[PATCH 2/2] drm/amdgpu: improve VM state machine documentation
Christian König
christian.koenig at amd.com
Sat Sep 1 18:04:54 UTC 2018
Am 01.09.2018 um 18:45 schrieb Kuehling, Felix:
>>>> @@ -1746,9 +1805,9 @@ int amdgpu_vm_bo_update(struct amdgpu_device *adev,
>>>> uint32_t mem_type = bo->tbo.mem.mem_type;
>>>>
>>>> if (!(bo->preferred_domains & amdgpu_mem_type_to_domain(mem_type)))
>>>> - list_add_tail(&bo_va->base.vm_status, &vm->evicted);
>>>> + amdgpu_vm_bo_evicted(&bo_va->base);
>>>> else
>>>> - list_add(&bo_va->base.vm_status, &vm->idle);
>>>> + amdgpu_vm_bo_idle(&bo_va->base);
>>> There is a small change in behaviour here for clearing
>>> bo_va->base.moved. Not sure if it matters.
>> Using list_add is minimal more efficient.
> What I meant was that the original code would not always set bo_va->base.moved = false (it's done a few lines above this under some conditions, but not always). The new code will always set bo_va->base.moved = false inside amdgpu_vm_bo_idle.
Good point as well, but that is indeed correct.
Regards,
Christian.
>
> Regards,
> Felix
>
> ________________________________________
> From: Christian König <ckoenig.leichtzumerken at gmail.com>
> Sent: Saturday, September 1, 2018 4:16:17 AM
> To: Kuehling, Felix; amd-gfx at lists.freedesktop.org
> Subject: Re: [PATCH 2/2] drm/amdgpu: improve VM state machine documentation
>
> Am 01.09.2018 um 01:51 schrieb Felix Kuehling:
>> Thanks for this. A few comments and a question inline.
>>
>> On 2018-08-31 09:27 AM, Christian König wrote:
>>> Since we have a lot of FAQ on the VM state machine try to improve the
>>> documentation by adding functions for each state move.
>>>
>>> Signed-off-by: Christian König <christian.koenig at amd.com>
>>> ---
>>> drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 107 ++++++++++++++++++++++++---------
>>> 1 file changed, 79 insertions(+), 28 deletions(-)
>>>
>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>>> index a9275a99d793..40c22635fefd 100644
>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>>> @@ -204,6 +204,69 @@ static unsigned amdgpu_vm_bo_size(struct amdgpu_device *adev, unsigned level)
>>> return AMDGPU_GPU_PAGE_ALIGN(amdgpu_vm_num_entries(adev, level) * 8);
>>> }
>>>
>>> +/**
>>> + * amdgpu_vm_bo_evicted - vm_bo is evicted
>>> + *
>>> + * @vm_bo: vm_bo which is evicted
>>> + *
>>> + * State for PDs/PTs and per VM BOs which are not at the location they should
>>> + * be.
>>> + */
>>> +static void amdgpu_vm_bo_evicted(struct amdgpu_vm_bo_base *vm_bo)
>>> +{
>>> + struct amdgpu_vm *vm = vm_bo->vm;
>>> + struct amdgpu_bo *bo = vm_bo->bo;
>>> +
>>> + vm_bo->moved = true;
>>> + if (bo->tbo.type == ttm_bo_type_kernel)
>>> + list_move(&vm_bo->vm_status, &vm->evicted);
>>> + else
>>> + list_move_tail(&vm_bo->vm_status, &vm->evicted);
>>> +}
>>> +
>>> +/**
>>> + * amdgpu_vm_bo_relocated - vm_bo is reloacted
>>> + *
>>> + * @vm_bo: vm_bo which is relocated
>>> + *
>>> + * State for PDs/PTs which needs to update their parent PD.
