On Wed, Jul 17, 2024 at 12:45 PM Mikulas Patocka <mpatocka@xxxxxxxxxx> wrote:
>
> Hi
>
> I am wondering why does do_resume need to call dm_suspend at all. Does
> anyone here remember why is this code path needed?
In our case, we have a sequence with load_table followed by a resume,
with no suspend first. The resume path suspends if needed, swaps
tables, then resumes. Removing the suspend here would break existing
userspace, I'd imagine. It seems like minimizing the suspended time
would also be a nice benefit.
>
> Mikulas
>
>
>
> On Wed, 17 Jul 2024, Khazhismel Kumykov wrote:
>
> > do_resume when loading a new map first calls dm_suspend, which could
> > silently fail. When we proceeded to dm_swap_table, we would bail out
> > with EINVAL. Instead, restore new_map and return ERESTARTSYS when
> > signaled.
> >
> > Signed-off-by: Khazhismel Kumykov <khazhy@xxxxxxxxxx>
> > ---
> > drivers/md/dm-ioctl.c | 18 ++++++++++++++++--
> > 1 file changed, 16 insertions(+), 2 deletions(-)
> >
> >
> > RFC as I am rather unfamiliar with the locking semantics here - whether
> > we do need to re-grab hash_lock to write to an hc we previously grabbed,
> > and whether the device becoming unhashed while we're in this function is
> > really something that needs to be checked. However, this patch does fix
> > the issue we were seeing - we'd get EINVAL when thread in ioctl was
> > signaled.
> >
> >
> > diff --git a/drivers/md/dm-ioctl.c b/drivers/md/dm-ioctl.c
> > index c2c07bfa6471..b81650c6d096 100644
> > --- a/drivers/md/dm-ioctl.c
> > +++ b/drivers/md/dm-ioctl.c
> > @@ -1181,8 +1181,22 @@ static int do_resume(struct dm_ioctl *param)
> > suspend_flags &= ~DM_SUSPEND_LOCKFS_FLAG;
> > if (param->flags & DM_NOFLUSH_FLAG)
> > suspend_flags |= DM_SUSPEND_NOFLUSH_FLAG;
> > - if (!dm_suspended_md(md))
> > - dm_suspend(md, suspend_flags);
> > + if (!dm_suspended_md(md)) {
> > + r = dm_suspend(md, suspend_flags);
> > + if (r == -EINTR)
> > + r = -ERESTARTSYS;
> > + if (r) {
> > + down_write(&_hash_lock);
> > + hc = dm_get_mdptr(md);
> > + if (!hc)
> > + r = -ENXIO;
> > + else
> > + hc->new_map = new_map;
Oh - I probably want to check if hc->new_map has become non-null in
the meantime and if so... pick a winner then put the loser? Presumably
the newest map should win if that happens / is possible. although the
concept seems suspect to me.
> > + up_write(&_hash_lock);
> > + dm_put(md);
> > + return r;
> > + }
> > + }
> >
> > old_size = dm_get_size(md);
> > old_map = dm_swap_table(md, new_map);
> > --
> > 2.45.2.993.g49e7a77208-goog
> >
>
