NULL pointer dereference in nvmet_rdma_queue_disconnect during bond failover

NULL pointer dereference in nvmet_rdma_queue_disconnect during bond failover

[Alex Lyakas]

Greetings NVMe community,

I am running kernel 5.1.6, which is the latest stable kernel.

I am testing a nvmf kernel target, configured on top of a bond interface, 
for high availability. The bond interface is created on top of two 
ConnectX-3 interfaces, which represent two ports of one ConnectX-3 VF (with 
this hardware a VF is dual-ported, i.e., a single VF yields two network 
interfaces). The bond is configured in active-backup mode. Exact bonding 
configuration is given in [1]. The nvmet target configuration doesn't have 
anything special and is given in [2].

I create a nvmf connection from a different machine to the nvmet target. 
Then I initiate bond failover, by disconnecting a cable that corresponds to 
the active bond slave. As a result, I get the following kernel panic:

[  268.036732] mlx4_en: b1s1: Link Down
[  268.036739] mlx4_en: b0s1: Link Down
[  268.036771] mlx4_en: b2s1: Link Down
[  268.138594] bebond: link status definitely down for interface b1s1, 
disabling it
[  268.138597] bebond: making interface b1s0 the new active one 53500 ms 
earlier
[  268.138671] RDMA CM addr change for ndev bebond used by id 
0000000019666fc8
[  268.138673] RDMA CM addr change for ndev bebond used by id 
000000007a8dd02e
[  268.138674] RDMA CM addr change for ndev bebond used by id 
00000000f825cc30
[  268.138675] RDMA CM addr change for ndev bebond used by id 
00000000c575ce3d
[  268.138733] BUG: unable to handle kernel NULL pointer dereference at 
0000000000000148
[  268.138764] #PF error: [normal kernel read fault]
[  268.138782] PGD 0 P4D 0
[  268.138795] Oops: 0000 [#1] SMP PTI
[  268.138811] CPU: 1 PID: 869 Comm: kworker/u4:5 Not tainted 
5.1.6-050106-generic #201905311031
[  268.138839] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 
1.10.2-1ubuntu1 04/01/2014
[  268.138885] Workqueue: rdma_cm cma_ndev_work_handler [rdma_cm]
[  268.138912] RIP: 0010:nvmet_rdma_queue_disconnect+0x19/0x80 [nvmet_rdma]
[  268.138937] Code: e8 bc fe ff ff e9 68 ff ff ff 0f 1f 80 00 00 00 00 66 
66 66 66 90 55 48 89 e5 53 48 89 fb 48 c7 c7 80 10 86 c0 e8 57 1d ff d1 <48> 
8b 93 48 01 00 00 48 8d 83 48 01 00 00 48 39 d0 74 3a 48 8b 8b
[  268.139020] RSP: 0018:ffffb28a0111be08 EFLAGS: 00010246
[  268.139712] RAX: 0000000000000000 RBX: 0000000000000000 RCX: 
0000000000000000
[  268.140348] RDX: ffff9cc2a7c15c00 RSI: 000000000000000e RDI: 
ffffffffc0861080
[  268.140764] RBP: ffffb28a0111be10 R08: ffff9cc2a7c15c00 R09: 
000000000000008c
[  268.141195] R10: 00000000000001ed R11: 0000000000000001 R12: 
ffff9cc2a7c54aa8
[  268.141616] R13: ffff9cc2a9b55800 R14: ffff9cc2a7c54a80 R15: 
0ffff9cc2a78ee60
[  268.142057] FS:  0000000000000000(0000) GS:ffff9cc2b9b00000(0000) 
knlGS:0000000000000000
[  268.142520] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[  268.142962] CR2: 0000000000000148 CR3: 00000001afe30004 CR4: 
00000000000606e0
[  268.143430] Call Trace:
[  268.143880]  nvmet_rdma_cm_handler+0x94/0x292 [nvmet_rdma]
[  268.144343]  cma_ndev_work_handler+0x45/0xb0 [rdma_cm]
[  268.144792]  process_one_work+0x20f/0x410
[  268.145246]  worker_thread+0x34/0x400
[  268.145689]  kthread+0x120/0x140
[  268.146141]  ? process_one_work+0x410/0x410
[  268.146595]  ? __kthread_parkme+0x70/0x70
[  268.147045]  ret_from_fork+0x35/0x40

This is 100% reproducible.

Thanks,
Alex.

[1]
echo +bebond >/sys/class/net/bonding_masters
echo "1" > /proc/sys/net/ipv6/conf/bebond/disable_ipv6
echo "1" > /sys/class/net/bebond/bonding/mode
echo "100" > /sys/class/net/bebond/bonding/miimon
echo "1" > /sys/class/net/bebond/bonding/fail_over_mac
echo "60000" > /sys/class/net/bebond/bonding/updelay
ifconfig b1s1 down
echo "+b1s1" > /sys/class/net/bebond/bonding/slaves
ifconfig b1s0 down
echo "+b1s0" > /sys/class/net/bebond/bonding/slaves
echo "b1s1" > /sys/class/net/bebond/bonding/primary
ip addr add 10.3.3.23/24 dev bebond

[2]
mkdir /sys/kernel/config/nvmet/subsystems/volume-55555555
echo 1 > 
/sys/kernel/config/nvmet/subsystems/volume-55555555/attr_allow_any_host
echo 000055555555 > 
/sys/kernel/config/nvmet/subsystems/volume-55555555/attr_serial
mkdir /sys/kernel/config/nvmet/subsystems/volume-55555555/namespaces/1
echo 0977dff3-6885-43b3-a948-000055555555 > 
/sys/kernel/config/nvmet/subsystems/volume-55555555/namespaces/1/device_uuid
echo -n /dev/loop0 > 
/sys/kernel/config/nvmet/subsystems/volume-55555555/namespaces/1/device_path
echo 1 > 
/sys/kernel/config/nvmet/subsystems/volume-55555555/namespaces/1/enable

mkdir /sys/kernel/config/nvmet/ports/1
echo -n "ipv4" > /sys/kernel/config/nvmet/ports/1/addr_adrfam
echo -n "rdma" > /sys/kernel/config/nvmet/ports/1/addr_trtype
echo -n  10.3.3.23 > /sys/kernel/config/nvmet/ports/1/addr_traddr
echo -n 4420 > /sys/kernel/config/nvmet/ports/1/addr_trsvcid
ln -s /sys/kernel/config/nvmet/subsystems/volume-55555555 
/sys/kernel/config/nvmet/ports/1/subsystems/

NULL pointer dereference in nvmet_rdma_queue_disconnect during bond failover

[Sagi Grimberg]

> Greetings NVMe community,
> 
> I am running kernel 5.1.6, which is the latest stable kernel.
> 
> I am testing a nvmf kernel target, configured on top of a bond 
> interface, for high availability. The bond interface is created on top 
> of two ConnectX-3 interfaces, which represent two ports of one 
> ConnectX-3 VF (with this hardware a VF is dual-ported, i.e., a single VF 
> yields two network interfaces). The bond is configured in active-backup 
> mode. Exact bonding configuration is given in [1]. The nvmet target 
> configuration doesn't have anything special and is given in [2].
> 
> I create a nvmf connection from a different machine to the nvmet target. 
> Then I initiate bond failover, by disconnecting a cable that corresponds 
> to the active bond slave. As a result, I get the following kernel panic:

Max sent a fix exactly for this. You can test that it works for you
when he sends v2.

Max, care to CC Alex when you send it?

NULL pointer dereference in nvmet_rdma_queue_disconnect during bond failover

[Max Gurtovoy]

On 6/6/2019 3:05 AM, Sagi Grimberg wrote:
>
>> Greetings NVMe community,
>>
>> I am running kernel 5.1.6, which is the latest stable kernel.
>>
>> I am testing a nvmf kernel target, configured on top of a bond 
>> interface, for high availability. The bond interface is created on 
>> top of two ConnectX-3 interfaces, which represent two ports of one 
>> ConnectX-3 VF (with this hardware a VF is dual-ported, i.e., a single 
>> VF yields two network interfaces). The bond is configured in 
>> active-backup mode. Exact bonding configuration is given in [1]. The 
>> nvmet target configuration doesn't have anything special and is given 
>> in [2].
>>
>> I create a nvmf connection from a different machine to the nvmet 
>> target. Then I initiate bond failover, by disconnecting a cable that 
>> corresponds to the active bond slave. As a result, I get the 
>> following kernel panic:
>
> Max sent a fix exactly for this. You can test that it works for you
> when he sends v2.
>
> Max, care to CC Alex when you send it?

Sure, No problem.


>
> _______________________________________________
> Linux-nvme mailing list
> Linux-nvme at lists.infradead.org
> https://eur03.safelinks.protection.outlook.com/?url=http%3A%2F%2Flists.infradead.org%2Fmailman%2Flistinfo%2Flinux-nvme&data=02%7C01%7Cmaxg%40mellanox.com%7C9879366f21df4e7f4ce608d6ea12c41f%7Ca652971c7d2e4d9ba6a4d149256f461b%7C0%7C0%7C636953763648159480&sdata=UtkpvcEM9%2BBBdj68Kx6bobMWkeGSp1Jz3yJXxH8MLgY%3D&reserved=0 
>

NULL pointer dereference in nvmet_rdma_queue_disconnect during bond failover

[Alex Lyakas]

Hi Max,

Has any patch been sent to resolve the kernel panic in nvmet that we are 
seeing?

Thanks,
Alex.


-----Original Message----- 
From: Max Gurtovoy
Sent: Thursday, June 06, 2019 10:31 AM
To: linux-nvme at lists.infradead.org
Subject: Re: NULL pointer dereference in nvmet_rdma_queue_disconnect during 
bond failover


On 6/6/2019 3:05 AM, Sagi Grimberg wrote:
>
>> Greetings NVMe community,
>>
>> I am running kernel 5.1.6, which is the latest stable kernel.
>>
>> I am testing a nvmf kernel target, configured on top of a bond interface, 
>> for high availability. The bond interface is created on top of two 
>> ConnectX-3 interfaces, which represent two ports of one ConnectX-3 VF 
>> (with this hardware a VF is dual-ported, i.e., a single VF yields two 
>> network interfaces). The bond is configured in active-backup mode. Exact 
>> bonding configuration is given in [1]. The nvmet target configuration 
>> doesn't have anything special and is given in [2].
>>
>> I create a nvmf connection from a different machine to the nvmet target. 
>> Then I initiate bond failover, by disconnecting a cable that corresponds 
>> to the active bond slave. As a result, I get the following kernel panic:
>
> Max sent a fix exactly for this. You can test that it works for you
> when he sends v2.
>
> Max, care to CC Alex when you send it?

Sure, No problem.


>
> _______________________________________________
> Linux-nvme mailing list
> Linux-nvme at lists.infradead.org
> https://eur03.safelinks.protection.outlook.com/?url=http%3A%2F%2Flists.infradead.org%2Fmailman%2Flistinfo%2Flinux-nvme&data=02%7C01%7Cmaxg%40mellanox.com%7C9879366f21df4e7f4ce608d6ea12c41f%7Ca652971c7d2e4d9ba6a4d149256f461b%7C0%7C0%7C636953763648159480&sdata=UtkpvcEM9%2BBBdj68Kx6bobMWkeGSp1Jz3yJXxH8MLgY%3D&reserved=0

_______________________________________________
Linux-nvme mailing list
Linux-nvme at lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme 

NULL pointer dereference in nvmet_rdma_queue_disconnect during bond failover

[Max Gurtovoy]

Hi Alex,

Not yet. Our fix is in the Initiator/Host side and it was merged.

This is on our plate.

In case you would like to send a patch to solve this, we'll review it 
of-course.

-Max.

On 7/3/2019 12:28 PM, Alex Lyakas wrote:
> Hi Max,
>
> Has any patch been sent to resolve the kernel panic in nvmet that we 
> are seeing?
>
> Thanks,
> Alex.
>
>
> -----Original Message----- From: Max Gurtovoy
> Sent: Thursday, June 06, 2019 10:31 AM
> To: linux-nvme at lists.infradead.org
> Subject: Re: NULL pointer dereference in nvmet_rdma_queue_disconnect 
> during bond failover
>
>
> On 6/6/2019 3:05 AM, Sagi Grimberg wrote:
>>
>>> Greetings NVMe community,
>>>
>>> I am running kernel 5.1.6, which is the latest stable kernel.
>>>
>>> I am testing a nvmf kernel target, configured on top of a bond 
>>> interface, for high availability. The bond interface is created on 
>>> top of two ConnectX-3 interfaces, which represent two ports of one 
>>> ConnectX-3 VF (with this hardware a VF is dual-ported, i.e., a 
>>> single VF yields two network interfaces). The bond is configured in 
>>> active-backup mode. Exact bonding configuration is given in [1]. The 
>>> nvmet target configuration doesn't have anything special and is 
>>> given in [2].
>>>
>>> I create a nvmf connection from a different machine to the nvmet 
>>> target. Then I initiate bond failover, by disconnecting a cable that 
>>> corresponds to the active bond slave. As a result, I get the 
>>> following kernel panic:
>>
>> Max sent a fix exactly for this. You can test that it works for you
>> when he sends v2.
>>
>> Max, care to CC Alex when you send it?
>
> Sure, No problem.
>
>
>>
>> _______________________________________________
>> Linux-nvme mailing list
>> Linux-nvme at lists.infradead.org
>> https://eur03.safelinks.protection.outlook.com/?url=http%3A%2F%2Flists.infradead.org%2Fmailman%2Flistinfo%2Flinux-nvme&data=02%7C01%7Cmaxg%40mellanox.com%7Cb0df220121534ce1fe4b08d6ff98e119%7Ca652971c7d2e4d9ba6a4d149256f461b%7C0%7C0%7C636977429409072070&sdata=nwWwML3PUy%2BZ0HJeJRKpNyLB0Nm%2BV%2BUxNuweaAWyeyA%3D&reserved=0 
>>
>
> _______________________________________________
> Linux-nvme mailing list
> Linux-nvme at lists.infradead.org
> https://eur03.safelinks.protection.outlook.com/?url=http%3A%2F%2Flists.infradead.org%2Fmailman%2Flistinfo%2Flinux-nvme&data=02%7C01%7Cmaxg%40mellanox.com%7Cb0df220121534ce1fe4b08d6ff98e119%7Ca652971c7d2e4d9ba6a4d149256f461b%7C0%7C0%7C636977429409072070&sdata=nwWwML3PUy%2BZ0HJeJRKpNyLB0Nm%2BV%2BUxNuweaAWyeyA%3D&reserved=0 
>

NULL pointer dereference in nvmet_rdma_queue_disconnect during bond failover

[Sagi Grimberg]

> Hi Alex,
> 
> Not yet. Our fix is in the Initiator/Host side and it was merged.
> 
> This is on our plate.
> 
> In case you would like to send a patch to solve this, we'll review it 
> of-course.

Does the attached untested patch help?
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 0001-nvmet-rdma-fix-bonding-failover-possible-NULL-deref.patch
Type: text/x-patch
Size: 8365 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/linux-nvme/attachments/20190703/c697655b/attachment-0001.bin>

NULL pointer dereference in nvmet_rdma_queue_disconnect during bond failover

[Sagi Grimberg]

>> Hi Alex,
>>
>> Not yet. Our fix is in the Initiator/Host side and it was merged.
>>
>> This is on our plate.
>>
>> In case you would like to send a patch to solve this, we'll review it 
>> of-course.
> 
> Does the attached untested patch help?

Alex? Max?

NULL pointer dereference in nvmet_rdma_queue_disconnect during bond failover

[Alex Lyakas]

Hi Sagi,

Which kernel this patch applies to?

At this point the environment I used for nvmf evaluation is not available 
for me. I will make an effort to test this patch, and get back to you.

Thanks,
Alex,


-----Original Message----- 
From: Sagi Grimberg
Sent: Friday, July 12, 2019 10:38 PM
To: Max Gurtovoy ; Alex Lyakas ; linux-nvme at lists.infradead.org ; Shlomi 
Nimrodi ; Israel Rukshin ; tomwu at mellanox.com
Subject: Re: NULL pointer dereference in nvmet_rdma_queue_disconnect during 
bond failover


>> Hi Alex,
>>
>> Not yet. Our fix is in the Initiator/Host side and it was merged.
>>
>> This is on our plate.
>>
>> In case you would like to send a patch to solve this, we'll review it 
>> of-course.
>
> Does the attached untested patch help?

Alex? Max? 

NULL pointer dereference in nvmet_rdma_queue_disconnect during bond failover

[Sagi Grimberg]

> Hi Sagi,
> 
> Which kernel this patch applies to?

its based on the nvme tree, but it should apply cleanly on upstream
5.2...

NULL pointer dereference in nvmet_rdma_queue_disconnect during bond failover

[Sagi Grimberg]

>> Hi Sagi,
>>
>> Which kernel this patch applies to?
> 
> its based on the nvme tree, but it should apply cleanly on upstream
> 5.2...

Alex, Max? did you retest this?

Re: NULL pointer dereference in nvmet_rdma_queue_disconnect during bond failover

[Sagi Grimberg]

>>> Hi Sagi,
>>>
>>> Which kernel this patch applies to?
>>
>> its based on the nvme tree, but it should apply cleanly on upstream
>> 5.2...
> 
> Alex, Max? did you retest this?

Raising this from the ashes...

Alex, did you test this patch?

_______________________________________________
Linux-nvme mailing list
Linux-nvme@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme

Re: NULL pointer dereference in nvmet_rdma_queue_disconnect during bond failover

[Alex Lyakas]

Hi Sagi,

>
>
> >>> Hi Sagi,
> >>>
> >>> Which kernel this patch applies to?
> >>
> >> its based on the nvme tree, but it should apply cleanly on upstream
> >> 5.2...
> >
> > Alex, Max? did you retest this?
>
> Raising this from the ashes...
>
> Alex, did you test this patch?

Raising from the ashes!

In short: this patch fixes the issue!

More details:

This patch doesn't apply on kernel 5.2. Moreover, I believe this patch
is incomplete, because nvmet_rdma_find_get_device() needs to be fixed
to treat cm_id->context as "struct nvmet_rdma_port" and not as "struct
nvmet_port".

However, since we are working with kernel modules from Mellanox OFED,
I tried applying this patch on OFED 4.7. I discovered that it already
has almost everything this patch introduces. Like "struct
nvmet_rdma_port" and the refactoring of nvmet_rdma_add_port into
nvmet_rdma_enable_port, and nvmet_rdma_remove_port to
nvmet_rdma_disable_port. I ended up with this patch [1].

Tested bond failover, and cm_id is destroyed and re-created as expected [2]

Israel, Max and other Mellanox folks: can we have this fix in OFED 4.9?

Thanks,
Alex.


[1]
diff -ru mlnx-nvme-4.7-orig/target/rdma.c mlnx-nvme-4.7/target/rdma.c
--- mlnx-nvme-4.7-orig/target/rdma.c    2020-01-15 09:58:59.000000000 +0200
+++ mlnx-nvme-4.7/target/rdma.c    2020-03-30 20:49:49.932479383 +0300
@@ -191,6 +191,7 @@
                 struct nvmet_rdma_rsp *r);
 static int nvmet_rdma_alloc_rsp(struct nvmet_rdma_device *ndev,
                 struct nvmet_rdma_rsp *r);
+static void nvmet_rdma_disable_port(struct nvmet_rdma_port *port);

 static const struct nvmet_fabrics_ops nvmet_rdma_ops;

@@ -1544,6 +1545,13 @@
         nvmet_rdma_queue_established(queue);
         break;
     case RDMA_CM_EVENT_ADDR_CHANGE:
+        if (!queue) {
+            struct nvmet_rdma_port *port = cm_id->context;
+
+            pr_warn("RDMA_CM_EVENT_ADDR_CHANGE: cm_id=%p schedule
enable_work\n", cm_id);
+            schedule_delayed_work(&port->enable_work, 0);
+            break;
+        }
     case RDMA_CM_EVENT_DISCONNECTED:
     case RDMA_CM_EVENT_TIMEWAIT_EXIT:
         nvmet_rdma_queue_disconnect(queue);
@@ -1598,6 +1606,8 @@
         return PTR_ERR(cm_id);
     }

+    pr_info("nvmet_rdma_enable_port: created cm_id=%p\n", cm_id);
+
     /*
      * Allow both IPv4 and IPv6 sockets to bind a single port
      * at the same time.
@@ -1620,7 +1630,7 @@
         goto out_destroy_id;
     }

-    port->cm_id = cm_id;
+    xchg(&port->cm_id, cm_id);
     if (cm_id->device)
         port->node_guid = cm_id->device->node_guid;

@@ -1640,6 +1650,7 @@
             struct nvmet_rdma_port, enable_work);
     int ret;

+    nvmet_rdma_disable_port(port);
     ret = nvmet_rdma_enable_port(port);
     if (ret)
         schedule_delayed_work(&port->enable_work, 5 * HZ);
@@ -1707,13 +1718,14 @@

 static void nvmet_rdma_disable_port(struct nvmet_rdma_port *port)
 {
-    struct rdma_cm_id *cm_id = port->cm_id;
+    struct rdma_cm_id *cm_id = xchg(&port->cm_id, NULL);
     struct nvmet_port *nport = port->nport;

+    pr_info("nvmet_rdma_disable_port: cm_id=%p\n", cm_id);
+
     if (nport->offload && cm_id)
         nvmet_rdma_destroy_xrqs(nport);

-    port->cm_id = NULL;
     if (cm_id)
         rdma_destroy_id(cm_id);
 }


[2]
Mar 30 21:57:48.030761 qa3-sn2 kernel: [95220.661707] bebond: making
interface be10G2 the new active one
Mar 30 21:57:48.030789 qa3-sn2 kernel: [95220.662003] RDMA CM addr
change for ndev bebond used by id ffff966432c63000
Mar 30 21:57:48.030793 qa3-sn2 kernel: [95220.662007] RDMA CM addr
change for ndev bebond used by id ffff966a6ee85800
Mar 30 21:57:48.030817 qa3-sn2 kernel: [95220.662010] RDMA CM addr
change for ndev bebond used by id ffff966a6ee87400
Mar 30 21:57:48.030821 qa3-sn2 kernel: [95220.662012] RDMA CM addr
change for ndev bebond used by id ffff966a6ee85400
Mar 30 21:57:48.030824 qa3-sn2 kernel: [95220.662015] RDMA CM addr
change for ndev bebond used by id ffff966a6ee83c00
Mar 30 21:57:48.030827 qa3-sn2 kernel: [95220.662017] RDMA CM addr
change for ndev bebond used by id ffff966a6ee84c00
Mar 30 21:57:48.030829 qa3-sn2 kernel: [95220.662025] nvmet_rdma:
RDMA_CM_EVENT_ADDR_CHANGE: cm_id=ffff966432c63000 schedule enable_work
Mar 30 21:57:48.030832 qa3-sn2 kernel: [95220.662069] nvmet_rdma:
nvmet_rdma_disable_port: cm_id=ffff966432c63000
Mar 30 21:57:48.030834 qa3-sn2 kernel: [95220.662093] nvmet_rdma:
nvmet_rdma_enable_port: created cm_id=ffff96658fdab800
Mar 30 21:57:48.030837 qa3-sn2 kernel: [95220.662120] nvmet_rdma:
enabling port 1 (10.3.3.3:4420)
Mar 30 21:57:50.266755 qa3-sn2 kernel: [95222.897752] nvmet: creating
controller 1 for subsystem
nqn.2011-04.com.zadarastorage:volume-00000010 for NQN
iqn.2011-04.com.zadarastorage:2:vc-1.

_______________________________________________
linux-nvme mailing list
linux-nvme@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme

Re: NULL pointer dereference in nvmet_rdma_queue_disconnect during bond failover

[Max Gurtovoy]

On 3/30/2020 10:02 PM, Alex Lyakas wrote:
> Hi Sagi,
>
>>
>>>>> Hi Sagi,
>>>>>
>>>>> Which kernel this patch applies to?
>>>> its based on the nvme tree, but it should apply cleanly on upstream
>>>> 5.2...
>>> Alex, Max? did you retest this?
>> Raising this from the ashes...
>>
>> Alex, did you test this patch?
> Raising from the ashes!
>
> In short: this patch fixes the issue!
>
> More details:
>
> This patch doesn't apply on kernel 5.2. Moreover, I believe this patch
> is incomplete, because nvmet_rdma_find_get_device() needs to be fixed
> to treat cm_id->context as "struct nvmet_rdma_port" and not as "struct
> nvmet_port".
>
> However, since we are working with kernel modules from Mellanox OFED,
> I tried applying this patch on OFED 4.7. I discovered that it already
> has almost everything this patch introduces. Like "struct
> nvmet_rdma_port" and the refactoring of nvmet_rdma_add_port into
> nvmet_rdma_enable_port, and nvmet_rdma_remove_port to
> nvmet_rdma_disable_port. I ended up with this patch [1].
>
> Tested bond failover, and cm_id is destroyed and re-created as expected [2]
>
> Israel, Max and other Mellanox folks: can we have this fix in OFED 4.9?

Alex,

We first need to fix this issue in upstream.

hopefully we can get to it soon.


>
> Thanks,
> Alex.
>
>
> [1]
> diff -ru mlnx-nvme-4.7-orig/target/rdma.c mlnx-nvme-4.7/target/rdma.c
> --- mlnx-nvme-4.7-orig/target/rdma.c    2020-01-15 09:58:59.000000000 +0200
> +++ mlnx-nvme-4.7/target/rdma.c    2020-03-30 20:49:49.932479383 +0300
> @@ -191,6 +191,7 @@
>                   struct nvmet_rdma_rsp *r);
>   static int nvmet_rdma_alloc_rsp(struct nvmet_rdma_device *ndev,
>                   struct nvmet_rdma_rsp *r);
> +static void nvmet_rdma_disable_port(struct nvmet_rdma_port *port);
>
>   static const struct nvmet_fabrics_ops nvmet_rdma_ops;
>
> @@ -1544,6 +1545,13 @@
>           nvmet_rdma_queue_established(queue);
>           break;
>       case RDMA_CM_EVENT_ADDR_CHANGE:
> +        if (!queue) {
> +            struct nvmet_rdma_port *port = cm_id->context;
> +
> +            pr_warn("RDMA_CM_EVENT_ADDR_CHANGE: cm_id=%p schedule
> enable_work\n", cm_id);
> +            schedule_delayed_work(&port->enable_work, 0);
> +            break;
> +        }
>       case RDMA_CM_EVENT_DISCONNECTED:
>       case RDMA_CM_EVENT_TIMEWAIT_EXIT:
>           nvmet_rdma_queue_disconnect(queue);
> @@ -1598,6 +1606,8 @@
>           return PTR_ERR(cm_id);
>       }
>
> +    pr_info("nvmet_rdma_enable_port: created cm_id=%p\n", cm_id);
> +
>       /*
>        * Allow both IPv4 and IPv6 sockets to bind a single port
>        * at the same time.
> @@ -1620,7 +1630,7 @@
>           goto out_destroy_id;
>       }
>
> -    port->cm_id = cm_id;
> +    xchg(&port->cm_id, cm_id);
>       if (cm_id->device)
>           port->node_guid = cm_id->device->node_guid;
>
> @@ -1640,6 +1650,7 @@
>               struct nvmet_rdma_port, enable_work);
>       int ret;
>
> +    nvmet_rdma_disable_port(port);
>       ret = nvmet_rdma_enable_port(port);
>       if (ret)
>           schedule_delayed_work(&port->enable_work, 5 * HZ);
> @@ -1707,13 +1718,14 @@
>
>   static void nvmet_rdma_disable_port(struct nvmet_rdma_port *port)
>   {
> -    struct rdma_cm_id *cm_id = port->cm_id;
> +    struct rdma_cm_id *cm_id = xchg(&port->cm_id, NULL);
>       struct nvmet_port *nport = port->nport;
>
> +    pr_info("nvmet_rdma_disable_port: cm_id=%p\n", cm_id);
> +
>       if (nport->offload && cm_id)
>           nvmet_rdma_destroy_xrqs(nport);
>
> -    port->cm_id = NULL;
>       if (cm_id)
>           rdma_destroy_id(cm_id);
>   }
>
>
> [2]
> Mar 30 21:57:48.030761 qa3-sn2 kernel: [95220.661707] bebond: making
> interface be10G2 the new active one
> Mar 30 21:57:48.030789 qa3-sn2 kernel: [95220.662003] RDMA CM addr
> change for ndev bebond used by id ffff966432c63000
> Mar 30 21:57:48.030793 qa3-sn2 kernel: [95220.662007] RDMA CM addr
> change for ndev bebond used by id ffff966a6ee85800
> Mar 30 21:57:48.030817 qa3-sn2 kernel: [95220.662010] RDMA CM addr
> change for ndev bebond used by id ffff966a6ee87400
> Mar 30 21:57:48.030821 qa3-sn2 kernel: [95220.662012] RDMA CM addr
> change for ndev bebond used by id ffff966a6ee85400
> Mar 30 21:57:48.030824 qa3-sn2 kernel: [95220.662015] RDMA CM addr
> change for ndev bebond used by id ffff966a6ee83c00
> Mar 30 21:57:48.030827 qa3-sn2 kernel: [95220.662017] RDMA CM addr
> change for ndev bebond used by id ffff966a6ee84c00
> Mar 30 21:57:48.030829 qa3-sn2 kernel: [95220.662025] nvmet_rdma:
> RDMA_CM_EVENT_ADDR_CHANGE: cm_id=ffff966432c63000 schedule enable_work
> Mar 30 21:57:48.030832 qa3-sn2 kernel: [95220.662069] nvmet_rdma:
> nvmet_rdma_disable_port: cm_id=ffff966432c63000
> Mar 30 21:57:48.030834 qa3-sn2 kernel: [95220.662093] nvmet_rdma:
> nvmet_rdma_enable_port: created cm_id=ffff96658fdab800
> Mar 30 21:57:48.030837 qa3-sn2 kernel: [95220.662120] nvmet_rdma:
> enabling port 1 (10.3.3.3:4420)
> Mar 30 21:57:50.266755 qa3-sn2 kernel: [95222.897752] nvmet: creating
> controller 1 for subsystem
> nqn.2011-04.com.zadarastorage:volume-00000010 for NQN
> iqn.2011-04.com.zadarastorage:2:vc-1.

_______________________________________________
linux-nvme mailing list
linux-nvme@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme

Re: NULL pointer dereference in nvmet_rdma_queue_disconnect during bond failover

[Sagi Grimberg]

Hey Alex,

>>> Alex, Max? did you retest this?
>>
>> Raising this from the ashes...
>>
>> Alex, did you test this patch?
> 
> Raising from the ashes!
> 
> In short: this patch fixes the issue!

Thanks for following up..

> 
> More details:
> 
> This patch doesn't apply on kernel 5.2. Moreover, I believe this patch
> is incomplete, because nvmet_rdma_find_get_device() needs to be fixed
> to treat cm_id->context as "struct nvmet_rdma_port" and not as "struct
> nvmet_port".

Does patch [1] apply on kernel 5.2?

> However, since we are working with kernel modules from Mellanox OFED,
> I tried applying this patch on OFED 4.7. I discovered that it already
> has almost everything this patch introduces. Like "struct
> nvmet_rdma_port" and the refactoring of nvmet_rdma_add_port into
> nvmet_rdma_enable_port, and nvmet_rdma_remove_port to
> nvmet_rdma_disable_port. I ended up with this patch [1].
> 
> Tested bond failover, and cm_id is destroyed and re-created as expected [2]
> 
> Israel, Max and other Mellanox folks: can we have this fix in OFED 4.9?
> 

For MOFED issues you can follow-up with Max and Israel offline. If you
can test upstream or even 5.2 stable that would be beneficial as I can
add your Tested-by tag.

Thanks.

[1]:
--
Author: Sagi Grimberg <sagi@grimberg.me>
Date:   Wed Jul 3 15:33:01 2019 -0700

     nvmet-rdma: fix bonding failover possible NULL deref

     RDMA_CM_EVENT_ADDR_CHANGE event occur in the case of bonding failover
     on normal as well as on listening cm_ids. Hence this event will
     immediately trigger a NULL dereference trying to disconnect a queue
     for a cm_id that actually belongs to the port.

     To fix this we provide a different handler for the listener cm_ids
     that will defer a work to disable+(re)enable the port which essentially
     destroys and setups another listener cm_id

     Reported-by: Alex Lyakas <alex@zadara.com>
     Signed-off-by: Sagi Grimberg <sagi@grimberg.me>

diff --git a/drivers/nvme/target/rdma.c b/drivers/nvme/target/rdma.c
index 9e1b8c61f54e..8dac89b7aa12 100644
--- a/drivers/nvme/target/rdma.c
+++ b/drivers/nvme/target/rdma.c
@@ -105,6 +105,13 @@ struct nvmet_rdma_queue {
         struct list_head        queue_list;
  };

+struct nvmet_rdma_port {
+       struct nvmet_port       *nport;
+       struct sockaddr_storage addr;
+       struct rdma_cm_id       *cm_id;
+       struct delayed_work     repair_work;
+};
+
  struct nvmet_rdma_device {
         struct ib_device        *device;
         struct ib_pd            *pd;
@@ -1272,6 +1279,7 @@ static int nvmet_rdma_cm_accept(struct rdma_cm_id 
*cm_id,
  static int nvmet_rdma_queue_connect(struct rdma_cm_id *cm_id,
                 struct rdma_cm_event *event)
  {
+       struct nvmet_rdma_port *port = cm_id->context;
         struct nvmet_rdma_device *ndev;
         struct nvmet_rdma_queue *queue;
         int ret = -EINVAL;
@@ -1287,7 +1295,7 @@ static int nvmet_rdma_queue_connect(struct 
rdma_cm_id *cm_id,
                 ret = -ENOMEM;
                 goto put_device;
         }
-       queue->port = cm_id->context;
+       queue->port = port->nport;

         if (queue->host_qid == 0) {
                 /* Let inflight controller teardown complete */
@@ -1412,7 +1420,7 @@ static void nvmet_rdma_queue_connect_fail(struct 
rdma_cm_id *cm_id,
  static int nvmet_rdma_device_removal(struct rdma_cm_id *cm_id,
                 struct nvmet_rdma_queue *queue)
  {
-       struct nvmet_port *port;
+       struct nvmet_rdma_port *port;

         if (queue) {
                 /*
@@ -1431,7 +1439,7 @@ static int nvmet_rdma_device_removal(struct 
rdma_cm_id *cm_id,
          * cm_id destroy. use atomic xchg to make sure
          * we don't compete with remove_port.
          */
-       if (xchg(&port->priv, NULL) != cm_id)
+       if (xchg(&port->cm_id, NULL) != cm_id)
                 return 0;

         /*
@@ -1462,6 +1470,13 @@ static int nvmet_rdma_cm_handler(struct 
rdma_cm_id *cm_id,
                 nvmet_rdma_queue_established(queue);
                 break;
         case RDMA_CM_EVENT_ADDR_CHANGE:
+               if (!queue) {
+                       struct nvmet_rdma_port *port = cm_id->context;
+
+                       schedule_delayed_work(&port->repair_work, 0);
+                       break;
+               }
+               /* FALLTHROUGH */
         case RDMA_CM_EVENT_DISCONNECTED:
         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
                 nvmet_rdma_queue_disconnect(queue);
@@ -1504,42 +1519,19 @@ static void nvmet_rdma_delete_ctrl(struct 
nvmet_ctrl *ctrl)
         mutex_unlock(&nvmet_rdma_queue_mutex);
  }

-static int nvmet_rdma_add_port(struct nvmet_port *port)
+static void nvmet_rdma_disable_port(struct nvmet_rdma_port *port)
  {
-       struct rdma_cm_id *cm_id;
-       struct sockaddr_storage addr = { };
-       __kernel_sa_family_t af;
-       int ret;
+       struct rdma_cm_id *cm_id = xchg(&port->cm_id, NULL);

-       switch (port->disc_addr.adrfam) {
-       case NVMF_ADDR_FAMILY_IP4:
-               af = AF_INET;
-               break;
-       case NVMF_ADDR_FAMILY_IP6:
-               af = AF_INET6;
-               break;
-       default:
-               pr_err("address family %d not supported\n",
-                               port->disc_addr.adrfam);
-               return -EINVAL;
-       }
-
-       if (port->inline_data_size < 0) {
-               port->inline_data_size = 
NVMET_RDMA_DEFAULT_INLINE_DATA_SIZE;
-       } else if (port->inline_data_size > 
NVMET_RDMA_MAX_INLINE_DATA_SIZE) {
-               pr_warn("inline_data_size %u is too large, reducing to 
%u\n",
-                       port->inline_data_size,
-                       NVMET_RDMA_MAX_INLINE_DATA_SIZE);
-               port->inline_data_size = NVMET_RDMA_MAX_INLINE_DATA_SIZE;
-       }
+       if (cm_id)
+               rdma_destroy_id(cm_id);
+}

-       ret = inet_pton_with_scope(&init_net, af, port->disc_addr.traddr,
-                       port->disc_addr.trsvcid, &addr);
-       if (ret) {
-               pr_err("malformed ip/port passed: %s:%s\n",
-                       port->disc_addr.traddr, port->disc_addr.trsvcid);
-               return ret;
-       }
+static int nvmet_rdma_enable_port(struct nvmet_rdma_port *port)
+{
+       struct sockaddr *addr = (struct sockaddr *)&port->addr;
+       struct rdma_cm_id *cm_id;
+       int ret;

         cm_id = rdma_create_id(&init_net, nvmet_rdma_cm_handler, port,
                         RDMA_PS_TCP, IB_QPT_RC);
@@ -1558,23 +1550,19 @@ static int nvmet_rdma_add_port(struct nvmet_port 
*port)
                 goto out_destroy_id;
         }

