Re: [PCIE device driver: tw686x] I have some problems with tw686x and I need help.

Re: [PCIE device driver: tw686x] I have some problems with tw686x and I need help.

[Krzysztof Hałasa]

Hello Singh,

I've added linux-media as well as the current TW686x driver maintainer
to Cc: list. Perhaps they will be better prepared to help you, the
driver differs much from what it was when I originally wrote it.

<zhaoxuegang@suntec.net> writes:

> First, I download source code from website
> https://github.com/torvalds/linux/tree/master/drivers/media/pci/tw686x,
> build the driver in my project. when i start the machine the machine
> printk the log
> "[ 5.557782] tw6869: PCI 0000:01:00.0, IRQ 170, MMIO 0xc8000000 (memcpy mode)
> [ 5.565010] tw686x 0000:01:00.0: enabling device (0000 -> 0002)".
> from the log we can see that pcie-bus read the ID is 6869, but out
> chip is tw6865. this is not our main problem.

The truth is I don't have a real TW6865, I simply assumed its PCI ID
would be 0x6865 because it worked this way for TW6864 and TW6869.
Could you please do a "lspci -vvn" on the machine with this card and
send me (with CC: linux-media@vger.kernel.org) this command's output
(you may omit other devices, I'm only interested in the TW chip
information). Perhaps the invalid identification can be fixed.

Thanks.

> After we booted the machine. then we lunch out application to display
> video that we capture from tw6865. But there are a lot of log , such
> as
> "[ 24.671551] tw686x 0000:01:00.0: video10: unexpected p-b buffer!
> [ 24.671561] tw686x 0000:01:00.0: video11: unexpected p-b buffer!
> [ 24.671566] tw686x 0000:01:00.0: video12: unexpected p-b buffer!
> [ 24.671570] tw686x 0000:01:00.0: video13: unexpected p-b buffer!"
> and the log print always.
> those log means tw6865 always reset dma channel.
> sometimes we can see the right picture, but the video moves slowly, about 2 frame per 5 seconds.
> after a few minute, the linux kernel crashed. I post the crash log
> below.

Unfortunately there is no meaningful stack dump. Perhaps the crashes
will go away when the "unexpected p-b buffer" issue is fixed so I
wouldn't worry too much about the dumps at the moment.

> Sometimes, after I booted the machine, and lunched the application. To
> the begin, we can see a few frame displaying on screen. after a few
> seconds, the monitor are freezed. and I use "cat /proc/interrupts" to
> look up tw6865's interrput. and we find that there is no tw6865
> interrupt any more. after a few minutes, most probely the machine will
> crash and reboot(crash log is same with the before which I have
> posted).
> we use command "dmesg" to look up log. we found that there ever
> happend "tw686x 0000:01:00.0: video10: FIFO error" or "we found that
> there ever happend" frequently.

This could be caused by the invalid identification (and handling) of
the chip.

OTOH, the "memcpy" mode of this driver may be too slow on your machine
(ARM usually doesn't have hardware-coherent cache memory and certain
cache sync operations are apparently very slow). You may want to try the
DMA mode instead (modprobe tw686x dma_mode=contig or sg). Be aware that
the DMA modes aren't as flexible as memcpy mode, the frame formats are
more limited.

I don't have experience with any ARM64, though.

Also I think the current code will OOPS on any video initialization
error (which will most probably happen on ARM in contig mode because
ARM machines don't allow that much continuous allocation). At least this
can be easily fixed, I'll attach the patch.
-- 
Krzysztof Halasa

Industrial Research Institute for Automation and Measurements PIAP
Al. Jerozolimskie 202, 02-486 Warsaw, Poland

[PATCH] TW686x: Fix OOPS on buffer alloc failure

[Krzysztof Hałasa]

Signed-off-by: Krzysztof Hałasa <khalasa@piap.pl>

diff --git a/drivers/media/pci/tw686x/tw686x-video.c b/drivers/media/pci/tw686x/tw686x-video.c
index c3fafa9..d637f47 100644
--- a/drivers/media/pci/tw686x/tw686x-video.c
+++ b/drivers/media/pci/tw686x/tw686x-video.c
@@ -1190,6 +1190,13 @@ int tw686x_video_init(struct tw686x_dev *dev)
 			return err;
 	}
 
