On Tue, Mar 10, 2026 at 11:57 AM Caleb Sander Mateos <[email protected]> wrote: > > On Mon, Mar 9, 2026 at 4:55 PM Sungwoo Kim <[email protected]> wrote: > > > > On Mon, Mar 9, 2026 at 11:31 AM Caleb Sander Mateos > > <[email protected]> wrote: > > > > > > On Sun, Mar 8, 2026 at 11:30 PM Sungwoo Kim <[email protected]> wrote: > > > > > > > > The numa_node can be < 0 since NUMA_NO_NODE = -1. However, > > > > struct blk_mq_hw_ctx{} defines numa_node as unsigned int. As a result, > > > > numa_node is set to UINT_MAX for NUMA_NO_NODE in blk_mq_alloc_hctx(). > > > > > > The node argument to blk_mq_alloc_hctx() comes from > > > blk_mq_alloc_and_init_hctx(), which is called by > > > blk_mq_alloc_and_init_hctx() with int node = blk_mq_get_hctx_node(set, > > > i). node = NUMA_NO_NODE would suggest that blk_mq_hw_queue_to_node() > > > doesn't find any CPU affinitized to the queue. Is that even possible? > > > > Thanks for your review, Celeb. > > While I'm flattered you consider me a celebrity, my name is Caleb :)
My apologies, Caleb. I'll be more careful. > > > > > blk_mq_hw_queue_to_node() can return NUMA_NO_NODE if the device queues > > exceed the > > number of CPUs. Afterward, it is adjusted on the caller side to > > numa_node = set->numa_node. > > I thought the NVMe driver capped the number of queues so every queue > is affinitized to some CPU (see nvme_max_io_queues()). What am I > missing? > > Best, > Caleb You are right, every hw queue should be affinitized. It seems I'm missing a lot. I'll look into the reasons and get back with V2. I appreciate your review again. Sungwoo.
