Re: [Xen-devel] [RFC PATCH] hvm/vpt: fix inconsistent views of vIOAPIC in vmx_intr_assist()

[Chao Gao <chao.gao@xxxxxxxxx>]

On Wed, Apr 12, 2017 at 02:45:36AM +0000, Xuquan (Quan Xu) wrote:
>On April 07, 2017 11:24 AM, Chao Gao wrote:
>>When injecting periodic timer interrupt in vmx_intr_assist(), multiple read
>>operation is operated during one event delivery and incur to inconsistent
>>views of vIOAPIC. For example, if a periodic timer interrupt is from PIT, when
>>set the corresponding bit in vIRR, the corresponding RTE is accessed in
>>pt_update_irq(). When this function returns, it accesses the RTE again to get
>>the vector it set in vIRR.  Between the two accesses, the content of RTE may
>>have been changed by another CPU for no protection method in use. This case
>>can incur the assertion failure in vmx_intr_assist().  Also after a periodic
>>timer interrupt is injected successfully, pt_irq_posted() will decrease the
>>count (pending_intr_nr). But if the corresponding RTE has been changed,
>>pt_irq_posted() will not decrease the count, resulting one more timer
>>interrupt.
>>
>>More discussion and history can be found in
>>1.https://lists.xenproject.org/archives/html/xen-devel/2017-03/msg00906.ht
>>ml
>>2.https://lists.xenproject.org/archives/html/xen-devel/2017-01/msg01019.ht
>>ml
>>
>>This patch simply uses irq_lock to avoid these inconsistent views of vIOAPIC.
>
>To vLAPIC, is it in a similar situation?

For vLAPIC, there is no such situation. The irq field in structure periodic_time
has different meanings for vLAPIC/vIOAPIC case. For vLAPIC case, it is the 
interrupt
vector. For vIOAPIC case, it is the irq number; to get interrupt vector and to
deliver this interrupt, accessing IOAPIC RTE is needed.

Thanks
Chao

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
https://lists.xen.org/xen-devel