diff options
Diffstat (limited to 'drivers/gpu/drm/amd/amdkfd/kfd_process.c')
| -rw-r--r-- | drivers/gpu/drm/amd/amdkfd/kfd_process.c | 85 |
1 files changed, 73 insertions, 12 deletions
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c index 51b1683ac5c1..4208e0f01064 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c @@ -1167,6 +1167,17 @@ static void kfd_process_free_notifier(struct mmu_notifier *mn) kfd_unref_process(container_of(mn, struct kfd_process, mmu_notifier)); } +static void kfd_process_notifier_release_internal(struct kfd_process *p) +{ + cancel_delayed_work_sync(&p->eviction_work); + cancel_delayed_work_sync(&p->restore_work); + + /* Indicate to other users that MM is no longer valid */ + p->mm = NULL; + + mmu_notifier_put(&p->mmu_notifier); +} + static void kfd_process_notifier_release(struct mmu_notifier *mn, struct mm_struct *mm) { @@ -1181,17 +1192,22 @@ static void kfd_process_notifier_release(struct mmu_notifier *mn, return; mutex_lock(&kfd_processes_mutex); + /* + * Do early return if table is empty. + * + * This could potentially happen if this function is called concurrently + * by mmu_notifier and by kfd_cleanup_pocesses. + * + */ + if (hash_empty(kfd_processes_table)) { + mutex_unlock(&kfd_processes_mutex); + return; + } hash_del_rcu(&p->kfd_processes); mutex_unlock(&kfd_processes_mutex); synchronize_srcu(&kfd_processes_srcu); - cancel_delayed_work_sync(&p->eviction_work); - cancel_delayed_work_sync(&p->restore_work); - - /* Indicate to other users that MM is no longer valid */ - p->mm = NULL; - - mmu_notifier_put(&p->mmu_notifier); + kfd_process_notifier_release_internal(p); } static const struct mmu_notifier_ops kfd_process_mmu_notifier_ops = { @@ -1200,6 +1216,43 @@ static const struct mmu_notifier_ops kfd_process_mmu_notifier_ops = { .free_notifier = kfd_process_free_notifier, }; +/* + * This code handles the case when driver is being unloaded before all + * mm_struct are released. We need to safely free the kfd_process and + * avoid race conditions with mmu_notifier that might try to free them. + * + */ +void kfd_cleanup_processes(void) +{ + struct kfd_process *p; + struct hlist_node *p_temp; + unsigned int temp; + HLIST_HEAD(cleanup_list); + + /* + * Move all remaining kfd_process from the process table to a + * temp list for processing. Once done, callback from mmu_notifier + * release will not see the kfd_process in the table and do early return, + * avoiding double free issues. + */ + mutex_lock(&kfd_processes_mutex); + hash_for_each_safe(kfd_processes_table, temp, p_temp, p, kfd_processes) { + hash_del_rcu(&p->kfd_processes); + synchronize_srcu(&kfd_processes_srcu); + hlist_add_head(&p->kfd_processes, &cleanup_list); + } + mutex_unlock(&kfd_processes_mutex); + + hlist_for_each_entry_safe(p, p_temp, &cleanup_list, kfd_processes) + kfd_process_notifier_release_internal(p); + + /* + * Ensures that all outstanding free_notifier get called, triggering + * the release of the kfd_process struct. + */ + mmu_notifier_synchronize(); +} + static int kfd_process_init_cwsr_apu(struct kfd_process *p, struct file *filep) { unsigned long offset; @@ -1330,7 +1383,7 @@ bool kfd_process_xnack_mode(struct kfd_process *p, bool supported) * per-process XNACK mode selection. But let the dev->noretry * setting still influence the default XNACK mode. */ - if (supported && KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 2)) + if (supported && KFD_SUPPORT_XNACK_PER_PROCESS(dev)) continue; /* GFXv10 and later GPUs do not support shader preemption @@ -1563,6 +1616,8 @@ err_free_pdd: int kfd_process_device_init_vm(struct kfd_process_device *pdd, struct file *drm_file) { + struct amdgpu_fpriv *drv_priv; + struct amdgpu_vm *avm; struct kfd_process *p; struct kfd_dev *dev; int ret; @@ -1573,10 +1628,15 @@ int kfd_process_device_init_vm(struct kfd_process_device *pdd, if (pdd->drm_priv) return -EBUSY; + ret = amdgpu_file_to_fpriv(drm_file, &drv_priv); + if (ret) + return ret; + avm = &drv_priv->vm; + p = pdd->process; dev = pdd->dev; - ret = amdgpu_amdkfd_gpuvm_acquire_process_vm(dev->adev, drm_file, + ret = amdgpu_amdkfd_gpuvm_acquire_process_vm(dev->adev, avm, &p->kgd_process_info, &p->ef); if (ret) { @@ -1593,7 +1653,7 @@ int kfd_process_device_init_vm(struct kfd_process_device *pdd, if (ret) goto err_init_cwsr; - ret = amdgpu_amdkfd_gpuvm_set_vm_pasid(dev->adev, drm_file, p->pasid); + ret = amdgpu_amdkfd_gpuvm_set_vm_pasid(dev->adev, avm, p->pasid); if (ret) goto err_set_pasid; @@ -1607,6 +1667,7 @@ err_init_cwsr: kfd_process_device_destroy_ib_mem(pdd); err_reserve_ib_mem: pdd->drm_priv = NULL; + amdgpu_amdkfd_gpuvm_destroy_cb(dev->adev, avm); return ret; } @@ -1978,8 +2039,8 @@ int kfd_reserved_mem_mmap(struct kfd_dev *dev, struct kfd_process *process, return -ENOMEM; } - vma->vm_flags |= VM_IO | VM_DONTCOPY | VM_DONTEXPAND - | VM_NORESERVE | VM_DONTDUMP | VM_PFNMAP; + vm_flags_set(vma, VM_IO | VM_DONTCOPY | VM_DONTEXPAND + | VM_NORESERVE | VM_DONTDUMP | VM_PFNMAP); /* Mapping pages to user process */ return remap_pfn_range(vma, vma->vm_start, PFN_DOWN(__pa(qpd->cwsr_kaddr)), |