On Wed, Feb 5, 2014 at 9:15 AM, Jens Axboe <axboe@xxxxxxxxx> wrote: > On Mon, Feb 03 2014, Puthikorn Voravootivat wrote: >> In verify phase, the rand_seed generated on replay does not match >> the written rand_seed. >> >> Multiple problems are causing this: >> 1. In verify phase fio does not set io_u->rand_seed to compare with >> hdr->rand_seed >> 2. In randrw scenario, fio log is stored in red-black tree in "sorted by LBA" >> order. Thus, it is imposible to replay the written order, or rather >> generate the seeds again in the same order. >> 3. In write phase, the code currently will generate rand_seed, write data >> and log rand_seed. When queuedepth > 1, it's possible the writes complete >> in a different order than rand_seed was generated. Thus when replaying >> the log, the generated rand_seed might not match what was written. >> 4. verify_backlog option will start verification before all the data has been >> written and it make rand_seed replay code broken with current design. >> >> Proposed fixes: >> 1. Use of existing verify_state to generate verify header. >> (and assumes this was the original intention of verify_state). And also >> adds code to replay rand_seed in verify phase. >> 2. If verifysort option is not enabled, store the write log in a list instead >> of the red-black tree. Otherwise, don't attempt to verify the rand_seed >> in the header. >> 3. In write phase, generate rand_seed, log rand_seed, write data. I.e. log >> IO transactions in the order generated, not completed. >> 4. Don't verify rand_seed when verify_backlog is enabled. > > Seems like a reasonable solution to it all. Thanks! I'll get this > applied. Awesome - I see that it's applied. Puthikorn has one more fix in flight and I'm going to suggest something you'll probably reject in another email. cheers, grant -- To unsubscribe from this list: send the line "unsubscribe fio" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
