Re: [PATCH] media: dvb_ringbuffer: Add memory barriers

From: Soeren Moch
Date: Tue May 10 2016 - 18:02:49 EST

Hi Mauro,

thanks for looking after this patch.

On 05/07/16 15:22, Mauro Carvalho Chehab wrote:
Hi Soeren,

Em Sun, 7 Feb 2016 20:22:36 +0100
Soeren Moch <smoch@xxxxxx> escreveu:

On 27.12.2015 21:41, Soeren Moch wrote:
Implement memory barriers according to Documentation/circular-buffers.txt:
- use smp_store_release() to update ringbuffer read/write pointers
- use smp_load_acquire() to load write pointer on reader side
- use ACCESS_ONCE() to load read pointer on writer side

This fixes data stream corruptions observed e.g. on an ARM Cortex-A9
quad core system with different types (PCI, USB) of DVB tuners.

Signed-off-by: Soeren Moch <smoch@xxxxxx>
Cc: stable@xxxxxxxxxxxxxxx # 3.14+

Mauro,

any news or comments on this?
Since this is a real fix for broken behaviour, can you pick this up, please?

The problem here is that I'm very reluctant to touch at the DVB core
without doing some tests myself, as things like locking can be
very sensible.

I agree. But this patch adds memory barriers (no locks) according to
Documentation/circular-buffers.txt. It should not be dangerous to
follow these guidelines.
Nevertheless, independent review and testing is always a good idea,
especially in core code.

I'll try to find some time to take a look on it for Kernel 4.8,

Thanks.

but I'd like to reproduce the bug locally.
>
Could you please provide me enough info to reproduce it (and
eventually some test MPEG-TS where you know this would happen)?

I used vdr with different types of DVB tuners on a TBS2910 board
(quad core ARM Cortex-A9, see arch/arm/boot/dts/imx6q-tbs2910.dts).
With more than one active cpu core I occasionally see data stream
corruptions in recorded ts streams. This is not the case for one
active cpu.

The problem occurs when dvb_dmxdev_buffer_write() and
dvb_dmxdev_buffer_read() are running simultaneously on different
cpu cores on architectures without strong write ordering (e.g.
on arm, not on x86). Here the write data pointer (written in
dvb_ringbuffer_write() ) can become visible to the other cpu core
(in dvb_ringbuffer_avail() ), before the actual ts packet data is
visible. dvb_ringbuffer_read_user() then can read old ts data,
although new ts data is already written into the ringbuffer
by the other cpu core.

With smp_store_release() and smp_load_acquire() the correct
write ordering is maintained, ts packet data is visible on the
reading cpu core before the updated write pointer.

I have two DekTek RF generators here, so I should be able to
play such TS and see what happens with and without the patch
on x86, arm32 and arm64.

On x86 you should not see any difference, since
smp_store_release() and smp_load_acquire() expand to
simple stores and loads there. On multi-core arm systems
you may see broken ts streams without patch, especially
when reading the dvr device very fast, so that the
dvb_ringbuffer is always close to empty.

Regards,
Soeren


Regards,
Mauro


Regards,
Soeren

---
Cc: Mauro Carvalho Chehab <mchehab@xxxxxxxxxxxxxxx>
Cc: linux-media@xxxxxxxxxxxxxxx
Cc: linux-kernel@xxxxxxxxxxxxxxx

Since smp_store_release() and smp_load_acquire() were introduced in linux-3.14,
a 3.14+ stable tag was added. Is it desired to apply a similar patch to older
stable kernels?
---
drivers/media/dvb-core/dvb_ringbuffer.c | 27 ++++++++++++++-------------
1 file changed, 14 insertions(+), 13 deletions(-)

diff --git a/drivers/media/dvb-core/dvb_ringbuffer.c b/drivers/media/dvb-core/dvb_ringbuffer.c
index 1100e98..58b5968 100644
--- a/drivers/media/dvb-core/dvb_ringbuffer.c
+++ b/drivers/media/dvb-core/dvb_ringbuffer.c
@@ -55,7 +55,7 @@ void dvb_ringbuffer_init(struct dvb_ringbuffer *rbuf, void *data, size_t len)

int dvb_ringbuffer_empty(struct dvb_ringbuffer *rbuf)
{
- return (rbuf->pread==rbuf->pwrite);
+ return (rbuf->pread == smp_load_acquire(&rbuf->pwrite));
}


