[PATCH v6 40/40] drm/msm: Add VM_BIND throttling

[Rob Clark <robin.clark@oss.qualcomm.com>]

A large number of (unsorted or separate) small (<2MB) mappings can cause
a lot of, probably unnecessary, prealloc pages.  Ie. a single 4k page
size mapping will pre-allocate 3 pages (for levels 2-4) for the
pagetable.  Which can chew up a large amount of unneeded memory.  So add
a mechanism to put an upper bound on the # of pre-alloc pages.

Signed-off-by: Rob Clark <robin.clark@oss.qualcomm.com>
---
 drivers/gpu/drm/msm/msm_gem_vma.c | 23 +++++++++++++++++++++--
 drivers/gpu/drm/msm/msm_gpu.h     |  3 +++
 2 files changed, 24 insertions(+), 2 deletions(-)

diff --git a/drivers/gpu/drm/msm/msm_gem_vma.c b/drivers/gpu/drm/msm/msm_gem_vma.c
index b6de87e5c3f7..83f6f95b4865 100644
--- a/drivers/gpu/drm/msm/msm_gem_vma.c
+++ b/drivers/gpu/drm/msm/msm_gem_vma.c
@@ -705,6 +705,8 @@ msm_vma_job_free(struct drm_sched_job *_job)
 
 	mmu->funcs->prealloc_cleanup(mmu, &job->prealloc);
 
+	atomic_sub(job->prealloc.count, &job->queue->in_flight_prealloc);
+
 	drm_sched_job_cleanup(_job);
 
 	job_foreach_bo (obj, job)
@@ -1087,10 +1089,11 @@ ops_are_same_pte(struct msm_vm_bind_op *first, struct msm_vm_bind_op *next)
  * them as a single mapping.  Otherwise the prealloc_count() will not realize
  * they can share pagetable pages and vastly overcount.
  */
-static void
+static int
 vm_bind_prealloc_count(struct msm_vm_bind_job *job)
 {
 	struct msm_vm_bind_op *first = NULL, *last = NULL;
+	int ret;
 
 	for (int i = 0; i < job->nr_ops; i++) {
 		struct msm_vm_bind_op *op = &job->ops[i];
@@ -1119,6 +1122,20 @@ vm_bind_prealloc_count(struct msm_vm_bind_job *job)
 
 	/* Flush the remaining range: */
 	prealloc_count(job, first, last);
+
+	/*
+	 * Now that we know the needed amount to pre-alloc, throttle on pending
+	 * VM_BIND jobs if we already have too much pre-alloc memory in flight
+	 */
+	ret = wait_event_interruptible(
+			to_msm_vm(job->vm)->sched.job_scheduled,
+			atomic_read(&job->queue->in_flight_prealloc) <= 1024);
+	if (ret)
+		return ret;
+
+	atomic_add(job->prealloc.count, &job->queue->in_flight_prealloc);
+
+	return 0;
 }
 
 /*
@@ -1389,7 +1406,9 @@ msm_ioctl_vm_bind(struct drm_device *dev, void *data, struct drm_file *file)
 	if (ret)
 		goto out_unlock;
 
-	vm_bind_prealloc_count(job);
+	ret = vm_bind_prealloc_count(job);
+	if (ret)
+		goto out_unlock;
 
 	struct drm_exec exec;
 	unsigned flags = DRM_EXEC_IGNORE_DUPLICATES | DRM_EXEC_INTERRUPTIBLE_WAIT;
diff --git a/drivers/gpu/drm/msm/msm_gpu.h b/drivers/gpu/drm/msm/msm_gpu.h
index 31b83e9e3673..5508885d865f 100644
--- a/drivers/gpu/drm/msm/msm_gpu.h
+++ b/drivers/gpu/drm/msm/msm_gpu.h
@@ -555,6 +555,8 @@ static inline int msm_gpu_convert_priority(struct msm_gpu *gpu, int prio,
  *             seqno, protected by submitqueue lock
  * @idr_lock:  for serializing access to fence_idr
  * @lock:      submitqueue lock for serializing submits on a queue
+ * @in_flight_prealloc: for VM_BIND queue, # of preallocated pgtable pages for
+ *             queued VM_BIND jobs
  * @ref:       reference count
  * @entity:    the submit job-queue
  */
@@ -569,6 +571,7 @@ struct msm_gpu_submitqueue {
 	struct idr fence_idr;
 	struct spinlock idr_lock;
 	struct mutex lock;
+	atomic_t in_flight_prealloc;
 	struct kref ref;
 	struct drm_sched_entity *entity;
 
-- 
2.49.0