Re: [Xen-devel] [PATCH v5 1/3] xen/pt: fix some pass-thru devices don't work across reboot

[Chao Gao <chao.gao@xxxxxxxxx>]

On Wed, Jan 16, 2019 at 11:38:23AM +0100, Roger Pau Monné wrote:
>On Wed, Jan 16, 2019 at 04:17:30PM +0800, Chao Gao wrote:
>> I find some pass-thru devices don't work any more across guest
>> reboot. Assigning it to another domain also meets the same issue. And
>> the only way to make it work again is un-binding and binding it to
>> pciback. Someone reported this issue one year ago [1].
>> 
>> If the device's driver doesn't disable MSI-X during shutdown or qemu is
>> killed/crashed before the domain shutdown, this domain's pirq won't be
>> unmapped. Then xen takes over this work, unmapping all pirq-s, when
>> destroying guest. But as pciback has already disabled meory decoding before
>> xen unmapping pirq, Xen has to sets the host_maskall flag and maskall bit
>> to mask a MSI rather than sets maskbit in MSI-x table. The call trace of
>> this process is:
>> 
>> ->arch_domain_destroy
>>     ->free_domain_pirqs
>>         ->unmap_domain_pirq (if pirq isn't unmapped by qemu)
>>             ->pirq_guest_force_unbind
>>                 ->__pirq_guest_unbind
>>                     ->mask_msi_irq(=desc->handler->disable())
>>                         ->the warning in msi_set_mask_bit()
>> 
>> The host_maskall bit will prevent guests from clearing the maskall bit
>> even the device is assigned to another guest later. Then guests cannot
>> receive MSIs from this device.
>> 
>> To fix this issue, a pirq is unmapped before memory decoding is disabled by
>> pciback. Specifically, when a device is detached from a guest, all 
>> established
>> mappings between pirq and msi are destroying before changing the ownership.
>> 
>> [1]: 
>> https://lists.xenproject.org/archives/html/xen-devel/2017-09/msg02520.html
>
>Thanks, I think the approach is fine, just a couple of comments.
>
>> Signed-off-by: Chao Gao <chao.gao@xxxxxxxxx>
>> ---
>> Changes in v5:
>>  - fix the potential infinite loop
>>  - assert that unmap_domain_pirq() won't fail
>>  - assert msi_list is empty after the loop in pci_unmap_msi
>>  - provide a stub for pt_irq_destroy_bind_msi() if !CONFIG_HVM to fix a
>>    compilation error when building PVShim
>> 
>> Changes in v4:
>>  - split out change to 'msix->warned' field
>>  - handle multiple msi cases
>>  - use list_first_entry_or_null to traverse 'pdev->msi_list'
>> ---
>>  xen/drivers/passthrough/io.c  | 57 ++++++++++++++++++++++++++------------
>>  xen/drivers/passthrough/pci.c | 64 
>> +++++++++++++++++++++++++++++++++++++++++++
>>  xen/include/xen/iommu.h       |  4 +++
>>  3 files changed, 107 insertions(+), 18 deletions(-)
>> 
>> diff --git a/xen/drivers/passthrough/io.c b/xen/drivers/passthrough/io.c
>> index a6eb8a4..56ee1ef 100644
>> --- a/xen/drivers/passthrough/io.c
>> +++ b/xen/drivers/passthrough/io.c
>> @@ -619,6 +619,42 @@ int pt_irq_create_bind(
>>      return 0;
>>  }
>>  
>> +static void pt_irq_destroy_bind_common(struct domain *d, struct pirq *pirq)
>> +{
>> +    struct hvm_pirq_dpci *pirq_dpci = pirq_dpci(pirq);
>> +
>> +    ASSERT(spin_is_locked(&d->event_lock));
>> +
>> +    if ( pirq_dpci && (pirq_dpci->flags & HVM_IRQ_DPCI_MAPPED) &&
>> +         list_empty(&pirq_dpci->digl_list) )
>> +    {
>> +        pirq_guest_unbind(d, pirq);
>> +        msixtbl_pt_unregister(d, pirq);
>> +        if ( pt_irq_need_timer(pirq_dpci->flags) )
>> +            kill_timer(&pirq_dpci->timer);
>> +        pirq_dpci->flags = 0;
>> +        /*
>> +         * See comment in pt_irq_create_bind's PT_IRQ_TYPE_MSI before the
>> +         * call to pt_pirq_softirq_reset.
>> +         */
>> +        pt_pirq_softirq_reset(pirq_dpci);
>> +
>> +        pirq_cleanup_check(pirq, d);
>> +    }
>> +}
>> +
>> +void pt_irq_destroy_bind_msi(struct domain *d, struct pirq *pirq)
>> +{
>> +    struct hvm_pirq_dpci *pirq_dpci = pirq_dpci(pirq);
>
>const
>
>> +
>> +    ASSERT(spin_is_locked(&d->event_lock));
>> +
>> +    if ( pirq_dpci && pirq_dpci->gmsi.posted )
>> +        pi_update_irte(NULL, pirq, 0);
>> +
>> +    pt_irq_destroy_bind_common(d, pirq);
>> +}
>> +
>>  int pt_irq_destroy_bind(
>>      struct domain *d, const struct xen_domctl_bind_pt_irq *pt_irq_bind)
>>  {
>> @@ -727,26 +763,11 @@ int pt_irq_destroy_bind(
>>          }
>>          else
>>              what = "bogus";
>> -    }
>> -    else if ( pirq_dpci && pirq_dpci->gmsi.posted )
>> -        pi_update_irte(NULL, pirq, 0);
>> -
>> -    if ( pirq_dpci && (pirq_dpci->flags & HVM_IRQ_DPCI_MAPPED) &&
>> -         list_empty(&pirq_dpci->digl_list) )
>> -    {
>> -        pirq_guest_unbind(d, pirq);
>> -        msixtbl_pt_unregister(d, pirq);
>> -        if ( pt_irq_need_timer(pirq_dpci->flags) )
>> -            kill_timer(&pirq_dpci->timer);
>> -        pirq_dpci->flags = 0;
>> -        /*
>> -         * See comment in pt_irq_create_bind's PT_IRQ_TYPE_MSI before the
>> -         * call to pt_pirq_softirq_reset.
>> -         */
>> -        pt_pirq_softirq_reset(pirq_dpci);
>>  
>> -        pirq_cleanup_check(pirq, d);
>> +        pt_irq_destroy_bind_common(d, pirq);
>>      }
>> +    else
>> +        pt_irq_destroy_bind_msi(d, pirq);
>>  
>>      spin_unlock(&d->event_lock);
>>  
>> diff --git a/xen/drivers/passthrough/pci.c b/xen/drivers/passthrough/pci.c
>> index 93c20b9..4f2be02 100644
>> --- a/xen/drivers/passthrough/pci.c
>> +++ b/xen/drivers/passthrough/pci.c
>> @@ -1514,6 +1514,68 @@ static int assign_device(struct domain *d, u16 seg, 
>> u8 bus, u8 devfn, u32 flag)
>>      return rc;
>>  }
>>  
>> +/*
>> + * Unmap established mappings between domain's pirq and device's MSI.
>> + * These mappings were set up by qemu/guest and are expected to be
>> + * destroyed when changing the device's ownership.
>> + */
>> +static void pci_unmap_msi(struct pci_dev *pdev)
>> +{
>> +    struct msi_desc *entry, *tmp;
>> +    struct domain *d = pdev->domain;
>> +
>> +    ASSERT(pcidevs_locked());
>> +    ASSERT(d);
>> +
>> +    spin_lock(&d->event_lock);
>> +    list_for_each_entry_safe(entry, tmp, &pdev->msi_list, list)
>> +    {
>> +        struct pirq *info;
>> +        int ret, pirq = 0;
>> +        unsigned int nr = entry->msi_attrib.type != PCI_CAP_ID_MSIX
>> +                          ? entry->msi.nvec : 1;
>
>I think you should mask the entry, like it's done in
>pt_irq_destroy_bind, see the call to guest_mask_msi_irq. That gives a
>consistent state between bind and unbind.
>
>> +
>> +        while ( nr-- )
>> +        {
>> +            struct hvm_pirq_dpci *pirq_dpci;
>
>Nit: you could reduce the scope of info by declaring it here AFAICT.
>
>> +
>> +            pirq = domain_irq_to_pirq(d, entry[nr].irq);
>> +            WARN_ON(pirq < 0);
>> +            if ( pirq <= 0 )
>> +                continue;
>> +
>> +            info = pirq_info(d, pirq);
>> +            if ( !info )
>> +                continue;
>> +
>> +            pirq_dpci = pirq_dpci(info);
>> +            if ( pirq_dpci &&
>> +                 (pirq_dpci->flags & HVM_IRQ_DPCI_MACH_MSI) &&
>> +                 (pirq_dpci->flags & HVM_IRQ_DPCI_GUEST_MSI) )
>> +                pt_irq_destroy_bind_msi(d, info);
>> +        }
>> +
>> +        if ( pirq > 0 )
>> +        {
>> +            /*
>> +             * The pirq is derived from an entry in msi_list rather than an
>> +             * arbitrary value passed down. There should be a irq (msi) 
>> mapped
>> +             * to this pirq. In this case, unmapping this pirq should 
>> succeed.
>> +             * Otherwise, something goes wrong.
>> +             */
>> +            ret = unmap_domain_pirq(d, pirq);
>> +            ASSERT(!ret);
>
>unmap_domain_pirq can fail, why not make pci_unmap_msi return an int
>and propagate the error to the caller?
>
>deassign_device returning an error should also be fine.
>
>> +        }
>> +    }
>> +    /*
>> +     * All pirq-s should have been unmapped and corresponding msi_desc
>> +     * entries should have been removed in the above loop.
>> +     */
>> +    ASSERT(list_empty(&pdev->msi_list));
>> +
>> +    spin_unlock(&d->event_lock);
>> +}
>> +
>>  /* caller should hold the pcidevs_lock */
>>  int deassign_device(struct domain *d, u16 seg, u8 bus, u8 devfn)
>>  {
>> @@ -1529,6 +1591,8 @@ int deassign_device(struct domain *d, u16 seg, u8 bus, 
>> u8 devfn)
>>      if ( !pdev )
>>          return -ENODEV;
>>  
>> +    pci_unmap_msi(pdev);
>
>Just want to make sure, since deassign_device will be called for both
>PV and HVM domains. AFAICT pci_unmap_msi is safe to call when the
>device is assigned to a PV guest, but would like your confirmation.

