[PATCH] drm/amdgpu: use HMM mirror callback to replace mmu notifier v3

christian.koenig@xxxxxxx (Christian König) · Fri, 14 Sep 2018 08:55:34 +0200

Am 13.09.2018 um 22:45 schrieb Philip Yang:
> On 2018-09-13 02:24 PM, Christian KÃ¶nig wrote:
>> Am 13.09.2018 um 20:00 schrieb Philip Yang:
>>> Replace our MMU notifier with hmm_mirror_ops.sync_cpu_device_pagetables
>>> callback. Enable CONFIG_HMM and CONFIG_HMM_MIRROR as a dependency in
>>> DRM_AMDGPU_USERPTR Kconfig.
>>>
>>> It supports both KFD userptr and gfx userptr paths.
>>>
>>> This depends on several HMM patchset from JÃ©rÃ´me Glisse queued for
>>> upstream. See
>>> http://172.27.226.38/root/kernel_amd/commits/hmm-dev-v01 (for AMD 
>>> intranet)
>>>
>>> Change-Id: Ie62c3c5e3c5b8521ab3b438d1eff2aa2a003835e
>>> Signed-off-by: Philip Yang <Philip.Yang at amd.com>
>>> ---
>>> Â  drivers/gpu/drm/amd/amdgpu/KconfigÂ Â Â Â  |Â  6 +--
>>> Â  drivers/gpu/drm/amd/amdgpu/MakefileÂ Â Â  |Â  2 +-
>>> Â  drivers/gpu/drm/amd/amdgpu/amdgpu_mn.c | 88 
>>> +++++++++++++++++++++++++++-------
>>> Â  3 files changed, 75 insertions(+), 21 deletions(-)
>>>
>>> diff --git a/drivers/gpu/drm/amd/amdgpu/Kconfig 
>>> b/drivers/gpu/drm/amd/amdgpu/Kconfig
>>> index 9221e54..960a633 100644
>>> --- a/drivers/gpu/drm/amd/amdgpu/Kconfig
>>> +++ b/drivers/gpu/drm/amd/amdgpu/Kconfig
>>> @@ -26,10 +26,10 @@ config DRM_AMDGPU_CIK
>>> Â  config DRM_AMDGPU_USERPTR
>>> Â Â Â Â Â  bool "Always enable userptr write support"
>>> Â Â Â Â Â  depends on DRM_AMDGPU
>>> -Â Â Â  select MMU_NOTIFIER
>>> +Â Â Â  select HMM_MIRROR
>>> Â Â Â Â Â  help
>>> -Â Â Â Â Â  This option selects CONFIG_MMU_NOTIFIER if it isn't already
>>> -Â Â Â Â Â  selected to enabled full userptr support.
>>> +Â Â Â Â Â  This option selects CONFIG_HMM and CONFIG_HMM_MIRROR if it
>>> +Â Â Â Â Â  isn't already selected to enabled full userptr support.
>>> Â  Â  config DRM_AMDGPU_GART_DEBUGFS
>>> Â Â Â Â Â  bool "Allow GART access through debugfs"
>>> diff --git a/drivers/gpu/drm/amd/amdgpu/Makefile 
>>> b/drivers/gpu/drm/amd/amdgpu/Makefile
>>> index 138cb78..c1e5d43 100644
>>> --- a/drivers/gpu/drm/amd/amdgpu/Makefile
>>> +++ b/drivers/gpu/drm/amd/amdgpu/Makefile
>>> @@ -171,7 +171,7 @@ endif
>>> Â  amdgpu-$(CONFIG_COMPAT) += amdgpu_ioc32.o
>>> Â  amdgpu-$(CONFIG_VGA_SWITCHEROO) += amdgpu_atpx_handler.o
>>> Â  amdgpu-$(CONFIG_ACPI) += amdgpu_acpi.o
>>> -amdgpu-$(CONFIG_MMU_NOTIFIER) += amdgpu_mn.o
>>> +amdgpu-$(CONFIG_HMM) += amdgpu_mn.o
>>> Â  Â  include $(FULL_AMD_PATH)/powerplay/Makefile
>>> Â  diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_mn.c 
>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_mn.c
>>> index e55508b..ea8671f6 100644
>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_mn.c
>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_mn.c
>>> @@ -46,6 +46,7 @@
>>> Â  #include <linux/firmware.h>
>>> Â  #include <linux/module.h>
>>> Â  #include <linux/mmu_notifier.h>
>>> +#include <linux/hmm.h>
>>
>> Can we now drop including linux/mmu_notifier.h?
> Yes, will use hmm_mirror_ops to replace gfx and kfd mmu_notifier_ops

Please drop that and implement the gfx and kfd operations directly.

Thanks,
Christian.

>>> Â  #include <linux/interval_tree.h>
>>> Â  #include <drm/drmP.h>
>>> Â  #include <drm/drm.h>
>>> @@ -66,6 +67,7 @@
>>> Â Â  * @objects: interval tree containing amdgpu_mn_nodes
>>> Â Â  * @read_lock: mutex for recursive locking of @lock
>>> Â Â  * @recursion: depth of recursion
>>> + * @mirror: HMM mirror function support
>>> Â Â  *
>>> Â Â  * Data for each amdgpu device and process address space.
>>> Â Â  */
>>> @@ -87,6 +89,9 @@ struct amdgpu_mn {
>>> Â Â Â Â Â  struct rb_root_cachedÂ Â Â  objects;
>>> Â Â Â Â Â  struct mutexÂ Â Â Â Â Â Â  read_lock;
>>> Â Â Â Â Â  atomic_tÂ Â Â Â Â Â Â  recursion;
>>> +
>>> +Â Â Â  /* HMM mirror */
>>> +Â Â Â  struct hmm_mirrorÂ Â Â  mirror;
>>> Â  };
>>> Â  Â  /**
>>> @@ -103,7 +108,7 @@ struct amdgpu_mn_node {
>>> Â  };
>>> Â  Â  /**
>>> - * amdgpu_mn_destroy - destroy the MMU notifier
>>> + * amdgpu_mn_destroy - destroy the HMM mirror
>>> Â Â  *
>>> Â Â  * @work: previously sheduled work item
>>> Â Â  *
>>> @@ -129,28 +134,27 @@ static void amdgpu_mn_destroy(struct 
>>> work_struct *work)
>>> Â Â Â Â Â  }
>>> Â Â Â Â Â  up_write(&amn->lock);
>>> Â Â Â Â Â  mutex_unlock(&adev->mn_lock);
>>> -Â Â Â  mmu_notifier_unregister_no_release(&amn->mn, amn->mm);
>>> +
>>> +Â Â Â  hmm_mirror_unregister(&amn->mirror);
>>> +
>>> Â Â Â Â Â  kfree(amn);
>>> Â  }
>>> Â  Â  /**
>>> Â Â  * amdgpu_mn_release - callback to notify about mm destruction
>>> Â Â  *
>>> - * @mn: our notifier
>>> - * @mm: the mm this callback is about
>>> + * @mirror: the HMM mirror (mm) this callback is about
>>> Â Â  *
>>> - * Shedule a work item to lazy destroy our notifier.
>>> + * Shedule a work item to lazy destroy HMM mirror.
