Am 17.10.22 um 16:30 schrieb Arvind Yadav:
-This is purely a timing issue. Here, sometimes Job free
is happening before the job is done.
To fix this issue moving 'dma_fence_cb' callback from
job(struct drm_sched_job) to scheduler fence (struct drm_sched_fence).
- Added drm_sched_fence_set_parent() function(and others *_parent_cb)
in sched_fence.c. Moved parent fence intilization and callback
installation into this (this just cleanup).
BUG: kernel NULL pointer dereference, address: 0000000000000088
#PF: supervisor read access in kernel mode
#PF: error_code(0x0000) - not-present page
PGD 0 P4D 0
Oops: 0000 [#1] PREEMPT SMP NOPTI
CPU: 2 PID: 0 Comm: swapper/2 Not tainted 6.0.0-rc2-custom #1
Arvind : [dma_fence_default_wait _START] timeout = -1
Hardware name: AMD Dibbler/Dibbler, BIOS RDB1107CC 09/26/2018
RIP: 0010:drm_sched_job_done.isra.0+0x11/0x140 [gpu_sched]
Code: 8b fe ff ff be 03 00 00 00 e8 7b da b7 e3 e9 d4 fe ff ff 66 0f 1f 44 00 00 0f
1f 44 00 00 55 48 89 e5 41 55 41 54 49 89 fc 53 <48> 8b 9f 88 00 00 00 f0 ff 8b
f0 00 00 00 48 8b 83 80 01 00 00 f0
RSP: 0018:ffffb1b1801d4d38 EFLAGS: 00010087
RAX: ffffffffc0aa48b0 RBX: ffffb1b1801d4d70 RCX: 0000000000000018
RDX: 000036c70afb7c1d RSI: ffff8a45ca413c60 RDI: 0000000000000000
RBP: ffffb1b1801d4d50 R08: 00000000000000b5 R09: 0000000000000000
R10: 0000000000000000 R11: 0000000000000000 R12: 0000000000000000
R13: ffffb1b1801d4d70 R14: ffff8a45c4160000 R15: ffff8a45c416a708
FS: 0000000000000000(0000) GS:ffff8a48a0a80000(0000) knlGS:0000000000000000
CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 0000000000000088 CR3: 000000014ad50000 CR4: 00000000003506e0
Call Trace:
<IRQ>
drm_sched_job_done_cb+0x12/0x20 [gpu_sched]
dma_fence_signal_timestamp_locked+0x7e/0x110
dma_fence_signal+0x31/0x60
amdgpu_fence_process+0xc4/0x140 [amdgpu]
gfx_v9_0_eop_irq+0x9d/0xd0 [amdgpu]
amdgpu_irq_dispatch+0xb7/0x210 [amdgpu]
amdgpu_ih_process+0x86/0x100 [amdgpu]
amdgpu_irq_handler+0x24/0x60 [amdgpu]
__handle_irq_event_percpu+0x4b/0x190
handle_irq_event_percpu+0x15/0x50
handle_irq_event+0x39/0x60
handle_edge_irq+0xaf/0x210
__common_interrupt+0x6e/0x110
common_interrupt+0xc1/0xe0
</IRQ>
<TASK>
Signed-off-by: Arvind Yadav <arvind.ya...@amd.com>
---
Changes in v2: Moving 'dma_fence_cb' callback from
job(struct drm_sched_job) to scheduler fence(struct drm_sched_fence)
instead of adding NULL check for s_fence.
Changes in v3: Added drm_sched_fence_set_parent() function(and others
*_parent_cb)
in sched_fence.c. Moved parent fence intilization and callback
installation into this (this just cleanup).
---
drivers/gpu/drm/scheduler/sched_fence.c | 53 +++++++++++++++++++++++++
drivers/gpu/drm/scheduler/sched_main.c | 38 +++++-------------
include/drm/gpu_scheduler.h | 12 +++++-
3 files changed, 72 insertions(+), 31 deletions(-)
diff --git a/drivers/gpu/drm/scheduler/sched_fence.c
b/drivers/gpu/drm/scheduler/sched_fence.c
index 7fd869520ef2..f6808f363261 100644
--- a/drivers/gpu/drm/scheduler/sched_fence.c
+++ b/drivers/gpu/drm/scheduler/sched_fence.c
@@ -77,6 +77,59 @@ static void drm_sched_fence_free_rcu(struct rcu_head *rcu)
if (!WARN_ON_ONCE(!fence))
kmem_cache_free(sched_fence_slab, fence);
}
Please add an empty line here.
