> On Thu, Jul 07, 2022 at 11:16:42AM +0200, Sergei Shtepa wrote: >> The module creates a block device for each snapshot image. >> To make a backup of a block device, the backup tool reads snapshot image. >> This snapshot image block device allows to mount a file system on it >> and perform the necessary preparation. If not for this requirement, >> it would be possible to implement reading via an additional ioctl. >> But that wouldn't be a good design, I think. > Ok, got it. It was just me who was confused. > >> Perhaps I have implemented this block device incorrectly? >> Processing requests of the snapshot image block device is started >> in the function snapimage_queue_rq(). And ends in the >> snapimage_queue_work() in another kernel thread. Therefore, when >> the request is initialized in snapimage_init_request(), a kernel worker >> is prepared. > I don't think it is wrong, but there is some potential for optimization. > > Is there a reson this is implemented as a blk-mq driver vs a bio > based driver that just implements ->submit_bio? The latter has > the advantage that you are always called in user context, and don't > need the extra workqueue offload. > > The block layer also generally assumes that blk-mq drivers don't > call submit_bio_noacct underneath, so if we can't I'd much > prefer the bio based driver approach here. > There was a goal to use the kernel innovations whenever possible. Of course, it makes sense to return to bio based if it allows to achieve better performance. This is not a matter of principle.
