PaperlayerView exact revision on arXiv

Source-linked AI summary

From Fleet to Lab: Revisiting the Security and Complexity of Industrial Rowhammer Mitigation

Hritvik Taneja, Moinuddin Qureshi

arXiv:2608.26072v1cs.CRcs.AR

TL;DR

Rowhammer mitigation must balance storage, mitigation overhead, and security across mode transitions. The paper analyzes Sigries and introduces FiRM, whose co-designed filters and fallbacks provide secure operation with lower complexity and storage. FiRM-P and FiRM-D retain zero slowdown for benign workloads while addressing probabilistic and deterministic mitigation needs.

  • Problem

    The paper examines how to provide secure memory-controller Rowhammer mitigation without the storage and CAM complexity of precise tracking or the frequent mitigations of sampling.

  • Method

    FiRM co-designs a filtering mechanism with its fallback so the activation budget remains bounded in both modes and during transitions, using probabilistic or deterministic variants.

  • Results

    FiRM-P and FiRM-D both incur zero slowdown for benign workloads; FiRM-P uses less than half Sigries’s storage, while FiRM-D uses less than two-thirds.

  • Takeaways & Limitations

    A principled filtered design can provide strong security while avoiding Sigries’s security tradeoff and CAM complexity.

Abstract

from arXiv · show

This paper studies efficient and secure Rowhammer mitigation at the Memory-Controller (MC). Rowhammer mitigation faces a fundamental tradeoff between tracking storage and mitigation rate: precise trackers (such as Misra-Gries) avoid unnecessary mitigations but require large CAM structures, whereas sampling-based schemes (such as PARA) require no storage but incur frequent mitigations even when not under attack. Microsoft recently deployed Sigries, an MC-side Rowhammer defense that combines an under-provisioned Misra-Gries tracker with a row-sampling fallback, in its Azure Cobalt 200 SoC. Sigries observed that the tracker-to-sampling transition can be insecure, and claimed the reverse transition is always safe. Our analysis shows that this transition is also vulnerable, and a Round-Robin Attack across sub-banks reduces the MTTF of Sigries to about 1 second, 8 orders of magnitude below the 13 years with PARA. Sigries also suffers from CAM complexity and high storage overheads. Our proposal, FiRM (Filtered Rowhammer Mitigation), is based on the insight that, for a secure design, the tracking-mode and sampling-mode should not be configured independently but co-designed to ensure the system remains secure not only in both modes but also during transitions. FiRM incurs zero slowdown for benign workloads, since they do not exceed the filtering threshold, and also replaces the complex CAM-based tracker with simple SRAM filters. To handle stressful patterns, we propose FiRM-P (probabilistic) and FiRM-D (deterministic). FiRM-P uses varying probabilities during transitions and steady state to ensure both security and low performance overhead. FiRM-D provides guaranteed deterministic security by modulating the rate of mitigation. Both FiRM-P and FiRM-D have less storage overhead than Sigries. Our paper shows that a principled approach can avoid both the insecurity and the complexity of Sigries.

I. INTRODUCTION

The paper identifies insecurity and complexity in Sigries and proposes FiRM, which co-designs filtering and fallback behavior to secure both modes and their transitions.

  • Sigries: Sigries combines an under-provisioned Misra-Gries CAM tracker with a PARA fallback, but the combination suffers from insecurity and complexity.The tracker handles benign workloads with minimal mitigation, while overflow triggers sampling-mode mitigation.
  • Sigries: The tracker-to-sampling and sampling-to-tracker transitions cannot be assessed independently because preceding modes can leave unmitigated activations.Sigries’s reverse transition is unsafe when rows carry hundreds of activations from prior sampling-mode.
  • Security analysis: 1 second versus 13 years: a Round-Robin Attack across multiple sub-banks reduces Sigries’s MTTF by eight orders of magnitude relative to PARA.The attack keeps vulnerability windows open at almost every instant.
  • FiRM: FiRM co-designs the filter and fallback so the attacker’s activation budget remains bounded across modes and transitions.The design replaces precise Misra-Gries tracking with simple SRAM filters that only need to detect whether benign traffic exceeds a threshold.
  • FiRM: FiRM-P uses elevated transition probabilities and lower steady-state probabilities, while FiRM-D modulates mitigation rates for deterministic security.Both variants avoid CAM complexity; FiRM-D also avoids reliance on an RNG.
  • FiRM: FiRM-P incurs zero slowdown for benign workloads and has similar performance as Sigries, while FiRM-P and FiRM-D use less storage than Sigries.FiRM-P requires less than half the storage, and FiRM-D requires less than two-thirds the storage of Sigries.

II. BACKGROUND AND MOTIVATION

This section establishes the Rowhammer threat and the memory-controller mitigation mechanisms, then contrasts probabilistic PARA with precise but complex Misra-Gries tracking.

  • DRAM architecture: 620K activations: an attacker can issue up to this many bank activations within the refresh window tREFW.DRAM retention requires every row to be refreshed within that window.
  • Mitigation mechanisms: Memory-controller mitigation tracks aggressor rows and refreshes their victim rows, using DRFM because the controller cannot directly refresh a victim row.DRFM samples an aggressor row and commands the DRAM to refresh associated victims.
  • Probabilistic mitigation: PARA mitigates neighboring rows with probability p on each activation, while DREAM delays DRFM to reduce the number of mitigations and associated stalls.At TRHD=3K, Naive-DRFM causes 28% loss whereas DREAM-DRFM causes 3.5% loss under continuous activations.
  • Precise tracking: Misra-Gries reduces benign-workload mitigation by tracking frequently activated rows, but requires a large associative CAM structure.The default configuration uses 416 entries per bank for TRHD=3K, creating area, complexity, and power costs.

E. Sigries: Rowhammer Mitigation in Production

Sigries is an MC-side defense that combines lightweight per-sub-bank tracking with PARA fallback to balance benign-workload performance and attack protection. Its design targets commercial-adoption requirements but retains substantial tracker complexity and transition assumptions.

  • Design goals: Sigries targets minimal overheads, no performance outliers, liveness, low hardware cost, flexibility, and configurable security guarantees for commercial adoption.
  • Design: Sigries divides each bank into sub-banks, each using an under-provisioned TinyMG tracker that switches to probabilistic PARA after spill-counter overflow.Heavy-mode lasts at least E windows and may be extended when shadow counters indicate an ongoing attack.
  • Security: Sigries’s modes are individually analyzed under a zero-unmitigated-activation assumption, with TMG set to TRHD/2 and p set to 20/TRHD.The design bounds the lite-to-heavy vulnerability window through a minimum heavy-mode epoch and attack-sensitive epoch extension.
  • Performance: Zero slowdown for benign applications is Sigries’s key performance advantage, while mitigation overhead remains small even under attack.
  • Storage and complexity: The estimated Sigries configuration uses 32-entry TinyMG trackers per sub-bank, while the design’s MG storage is summarized alongside Graphene in Table IV.The deployed parameters are not disclosed and are therefore estimated or inferred for TRHD of 3K.

F. Goal of This Paper

The paper examines whether production Sigries has additional vulnerabilities, how its exposure target affects system MTTF, and whether a lower-complexity dual-mode alternative can preserve strong security and performance. It focuses on the unsafe transition that Sigries claims is always safe and motivates a joint design of tracking and sampling.

  • Research questions: The paper asks whether Sigries has unknown vulnerabilities, because it is deployed in commercial systems rather than being only a theoretical design.
  • Research questions: It evaluates how Sigries’s one-hour-per-year vulnerability target affects the system’s Mean-Time-To-Failure.
  • Research questions: It seeks an MC-side mitigation with no vulnerability windows, lower CAM complexity, and performance matching Sigries.
  • Research questions: It asks whether a dual-mode mitigation can be fully deterministic and avoid probabilistic fallback.
  • Scope: The analysis targets Sigries’s heavy-mode-to-lite-mode transition and measures multi-sub-bank attacks against system MTTF.The lite-mode budget is TMG, whereas PARA heavy-mode is configured against TRHD; their composition can exceed the secure activation budget.

B. Understanding the Impact of the New Vulnerability

The second-transition vulnerability becomes systemically severe when attacks rotate across sub-banks, keeping vulnerability windows open for much of the time. The resulting MTTF analysis combines activation-budget accounting, reset-timing variation, and PARA escape probabilities.

  • Vulnerability analysis: e^20 = 4.9×10^8: the analyzed transition attack raises the victim failure probability relative to heavy-mode’s steady-state failure rate.Each aggressor’s escape probability is e−10, yielding victim failure probability e−20 under double-sided attack.
  • Vulnerability analysis: Two vulnerable transitions double Sigries’s exposure to two hours per year per sub-bank, while a per-sub-bank time target omits system-wide failure risk.
  • Round-Robin Attack: 70% of windows are vulnerable in the Round-Robin Attack, which targets 6144 sub-banks one at a time and cycles through them in 8760 windows.The attack waits for sub-banks to return to lite-mode before repeating.
  • Round-Robin Attack: Each sub-bank returns to lite-mode after E windows because the attack produces one anomalous window followed by E−1 quiet windows.The attack therefore bypasses the shadow-counter detector’s indication of a continuing attack.
  • Scope: The attack’s success is not dependent on the exact epoch size: extending each heavy-mode epoch by 10× would yield an MTTF of 10 seconds.
  • MTTF analysis: Reset timing makes A1 dynamic, while the activation budget divides into pre-reset activations, TinyMG-overflow activations, and PARA’s leftover budget.A1 varies because controller resets are periodic and unsynchronized with the attack.
  • MTTF analysis: A budget of B activations provides λ = p · B expected PARA mitigation chances, so reducing B causes exponential decay in sampling effectiveness.At p = 1/150, 1500 activations provide 10 chances with miss likelihood e−10, whereas 150 provide one chance with miss likelihood e−1.
  • MTTF: The MTTF is about 1 second under the Round-Robin Attack, versus 13 years for PARA under continuous attack, an eight-order-of-magnitude reduction.The calculation uses a 4.2% victim-failure probability per vulnerability window, 70% vulnerable windows, and 32-millisecond windows.

E. Takeaway: Duty Cycle Bounds Exposure, Not Failure

Sigries bounds how often vulnerability windows open but not the probability of failure within them. FiRM instead co-designs and partitions one activation budget across modes, avoiding this exposure-management weakness.

  • E. Takeaway: Duty Cycle Bounds Exposure, Not Failure: 4.2% rather than e−40 is the demonstrated in-window failure probability for Sigries, despite its E ≥8760 security condition.E bounds vulnerability-window frequency, not failure probability within an open window.
  • E. Takeaway: Duty Cycle Bounds Exposure, Not Failure: 70% vulnerable-window exposure from Round-Robin Attack makes Sigries’s in-window failure probability an unacceptably low practical guarantee.The attack raises the fraction of vulnerable windows across sub-banks.
  • E. Takeaway: Duty Cycle Bounds Exposure, Not Failure: FiRM partitions TRHD across the filter and fallback so every mode sequence remains within the activation threshold.This shared-budget design avoids allocating the full budget independently to each mode.
  • E. Takeaway: Duty Cycle Bounds Exposure, Not Failure: FiRM replaces Sigries’s CAM-based tracker with an untagged direct-mapped SRAM counter table because the filter only needs threshold information.Aliasing can over-count activations, which may cost performance but does not cost security.
  • E. Takeaway: Duty Cycle Bounds Exposure, Not Failure: FiRM develops probabilistic and deterministic variants, FiRM-P and FiRM-D, based on the type of fallback algorithm.FiRM-P uses a probabilistic fallback, while FiRM-D uses a deterministic fallback.

A. FiRM-P: Design and Operation

FiRM-P uses an SRAM filter to remain in lite-mode below a threshold and switches to probabilistic heavy-mode when that threshold is exceeded. Its transition probabilities are elevated for security, while steady-state sampling preserves comparable mitigation behavior to Sigries.

  • A. FiRM-P: Design and Operation: FiRM-P increments an SRAM counter and enters heavy-mode when the accessed count exceeds TF; otherwise it remains in lite-mode without mitigation.Heavy-mode lasts through the remainder of the current window and E + 2 subsequent full windows.
  • A. FiRM-P: Design and Operation: TF is constrained above by lite-mode security and below by benign behavior, so threshold selection must satisfy both requirements.Filter resets can let an attacker accrue TF activations on either side of a reset.
  • A. FiRM-P: Design and Operation: FiRM-P splits the 3K activation budget between TF and PARA, using p1=1/25, p2=1/150, and p3=1/60 across transition and steady-state windows.The sequence includes entry, bridge, steady-state, and exit behavior before the filter is reset and reused.
  • A. FiRM-P: Design and Operation: TF=1250 for TRHD=3K satisfies the security bound of 1500 with 500 activations left for heavy-mode.The selected threshold leaves a margin within the tREFW window.
  • A. FiRM-P: Design and Operation: At E=100 FiRM-P issues 11% more mitigations than Sigries, but by E=1000 the gap is within 1%, and at E ≥8760 rates are indistinguishable.E controls the amortization of elevated transition probabilities and is not a security parameter for FiRM-P.

C. Security Analysis

FiRM’s security proof partitions the activation budget across filter and fallback cases, covering lite-mode, heavy-mode, and both transitions. The design also reduces storage and lookup complexity relative to Sigries.

  • C. Security Analysis: FiRM’s five-case analysis shows that no row receives more than TRHD activations between victim refreshes.A refresh interval spans at most one filter reset and one mode boundary, enabling the budget accounting.
  • C. Security Analysis: 2·TF=2500 remains within TRHD in lite-mode because filter-counter aliasing only over-counts activations.Lite-mode security is deterministic and does not rely on fallback sampling.
  • C. Security Analysis: In heavy-mode, A=0 leaves the full budget to sampling, where p2=20/TRHD=1/150 is sufficient for security.Bridge and exit windows sample faster, covering boundaries between heavy-mode windows.
  • C. Security Analysis: Entry charges 2·TF=2500 activations to filtering and 500 to sampling, with p1=1/25 making the combined allowance exactly TRHD.The elevated entry probability accounts for activations accumulated before the fallback begins.
  • C. Security Analysis: FiRM-P has no vulnerability windows because every mode and transition stays within TRHD and the failure probability never exceeds e−20.Consequently, the Sigries multi-sub-bank attack has no schedulable vulnerability window against FiRM-P.
  • C. Security Analysis: FiRM requires 352 bytes per bank versus 864 for Sigries at equal entry count, a 2.5× storage reduction.The reduction comes from discarding tags and metadata rather than tracking fewer regions.
  • C. Security Analysis: FiRM replaces Sigries’s 32-way associative matching and eviction search with one SRAM read-modify-write and a threshold comparison.The simpler structure eliminates tags, associative lookup, and eviction policy.

E. Performance Impact for Benign Workloads

On the evaluated benign workloads, FiRM-P remains in lite-mode and matches Sigries’s zero slowdown, while PARA slows execution. Under attack, FiRM-P and Sigries have similar slowdown; FiRM-P can also be configured for different thresholds.

  • E. Performance Impact for Benign Workloads: The evaluation uses DRAMSim3 with twelve SPEC2017, six GAP, and four STREAM benchmarks in 8-core rate mode.Applications run for 250 million instructions each, using weighted speedup as the metric.
  • E. Performance Impact for Benign Workloads: PARA incurs an average slowdown of 4.4%, while Sigries and FiRM-P incur 0% slowdown for all benign workloads.Both Sigries and FiRM-P always remain in lite-mode in this evaluation.
  • E. Performance Impact for Benign Workloads: Under attacks, Sigries and FiRM-P both average 4.3% slowdown and reach 7.7% worst-case slowdown.Both schemes degenerate into PARA with p=1/150 when filtering is bypassed.
  • E. Performance Impact for Benign Workloads: FiRM-P is configurable across TRHD values, with 0.5% slowdown at TRHD=2K when operated below its designed TRHD.The same hardware can target TRHD=2K, 3K, or 4K with reconfigured parameters.

V. FIRM-D: DETERMINISTIC FILTERED MITIGATION

FiRM-D provides deterministic Rowhammer protection by sharing activation counters across gangs and modulating mitigation after a filtering threshold. Its paced design avoids benign-workload mitigations while limiting attack-time stalls.

  • FiRM-D design: FiRM-D maintains a counter over many rows and modulates mitigation so every row is refreshed within TRHD activations.Mitigation is zero below TF and higher thereafter, providing deterministic security without uniform mitigation rates.
  • FiRM-D substrate: DRFMab refreshes one sampled row in every bank, allowing one command to mitigate 32 gang members across banks.V DRFMab commands cover V rows per bank, while a rotation pointer selects the next row to refresh.
  • FiRM-D substrate: FiRM-D shares activation counters across gangs and banks, using a rotation pointer to identify the next row for mitigation.The substrate reduces storage by grouping multiple rows under each counter and preserves rotation order during gradual service.
  • Mitigation strategies: DREAM-C triggers a burst of V mitigations at TRHD/2, but its 6.6 µs stall violates the liveness requirement R3.Its advantage is zero mitigation while the activation counter remains below TRHD/2.
  • Mitigation strategies: Gradual pacing sends one mitigation every X = TRHD/(V + 1) activations, amortizing bursts but applying overhead even to benign workloads.The paced rate is constant rather than threshold-filtered.
  • FiRM-D design: FiRM-D avoids mitigation until TF and then paces refreshes faster than Gradual, balancing benign-workload efficiency with attack protection.For V=4, Figure 10 illustrates filtering followed by paced mitigation rather than bursty or always-on service.
  • Performance: FiRM-D and Sigries have zero slowdown in the reported workload comparison, while Gradual incurs an average slowdown of 38%.The comparison is normalized to a baseline without mitigation.

C. Parameters

This section sets FiRM-D’s deterministic mitigation parameters and explains how threshold filtering changes mitigation frequency. It also reports zero slowdown for the evaluated benign workloads.

  • Parameters: The three deterministic schemes are parameterized by filtering threshold TF and ACTs per DRFM, APD, for TRHD = 3K and V = 16.Table XI compares DREAM-C, Gradual, and FiRM-D under these assumptions.
  • Parameters: FiRM-D mitigates at 1-per-29 activations after filtering, versus Gradual’s 1-per-176 rate.FiRM-D compresses its V mitigations into the remaining budget TRHD−2·TF, making its attack-time rate 6× higher.
  • Parameters: Epoch-Based FiRM-D is presented as enabling better performance under attacks.The cited figure identifies the epoch-based variant but does not quantify the improvement.
  • Security: Security analysis preserves the rotation pointer across resets and transitions so each row is serviced within the bounded mitigation schedule.The argument accounts for at most one reset and one mode transition per refresh interval.
  • Performance: FiRM-D issues no mitigations in the benign workload suite because average ACTR reaches 474 while TF = 1250.Gradual slows down by 38% on average and up to 60%, whereas FiRM-D and Sigries have zero slowdown.
  • Performance: PRAC incurs 8.4% average slowdown, while Sigries, FiRM-P, and FiRM-D are free on benign workloads.Figure 13 compares these designs at TRHD = 3K.

