path: root/common/recipes-kernel/linux/linux-yocto-4.19.8/2724-drm-amdgpu-Add-KFD-VRAM-limit-checking.patch

From 275b387f2746831946aec9217b59fae318f12104 Mon Sep 17 00:00:00 2001
From: Felix Kuehling <Felix.Kuehling@amd.com>
Date: Mon, 19 Nov 2018 20:05:54 -0500
Subject: [PATCH 2724/2940] drm/amdgpu: Add KFD VRAM limit checking

We don't want KFD processes evicting each other over VRAM usage.
Therefore prevent overcommitting VRAM among KFD applications with
a per-GPU limit. Also leave enough room for page tables on top
of the application memory usage.

Acked-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Eric Huang <JinHuiEric.Huang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu.h           |  7 ++-
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c    | 50 ++++++++++------
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h    |  8 ++-
 .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c  | 60 +++++++++----------
 drivers/gpu/drm/amd/amdgpu/amdgpu_object.c    |  2 +-
 5 files changed, 74 insertions(+), 53 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
index 14a8871b9917..11287a5ffc88 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
@@ -83,6 +83,7 @@
 #include "amdgpu_bo_list.h"
 #include "amdgpu_gem.h"
 #include "amdgpu_doorbell.h"
+#include "amdgpu_amdkfd.h"
 #include "amdgpu_smu.h"
 #include "amdgpu_discovery.h"
 #include "amdgpu_mes.h"
@@ -950,6 +951,9 @@ struct amdgpu_device {
 	/* GDS */
 	struct amdgpu_gds		gds;
 
+	/* KFD */
+	struct amdgpu_kfd_dev		kfd;
+
 	/* display related functionality */
 	struct amdgpu_display_manager dm;
 
@@ -970,9 +974,6 @@ struct amdgpu_device {
 	atomic64_t visible_pin_size;
 	atomic64_t gart_pin_size;
 
-	/* amdkfd interface */
-	struct kfd_dev          *kfd;
-
 	/* soc15 register offset based on ip, instance and  segment */
 	uint32_t 		*reg_offset[MAX_HWIP][HWIP_MAX_INSTANCE];
 
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
index 77bccdb27c7a..06bdbf06a0a5 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
@@ -32,10 +32,20 @@ bool (*kgd2kfd_init_p)(unsigned int, const struct kgd2kfd_calls**);
 
 static const unsigned int compute_vmid_bitmap = 0xFF00;
 
+/* Total memory size in system memory and all GPU VRAM. Used to
+ * estimate worst case amount of memory to reserve for page tables
+ */
+uint64_t amdgpu_amdkfd_total_mem_size;
+
 int amdgpu_amdkfd_init(void)
 {
+	struct sysinfo si;
 	int ret;
 
+	si_meminfo(&si);
+	amdgpu_amdkfd_total_mem_size = si.totalram - si.totalhigh;
+	amdgpu_amdkfd_total_mem_size *= si.mem_unit;
+
 #ifdef CONFIG_HSA_AMD
 	ret = kgd2kfd_init(KFD_INTERFACE_VERSION, &kgd2kfd);
 	if (ret)
@@ -90,8 +100,11 @@ void amdgpu_amdkfd_device_probe(struct amdgpu_device *adev)
 		return;
 	}
 
-	adev->kfd = kgd2kfd->probe((struct kgd_dev *)adev,
-				   adev->pdev, kfd2kgd);
+	adev->kfd.dev = kgd2kfd->probe((struct kgd_dev *)adev,
+				       adev->pdev, kfd2kgd);
+
+	if (adev->kfd.dev)
+		amdgpu_amdkfd_total_mem_size += adev->gmc.real_vram_size;
 }
 
