[Intel-gfx] [PATCH i-g-t v3] lib/igt_device: Add support for accessing unbound VF PCI devices

[Intel-gfx] [PATCH i-g-t v3] lib/igt_device: Add support for accessing unbound VF PCI devices

[Janusz Krzysztofik]

The library provides igt_device_get_pci_device() function that allows to
get access to a PCI device from an open DRM device file descriptor.  It
can be used on VF devices as long as a DRM driver is bound to them.
However, SR-IOV tests may want to exercise VF PCI devices created by a PF
without binding any DRM driver to them.

While keeping the API of igt_device_get_pci_device() untouched, extend API
of its underlying helper __igt_device_get_pci_device() with an extra
argument for specifying VF ID of the requested PCI device and expose this
function as public.

v2: refresh on top of IGT libpciaccess wrappers and drop previously added
    but no longer needed error unwind path and recommendations for users
    on calling pci_system_cleanup() after use (Chris),
  - fix incorrect validation of snprintf() result and misaligned
    formatting of igt_warn_on_f() arguments.
v3: follow VF numbering convention of Linux PCI ABI (Chris),
  - fix and improve DOC.

Signed-off-by: Janusz Krzysztofik <janusz.krzysztofik@linux.intel.com>
Reviewed-by: Chris Wilson <chris.p.wilson@intel.com> # v2
---
 lib/igt_device.c | 33 +++++++++++++++++++++++++++------
 lib/igt_device.h |  1 +
 2 files changed, 28 insertions(+), 6 deletions(-)

diff --git a/lib/igt_device.c b/lib/igt_device.c
index c50bf4a1f7..46b7dbc490 100644
--- a/lib/igt_device.c
+++ b/lib/igt_device.c
@@ -149,9 +149,9 @@ struct igt_pci_addr {
 	unsigned int function;
 };
 
-static int igt_device_get_pci_addr(int fd, struct igt_pci_addr *pci)
+static int igt_device_get_pci_addr(int fd, int vf_id, struct igt_pci_addr *pci)
 {
-	char path[IGT_DEV_PATH_LEN];
+	char link[20], path[IGT_DEV_PATH_LEN];
 	char *buf;
 	int sysfs;
 	int len;
@@ -159,11 +159,21 @@ static int igt_device_get_pci_addr(int fd, struct igt_pci_addr *pci)
 	if (!igt_device_is_pci(fd))
 		return -ENODEV;
 
+	if (vf_id < 0)
+		len = snprintf(link, sizeof(link), "device");
+	else
+		len = snprintf(link, sizeof(link), "device/virtfn%u", vf_id);
+	if (igt_warn_on_f(len >= sizeof(link),
+			  "IGT bug: insufficient buffer space for rendering PCI device link name\n"))
+		return -ENOSPC;
+	else if (igt_debug_on_f(len < 0, "unexpected failure from snprintf()\n"))
+		return len;
+
 	sysfs = igt_sysfs_open(fd);
 	if (sysfs == -1)
 		return -ENOENT;
 
-	len = readlinkat(sysfs, "device", path, sizeof(path) - 1);
+	len = readlinkat(sysfs, link, path, sizeof(path) - 1);
 	close(sysfs);
 	if (len == -1)
 		return -ENOENT;
@@ -183,12 +193,23 @@ static int igt_device_get_pci_addr(int fd, struct igt_pci_addr *pci)
 	return 0;
 }
 
-static struct pci_device *__igt_device_get_pci_device(int fd)
+/**
+ * __igt_device_get_pci_device:
+ *
+ * @fd: DRM device file descriptor
+ * @vf_id: PCI virtual function number or -1 if native or PF itself
+ *
+ * Looks up a PCI interface of a DRM device or a VF PCI device of the DRM PF using libpciaccess.
+ *
+ * Returns:
+ * The pci_device, NULL on any failures.
+ */
+struct pci_device *__igt_device_get_pci_device(int fd, int vf_id)
 {
 	struct igt_pci_addr pci_addr;
 	struct pci_device *pci_dev;
 
-	if (igt_device_get_pci_addr(fd, &pci_addr)) {
+	if (igt_device_get_pci_addr(fd, vf_id, &pci_addr)) {
 		igt_warn("Unable to find device PCI address\n");
 		return NULL;
 	}
@@ -231,7 +252,7 @@ struct pci_device *igt_device_get_pci_device(int fd)
 {
 	struct pci_device *pci_dev;
 
-	pci_dev = __igt_device_get_pci_device(fd);
+	pci_dev = __igt_device_get_pci_device(fd, -1);
 	igt_require(pci_dev);
 
 	return pci_dev;
diff --git a/lib/igt_device.h b/lib/igt_device.h
index 278ba7a9b3..00da853e71 100644
--- a/lib/igt_device.h
+++ b/lib/igt_device.h
@@ -33,5 +33,6 @@ void igt_device_drop_master(int fd);
 
 int igt_device_get_card_index(int fd);
 struct pci_device *igt_device_get_pci_device(int fd);
+struct pci_device *__igt_device_get_pci_device(int fd, int vf_id);
 
 #endif /* __IGT_DEVICE_H__ */
-- 
2.25.1

Re: [Intel-gfx] [igt-dev] [PATCH i-g-t v3] lib/igt_device: Add support for accessing unbound VF PCI devices

[Michal Wajdeczko]

On 22.02.2022 16:11, Janusz Krzysztofik wrote:
> The library provides igt_device_get_pci_device() function that allows to
> get access to a PCI device from an open DRM device file descriptor.  It
> can be used on VF devices as long as a DRM driver is bound to them.
> However, SR-IOV tests may want to exercise VF PCI devices created by a PF
> without binding any DRM driver to them.
> 
> While keeping the API of igt_device_get_pci_device() untouched, extend API
> of its underlying helper __igt_device_get_pci_device() with an extra
> argument for specifying VF ID of the requested PCI device and expose this
> function as public.
> 
> v2: refresh on top of IGT libpciaccess wrappers and drop previously added
>     but no longer needed error unwind path and recommendations for users
>     on calling pci_system_cleanup() after use (Chris),
>   - fix incorrect validation of snprintf() result and misaligned
>     formatting of igt_warn_on_f() arguments.
> v3: follow VF numbering convention of Linux PCI ABI (Chris),
>   - fix and improve DOC.
> 
> Signed-off-by: Janusz Krzysztofik <janusz.krzysztofik@linux.intel.com>
> Reviewed-by: Chris Wilson <chris.p.wilson@intel.com> # v2
> ---
>  lib/igt_device.c | 33 +++++++++++++++++++++++++++------
>  lib/igt_device.h |  1 +
>  2 files changed, 28 insertions(+), 6 deletions(-)
> 
> diff --git a/lib/igt_device.c b/lib/igt_device.c
> index c50bf4a1f7..46b7dbc490 100644
> --- a/lib/igt_device.c
> +++ b/lib/igt_device.c
> @@ -149,9 +149,9 @@ struct igt_pci_addr {
>  	unsigned int function;
>  };
>  
> -static int igt_device_get_pci_addr(int fd, struct igt_pci_addr *pci)
> +static int igt_device_get_pci_addr(int fd, int vf_id, struct igt_pci_addr *pci)
>  {
> -	char path[IGT_DEV_PATH_LEN];
> +	char link[20], path[IGT_DEV_PATH_LEN];
>  	char *buf;
>  	int sysfs;
>  	int len;
> @@ -159,11 +159,21 @@ static int igt_device_get_pci_addr(int fd, struct igt_pci_addr *pci)
>  	if (!igt_device_is_pci(fd))
>  		return -ENODEV;
>  
> +	if (vf_id < 0)
> +		len = snprintf(link, sizeof(link), "device");
> +	else
> +		len = snprintf(link, sizeof(link), "device/virtfn%u", vf_id);
> +	if (igt_warn_on_f(len >= sizeof(link),
> +			  "IGT bug: insufficient buffer space for rendering PCI device link name\n"))
> +		return -ENOSPC;
> +	else if (igt_debug_on_f(len < 0, "unexpected failure from snprintf()\n"))
> +		return len;
> +
>  	sysfs = igt_sysfs_open(fd);
>  	if (sysfs == -1)
>  		return -ENOENT;
>  
> -	len = readlinkat(sysfs, "device", path, sizeof(path) - 1);
> +	len = readlinkat(sysfs, link, path, sizeof(path) - 1);
>  	close(sysfs);
>  	if (len == -1)
>  		return -ENOENT;
> @@ -183,12 +193,23 @@ static int igt_device_get_pci_addr(int fd, struct igt_pci_addr *pci)
>  	return 0;
>  }
>  
> -static struct pci_device *__igt_device_get_pci_device(int fd)
> +/**
> + * __igt_device_get_pci_device:
> + *
> + * @fd: DRM device file descriptor
> + * @vf_id: PCI virtual function number or -1 if native or PF itself

this param seems to be used here rather as 0-based "index" that
subsystem uses to list virtfn entries, while real VF "numbers" are
1-based, see PCI spec which says:

"VFs are numbered starting with 1 so the first VF associated with PF M
is VF M,1."

maybe we should update the wording to minimize any confusions?

