On 2023/6/29 13:28, Christoph Hellwig wrote: > On Wed, Jun 28, 2023 at 08:45:44PM +0800, chengming.zhou@xxxxxxxxx wrote: >> After these cleanup, __blk_mq_alloc_requests() is the only entry to >> alloc and init rq. > > I find the code a little hard to follow now, due to the optional > setting of the ctx. We also introduce really odd behavior here > if the caller for a hctx-specific allocation doesn't have free > tags, as we'll now run into the normal retry path. > > Is this really needed for your timestamp changes? If not I'd prefer > to skip it. > Thanks for your review! Since hctx-specific allocation path always has BLK_MQ_REQ_NOWAIT flag, it won't retry. But I agree, this makes the general __blk_mq_alloc_requests() more complex. The reason is blk_mq_rq_ctx_init() has some data->rq_flags initialization: ``` if (data->flags & BLK_MQ_REQ_PM) data->rq_flags |= RQF_PM; if (blk_queue_io_stat(q)) data->rq_flags |= RQF_IO_STAT; rq->rq_flags = data->rq_flags; ``` Because we need this data->rq_flags to tell if we need start_time_ns, we need to put these initialization in the callers of blk_mq_rq_ctx_init(). Now we basically have two callers, the 1st is general __blk_mq_alloc_requests(), the 2nd is the special blk_mq_alloc_request_hctx(). So I change the 2nd caller to reuse the 1st __blk_mq_alloc_requests(). Or we put these data->rq_flags initialization in blk_mq_alloc_request_hctx() too? Thanks.
