Re: [PATCH net 3/3] net/smc: Resolve the race between SMC-R link access and clear

From: Karsten Graul
Date: Tue Jan 11 2022 - 11:03:02 EST


On 11/01/2022 16:49, Wen Gu wrote:
> Thanks for your review.
>
> On 2022/1/11 4:40 pm, Karsten Graul wrote:
>> On 10/01/2022 10:26, Wen Gu wrote:
>>> @@ -1226,15 +1245,23 @@ void smcr_link_clear(struct smc_link *lnk, bool log)
>>>       smc_wr_free_link(lnk);
>>>       smc_ib_destroy_queue_pair(lnk);
>>>       smc_ib_dealloc_protection_domain(lnk);
>>> -    smc_wr_free_link_mem(lnk);
>>> -    smc_lgr_put(lnk->lgr); /* lgr_hold in smcr_link_init() */
>>>       smc_ibdev_cnt_dec(lnk);
>>>       put_device(&lnk->smcibdev->ibdev->dev);
>>>       smcibdev = lnk->smcibdev;
>>> -    memset(lnk, 0, sizeof(struct smc_link));
>>> -    lnk->state = SMC_LNK_UNUSED;
>>>       if (!atomic_dec_return(&smcibdev->lnk_cnt))
>>>           wake_up(&smcibdev->lnks_deleted);
>>
>> Same here, waiter should not be woken up until the link memory is actually freed.
>>
>
> OK, I will correct this as well.
>
> And similarly I want to move smc_ibdev_cnt_dec() and put_device() to
> __smcr_link_clear() as well to ensure that put link related resources
> only when link is actually cleared. What do you think?

I think that's a good idea, yes.