+	/* Initialize vc->dev and vc->ch for the error path first */
+	for (ch = 0; ch < max_channels(dev); ch++) {
+		struct tw686x_video_channel *vc = &dev->video_channels[ch];
+		vc->dev = dev;
+		vc->ch = ch;
+	}
+
 	for (ch = 0; ch < max_channels(dev); ch++) {
 		struct tw686x_video_channel *vc = &dev->video_channels[ch];
 		struct video_device *vdev;
@@ -1198,9 +1205,6 @@ int tw686x_video_init(struct tw686x_dev *dev)
 		spin_lock_init(&vc->qlock);
 		INIT_LIST_HEAD(&vc->vidq_queued);
 
-		vc->dev = dev;
-		vc->ch = ch;
-
 		/* default settings */
 		err = tw686x_set_standard(vc, V4L2_STD_NTSC);
 		if (err)

Re: [PATCH] TW686x: Fix OOPS on buffer alloc failure

[Ezequiel Garcia]

Hi Krzysztof,

On 10 May 2017 at 06:51, Krzysztof Hałasa <khalasa@piap.pl> wrote:
> Signed-off-by: Krzysztof Hałasa <khalasa@piap.pl>
>
> diff --git a/drivers/media/pci/tw686x/tw686x-video.c b/drivers/media/pci/tw686x/tw686x-video.c
> index c3fafa9..d637f47 100644
> --- a/drivers/media/pci/tw686x/tw686x-video.c
> +++ b/drivers/media/pci/tw686x/tw686x-video.c
> @@ -1190,6 +1190,13 @@ int tw686x_video_init(struct tw686x_dev *dev)
>                         return err;
>         }
>
> +       /* Initialize vc->dev and vc->ch for the error path first */
> +       for (ch = 0; ch < max_channels(dev); ch++) {
> +               struct tw686x_video_channel *vc = &dev->video_channels[ch];
> +               vc->dev = dev;
> +               vc->ch = ch;
> +       }
> +

I'm not sure where is the oops this commit fixes, care to explain it to me?

>         for (ch = 0; ch < max_channels(dev); ch++) {
>                 struct tw686x_video_channel *vc = &dev->video_channels[ch];
>                 struct video_device *vdev;
> @@ -1198,9 +1205,6 @@ int tw686x_video_init(struct tw686x_dev *dev)
>                 spin_lock_init(&vc->qlock);
>                 INIT_LIST_HEAD(&vc->vidq_queued);
>
> -               vc->dev = dev;
> -               vc->ch = ch;
> -
>                 /* default settings */
>                 err = tw686x_set_standard(vc, V4L2_STD_NTSC);
>                 if (err)

Thanks,
-- 
Ezequiel García, VanguardiaSur
www.vanguardiasur.com.ar

Re: [PATCH] TW686x: Fix OOPS on buffer alloc failure

[Krzysztof Hałasa]

Ezequiel Garcia <ezequiel@vanguardiasur.com.ar> writes:

>> +       /* Initialize vc->dev and vc->ch for the error path first */
>> +       for (ch = 0; ch < max_channels(dev); ch++) {
>> +               struct tw686x_video_channel *vc = &dev->video_channels[ch];
>> +               vc->dev = dev;
>> +               vc->ch = ch;
>> +       }
>> +
>
> I'm not sure where is the oops this commit fixes, care to explain it to me?

The error path apparently calls tw686x_video_free() which requires
vc->dev to be initialized. Now, the vc->dev is set for the channel being
currently initialized (unsuccesfully), but not for ones which haven't
been initialized yet. tw686x_video_free() iterates over the whole set.