G. Storage Analysis

FiRM reduces Sigries’ storage and avoids its CAM complexity while preserving zero benign-workload slowdown. FiRM-D trades higher attack-time slowdown for deterministic, bounded stalls and security.

  • Storage: FiRM-P uses less than half Sigries’ storage, while FiRM-D uses less than two-thirds, and both avoid CAM complexity.FiRM-D entries use an activation counter, rotation pointer, and mode bit, with a single global epoch counter.
  • Performance under attack: Sigries and FiRM-P incur 4.3% average slowdown under maximum-rate attack mitigations, whereas FiRM-D incurs 2×.FiRM-D refreshes every row in a gang and each DRFMab stalls all 32 banks.
  • Performance under attack: FiRM-D’s maximum rate is one DRFMab per 29 activations, so its attack-time degradation has no performance tail.The rate is bounded and known at design time, and degradation ceases with the attack.
  • Performance under attack: FiRM-D’s longest stall is 411 ns, 16× below DREAM-C and below the microsecond scale prohibited by R3.The design trades attack-time throughput for deterministic security.
  • Comparison with PRAC: PRAC slows benign workloads by 8.4%, while the three MC-side designs are free on benign workloads.The comparison is reported at TRHD = 3K.
  • Overall comparison: FiRM-P and FiRM-D replace Sigries’ complex precision tracking without requiring its security tradeoff or high storage overhead.The paper’s conclusion presents both designs as practical filtered mitigations with zero benign-workload slowdown.

APPENDIX A ON OPTIMIZING FOR THE BLIND-STATIC ATTACKER

The appendix argues that Sigries’ precision advantage matters only for a narrow Blind-Static regime, while informed or adaptive attackers receive no corresponding benefit. It also documents conservative PARA-rate estimation and analysis scope.

  • Attacker model: The evaluation assumes attackers know the mitigation algorithm and parameters, with RNG output as the only secret.Security requires no more than TRHD activations between consecutive refreshes; RowPress is excluded because page-closure policy handles it.
  • Tracker regime: Misra-Gries tracks patterns with at most C distinct rows exactly, whereas FiRM trips fallback whenever a region exceeds TF.With C = 32, Sigries remains in lite-mode and mitigates tracked rows once per TMG = 1500 activations.
  • Benign workloads: Across 22 benign workloads, neither Sigries nor FiRM issues a mitigation because the busiest activity remains below their thresholds.The hottest row receives 312 activations per tREFW, while FiRM’s filter threshold is TF = 1250.
  • Tracker regime: Decahammer against Sigries provides neither bit flips nor meaningful slowdown, while using C+1 rows yields 4.3% slowdown instead of 0.4%.The appendix therefore characterizes the precise-tracker regime as unattractive to persistent attackers seeking failure or slowdown.
  • Attacker taxonomy: Sigries’ tracker outperforms FiRM only against Blind-Static attackers, who have no incentive to run for meaningful time.The reported Rowhammer attack classes are Blind-Adaptive, Informed-Static, and Informed-Adaptive, and Sigries and FiRM behave identically in all three.
  • Residual case: For the residual Blind-Static case, FiRM-P remains within a small slowdown envelope: 4.3% on average and 7.7% at worst.The pattern creates no new worst case, and R1 and R2 remain unaffected.
  • PARA sampling rate: The PARA approximation p = 20/TRHD is conservative, sampling 2–5% more often than the precise recurrence-based method.Its coefficient is 20 versus 19.02–19.60, with negligible performance difference.
  • Scope: The Sigries analysis uses estimated parameters from the published design and does not attempt the Round-Robin Attack on a deployed system.The authors disclosed the transition concern to the Sigries authors and revised the reported MTTF after feedback.
Loading 2608.26072v1…