coresight: tmc-etr: Speed up for bounce buffer in flat mode

The AUX bounce buffer is allocated with API dma_alloc_coherent(), in the
low level's architecture code, e.g. for Arm64, it maps the memory with
the attribution "Normal non-cacheable"; this can be concluded from the
definition for pgprot_dmacoherent() in arch/arm64/include/asm/pgtable.h.

Later when access the AUX bounce buffer, since the memory mapping is
non-cacheable, it's low efficiency due to every load instruction must
reach out DRAM.

This patch changes to allocate pages with dma_alloc_noncoherent(), the
driver can access the memory via cacheable mapping; therefore, load
instructions can fetch data from cache lines rather than always read
data from DRAM, the driver can boost memory performance.  After using
the cacheable mapping, the driver uses dma_sync_single_for_cpu() to
invalidate cacheline prior to read bounce buffer so can avoid read stale
trace data.

By measurement the duration for function tmc_update_etr_buffer() with
ftrace function_graph tracer, it shows the performance significant
improvement for copying 4MiB data from bounce buffer:

  # echo tmc_etr_get_data_flat_buf > set_graph_notrace // avoid noise
  # echo tmc_update_etr_buffer > set_graph_function
  # echo function_graph > current_tracer

  before:

  # CPU  DURATION                  FUNCTION CALLS
  # |     |   |                     |   |   |   |
  2)               |    tmc_update_etr_buffer() {
  ...
  2) # 8148.320 us |    }

  after:

  # CPU  DURATION                  FUNCTION CALLS
  # |     |   |                     |   |   |   |
  2)               |  tmc_update_etr_buffer() {
  ...
  2) # 2525.420 us |  }

Signed-off-by: Leo Yan <leo.yan@linaro.org>
Reviewed-by: Suzuki K Poulose <suzuki.poulose@arm.com>
Link: https://lore.kernel.org/r/20210905032144.966766-1-leo.yan@linaro.org
Signed-off-by: Mathieu Poirier <mathieu.poirier@linaro.org>

drivers/hwtracing/coresight/coresight-tmc-etr.c

@@ -607,8 +607,9 @@ static int tmc_etr_alloc_flat_buf(struct tmc_drvdata *drvdata,
 	if (!flat_buf)
 		return -ENOMEM;
 
-	flat_buf->vaddr = dma_alloc_coherent(real_dev, etr_buf->size,
+	flat_buf->vaddr = dma_alloc_noncoherent(real_dev, etr_buf->size,
-					     &flat_buf->daddr, GFP_KERNEL);
+						&flat_buf->daddr,
+						DMA_FROM_DEVICE, GFP_KERNEL);
 	if (!flat_buf->vaddr) {
 		kfree(flat_buf);
 		return -ENOMEM;
@@ -629,14 +630,18 @@ static void tmc_etr_free_flat_buf(struct etr_buf *etr_buf)
 	if (flat_buf && flat_buf->daddr) {
 		struct device *real_dev = flat_buf->dev->parent;
 
-		dma_free_coherent(real_dev, flat_buf->size,
+		dma_free_noncoherent(real_dev, etr_buf->size,
-				  flat_buf->vaddr, flat_buf->daddr);
+				     flat_buf->vaddr, flat_buf->daddr,
+				     DMA_FROM_DEVICE);
 	}
 	kfree(flat_buf);
 }
 
 static void tmc_etr_sync_flat_buf(struct etr_buf *etr_buf, u64 rrp, u64 rwp)
 {
+	struct etr_flat_buf *flat_buf = etr_buf->private;
+	struct device *real_dev = flat_buf->dev->parent;
+
 	/*
 	 * Adjust the buffer to point to the beginning of the trace data
 	 * and update the available trace data.
@@ -646,6 +651,19 @@ static void tmc_etr_sync_flat_buf(struct etr_buf *etr_buf, u64 rrp, u64 rwp)
 		etr_buf->len = etr_buf->size;
 	else
 		etr_buf->len = rwp - rrp;
+
+	/*
+	 * The driver always starts tracing at the beginning of the buffer,
+	 * the only reason why we would get a wrap around is when the buffer
+	 * is full.  Sync the entire buffer in one go for this case.
+	 */
+	if (etr_buf->offset + etr_buf->len > etr_buf->size)
+		dma_sync_single_for_cpu(real_dev, flat_buf->daddr,
+					etr_buf->size, DMA_FROM_DEVICE);
+	else
+		dma_sync_single_for_cpu(real_dev,
+					flat_buf->daddr + etr_buf->offset,
+					etr_buf->len, DMA_FROM_DEVICE);
 }
 
 static ssize_t tmc_etr_get_data_flat_buf(struct etr_buf *etr_buf,