It seems it also happens in "memcpy" mode. I didn't test it before since
on my ARMv7 "memcpy" mode is unusable, it's way too slow. Also, does the
driver attempt to use consistent memory for entire buffers in this mode?
This may work on i686/x86_64 because the caches are coherent by design
and there is no difference between consistent and non-consistent RAM
(if one isn't using SWIOTLB etc).

tw6869: PCI 0000:07:00.0, IRQ 24, MMIO 0x1100000 (memcpy mode)
tw686x 0000:07:00.0: enabling device (0140 -> 0142)
tw686x 0000:07:00.0: dma0: unable to allocate P-buffer
Unable to handle kernel NULL pointer dereference at virtual address 00000000
PC is at _raw_spin_lock_irqsave+0x10/0x4c
LR is at tw686x_memcpy_dma_free+0x1c/0x124
pc : [<805a8b14>]    lr : [<7f04a3c0>]    psr: 20010093
sp : be915c80  ip : 00000000  fp : bea1b000
r10: 00000000  r9 : fffffff4  r8 : 0000b000
r7 : 00000000  r6 : 000003f0  r5 : 00000000  r4 : bf0e21f8
r3 : 7f04a3a4  r2 : 00000000  r1 : 00000000  r0 : 20010013
Flags: nzCv  IRQs off  FIQs on  Mode SVC_32  ISA ARM  Segment none
Control: 10c5387d  Table: 4e91804a  DAC: 00000051
Process udevd (pid: 88, stack limit = 0xbe914210)
(_raw_spin_lock_irqsave) from (tw686x_memcpy_dma_free+0x1c/0x124)
(tw686x_memcpy_dma_free) from (tw686x_video_free+0x50/0x78)
(tw686x_video_free) from (tw686x_video_init+0x478/0x5e8)
(tw686x_video_init) from (tw686x_probe+0x36c/0x3fc)
(tw686x_probe) from (pci_device_probe+0x88/0xf4)
(pci_device_probe) from (driver_probe_device+0x238/0x2d8)
(driver_probe_device) from (__driver_attach+0xac/0xb0)
(__driver_attach) from (bus_for_each_dev+0x6c/0xa0)
(bus_for_each_dev) from (bus_add_driver+0x1a0/0x218)
(bus_add_driver) from (driver_register+0x78/0xf8)
(driver_register) from (do_one_initcall+0x40/0x168)
(do_one_initcall) from (do_init_module+0x60/0x3a4)
(do_init_module) from (load_module+0x1c90/0x20e4)
(load_module) from (SyS_finit_module+0x8c/0x9c)
(SyS_finit_module) from (ret_fast_syscall+0x0/0x3c)
Code: e1a02000 e10f0000 f10c0080 f592f000 (e1923f9f)


With the patch:
tw6869: PCI 0000:07:00.0, IRQ 24, MMIO 0x1100000 (memcpy mode)
tw686x 0000:07:00.0: enabling device (0140 -> 0142)
tw686x 0000:07:00.0: dma0: unable to allocate P-buffer
tw686x 0000:07:00.0: can't register video
tw686x: probe of 0000:07:00.0 failed with error -12
--
Krzysztof Halasa

Industrial Research Institute for Automation and Measurements PIAP
Al. Jerozolimskie 202, 02-486 Warsaw, Poland

Re: [PATCH] TW686x: Fix OOPS on buffer alloc failure

[Ezequiel Garcia]

On 11 May 2017 at 04:41, Krzysztof Hałasa <khalasa@piap.pl> wrote:
> Ezequiel Garcia <ezequiel@vanguardiasur.com.ar> writes:
>
>>> +       /* Initialize vc->dev and vc->ch for the error path first */
>>> +       for (ch = 0; ch < max_channels(dev); ch++) {
>>> +               struct tw686x_video_channel *vc = &dev->video_channels[ch];
>>> +               vc->dev = dev;
>>> +               vc->ch = ch;
>>> +       }
>>> +
>>
>> I'm not sure where is the oops this commit fixes, care to explain it to me?
>
> The error path apparently calls tw686x_video_free() which requires
> vc->dev to be initialized. Now, the vc->dev is set for the channel being
> currently initialized (unsuccesfully), but not for ones which haven't
> been initialized yet. tw686x_video_free() iterates over the whole set.
>
> It seems it also happens in "memcpy" mode. I didn't test it before since
> on my ARMv7 "memcpy" mode is unusable, it's way too slow. Also, does the
> driver attempt to use consistent memory for entire buffers in this mode?

Yes, memcpy mode allocates a couple P and B dma-coherent bounce
buffers for each channel.

> This may work on i686/x86_64 because the caches are coherent by design
> and there is no difference between consistent and non-consistent RAM
> (if one isn't using SWIOTLB etc).
>
> tw6869: PCI 0000:07:00.0, IRQ 24, MMIO 0x1100000 (memcpy mode)
> tw686x 0000:07:00.0: enabling device (0140 -> 0142)
> tw686x 0000:07:00.0: dma0: unable to allocate P-buffer
> Unable to handle kernel NULL pointer dereference at virtual address 00000000
> PC is at _raw_spin_lock_irqsave+0x10/0x4c
> LR is at tw686x_memcpy_dma_free+0x1c/0x124
> pc : [<805a8b14>]    lr : [<7f04a3c0>]    psr: 20010093
> sp : be915c80  ip : 00000000  fp : bea1b000
> r10: 00000000  r9 : fffffff4  r8 : 0000b000
> r7 : 00000000  r6 : 000003f0  r5 : 00000000  r4 : bf0e21f8
> r3 : 7f04a3a4  r2 : 00000000  r1 : 00000000  r0 : 20010013
> Flags: nzCv  IRQs off  FIQs on  Mode SVC_32  ISA ARM  Segment none
> Control: 10c5387d  Table: 4e91804a  DAC: 00000051
> Process udevd (pid: 88, stack limit = 0xbe914210)
> (_raw_spin_lock_irqsave) from (tw686x_memcpy_dma_free+0x1c/0x124)
> (tw686x_memcpy_dma_free) from (tw686x_video_free+0x50/0x78)
> (tw686x_video_free) from (tw686x_video_init+0x478/0x5e8)
> (tw686x_video_init) from (tw686x_probe+0x36c/0x3fc)
> (tw686x_probe) from (pci_device_probe+0x88/0xf4)
> (pci_device_probe) from (driver_probe_device+0x238/0x2d8)
> (driver_probe_device) from (__driver_attach+0xac/0xb0)
> (__driver_attach) from (bus_for_each_dev+0x6c/0xa0)
> (bus_for_each_dev) from (bus_add_driver+0x1a0/0x218)
> (bus_add_driver) from (driver_register+0x78/0xf8)
> (driver_register) from (do_one_initcall+0x40/0x168)
> (do_one_initcall) from (do_init_module+0x60/0x3a4)
> (do_init_module) from (load_module+0x1c90/0x20e4)
> (load_module) from (SyS_finit_module+0x8c/0x9c)
> (SyS_finit_module) from (ret_fast_syscall+0x0/0x3c)
> Code: e1a02000 e10f0000 f10c0080 f592f000 (e1923f9f)
>
>
> With the patch:
> tw6869: PCI 0000:07:00.0, IRQ 24, MMIO 0x1100000 (memcpy mode)
> tw686x 0000:07:00.0: enabling device (0140 -> 0142)
> tw686x 0000:07:00.0: dma0: unable to allocate P-buffer
> tw686x 0000:07:00.0: can't register video
> tw686x: probe of 0000:07:00.0 failed with error -12

Oh, I see. Thanks for the details!

How about this one (untested) ?

diff --git a/drivers/media/pci/tw686x/tw686x-video.c
b/drivers/media/pci/tw686x/tw686x-video.c
index c3fafa97b2d0..77b8d2dbd995 100644
--- a/drivers/media/pci/tw686x/tw686x-video.c
+++ b/drivers/media/pci/tw686x/tw686x-video.c
@@ -86,6 +86,9 @@ static void tw686x_memcpy_dma_free(struct
tw686x_video_channel *vc,
        struct pci_dev *pci_dev;
        unsigned long flags;

+       /* Make sure this channel is initialized */
+       if (!dev)
+               return;
        /* Check device presence. Shouldn't really happen! */
        spin_lock_irqsave(&dev->lock, flags);
        pci_dev = dev->pci_dev;

Also, when submitting a patch please make sure you write a commit
log. Only trivial patches can have empty commit logs.

