>>> - return r;
>>> + if (i == bus->dev_count)
>>> + return;
>>>
>>> new_bus = kmalloc(sizeof(*bus) + ((bus->dev_count - 1) *
>>> sizeof(struct kvm_io_range)), GFP_KERNEL);
>>> - if (!new_bus)
>>> - return -ENOMEM;
>>> + if (!new_bus) {
>>> + pr_err("kvm: failed to shrink bus, removing it completely\n");
>>> + goto broken;
>>
>> The guest will fail in mysterious ways, if you do this (and
>> io_bus_unregister_dev can be called during runtime): in-kernel device
>> accesses will fail with unknown behaviour in the guest.
Actually, the next access to the BUS should result in -ENOMEM. And the
error message should be enough to then figure out what went wrong.
However, to hit this scenario at all feels very unlikely. So I would
like to avoid advanced allocation schemes.
>>
>> Can't you retry a handful of times with GFP_KERNEL before switching to GFP_ATOMIC?
>> (which in case fails the machine is likely to be crashing soon).
>
> The process can run in a cgroup, then kmalloc failure has nothing to
> do with overall memory consumption. Machine can be perfectly fine.
> Also, this very process can be chosen as an OOM kill target, then it
> needs to gracefully deal with kmalloc failure and proceed to a
> termination point.
> Generally retrying something in a loop does not look like a solid plan
> to deal with errors.
>
I agree, looping on memory allocations never feels like the right thing
to do.
--
Thanks,
David