msm: kgsl: Execute user profiling commands in an IB

Execute user profiling in an indirect buffer. This ensures that addresses
and values specified directly from the user don't end up in the
ringbuffer.

Change-Id: Ic0dedbadedcaab29ce5738a39c1ff6269261bae4
Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org>
Signed-off-by: Harshitha Sai Neelati <hsaine@codeaurora.org>

drivers/gpu/msm/adreno_ringbuffer.c

@@ -1,4 +1,4 @@
-/* Copyright (c) 2002,2007-2017, The Linux Foundation. All rights reserved.
+/* Copyright (c) 2002,2007-2017,2019, The Linux Foundation. All rights reserved.
  *
  * This program is free software; you can redistribute it and/or modify
  * it under the terms of the GNU General Public License version 2 and
@@ -259,6 +259,11 @@ static int _adreno_ringbuffer_probe(struct adreno_device *adreno_dev,
 		PAGE_SIZE, 0, KGSL_MEMDESC_PRIVILEGED, "pagetable_desc");
 	if (ret)
 		return ret;
+
+	/* allocate a chunk of memory to create user profiling IB1s */
+	kgsl_allocate_global(KGSL_DEVICE(adreno_dev), &rb->profile_desc,
+		PAGE_SIZE, KGSL_MEMFLAGS_GPUREADONLY, 0, "profile_desc");
+
 	return kgsl_allocate_global(KGSL_DEVICE(adreno_dev), &rb->buffer_desc,
 			KGSL_RB_SIZE, KGSL_MEMFLAGS_GPUREADONLY,
 			0, "ringbuffer");
@@ -303,7 +308,7 @@ static void _adreno_ringbuffer_close(struct adreno_device *adreno_dev,
 
 	kgsl_free_global(device, &rb->pagetable_desc);
 	kgsl_free_global(device, &rb->preemption_desc);
-
+	kgsl_free_global(device, &rb->profile_desc);
 	kgsl_free_global(device, &rb->buffer_desc);
 	kgsl_del_event_group(&rb->events);
 	memset(rb, 0, sizeof(struct adreno_ringbuffer));
@@ -737,6 +742,37 @@ static inline int _get_alwayson_counter(struct adreno_device *adreno_dev,
 	return (unsigned int)(p - cmds);
 }
 
+/* This is the maximum possible size for 64 bit targets */
+#define PROFILE_IB_DWORDS 4
+#define PROFILE_IB_SLOTS (PAGE_SIZE / (PROFILE_IB_DWORDS << 2))
+
+static int set_user_profiling(struct adreno_device *adreno_dev,
+		struct adreno_ringbuffer *rb, u32 *cmds, u64 gpuaddr)
+{
+	int dwords, index = 0;
+	u64 ib_gpuaddr;
+	u32 *ib;
+
+	if (!rb->profile_desc.hostptr)
+		return 0;
+
+	ib = ((u32 *) rb->profile_desc.hostptr) +
+		(rb->profile_index * PROFILE_IB_DWORDS);
+	ib_gpuaddr = rb->profile_desc.gpuaddr +
+		(rb->profile_index * (PROFILE_IB_DWORDS << 2));
+
+	dwords = _get_alwayson_counter(adreno_dev, ib, gpuaddr);
+
+	/* Make an indirect buffer for the request */
+	cmds[index++] = cp_mem_packet(adreno_dev, CP_INDIRECT_BUFFER_PFE, 2, 1);
+	index += cp_gpuaddr(adreno_dev, &cmds[index], ib_gpuaddr);
+	cmds[index++] = dwords;
+
+	rb->profile_index = (rb->profile_index + 1) % PROFILE_IB_SLOTS;
+
+	return index;
+}
+
 /* adreno_rindbuffer_submitcmd - submit userspace IBs to the GPU */
 int adreno_ringbuffer_submitcmd(struct adreno_device *adreno_dev,
 		struct kgsl_drawobj_cmd *cmdobj,
@@ -836,14 +872,12 @@ int adreno_ringbuffer_submitcmd(struct adreno_device *adreno_dev,
 	if (drawobj->flags & KGSL_DRAWOBJ_PROFILING &&
 		!adreno_is_a3xx(adreno_dev) && profile_buffer) {
 		user_profiling = true;
-		dwords += 6;
 
 		/*
-		 * REG_TO_MEM packet on A5xx and above needs another ordinal.
-		 * Add 2 more dwords since we do profiling before and after.
+		 * User side profiling uses two IB1s, one before with 4 dwords
+		 * per INDIRECT_BUFFER_PFE call
 		 */
-		if (!ADRENO_LEGACY_PM4(adreno_dev))
-			dwords += 2;
+		dwords += 8;
 
 		/*
 		 * we want to use an adreno_submit_time struct to get the
@@ -886,11 +920,11 @@ int adreno_ringbuffer_submitcmd(struct adreno_device *adreno_dev,
 	}
 
 	/*
-	 * Add cmds to read the GPU ticks at the start of command obj and
+	 * Add IB1 to read the GPU ticks at the start of command obj and
 	 * write it into the appropriate command obj profiling buffer offset
 	 */
 	if (user_profiling) {
-		cmds += _get_alwayson_counter(adreno_dev, cmds,
+		cmds += set_user_profiling(adreno_dev, rb, cmds,
 			cmdobj->profiling_buffer_gpuaddr +
 			offsetof(struct kgsl_drawobj_profiling_buffer,
 			gpu_ticks_submitted));
@@ -929,11 +963,11 @@ int adreno_ringbuffer_submitcmd(struct adreno_device *adreno_dev,
 	}
 
 	/*
-	 * Add cmds to read the GPU ticks at the end of command obj and
+	 * Add IB1 to read the GPU ticks at the end of command obj and
 	 * write it into the appropriate command obj profiling buffer offset
 	 */
 	if (user_profiling) {
-		cmds += _get_alwayson_counter(adreno_dev, cmds,
+		cmds += set_user_profiling(adreno_dev, rb, cmds,
 			cmdobj->profiling_buffer_gpuaddr +
 			offsetof(struct kgsl_drawobj_profiling_buffer,
 			gpu_ticks_retired));

drivers/gpu/msm/adreno_ringbuffer.h

@@ -1,4 +1,4 @@
-/* Copyright (c) 2002,2007-2016, The Linux Foundation. All rights reserved.
+/* Copyright (c) 2002,2007-2016,2019, The Linux Foundation. All rights reserved.
  *
  * This program is free software; you can redistribute it and/or modify
  * it under the terms of the GNU General Public License version 2 and
@@ -127,6 +127,18 @@ struct adreno_ringbuffer {
 	unsigned long sched_timer;
 	enum adreno_dispatcher_starve_timer_states starve_timer_state;
 	spinlock_t preempt_lock;
+	/**
+	 * @profile_desc: global memory to construct IB1s to do user side
+	 * profiling
+	 */
+	struct kgsl_memdesc profile_desc;
+	/**
+	 * @profile_index: Pointer to the next "slot" in profile_desc for a user
+	 * profiling IB1.  This allows for PAGE_SIZE / 16 = 256 simultaneous
+	 * commands per ringbuffer with user profiling enabled
+	 * enough.
+	 */
+	u32 profile_index;
 };
 
 /* Returns the current ringbuffer */