On Wed 18-07-18 12:17:24, Waiman Long wrote: > On 07/16/2018 05:09 AM, Michal Hocko wrote: > > On Fri 13-07-18 10:36:14, Dave Chinner wrote: > > [...] > >> By limiting the number of negative dentries in this case, internal > >> slab fragmentation is reduced such that reclaim cost never gets out > >> of control. While it appears to "fix" the symptoms, it doesn't > >> address the underlying problem. It is a partial solution at best but > >> at worst it's another opaque knob that nobody knows how or when to > >> tune. > > Would it help to put all the negative dentries into its own slab cache? > > > >> Very few microbenchmarks expose this internal slab fragmentation > >> problem because they either don't run long enough, don't create > >> memory pressure, or don't have access patterns that mix long and > >> short term slab objects together in a way that causes slab > >> fragmentation. Run some cold cache directory traversals (git > >> status?) at the same time you are creating negative dentries so you > >> create pinned partial pages in the slab cache and see how the > >> behaviour changes.... > > Agreed! Slab fragmentation is a real problem we are seeing for quite > > some time. We should try to address it rather than paper over it with > > weird knobs. > > I am aware that you don't like the limit knob that control how many > negative dentries are allowed as a percentage of total system memory. I > got comments in the past about doing some kind of auto-tuning. How about > consolidating the 2 knobs that I currently have in the patchset into a > single one with 3 possible values, like: > > 0 - no limiting > 1 - set soft limit to "a constant + 4 x max # of positive dentries" and > warn if exceeded > 2 - same limit but kill excess negative dentries after use. > > Does that kind of knob make more sense to you? Not really. See the pagecache limit story in http://lkml.kernel.org/r/20180719084538.GP7193@xxxxxxxxxxxxxx I might be overly sensitive but I got burnt a lot in the past. We should strive to make the reclaim seamless without asking admins to do our job. -- Michal Hocko SUSE Labs -- To unsubscribe from this list: send the line "unsubscribe linux-doc" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
