Re: [Xen-devel] [PATCH RFC v2 10/10] AMD/IOMMU: correct IRTE updating

From: Jan Beulich <JBeulich@suse.com>
To: Andrew Cooper <andrew.cooper3@citrix.com>,
	xen-devel <xen-devel@lists.xenproject.org>
Cc: Brian Woods <brian.woods@amd.com>,
	Suravee Suthikulpanit <suravee.suthikulpanit@amd.com>
Subject: Re: [Xen-devel] [PATCH RFC v2 10/10] AMD/IOMMU: correct IRTE updating
Date: Wed, 3 Jul 2019 08:55:28 +0000	[thread overview]
Message-ID: <2aacdea9-d5a1-eee6-6c1b-7dca6315d8e6@suse.com> (raw)
In-Reply-To: <f3798406-3072-14db-cb15-d88cc1762c55@citrix.com>

On 02.07.2019 17:08, Andrew Cooper wrote:
> On 27/06/2019 16:23, Jan Beulich wrote:
>> While for 32-bit IRTEs I think we can safely continue to assume that the
>> writes will translate to a single MOV, the use of CMPXCHG16B is more
> 
> The CMPXCHG16B here is stale.

Indeed, as is the 32-bit IRTE part of the sentence (now that I
use ACCESS_ONCE() already before this patch).

>> heavy handed than necessary for the 128-bit form, and the flushing
>> didn't get done along the lines of what the specification says. Mark
>> entries to be updated as not remapped (which will result in interrupt
>> requests to get target aborted, but the interrupts should be masked
>> anyway at that point in time), issue the flush, and only then write the
>> new entry. In the 128-bit IRTE case set RemapEn separately last, to that
>> the ordering of the writes of the two 64-bit halves won't matter.

This last sentence is stale too, and hence I've now removed it.

>> --- a/xen/drivers/passthrough/amd/iommu_intr.c
>> +++ b/xen/drivers/passthrough/amd/iommu_intr.c
>> @@ -238,8 +238,7 @@ static void update_intremap_entry(union
>>           break;
>>   
>>       case irte128:
>> -        ACCESS_ONCE(entry.ptr128->raw[0]) = 0;
>> -        barrier();
>> +        ASSERT(!entry.ptr128->full.remap_en);
>>           entry.ptr128->raw[1] =
>>               container_of(&full, union irte128, full)->raw[1];
>>           barrier();
>> @@ -308,6 +307,20 @@ static int update_intremap_entry_from_io
>>       }
>>   
>>       entry = get_intremap_entry(iommu->seg, req_id, offset);
>> +
>> +    /* The RemapEn fields match for all formats. */
>> +    while ( iommu->enabled && entry.ptr32->basic.remap_en )
> 
> Why while?  (and by this, what I mean is that this definitely needs a
> comment, because the code looks like it ought to be an if.)

Well - see the RFC remark after the description. I'd be happy to
change to if(), but only on solid grounds. Without clear
guarantees that no races between IRTE updates can occur, we need
to continue flushing as long as we find RemapEn to have got set
again after a flush. Note how the necessary lock guarding against
such is getting dropped and re-acquired in the loop bodies.

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