tree: git://anongit.freedesktop.org/drm-intel drm-intel-next-queued
head: c616d2387aeeb987f03eee848f04ffdc248c7aae
commit: c616d2387aeeb987f03eee848f04ffdc248c7aae [5/5] drm/i915/gt: Expand bad CS completion event debug
config: i386-randconfig-h003-20200213 (attached as .config)
compiler: gcc-7 (Debian 7.5.0-4) 7.5.0
reproduce:
git checkout c616d2387aeeb987f03eee848f04ffdc248c7aae
# save the attached .config to linux build tree
make ARCH=i386
If you fix the issue, kindly add following tag
Reported-by: kbuild test robot <lkp@xxxxxxxxx>
All errors (new ones prefixed by >>):
drivers/gpu/drm/i915/gt/intel_lrc.c: In function 'process_csb':
>> drivers/gpu/drm/i915/gt/intel_lrc.c:2335:16: error: unused variable 'regs' [-Werror=unused-variable]
const u32 *regs = rq->context->lrc_reg_state;
^~~~
cc1: all warnings being treated as errors
vim +/regs +2335 drivers/gpu/drm/i915/gt/intel_lrc.c
2222
2223 static void process_csb(struct intel_engine_cs *engine)
2224 {
2225 struct intel_engine_execlists * const execlists = &engine->execlists;
2226 const u32 * const buf = execlists->csb_status;
2227 const u8 num_entries = execlists->csb_size;
2228 u8 head, tail;
2229
2230 /*
2231 * As we modify our execlists state tracking we require exclusive
2232 * access. Either we are inside the tasklet, or the tasklet is disabled
2233 * and we assume that is only inside the reset paths and so serialised.
2234 */
2235 GEM_BUG_ON(!tasklet_is_locked(&execlists->tasklet) &&
2236 !reset_in_progress(execlists));
2237 GEM_BUG_ON(!intel_engine_in_execlists_submission_mode(engine));
2238
2239 /*
2240 * Note that csb_write, csb_status may be either in HWSP or mmio.
2241 * When reading from the csb_write mmio register, we have to be
2242 * careful to only use the GEN8_CSB_WRITE_PTR portion, which is
2243 * the low 4bits. As it happens we know the next 4bits are always
2244 * zero and so we can simply masked off the low u8 of the register
2245 * and treat it identically to reading from the HWSP (without having
2246 * to use explicit shifting and masking, and probably bifurcating
2247 * the code to handle the legacy mmio read).
2248 */
2249 head = execlists->csb_head;
2250 tail = READ_ONCE(*execlists->csb_write);
2251 if (unlikely(head == tail))
2252 return;
2253
2254 /*
2255 * Hopefully paired with a wmb() in HW!
2256 *
2257 * We must complete the read of the write pointer before any reads
2258 * from the CSB, so that we do not see stale values. Without an rmb
2259 * (lfence) the HW may speculatively perform the CSB[] reads *before*
2260 * we perform the READ_ONCE(*csb_write).
2261 */
2262 rmb();
2263
2264 ENGINE_TRACE(engine, "cs-irq head=%d, tail=%d\n", head, tail);
2265 do {
2266 bool promote;
2267
2268 if (++head == num_entries)
2269 head = 0;
2270
2271 /*
2272 * We are flying near dragons again.
2273 *
2274 * We hold a reference to the request in execlist_port[]
2275 * but no more than that. We are operating in softirq
2276 * context and so cannot hold any mutex or sleep. That
2277 * prevents us stopping the requests we are processing
2278 * in port[] from being retired simultaneously (the
2279 * breadcrumb will be complete before we see the
2280 * context-switch). As we only hold the reference to the
2281 * request, any pointer chasing underneath the request
2282 * is subject to a potential use-after-free. Thus we
2283 * store all of the bookkeeping within port[] as
2284 * required, and avoid using unguarded pointers beneath
2285 * request itself. The same applies to the atomic
2286 * status notifier.
2287 */
2288
2289 ENGINE_TRACE(engine, "csb[%d]: status=0x%08x:0x%08x\n",
2290 head, buf[2 * head + 0], buf[2 * head + 1]);
2291
2292 if (INTEL_GEN(engine->i915) >= 12)
2293 promote = gen12_csb_parse(execlists, buf + 2 * head);
2294 else
2295 promote = gen8_csb_parse(execlists, buf + 2 * head);
2296 if (promote) {
2297 struct i915_request * const *old = execlists->active;
2298
2299 GEM_BUG_ON(!assert_pending_valid(execlists, "promote"));
2300
2301 ring_set_paused(engine, 0);
2302
2303 /* Point active to the new ELSP; prevent overwriting */
2304 WRITE_ONCE(execlists->active, execlists->pending);
2305
2306 /* cancel old inflight, prepare for switch */
2307 trace_ports(execlists, "preempted", old);
2308 while (*old)
2309 execlists_schedule_out(*old++);
2310
2311 /* switch pending to inflight */
2312 WRITE_ONCE(execlists->active,
2313 memcpy(execlists->inflight,
2314 execlists->pending,
2315 execlists_num_ports(execlists) *
2316 sizeof(*execlists->pending)));
2317
2318 WRITE_ONCE(execlists->pending[0], NULL);
2319 } else {
2320 GEM_BUG_ON(!*execlists->active);
2321
2322 /* port0 completed, advanced to port1 */
2323 trace_ports(execlists, "completed", execlists->active);
2324
2325 /*
2326 * We rely on the hardware being strongly
2327 * ordered, that the breadcrumb write is
2328 * coherent (visible from the CPU) before the
2329 * user interrupt and CSB is processed.
2330 */
2331 if (GEM_SHOW_DEBUG() &&
2332 !i915_request_completed(*execlists->active) &&
2333 !reset_in_progress(execlists)) {
2334 struct i915_request *rq = *execlists->active;
> 2335 const u32 *regs = rq->context->lrc_reg_state;
2336
2337 ENGINE_TRACE(engine,
2338 "ring:{start:0x%08x, head:%04x, tail:%04x, ctl:%08x, mode:%08x}\n",
2339 ENGINE_READ(engine, RING_START),
2340 ENGINE_READ(engine, RING_HEAD) & HEAD_ADDR,
2341 ENGINE_READ(engine, RING_TAIL) & TAIL_ADDR,
2342 ENGINE_READ(engine, RING_CTL),
2343 ENGINE_READ(engine, RING_MI_MODE));
2344 ENGINE_TRACE(engine,
2345 "rq:{start:%08x, head:%04x, tail:%04x, seqno:%llx:%d, hwsp:%d}, ",
2346 i915_ggtt_offset(rq->ring->vma),
2347 rq->head, rq->tail,
2348 rq->fence.context,
2349 lower_32_bits(rq->fence.seqno),
2350 hwsp_seqno(rq));
2351 ENGINE_TRACE(engine,
2352 "ctx:{start:%08x, head:%04x, tail:%04x}, ",
2353 regs[CTX_RING_START],
2354 regs[CTX_RING_HEAD],
2355 regs[CTX_RING_TAIL]);
2356
2357 GEM_BUG_ON("context completed before request");
2358 }
2359
2360 execlists_schedule_out(*execlists->active++);
2361
2362 GEM_BUG_ON(execlists->active - execlists->inflight >
2363 execlists_num_ports(execlists));
2364 }
2365 } while (head != tail);
2366
2367 execlists->csb_head = head;
2368 set_timeslice(engine);
2369
2370 /*
2371 * Gen11 has proven to fail wrt global observation point between
2372 * entry and tail update, failing on the ordering and thus
2373 * we see an old entry in the context status buffer.
2374 *
2375 * Forcibly evict out entries for the next gpu csb update,
2376 * to increase the odds that we get a fresh entries with non
2377 * working hardware. The cost for doing so comes out mostly with
2378 * the wash as hardware, working or not, will need to do the
2379 * invalidation before.
2380 */
2381 invalidate_csb_entries(&buf[0], &buf[num_entries - 1]);
2382 }
2383
---
0-DAY CI Kernel Test Service, Intel Corporation
https://lists.01.org/hyperkitty/list/kbuild-all@xxxxxxxxxxxx
Attachment:
.config.gz
Description: application/gzip
_______________________________________________ dri-devel mailing list dri-devel@xxxxxxxxxxxxxxxxxxxxx https://lists.freedesktop.org/mailman/listinfo/dri-devel
