v2: I've added two more patches on top of "Cancel only job that actually got
stuck" patch. Hence it is a patch-series now.
Dmitry Osipenko (3):
gpu: host1x: Cancel only job that actually got stuck
gpu: host1x: Don't complete a completed job
gpu: host1x: Continue CDMA execution starting with a next job
drivers/gpu/host1x/cdma.c | 62 +++++++++++++--------------------
drivers/gpu/host1x/hw/cdma_hw.c | 13 -------
2 files changed, 24 insertions(+), 51 deletions(-)
--
2.18.0
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 <[email protected]>
Reviewed-by: Mikko Perttunen <[email protected]>
---
drivers/gpu/host1x/cdma.c | 33 +++++++--------------------------
1 file changed, 7 insertions(+), 26 deletions(-)
diff --git a/drivers/gpu/host1x/cdma.c b/drivers/gpu/host1x/cdma.c
index 91df51e631b2..75f339f5df6f 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
@@ -362,20 +360,15 @@ void host1x_cdma_update_sync_queue(struct host1x_cdma *cdma,
* properly for this buffer and resources are freed.
*/
- dev_dbg(dev, "%s: perform CPU incr on pending same ctx buffers\n",
- __func__);
+ dev_dbg(dev, "%s: perform CPU incr on pending buffers\n", __func__);
if (!list_empty(&cdma->sync_queue))
restart_addr = job->first_get;
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 +381,8 @@ 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
There is a chance that the last job has been completed at the time of
CDMA timeout handler invocation. In this case there is no need to complete
the completed job.
Signed-off-by: Dmitry Osipenko <[email protected]>
---
drivers/gpu/host1x/cdma.c | 22 ++++++++++++++--------
1 file changed, 14 insertions(+), 8 deletions(-)
diff --git a/drivers/gpu/host1x/cdma.c b/drivers/gpu/host1x/cdma.c
index 75f339f5df6f..6aa6fa1498e8 100644
--- a/drivers/gpu/host1x/cdma.c
+++ b/drivers/gpu/host1x/cdma.c
@@ -323,7 +323,7 @@ void host1x_cdma_update_sync_queue(struct host1x_cdma *cdma,
{
struct host1x *host1x = cdma_to_host1x(cdma);
u32 restart_addr, syncpt_incrs, syncpt_val;
- struct host1x_job *job = NULL;
+ struct host1x_job *job;
syncpt_val = host1x_syncpt_load(cdma->timeout.syncpt);
@@ -342,11 +342,16 @@ void host1x_cdma_update_sync_queue(struct host1x_cdma *cdma,
list_for_each_entry(job, &cdma->sync_queue, list) {
if (syncpt_val < job->syncpt_end)
- break;
+ goto syncpt_incr;
host1x_job_dump(dev, job);
}
+ /* all jobs have been completed */
+ job = NULL;
+
+syncpt_incr:
+
/*
* Increment with CPU the remaining syncpts of a partially executed job.
*
@@ -359,16 +364,16 @@ void host1x_cdma_update_sync_queue(struct host1x_cdma *cdma,
* modified NOP-ed PB slots). This lets things appear to have completed
* properly for this buffer and resources are freed.
*/
-
- dev_dbg(dev, "%s: perform CPU incr on pending buffers\n", __func__);
-
- if (!list_empty(&cdma->sync_queue))
+ if (job)
restart_addr = job->first_get;
else
restart_addr = cdma->last_pos;
/* do CPU increments for the remaining syncpts */
if (job) {
+ dev_dbg(dev, "%s: perform CPU incr on pending buffers\n",
+ __func__);
+
/* won't need a timeout when replayed */
job->timeout = 0;
@@ -381,9 +386,10 @@ 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);
- }
- dev_dbg(dev, "%s: finished sync_queue modification\n", __func__);
+ dev_dbg(dev, "%s: finished sync_queue modification\n",
+ __func__);
+ }
/* roll back DMAGET and start up channel again */
host1x_hw_cdma_resume(host1x, cdma, restart_addr);
--
2.18.0
Currently gathers of a hung job are getting NOP'ed and a restarted CDMA
executes the NOP'ed gathers. There shouldn't be a reason to not restart
CDMA execution starting with a next job, avoiding the unnecessary churning
with gathers NOP'ing.
