在 2018/12/12 20:24, Daniel Vetter 写道: > On Wed, Dec 12, 2018 at 12:40 PM Zhou, David(ChunMing) > <David1.Zhou@xxxxxxx> wrote: >> + Daniel Rakos and Jason Ekstrand. >> >> Below is the background, which is from Daniel R should be able to explain that's why: >> " ISVs, especially those coming from D3D12, are unsatisfied with the behavior of the Vulkan semaphores as they are unhappy with the fact that for every single dependency they need to use separate semaphores due to their binary nature. >> Compared to that a synchronization primitive like D3D12 monitored fences enable one of those to be used to track a sequence of operations by simply associating timeline values to the completion of individual operations. This allows them to track the lifetime and usage of resources and the ordered completion of sequences. >> Besides that, they also want to use a single synchronization primitive to be able to handle GPU-to-GPU and GPU-to-CPU dependencies, compared to using semaphores for the former and fences for the latter. >> In addition, compared to legacy semaphores, timeline semaphores are proposed to support wait-before-signal, i.e. allow enqueueing a semaphore wait operation with a wait value that is larger than any of the already enqueued signal values. This seems to be a hard requirement for ISVs. Without UMD-side queue batching, and even UMD-side queue batching doesn’t help the situation when such a semaphore is externally shared with another API. Thus in order to properly support wait-before-signal the KMD implementation has to also be able to support such dependencies. >> " > I was tangetially involved in that wg too, I understand the overall > use-case of vk timelines. I don't understand the exact corner case > here, because I wasn't deeply involved in the details. all details are here: https://gitlab.khronos.org/vulkan/vulkan/merge_requests/2696 -David > -Daniel > >> Btw, we already add test case to igt, and tested by many existing test, like libdrm unit test, igt related test, vulkan cts, and steam games. >> >> -David >>> -----Original Message----- >>> From: Daniel Vetter <daniel@xxxxxxxx> >>> Sent: Wednesday, December 12, 2018 7:15 PM >>> To: Koenig, Christian <Christian.Koenig@xxxxxxx> >>> Cc: Zhou, David(ChunMing) <David1.Zhou@xxxxxxx>; dri-devel <dri- >>> devel@xxxxxxxxxxxxxxxxxxxxx>; amd-gfx list <amd-gfx@xxxxxxxxxxxxxxxxxxxxx>; >>> intel-gfx <intel-gfx@xxxxxxxxxxxxxxxxxxxxx>; Christian König >>> <ckoenig.leichtzumerken@xxxxxxxxx> >>> Subject: Re: [Intel-gfx] [PATCH 03/10] drm/syncobj: add new >>> drm_syncobj_add_point interface v2 >>> >>> On Wed, Dec 12, 2018 at 12:08 PM Koenig, Christian >>> <Christian.Koenig@xxxxxxx> wrote: >>>> Am 12.12.18 um 11:49 schrieb Daniel Vetter: >>>>> On Fri, Dec 07, 2018 at 11:54:15PM +0800, Chunming Zhou wrote: >>>>>> From: Christian König <ckoenig.leichtzumerken@xxxxxxxxx> >>>>>> >>>>>> Use the dma_fence_chain object to create a timeline of fence >>>>>> objects instead of just replacing the existing fence. >>>>>> >>>>>> v2: rebase and cleanup >>>>>> >>>>>> Signed-off-by: Christian König <christian.koenig@xxxxxxx> >>>>> Somewhat jumping back into this. Not sure we discussed this already >>>>> or not. I'm a bit unclear on why we have to chain the fences in the >>> timeline: >>>>> - The timeline stuff is modelled after the WDDM2 monitored fences. >>> Which >>>>> really are just u64 counters in memory somewhere (I think could be >>>>> system ram or vram). Because WDDM2 has the memory management >>> entirely >>>>> separated from rendering synchronization it totally allows userspace to >>>>> create loops and deadlocks and everything else nasty