Michal

> + *
> + * Looks up a PCI interface of a DRM device or a VF PCI device of the DRM PF using libpciaccess.
> + *
> + * Returns:
> + * The pci_device, NULL on any failures.
> + */
> +struct pci_device *__igt_device_get_pci_device(int fd, int vf_id)
>  {
>  	struct igt_pci_addr pci_addr;
>  	struct pci_device *pci_dev;
>  
> -	if (igt_device_get_pci_addr(fd, &pci_addr)) {
> +	if (igt_device_get_pci_addr(fd, vf_id, &pci_addr)) {
>  		igt_warn("Unable to find device PCI address\n");
>  		return NULL;
>  	}
> @@ -231,7 +252,7 @@ struct pci_device *igt_device_get_pci_device(int fd)
>  {
>  	struct pci_device *pci_dev;
>  
> -	pci_dev = __igt_device_get_pci_device(fd);
> +	pci_dev = __igt_device_get_pci_device(fd, -1);
>  	igt_require(pci_dev);
>  
>  	return pci_dev;
> diff --git a/lib/igt_device.h b/lib/igt_device.h
> index 278ba7a9b3..00da853e71 100644
> --- a/lib/igt_device.h
> +++ b/lib/igt_device.h
> @@ -33,5 +33,6 @@ void igt_device_drop_master(int fd);
>  
>  int igt_device_get_card_index(int fd);
>  struct pci_device *igt_device_get_pci_device(int fd);
> +struct pci_device *__igt_device_get_pci_device(int fd, int vf_id);
>  
>  #endif /* __IGT_DEVICE_H__ */

Re: [Intel-gfx] [igt-dev] [PATCH i-g-t v3] lib/igt_device: Add support for accessing unbound VF PCI devices

[Janusz Krzysztofik]

Hi Michał,

Thanks for review.

