View Issue Details
| ID | Project | Category | View Status | Date Submitted | Last Update |
|---|---|---|---|---|---|
| 0000905 | channel: kernel/el7 | kernel-ml | public | 2019-03-05 00:31 | 2019-03-06 12:15 |
| Reporter | BS86 | Assigned To | burakkucat | ||
| Priority | normal | Severity | major | Reproducibility | always |
| Status | resolved | Resolution | fixed | ||
| Summary | 0000905: No schedulers present due to upstream drop of single queue schedulers | ||||
| Description | Kernel 5.0 dropped the single queue schedulers and the whole queue. In the past on kernel-ml and kernel-lt, forcing multiqueue scheduling with scsi_mod.use_blk_mq=y had no effect. On kernel-ml with 5.0, the effect is that now no schedulers are present at all. It seems like blk_mq or multiqueue schedulers are dropped during build or in the config. They should now be enabled. | ||||
| Additional Information | cat /sys/block/*/queue/scheduler none none [none] [none] [none] | ||||
| Tags | No tags attached. | ||||
|
|
Addition: cat /sys/block/*/queue/scheduler none #dm-0 none #dm-1 [none] #sda [none] #sdb [none] #sr0 |
|
|
Acknowledging your report. A diff of the 5.0.0 and the 4.20.13 configuration files, greping for SCHED, shows the following -- [kernel-ml-config]$ diff config-5.0.0-x86_64 config-4.20.13-x86_64 | grep SCHED > CONFIG_IOSCHED_NOOP=y > CONFIG_IOSCHED_DEADLINE=y > CONFIG_IOSCHED_CFQ=y > CONFIG_CFQ_GROUP_IOSCHED=y > CONFIG_DEFAULT_IOSCHED="deadline" [kernel-ml-config]$ That is quite crazy. I have no idea why those options were disabled. |
|
|
Those are the single-queue schedulers that are dropped with 5.0 The correct config paramters are: cat config.x86_64 |grep -i mq CONFIG_BLK_WBT_MQ=y CONFIG_BLK_MQ_PCI=y CONFIG_BLK_MQ_VIRTIO=y CONFIG_BLK_MQ_RDMA=y CONFIG_MQ_IOSCHED_DEADLINE=y CONFIG_MQ_IOSCHED_KYBER=y |
|
|
and not to forget: CONFIG_IOSCHED_BFQ=y CONFIG_BFQ_GROUP_IOSCHED=y For BFQ, the single-queue scheduler was: CONFIG_IOSCHED_BFQ_SQ=y CONFIG_BFQ_SQ_GROUP_IOSCHED=y |
|
|
@BS86 CONFIG_MQ_IOSCHED_KYBER and CONFIG_IOSCHED_BFQ can be =m. I suppose no problem making them modules? |
|
|
I don't think that there is a problem with module-ing them. It just narrows down the choice to None and MQ_deadline. |
|
|
An interesting article about I/O scheduler: https://www.linaro.org/blog/io-bandwidth-management-for-production-quality-services/ (not that this is directly related to the current bug report) |
|
|
We now have two configuration files based upon your request. The first has =y for each of the relevant options -- CONFIG_BLK_WBT_MQ=y CONFIG_BLK_MQ_PCI=y CONFIG_BLK_MQ_VIRTIO=y CONFIG_BLK_MQ_RDMA=y CONFIG_MQ_IOSCHED_DEADLINE=y CONFIG_MQ_IOSCHED_KYBER=y CONFIG_IOSCHED_BFQ=y CONFIG_BFQ_GROUP_IOSCHED=y The second is identical, with the exception that the three schedulers are now =m -- CONFIG_BLK_WBT_MQ=y CONFIG_BLK_MQ_PCI=y CONFIG_BLK_MQ_VIRTIO=y CONFIG_BLK_MQ_RDMA=y CONFIG_MQ_IOSCHED_DEADLINE=m CONFIG_MQ_IOSCHED_KYBER=m CONFIG_IOSCHED_BFQ=m CONFIG_BFQ_GROUP_IOSCHED=y Of the two, I would prefer the latter, as it does not increase the size of the core kernel image. However, I would like to know which of the above you would prefer. Please let me know and then I will build a new release. |
|
|
BFQ has significant better performance on rotational disks for desktop usage, so I would go with the first one. Personally, I use bfq for all rotational and mq_deadline for all non-rotational (even on nvme disks), but loading them during boot is also ok. I don't think that the growth in size is that big with them added. I want to add that I am not 100% sure if the list of options is complete to fix the problem. I am still in the process of learning my stuff with kernels. |
|
|
A kernel-ml-5.0.0-2.el7.elrepo package set has now been built and is currently synchronising across our mirror sites. kernel-ml-5.0.0-2.el7.elrepo.nosrc.rpm kernel-ml-5.0.0-2.el7.elrepo.x86_64.rpm kernel-ml-devel-5.0.0-2.el7.elrepo.x86_64.rpm kernel-ml-doc-5.0.0-2.el7.elrepo.noarch.rpm kernel-ml-headers-5.0.0-2.el7.elrepo.x86_64.rpm kernel-ml-tools-5.0.0-2.el7.elrepo.x86_64.rpm kernel-ml-tools-libs-5.0.0-2.el7.elrepo.x86_64.rpm kernel-ml-tools-libs-devel-5.0.0-2.el7.elrepo.x86_64.rpm perf-5.0.0-2.el7.elrepo.x86_64.rpm python-perf-5.0.0-2.el7.elrepo.x86_64.rpm [Build64R7 uploads]$ rpm -qp --changelog *.nosrc.rpm | head -n 6 * Tue Mar 05 2019 Alan Bartlett <ajb@elrepo.org> - 5.0.0-2 - CONFIG_BLK_WBT=y, CONFIG_BLK_WBT_MQ=y, - CONFIG_MQ_IOSCHED_DEADLINE=y, CONFIG_MQ_IOSCHED_KYBER=y, - CONFIG_IOSCHED_BFQ=y and CONFIG_BFQ_GROUP_IOSCHED=y - [https://elrepo.org/bugs/view.php?id=905] [Build64R7 uploads]$ |
|
|
Thanks. It works now: cat /sys/block/*/queue/scheduler none #dm-0 none #dm-1 [mq-deadline] kyber bfq none #sda mq-deadline kyber [bfq] none #sdb [mq-deadline] kyber bfq none #sr0 |
|
|
Thank you for confirming the success. Now closing as "resolved/fixed". |
| Date Modified | Username | Field | Change |
|---|---|---|---|
| 2019-03-05 00:31 | BS86 | New Issue | |
| 2019-03-05 00:31 | BS86 | Status | new => assigned |
| 2019-03-05 00:31 | BS86 | Assigned To | => burakkucat |
| 2019-03-05 00:37 | BS86 | Note Added: 0006238 | |
| 2019-03-05 10:40 | burakkucat | Note Added: 0006239 | |
| 2019-03-05 10:48 | BS86 | Note Added: 0006240 | |
| 2019-03-05 10:50 | BS86 | Note Added: 0006241 | |
| 2019-03-05 11:18 | toracat | Note Added: 0006242 | |
| 2019-03-05 11:47 | BS86 | Note Added: 0006243 | |
| 2019-03-05 12:04 | toracat | Note Added: 0006245 | |
| 2019-03-05 13:32 | burakkucat | Note Added: 0006248 | |
| 2019-03-05 14:13 | BS86 | Note Added: 0006249 | |
| 2019-03-05 14:56 | burakkucat | Note Added: 0006255 | |
| 2019-03-05 23:57 | BS86 | Note Added: 0006258 | |
| 2019-03-06 12:15 | burakkucat | Note Added: 0006259 | |
| 2019-03-06 12:15 | burakkucat | Status | assigned => resolved |
| 2019-03-06 12:15 | burakkucat | Resolution | open => fixed |
