Re: [PATCH] drm/msm: use proper memory barriers for updating tail/head

[Daniel Vetter]

On Tue, Oct 3, 2017 at 2:59 PM, Rob Clark  wrote:
> On Tue, Oct 3, 2017 at 4:08 AM, Daniel Vetter  wrote:
>> On Mon, Oct 02, 2017 at 10:37:11AM -0400, Rob Clark wrote:
>>> Fixes intermittent corruption of cmdstream dump.
>>>
>>> Signed-off-by: Rob Clark 
>>
>> Rule of memory barriers: They always need to come in pairs, and you should
>> put a comment to each one explaining where the other side is. I.e. for all
>> the store_release ones you add here you need to put a few read_acquire
>> ones somewhere else, and then explain how they link.
>>
>> One side only doesn't really work with barriers - except when you're lucky
>> and on x86 :-)
>
> fwiw, this is straight out of Documentation/circular-buffers.txt
>
> I guess it could read_acquire when checking the avail/free space on
> the consumer/producer side.. otoh if we see an old value there it is
> harmless (ie. we just fall into a wait_event()).  I assume that is why
> the documentation didn't bother with a read_acquire.

smp_load_acquire is the right name, and it's right there. In the
example at least :-)
-Daniel

>
> BR,
> -R
>
>> -Daniel
>>> ---
>>>  drivers/gpu/drm/msm/msm_rd.c | 4 ++--
>>>  1 file changed, 2 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/drivers/gpu/drm/msm/msm_rd.c b/drivers/gpu/drm/msm/msm_rd.c
>>> index 5c087a4228ff..918c65e19964 100644
>>> --- a/drivers/gpu/drm/msm/msm_rd.c
>>> +++ b/drivers/gpu/drm/msm/msm_rd.c
>>> @@ -120,7 +120,7 @@ static void rd_write(struct msm_rd_state *rd, const 
>>> void *buf, int sz)
>>>   n = min(sz, circ_space_to_end(>fifo));
>>>   memcpy(fptr, ptr, n);
>>>
>>> - fifo->head = (fifo->head + n) & (BUF_SZ - 1);
>>> + smp_store_release(>head, (fifo->head + n) & (BUF_SZ - 
>>> 1));
>>>   sz  -= n;
>>>   ptr += n;
>>>
>>> @@ -157,7 +157,7 @@ static ssize_t rd_read(struct file *file, char __user 
>>> *buf,
>>>   goto out;
>>>   }
>>>
>>> - fifo->tail = (fifo->tail + n) & (BUF_SZ - 1);
>>> + smp_store_release(>tail, (fifo->tail + n) & (BUF_SZ - 1));
>>>   *ppos += n;
>>>
>>>   wake_up_all(>fifo_event);
>>> --
>>> 2.13.5
>>>
>>> ___
>>> dri-devel mailing list
>>> dri-devel@lists.freedesktop.org
>>> https://lists.freedesktop.org/mailman/listinfo/dri-devel
>>
>> --
>> Daniel Vetter
>> Software Engineer, Intel Corporation
>> http://blog.ffwll.ch



-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
___
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Re: [PATCH] drm/msm: use proper memory barriers for updating tail/head

[Rob Clark]

On Tue, Oct 3, 2017 at 4:08 AM, Daniel Vetter  wrote:
> On Mon, Oct 02, 2017 at 10:37:11AM -0400, Rob Clark wrote:
>> Fixes intermittent corruption of cmdstream dump.
>>
>> Signed-off-by: Rob Clark 
>
> Rule of memory barriers: They always need to come in pairs, and you should
> put a comment to each one explaining where the other side is. I.e. for all
> the store_release ones you add here you need to put a few read_acquire
> ones somewhere else, and then explain how they link.
>
> One side only doesn't really work with barriers - except when you're lucky
> and on x86 :-)

fwiw, this is straight out of Documentation/circular-buffers.txt

I guess it could read_acquire when checking the avail/free space on
the consumer/producer side.. otoh if we see an old value there it is
harmless (ie. we just fall into a wait_event()).  I assume that is why
the documentation didn't bother with a read_acquire.

BR,
-R

> -Daniel
>> ---
>>  drivers/gpu/drm/msm/msm_rd.c | 4 ++--
>>  1 file changed, 2 insertions(+), 2 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/msm/msm_rd.c b/drivers/gpu/drm/msm/msm_rd.c
>> index 5c087a4228ff..918c65e19964 100644
>> --- a/drivers/gpu/drm/msm/msm_rd.c
>> +++ b/drivers/gpu/drm/msm/msm_rd.c
>> @@ -120,7 +120,7 @@ static void rd_write(struct msm_rd_state *rd, const void 
>> *buf, int sz)
>>   n = min(sz, circ_space_to_end(>fifo));
>>   memcpy(fptr, ptr, n);
>>
>> - fifo->head = (fifo->head + n) & (BUF_SZ - 1);
>> + smp_store_release(>head, (fifo->head + n) & (BUF_SZ - 
>> 1));
>>   sz  -= n;
>>   ptr += n;
>>
>> @@ -157,7 +157,7 @@ static ssize_t rd_read(struct file *file, char __user 
>> *buf,
>>   goto out;
>>   }
>>
>> - fifo->tail = (fifo->tail + n) & (BUF_SZ - 1);
>> + smp_store_release(>tail, (fifo->tail + n) & (BUF_SZ - 1));
>>   *ppos += n;
>>
>>   wake_up_all(>fifo_event);
>> --
>> 2.13.5
>>
>> ___
>> dri-devel mailing list
>> dri-devel@lists.freedesktop.org
>> https://lists.freedesktop.org/mailman/listinfo/dri-devel
>
> --
> Daniel Vetter
> Software Engineer, Intel Corporation
> http://blog.ffwll.ch
___
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Re: [PATCH] drm/msm: use proper memory barriers for updating tail/head

[Daniel Vetter]

On Mon, Oct 02, 2017 at 10:37:11AM -0400, Rob Clark wrote:
> Fixes intermittent corruption of cmdstream dump.
> 
> Signed-off-by: Rob Clark 

Rule of memory barriers: They always need to come in pairs, and you should
put a comment to each one explaining where the other side is. I.e. for all
the store_release ones you add here you need to put a few read_acquire
ones somewhere else, and then explain how they link.

One side only doesn't really work with barriers - except when you're lucky
and on x86 :-)
-Daniel
> ---
>  drivers/gpu/drm/msm/msm_rd.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/gpu/drm/msm/msm_rd.c b/drivers/gpu/drm/msm/msm_rd.c
> index 5c087a4228ff..918c65e19964 100644
> --- a/drivers/gpu/drm/msm/msm_rd.c
> +++ b/drivers/gpu/drm/msm/msm_rd.c
> @@ -120,7 +120,7 @@ static void rd_write(struct msm_rd_state *rd, const void 
> *buf, int sz)
>   n = min(sz, circ_space_to_end(>fifo));
>   memcpy(fptr, ptr, n);
>  
> - fifo->head = (fifo->head + n) & (BUF_SZ - 1);
> + smp_store_release(>head, (fifo->head + n) & (BUF_SZ - 1));
>   sz  -= n;
>   ptr += n;
>  
> @@ -157,7 +157,7 @@ static ssize_t rd_read(struct file *file, char __user 
> *buf,
>   goto out;
>   }
>  
> - fifo->tail = (fifo->tail + n) & (BUF_SZ - 1);
> + smp_store_release(>tail, (fifo->tail + n) & (BUF_SZ - 1));
>   *ppos += n;
>  
>   wake_up_all(>fifo_event);
> -- 
> 2.13.5
> 
> ___
> dri-devel mailing list
> dri-devel@lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/dri-devel

-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
___
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel

[PATCH] drm/msm: use proper memory barriers for updating tail/head

[Rob Clark]

Fixes intermittent corruption of cmdstream dump.

Signed-off-by: Rob Clark 
---
 drivers/gpu/drm/msm/msm_rd.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/drivers/gpu/drm/msm/msm_rd.c b/drivers/gpu/drm/msm/msm_rd.c
index 5c087a4228ff..918c65e19964 100644
--- a/drivers/gpu/drm/msm/msm_rd.c
+++ b/drivers/gpu/drm/msm/msm_rd.c
@@ -120,7 +120,7 @@ static void rd_write(struct msm_rd_state *rd, const void 
*buf, int sz)
n = min(sz, circ_space_to_end(>fifo));
memcpy(fptr, ptr, n);
 
-   fifo->head = (fifo->head + n) & (BUF_SZ - 1);
+   smp_store_release(>head, (fifo->head + n) & (BUF_SZ - 1));
sz  -= n;
ptr += n;
 
@@ -157,7 +157,7 @@ static ssize_t rd_read(struct file *file, char __user *buf,
goto out;
}
 
-   fifo->tail = (fifo->tail + n) & (BUF_SZ - 1);
+   smp_store_release(>tail, (fifo->tail + n) & (BUF_SZ - 1));
*ppos += n;
 
wake_up_all(>fifo_event);
-- 
2.13.5

___
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel