On 2023/4/19 08:03, Harris, James R wrote: Hi, Thanks for using the NEED_GET_DATA feature. > Hi, > > I’m working on adding NEED_GET_DATA support to the SPDK ublk server, to avoid allocating I/O buffers until they are actually needed. Yes, with NEED_GET_DATA, you don't have to preallocate I/O buffers for WRITE requests. > > It is very clear how this works for write commands with NEED_GET_DATA. We wait to allocate the buffer until we get UBLK_IO_RES_NEED_GET_DATA completion and submit again using UBLK_IO_NEED_GET_DATA. After we get the UBLK_IO_RES_OK completion from ublk, we submit the block request to the SPDK bdev layer. After it completes, we submit using UBLK_IO_COMMIT_AND_FETCH_REQ and can free the I/O buffer because the data has been committed. Yes, you are right. > > But how does this work for the read path? On a read, I can wait to allocate the buffer until I get the UBLK_IO_RES_OK completion. But after the read operation is completed and SPDK submits the UBLK_IO_COMMIT_AND_FETCH_REQ, how do I know when ublk has finished copying data out of the buffer so that I can reuse that buffer? > > I’m sure I’m missing something obvious, if anyone can provide a pointer I would appreciate it. In the WRITE path, data copy happens before the backend gets UBLK_IO_RES_OK. In the READ path, data copy happens after the backend submits UBLK_IO_COMMIT_AND_FETCH_REQ, but unfortunately we cannot know the time data copy happens because ublk driver does not notify the backend until next request(on this tag) comes. Regards, Zhang > > Thanks, > > Jim Harris >
