Hi Andrew,
On 2024/7/30 21:09, Andrew Cooper wrote:
> On 08/07/2024 12:41 pm, Jiqian Chen wrote:
>> If run Xen with PVH dom0 and hvm domU, hvm will map a pirq for
>> a passthrough device by using gsi, see qemu code
>> xen_pt_realize->xc_physdev_map_pirq and libxl code
>> pci_add_dm_done->xc_physdev_map_pirq. Then xc_physdev_map_pirq
>> will call into Xen, but in hvm_physdev_op, PHYSDEVOP_map_pirq
>> is not allowed because currd is PVH dom0 and PVH has no
>> X86_EMU_USE_PIRQ flag, it will fail at has_pirq check.
>>
>> So, allow PHYSDEVOP_map_pirq when dom0 is PVH and also allow
>> PHYSDEVOP_unmap_pirq for the removal device path to unmap pirq.
>> And add a new check to prevent (un)map when the subject domain
>> doesn't have a notion of PIRQ.
>>
>> So that the interrupt of a passthrough device can be
>> successfully mapped to pirq for domU with a notion of PIRQ
>> when dom0 is PVH
>>
>> Signed-off-by: Jiqian Chen <jiqian.c...@amd.com>
>> Signed-off-by: Huang Rui <ray.hu...@amd.com>
>> Signed-off-by: Jiqian Chen <jiqian.c...@amd.com>
>> ---
>> xen/arch/x86/hvm/hypercall.c | 6 ++++++
>> xen/arch/x86/physdev.c | 12 ++++++++++--
>> 2 files changed, 16 insertions(+), 2 deletions(-)
>>
>> diff --git a/xen/arch/x86/hvm/hypercall.c b/xen/arch/x86/hvm/hypercall.c
>> index 0fab670a4871..03ada3c880bd 100644
>> --- a/xen/arch/x86/hvm/hypercall.c
>> +++ b/xen/arch/x86/hvm/hypercall.c
>> @@ -71,8 +71,14 @@ long hvm_physdev_op(int cmd, XEN_GUEST_HANDLE_PARAM(void)
>> arg)
>>
>> switch ( cmd )
>> {
>> + /*
>> + * Only being permitted for management of other domains.
>> + * Further restrictions are enforced in do_physdev_op.
>> + */
>> case PHYSDEVOP_map_pirq:
>> case PHYSDEVOP_unmap_pirq:
>> + break;
>> +
>> case PHYSDEVOP_eoi:
>> case PHYSDEVOP_irq_status_query:
>> case PHYSDEVOP_get_free_pirq:
>> diff --git a/xen/arch/x86/physdev.c b/xen/arch/x86/physdev.c
>> index d6dd622952a9..9f30a8c63a06 100644
>> --- a/xen/arch/x86/physdev.c
>> +++ b/xen/arch/x86/physdev.c
>> @@ -323,7 +323,11 @@ ret_t do_physdev_op(int cmd,
>> XEN_GUEST_HANDLE_PARAM(void) arg)
>> if ( !d )
>> break;
>>
>> - ret = physdev_map_pirq(d, map.type, &map.index, &map.pirq, &msi);
>> + /* Only mapping when the subject domain has a notion of PIRQ */
>> + if ( !is_hvm_domain(d) || has_pirq(d) )
>> + ret = physdev_map_pirq(d, map.type, &map.index, &map.pirq,
>> &msi);
>> + else
>> + ret = -EOPNOTSUPP;
>>
>> rcu_unlock_domain(d);
>>
>> @@ -346,7 +350,11 @@ ret_t do_physdev_op(int cmd,
>> XEN_GUEST_HANDLE_PARAM(void) arg)
>> if ( !d )
>> break;
>>
>> - ret = physdev_unmap_pirq(d, unmap.pirq);
>> + /* Only unmapping when the subject domain has a notion of PIRQ */
>> + if ( !is_hvm_domain(d) || has_pirq(d) )
>> + ret = physdev_unmap_pirq(d, unmap.pirq);
>> + else
>> + ret = -EOPNOTSUPP;
>>
>> rcu_unlock_domain(d);
>>
>
> Gitlab is displeased with your offering.
>
> https://gitlab.com/xen-project/xen/-/pipelines/1393459622
>
> This breaks both {adl,zen3p}-pci-hvm-x86-64-gcc-debug, and given the:
>
> (XEN) [ 8.150305] HVM restore d1: CPU 0
> libxl: error: libxl_pci.c:1491:pci_add_dm_done: Domain
> 1:xc_physdev_map_pirq irq=18 (error=-1): Not supported
> libxl: error: libxl_pci.c:1809:device_pci_add_done: Domain
> 1:libxl__device_pci_add failed for PCI device 0:3:0.0 (rc -3)
> libxl: error: libxl_create.c:1962:domcreate_attach_devices: Domain
> 1:unable to add pci devices
> libxl: error: libxl_xshelp.c:206:libxl__xs_read_mandatory: xenstore read
> failed: `/libxl/1/type': No such file or directory
> libxl: warning: libxl_dom.c:49:libxl__domain_type: unable to get domain
> type for domid=1, assuming HVM
> libxl: error: libxl_domain.c:1616:domain_destroy_domid_cb: Domain
> 1:xc_domain_destroy failed: No such process
>
> I'd say that we're hitting the newly introduced -EOPNOTSUPP path.
>
> In the test scenario, dom0 is PV, and it's an HVM domU which is breaking.
>
> The sibling *-pci-pv-* tests (a PV domU) are working fine.
>
> Either way, I'm going to revert this for now because clearly the "the
> subject domain has a notion of PIRQ" hasn't been reasoned about
> correctly, and it's important to keep Gitlab CI green across the board.
OK, I will try to reproduce and investigate this issue, thanks.
>
> ~Andrew
--
Best regards,
Jiqian Chen.
