Re: linux-next regression: SNP Guest boot hangs with certain cpu/mem config combination
From: Kirill A. Shutemov
Date: Thu Mar 27 2025 - 10:45:18 EST
On Thu, Mar 27, 2025 at 10:35:33AM -0400, Jason Baron wrote:
>
>
> On 3/27/25 10:25 AM, Kirill A. Shutemov wrote:
> > !-------------------------------------------------------------------|
> > This Message Is From an External Sender
> > This message came from outside your organization.
> > |-------------------------------------------------------------------!
> >
> > On Thu, Mar 27, 2025 at 02:58:12PM +0200, Kirill A. Shutemov wrote:
> > > On Wed, Mar 26, 2025 at 05:30:35PM -0500, Tom Lendacky wrote:
> > > > On 3/25/25 08:33, Kirill A. Shutemov wrote:
> > > > > On Tue, Mar 25, 2025 at 02:40:00PM +0530, Aithal, Srikanth wrote:
> > > > > > Hello,
> > > > > >
> > > > > >
> > > > > > Starting linux-next build next-20250312, including recent build 20250324, we
> > > > > > are seeing an issue where the SNP guest boot hangs at the "boot smp config"
> > > > > > step:
> > > > > >
> > > > > >
> > > > > > [ 2.294722] smp: Bringing up secondary CPUs ...
> > > > > > [ 2.295211] smpboot: Parallel CPU startup disabled by the platform
> > > > > > [ 2.309687] smpboot: x86: Booting SMP configuration:
> > > > > > [ 2.310214] .... node #0, CPUs: #1 #2 #3 #4 #5 #6
> > > > > > #7 #8 #9 #10 #11 #12 #13 #14 #15 #16 #17 #18 #19 #20 #21
> > > > > > #22 #23 #24 #25 #26 #27 #28 #29 #30 #31 #32 #33 #34 #35 #36
> > > > > > #37 #38 #39 #40 #41 #42 #43 #44 #45 #46 #47 #48 #49 #50 #51
> > > > > > #52 #53 #54 #55 #56 #57 #58 #59 #60 #61 #62 #63 #64 #65 #66
> > > > > > #67 #68 #69 #70 #71 #72 #73 #74 #75 #76 #77 #78 #79 #80 #81
> > > > > > #82 #83 #84 #85 #86 #87 #88 #89 #90 #91 #92 #93 #94 #95 #96
> > > > > > #97 #98 #99 #100 #101 #102 #103 #104 #105 #106 #107 #108 #109 #110 #111
> > > > > > #112 #113 #114 #115 #116 #117 #118 #119 #120 #121 #122 #123 #124 #125 #126
> > > > > > #127 #128 #129 #130 #131 #132 #133 #134 #135 #136 #137 #138 #139 #140 #141
> > > > > > #142 #143 #144 #145 #146 #147 #148 #149 #150 #151 #152 #153 #154 #155 #156
> > > > > > #157 #158 #159 #160 #161 #162 #163 #164 #165 #166 #167 #168 #169 #170 #171
> > > > > > #172 #173 #174 #175 #176 #177 #178 #179 #180 #181 #182 #183 #184 #185 #186
> > > > > > #187 #188 #189 #190 #191 #192 #193 #194 #195 #196 #197 #198
> > > > > > --> The guest hangs forever at this point.
> > > > > >
> > > > > >
> > > > > > I have observed that certain vCPU and memory combinations work, while others
> > > > > > do not. The VM configuration I am using does not have any NUMA nodes.
> > > > > >
> > > > > > vcpus Mem SNP guest boot
> > > > > > <=240 19456M Boots fine
> > > > > > > =241,<255 19456M Hangs
> > > > > > 1-255 2048M Boots fine
> > > > > > 1-255 4096M Boots fine
> > > > > > > 71 8192M Hangs
> > > > > > > 41 6144M Hangs
> > > > > >
> > > > > > When I bisected this issue, it pointed to the following commit :
> > > > > >
> > > > > >
> > > > > > *commit 800f1059c99e2b39899bdc67a7593a7bea6375d8*
> > > > > > Author: Kirill A. Shutemov <kirill.shutemov@xxxxxxxxxxxxxxx>
> > > > > > Date: Mon Mar 10 10:28:55 2025 +0200
> > > > > >
> > > > > > mm/page_alloc: fix memory accept before watermarks gets initialized
> > > > >
> > > > > Hm. It is puzzling for me. I don't see how this commit can cause the hang.
> > > > >
> > > > > Could you track down where hang happens?
> > > >
> > > > Let me say that the guest config is key for this. Using that config, I
> > > > think you might be able to repro this on TDX. The config does turn off TDX
> > > > support, so I'm hoping that turning it on doesn't change anything.
> > > >
> > > > I've been able to track it down slightly... It is happening during the CPU
> > > > bringup trace points and it eventually gets to line 2273 in
> > > > rb_allocate_cpu_buffer() and never comes back from an alloc_pages_node()
> > > > call. That's as far as I've gotten so far. I'm not a mm expert so not sure
> > > > if I'll be able to progress much further.
> > >
> > > Urgh... It is deadlock on cpu_hotplug_lock :/
> > >
> > > _cpu_up() takes the lock on write and starts CPU bring up under it.
> > > If during CPU bringup we accept the last page in the zone, __accept_page()
> > > calls static_branch_dec() which takes the lock again.
> > >
> > > Oopsie.
> > >
> > > So the patch itself doesn't introduce a regression, but uncovers
> > > preexisting deadlock. With the patch we accept more pages during the boot
> > > and it triggers the deadlock.
> > >
> > > Let me think about the fix.
> >
> > + Static branch maintainers
> >
> > The only option I see so far is to drop static branch from this path.
> >
> > But I am not sure if it the only case were we use static branch from CPU
> > hotplug callbacks.
> >
> > Any other ideas?
>
>
> Hmmm, didn't take too close a look here, but there is the
> static_key_slow_dec_cpuslocked() variant, would that work here? Is the issue
> the caller may or may not have the cpu_hotplug lock?
Yes. This is generic page alloc path and it can be called with and without
the lock.
--
Kiryl Shutsemau / Kirill A. Shutemov