-       ret = rdma_bind_addr(cm_id, (struct sockaddr *)&addr);
+       ret = rdma_bind_addr(cm_id, addr);
         if (ret) {
-               pr_err("binding CM ID to %pISpcs failed (%d)\n",
-                       (struct sockaddr *)&addr, ret);
+               pr_err("binding CM ID to %pISpcs failed (%d)\n", addr, ret);
                 goto out_destroy_id;
         }

         ret = rdma_listen(cm_id, 128);
         if (ret) {
-               pr_err("listening to %pISpcs failed (%d)\n",
-                       (struct sockaddr *)&addr, ret);
+               pr_err("listening to %pISpcs failed (%d)\n", addr, ret);
                 goto out_destroy_id;
         }

-       pr_info("enabling port %d (%pISpcs)\n",
-               le16_to_cpu(port->disc_addr.portid), (struct sockaddr 
*)&addr);
-       port->priv = cm_id;
+       port->cm_id = cm_id;
         return 0;

  out_destroy_id:
@@ -1582,18 +1570,92 @@ static int nvmet_rdma_add_port(struct nvmet_port 
*port)
         return ret;
  }

-static void nvmet_rdma_remove_port(struct nvmet_port *port)
+static void nvmet_rdma_repair_port_work(struct work_struct *w)
  {
-       struct rdma_cm_id *cm_id = xchg(&port->priv, NULL);
+       struct nvmet_rdma_port *port = container_of(to_delayed_work(w),
+                       struct nvmet_rdma_port, repair_work);
+       int ret;

-       if (cm_id)
-               rdma_destroy_id(cm_id);
+       nvmet_rdma_disable_port(port);
+       ret = nvmet_rdma_enable_port(port);
+       if (ret)
+               schedule_delayed_work(&port->repair_work, 5 * HZ);
+}
+
+static int nvmet_rdma_add_port(struct nvmet_port *nport)
+{
+       struct nvmet_rdma_port *port;
+       __kernel_sa_family_t af;
+       int ret;
+
+       port = kzalloc(sizeof(*port), GFP_KERNEL);
+       if (!port)
+               return -ENOMEM;
+
+       nport->priv = port;
+       port->nport = nport;
+       INIT_DELAYED_WORK(&port->repair_work, nvmet_rdma_repair_port_work);
+
+       switch (nport->disc_addr.adrfam) {
+       case NVMF_ADDR_FAMILY_IP4:
+               af = AF_INET;
+               break;
+       case NVMF_ADDR_FAMILY_IP6:
+               af = AF_INET6;
+               break;
+       default:
+               pr_err("address family %d not supported\n",
+                               nport->disc_addr.adrfam);
+               ret = -EINVAL;
+               goto out_free_port;
+       }
+
+       if (nport->inline_data_size < 0) {
+               nport->inline_data_size = 
NVMET_RDMA_DEFAULT_INLINE_DATA_SIZE;
+       } else if (nport->inline_data_size > 
NVMET_RDMA_MAX_INLINE_DATA_SIZE) {
+               pr_warn("inline_data_size %u is too large, reducing to 
%u\n",
+                       nport->inline_data_size,
+                       NVMET_RDMA_MAX_INLINE_DATA_SIZE);
+               nport->inline_data_size = NVMET_RDMA_MAX_INLINE_DATA_SIZE;
+       }
+
+       ret = inet_pton_with_scope(&init_net, af, nport->disc_addr.traddr,
+                       nport->disc_addr.trsvcid, &port->addr);
+       if (ret) {
+               pr_err("malformed ip/port passed: %s:%s\n",
+                       nport->disc_addr.traddr, nport->disc_addr.trsvcid);
+               goto out_free_port;
+       }
+
+       ret = nvmet_rdma_enable_port(port);
+       if(ret)
+               goto out_free_port;
+
+       pr_info("enabling port %d (%pISpcs)\n",
+               le16_to_cpu(nport->disc_addr.portid),
+               (struct sockaddr *)&port->addr);
+
+       return 0;
+
+out_free_port:
+       kfree(port);
+       return ret;
+}
+
+static void nvmet_rdma_remove_port(struct nvmet_port *nport)
+{
+       struct nvmet_rdma_port *port = nport->priv;
+
+       cancel_delayed_work_sync(&port->repair_work);
+       nvmet_rdma_disable_port(port);
+       kfree(port);
  }

  static void nvmet_rdma_disc_port_addr(struct nvmet_req *req,
-               struct nvmet_port *port, char *traddr)
+               struct nvmet_port *nport, char *traddr)
  {
-       struct rdma_cm_id *cm_id = port->priv;
+       struct nvmet_rdma_port *port = nport->priv;
+       struct rdma_cm_id *cm_id = port->cm_id;

         if (inet_addr_is_any((struct sockaddr 
*)&cm_id->route.addr.src_addr)) {
                 struct nvmet_rdma_rsp *rsp =
@@ -1603,7 +1665,7 @@ static void nvmet_rdma_disc_port_addr(struct 
nvmet_req *req,

                 sprintf(traddr, "%pISc", addr);
         } else {
-               memcpy(traddr, port->disc_addr.traddr, NVMF_TRADDR_SIZE);
+               memcpy(traddr, nport->disc_addr.traddr, NVMF_TRADDR_SIZE);
         }
  }
--

_______________________________________________
linux-nvme mailing list
linux-nvme@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme

RE: NULL pointer dereference in nvmet_rdma_queue_disconnect during bond failover

[Shlomi Nimrodi]

++

-----Original Message-----
From: Sagi Grimberg <sagi@grimberg.me> 
Sent: Tuesday, March 31, 2020 03:22
To: Alex Lyakas <alex@zadara.com>
Cc: Tom Wu <tomwu@mellanox.com>; Max Gurtovoy <maxg@mellanox.com>; Israel Rukshin <israelr@mellanox.com>; linux-nvme <linux-nvme@lists.infradead.org>; Shlomi Nimrodi <shlomin@mellanox.com>
Subject: Re: NULL pointer dereference in nvmet_rdma_queue_disconnect during bond failover

Hey Alex,

>>> Alex, Max? did you retest this?
>>
>> Raising this from the ashes...
>>
>> Alex, did you test this patch?
> 
> Raising from the ashes!
> 
> In short: this patch fixes the issue!

Thanks for following up..

> 
> More details:
> 
> This patch doesn't apply on kernel 5.2. Moreover, I believe this patch 
> is incomplete, because nvmet_rdma_find_get_device() needs to be fixed 
> to treat cm_id->context as "struct nvmet_rdma_port" and not as "struct 
> nvmet_port".

Does patch [1] apply on kernel 5.2?

> However, since we are working with kernel modules from Mellanox OFED, 
> I tried applying this patch on OFED 4.7. I discovered that it already 
> has almost everything this patch introduces. Like "struct 
> nvmet_rdma_port" and the refactoring of nvmet_rdma_add_port into 
> nvmet_rdma_enable_port, and nvmet_rdma_remove_port to 
> nvmet_rdma_disable_port. I ended up with this patch [1].
> 
> Tested bond failover, and cm_id is destroyed and re-created as 
> expected [2]
> 
> Israel, Max and other Mellanox folks: can we have this fix in OFED 4.9?
> 

For MOFED issues you can follow-up with Max and Israel offline. If you can test upstream or even 5.2 stable that would be beneficial as I can add your Tested-by tag.

Thanks.

[1]:
--
Author: Sagi Grimberg <sagi@grimberg.me>
Date:   Wed Jul 3 15:33:01 2019 -0700

     nvmet-rdma: fix bonding failover possible NULL deref

     RDMA_CM_EVENT_ADDR_CHANGE event occur in the case of bonding failover
     on normal as well as on listening cm_ids. Hence this event will
     immediately trigger a NULL dereference trying to disconnect a queue
     for a cm_id that actually belongs to the port.

     To fix this we provide a different handler for the listener cm_ids
     that will defer a work to disable+(re)enable the port which essentially
     destroys and setups another listener cm_id

     Reported-by: Alex Lyakas <alex@zadara.com>
     Signed-off-by: Sagi Grimberg <sagi@grimberg.me>

diff --git a/drivers/nvme/target/rdma.c b/drivers/nvme/target/rdma.c index 9e1b8c61f54e..8dac89b7aa12 100644
--- a/drivers/nvme/target/rdma.c
+++ b/drivers/nvme/target/rdma.c
@@ -105,6 +105,13 @@ struct nvmet_rdma_queue {
         struct list_head        queue_list;
  };

+struct nvmet_rdma_port {
+       struct nvmet_port       *nport;
+       struct sockaddr_storage addr;
+       struct rdma_cm_id       *cm_id;
+       struct delayed_work     repair_work;
+};
+
  struct nvmet_rdma_device {
         struct ib_device        *device;
         struct ib_pd            *pd;
@@ -1272,6 +1279,7 @@ static int nvmet_rdma_cm_accept(struct rdma_cm_id *cm_id,
  static int nvmet_rdma_queue_connect(struct rdma_cm_id *cm_id,
                 struct rdma_cm_event *event)
  {
+       struct nvmet_rdma_port *port = cm_id->context;
         struct nvmet_rdma_device *ndev;
         struct nvmet_rdma_queue *queue;
         int ret = -EINVAL;
@@ -1287,7 +1295,7 @@ static int nvmet_rdma_queue_connect(struct rdma_cm_id *cm_id,
                 ret = -ENOMEM;
                 goto put_device;
         }
-       queue->port = cm_id->context;
+       queue->port = port->nport;

         if (queue->host_qid == 0) {
                 /* Let inflight controller teardown complete */
@@ -1412,7 +1420,7 @@ static void nvmet_rdma_queue_connect_fail(struct 
rdma_cm_id *cm_id,
  static int nvmet_rdma_device_removal(struct rdma_cm_id *cm_id,
                 struct nvmet_rdma_queue *queue)
  {
-       struct nvmet_port *port;
+       struct nvmet_rdma_port *port;

         if (queue) {
                 /*
@@ -1431,7 +1439,7 @@ static int nvmet_rdma_device_removal(struct 
rdma_cm_id *cm_id,
          * cm_id destroy. use atomic xchg to make sure
          * we don't compete with remove_port.
          */
-       if (xchg(&port->priv, NULL) != cm_id)
+       if (xchg(&port->cm_id, NULL) != cm_id)
                 return 0;

         /*
@@ -1462,6 +1470,13 @@ static int nvmet_rdma_cm_handler(struct 
rdma_cm_id *cm_id,
                 nvmet_rdma_queue_established(queue);
                 break;
         case RDMA_CM_EVENT_ADDR_CHANGE:
+               if (!queue) {
+                       struct nvmet_rdma_port *port = cm_id->context;
+
+                       schedule_delayed_work(&port->repair_work, 0);
+                       break;
+               }
+               /* FALLTHROUGH */
         case RDMA_CM_EVENT_DISCONNECTED:
         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
                 nvmet_rdma_queue_disconnect(queue);
@@ -1504,42 +1519,19 @@ static void nvmet_rdma_delete_ctrl(struct 
nvmet_ctrl *ctrl)
         mutex_unlock(&nvmet_rdma_queue_mutex);
  }

-static int nvmet_rdma_add_port(struct nvmet_port *port)
+static void nvmet_rdma_disable_port(struct nvmet_rdma_port *port)
  {
-       struct rdma_cm_id *cm_id;
-       struct sockaddr_storage addr = { };
-       __kernel_sa_family_t af;
-       int ret;
+       struct rdma_cm_id *cm_id = xchg(&port->cm_id, NULL);

-       switch (port->disc_addr.adrfam) {
-       case NVMF_ADDR_FAMILY_IP4:
-               af = AF_INET;
-               break;
-       case NVMF_ADDR_FAMILY_IP6:
-               af = AF_INET6;
-               break;
-       default:
-               pr_err("address family %d not supported\n",
-                               port->disc_addr.adrfam);
-               return -EINVAL;
-       }
-
-       if (port->inline_data_size < 0) {
-               port->inline_data_size = 
NVMET_RDMA_DEFAULT_INLINE_DATA_SIZE;
-       } else if (port->inline_data_size > 
NVMET_RDMA_MAX_INLINE_DATA_SIZE) {
-               pr_warn("inline_data_size %u is too large, reducing to 
%u\n",
-                       port->inline_data_size,
-                       NVMET_RDMA_MAX_INLINE_DATA_SIZE);
-               port->inline_data_size = NVMET_RDMA_MAX_INLINE_DATA_SIZE;
-       }
+       if (cm_id)
+               rdma_destroy_id(cm_id);
+}

-       ret = inet_pton_with_scope(&init_net, af, port->disc_addr.traddr,
-                       port->disc_addr.trsvcid, &addr);
-       if (ret) {
-               pr_err("malformed ip/port passed: %s:%s\n",
-                       port->disc_addr.traddr, port->disc_addr.trsvcid);
-               return ret;
-       }
+static int nvmet_rdma_enable_port(struct nvmet_rdma_port *port)
+{
+       struct sockaddr *addr = (struct sockaddr *)&port->addr;
+       struct rdma_cm_id *cm_id;
+       int ret;

         cm_id = rdma_create_id(&init_net, nvmet_rdma_cm_handler, port,
                         RDMA_PS_TCP, IB_QPT_RC);
@@ -1558,23 +1550,19 @@ static int nvmet_rdma_add_port(struct nvmet_port 
*port)
                 goto out_destroy_id;
         }

-       ret = rdma_bind_addr(cm_id, (struct sockaddr *)&addr);
+       ret = rdma_bind_addr(cm_id, addr);
         if (ret) {
-               pr_err("binding CM ID to %pISpcs failed (%d)\n",
-                       (struct sockaddr *)&addr, ret);
+               pr_err("binding CM ID to %pISpcs failed (%d)\n", addr, ret);
                 goto out_destroy_id;
         }

         ret = rdma_listen(cm_id, 128);
         if (ret) {
-               pr_err("listening to %pISpcs failed (%d)\n",
-                       (struct sockaddr *)&addr, ret);
+               pr_err("listening to %pISpcs failed (%d)\n", addr, ret);
                 goto out_destroy_id;
         }

-       pr_info("enabling port %d (%pISpcs)\n",
-               le16_to_cpu(port->disc_addr.portid), (struct sockaddr 
*)&addr);
-       port->priv = cm_id;
+       port->cm_id = cm_id;
         return 0;

  out_destroy_id:
@@ -1582,18 +1570,92 @@ static int nvmet_rdma_add_port(struct nvmet_port 
*port)
         return ret;
  }

-static void nvmet_rdma_remove_port(struct nvmet_port *port)
+static void nvmet_rdma_repair_port_work(struct work_struct *w)
  {
-       struct rdma_cm_id *cm_id = xchg(&port->priv, NULL);
+       struct nvmet_rdma_port *port = container_of(to_delayed_work(w),
+                       struct nvmet_rdma_port, repair_work);
+       int ret;

-       if (cm_id)
-               rdma_destroy_id(cm_id);
+       nvmet_rdma_disable_port(port);
+       ret = nvmet_rdma_enable_port(port);
+       if (ret)
+               schedule_delayed_work(&port->repair_work, 5 * HZ);
+}
+
+static int nvmet_rdma_add_port(struct nvmet_port *nport)
+{
+       struct nvmet_rdma_port *port;
+       __kernel_sa_family_t af;
+       int ret;
+
+       port = kzalloc(sizeof(*port), GFP_KERNEL);
+       if (!port)
+               return -ENOMEM;
+
+       nport->priv = port;
+       port->nport = nport;
+       INIT_DELAYED_WORK(&port->repair_work, nvmet_rdma_repair_port_work);
+
+       switch (nport->disc_addr.adrfam) {
+       case NVMF_ADDR_FAMILY_IP4:
+               af = AF_INET;
+               break;
+       case NVMF_ADDR_FAMILY_IP6:
+               af = AF_INET6;
+               break;
+       default:
+               pr_err("address family %d not supported\n",
+                               nport->disc_addr.adrfam);
+               ret = -EINVAL;
+               goto out_free_port;
+       }
+
+       if (nport->inline_data_size < 0) {
+               nport->inline_data_size = 
NVMET_RDMA_DEFAULT_INLINE_DATA_SIZE;
+       } else if (nport->inline_data_size > 
NVMET_RDMA_MAX_INLINE_DATA_SIZE) {
+               pr_warn("inline_data_size %u is too large, reducing to 
%u\n",
+                       nport->inline_data_size,
+                       NVMET_RDMA_MAX_INLINE_DATA_SIZE);
+               nport->inline_data_size = NVMET_RDMA_MAX_INLINE_DATA_SIZE;
+       }
+
+       ret = inet_pton_with_scope(&init_net, af, nport->disc_addr.traddr,
+                       nport->disc_addr.trsvcid, &port->addr);
+       if (ret) {
+               pr_err("malformed ip/port passed: %s:%s\n",
+                       nport->disc_addr.traddr, nport->disc_addr.trsvcid);
+               goto out_free_port;
+       }
+
+       ret = nvmet_rdma_enable_port(port);
+       if(ret)
+               goto out_free_port;
+
+       pr_info("enabling port %d (%pISpcs)\n",
+               le16_to_cpu(nport->disc_addr.portid),
+               (struct sockaddr *)&port->addr);
+
+       return 0;
+
+out_free_port:
+       kfree(port);
+       return ret;
+}
+
+static void nvmet_rdma_remove_port(struct nvmet_port *nport)
+{
+       struct nvmet_rdma_port *port = nport->priv;
+
+       cancel_delayed_work_sync(&port->repair_work);
+       nvmet_rdma_disable_port(port);
+       kfree(port);
  }

  static void nvmet_rdma_disc_port_addr(struct nvmet_req *req,
-               struct nvmet_port *port, char *traddr)
+               struct nvmet_port *nport, char *traddr)
  {
-       struct rdma_cm_id *cm_id = port->priv;
+       struct nvmet_rdma_port *port = nport->priv;
+       struct rdma_cm_id *cm_id = port->cm_id;

         if (inet_addr_is_any((struct sockaddr 
*)&cm_id->route.addr.src_addr)) {
                 struct nvmet_rdma_rsp *rsp =
@@ -1603,7 +1665,7 @@ static void nvmet_rdma_disc_port_addr(struct 
nvmet_req *req,

                 sprintf(traddr, "%pISc", addr);
         } else {
-               memcpy(traddr, port->disc_addr.traddr, NVMF_TRADDR_SIZE);
+               memcpy(traddr, nport->disc_addr.traddr, NVMF_TRADDR_SIZE);
         }
  }
--
_______________________________________________
linux-nvme mailing list
linux-nvme@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme

Re: NULL pointer dereference in nvmet_rdma_queue_disconnect during bond failover

[Alex Lyakas]

[-- Attachment #1: Type: text/plain, Size: 12348 bytes --]

Hi Sagi,

On Tue, Mar 31, 2020 at 3:21 AM Sagi Grimberg <sagi@grimberg.me> wrote:
>
> Hey Alex,
>
> >>> Alex, Max? did you retest this?
> >>
> >> Raising this from the ashes...
> >>
> >> Alex, did you test this patch?
> >
> > Raising from the ashes!
> >
> > In short: this patch fixes the issue!
>
> Thanks for following up..
>
> >
> > More details:
> >
> > This patch doesn't apply on kernel 5.2. Moreover, I believe this patch
> > is incomplete, because nvmet_rdma_find_get_device() needs to be fixed
> > to treat cm_id->context as "struct nvmet_rdma_port" and not as "struct
> > nvmet_port".
>
> Does patch [1] apply on kernel 5.2?
>
> > However, since we are working with kernel modules from Mellanox OFED,
> > I tried applying this patch on OFED 4.7. I discovered that it already
> > has almost everything this patch introduces. Like "struct
> > nvmet_rdma_port" and the refactoring of nvmet_rdma_add_port into
> > nvmet_rdma_enable_port, and nvmet_rdma_remove_port to
> > nvmet_rdma_disable_port. I ended up with this patch [1].
> >
> > Tested bond failover, and cm_id is destroyed and re-created as expected [2]
> >
> > Israel, Max and other Mellanox folks: can we have this fix in OFED 4.9?
> >
>
> For MOFED issues you can follow-up with Max and Israel offline. If you
> can test upstream or even 5.2 stable that would be beneficial as I can
> add your Tested-by tag.
>
This patch did not apply to latest 5.2 (5.2.21). All of 10 hunks
failed. I applied it manually, and also handled cm_id->context in
nvmet_rdma_find_get_device as I mentioned earlier. I am attaching the
patch that I tested on kernel 5.2.21 (target side). I confirm that
this patch fixes the bond failover issue.

Tested-by: Alex Lyakas <alex@zadara.com>

Max, will this help to deliver this fix upstream, so that we can get
it in MOFED 4.9?

Thanks,
Alex.



> Thanks.
>
> [1]:
> --
> Author: Sagi Grimberg <sagi@grimberg.me>
> Date:   Wed Jul 3 15:33:01 2019 -0700
>
>      nvmet-rdma: fix bonding failover possible NULL deref
>
>      RDMA_CM_EVENT_ADDR_CHANGE event occur in the case of bonding failover
>      on normal as well as on listening cm_ids. Hence this event will
>      immediately trigger a NULL dereference trying to disconnect a queue
>      for a cm_id that actually belongs to the port.
>
>      To fix this we provide a different handler for the listener cm_ids
>      that will defer a work to disable+(re)enable the port which essentially
>      destroys and setups another listener cm_id
>
>      Reported-by: Alex Lyakas <alex@zadara.com>
>      Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
>
> diff --git a/drivers/nvme/target/rdma.c b/drivers/nvme/target/rdma.c
> index 9e1b8c61f54e..8dac89b7aa12 100644
> --- a/drivers/nvme/target/rdma.c
> +++ b/drivers/nvme/target/rdma.c
> @@ -105,6 +105,13 @@ struct nvmet_rdma_queue {
>          struct list_head        queue_list;
>   };
>
> +struct nvmet_rdma_port {
> +       struct nvmet_port       *nport;
> +       struct sockaddr_storage addr;
> +       struct rdma_cm_id       *cm_id;
> +       struct delayed_work     repair_work;
> +};
> +
>   struct nvmet_rdma_device {
>          struct ib_device        *device;
>          struct ib_pd            *pd;
> @@ -1272,6 +1279,7 @@ static int nvmet_rdma_cm_accept(struct rdma_cm_id
> *cm_id,
>   static int nvmet_rdma_queue_connect(struct rdma_cm_id *cm_id,
>                  struct rdma_cm_event *event)
>   {
> +       struct nvmet_rdma_port *port = cm_id->context;
>          struct nvmet_rdma_device *ndev;
>          struct nvmet_rdma_queue *queue;
>          int ret = -EINVAL;
> @@ -1287,7 +1295,7 @@ static int nvmet_rdma_queue_connect(struct
> rdma_cm_id *cm_id,
>                  ret = -ENOMEM;
>                  goto put_device;
>          }
> -       queue->port = cm_id->context;
> +       queue->port = port->nport;
>
>          if (queue->host_qid == 0) {
>                  /* Let inflight controller teardown complete */
> @@ -1412,7 +1420,7 @@ static void nvmet_rdma_queue_connect_fail(struct
> rdma_cm_id *cm_id,
>   static int nvmet_rdma_device_removal(struct rdma_cm_id *cm_id,
>                  struct nvmet_rdma_queue *queue)
>   {
> -       struct nvmet_port *port;
> +       struct nvmet_rdma_port *port;
>
>          if (queue) {
>                  /*
> @@ -1431,7 +1439,7 @@ static int nvmet_rdma_device_removal(struct
> rdma_cm_id *cm_id,
>           * cm_id destroy. use atomic xchg to make sure
>           * we don't compete with remove_port.
>           */
> -       if (xchg(&port->priv, NULL) != cm_id)
> +       if (xchg(&port->cm_id, NULL) != cm_id)
>                  return 0;
>
>          /*
> @@ -1462,6 +1470,13 @@ static int nvmet_rdma_cm_handler(struct
> rdma_cm_id *cm_id,
>                  nvmet_rdma_queue_established(queue);
>                  break;
>          case RDMA_CM_EVENT_ADDR_CHANGE:
> +               if (!queue) {
> +                       struct nvmet_rdma_port *port = cm_id->context;
> +
> +                       schedule_delayed_work(&port->repair_work, 0);
> +                       break;
> +               }
> +               /* FALLTHROUGH */
>          case RDMA_CM_EVENT_DISCONNECTED:
>          case RDMA_CM_EVENT_TIMEWAIT_EXIT:
>                  nvmet_rdma_queue_disconnect(queue);
> @@ -1504,42 +1519,19 @@ static void nvmet_rdma_delete_ctrl(struct
> nvmet_ctrl *ctrl)
>          mutex_unlock(&nvmet_rdma_queue_mutex);
>   }
>
> -static int nvmet_rdma_add_port(struct nvmet_port *port)
> +static void nvmet_rdma_disable_port(struct nvmet_rdma_port *port)
>   {
> -       struct rdma_cm_id *cm_id;
> -       struct sockaddr_storage addr = { };
> -       __kernel_sa_family_t af;
> -       int ret;
> +       struct rdma_cm_id *cm_id = xchg(&port->cm_id, NULL);
>
> -       switch (port->disc_addr.adrfam) {
> -       case NVMF_ADDR_FAMILY_IP4:
> -               af = AF_INET;
> -               break;
> -       case NVMF_ADDR_FAMILY_IP6:
> -               af = AF_INET6;
> -               break;
> -       default:
> -               pr_err("address family %d not supported\n",
> -                               port->disc_addr.adrfam);
> -               return -EINVAL;
> -       }
> -
> -       if (port->inline_data_size < 0) {
> -               port->inline_data_size =
> NVMET_RDMA_DEFAULT_INLINE_DATA_SIZE;
> -       } else if (port->inline_data_size >
> NVMET_RDMA_MAX_INLINE_DATA_SIZE) {
> -               pr_warn("inline_data_size %u is too large, reducing to
> %u\n",
> -                       port->inline_data_size,
> -                       NVMET_RDMA_MAX_INLINE_DATA_SIZE);
> -               port->inline_data_size = NVMET_RDMA_MAX_INLINE_DATA_SIZE;
> -       }
> +       if (cm_id)
> +               rdma_destroy_id(cm_id);
> +}
>
> -       ret = inet_pton_with_scope(&init_net, af, port->disc_addr.traddr,
> -                       port->disc_addr.trsvcid, &addr);
> -       if (ret) {
> -               pr_err("malformed ip/port passed: %s:%s\n",
> -                       port->disc_addr.traddr, port->disc_addr.trsvcid);
> -               return ret;
> -       }
> +static int nvmet_rdma_enable_port(struct nvmet_rdma_port *port)
> +{
> +       struct sockaddr *addr = (struct sockaddr *)&port->addr;
> +       struct rdma_cm_id *cm_id;
> +       int ret;
>
>          cm_id = rdma_create_id(&init_net, nvmet_rdma_cm_handler, port,
>                          RDMA_PS_TCP, IB_QPT_RC);
> @@ -1558,23 +1550,19 @@ static int nvmet_rdma_add_port(struct nvmet_port
> *port)
>                  goto out_destroy_id;
>          }
>
> -       ret = rdma_bind_addr(cm_id, (struct sockaddr *)&addr);
> +       ret = rdma_bind_addr(cm_id, addr);
>          if (ret) {
> -               pr_err("binding CM ID to %pISpcs failed (%d)\n",
> -                       (struct sockaddr *)&addr, ret);
> +               pr_err("binding CM ID to %pISpcs failed (%d)\n", addr, ret);
>                  goto out_destroy_id;
>          }
>
>          ret = rdma_listen(cm_id, 128);
>          if (ret) {
> -               pr_err("listening to %pISpcs failed (%d)\n",
> -                       (struct sockaddr *)&addr, ret);
> +               pr_err("listening to %pISpcs failed (%d)\n", addr, ret);
>                  goto out_destroy_id;
>          }
>
> -       pr_info("enabling port %d (%pISpcs)\n",
> -               le16_to_cpu(port->disc_addr.portid), (struct sockaddr
> *)&addr);
> -       port->priv = cm_id;
> +       port->cm_id = cm_id;
>          return 0;
>
>   out_destroy_id:
> @@ -1582,18 +1570,92 @@ static int nvmet_rdma_add_port(struct nvmet_port
> *port)
>          return ret;
>   }
>
> -static void nvmet_rdma_remove_port(struct nvmet_port *port)
> +static void nvmet_rdma_repair_port_work(struct work_struct *w)
>   {
> -       struct rdma_cm_id *cm_id = xchg(&port->priv, NULL);
> +       struct nvmet_rdma_port *port = container_of(to_delayed_work(w),
> +                       struct nvmet_rdma_port, repair_work);
> +       int ret;
>
> -       if (cm_id)
> -               rdma_destroy_id(cm_id);
> +       nvmet_rdma_disable_port(port);
> +       ret = nvmet_rdma_enable_port(port);
> +       if (ret)
> +               schedule_delayed_work(&port->repair_work, 5 * HZ);
> +}
> +
> +static int nvmet_rdma_add_port(struct nvmet_port *nport)
> +{
> +       struct nvmet_rdma_port *port;
> +       __kernel_sa_family_t af;
> +       int ret;
> +
> +       port = kzalloc(sizeof(*port), GFP_KERNEL);
> +       if (!port)
> +               return -ENOMEM;
> +
> +       nport->priv = port;
> +       port->nport = nport;
> +       INIT_DELAYED_WORK(&port->repair_work, nvmet_rdma_repair_port_work);
> +
> +       switch (nport->disc_addr.adrfam) {
> +       case NVMF_ADDR_FAMILY_IP4:
> +               af = AF_INET;
> +               break;
> +       case NVMF_ADDR_FAMILY_IP6:
> +               af = AF_INET6;
> +               break;
> +       default:
> +               pr_err("address family %d not supported\n",
> +                               nport->disc_addr.adrfam);
> +               ret = -EINVAL;
> +               goto out_free_port;
> +       }
> +
> +       if (nport->inline_data_size < 0) {
> +               nport->inline_data_size =
> NVMET_RDMA_DEFAULT_INLINE_DATA_SIZE;
> +       } else if (nport->inline_data_size >
> NVMET_RDMA_MAX_INLINE_DATA_SIZE) {
> +               pr_warn("inline_data_size %u is too large, reducing to
> %u\n",
> +                       nport->inline_data_size,
> +                       NVMET_RDMA_MAX_INLINE_DATA_SIZE);
> +               nport->inline_data_size = NVMET_RDMA_MAX_INLINE_DATA_SIZE;
> +       }
> +
> +       ret = inet_pton_with_scope(&init_net, af, nport->disc_addr.traddr,
> +                       nport->disc_addr.trsvcid, &port->addr);
> +       if (ret) {
> +               pr_err("malformed ip/port passed: %s:%s\n",
> +                       nport->disc_addr.traddr, nport->disc_addr.trsvcid);
> +               goto out_free_port;
> +       }
> +
> +       ret = nvmet_rdma_enable_port(port);
> +       if(ret)
> +               goto out_free_port;
> +
> +       pr_info("enabling port %d (%pISpcs)\n",
> +               le16_to_cpu(nport->disc_addr.portid),
> +               (struct sockaddr *)&port->addr);
> +
> +       return 0;
> +
> +out_free_port:
> +       kfree(port);
> +       return ret;
> +}
> +
> +static void nvmet_rdma_remove_port(struct nvmet_port *nport)
> +{
> +       struct nvmet_rdma_port *port = nport->priv;
> +
> +       cancel_delayed_work_sync(&port->repair_work);
> +       nvmet_rdma_disable_port(port);
> +       kfree(port);
>   }
>
>   static void nvmet_rdma_disc_port_addr(struct nvmet_req *req,
> -               struct nvmet_port *port, char *traddr)
> +               struct nvmet_port *nport, char *traddr)
>   {
> -       struct rdma_cm_id *cm_id = port->priv;
> +       struct nvmet_rdma_port *port = nport->priv;
> +       struct rdma_cm_id *cm_id = port->cm_id;
>
>          if (inet_addr_is_any((struct sockaddr
> *)&cm_id->route.addr.src_addr)) {
>                  struct nvmet_rdma_rsp *rsp =
> @@ -1603,7 +1665,7 @@ static void nvmet_rdma_disc_port_addr(struct
> nvmet_req *req,
>
>                  sprintf(traddr, "%pISc", addr);
>          } else {
> -               memcpy(traddr, port->disc_addr.traddr, NVMF_TRADDR_SIZE);
> +               memcpy(traddr, nport->disc_addr.traddr, NVMF_TRADDR_SIZE);
>          }
>   }
> --

[-- Attachment #2: 0001-nvmet-rdma-fix-bonding-failover-possible-NULL-deref.5.2.patch --]
[-- Type: application/octet-stream, Size: 8809 bytes --]

diff --git a/drivers/nvme/target/rdma.c b/drivers/nvme/target/rdma.c
index 36d906a..ebe7c43 100644
--- a/drivers/nvme/target/rdma.c
+++ b/drivers/nvme/target/rdma.c
@@ -102,6 +102,13 @@ struct nvmet_rdma_queue {
 	struct list_head	queue_list;
 };
 
+struct nvmet_rdma_port {
+	struct nvmet_port       *nport;
+	struct sockaddr_storage addr;
+	struct rdma_cm_id       *cm_id;
+	struct delayed_work     repair_work;
+};
+
 struct nvmet_rdma_device {
 	struct ib_device	*device;
 	struct ib_pd		*pd;
@@ -914,7 +921,8 @@ static void nvmet_rdma_free_dev(struct kref *ref)
 static struct nvmet_rdma_device *
 nvmet_rdma_find_get_device(struct rdma_cm_id *cm_id)
 {
-	struct nvmet_port *port = cm_id->context;
+	struct nvmet_rdma_port *port = cm_id->context;
+	struct nvmet_port *nport = port->nport;
 	struct nvmet_rdma_device *ndev;
 	int inline_page_count;
 	int inline_sge_count;
@@ -931,17 +939,17 @@ static void nvmet_rdma_free_dev(struct kref *ref)
 	if (!ndev)
 		goto out_err;
 
-	inline_page_count = num_pages(port->inline_data_size);
+	inline_page_count = num_pages(nport->inline_data_size);
 	inline_sge_count = max(cm_id->device->attrs.max_sge_rd,
 				cm_id->device->attrs.max_recv_sge) - 1;
 	if (inline_page_count > inline_sge_count) {
 		pr_warn("inline_data_size %d cannot be supported by device %s. Reducing to %lu.\n",
-			port->inline_data_size, cm_id->device->name,
+			nport->inline_data_size, cm_id->device->name,
 			inline_sge_count * PAGE_SIZE);
-		port->inline_data_size = inline_sge_count * PAGE_SIZE;
+		nport->inline_data_size = inline_sge_count * PAGE_SIZE;
 		inline_page_count = inline_sge_count;
 	}
-	ndev->inline_data_size = port->inline_data_size;
+	ndev->inline_data_size = nport->inline_data_size;
 	ndev->inline_page_count = inline_page_count;
 	ndev->device = cm_id->device;
 	kref_init(&ndev->ref);
@@ -1267,6 +1275,7 @@ static int nvmet_rdma_cm_accept(struct rdma_cm_id *cm_id,
 static int nvmet_rdma_queue_connect(struct rdma_cm_id *cm_id,
 		struct rdma_cm_event *event)
 {
+	struct nvmet_rdma_port *port = cm_id->context;
 	struct nvmet_rdma_device *ndev;
 	struct nvmet_rdma_queue *queue;
 	int ret = -EINVAL;
@@ -1282,7 +1291,7 @@ static int nvmet_rdma_queue_connect(struct rdma_cm_id *cm_id,
 		ret = -ENOMEM;
 		goto put_device;
 	}
-	queue->port = cm_id->context;
+	queue->port = port->nport;
 
 	if (queue->host_qid == 0) {
 		/* Let inflight controller teardown complete */
@@ -1407,7 +1416,7 @@ static void nvmet_rdma_queue_connect_fail(struct rdma_cm_id *cm_id,
 static int nvmet_rdma_device_removal(struct rdma_cm_id *cm_id,
 		struct nvmet_rdma_queue *queue)
 {
-	struct nvmet_port *port;
+	struct nvmet_rdma_port *port;
 
 	if (queue) {
 		/*
@@ -1426,7 +1435,7 @@ static int nvmet_rdma_device_removal(struct rdma_cm_id *cm_id,
 	 * cm_id destroy. use atomic xchg to make sure
 	 * we don't compete with remove_port.
 	 */
-	if (xchg(&port->priv, NULL) != cm_id)
+	if (xchg(&port->cm_id, NULL) != cm_id)
 		return 0;
 
 	/*
@@ -1457,6 +1466,13 @@ static int nvmet_rdma_cm_handler(struct rdma_cm_id *cm_id,
 		nvmet_rdma_queue_established(queue);
 		break;
 	case RDMA_CM_EVENT_ADDR_CHANGE:
+		if (!queue) {
+			struct nvmet_rdma_port *port = cm_id->context;
+
+			schedule_delayed_work(&port->repair_work, 0);
+			break;
+		}
+		/* FALLTHROUGH */
 	case RDMA_CM_EVENT_DISCONNECTED:
 	case RDMA_CM_EVENT_TIMEWAIT_EXIT:
 		nvmet_rdma_queue_disconnect(queue);
@@ -1499,42 +1515,19 @@ static void nvmet_rdma_delete_ctrl(struct nvmet_ctrl *ctrl)
 	mutex_unlock(&nvmet_rdma_queue_mutex);
 }
 
