[Bug report] memset is stranger when i use SD_CARD_BENCH program with DMA.

[Norman Feske]

Hello Eun Seok, Kim,

btw, what is the polite way to greet you? Is Kim you first name?

Thank you for elaborating in more detail your modifications of the
'sd_card_bench' program. Given those information, I was able to
reproduce the problem. It turns out that it is actually unrelated to the
SD-card driver code or DMA for that matter. I reduced the test case to a
much smaller program that does not even touches any devices:

  1d7eeb00f6a89fb25159280133cf0f439a3a9390

This program produces the same artifacts as you observed. Here are the
things that I noted:

* The artifacts always appear at page boundaries, which suggests that
  the weird behavior may somehow be related to page-fault handling.
  This hypothesis is further supported by the little experiment of
  performing the memset + hexdump twice. The first time, the artifacts
  are visible but the second time, everything looks ok.

  I also did another experiment. To prevent page faults to occur during
  the 'memset' operation, I added a bit of code that touches each
  page of the buffer and thereby pages-in the corresponding page.
  With this eager page-fault handling in place, the artifacts
  disappear. You may try this out by removing 'if (0)'.

* The artifacts occur only when setting the 'cached' attribute of the
  'Attached_ram_dataspace' to false (as it should be for DMA buffers).
  When allocating a RAM dataspace with the cached attribute set, the
  artifacts disappear.

* The problem does occur on the Pandaboard but not on Qemu (pbxa9),
  which is typical for cache-related problems.

I am quite puzzled about the possible reason for the observed artifacts.
It seems to be a weird interplay between cache attributes and page-fault
handling on the ARMv7 platform.

I opened an new topic at our issue tracker:

  https://github.com/genodelabs/genode/issues/452

Would to mind taking the further discussion of this issue to there?

Best regards
Norman

-- 
Dr.-Ing. Norman Feske
Genode Labs

http://www.genode-labs.com · http://genode.org

Genode Labs GmbH · Amtsgericht Dresden · HRB 28424 · Sitz Dresden
Geschäftsführer: Dr.-Ing. Norman Feske, Christian Helmuth