From: Bruce Richardson <bruce.richardson@intel.com>
To: Ferruh Yigit <ferruh.yigit@intel.com>
Cc: Igor Ryzhov <iryzhov@nfware.com>,
dev@dpdk.org, David Marchand <david.marchand@6wind.com>,
"Liu, Yuanhan" <yuanhan.liu@intel.com>
Subject: Re: [dpdk-dev] rte_eth_dev_attach returns 0, although device is not attached
Date: Fri, 5 Aug 2016 13:29:42 +0100 [thread overview]
Message-ID: <20160805122942.GA33788@bricha3-MOBL3> (raw)
In-Reply-To: <57A3638D.2060300@intel.com>
On Thu, Aug 04, 2016 at 04:47:25PM +0100, Ferruh Yigit wrote:
> On 8/4/2016 3:54 PM, Igor Ryzhov wrote:
> >
> >> 4 авг. 2016 г., в 16:21, Ferruh Yigit <ferruh.yigit@intel.com
> >> <mailto:ferruh.yigit@intel.com>> написал(а):
> >>
> >> On 8/4/2016 12:51 PM, Igor Ryzhov wrote:
> >>> Hello Ferruh,
> >>>
> >>>> 4 авг. 2016 г., в 14:33, Ferruh Yigit <ferruh.yigit@intel.com
> >>>> <mailto:ferruh.yigit@intel.com>> написал(а):
> >>>>
> >>>> Hi Igor,
> >>>>
> >>>> On 8/3/2016 5:58 PM, Igor Ryzhov wrote:
> >>>>> Hello.
> >>>>>
> >>>>> Function rte_eth_dev_attach can return false positive result.
> >>>>> It happens because rte_eal_pci_probe_one returns zero if no driver
> >>>>> is found for the device:
> >>>>> ret = pci_probe_all_drivers(dev);
> >>>>> if (ret < 0)
> >>>>> goto err_return;
> >>>>> return 0;
> >>>>> (pci_probe_all_drivers returns 1 in that case)
> >>>>>
> >>>>> For example, it can be easily reproduced by trying to attach virtio
> >>>>> device, managed by kernel driver.
> >>>>
> >>>> You are right, and I did able to reproduce this issue with virtio as you
> >>>> suggest.
> >>>>
> >>>> But I wonder why rte_eth_dev_get_port_by_addr() is not catching this.
> >>>> Perhaps a dev->attached check needs to be added into this function.
> >>
> >> With a second check, rte_eth_dev_get_port_by_addr() catches it if the
> >> driver is missing.
> >>
> >> But for virtio case, problem is not missing driver.
> >> Problem is eth_virtio_dev_init() is returning a positive value on fail.
> >>
> >> Call stack is:
> >> rte_eal_pci_probe_one
> >> pci_probe_all_drivers
> >> rte_eal_pci_probe_one_driver
> >> rte_eth_dev_init
> >> eth_virtio_dev_init
> >>
> >> So rte_eal_pci_probe_one_driver() also returns positive value, as no
> >> driver found, and rte_eth_dev_get_port_by_addr() returns a valid
> >> port_id, since rte_eth_dev_init() allocated an eth_dev.
> >>
> >> Briefly, this can be fixed in virtio pmd, instead of eal pci.
> >>
> >>>>
> >>>>>
> >>>>> I think it should be:
> >>>>> ret = pci_probe_all_drivers(dev);
> >>>>> if (ret)
> >>>>> goto err_return;
> >>>>> return 0;
> >>>>
> >>>> Your proposal looks good to me. Will you send a patch?
> >>>
> >>
> >> Original code silently ignores the if driver is missing for that dev,
> >> although it is still questionable, I think we can keep this as it is.
> >>
> >>> Patch sent.
> >>
> >> Sorry for this, but can you please test with following modification in
> >> virtio:
> >> index 07d6449..c74eeee 100644
> >> --- a/drivers/net/virtio/virtio_ethdev.c
> >> +++ b/drivers/net/virtio/virtio_ethdev.c
> >> @@ -1156,7 +1156,7 @@ eth_virtio_dev_init(struct rte_eth_dev *eth_dev)
> >> if (pci_dev) {
> >> ret = vtpci_init(pci_dev, hw, &dev_flags);
> >> if (ret)
> >> - return ret;
> >> + return -1;
> >> }
> >>
> >> /* Reset the device although not necessary at startup */
> >
> > I think it's not a good change, because it will break the idea of this
> > patch - http://dpdk.org/browse/dpdk/commit/?id=ac5e1d83
>
> Yes, breaks this one, I wasn't aware of this patch. But in this patch,
> commit log says: "return 1 to tell the upper layer we
> don't take over this device.", I am not sure upper layer designed for this.
>
> >
> > Also, with your patch the application will not start, because
> > rte_eal_pci_probe will fail:
> >
> > if (ret < 0)
> > rte_exit(EXIT_FAILURE, "Requested device " PCI_PRI_FMT
> > " cannot be used\n", dev->addr.domain, dev->addr.bus,
> > dev->addr.devid, dev->addr.function);
>
> Yes it fails, and this looks like intended behavior. This failure is
> correct according code.
>
> >
> > And now I think that maybe we should change the way rte_eal_pci_probe works.
> > I think we shouldn't stop the application if just one of PCI devices is
> > not probed successfully.
>
> Agreed. Overall rte_exit() usage already discussed a few times.
>
> I think best option is:
> - don't exit app if rte_eal_pci_probe() fails, only print an error.
Whether or not the pci probe exits the app or not, I think it should signal
a serious error if the probe fails and a device was explicitly whitelisted on
the commandline. Given the user explicitly requested the device, a failure to
use is probably a problem which requires the user to fix before running the
app.
/Bruce
prev parent reply other threads:[~2016-08-05 12:29 UTC|newest]
Thread overview: 7+ messages / expand[flat|nested] mbox.gz Atom feed top
2016-08-03 16:58 Igor Ryzhov
2016-08-04 11:33 ` Ferruh Yigit
2016-08-04 11:51 ` Igor Ryzhov
2016-08-04 13:21 ` Ferruh Yigit
2016-08-04 14:54 ` Igor Ryzhov
2016-08-04 15:47 ` Ferruh Yigit
2016-08-05 12:29 ` Bruce Richardson [this message]
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20160805122942.GA33788@bricha3-MOBL3 \
--to=bruce.richardson@intel.com \
--cc=david.marchand@6wind.com \
--cc=dev@dpdk.org \
--cc=ferruh.yigit@intel.com \
--cc=iryzhov@nfware.com \
--cc=yuanhan.liu@intel.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).