IB/hfi1: Invalid NUMA node information can cause a divide by zero
authorMichael J. Ruhl <michael.j.ruhl@intel.com>
Thu, 16 Aug 2018 05:54:49 +0000 (22:54 -0700)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Sat, 15 Sep 2018 07:45:27 +0000 (09:45 +0200)
[ Upstream commit c513de490f808d8480346f9a58e6a4a5f3de12e7 ]

If the system BIOS does not supply NUMA node information to the
PCI devices, the NUMA node is selected by choosing the current
node.

This can lead to the following crash:

divide error: 0000 SMP
CPU: 0 PID: 4 Comm: kworker/0:0 Tainted: G          IOE
------------   3.10.0-693.21.1.el7.x86_64 #1
Hardware name: Intel Corporation S2600KP/S2600KP, BIOS
SE5C610.86B.01.01.0005.101720141054 10/17/2014
Workqueue: events work_for_cpu_fn
task: ffff880174480fd0 ti: ffff880174488000 task.ti: ffff880174488000
RIP: 0010: [<ffffffffc020ac69>] hfi1_dev_affinity_init+0x129/0x6a0 [hfi1]
RSP: 0018:ffff88017448bbf8  EFLAGS: 00010246
RAX: 0000000000000011 RBX: ffff88107ffba6c0 RCX: ffff88085c22e130
RDX: 0000000000000000 RSI: 0000000000000000 RDI: ffff880824ad0000
RBP: ffff88017448bc48 R08: 0000000000000011 R09: 0000000000000002
R10: ffff8808582b6ca0 R11: 0000000000003151 R12: ffff8808582b6ca0
R13: ffff8808582b6518 R14: ffff8808582b6010 R15: 0000000000000012
FS:  0000000000000000(0000) GS:ffff88085ec00000(0000) knlGS:0000000000000000
CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 00007efc707404f0 CR3: 0000000001a02000 CR4: 00000000001607f0
DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
DR3: 0000000000000000 DR6: 00000000ffff0ff0 DR7: 0000000000000400
Call Trace:
 hfi1_init_dd+0x14b3/0x27a0 [hfi1]
 ? pcie_capability_write_word+0x46/0x70
 ? hfi1_pcie_init+0xc0/0x200 [hfi1]
 do_init_one+0x153/0x4c0 [hfi1]
 ? sched_clock_cpu+0x85/0xc0
 init_one+0x1b5/0x260 [hfi1]
 local_pci_probe+0x4a/0xb0
 work_for_cpu_fn+0x1a/0x30
 process_one_work+0x17f/0x440
 worker_thread+0x278/0x3c0
 ? manage_workers.isra.24+0x2a0/0x2a0
 kthread+0xd1/0xe0
 ? insert_kthread_work+0x40/0x40
 ret_from_fork+0x77/0xb0
 ? insert_kthread_work+0x40/0x40

If the BIOS is not supplying NUMA information:
  - set the default table count to 1 for all possible nodes
  - select node 0 (instead of current NUMA) node to get consistent
    performance
  - generate an error indicating that the BIOS should be upgraded

Reviewed-by: Gary Leshner <gary.s.leshner@intel.com>
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Michael J. Ruhl <michael.j.ruhl@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Sasha Levin <alexander.levin@microsoft.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
drivers/infiniband/hw/hfi1/affinity.c

index b5fab55cc275068166369c184a759ab9a925849d..b197e925fe3639dfc33b4ce373876ee0a50f152a 100644 (file)
@@ -146,7 +146,7 @@ int node_affinity_init(void)
                while ((dev = pci_get_device(ids->vendor, ids->device, dev))) {
                        node = pcibus_to_node(dev->bus);
                        if (node < 0)
-                               node = numa_node_id();
+                               goto out;
 
                        hfi1_per_node_cntr[node]++;
                }
@@ -154,6 +154,18 @@ int node_affinity_init(void)
        }
 
        return 0;
+
+out:
+       /*
+        * Invalid PCI NUMA node information found, note it, and populate
+        * our database 1:1.
+        */
+       pr_err("HFI: Invalid PCI NUMA node. Performance may be affected\n");
+       pr_err("HFI: System BIOS may need to be upgraded\n");
+       for (node = 0; node < node_affinity.num_possible_nodes; node++)
+               hfi1_per_node_cntr[node] = 1;
+
+       return 0;
 }
 
 void node_affinity_destroy(void)
@@ -227,8 +239,14 @@ int hfi1_dev_affinity_init(struct hfi1_devdata *dd)
        const struct cpumask *local_mask;
        int curr_cpu, possible, i;
 
-       if (node < 0)
-               node = numa_node_id();
+       /*
+        * If the BIOS does not have the NUMA node information set, select
+        * NUMA 0 so we get consistent performance.
+        */
+       if (node < 0) {
+               dd_dev_err(dd, "Invalid PCI NUMA node. Performance may be affected\n");
+               node = 0;
+       }
        dd->node = node;
 
        local_mask = cpumask_of_node(dd->node);