using this - the >>>>> memory manager won't deadlock because these monitored fences >>> never leak >>>>> into the buffer manager. And if CS deadlock, gpu reset takes care of the >>>>> mess. >>>>> >>>>> - This has a few consequences, as in they seem to indeed work like a >>>>> memory location: Userspace incrementing out-of-order (because they >>> run >>>>> batches updating the same fence on different engines) is totally fine, >>>>> as is doing anything else "stupid". >>>>> >>>>> - Now on linux we can't allow anything, because we need to make sure >>> that >>>>> deadlocks don't leak into the memory manager. But as long as we block >>>>> until the underlying dma_fence has materialized, nothing userspace can >>>>> do will lead to such a deadlock. Even if userspace ends up submitting >>>>> jobs without enough built-in synchronization, leading to out-of-order >>>>> signalling of fences on that "timeline". And I don't think that would >>>>> pose a problem for us. >>>>> >>>>> Essentially I think we can look at timeline syncobj as a dma_fence >>>>> container indexed through an integer, and there's no need to enforce >>>>> that the timline works like a real dma_fence timeline, with all it's >>>>> guarantees. It's just a pile of (possibly, if userspace is stupid) >>>>> unrelated dma_fences. You could implement the entire thing in >>>>> userspace after all, except for the "we want to share these timeline >>>>> objects between processes" problem. >>>>> >>>>> tldr; I think we can drop the dma_fence_chain complexity completely. >>>>> Or at least I'm not really understanding why it's needed. >>>>> >>>>> Of course that means drivers cannot treat a drm_syncobj timeline as >>>>> a dma_fence timeline. But given the future fences stuff and all >>>>> that, that's already out of the window anyway. >>>>> >>>>> What am I missing? >>>> Good question, since that was exactly my initial idea as well. >>>> >>>> Key point is that our Vulcan guys came back and said that this >>>> wouldn't be sufficient, but I honestly don't fully understand why. >>> Hm, sounds like we really need those testscases (vk cts on top of mesa, igt) >>> so we can talk about the exact corner cases we care about and why. >>> >>> I guess one thing that might happen is that userspace leaves out a number >>> and never sets that fence, relying on the >= semantics of the monitored >>> fence to unblock that thread. E.g. when skipping a frame in one of the >>> auxiliary workloads. For that case we'd need to make sure we don't just wait >>> for the given fence to materialize, but also any fences later in the timeline. >>> >>> But we can't decide that without understanding the actual use-case that >>> needs to be supported at the other end of the stack, and how all the bits in >>> between should look like. >>> >>> I guess we're back to "uapi design without userspace doesn't make sense" ... >>> >>>> Anyway that's why David came up with using the fence array to wait for >>>> all previously added fences, which I then later on extended into this >>>> chain container. >>>> >>>> I have to admit that it is way more defensive implemented this way. E.g. >>>> there is much fewer things userspace can do wrong. >>>> >>>> The principal idea is that when they mess things up they are always >>>> going to wait more than necessary, but never less. >>> That seems against the spirit of vulkan, which is very much about "you get all >>> the pieces". It also might dig us a hole in the future, if we ever get around to >>> moving towards a WDDM2 style memory management model. For future >>> proving I think it would make sense if we implement the minimal uapi we >>> need