Am Freitag, den 13.01.2017, 14:24 +0100 schrieb Vlastimil Babka: > On 01/13/2017 12:51 PM, Lucas Stach wrote: > > Commit 73e64c51afc5 (mm, compaction: allow compaction for GFP_NOFS requests) > > changed compation to skip FS pages if not explicitly allowed to touch them, > > but missed to update the CMA compact_control. > > > > This leads to a very high isolation failure rate, crippling performance of > > CMA even on a lightly loaded system. Re-allow CMA to compact FS pages by > > setting the correct GFP flags, restoring CMA behavior and performance to > > the kernel 4.9 level. > > > > Fixes: 73e64c51afc5 (mm, compaction: allow compaction for GFP_NOFS requests) > > Signed-off-by: Lucas Stach <l.stach@xxxxxxxxxxxxxx> > > Acked-by: Vlastimil Babka <vbabka@xxxxxxx> > > It's true that this restores the behavior for CMA to 4.9. But it also reveals > that CMA always implicitly assumed to be called from non-fs context. That's > expectable for the original CMA use-case of drivers for devices such as cameras, > but I now wonder if there's danger when CMA gets invoked via dma-cma layer with > generic cma range for e.g. a disk device... I guess that would be another > argument for scoped GFP_NOFS, which should then be applied to adjust the > gfp_mask here. Or we could apply at least memalloc_noio_flags() right now, which > should already handle the disk device -> dma -> cma scenario? That's right. But I don't think we need to fix this for 4.10. The minimal fix in this patch brings things back to the old assumptions. As dma allocations already carry proper GFP flags it's just a matter of passing them through to CMA, to make things work correctly. I'll cook up a follow up patch for that, but I think this should wait for the next merge window to be applied. Regards, Lucas -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>