|
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] Re: [PATCH] x86/hvm: Don't save or load HVM context of a domain with no vCPUs
Le 18/11/2026 à 08:06, Lin Liu a écrit :
> XEN_DOMCTL_createdomain allocates d->vcpu[] but the vCPUs themselves are
> only created by a later XEN_DOMCTL_max_vcpus. In that window
> XEN_DOMCTL_gethvmcontext{,_partial} on the domain reach the per-domain
> save handlers with d->vcpu[0] == NULL. A dom0 tool polling xen-hvmctx
> across all domains while a VM was being restored took the host down:
>
> (XEN) HVM d25 save: PIC
> (XEN) RIP: e008:[<ffff82d04030dc98>]
> arch/x86/hvm/vpic.c#vpic_save+0x8/0x8f
> (XEN) rdi: 0000000000000000
> (XEN) Xen call trace:
> (XEN) [<ffff82d04030dc98>] R arch/x86/hvm/vpic.c#vpic_save+0x8/0x8f
> (XEN) [<ffff82d04021009f>] S arch/x86/hvm/save.c#hvm_save.cold+0x9b/0xf7
> (XEN) [<ffff82d04036d99c>] S arch_do_domctl+0x1a2c/0x28a6
> (XEN) [<ffff82d04025341a>] S do_domctl+0x139a/0x1642
> (XEN) [<ffff82d04033331e>] S pv_hypercall+0x3ae/0x59a
> (XEN) [<ffff82d0402003b5>] S lstar_enter+0x105/0x10a
> (XEN)
> (XEN) Panic on CPU 0:
> (XEN) FATAL PAGE FAULT
> (XEN) [error_code=0000]
> (XEN) Faulting linear address: 0000000000000018
>
> XEN_DOMCTL_sethvmcontext has the same problem on the load side, e.g.
> pit_load() uses d->vcpu[0] for the guest time.
>
> Refuse to save or load the whole context until vCPU 0 exists, and look up
> the vCPU in hvm_save_one() with domain_vcpu(), failing with -ENOENT as
> for other absent vCPUs.
>
> Fixes: d0d8980138c4 ("x86/hvm: remove redundant save functions")
> Assisted-by: Claude:claude-opus-5
> Signed-off-by: Lin Liu <lin.liu01@xxxxxxxxxx>
> ---
> xen/arch/x86/hvm/save.c | 15 ++++++++-------
> 1 file changed, 8 insertions(+), 7 deletions(-)
>
> diff --git a/xen/arch/x86/hvm/save.c b/xen/arch/x86/hvm/save.c
> index 8ab6405706..48f8229061 100644
> --- a/xen/arch/x86/hvm/save.c
> +++ b/xen/arch/x86/hvm/save.c
> @@ -139,12 +139,11 @@ int hvm_save_one(struct domain *d, unsigned int
> typecode, unsigned int instance,
> !hvm_sr_handlers[typecode].save )
> return -EINVAL;
>
> - if ( hvm_sr_handlers[typecode].kind != HVMSR_PER_VCPU )
> - v = d->vcpu[0];
> - else if ( instance >= d->max_vcpus || !d->vcpu[instance] )
> + /* Per-domain records are saved via vCPU 0, absent before max_vcpus. */
> + v = domain_vcpu(d, hvm_sr_handlers[typecode].kind == HVMSR_PER_VCPU
> + ? instance : 0);
> + if ( !v )
> return -ENOENT;
> - else
> - v = d->vcpu[instance];
> ctxt.size = hvm_sr_handlers[typecode].size;
> ctxt.data = xmalloc_bytes(ctxt.size);
> if ( !ctxt.data )
> @@ -203,7 +202,8 @@ int hvm_save(struct domain *d, hvm_domain_context_t *h)
> struct hvm_save_end end;
> unsigned int i;
>
> - if ( d->is_dying )
> + /* Per-domain handlers are passed vCPU 0. */
> + if ( d->is_dying || !d->vcpu[0] )
> return -EINVAL;
>
> hdr.magic = HVM_FILE_MAGIC;
> @@ -291,7 +291,8 @@ int hvm_load(struct domain *d, bool real,
> hvm_domain_context_t *h)
> struct hvm_save_descriptor *desc;
> int rc;
>
> - if ( d->is_dying )
> + /* Some per-domain loaders use vCPU 0. */
> + if ( d->is_dying || !d->vcpu[0] )
> return -EINVAL;
>
> /* Get at the save header, which must be first */
Reviewed-by: Teddy Astie <teddy.astie@xxxxxxxxxx>
We probably want to refactor parts of this function to avoid duplication
(repeated hvm_sr_handlers[typecode], hvm_sr_handlers[typecode].kind ==
HVMSR_PER_VCPU, ...). But that's better done in a separate patch.
Teddy
--
Teddy Astie | Vates XCP-ng Developer
XCP-ng & Xen Orchestra - Vates solutions
web: https://vates.tech
|
![]() |
Lists.xenproject.org is hosted with RackSpace, monitoring our |