The description you made in your last mail is perfect,
including the oops details, the backtrace and
the "With the patch" note. Doing a `git log --grep="fix.*oops"`
might give you some ideas on commit logs.

The subject of the patch should be:

"tw686x: Fix oops on buffer alloc failure"

If you `git log` on tw686x, you'll notice all the
commits use "tw686x" prefix.

When submitting fixes, it's useful to add a "Fixes" tag.
See https://static.lwn.net/kerneldoc/process/submitting-patches.html
for details.
-- 
Ezequiel García, VanguardiaSur
www.vanguardiasur.com.ar

Re: [PATCH] TW686x: Fix OOPS on buffer alloc failure

[Krzysztof Hałasa]

Ezequiel Garcia <ezequiel@vanguardiasur.com.ar> writes:

> How about this one (untested) ?
>
> diff --git a/drivers/media/pci/tw686x/tw686x-video.c
> b/drivers/media/pci/tw686x/tw686x-video.c
> index c3fafa97b2d0..77b8d2dbd995 100644
> --- a/drivers/media/pci/tw686x/tw686x-video.c
> +++ b/drivers/media/pci/tw686x/tw686x-video.c
> @@ -86,6 +86,9 @@ static void tw686x_memcpy_dma_free(struct
> tw686x_video_channel *vc,
>         struct pci_dev *pci_dev;
>         unsigned long flags;
>
> +       /* Make sure this channel is initialized */
> +       if (!dev)
> +               return;

Whatever you wish. Just make sure it doesn't bomb out by default, when
one happens to have such a card in his or her machine.
-- 
Krzysztof Halasa

Industrial Research Institute for Automation and Measurements PIAP
Al. Jerozolimskie 202, 02-486 Warsaw, Poland

Re: [PATCH] TW686x: Fix OOPS on buffer alloc failure

[Krzysztof Hałasa]

Hans Verkuil <hverkuil@xs4all.nl> writes:

> Any progress on this? I gather I can expect a new patch from someone?

Well, the issue is trivial and very easy to test, though not present
on common x86 hw. That patch I've sent fixes it, but I'm not the one who
decides.
-- 
Krzysztof Halasa

Industrial Research Institute for Automation and Measurements PIAP
Al. Jerozolimskie 202, 02-486 Warsaw, Poland

Re: [PATCH] TW686x: Fix OOPS on buffer alloc failure

[Ezequiel Garcia]

On 23 June 2017 at 05:18, Krzysztof Hałasa <khalasa@piap.pl> wrote:
> Hans Verkuil <hverkuil@xs4all.nl> writes:
>
>> Any progress on this? I gather I can expect a new patch from someone?
>
> Well, the issue is trivial and very easy to test, though not present
> on common x86 hw. That patch I've sent fixes it, but I'm not the one who
> decides.

If you can re-submit your patch addressing all the comments, I'd be happy
to Ack it.

As it stands, with the wrong subject style and without a commit log,
it's a NAK on my side.
-- 
Ezequiel García, VanguardiaSur
www.vanguardiasur.com.ar

Re: [PATCH] TW686x: Fix OOPS on buffer alloc failure

[Krzysztof Hałasa]

"zhaoxuegang" <zhaoxuegang@suntec.net> writes:

> In my opinion, the oops occur to tw686x_free_dma(struct tw686x_video_channel *vc, unsigned int pb).
> In function tw686x_video_init, if coherent-DMA is not enough, tw686x_alloc_dma will failed.
> Then tw686x_video_free will be called. but some channel's dev is null（in other words, vc->dev is null）

Exactly.
-- 
Krzysztof Halasa

Industrial Research Institute for Automation and Measurements PIAP
Al. Jerozolimskie 202, 02-486 Warsaw, Poland