gpu: host1x: Rewrite syncpoint interrupt handling

Move from the old, complex intr handling code to a new implementation
based on dma_fences. While there is a fair bit of churn to get there,
the new implementation is much simpler and likely faster as well due
to allowing signaling directly from interrupt context.
Signed-off-by: Mikko Perttunen <mperttunen@nvidia.com>
Signed-off-by: Thierry Reding <treding@nvidia.com>

drivers/gpu/host1x/debug.c

@@ -77,6 +77,7 @@ static int show_channel(struct host1x_channel *ch, void *data, bool show_fifo)
 
 static void show_syncpts(struct host1x *m, struct output *o, bool show_all)
 {
+	unsigned long irqflags;
 	struct list_head *pos;
 	unsigned int i;
 	int err;
@@ -92,10 +93,10 @@ static void show_syncpts(struct host1x *m, struct output *o, bool show_all)
 		u32 min = host1x_syncpt_load(m->syncpt + i);
 		unsigned int waiters = 0;
 
-		spin_lock(&m->syncpt[i].intr.lock);
-		list_for_each(pos, &m->syncpt[i].intr.wait_head)
+		spin_lock_irqsave(&m->syncpt[i].fences.lock, irqflags);
+		list_for_each(pos, &m->syncpt[i].fences.list)
 			waiters++;
-		spin_unlock(&m->syncpt[i].intr.lock);
+		spin_unlock_irqrestore(&m->syncpt[i].fences.lock, irqflags);
 
 		if (!kref_read(&m->syncpt[i].ref))
 			continue;

drivers/gpu/host1x/dev.c

@@ -516,7 +516,7 @@ static int host1x_probe(struct platform_device *pdev)
 			return PTR_ERR(host->regs);
 	}
 
-	syncpt_irq = platform_get_irq(pdev, 0);
+	host->syncpt_irq = platform_get_irq(pdev, 0);
 	if (syncpt_irq < 0)
 		return syncpt_irq;
 
@@ -578,7 +578,7 @@ static int host1x_probe(struct platform_device *pdev)
 		goto free_contexts;
 	}
 
-	err = host1x_intr_init(host, syncpt_irq);
+	err = host1x_intr_init(host);
 	if (err) {
 		dev_err(&pdev->dev, "failed to initialize interrupts\n");
 		goto deinit_syncpt;

drivers/gpu/host1x/dev.h

@@ -74,8 +74,7 @@ struct host1x_syncpt_ops {
 };
 
 struct host1x_intr_ops {
-	int (*init_host_sync)(struct host1x *host, u32 cpm,
-		void (*syncpt_thresh_work)(struct work_struct *work));
+	int (*init_host_sync)(struct host1x *host, u32 cpm);
 	void (*set_syncpt_threshold)(
 		struct host1x *host, unsigned int id, u32 thresh);
 	void (*enable_syncpt_intr)(struct host1x *host, unsigned int id);
@@ -125,6 +124,7 @@ struct host1x {
 	void __iomem *regs;
 	void __iomem *hv_regs; /* hypervisor region */
 	void __iomem *common_regs;
+	int syncpt_irq;
 	struct host1x_syncpt *syncpt;
 	struct host1x_syncpt_base *bases;
 	struct device *dev;
@@ -138,7 +138,6 @@ struct host1x {
 	dma_addr_t iova_end;
 
 	struct mutex intr_mutex;
-	int intr_syncpt_irq;
 
 	const struct host1x_syncpt_ops *syncpt_op;
 	const struct host1x_intr_ops *intr_op;
@@ -216,10 +215,9 @@ static inline void host1x_hw_syncpt_enable_protection(struct host1x *host)
 	return host->syncpt_op->enable_protection(host);
 }
 
-static inline int host1x_hw_intr_init_host_sync(struct host1x *host, u32 cpm,
-			void (*syncpt_thresh_work)(struct work_struct *))
+static inline int host1x_hw_intr_init_host_sync(struct host1x *host, u32 cpm)
 {
-	return host->intr_op->init_host_sync(host, cpm, syncpt_thresh_work);
+	return host->intr_op->init_host_sync(host, cpm);
 }
 
 static inline void host1x_hw_intr_set_syncpt_threshold(struct host1x *host,

drivers/gpu/host1x/fence.c

@@ -15,22 +15,6 @@
 #include "intr.h"
 #include "syncpt.h"
 
-static DEFINE_SPINLOCK(lock);
-
-struct host1x_syncpt_fence {
-	struct dma_fence base;
-
-	atomic_t signaling;
-
-	struct host1x_syncpt *sp;
-	u32 threshold;
-
-	struct host1x_waitlist *waiter;
-	void *waiter_ref;
-
-	struct delayed_work timeout_work;
-};
-
 static const char *host1x_syncpt_fence_get_driver_name(struct dma_fence *f)
 {
 	return "host1x";
@@ -49,11 +33,12 @@ static struct host1x_syncpt_fence *to_host1x_fence(struct dma_fence *f)
 static bool host1x_syncpt_fence_enable_signaling(struct dma_fence *f)
 {
 	struct host1x_syncpt_fence *sf = to_host1x_fence(f);
-	int err;
 
 	if (host1x_syncpt_is_expired(sf->sp, sf->threshold))
 		return false;
 
+	/* One reference for interrupt path, one for timeout path. */
+	dma_fence_get(f);
 	dma_fence_get(f);
 
 	/*
@@ -61,24 +46,13 @@ static bool host1x_syncpt_fence_enable_signaling(struct dma_fence *f)
 	 * reference to any fences for which 'enable_signaling' has been
 	 * called (and that have not been signalled).
 	 *
-	 * We provide a userspace API to create arbitrary syncpoint fences,
-	 * so we cannot normally guarantee that all fences get signalled.
+	 * We cannot (for now) normally guarantee that all fences get signalled.
 	 * As such, setup a timeout, so that long-lasting fences will get
 	 * reaped eventually.
 	 */
 	schedule_delayed_work(&sf->timeout_work, msecs_to_jiffies(30000));
 
-	err = host1x_intr_add_action(sf->sp->host, sf->sp, sf->threshold,
-				     HOST1X_INTR_ACTION_SIGNAL_FENCE, f,
-				     sf->waiter, &sf->waiter_ref);
-	if (err) {
-		cancel_delayed_work_sync(&sf->timeout_work);
-		dma_fence_put(f);
-		return false;
-	}
-
-	/* intr framework takes ownership of waiter */
-	sf->waiter = NULL;
+	host1x_intr_add_fence_locked(sf->sp->host, sf);
 
 	/*
 	 * The fence may get signalled at any time after the above call,
@@ -89,37 +63,32 @@ static bool host1x_syncpt_fence_enable_signaling(struct dma_fence *f)
 	return true;
 }
 
-static void host1x_syncpt_fence_release(struct dma_fence *f)
-{
-	struct host1x_syncpt_fence *sf = to_host1x_fence(f);
-
-	if (sf->waiter)
-		kfree(sf->waiter);
-
-	dma_fence_free(f);
-}
-
 static const struct dma_fence_ops host1x_syncpt_fence_ops = {
 	.get_driver_name = host1x_syncpt_fence_get_driver_name,
 	.get_timeline_name = host1x_syncpt_fence_get_timeline_name,
 	.enable_signaling = host1x_syncpt_fence_enable_signaling,
-	.release = host1x_syncpt_fence_release,
 };
 
 void host1x_fence_signal(struct host1x_syncpt_fence *f)
 {
-	if (atomic_xchg(&f->signaling, 1))
+	if (atomic_xchg(&f->signaling, 1)) {
+		/*
+		 * Already on timeout path, but we removed the fence before
+		 * timeout path could, so drop interrupt path reference.
+		 */
+		dma_fence_put(&f->base);
 		return;
+	}
 
-	/*
-	 * Cancel pending timeout work - if it races, it will
-	 * not get 'f->signaling' and return.
-	 */
-	cancel_delayed_work_sync(&f->timeout_work);
-
-	host1x_intr_put_ref(f->sp->host, f->sp->id, f->waiter_ref, false);
+	if (cancel_delayed_work(&f->timeout_work)) {
+		/*
+		 * We know that the timeout path will not be entered.
+		 * Safe to drop the timeout path's reference now.
+		 */
+		dma_fence_put(&f->base);
+	}
 
-	dma_fence_signal(&f->base);
+	dma_fence_signal_locked(&f->base);
 	dma_fence_put(&f->base);
 }
 