>>> Â Â  */
>>> -static void amdgpu_mn_release(struct mmu_notifier *mn,
>>> -Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  struct mm_struct *mm)
>>> +static void amdgpu_hmm_mirror_release(struct hmm_mirror *mirror)
>>> Â  {
>>> -Â Â Â  struct amdgpu_mn *amn = container_of(mn, struct amdgpu_mn, mn);
>>> +Â Â Â  struct amdgpu_mn *amn = container_of(mirror, struct amdgpu_mn, 
>>> mirror);
>>> Â  Â Â Â Â Â  INIT_WORK(&amn->work, amdgpu_mn_destroy);
>>> Â Â Â Â Â  schedule_work(&amn->work);
>>> Â  }
>>> Â  -
>>> Â  /**
>>> Â Â  * amdgpu_mn_lock - take the write side lock for this notifier
>>> Â Â  *
>>> @@ -355,12 +359,10 @@ static void 
>>> amdgpu_mn_invalidate_range_end(struct mmu_notifier *mn,
>>> Â  Â  static const struct mmu_notifier_ops amdgpu_mn_ops[] = {
>>> Â Â Â Â Â  [AMDGPU_MN_TYPE_GFX] = {
>>> -Â Â Â Â Â Â Â  .release = amdgpu_mn_release,
>>> Â Â Â Â Â Â Â Â Â  .invalidate_range_start = 
>>> amdgpu_mn_invalidate_range_start_gfx,
>>> Â Â Â Â Â Â Â Â Â  .invalidate_range_end = amdgpu_mn_invalidate_range_end,
>>> Â Â Â Â Â  },
>>> Â Â Â Â Â  [AMDGPU_MN_TYPE_HSA] = {
>>> -Â Â Â Â Â Â Â  .release = amdgpu_mn_release,
>>> Â Â Â Â Â Â Â Â Â  .invalidate_range_start = 
>>> amdgpu_mn_invalidate_range_start_hsa,
>>> Â Â Â Â Â Â Â Â Â  .invalidate_range_end = amdgpu_mn_invalidate_range_end,
>>> Â Â Â Â Â  },
>>> @@ -373,12 +375,63 @@ static const struct mmu_notifier_ops 
>>> amdgpu_mn_ops[] = {
>>> Â  #define AMDGPU_MN_KEY(mm, type) ((unsigned long)(mm) + (type))
>>> Â  Â  /**
>>> - * amdgpu_mn_get - create notifier context
>>> + * amdgpu_hmm_sync_cpu_device_pagetables - synchronize CPU/GPU page 
>>> tables
>>> + *
>>> + * @mirror: the hmm_mirror (mm) is about to update
>>> + * @update: the update start, end address
>>> + *
>>> + * This callback is called from mmu_notifiers when the CPU page 
>>> table is
>>> + * updated.
>>> + */
>>> +static int amdgpu_hmm_sync_cpu_device_pagetables(struct hmm_mirror 
>>> *mirror,
>>> +Â Â Â Â Â Â Â Â Â Â Â  const struct hmm_update *update)
>>> +{
>>> +Â Â Â  struct amdgpu_mn *amn;
>>> +Â Â Â  struct hmm *hmm;
>>> +Â Â Â  struct mm_struct *mm;
>>> +Â Â Â  unsigned long start;
>>> +Â Â Â  unsigned long end;
>>> +Â Â Â  unsigned long key;
>>> +Â Â Â  int r = 0;
>>> +
>>> +Â Â Â  start = update->start;
>>> +Â Â Â  end = update->end;
>>> +Â Â Â  hmm = mirror->hmm;
>>> +Â Â Â  mm = *(struct mm_struct **)hmm;
>>> +
>>> +Â Â Â  pr_debug("mirror %p start %lx end %lx\n", mirror, start, end);
>>
>> You should probably remove those pr_debug lines before upstreaming.
>>
>> Alternatively we could turn them into trace points.
>>
>>> +
>>> +Â Â Â  amn = container_of(mirror, struct amdgpu_mn, mirror);
>>> +Â Â Â  key = AMDGPU_MN_KEY(mm, amn->type);
>>> +
>>> +Â Â Â  hash_for_each_possible(amn->adev->mn_hash, amn, node, key)
>>> +Â Â Â Â Â Â Â  if (AMDGPU_MN_KEY(amn->mm, amn->type) == key) {
>>> +Â Â Â Â Â Â Â Â Â Â Â  r = amn->mn.ops->invalidate_range_start(&amn->mn, mm,
>>> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  start, end,
>>> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  update->blockable);
>>> + amn->mn.ops->invalidate_range_end(&amn->mn, mm,
>>> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  start, end);
>>> +Â Â Â Â Â Â Â Â Â Â Â  if (r) {
>>> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  DRM_ERROR("Failed to invalidate %lx\n", start);
>>> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  break;
>>> +Â Â Â Â Â Â Â Â Â Â Â  }
>>> +Â Â Â Â Â Â Â  }
>>
>> That looks fishy to me, why is this necessary?
>>
> Yeah, it is unnecessary, one process (mm) register once hmm mirror amn 
> for each adev,
> so one mm will get individual call back, do not need use mm to lookup 
> amn from adev->mn_hash.
>
>>> +
>>> +Â Â Â  return r;
>>> +}
>>> +
>>> +static struct hmm_mirror_ops amdgpu_hmm_mirror_ops = {
>>> +Â Â Â  .sync_cpu_device_pagetables = 
>>> amdgpu_hmm_sync_cpu_device_pagetables,
>>> +Â Â Â  .release = amdgpu_hmm_mirror_release
>>> +};
>>
>> You should probably clean that up as well and have a separate one for 
>> each AMDGPU_MN_TYPE_*.
>>
> Done, will submit v4.
>> Christian.
>>
>>> +
>>> +/**
>>> + * amdgpu_mn_get - create HMM mirror context
>>> Â Â  *
>>> Â Â  * @adev: amdgpu device pointer
>>> Â Â  * @type: type of MMU notifier context
>>> Â Â  *
>>> - * Creates a notifier context for current->mm.
>>> + * Creates a HMM mirror context for current->mm.
>>> Â Â  */
>>> Â  struct amdgpu_mn *amdgpu_mn_get(struct amdgpu_device *adev,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  enum amdgpu_mn_type type)
>>> @@ -413,7 +466,8 @@ struct amdgpu_mn *amdgpu_mn_get(struct 
>>> amdgpu_device *adev,
>>> Â Â Â Â Â  mutex_init(&amn->read_lock);
>>> Â Â Â Â Â  atomic_set(&amn->recursion, 0);
>>> Â  -Â Â Â  r = __mmu_notifier_register(&amn->mn, mm);
>>> +Â Â Â  amn->mirror.ops = &amdgpu_hmm_mirror_ops;
>>> +Â Â Â  r = hmm_mirror_register(&amn->mirror, mm);
>>> Â Â Â Â Â  if (r)
>>> Â Â Â Â Â Â Â Â Â  goto free_amn;
>>> Â  @@ -439,7 +493,7 @@ struct amdgpu_mn *amdgpu_mn_get(struct 
>>> amdgpu_device *adev,
>>> Â Â  * @bo: amdgpu buffer object
>>> Â Â  * @addr: userptr addr we should monitor
>>> Â Â  *
>>> - * Registers an MMU notifier for the given BO at the specified 
>>> address.
>>> + * Registers an HMM mirror for the given BO at the specified address.
>>> Â Â  * Returns 0 on success, -ERRNO if anything goes wrong.
>>> Â Â  */
>>> Â  int amdgpu_mn_register(struct amdgpu_bo *bo, unsigned long addr)
>>> @@ -495,11 +549,11 @@ int amdgpu_mn_register(struct amdgpu_bo *bo, 
>>> unsigned long addr)
>>> Â  }
>>> Â  Â  /**
>>> - * amdgpu_mn_unregister - unregister a BO for notifier updates
>>> + * amdgpu_mn_unregister - unregister a BO for HMM mirror updates
>>> Â Â  *
>>> Â Â  * @bo: amdgpu buffer object
>>> Â Â  *
>>> - * Remove any registration of MMU notifier updates from the buffer 
>>> object.
>>> + * Remove any registration of HMM mirror updates from the buffer 
>>> object.
>>> Â Â  */
>>> Â  void amdgpu_mn_unregister(struct amdgpu_bo *bo)
>>> Â  {
>>
>