Tested with a PV guest loaded by Pygrub. PV guest doesn't suffer the
msi-x issue I want to fix.

With these three patches applied, I got some error messages from Xen
and Dom0 as follow:

(XEN) irq.c:2176: dom3: forcing unbind of pirq 332
(XEN) irq.c:2176: dom3: forcing unbind of pirq 331
(XEN) irq.c:2176: dom3: forcing unbind of pirq 328
(XEN) irq.c:2148: dom3: pirq 359 not mapped
[ 2887.067685] xen:events: unmap irq failed -22
(XEN) irq.c:2148: dom3: pirq 358 not mapped
[ 2887.075917] xen:events: unmap irq failed -22
(XEN) irq.c:2148: dom3: pirq 357 not mapped

It seems, the cause of such error is that pirq-s are unmapped and forcibly
unbound on deassignment; subsequent unmapping pirq issued by dom0 fail.
From some aspects, this error is expected. Because with this patch,
pirq-s are expected to be mapped by qemu or dom0 kernel (for pv case) before
deassignment and mapping/binding pirq after deassignment should fail.

So what's your opinion on handling such error? We should figure out another
method to fix msi-x issue to avoid such error or suppress these errors in
qemu and linux kernel?

Thanks
Chao

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxxx
https://lists.xenproject.org/mailman/listinfo/xen-devel