On Sat, May 04, 2019 at 11:46:35AM +0200, Nicholas Mc Guire wrote: > On Tue, Apr 30, 2019 at 04:49:44PM +0300, Sakari Ailus wrote: > > Hi Nicholas, > > > > On Sun, Apr 07, 2019 at 04:16:02AM +0200, Nicholas Mc Guire wrote: > > > Allow the hrtimer subsystem to coalesce delay timers of lower accuracy > > > by providing a suitable range > > > > > > Signed-off-by: Nicholas Mc Guire <hofrat@xxxxxxxxxxx> > > > --- > > > > > > Problem located by an experimental coccinelle script > > > > > > hrtimers in atomic context have limited accuracy due to possible > > > context-switching and interruption so the accuracy is limited > > > anyway. Giving the hrtimer subsystem a reasonable range for merging > > > hrtimers helps to reduce the load on the hrtimer subsystem. As this > > > delays do not seem to mandate high accuracy the range of a factor > > > two seems acceptable. > > > > > > Patch was compile tested with: x86_64_defconfig + MEDIA_SUPPORT=m, > > > MEDIA_CAMERA_SUPPORT=y, MEDIA_CONTROLLER=y, VIDEO_V4L2_SUBDEV_API=y, > > > VIDEO_SMIAPP=m > > > (with a number of sparse warnings on sizeof() usage) > > > > > > Patch is against 5.1-rc3 (localversion-next is next-20190405) > > > > The delays are exact for the reason that they are user-visible delays in > > image capturing related use cases. Sometimes milliseconds matter, or at > > least adding more does not help. > > > > Actually it can be better iwith respect to jitter to let the hrtimer > subsystem use an existing timer event than to have a close by second event > and the accuracy is determined by the non-atomic context anyway - > so while the proposed delay extension might be excessive in your case > I would still suggest to try to get away from a range of 0 - even if > you only end up with (1000,1050) that would be an advantage for the > timer subsystem. Sounds reasonable to me. Would you like to send v2? Thanks. -- Sakari Ailus
