When it comes to head motion most of the standard fMRI papers refer to overall motion/displacement, measured from the first to the last volume. The criterion is typically ~3 - 5 mm or in other words, motion within the size of a voxel. Unfortunately this doesn't tell us much about data quality, as subjects might constantly be shaking their heads/shivering, which often results in rather drastic artefacts. This is probably more problematic than the overall motion (which is going to affect the data quality to some extent as well of course, reduced SNR, losing voxels from analysis if subjects moved out of plane, geometric deformations due to inhomogenous magnetic field).

Artrepair plots estimated motion between successive volumes ("fast motion", scan-to-scan motion). The default criterion is 0.5 mm/TR. Similar thresholds or even more conservative ones like 0.3 mm/TR have frequently been employed in the context of resting state fMRI for the last couple of years. So in your case, as the run is short and the number of affected volumes large, the run might have to be rejected. There are different opionions on how to treat data sets like that, some of them try to replace the "bad" volumes with a mean or some interpolated data, others add regressors to the design matrix, others discard these subjects entirely. In any case you should check whether motion is related to some other variable, as you might introduce some bias (for example larger head motion in patients).