On Mon, Dec 17, 2018 at 11:57:25AM +0100, Christoph Hellwig wrote: > On Mon, Dec 17, 2018 at 06:42:46PM +0800, Ming Lei wrote: > > From: Christoph Hellwig <hch@xxxxxx> > > > > Even though poll_queues are zero, nvme's mapping for HCTX_TYPE_POLL still > > may be setup via blk_mq_map_queues() which cause different mapping compared > > with HCTX_TYPE_DEFAULT's mapping built from managed irq affinity. > > > > This mapping will cause hctx->type to be over-written in blk_mq_map_swqueue(), > > then the whole mapping may become broken, for example, one same ctx can be > > mapped to different hctxs with same hctx type. This bad mapping has caused > > IO hang in simple dd test, as reported by Mike. > > > > This patch sets map->nr_queues as zero explictly if there is zero > > queues for such queue type, also maps to correct hctx if .nr_queues of the > > queue type is zero. > > > > Cc: Jeff Moyer <jmoyer@xxxxxxxxxx> > > Cc: Mike Snitzer <snitzer@xxxxxxxxxx> > > Cc: Christoph Hellwig <hch@xxxxxx> > > > > (don't handle zero .nr_queues map in blk_mq_map_swqueue()) > > Signed-off-by: Ming Lei <ming.lei@xxxxxxxxxx> > > I actually think we should split this into three patches: > > 1) skip zero-queue maps in blk_mq_map_swqueue > 2) skip zero-queue maps in blk_mq_map_queue > 3) don't duplicate maps in nvme > > sorry for misleading you with my quick WIP patch. > > I can send send patches for my two with a proper changelogs, but OK, I just want to fix the issue quick, :-) > I'll leave blk_mq_map_swqueue to you as I don't full understand > the rationale, even if it looks sensible to me. If maps won't be duplicated, skip zero-queue maps in blk_mq_map_swqueue() may not be necessary. However, it may make the code more robust/simple. Thanks, Ming
