On 3/23/2017 11:23 AM, Allain Legacy wrote: > This patch series submits an initial version of the AVP PMD from Wind River > Systems. The series includes shared header files, driver implementation, > and changes to documentation files in support of this new driver. The AVP > driver is a shared memory based device. It is intended to be used as a PMD > within a virtual machine running on a Wind River virtualization platform. > See: http://www.windriver.com/products/titanium-cloud/ > > It enables optimized packet throughput without requiring any packet > processing in qemu. This allowed us to provide our customers with a > significant performance increase for both DPDK and non-DPDK applications > in the VM. Since our AVP implementation supports VM live-migration it > is viewed as a better alternative to PCI passthrough or PCI SRIOV since > neither of those support VM live-migration without manual intervention > or significant performance penalties. > > Since the initial implementation of AVP devices, vhost-user has become part > of the qemu offering with a significant performance increase over the > original virtio implementation. However, vhost-user still does not achieve > the level of performance that the AVP device can provide to our customers > for DPDK based guests. > > A number of our customers have requested that we upstream the driver to > dpdk.org. > > v2: > * Fixed coding style violations that slipped in accidentally because of an > out of date checkpatch.pl from an older kernel. > > v3: > * Updated 17.05 release notes to add a section for this new PMD > * Added additional info to the AVP nic guide document to clarify the > benefit of using AVP over virtio. > * Fixed spelling error in debug log missed by local checkpatch.pl version > * Split the transmit patch to separate the stats functions as they > accidentally got squashed in the last patchset. > * Fixed debug log strings so that they exceed 80 characters rather than > span multiple lines. > * Renamed RTE_AVP_* defines that were in avp_ethdev.h to be AVP_* instead > * Replaced usage of RTE_WRITE32 and RTE_READ32 with rte_write32_relaxed > and rte_read32_relaxed. > * Declared rte_pci_id table as const > > v4: > * Split our interrupt handlers to a separate patch and moved to the end > of the series. > * Removed memset() from stats_get API > * Removed usage of RTE_AVP_ALIGNMENT > * Removed unnecessary parentheses in rte_avp_common.h > * Removed unneeded "goto unlock" where there are no statements in between > the goto and the end of the function. > * Re-tested with pktgen and found that rte_eth_tx_burst() is being called > with 0 packets even before starting traffic which resulted in > incrementing oerrors; fixed in transmit patch. > > v5: > * Updated documentation to remove references to ivshmem as it lead to > confusion about whether AVP is exactly like ivshmem or simply based on > how ivshmem exports memory to a VM via a PCI device. > * Restructured first set of patches to condense them down to a base patch > with the files needed to apply subsequent patches. > * Removed static prototypes from init/uninit functions in avp_ethdev.c > * Moved MAC addresses init to the device initialization patch because it > is setup by the avp_dev_create() function. > * Split the changes to the avp.ini features file so that features are > marked as enabled in the patch that actually enables them. > > Allain Legacy (14): > drivers/net: adds AVP PMD base files > net/avp: public header files > net/avp: debug log macros > net/avp: driver registration > net/avp: device initialization > net/avp: device configuration > net/avp: queue setup and release > net/avp: packet receive functions > net/avp: packet transmit functions > net/avp: device statistics operations > net/avp: device promiscuous functions > net/avp: device start and stop operations > net/avp: migration interrupt handling > doc: adds information related to the AVP PMD
Hi Allain, Overall PMD looks good to me. Only can you please update documentation based on tech board decision [1]. Repeating related part here: " It should be clearly stated that right now AVP PMD is limited to use with WindRiver hypervisor and for most cases virtio is still a preferred virtual device to use with QEMU. " Thanks, ferruh [1] http://dpdk.org/ml/archives/dev/2017-March/061009.html
