Re: [dpdk-dev] [PATCH] eal: disable IOVA mode detection by default

[Ferruh Yigit <ferruh.yigit@intel.com>]

On 10/31/2017 11:54 PM, Jerin Jacob wrote:
> -----Original Message-----
>> Date: Tue, 31 Oct 2017 23:21:18 -0700
>> From: Ferruh Yigit <ferruh.yigit@intel.com>
>> To: Jerin Jacob <jerin.jacob@caviumnetworks.com>
>> CC: Thomas Monjalon <thomas@monjalon.net>, Bruce Richardson
>>  <bruce.richardson@intel.com>, Sergio Gonzalez Monroy
>>  <sergio.gonzalez.monroy@intel.com>, dev@dpdk.org, Jianfeng Tan
>>  <jianfeng.tan@intel.com>, Santosh Shukla
>>  <santosh.shukla@caviumnetworks.com>
>> Subject: Re: [dpdk-dev] [PATCH] eal: disable IOVA mode detection by default
>> User-Agent: Mozilla/5.0 (Windows NT 10.0; WOW64; rv:52.0) Gecko/20100101
>>  Thunderbird/52.4.0
>>
>> On 10/31/2017 9:02 PM, Jerin Jacob wrote:
>>> -----Original Message-----
>>>> Date: Wed, 1 Nov 2017 01:07:26 +0000
>>>> From: Ferruh Yigit <ferruh.yigit@intel.com>
>>>> To: Thomas Monjalon <thomas@monjalon.net>, Bruce Richardson
>>>>  <bruce.richardson@intel.com>, Sergio Gonzalez Monroy
>>>>  <sergio.gonzalez.monroy@intel.com>
>>>> CC: dev@dpdk.org, Ferruh Yigit <ferruh.yigit@intel.com>, Jianfeng Tan
>>>>  <jianfeng.tan@intel.com>, Santosh Shukla
>>>>  <santosh.shukla@caviumnetworks.com>
>>>> Subject: [dpdk-dev] [PATCH] eal: disable IOVA mode detection by default
>>>> X-Mailer: git-send-email 2.13.6
>>>>
>>>> Fix kernel crash with KNI because KNI requires physical addresses.
>>>
>>> The actual fix would be to make KNI IOMMU aware based on the DPDK mode.
>>>
>>> ie. On slow path,
>>>
>>> /* Get iommu domain for iova to physical addr conversion */             
>>> if (rte_eal_iova_mode() == RTE_IOVA_VA)
>>> 	kni->iommu_domain = iommu_get_domain_for_dev(dev);
>>> else
>>> 	kni->iommu_domain = NULL;
>>>
>>> On fast path,
>>>
>>> static inline u64 kni_iova_to_phys(struct ... *kni, dma_addr_t dma_addr)    
>>> {                                                                               
>>>         /* Translation is installed only when IOMMU is present */               
>>>         if (kni->iommu_domain)                                                  
>>>                 return iommu_iova_to_phys(kni->iommu_domain, dma_addr);         
>>>         return dma_addr;                                                        
>>> } 
>>>
>>>>
>>>> A config option introduced to disable IOVA mode detection and to set it
>>>> to physical address by default. Disabling config option will enable IOVA
>>>> mode detection.
>>>>
>>>> When there is no intension to use KNI, it is safe to enable detection.
>>>>
>>>> Config option disable IOVA mode detection by default to be sure only who
>>>> is aware of result enable it.
>>>>
>>>> Fixes: 72d013644bd6 ("mem: honor IOVA mode in malloc virt2phy")
>>>>
>>>> Signed-off-by: Ferruh Yigit <ferruh.yigit@intel.com>
>>>> ---
>>>> Cc: Jianfeng Tan <jianfeng.tan@intel.com>
>>>> Cc: Santosh Shukla <santosh.shukla@caviumnetworks.com>
>>>> Cc: Thomas Monjalon <thomas@monjalon.net>
>>>> ---
>>>>  config/common_base                | 5 +++++
>>>>  lib/librte_eal/bsdapp/eal/eal.c   | 4 ++++
>>>>  lib/librte_eal/linuxapp/eal/eal.c | 4 ++++
>>>>  3 files changed, 13 insertions(+)
>>>>
>>>> diff --git a/config/common_base b/config/common_base
>>>> index 82ee75456..903e7685b 100644
>>>> --- a/config/common_base
>>>> +++ b/config/common_base
>>>> @@ -107,6 +107,11 @@ CONFIG_RTE_MALLOC_DEBUG=n
>>>>  CONFIG_RTE_EAL_NUMA_AWARE_HUGEPAGES=n
>>>>  
>>>>  #
>>>> +# Disabling PHYS_IOVA may crash kernel for KNI, use with caution
>>>> +#
>>>> +CONFIG_RTE_EAL_USE_PHYS_IOVA=y
>>>
>>> Defeat the purpose of all dynamic probing scheme.
>>> Either we can fix the KNI or revert the following patch for this release.
>>>
>>> http://dpdk.org/commit/f37dfab2
>>
>> This commit just enables IOVA VA mode for Intel drivers, that is how I can able
>> to observe the issue, but it is not the source of the problem. Reverting that
>> commit will not solve KNI crash with any other PMD that enables IOVA VA mode.
> 
> I don't understand why a PMD needs to enable IOVA_VA if it can support IOVA_PA.
> IMO, IOVA_VA should be enabled only for those device it can WORK ONLY on
> IOVA_VA mode. Forget about KNI, If we set CONFIG_RTE_EAL_USE_PHYS_IOVA
> as y then the normal stuff wont work for if PMD can operate only in
> IOVA_VA mode(like octeontx).

Hmm, I wasn't aware that octeontx only operate on IOVA VA mode, I got your
concern now.

But still that config option still enables choosing between KNI and "IOVA_VA
mode only" devices until KNI gets fixed. Unfortunately both won't work at same
time for now.

> Regarding the KNI crash, it can be avoid by first checking the exiting
> mode(rte_eal_iova_mode()). i.e since legacy driver like KNI need real
> physical address to work "now", it can grace full exit on the init time if
> mode == IOVA_VA;

Definitely agree to gracefully exit, but I was looking to make KNI work more
than just exit without crash.
>> Related to the KNI, iommu is not involved at all, I am not clear with your above
>> suggestion, physical address is required for kernelspace - userspace communication.
> 
> vhost-kernel addressed this case with IOVA as VA. I need to spend cycles
> on what takes to remove physical address dependency from KNI.
> But someone can add that support if it is required in future.

Since the iova patchset breaks the existing KNI, I would like to see KNI fix
part of the iova work.
It doesn't look right to me that a new feature breaking the existing one.

>> KNI uses physical address for two things:
>>
>> 1- Creates a buffer in userspace, a memzone, and shares its physical address to
>> the kernel, so that both kernel and app can access same buffer.
>>
>> The question is, even all devices supports IOVA VA mode, why creating a memzone
>> that has real physical address info is not possible anymore?
>> For KNI case this is not related at all with what devices supports.
>>
>> 2- For each mbuf that will be sent to kernel, dpdk app puts physical address of
>> that mbuf into shared buffer, so that kernel can access it. So that kernel can
>> access to mbuf data that can be coming from any mempool.
>> For this the physical address of the mbuf is required.
>>
>>
>> Overall KNI needs to know physical address of mbufs and memzone.
>>
>> This patch provides a way to have old behavior with a config option, so that KNI
>> can work and anyone who needs to create memzone with physical address can a way
>> to have it, and all PMDs will work fine.
>>
>> If there is no need for these, it is possible to disable config and dynamic
>> probing scheme is already there.
>>
>> Thanks,
>> ferruh