diff options
| -rw-r--r-- | kernel/gsp_sema_service.c | 41 |
1 files changed, 39 insertions, 2 deletions
diff --git a/kernel/gsp_sema_service.c b/kernel/gsp_sema_service.c index 93aad52..7b63c0f 100644 --- a/kernel/gsp_sema_service.c +++ b/kernel/gsp_sema_service.c @@ -12014,8 +12014,45 @@ static void svc_opdef_drop(struct nv_vgpu_device *vgpu, struct svc_channel *ch, { struct svc_opdef_ent *e = svc_opdef_find(vgpu, ch->h_object, entry_idx); - if (e) - svc_opdef_clear(e); + if (!e) + return; + /* + * STALL FIX (r186/r187 entry=82, va=0x20400fef0 pay=84): when the + * forwarding machinery takes ownership of an entry, the withheld set + * may NOT ride the forwarded candidate's rel[] (svc_opdef_release() + * returned before the release was appended), so a blind clear strands + * the guest's tracker word with NO LATE-FLUSH/LOSS/EXPIRED ever + * logged (crel freezes right after the launches; evtest3 + * cudaStreamSynchronize(st[2]) hangs). Counter releases are + * monotonic completion signals: apply them forward-only here - a + * record at/ahead of the word is a no-op, so the fwd edge cannot + * double-apply and an additive reduction is skipped. + */ + { + u8 i; + u32 delivered = 0; + + for (i = 0; i < e->n; i++) { + if (!svc_counter_va(e->rel[i].va)) + continue; + if (e->rel[i].kind == SVC_FWD_REL_906F && + e->rel[i].op == HOST_SEM_OP_REDUCTION) + continue; /* additive: not idempotent */ + svc_fwd_rel_apply_all(vgpu, ch, &e->rel[i], 1); + svc.opdef_delivered_late++; + delivered++; + pr_info_ratelimited("sema-opdef: DROP-APPLY h=0x%x entry=%u va=0x%llx pay=%u (counter release delivered at fwd-ownership handoff)\n", + ch->h_object, entry_idx, + e->rel[i].va, e->rel[i].payload); + } + if (delivered < e->n) { + svc.opdef_drop_lost += (u32)(e->n - delivered); + pr_info_ratelimited("sema-opdef: LOSS h=0x%x entry=%u %u of %u withheld release(s) dropped at fwd-ownership handoff\n", + ch->h_object, entry_idx, + (u32)(e->n - delivered), e->n); + } + } + svc_opdef_clear(e); } /* |