> On 8 May 2017, at 16.52, Jens Axboe wrote: > > On 05/08/2017 08:46 AM, Javier González wrote: >>> On 8 May 2017, at 16.23, Jens Axboe wrote: >>> >>> On 05/08/2017 08:20 AM, Javier González wrote: >>>>> On 8 May 2017, at 16.13, Jens Axboe wrote: >>>>> >>>>> On 05/08/2017 07:44 AM, Javier González wrote: >>>>>>> On 8 May 2017, at 14.27, Ming Lei wrote: >>>>>>> >>>>>>> On Mon, May 08, 2017 at 01:54:58PM +0200, Javier González wrote: >>>>>>>> Hi, >>>>>>>> >>>>>>>> I find an unusual added latency(~20-30ms) on blk_queue_enter when >>>>>>>> allocating a request directly from the NVMe driver through >>>>>>>> nvme_alloc_request. I could use some help confirming that this is a bug >>>>>>>> and not an expected side effect due to something else. >>>>>>>> >>>>>>>> I can reproduce this latency consistently on LightNVM when mixing I/O >>>>>>>> from pblk and I/O sent through an ioctl using liblightnvm, but I don't >>>>>>>> see anything on the LightNVM side that could impact the request >>>>>>>> allocation. >>>>>>>> >>>>>>>> When I have a 100% read workload sent from pblk, the max. latency is >>>>>>>> constant throughout several runs at ~80us (which is normal for the media >>>>>>>> we are using at bs=4k, qd=1). All pblk I/Os reach the nvme_nvm_submit_io >>>>>>>> function on lightnvm.c., which uses nvme_alloc_request. When we send a >>>>>>>> command from user space through an ioctl, then the max latency goes up >>>>>>>> to ~20-30ms. This happens independently from the actual command >>>>>>>> (IN/OUT). I tracked down the added latency down to the call >>>>>>>> percpu_ref_tryget_live in blk_queue_enter. Seems that the queue >>>>>>>> reference counter is not released as it should through blk_queue_exit in >>>>>>>> blk_mq_alloc_request. For reference, all ioctl I/Os reach the >>>>>>>> nvme_nvm_submit_user_cmd on lightnvm.c >>>>>>>> >>>>>>>> Do you have any idea about why this might happen? I can dig more into >>>>>>>> it, but first I wanted to make sure that I am not missing any obvious >>>>>>>> assumption, which would explain the reference counter to be held for a >>>>>>>> longer time. >>>>>>> >>>>>>> You need to check if the .q_usage_counter is working at atomic mode. >>>>>>> This counter is initialized as atomic mode, and finally switchs to >>>>>>> percpu mode via percpu_ref_switch_to_percpu() in blk_register_queue(). >>>>>> >>>>>> Thanks for commenting Ming. >>>>>> >>>>>> The .q_usage_counter is not working on atomic mode. The queue is >>>>>> initialized normally through blk_register_queue() and the counter is >>>>>> switched to percpu mode, as you mentioned. As I understand it, this is >>>>>> how it should be, right? >>>>> >>>>> That is how it should be, yes. You're not running with any heavy >>>>> debugging options, like lockdep or anything like that? >>>> >>>> No lockdep, KASAN, kmemleak or any of the other usual suspects. >>>> >>>> What's interesting is that it only happens when one of the I/Os comes >>>> from user space through the ioctl. If I have several pblk instances on >>>> the same device (which would end up allocating a new request in >>>> parallel, potentially on the same core), the latency spike does not >>>> trigger. >>>> >>>> I also tried to bind the read thread and the liblightnvm thread issuing >>>> the ioctl to different cores, but it does not help... >>> >>> How do I reproduce this? Off the top of my head, and looking at the code, >>> I have no idea what is going on here. >> >> Using LightNVM and liblightnvm [1] you can reproduce it by: >> >> 1. Instantiate a pblk instance on the first channel (luns 0 - 7): >> sudo nvme lnvm create -d nvme0n1 -n test0 -t pblk -b 0 -e 7 -f >> 2. Write 5GB to the test0 block device with a normal fio script >> 3. Read 5GB to verify that latencies are good (max. ~80-90us at bs=4k, qd=1) >> 4. Re-run 3. and in parallel send a command through liblightnvm to a >> different channel. A simple command is an erase (erase block 900 on >> channel 2, lun 0): >> sudo nvm_vblk line_erase /dev/nvme0n1 2 2 0 0 900 >> >> After 4. you should see a ~25-30ms latency on the read workload. >> >> I tried to reproduce the ioctl in a more generic way to reach >> __nvme_submit_user_cmd(), but SPDK steals the whole device. Also, qemu >> is not reliable for this kind of performance testing. >> >> If you have a suggestion on how I can mix an ioctl with normal block I/O >> read on a standard NVMe device, I'm happy to try it and see if I can >> reproduce the issue. > > Just to rule out this being any hardware related delays in processing > IO: > > 1) Does it reproduce with a simpler command, anything close to a no-op > that you can test? Yes. I tried with a 4KB read and with a fake command I drop right after allocation. > 2) What did you use to time the stall being blk_queue_enter()? > I have some debug code measuring time with ktime_get() in different places in the stack, and among other places, around blk_queue_enter(). I use them then to measure max latency and expose it through sysfs. I can see that the latency peak is recorded in the probe before blk_queue_enter() and not in the one after. I also did an experiment, where the normal I/O path allocates the request with BLK_MQ_REQ_NOWAIT. When running the experiment above, the read test fails since we reach: if (nowait) return -EBUSY; in blk_queue_enter. Javier