Re: domU suspend issue - freeze processes failed - Linux 6.16

["Marek Marczykowski-Górecki" <marmarek@invisiblethingslab.com>]

[-- Attachment #1: Type: text/plain, Size: 7241 bytes --]

On Wed, Sep 24, 2025 at 01:17:15PM +0300, Grygorii Strashko wrote:
> 
> 
> On 22.09.25 13:09, Marek Marczykowski-Górecki wrote:
> > On Fri, Aug 22, 2025 at 08:42:30PM +0200, Marek Marczykowski-Górecki wrote:
> > > On Fri, Aug 22, 2025 at 05:27:20PM +0200, Jürgen Groß wrote:
> > > > On 22.08.25 16:42, Marek Marczykowski-Górecki wrote:
> > > > > On Fri, Aug 22, 2025 at 04:39:33PM +0200, Marek Marczykowski-Górecki wrote:
> > > > > > Hi,
> > > > > > 
> > > > > > When suspending domU I get the following issue:
> > > > > > 
> > > > > >       Freezing user space processes
> > > > > >       Freezing user space processes failed after 20.004 seconds (1 tasks refusing to freeze, wq_busy=0):
> > > > > >       task:xl              state:D stack:0     pid:466   tgid:466   ppid:1      task_flags:0x400040 flags:0x00004006
> > > > > >       Call Trace:
> > > > > >        <TASK>
> > > > > >        __schedule+0x2f3/0x780
> > > > > >        schedule+0x27/0x80
> > > > > >        schedule_preempt_disabled+0x15/0x30
> > > > > >        __mutex_lock.constprop.0+0x49f/0x880
> > > > > >        unregister_xenbus_watch+0x216/0x230
> > > > > >        xenbus_write_watch+0xb9/0x220
> > > > > >        xenbus_file_write+0x131/0x1b0
> > > > > >        vfs_writev+0x26c/0x3d0
> > > > > >        ? do_writev+0xeb/0x110
> > > > > >        do_writev+0xeb/0x110
> > > > > >        do_syscall_64+0x84/0x2c0
> > > > > >        ? do_syscall_64+0x200/0x2c0
> > > > > >        ? generic_handle_irq+0x3f/0x60
> > > > > >        ? syscall_exit_work+0x108/0x140
> > > > > >        ? do_syscall_64+0x200/0x2c0
> > > > > >        ? __irq_exit_rcu+0x4c/0xe0
> > > > > >        entry_SYSCALL_64_after_hwframe+0x76/0x7e
> > > > > >       RIP: 0033:0x79b618138642
> > > > > >       RSP: 002b:00007fff9a192fc8 EFLAGS: 00000246 ORIG_RAX: 0000000000000014
> > > > > >       RAX: ffffffffffffffda RBX: 00000000024fd490 RCX: 000079b618138642
> > > > > >       RDX: 0000000000000003 RSI: 00007fff9a193120 RDI: 0000000000000014
> > > > > >       RBP: 00007fff9a193000 R08: 0000000000000000 R09: 0000000000000000
> > > > > >       R10: 0000000000000000 R11: 0000000000000246 R12: 0000000000000014
> > > > > >       R13: 00007fff9a193120 R14: 0000000000000003 R15: 0000000000000000
> > > > > >        </TASK>
> > > > > >       OOM killer enabled.
> > > > > >       Restarting tasks: Starting
> > > > > >       Restarting tasks: Done
> > > > > >       xen:manage: do_suspend: freeze processes failed -16
> > > > > > 
> > > > > > The process in question is `xl devd` daemon. It's a domU serving a
> > > > > > xenvif backend.
> > > > > > 
> > > > > > I noticed it on 6.16.1, but looking at earlier test logs I see it with
> > > > > > 6.16-rc6 already (but interestingly, not 6.16-rc2 yet? feels weird given
> > > > > > seemingly no relevant changes between rc2 and rc6).
> > > > > 
> > > > > I forgot to include link for (a little) more details:
> > > > > https://github.com/QubesOS/qubes-linux-kernel/pull/1157
> > > > > 
> > > > > Especially, there is another call trace with panic_on_warn enabled -
> > > > > slightly different, but looks related.
> > > > > 
> > > > 
> > > > I'm pretty sure the PV variant for suspending is just wrong: it is calling
> > > > dpm_suspend_start() from do_suspend() without taking the required
> > > > system_transition_mutex, resulting in the WARN() in pm_restrict_gfp_mask().
> > > > 
> > > > It might be as easy as just adding the mutex() call to do_suspend(), but I'm
> > > > really not sure that will be a proper fix.
> > > 
> > > Hm, this might explain the second call trace, but not the freeze failure
> > > quoted here above, I think?
> > 
> > While the patch I sent appears to fix this particular issue, it made me
> > wonder: is there any fundamental reason why do_suspend() is not using
> > pm_suspend() and register Xen-specific actions via platform_suspend_ops
> > (and maybe syscore_ops)? From a brief look at the code, it should
> > theoretically be possible, and should avoid issues like this.
> > 
> > I tried to do a quick&dirty attempt at that[1], and it failed (panic). I
> > surely made several mistakes there (and also left a ton of todo
> > comments). But before spending any more time at that, I'd like to ask
> > if this is a viable option at all.
> 
> I think it might, but be careful with this, because there are two "System Low power" paths in Linux
> 1) Suspend2RAM and Co
> 2) Hybernation
> 
> While "Suspend2RAM and Co" path is relatively straight forward and expected to be always
> started through pm_suspend(). In general, it's expected to happen
>  - from sysfs (User space)
>  - from autosuspend (wakelocks).
> 
> the "hibernation" path is more complicated:(
> - Genuine Linux hybernation hibernate()/hibernate_quiet_exec()

