[PATCH v1] gpu: host1x: Cancel only job that actually got stuck

[PATCH v1] gpu: host1x: Cancel only job that actually got stuck

[Dmitry Osipenko]

Host1x doesn't have information about jobs inter-dependency, that is
something that will become available once host1x will get a proper
jobs scheduler implementation. Currently a hang job causes other unrelated
jobs to be canceled, that is a relic from downstream driver which is
irrelevant to upstream. Let's cancel only the hanging job and not to touch
other jobs in queue.

Signed-off-by: Dmitry Osipenko <digetx@gmail.com>
---
 drivers/gpu/host1x/cdma.c | 30 ++++++------------------------
 1 file changed, 6 insertions(+), 24 deletions(-)

diff --git a/drivers/gpu/host1x/cdma.c b/drivers/gpu/host1x/cdma.c
index 91df51e631b2..4d94af4a315f 100644
--- a/drivers/gpu/host1x/cdma.c
+++ b/drivers/gpu/host1x/cdma.c
@@ -348,13 +348,11 @@ void host1x_cdma_update_sync_queue(struct host1x_cdma *cdma,
 	}
 
 	/*
-	 * Walk the sync_queue, first incrementing with the CPU syncpts that
-	 * are partially executed (the first buffer) or fully skipped while
-	 * still in the current context (slots are also NOP-ed).
+	 * Increment with CPU the remaining syncpts of a partially executed job.
 	 *
-	 * At the point contexts are interleaved, syncpt increments must be
-	 * done inline with the pushbuffer from a GATHER buffer to maintain
-	 * the order (slots are modified to be a GATHER of syncpt incrs).
+	 * Syncpt increments must be done inline with the pushbuffer from a
+	 * GATHER buffer to maintain the order (slots are modified to be a
+	 * GATHER of syncpt incrs).
 	 *
 	 * Note: save in restart_addr the location where the timed out buffer
 	 * started in the PB, so we can start the refetch from there (with the
@@ -370,12 +368,8 @@ void host1x_cdma_update_sync_queue(struct host1x_cdma *cdma,
 	else
 		restart_addr = cdma->last_pos;
 
-	/* do CPU increments as long as this context continues */
-	list_for_each_entry_from(job, &cdma->sync_queue, list) {
-		/* different context, gets us out of this loop */
-		if (job->client != cdma->timeout.client)
-			break;
-
+	/* do CPU increments for the remaining syncpts */
+	if (job) {
 		/* won't need a timeout when replayed */
 		job->timeout = 0;
 
@@ -388,20 +382,8 @@ void host1x_cdma_update_sync_queue(struct host1x_cdma *cdma,
 		host1x_hw_cdma_timeout_cpu_incr(host1x, cdma, job->first_get,
 						syncpt_incrs, job->syncpt_end,
 						job->num_slots);
-
-		syncpt_val += syncpt_incrs;
 	}
 
-	/*
-	 * The following sumbits from the same client may be dependent on the
-	 * failed submit and therefore they may fail. Force a small timeout
-	 * to make the queue cleanup faster.
-	 */
-
-	list_for_each_entry_from(job, &cdma->sync_queue, list)
-		if (job->client == cdma->timeout.client)
-			job->timeout = min_t(unsigned int, job->timeout, 500);
-
 	dev_dbg(dev, "%s: finished sync_queue modification\n", __func__);
 
 	/* roll back DMAGET and start up channel again */
-- 
2.18.0

Re: [PATCH v1] gpu: host1x: Cancel only job that actually got stuck

[Mikko Perttunen]

Reviewed-by: Mikko Perttunen <mperttunen@nvidia.com>

On 05.08.2018 20:01, Dmitry Osipenko wrote:
> Host1x doesn't have information about jobs inter-dependency, that is
> something that will become available once host1x will get a proper
> jobs scheduler implementation. Currently a hang job causes other unrelated
> jobs to be canceled, that is a relic from downstream driver which is
> irrelevant to upstream. Let's cancel only the hanging job and not to touch
> other jobs in queue.
> 
> Signed-off-by: Dmitry Osipenko <digetx@gmail.com>
> ---
>   drivers/gpu/host1x/cdma.c | 30 ++++++------------------------
>   1 file changed, 6 insertions(+), 24 deletions(-)
> 
> diff --git a/drivers/gpu/host1x/cdma.c b/drivers/gpu/host1x/cdma.c
> index 91df51e631b2..4d94af4a315f 100644
> --- a/drivers/gpu/host1x/cdma.c
> +++ b/drivers/gpu/host1x/cdma.c
> @@ -348,13 +348,11 @@ void host1x_cdma_update_sync_queue(struct host1x_cdma *cdma,
>   	}
>   
>   	/*
> -	 * Walk the sync_queue, first incrementing with the CPU syncpts that
> -	 * are partially executed (the first buffer) or fully skipped while
> -	 * still in the current context (slots are also NOP-ed).
> +	 * Increment with CPU the remaining syncpts of a partially executed job.
>   	 *
> -	 * At the point contexts are interleaved, syncpt increments must be
> -	 * done inline with the pushbuffer from a GATHER buffer to maintain
> -	 * the order (slots are modified to be a GATHER of syncpt incrs).
> +	 * Syncpt increments must be done inline with the pushbuffer from a
> +	 * GATHER buffer to maintain the order (slots are modified to be a
> +	 * GATHER of syncpt incrs).
>   	 *
>   	 * Note: save in restart_addr the location where the timed out buffer
>   	 * started in the PB, so we can start the refetch from there (with the
> @@ -370,12 +368,8 @@ void host1x_cdma_update_sync_queue(struct host1x_cdma *cdma,
>   	else
>   		restart_addr = cdma->last_pos;
>   
> -	/* do CPU increments as long as this context continues */
> -	list_for_each_entry_from(job, &cdma->sync_queue, list) {
> -		/* different context, gets us out of this loop */
> -		if (job->client != cdma->timeout.client)
> -			break;
> -
> +	/* do CPU increments for the remaining syncpts */
> +	if (job) {
>   		/* won't need a timeout when replayed */
>   		job->timeout = 0;
>   
> @@ -388,20 +382,8 @@ void host1x_cdma_update_sync_queue(struct host1x_cdma *cdma,
>   		host1x_hw_cdma_timeout_cpu_incr(host1x, cdma, job->first_get,
>   						syncpt_incrs, job->syncpt_end,
>   						job->num_slots);
> -
> -		syncpt_val += syncpt_incrs;
>   	}
>   
> -	/*
> -	 * The following sumbits from the same client may be dependent on the
> -	 * failed submit and therefore they may fail. Force a small timeout
> -	 * to make the queue cleanup faster.
> -	 */
> -
> -	list_for_each_entry_from(job, &cdma->sync_queue, list)
> -		if (job->client == cdma->timeout.client)
> -			job->timeout = min_t(unsigned int, job->timeout, 500);
> -
>   	dev_dbg(dev, "%s: finished sync_queue modification\n", __func__);
>   
>   	/* roll back DMAGET and start up channel again */
>