Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Subject: Re: [PATCH] powerpc/drmem: cache LMBs in xarray to accelerate lookup
To:     Scott Cheloha <cheloha@linux.ibm.com>,
        Nathan Lynch <nathanl@linux.ibm.com>
Cc:     Rick Lindsley <ricklind@linux.vnet.ibm.com>,
        linuxppc-dev@lists.ozlabs.org, linux-kernel@vger.kernel.org,
        Michael Ellerman <mpe@ellerman.id.au>
References: <20200128221113.17158-1-cheloha@linux.ibm.com>
 <87pnf3i188.fsf@linux.ibm.com>
 <20200129181013.lz6q5lpntnhwclqi@rascal.austin.ibm.com>
From:   "Fontenot, Nathan" <ndfont@gmail.com>
Message-ID: <4dfb2f93-7af8-8c5f-854c-22afead18a8c@gmail.com>
Date:   Thu, 30 Jan 2020 10:09:32 -0600
User-Agent: Mozilla/5.0 (Windows NT 10.0; WOW64; rv:68.0) Gecko/20100101
 Thunderbird/68.4.2
MIME-Version: 1.0
In-Reply-To: <20200129181013.lz6q5lpntnhwclqi@rascal.austin.ibm.com>
Content-Type: text/plain; charset=utf-8
Content-Language: en-US
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On 1/29/2020 12:10 PM, Scott Cheloha wrote:
> On Tue, Jan 28, 2020 at 05:56:55PM -0600, Nathan Lynch wrote:
>> Scott Cheloha <cheloha@linux.ibm.com> writes:
>>> LMB lookup is currently an O(n) linear search.  This scales poorly when
>>> there are many LMBs.
>>>
>>> If we cache each LMB by both its base address and its DRC index
>>> in an xarray we can cut lookups to O(log n), greatly accelerating
>>> drmem initialization and memory hotplug.
>>>
>>> This patch introduces two xarrays of of LMBs and fills them during
>>> drmem initialization.  The patch also adds two interfaces for LMB
>>> lookup.
>>
>> Good but can you replace the array of LMBs altogether
>> (drmem_info->lmbs)? xarray allows iteration over the members if needed.
> 
> I don't think we can without potentially changing the current behavior.
> 
> The current behavior in dlpar_memory_{add,remove}_by_ic() is to advance
> linearly through the array from the LMB with the matching DRC index.
> 
> Iteration through the xarray via xa_for_each_start() will return LMBs
> indexed with monotonically increasing DRC indices.> 
> Are they equivalent?  Or can we have an LMB with a smaller DRC index
> appear at a greater offset in the array?
> 
> If the following condition is possible:
> 
> 	drmem_info->lmbs[i].drc_index > drmem_info->lmbs[j].drc_index
> 
> where i < j, then we have a possible behavior change because
> xa_for_each_start() may not return a contiguous array slice.  It might
> "leap backwards" in the array.  Or it might skip over a chunk of LMBs.
> 

The LMB array should have each LMB in monotonically increasing DRC Index
value. Note that this is set up based on the DT property but I don't recall
ever seeing the DT specify LMBs out of order or not being contiguous.

I am not very familiar with xarrays but it appears this should be possible.

-Nathan