Here is the an example to describe existing issue:
There is a Write request with 64KiB data length, and this IO is cross the IO
boundary. We assume that the parent IO will have 2 children requests, one is
33KiB length, the other one is 31KiB. Here is the view of parent iovs, the
first 33KiB length data has 33 iovs:
iov.[0].iov_length = 1024;
.
.
iov.[31].iov_length = 256;
iov.[32].iov_length = 768;
.
.
iov.[64].iov_length = 1024;
In function _spdk_bdev_io_split(), then you can see that for the 33KiB length
child request, exiting code will run out of child child_iov space and return
error due to last one data buffer is not block size aligned.
Here we can rewind the existing offset to last block size aligned buffer to
avoid the error case, for backend which need aligned data buffer such as
AIO backend, the request will go through spdk_bdev_io_get_buf() again to
do the data copy, otherwise for those backend devices such as NVMe with
hardware SGL support, 256 data segment is fine for them.
Change-Id: I96ebdf29829d86f9b38fab28a7406eedc9fa44ef
Signed-off-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/453604
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com>
efd7b514d4