tools/cuda-ab: memset_probe + wedge_recover forensics clients (session 10)

memset_probe: linear 8 MB cudaMemset + D2H with first-bad offset and a
4 KB page map (correct / other / malloc-prefill) - separates 'fill
under-delivered' from 'readback delivered wrong pages'; standalone green,
bisects the memcpy_roundtrip memset phase.

wedge_recover: launch+verify loop that on failure sleeps N seconds and
retries the same iteration (same context), reporting per-iteration
attempts and launch times - tests whether an idle gap heals the per-boot
no-execute wedge (r409 log shows a ~37 s client-side gap healed a wedged
run).

Also ships the missing inplace_scale run binary (its source landed in
7685d1b3).

1 files changed, 87 insertions, 0 deletions

diff --git a/tools/cuda-ab/memset_probe.cu b/tools/cuda-ab/memset_probe.cu
new file mode 100644
index 0000000..a6a12e0
--- /dev/null
+++ b/tools/cuda-ab/memset_probe.cu
@@ -0,0 +1,87 @@
+// SPDX-License-Identifier: BSD-3-Clause
+/*
+ * memset_probe.cu - isolate the memcpy_roundtrip memset-phase failure.
+ *
+ * cudaMemset(8 MB, 0x5A) -> D2H -> byte verify, with a run-length report:
+ * the first bad offset, and a coverage map at 4 KB granularity (correct /
+ * wrong / untouched-malloc-prefill) so "fill under-delivered" is
+ * distinguishable from "readback delivered wrong pages".
+ *
+ * Build: nvcc -O2 --cudart=static -o memset_probe memset_probe.cu
+ */
+
+#include <cstdio>
+#include <cstdlib>
+#include <cuda_runtime.h>
+
+#define CHECK(x)                                                         \
+	do {                                                             \
+		cudaError_t e_ = (x);                                    \
+		if (e_ != cudaSuccess) {                                 \
+			printf("CUDA_ERROR %s at line %d\n",             \
+			       cudaGetErrorString(e_), __LINE__);        \
+			printf("STATUS FAIL\n");                         \
+			return 1;                                        \
+		}                                                        \
+	} while (0)
+
+#define PAGE 4096
+#define MemsetByte 0x5A
+
+int main(void)
+{
+	const size_t bytes = 8u << 20;
+	unsigned char *h;
+	unsigned char *d;
+	size_t i;
+	int bad = 0;
+	size_t first_bad = (size_t)-1;
+	size_t n_pages, p;
+
+	setvbuf(stdout, NULL, _IONBF, 0);
+	h = (unsigned char *)malloc(bytes);
+	if (!h)
+		return 1;
+	memset(h, 0xEE, bytes);
+
+	CHECK(cudaMalloc(&d, bytes));
+	printf("DPTR d=%p bytes=%zu\n", (void *)d, bytes);
+
+	CHECK(cudaMemset(d, MemsetByte, bytes));
+	CHECK(cudaDeviceSynchronize());
+	CHECK(cudaMemcpy(h, d, bytes, cudaMemcpyDeviceToHost));
+
+	for (i = 0; i < bytes; i++) {
+		if (h[i] != (unsigned char)MemsetByte) {
+			if (first_bad == (size_t)-1)
+				first_bad = i;
+			bad++;
+		}
+	}
+	printf("FIRST_BAD off=%zu (0x%zx) bad=%d total=%zu\n",
+	       first_bad, first_bad, bad, bytes);
+	n_pages = bytes / PAGE;
+	for (p = 0; p < n_pages; p++) {
+		int good = 1, prefill = 1;
+
+		for (i = 0; i < PAGE; i++) {
+			if (h[p * PAGE + i] != (unsigned char)MemsetByte)
+				good = 0;
+			if (h[p * PAGE + i] != 0xEE)
+				prefill = 0;
+			if (!good && !prefill)
+				break;
+		}
+		if (!good) {
+			printf("PAGE %zu %s b0=%02x b1=%02x b2=%02x b3=%02x b4095=%02x\n",
+			       p, prefill ? "PREFILL" : "OTHER",
+			       h[p * PAGE], h[p * PAGE + 1],
+			       h[p * PAGE + 2], h[p * PAGE + 3],
+			       h[p * PAGE + 4095]);
+			if (p > 24)
+				break;
+		}
+	}
+	printf("STATUS %s\n", bad == 0 ? "OK" : "FAIL");
+	return 0;
+}