 /**
@@ -131,7 +144,8 @@ void amdgpu_amdkfd_device_init(struct amdgpu_device *adev)
 {
 	int i;
 	int last_valid_bit;
-	if (adev->kfd) {
+
+	if (adev->kfd.dev) {
 		struct kgd2kfd_shared_resources gpu_resources = {
 			.compute_vmid_bitmap = compute_vmid_bitmap,
 			.num_pipe_per_mec = adev->gfx.mec.num_pipe_per_mec,
@@ -191,37 +205,37 @@ void amdgpu_amdkfd_device_init(struct amdgpu_device *adev)
 		gpu_resources.reserved_doorbell_mask = 0x1f0;
 		gpu_resources.reserved_doorbell_val  = 0x0f0;
 
-		kgd2kfd->device_init(adev->kfd, &gpu_resources);
+		kgd2kfd->device_init(adev->kfd.dev, &gpu_resources);
 	}
 }
 
 void amdgpu_amdkfd_device_fini(struct amdgpu_device *adev)
 {
-	if (adev->kfd) {
-		kgd2kfd->device_exit(adev->kfd);
-		adev->kfd = NULL;
+	if (adev->kfd.dev) {
+		kgd2kfd->device_exit(adev->kfd.dev);
+		adev->kfd.dev = NULL;
 	}
 }
 
 void amdgpu_amdkfd_interrupt(struct amdgpu_device *adev,
 		const void *ih_ring_entry)
 {
-	if (adev->kfd)
-		kgd2kfd->interrupt(adev->kfd, ih_ring_entry);
+	if (adev->kfd.dev)
+		kgd2kfd->interrupt(adev->kfd.dev, ih_ring_entry);
 }
 
 void amdgpu_amdkfd_suspend(struct amdgpu_device *adev)
 {
-	if (adev->kfd)
-		kgd2kfd->suspend(adev->kfd);
+	if (adev->kfd.dev)
+		kgd2kfd->suspend(adev->kfd.dev);
 }
 
 int amdgpu_amdkfd_resume(struct amdgpu_device *adev)
 {
 	int r = 0;
 
-	if (adev->kfd)
-		r = kgd2kfd->resume(adev->kfd);
+	if (adev->kfd.dev)
+		r = kgd2kfd->resume(adev->kfd.dev);
 
 	return r;
 }
@@ -230,8 +244,8 @@ int amdgpu_amdkfd_pre_reset(struct amdgpu_device *adev)
 {
 	int r = 0;
 
-	if (adev->kfd)
-		r = kgd2kfd->pre_reset(adev->kfd);
+	if (adev->kfd.dev)
+		r = kgd2kfd->pre_reset(adev->kfd.dev);
 
 	return r;
 }
@@ -240,8 +254,8 @@ int amdgpu_amdkfd_post_reset(struct amdgpu_device *adev)
 {
 	int r = 0;
 
-	if (adev->kfd)
-		r = kgd2kfd->post_reset(adev->kfd);
+	if (adev->kfd.dev)
+		r = kgd2kfd->post_reset(adev->kfd.dev);
 
 	return r;
 }
@@ -489,7 +503,7 @@ void amdgpu_amdkfd_set_compute_idle(struct kgd_dev *kgd, bool idle)
 
 bool amdgpu_amdkfd_is_kfd_vmid(struct amdgpu_device *adev, u32 vmid)
 {
-	if (adev->kfd) {
+	if (adev->kfd.dev) {
 		if ((1 << vmid) & compute_vmid_bitmap)
 			return true;
 	}
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
index 573df02e74b1..63895bcbc5c5 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
@@ -34,6 +34,7 @@
 #include "amdgpu_vm.h"
 
 extern const struct kgd2kfd_calls *kgd2kfd;
+extern uint64_t amdgpu_amdkfd_total_mem_size;
 
 struct amdgpu_device;
 
@@ -76,6 +77,11 @@ struct amdgpu_amdkfd_fence {
 	char timeline_name[TASK_COMM_LEN];
 };
 
