On Wed, Aug 31, 2022 at 12:33:55PM -0700, Ashutosh Dixit wrote:
1. Do all ce->stats updates and reads under guc->timestamp.lock
Other than the question about READ_ONCE/WRITE_ONCE, I am not sure I
understand what's broken in the locking logic here. Can you please add
some description? thanks
2. Pin context image before reading
3. Merge __guc_context_update_clks and guc_context_update_stats into a
single function
4. Call lrc_update_runtime() unconditionally in guc_context_update_stats
5. Seems no need to update ce->stats.active with this approach
Some of the above steps may not be correct or complete but the idea is to
discuss/improve the original patch.
Cc: Umesh Nerlige Ramappa <[email protected]>
Signed-off-by: Ashutosh Dixit <[email protected]>
---
drivers/gpu/drm/i915/gt/intel_context.c | 2 +-
drivers/gpu/drm/i915/gt/intel_context_types.h | 2 +-
.../gpu/drm/i915/gt/uc/intel_guc_submission.c | 41 ++++++++++---------
3 files changed, 24 insertions(+), 21 deletions(-)
diff --git a/drivers/gpu/drm/i915/gt/intel_context.c
b/drivers/gpu/drm/i915/gt/intel_context.c
index e2d70a9fdac0..c004f676de27 100644
--- a/drivers/gpu/drm/i915/gt/intel_context.c
+++ b/drivers/gpu/drm/i915/gt/intel_context.c
@@ -581,7 +581,7 @@ u64 intel_context_get_total_runtime_ns(struct intel_context
*ce)
u64 total, active;
if (ce->ops->update_stats)
- ce->ops->update_stats(ce);
+ return ce->ops->update_stats(ce);
This is an improvement that we can add and eventually, we can make this
a GuC specific vfunc. Doing so may also remove the
COPS_RUNTIME_ACTIVE_TOTAL option that I added to GuC specific context
ops.
total = ce->stats.runtime.total;
if (ce->ops->flags & COPS_RUNTIME_CYCLES)
diff --git a/drivers/gpu/drm/i915/gt/intel_context_types.h
b/drivers/gpu/drm/i915/gt/intel_context_types.h
index f7ff4c7d81c7..699435bfff99 100644
--- a/drivers/gpu/drm/i915/gt/intel_context_types.h
+++ b/drivers/gpu/drm/i915/gt/intel_context_types.h
@@ -59,7 +59,7 @@ struct intel_context_ops {
void (*sched_disable)(struct intel_context *ce);
- void (*update_stats)(struct intel_context *ce);
+ u64 (*update_stats)(struct intel_context *ce);
void (*reset)(struct intel_context *ce);
void (*destroy)(struct kref *kref);
diff --git a/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c
b/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c
index bee8cf10f749..40d0edaf2e0a 100644
--- a/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c
+++ b/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c
@@ -1376,7 +1376,6 @@ static void __update_guc_busyness_stats(struct intel_guc
*guc)
spin_unlock_irqrestore(&guc->timestamp.lock, flags);
}
-static void __guc_context_update_clks(struct intel_context *ce);
static void guc_timestamp_ping(struct work_struct *wrk)
{
struct intel_guc *guc = container_of(wrk, typeof(*guc),
@@ -1401,7 +1400,8 @@ static void guc_timestamp_ping(struct work_struct *wrk)
/* adjust context stats for overflow */
xa_for_each(&guc->context_lookup, index, ce)
- __guc_context_update_clks(ce);
+ if (ce->ops->update_stats)
+ ce->ops->update_stats(ce);
context pinning logic needs to be separated for this (ping) path vs. the
query path - intel_context_get_total_runtime_ns().
intel_gt_reset_unlock(gt, srcu);
@@ -1476,17 +1476,21 @@ void intel_guc_busyness_unpark(struct intel_gt *gt)
guc->timestamp.ping_delay);
}
-static void __guc_context_update_clks(struct intel_context *ce)
+static u64 guc_context_update_stats(struct intel_context *ce)
{
struct intel_guc *guc = ce_to_guc(ce);
struct intel_gt *gt = ce->engine->gt;
u32 *pphwsp, last_switch, engine_id;
- u64 start_gt_clk, active;
unsigned long flags;
+ u64 total, active = 0;
ktime_t unused;
+ intel_context_pin(ce);
intel_context_pin can sleep and we are not allowed to sleep in this path
- intel_context_get_total_runtime_ns(), however we can sleep in the ping
worker path, so ideally we want to separate it out for the 2 paths.
spin_lock_irqsave(&guc->timestamp.lock, flags);
+ lrc_update_runtime(ce);
+ total = ce->stats.runtime.total;
For long running contexts and frequenct queries, calling this ^ when a
context is active does not add value. I would just call it in the else
like before.
+
/*
* GPU updates ce->lrc_reg_state[CTX_TIMESTAMP] when context is switched
* out, however GuC updates PPHWSP offsets below. Hence KMD (CPU)
@@ -1502,27 +1506,26 @@ static void __guc_context_update_clks(struct
intel_context *ce)
guc_update_pm_timestamp(guc, &unused);
if (engine_id != 0xffffffff && last_switch) {
- start_gt_clk = READ_ONCE(ce->stats.runtime.start_gt_clk);
- __extend_last_switch(guc, &start_gt_clk, last_switch);
- active = intel_gt_clock_interval_to_ns(gt,
guc->timestamp.gt_stamp - start_gt_clk);
- WRITE_ONCE(ce->stats.runtime.start_gt_clk, start_gt_clk);
- WRITE_ONCE(ce->stats.active, active);
- } else {
- lrc_update_runtime(ce);
+ __extend_last_switch(guc, &ce->stats.runtime.start_gt_clk,
last_switch);
+ active = intel_gt_clock_interval_to_ns(gt,
+ guc->timestamp.gt_stamp -
ce->stats.runtime.start_gt_clk);
Makes sense. Earlier it was a rmw, now it's just a write to
ce->stats.runtime.start_gt_clk.
}
spin_unlock_irqrestore(&guc->timestamp.lock, flags);
+ intel_context_unpin(ce);
+
+ return total + active;
return ce->stats.runtime.total + active;
and then drop the local copy of total by bringing back the else{}.
}
-static void guc_context_update_stats(struct intel_context *ce)
+void lrc_update_runtime_locked(struct intel_context *ce)
{
- if (!intel_context_pin_if_active(ce)) {
- WRITE_ONCE(ce->stats.runtime.start_gt_clk, 0);
- WRITE_ONCE(ce->stats.active, 0);
- return;
- }
+ struct intel_guc *guc = ce_to_guc(ce);
+ unsigned long flags;
- __guc_context_update_clks(ce);
+ intel_context_pin(ce);
+ spin_lock_irqsave(&guc->timestamp.lock, flags);
+ lrc_update_runtime(ce);
+ spin_unlock_irqrestore(&guc->timestamp.lock, flags);
intel_context_unpin(ce);
From where lrc_update_runtime_locked is being called, I would assume
that the context is already pinned (until unpin_guc_id is called).
Thanks,
Umesh
}
@@ -2780,7 +2783,7 @@ static void guc_context_unpin(struct intel_context *ce)
{
struct intel_guc *guc = ce_to_guc(ce);
- lrc_update_runtime(ce);
+ lrc_update_runtime_locked(ce);
unpin_guc_id(guc, ce);
lrc_unpin(ce);
--
2.34.1