@@ -129,17 +98,24 @@ static void do_fence_timeout(struct work_struct *work)
 	struct host1x_syncpt_fence *f =
 		container_of(dwork, struct host1x_syncpt_fence, timeout_work);
 
-	if (atomic_xchg(&f->signaling, 1))
+	if (atomic_xchg(&f->signaling, 1)) {
+		/* Already on interrupt path, drop timeout path reference. */
+		dma_fence_put(&f->base);
 		return;
+	}
 
-	/*
-	 * Cancel pending timeout work - if it races, it will
-	 * not get 'f->signaling' and return.
-	 */
-	host1x_intr_put_ref(f->sp->host, f->sp->id, f->waiter_ref, true);
+	if (host1x_intr_remove_fence(f->sp->host, f)) {
+		/*
+		 * Managed to remove fence from queue, so it's safe to drop
+		 * the interrupt path's reference.
+		 */
+		dma_fence_put(&f->base);
+	}
 
 	dma_fence_set_error(&f->base, -ETIMEDOUT);
 	dma_fence_signal(&f->base);
+
+	/* Drop timeout path reference. */
 	dma_fence_put(&f->base);
 }
 
@@ -151,16 +127,10 @@ struct dma_fence *host1x_fence_create(struct host1x_syncpt *sp, u32 threshold)
 	if (!fence)
 		return ERR_PTR(-ENOMEM);
 
-	fence->waiter = kzalloc(sizeof(*fence->waiter), GFP_KERNEL);
-	if (!fence->waiter) {
-		kfree(fence);
-		return ERR_PTR(-ENOMEM);
-	}
-
 	fence->sp = sp;
 	fence->threshold = threshold;
 
-	dma_fence_init(&fence->base, &host1x_syncpt_fence_ops, &lock,
+	dma_fence_init(&fence->base, &host1x_syncpt_fence_ops, &sp->fences.lock,
 		       dma_fence_context_alloc(1), 0);
 
 	INIT_DELAYED_WORK(&fence->timeout_work, do_fence_timeout);

drivers/gpu/host1x/fence.h

@@ -6,7 +6,23 @@
 #ifndef HOST1X_FENCE_H
 #define HOST1X_FENCE_H
 
-struct host1x_syncpt_fence;
+struct host1x_syncpt_fence {
+	struct dma_fence base;
+
+	atomic_t signaling;
+
+	struct host1x_syncpt *sp;
+	u32 threshold;
+
+	struct delayed_work timeout_work;
+
+	struct list_head list;
+};
+
+struct host1x_fence_list {
+	spinlock_t lock;
+	struct list_head list;
+};
 
 void host1x_fence_signal(struct host1x_syncpt_fence *fence);
 

drivers/gpu/host1x/hw/intr_hw.c

@@ -13,23 +13,6 @@
 #include "../intr.h"
 #include "../dev.h"
 
-/*
- * Sync point threshold interrupt service function
- * Handles sync point threshold triggers, in interrupt context
- */
-static void host1x_intr_syncpt_handle(struct host1x_syncpt *syncpt)
-{
-	unsigned int id = syncpt->id;
-	struct host1x *host = syncpt->host;
-
-	host1x_sync_writel(host, BIT(id % 32),
-		HOST1X_SYNC_SYNCPT_THRESH_INT_DISABLE(id / 32));
-	host1x_sync_writel(host, BIT(id % 32),
-		HOST1X_SYNC_SYNCPT_THRESH_CPU0_INT_STATUS(id / 32));
-
-	schedule_work(&syncpt->intr.work);
-}
-
 static irqreturn_t syncpt_thresh_isr(int irq, void *dev_id)
 {
 	struct host1x *host = dev_id;
@@ -39,17 +22,20 @@ static irqreturn_t syncpt_thresh_isr(int irq, void *dev_id)
 	for (i = 0; i < DIV_ROUND_UP(host->info->nb_pts, 32); i++) {
 		reg = host1x_sync_readl(host,
 			HOST1X_SYNC_SYNCPT_THRESH_CPU0_INT_STATUS(i));
-		for_each_set_bit(id, &reg, 32) {
-			struct host1x_syncpt *syncpt =
-				host->syncpt + (i * 32 + id);
-			host1x_intr_syncpt_handle(syncpt);
-		}
+
+		host1x_sync_writel(host, reg,
+			HOST1X_SYNC_SYNCPT_THRESH_INT_DISABLE(i));
+		host1x_sync_writel(host, reg,
+			HOST1X_SYNC_SYNCPT_THRESH_CPU0_INT_STATUS(i));
+
+		for_each_set_bit(id, &reg, 32)
+			host1x_intr_handle_interrupt(host, i * 32 + id);
 	}
 
