video: tegra: host: Change sync queue to a list

Sync queue is the list of jobs still in flight. As context priorities
requires possibility to insert a job in the middle of the queue, the
structure needs to be changed into a linked list.

Bug 926690

Change-Id: Id257a11f18476c70dd69e36ba44ed2d380c80040
Signed-off-by: Terje Bergstrom <tbergstrom@nvidia.com>
Reviewed-on: http://git-master/r/83127

5 files changed, 30 insertions, 88 deletions

diff --git a/drivers/video/tegra/host/dev.h b/drivers/video/tegra/host/dev.h
index 7b6fd99746dc..67d78aabfa6a 100644
--- a/drivers/video/tegra/host/dev.h
+++ b/drivers/video/tegra/host/dev.h
@@ -47,8 +47,6 @@ struct nvhost_master {
 	struct nvhost_channel *channels;
 	u32 nb_channels;
 
-	u32 sync_queue_size;
-
 	struct nvhost_chip_support op;
 
 	atomic_t clientid;
diff --git a/drivers/video/tegra/host/host1x/host1x_cdma.c b/drivers/video/tegra/host/host1x/host1x_cdma.c
index 65a72801528f..1006fbd90f88 100644
--- a/drivers/video/tegra/host/host1x/host1x_cdma.c
+++ b/drivers/video/tegra/host/host1x/host1x_cdma.c
@@ -655,8 +655,6 @@ int host1x_init_cdma_support(struct nvhost_master *host)
 	host->op.cdma.timeout_cpu_incr = cdma_timeout_cpu_incr;
 	host->op.cdma.timeout_pb_incr = cdma_timeout_pb_incr;
 
-	host->sync_queue_size = NVHOST_SYNC_QUEUE_SIZE;
-
 	host->op.push_buffer.reset = push_buffer_reset;
 	host->op.push_buffer.init = push_buffer_init;
 	host->op.push_buffer.destroy = push_buffer_destroy;
diff --git a/drivers/video/tegra/host/nvhost_cdma.c b/drivers/video/tegra/host/nvhost_cdma.c
index 6446befbdd58..908d8bb1da04 100644
--- a/drivers/video/tegra/host/nvhost_cdma.c
+++ b/drivers/video/tegra/host/nvhost_cdma.c
@@ -33,36 +33,11 @@
  * TODO:
  *   stats
  *     - for figuring out what to optimize further
- *   resizable push buffer & sync queue
+ *   resizable push buffer
  *     - some channels hardly need any, some channels (3d) could use more
  */
 
 /**
- * kfifo_save - save current out pointer
- * @fifo: address of the fifo to be used
- */
-#define	kfifo_save(fifo) \
-__kfifo_uint_must_check_helper( \
-({ \
-	typeof((fifo) + 1) __tmp = (fifo); \
-	struct __kfifo *__kfifo = &__tmp->kfifo; \
-	__kfifo->out; \
-}) \
-)
-
-/**
- * kfifo_restore - restore previously saved pointer
- * @fifo: address of the fifo to be used
- * @out: output pointer
- */
-#define	kfifo_restore(fifo, restore) \
-(void)({ \
-	typeof((fifo) + 1) __tmp = (fifo); \
-	struct __kfifo *__kfifo = &__tmp->kfifo; \
-	__kfifo->out = (restore); \
-})
-
-/**
  * Add an entry to the sync queue.
  */
 static void add_to_sync_queue(struct nvhost_cdma *cdma,
@@ -75,13 +50,12 @@ static void add_to_sync_queue(struct nvhost_cdma *cdma,
 	job->first_get = first_get;
 	job->num_slots = nr_slots;
 	nvhost_job_get(job);
-	kfifo_in(&cdma->sync_queue, (void *)&job, 1);
+	list_add_tail(&job->list, &cdma->sync_queue);
 }
 
 /**
  * Return the status of the cdma's sync queue or push buffer for the given event
  *  - sq empty: returns 1 for empty, 0 for not empty (as in "1 empty queue" :-)
- *  - sq space: returns the number of handles that can be stored in the queue
  *  - pb space: returns the number of free slots in the channel's push buffer
  * Must be called with the cdma lock held.
  */
@@ -90,9 +64,7 @@ static unsigned int cdma_status_locked(struct nvhost_cdma *cdma,
 {
 	switch (event) {
 	case CDMA_EVENT_SYNC_QUEUE_EMPTY:
-		return kfifo_len(&cdma->sync_queue) == 0 ? 1 : 0;
-	case CDMA_EVENT_SYNC_QUEUE_SPACE:
-		return kfifo_avail(&cdma->sync_queue);
+		return list_empty(&cdma->sync_queue) ? 1 : 0;
 	case CDMA_EVENT_PUSH_BUFFER_SPACE: {
 		struct push_buffer *pb = &cdma->push_buffer;
 		BUG_ON(!cdma_pb_op(cdma).space);
@@ -107,7 +79,6 @@ static unsigned int cdma_status_locked(struct nvhost_cdma *cdma,
  * Sleep (if necessary) until the requested event happens
  *   - CDMA_EVENT_SYNC_QUEUE_EMPTY : sync queue is completely empty.
  *     - Returns 1
- *   - CDMA_EVENT_SYNC_QUEUE_SPACE : there is space in the sync queue.
  *   - CDMA_EVENT_PUSH_BUFFER_SPACE : there is space in the push buffer
  *     - Return the amount of space (> 0)
  * Must be called with the cdma lock held.
@@ -181,6 +152,8 @@ static void update_cdma_locked(struct nvhost_cdma *cdma)
 {
 	bool signal = false;
 	struct nvhost_master *dev = cdma_to_dev(cdma);
+	struct nvhost_syncpt *sp = &dev->syncpt;
+	struct nvhost_job *job, *n;
 
 	BUG_ON(!cdma->running);
 
@@ -188,18 +161,7 @@ static void update_cdma_locked(struct nvhost_cdma *cdma)
 	 * Walk the sync queue, reading the sync point registers as necessary,
 	 * to consume as many sync queue entries as possible without blocking
 	 */
-	for (;;) {
-		struct nvhost_syncpt *sp = &dev->syncpt;
-		struct nvhost_job *job;
-		int result;
-
-		result = kfifo_peek(&cdma->sync_queue, &job);
-		if (!result) {
-			if (cdma->event == CDMA_EVENT_SYNC_QUEUE_EMPTY)
-				signal = true;
-			break;
-		}
-
+	list_for_each_entry_safe(job, n, &cdma->sync_queue, list) {
 		BUG_ON(job->syncpt_id == NVSYNCPT_INVALID);
 
 		/* Check whether this syncpt has completed, and bail if not */
@@ -227,12 +189,14 @@ static void update_cdma_locked(struct nvhost_cdma *cdma)
 				signal = true;
 		}
 
+		list_del(&job->list);
 		nvhost_job_put(job);
-		kfifo_skip(&cdma->sync_queue);
-		if (cdma->event == CDMA_EVENT_SYNC_QUEUE_SPACE)
-			signal = true;
 	}
 
+	if (list_empty(&cdma->sync_queue) &&
+				cdma->event == CDMA_EVENT_SYNC_QUEUE_EMPTY)
+			signal = true;
+
 	/* Wake up CdmaWait() if the requested event happened */
 	if (signal) {
 		cdma->event = CDMA_EVENT_NONE;
@@ -246,18 +210,14 @@ void nvhost_cdma_update_sync_queue(struct nvhost_cdma *cdma,
 	u32 get_restart;
 	u32 syncpt_incrs;
 	bool exec_ctxsave;
-	unsigned int queue_restore;
 	struct nvhost_job *job = NULL;
-	int result;
 	u32 syncpt_val;
 
 	syncpt_val = nvhost_syncpt_update_min(syncpt, cdma->timeout.syncpt_id);
-	queue_restore = kfifo_save(&cdma->sync_queue);
 
 	dev_dbg(dev,
-		"%s: starting cleanup (thresh %d, queue length %d)\n",
-		__func__,
-		syncpt_val, kfifo_len(&cdma->sync_queue));
+		"%s: starting cleanup (thresh %d)\n",
+		__func__, syncpt_val);
 
