Re: [PATCH v3] Fix loading of module radeonfb on PowerMac

From: Bartlomiej Zolnierkiewicz
Date: Wed Jan 31 2018 - 06:58:03 EST


On Tuesday, January 30, 2018 02:14:10 PM Mathieu Malaterre wrote:
> Bartlomiej,
>
> On Wed, Jan 3, 2018 at 3:47 PM, Bartlomiej Zolnierkiewicz
> <b.zolnierkie@xxxxxxxxxxx> wrote:
> >
> > On Thursday, December 21, 2017 11:07:56 PM Mathieu Malaterre wrote:
> >> When the linux kernel is build with (typical kernel ship with Debian
> >> installer):
> >>
> >> CONFIG_FB_OF=y
> >> CONFIG_VT_HW_CONSOLE_BINDING=y
> >> CONFIG_FB_RADEON=m
> >>
> >> The offb driver takes precedence over module radeonfb. It is then
> >> impossible to load the module, error reported is:
> >>
> >> [ 96.551486] radeonfb 0000:00:10.0: enabling device (0006 -> 0007)
> >> [ 96.551526] radeonfb 0000:00:10.0: BAR 0: can't reserve [mem 0x98000000-0x9fffffff pref]
> >> [ 96.551531] radeonfb (0000:00:10.0): cannot request region 0.
> >> [ 96.551545] radeonfb: probe of 0000:00:10.0 failed with error -16
> >>
> >> This patch reproduce the behavior of the module radeon, so as to make it
> >> possible to load radeonfb when offb is first loaded.
> >>
> >> It should be noticed that `offb_destroy` is never called which explain the
> >> need to skip error detection on the radeon side.
> >
> > This still needs to be explained more, from my last mail:
> >
> > "The last put_fb_info() on fb_info should call ->fb_destroy
> > (offb_destroy in our case) and remove_conflicting_framebuffers()
> > is calling put_fb_info() so there is some extra reference on
> > fb_info somewhere preventing it from going away.
> >
> > Please look into fixing this."
>
> I am not familiar with the fb stuff internals but here is what I see:
>
> # modprobe radeonfb
>
> leads to:
>
> [ 52.058546] bus: 'pci': add driver radeonfb
> [ 52.058588] bus: 'pci': driver_probe_device: matched device
> 0000:00:10.0 with driver radeonfb
> [ 52.058595] bus: 'pci': really_probe: probing driver radeonfb with
> device 0000:00:10.0
> [ 52.058608] devices_kset: Moving 0000:00:10.0 to end of list
> [ 52.058613] radeonfb_pci_register BEGIN
> [ 52.058634] radeonfb 0000:00:10.0: enabling device (0006 -> 0007)
> <at this point radeon_kick_out_firmware_fb is called>
> [ 52.058666] checking generic (9c008000 96000) vs hw (98000000 8000000)
> [ 52.058667] fb: switching to radeonfb from OFfb ATY,RockHo
> [ 52.058844] Console: switching to colour dummy device 80x25
> [ 52.058860] device: 'fb0': device_unregister
> [ 52.058956] PM: Removing info for No Bus:fb0
> [ 52.059014] device: 'fb0': device_create_release
> <a call to do_unregister_framebuffer is done>
> <put_fb_info is done with a count=2 and dev=NULL>
> [ 52.059048] device: 'vtcon1': device_unregister
> [ 52.059076] PM: Removing info for No Bus:vtcon1
> [ 52.059091] device: 'vtcon1': device_create_release
> [ 52.059107] radeonfb 0000:00:10.0: BAR 0: can't reserve [mem
> 0x98000000-0x9fffffff pref]
> [ 52.256151] aper_base: 98000000 MC_FB_LOC to: 9bff9800, MC_AGP_LOC
> to: ffffa000
> [ 52.256157] radeonfb (0000:00:10.0): Found 32768k of DDR 64 bits
> wide videoram
>
> I can confirm that offb_destroy is never called (not sure exactly
> why), but in any case the call to radeon_kick_out_firmware_fb happen
> much earlier, at least before the put_fb_info.

It is okay, put_fb_info() is called indirectly by radeon_kick_out_firmware_fb()

radeon_kick_out_firmware_fb()
remove_conflicting_framebuffers()
do_remove_conflicting_framebuffers()
do_unregister_framebuffer()
put_fb_info()

offb_destroy() is not called because there is an extra reference on old
fb_info (->count == 2):

static void put_fb_info(struct fb_info *fb_info)
{
if (!atomic_dec_and_test(&fb_info->count))
return;
if (fb_info->fbops->fb_destroy)
fb_info->fbops->fb_destroy(fb_info);
}

The question is why there is an extra reference, probably user-space
is still holding the fb_info reference obtained in fb_open() call and
fb_release() is never called. Besides not calling fbops->fb_destroy()
this also causes missing call of fbops->fb_release() (in fb_release())
which some fb drivers are implementing (but not offb.c).

> Could you describe a bit more the chain of calls you were thinking of ?

Please add WARN_ON(1) to get_fb_info() and put_fb_info() so we can check
from the stacktrace if it is actually fb_open() that holds the extra
old fb_info reference.

drivers/video/fbdev/core/fbmem.c:

static struct fb_info *get_fb_info(unsigned int idx)
{
struct fb_info *fb_info;

if (idx >= FB_MAX)
return ERR_PTR(-ENODEV);

mutex_lock(&registration_lock);
fb_info = registered_fb[idx];
if (fb_info)
atomic_inc(&fb_info->count);

if (fb_info)
WARN_ON(1);

mutex_unlock(&registration_lock);

return fb_info;
}

static void put_fb_info(struct fb_info *fb_info)
{
WARN_ON(1);

if (!atomic_dec_and_test(&fb_info->count))
return;
if (fb_info->fbops->fb_destroy)
fb_info->fbops->fb_destroy(fb_info);
}

> >> Signed-off-by: Mathieu Malaterre <malat@xxxxxxxxxx>
> >> Link: https://bugs.debian.org/826629#57
> >> Link: https://bugzilla.kernel.org/show_bug.cgi?id=119741
> >> Suggested-by: Lennart Sorensen <lsorense@xxxxxxxxxxxxxxxxxxx>
> >> ---
> >> v2: Only fails when CONFIG_PCC is not set
> >> v3: Only fails when CONFIG_FB_OF is not set, CONFIG_PCC was too broad. Since the conflicts in region is due to OFfb explicitly refers to it.
> >
> > It seems that there may still be configurations when this is
> > incorrect -> when offb drives primary (non-radeon) card and radeonfb
> > drives secondary (radeon) card..
> >
> >> drivers/video/fbdev/aty/radeon_base.c | 26 ++++++++++++++++++++++++++
> >> 1 file changed, 26 insertions(+)
> >>
> >> diff --git a/drivers/video/fbdev/aty/radeon_base.c b/drivers/video/fbdev/aty/radeon_base.c
> >> index 4d77daeecf99..221879196531 100644
> >> --- a/drivers/video/fbdev/aty/radeon_base.c
> >> +++ b/drivers/video/fbdev/aty/radeon_base.c
> >> @@ -2259,6 +2259,22 @@ static const struct bin_attribute edid2_attr = {
> >> .read = radeon_show_edid2,
> >> };
> >>
> >> +static int radeon_kick_out_firmware_fb(struct pci_dev *pdev)
> >> +{
> >> + struct apertures_struct *ap;
> >> +
> >> + ap = alloc_apertures(1);
> >> + if (!ap)
> >> + return -ENOMEM;
> >> +
> >> + ap->ranges[0].base = pci_resource_start(pdev, 0);
> >> + ap->ranges[0].size = pci_resource_len(pdev, 0);
> >> +
> >> + remove_conflicting_framebuffers(ap, KBUILD_MODNAME, false);
> >> + kfree(ap);
> >> +
> >> + return 0;
> >> +}
> >>
> >> static int radeonfb_pci_register(struct pci_dev *pdev,
> >> const struct pci_device_id *ent)
> >> @@ -2312,19 +2328,27 @@ static int radeonfb_pci_register(struct pci_dev *pdev,
> >> rinfo->fb_base_phys = pci_resource_start (pdev, 0);
> >> rinfo->mmio_base_phys = pci_resource_start (pdev, 2);
> >>
> >> + ret = radeon_kick_out_firmware_fb(pdev);
> >> + if (ret)
> >> + return ret;
> >> +
> >> /* request the mem regions */
> >> ret = pci_request_region(pdev, 0, "radeonfb framebuffer");
> >> if (ret < 0) {
> >> +#ifndef CONFIG_FB_OF
> >> printk( KERN_ERR "radeonfb (%s): cannot request region 0.\n",
> >> pci_name(rinfo->pdev));
> >> goto err_release_fb;
> >> +#endif
> >> }
> >>
> >> ret = pci_request_region(pdev, 2, "radeonfb mmio");
> >> if (ret < 0) {
> >> +#ifndef CONFIG_FB_OF
> >> printk( KERN_ERR "radeonfb (%s): cannot request region 2.\n",
> >> pci_name(rinfo->pdev));
> >> goto err_release_pci0;
> >> +#endif
> >> }
> >>
> >> /* map the regions */
> >> @@ -2509,10 +2533,12 @@ static int radeonfb_pci_register(struct pci_dev *pdev,
> >> iounmap(rinfo->mmio_base);
> >> err_release_pci2:
> >> pci_release_region(pdev, 2);
> >> +#ifndef CONFIG_FB_OF
> >> err_release_pci0:
> >> pci_release_region(pdev, 0);
> >> err_release_fb:
> >> framebuffer_release(info);
> >> +#endif
> >> err_disable:
> >> err_out:
> >> return ret;
> >
> > Best regards,
> > --
> > Bartlomiej Zolnierkiewicz
> > Samsung R&D Institute Poland
> > Samsung Electronics
> >
>
> Thanks,
> -M

Best regards,
--
Bartlomiej Zolnierkiewicz
Samsung R&D Institute Poland
Samsung Electronics