+struct amdgpu_kfd_dev {
+	struct kfd_dev *dev;
+	uint64_t vram_used;
+};
+
 struct amdgpu_amdkfd_fence *amdgpu_amdkfd_fence_create(u64 context,
 						       struct mm_struct *mm);
 bool amdkfd_fence_check_mm(struct dma_fence *f, struct mm_struct *mm);
@@ -194,6 +200,6 @@ int amdgpu_amdkfd_gpuvm_get_vm_fault_info(struct kgd_dev *kgd,
 					      struct kfd_vm_fault_info *info);
 
 void amdgpu_amdkfd_gpuvm_init_mem_limits(void);
-void amdgpu_amdkfd_unreserve_system_memory_limit(struct amdgpu_bo *bo);
+void amdgpu_amdkfd_unreserve_memory_limit(struct amdgpu_bo *bo);
 
 #endif /* AMDGPU_AMDKFD_H_INCLUDED */
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
index 3162408e4e73..ed12c61e3737 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
@@ -110,17 +110,17 @@ void amdgpu_amdkfd_gpuvm_init_mem_limits(void)
 		(kfd_mem_limit.max_ttm_mem_limit >> 20));
 }
 
-static int amdgpu_amdkfd_reserve_system_mem_limit(struct amdgpu_device *adev,
+static int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev,
 		uint64_t size, u32 domain, bool sg)
 {
-	size_t acc_size, system_mem_needed, ttm_mem_needed;
+	size_t acc_size, system_mem_needed, ttm_mem_needed, vram_needed;
+	uint64_t reserved_for_pt = amdgpu_amdkfd_total_mem_size >> 9;
 	int ret = 0;
 
 	acc_size = ttm_bo_dma_acc_size(&adev->mman.bdev, size,
 				       sizeof(struct amdgpu_bo));
 
-	spin_lock(&kfd_mem_limit.mem_limit_lock);
-
+	vram_needed = 0;
 	if (domain == AMDGPU_GEM_DOMAIN_GTT) {
 		/* TTM GTT memory */
 		system_mem_needed = acc_size + size;
@@ -133,23 +133,30 @@ static int amdgpu_amdkfd_reserve_system_mem_limit(struct amdgpu_device *adev,
 		/* VRAM and SG */
 		system_mem_needed = acc_size;
 		ttm_mem_needed = acc_size;
+		if (domain == AMDGPU_GEM_DOMAIN_VRAM)
+			vram_needed = size;
 	}
 
+	spin_lock(&kfd_mem_limit.mem_limit_lock);
+
 	if ((kfd_mem_limit.system_mem_used + system_mem_needed >
-		kfd_mem_limit.max_system_mem_limit) ||
-		(kfd_mem_limit.ttm_mem_used + ttm_mem_needed >
-		kfd_mem_limit.max_ttm_mem_limit))
+	     kfd_mem_limit.max_system_mem_limit) ||
+	    (kfd_mem_limit.ttm_mem_used + ttm_mem_needed >
+	     kfd_mem_limit.max_ttm_mem_limit) ||
+	    (adev->kfd.vram_used + vram_needed >
+	     adev->gmc.real_vram_size - reserved_for_pt)) {
 		ret = -ENOMEM;
-	else {
+	} else {
 		kfd_mem_limit.system_mem_used += system_mem_needed;
 		kfd_mem_limit.ttm_mem_used += ttm_mem_needed;
+		adev->kfd.vram_used += vram_needed;
 	}
 
 	spin_unlock(&kfd_mem_limit.mem_limit_lock);
 	return ret;
 }
 
-static void unreserve_system_mem_limit(struct amdgpu_device *adev,
+static void unreserve_mem_limit(struct amdgpu_device *adev,
 		uint64_t size, u32 domain, bool sg)
 {
 	size_t acc_size;
@@ -167,6 +174,11 @@ static void unreserve_system_mem_limit(struct amdgpu_device *adev,
 	} else {
 		kfd_mem_limit.system_mem_used -= acc_size;
 		kfd_mem_limit.ttm_mem_used -= acc_size;
+		if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
+			adev->kfd.vram_used -= size;
+			WARN_ONCE(adev->kfd.vram_used < 0,
+				  "kfd VRAM memory accounting unbalanced");
+		}
 	}
 	WARN_ONCE(kfd_mem_limit.system_mem_used < 0,
 		  "kfd system memory accounting unbalanced");
@@ -176,29 +188,18 @@ static void unreserve_system_mem_limit(struct amdgpu_device *adev,
 	spin_unlock(&kfd_mem_limit.mem_limit_lock);
 }
 
-void amdgpu_amdkfd_unreserve_system_memory_limit(struct amdgpu_bo *bo)
+void amdgpu_amdkfd_unreserve_memory_limit(struct amdgpu_bo *bo)
 {
-	spin_lock(&kfd_mem_limit.mem_limit_lock);
+	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
+	u32 domain = bo->preferred_domains;
+	bool sg = (bo->preferred_domains == AMDGPU_GEM_DOMAIN_CPU);
 
 	if (bo->flags & AMDGPU_AMDKFD_USERPTR_BO) {
-		kfd_mem_limit.system_mem_used -=
-			(bo->tbo.acc_size + amdgpu_bo_size(bo));
-		kfd_mem_limit.ttm_mem_used -= bo->tbo.acc_size;
-	} else if (bo->preferred_domains == AMDGPU_GEM_DOMAIN_GTT) {
-		kfd_mem_limit.system_mem_used -=
-			(bo->tbo.acc_size + amdgpu_bo_size(bo));
-		kfd_mem_limit.ttm_mem_used -=
-			(bo->tbo.acc_size + amdgpu_bo_size(bo));
-	} else {
-		kfd_mem_limit.system_mem_used -= bo->tbo.acc_size;
-		kfd_mem_limit.ttm_mem_used -= bo->tbo.acc_size;
+		domain = AMDGPU_GEM_DOMAIN_CPU;
+		sg = false;
 	}
-	WARN_ONCE(kfd_mem_limit.system_mem_used < 0,
-		  "kfd system memory accounting unbalanced");
-	WARN_ONCE(kfd_mem_limit.ttm_mem_used < 0,
-		  "kfd TTM memory accounting unbalanced");
 
-	spin_unlock(&kfd_mem_limit.mem_limit_lock);
+	unreserve_mem_limit(adev, amdgpu_bo_size(bo), domain, sg);
 }
 
 
@@ -1226,8 +1227,7 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
 
 	amdgpu_sync_create(&(*mem)->sync);
 
-	ret = amdgpu_amdkfd_reserve_system_mem_limit(adev, size,
-						     alloc_domain, false);
+	ret = amdgpu_amdkfd_reserve_mem_limit(adev, size, alloc_domain, false);
 	if (ret) {
 		pr_debug("Insufficient system memory\n");
 		goto err_reserve_limit;
@@ -1280,7 +1280,7 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
 	/* Don't unreserve system mem limit twice */
 	goto err_reserve_limit;
 err_bo_create:
-	unreserve_system_mem_limit(adev, size, alloc_domain, false);
+	unreserve_mem_limit(adev, size, alloc_domain, false);
 err_reserve_limit:
 	mutex_destroy(&(*mem)->lock);
 	kfree(*mem);
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
index bc028b41b3e6..2401d61cb741 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
@@ -83,7 +83,7 @@ static void amdgpu_bo_destroy(struct ttm_buffer_object *tbo)
 		amdgpu_bo_subtract_pin_size(bo);
 
 	if (bo->kfd_bo)
-		amdgpu_amdkfd_unreserve_system_memory_limit(bo);
+		amdgpu_amdkfd_unreserve_memory_limit(bo);
 
 	amdgpu_bo_kunmap(bo);
 
-- 
2.17.1