nvhost: Fix tegra_host/status debug output

Add tracking of mapping between push buffer gathers and nvmap handles.
Use the mapping to access contents of gather buffers.

Bug 840976

Change-Id: Ic6fe2fc7a83e8f14cd606a9e638b6420aa714495
Signed-off-by: Terje Bergstrom <tbergstrom@nvidia.com>
Reviewed-on: http://git-master/r/44779
Reviewed-by: Varun Colbert <vcolbert@nvidia.com>
Tested-by: Varun Colbert <vcolbert@nvidia.com>

7 files changed, 102 insertions, 39 deletions

diff --git a/drivers/video/tegra/host/chip_support.h b/drivers/video/tegra/host/chip_support.h
index a0b0f0278a0f..eae99671a20d 100644
--- a/drivers/video/tegra/host/chip_support.h
+++ b/drivers/video/tegra/host/chip_support.h
@@ -61,6 +61,7 @@ struct nvhost_chip_support {
 		int (*init)(struct push_buffer *);
 		void (*destroy)(struct push_buffer *);
 		void (*push_to)(struct push_buffer *,
+				struct nvmap_handle *,
 				u32 op1, u32 op2);
 		void (*pop_from)(struct push_buffer *,
 				 unsigned int slots);
diff --git a/drivers/video/tegra/host/nvhost_cdma.c b/drivers/video/tegra/host/nvhost_cdma.c
index 24c0c0fff0c0..fcce8334f272 100644
--- a/drivers/video/tegra/host/nvhost_cdma.c
+++ b/drivers/video/tegra/host/nvhost_cdma.c
@@ -408,17 +408,28 @@ void nvhost_cdma_begin(struct nvhost_cdma *cdma)
  */
 void nvhost_cdma_push(struct nvhost_cdma *cdma, u32 op1, u32 op2)
 {
+	nvhost_cdma_push_gather(cdma, NULL, op1, op2);
+}
+
+/**
+ * Push two words into a push buffer slot
+ * Blocks as necessary if the push buffer is full.
+ */
+void nvhost_cdma_push_gather(struct nvhost_cdma *cdma,
+		struct nvmap_handle *handle, u32 op1, u32 op2)
+{
 	u32 slots_free = cdma->slots_free;
 	struct push_buffer *pb = &cdma->push_buffer;
 	BUG_ON(!cdma_pb_op(cdma).push_to);
 	BUG_ON(!cdma_op(cdma).kick);
 	if (slots_free == 0) {
 		cdma_op(cdma).kick(cdma);
-		slots_free = nvhost_cdma_wait(cdma, CDMA_EVENT_PUSH_BUFFER_SPACE);
+		slots_free = nvhost_cdma_wait(cdma,
+				CDMA_EVENT_PUSH_BUFFER_SPACE);
 	}
 	cdma->slots_free = slots_free - 1;
 	cdma->slots_used++;
-	cdma_pb_op(cdma).push_to(pb, op1, op2);
+	cdma_pb_op(cdma).push_to(pb, handle, op1, op2);
 }
 
 /**
diff --git a/drivers/video/tegra/host/nvhost_cdma.h b/drivers/video/tegra/host/nvhost_cdma.h
index e48d733850a9..45c2f7c57a7f 100644
--- a/drivers/video/tegra/host/nvhost_cdma.h
+++ b/drivers/video/tegra/host/nvhost_cdma.h
@@ -47,11 +47,12 @@
 
 
 struct push_buffer {
-	struct nvmap_handle_ref *mem; /* handle to pushbuffer memory */
-	u32 *mapped;		/* mapped pushbuffer memory */
-	u32 phys;		/* physical address of pushbuffer */
-	u32 fence;		/* index we've written */
-	u32 cur;		/* index to write to */
+	struct nvmap_handle_ref *mem;	/* handle to pushbuffer memory */
+	u32 *mapped;			/* mapped pushbuffer memory */
+	u32 phys;			/* physical address of pushbuffer */
+	u32 fence;			/* index we've written */
+	u32 cur;			/* index to write to */
+	struct nvmap_handle **handles;	/* nvmap handle for each opcode pair */
 };
 
 struct sync_queue {
@@ -93,6 +94,8 @@ void	nvhost_cdma_deinit(struct nvhost_cdma *cdma);
 void	nvhost_cdma_stop(struct nvhost_cdma *cdma);
 void	nvhost_cdma_begin(struct nvhost_cdma *cdma);
 void	nvhost_cdma_push(struct nvhost_cdma *cdma, u32 op1, u32 op2);
+void	nvhost_cdma_push_gather(struct nvhost_cdma *cdma,
+		struct nvmap_handle *handle, u32 op1, u32 op2);
 void	nvhost_cdma_end(struct nvhost_cdma *cdma,
 		struct nvmap_client *user_nvmap,
 		u32 sync_point_id, u32 sync_point_value,
diff --git a/drivers/video/tegra/host/t20/3dctx_t20.c b/drivers/video/tegra/host/t20/3dctx_t20.c
index f3f1e9475efa..dadfbed3434a 100644
--- a/drivers/video/tegra/host/t20/3dctx_t20.c
+++ b/drivers/video/tegra/host/t20/3dctx_t20.c
@@ -392,7 +392,8 @@ static void __init setup_save(const struct ctx_saver *saver, u32 *ptr)
 static void save_push_v0(struct nvhost_cdma *cdma,
 			struct nvhost_hwctx *ctx)
 {
-	nvhost_cdma_push(cdma,
+	nvhost_cdma_push_gather(cdma,
+			nvmap_ref_to_handle(save_buf),
 			nvhost_opcode_gather(save_size),
 			save_phys);
 }
@@ -524,7 +525,8 @@ static void save_push_v1(struct nvhost_cdma *cdma,
 			nvhost_opcode_nonincr(0x904, 1),
 			ctx->restore_phys);
 	/* gather the save buffer */
-	nvhost_cdma_push(cdma,
+	nvhost_cdma_push_gather(cdma,
+			nvmap_ref_to_handle(save_buf),
 			nvhost_opcode_gather(save_size),
 			save_phys);
 }
diff --git a/drivers/video/tegra/host/t20/cdma_t20.c b/drivers/video/tegra/host/t20/cdma_t20.c
index 77f514eb4e01..eaba1c78af92 100644
--- a/drivers/video/tegra/host/t20/cdma_t20.c
+++ b/drivers/video/tegra/host/t20/cdma_t20.c
@@ -20,6 +20,7 @@
  * 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301, USA.
  */
 
+#include <linux/slab.h>
 #include "../nvhost_cdma.h"
 #include "../dev.h"
 
@@ -53,6 +54,7 @@ static int t20_push_buffer_init(struct push_buffer *pb)
 	pb->mem = NULL;
 	pb->mapped = NULL;
 	pb->phys = 0;
+	pb->handles = NULL;
 
 	BUG_ON(!cdma_pb_op(cdma).reset);
 	cdma_pb_op(cdma).reset(pb);
@@ -75,6 +77,11 @@ static int t20_push_buffer_init(struct push_buffer *pb)
 		goto fail;
 	}
 
+	/* memory for storing nvmap handles for each opcode pair */
+	pb->handles = kzalloc(PUSH_BUFFER_SIZE/2, GFP_KERNEL);
+	if (!pb->handles)
+		goto fail;
+
 	/* put the restart at the end of pushbuffer memory */
 	*(pb->mapped + (PUSH_BUFFER_SIZE >> 2)) = nvhost_opcode_restart(pb->phys);
 
@@ -101,22 +108,27 @@ static void t20_push_buffer_destroy(struct push_buffer *pb)
 	if (pb->mem)
 		nvmap_free(nvmap, pb->mem);
 
+	kfree(pb->handles);
+
 	pb->mem = NULL;
 	pb->mapped = NULL;
 	pb->phys = 0;
+	pb->handles = 0;
 }
 
 /**
  * Push two words to the push buffer
  * Caller must ensure push buffer is not full
  */
-static void t20_push_buffer_push_to(struct push_buffer *pb, u32 op1, u32 op2)
+static void t20_push_buffer_push_to(struct push_buffer *pb,
+			struct nvmap_handle *handle, u32 op1, u32 op2)
 {
 	u32 cur = pb->cur;
 	u32 *p = (u32*)((u32)pb->mapped + cur);
 	BUG_ON(cur == pb->fence);
 	*(p++) = op1;
 	*(p++) = op2;
+	pb->handles[cur/8] = handle;
 	pb->cur = (cur + 8) & (PUSH_BUFFER_SIZE - 1);
 	/* printk("push_to_push_buffer: op1=%08x; op2=%08x; cur=%x\n", op1, op2, pb->cur); */
 }
diff --git a/drivers/video/tegra/host/t20/channel_t20.c b/drivers/video/tegra/host/t20/channel_t20.c
index a3fb15212820..fdbf6ba7355d 100644
--- a/drivers/video/tegra/host/t20/channel_t20.c
+++ b/drivers/video/tegra/host/t20/channel_t20.c
@@ -232,7 +232,8 @@ static int t20_channel_submit(struct nvhost_channel *channel,
 
