[Intel-gfx] [PATCH v2 03/38] drm/i915: Add some selftests for sg_table manipulation

[Chris Wilson]

On Wed, Feb 01, 2017 at 11:17:39AM +0000, Tvrtko Ursulin wrote:
> >+static noinline int expect_pfn_sg(struct pfn_table *pt,
> 
> Why noinline?

So they show up in perf individually.

> >+
> >+			for (npages = npages_funcs; *npages; npages++) {
> >+				prandom_seed_state(&prng,
> >+						   i915_selftest.random_seed);
> >+				if (!alloc_table(&pt, sz, sz, *npages, &prng))
> >+					return 0; /* out of memory, give up */
> 
> You don't have skip status? Sounds not ideal to silently abort.

It runs until we use all physical memory, if left to its own devices. It's
not a skip if we have already completed some tests. ENOMEM of the test
setup itself is not what I'm testing for here, the test is for the
iterators.
 
> >+
> >+				prandom_seed_state(&prng,
> >+						   i915_selftest.random_seed);
> >+				err = expect_pfn_sgtable(&pt, *npages, &prng,
> >+							 "sg_alloc_table",
> >+							 end_time);
> 
> Random numbers you use are guaranteed to be the same sequence after
> you re-set the seed? Probably yes since otherwise this wouldn't have
> ever worked.. I just remember some discussion on what source we use
> and it looked like we might be using proper random numbers on some
> CPUs, or even urandom which I didn't think has that property.

It's a completely deterministic prng. get_random_int() is the urandom
equivalent.

> >+static int igt_sg_trim(void *ignored)
> >+{
> >+	IGT_TIMEOUT(end_time);
> >+	const unsigned long max = PAGE_SIZE; /* not prime! */
> >+	struct pfn_table pt;
> >+	unsigned long prime;
> >+
> >+	for_each_prime_number(prime, max) {
> >+		const npages_fn_t *npages;
> >+		int err;
> >+
> >+		for (npages = npages_funcs; *npages; npages++) {
> >+			struct rnd_state prng;
> >+
> >+			prandom_seed_state(&prng, i915_selftest.random_seed);
> >+			if (!alloc_table(&pt, prime, max, *npages, &prng))
> >+				return 0; /* out of memory, give up */
> >+
> >+			err = 0;
> >+			if (i915_sg_trim(&pt.st)) {
> >+				if (pt.st.orig_nents != prime ||
> >+				    pt.st.nents != prime) {
> >+					pr_err("i915_sg_trim failed (nents %u, orig_nents %u), expected %lu\n",
> >+					       pt.st.nents, pt.st.orig_nents, prime);
> >+					err = -EINVAL;
> >+				} else {
> >+					prandom_seed_state(&prng,
> >+							   i915_selftest.random_seed);
> >+					err = expect_pfn_sgtable(&pt,
> >+								 *npages, &prng,
> >+								 "i915_sg_trim",
> >+								 end_time);
> >+				}
> >+			}
> 
> Similar to alloc_table failures above - no log or action when
> i915_sg_trim fails due out of memory?

No, simply because that's an expected and acceptable result. The
question should be whether we always want to check after sg_trim.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre