When userspace takes drm_master, the drm_client buffer is no more
visible, so drm_log shouldn't waste CPU cycle to draw on it.

Signed-off-by: Jocelyn Falempe <jfalempe@redhat.com>
---
 drivers/gpu/drm/drm_log.c | 10 ++++++++--
 1 file changed, 8 insertions(+), 2 deletions(-)

diff --git a/drivers/gpu/drm/drm_log.c b/drivers/gpu/drm/drm_log.c
index 376ee173d99d9..226e206e8b6a3 100644
--- a/drivers/gpu/drm/drm_log.c
+++ b/drivers/gpu/drm/drm_log.c
@@ -18,6 +18,7 @@
 #include <drm/drm_print.h>
 
 #include "drm_draw.h"
+#include "drm_internal.h"
 #include "drm_log.h"
 
 MODULE_AUTHOR("Jocelyn Falempe");
@@ -308,8 +309,13 @@ static void drm_log_write_thread(struct console *con, struct nbcon_write_context
 	if (!dlog->probed)
 		drm_log_init_client(dlog);
 
-	for (i = 0; i < dlog->n_scanout; i++)
-		drm_log_draw_kmsg_record(&dlog->scanout[i], wctxt->outbuf, wctxt->len);
+	/* Check that we are still the master before drawing */
+	if (drm_master_internal_acquire(dlog->client.dev)) {
+		drm_master_internal_release(dlog->client.dev);
+
+		for (i = 0; i < dlog->n_scanout; i++)
+			drm_log_draw_kmsg_record(&dlog->scanout[i], wctxt->outbuf, wctxt->len);
+	}
 }
 
 static void drm_log_lock(struct console *con, unsigned long *flags)
-- 
2.47.0

Hi

Am 08.11.24 um 09:10 schrieb Jocelyn Falempe:
> When userspace takes drm_master, the drm_client buffer is no more
> visible, so drm_log shouldn't waste CPU cycle to draw on it.
>
> Signed-off-by: Jocelyn Falempe <jfalempe@redhat.com>
> ---
>   drivers/gpu/drm/drm_log.c | 10 ++++++++--
>   1 file changed, 8 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/gpu/drm/drm_log.c b/drivers/gpu/drm/drm_log.c
> index 376ee173d99d9..226e206e8b6a3 100644
> --- a/drivers/gpu/drm/drm_log.c
> +++ b/drivers/gpu/drm/drm_log.c
> @@ -18,6 +18,7 @@
>   #include <drm/drm_print.h>
>   
>   #include "drm_draw.h"
> +#include "drm_internal.h"
>   #include "drm_log.h"
>   
>   MODULE_AUTHOR("Jocelyn Falempe");
> @@ -308,8 +309,13 @@ static void drm_log_write_thread(struct console *con, struct nbcon_write_context
>   	if (!dlog->probed)
>   		drm_log_init_client(dlog);
>   
> -	for (i = 0; i < dlog->n_scanout; i++)
> -		drm_log_draw_kmsg_record(&dlog->scanout[i], wctxt->outbuf, wctxt->len);
> +	/* Check that we are still the master before drawing */
> +	if (drm_master_internal_acquire(dlog->client.dev)) {

Just a note: it would be better to track this state in the client code 
and handle these locks automatically. But it's ok for now. It would 
require an overhaul of the fbdev helpers as well.

> +		drm_master_internal_release(dlog->client.dev);

Don't you have to release after drawing?

Best regards
Thomas


> +
> +		for (i = 0; i < dlog->n_scanout; i++)
> +			drm_log_draw_kmsg_record(&dlog->scanout[i], wctxt->outbuf, wctxt->len);
> +	}
>   }
>   
>   static void drm_log_lock(struct console *con, unsigned long *flags)

-- 
--
Thomas Zimmermann
Graphics Driver Developer
SUSE Software Solutions Germany GmbH
Frankenstrasse 146, 90461 Nuernberg, Germany
GF: Ivo Totev, Andrew Myers, Andrew McDonald, Boudien Moerman
HRB 36809 (AG Nuernberg)

On 08/11/2024 14:33, Thomas Zimmermann wrote:
> Hi
> 
> Am 08.11.24 um 09:10 schrieb Jocelyn Falempe:
>> When userspace takes drm_master, the drm_client buffer is no more
>> visible, so drm_log shouldn't waste CPU cycle to draw on it.
>>
>> Signed-off-by: Jocelyn Falempe <jfalempe@redhat.com>
>> ---
>>   drivers/gpu/drm/drm_log.c | 10 ++++++++--
>>   1 file changed, 8 insertions(+), 2 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/drm_log.c b/drivers/gpu/drm/drm_log.c
>> index 376ee173d99d9..226e206e8b6a3 100644
>> --- a/drivers/gpu/drm/drm_log.c
>> +++ b/drivers/gpu/drm/drm_log.c
>> @@ -18,6 +18,7 @@
>>   #include <drm/drm_print.h>
>>   #include "drm_draw.h"
>> +#include "drm_internal.h"
>>   #include "drm_log.h"
>>   MODULE_AUTHOR("Jocelyn Falempe");
>> @@ -308,8 +309,13 @@ static void drm_log_write_thread(struct console 
>> *con, struct nbcon_write_context
>>       if (!dlog->probed)
>>           drm_log_init_client(dlog);
>> -    for (i = 0; i < dlog->n_scanout; i++)
>> -        drm_log_draw_kmsg_record(&dlog->scanout[i], wctxt->outbuf, 
>> wctxt->len);
>> +    /* Check that we are still the master before drawing */
>> +    if (drm_master_internal_acquire(dlog->client.dev)) {
> 
> Just a note: it would be better to track this state in the client code 
> and handle these locks automatically. But it's ok for now. It would 
> require an overhaul of the fbdev helpers as well.
> 
>> +        drm_master_internal_release(dlog->client.dev);
> 
> Don't you have to release after drawing?

I'm not sure, the drawing code will only call 
drm_client_buffer_vmap_local() / unmap
and drm_client_framebuffer_flush(), and they don't require the master 
lock to be taken. I think master lock is needed only for modesetting.

Also the dlog->lock is taken by the console thread, so there are no race 
conditions with the drm_client callbacks (hotplug/suspend/resume).

> 
> Best regards
> Thomas
> 
> 
>> +
>> +        for (i = 0; i < dlog->n_scanout; i++)
>> +            drm_log_draw_kmsg_record(&dlog->scanout[i], wctxt- 
>> >outbuf, wctxt->len);
>> +    }
>>   }
>>   static void drm_log_lock(struct console *con, unsigned long *flags)
>