On Tuesday, 22 February 2022 17:16:54 CET Michal Wajdeczko wrote:
> 
> On 22.02.2022 16:11, Janusz Krzysztofik wrote:
> > The library provides igt_device_get_pci_device() function that allows to
> > get access to a PCI device from an open DRM device file descriptor.  It
> > can be used on VF devices as long as a DRM driver is bound to them.
> > However, SR-IOV tests may want to exercise VF PCI devices created by a PF
> > without binding any DRM driver to them.
> > 
> > While keeping the API of igt_device_get_pci_device() untouched, extend API
> > of its underlying helper __igt_device_get_pci_device() with an extra
> > argument for specifying VF ID of the requested PCI device and expose this
> > function as public.
> > 
> > v2: refresh on top of IGT libpciaccess wrappers and drop previously added
> >     but no longer needed error unwind path and recommendations for users
> >     on calling pci_system_cleanup() after use (Chris),
> >   - fix incorrect validation of snprintf() result and misaligned
> >     formatting of igt_warn_on_f() arguments.
> > v3: follow VF numbering convention of Linux PCI ABI (Chris),
> >   - fix and improve DOC.
> > 
> > Signed-off-by: Janusz Krzysztofik <janusz.krzysztofik@linux.intel.com>
> > Reviewed-by: Chris Wilson <chris.p.wilson@intel.com> # v2
> > ---
> >  lib/igt_device.c | 33 +++++++++++++++++++++++++++------
> >  lib/igt_device.h |  1 +
> >  2 files changed, 28 insertions(+), 6 deletions(-)
> > 
> > diff --git a/lib/igt_device.c b/lib/igt_device.c
> > index c50bf4a1f7..46b7dbc490 100644
> > --- a/lib/igt_device.c
> > +++ b/lib/igt_device.c
> > @@ -149,9 +149,9 @@ struct igt_pci_addr {
> >  	unsigned int function;
> >  };
> >  
> > -static int igt_device_get_pci_addr(int fd, struct igt_pci_addr *pci)
> > +static int igt_device_get_pci_addr(int fd, int vf_id, struct igt_pci_addr *pci)
> >  {
> > -	char path[IGT_DEV_PATH_LEN];
> > +	char link[20], path[IGT_DEV_PATH_LEN];
> >  	char *buf;
> >  	int sysfs;
> >  	int len;
> > @@ -159,11 +159,21 @@ static int igt_device_get_pci_addr(int fd, struct igt_pci_addr *pci)
> >  	if (!igt_device_is_pci(fd))
> >  		return -ENODEV;
> >  
> > +	if (vf_id < 0)
> > +		len = snprintf(link, sizeof(link), "device");
> > +	else
> > +		len = snprintf(link, sizeof(link), "device/virtfn%u", vf_id);
> > +	if (igt_warn_on_f(len >= sizeof(link),
> > +			  "IGT bug: insufficient buffer space for rendering PCI device link name\n"))
> > +		return -ENOSPC;
> > +	else if (igt_debug_on_f(len < 0, "unexpected failure from snprintf()\n"))
> > +		return len;
> > +
> >  	sysfs = igt_sysfs_open(fd);
> >  	if (sysfs == -1)
> >  		return -ENOENT;
> >  
> > -	len = readlinkat(sysfs, "device", path, sizeof(path) - 1);
> > +	len = readlinkat(sysfs, link, path, sizeof(path) - 1);
> >  	close(sysfs);
> >  	if (len == -1)
> >  		return -ENOENT;
> > @@ -183,12 +193,23 @@ static int igt_device_get_pci_addr(int fd, struct igt_pci_addr *pci)
> >  	return 0;
> >  }
> >  
> > -static struct pci_device *__igt_device_get_pci_device(int fd)
> > +/**
> > + * __igt_device_get_pci_device:
> > + *
> > + * @fd: DRM device file descriptor
> > + * @vf_id: PCI virtual function number or -1 if native or PF itself
> 
> this param seems to be used here rather as 0-based "index" that
> subsystem uses to list virtfn entries, while real VF "numbers" are
> 1-based, see PCI spec which says:
> 
> "VFs are numbered starting with 1 so the first VF associated with PF M
> is VF M,1."

OK, since v2 using 1-based VF numbering has been positively reviewed by Chris
despite his comment on inconsistency with Linux PCI ABI then let's consider
the 0-based VF numbering used in that ABI as non-standard and merge v2 of this 
patch with some DOC improvements from v3 if there are no more comments.

Thanks,
Janusz

> 
> maybe we should update the wording to minimize any confusions?
> 
> Michal
> 
> > + *
> > + * Looks up a PCI interface of a DRM device or a VF PCI device of the DRM PF using libpciaccess.
> > + *
> > + * Returns:
> > + * The pci_device, NULL on any failures.
> > + */
> > +struct pci_device *__igt_device_get_pci_device(int fd, int vf_id)
> >  {
> >  	struct igt_pci_addr pci_addr;
> >  	struct pci_device *pci_dev;
> >  
> > -	if (igt_device_get_pci_addr(fd, &pci_addr)) {
> > +	if (igt_device_get_pci_addr(fd, vf_id, &pci_addr)) {
> >  		igt_warn("Unable to find device PCI address\n");
> >  		return NULL;
> >  	}
> > @@ -231,7 +252,7 @@ struct pci_device *igt_device_get_pci_device(int fd)
> >  {
> >  	struct pci_device *pci_dev;
> >  
> > -	pci_dev = __igt_device_get_pci_device(fd);
> > +	pci_dev = __igt_device_get_pci_device(fd, -1);
> >  	igt_require(pci_dev);
> >  
> >  	return pci_dev;
> > diff --git a/lib/igt_device.h b/lib/igt_device.h
> > index 278ba7a9b3..00da853e71 100644
> > --- a/lib/igt_device.h
> > +++ b/lib/igt_device.h
> > @@ -33,5 +33,6 @@ void igt_device_drop_master(int fd);
> >  
> >  int igt_device_get_card_index(int fd);
> >  struct pci_device *igt_device_get_pci_device(int fd);
> > +struct pci_device *__igt_device_get_pci_device(int fd, int vf_id);
> >  
> >  #endif /* __IGT_DEVICE_H__ */
>