From: Felix Kuehling <felix.kuehling@amd.com>
To: "Christian König" <ckoenig.leichtzumerken@gmail.com>,
friedrich.vock@gmx.de, bas@basnieuwenhuizen.nl,
ishitatsuyuki@gmail.com
Cc: amd-gfx@lists.freedesktop.org
Subject: Re: [PATCH 2/3] drm/amdgpu: sync to KFD fences before clearing PTEs
Date: Wed, 21 Aug 2024 16:01:07 -0400 [thread overview]
Message-ID: <a511a28c-13d7-452f-96bd-911148c4d175@amd.com> (raw)
In-Reply-To: <20240821120324.4583-2-christian.koenig@amd.com>
On 2024-08-21 08:03, Christian König wrote:
> This patch tries to solve the basic problem we also need to sync to
> the KFD fences of the BO because otherwise it can be that we clear
> PTEs while the KFD queues are still running.
This is going to trigger a lot of phantom KFD evictions and will tank
performance. It's probably not what you intended.
Regards,
Felix
>
> Signed-off-by: Christian König <christian.koenig@amd.com>
> ---
> drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c | 30 ++++++++++++++++++++++++
> drivers/gpu/drm/amd/amdgpu/amdgpu_sync.h | 1 +
> drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 6 +++++
> 3 files changed, 37 insertions(+)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c
> index bdf1ef825d89..c586ab4c911b 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c
> @@ -260,6 +260,36 @@ int amdgpu_sync_resv(struct amdgpu_device *adev, struct amdgpu_sync *sync,
> return 0;
> }
>
> +/**
> + * amdgpu_sync_kfd - sync to KFD fences
> + *
> + * @sync: sync object to add KFD fences to
> + * @resv: reservation object with KFD fences
> + *
> + * Extract all KFD fences and add them to the sync object.
> + */
> +int amdgpu_sync_kfd(struct amdgpu_sync *sync, struct dma_resv *resv)
> +{
> + struct dma_resv_iter cursor;
> + struct dma_fence *f;
> + int r = 0;
> +
> + dma_resv_iter_begin(&cursor, resv, DMA_RESV_USAGE_BOOKKEEP);
> + dma_resv_for_each_fence_unlocked(&cursor, f) {
> + void *fence_owner = amdgpu_sync_get_owner(f);
> +
> + if (fence_owner != AMDGPU_FENCE_OWNER_KFD)
> + continue;
> +
> + r = amdgpu_sync_fence(sync, f);
> + if (r)
> + break;
> + }
> + dma_resv_iter_end(&cursor);
> +
> + return r;
> +}
> +
> /* Free the entry back to the slab */
> static void amdgpu_sync_entry_free(struct amdgpu_sync_entry *e)
> {
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.h
> index cf1e9e858efd..e3272dce798d 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.h
> @@ -51,6 +51,7 @@ int amdgpu_sync_fence(struct amdgpu_sync *sync, struct dma_fence *f);
> int amdgpu_sync_resv(struct amdgpu_device *adev, struct amdgpu_sync *sync,
> struct dma_resv *resv, enum amdgpu_sync_mode mode,
> void *owner);
> +int amdgpu_sync_kfd(struct amdgpu_sync *sync, struct dma_resv *resv);
> struct dma_fence *amdgpu_sync_peek_fence(struct amdgpu_sync *sync,
> struct amdgpu_ring *ring);
> struct dma_fence *amdgpu_sync_get_fence(struct amdgpu_sync *sync);
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> index ba99d428610a..13d429b91327 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> @@ -1168,6 +1168,12 @@ int amdgpu_vm_bo_update(struct amdgpu_device *adev, struct amdgpu_bo_va *bo_va,
> AMDGPU_SYNC_EQ_OWNER, vm);
> if (r)
> goto error_free;
> + if (bo) {
> + r = amdgpu_sync_kfd(&sync, bo->tbo.base.resv);
> + if (r)
> + goto error_free;
> + }
> +
> } else {
> struct drm_gem_object *obj = &bo->tbo.base;
>
next prev parent reply other threads:[~2024-08-21 20:01 UTC|newest]
Thread overview: 11+ messages / expand[flat|nested] mbox.gz Atom feed top
2024-08-21 12:03 [PATCH 1/3] drm/amdgpu: re-work VM syncing Christian König
2024-08-21 12:03 ` [PATCH 2/3] drm/amdgpu: sync to KFD fences before clearing PTEs Christian König
2024-08-21 20:01 ` Felix Kuehling [this message]
2024-08-22 9:07 ` Christian König
2024-08-28 22:40 ` Felix Kuehling
2024-08-29 9:48 ` Christian König
2024-08-21 12:03 ` [PATCH 3/3] drm/amdgpu: stop syncing PRT map operations Christian König
2024-08-21 17:39 ` [PATCH 1/3] drm/amdgpu: re-work VM syncing Friedrich Vock
2024-08-21 20:46 ` Felix Kuehling
2024-08-22 7:28 ` Friedrich Vock
2024-08-28 22:29 ` Felix Kuehling
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=a511a28c-13d7-452f-96bd-911148c4d175@amd.com \
--to=felix.kuehling@amd.com \
--cc=amd-gfx@lists.freedesktop.org \
--cc=bas@basnieuwenhuizen.nl \
--cc=ckoenig.leichtzumerken@gmail.com \
--cc=friedrich.vock@gmx.de \
--cc=ishitatsuyuki@gmail.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox