tegra: host: move stale wait checking into the kernel

The kernel now receives wait tracking data (similar to gathers and
relocs) and compares the current syncpt with the threshold value.

If it's old, it gets a kernel mapping and rewrites the method data
to use a kernel reserved syncpt that is always 0 (so trivially pops
when seen by the HW).

Patch has dependency to the user-space patches

Submitted on behalf of: Chris Johnson <cjohnson@nvidia.com>
original work by: Chris Johnson <cjohnson@nvidia.com>

Change-Id: I4d4e5d3b49cab860485c4172f87247f5b4f5ea6e

8 files changed, 177 insertions, 3 deletions

diff --git a/arch/arm/mach-tegra/include/mach/nvhost.h b/arch/arm/mach-tegra/include/mach/nvhost.h
index c72666ae0d37..4bec3127a545 100644
--- a/arch/arm/mach-tegra/include/mach/nvhost.h
+++ b/arch/arm/mach-tegra/include/mach/nvhost.h
@@ -81,6 +81,8 @@ struct nvhost_submit_hdr {
 	__u32 syncpt_incrs;
 	__u32 num_cmdbufs;
 	__u32 num_relocs;
+	__u32 num_waitchks;
+	__u32 waitchk_mask;
 };
 
 struct nvhost_cmdbuf {
@@ -96,6 +98,13 @@ struct nvhost_reloc {
 	__u32 target_offset;
 };
 
+struct nvhost_waitchk {
+	__u32 mem;
+	__u32 offset;
+	__u32 syncpt_id;
+	__u32 thresh;
+};
+
 struct nvhost_get_param_args {
 	__u32 value;
 };
diff --git a/arch/arm/mach-tegra/include/mach/nvmap.h b/arch/arm/mach-tegra/include/mach/nvmap.h
index 7422d1a44d8a..7a79748e5433 100644
--- a/arch/arm/mach-tegra/include/mach/nvmap.h
+++ b/arch/arm/mach-tegra/include/mach/nvmap.h
@@ -97,6 +97,10 @@ int nvmap_pin_array(struct nvmap_client *client, struct nvmap_handle *gather,
 void nvmap_unpin_handles(struct nvmap_client *client,
 			 struct nvmap_handle **h, int nr);
 
+int nvmap_patch_wait(struct nvmap_client *client,
+		     struct nvmap_handle *patch,
+		     u32 patch_offset, u32 patch_value);
+
 struct nvmap_platform_carveout {
 	const char *name;
 	unsigned int usage_mask;
diff --git a/drivers/video/tegra/host/dev.c b/drivers/video/tegra/host/dev.c
index 20a4eda0fb53..2f3755870e48 100644
--- a/drivers/video/tegra/host/dev.c
+++ b/drivers/video/tegra/host/dev.c
@@ -50,6 +50,8 @@ struct nvhost_channel_userctx {
 	u32 syncpt_incrs;
 	u32 cmdbufs_pending;
 	u32 relocs_pending;
+	u32 waitchk_pending;
+	u32 waitchk_ref;
 	struct nvmap_handle_ref *gather_mem;
 	struct nvhost_op_pair *gathers;
 	int num_gathers;
@@ -57,6 +59,9 @@ struct nvhost_channel_userctx {
 	struct nvmap_pinarray_elem pinarray[NVHOST_MAX_HANDLES];
 	struct nvmap_handle *unpinarray[NVHOST_MAX_HANDLES];
 	struct nvmap_client *nvmap;
+	struct nvhost_waitchk waitchks[NVHOST_MAX_WAIT_CHECKS];
+	u32 num_waitchks;
+	u32 waitchk_mask;
 };
 
 struct nvhost_ctrl_userctx {
@@ -141,6 +146,7 @@ static void reset_submit(struct nvhost_channel_userctx *ctx)
 {
 	ctx->cmdbufs_pending = 0;
 	ctx->relocs_pending = 0;
+	ctx->waitchk_pending = 0;
 }
 
 static ssize_t nvhost_channelwrite(struct file *filp, const char __user *buf,
@@ -152,7 +158,7 @@ static ssize_t nvhost_channelwrite(struct file *filp, const char __user *buf,
 
