[LU-17143] Use migrate_copy_data() for all data movement

• Type: Improvement
• Resolution: Unresolved
• Priority: Minor
• Severity: 3
• Assignee: Feng Lei
• Reporter: Andreas Dilger

Description

The "lfs mirror resync" code is using its own implementation of data copying instead of migrate_copy_data(), which means that bug fixes and improvements to the data copying code (eg. bandwidth limit, stats reporting, buffer sizing, DIO usage, etc.) need to be implemented in two places.

It would be better to have a single function to copy file data shared between the different operations.

Issue Links

• is related to: LU-9961 FLR2: Relocating individual objects to a new OST
• is related to: LU-16789 "lfs migrate" to use AIO/DIO or io_uring (kernel 5.1+)

Comments

[Feng Lei]

Not able to optimize 'lfs mirror resync'.

'lfs mirror resync' command reads data from the default mirror, and writes to (multiple) stale mirror(s). The fd is protected with a lease lock to detect whether the file is changed during resyncing. This means only 1 fd can be written to during the resyncing, or the lease lock will be lost. So:

• We cannot write to multiple mirrors by open the file for multiple times and set different mirror id for them.
• If we want to read data once and write to multiple stale mirrors by set/clear mirror id, it is hard to control the set/clear action precisely for each request because many requests are sharing the same fd. But if we control the mirror id for each batch of requests, I'm not confident that the performance can be improved.
• Another way is to resync one stale mirror as a whole, then resync another stale mirror as a whole. For 2-mirror file it works well. But for N-mirrors file, we may need to read the file for (N-1) times. Hard to say this is an optimization.

[Patrick Farrell]

migrate_copy_data is by far the most complex of the various copy functions, so we should start there, though it doesn't have any parallelism in it.