Re: [PATCH] drm/amd/amdgpu: remove vram_page_split kernel option (v3)

["Christian König" <ckoenig.leichtzumerken-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org>]

Am 06.06.19 um 13:51 schrieb StDenis, Tom:
> On 2019-06-06 7:49 a.m., Christian König wrote:
>> Am 06.06.19 um 12:50 schrieb StDenis, Tom:
>>> This option is no longer needed.  The default code paths
>>> are now the only option.
>>>
>>> v2: Add HPAGE support and a default for non contiguous maps
>>> v3: Misread 512 pages as MiB ...
>>>
>>> Signed-off-by: Tom St Denis <tom.stdenis@amd.com>
>>> ---
>>>    drivers/gpu/drm/amd/amdgpu/amdgpu.h          |  1 -
>>>    drivers/gpu/drm/amd/amdgpu/amdgpu_device.c   |  7 -------
>>>    drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c      |  8 --------
>>>    drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c | 14 +++++++++-----
>>>    4 files changed, 9 insertions(+), 21 deletions(-)
>>>
>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
>>> index 19a00282e34c..e54f31865f60 100644
>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
>>> @@ -142,7 +142,6 @@ extern uint amdgpu_sdma_phase_quantum;
>>>    extern char *amdgpu_disable_cu;
>>>    extern char *amdgpu_virtual_display;
>>>    extern uint amdgpu_pp_feature_mask;
>>> -extern int amdgpu_vram_page_split;
>>>    extern int amdgpu_ngg;
>>>    extern int amdgpu_prim_buf_per_se;
>>>    extern int amdgpu_pos_buf_per_se;
>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>>> index 0212c9ee317c..2e13b8ef6681 100644
>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>>> @@ -973,13 +973,6 @@ static int amdgpu_device_check_arguments(struct
>>> amdgpu_device *adev)
>>>          amdgpu_device_check_block_size(adev);
>>>    -    if (amdgpu_vram_page_split != -1 && (amdgpu_vram_page_split <
>>> 16 ||
>>> -        !is_power_of_2(amdgpu_vram_page_split))) {
>>> -        dev_warn(adev->dev, "invalid VRAM page split (%d)\n",
>>> -             amdgpu_vram_page_split);
>>> -        amdgpu_vram_page_split = 1024;
>>> -    }
>>> -
>>>        ret = amdgpu_device_get_job_timeout_settings(adev);
>>>        if (ret) {
>>>            dev_err(adev->dev, "invalid lockup_timeout parameter
>>> syntax\n");
>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
>>> index 1f38d6fc1fe3..ef22a2a25448 100644
>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
>>> @@ -108,7 +108,6 @@ int amdgpu_vm_fragment_size = -1;
>>>    int amdgpu_vm_block_size = -1;
>>>    int amdgpu_vm_fault_stop = 0;
>>>    int amdgpu_vm_debug = 0;
>>> -int amdgpu_vram_page_split = 512;
>>>    int amdgpu_vm_update_mode = -1;
>>>    int amdgpu_exp_hw_support = 0;
>>>    int amdgpu_dc = -1;
>>> @@ -342,13 +341,6 @@ module_param_named(vm_debug, amdgpu_vm_debug,
>>> int, 0644);
>>>    MODULE_PARM_DESC(vm_update_mode, "VM update using CPU (0 = never
>>> (default except for large BAR(LB)), 1 = Graphics only, 2 = Compute
>>> only (default for LB), 3 = Both");
>>>    module_param_named(vm_update_mode, amdgpu_vm_update_mode, int, 0444);
>>>    -/**
>>> - * DOC: vram_page_split (int)
>>> - * Override the number of pages after we split VRAM allocations
>>> (default 512, -1 = disable). The default is 512.
>>> - */
>>> -MODULE_PARM_DESC(vram_page_split, "Number of pages after we split
>>> VRAM allocations (default 512, -1 = disable)");
>>> -module_param_named(vram_page_split, amdgpu_vram_page_split, int, 0444);
>>> -
>>>    /**
>>>     * DOC: exp_hw_support (int)
>>>     * Enable experimental hw support (1 = enable). The default is 0
>>> (disabled).
>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c
>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c
>>> index ec9ea3fdbb4a..8aea2f21b202 100644
>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c
>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c
>>> @@ -284,17 +284,21 @@ static int amdgpu_vram_mgr_new(struct
>>> ttm_mem_type_manager *man,
>>>        if (!lpfn)
>>>            lpfn = man->size;
>>>    -    if (place->flags & TTM_PL_FLAG_CONTIGUOUS ||
>>> -        amdgpu_vram_page_split == -1) {
>>> +    if (place->flags & TTM_PL_FLAG_CONTIGUOUS) {
>>>            pages_per_node = ~0ul;
>>>            num_nodes = 1;
>>>        } else {
>>> -        pages_per_node = max((uint32_t)amdgpu_vram_page_split,
>>> -                     mem->page_alignment);
>>> +#ifdef CONFIG_TRANSPARENT_HUGEPAGE
>>> +        pages_per_node = HPAGE_PMD_NR;
>>> +#else
>>> +        /* default to 2MB */
>>> +        pages_per_node = (2UL << (20UL - PAGE_SHIFT));
>>> +#endif
>>> +        pages_per_node = max((uint32_t)pages_per_node,
>>> mem->page_alignment);
>>>            num_nodes = DIV_ROUND_UP(mem->num_pages, pages_per_node);
>>>        }
>>>    -    nodes = kvmalloc_array(num_nodes, sizeof(*nodes),
>>> +    nodes = kvmalloc_array((uint32_t)num_nodes, sizeof(*nodes),
>> You can probably drop that cast here, apart from that the patch is
>> Reviewed-by: Christian König <christian.koenig@amd.com>.
>
> Without the cast it produces a somewhat lengthy warning diagnostic.  The
> original code had the cast as well.

Strange, the original code hat the case because amdgpu_vram_page_split 
was a signed int, but num_nodes should be unsigned anyway.

So that shouldn't be necessary any more. Anyway keeping the cast works 
for me as well.

>> Do you want to look into test/benchmark as well?
> Sure, is the plan to cleave all that out too?

Well first of all dig a bit into what those still do. If I'm not 
completely mistaken at least one of them is redundant now.

Christian.

>
>
> Thanks,
>
> Tom
>
>
> _______________________________________________
> amd-gfx mailing list
> amd-gfx@lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx