Hi On Wed, Jun 26, 2013 at 2:11 PM, Tejun Heo <tj@xxxxxxxxxx> wrote: > On Wed, Jun 26, 2013 at 2:08 PM, Vivek Goyal <vgoyal@xxxxxxxxxx> wrote: >> I don't have an objection to not listing stats of devices which are >> zero, but wondering why all the devices of system are showing in >> cgroup stat. >> >> Don't we add a blkg to a blkcg only if that cgroup did some IO to >> that particular device. If yes, then only those devices should >> show to which cgroup did some IO and that should be non-zero. >> >> Is it because of hierarchical support where if a child does IO >> to device, we will add an blkg instance to parent's cgroup? In >> that case hierarchical stats will still be non-zero but group >> local stat can be zero. > > I wondered that too. Maybe they're configuring all combinations? Anatol? We use an io scheduler similar to cfq. So I added a log to find place where the blkg is created. I found that the structure is created on data write to CGROUP/weight_device CFQ file. Now let's look at cfqg_set_weight_device() function (I am looking at 3.5 sources but seems HEAD has the same issue). When we set weight for a device in that group then an instance of cfq_group is created for it. Here is the codepath from write() syscall to cfq_pd_init() function: [ 119.584630] [<ffffffff805001ba>] cfq_pd_init+0x10a/0x110 [ 119.584634] [<ffffffff804fce00>] blkg_alloc+0x110/0x130 [ 119.584639] [<ffffffff804fd098>] __blkg_lookup_create+0x278/0x3c0 [ 119.584643] [<ffffffff804fd1fb>] blkg_lookup_create+0x1b/0x40 [ 119.584647] [<ffffffff804fd433>] blkg_conf_prep+0x213/0x270 [ 119.584669] [<ffffffff805008ca>] cfqg_set_weight_device+0x4a/0xd0 [ 119.584678] [<ffffffff802c1118>] cgroup_write_string.isra.17+0xc8/0x130 [ 119.584687] [<ffffffff802c12d9>] cgroup_file_write+0x159/0x1e0 [ 119.584707] [<ffffffff8038aa4f>] vfs_write+0xaf/0x160 [ 119.584711] [<ffffffff8038add6>] sys_write+0x76/0x100 [ 119.584716] [<ffffffff807da672>] system_call_fastpath+0x16/0x1b Bottom line: when we set weight for a device then it creates cfq_group and initializes device stats with zero. Later that zero stat is shown even if no activity happened on the device. Both CFQ and our custom IO scheduler have this problem. And because we write set weight for all devices in all cgroups we endup with a lot of zero useless stats. What is the right fix in this situation? -- To unsubscribe from this list: send the line "unsubscribe cgroups" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
