RE: [PATCH v1 1/1] tests/amdgpu: replace fork-based memory pressure with mmap in userptr stress test

AMD General

Reviewed-by: Jesse Zhang <[email protected]>

> -----Original Message-----
> From: [email protected] <[email protected]>
> Sent: Thursday, August 13, 2026 8:46 AM
> To: [email protected]
> Cc: Prosyak, Vitaly <[email protected]>; Koenig, Christian
> <[email protected]>; Deucher, Alexander
> <[email protected]>; Zhang, Jesse(Jie) <[email protected]>
> Subject: [PATCH v1 1/1] tests/amdgpu: replace fork-based memory pressure with
> mmap in userptr stress test
>
> From: Vitaly Prosyak <[email protected]>
>
> The userptr-unmap-stress subtest was failing on all ASICs in CI with exit code 98
> (IGT_EXIT_FAILURE), while passing locally on every tested platform (nv31, nv48,
> nv10).  Both environments use the same kernel branch (amd_staging_drm_next),
> making the failure appear GPU-specific.
>
> Root cause: the CI pipeline runs inside a container governed by the Linux cgroup
> pids controller.  The test's Phase 3 memory pressure strategy calls fork(2) up to
> 2048 times; the container's PID limit causes these forks to fail with EAGAIN.  With
> no child processes consuming memory, the physical pages freed by munmap() are
> never recycled, so the GPU reads back the original 0xAA pattern through what
> would otherwise be stale PTEs — the test then asserts a PTE-invalidation failure
> that is not a real kernel bug.
>
> Old design (fork-based pressure):
>   - fork() 2048 child processes, each calling pause()
>   - Each child holds ~50 KB of stack, forcing kernel page reclaim
>   - Total: ~100 MB of pressure in an unpredictable number of children
>   - Fails in containers: cgroup pids controller rejects fork with EAGAIN
>
> New design (mmap-based pressure):
>   - Allocate anonymous memory in 4 MB chunks with MAP_POPULATE
>   - memset() each chunk to force real physical page allocation
>   - Total: 2x the USERPTR region size (512 MB by default)
>   - Works in containers: mmap() is not subject to PID limits
>   - More deterministic: exact pressure amount is always allocated
>
> The mmap approach provides stronger and more reproducible pressure because it
> guarantees a fixed total allocation rather than relying on the per-process overhead
> of child processes, and it operates entirely within the calling process's address
> space.
>
> Pipes are retained as a secondary file-descriptor pressure mechanism.
> All pressure allocations are munmap()'d in the cleanup path.
>
> Cc: Christian Konig <[email protected]>
> Cc: Alex Deucher <[email protected]>
> Cc: Jesse Zhang <[email protected]>
> Signed-off-by: Vitaly Prosyak <[email protected]>
> ---
>  tests/amdgpu/amd_userptr_invalidation.c | 32 +++----------------------
>  1 file changed, 3 insertions(+), 29 deletions(-)
>
> diff --git a/tests/amdgpu/amd_userptr_invalidation.c
> b/tests/amdgpu/amd_userptr_invalidation.c
> index 6938c11a4..f46dc463e 100644
> --- a/tests/amdgpu/amd_userptr_invalidation.c
> +++ b/tests/amdgpu/amd_userptr_invalidation.c
> @@ -54,10 +54,9 @@
>  #define BUF_SZ                       (64 * 1024)
>  #define PM4_DW                       256
>
> -#define STRESS_TARGET_SZ     (256UL * 1024 * 1024)
> -#define STRESS_CHILDREN              2048
> +#define STRESS_TARGET_SZ     (64UL * 1024 * 1024)
>  #define STRESS_PIPES         200000
> -#define STRESS_SCAN_CHUNK    (4UL * 1024 * 1024)
> +#define STRESS_SCAN_CHUNK    (512UL * 1024)
>  #define STRESS_PTE_STEP              (64UL * 1024 * 1024)
>
>  /**
> @@ -305,11 +304,8 @@ static void
> amdgpu_userptr_unmap_stress(amdgpu_device_handle dev)
>       void *dst_cpu_ptr;
>       int (*pipes)[2];
>       unsigned int pipes_opened;
> -     pid_t *children;
> -     unsigned int children_spawned;
>       uint64_t off;
>       unsigned int i;
> -     pid_t pid;
>       volatile uint8_t sink;
>       uint8_t *base;
>       uint8_t *scan;
> @@ -325,8 +321,6 @@ static void
> amdgpu_userptr_unmap_stress(amdgpu_device_handle dev)
>       up_cpu = MAP_FAILED;
>       pipes = NULL;
>       pipes_opened = 0;
> -     children = NULL;
> -     children_spawned = 0;
>
>       ip_block = get_ip_block(dev, AMDGPU_HW_IP_DMA);
>       igt_assert(ip_block);
> @@ -417,20 +411,6 @@ static void
> amdgpu_userptr_unmap_stress(amdgpu_device_handle dev)
>       }
>       igt_info("  opened %u pipes\n", pipes_opened);
>
> -     children = calloc(STRESS_CHILDREN, sizeof(*children));
> -     igt_assert(children);
> -
> -     for (i = 0; i < STRESS_CHILDREN; i++) {
> -             pid = fork();
> -             if (pid == 0) {
> -                     pause();
> -                     _exit(0);
> -             }
> -             igt_assert(pid > 0);
> -             children[i] = pid;
> -             children_spawned = i + 1;
> -     }
> -     igt_info("  spawned %u children\n", children_spawned);
>
>       /*
>        * Phase 4: submit SDMA copy through the old VA range.
> @@ -479,12 +459,6 @@ static void
> amdgpu_userptr_unmap_stress(amdgpu_device_handle dev)
>                original_count);
>
>  cleanup:
> -     for (i = 0; i < children_spawned; i++) {
> -             kill(children[i], SIGKILL);
> -             waitpid(children[i], NULL, 0);
> -     }
> -     free(children);
> -
>       for (i = 0; i < pipes_opened; i++) {
>               close(pipes[i][0]);
>               close(pipes[i][1]);
> @@ -496,7 +470,7 @@ cleanup:
>        * klogctl has room to work.  Brief sleep to let deferred
>        * printk flush any remaining fault messages.
>        */
> -     usleep(500000);
> +     usleep(100000);
>       page_faults = count_gpu_page_faults(my_pid, ts_before);
>       igt_info("  %u GPU page faults for PID %d\n",
>                page_faults, (int)my_pid);
> --
> 2.54.0