IIUC hibernation is very different as it puts Linux in charge of dumping
all the state to the disk. In case of Xen, the primary use case for
suspend is preparing VM for Xen toolstack serializing its state to disk
(or migrating to another host). 
Additionally, VM suspend may be used as preparation for host suspend
(this is what I actually do here). This is especially relevant if the VM
has some PCI passthrough - to properly suspend (and resume) devices
across host suspend.

> I'm not sure what path Xen originally implemented :( It seems like "suspend2RAM",
> but, at the same time "hybernation" specific staff is used, like PMSG_FREEZE/PMSG_THAW/PMSG_RESTORE.
> As result, Linux suspend/hybernation code moves forward while Xen stays behind and unsync.

Yeah, I think it's supposed to be suspend2RAM. TBH the
PMSG_FREEZE/PMSG_THAW/PMSG_RESTORE confuses me too and Qubes OS has a
patch[2] to switch it to PMSG_SUSPEND/PMSG_RESUME.

> So it sounds reasonable to avoid custom implementation, but may be not easy :(
> 
> Suspending Xen features can be split between suspend stages, but
> not sure if platform_suspend_ops can be used.
> 
> Generic suspend stages list
> - freeze
> - prepare
> - suspend
> - suspend_late
> - suspend_noirq (SPIs disabled, except wakeups)
>   [most of Xen specific staff has to be suspended at this point]
> - disable_secondary_cpus
> - arch disable IRQ (from this point no IRQs allowed, no timers, no scheduling)
> - syscore_suspend
>   [rest here]
> - platform->enter() (suspended)
> 
> You can't just overwrite platform_suspend_ops, because ARM64 is expected to enter
> suspend through PSCI FW interface:
> drivers/firmware/psci/psci.c
>  static const struct platform_suspend_ops psci_suspend_ops = {

Does this apply to a VM on ARM64 too? At least on x86, the VM is
supposed to make a hypercall to tell Xen it suspended (the hypercall
will return only on resume).

> As an option, some Xen components could be converted to use syscore_ops (but not xenstore),
> and some might need to use DD(dev_pm_ops).
> 
> > 
> > [1] https://github.com/marmarek/linux/commit/47cfdb991c85566c9c333570511e67bf477a5da6
> 
> -- 
> Best regards,
> -grygorii
> 

[2] https://github.com/QubesOS/qubes-linux-kernel/blob/main/xen-pm-use-suspend.patch

-- 
Best Regards,
Marek Marczykowski-Górecki
Invisible Things Lab

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 488 bytes --]