Hi Reinette,
On 5/26/2026 11:25 PM, Reinette Chatre wrote:
Hi Chenyu,
On 5/25/26 4:59 AM, Chen, Yu C wrote:
Hi Ilpo,
On 5/25/2026 6:01 PM, Ilpo Järvinen wrote:
On Sat, 23 May 2026, Chen Yu wrote:
The shareable_bits indicates which cache portions might be shared with
other agents (e.g., GPU, other I/O devices). On some platforms, all L3
cache ways are marked as potentially shared.
The L3_CAT test should fall back to using the full CBM mask when no
exclusive portion exists, since the test logic (writing different
schemata and measuring cache misses) still works. The shareable mask
How do you define "still works"? Allowing more interference will impact the
number of cache misses that is used to determine effectiveness of cache
allocation, no?
is a hint about potential interference, not a hard exclusion.
Reported-by: Hongyu Ning <[email protected]>
Signed-off-by: Chen Yu <[email protected]>
---
tools/testing/selftests/resctrl/cat_test.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/tools/testing/selftests/resctrl/cat_test.c
b/tools/testing/selftests/resctrl/cat_test.c
index f00b622c1460..b6994ba4504b 100644
--- a/tools/testing/selftests/resctrl/cat_test.c
+++ b/tools/testing/selftests/resctrl/cat_test.c
@@ -248,7 +248,7 @@ static int cat_run_test(const struct resctrl_test *test,
const struct user_param
/* Get the largest contiguous exclusive portion of the cache */
ret = get_mask_no_shareable(test->resource, &long_mask);
if (ret)
- return ret;
+ long_mask = full_cache_mask;
Wouldn't it warrant to log something when the fallback occurs? It
certainly looks something that makes the results less reliable (so it
would be nice to know about it when looking at the test results).
Got it. Let me add the following message:
ksft_print_msg("No exclusive cache portion found, fall back to full cache
mask\n");
Allowing more interference will cause the test to fail while the system behaves
correctly.
This log message may help a knowledgeable user running the test manually but I
expect would
be confusing to folks not familiar with resctrl and definitely does not help
test automation
that just looks for the pass/fail.
The test can be fixed by skipping the test, instead of failing, if there are no
bits
available for the test to run as intended. Of course, this will prevent CAT
mechanics to
be tested on this platform so to test CAT on this platform there is a need a
new functional
test since a performance test is not appropriate.
Got it.
My thinking was that shareable_bits only describes portions that may
be shared with non-CPU agents, and on an otherwise idle test system
the CAT miss-rate trend can still be observed even if all ways are
marked shareable. But I agree that this is not a good assumption for
kselftest pass/fail.
So I agree that falling back to the full CBM is not the right fix for
this test. I will change the patch to skip L3_CAT when no exclusive
cache portion is available, with a message explaining that the performance
test cannot run reliably because all cache portions are shareable.
For platforms where all cache ways are marked shareable, do you think
providing a extra parameter to ask cat_test to "ignore" the share
restriction is acceptable?
thanks,
Chenyu
