Re: [dpdk-dev] [dpdk-stable] [PATCH v2] bus/pci: fix driver detach clear

[David Marchand <david.marchand@redhat.com>]

On Wed, Nov 20, 2019 at 10:48 AM Matan Azrad <matan@mellanox.com> wrote:
>
> When a rte_device is unplugged, the driver should be detached from the
> device.
>
> The PCI detach driver operation wrongly didn't clear the driver from the
> device structure what remain the device in probe state from the EAL
> point of view.
>
> For example, when a device is removed twice using rte_dev_remove, it
> cause a crash in EAL.

I can see a crash when using port detach in testpmd with a virtio pci device.

testpmd> port attach 0000:07:00.0
Attaching a new port...
EAL: PCI device 0000:07:00.0 on NUMA socket -1
EAL:   Invalid NUMA socket, default to 0
EAL:   probe driver: 1af4:1041 net_virtio
Port 1 is attached. Now total ports is 2
Done
testpmd> port close 1
Closing ports...
EAL: Releasing pci mapped resource for 0000:07:00.0
EAL: Calling pci_unmap_resource for 0000:07:00.0 at 0x2200006000
Done
testpmd> port detach 1
Removing a device...

Breakpoint 1, local_dev_remove (dev=0x1de64b0) at
/root/dpdk/lib/librte_eal/common/eal_common_dev.c:315
315        if (dev->bus->unplug == NULL) {
Missing separate debuginfos, use: debuginfo-install
glibc-2.17-292.el7.x86_64 libgcc-4.8.5-39.el7.x86_64
libpcap-1.5.3-11.el7.x86_64 numactl-libs-2.0.12-3.el7.x86_64
(gdb) p *dev
$1 = {next = {tqe_next = 0x0, tqe_prev = 0x0}, name = 0x1cf8078
"0000:07:00.0", driver = 0x16c68f0 <rte_virtio_pmd+16>, bus =
0x16b2640 <rte_pci_bus>, numa_node = 0, devargs = 0x1cf8060}
(gdb) c
Continuing.
Device of port 1 is detached
Now total ports is 1
Done


On the first detach, the pci bus frees the rte_pci_device which embeds
the rte_device object.

static int
pci_unplug(struct rte_device *dev)
{
        struct rte_pci_device *pdev;
        int ret;

        pdev = RTE_DEV_TO_PCI(dev);
        ret = rte_pci_detach_dev(pdev);
        if (ret == 0) {
                rte_pci_remove_device(pdev);
                rte_devargs_remove(dev->devargs);
                free(pdev);
        }
        return ret;
}



testpmd> port detach 1
Removing a device...

Breakpoint 1, local_dev_remove (dev=0x1de64b0) at
/root/dpdk/lib/librte_eal/common/eal_common_dev.c:315
315        if (dev->bus->unplug == NULL) {
(gdb) p *dev
$2 = {next = {tqe_next = 0x0, tqe_prev = 0x0}, name = 0xa <Address 0xa
out of bounds>, driver = 0x0, bus = 0x4637, numa_node = 1, devargs =
0x40000002e040018}
(gdb) c
Continuing.

Program received signal SIGSEGV, Segmentation fault.
0x00000000007c1ddd in local_dev_remove (dev=0x1de64b0) at
/root/dpdk/lib/librte_eal/common/eal_common_dev.c:315
315        if (dev->bus->unplug == NULL) {


On the second detach, testpmd passes the same rte_device pointer it
extracts from rte_eth_devices, but the malloc'd location has been
reused (with watchpoint on the location, I found somewhere around
rte_mp_request_sync/opendir()), and then *crunch* on dev->bus.


From my pov:
- testpmd is wrongly reusing a pointer coming from rte_eth_devices[],
without caring about the port state (this is what your second patch
fixes),
- testpmd is directly kicking pointers in rte_eth_devices[] (setting
->device = NULL for its own logic), which is bad too,
- this patch just hides the reuse of a freed pointer,


-- 
David Marchand