[0/3] drm/panfrost: Bug fixes for lock_region

Message

Alyssa Rosenzweig Aug. 20, 2021, 9:31 p.m. UTC

Chris Morgan reported UBSAN errors in panfrost and tracked them down to
the size computation in lock_region. This calculation is overcomplicated
(seemingly cargo culted from kbase) and can be simplified with kernel
helpers and some mathematical identities. The first patch in the series
rewrites the calculation in a form avoiding undefined behaviour; Chris
confirms it placates UBSAN.

While researching this function, I noticed a pair of other potential
bugs: Bifrost can lock more than 4GiB at a time, but must lock at least
32KiB at a time. The latter patches in the series handle these cases.

The size computation was unit-tested in userspace. Relevant code below,
just missing some copypaste definitions for fls64/clamp/etc:

	#define MIN_LOCK (1ULL << 12)
	#define MAX_LOCK (1ULL << 48)

	struct {
		uint64_t size;
		uint8_t encoded;
	} tests[] = {
		/* Clamping */
		{ 0, 11 },
		{ 1, 11 },
		{ 2, 11 },
		{ 4095, 11 },
		/* Power of two */
		{ 4096, 11 },
		/* Round up */
		{ 4097, 12 },
		{ 8192, 12 },
		{ 16384, 13 },
		{ 16385, 14 },
		/* Maximum */
		{ ~0ULL, 47 },
	};

	static uint8_t region_width(uint64_t size)
	{
		size = clamp(size, MIN_LOCK, MAX_LOCK);
		return fls64(size - 1) - 1;
	}

	int main(int argc, char **argv)
	{
		for (unsigned i = 0; i < ARRAY_SIZE(tests); ++i) {
			uint64_t test = tests[i].size;
			uint8_t expected = tests[i].encoded;
			uint8_t actual = region_width(test);

			assert(expected == actual);
		}
	}

Alyssa Rosenzweig (3):
  drm/panfrost: Simplify lock_region calculation
  drm/panfrost: Use u64 for size in lock_region
  drm/panfrost: Clamp lock region to Bifrost minimum

 drivers/gpu/drm/panfrost/panfrost_mmu.c  | 31 +++++++++---------------
 drivers/gpu/drm/panfrost/panfrost_regs.h |  2 ++
 2 files changed, 13 insertions(+), 20 deletions(-)