On Tue, Mar 21, 2023 at 11:44:59AM -0700, Bart Van Assche wrote: > On 3/20/23 23:05, Christoph Hellwig wrote: >> In fact I wonder how you managed to get into __bio_split_to_limits >> wtih a NULL current->bio_list at all. > > Hi Christoph, > > This patch series is what I came up with after having observed an UNALIGNED > WRITE COMMAND response with a 5.15 kernel. In that kernel version (but > probably not in the latest upstream kernel) the device mapper core submits > split bios in the wrong order. I will check again whether this patch is > really necessary. How did that escape the upper layer serialization of zoned writes? But my point is that all ->submit_bio instances as well as blk_mq_submit_bio are only called with a non-NULL current->bio_list. So I don't think you can reach the NULL current->bio_list case in this patch. Note that even without any specific zone device isss, submitting bios in LBA oders works best for both spinning and solid state media, but all of the arguments here don't add up.
