1 files changed, 102 insertions, 69 deletions

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
index b64938ed8cb6..55e0284b2bdd 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
@@ -58,6 +58,7 @@
 #include "amdgpu_amdkfd.h"
 #include "amdgpu_sdma.h"
 #include "amdgpu_ras.h"
+#include "amdgpu_hmm.h"
 #include "amdgpu_atomfirmware.h"
 #include "amdgpu_res_cursor.h"
 #include "bif/bif_4_1_d.h"
@@ -189,7 +190,6 @@ static int amdgpu_ttm_map_buffer(struct ttm_buffer_object *bo,
 	struct amdgpu_device *adev = ring->adev;
 	unsigned offset, num_pages, num_dw, num_bytes;
 	uint64_t src_addr, dst_addr;
-	struct dma_fence *fence;
 	struct amdgpu_job *job;
 	void *cpu_addr;
 	uint64_t flags;
@@ -229,7 +229,9 @@ static int amdgpu_ttm_map_buffer(struct ttm_buffer_object *bo,
 	num_dw = ALIGN(adev->mman.buffer_funcs->copy_num_dw, 8);
 	num_bytes = num_pages * 8 * AMDGPU_GPU_PAGES_IN_CPU_PAGE;
 
-	r = amdgpu_job_alloc_with_ib(adev, num_dw * 4 + num_bytes,
+	r = amdgpu_job_alloc_with_ib(adev, &adev->mman.entity,
+				     AMDGPU_FENCE_OWNER_UNDEFINED,
+				     num_dw * 4 + num_bytes,
 				     AMDGPU_IB_POOL_DELAYED, &job);
 	if (r)
 		return r;
@@ -269,18 +271,8 @@ static int amdgpu_ttm_map_buffer(struct ttm_buffer_object *bo,
 		}
 	}
 
-	r = amdgpu_job_submit(job, &adev->mman.entity,
-			      AMDGPU_FENCE_OWNER_UNDEFINED, &fence);
-	if (r)
-		goto error_free;
-
-	dma_fence_put(fence);
-
-	return r;
-
-error_free:
-	amdgpu_job_free(job);
-	return r;
+	dma_fence_put(amdgpu_job_submit(job));
+	return 0;
 }
 
 /**
@@ -381,7 +373,7 @@ static int amdgpu_move_blit(struct ttm_buffer_object *bo,
 	dst.offset = 0;
 
 	r = amdgpu_ttm_copy_mem_to_mem(adev, &src, &dst,
-				       new_mem->num_pages << PAGE_SHIFT,
+				       new_mem->size,
 				       amdgpu_bo_encrypted(abo),
 				       bo->base.resv, &fence);
 	if (r)
@@ -424,7 +416,7 @@ error:
 static bool amdgpu_mem_visible(struct amdgpu_device *adev,
 			       struct ttm_resource *mem)
 {
-	u64 mem_size = (u64)mem->num_pages << PAGE_SHIFT;
+	u64 mem_size = (u64)mem->size;
 	struct amdgpu_res_cursor cursor;
 	u64 end;
 
@@ -571,7 +563,7 @@ static int amdgpu_ttm_io_mem_reserve(struct ttm_device *bdev,
 				     struct ttm_resource *mem)
 {
 	struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
-	size_t bus_size = (size_t)mem->num_pages << PAGE_SHIFT;
+	size_t bus_size = (size_t)mem->size;
 
 	switch (mem->mem_type) {
 	case TTM_PL_SYSTEM:
@@ -691,9 +683,8 @@ int amdgpu_ttm_tt_get_user_pages(struct amdgpu_bo *bo, struct page **pages,
 	}
 
 	readonly = amdgpu_ttm_tt_is_readonly(ttm);
-	r = amdgpu_hmm_range_get_pages(&bo->notifier, mm, pages, start,
-				       ttm->num_pages, range, readonly,
-				       true, NULL);
+	r = amdgpu_hmm_range_get_pages(&bo->notifier, start, ttm->num_pages,
+				       readonly, NULL, pages, range);
 out_unlock:
 	mmap_read_unlock(mm);
 	if (r)
@@ -704,8 +695,19 @@ out_unlock:
 	return r;
 }
 
+/* amdgpu_ttm_tt_discard_user_pages - Discard range and pfn array allocations
+ */
+void amdgpu_ttm_tt_discard_user_pages(struct ttm_tt *ttm,
+				      struct hmm_range *range)
+{
+	struct amdgpu_ttm_tt *gtt = (void *)ttm;
+
+	if (gtt && gtt->userptr && range)
+		amdgpu_hmm_range_get_pages_done(range);
+}
+
 /*
- * amdgpu_ttm_tt_userptr_range_done - stop HMM track the CPU page table change
+ * amdgpu_ttm_tt_get_user_pages_done - stop HMM track the CPU page table change
  * Check if the pages backing this ttm range have been invalidated
  *
  * Returns: true if pages are still valid
@@ -723,10 +725,6 @@ bool amdgpu_ttm_tt_get_user_pages_done(struct ttm_tt *ttm,
 
 	WARN_ONCE(!range->hmm_pfns, "No user pages to check\n");
 
-	/*
-	 * FIXME: Must always hold notifier_lock for this, and must
-	 * not ignore the return code.
-	 */
 	return !amdgpu_hmm_range_get_pages_done(range);
 }
 #endif
@@ -1154,8 +1152,9 @@ int amdgpu_ttm_tt_get_userptr(const struct ttm_buffer_object *tbo,
  * @addr:  The address in the current tasks VM space to use
  * @flags: Requirements of userptr object.
  *
- * Called by amdgpu_gem_userptr_ioctl() to bind userptr pages
- * to current task
+ * Called by amdgpu_gem_userptr_ioctl() and kfd_ioctl_alloc_memory_of_gpu() to
+ * bind userptr pages to current task and by kfd_ioctl_acquire_vm() to
+ * initialize GPU VM for a KFD process.
  */
 int amdgpu_ttm_tt_set_userptr(struct ttm_buffer_object *bo,
 			      uint64_t addr, uint32_t flags)
@@ -1394,7 +1393,8 @@ static void amdgpu_ttm_vram_mm_access(struct amdgpu_device *adev, loff_t pos,
 }
 
 static int amdgpu_ttm_access_memory_sdma(struct ttm_buffer_object *bo,
-					unsigned long offset, void *buf, int len, int write)
+					unsigned long offset, void *buf,
+					int len, int write)
 {
 	struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
 	struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
@@ -1418,26 +1418,27 @@ static int amdgpu_ttm_access_memory_sdma(struct ttm_buffer_object *bo,
 		memcpy(adev->mman.sdma_access_ptr, buf, len);
 
 	num_dw = ALIGN(adev->mman.buffer_funcs->copy_num_dw, 8);
-	r = amdgpu_job_alloc_with_ib(adev, num_dw * 4, AMDGPU_IB_POOL_DELAYED, &job);
+	r = amdgpu_job_alloc_with_ib(adev, &adev->mman.entity,
+				     AMDGPU_FENCE_OWNER_UNDEFINED,
+				     num_dw * 4, AMDGPU_IB_POOL_DELAYED,
+				     &job);
 	if (r)
 		goto out;
 
 	amdgpu_res_first(abo->tbo.resource, offset, len, &src_mm);
-	src_addr = amdgpu_ttm_domain_start(adev, bo->resource->mem_type) + src_mm.start;
+	src_addr = amdgpu_ttm_domain_start(adev, bo->resource->mem_type) +
+		src_mm.start;
 	dst_addr = amdgpu_bo_gpu_offset(adev->mman.sdma_access_bo);
 	if (write)
 		swap(src_addr, dst_addr);
 
-	amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_addr, dst_addr, PAGE_SIZE, false);
+	amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_addr, dst_addr,
+				PAGE_SIZE, false);
 
 	amdgpu_ring_pad_ib(adev->mman.buffer_funcs_ring, &job->ibs[0]);
 	WARN_ON(job->ibs[0].length_dw > num_dw);
 
-	r = amdgpu_job_submit(job, &adev->mman.entity, AMDGPU_FENCE_OWNER_UNDEFINED, &fence);
-	if (r) {
-		amdgpu_job_free(job);
-		goto out;
-	}
+	fence = amdgpu_job_submit(job);
 
 	if (!dma_fence_wait_timeout(fence, false, adev->sdma_timeout))
 		r = -ETIMEDOUT;
@@ -1537,6 +1538,23 @@ static void amdgpu_ttm_fw_reserve_vram_fini(struct amdgpu_device *adev)
 		NULL, &adev->mman.fw_vram_usage_va);
 }
 
+/*
+ * Driver Reservation functions
+ */
+/**
+ * amdgpu_ttm_drv_reserve_vram_fini - free drv reserved vram
+ *
+ * @adev: amdgpu_device pointer
+ *
+ * free drv reserved vram if it has been reserved.
+ */
+static void amdgpu_ttm_drv_reserve_vram_fini(struct amdgpu_device *adev)
+{
+	amdgpu_bo_free_kernel(&adev->mman.drv_vram_usage_reserved_bo,
+						  NULL,
+						  &adev->mman.drv_vram_usage_va);
+}
+
 /**
  * amdgpu_ttm_fw_reserve_vram_init - create bo vram reservation from fw
  *
@@ -1558,11 +1576,35 @@ static int amdgpu_ttm_fw_reserve_vram_init(struct amdgpu_device *adev)
 	return amdgpu_bo_create_kernel_at(adev,
 					  adev->mman.fw_vram_usage_start_offset,
 					  adev->mman.fw_vram_usage_size,
-					  AMDGPU_GEM_DOMAIN_VRAM,
 					  &adev->mman.fw_vram_usage_reserved_bo,
 					  &adev->mman.fw_vram_usage_va);
 }
 
+/**
+ * amdgpu_ttm_drv_reserve_vram_init - create bo vram reservation from driver
+ *
+ * @adev: amdgpu_device pointer
+ *
+ * create bo vram reservation from drv.
+ */
+static int amdgpu_ttm_drv_reserve_vram_init(struct amdgpu_device *adev)
+{
+	u64 vram_size = adev->gmc.visible_vram_size;
+
+	adev->mman.drv_vram_usage_va = NULL;
+	adev->mman.drv_vram_usage_reserved_bo = NULL;
+
+	if (adev->mman.drv_vram_usage_size == 0 ||
+	    adev->mman.drv_vram_usage_size > vram_size)
+		return 0;
+
+	return amdgpu_bo_create_kernel_at(adev,
+					  adev->mman.drv_vram_usage_start_offset,
+					  adev->mman.drv_vram_usage_size,
+					  &adev->mman.drv_vram_usage_reserved_bo,
+					  &adev->mman.drv_vram_usage_va);
+}
+
 /*
  * Memoy training reservation functions
  */
@@ -1639,7 +1681,6 @@ static int amdgpu_ttm_reserve_tmr(struct amdgpu_device *adev)
 		ret = amdgpu_bo_create_kernel_at(adev,
 					 ctx->c2p_train_data_offset,
 					 ctx->train_data_size,
-					 AMDGPU_GEM_DOMAIN_VRAM,
 					 &ctx->c2p_bo,
 					 NULL);
 		if (ret) {
@@ -1653,7 +1694,6 @@ static int amdgpu_ttm_reserve_tmr(struct amdgpu_device *adev)
 	ret = amdgpu_bo_create_kernel_at(adev,
 				adev->gmc.real_vram_size - adev->mman.discovery_tmr_size,
 				adev->mman.discovery_tmr_size,
-				AMDGPU_GEM_DOMAIN_VRAM,
 				&adev->mman.discovery_memory,
 				NULL);
 	if (ret) {
@@ -1731,6 +1771,14 @@ int amdgpu_ttm_init(struct amdgpu_device *adev)
 	}
 
 	/*
+	 *The reserved vram for driver must be pinned to the specified
+	 *place on the VRAM, so reserve it early.
+	 */
+	r = amdgpu_ttm_drv_reserve_vram_init(adev);
+	if (r)
+		return r;
+
+	/*
 	 * only NAVI10 and onwards ASIC support for IP discovery.
 	 * If IP discovery enabled, a block of memory should be
 	 * reserved for IP discovey.
@@ -1746,21 +1794,18 @@ int amdgpu_ttm_init(struct amdgpu_device *adev)
 	 * avoid display artifacts while transitioning between pre-OS
 	 * and driver.  */
 	r = amdgpu_bo_create_kernel_at(adev, 0, adev->mman.stolen_vga_size,
-				       AMDGPU_GEM_DOMAIN_VRAM,
 				       &adev->mman.stolen_vga_memory,
 				       NULL);
 	if (r)
 		return r;
 	r = amdgpu_bo_create_kernel_at(adev, adev->mman.stolen_vga_size,
 				       adev->mman.stolen_extended_size,
-				       AMDGPU_GEM_DOMAIN_VRAM,
 				       &adev->mman.stolen_extended_memory,
 				       NULL);
 	if (r)
 		return r;
 	r = amdgpu_bo_create_kernel_at(adev, adev->mman.stolen_reserved_offset,
 				       adev->mman.stolen_reserved_size,
-				       AMDGPU_GEM_DOMAIN_VRAM,
 				       &adev->mman.stolen_reserved_memory,
 				       NULL);
 	if (r)
@@ -1855,6 +1900,7 @@ void amdgpu_ttm_fini(struct amdgpu_device *adev)
 	amdgpu_bo_free_kernel(&adev->mman.sdma_access_bo, NULL,
 					&adev->mman.sdma_access_ptr);
 	amdgpu_ttm_fw_reserve_vram_fini(adev);
+	amdgpu_ttm_drv_reserve_vram_fini(adev);
 
 	if (drm_dev_enter(adev_to_drm(adev), &idx)) {
 
@@ -1936,7 +1982,9 @@ static int amdgpu_ttm_prepare_job(struct amdgpu_device *adev,
 		AMDGPU_IB_POOL_DELAYED;
 	int r;
 
-	r = amdgpu_job_alloc_with_ib(adev, num_dw * 4, pool, job);
+	r = amdgpu_job_alloc_with_ib(adev, &adev->mman.entity,
+				     AMDGPU_FENCE_OWNER_UNDEFINED,
+				     num_dw * 4, pool, job);
 	if (r)
 		return r;
 
@@ -1946,17 +1994,11 @@ static int amdgpu_ttm_prepare_job(struct amdgpu_device *adev,
 							adev->gart.bo);
 		(*job)->vm_needs_flush = true;
 	}
-	if (resv) {
-		r = amdgpu_sync_resv(adev, &(*job)->sync, resv,
-				     AMDGPU_SYNC_ALWAYS,
-				     AMDGPU_FENCE_OWNER_UNDEFINED);
-		if (r) {
-			DRM_ERROR("sync failed (%d).\n", r);
-			amdgpu_job_free(*job);
-			return r;
-		}
-	}
-	return 0;
+	if (!resv)
+		return 0;
+
+	return drm_sched_job_add_resv_dependencies(&(*job)->base, resv,
+						   DMA_RESV_USAGE_BOOKKEEP);
 }
 
 int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset,
@@ -2001,8 +2043,7 @@ int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset,
 	if (direct_submit)
 		r = amdgpu_job_submit_direct(job, ring, fence);
 	else
-		r = amdgpu_job_submit(job, &adev->mman.entity,
-				      AMDGPU_FENCE_OWNER_UNDEFINED, fence);
+		*fence = amdgpu_job_submit(job);
 	if (r)
 		goto error_free;
 
@@ -2047,16 +2088,8 @@ static int amdgpu_ttm_fill_mem(struct amdgpu_ring *ring, uint32_t src_data,
 
 	amdgpu_ring_pad_ib(ring, &job->ibs[0]);
 	WARN_ON(job->ibs[0].length_dw > num_dw);
-	r = amdgpu_job_submit(job, &adev->mman.entity,
-			      AMDGPU_FENCE_OWNER_UNDEFINED, fence);
-	if (r)
-		goto error_free;
-
+	*fence = amdgpu_job_submit(job);
 	return 0;
-
-error_free:
-	amdgpu_job_free(job);
-	return r;
 }
 
 int amdgpu_fill_buffer(struct amdgpu_bo *bo,
@@ -2272,9 +2305,9 @@ static ssize_t amdgpu_iomem_read(struct file *f, char __user *buf,
 		if (p->mapping != adev->mman.bdev.dev_mapping)
 			return -EPERM;
 
-		ptr = kmap(p);
+		ptr = kmap_local_page(p);
 		r = copy_to_user(buf, ptr + off, bytes);
-		kunmap(p);
+		kunmap_local(ptr);
 		if (r)
 			return -EFAULT;
 
@@ -2323,9 +2356,9 @@ static ssize_t amdgpu_iomem_write(struct file *f, const char __user *buf,
 		if (p->mapping != adev->mman.bdev.dev_mapping)
 			return -EPERM;
 
-		ptr = kmap(p);
+		ptr = kmap_local_page(p);
 		r = copy_from_user(ptr + off, buf, bytes);
-		kunmap(p);
+		kunmap_local(ptr);
 		if (r)
 			return -EFAULT;