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.orgSigned-off-by: Mathieu Poirier <mathieu.poirier@linaro.org>