On 9/29/2020 5:02 PM, Thomas Monjalon wrote:
29/09/2020 17:50, Ferruh Yigit:
On 9/29/2020 12:58 PM, Wang, Haiyue wrote:
From: Thomas Monjalon <tho...@monjalon.net>
29/09/2020 12:26, Maxime Coquelin:
On 9/29/20 1:14 AM, Thomas Monjalon wrote:
The function rte_eth_dev_release_port() was resetting partially
the struct rte_eth_dev. The drivers were completing it
with more pointers set to NULL in the close or remove operations.

A full memset is done so most of those assignments become useless.
[...]
With this patch, I get following segfault at init time with Virtio PMD:

          Program received signal SIGSEGV, Segmentation fault.
          0x0000000000854c9b in rte_eth_dev_callback_register (port_id=32,
              event=RTE_ETH_EVENT_UNKNOWN, cb_fn=0x4b24de
<eth_event_callback>,
              cb_arg=0x0) at ../lib/librte_ethdev/rte_ethdev.c:4042
          4042
TAILQ_INSERT_TAIL(&(dev->link_intr_cbs),

Yes this is because after closing a port, everything is resetted,
including .link_intr_cbs which is set only once in a constructor:
http://git.dpdk.org/dpdk/commit/?id=9ec0b3869d8

I can change this patch to selectively set pointers to NULL.

Or if we prefer a big memset 0, we need to rework how RTE_ETH_ALL
is managed to register a callback for any event.
Instead of setting the callback for all ports, we could have
a special catch-call callback list which is called for all events.
This way we could revert initializing .link_intr_cbs in eth_dev_get().



Move 'struct rte_eth_dev_cb_list link_intr_cbs' to the end of 'struct 
rte_eth_dev',
and starting from link_intr_cbs, these members will be kept after closed ? :-)

memset(eth_dev, 0, offsetof(struct rte_eth_dev, link_intr_cbs));


This is similar version of a previous patch:
https://patches.dpdk.org/patch/65795/

I forgot this patch :)

That one is also waiting because I remember it was not safe for hotplug.

I don't think there is a safety issue.
It makes harder to play with dangling pointers, which is good.
Note: the .device pointer was already resetted by
rte_eth_dev_pci_generic_remove(), and generalized in patch 1.


As far as I remember, resetting '.device' in the 'close()' was causing some issues on hot remove. Yes patch one already does it, I will test the hot remove after close with it.

Instead of a big memset, I am for selectively set pointers to NULL.

I will prepare the patch with selective resets.



Reply via email to