Re: [PATCH v2 4/7] x86/vmx: add do_vmtrace_op

[Michał Leszczyński <michal.leszczynski@xxxxxxx>]

----- 23 cze 2020 o 19:24, Andrew Cooper andrew.cooper3@xxxxxxxxxx napisał(a):
> On 23/06/2020 09:51, Jan Beulich wrote:
>> I'd still like to see an explicit confirmation by him that this
>> use of memory is indeed what he has intended. There are much smaller
>> amounts of memory which we allocate on demand, just to avoid
>> allocating some without then ever using it.
> 
> PT is a debug/diagnostic tool.  Its not something you'd run in
> production against a production VM.
> 
> It's off by default (by virtue of having to explicitly ask to use it in
> the first place), and those who've asked for it don't want to be finding
> -ENOMEM after the domain has been running for a few seconds (or midway
> through the vcpus), when they inveterately want to map the rings.
> 
> Those who request buffers in the first place and forget about them are
> not semantically different from those who ask for a silly shadow memory
> limit, or typo the guest memory and give it too much.  Its a admin
> error, not a safety/correctness issue.
> 
> ~Andrew


Absolutely +1.

Assuming that somebody wants to perform some advanced scenario and is trying
to run many domains (e.g. 20), it's much better to have 19 domains
working fine and 1 prematurely crashing because of -ENOMEM,
rather than have all 20 domains randomly crashing in runtime because
it turned out there is a shortage of memory.


Best regards,
Michał Leszczyński
CERT Polska