Re: [PATCH] ceph: parallelize object copy in ceph_do_objects_copy()

Hi Xiubo,

Thanks for the patch, the idea is sound but there are a few correctness
issues that need fixing:

The drain loop walks past the valid window on error — it dereferences
request pointers beyond the tail of the in-flight window, which are
either freed or uninitialized depending on where the failure occurred.

The window size is read from copyfrom_max_inflight on every iteration.
A concurrent sysfs write can change it mid-flight, causing the loop
bounds to shift while requests are already in-flight.

If a request allocation fails partway through, the error return masks
any I/O errors that were already recorded for earlier in-flight
requests.

There's also a design issue for mid-file copies: parallel writes can
modify the destination beyond the position implied by a short source
count before the short is detected.

Could you respin with these fixed?

--
Alex Markuze