 	/* gather restore buffer */
 	if (need_restore)
-		nvhost_cdma_push(&channel->cdma,
+		nvhost_cdma_push_gather(&channel->cdma,
+			nvmap_ref_to_handle(channel->cur_ctx->restore),
 			nvhost_opcode_gather(channel->cur_ctx->restore_size),
 			channel->cur_ctx->restore_phys);
 
@@ -269,10 +270,13 @@ static int t20_channel_submit(struct nvhost_channel *channel,
 	}
 	else {
 		/* push user gathers */
-		for ( ; gather != gather_end; gather += 2)
-			nvhost_cdma_push(&channel->cdma,
-					nvhost_opcode_gather(gather[0]),
-					gather[1]);
+		int i = 0;
+		for ( ; i < gather_end-gather; i += 2) {
+			nvhost_cdma_push_gather(&channel->cdma,
+					unpins[i/2],
+					nvhost_opcode_gather(gather[i]),
+					gather[i+1]);
+		}
 	}
 
 	/* end CDMA submit & stash pinned hMems into sync queue */
diff --git a/drivers/video/tegra/host/t20/debug_t20.c b/drivers/video/tegra/host/t20/debug_t20.c
index 0c77745b8c41..d3329bf2ed9e 100644
--- a/drivers/video/tegra/host/t20/debug_t20.c
+++ b/drivers/video/tegra/host/t20/debug_t20.c
@@ -19,12 +19,14 @@
 
 #include <linux/debugfs.h>
 #include <linux/seq_file.h>
+#include <linux/mm.h>
 
 #include <asm/io.h>
 
 #include "../dev.h"
 #include "../debug.h"
 #include "../nvhost_cdma.h"
+#include "../../nvmap/nvmap.h"
 
 #include "hardware_t20.h"
 
@@ -34,7 +36,7 @@ enum {
 	NVHOST_DBG_STATE_GATHER = 2
 };
 
-static int show_channel_command(struct output *o, u32 val, int *count)
+static int show_channel_command(struct output *o, u32 addr, u32 val, int *count)
 {
 	unsigned mask;
 	unsigned subop;
@@ -104,11 +106,11 @@ static int show_channel_command(struct output *o, u32 val, int *count)
  * TODO: This uses ioremap_xxx on memory which is deprecated.
  * Also, it won't work properly with SMMU.
  */
-static void show_channel_gather(struct output *o, u32 phys_addr,
-				phys_addr_t words);
+static void show_channel_gather(struct output *o, u32 addr, u32 phys_addr,
+		phys_addr_t words, struct nvhost_cdma *cdma);
 
 static void show_channel_word(struct output *o, int *state, int *count,
-				     u32 addr, u32 val)
+		u32 addr, u32 val, struct nvhost_cdma *cdma)
 {
 	switch (*state) {
 	case NVHOST_DBG_STATE_CMD:
@@ -117,7 +119,7 @@ static void show_channel_word(struct output *o, int *state, int *count,
 		else
 			nvhost_debug_output(o, "%08x:", val);
 
-		*state = show_channel_command(o, val, count);
+		*state = show_channel_command(o, addr, val, count);
 		if (*state == NVHOST_DBG_STATE_DATA && *count == 0) {
 			*state = NVHOST_DBG_STATE_CMD;
 			nvhost_debug_output(o, "])\n");
@@ -134,38 +136,54 @@ static void show_channel_word(struct output *o, int *state, int *count,
 	case NVHOST_DBG_STATE_GATHER:
 		*state = NVHOST_DBG_STATE_CMD;
 		nvhost_debug_output(o, "%08x]):\n", val);
-		show_channel_gather(o, val, *count);
+		if (cdma)
+			show_channel_gather(o, addr, val, *count, cdma);
 		break;
 	}
 }
 
-static void show_channel_gather(struct output *o, phys_addr_t phys_addr,
-				u32 words)
+static void show_channel_gather(struct output *o, u32 addr,
+		phys_addr_t phys_addr,
+		u32 words, struct nvhost_cdma *cdma)
 {
+	/* Calculate page boundary and offset */
 	phys_addr_t map_base = phys_addr & PAGE_MASK;
-	phys_addr_t map_end = (phys_addr + words * 4 + PAGE_SIZE - 1) & PAGE_MASK;
-	phys_addr_t map_size = map_end - map_base;
 	phys_addr_t map_offset = phys_addr - map_base;
-	void *map_addr = ioremap_nocache(map_base, map_size);
-	int state = NVHOST_DBG_STATE_CMD;
-	int count, i;
 
+	/* Map dmaget cursor to corresponding nvmap_handle */
+	struct push_buffer *pb = &cdma->push_buffer;
+	u32 cur = addr - pb->phys;
+	struct nvmap_handle *h = pb->handles[cur/8];
+
+	/* Create a fake nvmap_handle_ref - nvmap_mmap requires it
+	 * but accesses only the first field - nvmap_handle */
+	struct nvmap_handle_ref ref = {.handle = h};
+
+	u32 *map_addr;
+	int state, count, i;
+
+	map_addr = nvmap_mmap(&ref);
 	if (!map_addr)
 		return;
+
+	/* GATHER buffer starts always with commands */
+	state = NVHOST_DBG_STATE_CMD;
 	for (i = 0; i < words; i++)
-		show_channel_word(o, &state, &count, phys_addr + i * 4,
-				readl(map_addr + map_offset + i * 4));
-	iounmap(map_addr);
+		show_channel_word(o, &state, &count,
+				phys_addr + i * 4,
+				*(map_addr + map_offset/4 + i), cdma);
+
+	nvmap_munmap(&ref, map_addr);
 }
 
 static void show_channel_pair(struct output *o, u32 addr,
-				u32 w0, u32 w1)
+		u32 w0, u32 w1, struct nvhost_cdma *cdma)
 {
 	int state = NVHOST_DBG_STATE_CMD;
 	int count;
 
-	show_channel_word(o, &state, &count, addr, w0);
-	show_channel_word(o, &state, &count, addr, w1);
+	show_channel_word(o, &state, &count, addr, w0, cdma);
+	show_channel_word(o, &state, &count, addr+4, w1, cdma);
 }
 
 /**
@@ -182,10 +200,20 @@ static void cdma_peek(struct nvhost_cdma *cdma,
 	out[1] = p[offset + 1];
 }
 
+u32 previous_oppair(struct nvhost_cdma *cdma, u32 cur)
+{
+	u32 pb = cdma->push_buffer.phys;
+	u32 prev = cur-8;
+	if (prev < pb)
+		prev += PUSH_BUFFER_SIZE;
+	return prev;
+}
+
 static void t20_debug_show_channel_cdma(struct nvhost_master *m,
 					struct output *o, int chid)
 {
 	struct nvhost_channel *channel = m->channels + chid;
+	struct nvhost_cdma *cdma = &channel->cdma;
 	u32 dmaput, dmaget, dmactrl;
 	u32 cbstat, cbread;
 	u32 val, base, baseval;
@@ -198,7 +226,8 @@ static void t20_debug_show_channel_cdma(struct nvhost_master *m,
 	cbstat = readl(m->aperture + HOST1X_SYNC_CBSTAT(chid));
 
 	nvhost_debug_output(o, "%d-%s (%d): ", chid,
-			    channel->mod.name, atomic_read(&channel->mod.refcount));
+			    channel->mod.name,
+			    atomic_read(&channel->mod.refcount));
 
 	if ((dmactrl & 1) || !channel->cdma.push_buffer.mapped) {
 		nvhost_debug_output(o, "inactive\n\n");
@@ -223,13 +252,14 @@ static void t20_debug_show_channel_cdma(struct nvhost_master *m,
 		break;
 
 	default:
-		nvhost_debug_output(o, "active class %02x, offset %04x, val %08x\n",
+		nvhost_debug_output(o,
+				"active class %02x, offset %04x, val %08x\n",
 			cbstat >> 16, cbstat & 0xffff, cbread);
 		break;
 	}
 
-	cdma_peek(&channel->cdma, dmaget, -1, pbw);
-	show_channel_pair(o, chid, pbw[0], pbw[1]);
+	cdma_peek(cdma, dmaget, -1, pbw);
+	show_channel_pair(o, previous_oppair(cdma, dmaget), pbw[0], pbw[1], &channel->cdma);
 	nvhost_debug_output(o, "\n");
 }
 
@@ -264,7 +294,7 @@ void t20_debug_show_channel_fifo(struct nvhost_master *m,
 			m->aperture + HOST1X_SYNC_CFPEEK_CTRL);
 		val = readl(m->aperture + HOST1X_SYNC_CFPEEK_READ);
 
-		show_channel_word(o, &state, &count, 0, val);
+		show_channel_word(o, &state, &count, 0, val, NULL);
 
 		if (rd_ptr == end)
 			rd_ptr = start;