 	while (remaining) {
 		size_t consumed;
-		if (!priv->relocs_pending && !priv->cmdbufs_pending) {
+		if (!priv->relocs_pending && !priv->cmdbufs_pending && !priv->waitchk_pending) {
 			consumed = sizeof(struct nvhost_submit_hdr);
 			if (remaining < consumed)
 				break;
@@ -167,6 +173,7 @@ static ssize_t nvhost_channelwrite(struct file *filp, const char __user *buf,
 			/* leave room for ctx switch */
 			priv->num_gathers = 2;
 			priv->pinarray_size = 0;
+			priv->waitchk_mask |= priv->waitchk_ref;
 		} else if (priv->cmdbufs_pending) {
 			struct nvhost_cmdbuf cmdbuf;
 			consumed = sizeof(cmdbuf);
@@ -192,6 +199,18 @@ static ssize_t nvhost_channelwrite(struct file *filp, const char __user *buf,
 			}
 			priv->pinarray_size += numrelocs;
 			priv->relocs_pending -= numrelocs;
+		} else if (priv->waitchk_pending) {
+			struct nvhost_waitchk *waitp;
+			consumed = sizeof(struct nvhost_waitchk);
+			if (remaining < consumed)
+				break;
+			waitp = &priv->waitchks[priv->num_waitchks];
+			if (copy_from_user(waitp, buf, consumed)) {
+				err = -EFAULT;
+				break;
+			}
+			priv->num_waitchks++;
+			priv->waitchk_pending--;
 		} else {
 			err = -EFAULT;
 			break;
@@ -219,7 +238,7 @@ static int nvhost_ioctl_channel_flush(struct nvhost_channel_userctx *ctx,
 	int num_unpin;
 	int err;
 
-	if (ctx->relocs_pending || ctx->cmdbufs_pending) {
+	if (ctx->relocs_pending || ctx->cmdbufs_pending || ctx->waitchk_pending) {
 		reset_submit(ctx);
 		dev_err(&ctx->ch->dev->pdev->dev, "channel submit out of sync\n");
 		return -EFAULT;
@@ -254,6 +273,23 @@ static int nvhost_ioctl_channel_flush(struct nvhost_channel_userctx *ctx,
 		return err;
 	}
 
+	/* remove stale waits */
+	if (ctx->num_waitchks) {
+		err = nvhost_syncpt_wait_check(ctx->nvmap,
+				&ctx->ch->dev->syncpt, ctx->waitchk_mask,
+				ctx->waitchks, ctx->num_waitchks);
+		if (err) {
+			dev_warn(&ctx->ch->dev->pdev->dev,
+				"nvhost_syncpt_wait_check failed: %d\n", err);
+			mutex_unlock(&ctx->ch->submitlock);
+			nvmap_unpin_handles(ctx->nvmap, ctx->unpinarray, num_unpin);
+			nvhost_module_idle(&ctx->ch->mod);
+			return err;
+		}
+		ctx->num_waitchks = 0;
+		ctx->waitchk_mask = 0;
+	}
+
 	/* context switch */
 	if (ctx->ch->cur_ctx != ctx->hwctx) {
 		struct nvhost_hwctx *hw = ctx->hwctx;
@@ -337,6 +373,8 @@ static long nvhost_channelctl(struct file *filp,
 		err = nvhost_ioctl_channel_flush(priv, (void *)buf);
 		break;
 	case NVHOST_IOCTL_CHANNEL_GET_SYNCPOINTS:
+		/* host syncpt ID is used by the RM (and never be given out) */
+		BUG_ON(priv->ch->desc->syncpts & (1 << NVSYNCPT_GRAPHICS_HOST));
 		((struct nvhost_get_param_args *)buf)->value =
 			priv->ch->desc->syncpts;
 		break;
diff --git a/drivers/video/tegra/host/nvhost_channel.h b/drivers/video/tegra/host/nvhost_channel.h
index c62d7397a192..492d4a2feb69 100644
--- a/drivers/video/tegra/host/nvhost_channel.h
+++ b/drivers/video/tegra/host/nvhost_channel.h
@@ -32,6 +32,7 @@
 
 #define NVHOST_CHANNEL_BASE 0
 #define NVHOST_NUMCHANNELS (NV_HOST1X_CHANNELS - 1)
+#define NVHOST_MAX_WAIT_CHECKS 256
 #define NVHOST_MAX_GATHERS 512
 #define NVHOST_MAX_HANDLES 1280
 
diff --git a/drivers/video/tegra/host/nvhost_hardware.h b/drivers/video/tegra/host/nvhost_hardware.h
index f69f467dd64e..3309ae4d8532 100644
--- a/drivers/video/tegra/host/nvhost_hardware.h
+++ b/drivers/video/tegra/host/nvhost_hardware.h
@@ -141,6 +141,12 @@ enum {
 	NV_CLASS_HOST_INDDATA = 0x2e
 };
 
+static inline u32 nvhost_class_host_wait_syncpt(
+	unsigned indx, unsigned threshold)
+{
+	return (indx << 24) | (threshold & 0xffffff);
+}
+
 static inline u32 nvhost_class_host_wait_syncpt_base(
 	unsigned indx, unsigned base_indx, unsigned offset)
 {
diff --git a/drivers/video/tegra/host/nvhost_syncpt.c b/drivers/video/tegra/host/nvhost_syncpt.c
index dd2ab0d379e0..1881716ed428 100644
--- a/drivers/video/tegra/host/nvhost_syncpt.c
+++ b/drivers/video/tegra/host/nvhost_syncpt.c
@@ -226,7 +226,7 @@ done:
 }
 
 static const char *s_syncpt_names[32] = {
-	"", "", "", "", "", "", "", "", "", "", "", "",
+	"gfx_host", "", "", "", "", "", "", "", "", "", "", "",
 	"vi_isp_0", "vi_isp_1", "vi_isp_2", "vi_isp_3", "vi_isp_4", "vi_isp_5",
 	"2d_0", "2d_1",
 	"", "",
@@ -254,3 +254,63 @@ void nvhost_syncpt_debug(struct nvhost_syncpt *sp)
 
 	}
 }
+
+/* returns true, if a <= b < c using wrapping comparison */
+static inline bool nvhost_syncpt_is_between(u32 a, u32 b, u32 c)
+{
+	return b-a < c-a;
+}
+
+/* returns true, if x >= y (mod 1 << 32) */
+static bool nvhost_syncpt_wrapping_comparison(u32 x, u32 y)
+{
+	return nvhost_syncpt_is_between(y, x, (1UL<<31UL)+y);
+}
+
+/* check for old WAITs to be removed (avoiding a wrap) */
+int nvhost_syncpt_wait_check(struct nvmap_client *nvmap,
+			struct nvhost_syncpt *sp, u32 waitchk_mask,
+			struct nvhost_waitchk *waitp, u32 waitchks)
+{
+	u32 idx;
+	int err = 0;
+
+	/* get current syncpt values */
+	for (idx = 0; idx < NV_HOST1X_SYNCPT_NB_PTS; idx++) {
+		if (BIT(idx) & waitchk_mask) {
+			nvhost_syncpt_update_min(sp, idx);
+		}
+	}
+
+	BUG_ON(!waitp);
+
+	/* compare syncpt vs wait threshold */
+	while (waitchks) {
+		u32 syncpt, override;
+
+		BUG_ON(waitp->syncpt_id > NV_HOST1X_SYNCPT_NB_PTS);
+
+		syncpt = atomic_read(&sp->min_val[waitp->syncpt_id]);
+		if (nvhost_syncpt_wrapping_comparison(syncpt, waitp->thresh)) {
+
+			/* wait has completed already, so can be removed */
+			dev_dbg(&syncpt_to_dev(sp)->pdev->dev,
+					"drop WAIT id %d (%s) thresh 0x%x, syncpt 0x%x\n",
+					waitp->syncpt_id,  nvhost_syncpt_name(waitp->syncpt_id),
+					waitp->thresh, syncpt);
+
+			/* move wait to a kernel reserved syncpt (that's always 0) */
+			override = nvhost_class_host_wait_syncpt(NVSYNCPT_GRAPHICS_HOST, 0);
+
+			/* patch the wait */
+			err = nvmap_patch_wait(nvmap,
+						(struct nvmap_handle *)waitp->mem,
+						waitp->offset, override);
+			if (err)
+				break;
+		}
+		waitchks--;
+		waitp++;
+	}
+	return err;
+}
diff --git a/drivers/video/tegra/host/nvhost_syncpt.h b/drivers/video/tegra/host/nvhost_syncpt.h
index f161f2051406..b4ce3c6ee6d4 100644
--- a/drivers/video/tegra/host/nvhost_syncpt.h
+++ b/drivers/video/tegra/host/nvhost_syncpt.h
@@ -25,10 +25,13 @@
 
 #include <linux/kernel.h>
 #include <linux/sched.h>
+#include <mach/nvhost.h>
+#include <mach/nvmap.h>
 #include <asm/atomic.h>
 
 #include "nvhost_hardware.h"
 
+#define NVSYNCPT_GRAPHICS_HOST		     (0)
 #define NVSYNCPT_VI_ISP_0		     (12)
 #define NVSYNCPT_VI_ISP_1		     (13)
 #define NVSYNCPT_VI_ISP_2		     (14)
@@ -142,6 +145,9 @@ static inline int nvhost_syncpt_wait(struct nvhost_syncpt *sp, u32 id, u32 thres
 	return nvhost_syncpt_wait_timeout(sp, id, thresh, MAX_SCHEDULE_TIMEOUT);
 }
 
+int nvhost_syncpt_wait_check(struct nvmap_client *nvmap,
+			struct nvhost_syncpt *sp, u32 mask,
+			struct nvhost_waitchk *waitp, u32 num_waits);
 
 const char *nvhost_syncpt_name(u32 id);
 
diff --git a/drivers/video/tegra/nvmap/nvmap.c b/drivers/video/tegra/nvmap/nvmap.c
index 1a70f43d79f5..380f70ec187c 100644
--- a/drivers/video/tegra/nvmap/nvmap.c
+++ b/drivers/video/tegra/nvmap/nvmap.c
@@ -724,3 +724,53 @@ void nvmap_free(struct nvmap_client *client, struct nvmap_handle_ref *r)
 {
 	nvmap_free_handle_id(client, nvmap_ref_to_id(r));
 }
+
+/*
+ * create a mapping to the user's buffer and write it
+ * (uses similar logic from nvmap_reloc_pin_array to map the cmdbuf)
+ */
+int nvmap_patch_wait(struct nvmap_client *client,
+				struct nvmap_handle *patch,
+				u32 patch_offset, u32 patch_value)
+{
+	unsigned long phys;
+	unsigned int pfn, last_pfn = 0;
+	void *addr;
+	pte_t **pte;
+
+	if (patch_offset >= patch->size) {
+		nvmap_warn(client, "read/write outside of handle\n");
+		return -EFAULT;
+	}
+
+	pte = nvmap_alloc_pte(client->dev, &addr);
+	if (IS_ERR(pte))
+		return PTR_ERR(pte);
+
+	/* derive physaddr of cmdbuf WAIT to patch */
+	if (patch->heap_pgalloc) {
+		unsigned int page = patch_offset >> PAGE_SHIFT;
+		phys = page_to_phys(patch->pgalloc.pages[page]);
+		phys += (patch_offset & ~PAGE_MASK);
+	} else {
+		phys = patch->carveout->base + patch_offset;
+	}
+
+	pfn = __phys_to_pfn(phys);
+
+	/* write PTE, so addr points to cmdbuf PFN */
+	if (pfn != last_pfn) {
+		pgprot_t prot = nvmap_pgprot(patch, pgprot_kernel);
+		unsigned long kaddr = (unsigned long)addr;
+		set_pte_at(&init_mm, kaddr, *pte, pfn_pte(pfn, prot));
+		flush_tlb_kernel_page(kaddr);
+		last_pfn = pfn;
+	}
+
+	/* write patch_value to addr + page offset */
+	__raw_writel(patch_value, addr + (phys & ~PAGE_MASK));
+
+	nvmap_free_pte(client->dev, pte);
+	wmb();
+	return 0;
+}