On Fri, May 03, 2024 at 11:00:18PM +0800, Chen Yu wrote:
> On 2024-05-03 at 16:47:49 +0300, Kirill A. Shutemov wrote:
> > On Fri, May 03, 2024 at 12:31:12PM +0200, Ard Biesheuvel wrote:
> > > On Wed, 24 Apr 2024 at 19:12, Ard Biesheuvel <ardb@xxxxxxxxxx> wrote:
> > > >
> > > > On Mon, 22 Apr 2024 at 16:40, Chen Yu <yu.c.chen@xxxxxxxxx> wrote:
> > > > >
> > > > > On 2024-04-11 at 08:49:07 +0800, Chen Yu wrote:
> > > > > > Commit 50e782a86c98 ("efi/unaccepted: Fix soft lockups caused
> > > > > > by parallel memory acceptance") has released the spinlock so
> > > > > > other CPUs can do memory acceptance in parallel and not
> > > > > > triggers softlockup on other CPUs.
> > > > > >
> > > > > > However the softlock up was intermittent shown up if the memory
> > > > > > of the TD guest is large, and the timeout of softlockup is set
> > > > > > to 1 second.
> > > > > >
> > > > > > The symptom is:
> > > > > > When the local irq is enabled at the end of accept_memory(),
> > > > > > the softlockup detects that the watchdog on single CPU has
> > > > > > not been fed for a while. That is to say, even other CPUs
> > > > > > will not be blocked by spinlock, the current CPU might be
> > > > > > stunk with local irq disabled for a while, which hurts not
> > > > > > only nmi watchdog but also softlockup.
> > > > > >
> > > > > > Chao Gao pointed out that the memory accept could be time
> > > > > > costly and there was similar report before. Thus to avoid
> > > > > > any softlocup detection during this stage, give the
> > > > > > softlockup a flag to skip the timeout check at the end of
> > > > > > accept_memory(), by invoking touch_softlockup_watchdog().
> > > > > >
> > > > > > Fixes: 50e782a86c98 ("efi/unaccepted: Fix soft lockups caused by parallel memory acceptance")
> > > > > > Reported-by: "Hossain, Md Iqbal" <md.iqbal.hossain@xxxxxxxxx>
> > > > > > Reviewed-by: Kirill A. Shutemov <kirill.shutemov@xxxxxxxxxxxxxxx>
> > > > > > Signed-off-by: Chen Yu <yu.c.chen@xxxxxxxxx>
> > > > > > ---
> > > > > > v1 -> v2:
> > > > > > Refine the commit log and add fixes tag/reviewed-by tag from Kirill.
> > > > >
> > > > > Gently pinging about this patch.
> > > > >
> > > >
> > > > Queued up in efi/urgent now, thanks.
> > >
> > > OK, I was about to send this patch to Linus (and I am still going to).
> > >
> > > However, I do wonder if sprinkling touch_softlockup_watchdog() left
> > > and right is really the right solution here.
> > >
> > > Looking at the backtrace, this is a page fault originating in user
> > > space. So why do we end up calling into the hypervisor to accept a
> > > chunk of memory large enough to trigger the softlockup watchdog? Or is
> > > the hypercall simply taking a disproportionate amount of time?
> >
> > Note that softlockup timeout was set to 1 second to trigger this. So this
> > is exaggerated case.
> >
> > > And AIUI, touch_softlockup_watchdog() hides the fact that we are
> > > hogging the CPU for way too long - is there any way we can at least
> > > yield the CPU on this condition?
> >
> > Not really. There's no magic entity that handles accept. It is done by
> > CPU.
> >
> > There's a feature in pipeline that makes page accept interruptable in TDX
> > guest. It can help some cases. But if ended up in this codepath from
> > non-preemptable context, it won't help.
> >
>
> Is it possible to enable the local irq for a little while after
> each arch_accept_memory(phys_start, phys_end),
> and even split the [phys_start,phys_end] to smaller regions?
> so the watchdog can be fed on time/tick is normal. But currently
> the softlock fed at the end seems to be more easier to implement.
That's what I did initially. But Vlastimil correctly pointed that it will
lead to deadlock.
https://lore.kernel.org/all/088593ea-e001-fa87-909f-a196b1373ca4@xxxxxxx/
--
Kiryl Shutsemau / Kirill A. Shutemov
