Re: [Xen-devel] [PATCH 2/3] x86/smp: use a dedicated scratch cpumask in send_IPI_mask

From: Jan Beulich <jbeulich@suse.com>
To: "Roger Pau Monné" <roger.pau@citrix.com>
Cc: xen-devel@lists.xenproject.org,
	Sander Eikelenboom <linux@eikelenboom.it>, Wei Liu <wl@xen.org>,
	Andrew Cooper <andrew.cooper3@citrix.com>
Subject: Re: [Xen-devel] [PATCH 2/3] x86/smp: use a dedicated scratch cpumask in send_IPI_mask
Date: Thu, 13 Feb 2020 14:35:32 +0100	[thread overview]
Message-ID: <d7ef9355-8661-1a20-9f27-24db633eed78@suse.com> (raw)
In-Reply-To: <20200213114154.GX4679@Air-de-Roger>

On 13.02.2020 12:41, Roger Pau Monné wrote:
> On Thu, Feb 13, 2020 at 11:19:02AM +0100, Jan Beulich wrote:
>> On 13.02.2020 11:03, Roger Pau Monné wrote:
>>> On Thu, Feb 13, 2020 at 10:59:29AM +0100, Jan Beulich wrote:
>>>> On 12.02.2020 17:49, Roger Pau Monne wrote:
>>>>> Using scratch_cpumask in send_IPI_mak is not safe because it can be
>>>>> called from interrupt context, and hence Xen would have to make sure
>>>>> all the users of the scratch cpumask disable interrupts while using
>>>>> it.
>>>>>
>>>>> Instead introduce a new cpumask to be used by send_IPI_mask, and
>>>>> disable interrupts while using.
>>>>
>>>> My first thought here was: What about NMI or #MC context? Even if
>>>> not using the function today (which I didn't check), there shouldn't
>>>> be a latent issue introduced here preventing them from doing so in
>>>> the future. Instead I think you want to allocate the scratch mask
>>>> dynamically (at least if caller context doesn't allow use of the
>>>> generic one), and simply refrain from coalescing IPIs if this
>>>> allocations fails.
>>>
>>> Hm, isn't this going to be quite expensive, and hence render the
>>> benefit of using the shorthand moot?
>>
>> Depends on how many CPUs there are, i.e. how long the loop otherwise
>> would be. When xmalloc() doesn't need to turn to the page allocator,
>> I think it won't be overly slow. Another option would be to avoid
>> coalescing in a slightly different way (without having to fiddle
>> with the scratch mask) when called in interrupt context.
> 
> What about preventing the mask usage when in nmi context?
> 
> I could introduce something like in_nmi and just avoid the scratch
> mask usage in that case (and the shorthand).

Right, allocation isn't permitted in NMI context anyway. As to
#MC context - I'm afraid we don't have any indicator of this so
far, which is a problem (here and maybe also elsewhere).

Jan

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