DPDK patches and discussions
 help / color / mirror / Atom feed
From: fengchengwen <fengchengwen@huawei.com>
To: Konstantin Ananyev <konstantin.v.ananyev@yandex.ru>,
	<dev@dpdk.org>,
	Konstantin Ananyev <konstantin.ananyev@huawei.com>,
	Honnappa Nagarahalli <Honnappa.Nagarahalli@arm.com>,
	Stephen Hemminger <stephen@networkplumber.org>,
	Ruifeng Wang <Ruifeng.Wang@arm.com>,
	"Ajit Khaparde (ajit.khaparde@broadcom.com)"
	<ajit.khaparde@broadcom.com>
Subject: Re: [PATCH 2/2] ethdev: fix race condition in fast-path ops setup
Date: Mon, 27 Feb 2023 10:56:35 +0800	[thread overview]
Message-ID: <6618689f-373c-354d-8424-3186792bf68d@huawei.com> (raw)
In-Reply-To: <3dc3b3d3-c80f-a361-8780-b1b3e48d843e@yandex.ru>

On 2023/2/27 1:22, Konstantin Ananyev wrote:
> 
>>>>>>>>>>> If ethdev enqueue or dequeue function is called during
>>>>>>>>>>> eth_dev_fp_ops_setup(), it may get pre-empted after setting the
>>>>>>>>>>> function pointers, but before setting the pointer to port data.
>>>>>>>>>>> In this case the newly registered enqueue/dequeue function will
>>>>>>>>>>> use dummy port data and end up in seg fault.
>>>>>>>>>>>
>>>>>>>>>>> This patch moves the updation of each data pointers before
>>>>>>>>>>> updating corresponding function pointers.
>>>>>>>>>>>
>>>>>>>>>>> Fixes: c87d435a4d79 ("ethdev: copy fast-path API into separate
>>>>>>>>>>> structure")
>>>>>>>>>>> Cc: stable@dpdk.org
>>>>>>>>
>>>>>>>> Why is something calling enqueue/dequeue when device is not fully
>>>>>> started.
>>>>>>>> A correctly written application would not call rx/tx burst until
>>>>>>>> after ethdev start had finished.
>>>>>>>
>>>>>>> Please refer the eb0d471a894 (ethdev: add proactive error handling
>>>>>>> mode), when driver recover itself, the application may still invoke
>>>>>> enqueue/dequeue API.
>>>>>>
>>>>>> Right now DPDK ethdev layer *does not* provide synchronization
>>>>>> mechanisms between data-path and control-path functions.
>>>>>> That was a deliberate deisgn choice. If we want to change that rule, then I
>>>>>> suppose we need a community consensus for it.
>>>>>> I think that if the driver wants to provide some sort of error recovery
>>>>>> procedure, then it has to provide some synchronization mechanism inside it
>>>>>> between data-path and control-path functions.
>>>>>> Actually looking at eb0d471a894 (ethdev: add proactive error handling
>>>>>> mode), and following patches I wonder how it creeped in?
>>>>>> It seems we just introduced a loophole for race condition with this
>>>>>> approach...
>>>>
>>>> Could you try to describe the specific scenario of loophole ?
>>>
>>> Ok, as I understand the existing mechanism:
>>>
>>> When PMD wants to start a recovery it has to:
>>>   - invoke  rte_eth_dev_callback_process(RTE_ETH_EVENT_ERR_RECOVERING);
>>>     That supposed to call user provided callback. After callback is finished PMD assumes
>>>     that user is aware that recovery is about to start and should make some precautions.
>>> - when recovery is finished it invokes another callback:
>>>    RTE_ETH_EVENT_RECOVERY_(SUCCESS/FAILED). After that user either can continue to
>>>    use port or have to treat is as faulty.
>>>
>>> The idea is ok in principle, but there is a problem.
>>>
>>> lib/ethdev/rte_ethdev.h:
>>>             /** Port recovering from a hardware or firmware error.
>>>           * If PMD supports proactive error recovery,
>>>           * it should trigger this event to notify application
>>>           * that it detected an error and the recovery is being started.
>>>
>>> <<< !!!!!
>>>           * Upon receiving the event, the application should not invoke any control path API
>>>           * (such as rte_eth_dev_configure/rte_eth_dev_stop...) until receiving
>>>           * RTE_ETH_EVENT_RECOVERY_SUCCESS or RTE_ETH_EVENT_RECOVERY_FAILED event.
>>>           * The PMD will set the data path pointers to dummy functions,
>>>           * and re-set the data path pointers to non-dummy functions
>>>           * before reporting RTE_ETH_EVENT_RECOVERY_SUCCESS event.
>>> <<< !!!!!
>>>
>>> That part is just wrong I believe.
>>> It should be:
>>> Upon receiving the event, the application should not invoke any *both control and data-path* API
>>> until receiving  RTE_ETH_EVENT_RECOVERY_SUCCESS or RTE_ETH_EVENT_RECOVERY_FAILED event.
>>> Resetting data path pointers to dummy functions by PMD *before* invoking
>>> rte_eth_dev_callback_process(RTE_ETH_EVENT_ERR_RECOVERING);
>>> introduces a race-condition with data-path threads, as such thread could already be inside RX/TX function
>>> or can already read RX/TX function/data pointers and be about to use them.
>>
>> Current practices: the PMDs already add some delay after set Rx/Tx callback to dummy, and plus the DPDK
>> worker thread is busypolling, the probability of occurence in reality is zero. But in theoretically exist
>> the above race-condition.
> 
> 
> Adding delay might make a problem a bit less reproducible,
> but it doesn't fix it.
> The bug is still there.
> 
> 
>>
>>> And right now rte_ethdev layer doesn't provide any mechanism to check it or wait when they'll finish, etc.
>>
>> Yes
>>
>>>
>>> So, probably the simplest way to fix it with existing DPDK design:
>>> - user level callback  RTE_ETH_EVENT_ERR_RECOVERING should return only after it ensures that *all*
>>>    application threads (and processes) stopped using either control or data-path functions for that port
>>
>> Agree
>>
>>>    (yes it means that application that wants to use this feature has to provide its own synchronization mechanism
>>>    around data-path functions (RX/TX) that it is going to use).
>>> - after that PMD is safe to reset rte_eth_fp_ops[] values to dummy ones.
>>>
>>> And message to all PMD developers:
>>> *please stop updating rte_eth_fp_ops[] on your own*.
>>> That's a bad practice and it is not supposed to do things that way.
>>> There is a special API provided for these purposes:
>>> eth_dev_fp_ops_reset(), eth_dev_fp_ops_setup(), so use it.
>>
>> This two function is in private.h, so it should be expose to public header file.
> 
> You mean we need to move these functions declarations into ethdev_driver.h?
> If so, then yes, I think we probably do.
> 
> 
>>>
>>> BTW,  I don't see any implementation for RTE_ETH_EVENT_ERR_RECOVERING within
>>> either testpmd or any other example apps.
>>> Am I missing something?
>>
>> Currently it just promote the event.
> 
> 
> Ok, can I suggest then to add a proper usage for into in testpmd?

our team will do that, thanks.

> It looks really strange that we add new feature into ethdev (and 2 PMDs),
> but didn't provide any way for users to test it.
> 
>>
>>> If not, then probably it could be a good starting point - let's incorporate it inside testpmd
>>> (new forwarding engine probably) so everyone can test/try it.
>>>
>>>           * It means that the application cannot send or receive any packets
>>>           * during this period.
>>>           * @note Before the PMD reports the recovery result,
>>>           * the PMD may report the RTE_ETH_EVENT_ERR_RECOVERING event again,
>>>           * because a larger error may occur during the recovery.
>>>           */
>>>          RTE_ETH_EVENT_ERR_RECOVERING,
>>>
>>>>>> It probably needs to be either deprecated or reworked.
>>>>> Looking at the commit, it does not say anything about the data plane functions which probably means, the error recovery is
>>>> happening within the data plane thread. What happens to other data plane threads that are polling the same port on which the error
>>>> recovery is happening?
>>>>
>>>> The commit log says: "the PMD sets the data path pointers to dummy functions".
>>>>
>>>> So the data plane threads will receive non-packet and send zero with port which in error recovery.
>>>>
>>>>>
>>>>> Also, the commit log says that while the error recovery is under progress, the application should not call any control plane APIs. Does
>>>> that mean, the application has to check for error condition every time it calls a control plane API?
>>>>
>>>> If application has not register event (RTE_ETH_EVENT_ERR_RECOVERING) callback, it could calls control plane API, but it will return
>>>> failed.
>>>> If application has register above callback, it can wait for recovery result, or direct call without wait but this will return failed.
>>>>
>>>>>
>>>>> The commit message also says that "PMD makes sure the control path operations failed with retcode -EBUSY". It does not say how it
>>>> does this. But, any communication from the PMD thread to control plane thread may introduce race conditions if not done correctly.
>>>>
>>>> First there are no PMD thread, do you mean eal-intr-thread ?
>>>>
>>>> As for this question, you can see PMDs which already implement it, they both provides mutual exclusion protection.
>>>>
>>>>>
>>>>>>
>>>>>>>
>>>>>>>>
>>>>>>>> Would something like this work better?
>>>>>>>>
>>>>>>>> Note: there is another bug in current code. The check for link state
>>>>>>>> interrupt and link_ops could return -ENOTSUP and leave device in
>>>>>> indeterminate state.
>>>>>>>> The check should be done before calling PMD.
>>>>>>>>
>>>>>>>> diff --git a/lib/ethdev/rte_ethdev.c b/lib/ethdev/rte_ethdev.c index
>>>>>>>> 0266cc82acb6..d6c163ed85e7 100644
>>>>>>>> --- a/lib/ethdev/rte_ethdev.c
>>>>>>>> +++ b/lib/ethdev/rte_ethdev.c
>>>>>>>> @@ -1582,6 +1582,14 @@ rte_eth_dev_start(uint16_t port_id)
>>>>>>>>           return 0;
>>>>>>>>       }
>>>>>>>>
>>>>>>>> +    if (dev->data->dev_conf.intr_conf.lsc == 0 &&
>>>>>>>> +        dev->dev_ops->link_update == NULL) {
>>>>>>>> +        RTE_ETHDEV_LOG(INFO,
>>>>>>>> +                   "Device with port_id=%"PRIu16" link update not
>>>>>> supported\n",
>>>>>>>> +                   port_id);
>>>>>>>> +            return -ENOTSUP;
>>>>>>>> +    }
>>>>>>>> +
>>>>>>>>       ret = rte_eth_dev_info_get(port_id, &dev_info);
>>>>>>>>       if (ret != 0)
>>>>>>>>           return ret;
>>>>>>>> @@ -1591,9 +1599,7 @@ rte_eth_dev_start(uint16_t port_id)
>>>>>>>>           eth_dev_mac_restore(dev, &dev_info);
>>>>>>>>
>>>>>>>>       diag = (*dev->dev_ops->dev_start)(dev);
>>>>>>>> -    if (diag == 0)
>>>>>>>> -        dev->data->dev_started = 1;
>>>>>>>> -    else
>>>>>>>> +    if (diag != 0)
>>>>>>>>           return eth_err(port_id, diag);
>>>>>>>>
>>>>>>>>       ret = eth_dev_config_restore(dev, &dev_info, port_id); @@ -1611,16
>>>>>>>> +1617,18 @@ rte_eth_dev_start(uint16_t port_id)
>>>>>>>>           return ret;
>>>>>>>>       }
>>>>>>>>
>>>>>>>> -    if (dev->data->dev_conf.intr_conf.lsc == 0) {
>>>>>>>> -        if (*dev->dev_ops->link_update == NULL)
>>>>>>>> -            return -ENOTSUP;
>>>>>>>> -        (*dev->dev_ops->link_update)(dev, 0);
>>>>>>>> -    }
>>>>>>>> -
>>>>>>>>       /* expose selection of PMD fast-path functions */
>>>>>>>>       eth_dev_fp_ops_setup(rte_eth_fp_ops + port_id, dev);
>>>>>>>>
>>>>>>>> +    /* ensure state is set before marking device ready */
>>>>>>>> +    rte_smp_wmb();
>>>>>>>> +
>>>>>>>>       rte_ethdev_trace_start(port_id);
>>>>>>>> +
>>>>>>>> +    /* Update current link state */
>>>>>>>> +    if (dev->data->dev_conf.intr_conf.lsc == 0)
>>>>>>>> +        (*dev->dev_ops->link_update)(dev, 0);
>>>>>>>> +
>>>>>>>>       return 0;
>>>>>>>>   }
>>>>>>>>
>>>>>>>>
>>>>>>>> .
>>>>>>>>
>>>>>
> 
> .

  reply	other threads:[~2023-02-27  2:56 UTC|newest]

Thread overview: 29+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2023-02-20  6:08 [PATCH 1/2] eventdev: fix race condition in fast-path set function Ashok Kaladi
2023-02-20  6:08 ` [PATCH 2/2] ethdev: fix race condition in fast-path ops setup Ashok Kaladi
2023-02-20  6:57   ` fengchengwen
2023-02-21  7:24     ` Ruifeng Wang
2023-02-21 17:00       ` Stephen Hemminger
2023-02-22  1:07         ` fengchengwen
2023-02-22  9:41           ` Ruifeng Wang
2023-02-22 10:41           ` Konstantin Ananyev
2023-02-22 22:48             ` Honnappa Nagarahalli
2023-02-23  1:15               ` Stephen Hemminger
2023-02-23  4:47                 ` Honnappa Nagarahalli
2023-02-23  4:40             ` Honnappa Nagarahalli
2023-02-23  8:23               ` fengchengwen
2023-02-23 13:31                 ` Konstantin Ananyev
2023-02-25  1:32                   ` fengchengwen
2023-02-26 17:22                     ` Konstantin Ananyev
2023-02-27  2:56                       ` fengchengwen [this message]
2023-02-27 19:08                         ` Konstantin Ananyev
2023-03-03 17:19                       ` Ferruh Yigit
2023-03-06  1:57                         ` fengchengwen
2023-03-06  6:13                         ` Ruifeng Wang
2023-03-06 10:32                           ` Konstantin Ananyev
2023-03-06 11:17                             ` Ajit Khaparde
2023-03-06 11:57                             ` Ferruh Yigit
2023-03-06 12:36                               ` Konstantin Ananyev
2023-02-28 23:57                   ` Honnappa Nagarahalli
2023-02-20  7:01   ` fengchengwen
2023-02-20  9:44   ` Konstantin Ananyev
2023-03-03 16:49   ` Ferruh Yigit

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=6618689f-373c-354d-8424-3186792bf68d@huawei.com \
    --to=fengchengwen@huawei.com \
    --cc=Honnappa.Nagarahalli@arm.com \
    --cc=Ruifeng.Wang@arm.com \
    --cc=ajit.khaparde@broadcom.com \
    --cc=dev@dpdk.org \
    --cc=konstantin.ananyev@huawei.com \
    --cc=konstantin.v.ananyev@yandex.ru \
    --cc=stephen@networkplumber.org \
    /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).