 	return IRQ_HANDLED;
 }
 
-static void _host1x_intr_disable_all_syncpt_intrs(struct host1x *host)
+static void host1x_intr_disable_all_syncpt_intrs(struct host1x *host)
 {
 	unsigned int i;
 
@@ -90,45 +76,38 @@ static void intr_hw_init(struct host1x *host, u32 cpm)
 }
 
 static int
-_host1x_intr_init_host_sync(struct host1x *host, u32 cpm,
-			    void (*syncpt_thresh_work)(struct work_struct *))
+host1x_intr_init_host_sync(struct host1x *host, u32 cpm)
 {
-	unsigned int i;
 	int err;
 
 	host1x_hw_intr_disable_all_syncpt_intrs(host);
 
-	for (i = 0; i < host->info->nb_pts; i++)
-		INIT_WORK(&host->syncpt[i].intr.work, syncpt_thresh_work);
-
-	err = devm_request_irq(host->dev, host->intr_syncpt_irq,
+	err = devm_request_irq(host->dev, host->syncpt_irq,
 			       syncpt_thresh_isr, IRQF_SHARED,
 			       "host1x_syncpt", host);
-	if (err < 0) {
-		WARN_ON(1);
+	if (err < 0)
 		return err;
-	}
 
 	intr_hw_init(host, cpm);
 
 	return 0;
 }
 
-static void _host1x_intr_set_syncpt_threshold(struct host1x *host,
+static void host1x_intr_set_syncpt_threshold(struct host1x *host,
 					      unsigned int id,
 					      u32 thresh)
 {
 	host1x_sync_writel(host, thresh, HOST1X_SYNC_SYNCPT_INT_THRESH(id));
 }
 
-static void _host1x_intr_enable_syncpt_intr(struct host1x *host,
+static void host1x_intr_enable_syncpt_intr(struct host1x *host,
 					    unsigned int id)
 {
 	host1x_sync_writel(host, BIT(id % 32),
 		HOST1X_SYNC_SYNCPT_THRESH_INT_ENABLE_CPU0(id / 32));
 }
 
-static void _host1x_intr_disable_syncpt_intr(struct host1x *host,
+static void host1x_intr_disable_syncpt_intr(struct host1x *host,
 					     unsigned int id)
 {
 	host1x_sync_writel(host, BIT(id % 32),
@@ -137,23 +116,10 @@ static void _host1x_intr_disable_syncpt_intr(struct host1x *host,
 		HOST1X_SYNC_SYNCPT_THRESH_CPU0_INT_STATUS(id / 32));
 }
 
-static int _host1x_free_syncpt_irq(struct host1x *host)
-{
-	unsigned int i;
-
-	devm_free_irq(host->dev, host->intr_syncpt_irq, host);
-
-	for (i = 0; i < host->info->nb_pts; i++)
-		cancel_work_sync(&host->syncpt[i].intr.work);
-
-	return 0;
-}
-
 static const struct host1x_intr_ops host1x_intr_ops = {
-	.init_host_sync = _host1x_intr_init_host_sync,
-	.set_syncpt_threshold = _host1x_intr_set_syncpt_threshold,
-	.enable_syncpt_intr = _host1x_intr_enable_syncpt_intr,
-	.disable_syncpt_intr = _host1x_intr_disable_syncpt_intr,
-	.disable_all_syncpt_intrs = _host1x_intr_disable_all_syncpt_intrs,
-	.free_syncpt_irq = _host1x_free_syncpt_irq,
+	.init_host_sync = host1x_intr_init_host_sync,
+	.set_syncpt_threshold = host1x_intr_set_syncpt_threshold,
+	.enable_syncpt_intr = host1x_intr_enable_syncpt_intr,
+	.disable_syncpt_intr = host1x_intr_disable_syncpt_intr,
+	.disable_all_syncpt_intrs = host1x_intr_disable_all_syncpt_intrs,
 };

drivers/gpu/host1x/intr.h

@@ -2,87 +2,17 @@
 /*
  * Tegra host1x Interrupt Management
  *
- * Copyright (c) 2010-2013, NVIDIA Corporation.
+ * Copyright (c) 2010-2021, NVIDIA Corporation.
  */
 
 #ifndef __HOST1X_INTR_H
 #define __HOST1X_INTR_H
 
-#include <linux/interrupt.h>
-#include <linux/workqueue.h>
-
-struct host1x_syncpt;
 struct host1x;
-
-enum host1x_intr_action {
-	/*
-	 * Perform cleanup after a submit has completed.
-	 * 'data' points to a channel
-	 */
-	HOST1X_INTR_ACTION_SUBMIT_COMPLETE = 0,
-
-	/*
-	 * Wake up a  task.
-	 * 'data' points to a wait_queue_head_t
-	 */
-	HOST1X_INTR_ACTION_WAKEUP,
-
-	/*
-	 * Wake up a interruptible task.
-	 * 'data' points to a wait_queue_head_t
-	 */
-	HOST1X_INTR_ACTION_WAKEUP_INTERRUPTIBLE,
-
-	HOST1X_INTR_ACTION_SIGNAL_FENCE,
-
-	HOST1X_INTR_ACTION_COUNT
-};
-
-struct host1x_syncpt_intr {
-	spinlock_t lock;
-	struct list_head wait_head;
-	char thresh_irq_name[12];
-	struct work_struct work;
-};
-
-struct host1x_waitlist {
-	struct list_head list;
-	struct kref refcount;
-	u32 thresh;
-	enum host1x_intr_action action;
-	atomic_t state;
-	void *data;
-	int count;
-};
-
-/*
- * Schedule an action to be taken when a sync point reaches the given threshold.
- *
- * @id the sync point
- * @thresh the threshold
- * @action the action to take
- * @data a pointer to extra data depending on action, see above
- * @waiter waiter structure - assumes ownership
- * @ref must be passed if cancellation is possible, else NULL
- *
- * This is a non-blocking api.
- */
-int host1x_intr_add_action(struct host1x *host, struct host1x_syncpt *syncpt,
-			   u32 thresh, enum host1x_intr_action action,
-			   void *data, struct host1x_waitlist *waiter,
-			   void **ref);
-
-/*
- * Unreference an action submitted to host1x_intr_add_action().
- * You must call this if you passed non-NULL as ref.
- * @ref the ref returned from host1x_intr_add_action()
- * @flush wait until any pending handlers have completed before returning.
- */
-void host1x_intr_put_ref(struct host1x *host, unsigned int id, void *ref,
-			 bool flush);
+struct host1x_syncpt_fence;
 
 /* Initialize host1x sync point interrupt */
-int host1x_intr_init(struct host1x *host, unsigned int irq_sync);
+int host1x_intr_init(struct host1x *host);
 
 /* Deinitialize host1x sync point interrupt */
 void host1x_intr_deinit(struct host1x *host);
@@ -93,5 +23,10 @@ void host1x_intr_start(struct host1x *host);
 /* Disable host1x sync point interrupt */
 void host1x_intr_stop(struct host1x *host);
 
-irqreturn_t host1x_syncpt_thresh_fn(void *dev_id);
+void host1x_intr_handle_interrupt(struct host1x *host, unsigned int id);
+
+void host1x_intr_add_fence_locked(struct host1x *host, struct host1x_syncpt_fence *fence);
+
+bool host1x_intr_remove_fence(struct host1x *host, struct host1x_syncpt_fence *fence);
+
 #endif

drivers/gpu/host1x/syncpt.h

@@ -14,6 +14,7 @@
 #include <linux/kref.h>
 #include <linux/sched.h>
 
+#include "fence.h"
 #include "intr.h"
 
 struct host1x;
@@ -39,7 +40,7 @@ struct host1x_syncpt {
 	struct host1x_syncpt_base *base;
 
 	/* interrupt data */
-	struct host1x_syncpt_intr intr;
+	struct host1x_fence_list fences;
 
 	/*
 	 * If a submission incrementing this syncpoint fails, lock it so that