Re: [PATCH] Skip vfio in the scenario of non-privileged mode

[Stephen Hemminger <stephen@networkplumber.org>]

On Wed, 22 Jan 2025 16:15:03 +0800
Yang Ming <ming.1.yang@nokia-sbell.com> wrote:

> On 2025/1/18 00:47, Stephen Hemminger wrote:
> > Caution: This is an external email. Please be very careful when clicking links or opening attachments. See http://nok.it/nsb for additional information.
> >
> > On Fri, 17 Jan 2025 15:28:47 +0800
> > Yang Ming <ming.1.yang@nokia-sbell.com> wrote:
> >  
> >> DPDK detect vfio container according the existence of vfio
> >> module. But for container with non-privileged mode, there is
> >> possibility that no VFIO_DIR(/dev/vfio) mapping from host to
> >> container when host have both Intel NIC and Mellanox NIC but
> >> this conntainer only allocate VFs from Mellanox NIC.
> >> In this case, vfio kernel module has already been loaded from
> >> the host.
> >> This scenario will cause the error log occurs in DPDK primary
> >> process as below:
> >> 'EAL:   cannot open VFIO container, error 2 (No such file or
> >> directory)'
> >> 'EAL: VFIO support could not be initialized'
> >> Because `rte_vfio_enable()` call `rte_vfio_get_container_fd()`
> >> to execute `vfio_container_fd = open(VFIO_CONTAINER_PATH,
> >> O_RDWR);` but VFIO_CONTAINER_PATH(/dev/vfio/vfio) doesn't exist
> >> in this container.
> >> This scenario will also lead to the delay of DPDK secondary
> >> process because `default_vfio_cfg->vfio_enabled = 0` and
> >> `default_vfio_cfg->vfio_container_fd = -1`, socket error will
> >> be set in DPDK primary process when it sync this info to
> >> the secondary process.
> >> This patch use to skip this kind of useless detection for this
> >> scenario.
> >>
> >> Signed-off-by: Yang Ming <ming.1.yang@nokia-sbell.com>
> >> ---
> >>   lib/eal/linux/eal_vfio.c | 11 +++++++++++
> >>   1 file changed, 11 insertions(+)
> >>
> >> diff --git a/lib/eal/linux/eal_vfio.c b/lib/eal/linux/eal_vfio.c
> >> index 7132e24cba..1679d29263 100644
> >> --- a/lib/eal/linux/eal_vfio.c
> >> +++ b/lib/eal/linux/eal_vfio.c
> >> @@ -7,6 +7,7 @@
> >>   #include <fcntl.h>
> >>   #include <unistd.h>
> >>   #include <sys/ioctl.h>
> >> +#include <dirent.h>
> >>   
> >>   #include <rte_errno.h>
> >>   #include <rte_log.h>
> >> @@ -1083,6 +1084,7 @@ rte_vfio_enable(const char *modname)
> >>   	/* initialize group list */
> >>   	int i, j;
> >>   	int vfio_available;
> >> +	DIR *dir;
> >>   	const struct internal_config *internal_conf =
> >>   		eal_get_internal_configuration();
> >>   
> >> @@ -1119,6 +1121,15 @@ rte_vfio_enable(const char *modname)
> >>   		return 0;
> >>   	}
> >>   
> >> +	/* return 0 if VFIO directory not exist for container with non-privileged mode */
> >> +	dir = opendir(VFIO_DIR);
> >> +	if (dir == NULL) {
> >> +		EAL_LOG(DEBUG,
> >> +			"VFIO directory not exist, skipping VFIO support...");
> >> +		return 0;
> >> +	}
> >> +	closedir(dir);  
> > You need to test the non-container cases.
> > If vfio is loaded /dev/vfio is a character device (not a directory)
> >
> > Also looks suspicious that VFIO_DIR is defined but never used currently.
> >  
> Hi Stephen,
> For non-container test, /dev/vfio/vfio will be character device, not 
> /dev/vfio.
> Here is the command result on my testing environment with Intel NIC.
> 
> [root@computer-1 testuser]# ls -l /dev/vfio
> total 0
> crw-rw-rw-. 1 root root 10, 196 Jan 22 01:50 vfio
> [root@computer-1 testuser]# dpdk-devbind.py -b vfio-pci 0000:04:10.2
> [root@computer-1 testuser]# ls -l /dev/vfio
> total 0
> crw-------. 1 root root 239,   0 Jan 22 01:52 59
> crw-rw-rw-. 1 root root  10, 196 Jan 22 01:50 vfio
> [root@computer-1 testuser]# dpdk-devbind.py -b ixgbevf 0000:04:10.2
> [root@computer-1 testuser]# ls -l /dev/vfio
> total 0
> crw-rw-rw-. 1 root root 10, 196 Jan 22 01:50 vfio
> 
> Can you confirm your test scenario?
> 
> 

When vfio-pci is loaded but no device bound:
$ ls -l /dev/vfio
total 0
crw-rw-rw- 1 root root 10, 196 Feb 26 05:39 vfio

After binding device
$ ls -l /dev/vfio
total 0
crw------- 1 root root 511,   0 Feb 26 05:42 15
crw-rw-rw- 1 root root  10, 196 Feb 26 05:39 vfio

So testing for /dev/vfio is good indication that module is loaded.
Not sure what I was thinking earlier.