diff options
Diffstat (limited to 'kernel/gsp_phys_booter.c')
| -rw-r--r-- | kernel/gsp_phys_booter.c | 134 |
1 files changed, 132 insertions, 2 deletions
diff --git a/kernel/gsp_phys_booter.c b/kernel/gsp_phys_booter.c index 5ca854d..effff74 100644 --- a/kernel/gsp_phys_booter.c +++ b/kernel/gsp_phys_booter.c @@ -17645,6 +17645,34 @@ MODULE_PARM_DESC(sema_gl_mme_nop_drop, "Slots whose call runs are dropped entirely (proven-safe retire, SA[slot] stubbed) instead of retargeted to the CALL_MME_DATA port (u64 bitmask; default 0)"); /* + * SESSION 153-CONTINUED-82: hold real CALL runs until the prelude latch. + * The r1220 (all-real, nop_drop=1<<57, sa_stub_all=N) wedge: the window's + * real CALL runs executed in the SAME launch that emitted only the first + * PART of a fresh capture generation (`MME-PRELUDE part gen=3 off=1776 + * words=1776/3083 ... (continuation next launch)`), so the called macros + * fetched IR code that was not resident yet - `MME-FAST launch ... + * pc_valid=1 pc=513`, priv-error poison, `Shader Program Header 18 Error`, + * `Class Error ... Offset 0000238c, Data 00020020` (the r1116 lane-poison + * class) and the boot wedged. The prelude latch (gsp_phys_gr_mme_prelude_* + * gen/h) is set only when the FINAL part has gone out; with this knob the + * launch builder holds (elides, the proven-safe drop shape) every remaining + * CALL_MME_MACRO/DATA run of the emitted window while EITHER the latch for + * this generation+channel is not set after this launch, OR this launch + * uploaded generation code for the first time (latch invalid before the + * emission). The r1222 all-real boot proved the second clause necessary: + * the drain-completing launch (part-2 IR uploads + window real calls in the + * SAME block, latch set by its own emission) still hit ILLEGAL_OPCODE - + * real calls resume on the next launch, when only the idempotent keep-forced + * re-emission of the already-resident prefix is in flight. Default 1 (the + * fix); A/B with 0 to reproduce the r1209/r1220/r1222 wedge. + */ +static bool gsp_phys_rmcall_gr_mme_hold_calls = true; +module_param_named(sema_gl_mme_hold_calls, gsp_phys_rmcall_gr_mme_hold_calls, + bool, 0644); +MODULE_PARM_DESC(sema_gl_mme_hold_calls, + "Hold (elide) window CALL_MME_MACRO/DATA runs until the MME prelude latches for the current generation+channel (default 1; 0 = legacy, real calls run against a partial prelude)"); + +/* * The same stubbed-slot predicate the launch builder applies (see the * sa_stub emit decision): a slot is stubbed when stubbing is on and either * the all-slots knob is set, the capture holds no SA entry, or the IR @@ -17939,6 +17967,9 @@ static u32 gsp_phys_rmcall_build_gr_launch(u32 *pb, */ bool emit_prelude = false; bool pre_done = false; + bool mme_hold_calls = false; + bool latch_before = false; + u32 pre_words_out = 0; const u32 *pl = NULL; u32 pl_n = 0, pl_gen = 0, pl_room = 0; u32 block_cap = READ_ONCE(gsp_phys_rmcall_gr_block_cap); @@ -18159,6 +18190,23 @@ static u32 gsp_phys_rmcall_build_gr_launch(u32 *pb, */ u32 fixed = 2u + sa_block + win3d_n + 8u; + /* + * SESSION 153-CONTINUED-82: latch state BEFORE + * this launch's emission (see the hold decision + * below). A launch that first uploads a + * generation's MME code must not run real CALL + * runs in the same block (the r1222 evidence: + * the drain-completing launch bundled part-2's + * IR uploads with the window's real calls and + * hit ILLEGAL_OPCODE even though the latch was + * set by its own emission step). + */ + latch_before = + (pl_gen == READ_ONCE( + gsp_phys_gr_mme_prelude_gen) && + req->h_object == READ_ONCE( + gsp_phys_gr_mme_prelude_h)); + pl_room = (fixed + 24u <= block_cap) ? block_cap - fixed : 0u; if (pl_room > pl_n) @@ -18251,6 +18299,7 @@ static u32 gsp_phys_rmcall_build_gr_launch(u32 *pb, } WRITE_ONCE(gsp_phys_gr_mme_pre_emit_off, doff); pre_done = (doff >= pl_n); + pre_words_out = pwords; /* seqcheck: a mid-copy append must not be recorded as * emitted (the copy may be a torn prefix). */ gsp_sema_mme_prelude_get(NULL, &g1); @@ -18308,6 +18357,36 @@ static u32 gsp_phys_rmcall_build_gr_launch(u32 *pb, } } /* + * SESSION 153-CONTINUED-82: hold real CALL runs until the + * prelude latch for this generation+channel is set (see the + * sema_gl_mme_hold_calls comment). Evaluated AFTER the + * emission step above, so a launch that completed the drain + * in this same block has just written the latch. Real calls + * stay held when EITHER + * - the latch for this generation+channel is not valid after + * this launch (partial drain, raced append, or a fresh + * generation whose code is not fully resident), or + * - this launch uploaded generation code for the FIRST time + * (emitted words with the latch invalid before the step) - + * the r1222 evidence: the drain-completing launch bundled + * part-2's IR uploads with the window's real calls and hit + * ILLEGAL_OPCODE even though its own emission set the + * latch. A later launch (latch already valid, only the + * idempotent keep-forced re-emission of the prefix) runs + * the real calls against the now-resident program. + */ + if (pl && pl_n && pl_gen && + READ_ONCE(gsp_phys_rmcall_gr_mme_prelude) && + READ_ONCE(gsp_phys_rmcall_gr_mme_hold_calls)) { + u32 lg = READ_ONCE(gsp_phys_gr_mme_prelude_gen); + u32 lh = READ_ONCE(gsp_phys_gr_mme_prelude_h); + bool latch_after = (pl_gen == lg && + req->h_object == lh); + + mme_hold_calls = !latch_after || + (pre_words_out && !latch_before); + } + /* * SESSION 153-CONTINUED-63: emit the SA-RAM completion computed * above - a one-instruction end-noop stub in the IR plus one * ONE_INC 0x11c run loading [slot, stub] for every called slot @@ -18481,8 +18560,59 @@ static u32 gsp_phys_rmcall_build_gr_launch(u32 *pb, } } } - memcpy(p, win3d, (size_t)win3d_n * 4); - p += win3d_n; + if (mme_hold_calls) { + /* + * SESSION 153-CONTINUED-82: the prelude latch for + * this generation+channel is not set - elide every + * remaining CALL_MME_MACRO/DATA run of the window + * (the proven-safe drop shape; see the + * sema_gl_mme_hold_calls comment). Runs touching + * the MME macro method space 0x3800..0x39fc are the + * ones that execute a macro; the drop decision is + * the same walk gsp_phys_gr_mme_drop_calls uses. + */ + u32 r = 0, w = 0, held = 0; + + while (r < win3d_n) { + u32 h = win3d[r]; + u32 op = (h >> 29) & 7u; + u32 cnt = (h >> 16) & 0xfffu; + u32 span = 1u; + bool iscall = false; + u32 k; + + if ((op == 1u || op == 3u || op == 5u) && + cnt && r + 1u + cnt <= win3d_n) + span = 1u + cnt; + for (k = 0; k < span; k++) { + u32 m = ((h & 0xfffu) << 2) + + 4u * k; + + if (k && op != 1u && op != 3u) + break; + if (m >= 0x3800u && m <= 0x39fcu) { + iscall = true; + break; + } + } + if (iscall) { + held += span; + r += span; + continue; + } + memcpy(&p[w], &win3d[r], (size_t)span * 4u); + w += span; + r += span; + } + if (held) + pr_info("vgsp-gr: MME-CALL-HOLD %u word(s) of real CALL runs held (prelude gen=%u not latched for h=0x%x; window=%u->%u) push=0x%llx\n", + held, pl_gen, req->h_object, + win3d_n, w, req->push_va); + p += w; + } else { + memcpy(p, win3d, (size_t)win3d_n * 4); + p += win3d_n; + } p = gsp_phys_rmcall_push_gr(p, GSP_PHYS_RMCALL_GR_SET_REPORT_SEMAPHORE_A, (u32)(sem_va >> 32)); |