[PATCH 06/25] drm/amdgpu: Add KFD eviction fence

[Felix Kuehling]

[+Harish, forgot to acknowledge him in the commit description, will fix
that in v2]

Harish, please see Christian's question below in amd_kfd_fence_signal.
Did I understand this correctly?

Regards,
  Felix

On 2018-01-28 06:42 PM, Felix Kuehling wrote:
> On 2018-01-27 04:16 AM, Christian König wrote:
>> Am 27.01.2018 um 02:09 schrieb Felix Kuehling:
> [snip]
>>> +struct amdgpu_amdkfd_fence *amdgpu_amdkfd_fence_create(u64 context,
>>> +                               void *mm)
>>> +{
>>> +    struct amdgpu_amdkfd_fence *fence = NULL;
>>> +
>>> +    fence = kzalloc(sizeof(*fence), GFP_KERNEL);
>>> +    if (fence == NULL)
>>> +        return NULL;
>>> +
>>> +    /* mm_struct mm is used as void pointer to identify the parent
>>> +     * KFD process. Don't dereference it. Fence and any threads using
>>> +     * mm is guranteed to be released before process termination.
>>> +     */
>>> +    fence->mm = mm;
>> That won't work. Fences can live much longer than any process who
>> created them.
>>
>> I've already found a fence in a BO still living hours after the
>> process was killed and the pid long recycled.
>>
>> I suggest to make fence->mm a real mm_struct pointer with reference
>> counting and then set it to NULL and drop the reference in
>> enable_signaling.
> I agree. But enable_signaling may be too early to drop the reference.
> amd_kfd_fence_check_mm could still be called later from
> amdgpu_ttm_bo_eviction_valuable, as long as the fence hasn't signaled yet.
>
> The safe place is problably in amd_kfd_fence_release.
>
>>> +    get_task_comm(fence->timeline_name, current);
>>> +    spin_lock_init(&fence->lock);
>>> +
>>> +    dma_fence_init(&fence->base, &amd_kfd_fence_ops, &fence->lock,
>>> +           context, atomic_inc_return(&fence_seq));
>>> +
>>> +    return fence;
>>> +}
>>> +
>>> +struct amdgpu_amdkfd_fence *to_amdgpu_amdkfd_fence(struct dma_fence *f)
>>> +{
>>> +    struct amdgpu_amdkfd_fence *fence;
>>> +
>>> +    if (!f)
>>> +        return NULL;
>>> +
>>> +    fence = container_of(f, struct amdgpu_amdkfd_fence, base);
>>> +    if (fence && f->ops == &amd_kfd_fence_ops)
>>> +        return fence;
>>> +
>>> +    return NULL;
>>> +}
>>> +
>>> +static const char *amd_kfd_fence_get_driver_name(struct dma_fence *f)
>>> +{
>>> +    return "amdgpu_amdkfd_fence";
>>> +}
>>> +
>>> +static const char *amd_kfd_fence_get_timeline_name(struct dma_fence *f)
>>> +{
>>> +    struct amdgpu_amdkfd_fence *fence = to_amdgpu_amdkfd_fence(f);
>>> +
>>> +    return fence->timeline_name;
>>> +}
>>> +
>>> +/**
>>> + * amd_kfd_fence_enable_signaling - This gets called when TTM wants
>>> to evict
>>> + *  a KFD BO and schedules a job to move the BO.
>>> + *  If fence is already signaled return true.
>>> + *  If fence is not signaled schedule a evict KFD process work item.
>>> + */
>>> +static bool amd_kfd_fence_enable_signaling(struct dma_fence *f)
>>> +{
>>> +    struct amdgpu_amdkfd_fence *fence = to_amdgpu_amdkfd_fence(f);
>>> +
>>> +    if (!fence)
>>> +        return false;
>>> +
>>> +    if (dma_fence_is_signaled(f))
>>> +        return true;
>>> +
>>> +    if (!kgd2kfd->schedule_evict_and_restore_process(
>>> +                (struct mm_struct *)fence->mm, f))
>>> +        return true;
>>> +
>>> +    return false;
>>> +}
>>> +
>>> +static int amd_kfd_fence_signal(struct dma_fence *f)
>>> +{
>>> +    unsigned long flags;
>>> +    int ret;
>>> +
>>> +    spin_lock_irqsave(f->lock, flags);
>>> +    /* Set enabled bit so cb will called */
>>> +    set_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT, &f->flags);
>> Mhm, why is that necessary?
> This only gets called from fence_release below. I think this is to avoid
> needlessly scheduling an eviction/restore cycle when an eviction fence
> gets destroyed that hasn't been triggered before, probably during
> process termination.
>
> Harish, do you remember any other reason for this?
>
>>> +    ret = dma_fence_signal_locked(f);
>>> +    spin_unlock_irqrestore(f->lock, flags);
>>> +
>>> +    return ret;
>>> +}
>>> +
>>> +/**
>>> + * amd_kfd_fence_release - callback that fence can be freed
>>> + *
>>> + * @fence: fence
>>> + *
>>> + * This function is called when the reference count becomes zero.
>>> + * It just RCU schedules freeing up the fence.
>>> + */
>>> +static void amd_kfd_fence_release(struct dma_fence *f)
>>> +{
>>> +    struct amdgpu_amdkfd_fence *fence = to_amdgpu_amdkfd_fence(f);
>>> +    /* Unconditionally signal the fence. The process is getting
>>> +     * terminated.
>>> +     */
>>> +    if (WARN_ON(!fence))
>>> +        return; /* Not an amdgpu_amdkfd_fence */
>>> +
>>> +    amd_kfd_fence_signal(f);
>>> +    kfree_rcu(f, rcu);
>>> +}
>>> +
>>> +/**
>>> + * amd_kfd_fence_check_mm - Check if @mm is same as that of the
>>> fence @f
>>> + *  if same return TRUE else return FALSE.
>>> + *
>>> + * @f: [IN] fence
>>> + * @mm: [IN] mm that needs to be verified
>>> + */
>>> +bool amd_kfd_fence_check_mm(struct dma_fence *f, void *mm)
>>> +{
>>> +    struct amdgpu_amdkfd_fence *fence = to_amdgpu_amdkfd_fence(f);
>>> +
>>> +    if (!fence)
>>> +        return false;
>>> +    else if (fence->mm == mm)
>>> +        return true;
>>> +
>>> +    return false;
>>> +}
>>> +
>>> +const struct dma_fence_ops amd_kfd_fence_ops = {
>>> +    .get_driver_name = amd_kfd_fence_get_driver_name,
>>> +    .get_timeline_name = amd_kfd_fence_get_timeline_name,
>>> +    .enable_signaling = amd_kfd_fence_enable_signaling,
>>> +    .signaled = NULL,
>>> +    .wait = dma_fence_default_wait,
>>> +    .release = amd_kfd_fence_release,
>>> +};
>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
>>> index 65d5a4e..ca00dd2 100644
>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
>>> @@ -36,8 +36,9 @@
>>>   #define AMDGPU_MAX_UVD_ENC_RINGS    2
>>>     /* some special values for the owner field */
>>> -#define AMDGPU_FENCE_OWNER_UNDEFINED    ((void*)0ul)
>>> -#define AMDGPU_FENCE_OWNER_VM        ((void*)1ul)
>>> +#define AMDGPU_FENCE_OWNER_UNDEFINED    ((void *)0ul)
>>> +#define AMDGPU_FENCE_OWNER_VM        ((void *)1ul)
>>> +#define AMDGPU_FENCE_OWNER_KFD        ((void *)2ul)
>>>     #define AMDGPU_FENCE_FLAG_64BIT         (1 << 0)
>>>   #define AMDGPU_FENCE_FLAG_INT           (1 << 1)
>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c
>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c
>>> index df65c66..0cb31d9 100644
>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c
>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c
>>> @@ -31,6 +31,7 @@
>>>   #include <drm/drmP.h>
>>>   #include "amdgpu.h"
>>>   #include "amdgpu_trace.h"
>>> +#include "amdgpu_amdkfd.h"
>>>     struct amdgpu_sync_entry {
>>>       struct hlist_node    node;
>>> @@ -86,10 +87,18 @@ static bool amdgpu_sync_same_dev(struct
>>> amdgpu_device *adev,
>>>   static void *amdgpu_sync_get_owner(struct dma_fence *f)
>>>   {
>>>       struct drm_sched_fence *s_fence = to_drm_sched_fence(f);
>>> +    struct amdgpu_amdkfd_fence *kfd_fence;
>>> +
>>> +    if (!f)
>>> +        return AMDGPU_FENCE_OWNER_UNDEFINED;
>> When you add the extra NULL check here then please move the
>> to_drm_sched_fence() after it as well.
> Yeah, makes sense.
>
> Regards,
>   Felix
>
>> Christian.
>>
>>>         if (s_fence)
>>>           return s_fence->owner;
>>>   +    kfd_fence = to_amdgpu_amdkfd_fence(f);
>>> +    if (kfd_fence)
>>> +        return AMDGPU_FENCE_OWNER_KFD;
>>> +
>>>       return AMDGPU_FENCE_OWNER_UNDEFINED;
>>>   }
>>>   @@ -204,11 +213,18 @@ int amdgpu_sync_resv(struct amdgpu_device *adev,
>>>       for (i = 0; i < flist->shared_count; ++i) {
>>>           f = rcu_dereference_protected(flist->shared[i],
>>>                             reservation_object_held(resv));
>>> +        /* We only want to trigger KFD eviction fences on
>>> +         * evict or move jobs. Skip KFD fences otherwise.
>>> +         */
>>> +        fence_owner = amdgpu_sync_get_owner(f);
>>> +        if (fence_owner == AMDGPU_FENCE_OWNER_KFD &&
>>> +            owner != AMDGPU_FENCE_OWNER_UNDEFINED)
>>> +            continue;
>>> +
>>>           if (amdgpu_sync_same_dev(adev, f)) {
>>>               /* VM updates are only interesting
>>>                * for other VM updates and moves.
>>>                */
>>> -            fence_owner = amdgpu_sync_get_owner(f);
>>>               if ((owner != AMDGPU_FENCE_OWNER_UNDEFINED) &&
>>>                   (fence_owner != AMDGPU_FENCE_OWNER_UNDEFINED) &&
>>>                   ((owner == AMDGPU_FENCE_OWNER_VM) !=
>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
>>> index e4bb435..c3f33d3 100644
>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
>>> @@ -46,6 +46,7 @@
>>>   #include "amdgpu.h"
>>>   #include "amdgpu_object.h"
>>>   #include "amdgpu_trace.h"
>>> +#include "amdgpu_amdkfd.h"
>>>   #include "bif/bif_4_1_d.h"
>>>     #define DRM_FILE_PAGE_OFFSET (0x100000000ULL >> PAGE_SHIFT)
>>> @@ -1170,6 +1171,23 @@ static bool
>>> amdgpu_ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
>>>   {
>>>       unsigned long num_pages = bo->mem.num_pages;
>>>       struct drm_mm_node *node = bo->mem.mm_node;
>>> +    struct reservation_object_list *flist;
>>> +    struct dma_fence *f;
>>> +    int i;
>>> +
>>> +    /* If bo is a KFD BO, check if the bo belongs to the current
>>> process.
>>> +     * If true, then return false as any KFD process needs all its
>>> BOs to
>>> +     * be resident to run successfully
>>> +     */
>>> +    flist = reservation_object_get_list(bo->resv);
>>> +    if (flist) {
>>> +        for (i = 0; i < flist->shared_count; ++i) {
>>> +            f = rcu_dereference_protected(flist->shared[i],
>>> +                reservation_object_held(bo->resv));
>>> +            if (amd_kfd_fence_check_mm(f, current->mm))
>>> +                return false;
>>> +        }
>>> +    }
>>>         switch (bo->mem.mem_type) {
>>>       case TTM_PL_TT:
>>> diff --git a/drivers/gpu/drm/amd/include/kgd_kfd_interface.h
>>> b/drivers/gpu/drm/amd/include/kgd_kfd_interface.h
>>> index 94eab54..9e35249 100644
>>> --- a/drivers/gpu/drm/amd/include/kgd_kfd_interface.h
>>> +++ b/drivers/gpu/drm/amd/include/kgd_kfd_interface.h
>>> @@ -30,6 +30,7 @@
>>>     #include <linux/types.h>
>>>   #include <linux/bitmap.h>
>>> +#include <linux/dma-fence.h>
>>>     struct pci_dev;
>>>   @@ -286,6 +287,9 @@ struct kfd2kgd_calls {
>>>    *
>>>    * @resume: Notifies amdkfd about a resume action done to a kgd device
>>>    *
>>> + * @schedule_evict_and_restore_process: Schedules work queue that
>>> will prepare
>>> + * for safe eviction of KFD BOs that belong to the specified process.
>>> + *
>>>    * This structure contains function callback pointers so the kgd
>>> driver
>>>    * will notify to the amdkfd about certain status changes.
>>>    *
>>> @@ -300,6 +304,8 @@ struct kgd2kfd_calls {
>>>       void (*interrupt)(struct kfd_dev *kfd, const void *ih_ring_entry);
>>>       void (*suspend)(struct kfd_dev *kfd);
>>>       int (*resume)(struct kfd_dev *kfd);
>>> +    int (*schedule_evict_and_restore_process)(struct mm_struct *mm,
>>> +            struct dma_fence *fence);
>>>   };
>>>     int kgd2kfd_init(unsigned interface_version,