2014-10-25 00:48, Ouyang, Changchun: > From: Thomas Monjalon [mailto:thomas.monjalon at 6wind.com] > > 2014-10-24 16:38, Ouyang Changchun: > > > For zero copy, it need check whether RX descriptor num meets the least > > > requirement when using vector PMD Rx function, and give user more > > > hints if it fails to meet the least requirement. > > [...] > > > --- a/examples/vhost/main.c > > > +++ b/examples/vhost/main.c > > > @@ -131,6 +131,10 @@ > > > #define RTE_TEST_RX_DESC_DEFAULT_ZCP 32 /* legacy: 32, DPDK virt FE: > > > 128. */ > > > #define RTE_TEST_TX_DESC_DEFAULT_ZCP 64 /* legacy: 64, DPDK virt FE: > > > 64. */ > > > > > > +#ifdef RTE_IXGBE_INC_VECTOR > > > +#define VPMD_RX_BURST 32 > > > +#endif > > > + > > > /* Get first 4 bytes in mbuf headroom. */ #define > > > MBUF_HEADROOM_UINT32(mbuf) (*(uint32_t *)((uint8_t *)(mbuf) \ > > > + sizeof(struct rte_mbuf))) > > > @@ -792,6 +796,19 @@ us_vhost_parse_args(int argc, char **argv) > > > return -1; > > > } > > > > > > +#ifdef RTE_IXGBE_INC_VECTOR > > > + if ((zero_copy == 1) && (num_rx_descriptor <= VPMD_RX_BURST)) { > > > + RTE_LOG(INFO, VHOST_PORT, > > > + "The RX desc num: %d is too small for PMD to work\n" > > > + "properly, please enlarge it to bigger than %d if\n" > > > + "possible by the option: '--rx-desc-num <number>'\n" > > > + "One alternative is disabling RTE_IXGBE_INC_VECTOR\n" > > > + "in config file and rebuild the libraries.\n", > > > + num_rx_descriptor, VPMD_RX_BURST); > > > + return -1; > > > + } > > > +#endif > > > + > > > return 0; > > > } > > > > I feel there is a design problem here. > > An application shouldn't have to care about the underlying driver. > > For most of other applications, as their descriptor numbers are set as big > enough(1024 or so) , > So there is no need to check the descriptor number at the early stage of > running. > > But for vhost zero copy(note vhost one copy also has 1024 descriptor number) > has the default > descriptor number of 32. > Why use 32? > because vhost zero copy implementation (working as backend) need support dpdk > based app which use pmd virtio-net driver, > And also need support linux legacy virtio-net based application. > When it is the linux legacy virtio-net case, on one side the qemu has hard > code to confine the total virtio descriptor size to 256, > On other side, legacy virtio use half of them as virtio header, and then only > another half i.e. 128 descriptors are available to use as real buffer. > > In PMD mode, all HW descriptors need to be filled DMA address in the rx > initial stage, otherwise there is probably exceptional in rx process. > Based on that, we need use really limited virtio buffer to fully fill all hw > descriptor DMA address, > Or in other word, the available virtio descriptor size will determine the > total mbuf size and hw descriptor size in the case of zero copy, > > Tune and find that 32 is the suitable value for vhost zero copy to work > properly when it legacy linux virtio case. > Another factor to reduce the value to 32, is that mempool use ring to > accommodate the mbuf, it cost one to flag the ring head/tail, > And there are some other overheads like temporary mbufs(size as RX_BURST) > when rx. > Note that number descriptor should need power 2. > > Why the change occur at this moment? > Recently the default rx function is modified into vector RX function, while > it use non-vector mode (scalar mode) Rx previously, > Vector RX function need more than 32 descriptor to work properly, but scalar > mode RX hasn't this limitation. > > As the RX function is changeable(you can use vector mode or non-vector), and > descriptor number can also be changed. > So here in the vhost app, check if they match to make sure all things could > work normally, and give some hints if they don't match. > > Hope the above could make it a bit clearer. :-)
Thank you for your explanation. Your fix shows that driver and application are tightly linked. It's a design issue. As I said: "An application shouldn't have to care about the underlying driver." I didn't dig enough in vhost to suggest a good fix but I'm sure someone could have an idea. -- Thomas
