Skip to content

Fixed-priority MMU miss admission can starve higher-numbered ports #483

Closed

Description

@JiangboWANGfr

A SimX test launched 33 threads in one block with virtual memory enabled and 32 threads per warp. A load from the second warp waited for a TLB miss slot while lower-numbered ports kept accepting requests. Page walks continued to complete, but the warp's issue watchdog fired after 200,000 cycles.

Minimal reproduction

The four-port RTL test in #482 isolates the MMU. Each port continuously requests a different stream of cold pages. The MMU has two miss slots. After an initial stall, the test lets one page walk proceed every eight cycles. With walks completing and capacity becoming available, each requesting port should make progress.

From a clean checkout of upstream commit 7300bc849 with the Vortex toolchain installed, copy only the test harness from the PR and run it against the original RTL.

git fetch origin refs/pull/482/head
git restore --source=f434cfc7b --worktree -- hw/unittest/mmu_park
mkdir -p build/rv32
cd build/rv32
../../configure --xlen=32
make -C hw/unittest/mmu_park run
Accepted misses after 256 cycles Port 0 Port 1 Port 2 Port 3
Upstream 7300bc849 32 0 0 0
With the fix in #482 8 8 8 8

RV32 and RV64 produced the same counts. The RTL test ran on Ubuntu 24.04 using the prebuilt Verilator from the Vortex v3.0.2 toolchain.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions