[PATCH 09/25] drm/amdgpu: Add GPUVM memory management functions for KFD

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Wed, Feb 7, 2018 at 3:32 AM, Felix Kuehling <Felix.Kuehling at amd.com> wrote:
> v2:
> * Removed unused flags from struct kgd_mem
> * Updated some comments
> * Added a check to unmap_memory_from_gpu whether BO was mapped
>
> Signed-off-by: Felix Kuehling <Felix.Kuehling at amd.com>
> ---
>  drivers/gpu/drm/amd/amdgpu/Makefile               |    1 +
>  drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c        |    1 +
>  drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h        |   91 +-
>  drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v7.c |   66 +-
>  drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v8.c |   67 +-
>  drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c  | 1501 +++++++++++++++++++++
>  drivers/gpu/drm/amd/amdgpu/amdgpu_object.c        |    4 +
>  drivers/gpu/drm/amd/amdgpu/amdgpu_object.h        |    2 +
>  drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c           |    7 +
>  drivers/gpu/drm/amd/include/kgd_kfd_interface.h   |   77 ++
>  10 files changed, 1813 insertions(+), 4 deletions(-)
>  create mode 100644 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/Makefile b/drivers/gpu/drm/amd/amdgpu/Makefile
> index 43dc3f9..180b2a8 100644
> --- a/drivers/gpu/drm/amd/amdgpu/Makefile
> +++ b/drivers/gpu/drm/amd/amdgpu/Makefile
> @@ -131,6 +131,7 @@ amdgpu-y += \
>  amdgpu-y += \
>          amdgpu_amdkfd.o \
>          amdgpu_amdkfd_fence.o \
> +        amdgpu_amdkfd_gpuvm.o \
>          amdgpu_amdkfd_gfx_v8.o
>
>  # add cgs
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
> index 25c2aed..01fb142 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
> @@ -58,6 +58,7 @@ int amdgpu_amdkfd_init(void)
>  #else
>         ret = -ENOENT;
>  #endif
> +       amdgpu_amdkfd_gpuvm_init_mem_limits();
>
>         return ret;
>  }
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
> index 9bed9fc..87fb4e6 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
> @@ -28,15 +28,41 @@
>  #include <linux/types.h>
>  #include <linux/mmu_context.h>
>  #include <kgd_kfd_interface.h>
> +#include <drm/ttm/ttm_execbuf_util.h>
> +#include "amdgpu_sync.h"
> +#include "amdgpu_vm.h"
>
>  extern const struct kgd2kfd_calls *kgd2kfd;
>
>  struct amdgpu_device;
>
> +struct kfd_bo_va_list {
> +       struct list_head bo_list;
> +       struct amdgpu_bo_va *bo_va;
> +       void *kgd_dev;
> +       bool is_mapped;
> +       uint64_t va;
> +       uint64_t pte_flags;
> +};
> +
>  struct kgd_mem {
> +       struct mutex lock;
>         struct amdgpu_bo *bo;
> -       uint64_t gpu_addr;
> -       void *cpu_ptr;
> +       struct list_head bo_va_list;
> +       /* protected by amdkfd_process_info.lock */
> +       struct ttm_validate_buffer validate_list;
> +       struct ttm_validate_buffer resv_list;
> +       uint32_t domain;
> +       unsigned int mapped_to_gpu_memory;
> +       uint64_t va;
> +
> +       uint32_t mapping_flags;
> +
> +       struct amdkfd_process_info *process_info;
> +
> +       struct amdgpu_sync sync;
> +
> +       bool aql_queue;
>  };
>
>  /* KFD Memory Eviction */
> @@ -52,6 +78,41 @@ struct amdgpu_amdkfd_fence *amdgpu_amdkfd_fence_create(u64 context,
>  bool amd_kfd_fence_check_mm(struct dma_fence *f, struct mm_struct *mm);
>  struct amdgpu_amdkfd_fence *to_amdgpu_amdkfd_fence(struct dma_fence *f);
>
> +struct amdkfd_process_info {
> +       /* List head of all VMs that belong to a KFD process */
> +       struct list_head vm_list_head;
> +       /* List head for all KFD BOs that belong to a KFD process. */
> +       struct list_head kfd_bo_list;
> +       /* Lock to protect kfd_bo_list */
> +       struct mutex lock;
> +
> +       /* Number of VMs */
> +       unsigned int n_vms;
> +       /* Eviction Fence */
> +       struct amdgpu_amdkfd_fence *eviction_fence;
> +};
> +
> +/* struct amdkfd_vm -
> + * For Memory Eviction KGD requires a mechanism to keep track of all KFD BOs
> + * belonging to a KFD process. All the VMs belonging to the same process point
> + * to the same amdkfd_process_info.
> + */
> +struct amdkfd_vm {
> +       /* Keep base as the first parameter for pointer compatibility between
> +        * amdkfd_vm and amdgpu_vm.
> +        */
> +       struct amdgpu_vm base;
> +
> +       /* List node in amdkfd_process_info.vm_list_head*/
> +       struct list_head vm_list_node;
> +
> +       struct amdgpu_device *adev;
> +       /* Points to the KFD process VM info*/
> +       struct amdkfd_process_info *process_info;
> +
> +       uint64_t pd_phys_addr;
> +};
> +
>  int amdgpu_amdkfd_init(void);
>  void amdgpu_amdkfd_fini(void);
>
> @@ -96,4 +157,30 @@ uint64_t amdgpu_amdkfd_get_vram_usage(struct kgd_dev *kgd);
>                 valid;                                                  \
>         })
>
> +/* GPUVM API */
> +int amdgpu_amdkfd_gpuvm_create_process_vm(struct kgd_dev *kgd, void **vm,
> +                                         void **process_info,
> +                                         struct dma_fence **ef);
> +void amdgpu_amdkfd_gpuvm_destroy_process_vm(struct kgd_dev *kgd, void *vm);
> +uint32_t amdgpu_amdkfd_gpuvm_get_process_page_dir(void *vm);
> +int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
> +               struct kgd_dev *kgd, uint64_t va, uint64_t size,
> +               void *vm, struct kgd_mem **mem,
> +               uint64_t *offset, uint32_t flags);
> +int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(
> +               struct kgd_dev *kgd, struct kgd_mem *mem);
> +int amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
> +               struct kgd_dev *kgd, struct kgd_mem *mem, void *vm);
> +int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
> +               struct kgd_dev *kgd, struct kgd_mem *mem, void *vm);
> +int amdgpu_amdkfd_gpuvm_sync_memory(
> +               struct kgd_dev *kgd, struct kgd_mem *mem, bool intr);
> +int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct kgd_dev *kgd,
> +               struct kgd_mem *mem, void **kptr, uint64_t *size);
> +int amdgpu_amdkfd_gpuvm_restore_process_bos(void *process_info,
> +                                           struct dma_fence **ef);
> +
> +void amdgpu_amdkfd_gpuvm_init_mem_limits(void);
> +void amdgpu_amdkfd_unreserve_system_memory_limit(struct amdgpu_bo *bo);
> +
>  #endif /* AMDGPU_AMDKFD_H_INCLUDED */
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v7.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v7.c
> index 1362181..65783d1 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v7.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v7.c
> @@ -143,6 +143,10 @@ static uint16_t get_atc_vmid_pasid_mapping_pasid(struct kgd_dev *kgd,
>  static uint16_t get_fw_version(struct kgd_dev *kgd, enum kgd_engine_type type);
>  static void set_scratch_backing_va(struct kgd_dev *kgd,
>                                         uint64_t va, uint32_t vmid);
> +static void set_vm_context_page_table_base(struct kgd_dev *kgd, uint32_t vmid,
> +               uint32_t page_table_base);
> +static int invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid);
> +static int invalidate_tlbs_vmid(struct kgd_dev *kgd, uint16_t vmid);
>
>  /* Because of REG_GET_FIELD() being used, we put this function in the
>   * asic specific file.
> @@ -199,7 +203,20 @@ static const struct kfd2kgd_calls kfd2kgd = {
>         .set_scratch_backing_va = set_scratch_backing_va,
>         .get_tile_config = get_tile_config,
>         .get_cu_info = get_cu_info,
> -       .get_vram_usage = amdgpu_amdkfd_get_vram_usage
> +       .get_vram_usage = amdgpu_amdkfd_get_vram_usage,
> +       .create_process_vm = amdgpu_amdkfd_gpuvm_create_process_vm,
> +       .destroy_process_vm = amdgpu_amdkfd_gpuvm_destroy_process_vm,
> +       .get_process_page_dir = amdgpu_amdkfd_gpuvm_get_process_page_dir,
> +       .set_vm_context_page_table_base = set_vm_context_page_table_base,
> +       .alloc_memory_of_gpu = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu,
> +       .free_memory_of_gpu = amdgpu_amdkfd_gpuvm_free_memory_of_gpu,
> +       .map_memory_to_gpu = amdgpu_amdkfd_gpuvm_map_memory_to_gpu,
> +       .unmap_memory_to_gpu = amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu,
> +       .sync_memory = amdgpu_amdkfd_gpuvm_sync_memory,
> +       .map_gtt_bo_to_kernel = amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel,
> +       .restore_process_bos = amdgpu_amdkfd_gpuvm_restore_process_bos,
> +       .invalidate_tlbs = invalidate_tlbs,
> +       .invalidate_tlbs_vmid = invalidate_tlbs_vmid,
>  };
>
>  struct kfd2kgd_calls *amdgpu_amdkfd_gfx_7_get_functions(void)
> @@ -855,3 +872,50 @@ static uint16_t get_fw_version(struct kgd_dev *kgd, enum kgd_engine_type type)
>         return hdr->common.ucode_version;
>  }
>
> +static void set_vm_context_page_table_base(struct kgd_dev *kgd, uint32_t vmid,
> +                       uint32_t page_table_base)
> +{
> +       struct amdgpu_device *adev = get_amdgpu_device(kgd);
> +
> +       if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
> +               pr_err("trying to set page table base for wrong VMID\n");
> +               return;
> +       }
> +       WREG32(mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + vmid - 8, page_table_base);
> +}
> +
> +static int invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid)
> +{
> +       struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
> +       int vmid;
> +       unsigned int tmp;
> +
> +       for (vmid = 0; vmid < 16; vmid++) {
> +               if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid))
> +                       continue;
> +
> +               tmp = RREG32(mmATC_VMID0_PASID_MAPPING + vmid);
> +               if ((tmp & ATC_VMID0_PASID_MAPPING__VALID_MASK) &&
> +                       (tmp & ATC_VMID0_PASID_MAPPING__PASID_MASK) == pasid) {
> +                       WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid);
> +                       RREG32(mmVM_INVALIDATE_RESPONSE);
> +                       break;
> +               }
> +       }
> +
> +       return 0;
> +}
> +
> +static int invalidate_tlbs_vmid(struct kgd_dev *kgd, uint16_t vmid)
> +{
> +       struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
> +
> +       if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
> +               pr_err("non kfd vmid\n");
> +               return 0;
> +       }
> +
> +       WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid);
> +       RREG32(mmVM_INVALIDATE_RESPONSE);
> +       return 0;
> +}
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v8.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v8.c
> index 5130eac..1b5bf13 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v8.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v8.c
> @@ -101,6 +101,10 @@ static uint16_t get_atc_vmid_pasid_mapping_pasid(struct kgd_dev *kgd,
>  static uint16_t get_fw_version(struct kgd_dev *kgd, enum kgd_engine_type type);
>  static void set_scratch_backing_va(struct kgd_dev *kgd,
>                                         uint64_t va, uint32_t vmid);
> +static void set_vm_context_page_table_base(struct kgd_dev *kgd, uint32_t vmid,
> +               uint32_t page_table_base);
> +static int invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid);
> +static int invalidate_tlbs_vmid(struct kgd_dev *kgd, uint16_t vmid);
>
>  /* Because of REG_GET_FIELD() being used, we put this function in the
>   * asic specific file.
> @@ -159,7 +163,20 @@ static const struct kfd2kgd_calls kfd2kgd = {
>         .set_scratch_backing_va = set_scratch_backing_va,
>         .get_tile_config = get_tile_config,
>         .get_cu_info = get_cu_info,
> -       .get_vram_usage = amdgpu_amdkfd_get_vram_usage
> +       .get_vram_usage = amdgpu_amdkfd_get_vram_usage,
> +       .create_process_vm = amdgpu_amdkfd_gpuvm_create_process_vm,
> +       .destroy_process_vm = amdgpu_amdkfd_gpuvm_destroy_process_vm,
> +       .get_process_page_dir = amdgpu_amdkfd_gpuvm_get_process_page_dir,
> +       .set_vm_context_page_table_base = set_vm_context_page_table_base,
> +       .alloc_memory_of_gpu = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu,
> +       .free_memory_of_gpu = amdgpu_amdkfd_gpuvm_free_memory_of_gpu,
> +       .map_memory_to_gpu = amdgpu_amdkfd_gpuvm_map_memory_to_gpu,
> +       .unmap_memory_to_gpu = amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu,
> +       .sync_memory = amdgpu_amdkfd_gpuvm_sync_memory,
> +       .map_gtt_bo_to_kernel = amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel,
> +       .restore_process_bos = amdgpu_amdkfd_gpuvm_restore_process_bos,
> +       .invalidate_tlbs = invalidate_tlbs,
> +       .invalidate_tlbs_vmid = invalidate_tlbs_vmid,
>  };
>
>  struct kfd2kgd_calls *amdgpu_amdkfd_gfx_8_0_get_functions(void)
> @@ -816,3 +833,51 @@ static uint16_t get_fw_version(struct kgd_dev *kgd, enum kgd_engine_type type)
>         /* Only 12 bit in use*/
>         return hdr->common.ucode_version;
>  }
> +
> +static void set_vm_context_page_table_base(struct kgd_dev *kgd, uint32_t vmid,
> +               uint32_t page_table_base)
> +{
> +       struct amdgpu_device *adev = get_amdgpu_device(kgd);
> +
> +       if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
> +               pr_err("trying to set page table base for wrong VMID\n");
> +               return;
> +       }
> +       WREG32(mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + vmid - 8, page_table_base);
> +}
> +
> +static int invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid)
> +{
> +       struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
> +       int vmid;
> +       unsigned int tmp;
> +
> +       for (vmid = 0; vmid < 16; vmid++) {
> +               if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid))
> +                       continue;
> +
> +               tmp = RREG32(mmATC_VMID0_PASID_MAPPING + vmid);
> +               if ((tmp & ATC_VMID0_PASID_MAPPING__VALID_MASK) &&
> +                       (tmp & ATC_VMID0_PASID_MAPPING__PASID_MASK) == pasid) {
> +                       WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid);
> +                       RREG32(mmVM_INVALIDATE_RESPONSE);
> +                       break;
> +               }
> +       }
> +
> +       return 0;
> +}
> +
> +static int invalidate_tlbs_vmid(struct kgd_dev *kgd, uint16_t vmid)
> +{
> +       struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
> +
> +       if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
> +               pr_err("non kfd vmid %d\n", vmid);
> +               return -EINVAL;
> +       }
> +
> +       WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid);
> +       RREG32(mmVM_INVALIDATE_RESPONSE);
> +       return 0;
> +}
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
> new file mode 100644
> index 0000000..9703fd0
> --- /dev/null
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
> @@ -0,0 +1,1501 @@
> +/*
> + * Copyright 2014-2018 Advanced Micro Devices, Inc.
> + *
> + * Permission is hereby granted, free of charge, to any person obtaining a
> + * copy of this software and associated documentation files (the "Software"),
> + * to deal in the Software without restriction, including without limitation
> + * the rights to use, copy, modify, merge, publish, distribute, sublicense,
> + * and/or sell copies of the Software, and to permit persons to whom the
> + * Software is furnished to do so, subject to the following conditions:
> + *
> + * The above copyright notice and this permission notice shall be included in
> + * all copies or substantial portions of the Software.
> + *
> + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
> + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
> + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
> + * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
> + * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
> + * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
> + * OTHER DEALINGS IN THE SOFTWARE.
> + */
> +
> +#define pr_fmt(fmt) "kfd2kgd: " fmt
> +
> +#include <linux/list.h>
> +#include <drm/drmP.h>
> +#include "amdgpu_object.h"
> +#include "amdgpu_vm.h"
> +#include "amdgpu_amdkfd.h"
> +
> +/* Special VM and GART address alignment needed for VI pre-Fiji due to
> + * a HW bug.
> + */
> +#define VI_BO_SIZE_ALIGN (0x8000)
> +
> +/* Impose limit on how much memory KFD can use */
> +static struct {
> +       uint64_t max_system_mem_limit;
> +       int64_t system_mem_used;
> +       spinlock_t mem_limit_lock;
> +} kfd_mem_limit;
> +
> +/* Struct used for amdgpu_amdkfd_bo_validate */
> +struct amdgpu_vm_parser {
> +       uint32_t        domain;
> +       bool            wait;
> +};
> +
> +static const char * const domain_bit_to_string[] = {
> +               "CPU",
> +               "GTT",
> +               "VRAM",
> +               "GDS",
> +               "GWS",
> +               "OA"
> +};
> +
> +#define domain_string(domain) domain_bit_to_string[ffs(domain)-1]
> +
> +
> +
> +static inline struct amdgpu_device *get_amdgpu_device(struct kgd_dev *kgd)
> +{
> +       return (struct amdgpu_device *)kgd;
> +}
> +
> +static bool check_if_add_bo_to_vm(struct amdgpu_vm *avm,
> +               struct kgd_mem *mem)
> +{
> +       struct kfd_bo_va_list *entry;
> +
> +       list_for_each_entry(entry, &mem->bo_va_list, bo_list)
> +               if (entry->bo_va->base.vm == avm)
> +                       return false;
> +
> +       return true;
> +}
> +
> +/* Set memory usage limits. Current, limits are
> + *  System (kernel) memory - 3/8th System RAM
> + */
> +void amdgpu_amdkfd_gpuvm_init_mem_limits(void)
> +{
> +       struct sysinfo si;
> +       uint64_t mem;
> +
> +       si_meminfo(&si);
> +       mem = si.totalram - si.totalhigh;
> +       mem *= si.mem_unit;
> +
> +       spin_lock_init(&kfd_mem_limit.mem_limit_lock);
> +       kfd_mem_limit.max_system_mem_limit = (mem >> 1) - (mem >> 3);
> +       pr_debug("Kernel memory limit %lluM\n",
> +               (kfd_mem_limit.max_system_mem_limit >> 20));
> +}
> +
> +static int amdgpu_amdkfd_reserve_system_mem_limit(struct amdgpu_device *adev,
> +                                             uint64_t size, u32 domain)
> +{
> +       size_t acc_size;
> +       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);
> +       if (domain == AMDGPU_GEM_DOMAIN_GTT) {
> +               if (kfd_mem_limit.system_mem_used + (acc_size + size) >
> +                       kfd_mem_limit.max_system_mem_limit) {
> +                       ret = -ENOMEM;
> +                       goto err_no_mem;
> +               }
> +               kfd_mem_limit.system_mem_used += (acc_size + size);
> +       }
> +err_no_mem:
> +       spin_unlock(&kfd_mem_limit.mem_limit_lock);
> +       return ret;
> +}
> +
> +static void unreserve_system_mem_limit(struct amdgpu_device *adev,
> +                                      uint64_t size, u32 domain)
> +{
> +       size_t acc_size;
> +
> +       acc_size = ttm_bo_dma_acc_size(&adev->mman.bdev, size,
> +                                      sizeof(struct amdgpu_bo));
> +
> +       spin_lock(&kfd_mem_limit.mem_limit_lock);
> +       if (domain == AMDGPU_GEM_DOMAIN_GTT)
> +               kfd_mem_limit.system_mem_used -= (acc_size + size);
> +       WARN_ONCE(kfd_mem_limit.system_mem_used < 0,
> +                 "kfd system memory accounting unbalanced");
> +
> +       spin_unlock(&kfd_mem_limit.mem_limit_lock);
> +}
> +
> +void amdgpu_amdkfd_unreserve_system_memory_limit(struct amdgpu_bo *bo)
> +{
> +       spin_lock(&kfd_mem_limit.mem_limit_lock);
> +
> +       if (bo->preferred_domains == AMDGPU_GEM_DOMAIN_GTT) {
> +               kfd_mem_limit.system_mem_used -=
> +                       (bo->tbo.acc_size + amdgpu_bo_size(bo));
> +       }
> +       WARN_ONCE(kfd_mem_limit.system_mem_used < 0,
> +                 "kfd system memory accounting unbalanced");
> +
> +       spin_unlock(&kfd_mem_limit.mem_limit_lock);
> +}
> +
> +
> +/* amdgpu_amdkfd_remove_eviction_fence - Removes eviction fence(s) from BO's
> + *  reservation object.
> + *
> + * @bo: [IN] Remove eviction fence(s) from this BO
> + * @ef: [IN] If ef is specified, then this eviction fence is removed if it
> + *  is present in the shared list.
> + * @ef_list: [OUT] Returns list of eviction fences. These fences are removed
> + *  from BO's reservation object shared list.
> + * @ef_count: [OUT] Number of fences in ef_list.
> + *
> + * NOTE: If called with ef_list, then amdgpu_amdkfd_add_eviction_fence must be
> + *  called to restore the eviction fences and to avoid memory leak. This is
> + *  useful for shared BOs.
> + * NOTE: Must be called with BO reserved i.e. bo->tbo.resv->lock held.
> + */
> +static int amdgpu_amdkfd_remove_eviction_fence(struct amdgpu_bo *bo,
> +                                       struct amdgpu_amdkfd_fence *ef,
> +                                       struct amdgpu_amdkfd_fence ***ef_list,
> +                                       unsigned int *ef_count)
> +{
> +       struct reservation_object_list *fobj;
> +       struct reservation_object *resv;
> +       unsigned int i = 0, j = 0, k = 0, shared_count;
> +       unsigned int count = 0;
> +       struct amdgpu_amdkfd_fence **fence_list;
> +
> +       if (!ef && !ef_list)
> +               return -EINVAL;
> +
> +       if (ef_list) {
> +               *ef_list = NULL;
> +               *ef_count = 0;
> +       }
> +
> +       resv = bo->tbo.resv;
> +       fobj = reservation_object_get_list(resv);
> +
> +       if (!fobj)
> +               return 0;
> +
> +       preempt_disable();
> +       write_seqcount_begin(&resv->seq);
> +
> +       /* Go through all the shared fences in the resevation object. If
> +        * ef is specified and it exists in the list, remove it and reduce the
> +        * count. If ef is not specified, then get the count of eviction fences
> +        * present.
> +        */
> +       shared_count = fobj->shared_count;
> +       for (i = 0; i < shared_count; ++i) {
> +               struct dma_fence *f;
> +
> +               f = rcu_dereference_protected(fobj->shared[i],
> +                                             reservation_object_held(resv));
> +
> +               if (ef) {
> +                       if (f->context == ef->base.context) {
> +                               dma_fence_put(f);
> +                               fobj->shared_count--;
> +                       } else
> +                               RCU_INIT_POINTER(fobj->shared[j++], f);
> +
> +               } else if (to_amdgpu_amdkfd_fence(f))
> +                       count++;
> +       }
> +       write_seqcount_end(&resv->seq);
> +       preempt_enable();
> +
> +       if (ef || !count)
> +               return 0;
> +
> +       /* Alloc memory for count number of eviction fence pointers. Fill the
> +        * ef_list array and ef_count
> +        */
> +       fence_list = kcalloc(count, sizeof(struct amdgpu_amdkfd_fence *),
> +                            GFP_KERNEL);
> +       if (!fence_list)
> +               return -ENOMEM;
> +
> +       preempt_disable();
> +       write_seqcount_begin(&resv->seq);
> +
> +       j = 0;
> +       for (i = 0; i < shared_count; ++i) {
> +               struct dma_fence *f;
> +               struct amdgpu_amdkfd_fence *efence;
> +
> +               f = rcu_dereference_protected(fobj->shared[i],
> +                       reservation_object_held(resv));
> +
> +               efence = to_amdgpu_amdkfd_fence(f);
> +               if (efence) {
> +                       fence_list[k++] = efence;
> +                       fobj->shared_count--;
> +               } else
> +                       RCU_INIT_POINTER(fobj->shared[j++], f);
> +       }
> +
> +       write_seqcount_end(&resv->seq);
> +       preempt_enable();
> +
> +       *ef_list = fence_list;
> +       *ef_count = k;
> +
> +       return 0;
> +}
> +
> +/* amdgpu_amdkfd_add_eviction_fence - Adds eviction fence(s) back into BO's
> + *  reservation object.
> + *
> + * @bo: [IN] Add eviction fences to this BO
> + * @ef_list: [IN] List of eviction fences to be added
> + * @ef_count: [IN] Number of fences in ef_list.
> + *
> + * NOTE: Must call amdgpu_amdkfd_remove_eviction_fence before calling this
> + *  function.
> + */
> +static void amdgpu_amdkfd_add_eviction_fence(struct amdgpu_bo *bo,
> +                               struct amdgpu_amdkfd_fence **ef_list,
> +                               unsigned int ef_count)
> +{
> +       int i;
> +
> +       if (!ef_list || !ef_count)
> +               return;
> +
> +       for (i = 0; i < ef_count; i++) {
> +               amdgpu_bo_fence(bo, &ef_list[i]->base, true);
> +               /* Readding the fence takes an additional reference. Drop that
> +                * reference.
> +                */
> +               dma_fence_put(&ef_list[i]->base);
> +       }
> +
> +       kfree(ef_list);
> +}
> +
> +static int amdgpu_amdkfd_bo_validate(struct amdgpu_bo *bo, uint32_t domain,
> +                                    bool wait)
> +{
> +       struct ttm_operation_ctx ctx = { false, false };
> +       int ret;
> +
> +       if (WARN(amdgpu_ttm_tt_get_usermm(bo->tbo.ttm),
> +                "Called with userptr BO"))
> +               return -EINVAL;
> +
> +       amdgpu_ttm_placement_from_domain(bo, domain);
> +
> +       ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
> +       if (ret)
> +               goto validate_fail;
> +       if (wait) {
> +               struct amdgpu_amdkfd_fence **ef_list;
> +               unsigned int ef_count;
> +
> +               ret = amdgpu_amdkfd_remove_eviction_fence(bo, NULL, &ef_list,
> +                                                         &ef_count);
> +               if (ret)
> +                       goto validate_fail;
> +
> +               ttm_bo_wait(&bo->tbo, false, false);
> +               amdgpu_amdkfd_add_eviction_fence(bo, ef_list, ef_count);
> +       }
> +
> +validate_fail:
> +       return ret;
> +}
> +
> +static int amdgpu_amdkfd_validate(void *param, struct amdgpu_bo *bo)
> +{
> +       struct amdgpu_vm_parser *p = param;
> +
> +       return amdgpu_amdkfd_bo_validate(bo, p->domain, p->wait);
> +}
> +
> +/* vm_validate_pt_pd_bos - Validate page table and directory BOs
> + *
> + * Page directories are not updated here because huge page handling
> + * during page table updates can invalidate page directory entries
> + * again. Page directories are only updated after updating page
> + * tables.
> + */
> +static int vm_validate_pt_pd_bos(struct amdkfd_vm *vm)
> +{
> +       struct amdgpu_bo *pd = vm->base.root.base.bo;
> +       struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
> +       struct amdgpu_vm_parser param;
> +       uint64_t addr, flags = AMDGPU_PTE_VALID;
> +       int ret;
> +
> +       param.domain = AMDGPU_GEM_DOMAIN_VRAM;
> +       param.wait = false;
> +
> +       ret = amdgpu_vm_validate_pt_bos(adev, &vm->base, amdgpu_amdkfd_validate,
> +                                       &param);
> +       if (ret) {
> +               pr_err("amdgpu: failed to validate PT BOs\n");
> +               return ret;
> +       }
> +
> +       ret = amdgpu_amdkfd_validate(&param, pd);
> +       if (ret) {
> +               pr_err("amdgpu: failed to validate PD\n");
> +               return ret;
> +       }
> +
> +       addr = amdgpu_bo_gpu_offset(vm->base.root.base.bo);
> +       amdgpu_gart_get_vm_pde(adev, -1, &addr, &flags);
> +       vm->pd_phys_addr = addr;
> +
> +       if (vm->base.use_cpu_for_update) {
> +               ret = amdgpu_bo_kmap(pd, NULL);
> +               if (ret) {
> +                       pr_err("amdgpu: failed to kmap PD, ret=%d\n", ret);
> +                       return ret;
> +               }
> +       }
> +
> +       return 0;
> +}
> +
> +static int sync_vm_fence(struct amdgpu_device *adev, struct amdgpu_sync *sync,
> +                        struct dma_fence *f)
> +{
> +       int ret = amdgpu_sync_fence(adev, sync, f, false);
> +
> +       /* Sync objects can't handle multiple GPUs (contexts) updating
> +        * sync->last_vm_update. Fortunately we don't need it for
> +        * KFD's purposes, so we can just drop that fence.
> +        */
> +       if (sync->last_vm_update) {
> +               dma_fence_put(sync->last_vm_update);
> +               sync->last_vm_update = NULL;
> +       }
> +
> +       return ret;
> +}
> +
> +static int vm_update_pds(struct amdgpu_vm *vm, struct amdgpu_sync *sync)
> +{
> +       struct amdgpu_bo *pd = vm->root.base.bo;
> +       struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
> +       int ret;
> +
> +       ret = amdgpu_vm_update_directories(adev, vm);
> +       if (ret)
> +               return ret;
> +
> +       return sync_vm_fence(adev, sync, vm->last_update);
> +}
> +
> +/* add_bo_to_vm - Add a BO to a VM
> + *
> + * Everything that needs to bo done only once when a BO is first added
> + * to a VM. It can later be mapped and unmapped many times without
> + * repeating these steps.
> + *
> + * 1. Allocate and initialize BO VA entry data structure
> + * 2. Add BO to the VM
> + * 3. Determine ASIC-specific PTE flags
> + * 4. Alloc page tables and directories if needed
> + * 4a.  Validate new page tables and directories
> + */
> +static int add_bo_to_vm(struct amdgpu_device *adev, struct kgd_mem *mem,
> +               struct amdgpu_vm *avm, bool is_aql,
> +               struct kfd_bo_va_list **p_bo_va_entry)
> +{
> +       int ret;
> +       struct kfd_bo_va_list *bo_va_entry;
> +       struct amdkfd_vm *kvm = container_of(avm,
> +                                            struct amdkfd_vm, base);
> +       struct amdgpu_bo *pd = avm->root.base.bo;
> +       struct amdgpu_bo *bo = mem->bo;
> +       uint64_t va = mem->va;
> +       struct list_head *list_bo_va = &mem->bo_va_list;
> +       unsigned long bo_size = bo->tbo.mem.size;
> +
> +       if (!va) {
> +               pr_err("Invalid VA when adding BO to VM\n");
> +               return -EINVAL;
> +       }
> +
> +       if (is_aql)
> +               va += bo_size;
> +
> +       bo_va_entry = kzalloc(sizeof(*bo_va_entry), GFP_KERNEL);
> +       if (!bo_va_entry)
> +               return -ENOMEM;
> +
> +       pr_debug("\t add VA 0x%llx - 0x%llx to vm %p\n", va,
> +                       va + bo_size, avm);
> +
> +       /* Add BO to VM internal data structures*/
> +       bo_va_entry->bo_va = amdgpu_vm_bo_add(adev, avm, bo);
> +       if (!bo_va_entry->bo_va) {
> +               ret = -EINVAL;
> +               pr_err("Failed to add BO object to VM. ret == %d\n",
> +                               ret);
> +               goto err_vmadd;
> +       }
> +
> +       bo_va_entry->va = va;
> +       bo_va_entry->pte_flags = amdgpu_vm_get_pte_flags(adev,
> +                                                        mem->mapping_flags);
> +       bo_va_entry->kgd_dev = (void *)adev;
> +       list_add(&bo_va_entry->bo_list, list_bo_va);
> +
> +       if (p_bo_va_entry)
> +               *p_bo_va_entry = bo_va_entry;
> +
> +       /* Allocate new page tables if neeeded and validate
> +        * them. Clearing of new page tables and validate need to wait
> +        * on move fences. We don't want that to trigger the eviction
> +        * fence, so remove it temporarily.
> +        */
> +       amdgpu_amdkfd_remove_eviction_fence(pd,
> +                                       kvm->process_info->eviction_fence,
> +                                       NULL, NULL);
> +
> +       ret = amdgpu_vm_alloc_pts(adev, avm, va, amdgpu_bo_size(bo));
> +       if (ret) {
> +               pr_err("Failed to allocate pts, err=%d\n", ret);
> +               goto err_alloc_pts;
> +       }
> +
> +       ret = vm_validate_pt_pd_bos(kvm);
> +       if (ret) {
> +               pr_err("validate_pt_pd_bos() failed\n");
> +               goto err_alloc_pts;
> +       }
> +
> +       /* Add the eviction fence back */
> +       amdgpu_bo_fence(pd, &kvm->process_info->eviction_fence->base, true);
> +
> +       return 0;
> +
> +err_alloc_pts:
> +       amdgpu_bo_fence(pd, &kvm->process_info->eviction_fence->base, true);
> +       amdgpu_vm_bo_rmv(adev, bo_va_entry->bo_va);
> +       list_del(&bo_va_entry->bo_list);
> +err_vmadd:
> +       kfree(bo_va_entry);
> +       return ret;
> +}
> +
> +static void remove_bo_from_vm(struct amdgpu_device *adev,
> +               struct kfd_bo_va_list *entry, unsigned long size)
> +{
> +       pr_debug("\t remove VA 0x%llx - 0x%llx in entry %p\n",
> +                       entry->va,
> +                       entry->va + size, entry);
> +       amdgpu_vm_bo_rmv(adev, entry->bo_va);
> +       list_del(&entry->bo_list);
> +       kfree(entry);
> +}
> +
> +static void add_kgd_mem_to_kfd_bo_list(struct kgd_mem *mem,
> +                               struct amdkfd_process_info *process_info)
> +{
> +       struct ttm_validate_buffer *entry = &mem->validate_list;
> +       struct amdgpu_bo *bo = mem->bo;
> +
> +       INIT_LIST_HEAD(&entry->head);
> +       entry->shared = true;
> +       entry->bo = &bo->tbo;
> +       mutex_lock(&process_info->lock);
> +       list_add_tail(&entry->head, &process_info->kfd_bo_list);
> +       mutex_unlock(&process_info->lock);
> +}
> +
> +/* Reserving a BO and its page table BOs must happen atomically to
> + * avoid deadlocks. Some operations update multiple VMs at once. Track
> + * all the reservation info in a context structure. Optionally a sync
> + * object can track VM updates.
> + */
> +struct bo_vm_reservation_context {
> +       struct amdgpu_bo_list_entry kfd_bo; /* BO list entry for the KFD BO */
> +       unsigned int n_vms;                 /* Number of VMs reserved       */
> +       struct amdgpu_bo_list_entry *vm_pd; /* Array of VM BO list entries  */
> +       struct ww_acquire_ctx ticket;       /* Reservation ticket           */
> +       struct list_head list, duplicates;  /* BO lists                     */
> +       struct amdgpu_sync *sync;           /* Pointer to sync object       */
> +       bool reserved;                      /* Whether BOs are reserved     */
> +};
> +
> +enum bo_vm_match {
> +       BO_VM_NOT_MAPPED = 0,   /* Match VMs where a BO is not mapped */
> +       BO_VM_MAPPED,           /* Match VMs where a BO is mapped     */
> +       BO_VM_ALL,              /* Match all VMs a BO was added to    */
> +};
> +
> +/**
> + * reserve_bo_and_vm - reserve a BO and a VM unconditionally.
> + * @mem: KFD BO structure.
> + * @vm: the VM to reserve.
> + * @ctx: the struct that will be used in unreserve_bo_and_vms().
> + */
> +static int reserve_bo_and_vm(struct kgd_mem *mem,
> +                             struct amdgpu_vm *vm,
> +                             struct bo_vm_reservation_context *ctx)
> +{
> +       struct amdgpu_bo *bo = mem->bo;
> +       int ret;
> +
> +       WARN_ON(!vm);
> +
> +       ctx->reserved = false;
> +       ctx->n_vms = 1;
> +       ctx->sync = &mem->sync;
> +
> +       INIT_LIST_HEAD(&ctx->list);
> +       INIT_LIST_HEAD(&ctx->duplicates);
> +
> +       ctx->vm_pd = kcalloc(ctx->n_vms, sizeof(*ctx->vm_pd), GFP_KERNEL);
> +       if (!ctx->vm_pd)
> +               return -ENOMEM;
> +
> +       ctx->kfd_bo.robj = bo;
> +       ctx->kfd_bo.priority = 0;
> +       ctx->kfd_bo.tv.bo = &bo->tbo;
> +       ctx->kfd_bo.tv.shared = true;
> +       ctx->kfd_bo.user_pages = NULL;
> +       list_add(&ctx->kfd_bo.tv.head, &ctx->list);
> +
> +       amdgpu_vm_get_pd_bo(vm, &ctx->list, &ctx->vm_pd[0]);
> +
> +       ret = ttm_eu_reserve_buffers(&ctx->ticket, &ctx->list,
> +                                    false, &ctx->duplicates);
> +       if (!ret)
> +               ctx->reserved = true;
> +       else {
> +               pr_err("Failed to reserve buffers in ttm\n");
> +               kfree(ctx->vm_pd);
> +               ctx->vm_pd = NULL;
> +       }
> +
> +       return ret;
> +}
> +
> +/**
> + * reserve_bo_and_cond_vms - reserve a BO and some VMs conditionally
> + * @mem: KFD BO structure.
> + * @vm: the VM to reserve. If NULL, then all VMs associated with the BO
> + * is used. Otherwise, a single VM associated with the BO.
> + * @map_type: the mapping status that will be used to filter the VMs.
> + * @ctx: the struct that will be used in unreserve_bo_and_vms().
> + *
> + * Returns 0 for success, negative for failure.
> + */
> +static int reserve_bo_and_cond_vms(struct kgd_mem *mem,
> +                               struct amdgpu_vm *vm, enum bo_vm_match map_type,
> +                               struct bo_vm_reservation_context *ctx)
> +{
> +       struct amdgpu_bo *bo = mem->bo;
> +       struct kfd_bo_va_list *entry;
> +       unsigned int i;
> +       int ret;
> +
> +       ctx->reserved = false;
> +       ctx->n_vms = 0;
> +       ctx->vm_pd = NULL;
> +       ctx->sync = &mem->sync;
> +
> +       INIT_LIST_HEAD(&ctx->list);
> +       INIT_LIST_HEAD(&ctx->duplicates);
> +
> +       list_for_each_entry(entry, &mem->bo_va_list, bo_list) {
> +               if ((vm && vm != entry->bo_va->base.vm) ||
> +                       (entry->is_mapped != map_type
> +                       && map_type != BO_VM_ALL))
> +                       continue;
> +
> +               ctx->n_vms++;
> +       }
> +
> +       if (ctx->n_vms != 0) {
> +               ctx->vm_pd = kcalloc(ctx->n_vms, sizeof(*ctx->vm_pd),
> +                                    GFP_KERNEL);
> +               if (!ctx->vm_pd)
> +                       return -ENOMEM;
> +       }
> +
> +       ctx->kfd_bo.robj = bo;
> +       ctx->kfd_bo.priority = 0;
> +       ctx->kfd_bo.tv.bo = &bo->tbo;
> +       ctx->kfd_bo.tv.shared = true;
> +       ctx->kfd_bo.user_pages = NULL;
> +       list_add(&ctx->kfd_bo.tv.head, &ctx->list);
> +
> +       i = 0;
> +       list_for_each_entry(entry, &mem->bo_va_list, bo_list) {
> +               if ((vm && vm != entry->bo_va->base.vm) ||
> +                       (entry->is_mapped != map_type
> +                       && map_type != BO_VM_ALL))
> +                       continue;
> +
> +               amdgpu_vm_get_pd_bo(entry->bo_va->base.vm, &ctx->list,
> +                               &ctx->vm_pd[i]);
> +               i++;
> +       }
> +
> +       ret = ttm_eu_reserve_buffers(&ctx->ticket, &ctx->list,
> +                                    false, &ctx->duplicates);
> +       if (!ret)
> +               ctx->reserved = true;
> +       else
> +               pr_err("Failed to reserve buffers in ttm.\n");
> +
> +       if (ret) {
> +               kfree(ctx->vm_pd);
> +               ctx->vm_pd = NULL;
> +       }
> +
> +       return ret;
> +}
> +
> +/**
> + * unreserve_bo_and_vms - Unreserve BO and VMs from a reservation context
> + * @ctx: Reservation context to unreserve
> + * @wait: Optionally wait for a sync object representing pending VM updates
> + * @intr: Whether the wait is interruptible
> + *
> + * Also frees any resources allocated in
> + * reserve_bo_and_(cond_)vm(s). Returns the status from
> + * amdgpu_sync_wait.
> + */
> +static int unreserve_bo_and_vms(struct bo_vm_reservation_context *ctx,
> +                                bool wait, bool intr)
> +{
> +       int ret = 0;
> +
> +       if (wait)
> +               ret = amdgpu_sync_wait(ctx->sync, intr);
> +
> +       if (ctx->reserved)
> +               ttm_eu_backoff_reservation(&ctx->ticket, &ctx->list);
> +       kfree(ctx->vm_pd);
> +
> +       ctx->sync = NULL;
> +
> +       ctx->reserved = false;
> +       ctx->vm_pd = NULL;
> +
> +       return ret;
> +}
> +
> +static int unmap_bo_from_gpuvm(struct amdgpu_device *adev,
> +                               struct kfd_bo_va_list *entry,
> +                               struct amdgpu_sync *sync)
> +{
> +       struct amdgpu_bo_va *bo_va = entry->bo_va;
> +       struct amdgpu_vm *vm = bo_va->base.vm;
> +       struct amdkfd_vm *kvm = container_of(vm, struct amdkfd_vm, base);
> +       struct amdgpu_bo *pd = vm->root.base.bo;
> +
> +       /* Remove eviction fence from PD (and thereby from PTs too as
> +        * they share the resv. object). Otherwise during PT update
> +        * job (see amdgpu_vm_bo_update_mapping), eviction fence would
> +        * get added to job->sync object and job execution would
> +        * trigger the eviction fence.
> +        */
> +       amdgpu_amdkfd_remove_eviction_fence(pd,
> +                                           kvm->process_info->eviction_fence,
> +                                           NULL, NULL);
> +       amdgpu_vm_bo_unmap(adev, bo_va, entry->va);
> +
> +       amdgpu_vm_clear_freed(adev, vm, &bo_va->last_pt_update);
> +
> +       /* Add the eviction fence back */
> +       amdgpu_bo_fence(pd, &kvm->process_info->eviction_fence->base, true);
> +
> +       sync_vm_fence(adev, sync, bo_va->last_pt_update);
> +
> +       return 0;
> +}
> +
> +static int update_gpuvm_pte(struct amdgpu_device *adev,
> +               struct kfd_bo_va_list *entry,
> +               struct amdgpu_sync *sync)
> +{
> +       int ret;
> +       struct amdgpu_vm *vm;
> +       struct amdgpu_bo_va *bo_va;
> +       struct amdgpu_bo *bo;
> +
> +       bo_va = entry->bo_va;
> +       vm = bo_va->base.vm;
> +       bo = bo_va->base.bo;
> +
> +       /* Update the page tables  */
> +       ret = amdgpu_vm_bo_update(adev, bo_va, false);
> +       if (ret) {
> +               pr_err("amdgpu_vm_bo_update failed\n");
> +               return ret;
> +       }
> +
> +       return sync_vm_fence(adev, sync, bo_va->last_pt_update);
> +}
> +
> +static int map_bo_to_gpuvm(struct amdgpu_device *adev,
> +               struct kfd_bo_va_list *entry, struct amdgpu_sync *sync)
> +{
> +       int ret;
> +
> +       /* Set virtual address for the allocation */
> +       ret = amdgpu_vm_bo_map(adev, entry->bo_va, entry->va, 0,
> +                              amdgpu_bo_size(entry->bo_va->base.bo),
> +                              entry->pte_flags);
> +       if (ret) {
> +               pr_err("Failed to map VA 0x%llx in vm. ret %d\n",
> +                               entry->va, ret);
> +               return ret;
> +       }
> +
> +       ret = update_gpuvm_pte(adev, entry, sync);
> +       if (ret) {
> +               pr_err("update_gpuvm_pte() failed\n");
> +               goto update_gpuvm_pte_failed;
> +       }
> +
> +       return 0;
> +
> +update_gpuvm_pte_failed:
> +       unmap_bo_from_gpuvm(adev, entry, sync);
> +       return ret;
> +}
> +
> +static int process_validate_vms(struct amdkfd_process_info *process_info)
> +{
> +       struct amdkfd_vm *peer_vm;
> +       int ret;
> +
> +       list_for_each_entry(peer_vm, &process_info->vm_list_head,
> +                           vm_list_node) {
> +               ret = vm_validate_pt_pd_bos(peer_vm);
> +               if (ret)
> +                       return ret;
> +       }
> +
> +       return 0;
> +}
> +
> +static int process_update_pds(struct amdkfd_process_info *process_info,
> +                             struct amdgpu_sync *sync)
> +{
> +       struct amdkfd_vm *peer_vm;
> +       int ret;
> +
> +       list_for_each_entry(peer_vm, &process_info->vm_list_head,
> +                           vm_list_node) {
> +               ret = vm_update_pds(&peer_vm->base, sync);
> +               if (ret)
> +                       return ret;
> +       }
> +
> +       return 0;
> +}
> +
> +int amdgpu_amdkfd_gpuvm_create_process_vm(struct kgd_dev *kgd, void **vm,
> +                                         void **process_info,
> +                                         struct dma_fence **ef)
> +{
> +       int ret;
> +       struct amdkfd_vm *new_vm;
> +       struct amdkfd_process_info *info;
> +       struct amdgpu_device *adev = get_amdgpu_device(kgd);
> +
> +       new_vm = kzalloc(sizeof(*new_vm), GFP_KERNEL);
> +       if (!new_vm)
> +               return -ENOMEM;
> +
> +       /* Initialize the VM context, allocate the page directory and zero it */
> +       ret = amdgpu_vm_init(adev, &new_vm->base, AMDGPU_VM_CONTEXT_COMPUTE, 0);
> +       if (ret) {
> +               pr_err("Failed init vm ret %d\n", ret);
> +               goto vm_init_fail;
> +       }
> +       new_vm->adev = adev;
> +
> +       if (!*process_info) {
> +               info = kzalloc(sizeof(*info), GFP_KERNEL);
> +               if (!info) {
> +                       ret = -ENOMEM;
> +                       goto alloc_process_info_fail;
> +               }
> +
> +               mutex_init(&info->lock);
> +               INIT_LIST_HEAD(&info->vm_list_head);
> +               INIT_LIST_HEAD(&info->kfd_bo_list);
> +
> +               info->eviction_fence =
> +                       amdgpu_amdkfd_fence_create(dma_fence_context_alloc(1),
> +                                                  current->mm);
> +               if (!info->eviction_fence) {
> +                       pr_err("Failed to create eviction fence\n");
> +                       goto create_evict_fence_fail;
> +               }
> +
> +               *process_info = info;
> +               *ef = dma_fence_get(&info->eviction_fence->base);
> +       }
> +
> +       new_vm->process_info = *process_info;
> +
> +       mutex_lock(&new_vm->process_info->lock);
> +       list_add_tail(&new_vm->vm_list_node,
> +                       &(new_vm->process_info->vm_list_head));
> +       new_vm->process_info->n_vms++;
> +       mutex_unlock(&new_vm->process_info->lock);
> +
> +       *vm = (void *) new_vm;
> +
> +       pr_debug("Created process vm %p\n", *vm);
> +
> +       return ret;
> +
> +create_evict_fence_fail:
need to destroy mutex inside info

> +       kfree(info);
> +alloc_process_info_fail:
> +       amdgpu_vm_fini(adev, &new_vm->base);
> +vm_init_fail:
> +       kfree(new_vm);
> +       return ret;
> +
> +}
> +
> +void amdgpu_amdkfd_gpuvm_destroy_process_vm(struct kgd_dev *kgd, void *vm)
> +{
> +       struct amdgpu_device *adev = get_amdgpu_device(kgd);
> +       struct amdkfd_vm *kfd_vm = (struct amdkfd_vm *) vm;
> +       struct amdgpu_vm *avm = &kfd_vm->base;
> +       struct amdgpu_bo *pd;
> +       struct amdkfd_process_info *process_info;
> +
> +       if (WARN_ON(!kgd || !vm))
> +               return;
> +
> +       pr_debug("Destroying process vm %p\n", vm);
> +       /* Release eviction fence from PD */
> +       pd = avm->root.base.bo;
> +       amdgpu_bo_reserve(pd, false);
> +       amdgpu_bo_fence(pd, NULL, false);
> +       amdgpu_bo_unreserve(pd);
> +
> +       process_info = kfd_vm->process_info;
> +
> +       mutex_lock(&process_info->lock);
> +       process_info->n_vms--;
> +       list_del(&kfd_vm->vm_list_node);
> +       mutex_unlock(&process_info->lock);
> +
> +       /* Release per-process resources */
> +       if (!process_info->n_vms) {
> +               WARN_ON(!list_empty(&process_info->kfd_bo_list));
> +
> +               dma_fence_put(&process_info->eviction_fence->base);
I think we need to destroy the process_info mutex here before the free
> +               kfree(process_info);
> +       }
> +
> +       /* Release the VM context */
> +       amdgpu_vm_fini(adev, avm);
> +       kfree(vm);
> +}
> +
> +uint32_t amdgpu_amdkfd_gpuvm_get_process_page_dir(void *vm)
> +{
> +       struct amdkfd_vm *avm = (struct amdkfd_vm *)vm;
> +
> +       return avm->pd_phys_addr >> AMDGPU_GPU_PAGE_SHIFT;
> +}
> +
> +int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
> +               struct kgd_dev *kgd, uint64_t va, uint64_t size,
> +               void *vm, struct kgd_mem **mem,
> +               uint64_t *offset, uint32_t flags)
> +{
> +       struct amdgpu_device *adev = get_amdgpu_device(kgd);
> +       struct amdkfd_vm *kfd_vm = (struct amdkfd_vm *)vm;
> +       struct amdgpu_bo *bo;
> +       int byte_align;
> +       u32 alloc_domain;
> +       u64 alloc_flags;
> +       uint32_t mapping_flags;
> +       int ret;
> +
> +       /*
> +        * Check on which domain to allocate BO
> +        */
> +       if (flags & ALLOC_MEM_FLAGS_VRAM) {
> +               alloc_domain = AMDGPU_GEM_DOMAIN_VRAM;
> +               alloc_flags = AMDGPU_GEM_CREATE_VRAM_CLEARED;
> +               alloc_flags |= (flags & ALLOC_MEM_FLAGS_PUBLIC) ?
> +                       AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED :
> +                       AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
> +       } else if (flags & ALLOC_MEM_FLAGS_GTT) {
> +               alloc_domain = AMDGPU_GEM_DOMAIN_GTT;
> +               alloc_flags = 0;
> +       } else {
> +               return -EINVAL;
> +       }
> +
> +       *mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
> +       if (!*mem)
> +               return -ENOMEM;
> +       INIT_LIST_HEAD(&(*mem)->bo_va_list);
> +       mutex_init(&(*mem)->lock);
> +       (*mem)->aql_queue     = !!(flags & ALLOC_MEM_FLAGS_AQL_QUEUE_MEM);
> +
> +       /* Workaround for AQL queue wraparound bug. Map the same
> +        * memory twice. That means we only actually allocate half
> +        * the memory.
> +        */
> +       if ((*mem)->aql_queue)
> +               size = size >> 1;
> +
> +       /* Workaround for TLB bug on older VI chips */
> +       byte_align = (adev->family == AMDGPU_FAMILY_VI &&
> +                       adev->asic_type != CHIP_FIJI &&
> +                       adev->asic_type != CHIP_POLARIS10 &&
> +                       adev->asic_type != CHIP_POLARIS11) ?
> +                       VI_BO_SIZE_ALIGN : 1;
> +
> +       mapping_flags = AMDGPU_VM_PAGE_READABLE;
> +       if (flags & ALLOC_MEM_FLAGS_WRITABLE)
> +               mapping_flags |= AMDGPU_VM_PAGE_WRITEABLE;
> +       if (flags & ALLOC_MEM_FLAGS_EXECUTABLE)
> +               mapping_flags |= AMDGPU_VM_PAGE_EXECUTABLE;
> +       if (flags & ALLOC_MEM_FLAGS_COHERENT)
> +               mapping_flags |= AMDGPU_VM_MTYPE_UC;
> +       else
> +               mapping_flags |= AMDGPU_VM_MTYPE_NC;
> +       (*mem)->mapping_flags = mapping_flags;
> +
> +       amdgpu_sync_create(&(*mem)->sync);
> +
> +       ret = amdgpu_amdkfd_reserve_system_mem_limit(adev, size, alloc_domain);
> +       if (ret) {
> +               pr_debug("Insufficient system memory\n");
> +               goto err_bo_create;
I suggest to change that to "goto err_reserve_system_mem" and add that
label after err_bo_create label

> +       }
> +
> +       pr_debug("\tcreate BO VA 0x%llx size 0x%llx domain %s\n",
> +                       va, size, domain_string(alloc_domain));
> +
> +       ret = amdgpu_bo_create(adev, size, byte_align, false,
> +                               alloc_domain, alloc_flags, NULL, NULL, 0, &bo);
> +       if (ret) {
> +               pr_debug("Failed to create BO on domain %s. ret %d\n",
> +                               domain_string(alloc_domain), ret);
> +               unreserve_system_mem_limit(adev, size, alloc_domain);
Move the above line under "err_bo_create:" label

> +               goto err_bo_create;
> +       }
> +       bo->kfd_bo = *mem;
> +       (*mem)->bo = bo;
> +
> +       (*mem)->va = va;
> +       (*mem)->domain = alloc_domain;
> +       (*mem)->mapped_to_gpu_memory = 0;
> +       (*mem)->process_info = kfd_vm->process_info;
> +       add_kgd_mem_to_kfd_bo_list(*mem, kfd_vm->process_info);
> +
> +       if (offset)
> +               *offset = amdgpu_bo_mmap_offset(bo);
> +
> +       return 0;
> +
> +err_bo_create:
Need to destroy mutex before freeing mem

> +       kfree(*mem);
> +       return ret;
> +}
> +
> +int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(
> +               struct kgd_dev *kgd, struct kgd_mem *mem)
> +{
> +       struct amdkfd_process_info *process_info = mem->process_info;
> +       unsigned long bo_size = mem->bo->tbo.mem.size;
> +       struct kfd_bo_va_list *entry, *tmp;
> +       struct bo_vm_reservation_context ctx;
> +       struct ttm_validate_buffer *bo_list_entry;
> +       int ret;
> +
> +       mutex_lock(&mem->lock);
> +
> +       if (mem->mapped_to_gpu_memory > 0) {
> +               pr_debug("BO VA 0x%llx size 0x%lx is still mapped.\n",
> +                               mem->va, bo_size);
> +               mutex_unlock(&mem->lock);
> +               return -EBUSY;
> +       }
> +
> +       mutex_unlock(&mem->lock);
> +       /* lock is not needed after this, since mem is unused and will
> +        * be freed anyway
> +        */
> +
> +       /* Make sure restore workers don't access the BO any more */
> +       bo_list_entry = &mem->validate_list;
> +       mutex_lock(&process_info->lock);
> +       list_del(&bo_list_entry->head);
> +       mutex_unlock(&process_info->lock);
> +
> +       ret = reserve_bo_and_cond_vms(mem, NULL, BO_VM_ALL, &ctx);
> +       if (unlikely(ret))
> +               return ret;
> +
> +       /* The eviction fence should be removed by the last unmap.
> +        * TODO: Log an error condition if the bo still has the eviction fence
> +        * attached
> +        */
> +       amdgpu_amdkfd_remove_eviction_fence(mem->bo,
> +                                       process_info->eviction_fence,
> +                                       NULL, NULL);
> +       pr_debug("Release VA 0x%llx - 0x%llx\n", mem->va,
> +               mem->va + bo_size * (1 + mem->aql_queue));
> +
> +       /* Remove from VM internal data structures */
> +       list_for_each_entry_safe(entry, tmp, &mem->bo_va_list, bo_list)
> +               remove_bo_from_vm((struct amdgpu_device *)entry->kgd_dev,
> +                               entry, bo_size);
> +
> +       ret = unreserve_bo_and_vms(&ctx, false, false);
> +
> +       /* Free the sync object */
> +       amdgpu_sync_free(&mem->sync);
> +
> +       /* Free the BO*/
> +       amdgpu_bo_unref(&mem->bo);
Destroy the mem mutex before freeing the structure itself

> +       kfree(mem);
> +
> +       return ret;
> +}
> +
> +int amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
> +               struct kgd_dev *kgd, struct kgd_mem *mem, void *vm)
> +{
> +       struct amdgpu_device *adev = get_amdgpu_device(kgd);
> +       struct amdkfd_vm *kfd_vm = (struct amdkfd_vm *)vm;
> +       int ret;
> +       struct amdgpu_bo *bo;
> +       uint32_t domain;
> +       struct kfd_bo_va_list *entry;
> +       struct bo_vm_reservation_context ctx;
> +       struct kfd_bo_va_list *bo_va_entry = NULL;
> +       struct kfd_bo_va_list *bo_va_entry_aql = NULL;
> +       unsigned long bo_size;
> +
> +       /* Make sure restore is not running concurrently.
> +        */
> +       mutex_lock(&mem->process_info->lock);
> +
> +       mutex_lock(&mem->lock);
> +
> +       bo = mem->bo;
> +
> +       if (!bo) {
> +               pr_err("Invalid BO when mapping memory to GPU\n");
> +               ret = -EINVAL;
> +               goto out;
> +       }
> +
> +       domain = mem->domain;
> +       bo_size = bo->tbo.mem.size;
> +
> +       pr_debug("Map VA 0x%llx - 0x%llx to vm %p domain %s\n",
> +                       mem->va,
> +                       mem->va + bo_size * (1 + mem->aql_queue),
> +                       vm, domain_string(domain));
> +
> +       ret = reserve_bo_and_vm(mem, vm, &ctx);
> +       if (unlikely(ret))
> +               goto out;
> +
> +       if (check_if_add_bo_to_vm((struct amdgpu_vm *)vm, mem)) {
> +               ret = add_bo_to_vm(adev, mem, (struct amdgpu_vm *)vm, false,
> +                               &bo_va_entry);
> +               if (ret)
> +                       goto add_bo_to_vm_failed;
> +               if (mem->aql_queue) {
> +                       ret = add_bo_to_vm(adev, mem, (struct amdgpu_vm *)vm,
> +                                       true, &bo_va_entry_aql);
> +                       if (ret)
> +                               goto add_bo_to_vm_failed_aql;
> +               }
> +       } else {
> +               ret = vm_validate_pt_pd_bos((struct amdkfd_vm *)vm);
> +               if (unlikely(ret))
> +                       goto add_bo_to_vm_failed;
> +       }
> +
> +       if (mem->mapped_to_gpu_memory == 0) {
> +               /* Validate BO only once. The eviction fence gets added to BO
> +                * the first time it is mapped. Validate will wait for all
> +                * background evictions to complete.
> +                */
> +               ret = amdgpu_amdkfd_bo_validate(bo, domain, true);
> +               if (ret) {
> +                       pr_debug("Validate failed\n");
> +                       goto map_bo_to_gpuvm_failed;
> +               }
> +       }
> +
> +       list_for_each_entry(entry, &mem->bo_va_list, bo_list) {
> +               if (entry->bo_va->base.vm == vm && !entry->is_mapped) {
> +                       pr_debug("\t map VA 0x%llx - 0x%llx in entry %p\n",
> +                                       entry->va, entry->va + bo_size,
> +                                       entry);
> +
> +                       ret = map_bo_to_gpuvm(adev, entry, ctx.sync);
> +                       if (ret) {
> +                               pr_err("Failed to map radeon bo to gpuvm\n");
> +                               goto map_bo_to_gpuvm_failed;
> +                       }
> +
> +                       ret = vm_update_pds(vm, ctx.sync);
> +                       if (ret) {
> +                               pr_err("Failed to update page directories\n");
> +                               goto map_bo_to_gpuvm_failed;
> +                       }
> +
> +                       entry->is_mapped = true;
> +                       mem->mapped_to_gpu_memory++;
> +                       pr_debug("\t INC mapping count %d\n",
> +                                       mem->mapped_to_gpu_memory);
> +               }
> +       }
> +
> +       if (!amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) && !bo->pin_count)
> +               amdgpu_bo_fence(bo,
> +                               &kfd_vm->process_info->eviction_fence->base,
> +                               true);
> +       ret = unreserve_bo_and_vms(&ctx, false, false);
> +
> +       goto out;
> +
> +map_bo_to_gpuvm_failed:
> +       if (bo_va_entry_aql)
> +               remove_bo_from_vm(adev, bo_va_entry_aql, bo_size);
> +add_bo_to_vm_failed_aql:
> +       if (bo_va_entry)
> +               remove_bo_from_vm(adev, bo_va_entry, bo_size);
> +add_bo_to_vm_failed:
> +       unreserve_bo_and_vms(&ctx, false, false);
> +
> +out:
> +       mutex_unlock(&mem->process_info->lock);
> +       mutex_unlock(&mem->lock);
> +       return ret;
> +}
> +
> +int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
> +               struct kgd_dev *kgd, struct kgd_mem *mem, void *vm)
> +{
> +       struct amdgpu_device *adev = get_amdgpu_device(kgd);
> +       struct amdkfd_process_info *process_info =
> +               ((struct amdkfd_vm *)vm)->process_info;
> +       unsigned long bo_size = mem->bo->tbo.mem.size;
> +       struct kfd_bo_va_list *entry;
> +       struct bo_vm_reservation_context ctx;
> +       int ret;
> +
> +       mutex_lock(&mem->lock);
> +
> +       ret = reserve_bo_and_cond_vms(mem, vm, BO_VM_MAPPED, &ctx);
> +       if (unlikely(ret))
> +               goto out;
> +       /* If no VMs were reserved, it means the BO wasn't actually mapped */
> +       if (ctx.n_vms == 0) {
> +               ret = -EINVAL;
> +               goto unreserve_out;
> +       }
> +
> +       ret = vm_validate_pt_pd_bos((struct amdkfd_vm *)vm);
> +       if (unlikely(ret))
> +               goto unreserve_out;
> +
> +       pr_debug("Unmap VA 0x%llx - 0x%llx from vm %p\n",
> +               mem->va,
> +               mem->va + bo_size * (1 + mem->aql_queue),
> +               vm);
> +
> +       list_for_each_entry(entry, &mem->bo_va_list, bo_list) {
> +               if (entry->bo_va->base.vm == vm && entry->is_mapped) {
> +                       pr_debug("\t unmap VA 0x%llx - 0x%llx from entry %p\n",
> +                                       entry->va,
> +                                       entry->va + bo_size,
> +                                       entry);
> +
> +                       ret = unmap_bo_from_gpuvm(adev, entry, ctx.sync);
> +                       if (ret == 0) {
> +                               entry->is_mapped = false;
> +                       } else {
> +                               pr_err("failed to unmap VA 0x%llx\n",
> +                                               mem->va);
> +                               goto unreserve_out;
> +                       }
> +
> +                       mem->mapped_to_gpu_memory--;
> +                       pr_debug("\t DEC mapping count %d\n",
> +                                       mem->mapped_to_gpu_memory);
> +               }
> +       }
> +
> +       /* If BO is unmapped from all VMs, unfence it. It can be evicted if
> +        * required.
> +        */
> +       if (mem->mapped_to_gpu_memory == 0 &&
> +           !amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm) && !mem->bo->pin_count)
> +               amdgpu_amdkfd_remove_eviction_fence(mem->bo,
> +                                               process_info->eviction_fence,
> +                                                   NULL, NULL);
> +
> +unreserve_out:
> +       unreserve_bo_and_vms(&ctx, false, false);
> +out:
> +       mutex_unlock(&mem->lock);
> +       return ret;
> +}
> +
> +int amdgpu_amdkfd_gpuvm_sync_memory(
> +               struct kgd_dev *kgd, struct kgd_mem *mem, bool intr)
> +{
> +       struct amdgpu_sync sync;
> +       int ret;
> +
> +       amdgpu_sync_create(&sync);
> +
> +       mutex_lock(&mem->lock);
> +       amdgpu_sync_clone(&mem->sync, &sync);
> +       mutex_unlock(&mem->lock);
> +
> +       ret = amdgpu_sync_wait(&sync, intr);
> +       amdgpu_sync_free(&sync);
> +       return ret;
> +}
> +
> +int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct kgd_dev *kgd,
> +               struct kgd_mem *mem, void **kptr, uint64_t *size)
> +{
> +       int ret;
> +       struct amdgpu_bo *bo = mem->bo;
> +
> +       if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
> +               pr_err("userptr can't be mapped to kernel\n");
> +               return -EINVAL;
> +       }
> +
> +       /* delete kgd_mem from kfd_bo_list to avoid re-validating
> +        * this BO in BO's restoring after eviction.
> +        */
> +       mutex_lock(&mem->process_info->lock);
> +
> +       ret = amdgpu_bo_reserve(bo, true);
> +       if (ret) {
> +               pr_err("Failed to reserve bo. ret %d\n", ret);
> +               goto bo_reserve_failed;
> +       }
> +
> +       ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT, NULL);
> +       if (ret) {
> +               pr_err("Failed to pin bo. ret %d\n", ret);
> +               goto pin_failed;
> +       }
> +
> +       ret = amdgpu_bo_kmap(bo, kptr);
> +       if (ret) {
> +               pr_err("Failed to map bo to kernel. ret %d\n", ret);
> +               goto kmap_failed;
> +       }
> +
> +       amdgpu_amdkfd_remove_eviction_fence(
> +               bo, mem->process_info->eviction_fence, NULL, NULL);
> +       list_del_init(&mem->validate_list.head);
> +
> +       if (size)
> +               *size = amdgpu_bo_size(bo);
> +
> +       amdgpu_bo_unreserve(bo);
> +
> +       mutex_unlock(&mem->process_info->lock);
> +       return 0;
> +
> +kmap_failed:
> +       amdgpu_bo_unpin(bo);
> +pin_failed:
> +       amdgpu_bo_unreserve(bo);
> +bo_reserve_failed:
> +       mutex_unlock(&mem->process_info->lock);
> +
> +       return ret;
> +}
> +
> +/** amdgpu_amdkfd_gpuvm_restore_process_bos - Restore all BOs for the given
> + *   KFD process identified by process_info
> + *
> + * @process_info: amdkfd_process_info of the KFD process
> + *
> + * After memory eviction, restore thread calls this function. The function
> + * should be called when the Process is still valid. BO restore involves -
> + *
> + * 1.  Release old eviction fence and create new one
> + * 2.  Get two copies of PD BO list from all the VMs. Keep one copy as pd_list.
> + * 3   Use the second PD list and kfd_bo_list to create a list (ctx.list) of
> + *     BOs that need to be reserved.
> + * 4.  Reserve all the BOs
> + * 5.  Validate of PD and PT BOs.
> + * 6.  Validate all KFD BOs using kfd_bo_list and Map them and add new fence
> + * 7.  Add fence to all PD and PT BOs.
> + * 8.  Unreserve all BOs
> + */
> +int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence **ef)
> +{
> +       struct amdgpu_bo_list_entry *pd_bo_list;
> +       struct amdkfd_process_info *process_info = info;
> +       struct amdkfd_vm *peer_vm;
> +       struct kgd_mem *mem;
> +       struct bo_vm_reservation_context ctx;
> +       struct amdgpu_amdkfd_fence *new_fence;
> +       int ret = 0, i;
> +       struct list_head duplicate_save;
> +       struct amdgpu_sync sync_obj;
> +
> +       INIT_LIST_HEAD(&duplicate_save);
> +       INIT_LIST_HEAD(&ctx.list);
> +       INIT_LIST_HEAD(&ctx.duplicates);
> +
> +       pd_bo_list = kcalloc(process_info->n_vms,
> +                            sizeof(struct amdgpu_bo_list_entry),
> +                            GFP_KERNEL);
> +       if (!pd_bo_list)
> +               return -ENOMEM;
> +
> +       i = 0;
> +       mutex_lock(&process_info->lock);
> +       list_for_each_entry(peer_vm, &process_info->vm_list_head,
> +                       vm_list_node)
> +               amdgpu_vm_get_pd_bo(&peer_vm->base, &ctx.list,
> +                                   &pd_bo_list[i++]);
> +
> +       /* Reserve all BOs and page tables/directory. Add all BOs from
> +        * kfd_bo_list to ctx.list
> +        */
> +       list_for_each_entry(mem, &process_info->kfd_bo_list,
> +                           validate_list.head) {
> +
> +               list_add_tail(&mem->resv_list.head, &ctx.list);
> +               mem->resv_list.bo = mem->validate_list.bo;
> +               mem->resv_list.shared = mem->validate_list.shared;
> +       }
> +
> +       ret = ttm_eu_reserve_buffers(&ctx.ticket, &ctx.list,
> +                                    false, &duplicate_save);
> +       if (ret) {
> +               pr_debug("Memory eviction: TTM Reserve Failed. Try again\n");
> +               goto ttm_reserve_fail;
> +       }
> +
> +       amdgpu_sync_create(&sync_obj);
> +
> +       /* Validate PDs and PTs */
> +       ret = process_validate_vms(process_info);
> +       if (ret)
> +               goto validate_map_fail;
> +
> +       /* Wait for PD/PTs validate to finish */
> +       /* FIXME: I think this isn't needed */
> +       list_for_each_entry(peer_vm, &process_info->vm_list_head,
> +                           vm_list_node) {
> +               struct amdgpu_bo *bo = peer_vm->base.root.base.bo;
> +
> +               ttm_bo_wait(&bo->tbo, false, false);
> +       }
> +
> +       /* Validate BOs and map them to GPUVM (update VM page tables). */
> +       list_for_each_entry(mem, &process_info->kfd_bo_list,
> +                           validate_list.head) {
> +
> +               struct amdgpu_bo *bo = mem->bo;
> +               uint32_t domain = mem->domain;
> +               struct kfd_bo_va_list *bo_va_entry;
> +
> +               ret = amdgpu_amdkfd_bo_validate(bo, domain, false);
> +               if (ret) {
> +                       pr_debug("Memory eviction: Validate BOs failed. Try again\n");
> +                       goto validate_map_fail;
> +               }
> +
> +               list_for_each_entry(bo_va_entry, &mem->bo_va_list,
> +                                   bo_list) {
> +                       ret = update_gpuvm_pte((struct amdgpu_device *)
> +                                             bo_va_entry->kgd_dev,
> +                                             bo_va_entry,
> +                                             &sync_obj);
> +                       if (ret) {
> +                               pr_debug("Memory eviction: update PTE failed. Try again\n");
> +                               goto validate_map_fail;
> +                       }
> +               }
> +       }
> +
> +       /* Update page directories */
> +       ret = process_update_pds(process_info, &sync_obj);
> +       if (ret) {
> +               pr_debug("Memory eviction: update PDs failed. Try again\n");
> +               goto validate_map_fail;
> +       }
> +
> +       amdgpu_sync_wait(&sync_obj, false);
> +
> +       /* Release old eviction fence and create new one, because fence only
> +        * goes from unsignaled to signaled, fence cannot be reused.
> +        * Use context and mm from the old fence.
> +        */
> +       new_fence = amdgpu_amdkfd_fence_create(
> +                               process_info->eviction_fence->base.context,
> +                               process_info->eviction_fence->mm);
> +       if (!new_fence) {
> +               pr_err("Failed to create eviction fence\n");
> +               ret = -ENOMEM;
> +               goto validate_map_fail;
> +       }
> +       dma_fence_put(&process_info->eviction_fence->base);
> +       process_info->eviction_fence = new_fence;
> +       *ef = dma_fence_get(&new_fence->base);
> +
> +       /* Wait for validate to finish and attach new eviction fence */
> +       list_for_each_entry(mem, &process_info->kfd_bo_list,
> +               validate_list.head)
> +               ttm_bo_wait(&mem->bo->tbo, false, false);
> +       list_for_each_entry(mem, &process_info->kfd_bo_list,
> +               validate_list.head)
> +               amdgpu_bo_fence(mem->bo,
> +                       &process_info->eviction_fence->base, true);
> +
> +       /* Attach eviction fence to PD / PT BOs */
> +       list_for_each_entry(peer_vm, &process_info->vm_list_head,
> +                           vm_list_node) {
> +               struct amdgpu_bo *bo = peer_vm->base.root.base.bo;
> +
> +               amdgpu_bo_fence(bo, &process_info->eviction_fence->base, true);
> +       }
> +
> +validate_map_fail:
> +       ttm_eu_backoff_reservation(&ctx.ticket, &ctx.list);
> +       amdgpu_sync_free(&sync_obj);
> +ttm_reserve_fail:
> +       mutex_unlock(&process_info->lock);
> +       kfree(pd_bo_list);
> +       return ret;
> +}
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
> index 5c4c3e0..f608ecf 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
> @@ -36,6 +36,7 @@
>  #include <drm/drm_cache.h>
>  #include "amdgpu.h"
>  #include "amdgpu_trace.h"
> +#include "amdgpu_amdkfd.h"
>
>  static bool amdgpu_need_backup(struct amdgpu_device *adev)
>  {
> @@ -54,6 +55,9 @@ static void amdgpu_ttm_bo_destroy(struct ttm_buffer_object *tbo)
>         struct amdgpu_device *adev = amdgpu_ttm_adev(tbo->bdev);
>         struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
>
> +       if (bo->kfd_bo)
> +               amdgpu_amdkfd_unreserve_system_memory_limit(bo);
> +
>         amdgpu_bo_kunmap(bo);
>
>         drm_gem_object_release(&bo->gem_base);
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.h
> index 33615e2..ba5330a 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.h
> @@ -92,6 +92,8 @@ struct amdgpu_bo {
>                 struct list_head        mn_list;
>                 struct list_head        shadow_list;
>         };
> +
> +       struct kgd_mem                  *kfd_bo;
>  };
>
>  static inline struct amdgpu_bo *ttm_to_amdgpu_bo(struct ttm_buffer_object *tbo)
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
> index c3f33d3..76ee968 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
> @@ -261,6 +261,13 @@ static int amdgpu_verify_access(struct ttm_buffer_object *bo, struct file *filp)
>  {
>         struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
>
> +       /*
> +        * Don't verify access for KFD BOs. They don't have a GEM
> +        * object associated with them.
> +        */
> +       if (abo->kfd_bo)
> +               return 0;
> +
>         if (amdgpu_ttm_tt_get_usermm(bo->ttm))
>                 return -EPERM;
>         return drm_vma_node_verify_access(&abo->gem_base.vma_node,
> diff --git a/drivers/gpu/drm/amd/include/kgd_kfd_interface.h b/drivers/gpu/drm/amd/include/kgd_kfd_interface.h
> index 36c706a..5984fec 100644
> --- a/drivers/gpu/drm/amd/include/kgd_kfd_interface.h
> +++ b/drivers/gpu/drm/amd/include/kgd_kfd_interface.h
> @@ -127,6 +127,25 @@ struct tile_config {
>         uint32_t num_ranks;
>  };
>
> +
> +/*
> + * Allocation flag domains
> + */
> +#define ALLOC_MEM_FLAGS_VRAM           (1 << 0)
> +#define ALLOC_MEM_FLAGS_GTT            (1 << 1)
> +#define ALLOC_MEM_FLAGS_USERPTR                (1 << 2) /* TODO */
> +#define ALLOC_MEM_FLAGS_DOORBELL       (1 << 3) /* TODO */
> +
> +/*
> + * Allocation flags attributes/access options.
> + */
> +#define ALLOC_MEM_FLAGS_WRITABLE       (1 << 31)
> +#define ALLOC_MEM_FLAGS_EXECUTABLE     (1 << 30)
> +#define ALLOC_MEM_FLAGS_PUBLIC         (1 << 29)
> +#define ALLOC_MEM_FLAGS_NO_SUBSTITUTE  (1 << 28) /* TODO */
> +#define ALLOC_MEM_FLAGS_AQL_QUEUE_MEM  (1 << 27)
> +#define ALLOC_MEM_FLAGS_COHERENT       (1 << 26) /* For GFXv9 or later */
> +
>  /**
>   * struct kfd2kgd_calls
>   *
> @@ -186,6 +205,41 @@ struct tile_config {
>   *
>   * @get_vram_usage: Returns current VRAM usage
>   *
> + * @create_process_vm: Create a VM address space for a given process and GPU
> + *
> + * @destroy_process_vm: Destroy a VM
> + *
> + * @get_process_page_dir: Get physical address of a VM page directory
> + *
> + * @set_vm_context_page_table_base: Program page table base for a VMID
> + *
> + * @alloc_memory_of_gpu: Allocate GPUVM memory
> + *
> + * @free_memory_of_gpu: Free GPUVM memory
> + *
> + * @map_memory_to_gpu: Map GPUVM memory into a specific VM address
> + * space. Allocates and updates page tables and page directories as
> + * needed. This function may return before all page table updates have
> + * completed. This allows multiple map operations (on multiple GPUs)
> + * to happen concurrently. Use sync_memory to synchronize with all
> + * pending updates.
> + *
> + * @unmap_memor_to_gpu: Unmap GPUVM memory from a specific VM address space
> + *
> + * @sync_memory: Wait for pending page table updates to complete
> + *
> + * @map_gtt_bo_to_kernel: Map a GTT BO for kernel access
> + * Pins the BO, maps it to kernel address space. Such BOs are never evicted.
> + * The kernel virtual address remains valid until the BO is freed.
> + *
> + * @restore_process_bos: Restore all BOs that belong to the
> + * process. This is intended for restoring memory mappings after a TTM
> + * eviction.
> + *
> + * @invalidate_tlbs: Invalidate TLBs for a specific PASID
> + *
> + * @invalidate_tlbs_vmid: Invalidate TLBs for a specific VMID
> + *
>   * This structure contains function pointers to services that the kgd driver
>   * provides to amdkfd driver.
>   *
> @@ -275,6 +329,29 @@ struct kfd2kgd_calls {
>         void (*get_cu_info)(struct kgd_dev *kgd,
>                         struct kfd_cu_info *cu_info);
>         uint64_t (*get_vram_usage)(struct kgd_dev *kgd);
> +
> +       int (*create_process_vm)(struct kgd_dev *kgd, void **vm,
> +                       void **process_info, struct dma_fence **ef);
> +       void (*destroy_process_vm)(struct kgd_dev *kgd, void *vm);
> +       uint32_t (*get_process_page_dir)(void *vm);
> +       void (*set_vm_context_page_table_base)(struct kgd_dev *kgd,
> +                       uint32_t vmid, uint32_t page_table_base);
> +       int (*alloc_memory_of_gpu)(struct kgd_dev *kgd, uint64_t va,
> +                       uint64_t size, void *vm,
> +                       struct kgd_mem **mem, uint64_t *offset,
> +                       uint32_t flags);
> +       int (*free_memory_of_gpu)(struct kgd_dev *kgd, struct kgd_mem *mem);
> +       int (*map_memory_to_gpu)(struct kgd_dev *kgd, struct kgd_mem *mem,
> +                       void *vm);
> +       int (*unmap_memory_to_gpu)(struct kgd_dev *kgd, struct kgd_mem *mem,
> +                       void *vm);
> +       int (*sync_memory)(struct kgd_dev *kgd, struct kgd_mem *mem, bool intr);
> +       int (*map_gtt_bo_to_kernel)(struct kgd_dev *kgd, struct kgd_mem *mem,
> +                       void **kptr, uint64_t *size);
> +       int (*restore_process_bos)(void *process_info, struct dma_fence **ef);
> +
> +       int (*invalidate_tlbs)(struct kgd_dev *kgd, uint16_t pasid);
> +       int (*invalidate_tlbs_vmid)(struct kgd_dev *kgd, uint16_t vmid);
>  };
>
>  /**
> --
> 2.7.4
>

Hi Felix,
I wrote some minor comments. If you don't object to them, I'll just
add them to the patch to save you the trouble of re-sending.


This patch is:
Acked-by: Oded Gabbay <oded.gabbay at gmail.com>


[Index of Archives]     [Linux USB Devel]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]

  Powered by Linux