On Tue 27-06-23 16:57:53, David Hildenbrand wrote:
> On 27.06.23 16:17, Michal Hocko wrote:
> > On Tue 27-06-23 15:14:11, David Hildenbrand wrote:
> > > On 27.06.23 14:40, Michal Hocko wrote:
> > > > On Tue 27-06-23 13:22:18, David Hildenbrand wrote:
> > > > > John Hubbard writes [1]:
> > > > >
> > > > > Some device drivers add memory to the system via memory hotplug.
> > > > > When the driver is unloaded, that memory is hot-unplugged.
> > > > >
> > > > > However, memory hot unplug can fail. And these days, it fails a
> > > > > little too easily, with respect to the above case. Specifically, if
> > > > > a signal is pending on the process, hot unplug fails.
> > > > >
> > > > > [...]
> > > > >
> > > > > So in this case, other things (unmovable pages, un-splittable huge
> > > > > pages) can also cause the above problem. However, those are
> > > > > demonstrably less common than simply having a pending signal. I've
> > > > > got bug reports from users who can trivially reproduce this by
> > > > > killing their process with a "kill -9", for example.
> > > >
> > > > This looks like a bug of the said driver no? If the tear down process is
> > > > killed it could very well happen right before offlining so you end up in
> > > > the very same state. Or what am I missing?
> > >
> > > IIUC (John can correct me if I am wrong):
> > >
> > > 1) The process holds the device node open
> > > 2) The process gets killed or quits
> > > 3) As the process gets torn down, it closes the device node
> > > 4) Closing the device node results in the driver removing the device and
> > > calling offline_and_remove_memory()
> > >
> > > So it's not a "tear down process" that triggers that offlining_removal
> > > somehow explicitly, it's just a side-product of it letting go of the device
> > > node as the process gets torn down.
> >
> > Isn't that just fragile? The operation might fail for other reasons. Why
> > cannot there be a hold on the resource to control the tear down
> > explicitly?
>
> I'll let John comment on that. But from what I understood, in most setups
> where ZONE_MOVABLE gets used for hotplugged memory
> offline_and_remove_memory() succeeds and allows for reusing the device later
> without a reboot.
>
> For the cases where it doesn't work, a reboot is required.
Then the solution should be really robust and means to handle the
failure - e.g. by retrying or alerting the admin.
> > > > > Especially with ZONE_MOVABLE, offlining is supposed to work in most
> > > > > cases when offlining actually hotplugged (not boot) memory, and only fail
> > > > > in rare corner cases (e.g., some driver holds a reference to a page in
> > > > > ZONE_MOVABLE, turning it unmovable).
> > > > >
> > > > > In these corner cases we really don't want to be stuck forever in
> > > > > offline_and_remove_memory(). But in the general cases, we really want to
> > > > > do our best to make memory offlining succeed -- in a reasonable
> > > > > timeframe.
> > > > >
> > > > > Reliably failing in the described case when there is a fatal signal pending
> > > > > is sub-optimal. The pending signal check is mostly only relevant when user
> > > > > space explicitly triggers offlining of memory using sysfs device attributes
> > > > > ("state" or "online" attribute), but not when coming via
> > > > > offline_and_remove_memory().
> > > > >
> > > > > So let's use a timer instead and ignore fatal signals, because they are
> > > > > not really expressive for offline_and_remove_memory() users. Let's default
> > > > > to 30 seconds if no timeout was specified, and limit the timeout to 120
> > > > > seconds.
> > > >
> > > > I really hate having timeouts back. They just proven to be hard to get
> > > > right and it is essentially a policy implemented in the kernel. They
> > > > simply do not belong to the kernel space IMHO.
> > >
> > > As much as I agree with you in terms of offlining triggered from user space
> > > (e.g., write "state" or "online" attribute) where user-space is actually in
> > > charge and can do something reasonable (timeout, retry, whatever), in these
> > > the offline_and_remove_memory() case it's the driver that wants a
> > > best-effort memory offlining+removal.
> > >
> > > If it times out, virtio-mem will simply try another block or retry later.
> > > Right now, it could get stuck forever in offline_and_remove_memory(), which
> > > is obviously "not great". Fortunately, for virtio-mem it's configurable and
> > > we use the alloc_contig_range()-method for now as default.
> >
> > It seems that offline_and_remove_memory is using a wrong operation then.
> > If it wants an opportunistic offlining with some sort of policy. Timeout
> > might be just one policy to use but failure mode or a retry count might
> > be a better fit for some users. So rather than (ab)using offline_pages,
> > would be make more sense to extract basic offlining steps and allow
> > drivers like virtio-mem to reuse them and define their own policy?
>
> virtio-mem, in default operation, does that: use alloc_contig_range() to
> logically unplug ("fake offline") that memory and then just trigger
> offline_and_remove_memory() to make it "officially offline".
>
> In that mode, offline_and_remove_memory() cannot really timeout and is
> almost always going to succeed (except memory notifiers and some hugetlb
> dissolving).
>
> Right now we also allow the admin to configure ordinary offlining directly
> (without prior fake offlining) when bigger memory blocks are used:
> offline_pages() is more reliable than alloc_contig_range(), for example,
> because it disables the PCP and the LRU cache, and retries more often (well,
> unfortunately then also forever). It has a higher chance of succeeding
> especially when bigger blocks of memory are offlined+removed.
>
> Maybe we should make the alloc_contig_range()-based mechanism more
> configurable and make it the only mode in virtio-mem, such that we don't
> have to mess with offline_and_remove_memory() endless loops -- at least for
> virtio-mem.
Yes, that sounds better than hooking up into offline_pages the way this
patch is doing.
--
Michal Hocko
SUSE Labs