 	/*
 	 * Move the sync_queue read pointer to the first entry that hasn't
@@ -270,11 +230,11 @@ void nvhost_cdma_update_sync_queue(struct nvhost_cdma *cdma,
 		"%s: skip completed buffers still in sync_queue\n",
 		__func__);
 
-	result = kfifo_peek(&cdma->sync_queue, &job);
-	while (result && syncpt_val >= job->syncpt_end) {
+	list_for_each_entry(job, &cdma->sync_queue, list) {
+		if (syncpt_val < job->syncpt_end)
+			break;
+
 		nvhost_job_dump(dev, job);
-		kfifo_skip(&cdma->sync_queue);
-		result = kfifo_peek(&cdma->sync_queue, &job);
 	}
 
 	/*
@@ -297,11 +257,11 @@ void nvhost_cdma_update_sync_queue(struct nvhost_cdma *cdma,
 		__func__);
 
 	get_restart = cdma->last_put;
-	if (kfifo_len(&cdma->sync_queue) > 0)
+	if (!list_empty(&cdma->sync_queue))
 		get_restart = job->first_get;
 
 	/* do CPU increments as long as this context continues */
-	while (result && job->clientid == cdma->timeout.clientid) {
+	list_for_each_entry_from(job, &cdma->sync_queue, list) {
 		/* different context, gets us out of this loop */
 		if (job->clientid != cdma->timeout.clientid)
 			break;
@@ -323,8 +283,6 @@ void nvhost_cdma_update_sync_queue(struct nvhost_cdma *cdma,
 				job->num_slots);
 
 		syncpt_val += syncpt_incrs;
-		kfifo_skip(&cdma->sync_queue);
-		result = kfifo_peek(&cdma->sync_queue, &job);
 	}
 
 	dev_dbg(dev,
@@ -334,7 +292,7 @@ void nvhost_cdma_update_sync_queue(struct nvhost_cdma *cdma,
 	exec_ctxsave = false;
 
 	/* setup GPU increments */
-	while (result) {
+	list_for_each_entry_from(job, &cdma->sync_queue, list) {
 		/* same context, increment in the pushbuffer */
 		if (job->clientid == cdma->timeout.clientid) {
 			/* won't need a timeout when replayed */
@@ -360,16 +318,11 @@ void nvhost_cdma_update_sync_queue(struct nvhost_cdma *cdma,
 		}
 
 		nvhost_job_dump(dev, job);
-
-		kfifo_skip(&cdma->sync_queue);
-		result = kfifo_peek(&cdma->sync_queue, &job);
 	}
 
 	dev_dbg(dev,
 		"%s: finished sync_queue modification\n", __func__);
 
-	kfifo_restore(&cdma->sync_queue, queue_restore);
-
 	/* roll back DMAGET and start up channel again */
 	cdma_op(cdma).timeout_teardown_end(cdma, get_restart);
 
@@ -388,12 +341,7 @@ int nvhost_cdma_init(struct nvhost_cdma *cdma)
 	mutex_init(&cdma->lock);
 	sema_init(&cdma->sem, 0);
 
-	err = kfifo_alloc(&cdma->sync_queue,
-			cdma_to_dev(cdma)->sync_queue_size
-				* sizeof(struct nvhost_job *),
-			GFP_KERNEL);
-	if (err)
-		return err;
+	INIT_LIST_HEAD(&cdma->sync_queue);
 
 	cdma->event = CDMA_EVENT_NONE;
 	cdma->running = false;
@@ -414,7 +362,6 @@ void nvhost_cdma_deinit(struct nvhost_cdma *cdma)
 
 	BUG_ON(!cdma_pb_op(cdma).destroy);
 	BUG_ON(cdma->running);
-	kfifo_free(&cdma->sync_queue);
 	cdma_pb_op(cdma).destroy(pb);
 	cdma_op(cdma).timeout_destroy(cdma);
 }
@@ -482,23 +429,20 @@ void nvhost_cdma_push_gather(struct nvhost_cdma *cdma,
 
 /**
  * End a cdma submit
- * Kick off DMA, add a contiguous block of memory handles to the sync queue,
- * and a number of slots to be freed from the pushbuffer.
- * Blocks as necessary if the sync queue is full.
- * The handles for a submit must all be pinned at the same time, but they
- * can be unpinned in smaller chunks.
+ * Kick off DMA, add job to the sync queue, and a number of slots to be freed
+ * from the pushbuffer. The handles for a submit must all be pinned at the same
+ * time, but they can be unpinned in smaller chunks.
  */
 void nvhost_cdma_end(struct nvhost_cdma *cdma,
 		struct nvhost_job *job)
 {
-	bool was_idle = kfifo_len(&cdma->sync_queue) == 0;
+	bool was_idle = list_empty(&cdma->sync_queue);
 
 	BUG_ON(!cdma_op(cdma).kick);
 	cdma_op(cdma).kick(cdma);
 
 	BUG_ON(job->syncpt_id == NVSYNCPT_INVALID);
 
-	nvhost_cdma_wait_locked(cdma, CDMA_EVENT_SYNC_QUEUE_SPACE);
 	add_to_sync_queue(cdma,
 			job,
 			cdma->slots_used,
diff --git a/drivers/video/tegra/host/nvhost_cdma.h b/drivers/video/tegra/host/nvhost_cdma.h
index a5ef8da29e78..48c8cb2e8a26 100644
--- a/drivers/video/tegra/host/nvhost_cdma.h
+++ b/drivers/video/tegra/host/nvhost_cdma.h
@@ -28,7 +28,7 @@
 
 #include <linux/nvhost.h>
 #include <mach/nvmap.h>
-#include <linux/kfifo.h>
+#include <linux/list.h>
 
 #include "nvhost_acm.h"
 
@@ -87,7 +87,6 @@ struct buffer_timeout {
 enum cdma_event {
 	CDMA_EVENT_NONE,		/* not waiting for any event */
 	CDMA_EVENT_SYNC_QUEUE_EMPTY,	/* wait for empty sync queue */
-	CDMA_EVENT_SYNC_QUEUE_SPACE,	/* wait for space in sync queue */
 	CDMA_EVENT_PUSH_BUFFER_SPACE	/* wait for space in push buffer */
 };
 
@@ -101,7 +100,7 @@ struct nvhost_cdma {
 	unsigned int last_put;		/* last value written to DMAPUT */
 	struct push_buffer push_buffer;	/* channel's push buffer */
 	struct syncpt_buffer syncpt_buffer; /* syncpt incr buffer */
-	DECLARE_KFIFO_PTR(sync_queue, struct nvhost_job *); /* job queue */
+	struct list_head sync_queue;	/* job queue */
 	struct buffer_timeout timeout;	/* channel's timeout state/wq */
 	bool running;
 	bool torndown;
diff --git a/drivers/video/tegra/host/nvhost_job.h b/drivers/video/tegra/host/nvhost_job.h
index d00d60f2164a..09b96abb2cfd 100644
--- a/drivers/video/tegra/host/nvhost_job.h
+++ b/drivers/video/tegra/host/nvhost_job.h
@@ -38,6 +38,9 @@ struct nvhost_job {
 	/* When refcount goes to zero, job can be freed */
 	struct kref ref;
 
+	/* List entry */
+	struct list_head list;
+
 	/* Channel where job is submitted to */
 	struct nvhost_channel *ch;