Skip to contentSkip to conversation

net/failsafe: fix fd leak

PATCH Applied wangyunjian 5 participants 23 messages Apr 27, 2020 – May 11, 2020 Network drivers

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 net/failsafe: fix fd leak

From: Yunjian Wang <[email protected]>

Zero is a valid fd. The fd won't be closed thus leading fd leak, when it is zero.

Fixes f234e5bd996d ("net/failsafe: register slaves Rx interrupts")
Fixes 9e0360aebf23 ("net/failsafe: register as Rx interrupt mode")
---
 drivers/net/failsafe/failsafe_intr.c | 2 +-
 drivers/net/failsafe/failsafe_ops.c  | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)
drivers/net/failsafe/failsafe_intr.c +1 -1
index d8728fe7e..602c04033 100644
@@ -393,7 +393,7 @@ fs_rx_event_proxy_uninstall(struct fs_priv *priv)
 		free(priv->rxp.evec);
 		priv->rxp.evec = NULL;
 	}
-	if (priv->rxp.efd > 0) {
+	if (priv->rxp.efd >= 0) {
 		close(priv->rxp.efd);
 		priv->rxp.efd = -1;
 	}
drivers/net/failsafe/failsafe_ops.c +1 -1
index 50f2aca4e..e1d08e46c 100644
@@ -380,7 +380,7 @@ fs_rx_queue_release(void *queue)
 	rxq = queue;
 	dev = &rte_eth_devices[rxq->priv->data->port_id];
 	fs_lock(dev, 0);
-	if (rxq->event_fd > 0)
+	if (rxq->event_fd >= 0)
 		close(rxq->event_fd);
 	FOREACH_SUBDEV_STATE(sdev, i, dev, DEV_ACTIVE) {
 		if (ETH(sdev)->data->rx_queues != NULL &&
Signature · 1 line
2.19.1
Acked-by on v1 1/1
#1 lore raw
Quoted 9 lines · wangyunjian

From: Yunjian Wang <[email protected]>

Zero is a valid fd. The fd won't be closed thus leading fd leak, when it is zero.

Fixes f234e5bd996d ("net/failsafe: register slaves Rx interrupts")
Fixes 9e0360aebf23 ("net/failsafe: register as Rx interrupt mode")

Hello Yunjian,

Nothing prevents a DPDK app from closing 0 and getting it from another call, good catch.

Quoted 1 line
Quoted 35 lines
---
 drivers/net/failsafe/failsafe_intr.c | 2 +-
 drivers/net/failsafe/failsafe_ops.c  | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)
drivers/net/failsafe/failsafe_intr.c +1 -1
index d8728fe7e..602c04033 100644
@@ -393,7 +393,7 @@ fs_rx_event_proxy_uninstall(struct fs_priv *priv)
 		free(priv->rxp.evec);
 		priv->rxp.evec = NULL;
 	}
-	if (priv->rxp.efd > 0) {
+	if (priv->rxp.efd >= 0) {
 		close(priv->rxp.efd);
 		priv->rxp.efd = -1;
 	}
drivers/net/failsafe/failsafe_ops.c +1 -1
index 50f2aca4e..e1d08e46c 100644
@@ -380,7 +380,7 @@ fs_rx_queue_release(void *queue)
 	rxq = queue;
 	dev = &rte_eth_devices[rxq->priv->data->port_id];
 	fs_lock(dev, 0);
-	if (rxq->event_fd > 0)
+	if (rxq->event_fd >= 0)
 		close(rxq->event_fd);
 	FOREACH_SUBDEV_STATE(sdev, i, dev, DEV_ACTIVE) {
 		if (ETH(sdev)->data->rx_queues != NULL &&
Signature · 1 line
2.19.1
Signature · 1 line
Gaëtan
Applied
#2 lore raw
Quoted 11 lines · Gaëtan Rivet
Quoted 9 lines · wangyunjian

From: Yunjian Wang <[email protected]>

Zero is a valid fd. The fd won't be closed thus leading fd leak, when it is zero.

Fixes f234e5bd996d ("net/failsafe: register slaves Rx interrupts")
Fixes 9e0360aebf23 ("net/failsafe: register as Rx interrupt mode")
Quoted 8 lines

Hello Yunjian,

Nothing prevents a DPDK app from closing 0 and getting it from another call, good catch.

Quoted 1 line

Applied to dpdk-next-net/master, thanks.

#3 lore raw

Hi,

Quoted 23 lines

-----Original Message-----
From: dev <[email protected]> On Behalf Of Ferruh Yigit
Sent: Monday, April 27, 2020 7:56 PM
To: Gaëtan Rivet <[email protected]>; wangyunjian <[email protected]>
Cc: [email protected]; [email protected]; [email protected]; [email protected]
Subject: Re: [dpdk-dev] [dpdk-stable] [PATCH] net/failsafe: fix fd leak

Quoted 13 lines · Gaëtan Rivet
Quoted 9 lines · wangyunjian

From: Yunjian Wang <[email protected]>

Zero is a valid fd. The fd won't be closed thus leading fd leak, when it is zero.

Fixes f234e5bd996d ("net/failsafe: register slaves Rx interrupts")
Fixes 9e0360aebf23 ("net/failsafe: register as Rx interrupt mode")

Hello Yunjian,

Quoted 8 lines
Quoted 6 lines

Nothing prevents a DPDK app from closing 0 and getting it from another call, good catch.

Quoted 1 line

Applied to dpdk-next-net/master, thanks.

This patch is causing Testpmd to quit when I issue a "port stop" command. Testpmd log:

"""
x86_64-native-linuxapp-gcc/build/app/test-pmd/testpmd -n 4 -- -i --forward-mode=mac
EAL: Detected 8 lcore(s)
EAL: Detected 1 NUMA nodes
EAL: Multi-process socket /var/run/dpdk/rte/mp_socket
EAL: Selected IOVA mode 'PA'
EAL: No available hugepages reported in hugepages-1048576kB
EAL: Probing VFIO support...
EAL: PCI device 0002:00:02.0 on NUMA socket 0
EAL: probe driver: 15b3:1004 net_mlx4
Interactive-mode selected
Set mac packet forwarding mode
Warning: NUMA should be configured manually by using --port-numa-config and --ring-numa-config parameters along with --numa.
testpmd: create a new mbuf pool <mbuf_pool_socket_0>: n=203456, size=2176, socket=0
testpmd: preferred mempool ops selected: ring_mp_mc

Warning! port-topology=paired and odd forward ports number, the last port will pair with itself.

Configuring Port 1 (socket 0)
Port 1: 00:15:5D:26:2B:00
Checking link statuses...
Done
testpmd> port stop 1
Stopping ports...
Checking link statuses...
Done
testpmd>
Stopping port 1...
Stopping ports...
Done

Shutting down port 1...
Closing ports...
Done

Bye...
"""

My terminal gets broken at this point, and I have to reinitialize it with a "reset".

- Ali

#4 lore raw
Quoted 71 lines · Ali Alnubani

Hi,

Quoted 23 lines

-----Original Message-----
From: dev <[email protected]> On Behalf Of Ferruh Yigit
Sent: Monday, April 27, 2020 7:56 PM
To: Gaëtan Rivet <[email protected]>; wangyunjian <[email protected]>
Cc: [email protected]; [email protected]; [email protected]; [email protected]
Subject: Re: [dpdk-dev] [dpdk-stable] [PATCH] net/failsafe: fix fd leak

Quoted 13 lines · Gaëtan Rivet
Quoted 9 lines · wangyunjian

From: Yunjian Wang <[email protected]>

Zero is a valid fd. The fd won't be closed thus leading fd leak, when it is zero.

Fixes f234e5bd996d ("net/failsafe: register slaves Rx interrupts")
Fixes 9e0360aebf23 ("net/failsafe: register as Rx interrupt mode")

Hello Yunjian,

Quoted 8 lines
Quoted 6 lines

Nothing prevents a DPDK app from closing 0 and getting it from another call, good catch.

Quoted 1 line

Applied to dpdk-next-net/master, thanks.

This patch is causing Testpmd to quit when I issue a "port stop" command. Testpmd log:

"""
x86_64-native-linuxapp-gcc/build/app/test-pmd/testpmd -n 4 -- -i --forward-mode=mac
EAL: Detected 8 lcore(s)
EAL: Detected 1 NUMA nodes
EAL: Multi-process socket /var/run/dpdk/rte/mp_socket
EAL: Selected IOVA mode 'PA'
EAL: No available hugepages reported in hugepages-1048576kB
EAL: Probing VFIO support...
EAL: PCI device 0002:00:02.0 on NUMA socket 0
EAL: probe driver: 15b3:1004 net_mlx4
Interactive-mode selected
Set mac packet forwarding mode
Warning: NUMA should be configured manually by using --port-numa-config and --ring-numa-config parameters along with --numa.
testpmd: create a new mbuf pool <mbuf_pool_socket_0>: n=203456, size=2176, socket=0
testpmd: preferred mempool ops selected: ring_mp_mc

Warning! port-topology=paired and odd forward ports number, the last port will pair with itself.

Configuring Port 1 (socket 0)
Port 1: 00:15:5D:26:2B:00
Checking link statuses...
Done
testpmd> port stop 1
Stopping ports...
Checking link statuses...
Done
testpmd>
Stopping port 1...
Stopping ports...
Done

Shutting down port 1...
Closing ports...
Done

Quoted 6 lines

Bye...
"""

My terminal gets broken at this point, and I have to reinitialize it with a "reset".

- Ali

Hi Ali,

Thanks for the report, I am looking into it.

Are you testing failsafe on Azure?

I see a segfault currently at startup, so in any case there are fixes to be pushed. I'll see afterward if I need a specific platform to reproduce your bug.

Regards,

Signature · 1 line
Gaëtan
#5 lore raw
Quoted 68 lines · On Mon, 4 May 2020 18:22:26 +0200 Gaëtan Rivet <[email protected]> wrote:
Quoted 67 lines · Ali Alnubani

Hi,

Quoted 23 lines

-----Original Message-----
From: dev <[email protected]> On Behalf Of Ferruh Yigit
Sent: Monday, April 27, 2020 7:56 PM
To: Gaëtan Rivet <[email protected]>; wangyunjian <[email protected]>
Cc: [email protected]; [email protected]; [email protected]; [email protected]
Subject: Re: [dpdk-dev] [dpdk-stable] [PATCH] net/failsafe: fix fd leak

Quoted 13 lines · Gaëtan Rivet
Quoted 9 lines · wangyunjian

From: Yunjian Wang <[email protected]>

Zero is a valid fd. The fd won't be closed thus leading fd leak, when it is zero.

Fixes f234e5bd996d ("net/failsafe: register slaves Rx interrupts")
Fixes 9e0360aebf23 ("net/failsafe: register as Rx interrupt mode")

Hello Yunjian,

Quoted 8 lines
Quoted 6 lines

Nothing prevents a DPDK app from closing 0 and getting it from another call, good catch.

Quoted 1 line

Applied to dpdk-next-net/master, thanks.

This patch is causing Testpmd to quit when I issue a "port stop" command. Testpmd log:

"""
x86_64-native-linuxapp-gcc/build/app/test-pmd/testpmd -n 4 -- -i --forward-mode=mac
EAL: Detected 8 lcore(s)
EAL: Detected 1 NUMA nodes
EAL: Multi-process socket /var/run/dpdk/rte/mp_socket
EAL: Selected IOVA mode 'PA'
EAL: No available hugepages reported in hugepages-1048576kB
EAL: Probing VFIO support...
EAL: PCI device 0002:00:02.0 on NUMA socket 0
EAL: probe driver: 15b3:1004 net_mlx4
Interactive-mode selected
Set mac packet forwarding mode
Warning: NUMA should be configured manually by using --port-numa-config and --ring-numa-config parameters along with --numa.
testpmd: create a new mbuf pool <mbuf_pool_socket_0>: n=203456, size=2176, socket=0
testpmd: preferred mempool ops selected: ring_mp_mc

Warning! port-topology=paired and odd forward ports number, the last port will pair with itself.

Configuring Port 1 (socket 0)
Port 1: 00:15:5D:26:2B:00
Checking link statuses...
Done
testpmd> port stop 1
Stopping ports...
Checking link statuses...
Done
testpmd>
Stopping port 1...
Stopping ports...
Done

Quoted 8 lines
Quoted 8 lines

Shutting down port 1...
Closing ports...
Done

Bye...
"""

My terminal gets broken at this point, and I have to reinitialize it with a "reset".

The problem is that you did not blacklist the PCI address of the Mellanox device associated with your login session (normally this is the PCI device associated with eth0).

By default, DPDK will take over all VF devices it finds as part of the Mellanox device startup. This means the traffic that was going to the VF associated with eth0 (your ssh) is now going to DPDK; which is not what you want.

The solution is to either use blacklist (-b option) or whitelist (-w option) to get only the PCI devices you want to be part of the DPDK.

#6 lore raw

Hi Stephen,

Quoted 103 lines

-----Original Message-----
From: Stephen Hemminger <[email protected]>
Sent: Monday, May 4, 2020 7:29 PM
To: Gaëtan Rivet <[email protected]>
Cc: Ali Alnubani <[email protected]>; Ferruh Yigit <[email protected]>; wangyunjian <[email protected]>; [email protected]; [email protected]; [email protected]; [email protected]; Raslan Darawsheh <[email protected]>
Subject: Re: [dpdk-dev] [dpdk-stable] [PATCH] net/failsafe: fix fd leak

Quoted 32 lines · On Mon, 4 May 2020 18:22:26 +0200 Gaëtan Rivet <[email protected]> wrote:
Quoted 31 lines · Ali Alnubani

Hi,

Quoted 29 lines

-----Original Message-----
From: dev <[email protected]> On Behalf Of Ferruh Yigit
Sent: Monday, April 27, 2020 7:56 PM
To: Gaëtan Rivet <[email protected]>; wangyunjian <[email protected]>
Cc: [email protected]; [email protected]; [email protected]; [email protected]
Subject: Re: [dpdk-dev] [dpdk-stable] [PATCH] net/failsafe: fix fd leak

Quoted 18 lines · Gaëtan Rivet
Quoted 11 lines · wangyunjian

From: Yunjian Wang <[email protected]>

Zero is a valid fd. The fd won't be closed thus leading fd leak, when it is zero.

Fixes f234e5bd996d ("net/failsafe: register slaves Rx

interrupts")

Fixes 9e0360aebf23 ("net/failsafe: register as Rx interrupt

mode")
Cc: [email protected]

Hello Yunjian,

Nothing prevents a DPDK app from closing 0 and getting it from another call, good catch.

Quoted 7 lines
Quoted 7 lines
Quoted 5 lines
Quoted 3 lines
Quoted 1 line

Applied to dpdk-next-net/master, thanks.

This patch is causing Testpmd to quit when I issue a "port stop" command.

Testpmd log:

Quoted 15 lines
Quoted 15 lines

"""
x86_64-native-linuxapp-gcc/build/app/test-pmd/testpmd -n 4 -- -i --forward-mode=mac
EAL: Detected 8 lcore(s)
EAL: Detected 1 NUMA nodes
EAL: Multi-process socket /var/run/dpdk/rte/mp_socket
EAL: Selected IOVA mode 'PA'
EAL: No available hugepages reported in hugepages-1048576kB
EAL: Probing VFIO support...
EAL: PCI device 0002:00:02.0 on NUMA socket 0
EAL: probe driver: 15b3:1004 net_mlx4
Interactive-mode selected
Set mac packet forwarding mode
Warning: NUMA should be configured manually by using --port-numa-config

and --ring-numa-config parameters along with --numa.

Quoted 5 lines
Quoted 5 lines

testpmd: create a new mbuf pool <mbuf_pool_socket_0>: n=203456, size=2176, socket=0
testpmd: preferred mempool ops selected: ring_mp_mc

Warning! port-topology=paired and odd forward ports number, the last

port will pair with itself.

Quoted 14 lines
Quoted 14 lines

Configuring Port 1 (socket 0)
Port 1: 00:15:5D:26:2B:00
Checking link statuses...
Done
testpmd> port stop 1
Stopping ports...
Checking link statuses...
Done
testpmd>
Stopping port 1...
Stopping ports...
Done

Quoted 8 lines
Quoted 8 lines

Shutting down port 1...
Closing ports...
Done

Bye...
"""

My terminal gets broken at this point, and I have to reinitialize it with a

"reset".

The problem is that you did not blacklist the PCI address of the Mellanox device associated with your login session (normally this is the PCI device associated with eth0).

Quoted 6 lines

By default, DPDK will take over all VF devices it finds as part of the Mellanox device startup. This means the traffic that was going to the VF associated with eth0 (your ssh) is now going to DPDK; which is not what you want.

The solution is to either use blacklist (-b option) or whitelist (-w option) to get only the PCI devices you want to be part of the DPDK.

I'm confused.
I don't think my login interface was being whitelisted in DPDK because my issue isn't that I lose my connection, it's that testpmd quits when I stop a port.

Regards,
Ali

#7 lore raw
In reply to Gaëtan Rivet
Quoted 96 lines

-----Original Message-----
From: Gaëtan Rivet <[email protected]>
Sent: Monday, May 4, 2020 7:22 PM
To: Ali Alnubani <[email protected]>
Cc: Ferruh Yigit <[email protected]>; wangyunjian <[email protected]>; [email protected]; [email protected]; [email protected]; [email protected]; Raslan Darawsheh <[email protected]>
Subject: Re: [dpdk-dev] [dpdk-stable] [PATCH] net/failsafe: fix fd leak

Quoted 31 lines · Ali Alnubani

Hi,

Quoted 29 lines

-----Original Message-----
From: dev <[email protected]> On Behalf Of Ferruh Yigit
Sent: Monday, April 27, 2020 7:56 PM
To: Gaëtan Rivet <[email protected]>; wangyunjian <[email protected]>
Cc: [email protected]; [email protected]; [email protected]; [email protected]
Subject: Re: [dpdk-dev] [dpdk-stable] [PATCH] net/failsafe: fix fd leak

Quoted 18 lines · Gaëtan Rivet
Quoted 11 lines · wangyunjian

From: Yunjian Wang <[email protected]>

Zero is a valid fd. The fd won't be closed thus leading fd leak, when it is zero.

Fixes f234e5bd996d ("net/failsafe: register slaves Rx

interrupts")

Fixes 9e0360aebf23 ("net/failsafe: register as Rx interrupt

mode")
Cc: [email protected]

Hello Yunjian,

Nothing prevents a DPDK app from closing 0 and getting it from another call, good catch.

Quoted 7 lines
Quoted 5 lines
Quoted 3 lines
Quoted 1 line

Applied to dpdk-next-net/master, thanks.

This patch is causing Testpmd to quit when I issue a "port stop" command.

Testpmd log:

Quoted 15 lines

"""
x86_64-native-linuxapp-gcc/build/app/test-pmd/testpmd -n 4 -- -i --forward-mode=mac
EAL: Detected 8 lcore(s)
EAL: Detected 1 NUMA nodes
EAL: Multi-process socket /var/run/dpdk/rte/mp_socket
EAL: Selected IOVA mode 'PA'
EAL: No available hugepages reported in hugepages-1048576kB
EAL: Probing VFIO support...
EAL: PCI device 0002:00:02.0 on NUMA socket 0
EAL: probe driver: 15b3:1004 net_mlx4
Interactive-mode selected
Set mac packet forwarding mode
Warning: NUMA should be configured manually by using --port-numa-config

and --ring-numa-config parameters along with --numa.

Quoted 5 lines

testpmd: create a new mbuf pool <mbuf_pool_socket_0>: n=203456, size=2176, socket=0
testpmd: preferred mempool ops selected: ring_mp_mc

Warning! port-topology=paired and odd forward ports number, the last port

will pair with itself.

Quoted 14 lines

Configuring Port 1 (socket 0)
Port 1: 00:15:5D:26:2B:00
Checking link statuses...
Done
testpmd> port stop 1
Stopping ports...
Checking link statuses...
Done
testpmd>
Stopping port 1...
Stopping ports...
Done

Quoted 8 lines

Shutting down port 1...
Closing ports...
Done

Bye...
"""

My terminal gets broken at this point, and I have to reinitialize it with a

"reset".

Quoted 1 line
Quoted 7 lines
Quoted 1 line

- Ali

Hi Ali,

Thanks for the report, I am looking into it.

Are you testing failsafe on Azure?

This reproduces with Failsafe, but not necessarily on Azure. You can try to reproduce on any platform if you pass something like '-w 00:00.0 --vdev="net_failsafe0,dev(0000:08:00.0)"'.

Quoted 8 lines

I see a segfault currently at startup, so in any case there are fixes to be pushed.
I'll see afterward if I need a specific platform to reproduce your bug.

Regards,
--
Gaëtan

Regards,
Ali

#8 lore raw
Quoted 98 lines · Ali Alnubani
Quoted 98 lines

-----Original Message-----
From: Gaëtan Rivet <[email protected]>
Sent: Monday, May 4, 2020 7:22 PM
To: Ali Alnubani <[email protected]>
Cc: Ferruh Yigit <[email protected]>; wangyunjian <[email protected]>; [email protected]; [email protected]; [email protected]; [email protected]; Raslan Darawsheh <[email protected]>
Subject: Re: [dpdk-dev] [dpdk-stable] [PATCH] net/failsafe: fix fd leak

Quoted 31 lines · Ali Alnubani

Hi,

Quoted 29 lines

-----Original Message-----
From: dev <[email protected]> On Behalf Of Ferruh Yigit
Sent: Monday, April 27, 2020 7:56 PM
To: Gaëtan Rivet <[email protected]>; wangyunjian <[email protected]>
Cc: [email protected]; [email protected]; [email protected]; [email protected]
Subject: Re: [dpdk-dev] [dpdk-stable] [PATCH] net/failsafe: fix fd leak

Quoted 18 lines · Gaëtan Rivet
Quoted 11 lines · wangyunjian

From: Yunjian Wang <[email protected]>

Zero is a valid fd. The fd won't be closed thus leading fd leak, when it is zero.

Fixes f234e5bd996d ("net/failsafe: register slaves Rx

interrupts")

Fixes 9e0360aebf23 ("net/failsafe: register as Rx interrupt

mode")
Cc: [email protected]

Hello Yunjian,

Nothing prevents a DPDK app from closing 0 and getting it from another call, good catch.

Quoted 7 lines
Quoted 5 lines
Quoted 3 lines
Quoted 1 line

Applied to dpdk-next-net/master, thanks.

This patch is causing Testpmd to quit when I issue a "port stop" command.

Testpmd log:

Quoted 15 lines

"""
x86_64-native-linuxapp-gcc/build/app/test-pmd/testpmd -n 4 -- -i --forward-mode=mac
EAL: Detected 8 lcore(s)
EAL: Detected 1 NUMA nodes
EAL: Multi-process socket /var/run/dpdk/rte/mp_socket
EAL: Selected IOVA mode 'PA'
EAL: No available hugepages reported in hugepages-1048576kB
EAL: Probing VFIO support...
EAL: PCI device 0002:00:02.0 on NUMA socket 0
EAL: probe driver: 15b3:1004 net_mlx4
Interactive-mode selected
Set mac packet forwarding mode
Warning: NUMA should be configured manually by using --port-numa-config

and --ring-numa-config parameters along with --numa.

Quoted 5 lines

testpmd: create a new mbuf pool <mbuf_pool_socket_0>: n=203456, size=2176, socket=0
testpmd: preferred mempool ops selected: ring_mp_mc

Warning! port-topology=paired and odd forward ports number, the last port

will pair with itself.

Quoted 14 lines

Configuring Port 1 (socket 0)
Port 1: 00:15:5D:26:2B:00
Checking link statuses...
Done
testpmd> port stop 1
Stopping ports...
Checking link statuses...
Done
testpmd>
Stopping port 1...
Stopping ports...
Done

Quoted 8 lines

Shutting down port 1...
Closing ports...
Done

Bye...
"""

My terminal gets broken at this point, and I have to reinitialize it with a

"reset".

Quoted 2 lines

- Ali

Quoted 8 lines
Quoted 5 lines

Hi Ali,

Thanks for the report, I am looking into it.

Are you testing failsafe on Azure?

This reproduces with Failsafe, but not necessarily on Azure. You can try to reproduce on any platform if you pass something like '-w 00:00.0 --vdev="net_failsafe0,dev(0000:08:00.0)"'.

Hi,

Indeed, I am able to reproduce the issue using this command:
bash> ./build/app/dpdk-testpmd -n4 -m 4096 --no-huge --vdev='net_failsafe0,dev(net_ring0)' -- -i
(no need of PCI bus nor hugepages to validate failsafe sometimes).

I was asking about Azure because you did not give the command line options for fail-safe, so I assumed it had been probed automagically.

I made a fix, will send soon.

Regards,

Signature · 1 line
Gaëtan
#9 lore raw

cover failsafe & ring fixes

In reply to Ali Alnubani

Some issues seen on next-net.

Gaetan Rivet (3):
net/failsafe: avoid crash on malformed eth_dev net/ring: fix eth_dev device pointer on allocation net/failsafe: fix default service proxy state

 drivers/net/failsafe/failsafe.c         |  1 +
 drivers/net/failsafe/failsafe_ether.c   |  5 ++++
 drivers/net/failsafe/failsafe_private.h |  8 ++++++
 drivers/net/ring/rte_eth_ring.c         | 36 ++++++++++++++++++++-----
 4 files changed, 43 insertions(+), 7 deletions(-)
Signature · 1 line
2.26.2
#10 lore raw

PATCH 1/3 net/failsafe: avoid crash on malformed eth_dev

Some PMD do not respect the eth_dev API when allocating their rte_eth_dev. As a result, on device add event resulting from rte_eth_dev_probing_finish() call, the eth_dev processed is incomplete.

The segfault is a good way to focus the developer on the issue, but does not inspire confidence. Instead, warn the user of the error repeatedly.

The failsafe PMD can warn of the issue and continue. It will repeatedly attempt to initialize the failed port and complain about it, which should result in the same developer focus but with less crashing.

---
 drivers/net/failsafe/failsafe_ether.c | 5 +++++
 1 file changed, 5 insertions(+)
drivers/net/failsafe/failsafe_ether.c +5 -0
index 93deacd13..2b748bd8b 100644
@@ -623,6 +623,11 @@ failsafe_eth_new_event_callback(uint16_t port_id,
 	FOREACH_SUBDEV_STATE(sdev, i, fs_dev, DEV_PARSED) {
 		if (sdev->state >= DEV_PROBED)
 			continue;
+		if (dev->device == NULL) {
+			WARN("Trying to probe malformed device %s.\n",
+			     sdev->devargs.name);
+			continue;
+		}
 		if (strcmp(sdev->devargs.name, dev->device->name) != 0)
 			continue;
 		rte_eth_dev_owner_set(port_id, &PRIV(fs_dev)->my_owner);
Signature · 1 line
2.26.2
Applied
#11 lore raw

Re: [dpdk-dev] [PATCH v1 1/3] net/failsafe: avoid crash on malformed eth_dev

Quoted 4 lines · Gaetan Rivet

Some PMD do not respect the eth_dev API when allocating their rte_eth_dev. As a result, on device add event resulting from rte_eth_dev_probing_finish() call, the eth_dev processed is incomplete.

Quoted 8 lines

The segfault is a good way to focus the developer on the issue, but does not inspire confidence. Instead, warn the user of the error repeatedly.

The failsafe PMD can warn of the issue and continue. It will repeatedly attempt to initialize the failed port and complain about it, which should result in the same developer focus but with less crashing.

Applied to dpdk-next-net/master, thanks.

#12 lore raw

PATCH 2/3 net/ring: fix eth_dev device pointer on allocation

In reply to Gaetan Rivet

When a net_ring device is allocated, its device pointer is not set before calling rte_eth_dev_probing_finish, which is incorrect.

The following:
commit: 96cb19521147 ("net/ring: use EAL APIs in PMD specific API")
commit: a6992e961050 ("net/ring: set ethernet device field")

already attempted to fix this issue in 17.08, which was fine at the time. Adding the hook rte_eth_dev_probing_finish() however created this bug, as the eth_dev exposed when this hook is executed is expected to be complete.

Remove the prior attempts to fix the issue in rte_pmd_ring_probe() and write the pointer properly in do_eth_dev_ring_create().

Fixes fbe90cdd776c ("ethdev: add probing finish function")
---
 drivers/net/ring/rte_eth_ring.c | 36 ++++++++++++++++++++++++++-------
 1 file changed, 29 insertions(+), 7 deletions(-)
drivers/net/ring/rte_eth_ring.c +29 -7
index 41acbc513..ad27f9d89 100644
@@ -244,6 +244,26 @@ static const struct eth_dev_ops ops = {
 	.mac_addr_add = eth_mac_addr_add,
 };
 
+static int
+dev_name_cmp(const struct rte_device *dev, const void *_name)
+{
+	const char *name = _name;
+
+	return strcmp(dev->name, name);
+}
+
+static struct rte_device *
+ring_device_from_name(const char *name)
+{
+	struct rte_bus *vdev_bus;
+
+	vdev_bus = rte_bus_find_by_name("vdev");
+	if (vdev_bus == NULL)
+		return NULL;
+
+	return vdev_bus->find_device(NULL, dev_name_cmp, name);
+}
+
 static int
 do_eth_dev_ring_create(const char *name,
 		struct rte_ring * const rx_queues[],
@@ -294,6 +314,7 @@ do_eth_dev_ring_create(const char *name,
 	 * - store queue data in internals,
 	 * - store numa_node info in eth_dev_data
 	 * - point eth_dev_data to internals
+	 * - store EAL device in eth_dev,
 	 * - and point eth_dev structure to new eth_dev_data structure
 	 */
 
@@ -325,10 +346,17 @@ do_eth_dev_ring_create(const char *name,
 	data->kdrv = RTE_KDRV_NONE;
 	data->numa_node = numa_node;
 
-	/* finally assign rx and tx ops */
+	/* assign rx and tx ops */
 	eth_dev->rx_pkt_burst = eth_ring_rx;
 	eth_dev->tx_pkt_burst = eth_ring_tx;
 
+	/* finally set the rte_device pointer in eth_dev. */
+	eth_dev->device = ring_device_from_name(name);
+	if (eth_dev->device == NULL) {
+		rte_errno = ENODEV;
+		goto error;
+	}
+
 	rte_eth_dev_probing_finish(eth_dev);
 	*eth_dev_p = eth_dev;
 
@@ -584,9 +612,6 @@ rte_pmd_ring_probe(struct rte_vdev_device *dev)
 							  DEV_ATTACH, &eth_dev);
 			}
 
-			if (eth_dev)
-				eth_dev->device = &dev->device;
-
 			return ret;
 		}
 
@@ -644,9 +669,6 @@ rte_pmd_ring_probe(struct rte_vdev_device *dev)
 		}
 	}
 
-	if (eth_dev)
-		eth_dev->device = &dev->device;
-
 out_free:
 	rte_kvargs_free(kvlist);
 	rte_free(info);
Signature · 1 line
2.26.2
#13 lore raw

Re: [dpdk-dev] [PATCH v1 2/3] net/ring: fix eth_dev device pointer on allocation

Quoted 20 lines · Gaetan Rivet

When a net_ring device is allocated, its device pointer is not set before calling rte_eth_dev_probing_finish, which is incorrect.

The following:
commit: 96cb19521147 ("net/ring: use EAL APIs in PMD specific API")
commit: a6992e961050 ("net/ring: set ethernet device field")

already attempted to fix this issue in 17.08, which was fine at the time. Adding the hook rte_eth_dev_probing_finish() however created this bug, as the eth_dev exposed when this hook is executed is expected to be complete.

Remove the prior attempts to fix the issue in rte_pmd_ring_probe() and write the pointer properly in do_eth_dev_ring_create().

Fixes fbe90cdd776c ("ethdev: add probing finish function")

<...>

Quoted 8 lines
drivers/net/ring/rte_eth_ring.c +1 -1
@@ -325,10 +346,17 @@ do_eth_dev_ring_create(const char *name,
 	data->kdrv = RTE_KDRV_NONE;
 	data->numa_node = numa_node;
 
-	/* finally assign rx and tx ops */
+	/* assign rx and tx ops */
 	eth_dev->rx_pkt_burst = eth_ring_rx;
 	eth_dev->tx_pkt_burst = eth_ring_tx;

drivers/net/ring/rte_eth_ring.c do_eth_dev_ring_create(const char *name, · line 361

Quoted 10 lines
drivers/net/ring/rte_eth_ring.c +7 -0
 
+	/* finally set the rte_device pointer in eth_dev. */
+	eth_dev->device = ring_device_from_name(name);
+	if (eth_dev->device == NULL) {
+		rte_errno = ENODEV;
+		goto error;
+	}
+
 	rte_eth_dev_probing_finish(eth_dev);
 	*eth_dev_p = eth_dev;

Why not move the 'rte_eth_dev_probing_finish()' to the 'rte_pmd_ring_probe()', below where 'eth_dev->device' set?

#14 lore raw

Re: [dpdk-dev] [PATCH v1 2/3] net/ring: fix eth_dev device pointer on allocation

Quoted 33 lines · Ferruh Yigit
Quoted 20 lines · Gaetan Rivet

When a net_ring device is allocated, its device pointer is not set before calling rte_eth_dev_probing_finish, which is incorrect.

The following:
commit: 96cb19521147 ("net/ring: use EAL APIs in PMD specific API")
commit: a6992e961050 ("net/ring: set ethernet device field")

already attempted to fix this issue in 17.08, which was fine at the time. Adding the hook rte_eth_dev_probing_finish() however created this bug, as the eth_dev exposed when this hook is executed is expected to be complete.

Remove the prior attempts to fix the issue in rte_pmd_ring_probe() and write the pointer properly in do_eth_dev_ring_create().

Fixes fbe90cdd776c ("ethdev: add probing finish function")

<...>

Quoted 9 lines
drivers/net/ring/rte_eth_ring.c +1 -1
@@ -325,10 +346,17 @@ do_eth_dev_ring_create(const char *name,
 	data->kdrv = RTE_KDRV_NONE;
 	data->numa_node = numa_node;
 
-	/* finally assign rx and tx ops */
+	/* assign rx and tx ops */
 	eth_dev->rx_pkt_burst = eth_ring_rx;
 	eth_dev->tx_pkt_burst = eth_ring_tx;
 
Quoted 12 lines

drivers/net/ring/rte_eth_ring.c do_eth_dev_ring_create(const char *name, · line 361

Quoted 9 lines
drivers/net/ring/rte_eth_ring.c +7 -0
+	/* finally set the rte_device pointer in eth_dev. */
+	eth_dev->device = ring_device_from_name(name);
+	if (eth_dev->device == NULL) {
+		rte_errno = ENODEV;
+		goto error;
+	}
+
 	rte_eth_dev_probing_finish(eth_dev);
 	*eth_dev_p = eth_dev;

Why not move the 'rte_eth_dev_probing_finish()' to the 'rte_pmd_ring_probe()', below where 'eth_dev->device' set?

Hi Ferruh,

Sure it would work. The reason I did it this way is two-fold:

* I disliked having two places where eth_dev->device was conditionally set. It makes it harder to read rte_pmd_ring_probe.

* I was actually thinking, doing this patch, that we should modify rte_eth_dev_allocate() to take an rte_device as parameter, as all eth_dev are meant to be backed by an rte_device. Keeping this in mind, I meant to move writing the pointer closer to the rte_eth_dev_allocate() call.

But you are right that it is needlessly verbose, using vdev_bus->find_device() to do this stuff. I'm ok with changing it as you described if you prefer.

Signature · 1 line
Gaëtan
#15 lore raw

Re: [dpdk-dev] [dpdk-stable] [PATCH v1 2/3] net/ring: fix eth_dev device pointer on allocation

Quoted 47 lines · Gaëtan Rivet
Quoted 33 lines · Ferruh Yigit
Quoted 20 lines · Gaetan Rivet

When a net_ring device is allocated, its device pointer is not set before calling rte_eth_dev_probing_finish, which is incorrect.

The following:
commit: 96cb19521147 ("net/ring: use EAL APIs in PMD specific API")
commit: a6992e961050 ("net/ring: set ethernet device field")

already attempted to fix this issue in 17.08, which was fine at the time. Adding the hook rte_eth_dev_probing_finish() however created this bug, as the eth_dev exposed when this hook is executed is expected to be complete.

Remove the prior attempts to fix the issue in rte_pmd_ring_probe() and write the pointer properly in do_eth_dev_ring_create().

Fixes fbe90cdd776c ("ethdev: add probing finish function")

<...>

Quoted 9 lines
drivers/net/ring/rte_eth_ring.c +1 -1
@@ -325,10 +346,17 @@ do_eth_dev_ring_create(const char *name,
 	data->kdrv = RTE_KDRV_NONE;
 	data->numa_node = numa_node;
 
-	/* finally assign rx and tx ops */
+	/* assign rx and tx ops */
 	eth_dev->rx_pkt_burst = eth_ring_rx;
 	eth_dev->tx_pkt_burst = eth_ring_tx;
 
Quoted 12 lines

drivers/net/ring/rte_eth_ring.c do_eth_dev_ring_create(const char *name, · line 361

Quoted 9 lines
drivers/net/ring/rte_eth_ring.c +7 -0
+	/* finally set the rte_device pointer in eth_dev. */
+	eth_dev->device = ring_device_from_name(name);
+	if (eth_dev->device == NULL) {
+		rte_errno = ENODEV;
+		goto error;
+	}
+
 	rte_eth_dev_probing_finish(eth_dev);
 	*eth_dev_p = eth_dev;

Why not move the 'rte_eth_dev_probing_finish()' to the 'rte_pmd_ring_probe()', below where 'eth_dev->device' set?

Quoted 6 lines

Hi Ferruh,

Sure it would work. The reason I did it this way is two-fold:

* I disliked having two places where eth_dev->device was conditionally set. It makes it harder to read rte_pmd_ring_probe.

Agree, what about using a 'goto' to have the assignment and 'rte_eth_dev_probing_finish()' in a single place?
But check seems needed since creation may failed at that stage, if you think better check can be done on 'ret' instead of 'eth_dev'...

Quoted 6 lines

* I was actually thinking, doing this patch, that we should modify rte_eth_dev_allocate() to take an rte_device as parameter, as all eth_dev are meant to be backed by an rte_device. Keeping this in mind, I meant to move writing the pointer closer to the rte_eth_dev_allocate() call.

That is a bigger change, may affect many (if not all) PMDs, so I think this can be considered when that change is available.

And although that change may fix the issues that 'eth_dev->device' is not set, which we had a few times before, not sure it worth to change all PMDs and ethdev API directly couple with rte_device, instead of PMD being the glue. Can be discussed more on its own patch.

Quoted 5 lines

But you are right that it is needlessly verbose, using vdev_bus->find_device() to do this stuff. I'm ok with changing it as you described if you prefer.

That was the concern, that is too much code to take a value which is already available a few level up the stack.

#16 lore raw

Re: [dpdk-dev] [dpdk-stable] [PATCH v1 2/3] net/ring: fix eth_dev device pointer on allocation

Quoted 76 lines · Ferruh Yigit
Quoted 47 lines · Gaëtan Rivet
Quoted 33 lines · Ferruh Yigit
Quoted 20 lines · Gaetan Rivet

When a net_ring device is allocated, its device pointer is not set before calling rte_eth_dev_probing_finish, which is incorrect.

The following:
commit: 96cb19521147 ("net/ring: use EAL APIs in PMD specific API")
commit: a6992e961050 ("net/ring: set ethernet device field")

already attempted to fix this issue in 17.08, which was fine at the time. Adding the hook rte_eth_dev_probing_finish() however created this bug, as the eth_dev exposed when this hook is executed is expected to be complete.

Remove the prior attempts to fix the issue in rte_pmd_ring_probe() and write the pointer properly in do_eth_dev_ring_create().

Fixes fbe90cdd776c ("ethdev: add probing finish function")

<...>

Quoted 9 lines
drivers/net/ring/rte_eth_ring.c +1 -1
@@ -325,10 +346,17 @@ do_eth_dev_ring_create(const char *name,
 	data->kdrv = RTE_KDRV_NONE;
 	data->numa_node = numa_node;
 
-	/* finally assign rx and tx ops */
+	/* assign rx and tx ops */
 	eth_dev->rx_pkt_burst = eth_ring_rx;
 	eth_dev->tx_pkt_burst = eth_ring_tx;
 
Quoted 12 lines

drivers/net/ring/rte_eth_ring.c do_eth_dev_ring_create(const char *name, · line 361

Quoted 9 lines
drivers/net/ring/rte_eth_ring.c +7 -0
+	/* finally set the rte_device pointer in eth_dev. */
+	eth_dev->device = ring_device_from_name(name);
+	if (eth_dev->device == NULL) {
+		rte_errno = ENODEV;
+		goto error;
+	}
+
 	rte_eth_dev_probing_finish(eth_dev);
 	*eth_dev_p = eth_dev;

Why not move the 'rte_eth_dev_probing_finish()' to the 'rte_pmd_ring_probe()', below where 'eth_dev->device' set?

Quoted 6 lines

Hi Ferruh,

Sure it would work. The reason I did it this way is two-fold:

* I disliked having two places where eth_dev->device was conditionally set. It makes it harder to read rte_pmd_ring_probe.

Agree, what about using a 'goto' to have the assignment and 'rte_eth_dev_probing_finish()' in a single place?
But check seems needed since creation may failed at that stage, if you think better check can be done on 'ret' instead of 'eth_dev'...

Quoted 6 lines

* I was actually thinking, doing this patch, that we should modify rte_eth_dev_allocate() to take an rte_device as parameter, as all eth_dev are meant to be backed by an rte_device. Keeping this in mind, I meant to move writing the pointer closer to the rte_eth_dev_allocate() call.

That is a bigger change, may affect many (if not all) PMDs, so I think this can be considered when that change is available.

And although that change may fix the issues that 'eth_dev->device' is not set, which we had a few times before, not sure it worth to change all PMDs and ethdev API directly couple with rte_device, instead of PMD being the glue. Can be discussed more on its own patch.

Quoted 1 line
Quoted 7 lines
Quoted 4 lines

But you are right that it is needlessly verbose, using vdev_bus->find_device() to do this stuff. I'm ok with changing it as you described if you prefer.

That was the concern, that is too much code to take a value which is already available a few level up the stack.

Ok, future-proofing is a bad habit so let's not do it.

I'm not a fan of the goto for the 'happy path' though. Another simple way would be to bring the vdev pointer with me as we go down the stack.

I will send a v2 shortly, if it's too ugly to move the pointer down I'll use the goto, or if you tell me you'd prefer it.

Signature · 1 line
Gaëtan
#17 lore raw

PATCH v2 1/1 net/ring: fix eth_dev device pointer on allocation

When a net_ring device is allocated, its device pointer is not set before calling rte_eth_dev_probing_finish, which is incorrect.

The following:
commit: 96cb19521147 ("net/ring: use EAL APIs in PMD specific API")
commit: a6992e961050 ("net/ring: set ethernet device field")

already fixed the same issue in 17.08, which was fine at the time. Adding the hook rte_eth_dev_probing_finish() however created this bug, as the eth_dev exposed when this hook is executed is expected to be complete.

Remove the prior attempts to fix the issue in rte_pmd_ring_probe() and write the pointer properly in do_eth_dev_ring_create().

Fixes fbe90cdd776c ("ethdev: add probing finish function")
---
 drivers/net/ring/rte_eth_ring.c | 29 +++++++++++++++--------------
 1 file changed, 15 insertions(+), 14 deletions(-)
drivers/net/ring/rte_eth_ring.c +15 -14
index 41acbc513..f0fafa0c0 100644
@@ -246,6 +246,7 @@ static const struct eth_dev_ops ops = {
 
 static int
 do_eth_dev_ring_create(const char *name,
+		struct rte_vdev_device *vdev,
 		struct rte_ring * const rx_queues[],
 		const unsigned int nb_rx_queues,
 		struct rte_ring *const tx_queues[],
@@ -291,12 +292,15 @@ do_eth_dev_ring_create(const char *name,
 	}
 
 	/* now put it all together
+	 * - store EAL device in eth_dev,
 	 * - store queue data in internals,
 	 * - store numa_node info in eth_dev_data
 	 * - point eth_dev_data to internals
 	 * - and point eth_dev structure to new eth_dev_data structure
 	 */
 
+	eth_dev->device = &vdev->device;
+
 	data = eth_dev->data;
 	data->rx_queues = rx_queues_local;
 	data->tx_queues = tx_queues_local;
@@ -408,7 +412,9 @@ rte_eth_from_ring(struct rte_ring *r)
 }
 
 static int
-eth_dev_ring_create(const char *name, const unsigned int numa_node,
+eth_dev_ring_create(const char *name,
+		struct rte_vdev_device *vdev,
+		const unsigned int numa_node,
 		enum dev_action action, struct rte_eth_dev **eth_dev)
 {
 	/* rx and tx are so-called from point of view of first port.
@@ -438,7 +444,7 @@ eth_dev_ring_create(const char *name, const unsigned int numa_node,
 			return -1;
 	}
 
-	if (do_eth_dev_ring_create(name, rxtx, num_rings, rxtx, num_rings,
+	if (do_eth_dev_ring_create(name, vdev, rxtx, num_rings, rxtx, num_rings,
 		numa_node, action, eth_dev) < 0)
 		return -1;
 
@@ -560,12 +566,12 @@ rte_pmd_ring_probe(struct rte_vdev_device *dev)
 	PMD_LOG(INFO, "Initializing pmd_ring for %s", name);
 
 	if (params == NULL || params[0] == '\0') {
-		ret = eth_dev_ring_create(name, rte_socket_id(), DEV_CREATE,
+		ret = eth_dev_ring_create(name, dev, rte_socket_id(), DEV_CREATE,
 				&eth_dev);
 		if (ret == -1) {
 			PMD_LOG(INFO,
 				"Attach to pmd_ring for %s", name);
-			ret = eth_dev_ring_create(name, rte_socket_id(),
+			ret = eth_dev_ring_create(name, dev, rte_socket_id(),
 						  DEV_ATTACH, &eth_dev);
 		}
 	} else {
@@ -574,19 +580,16 @@ rte_pmd_ring_probe(struct rte_vdev_device *dev)
 		if (!kvlist) {
 			PMD_LOG(INFO,
 				"Ignoring unsupported parameters when creatingrings-backed ethernet device");
-			ret = eth_dev_ring_create(name, rte_socket_id(),
+			ret = eth_dev_ring_create(name, dev, rte_socket_id(),
 						  DEV_CREATE, &eth_dev);
 			if (ret == -1) {
 				PMD_LOG(INFO,
 					"Attach to pmd_ring for %s",
 					name);
-				ret = eth_dev_ring_create(name, rte_socket_id(),
+				ret = eth_dev_ring_create(name, dev, rte_socket_id(),
 							  DEV_ATTACH, &eth_dev);
 			}
 
-			if (eth_dev)
-				eth_dev->device = &dev->device;
-
 			return ret;
 		}
 
@@ -597,7 +600,7 @@ rte_pmd_ring_probe(struct rte_vdev_device *dev)
 			if (ret < 0)
 				goto out_free;
 
-			ret = do_eth_dev_ring_create(name,
+			ret = do_eth_dev_ring_create(name, dev,
 				internal_args->rx_queues,
 				internal_args->nb_rx_queues,
 				internal_args->tx_queues,
@@ -627,6 +630,7 @@ rte_pmd_ring_probe(struct rte_vdev_device *dev)
 
 			for (info->count = 0; info->count < info->total; info->count++) {
 				ret = eth_dev_ring_create(info->list[info->count].name,
+							  dev,
 							  info->list[info->count].node,
 							  info->list[info->count].action,
 							  &eth_dev);
@@ -635,7 +639,7 @@ rte_pmd_ring_probe(struct rte_vdev_device *dev)
 					PMD_LOG(INFO,
 						"Attach to pmd_ring for %s",
 						name);
-					ret = eth_dev_ring_create(name,
+					ret = eth_dev_ring_create(name, dev,
 							info->list[info->count].node,
 							DEV_ATTACH,
 							&eth_dev);
@@ -644,9 +648,6 @@ rte_pmd_ring_probe(struct rte_vdev_device *dev)
 		}
 	}
 
-	if (eth_dev)
-		eth_dev->device = &dev->device;
-
 out_free:
 	rte_kvargs_free(kvlist);
 	rte_free(info);
Signature · 1 line
2.26.2
Reviewed-by on v2 1/1
#18 lore raw

Re: [dpdk-dev] [PATCH v2] net/ring: fix eth_dev device pointer on allocation

Quoted 12 lines · Gaetan Rivet

When a net_ring device is allocated, its device pointer is not set before calling rte_eth_dev_probing_finish, which is incorrect.

The following:
commit: 96cb19521147 ("net/ring: use EAL APIs in PMD specific API")
commit: a6992e961050 ("net/ring: set ethernet device field")

already fixed the same issue in 17.08, which was fine at the time. Adding the hook rte_eth_dev_probing_finish() however created this bug, as the eth_dev exposed when this hook is executed is expected to be complete.

Quoted 6 lines

Remove the prior attempts to fix the issue in rte_pmd_ring_probe() and write the pointer properly in do_eth_dev_ring_create().

Fixes fbe90cdd776c ("ethdev: add probing finish function")

I would prefer moving the assignment up in the stack where 'device' is available, instead of moving the variable down in the stack to assign it, but both does the work ...

Applied
#19 lore raw

Re: [dpdk-dev] [PATCH v2] net/ring: fix eth_dev device pointer on allocation

Quoted 20 lines · Ferruh Yigit
Quoted 12 lines · Gaetan Rivet

When a net_ring device is allocated, its device pointer is not set before calling rte_eth_dev_probing_finish, which is incorrect.

The following:
commit: 96cb19521147 ("net/ring: use EAL APIs in PMD specific API")
commit: a6992e961050 ("net/ring: set ethernet device field")

already fixed the same issue in 17.08, which was fine at the time. Adding the hook rte_eth_dev_probing_finish() however created this bug, as the eth_dev exposed when this hook is executed is expected to be complete.

Quoted 6 lines

Remove the prior attempts to fix the issue in rte_pmd_ring_probe() and write the pointer properly in do_eth_dev_ring_create().

Fixes fbe90cdd776c ("ethdev: add probing finish function")
Quoted 6 lines

I would prefer moving the assignment up in the stack where 'device' is available, instead of moving the variable down in the stack to assign it, but both does the work ...

Applied to dpdk-next-net/master, thanks.

#20 lore raw

PATCH 3/3 net/failsafe: fix default service proxy state

In reply to Gaetan Rivet

The service proxy is initialized at 0. This is assuming that all of its fields are invalid at 0. The issue is that a file descriptor at 0 is a valid one.

The value -1 is used as sentinel during cleanup. Initialize the RX proxy file descriptor to -1.

Fixes 366226dd859f ("net/failsafe: fix fd leak")
Cc Ali Alnubani <[email protected]>
---
 drivers/net/failsafe/failsafe.c         | 1 +
 drivers/net/failsafe/failsafe_private.h | 8 ++++++++
 2 files changed, 9 insertions(+)
drivers/net/failsafe/failsafe.c +1 -0
index 8af31d71b..72362f35d 100644
@@ -190,6 +190,7 @@ fs_eth_dev_create(struct rte_vdev_device *vdev)
 	}
 	priv = PRIV(dev);
 	priv->data = dev->data;
+	priv->rxp = FS_RX_PROXY_INIT;
 	dev->dev_ops = &failsafe_ops;
 	dev->data->mac_addrs = &PRIV(dev)->mac_addrs[0];
 	dev->data->dev_link = eth_link;
drivers/net/failsafe/failsafe_private.h +8 -0
index 8e9706aef..651578a12 100644
@@ -58,6 +58,14 @@ struct rx_proxy {
 	enum rxp_service_state sstate;
 };
 
+#define FS_RX_PROXY_INIT (struct rx_proxy){ \
+	.efd = -1, \
+	.evec = NULL, \
+	.sid = 0, \
+	.scid = 0, \
+	.sstate = SS_NO_SERVICE, \
+}
+
 struct rxq {
 	struct fs_priv *priv;
 	uint16_t qid;
Signature · 1 line
2.26.2
Tested-by on v1 3/3
#21 lore raw

Re: [dpdk-dev] [PATCH v1 3/3] net/failsafe: fix default service proxy state

Thanks Gaetan.

Quoted 10 lines

-----Original Message-----
From: Gaetan Rivet <[email protected]>
Sent: Tuesday, May 5, 2020 10:11 PM
To: [email protected]
Cc: [email protected]; Ali Alnubani <[email protected]>
Subject: [PATCH v1 3/3] net/failsafe: fix default service proxy state

The service proxy is initialized at 0. This is assuming that all of its fields are invalid at 0. The issue is that a file descriptor at 0 is a valid one.

Quoted 7 lines

The value -1 is used as sentinel during cleanup. Initialize the RX proxy file descriptor to -1.

Fixes 366226dd859f ("net/failsafe: fix fd leak")
Cc Ali Alnubani <[email protected]>
#22 lore raw

Re: [dpdk-dev] [PATCH v1 3/3] net/failsafe: fix default service proxy state

In reply to Gaetan Rivet
Quoted 9 lines · Gaetan Rivet

The service proxy is initialized at 0. This is assuming that all of its fields are invalid at 0. The issue is that a file descriptor at 0 is a valid one.

The value -1 is used as sentinel during cleanup. Initialize the RX proxy file descriptor to -1.

Fixes 366226dd859f ("net/failsafe: fix fd leak")

I have squashed this to the commit in the fixes line, to not leave any commit that can cause crash, and kept both sign-offs and Ali's test tag.

Squashed into relevant commit in next-net, thanks.