In FreeBSD we create a struct bio object for each I/O request that goes to a disk. So, we create bio-s for all vdev zio-s that are to be passed down including read, write, ioctl (disk cache flush) zio-s. Previously, we freed those bio-s in a callback from the I/O subsystem. But since the ABD change we had to move that to the io_done stage, because the callback's context is too restrictive for abd_return_buf / abd_return_buf_copy (at least for the current ABD implementation which is taken as is from illumos). So, currently we are leaking bio-s created for ioctl zio-s because the io_done stage is not called for them.
There is probably another way to fix this problem as the ioctl zio-s have nothing to do with ABD. So, we could free their bio-s in the callback as we did before and handle the regular i/o bio-s in the io_done stage as we do now. This should work, but at the cost of the less uniform bio handling code. So, I wanted to see which of the solutions would be better from the point of view of the general zio pipeline architecture. Thank you. On 14/09/2017 18:48, George Wilson wrote: > Andriy, > > The ZIO_STAGE_VDEV_IO_DONE is not necessary for ZIO_IOCTL_PIPELINE and that's > why it was removed. At least in illumos, there is no functional reason for > calling it. To add it back would require a small amount of change which should > not be a big deal. Can you provide some context around the FreeBSD bug? > > Thanks, > George > > > On Thu, Sep 14, 2017 at 12:45 AM Andriy Gapon <a...@freebsd.org > <mailto:a...@freebsd.org>> wrote: > > > Does anyone know why ZIO_IOCTL_PIPELINE does not include > ZIO_STAGE_VDEV_IO_DONE? > Or, in other words, why ZIO_IOCTL_PIPELINE actively excludes it by not > using > ZIO_VDEV_IO_STAGES? > > It seems that ZIO_IOCTL_PIPELINE had ZIO_VDEV_IO_STAGES until this commit: > > commit e14bb3258d05c1b1077e2db7cf77088924e56919 > > Author: Jeff Bonwick <jeff.bonw...@sun.com> > > Date: Mon Sep 29 18:13:58 2008 -0700 > > 6754011 SPA 3.0: lock breakup, i/o pipeline refactoring, device > failure > handling > > 6667208 zfs/zpool commands on failed pool should not hang > > 6430480 grabbing config lock as writer during I/O load can take > excessively long > > Of course, the commit message does not have any detailed explanations and the > referenced bugs are not publicly accessible. And it was almost 9 years ago. > > I wonder if the change was because of anything specific to illumos vdev_disk. > I think that it would be totally okay to enable ZIO_STAGE_VDEV_IO_DONE with > FreeBSD vdev_geom. In fact, right now ZFS/FreeBSD has a bug because the done > stage is not executed. > > -- > Andriy Gapon > > *illumos-zfs* | Archives > <https://illumos.topicbox.com/groups/zfs/discussions/T7be335e1da154b96-M0cc0a74989d1c00edb4b391a> > | Powered by Topicbox <https://topicbox.com> -- Andriy Gapon ------------------------------------------ openzfs-developer Archives: https://openzfs.topicbox.com/groups/developer/discussions/T4b8a967c63e22769-Meab287e30d2674efda496750 Powered by Topicbox: https://topicbox.com
