value. The compression operation, by default, will compress every 2048 tested bits into 384 full-entropy bits.

The hardware conditioning can also be bypassed and replaced in normal operation with a firmware-defined conditioning algorithm. This firmware conditioning algorithm can be disabled on boot for security purposes.

The firmware override function has the capability to completely override the hardware health tests and the conditioner paths. In the case of health tests, firmware can turn off one or all of the health tests and perform the tests in firmware. A data path is provided in the hardware such that the inbound entropy can be trapped in the pre-conditioner FIFO. Once a pre-determined threshold of entropy has been reached in this FIFO, the firmware can then read the entropy bits out of the FIFO. The exact mechanism for this functionality starts with setting the FW_OV_MODE field in the FW_OV_CONTROL register. This will enable firmware to monitor post-health test entropy bits by reading from the FW_OV_RD_DATA register. Firmware can use the OBSERVE_FIFO_THRESH and OBSERVE_FIFO_DEPTH to determine the state of the OBSERVE FIFO. At this point, firmware can do additional health checks on the entropy. Optionally, firmware can do the conditioning function, assuming the hardware is configured to bypass the conditioner block. Once firmware has processed the entropy, it can then write the results back into the FW_OV_WR_DATA register (pre-conditioner FIFO). The FW_OV_ENTROPY_INSERT in the FW_OV_CONTROL register will enable inserting entropy bits back into the entropy flow. The firmware override control fields will be set such that the new entropy will resume normal flow operation.

An additional feature of the firmware override function is to insert entropy bits into the flow and still use the condtioning function in the hardware. Setting the FW_OV_INSERT_START field in the FW_OV_SHA3_START register will prepare the hardware for this flow. Once this field is set true, the FW_OV_WR_DATA register can be written with entropy bits. The FW_OV_WR_FIFO_FULL register should be monitored after each write to ensure data is not dropped. Once all of the data has been written, the FW_OV_INSERT_START field should be set to false. The normal SHA3 processing will continue and finally push the conditioned entropy through the module.

Health checks are performed on the input raw data from the PTRNG noise source when in that mode. There are four health tests that will be performed: repetitive count, adaptive proportion, bucket, and Markov tests. Each test has a pair of threshold values that determine that pass/fail of the test, one threshold for boot-time mode, and one for normal mode. By default, all tests are enabled, but can be turn off in the CONF register. Because of the variability of the PTRNG noise source, there are several registers that log statistics associated with the health tests. For example, the adaptive proportion test has a high watermark register that logs the highest measured number of ones. The ADAPTP_HI_WATERMARKS register has an entry for both normal and boot-time modes. This register allows for determining how close the threshold value should be set to the fail over value. Specific to the adaptive proportion test, there is also the ADAPTP_LO_WATERMARKS register, which will hold the lowest number of ones measured. To help understand how well the thresholds work through time, a running count of test fails is kept in the ADAPTP_HI_TOTAL_FAILS register. The above example for the adaptive proportion test also applies to the other health tests, with the exception of the low watermark registers. See the timing diagrams below for more details on how the health tests work. It should be noted that for all error counter registers, they are sized for 16 bits, which prevents any case where counters might wrap.

Vendor-specific tests are supported through an external health test interface (xht). This is the same interface that is used for the internal health tests. Below is a description of this interface:

• entropy_bit: 4-bit wide bus of entropy to be tested.
• entropy_bit_valid: indication of when the entropy is valid.
• clear: signal to clear counters, and is register driven.
• active: signal to indicate when the test should run, and is register driven.
• thresh_hi: field to indicate what high threshold the test should use, and is register driven.
• thresh_lo: field to indicate what low threshold the test should use, and is register driven.
• window: field to indicate what the size of the test window is, and is register driven.
• test_cnt: generic test count result, to be read from a register.
• test_fail_hi_pulse: indication that a high threshold comparison failed, to be read from a register.
• test_fail_lo_pulse: indication that a low threshold comparison failed, to be read from a register.

The ALERT_THRESHOLD register determines how many fails can occur before an alert is issued. By default, the current threshold is set to two, such that the occurrence of two failing test cycles back-to-back would provide a very low α value. The ALERT_FAIL_COUNTS register holds the total number of fails, plus all of the individual contributing failing tests. Setting the ALERT_THRESHOLD register to zero will disable alert generation.

Firmware has a path to read entropy from the ENTROPY_SRC block. The ENTROPY_CONTROL register allows firmware to set the internal multiplexers to steer entropy data to the ENTROPY_DATA register. The control bit ES_TYPE sets whether the entropy will come from the conditioning block or be sourced through the bypass path. A status bit will be set that can either be polled or generate an interrupt when the entropy bits are available to be read from the ENTROPY_DATA register. The firmware needs to read the ENTROPY_DATA register twelve times in order to cleanly evacuate the 384-bit seed from the hardware path (12*32bits=384bits total). The firmware will directly read out of the main entropy FIFO, and when the control bit ES_ROUTE is set, no entropy is being passed to the block hardware interface.

If the esfinal FIFO fills up, additional entropy that has been health checked will be dropped before entering the conditioner. This drop point will save on conditioner power, and still preserve esfinal FIFO entropy that has already been collected.

The above process will be repeated for as long as entropy bits are to be collected and processed.

At any time, the ENABLE field can be cleared to halt the entropy generation (and health check testing). See the Programmers Guide section for more details on the ENTROPY_SRC block disable sequence.

Block Diagram

Hardware Interfaces

Referring to the Comportable guideline for peripheral device functionality, the module entropy_src has the following hardware interfaces defined.

Primary Clock: clk_i

Other Clocks: none

Bus Device Interfaces (TL-UL): tl

Bus Host Interfaces (TL-UL): none

Peripheral Pins for Chip IO: none

Interrupts:

Security Alerts:

Security Countermeasures:

Design Details

Initialization

After power-up, the ENTROPY_SRC block is disabled.

For simplicity of initialization, only a single register write is needed to start functional operation of the ENTROPY_SRC block. This assumes that proper defaults are chosen for thresholds, sampling rate, and other registers.

For security reasons, a configuration and control register locking function is performed by the REGEN register. Clearing the bit in this register will prevent future modification of the CONF register or other writeable registers by firmware.

Entropy Processing

When enabled, the ENTROPY_SRC block will generate entropy bits continuously. The es_entropy_valid bit in the ENTROPY_SRC_INTR_STATE register will indicate to the firmware when entropy bits can read from the ENTROPY_DATA register. The firmware will do 32-bit register reads of the ENTROPY_DATA register to retrieve the entropy bits. Each read will automatically pop an entry from the entropy unpacker block. A full twelve 32-bit words need to be read at a time.

The hardware entropy interface will move entropy bits out of the ENTROPY FIFO when it is not empty, and the downstream hardware is ready. If firmware is not currently reading entropy bits, all processed entropy bits will flow to the hardware entropy interface.

Security

All module assets and countermeasures performed by hardware are listed in the hjson countermeasures section. Labels for each instance of asset and coutermeasure are located throughout the RTL source code.

For all of the health test threshold registers, these registers could be protected with shadow registers. A design choice was made here to not use shadow registers and save on silicon cost. The threshold registers are protected by software. It is expected that software will read the threshold registers on a periodic basis, and compare these values to what was originally programmed into the threshold registers.

Bus integrity checking is performed for the final seed delivery to CSRNG. This is done to make sure repeated values are not occurring. Only 64 bits (out of 384 bits) are checked, since this is statistically significant, and more checking would cost more silicon.

Interrupts

The ENTROPY_SRC module has several interrupts: es_entropy_valid, es_health_test_failed, and es_fifo_err.

The es_entropy_valid interrupt should be asserted when an entropy source has been implemented that is relatively slow.

The es_health_test_failed interrupt will trigger when the internal health test fails and exceeds the alert threshold.

The es_fifo_err interrupt will fire when an internal FIFO has a malfunction. The conditions that cause this to happen are either when there is a push to a full FIFO or a pull from an empty FIFO.

Main State Machine Diagram

The following diagram shows how the main state machine state is constructed. The larger circles show the how the overall state machine transitions. The sub-state machines with smaller circles show more detail about how the large circles operate.

Entropy Source Hardware Interface

The following waveform shows an example of how the entropy source hardware interface works, which is much like a FIFO.

PTRNG Hardware Interface

The following waveform shows an example of what the PTRNG timing looks like.

Repetition Count Test

The following waveform shows how a sampling of a data pattern will be tested by the Repetition Count test. Operating on each bit stream, this test will count when a signal is at a stuck level. This NIST test is intended to signal a catastrophic failure with the PTRNG noise source.

Adaptive Proportion Test

This NIST-defined test is intended to detect statistical bias in the raw entropy data. The test counts the number of 1’s in a given sample, and applies thresholds to reject samples which deviate too far from the ideal mean of 50%.

Depending on the value of the CONF.THRESHOLD_SCOPE field, the thresholds can either be applied collectively to the all RNG inputs, or the thresholds can be applied on a line-by-line basis. Setting CONF.THRESHOLD_SCOPE to kMuBi4True will apply the thresholds to the aggregated RNG stream. This can be useful for lowering the likelihood of coincidental test failures (higher α). Meanwhile, setting CONF.THRESHOLD_SCOPE to kMuBi4False will apply thresholds on a line-by-line basis which allows the ENTROPY_SRC to detect single line failures.

The following waveform shows how a sampling of a data pattern will be tested by the Adaptive Proportion test. In this example, the sum is taken over all RNG lines (i.e., CONF.THRESHOLD_SCOPE is True).

Bucket Test

The following waveform shows how a sampling of a data pattern will be tested by the Bucket test. Operating on all four bit streams, this test will identify the symbol and sort it into bin counters, or “buckets”. This test is intended to find bias with a symbol or symbols.

Markov Test

The following waveform shows how a sampling of a data pattern will be tested by the Markov test.

The test aims to detect either:

1. Oversampling of AST/RNG outputs leading to “clustered” input values that eventually change, but often are just repeats of the previous sample. For example the string: “00111111000011000111000111000001111” has roughly equal numbers of 1’s and 0’s, but no good entropy source should generate such strings, because each bit is likely just a repeat of the previous one.

2. Wild oscillations of the RNG, in a distinctly non-random way. For instance the string: “010101010101010101” has almost zero entropy, even though the number of 1’s and 0’s appears unbiased.

The test counts the number of changes in the a fixed number of RNG samples, and comparing the number of “01”/“10” pairs to the number of “00”/“11” pairs. On average, the number of switching (e.g., “01”) vs. non-switching (e.g., “00”) pairs should be 50% of the total, with a variance proportional to the sample size.

Like the Adaptive Proportion test, the Markov Test can be computed either cumulatively (summing the results over all RNG lines) or on a per-line basis. In this example, the RNG lines are scored individually (i.e., CONF.THRESHOLD_SCOPE is False).

Programmers Guide

Initialization

The following code snippet demonstrates initializing the ENTROPY_SRC block for entropy bit generation.

void entropy_src_init(unsigned int rate) {

  // set the configuration enable bits (rng mode, all health tests enabled)
  *CONF_REG = 0x1;

  // the ENTROPY_SRC interrupts can optionally be enabled
}

Entropy Processing

The following code snippet demonstrates reading entropy bits from the ENTROPY_SRC block.

int entropy_src_entropy(unsigned int numEntropyBits) {

  // read numEntropyBits, check for ENTROPY_SRC_INTR_STATE bit 0
  return *ENTROPY_DATA_REG;

}

Note that when software makes frequent re-seed requests to CSRNG, any stored up entropy seeds in the final entropy FIFO will quickly consumed. Once the FIFO is empty, subsequent entropy seed requests will have to wait the worst case latency time while new entropy is being created.

Entropy Source Module Disable

A useful feature for the ENTROPY_SRC block is the ability to disable it in a graceful matter. Since there exists another feature to avoid power spikes between ENTROPY_SRC and CSRNG, software needs to monitor the disabling process. Bit 16 in the DEBUG_STATUS should be polled after the ENTROPY_SRC enable bits are cleared in the CONF register. After the handshakes with CSRNG are finished, the above bit should be set and the ENTROPY_SRC block can be safely enabled again.

Error conditions

Need to alert the system of a FIFO overflow condition.

Register Table