-static int nvmet_rdma_add_port(struct nvmet_port *port)
+static void nvmet_rdma_disable_port(struct nvmet_rdma_port *port)
 {
-	struct rdma_cm_id *cm_id;
-	struct sockaddr_storage addr = { };
-	__kernel_sa_family_t af;
-	int ret;
-
-	switch (port->disc_addr.adrfam) {
-	case NVMF_ADDR_FAMILY_IP4:
-		af = AF_INET;
-		break;
-	case NVMF_ADDR_FAMILY_IP6:
-		af = AF_INET6;
-		break;
-	default:
-		pr_err("address family %d not supported\n",
-				port->disc_addr.adrfam);
-		return -EINVAL;
-	}
+	struct rdma_cm_id *cm_id = xchg(&port->cm_id, NULL);
 
-	if (port->inline_data_size < 0) {
-		port->inline_data_size = NVMET_RDMA_DEFAULT_INLINE_DATA_SIZE;
-	} else if (port->inline_data_size > NVMET_RDMA_MAX_INLINE_DATA_SIZE) {
-		pr_warn("inline_data_size %u is too large, reducing to %u\n",
-			port->inline_data_size,
-			NVMET_RDMA_MAX_INLINE_DATA_SIZE);
-		port->inline_data_size = NVMET_RDMA_MAX_INLINE_DATA_SIZE;
-	}
+	if (cm_id)
+		rdma_destroy_id(cm_id);
+}
 
-	ret = inet_pton_with_scope(&init_net, af, port->disc_addr.traddr,
-			port->disc_addr.trsvcid, &addr);
-	if (ret) {
-		pr_err("malformed ip/port passed: %s:%s\n",
-			port->disc_addr.traddr, port->disc_addr.trsvcid);
-		return ret;
-	}
+static int nvmet_rdma_enable_port(struct nvmet_rdma_port *port)
+{
+	struct sockaddr *addr = (struct sockaddr *)&port->addr;
+	struct rdma_cm_id *cm_id;
+	int ret;
 
 	cm_id = rdma_create_id(&init_net, nvmet_rdma_cm_handler, port,
 			RDMA_PS_TCP, IB_QPT_RC);
@@ -1553,23 +1546,19 @@ static int nvmet_rdma_add_port(struct nvmet_port *port)
 		goto out_destroy_id;
 	}
 
-	ret = rdma_bind_addr(cm_id, (struct sockaddr *)&addr);
+	ret = rdma_bind_addr(cm_id, addr);
 	if (ret) {
-		pr_err("binding CM ID to %pISpcs failed (%d)\n",
-			(struct sockaddr *)&addr, ret);
+		pr_err("binding CM ID to %pISpcs failed (%d)\n", addr, ret);
 		goto out_destroy_id;
 	}
 
 	ret = rdma_listen(cm_id, 128);
 	if (ret) {
-		pr_err("listening to %pISpcs failed (%d)\n",
-			(struct sockaddr *)&addr, ret);
+		pr_err("listening to %pISpcs failed (%d)\n", addr, ret);
 		goto out_destroy_id;
 	}
 
-	pr_info("enabling port %d (%pISpcs)\n",
-		le16_to_cpu(port->disc_addr.portid), (struct sockaddr *)&addr);
-	port->priv = cm_id;
+	port->cm_id = cm_id;
 	return 0;
 
 out_destroy_id:
@@ -1577,18 +1566,92 @@ static int nvmet_rdma_add_port(struct nvmet_port *port)
 	return ret;
 }
 
-static void nvmet_rdma_remove_port(struct nvmet_port *port)
+static void nvmet_rdma_repair_port_work(struct work_struct *w)
 {
-	struct rdma_cm_id *cm_id = xchg(&port->priv, NULL);
+	struct nvmet_rdma_port *port = container_of(to_delayed_work(w),
+					struct nvmet_rdma_port, repair_work);
+	int ret;
 
-	if (cm_id)
-		rdma_destroy_id(cm_id);
+	nvmet_rdma_disable_port(port);
+	ret = nvmet_rdma_enable_port(port);
+	if (ret)
+		schedule_delayed_work(&port->repair_work, 5 * HZ);
+}
+
+static int nvmet_rdma_add_port(struct nvmet_port *nport)
+{
+	struct nvmet_rdma_port *port;
+	__kernel_sa_family_t af;
+	int ret;
+
+	port = kzalloc(sizeof(*port), GFP_KERNEL);
+	if (!port)
+		return -ENOMEM;
+
+	nport->priv = port;
+	port->nport = nport;
+	INIT_DELAYED_WORK(&port->repair_work, nvmet_rdma_repair_port_work);
+
+	switch (nport->disc_addr.adrfam) {
+		case NVMF_ADDR_FAMILY_IP4:
+			af = AF_INET;
+			break;
+		case NVMF_ADDR_FAMILY_IP6:
+			af = AF_INET6;
+			break;
+		default:
+			pr_err("address family %d not supported\n",
+							nport->disc_addr.adrfam);
+			ret = -EINVAL;
+			goto out_free_port;
+		}
+
+		if (nport->inline_data_size < 0) {
+			nport->inline_data_size = NVMET_RDMA_DEFAULT_INLINE_DATA_SIZE;
+		} else if (nport->inline_data_size > NVMET_RDMA_MAX_INLINE_DATA_SIZE) {
+			pr_warn("inline_data_size %u is too large, reducing to %u\n",
+					nport->inline_data_size,
+					NVMET_RDMA_MAX_INLINE_DATA_SIZE);
+			nport->inline_data_size = NVMET_RDMA_MAX_INLINE_DATA_SIZE;
+		}
+
+		ret = inet_pton_with_scope(&init_net, af, nport->disc_addr.traddr,
+						nport->disc_addr.trsvcid, &port->addr);
+		if (ret) {
+			pr_err("malformed ip/port passed: %s:%s\n",
+					nport->disc_addr.traddr, nport->disc_addr.trsvcid);
+			goto out_free_port;
+		}
+
+		ret = nvmet_rdma_enable_port(port);
+		if(ret)
+			goto out_free_port;
+
+		pr_info("enabling port %d (%pISpcs)\n",
+				le16_to_cpu(nport->disc_addr.portid),
+				(struct sockaddr *)&port->addr);
+
+		return 0;
+
+out_free_port:
+		kfree(port);
+		return ret;
+}
+
+static void nvmet_rdma_remove_port(struct nvmet_port *nport)
+{
+	struct nvmet_rdma_port *port = nport->priv;
+
+	cancel_delayed_work_sync(&port->repair_work);
+	nvmet_rdma_disable_port(port);
+	kfree(port);
 }
 
 static void nvmet_rdma_disc_port_addr(struct nvmet_req *req,
-		struct nvmet_port *port, char *traddr)
+		struct nvmet_port *nport, char *traddr)
 {
-	struct rdma_cm_id *cm_id = port->priv;
+	struct nvmet_rdma_port *port = nport->priv;
+	struct rdma_cm_id *cm_id = port->cm_id;
 
 	if (inet_addr_is_any((struct sockaddr *)&cm_id->route.addr.src_addr)) {
 		struct nvmet_rdma_rsp *rsp =
@@ -1598,7 +1661,7 @@ static void nvmet_rdma_disc_port_addr(struct nvmet_req *req,
 
 		sprintf(traddr, "%pISc", addr);
 	} else {
-		memcpy(traddr, port->disc_addr.traddr, NVMF_TRADDR_SIZE);
+		memcpy(traddr, nport->disc_addr.traddr, NVMF_TRADDR_SIZE);
 	}
 }
 

[-- Attachment #3: Type: text/plain, Size: 158 bytes --]

_______________________________________________
linux-nvme mailing list
linux-nvme@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme

Re: NULL pointer dereference in nvmet_rdma_queue_disconnect during bond failover

[Max Gurtovoy]

On 4/2/2020 12:13 PM, Alex Lyakas wrote:
> Hi Sagi,
>
> On Tue, Mar 31, 2020 at 3:21 AM Sagi Grimberg <sagi@grimberg.me> wrote:
>> Hey Alex,
>>
>>>>> Alex, Max? did you retest this?
>>>> Raising this from the ashes...
>>>>
>>>> Alex, did you test this patch?
>>> Raising from the ashes!
>>>
>>> In short: this patch fixes the issue!
>> Thanks for following up..
>>
>>> More details:
>>>
>>> This patch doesn't apply on kernel 5.2. Moreover, I believe this patch
>>> is incomplete, because nvmet_rdma_find_get_device() needs to be fixed
>>> to treat cm_id->context as "struct nvmet_rdma_port" and not as "struct
>>> nvmet_port".
>> Does patch [1] apply on kernel 5.2?
>>
>>> However, since we are working with kernel modules from Mellanox OFED,
>>> I tried applying this patch on OFED 4.7. I discovered that it already
>>> has almost everything this patch introduces. Like "struct
>>> nvmet_rdma_port" and the refactoring of nvmet_rdma_add_port into
>>> nvmet_rdma_enable_port, and nvmet_rdma_remove_port to
>>> nvmet_rdma_disable_port. I ended up with this patch [1].
>>>
>>> Tested bond failover, and cm_id is destroyed and re-created as expected [2]
>>>
>>> Israel, Max and other Mellanox folks: can we have this fix in OFED 4.9?
>>>
>> For MOFED issues you can follow-up with Max and Israel offline. If you
>> can test upstream or even 5.2 stable that would be beneficial as I can
>> add your Tested-by tag.
>>
> This patch did not apply to latest 5.2 (5.2.21). All of 10 hunks
> failed. I applied it manually, and also handled cm_id->context in
> nvmet_rdma_find_get_device as I mentioned earlier. I am attaching the
> patch that I tested on kernel 5.2.21 (target side). I confirm that
> this patch fixes the bond failover issue.
>
> Tested-by: Alex Lyakas <alex@zadara.com>
>
> Max, will this help to deliver this fix upstream, so that we can get
> it in MOFED 4.9?
>
> Thanks,
> Alex.

Alex,

Thanks for testing this.

Waiting for Sagi's official rebased version for doing full review.

We can take it on us to send the rebased, reviewed and validated version 
(Sagi - let me know what you prefer).

For mofed - this is not the forum to discuss Mellanox SW release (let's 
take it offline).



