From: Dave Hansen <dave.hansen@intel.com>
To: Shaokun Zhang <zhangshaokun@hisilicon.com>,
linux-kernel@vger.kernel.org, netdev@vger.kernel.org
Cc: Yuqi Jin <jinyuqi@huawei.com>,
Rusty Russell <rusty@rustcorp.com.au>,
Andrew Morton <akpm@linux-foundation.org>,
Juergen Gross <jgross@suse.com>,
Paul Burton <paul.burton@mips.com>,
Michal Hocko <mhocko@suse.com>,
Michael Ellerman <mpe@ellerman.id.au>,
Mike Rapoport <rppt@linux.ibm.com>,
Anshuman Khandual <anshuman.khandual@arm.com>
Subject: Re: [PATCH v6] lib: optimize cpumask_local_spread()
Date: Fri, 13 Nov 2020 08:02:29 -0800 [thread overview]
Message-ID: <5e8b0304-4de1-4bdc-41d2-79fa5464fbc7@intel.com> (raw)
In-Reply-To: <eec4c1b6-8dad-9d07-7ef4-f0fbdcff1785@hisilicon.com>
On 11/12/20 6:06 PM, Shaokun Zhang wrote:
>>> On Huawei Kunpeng 920 server, there are 4 NUMA node(0 - 3) in the 2-cpu
>>> system(0 - 1). The topology of this server is followed:
>>
>> This is with a feature enabled that Intel calls sub-NUMA-clustering
>> (SNC), right? Explaining *that* feature would also be great context for
>
> Correct,
>
>> why this gets triggered on your system and not normally on others and
>> why nobody noticed this until now.
>
> This is on intel 6248 platform:
I have no idea what a "6248 platform" is.
>>> +static void calc_node_distance(int *node_dist, int node)
>>> +{
>>> + int i;
>>> +
>>> + for (i = 0; i < nr_node_ids; i++)
>>> + node_dist[i] = node_distance(node, i);
>>> +}
>>
>> This appears to be the only place node_dist[] is written. That means it
>> always contains a one-dimensional slice of the two-dimensional data
>> represented by node_distance().
>>
>> Why is a copy of this data needed?
>
> It is used to store the distance with the @node for later, apologies that I
> can't follow your question correctly.
Right, the data that you store is useful. *But*, it's also a verbatim
copy of the data from node_distance(). Why not just use node_distance()
directly in your code rather than creating a partial copy of it in the
local node_dist[] array?
>>> unsigned int cpumask_local_spread(unsigned int i, int node)
>>> {
>>> - int cpu, hk_flags;
>>> + static DEFINE_SPINLOCK(spread_lock);
>>> + static int node_dist[MAX_NUMNODES];
>>> + static bool used[MAX_NUMNODES];
>>
>> Not to be *too* picky, but there is a reason we declare nodemask_t as a
>> bitmap and not an array of bools. Isn't this just wasteful?
>>
>>> + unsigned long flags;
>>> + int cpu, hk_flags, j, id;
>>> const struct cpumask *mask;
>>>
>>> hk_flags = HK_FLAG_DOMAIN | HK_FLAG_MANAGED_IRQ;
>>> @@ -220,20 +256,28 @@ unsigned int cpumask_local_spread(unsigned int i, int node)
>>> return cpu;
>>> }
>>> } else {
>>> - /* NUMA first. */
>>> - for_each_cpu_and(cpu, cpumask_of_node(node), mask) {
>>> - if (i-- == 0)
>>> - return cpu;
>>> - }
>>> + spin_lock_irqsave(&spread_lock, flags);
>>> + memset(used, 0, nr_node_ids * sizeof(bool));
>>> + calc_node_distance(node_dist, node);
>>> + /* Local node first then the nearest node is used */
>>
>> Is this comment really correct? This makes it sound like there is only
>
> I think it is correct, that's what we want to choose the nearest node.
>
>> fallback to a single node. Doesn't the _code_ fall back basically
>> without limit?
>
> If I follow your question correctly, without this patch, if the local
> node is used up, one random node will be choosed, right? Now we firstly
> choose the nearest node by the distance, if all nodes has been choosen,
> it will return the initial solution.
The comment makes it sound like the code does:
1. Do the local node
2. Do the next nearest node
3. Stop
In reality, I *think* it's more of a loop where it search
ever-increasing distances away from the local node.
I just think the comment needs to be made more precise.
>>> + for (j = 0; j < nr_node_ids; j++) {
>>> + id = find_nearest_node(node_dist, used);
>>> + if (id < 0)
>>> + break;
>>>
>>> - for_each_cpu(cpu, mask) {
>>> - /* Skip NUMA nodes, done above. */
>>> - if (cpumask_test_cpu(cpu, cpumask_of_node(node)))
>>> - continue;
>>> + for_each_cpu_and(cpu, cpumask_of_node(id), mask)
>>> + if (i-- == 0) {
>>> + spin_unlock_irqrestore(&spread_lock,
>>> + flags);
>>> + return cpu;
>>> + }
>>> + used[id] = 1;
>>> + }
>>> + spin_unlock_irqrestore(&spread_lock, flags);
>>
>> The existing code was pretty sparsely commented. This looks to me to
>> make it more complicated and *less* commented. Not the best combo.
>
> Apologies for the bad comments, hopefully I describe it clearly by the above
> explantion.
Do you want to take another pass at submitting this patch?
next prev parent reply other threads:[~2020-11-13 16:02 UTC|newest]
Thread overview: 7+ messages / expand[flat|nested] mbox.gz Atom feed top
2020-11-03 13:39 [PATCH v6] lib: optimize cpumask_local_spread() Shaokun Zhang
2020-11-04 16:10 ` Dave Hansen
2020-11-13 2:06 ` Shaokun Zhang
2020-11-13 16:02 ` Dave Hansen [this message]
2020-11-16 7:59 ` Shaokun Zhang
2020-11-16 14:48 ` Dave Hansen
2020-11-17 1:12 ` Shaokun Zhang
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=5e8b0304-4de1-4bdc-41d2-79fa5464fbc7@intel.com \
--to=dave.hansen@intel.com \
--cc=akpm@linux-foundation.org \
--cc=anshuman.khandual@arm.com \
--cc=jgross@suse.com \
--cc=jinyuqi@huawei.com \
--cc=linux-kernel@vger.kernel.org \
--cc=mhocko@suse.com \
--cc=mpe@ellerman.id.au \
--cc=netdev@vger.kernel.org \
--cc=paul.burton@mips.com \
--cc=rppt@linux.ibm.com \
--cc=rusty@rustcorp.com.au \
--cc=zhangshaokun@hisilicon.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).