From mboxrd@z Thu Jan 1 00:00:00 1970 From: Gustavo Padovan Subject: Re: [RFC 0/5] rework fences on struct sync_file Date: Fri, 24 Jun 2016 12:19:32 -0300 Message-ID: <20160624151932.GG2508@joana> References: <1466695790-2833-1-git-send-email-gustavo@padovan.org> <576CFD0B.6000501@amd.com> <20160624131724.GA2503@joana> <576D4032.3000001@vodafone.de> <20160624145933.GF2508@joana> <576D4D3F.8080808@amd.com> Mime-Version: 1.0 Content-Type: text/plain; charset=iso-8859-1 Content-Transfer-Encoding: QUOTED-PRINTABLE Return-path: Content-Disposition: inline In-Reply-To: <576D4D3F.8080808@amd.com> Sender: linux-kernel-owner@vger.kernel.org To: Christian =?iso-8859-1?Q?K=F6nig?= Cc: Christian =?iso-8859-1?Q?K=F6nig?= , Gustavo Padovan , dri-devel@lists.freedesktop.org, marcheu@google.com, Daniel Stone , seanpaul@google.com, Daniel Vetter , linux-kernel@vger.kernel.org, laurent.pinchart@ideasonboard.com, John Harrison , m.chehab@samsung.com List-Id: dri-devel@lists.freedesktop.org 2016-06-24 Christian K=F6nig : > Am 24.06.2016 um 16:59 schrieb Gustavo Padovan: > > 2016-06-24 Christian K=F6nig : > >=20 > > > Am 24.06.2016 um 15:17 schrieb Gustavo Padovan: > > > > Hi Christian, > > > >=20 > > > > 2016-06-24 Christian K=F6nig : > > > >=20 > > > > > Am 23.06.2016 um 17:29 schrieb Gustavo Padovan: > > > > > > From: Gustavo Padovan > > > > > >=20 > > > > > > Hi all, > > > > > >=20 > > > > > > This is an attempt to improve fence support on Sync File. T= he basic idea > > > > > > is to have only sync_file->fence and store all fences there= , either as > > > > > > normal fences or fence_arrays. That way we can remove some = potential > > > > > > duplication when using fence_array with sync_file: the dupl= ication of the array > > > > > > of fences and the duplication of fence_add_callback() for a= ll fences. > > > > > >=20 > > > > > > Now when creating a new sync_file during the merge process = sync_file_set_fence() > > > > > > will set sync_file->fence based on the number of fences for= that sync_file. If > > > > > > there is more than one fence a fence_array is created. One = important advantage > > > > > > approach is that we only add one fence callback now, no mat= ter how many fences > > > > > > there are in a sync_file - the individual callbacks are add= ed by fence_array. > > > > > >=20 > > > > > > Two fence ops had to be created to help abstract the differ= ence between handling > > > > > > fences and fences_arrays: .teardown() and .get_fences(). Th= e former run needed > > > > > > on fence_array, and the latter just return a copy of all fe= nces in the fence. > > > > > > I'm not so sure about adding those two, speacially .get_fen= ces(). What do you > > > > > > think? > > > > > Clearly not a good idea to add this a fence ops, cause those = are specialized > > > > > functions for only a certain fence implementation (the fence_= array). > > > > Are you refering only to .get_fences()? > > > That comment was only for the get_fences() operation, but the tea= rdown() > > > callback looks very suspicious to me as well. > > >=20 > > > Can you explain once more why that should be necessary? > > When the sync_file owner exits we need to clean up it and that mean= s releasing > > the fence too, however with fence_array we can't just call fence_pu= t() > > as a extra reference to array->base for each fence is held when ena= bling > > signalling. Thus we need a prior step, that I called teardown(), to > > remove the callback for not signaled fences and put the extra > > references. > >=20 > > Another way to do this would be: > >=20 > > if (fence_is_array(sync_file->fence)) > > fence_array_destroy(to_fence_array(sync_file->fence)); > > else > > fence_put(sync_file_fence); > >=20 > > This would avoid the extra ops, maybe we should go this way. >=20 > NAK on both approaches. The fence array grabs another reference on it= self > for each callback it registers, so this isn't necessary: >=20 > > for (i =3D 0; i < array->num_fences; ++i) { > > cb[i].array =3D array; > > /* > > * As we may report that the fence is signaled befo= re all > > * callbacks are complete, we need to take an addit= ional > > * reference count on the array so that we do not f= ree > > it too > > * early. The core fence handling will only hold th= e > > reference > > * until we signal the array as complete (but that = is now > > * insufficient). > > */ > > fence_get(&array->base); > > if (fence_add_callback(array->fences[i], &cb[i].cb, > > fence_array_cb_func)) { > > fence_put(&array->base); > > if (atomic_dec_and_test(&array->num_pending= )) > > return false; > > } > > } >=20 > So you can just use fence_remove_callback() and then fence_put() with= out > worrying about the reference. Yes. That is what I have in mind for fence_array_destroy() in the snippet of code in the last e-mail. That plus the last fence_put() to release the fence_array(). Gustavo