[dpdk-dev] [PATCH v4] eal: add VFIO-PCI SR-IOV support
Wang, Haiyue
haiyue.wang at intel.com
Mon Apr 13 18:45:01 CEST 2020
Hi Andrew,
> -----Original Message-----
> From: Andrew Rybchenko <arybchenko at solarflare.com>
> Sent: Monday, April 13, 2020 23:38
> To: Wang, Haiyue <haiyue.wang at intel.com>; dev at dpdk.org; thomas at monjalon.net; vattunuru at marvell.com;
> jerinj at marvell.com; alex.williamson at redhat.com; david.marchand at redhat.com
> Subject: Re: [dpdk-dev] [PATCH v4] eal: add VFIO-PCI SR-IOV support
>
> On 4/13/20 11:29 AM, Haiyue Wang wrote:
> > The kernel module vfio-pci introduces the VF token to enable SR-IOV
> > support since 5.7.
> >
> > The VF token can be set by a vfio-pci based PF driver and must be known
> > by the vfio-pci based VF driver in order to gain access to the device.
> >
> > An example VF token option would take this form:
> >
> > 1. Install vfio-pci with option 'enable_sriov=1'
> >
> > 2. ./usertools/dpdk-devbind.py -b vfio-pci 0000:87:00.0
> >
> > 3. echo 2 > /sys/bus/pci/devices/0000:87:00.0/sriov_numvfs
> >
> > 4. Start the PF:
> > ./x86_64-native-linux-gcc/app/testpmd -l 22-25 -n 4 \
> > -w 87:00.0,vf_token=2ab74924-c335-45f4-9b16-8569e5b08258 \
> > --file-prefix=pf -- -i
>
> Should I get a token from my head? Any?
>
> > 5. Start the VF:
> > ./x86_64-native-linux-gcc/app/testpmd -l 26-29 -n 4 \
> > -w 87:02.0,vf_token=2ab74924-c335-45f4-9b16-8569e5b08258 \
> > --file-prefix=vf1 -- -i
> >
> > Signed-off-by: Haiyue Wang <haiyue.wang at intel.com>
> > Acked-by: Vamsi Attunuru <vattunuru at marvell.com>
> > ---
> > v4: 1. Ignore rte_vfio_setup_device ABI check since it is
> > for Linux driver use.
> >
> > v3: https://patchwork.dpdk.org/patch/68254/
> > Fix the Travis build failed:
> > (1). rte_uuid.h:97:55: error: unknown type name ‘size_t’
> > (2). rte_uuid.h:58:2: error: implicit declaration of function ‘memcpy’
> >
> > v2: https://patchwork.dpdk.org/patch/68240/
> > Fix the FreeBSD build error.
> >
> > v1: https://patchwork.dpdk.org/patch/68237/
> > Update the commit message.
> >
> > RFC v2: https://patchwork.dpdk.org/patch/68114/
> > Based on Vamsi's RFC v1, and Alex's patch for Qemu
> > [https://lore.kernel.org/lkml/20200204161737.34696b91@w520.home/]:
> > Use the devarg to pass-down the VF token.
> >
> > RFC v1: https://patchwork.dpdk.org/patch/66281/ by Vamsi.
> > ---
> > devtools/libabigail.abignore | 3 ++
> > drivers/bus/pci/linux/pci_vfio.c | 56 +++++++++++++++++++++++++++++--
> > lib/librte_eal/freebsd/eal.c | 3 +-
> > lib/librte_eal/include/rte_uuid.h | 2 ++
> > lib/librte_eal/include/rte_vfio.h | 8 ++++-
> > lib/librte_eal/linux/eal_vfio.c | 20 +++++++++--
> > 6 files changed, 85 insertions(+), 7 deletions(-)
> >
> > diff --git a/devtools/libabigail.abignore b/devtools/libabigail.abignore
> > index a59df8f13..d918746b4 100644
> > --- a/devtools/libabigail.abignore
> > +++ b/devtools/libabigail.abignore
> > @@ -11,3 +11,6 @@
> > type_kind = enum
> > name = rte_crypto_asym_xform_type
> > changed_enumerators = RTE_CRYPTO_ASYM_XFORM_TYPE_LIST_END
> > +; Ignore this function which is only relevant to linux for driver
> > +[suppress_type]
> > + name = rte_vfio_setup_device
> > diff --git a/drivers/bus/pci/linux/pci_vfio.c b/drivers/bus/pci/linux/pci_vfio.c
> > index 64cd84a68..7f99337c7 100644
> > --- a/drivers/bus/pci/linux/pci_vfio.c
> > +++ b/drivers/bus/pci/linux/pci_vfio.c
> > @@ -11,6 +11,7 @@
> > #include <sys/mman.h>
> > #include <stdbool.h>
> >
> > +#include <rte_devargs.h>
> > #include <rte_log.h>
> > #include <rte_pci.h>
> > #include <rte_bus_pci.h>
> > @@ -644,11 +645,59 @@ pci_vfio_msix_is_mappable(int vfio_dev_fd, int msix_region)
> > return ret;
> > }
> >
> > +static void
> > +vfio_pci_vf_token_arg(struct rte_devargs *devargs, rte_uuid_t uu)
> > +{
> > +#define VF_TOKEN_ARG "vf_token="
> > + char c, *p, *vf_token;
> > +
> > + if (devargs == NULL)
> > + return;
> > +
> > + p = strstr(devargs->args, VF_TOKEN_ARG);
> > + if (!p)
> > + return;
> > +
> > + vf_token = p + strlen(VF_TOKEN_ARG);
> > + if (strlen(vf_token) < (RTE_UUID_STRLEN - 1))
> > + return;
> > +
> > + c = vf_token[RTE_UUID_STRLEN - 1];
> > + if (c != '\0' && c != ',')
> > + return;
> > +
> > + vf_token[RTE_UUID_STRLEN - 1] = '\0';
>
> Is it possible to parse and handle devargs using rte_kvargs.h?
>
Since it needs to remove the 'vf_token', as 'vf_token' is not a
valid PMD related args, so need to parse and delete it.
rte_kvargs_parse(const char *args, const char * const valid_keys[])
> > + if (rte_uuid_parse(vf_token, uu)) {
> > + RTE_LOG(ERR, EAL,
> > + "The VF token is not a valid uuid : %s\n", vf_token);
> > + vf_token[RTE_UUID_STRLEN - 1] = c;
> > + return;
>
> I think that the function must return error which is handled
> by the caller when something bad happens (e.g. invalid
> UUID).
>
Yes, make sense, will add the error handling.
> > + }
> > +
> > + RTE_LOG(DEBUG, EAL,
> > + "The VF token is found : %s\n", vf_token);
> > +
> > + vf_token[RTE_UUID_STRLEN - 1] = c;
> > +
> > + /* Purge this vfio-pci specific token from the device arguments */
> > + if (c != '\0') {
> > + /* 1. Handle the case : 'vf_token=uuid,arg1=val1' */
> > + memmove(p, vf_token + RTE_UUID_STRLEN,
> > + strlen(vf_token + RTE_UUID_STRLEN) + 1);
> > + } else {
> > + /* 2. Handle the case : 'arg1=val1,vf_token=uuid' */
> > + if (p != devargs->args)
> > + p--;
> > +
> > + *p = '\0';
> > + }
>
> Is it really required to purge? Why? If yes, it should be explained in
> the comment above.
Please see above reply.
>
> > +}
> >
> > static int
> > pci_vfio_map_resource_primary(struct rte_pci_device *dev)
> > {
> > struct vfio_device_info device_info = { .argsz = sizeof(device_info) };
> > + rte_uuid_t vf_token = RTE_UUID_INIT(0, 0, 0, 0, 0ULL);
>
> May be it would be better if vfio_pci_vf_token_arg()
> initializes it anyway instead of duplication init
> in two places?
+1, will update it.
>
> > char pci_addr[PATH_MAX] = {0};
> > int vfio_dev_fd;
> > struct rte_pci_addr *loc = &dev->addr;
> > @@ -668,8 +717,9 @@ pci_vfio_map_resource_primary(struct rte_pci_device *dev)
> > snprintf(pci_addr, sizeof(pci_addr), PCI_PRI_FMT,
> > loc->domain, loc->bus, loc->devid, loc->function);
> >
> > + vfio_pci_vf_token_arg(dev->device.devargs, vf_token);
> > ret = rte_vfio_setup_device(rte_pci_get_sysfs_path(), pci_addr,
> > - &vfio_dev_fd, &device_info);
> > + &vfio_dev_fd, &device_info, vf_token);
> > if (ret)
> > return ret;
> >
> > @@ -797,6 +847,7 @@ static int
> > pci_vfio_map_resource_secondary(struct rte_pci_device *dev)
> > {
> > struct vfio_device_info device_info = { .argsz = sizeof(device_info) };
> > + rte_uuid_t vf_token = RTE_UUID_INIT(0, 0, 0, 0, 0ULL);
> > char pci_addr[PATH_MAX] = {0};
> > int vfio_dev_fd;
> > struct rte_pci_addr *loc = &dev->addr;
> > @@ -830,8 +881,9 @@ pci_vfio_map_resource_secondary(struct rte_pci_device *dev)
> > return -1;
> > }
> >
> > + vfio_pci_vf_token_arg(dev->device.devargs, vf_token);
> > ret = rte_vfio_setup_device(rte_pci_get_sysfs_path(), pci_addr,
> > - &vfio_dev_fd, &device_info);
> > + &vfio_dev_fd, &device_info, vf_token);
> > if (ret)
> > return ret;
> >
> > diff --git a/lib/librte_eal/freebsd/eal.c b/lib/librte_eal/freebsd/eal.c
> > index 6ae37e7e6..a92584795 100644
> > --- a/lib/librte_eal/freebsd/eal.c
> > +++ b/lib/librte_eal/freebsd/eal.c
> > @@ -995,7 +995,8 @@ rte_eal_vfio_intr_mode(void)
> > int rte_vfio_setup_device(__rte_unused const char *sysfs_base,
> > __rte_unused const char *dev_addr,
> > __rte_unused int *vfio_dev_fd,
> > - __rte_unused struct vfio_device_info *device_info)
> > + __rte_unused struct vfio_device_info *device_info,
> > + __rte_unused rte_uuid_t vf_token)
> > {
> > return -1;
> > }
> > diff --git a/lib/librte_eal/include/rte_uuid.h b/lib/librte_eal/include/rte_uuid.h
> > index 044afbdfa..8b42e070a 100644
> > --- a/lib/librte_eal/include/rte_uuid.h
> > +++ b/lib/librte_eal/include/rte_uuid.h
> > @@ -15,6 +15,8 @@ extern "C" {
> > #endif
> >
> > #include <stdbool.h>
> > +#include <stddef.h>
> > +#include <string.h>
> >
> > /**
> > * Struct describing a Universal Unique Identifier
> > diff --git a/lib/librte_eal/include/rte_vfio.h b/lib/librte_eal/include/rte_vfio.h
> > index 20ed8c45a..1f9e22d82 100644
> > --- a/lib/librte_eal/include/rte_vfio.h
> > +++ b/lib/librte_eal/include/rte_vfio.h
> > @@ -16,6 +16,8 @@ extern "C" {
> >
> > #include <stdint.h>
> >
> > +#include <rte_uuid.h>
> > +
> > /*
> > * determine if VFIO is present on the system
> > */
> > @@ -102,13 +104,17 @@ struct vfio_device_info;
> > * @param device_info
> > * Device information.
> > *
> > + * @param vf_token
> > + * VF token.
>
> Such comments are useles and just eat space adding nothing
> useful. Please, make it useful and explain what is behind the
> parameter, when it is necessary, why? Should it be specified
> for PF case, VF case, both?
>
Will add more comments, yes for both PF and VF, as Alex's linux patch
explained.
> > + *
> > * @return
> > * 0 on success.
> > * <0 on failure.
> > * >1 if the device cannot be managed this way.
> > */
> > int rte_vfio_setup_device(const char *sysfs_base, const char *dev_addr,
> > - int *vfio_dev_fd, struct vfio_device_info *device_info);
> > + int *vfio_dev_fd, struct vfio_device_info *device_info,
> > + rte_uuid_t vf_token);
>
> "rte_uuid_t vf_token" looks confusing. Shouldn't it be
> "rte_uuid_t *vf_token"?
This is UUID API design and type definition:
bool rte_uuid_is_null(const rte_uuid_t uu);
DPDK: typedef unsigned char rte_uuid_t[16];
vs
Linux: typedef struct {
__u8 b[UUID_SIZE];
} uuid_t;
>
> >
> > /**
> > * Release a device mapped to a VFIO-managed I/O MMU group.
> > diff --git a/lib/librte_eal/linux/eal_vfio.c b/lib/librte_eal/linux/eal_vfio.c
> > index 4502aefed..916082b5d 100644
> > --- a/lib/librte_eal/linux/eal_vfio.c
> > +++ b/lib/librte_eal/linux/eal_vfio.c
> > @@ -702,7 +702,8 @@ rte_vfio_clear_group(int vfio_group_fd)
> >
> > int
> > rte_vfio_setup_device(const char *sysfs_base, const char *dev_addr,
> > - int *vfio_dev_fd, struct vfio_device_info *device_info)
> > + int *vfio_dev_fd, struct vfio_device_info *device_info,
> > + rte_uuid_t vf_token)
> > {
> > struct vfio_group_status group_status = {
> > .argsz = sizeof(group_status)
> > @@ -712,6 +713,7 @@ rte_vfio_setup_device(const char *sysfs_base, const char *dev_addr,
> > int vfio_container_fd;
> > int vfio_group_fd;
> > int iommu_group_num;
> > + char dev[PATH_MAX];
>
> Why PATH_MAX?
Based on Vamsi's RFC v1, and found that it looked a little reasonable, ' char pci_addr[PATH_MAX] = {0}; '
static int
pci_vfio_map_resource_primary(struct rte_pci_device *dev)
{
struct vfio_device_info device_info = { .argsz = sizeof(device_info) };
char pci_addr[PATH_MAX] = {0}; <----
>
> > int i, ret;
> >
> > /* get group number */
> > @@ -895,8 +897,19 @@ rte_vfio_setup_device(const char *sysfs_base, const char *dev_addr,
> > t->type_id, t->name);
> > }
> >
> > + if (!rte_uuid_is_null(vf_token)) {
> > + char vf_token_str[RTE_UUID_STRLEN];
> > +
> > + rte_uuid_unparse(vf_token, vf_token_str, sizeof(vf_token_str));
> > + snprintf(dev, sizeof(dev),
> > + "%s vf_token=%s", dev_addr, vf_token_str);
> > + } else {
> > + snprintf(dev, sizeof(dev),
> > + "%s", dev_addr);
> > + }
> > +
> > /* get a file descriptor for the device */
> > - *vfio_dev_fd = ioctl(vfio_group_fd, VFIO_GROUP_GET_DEVICE_FD, dev_addr);
> > + *vfio_dev_fd = ioctl(vfio_group_fd, VFIO_GROUP_GET_DEVICE_FD, dev);
> > if (*vfio_dev_fd < 0) {
> > /* if we cannot get a device fd, this implies a problem with
> > * the VFIO group or the container not having IOMMU configured.
> > @@ -2081,7 +2094,8 @@ int
> > rte_vfio_setup_device(__rte_unused const char *sysfs_base,
> > __rte_unused const char *dev_addr,
> > __rte_unused int *vfio_dev_fd,
> > - __rte_unused struct vfio_device_info *device_info)
> > + __rte_unused struct vfio_device_info *device_info,
> > + __rte_unused rte_uuid_t vf_token)
> > {
> > return -1;
> > }
> >
More information about the dev
mailing list