I/O to file classified as metadata with xfs filesystem on kernel 6.9 and newer

[katlapinka]

Description

I have IO class config with classification by file size:

IO class id,IO class name,Eviction priority,Allocation
0,unclassified,255,1.00
1,file_size:eq:1122304,1,1.00
2,file_size:lt:1122304,1,1.00
3,file_size:gt:1122304,1,1.00
4,file_size:le:561152,1,1.00
5,file_size:ge:2244608,1,1.00
6,metadata,1,0.00

Writes or reads from files created on exported object (file sizes are related to those from config file) appears to be classified as metadata writes/reads. It happens only when block size is set to file size. Tested on few different kernels:

• kernel 5.14.0-533.el9.x86_64 -> everything is fine; I/O is classified correctly to corresponding IO classes
• kernel 6.9.7-1.el9.elrepo.x86_64 -> I/O to files is classified as metadata
• kernel 6.13.0-1.el8.elrepo.x86_64 -> I/O to files is classified as metadata

This bug occurs only with use of xfs filesystem. With ext3/ext4 everything worked fine.

Expected Behavior

I/O is classified by file size

Actual Behavior

I/O is classified as metadata

Steps to Reproduce

1. Start cache in wb mode
2. Disable udev
3. Set seq cutoff to never and cleaning policy to nop
4. Add core
5. Create xfs filesystem on core
6. Mount core
7. Create files with sizes correlated to those from IO class config
8. sync, drop caches, purge cache, reset counters
9. Write to file using dd with bs=filesize and count=1 (or simply cat a file)
10. Print io class statistics

Your Environment

• OpenCAS version (commit hash or tag): 7ee78ac
• Operating System: CentOS 8/9
• Cache configuration:
• • Cache mode: wb
• • Cache line size: (default: 4)
• • Promotion policy: (default: always)
• • Cleaning policy: nop
• • Sequential cutoff policy: never
• Other (e.g. lsblk, casadm -P, casadm -L)