Re: [PATCH] coresight: tmc-etr: Fix updating buffer in not-snapshot mode.

[Mathieu Poirier]

Good day Yabin,

With this patch you are addressing a long time itch I had - please read on.

On Mon, 5 Aug 2019 at 17:37, Yabin Cui <yabinc@xxxxxxxxxx> wrote:
>
> TMC etr always copies all available data to perf aux buffer, which
> may exceed the available space in perf aux buffer. It isn't suitable
> for not-snapshot mode, because:
> 1) It may overwrite previously written data.
> 2) It may make the perf_event_mmap_page->aux_head report having more
> or less data than the reality.
>
> Signed-off-by: Yabin Cui <yabinc@xxxxxxxxxx>
> ---
>  drivers/hwtracing/coresight/coresight-tmc-etr.c | 12 ++++++++----
>  1 file changed, 8 insertions(+), 4 deletions(-)
>
> diff --git a/drivers/hwtracing/coresight/coresight-tmc-etr.c b/drivers/hwtracing/coresight/coresight-tmc-etr.c
> index 17006705287a..697e68d492af 100644
> --- a/drivers/hwtracing/coresight/coresight-tmc-etr.c
> +++ b/drivers/hwtracing/coresight/coresight-tmc-etr.c
> @@ -1410,9 +1410,10 @@ static void tmc_free_etr_buffer(void *config)
>   * tmc_etr_sync_perf_buffer: Copy the actual trace data from the hardware
>   * buffer to the perf ring buffer.
>   */
> -static void tmc_etr_sync_perf_buffer(struct etr_perf_buffer *etr_perf)
> +static void tmc_etr_sync_perf_buffer(struct etr_perf_buffer *etr_perf,
> +                                    unsigned long to_copy)
>  {
> -       long bytes, to_copy;
> +       long bytes;
>         long pg_idx, pg_offset, src_offset;
>         unsigned long head = etr_perf->head;
>         char **dst_pages, *src_buf;
> @@ -1423,7 +1424,6 @@ static void tmc_etr_sync_perf_buffer(struct etr_perf_buffer *etr_perf)
>         pg_offset = head & (PAGE_SIZE - 1);
>         dst_pages = (char **)etr_perf->pages;
>         src_offset = etr_buf->offset;
> -       to_copy = etr_buf->len;
>
>         while (to_copy > 0) {
>                 /*
> @@ -1501,7 +1501,11 @@ tmc_update_etr_buffer(struct coresight_device *csdev,
>         spin_unlock_irqrestore(&drvdata->spinlock, flags);
>
>         size = etr_buf->len;
> -       tmc_etr_sync_perf_buffer(etr_perf);
> +       if (!etr_perf->snapshot && size > handle->size) {
> +               size = handle->size;
> +               lost = true;
> +       }

Perfect - this is in line with what is done for ETB and ETF.

> +       tmc_etr_sync_perf_buffer(etr_perf, size);

Here tmc_etr_sync_perf_buffer() will copy data to the perf ring buffer
starting at @etr_perf->offset for @size, clipping the _end_ of the
trace data accumulated in the trace buffer.  This is contrary to what
is done for ETB and ETF where the equivalent of @etr_perf->offset is
moved forward (clipping the _beginning_ of the trace data) in order to
keep as much of the end as possible.

I would rather enact the same heuristic here.

Thanks,
Mathieu

>
>         /*
>          * In snapshot mode we simply increment the head by the number of byte
> --
> 2.22.0.770.g0f2c4a37fd-goog
>