>>> + */
>>> +static void amdgpu_vm_bo_relocated(struct amdgpu_vm_bo_base *vm_bo)
>>> +{
>>> + list_move(&vm_bo->vm_status, &vm_bo->vm->relocated);
>>> +}
>>> +
>>> +/**
>>> + * amdgpu_vm_bo_moved - vm_bo is moved
>>> + *
>>> + * @vm_bo: vm_bo which is moved
>>> + *
>>> + * State for per VM and normal BOs which are moved, but that change is not yet
>>> + * reflected in the page tables.
>> I have a question here. Why does amdgpu_cs_vm_handling call
>> amdgpu_vm_bo_update manually for its BO list entries? Wouldn't it be
>> enough to just call amdgpu_vm_handle_moved?
> No, it is still possible that the mapping of the BO was never updated
> because the VM page tables where evicted.
>
> We still need to make sure that all BOs of the current submission are
> updated correctly.
>
>>> + */
>>> +static void amdgpu_vm_bo_moved(struct amdgpu_vm_bo_base *vm_bo)
>>> +{
>>> + struct amdgpu_vm *vm = vm_bo->vm;
>>> +
>>> + spin_lock(&vm->moved_lock);
>>> + list_move(&vm_bo->vm_status, &vm->moved);
>>> + spin_unlock(&vm->moved_lock);
>> If vm->moved_lock protects the moved list, do we also need to take it
>> whenever something is moved from that list?
> Correct, yes.
>
>> That could potentially be
>> any list_move operation that uses vm_bo->vm_status. I found one case
>> below where that may not be handled correctly.
>>
>>> +}
>>> +
>>> +/**
>>> + * amdgpu_vm_bo_idle - vm_bo is idle
>>> + *
>>> + * @vm_bo: vm_bo which is now idle
>>> + *
>>> + * State for PDs/PTs and per VM BOs which have gone through the state machine
>>> + * and are now idle.
>>> + */
>>> +static void amdgpu_vm_bo_idle(struct amdgpu_vm_bo_base *vm_bo)
>>> +{
>>> + list_move(&vm_bo->vm_status, &vm_bo->vm->idle);
>>> + vm_bo->moved = false;
>>> +}
>>> +
>>> /**
>>> * amdgpu_vm_bo_base_init - Adds bo to the list of bos associated with the vm
>>> *
>>> @@ -232,9 +295,9 @@ static void amdgpu_vm_bo_base_init(struct amdgpu_vm_bo_base *base,
>>>
>>> vm->bulk_moveable = false;
>>> if (bo->tbo.type == ttm_bo_type_kernel)
>>> - list_move(&base->vm_status, &vm->relocated);
>>> + amdgpu_vm_bo_relocated(base);
>>> else
>>> - list_move(&base->vm_status, &vm->idle);
>>> + amdgpu_vm_bo_idle(base);
>>>
>>> if (bo->preferred_domains &
>>> amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type))
>>> @@ -245,8 +308,7 @@ static void amdgpu_vm_bo_base_init(struct amdgpu_vm_bo_base *base,
>>> * is currently evicted. add the bo to the evicted list to make sure it
>>> * is validated on next vm use to avoid fault.
>>> * */
>>> - list_move_tail(&base->vm_status, &vm->evicted);
>>> - base->moved = true;
>>> + amdgpu_vm_bo_evicted(base);
>>> }
>>>
>>> /**
>>> @@ -342,9 +404,7 @@ int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
>>> break;
>>>
>>> if (bo->tbo.type != ttm_bo_type_kernel) {
>>> - spin_lock(&vm->moved_lock);
>>> - list_move(&bo_base->vm_status, &vm->moved);
>>> - spin_unlock(&vm->moved_lock);
>>> + amdgpu_vm_bo_moved(bo_base);
>>> } else {
>>> if (vm->use_cpu_for_update)
>>> r = amdgpu_bo_kmap(bo, NULL);
>>> @@ -352,7 +412,7 @@ int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
>>> r = amdgpu_ttm_alloc_gart(&bo->tbo);
>>> if (r)
>>> break;
>>> - list_move(&bo_base->vm_status, &vm->relocated);
>>> + amdgpu_vm_bo_relocated(bo_base);
>>> }
>>> }
>>>
>>> @@ -1123,8 +1183,7 @@ int amdgpu_vm_update_directories(struct amdgpu_device *adev,
>>> bo_base = list_first_entry(&vm->relocated,
>>> struct amdgpu_vm_bo_base,
>>> vm_status);
>>> - bo_base->moved = false;
>>> - list_move(&bo_base->vm_status, &vm->idle);
>>> + amdgpu_vm_bo_idle(bo_base);
>>>
>>> bo = bo_base->bo->parent;
>>> if (!bo)
>>> @@ -1243,7 +1302,7 @@ static void amdgpu_vm_handle_huge_pages(struct amdgpu_pte_update_params *p,
>>> if (entry->huge) {
>>> /* Add the entry to the relocated list to update it. */
>>> entry->huge = false;
>>> - list_move(&entry->base.vm_status, &p->vm->relocated);
>>> + amdgpu_vm_bo_relocated(&entry->base);
>>> }
>>> return;
>>> }
>>> @@ -1746,9 +1805,9 @@ int amdgpu_vm_bo_update(struct amdgpu_device *adev,
>>> uint32_t mem_type = bo->tbo.mem.mem_type;
>>>
>>> if (!(bo->preferred_domains & amdgpu_mem_type_to_domain(mem_type)))
>>> - list_add_tail(&bo_va->base.vm_status, &vm->evicted);
>>> + amdgpu_vm_bo_evicted(&bo_va->base);
>>> else
>>> - list_add(&bo_va->base.vm_status, &vm->idle);
>>> + amdgpu_vm_bo_idle(&bo_va->base);
>> There is a small change in behaviour here for clearing
>> bo_va->base.moved. Not sure if it matters.
> Using list_add is minimal more efficient.
>
>>> }
>>>
>>> list_splice_init(&bo_va->invalids, &bo_va->valids);
>>> @@ -2472,28 +2531,20 @@ void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev,
>>>
>>> list_for_each_entry(bo_base, &bo->va, bo_list) {
>>> struct amdgpu_vm *vm = bo_base->vm;
>>> - bool was_moved = bo_base->moved;
>>>
>>> - bo_base->moved = true;
>>> if (evicted && bo->tbo.resv == vm->root.base.bo->tbo.resv) {
>>> - if (bo->tbo.type == ttm_bo_type_kernel)
>>> - list_move(&bo_base->vm_status, &vm->evicted);
>>> - else
>>> - list_move_tail(&bo_base->vm_status,
>>> - &vm->evicted);
>>> + amdgpu_vm_bo_evicted(bo_base);
>> I think here it's possible that the BO was on the moved list. I think
>> that means amdgpu_vm_bo_evicted should take the moved_lock just in case.
> Good point, probably best if I split up the moved list into per VM BOs
> and all other BOs.
>
> That should make the locking much more clear.
>
> Regards,
> Christian.
>
>> Regards,
>> Felix
>>
>>> continue;
>>> }
>>>
>>> - if (was_moved)
>>> + if (bo_base->moved)
>>> continue;
>>>
>>> - if (bo->tbo.type == ttm_bo_type_kernel) {
>>> - list_move(&bo_base->vm_status, &vm->relocated);
>>> - } else {
>>> - spin_lock(&bo_base->vm->moved_lock);
>>> - list_move(&bo_base->vm_status, &vm->moved);
>>> - spin_unlock(&bo_base->vm->moved_lock);
>>> - }
>>> + bo_base->moved = true;
>>> + if (bo->tbo.type == ttm_bo_type_kernel)
>>> + amdgpu_vm_bo_relocated(bo_base);
>>> + else
>>> + amdgpu_vm_bo_moved(bo_base);
>>> }
>>> }
>>>
More information about the amd-gfx
mailing list