On Tue, Oct 08, 2019 at 10:20:39AM +0200, Jan Kara wrote:
> On Tue 08-10-19 05:38:59, Roman Gushchin wrote:
> > On Tue, Oct 08, 2019 at 03:06:31PM +1100, Dave Chinner wrote:
> > > On Fri, Oct 04, 2019 at 03:11:04PM -0700, Roman Gushchin wrote:
> > > > This is a RFC patch, which is not intended to be merged as is,
> > > > but hopefully will start a discussion which can result in a good
> > > > solution for the described problem.
> > > >
> > > > --
> > > >
> > > > We've noticed that the number of dying cgroups on our production hosts
> > > > tends to grow with the uptime. This time it's caused by the writeback
> > > > code.
> > > >
> > > > An inode which is getting dirty for the first time is associated
> > > > with the wb structure (look at __inode_attach_wb()). It can later
> > > > be switched to another wb under some conditions (e.g. some other
> > > > cgroup is writing a lot of data to the same inode), but generally
> > > > stays associated up to the end of life of the inode structure.
> > > >
> > > > The problem is that the wb structure holds a reference to the original
> > > > memory cgroup. So if the inode was dirty once, it has a good chance
> > > > to pin down the original memory cgroup.
> > > >
> > > > An example from the real life: some service runs periodically and
> > > > updates rpm packages. Each time in a new memory cgroup. Installed
> > > > .so files are heavily used by other cgroups, so corresponding inodes
> > > > tend to stay alive for a long. So do pinned memory cgroups.
> > > > In production I've seen many hosts with 1-2 thousands of dying
> > > > cgroups.
> > > >
> > > > This is not the first problem with the dying memory cgroups. As
> > > > always, the problem is with their relative size: memory cgroups
> > > > are large objects, easily 100x-1000x larger that inodes. So keeping
> > > > a couple of thousands of dying cgroups in memory without a good reason
> > > > (what we easily do with inodes) is quite costly (and is measured
> > > > in tens and hundreds of Mb).
> > > >
> > > > One possible approach to this problem is to switch inodes associated
> > > > with dying wbs to the root wb. Switching is a best effort operation
> > > > which can fail silently, so unfortunately we can't run once over a
> > > > list of associated inodes (even if we'd have such a list). So we
> > > > really have to scan all inodes.
> > > >
> > > > In the proposed patch I schedule a work on each memory cgroup
> > > > deletion, which is probably too often. Alternatively, we can do it
> > > > periodically under some conditions (e.g. the number of dying memory
> > > > cgroups is larger than X). So it's basically a gc run.
> > > >
> > > > I wonder if there are any better ideas?
> > > >
> > > > Signed-off-by: Roman Gushchin <guro@xxxxxx>
> > > > ---
> > > > fs/fs-writeback.c | 29 +++++++++++++++++++++++++++++
> > > > mm/memcontrol.c | 5 +++++
> > > > 2 files changed, 34 insertions(+)
> > > >
> > > > diff --git a/fs/fs-writeback.c b/fs/fs-writeback.c
> > > > index 542b02d170f8..4bbc9a200b2c 100644
> > > > --- a/fs/fs-writeback.c
> > > > +++ b/fs/fs-writeback.c
> > > > @@ -545,6 +545,35 @@ static void inode_switch_wbs(struct inode *inode, int new_wb_id)
> > > > up_read(&bdi->wb_switch_rwsem);
> > > > }
> > > >
> > > > +static void reparent_dirty_inodes_one_sb(struct super_block *sb, void *arg)
> > > > +{
> > > > + struct inode *inode, *next;
> > > > +
> > > > + spin_lock(&sb->s_inode_list_lock);
> > > > + list_for_each_entry_safe(inode, next, &sb->s_inodes, i_sb_list) {
> > > > + spin_lock(&inode->i_lock);
> > > > + if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
> > > > + spin_unlock(&inode->i_lock);
> > > > + continue;
> > > > + }
> > > > +
> > > > + if (inode->i_wb && wb_dying(inode->i_wb)) {
> > > > + spin_unlock(&inode->i_lock);
> > > > + inode_switch_wbs(inode, root_mem_cgroup->css.id);
> > > > + continue;
> > > > + }
> > > > +
> > > > + spin_unlock(&inode->i_lock);
> > > > + }
> > > > + spin_unlock(&sb->s_inode_list_lock);
> > >
> > > No idea what the best solution is, but I think this is fundamentally
> > > unworkable. It's not uncommon to have a hundred million cached
> > > inodes these days, often on a single filesystem. Anything that
> > > requires a brute-force system wide inode scan, especially without
> > > conditional reschedule points, is largely a non-starter.
> > >
> > > Also, inode_switch_wbs() is not guaranteed to move the inode to the
> > > destination wb. There can only be WB_FRN_MAX_IN_FLIGHT (1024)
> > > switches in flight at once and switches are run via RCU callbacks,
> > > so I suspect that using inode_switch_wbs() for bulk re-assignment is
> > > going to be a lot more complex than just finding inodes to call
> > > inode_switch_wbs() on....
> >
> > We can schedule it only if the number of dying cgroups exceeds a certain
> > number (like 100), which will make it relatively rare event. Maybe we can
> > add some other conditions, e.g. count the number of inodes associated with
> > a wb and skip scanning if it's zero.
> >
> > Alternatively the wb structure can keep the list of associated inodes,
> > and scan only them, but then it's not trivial to implement without
> > additional complication of already quite complex locking scheme.
> > And because inode_switch_wbs() can fail, we can't guarantee that a single
> > pass over such a list will be enough. That means the we need to schedule
> > scans periodically until all inodes will be switched.
> >
> > So I really don't know which option is better, but at the same time
> > doing nothing isn't the option too. Somehow the problem should be solved.
>
> I agree with Dave that scanning all inodes in the system can get really
> expensive quickly. So what I rather think we could do is create another 'IO
> list' (linked by inode->i_io_list) where we would put inodes that reference
> the wb but are not in any other IO list of the wb. And then we would
> switch inodes on this list when the wb is dying... One would have to be
> somewhat careful with properly draining this list since new inodes can be
> added to it while we work on it but otherwise I don't see any complication
> with this.
Good idea!
I've mastered something like this, will test it for a day or two and post here.
Thank you!
