On 11/1/2017 2:54 PM, Jerin Jacob wrote:
-----Original Message-----
Date: Tue, 31 Oct 2017 23:21:18 -0700
From: Ferruh Yigit <ferruh.yi...@intel.com>
To: Jerin Jacob <jerin.ja...@caviumnetworks.com>
CC: Thomas Monjalon <tho...@monjalon.net>, Bruce Richardson
<bruce.richard...@intel.com>, Sergio Gonzalez Monroy
<sergio.gonzalez.mon...@intel.com>, dev@dpdk.org, Jianfeng Tan
<jianfeng....@intel.com>, Santosh Shukla
<santosh.shu...@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.yi...@intel.com>
To: Thomas Monjalon <tho...@monjalon.net>, Bruce Richardson
<bruce.richard...@intel.com>, Sergio Gonzalez Monroy
<sergio.gonzalez.mon...@intel.com>
CC: dev@dpdk.org, Ferruh Yigit <ferruh.yi...@intel.com>, Jianfeng Tan
<jianfeng....@intel.com>, Santosh Shukla
<santosh.shu...@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.yi...@intel.com>
---
Cc: Jianfeng Tan <jianfeng....@intel.com>
Cc: Santosh Shukla <santosh.shu...@caviumnetworks.com>
Cc: Thomas Monjalon <tho...@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.
IOVA_VA is also necessary if we don't want/have root privilege to run
DPDK app.
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).
Yes, agreed.
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;
For the KNI crash, do we really need to correct the case: make those
PMDs (requires IOVA_VA) also work well with KNI?
- If no, we just need to report RTE_PCI_DRV_IOVA_AS_VA smartly.
- If yes, we shall make KNI work with VA.
Thanks,
Jianfeng