+/**
+ * drm_sched_job_done_cb - the callback for a done job
+ * @f: fence
+ * @cb: fence callbacks
+ */
+static void drm_sched_job_done_cb(struct dma_fence *f, struct dma_fence_cb *cb)
Probably best to rename this to something like drm_sched_fence_parent_cb().
+{
+ struct drm_sched_fence *s_fence = container_of(cb, struct
drm_sched_fence,
+ cb);
+ struct drm_gpu_scheduler *sched = s_fence->sched;
+
+ atomic_dec(&sched->hw_rq_count);
+ atomic_dec(sched->score);
+
+ dma_fence_get(&s_fence->finished);
We should probably make sure that this reference is taken before
installing the callback.
+ drm_sched_fence_finished(s_fence);
+ dma_fence_put(&s_fence->finished);
+ wake_up_interruptible(&sched->wake_up_worker);
+}
+
+int drm_sched_fence_add_parent_cb(struct dma_fence *fence,
+ struct drm_sched_fence *s_fence)
+{
+ return dma_fence_add_callback(fence, &s_fence->cb,
+ drm_sched_job_done_cb);
+}
+
+bool drm_sched_fence_remove_parent_cb(struct drm_sched_fence *s_fence)
+{
+ return dma_fence_remove_callback(s_fence->parent,
+ &s_fence->cb);
+}
Do we really need separate functions for that?
+
+/**
+ * drm_sched_fence_set_parent - set the parent fence and add the callback
+ * fence: pointer to the hw fence
+ * @s_fence: pointer to the fence
Reverse the parameter order, s_fence is the object we work on.
+ *
+ * Set the parent fence and intall the callback for a done job.
You need to document that we take the reference of the parent fence.
+ */
+int drm_sched_fence_set_parent(struct dma_fence *fence,
+ struct drm_sched_fence *s_fence)
+{
+ if (s_fence->parent &&
+ dma_fence_remove_callback(s_fence->parent, &s_fence->cb))
+ dma_fence_put(s_fence->parent);
+
+ s_fence->parent = dma_fence_get(fence);
+ /* Drop for original kref_init of the fence */
+ dma_fence_put(fence);
This leaks the reference to the old parent and the get/put dance is not
optimal either.
Better do something like this.
/* We keep the reference of the parent fence here. */
swap(s_fence->parent, fence);
dma_fence_put(fence);
+ return dma_fence_add_callback(fence, &s_fence->cb,
+ drm_sched_job_done_cb);
+}
When installing the callback fails we usually call the callback function
instead of returning the error.
/**
* drm_sched_fence_free - free up an uninitialized fence
diff --git a/drivers/gpu/drm/scheduler/sched_main.c
b/drivers/gpu/drm/scheduler/sched_main.c
index 4cc59bae38dd..cfb52e15f5b0 100644
--- a/drivers/gpu/drm/scheduler/sched_main.c
+++ b/drivers/gpu/drm/scheduler/sched_main.c
@@ -253,13 +253,12 @@ drm_sched_rq_select_entity_fifo(struct drm_sched_rq *rq)
/**
* drm_sched_job_done - complete a job
- * @s_job: pointer to the job which is done
+ * @s_fence: pointer to the fence of a done job
*
* Finish the job's fence and wake up the worker thread.
*/
-static void drm_sched_job_done(struct drm_sched_job *s_job)
+static void drm_sched_job_done(struct drm_sched_fence *s_fence)
{
- struct drm_sched_fence *s_fence = s_job->s_fence;
struct drm_gpu_scheduler *sched = s_fence->sched;
atomic_dec(&sched->hw_rq_count);
@@ -273,18 +272,6 @@ static void drm_sched_job_done(struct drm_sched_job *s_job)
wake_up_interruptible(&sched->wake_up_worker);
}
-/**
- * drm_sched_job_done_cb - the callback for a done job
- * @f: fence
- * @cb: fence callbacks
- */
-static void drm_sched_job_done_cb(struct dma_fence *f, struct dma_fence_cb *cb)
-{
- struct drm_sched_job *s_job = container_of(cb, struct drm_sched_job,
cb);
-
- drm_sched_job_done(s_job);
-}
-
/**
* drm_sched_dependency_optimized - test if the dependency can be optimized
*
@@ -505,8 +492,7 @@ void drm_sched_stop(struct drm_gpu_scheduler *sched, struct
drm_sched_job *bad)
list_for_each_entry_safe_reverse(s_job, tmp, &sched->pending_list,
list) {
if (s_job->s_fence->parent &&
- dma_fence_remove_callback(s_job->s_fence->parent,
- &s_job->cb)) {
+ drm_sched_fence_remove_parent_cb(s_job->s_fence)) {
dma_fence_put(s_job->s_fence->parent);
s_job->s_fence->parent = NULL;
Better just call drm_sched_fence_set_parent() with NULL here to clear
the currently installed parent.
