Re: [Xen-devel] [PATCH v2 1/2] x86/mm: factor out the code for shattering an l3 PTE

From: Julien Grall <jgrall@amazon.com>
To: "Xia, Hongyan" <hongyxia@amazon.com>,
	"jbeulich@suse.com" <jbeulich@suse.com>
Cc: "andrew.cooper3@citrix.com" <andrew.cooper3@citrix.com>,
	"xen-devel@lists.xenproject.org" <xen-devel@lists.xenproject.org>,
	"wl@xen.org" <wl@xen.org>,
	"roger.pau@citrix.com" <roger.pau@citrix.com>
Subject: Re: [Xen-devel] [PATCH v2 1/2] x86/mm: factor out the code for shattering an l3 PTE
Date: Wed, 11 Dec 2019 11:47:19 +0000	[thread overview]
Message-ID: <275b6932-5f79-033a-0bb8-a3a41159b979@amazon.com> (raw)
In-Reply-To: <eeb762add75c87db90a3c93a35a7f2149e81c6f7.camel@amazon.com>

Hi Hongyan,

On 11/12/2019 11:28, Xia, Hongyan wrote:
> On Wed, 2019-12-11 at 11:10 +0000, Julien Grall wrote:
>> Hi Hongyan,
>> ...
>>
>> While this involves more instructions, how often do we expect the
>> code
>> to be called?
>>
>> Cheers,
>>
> 
> I don't expect this to be called very often in the current Xen.
> Although with direct map removal, a lot of the memory allocations
> (mostly xenheap allocations) will be mapped and unmapped on-demand and
> there is a much higher change of merging/shattering.

Thank you for the explanation. In order to merge/shatter, you need the 
buddy allocator to ensure all the xenheap are allocated contiguously. I 
am pretty unconvinced there are a lot of page allocated via xenheap. So 
the chance to have contiguous xenheap allocation is very limited.

But the merging/shattering can be counterproductive. An example short 
memory allocation (we have a few places like that):
    xmalloc(...);
    do something
    xfree(...);

We would end up to merge and then a few ms later shatter again. So it 
feels to me, that merging is probably not worth it (I am planning to 
discuss with Andrew today about it).

> 
> However, the series moved all PTEs from xenheap to domheap, and we
> might see other things moved to domheap in the future, so we might not
> have many things left on xenheap anyway.

Typesafe is an important part of making our code base more secure than 
basic C (such as not mixing type).

In this case, I think if we start to merge/shatter a lot, then we have a 
bigger problem and we may want to consider to remove it (see above) So 
it feels the optimization is not worth it.

Note that I am not maintaining this code, so the final call is on Andrew 
and Jan.

Cheers,

-- 

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