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;
MIME-Version: 1.0
References: <1543892757-4323-1-git-send-email-kernelfans@gmail.com> <20181204072251.GT31738@dhcp22.suse.cz>
In-Reply-To: <20181204072251.GT31738@dhcp22.suse.cz>
From:   Pingfan Liu <kernelfans@gmail.com>
Date:   Tue, 4 Dec 2018 16:20:32 +0800
Message-ID: <CAFgQCTv56drDBx-sTr6KdeQNKJnojG3g_a-k8wKe_q2y9w9NtA@mail.gmail.com>
Subject: Re: [PATCH] mm/alloc: fallback to first node if the wanted node offline
To:     mhocko@kernel.org
Cc:     linux-mm@kvack.org, linux-kernel@vger.kernel.org,
        Andrew Morton <akpm@linux-foundation.org>,
        Vlastimil Babka <vbabka@suse.cz>,
        Mike Rapoport <rppt@linux.vnet.ibm.com>,
        Bjorn Helgaas <bhelgaas@google.com>,
        Jonathan Cameron <Jonathan.Cameron@huawei.com>
Content-Type: text/plain; charset="UTF-8"
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Tue, Dec 4, 2018 at 3:22 PM Michal Hocko <mhocko@kernel.org> wrote:
>
> On Tue 04-12-18 11:05:57, Pingfan Liu wrote:
> > During my test on some AMD machine, with kexec -l nr_cpus=x option, the
> > kernel failed to bootup, because some node's data struct can not be allocated,
> > e.g, on x86, initialized by init_cpu_to_node()->init_memory_less_node(). But
> > device->numa_node info is used as preferred_nid param for
> > __alloc_pages_nodemask(), which causes NULL reference
> >   ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
> > This patch tries to fix the issue by falling back to the first online node,
> > when encountering such corner case.
>
> We have seen similar issues already and the bug was usually that the
> zonelists were not initialized yet or the node is completely bogus.
> Zonelists should be initialized by build_all_zonelists quite early so I
> am wondering whether the later is the case. What is the actual node
> number the device is associated with?
>
The device's node num is 2. And in my case, I used nr_cpus param. Due
to init_cpu_to_node() initialize all the possible node.  It is hard
for me to figure out without this param, how zonelists is accessed
before page allocator works.

> Your patch is not correct btw, because we want to fallback into the node in
> the distance order rather into the first online node.
> --
What about this:
+extern int find_next_best_node(int node, nodemask_t *used_node_mask);
+
 /*
  * We get the zone list from the current node and the gfp_mask.
  * This zone list contains a maximum of MAXNODES*MAX_NR_ZONES zones.
@@ -453,6 +455,11 @@ static inline int gfp_zonelist(gfp_t flags)
  */
 static inline struct zonelist *node_zonelist(int nid, gfp_t flags)
 {
+       if (unlikely(!node_online(nid))) {
+               nodemask_t used_mask;
+               nodes_complement(used_mask, node_online_map);
+               nid = find_next_best_node(nid, &used_mask);
+       }
        return NODE_DATA(nid)->node_zonelists + gfp_zonelist(flags);
 }

I just finished the compiling, not test it yet, since the machine is
not on hand yet. It needs some time to get it again.

Thanks,
Pingfan