This moves all this dance into the scheduler fence code.
atomic_dec(&sched->hw_rq_count);
@@ -576,15 +562,14 @@ void drm_sched_start(struct drm_gpu_scheduler *sched,
bool full_recovery)
continue;
if (fence) {
- r = dma_fence_add_callback(fence, &s_job->cb,
- drm_sched_job_done_cb);
+ r = drm_sched_fence_add_parent_cb(fence,
s_job->s_fence);
if (r == -ENOENT)
- drm_sched_job_done(s_job);
+ drm_sched_job_done(s_job->s_fence);
else if (r)
DRM_DEV_ERROR(sched->dev, "fence add callback failed
(%d)\n",
Completely nuke that here. All of this should be done in the single
drm_sched_fence_set_parent() function.
And an error message is completely superfluous. We just need to handle
the case that the callback can't be installed because the fence is
already signaled.
Regards,
Christian.
r);
} else
- drm_sched_job_done(s_job);
+ drm_sched_job_done(s_job->s_fence);
}
if (full_recovery) {
@@ -1049,14 +1034,9 @@ static int drm_sched_main(void *param)
drm_sched_fence_scheduled(s_fence);
if (!IS_ERR_OR_NULL(fence)) {
- s_fence->parent = dma_fence_get(fence);
- /* Drop for original kref_init of the fence */
- dma_fence_put(fence);
-
- r = dma_fence_add_callback(fence, &sched_job->cb,
- drm_sched_job_done_cb);
+ r = drm_sched_fence_set_parent(fence, s_fence);
if (r == -ENOENT)
- drm_sched_job_done(sched_job);
+ drm_sched_job_done(s_fence);
else if (r)
DRM_DEV_ERROR(sched->dev, "fence add callback failed
(%d)\n",
r);
@@ -1064,7 +1044,7 @@ static int drm_sched_main(void *param)
if (IS_ERR(fence))
dma_fence_set_error(&s_fence->finished,
PTR_ERR(fence));
- drm_sched_job_done(sched_job);
+ drm_sched_job_done(s_fence);
}
wake_up(&sched->job_scheduled);
diff --git a/include/drm/gpu_scheduler.h b/include/drm/gpu_scheduler.h
index 1f7d9dd1a444..7258e2fa195f 100644
--- a/include/drm/gpu_scheduler.h
+++ b/include/drm/gpu_scheduler.h
@@ -281,6 +281,10 @@ struct drm_sched_fence {
* @owner: job owner for debugging
*/
void *owner;
+ /**
+ * @cb: callback
+ */
+ struct dma_fence_cb cb;
};
struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f);
@@ -300,7 +304,6 @@ struct drm_sched_fence *to_drm_sched_fence(struct dma_fence
*f);
* be scheduled further.
* @s_priority: the priority of the job.
* @entity: the entity to which this job belongs.
- * @cb: the callback for the parent fence in s_fence.
*
* A job is created by the driver using drm_sched_job_init(), and
* should call drm_sched_entity_push_job() once it wants the scheduler
@@ -325,7 +328,6 @@ struct drm_sched_job {
atomic_t karma;
enum drm_sched_priority s_priority;
struct drm_sched_entity *entity;
- struct dma_fence_cb cb;
/**
* @dependencies:
*
@@ -559,6 +561,12 @@ void drm_sched_fence_free(struct drm_sched_fence *fence);
void drm_sched_fence_scheduled(struct drm_sched_fence *fence);
void drm_sched_fence_finished(struct drm_sched_fence *fence);
+int drm_sched_fence_add_parent_cb(struct dma_fence *fence,
+ struct drm_sched_fence *s_fence);
+bool drm_sched_fence_remove_parent_cb(struct drm_sched_fence *s_fence);
+int drm_sched_fence_set_parent(struct dma_fence *fence,
+ struct drm_sched_fence *s_fence);
+
unsigned long drm_sched_suspend_timeout(struct drm_gpu_scheduler *sched);
void drm_sched_resume_timeout(struct drm_gpu_scheduler *sched,
unsigned long remaining);