@@ -64,7 +64,7 @@ ssize_t dvb_ringbuffer_free(struct dvb_ringbuffer *rbuf)
{
ssize_t free;

- free = rbuf->pread - rbuf->pwrite;
+ free = ACCESS_ONCE(rbuf->pread) - rbuf->pwrite;
if (free <= 0)
free += rbuf->size;
return free-1;
@@ -76,7 +76,7 @@ ssize_t dvb_ringbuffer_avail(struct dvb_ringbuffer *rbuf)
{
ssize_t avail;

- avail = rbuf->pwrite - rbuf->pread;
+ avail = smp_load_acquire(&rbuf->pwrite) - rbuf->pread;
if (avail < 0)
avail += rbuf->size;
return avail;
@@ -86,14 +86,15 @@ ssize_t dvb_ringbuffer_avail(struct dvb_ringbuffer *rbuf)

void dvb_ringbuffer_flush(struct dvb_ringbuffer *rbuf)
{
- rbuf->pread = rbuf->pwrite;
+ smp_store_release(&rbuf->pread, smp_load_acquire(&rbuf->pwrite));
rbuf->error = 0;
}
EXPORT_SYMBOL(dvb_ringbuffer_flush);

void dvb_ringbuffer_reset(struct dvb_ringbuffer *rbuf)
{
- rbuf->pread = rbuf->pwrite = 0;
+ smp_store_release(&rbuf->pread, 0);
+ smp_store_release(&rbuf->pwrite, 0);
rbuf->error = 0;
}

@@ -119,12 +120,12 @@ ssize_t dvb_ringbuffer_read_user(struct dvb_ringbuffer *rbuf, u8 __user *buf, si
return -EFAULT;
buf += split;
todo -= split;
- rbuf->pread = 0;
+ smp_store_release(&rbuf->pread, 0);
}
if (copy_to_user(buf, rbuf->data+rbuf->pread, todo))
return -EFAULT;

- rbuf->pread = (rbuf->pread + todo) % rbuf->size;
+ smp_store_release(&rbuf->pread, (rbuf->pread + todo) % rbuf->size);

return len;
}
@@ -139,11 +140,11 @@ void dvb_ringbuffer_read(struct dvb_ringbuffer *rbuf, u8 *buf, size_t len)
memcpy(buf, rbuf->data+rbuf->pread, split);
buf += split;
todo -= split;
- rbuf->pread = 0;
+ smp_store_release(&rbuf->pread, 0);
}
memcpy(buf, rbuf->data+rbuf->pread, todo);

- rbuf->pread = (rbuf->pread + todo) % rbuf->size;
+ smp_store_release(&rbuf->pread, (rbuf->pread + todo) % rbuf->size);
}


@@ -158,10 +159,10 @@ ssize_t dvb_ringbuffer_write(struct dvb_ringbuffer *rbuf, const u8 *buf, size_t
memcpy(rbuf->data+rbuf->pwrite, buf, split);
buf += split;
todo -= split;
- rbuf->pwrite = 0;
+ smp_store_release(&rbuf->pwrite, 0);
}
memcpy(rbuf->data+rbuf->pwrite, buf, todo);
- rbuf->pwrite = (rbuf->pwrite + todo) % rbuf->size;
+ smp_store_release(&rbuf->pwrite, (rbuf->pwrite + todo) % rbuf->size);

return len;
}
@@ -181,12 +182,12 @@ ssize_t dvb_ringbuffer_write_user(struct dvb_ringbuffer *rbuf,
return len - todo;
buf += split;
todo -= split;
- rbuf->pwrite = 0;
+ smp_store_release(&rbuf->pwrite, 0);
}
status = copy_from_user(rbuf->data+rbuf->pwrite, buf, todo);
if (status)
return len - todo;
- rbuf->pwrite = (rbuf->pwrite + todo) % rbuf->size;
+ smp_store_release(&rbuf->pwrite, (rbuf->pwrite + todo) % rbuf->size);

return len;
}


--
To unsubscribe from this list: send the line "unsubscribe linux-media" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html