diff options
| author | Arthur Hanson Rasmusson <arthur@vgpu.io> | 2026-11-16 09:50:31 -0500 |
|---|---|---|
| committer | Arthur Hanson Rasmusson <arthur@vgpu.io> | 2026-11-16 09:50:31 -0500 |
| commit | ed8a9c6b8e061fe772fc0050ab83c2dec2db8d1e (patch) | |
| tree | 458d2ea063d8a8a6996cd15891e3242cfe93ece5 /kernel | |
| parent | f2c73f4e52d24188a9edb0e8190c50f87f19181f (diff) | |
| download | open-iov-ed8a9c6b8e061fe772fc0050ab83c2dec2db8d1e.tar.gz open-iov-ed8a9c6b8e061fe772fc0050ab83c2dec2db8d1e.zip | |
vgsp: eng89 - live-derived GS map, draw-move anchor, raster enables, async capture
Kernel (deterministic fixes, each with an A/B knob):
- sema_gl_sph_gs_maps=6: derive the whole 216-bit GS IMAP/OMAP blocks
from the bound VTG image's OMAP and the PS image's IMAP at authoring
time (SPH-GS-MAP receipt; SPH-GS-SKIP names ps-invalid/gs-invalid/
vtg-overlap/vtg-short). maps=4 stays as the fallback.
- DS-MAP census line: dump stage-4/5 w5/w10/w13/w17 next to DS-IMAGE so
the fetched maps are visible without a new probe.
- MME pack carrier: reject maxa > 4096 (the IR pack array) instead of
64K, with MME-PACK-REJECT reason=bound.
- MME-DRAW-MOVE: anchor the slot-34 move after a 0x3810 VERTEX_STREAM
run. c848 moved CALL(34) to 0x204, ahead of the prepended stream bind,
so the 8x3 fetch used stale stream state.
- sema_gl_raster_enable=1: emit SET_RASTER_ENABLE=1 and
SET_GLOBAL_RENDER_ENABLE MODE_TRUE in the raster prelude (fresh replay
contexts may carry 0 for both, discarding every fragment).
Guest tooling (unblocks the render loop and capture):
- sdl_cap.c: asynchronous PBO+fence capture with a 5 s abandonment
bound; no glFinish/blocking readback (c899 wedged gzdoom in an
uninterruptible nv ioctl); clean sentinel from a destructor.
- gz_pbuf.c: request depth 24/stencil 8, log GL_DEPTH_BITS, and swap
with glFlush only.
- mission-gz.sh: gl_frames probe off by default; GZDoom launched
detached with a bounded wait, per-attempt X restart and retry;
capture in /dev/shm with GZCAP_MAXF.
- gzdoom_pbuf_run.sh: gl_pipeline_depth=2, gl_finishbeforeswap=0,
GZCAP_MAXF default, rc-file support for the detached run.
- run_cycles_long.sh: durable multi-cycle runner (the previous one lived
in /tmp and was lost on reboot); pushes tools, runs the mission, grabs
dmesg + guest logs + capture per cycle, stops on pixels.
Test matrix: params-r1624-maps{4,6}.txt.
Diffstat (limited to 'kernel')
| -rw-r--r-- | kernel/gsp_phys_booter.c | 188 | ||||
| -rw-r--r-- | kernel/gsp_sema_service.c | 11 |
2 files changed, 187 insertions, 12 deletions
diff --git a/kernel/gsp_phys_booter.c b/kernel/gsp_phys_booter.c index 55082e6..e61f3ea 100644 --- a/kernel/gsp_phys_booter.c +++ b/kernel/gsp_phys_booter.c @@ -18811,11 +18811,15 @@ static void gsp_phys_gr_mme_carry_arm(bool split, u32 poff, u32 maxa, /* * CONT 549: reject absurd pack totals. c890 decoded the packed * program's length word to 0x3D2F5F73-like garbage (expected 2799) - * and armed a continuation on it. A real golden IR is a few thousand - * words; anything past 64K cannot be a valid continuation total. + * and armed a continuation on it. eng89 tightens the bound to the + * IR pack array itself (GSP_PHYS_GR_RAW_WORDS = 4096): a total past + * the array can never complete and would make the builder's + * memcpy(&gsp_phys_gr_mme_ir_pack[poff], ..., chunk*4) overrun. */ - if (maxa == 0u || maxa > 0x10000u) + if (maxa == 0u || maxa > 4096u) { + pr_info("vgsp-gr: MME-PACK-REJECT reason=bound maxa=%u\n", maxa); return; + } if (!READ_ONCE(gsp_phys_rmcall_gr_mme_pack_carry) || !split || poff >= maxa || ((h_object & 0xfffff000u) != 0xbeef0000u && @@ -20483,7 +20487,7 @@ static u32 gsp_phys_rmcall_gr_sph_gs_maps; module_param_named(sema_gl_sph_gs_maps, gsp_phys_rmcall_gr_sph_gs_maps, uint, 0644); MODULE_PARM_DESC(sema_gl_sph_gs_maps, - "Synth GS SPH attribute-map variant: 0=pass-through(POS+COLOR both), 1=zero maps, 2=PS-exact(POS.W/POS.XYZW), 3=IMAP POS+COLOR / OMAP POS.XYZW (default 0)"); + "Synth GS SPH attribute-map variant: 0=pass-through(POS+COLOR both), 1=zero maps, 2=PS-exact(POS.W/POS.XYZW), 3=IMAP POS+COLOR / OMAP POS.XYZW, 4=IMAP=prior OMAP/OMAP=PS exact, 5=mesh-dummy, 6=live-derived from the bound VTG+PS images (default 0)"); u32 gsp_phys_rmcall_gr_sph_gs_maps_value(void) { @@ -20535,6 +20539,24 @@ MODULE_PARM_DESC(sema_gl_raster_prelude, "Emit CT_WRITE(0..7)=0x1111 + depth/stencil/scissor disable ahead of replayed 3D windows (default 0; A/B)"); /* + * eng89 store gate: the raster/global render enables are client channel-init + * state that a fresh replay context may not carry. If SET_RASTER_ENABLE is + * 0, every primitive is discarded in the backend regardless of a clean DS + * and a bound VB; if SET_GLOBAL_RENDER_ENABLE's MODE is FALSE the launch + * renders nothing. Emit the canonical enabled values ahead of the window: + * 0x037c SET_RASTER_ENABLE <- 1 + * 0x0130 SET_GLOBAL_RENDER_ENABLE_A <- 0 (offset upper) + * 0x0134 SET_GLOBAL_RENDER_ENABLE_B <- 0 (offset lower) + * 0x0138 SET_GLOBAL_RENDER_ENABLE_C <- 1 (MODE_TRUE) + * Harmless when the state is already enabled. Default 1; A/B with 0. + */ +static bool gsp_phys_rmcall_gr_raster_en = 1; +module_param_named(sema_gl_raster_enable, gsp_phys_rmcall_gr_raster_en, + bool, 0644); +MODULE_PARM_DESC(sema_gl_raster_enable, + "Emit SET_RASTER_ENABLE=1 and SET_GLOBAL_RENDER_ENABLE MODE_TRUE ahead of replayed 3D windows (default 1)"); + +/* * eng67 DRAW-STATE finding (c797): the release launch has RT0/RT1 bound * (alias 0xa000000, 1920x1080, fmt 0xe6), vertex stream 0 bound at * 0x1e305dc and the DRAW_IAB emitted its 8x3-vertex draws (0xd74=0x15), @@ -25481,6 +25503,7 @@ if (gsp_phys_gr_slot_kept(j, sa_mask, u32 rr = 0, first = 0; u32 s34 = 0, s34span = 0, leadsp = 0; u32 prev = 0, prevsp = 0, prevm = 0; + u32 vs_end = 0; while (rr < win3d_n) { u32 h = win3d[rr]; @@ -25508,8 +25531,36 @@ if (gsp_phys_gr_slot_kept(j, sa_mask, if ((m - 0x3800u) / 8u == 34u) slot34 = true; } - if (call && !first) - first = rr; + /* + * eng89: never move the DRAW ahead of + * the prepended VERTEX_STREAM bind + * (CALL slot 2, macro 0x3810). The + * release_fill prepends that run so a + * multi-page VB is bound before the + * draw; c848 moved CALL(34) to 0x204, + * i.e. *before* the 0x3810 run, so the + * 8x3 vertex fetch used the previous + * (stale/zero) stream state. Anchor + * the move at the first run after the + * 0x3810 span instead. + */ + if (call) { + bool vs2 = false; + + for (k = 0; k < nk; k++) { + u32 m = m0 + 4u * k; + + if (m >= 0x3800u && + m < 0x3a00u && + ((m - 0x3800u) / 8u) == 2u) + vs2 = true; + } + if (vs2) + vs_end = rr + span; + if (!first) + first = vs2 ? rr + span : + rr; + } if (slot34 && !s34) { s34 = rr; s34span = span; @@ -25522,6 +25573,8 @@ if (gsp_phys_gr_slot_kept(j, sa_mask, prevm = m0; rr += span; } + if (vs_end && first < vs_end) + first = vs_end; if (s34 && first && s34 > first + leadsp) { mme_draw_dst = first; mme_draw_src = s34 - leadsp; @@ -29001,10 +29054,36 @@ static u32 gsp_phys_gr_pair_n; * Diagnostic only; the pair arrays are written under gsp_phys_rm_lock, so * READ_ONCE keeps the census from tearing a line during a live submit. */ +/* + * eng89: bit accessors for the SPH V4T1 attribute-map blocks. Word N bit B + * of the little-endian image is byte (N*4 + B/8), bit (B%8). IMAP occupies + * absolute bits 184..399 and OMAP 424..639 (OMAP = IMAP + 240), so mapping + * the whole 216-bit block between stages needs an unaligned bit copy (the + * blocks are 8 bits off a halfword). x86-only driver: LE byte order. + */ +static inline u32 gsp_sph_bit_get(const u8 *p, u32 bit) +{ + return (p[bit >> 3] >> (bit & 7u)) & 1u; +} + +static inline void gsp_sph_bit_put(u8 *p, u32 bit, u32 v) +{ + if (v) + p[bit >> 3] |= (u8)(1u << (bit & 7u)); + else + p[bit >> 3] &= (u8)~(1u << (bit & 7u)); +} + +/* Attribute-map block geometry (SPH V4T1): IMAP at bit 184, OMAP at 424. */ +#define GSP_SPH_IMAP_BIT 184u +#define GSP_SPH_OMAP_BIT 424u +#define GSP_SPH_MAP_BITS 216u + static void gsp_phys_ds_image_census(const char *tag, u32 h_object, const u32 s[6], const u32 u[6]) { u32 w0[6] = { 0 }, code[6] = { 0 }; + u32 w5[6] = { 0 }, w10[6] = { 0 }, w13[6] = { 0 }, w17[6] = { 0 }; u32 si, pi, n = READ_ONCE(gsp_phys_gr_pair_n); for (si = 0; si < 6u; si++) { @@ -29022,10 +29101,14 @@ static void gsp_phys_ds_image_census(const char *tag, u32 h_object, continue; ph = READ_ONCE(gsp_phys_gr_pair_hpa[pi]) + (va - pv); if (pfn_valid((unsigned long)(ph >> PAGE_SHIFT))) { - w0[si] = get_unaligned_le32( - (const u8 *)__va(ph)); - code[si] = get_unaligned_le32( - (const u8 *)__va(ph) + 0x80u); + const u8 *ip = (const u8 *)__va(ph); + + w0[si] = get_unaligned_le32(ip); + code[si] = get_unaligned_le32(ip + 0x80u); + w5[si] = get_unaligned_le32(ip + 20u); + w10[si] = get_unaligned_le32(ip + 40u); + w13[si] = get_unaligned_le32(ip + 52u); + w17[si] = get_unaligned_le32(ip + 68u); } break; } @@ -29034,6 +29117,9 @@ static void gsp_phys_ds_image_census(const char *tag, u32 h_object, tag, h_object, w0[0], code[0], w0[1], code[1], w0[2], code[2], w0[3], code[3], w0[4], code[4], w0[5], code[5]); + pr_info_ratelimited("vgsp-gr: DS-MAP %s h=0x%x st4 w5=%08x w10=%08x w13=%08x w17=%08x st5 w5=%08x w10=%08x w13=%08x\n", + tag, h_object, w5[4], w10[4], w13[4], w17[4], + w5[5], w10[5], w13[5]); } /* @@ -33075,6 +33161,72 @@ gr_page_unmapped: gsp_phys_gr_sph_alias_receipt( fam, fam + 0x900u, 0, vpg->hpa, "auth"); + + /* + * eng89: live-derived GS + * attribute map (maps=6). + * The DS stage-4 check2 + * compares the GS IMAP with + * the previous stage's OMAP + * and the GS OMAP with the PS + * IMAP. Hardcoded maps=4 is + * exact only for one family + * pair; derive the whole + * 216-bit blocks from THIS + * family's VTG image (vsrc) + * and PS image (fam+0x0) so + * the map is exact for any + * pair. The GS header is + * header-only (the stage is + * bound with ENABLE=0); only + * the map words change. + */ + if (gsp_phys_rmcall_gr_sph_gs_maps_value() == 6u && + gsp_phys_rmcall_gr_sph_gs_img_value()) { + u32 gs_off = gsp_phys_rmcall_gr_sph_gs_addr_value() ? 0xe80u : 0x780u; + u8 *page = (u8 *)__va(vpg->hpa); + u8 *gs = page + gs_off; + u32 gw0 = get_unaligned_le32(gs); + u32 pw0 = get_unaligned_le32(page); + u32 bi; + bool pok = ((pw0 >> 10) & 0xfu) == 5u; + bool gok = ((gw0 >> 10) & 0xfu) == 4u; + bool ovl = (vlen > 0x580u) && (gs_off == 0xe80u); + + if (!pok || !gok || ovl || + vlen < 80u || gs_off + 128u > GSP_PHYS_PAGE_SIZE) { + if (sph_auth_log < 64u) { + sph_auth_log++; + pr_info("vgsp-gr: SPH-GS-SKIP off=0x%x reason=%s vlen=%u pw0=%08x gw0=%08x\n", + gs_off, + !pok ? "ps-invalid" : + !gok ? "gs-invalid" : + ovl ? "vtg-overlap" : + vlen < 80u ? "vtg-short" : "bounds", + vlen, pw0, gw0); + } + } else { + for (bi = 0; bi < GSP_SPH_MAP_BITS; bi++) { + gsp_sph_bit_put(gs, GSP_SPH_IMAP_BIT + bi, + gsp_sph_bit_get(vsrc, GSP_SPH_OMAP_BIT + bi)); + gsp_sph_bit_put(gs, GSP_SPH_OMAP_BIT + bi, + gsp_sph_bit_get(page, GSP_SPH_IMAP_BIT + bi)); + } + dma_wmb(); + if (sph_auth_log < 64u) { + sph_auth_log++; + pr_info("vgsp-gr: SPH-GS-MAP src=live off=0x%x w5=%08x w10=%08x w13=%08x w17=%08x\n", + gs_off, + get_unaligned_le32(gs + 20u), + get_unaligned_le32(gs + 40u), + get_unaligned_le32(gs + 52u), + get_unaligned_le32(gs + 68u)); + } + gsp_phys_gr_sph_alias_receipt( + fam, fam + (u64)gs_off, 0, + vpg->hpa + gs_off, "gsmap"); + } + } } } } @@ -33386,6 +33538,22 @@ gr_page_unmapped: in[nfin++] = 0x20000000u | (1u << 16) | (0x0e00u >> 2); in[nfin++] = 0u; + if (READ_ONCE(gsp_phys_rmcall_gr_raster_en)) { + in[nfin++] = 0x20000000u | (1u << 16) | + (0x037cu >> 2); + in[nfin++] = 1u; /* RASTER_ENABLE TRUE */ + in[nfin++] = 0x20000000u | (1u << 16) | + (0x0130u >> 2); + in[nfin++] = 0u; + in[nfin++] = 0x20000000u | (1u << 16) | + (0x0134u >> 2); + in[nfin++] = 0u; + in[nfin++] = 0x20000000u | (1u << 16) | + (0x0138u >> 2); + in[nfin++] = 1u; /* MODE_TRUE */ + pr_info_ratelimited("vgsp-gr: RASTER-ENABLE emitted raster=1 gre=1 h=0x%x push=0x%llx\n", + req->h_object, req->push_va); + } } /* * eng67: full-RT viewport for the replay context (see diff --git a/kernel/gsp_sema_service.c b/kernel/gsp_sema_service.c index 455adff..69b7341 100644 --- a/kernel/gsp_sema_service.c +++ b/kernel/gsp_sema_service.c @@ -13127,9 +13127,16 @@ static void svc_gl_flush(struct nv_vgpu_device *vgpu, gs_words[17] = 0u; } else if (gs_maps == 3u) { gs_words[17] = 0u; - } else if (gs_maps == 4u) { + } else if (gs_maps == 4u || gs_maps == 6u) { /* IMAP = previous OMAP; OMAP = - * PS IMAP exactly (POSITION.W). */ + * PS IMAP exactly (POSITION.W). + * maps=6 is the live-derived + * variant: the booter rewrites + * the whole 216-bit blocks from + * the bound VTG/PS images at + * authoring time; these words + * are the safe fallback when the + * booter cannot derive. */ gs_words[5] = 0xf0000000u; gs_words[10] = 0x0000000fu; gs_words[13] = 0x00008000u; |
