/proc/diskstats information is inaccurate

[peppaJoeng]

Background:

The /proc/diskstats information is assembled for a container by reading the blkio subsystem in the cgroup, in the proc_diskstats_read function.

Problem:

However, the blkio cgroup subsystem can be inaccurate. For instance, if a container operates on the disk /dev/sdc, the host's /proc/diskstats may show a lot of operations on /dev/sdc. But the blkio cgroup may not present this information accurately, as blkio.io_serviced_recursive, blkio.io_service_time_recursive, and other data related to these operations may be empty.

Proposed Solution:

To address this issue, one approach is to read the host's /proc/diskstats to assemble the container's /proc/diskstats.
This can be accomplished by configuring independent disks for each container, such as using /dev/sdc as the data disk of container A and /dev/sdd as the data disk of container B. This practice is also widely adopted in the industry to isolate container resources.
To support this approach, lxcfs may need to isolate /proc/partitions(I don't know why the community doesn't support it so far) and modify the diskstats assembly logic to read the corresponding disk data from the host's /proc/diskstats according to the disk used by the container and reassemble it.
A better approach may be to use an option (--enable-host-diskstats) to switch between the old and new solutions, ensuring compatibility.
Furthermore, given the limited accuracy of blkio cgroups in most cases, it may be redundant to use them to assemble /proc/diskstats.
Thank you for your attention！

[peppaJoeng]

Any ideas @mihalicyn ?

[peppaJoeng]

Can anyone give me some advice? I tried to implement a solution, a simpler method of disk isolation that does not depend on cgroup - only display the data of the disk used by the current container in /proc/diskstats. @stgraber @mihalicyn

Please help to review my solution, #601

[hallyn]

I assume there is a reason why it would not suffice to simply have the container manager bind mount /proc/diskstats from the host into the container after the lxcfs mounts?

[peppaJoeng]

In practice, it is found that the disk data recorded by cgroup is not accurate and incomplete, such as the fields rq_ticks and ios_pgr.
Especially in production environments, we usually configure independent disks for containers to ensure service quality. For example, partitioning is a common operation if a large disk is to be allocated to 2 or more containers. Therefore, we can consider the host's /proc/diskstats data.