regression caused by #682

[XinfengZhang]

there are one regression caused by #682 , current finding:
libva info: VA-API version 1.19.0
libva error: vaGetDriverNames() failed with operation failed
libva error: va_new_opendriver:715: Out of memory

from app side
vaInitialize, err=0x2

xdpyinfo | grep DRI
Windows-DRI

it maybe not a libva issue, maybe just this change expose potential issues.
I will paste more information after some investigation

[evelikov]

Seems unlikely to be caused by the PR but if you can provide meaningful details I can take a look.

[XinfengZhang]

just have a quick check:
the num_drivers is initialized as 20
then
after vaStatus = pDisplayContext->vaGetDriverNames(pDisplayContext, drivers, &num_drivers);
it is 4
looks some error happened in vaGetDriverNames,

after const char * const *va_drivers = map[i].va_driver;
https://github.com/intel/libva/blob/master/va/drm/va_drm_utils.c#L90

the va_drivers include 4 strings,
"iHD" "i965" "pvrsrvkm" "pvr"
strange part is map[0].va_driver still only have first two;

[XinfengZhang]

please ignore previous comments, previous debug info looks a env issue.

let re-start the issue debug:

the issue actually is:
there are two vaInitialize call,
1st success.
2nd failed with
libva error: vaGetDriverNames() failed with operation failed
libva error: va_new_opendriver:715: Out of memory

the second time,
va_drm_authenticate failed with
the ctx->display_type = VA_DISPLAY_DRM
drmGetMagic get magic num : magic = 2

so, the call failed , and the num_drivers is still 20
and there are no env set.

after revert #682, 2 initialization both success.

[XinfengZhang]

and before #682, the ctx->display_type also is 48 VA_DISPLAY_DRM
but the magic = 0x7fff;
the authentication still failed
but there are such statement
if (vaStatus != VA_STATUS_SUCCESS) {
num_candidates = 1;
}

then , it still work....

[evelikov]

Not sure I follow - the formatting is all over the place making it almost impossible to understand. Do you have a simple reproducer to share? Ideally we should add it to the libva test suite.

The latter living in libva-utils, which makes zero sense and should be moved, but I digress.

[XinfengZhang]

there are 2 problems
1st. problem:
could be reproduced with "vainfo --display x11"
there are no DRI2 & DRI3 support, then both va_DRI3_GetDriverNames & va_DRI2_GetDriverNames will failed.
then
https://github.com/intel/libva/blob/master/va/va.c#L686 , error message shows. but the num_drivers is still 20
then it will failed again in
https://github.com/intel/libva/blob/master/va/va.c#L715

I suppose we should set the num_drivers to be zero , if https://github.com/intel/libva/blob/master/va/va.c#L685 failed.

[XinfengZhang]

2nd problem:
could be reproduced with "vainfo --display drm"
with this patch:

diff --git a/va/drm/va_drm_utils.c b/va/drm/va_drm_utils.c
index 5f378d5..7846a6c 100644
--- a/va/drm/va_drm_utils.c
+++ b/va/drm/va_drm_utils.c
@@ -91,13 +91,18 @@ VA_DRM_GetDriverNames(VADriverContextP ctx, char **drivers, unsigned *num_driver

         while (va_drivers[count]) {
             if (count < MAX_NAMES && count < *num_drivers)

•            {
                 drivers[count] = strdup(va_drivers[count]);
  

•                printf("driver candidate %d, %s, %s\n", count, drivers[count], va_drivers[count]);
  

•            }
  

•            printf("driver candidate1 %d, %s, %s\n", count, drivers[count], va_drivers[count]);
             count++;
         }
         break;
     }
  

  }

using above patch: we could see:
driver candidate 0, iHD, iHD
driver candidate1 0, iHD, iHD
driver candidate 1, i965, i965
driver candidate1 1, i965, i965
driver candidate1 2, (null), pvrsrvkm
driver candidate1 3, (null), pvr

a simple fix is move the count++ under the protection of if (count < MAX_NAMES

[XinfengZhang]

about the issue I mentioned in previous comments:

and before #682, the ctx->display_type also is 48 VA_DISPLAY_DRM but the magic = 0x7fff; the authentication still failed but there are such statement if (vaStatus != VA_STATUS_SUCCESS) { num_candidates = 1; }

then , it still work....

it is very hard to reproduce, because it actually about such case:
drm authentication failed in va_DisplayContextConnect
https://github.com/intel/libva/blob/master/va/drm/va_drm.c#L62

and then initialize failed

but before #682 , even the authentication failed, the code still could run under this protection:
if (vaStatus != VA_STATUS_SUCCESS) { num_candidates = 1; }

so, it will still get the first candidate name.

it maybe is a application issue, need to understand why the authentication failed.
I guess in below function:

bool
va_drm_authenticate(int fd, uint32_t magic)
{
/* XXX: try to authenticate through Wayland, etc. */
#ifdef HAVE_VA_X11
if (va_drm_authenticate_x11(fd, magic))
return true;
#endif

/* Default: root + master privs are needed for the following call */
return drmAuthMagic(fd, magic) == 0;

}

suppose it is because of the second vaInitialize.

[XinfengZhang]

a simple sample to reproduce the issue, mentioned here #722 (comment) :

#include <unistd.h>
#include <termios.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <fcntl.h>
#include <va/va.h>
#include <va/va_drm.h>
void main()
{

int fd0,fd1;
VADisplay va_dpy0, va_dpy1;
int major_version, minor_version;
fd0 = open("/dev/dri/card0", O_RDWR);
fd1 = open("/dev/dri/card0", O_RDWR);

va_dpy0 = vaGetDisplayDRM(fd0);
vaInitialize(va_dpy0, &major_version, &minor_version);

va_dpy1 = vaGetDisplayDRM(fd1);
vaInitialize(va_dpy1, &major_version, &minor_version);

vaTerminate(va_dpy0);
vaTerminate(va_dpy1);

close(fd0);
close(fd1);
}

second vaInitialize will cause drmAuthMagic failure with errno=13

[evelikov]

Thanks for the example. Off the top of my head this behaviour is normal and expected.

In more detail: the first client (regardless if that's libva or Xorg or Weston or...) is always master and authenticated. Since the first client has not a) closed the fd or b) explicitly dropped the master - any other client who tries to issue ioctls that requires master will fail with EACCESS aka 13.

This is part of the DRM fundamentals - trying to relax that this will open a massive security hole. I would strongly recommend using the render node, unless you absolutely need to use the card node.

What program/project is that? If open-source I can send a fix their way.

[XinfengZhang]

the 3rd issue should be expected. I will sync with that app owners
but I suppose 1st and 2nd issue still need to be addressed.
even these 2 issues does not impact use case.

[evelikov]

1st. problem:
could be reproduced with "vainfo --display x11"
there are no DRI2 & DRI3 support, then both va_DRI3_GetDriverNames & va_DRI2_GetDriverNames will failed.

What do you mean there are no DRI2 & DRI3 support? Can you provide an example of what hardware/software I need?

[evelikov]

2nd problem: could be reproduced with "vainfo --display drm" with this patch:

The automatic markdown conversion has rendered the patch useless. Can you wrap it in backticks as per the documentation