>
>
>> Thanks.
>>
>> [1]:
>> --
>> Author: Sagi Grimberg <sagi@grimberg.me>
>> Date:   Wed Jul 3 15:33:01 2019 -0700
>>
>>       nvmet-rdma: fix bonding failover possible NULL deref
>>
>>       RDMA_CM_EVENT_ADDR_CHANGE event occur in the case of bonding failover
>>       on normal as well as on listening cm_ids. Hence this event will
>>       immediately trigger a NULL dereference trying to disconnect a queue
>>       for a cm_id that actually belongs to the port.
>>
>>       To fix this we provide a different handler for the listener cm_ids
>>       that will defer a work to disable+(re)enable the port which essentially
>>       destroys and setups another listener cm_id
>>
>>       Reported-by: Alex Lyakas <alex@zadara.com>
>>       Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
>>
>> diff --git a/drivers/nvme/target/rdma.c b/drivers/nvme/target/rdma.c
>> index 9e1b8c61f54e..8dac89b7aa12 100644
>> --- a/drivers/nvme/target/rdma.c
>> +++ b/drivers/nvme/target/rdma.c
>> @@ -105,6 +105,13 @@ struct nvmet_rdma_queue {
>>           struct list_head        queue_list;
>>    };
>>
>> +struct nvmet_rdma_port {
>> +       struct nvmet_port       *nport;
>> +       struct sockaddr_storage addr;
>> +       struct rdma_cm_id       *cm_id;
>> +       struct delayed_work     repair_work;
>> +};
>> +
>>    struct nvmet_rdma_device {
>>           struct ib_device        *device;
>>           struct ib_pd            *pd;
>> @@ -1272,6 +1279,7 @@ static int nvmet_rdma_cm_accept(struct rdma_cm_id
>> *cm_id,
>>    static int nvmet_rdma_queue_connect(struct rdma_cm_id *cm_id,
>>                   struct rdma_cm_event *event)
>>    {
>> +       struct nvmet_rdma_port *port = cm_id->context;
>>           struct nvmet_rdma_device *ndev;
>>           struct nvmet_rdma_queue *queue;
>>           int ret = -EINVAL;
>> @@ -1287,7 +1295,7 @@ static int nvmet_rdma_queue_connect(struct
>> rdma_cm_id *cm_id,
>>                   ret = -ENOMEM;
>>                   goto put_device;
>>           }
>> -       queue->port = cm_id->context;
>> +       queue->port = port->nport;
>>
>>           if (queue->host_qid == 0) {
>>                   /* Let inflight controller teardown complete */
>> @@ -1412,7 +1420,7 @@ static void nvmet_rdma_queue_connect_fail(struct
>> rdma_cm_id *cm_id,
>>    static int nvmet_rdma_device_removal(struct rdma_cm_id *cm_id,
>>                   struct nvmet_rdma_queue *queue)
>>    {
>> -       struct nvmet_port *port;
>> +       struct nvmet_rdma_port *port;
>>
>>           if (queue) {
>>                   /*
>> @@ -1431,7 +1439,7 @@ static int nvmet_rdma_device_removal(struct
>> rdma_cm_id *cm_id,
>>            * cm_id destroy. use atomic xchg to make sure
>>            * we don't compete with remove_port.
>>            */
>> -       if (xchg(&port->priv, NULL) != cm_id)
>> +       if (xchg(&port->cm_id, NULL) != cm_id)
>>                   return 0;
>>
>>           /*
>> @@ -1462,6 +1470,13 @@ static int nvmet_rdma_cm_handler(struct
>> rdma_cm_id *cm_id,
>>                   nvmet_rdma_queue_established(queue);
>>                   break;
>>           case RDMA_CM_EVENT_ADDR_CHANGE:
>> +               if (!queue) {
>> +                       struct nvmet_rdma_port *port = cm_id->context;
>> +
>> +                       schedule_delayed_work(&port->repair_work, 0);
>> +                       break;
>> +               }
>> +               /* FALLTHROUGH */
>>           case RDMA_CM_EVENT_DISCONNECTED:
>>           case RDMA_CM_EVENT_TIMEWAIT_EXIT:
>>                   nvmet_rdma_queue_disconnect(queue);
>> @@ -1504,42 +1519,19 @@ static void nvmet_rdma_delete_ctrl(struct
>> nvmet_ctrl *ctrl)
>>           mutex_unlock(&nvmet_rdma_queue_mutex);
>>    }
>>
>> -static int nvmet_rdma_add_port(struct nvmet_port *port)
>> +static void nvmet_rdma_disable_port(struct nvmet_rdma_port *port)
>>    {
>> -       struct rdma_cm_id *cm_id;
>> -       struct sockaddr_storage addr = { };
>> -       __kernel_sa_family_t af;
>> -       int ret;
>> +       struct rdma_cm_id *cm_id = xchg(&port->cm_id, NULL);
>>
>> -       switch (port->disc_addr.adrfam) {
>> -       case NVMF_ADDR_FAMILY_IP4:
>> -               af = AF_INET;
>> -               break;
>> -       case NVMF_ADDR_FAMILY_IP6:
>> -               af = AF_INET6;
>> -               break;
>> -       default:
>> -               pr_err("address family %d not supported\n",
>> -                               port->disc_addr.adrfam);
>> -               return -EINVAL;
>> -       }
>> -
>> -       if (port->inline_data_size < 0) {
>> -               port->inline_data_size =
>> NVMET_RDMA_DEFAULT_INLINE_DATA_SIZE;
>> -       } else if (port->inline_data_size >
>> NVMET_RDMA_MAX_INLINE_DATA_SIZE) {
>> -               pr_warn("inline_data_size %u is too large, reducing to
>> %u\n",
>> -                       port->inline_data_size,
>> -                       NVMET_RDMA_MAX_INLINE_DATA_SIZE);
>> -               port->inline_data_size = NVMET_RDMA_MAX_INLINE_DATA_SIZE;
>> -       }
>> +       if (cm_id)
>> +               rdma_destroy_id(cm_id);
>> +}
>>
>> -       ret = inet_pton_with_scope(&init_net, af, port->disc_addr.traddr,
>> -                       port->disc_addr.trsvcid, &addr);
>> -       if (ret) {
>> -               pr_err("malformed ip/port passed: %s:%s\n",
>> -                       port->disc_addr.traddr, port->disc_addr.trsvcid);
>> -               return ret;
>> -       }
>> +static int nvmet_rdma_enable_port(struct nvmet_rdma_port *port)
>> +{
>> +       struct sockaddr *addr = (struct sockaddr *)&port->addr;
>> +       struct rdma_cm_id *cm_id;
>> +       int ret;
>>
>>           cm_id = rdma_create_id(&init_net, nvmet_rdma_cm_handler, port,
>>                           RDMA_PS_TCP, IB_QPT_RC);
>> @@ -1558,23 +1550,19 @@ static int nvmet_rdma_add_port(struct nvmet_port
>> *port)
>>                   goto out_destroy_id;
>>           }
>>
>> -       ret = rdma_bind_addr(cm_id, (struct sockaddr *)&addr);
>> +       ret = rdma_bind_addr(cm_id, addr);
>>           if (ret) {
>> -               pr_err("binding CM ID to %pISpcs failed (%d)\n",
>> -                       (struct sockaddr *)&addr, ret);
>> +               pr_err("binding CM ID to %pISpcs failed (%d)\n", addr, ret);
>>                   goto out_destroy_id;
>>           }
>>
>>           ret = rdma_listen(cm_id, 128);
>>           if (ret) {
>> -               pr_err("listening to %pISpcs failed (%d)\n",
>> -                       (struct sockaddr *)&addr, ret);
>> +               pr_err("listening to %pISpcs failed (%d)\n", addr, ret);
>>                   goto out_destroy_id;
>>           }
>>
>> -       pr_info("enabling port %d (%pISpcs)\n",
>> -               le16_to_cpu(port->disc_addr.portid), (struct sockaddr
>> *)&addr);
>> -       port->priv = cm_id;
>> +       port->cm_id = cm_id;
>>           return 0;
>>
>>    out_destroy_id:
>> @@ -1582,18 +1570,92 @@ static int nvmet_rdma_add_port(struct nvmet_port
>> *port)
>>           return ret;
>>    }
>>
>> -static void nvmet_rdma_remove_port(struct nvmet_port *port)
>> +static void nvmet_rdma_repair_port_work(struct work_struct *w)
>>    {
>> -       struct rdma_cm_id *cm_id = xchg(&port->priv, NULL);
>> +       struct nvmet_rdma_port *port = container_of(to_delayed_work(w),
>> +                       struct nvmet_rdma_port, repair_work);
>> +       int ret;
>>
>> -       if (cm_id)
>> -               rdma_destroy_id(cm_id);
>> +       nvmet_rdma_disable_port(port);
>> +       ret = nvmet_rdma_enable_port(port);
>> +       if (ret)
>> +               schedule_delayed_work(&port->repair_work, 5 * HZ);
>> +}
>> +
>> +static int nvmet_rdma_add_port(struct nvmet_port *nport)
>> +{
>> +       struct nvmet_rdma_port *port;
>> +       __kernel_sa_family_t af;
>> +       int ret;
>> +
>> +       port = kzalloc(sizeof(*port), GFP_KERNEL);
>> +       if (!port)
>> +               return -ENOMEM;
>> +
>> +       nport->priv = port;
>> +       port->nport = nport;
>> +       INIT_DELAYED_WORK(&port->repair_work, nvmet_rdma_repair_port_work);
>> +
>> +       switch (nport->disc_addr.adrfam) {
>> +       case NVMF_ADDR_FAMILY_IP4:
>> +               af = AF_INET;
>> +               break;
>> +       case NVMF_ADDR_FAMILY_IP6:
>> +               af = AF_INET6;
>> +               break;
>> +       default:
>> +               pr_err("address family %d not supported\n",
>> +                               nport->disc_addr.adrfam);
>> +               ret = -EINVAL;
>> +               goto out_free_port;
>> +       }
>> +
>> +       if (nport->inline_data_size < 0) {
>> +               nport->inline_data_size =
>> NVMET_RDMA_DEFAULT_INLINE_DATA_SIZE;
>> +       } else if (nport->inline_data_size >
>> NVMET_RDMA_MAX_INLINE_DATA_SIZE) {
>> +               pr_warn("inline_data_size %u is too large, reducing to
>> %u\n",
>> +                       nport->inline_data_size,
>> +                       NVMET_RDMA_MAX_INLINE_DATA_SIZE);
>> +               nport->inline_data_size = NVMET_RDMA_MAX_INLINE_DATA_SIZE;
>> +       }
>> +
>> +       ret = inet_pton_with_scope(&init_net, af, nport->disc_addr.traddr,
>> +                       nport->disc_addr.trsvcid, &port->addr);
>> +       if (ret) {
>> +               pr_err("malformed ip/port passed: %s:%s\n",
>> +                       nport->disc_addr.traddr, nport->disc_addr.trsvcid);
>> +               goto out_free_port;
>> +       }
>> +
>> +       ret = nvmet_rdma_enable_port(port);
>> +       if(ret)
>> +               goto out_free_port;
>> +
>> +       pr_info("enabling port %d (%pISpcs)\n",
>> +               le16_to_cpu(nport->disc_addr.portid),
>> +               (struct sockaddr *)&port->addr);
>> +
>> +       return 0;
>> +
>> +out_free_port:
>> +       kfree(port);
>> +       return ret;
>> +}
>> +
>> +static void nvmet_rdma_remove_port(struct nvmet_port *nport)
>> +{
>> +       struct nvmet_rdma_port *port = nport->priv;
>> +
>> +       cancel_delayed_work_sync(&port->repair_work);
>> +       nvmet_rdma_disable_port(port);
>> +       kfree(port);
>>    }
>>
>>    static void nvmet_rdma_disc_port_addr(struct nvmet_req *req,
>> -               struct nvmet_port *port, char *traddr)
>> +               struct nvmet_port *nport, char *traddr)
>>    {
>> -       struct rdma_cm_id *cm_id = port->priv;
>> +       struct nvmet_rdma_port *port = nport->priv;
>> +       struct rdma_cm_id *cm_id = port->cm_id;
>>
>>           if (inet_addr_is_any((struct sockaddr
>> *)&cm_id->route.addr.src_addr)) {
>>                   struct nvmet_rdma_rsp *rsp =
>> @@ -1603,7 +1665,7 @@ static void nvmet_rdma_disc_port_addr(struct
>> nvmet_req *req,
>>
>>                   sprintf(traddr, "%pISc", addr);
>>           } else {
>> -               memcpy(traddr, port->disc_addr.traddr, NVMF_TRADDR_SIZE);
>> +               memcpy(traddr, nport->disc_addr.traddr, NVMF_TRADDR_SIZE);
>>           }
>>    }
>> --

_______________________________________________
linux-nvme mailing list
linux-nvme@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme

Re: NULL pointer dereference in nvmet_rdma_queue_disconnect during bond failover

[Sagi Grimberg]

>> This patch did not apply to latest 5.2 (5.2.21). All of 10 hunks
>> failed.

It applies on branch linux-5.2.y in the stable tree. What are you
using?

>> I applied it manually, and also handled cm_id->context in
>> nvmet_rdma_find_get_device as I mentioned earlier. I am attaching the
>> patch that I tested on kernel 5.2.21 (target side). I confirm that
>> this patch fixes the bond failover issue.
>>
>> Tested-by: Alex Lyakas <alex@zadara.com>

Cool.

>>
>> Max, will this help to deliver this fix upstream, so that we can get
>> it in MOFED 4.9?
>>
>> Thanks,
>> Alex.
> 
> Alex,
> 
> Thanks for testing this.
> 
> Waiting for Sagi's official rebased version for doing full review.

I'll send a patch...

_______________________________________________
linux-nvme mailing list
linux-nvme@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme

RE: NULL pointer dereference in nvmet_rdma_queue_disconnect during bond failover

[Shlomi Nimrodi]

++

-----Original Message-----
From: Sagi Grimberg <sagi@grimberg.me> 
Sent: Thursday, April 2, 2020 18:27
To: Max Gurtovoy <maxg@mellanox.com>; Alex Lyakas <alex@zadara.com>
Cc: Tom Wu <tomwu@mellanox.com>; Israel Rukshin <israelr@mellanox.com>; linux-nvme <linux-nvme@lists.infradead.org>; Shlomi Nimrodi <shlomin@mellanox.com>
Subject: Re: NULL pointer dereference in nvmet_rdma_queue_disconnect during bond failover


>> This patch did not apply to latest 5.2 (5.2.21). All of 10 hunks 
>> failed.

It applies on branch linux-5.2.y in the stable tree. What are you using?

>> I applied it manually, and also handled cm_id->context in 
>> nvmet_rdma_find_get_device as I mentioned earlier. I am attaching the 
>> patch that I tested on kernel 5.2.21 (target side). I confirm that 
>> this patch fixes the bond failover issue.
>>
>> Tested-by: Alex Lyakas <alex@zadara.com>

Cool.

>>
>> Max, will this help to deliver this fix upstream, so that we can get 
>> it in MOFED 4.9?
>>
>> Thanks,
>> Alex.
> 
> Alex,
> 
> Thanks for testing this.
> 
> Waiting for Sagi's official rebased version for doing full review.

I'll send a patch...
_______________________________________________
linux-nvme mailing list
linux-nvme@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme