From 7f7eebab6de93b7a359723346b34793a0f18325e Mon Sep 17 00:00:00 2001
From: Harish Kasiviswanathan <Harish.Kasiviswanathan@amd.com>
Date: Fri, 18 Aug 2017 16:50:54 -0400
Subject: [PATCH] Fix hsaKmtQueryPointerInfo for scratch memory

Change-Id: I35a0f1a81f8b0ac6e99c2e1572829eb32d3bc95b
Signed-off-by: Harish Kasiviswanathan <Harish.Kasiviswanathan@amd.com>


[ROCm/ROCR-Runtime commit: 00250f7686373b12e21acc77ee452a57e4178e76]
---
 projects/rocr-runtime/src/fmm.c | 46 ++++++++++++++++-----------------
 projects/rocr-runtime/src/fmm.h |  1 -
 2 files changed, 23 insertions(+), 24 deletions(-)

diff --git a/projects/rocr-runtime/src/fmm.c b/projects/rocr-runtime/src/fmm.c
index 381898d35f..11963b66ec 100644
--- a/projects/rocr-runtime/src/fmm.c
+++ b/projects/rocr-runtime/src/fmm.c
@@ -656,6 +656,23 @@ static manageable_aperture_t *fmm_get_aperture(HsaApertureInfo info)
 		return NULL;
 	}
 }
+
+static manageable_aperture_t *fmm_is_scratch_aperture(const void *address)
+{
+	uint32_t i;
+
+	for (i = 0; i < gpu_mem_count; i++) {
+		if (gpu_mem[i].gpu_id == NON_VALID_GPU_ID)
+			continue;
+
+		if ((address >= gpu_mem[i].scratch_physical.base) &&
+			(address <= gpu_mem[i].scratch_physical.limit))
+			return &gpu_mem[i].scratch_physical;
+
+	}
+	return NULL;
+}
+
 static manageable_aperture_t *fmm_find_aperture(const void *address,
 						HsaApertureInfo *info)
 {
@@ -666,8 +683,12 @@ static manageable_aperture_t *fmm_find_aperture(const void *address,
 	if (is_dgpu) {
 		if (address >= svm.dgpu_aperture.base &&
 			address <= svm.dgpu_aperture.limit) {
-			aperture = &svm.dgpu_aperture;
-			_info.type = HSA_APERTURE_DGPU;
+
+			aperture = fmm_is_scratch_aperture(address);
+			if (!aperture) {
+				aperture = &svm.dgpu_aperture;
+				_info.type = HSA_APERTURE_DGPU;
+			}
 		} else if (address >= svm.dgpu_alt_aperture.base &&
 			address <= svm.dgpu_alt_aperture.limit) {
 			aperture = &svm.dgpu_alt_aperture;
@@ -753,27 +774,6 @@ err_object_allocation_failed:
 	return NULL;
 }
 
-bool fmm_is_inside_some_aperture(void *address)
-{
-	uint32_t i;
-
-	for (i = 0; i < gpu_mem_count; i++) {
-		if (gpu_mem[i].gpu_id == NON_VALID_GPU_ID)
-			continue;
-		if ((address >= gpu_mem[i].lds_aperture.base) &&
-				(address <= gpu_mem[i].lds_aperture.limit))
-			return true;
-		if ((address >= gpu_mem[i].gpuvm_aperture.base) &&
-				(address <= gpu_mem[i].gpuvm_aperture.limit))
-			return true;
-		if ((address >= gpu_mem[i].scratch_aperture.base) &&
-				(address <= gpu_mem[i].scratch_aperture.limit))
-			return true;
-	}
-
-	return false;
-}
-
 #ifdef DEBUG_PRINT_APERTURE
 static void aperture_print(aperture_t *app)
 {
diff --git a/projects/rocr-runtime/src/fmm.h b/projects/rocr-runtime/src/fmm.h
index 3bc0fb2040..c874f6fe01 100644
--- a/projects/rocr-runtime/src/fmm.h
+++ b/projects/rocr-runtime/src/fmm.h
@@ -54,7 +54,6 @@ void *fmm_allocate_doorbell(uint32_t gpu_id, uint64_t MemorySizeInBytes, uint64_
 void *fmm_allocate_host(uint32_t node_id, uint64_t MemorySizeInBytes,
 			HsaMemFlags flags);
 void fmm_print(uint32_t node);
-bool fmm_is_inside_some_aperture(void *address);
 void fmm_release(void *address);
 int fmm_map_to_gpu(void *address, uint64_t size, uint64_t *gpuvm_address);
 int fmm_unmap_from_gpu(void *address);