Signed-off-by: Dmitry Osipenko <[email protected]>
---
drivers/gpu/host1x/cdma.c | 23 +++++++++++------------
drivers/gpu/host1x/hw/cdma_hw.c | 13 -------------
2 files changed, 11 insertions(+), 25 deletions(-)
diff --git a/drivers/gpu/host1x/cdma.c b/drivers/gpu/host1x/cdma.c
index 6aa6fa1498e8..9e4f01c7f663 100644
--- a/drivers/gpu/host1x/cdma.c
+++ b/drivers/gpu/host1x/cdma.c
@@ -323,7 +323,7 @@ void host1x_cdma_update_sync_queue(struct host1x_cdma *cdma,
{
struct host1x *host1x = cdma_to_host1x(cdma);
u32 restart_addr, syncpt_incrs, syncpt_val;
- struct host1x_job *job;
+ struct host1x_job *job, *next_job = NULL;
syncpt_val = host1x_syncpt_load(cdma->timeout.syncpt);
@@ -341,8 +341,13 @@ void host1x_cdma_update_sync_queue(struct host1x_cdma *cdma,
__func__);
list_for_each_entry(job, &cdma->sync_queue, list) {
- if (syncpt_val < job->syncpt_end)
+ if (syncpt_val < job->syncpt_end) {
+
+ if (!list_is_last(&job->list, &cdma->sync_queue))
+ next_job = list_next_entry(job, list);
+
goto syncpt_incr;
+ }
host1x_job_dump(dev, job);
}
@@ -355,17 +360,11 @@ void host1x_cdma_update_sync_queue(struct host1x_cdma *cdma,
/*
* Increment with CPU the remaining syncpts of a partially executed job.
*
- * 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
- * modified NOP-ed PB slots). This lets things appear to have completed
- * properly for this buffer and resources are freed.
+ * CDMA will continue execution starting with the next job or will get
+ * into idle state.
*/
- if (job)
- restart_addr = job->first_get;
+ if (next_job)
+ restart_addr = next_job->first_get;
else
restart_addr = cdma->last_pos;
diff --git a/drivers/gpu/host1x/hw/cdma_hw.c b/drivers/gpu/host1x/hw/cdma_hw.c
index ce320534cbed..bc203532ae6d 100644
--- a/drivers/gpu/host1x/hw/cdma_hw.c
+++ b/drivers/gpu/host1x/hw/cdma_hw.c
@@ -40,7 +40,6 @@ static void cdma_timeout_cpu_incr(struct host1x_cdma *cdma, u32 getptr,
u32 syncpt_incrs, u32 syncval, u32 nr_slots)
{
struct host1x *host1x = cdma_to_host1x(cdma);
- struct push_buffer *pb = &cdma->push_buffer;
unsigned int i;
for (i = 0; i < syncpt_incrs; i++)
@@ -48,18 +47,6 @@ static void cdma_timeout_cpu_incr(struct host1x_cdma *cdma, u32 getptr,
/* after CPU incr, ensure shadow is up to date */
host1x_syncpt_load(cdma->timeout.syncpt);
-
- /* NOP all the PB slots */
- while (nr_slots--) {
- u32 *p = (u32 *)(pb->mapped + getptr);
- *(p++) = HOST1X_OPCODE_NOP;
- *(p++) = HOST1X_OPCODE_NOP;
- dev_dbg(host1x->dev, "%s: NOP at %pad+%#x\n", __func__,
- &pb->dma, getptr);
- getptr = (getptr + 8) & (pb->size - 1);
- }
-
- wmb();
}
/*
--
2.18.0
On 07.08.2018 16:07, Dmitry Osipenko wrote:
> Currently gathers of a hung job are getting NOP'ed and a restarted CDMA
> executes the NOP'ed gathers. There shouldn't be a reason to not restart
> CDMA execution starting with a next job, avoiding the unnecessary churning
> with gathers NOP'ing.
>
> Signed-off-by: Dmitry Osipenko <[email protected]>
> ---
> drivers/gpu/host1x/cdma.c | 23 +++++++++++------------
> drivers/gpu/host1x/hw/cdma_hw.c | 13 -------------
> 2 files changed, 11 insertions(+), 25 deletions(-)
>
> diff --git a/drivers/gpu/host1x/cdma.c b/drivers/gpu/host1x/cdma.c
> index 6aa6fa1498e8..9e4f01c7f663 100644
> --- a/drivers/gpu/host1x/cdma.c
> +++ b/drivers/gpu/host1x/cdma.c
> @@ -323,7 +323,7 @@ void host1x_cdma_update_sync_queue(struct host1x_cdma *cdma,
> {
> struct host1x *host1x = cdma_to_host1x(cdma);
> u32 restart_addr, syncpt_incrs, syncpt_val;
> - struct host1x_job *job;
> + struct host1x_job *job, *next_job = NULL;
>
> syncpt_val = host1x_syncpt_load(cdma->timeout.syncpt);
>
> @@ -341,8 +341,13 @@ void host1x_cdma_update_sync_queue(struct host1x_cdma *cdma,
> __func__);
>
> list_for_each_entry(job, &cdma->sync_queue, list) {
> - if (syncpt_val < job->syncpt_end)
> + if (syncpt_val < job->syncpt_end) {
> +
> + if (!list_is_last(&job->list, &cdma->sync_queue))
> + next_job = list_next_entry(job, list);
> +
> goto syncpt_incr;
> + }
>
> host1x_job_dump(dev, job);
> }
> @@ -355,17 +360,11 @@ void host1x_cdma_update_sync_queue(struct host1x_cdma *cdma,
> /*
> * Increment with CPU the remaining syncpts of a partially executed job.
> *
> - * 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
> - * modified NOP-ed PB slots). This lets things appear to have completed
> - * properly for this buffer and resources are freed.
> + * CDMA will continue execution starting with the next job or will get
> + * into idle state.
> */
> - if (job)
> - restart_addr = job->first_get;
> + if (next_job)
> + restart_addr = next_job->first_get;
> else
> restart_addr = cdma->last_pos;
>
> diff --git a/drivers/gpu/host1x/hw/cdma_hw.c b/drivers/gpu/host1x/hw/cdma_hw.c
> index ce320534cbed..bc203532ae6d 100644
> --- a/drivers/gpu/host1x/hw/cdma_hw.c
> +++ b/drivers/gpu/host1x/hw/cdma_hw.c
> @@ -40,7 +40,6 @@ static void cdma_timeout_cpu_incr(struct host1x_cdma *cdma, u32 getptr,
> u32 syncpt_incrs, u32 syncval, u32 nr_slots)
> {
> struct host1x *host1x = cdma_to_host1x(cdma);
The *host1x is now unused and should be removed as well, I'll fix it in v3.
Mikko, could you take a look at the patches?
Reviewed-by: Mikko Perttunen <[email protected]>
On 07/08/2018 22.07, Dmitry Osipenko wrote:
> There is a chance that the last job has been completed at the time of
> CDMA timeout handler invocation. In this case there is no need to complete
> the completed job.
>
> Signed-off-by: Dmitry Osipenko <[email protected]>
> ---
> drivers/gpu/host1x/cdma.c | 22 ++++++++++++++--------
> 1 file changed, 14 insertions(+), 8 deletions(-)
>
> diff --git a/drivers/gpu/host1x/cdma.c b/drivers/gpu/host1x/cdma.c
> index 75f339f5df6f..6aa6fa1498e8 100644
> --- a/drivers/gpu/host1x/cdma.c
> +++ b/drivers/gpu/host1x/cdma.c
> @@ -323,7 +323,7 @@ void host1x_cdma_update_sync_queue(struct host1x_cdma *cdma,
> {
> struct host1x *host1x = cdma_to_host1x(cdma);
> u32 restart_addr, syncpt_incrs, syncpt_val;
> - struct host1x_job *job = NULL;
> + struct host1x_job *job;
>
> syncpt_val = host1x_syncpt_load(cdma->timeout.syncpt);
>
> @@ -342,11 +342,16 @@ void host1x_cdma_update_sync_queue(struct host1x_cdma *cdma,
>
> list_for_each_entry(job, &cdma->sync_queue, list) {
> if (syncpt_val < job->syncpt_end)
> - break;
> + goto syncpt_incr;
>
> host1x_job_dump(dev, job);
> }
>
> + /* all jobs have been completed */
> + job = NULL;
> +
> +syncpt_incr:
> +
> /*
> * Increment with CPU the remaining syncpts of a partially executed job.
> *
> @@ -359,16 +364,16 @@ void host1x_cdma_update_sync_queue(struct host1x_cdma *cdma,
> * modified NOP-ed PB slots). This lets things appear to have completed
> * properly for this buffer and resources are freed.
> */
> -
> - dev_dbg(dev, "%s: perform CPU incr on pending buffers\n", __func__);
> -
> - if (!list_empty(&cdma->sync_queue))
> + if (job)
> restart_addr = job->first_get;
> else
> restart_addr = cdma->last_pos;
>
> /* do CPU increments for the remaining syncpts */
> if (job) {
> + dev_dbg(dev, "%s: perform CPU incr on pending buffers\n",
> + __func__);
> +
> /* won't need a timeout when replayed */
> job->timeout = 0;
>
> @@ -381,9 +386,10 @@ 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);
> - }
>
> - dev_dbg(dev, "%s: finished sync_queue modification\n", __func__);
> + dev_dbg(dev, "%s: finished sync_queue modification\n",
> + __func__);
> + }
>
> /* roll back DMAGET and start up channel again */
> host1x_hw_cdma_resume(host1x, cdma, restart_addr);
>
Reviewed-by: Mikko Perttnuen <[email protected]>
On 07/08/2018 22.07, Dmitry Osipenko wrote:
> Currently gathers of a hung job are getting NOP'ed and a restarted CDMA
> executes the NOP'ed gathers. There shouldn't be a reason to not restart
> CDMA execution starting with a next job, avoiding the unnecessary churning
> with gathers NOP'ing.
>
> Signed-off-by: Dmitry Osipenko <[email protected]>
> ---
> drivers/gpu/host1x/cdma.c | 23 +++++++++++------------
> drivers/gpu/host1x/hw/cdma_hw.c | 13 -------------
> 2 files changed, 11 insertions(+), 25 deletions(-)
>
> diff --git a/drivers/gpu/host1x/cdma.c b/drivers/gpu/host1x/cdma.c
> index 6aa6fa1498e8..9e4f01c7f663 100644
> --- a/drivers/gpu/host1x/cdma.c
> +++ b/drivers/gpu/host1x/cdma.c
> @@ -323,7 +323,7 @@ void host1x_cdma_update_sync_queue(struct host1x_cdma *cdma,
> {
> struct host1x *host1x = cdma_to_host1x(cdma);
> u32 restart_addr, syncpt_incrs, syncpt_val;
> - struct host1x_job *job;
> + struct host1x_job *job, *next_job = NULL;
>
> syncpt_val = host1x_syncpt_load(cdma->timeout.syncpt);
>
> @@ -341,8 +341,13 @@ void host1x_cdma_update_sync_queue(struct host1x_cdma *cdma,
> __func__);
>
> list_for_each_entry(job, &cdma->sync_queue, list) {
> - if (syncpt_val < job->syncpt_end)
> + if (syncpt_val < job->syncpt_end) {
> +
> + if (!list_is_last(&job->list, &cdma->sync_queue))
> + next_job = list_next_entry(job, list);
> +
> goto syncpt_incr;
> + }
>
> host1x_job_dump(dev, job);
> }
> @@ -355,17 +360,11 @@ void host1x_cdma_update_sync_queue(struct host1x_cdma *cdma,
> /*
> * Increment with CPU the remaining syncpts of a partially executed job.
> *
> - * 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
> - * modified NOP-ed PB slots). This lets things appear to have completed
> - * properly for this buffer and resources are freed.
> + * CDMA will continue execution starting with the next job or will get
> + * into idle state.
> */
> - if (job)
> - restart_addr = job->first_get;
> + if (next_job)
> + restart_addr = next_job->first_get;
> else
> restart_addr = cdma->last_pos;
>
> diff --git a/drivers/gpu/host1x/hw/cdma_hw.c b/drivers/gpu/host1x/hw/cdma_hw.c
> index ce320534cbed..bc203532ae6d 100644
> --- a/drivers/gpu/host1x/hw/cdma_hw.c
> +++ b/drivers/gpu/host1x/hw/cdma_hw.c
> @@ -40,7 +40,6 @@ static void cdma_timeout_cpu_incr(struct host1x_cdma *cdma, u32 getptr,
> u32 syncpt_incrs, u32 syncval, u32 nr_slots)
> {
> struct host1x *host1x = cdma_to_host1x(cdma);
> - struct push_buffer *pb = &cdma->push_buffer;
> unsigned int i;
>
> for (i = 0; i < syncpt_incrs; i++)
> @@ -48,18 +47,6 @@ static void cdma_timeout_cpu_incr(struct host1x_cdma *cdma, u32 getptr,
>
> /* after CPU incr, ensure shadow is up to date */
> host1x_syncpt_load(cdma->timeout.syncpt);
> -
> - /* NOP all the PB slots */
> - while (nr_slots--) {
> - u32 *p = (u32 *)(pb->mapped + getptr);
> - *(p++) = HOST1X_OPCODE_NOP;
> - *(p++) = HOST1X_OPCODE_NOP;
> - dev_dbg(host1x->dev, "%s: NOP at %pad+%#x\n", __func__,
> - &pb->dma, getptr);
> - getptr = (getptr + 8) & (pb->size - 1);
> - }
> -
> - wmb();
> }
>
> /*
>
On 10/9/18 8:10 AM, Mikko Perttunen wrote:
> Reviewed-by: Mikko Perttnuen <[email protected]>
Thank you :)