Unverified Commit 6713c5e4 authored by Luna Nova's avatar Luna Nova
Browse files

rocmPackages.rocm-runtime: apply fix for segfault when queue allocation fails 

Found by @06kellyjac, possibly triggered by bug in linux-6.18.4

https://lore.kernel.org/linux-iommu/870872aa-28e9-412a-bac6-8020bf560e4f@amd.com/t/

Resolving this *will not fix the underlying issue*, but does transmute a
segfault into an actual error result.

ROCm error: out of memory
  current device: 0, in function stream at /build/source/ggml/src/ggml-hip/../ggml-cuda/common.cuh:1345
  hipStreamCreateWithFlags(&streams[device][stream], 0x01)
parent b17c41cc

pkgs/development/rocm-modules/rocm-runtime/default.nix

+3 −0
Original line number Diff line number Diff line
@@ -53,6 +53,9 @@ stdenv.mkDerivation (finalAttrs: { 
  ];



  patches = [

    # Vendored upstream PR for fix for segfault when queue allocation fails

    # https://github.com/ROCm/rocm-systems/pull/2850

    ./queue-failure.patch

    (fetchpatch {

      # [PATCH] rocr: Extend HIP ISA compatibility check

      sha256 = "sha256-8r2Lb5lBfFaZC3knCxfXGcnkzNv6JxOKyJn2rD5gus4=";

pkgs/development/rocm-modules/rocm-runtime/queue-failure.patch

0 → 100644
+71 −0
Original line number Diff line number Diff line 
From b7717446c85d08b2d7c0c60ba3ac0eff11ee6120 Mon Sep 17 00:00:00 2001

From: Luna Nova <git@lunnova.dev>

Date: Tue, 20 Jan 2026 12:55:45 -0800

Subject: [PATCH 1/2] rocm-runtime: fix crash in QueueCreate due to trying to

 free non allocated scratch



if (scratch.main_queue_base != nullptr) before calling ReleaseQueueMainScratch

because ReleaseQueueMainScratch is only valid if main_queue_base is set

and the scope guard can fire for an error allocating the queue.

---

 .../runtime/hsa-runtime/core/runtime/amd_gpu_agent.cpp        | 4 +++-

 1 file changed, 3 insertions(+), 1 deletion(-)



diff --git a/runtime/hsa-runtime/core/runtime/amd_gpu_agent.cpp b/runtime/hsa-runtime/core/runtime/amd_gpu_agent.cpp

index 01b01fe869..83db40dacc 100644

--- a/runtime/hsa-runtime/core/runtime/amd_gpu_agent.cpp

+++ b/runtime/hsa-runtime/core/runtime/amd_gpu_agent.cpp

@@ -1792,7 +1792,9 @@ hsa_status_t GpuAgent::QueueCreate(size_t size, hsa_queue_type32_t queue_type, u

   scratch.main_queue_base = nullptr;

   scratch.main_queue_process_offset = 0;

 

-  MAKE_NAMED_SCOPE_GUARD(scratchGuard, [&]() { ReleaseQueueMainScratch(scratch); });

+  MAKE_NAMED_SCOPE_GUARD(scratchGuard, [&]() {

+    if (scratch.main_queue_base != nullptr) ReleaseQueueMainScratch(scratch);

+  });

 

   if (scratch.main_size != 0) {

     AcquireQueueMainScratch(scratch);

-- 

2.52.0





From 9c1746cd76a703e4d2321dc2ffe85fc61bfd2f21 Mon Sep 17 00:00:00 2001

From: Luna Nova <git@lunnova.dev>

Date: Tue, 20 Jan 2026 13:00:32 -0800

Subject: [PATCH 2/2] rocm-runtime: log for errors in QueueCreate



---

 .../runtime/hsa-runtime/core/runtime/amd_gpu_agent.cpp   | 9 ++++++++-

 1 file changed, 8 insertions(+), 1 deletion(-)



diff --git a/runtime/hsa-runtime/core/runtime/amd_gpu_agent.cpp b/runtime/hsa-runtime/core/runtime/amd_gpu_agent.cpp

index 83db40dacc..ae68732eb5 100644

--- a/runtime/hsa-runtime/core/runtime/amd_gpu_agent.cpp

+++ b/runtime/hsa-runtime/core/runtime/amd_gpu_agent.cpp

@@ -1799,6 +1799,9 @@ hsa_status_t GpuAgent::QueueCreate(size_t size, hsa_queue_type32_t queue_type, u

   if (scratch.main_size != 0) {

     AcquireQueueMainScratch(scratch);

     if (scratch.main_queue_base == nullptr) {

+      LogPrint(HSA_AMD_LOG_FLAG_INFO,

+               "Failed to allocate scratch memory for queue, size=%zu, node=%u",

+               scratch.main_size, node_id());

       return HSA_STATUS_ERROR_OUT_OF_RESOURCES;

     }

   }

@@ -1827,7 +1830,11 @@ hsa_status_t GpuAgent::QueueCreate(size_t size, hsa_queue_type32_t queue_type, u

             node_id()));

   }

 

-  if (!shared_queue) return HSA_STATUS_ERROR_OUT_OF_RESOURCES;

+  if (!shared_queue) {

+    LogPrint(HSA_AMD_LOG_FLAG_INFO,

+             "Failed to allocate shared queue descriptor memory, node=%u", node_id());

+    return HSA_STATUS_ERROR_OUT_OF_RESOURCES;

+  }

 

   auto aql_queue = new AqlQueue(shared_queue, this, size, node_id(), scratch, event_callback, data,

                                 flags);

-- 

2.52.0