Re: [Xen-devel] [PATCH 2/2] AMD/IOMMU: Render IO_PAGE_FAULT errors in a more useful manner

[Andrew Cooper <andrew.cooper3@xxxxxxxxxx>]

On 27/11/2019 09:40, Roger Pau Monné wrote:
> On Tue, Nov 26, 2019 at 03:01:12PM +0000, Andrew Cooper wrote:
>> Print the PCI coordinates in its common format and use d%u notation for the
>> domain.  As well as printing flags, decode them.  IO_PAGE_FAULT is used for
>> interrupt remapping errors as well as DMA remapping errors.
>>
>> Before:
>>   (XEN) AMD-Vi: IO_PAGE_FAULT: domain = 0, device id = 0xa1, fault address = 
>> 0xbf695000, flags = 0x10
>>   (XEN) AMD-Vi: IO_PAGE_FAULT: domain = 0, device id = 0xa1, fault address = 
>> 0xbf695040, flags = 0x10
>>   (XEN) AMD-Vi: IO_PAGE_FAULT: domain = 0, device id = 0xa1, fault address = 
>> 0xfffffff0, flags = 0x30
>>   (XEN) AMD-Vi: IO_PAGE_FAULT: domain = 0, device id = 0xa1, fault address = 
>> 0x100000000, flags = 0x30
>>   (XEN) AMD-Vi: IO_PAGE_FAULT: domain = 0, device id = 0xa1, fault address = 
>> 0x100000040, flags = 0x30
>>
>> After:
>>   (XEN) AMD-Vi: IO_PAGE_FAULT: 0000:00:14.1 d0 addr 00000000bf5fc000 flags 
>> 0x10 PR
>>   (XEN) AMD-Vi: IO_PAGE_FAULT: 0000:00:14.1 d0 addr 00000000bf5fc040 flags 
>> 0x10 PR
>>   (XEN) AMD-Vi: IO_PAGE_FAULT: 0000:00:14.1 d0 addr 00000000fffffff0 flags 
>> 0x30 RW PR
>>   (XEN) AMD-Vi: IO_PAGE_FAULT: 0000:00:14.1 d0 addr 0000000100000000 flags 
>> 0x30 RW PR
>>   (XEN) AMD-Vi: IO_PAGE_FAULT: 0000:00:14.1 d0 addr 0000000100000040 flags 
>> 0x30 RW PR
> Nit: I would place the domain id information at the beginning (since
> that's more similar to gprintk format), and maybe drop the AMD-Vi
> prefix, it's not very useful IMO:
>
> (XEN) d0 IO_PAGE_FAULT 0000:00:14.1 addr 0000000100000040 flags 0x30 RW PR
>
> But I'm not specially concerned.

So I debated not using d%d format.  This is the DTE's "domain_id"
(a.k.a. Tag in the IO-TLB) field which by convention we set to the domid
of the owning device, but isn't necessarily the best option.

In particular, it might be wise to use domid + 1 and choke if we ever
find 0 in use.

>
>> +        uint64_t addr = *(uint64_t *)(entry + 2);
>> +
>> +        printk(XENLOG_ERR "AMD-Vi: %s: %04x:%02x:%02x.%u d%d addr 
>> %016"PRIx64
>> +               " flags %#x%s%s%s%s%s%s%s%s%s%s\n",
>> +               code_str, iommu->seg, PCI_BUS(device_id), 
>> PCI_SLOT(device_id),
>> +               PCI_FUNC(device_id), domain_id, addr, flags,
>> +               (flags & 0xe00) ? " ??" : "",
>> +               (flags & 0x100) ? " TR" : "",
>> +               (flags & 0x080) ? " RZ" : "",
>> +               (flags & 0x040) ? " PE" : "",
>> +               (flags & 0x020) ? " RW" : "",
>> +               (flags & 0x010) ? " PR" : "",
>> +               (flags & 0x008) ? " I" : "",
>> +               (flags & 0x004) ? " US" : "",
>> +               (flags & 0x002) ? " NX" : "",
>> +               (flags & 0x001) ? " GN" : "");
> I wold rather have those added with proper defined names to
> amd-iommu-defs.h.

All of this is in desperate need of turning into real C structs, rather
than being opencoded in terms of u32[] and offsets/shifts/masks, but
such a change definitely isn't appropriate for backport.

~Andrew

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