On 1/14/20 3:16 AM, Linus Torvalds wrote: > On Mon, Jan 13, 2020 at 4:29 PM Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> wrote: >> >> From: Vlastimil Babka <vbabka@xxxxxxx> >> Subject: mm, thp: tweak reclaim/compaction effort of local-only and all-node allocations > > I absolutely _detest_ how we've had this pattern of trying to change > the THP logic in a late -rc game. > > Considering how unsuccessful some of the earlier attempts have been, > late -rc kernels really are *not* the time to make changes like this. Agreed, myself would have been much more comfortable in rc1. > I'm going to take it this time, but Thanks. > (a) I'm going to revert *immediately* if anybody even peeps about the > new behavior > > (b) I want this to stop. Andrew - send THP "tweaks" during the merge > window, and not like some late-rc "fixes", when there are absolutely > _zero_ performance numbers or anything else critical associated with > the patch. > > IOW, there is a decided lack of "bugfix" here. And if there actually > _were_ performance numbers, they should have been in the explanation. FWIW Mel did some tests [1] which showed some improvement to THP success rate and no significant extra swapping that was feared. We didn't get confirmation from Andrea about his real life workloads though. [1] https://lore.kernel.org/linux-mm/20191113112042.GG28938@xxxxxxx/ > So stop randomly messing with the THP allocation code, since it's > _known_ to be fragile with subtle side effects. > > Linus >
