Re: [v4] net/gve: check driver compatibility

[Rushil Gupta <rushilg@google.com>]

[-- Attachment #1: Type: text/plain, Size: 3716 bytes --]

1. This is the excerpt from the google's virtual nic spec:
"In addition to the device-owned register file, vector table, and
doorbells, the gVNIC device uses *DMA* (which in most cases amounts to
ordinary memory access by host software since we're dealing with a virtual
device, but guests must assume the device could be backed by actual
hardware) to access physical memory. The following are all located in
physical memory: Admin queue - 4096-byte command queue used for configuring
gVNIC.
Some commands require an additional dma memory region to be passed to the
device. These memory regions are allocated to execute the command and freed
when the command completes."
The calloc by default doesn't allow memory to be shared between the dpdk
process and hypervisor (where virtual device lives); so that's the reason
it doesn't work.

2. I also have a query: RHEL8 compilation in ci/Intel-compilation context
fails due to; is this because of if `not is_linux`

meson.build:67:0: ERROR: Include dir lib/eal/linux/include does not exist.

Passes: http://patchwork.dpdk.org/project/dpdk/patch/20230508191552.104540-1-rushilg@google.com/

Fails: http://patchwork.dpdk.org/project/dpdk/patch/20230519204618.1507956-1-rushilg@google.com/


On Mon, May 22, 2023 at 1:52 AM Ferruh Yigit <ferruh.yigit@amd.com> wrote:

> On 5/19/2023 9:46 PM, Rushil Gupta wrote:
> > +static int
> > +gve_verify_driver_compatibility(struct gve_priv *priv)
> > +{
> > +     const struct rte_memzone *driver_info_mem;
> > +     struct gve_driver_info *driver_info;
> > +     int err;
> > +
> > +     driver_info_mem =
> rte_memzone_reserve_aligned("verify_driver_compatibility",
> > +                     sizeof(struct gve_driver_info),
> > +                     rte_socket_id(),
> > +                     RTE_MEMZONE_IOVA_CONTIG, PAGE_SIZE);
> > +
> > +     if (driver_info_mem == NULL) {
> > +             PMD_DRV_LOG(ERR,
> > +                         "Could not alloc memzone for driver
> compatibility");
> > +             return -ENOMEM;
> > +     }
> > +     driver_info = (struct gve_driver_info *)driver_info_mem->addr;
> > +
> > +     *driver_info = (struct gve_driver_info) {
> > +             .os_type = 5, /* DPDK */
> > +             .driver_major = GVE_VERSION_MAJOR,
> > +             .driver_minor = GVE_VERSION_MINOR,
> > +             .driver_sub = GVE_VERSION_SUB,
> > +             .os_version_major = cpu_to_be32(DPDK_VERSION_MAJOR),
> > +             .os_version_minor = cpu_to_be32(DPDK_VERSION_MINOR),
> > +             .os_version_sub = cpu_to_be32(DPDK_VERSION_SUB),
> > +             .driver_capability_flags = {
> > +                     cpu_to_be64(GVE_DRIVER_CAPABILITY_FLAGS1),
> > +                     cpu_to_be64(GVE_DRIVER_CAPABILITY_FLAGS2),
> > +                     cpu_to_be64(GVE_DRIVER_CAPABILITY_FLAGS3),
> > +                     cpu_to_be64(GVE_DRIVER_CAPABILITY_FLAGS4),
> > +             },
> > +     };
> > +
> > +     populate_driver_version_strings((char
> *)driver_info->os_version_str1,
> > +                     (char *)driver_info->os_version_str2);
> > +
> > +     err = gve_adminq_verify_driver_compatibility(priv,
> > +             sizeof(struct gve_driver_info), (dma_addr_t)driver_info);
>
> Back to previous discussion, other commands pass physical address to the
> admin command, but this pass virtual address.
> To follow the same semantic, shouldn't above be 'driver_info_mem.iova'?
>
> I asked before but not able to get an answer, what is the memory type
> requirement for device?
> Why virtual address obtained via 'calloc()' is not working, but virtual
> address from hugepages are working?
>

[-- Attachment #2: Type: text/html, Size: 5038 bytes --]