Re: port-xen/57199: Pure PVH i386 guests hang on disk activity

[Manuel Bouyer <bouyer%antioche.eu.org@localhost>]

On Mon, Jul 15, 2024 at 05:33:17PM +0000, Taylor R Campbell wrote:
> I can think of two ways this patch could have an impact:
> 
> 1. Some Xen driver relies on write-combining memory (i.e.,
>    `prefetchable' in PCIese and bus_dmaese), or on non-temporal
>    stores.  This seems unlikely.
> 
> 2. This is a single-(v)CPU system which has patched out the lock
>    prefix in membar_sync.
> 
> Unless (1) is happening, I doubt there's any reason to need mfence,
> lfence, or sfence -- except in the circumstances of (1), mfence is
> just a more expensive version of a locked-add for store-before-load
> ordering, and lfence and sfence are never necessary.  See, e.g., the
> AMD memory access ordering rules table:
> 
> AMD64 Architecture Programmer's Manual, Volume 2: System Programming,
> 24593--Rev. 3.38--November 2021, Sec. 7.4.2 Memory Barrier Interaction
> with Memory Types, Table 7-3, p. 196.
> https://web.archive.org/web/20220625040004/https://www.amd.com/system/files/TechDocs/24593.pdf#page=256
> 
> 
> Is this a single-(v)CPU system?  Can you enter crash(8) or drop into
> ddb and disassemble the membar_sync function?  I bet you'll find no
> lock prefix there, which would explain the hangs.
> 
> If my hypothesis about (2) is correct, the right thing is probably
> either to make xen_mb be an assembly stub that does

It is indeed a single-vCPU system, and in PV kernels we're probably not
running hotpatch.

> 
> 	lock
> 	addq $0,-8(%rsp)
> 
> (without the membar_sync hotpatching), or to make xen_mb be inline asm
> to do the same.

I misread the linux code in this area; mb() is not the same as smp_mb().
Linux is in fact not used *fence instructions for virt_*mb(), but
a lock addl for virt_mb() and just barrier() for virt_[rw]mb()

So just adding a lock addl xen_mb should be enough

-- 
Manuel Bouyer <bouyer%antioche.eu.org@localhost>
     NetBSD: 26 ans d'experience feront toujours la difference
--