for vk timelines, not the strictest guarantees we can get away with >>> (without performance impact) with current drivers. >>> -Daniel >>> >>> >>>> Christian. >>>> >>>>> -Daniel >>>>> >>>>>> --- >>>>>> drivers/gpu/drm/drm_syncobj.c | 37 >>> +++++++++++++++++++++++++++++++++++ >>>>>> include/drm/drm_syncobj.h | 5 +++++ >>>>>> 2 files changed, 42 insertions(+) >>>>>> >>>>>> diff --git a/drivers/gpu/drm/drm_syncobj.c >>>>>> b/drivers/gpu/drm/drm_syncobj.c index e19525af0cce..51f798e2194f >>>>>> 100644 >>>>>> --- a/drivers/gpu/drm/drm_syncobj.c >>>>>> +++ b/drivers/gpu/drm/drm_syncobj.c >>>>>> @@ -122,6 +122,43 @@ static void drm_syncobj_remove_wait(struct >>> drm_syncobj *syncobj, >>>>>> spin_unlock(&syncobj->lock); >>>>>> } >>>>>> >>>>>> +/** >>>>>> + * drm_syncobj_add_point - add new timeline point to the syncobj >>>>>> + * @syncobj: sync object to add timeline point do >>>>>> + * @chain: chain node to use to add the point >>>>>> + * @fence: fence to encapsulate in the chain node >>>>>> + * @point: sequence number to use for the point >>>>>> + * >>>>>> + * Add the chain node as new timeline point to the syncobj. >>>>>> + */ >>>>>> +void drm_syncobj_add_point(struct drm_syncobj *syncobj, >>>>>> + struct dma_fence_chain *chain, >>>>>> + struct dma_fence *fence, >>>>>> + uint64_t point) { >>>>>> + struct syncobj_wait_entry *cur, *tmp; >>>>>> + struct dma_fence *prev; >>>>>> + >>>>>> + dma_fence_get(fence); >>>>>> + >>>>>> + spin_lock(&syncobj->lock); >>>>>> + >>>>>> + prev = rcu_dereference_protected(syncobj->fence, >>>>>> + lockdep_is_held(&syncobj->lock)); >>>>>> + dma_fence_chain_init(chain, prev, fence, point); >>>>>> + rcu_assign_pointer(syncobj->fence, &chain->base); >>>>>> + >>>>>> + list_for_each_entry_safe(cur, tmp, &syncobj->cb_list, node) { >>>>>> + list_del_init(&cur->node); >>>>>> + syncobj_wait_syncobj_func(syncobj, cur); >>>>>> + } >>>>>> + spin_unlock(&syncobj->lock); >>>>>> + >>>>>> + /* Walk the chain once to trigger garbage collection */ >>>>>> + dma_fence_chain_for_each(prev, fence); } >>>>>> +EXPORT_SYMBOL(drm_syncobj_add_point); >>>>>> + >>>>>> /** >>>>>> * drm_syncobj_replace_fence - replace fence in a sync object. >>>>>> * @syncobj: Sync object to replace fence in diff --git >>>>>> a/include/drm/drm_syncobj.h b/include/drm/drm_syncobj.h index >>>>>> 7c6ed845c70d..8acb4ae4f311 100644 >>>>>> --- a/include/drm/drm_syncobj.h >>>>>> +++ b/include/drm/drm_syncobj.h >>>>>> @@ -27,6 +27,7 @@ >>>>>> #define __DRM_SYNCOBJ_H__ >>>>>> >>>>>> #include "linux/dma-fence.h" >>>>>> +#include "linux/dma-fence-chain.h" >>>>>> >>>>>> /** >>>>>> * struct drm_syncobj - sync object. >>>>>> @@ -110,6 +111,10 @@ drm_syncobj_fence_get(struct drm_syncobj >>>>>> *syncobj) >>>>>> >>>>>> struct drm_syncobj *drm_syncobj_find(struct drm_file *file_private, >>>>>> u32 handle); >>>>>> +void drm_syncobj_add_point(struct drm_syncobj *syncobj, >>>>>> + struct dma_fence_chain *chain, >>>>>> + struct dma_fence *fence, >>>>>> + uint64_t point); >>>>>> void drm_syncobj_replace_fence(struct drm_syncobj *syncobj, >>>>>> struct dma_fence *fence); >>>>>> int drm_syncobj_find_fence(struct drm_file *file_private, >>>>>> -- >>>>>> 2.17.1 >>>>>> >>>>>> _______________________________________________ >>>>>> Intel-gfx mailing list >>>>>> Intel-gfx@xxxxxxxxxxxxxxxxxxxxx >>>>>> https://lists.freedesktop.org/mailman/listinfo/intel-gfx >>> >>> -- >>> Daniel Vetter >>> Software Engineer, Intel Corporation >>> +41 (0) 79 365 57 48 - http://blog.ffwll.ch > > _______________________________________________ dri-devel mailing list dri-devel@xxxxxxxxxxxxxxxxxxxxx https://lists.freedesktop.org/mailman/listinfo/dri-devel
