| Commit message (Collapse) | Author | Age |
| |
|
|
|
|
|
|
|
|
|
|
|
|
| |
- gsp_phys_booter: phys_gr_dump_push_all scans every beef and 0x1000013
push for the MME IR/SA ports (0x114/0x118/0x11c/0x120) with a one-line
PUSH-SCAN receipt and full-dumps upload-bearing pushes. c1030/c1035
receipts: 14895-word beef push m114=0/m11c=0; draw push 0x400f9d8
m114=0 m1b00=6 (CALL(0) 0x3800 data=0xd48 present, no macro body).
- cycle-r1780: kill QEMU by exact pidfile PID (never pkill -f), refuse a
foreign qemu, poll the pidfile after daemonize, virtio-vga display.
- params-r1624: enable dump-all + any-channel prelude capture + draw
census; fix the stale gsp_phys_gr_dump_push parameter name.
- tools/guest-diag: demo-video.sh (Xvfb+XFCE desktop, nvidia-smi, GZDoom
segment, honest CUDA battery) and qmp_record.py (QMP screendump).
|
| | |
|
| | |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
per-draw MME macro body is the gate
- 11 wiki pages refreshed: the offscreen EGL/GBM shim gives the guest a real
NVIDIA GL context without Xorg/DDX; the real GZDoom draw window (12068 B
at push 0x402f9d8, op=4 CALL(0) 0x3800 imm 0xd48 plus slots 44/57/58) is
recognised, admitted bind-free, chunked and replayed verbatim.
- CALL(0) executes without trapping (eng89t resolve through client SA[0]
then golden, not-resident calls elided); c1026/c1027 entered the captured
macro and c1027 reached the first DS with real stage data. Trap 0x80001528
decoded as a sticky TRAPPED_ADDR latch, not a cause/PC.
- Remaining gate recorded precisely: the client's per-draw MME macro body was
never captured (the prelude carries only the 12 setup macros), with the
synthetic golden body the surviving wedge; stores still wrote=0, no px.
- STATE-CURRENT.md snapshot prepended; wiki published to open-iov.org
(revs 24221-24231; 68 pages audited, 0 problems).
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
MME-CALL-RESOLVE names the target served for CALL_MME_MACRO(0):
imm-as-IR-address when the captured upload covers it, else the client
SA[0] (present bit; entry 0 is a valid IR address), else golden SA[0].
On the live 570.144 capture the CALL imm (0xd48) is a macro
length/argument, not an address: the client's SA[0] is IR 0 and the
59580-byte prelude holds only 12 x 0x114 uploads / 573 code words.
The c1025 wedge came from releasing CALL(0) on a launch whose packed
program was not re-emitted after the per-fire channel re-register
(MME-PRELUDE part words=0/647 -> badf1201 pc=513 mac=0xf1). With
sema_gl_call0_unmapped_drop=1 (default) an op-4 CALL(0) whose resolved
body is not resident for THIS launch (no prelude words here and no
IRAM-gate live generation) is elided to a NOP and logged as
MME-CALL-UNMAPPED; the ladder continues instead of trapping.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The real 570.144 direct-GL draw window (12068 B @ push 0x402f9d8) carries
no 0x3910/0x3914 write. Its draw is an op=4 IMMD to method 0x3800
(imm=0xd48=3400) plus CALL slots 44/57/58 (0x3960/0x39c8/0x39d0), and the
window has no colour-target bind of its own, so RT-gated admission sent it
down the compute-QMD lane. Recognize and admit the real marker: the walker,
chunk check and window marker accept 0x3800/0x3960/0x39c8/0x39d0 (0x3910/
0x3914 kept); the call-0 admission predicate scans the admitted window
(a pre-window run header can swallow the window head); the draw chunk is
forced first (draw_slice=1, order 0) and the window replayed verbatim.
The booter's DRAW_IAB window predicate accepts the op=4 CALL(0) and the
secondaries, the census names slot 0 (MME-CALL-ARGS slot=0 data=), and the
final emitted window gets MME-EXEC-CALL slot=0 (the physical GR front-end
executes the CALL; the kernel has no CALL interpreter, the 0x114 prelude
uploads precede the window in the same launch).
Guest harness: add tools/guest-diag/gz_es.c and an on-demand build in
push_gl_tools.sh; GZ_EGL=1 switches gzdoom_pbuf_run.sh/mission-gz.sh to
SDL offscreen + real NVIDIA EGL/GLES (default unchanged); push_gl_tools.sh
creates the NVIDIA EGL vendor ICD only when missing.
Receipts: c1025 GL-FLUSH admitted=1, DRAW-WRITE m=0x3800 op=4 data=0xd48,
3D mark admitted (CALL(0) 0x3800 draw marker), 3D stream chunked, 3D chunk
launch draw=1, MME-CALL-ARGS slot=0 data=0xd48, MME-EXEC-CALL slot=0
op=4 imm=0xd48. Exact 0x402f9d8 post-fix replay still unverified (run
wedged on an earlier draw window); stores/px remain open.
|
| |
|
|
|
|
|
|
|
|
| |
Add a diagnostic-only module param (default 0) that dumps the launch
request's raw guest push (req->raw, raw_words dwords, raw[0] = push_va)
to dmesg in 16-dword PUSH[i] rows, gated to pushes >= 512 dwords. This
lets the real client method stream be scanned offline for draw encodings
(CALL_MME_MACRO slots 0x3800/0x3910/0x3914/0x3960/0x39c8/0x39d0, BEGIN
0x1618/0xd78, run/shape census) without a guest /dev/mem read. No
behaviour change when clear.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
to end, guest submission is the wall
- 11 wiki pages refreshed: the real frame pushes are admitted (persisted
colour-target bind, draw markers inside counted runs, 40-chunk full-window
replay, 128-slot ring); the beef draw lane latches a clean DS at the
live-derived GS map; capture-less boots get the golden 2799-word MME pack;
CALL hold/carry fire/SA re-arm fire with the latch kept across the channel
recycle.
- Decisive finding recorded as guest-side: 36 boots (c990-c1019) show no
non-null CALL(34) draw macro from the client (interior 0x3910/0x3914
writes carry data=0x0) plus guest EGL/dri2 failures and OOM kills.
- GZDoom video delivered via the guest software GL path (Mesa llvmpipe) is
documented as a non-vGPU deliverable.
- STATE-CURRENT.md snapshot prepended; wiki published to open-iov.org
(revs 24210-24220; 68 pages audited, 0 problems).
|
| |
|
|
| |
through hold/save
|
| |
|
|
| |
admission; admit a header-DRAW window
|
| |
|
|
|
|
|
|
|
| |
The c996-c1001 DRAW-SCAN false-positive: a 4353-word op=1 state run whose
m0=0x804 has 0x3910 at interior offset 0xC43, so the scan claimed the draw
was present, forced a 4353-word 'draw slice', and the real CALL header
(op=5 cnt=10 m0=0x3910) never reached the held-call path. A CALL can only
be a run header, so match m0 only (booter window scan and the service
walker/chunk scan).
|
| |
|
|
| |
clear damage on REARM
|
| |
|
|
| |
chunk of an admitted window
|
| |
|
|
|
|
|
|
| |
c970-c983 showed up to 105 CAND-OVERWRITE per boot with 40 chunks
queued against a 48-slot ring; the chunk carrying the DRAW_IAB can be
evicted before it is forwarded or saved (census slot34=0 everywhere).
Raise the ring to 128 so a full 40-chunk window plus the module lane
fits without eviction.
|
| |
|
|
|
|
|
|
|
|
|
| |
Every c970-c977 launch logged 'MME-PRELUDE copy raced an append
(gen 1196248389->0); will re-emit next launch': the seqcheck compares the
capture generation before/after the copy, and the synthetic golden gen can
never match the empty capture's 0, so the prelude latch never set, the
IRAM gate never verified, the carriers stayed load-only and no CALL/SA
release could run - despite MME-GOLDEN overlay maxa=2799 and a complete
pack. Accept the latch when g0 is the synthetic golden generation (the
golden image is immutable and pl_n==0, so there is nothing to race).
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
c963 got a CLEAN DS (ds_after=80000000) on the admitted DRAW chunk and
then poisoned: badf1201/pc=513/mac=0xf1. Root cause (audit): the beef
client's MME upload was absent, so pl_n=0/pl_gen=0 and every prelude /
pack / CALL-hold / rearm gate was skipped; the DRAW CALL ran against an
EMPTY MME program. The DS proves the draw path itself is correct.
Give a beef launch with no capture a synthetic generation ('MMEG') when
the prelude pack + hold calls + golden + golden_force knobs are on. The
existing machinery then builds the pack from the golden image alone
(2799 words, slot 34 -> 0x265), emits it in split parts, holds the real
CALL runs until the latch is set, rearms the SA and releases the saved
window. Module/X lanes are untouched (beef family gate).
Expected receipts: MME-GOLDEN overlay maxa=2799, MME-PRELUDE pack parts,
MME-CALL-HOLD window=..., MME-PACK-REARM, MME-CARRY-FIRE, mac=0x22,
GUEST-WIT surface stores, px0!=0.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
| |
eng89i admitted the 14837-word window and chunked it into 6 chunks, but
6*384 words covers only 18% of the window - the DRAW_IAB call at the end
was never replayed. Raise SVC_GL_STREAM_CHUNK_MAX 6 -> 40 (39 chunks for
14837 words at the proven 384-word chunk) and the candidate ring 32 -> 48
so the chunks are not evicted before the pump drains them.
Context: c952-c957 submitted chunk 0 and got a FECS priv-error (badf1201,
mme_pc=513 mac=0xf1) before the release; c956 reached MME-CARRY-ARM
(off=1632/2799) and CALL-HOLD window=121, then CANCEL reason=
chan-ctx-reset. The chunk-0 poison and the carry fire remain the next
gates after full-window coverage.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
c946-c951 admitted the big 14865-word draw window but refused it at
raw_ok=0: the raw[] capture kept only the last 4096 words of a 14895-word
push. Raise GSP_PHYS_GR_RAW_WORDS/SVC_GR_PUSH_WORDS to 16384 (64 KB
static each; the IR pack stays 4096 and its maxa bound unchanged) and the
candidate ring 8 -> 32 so a three-chunk window is not truncated.
Add WIN-START-SNAP: c949-c951 admitted start=0x78 where w0=0x3f800000
parses as op=1 count=0xf80 (data, not a method header; the real BEGIN
run is at 0xe8). If the start word is not a plausible method header, scan
forward up to 0x400 bytes for the first plausible one and replay from
there; bounded receipt names old/new.
c949 was the first boot with guest surface stores from the engine
(GUEST-WIT 601/2030 pages, va=0x120400000, post=0xff000000); the beef
big-window replay is now the remaining gate.
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
c940-c945: VP-METHOD 0x1618 op=1 in push 0x400108c while every
RT-ADMIT-BLOCK kept has_draw=0. The marker sits inside a counted INC run,
not at a run header, so the guard never saw it. Accept a marker covered
by the run's method span (m <= target < m + cnt*4 + 4) for the 0x1618 /
0xd78 / 0x3910 family in ops 1/3/5, keeping the op=4/5 exact match.
Context: c944 was the first boot with surface stores - GUEST-WIT 601/2027
pages changed at va=0x120400000 post=0xff000000 on the X/2D lane - so the
engine does write the surface; the beef draw window is the remaining
admission gate.
|
| |
|
|
|
|
|
|
|
|
| |
RT-ADMIT-BLOCK on c934-c938 showed the persisted bind (mask=1,
pchan=0x1000013, B=0x20400000 W=0x780 H=0x438) was always refused by the
draw-marker guard: the client encodes BEGIN 0x1618 as an immd13 (op 4)
and CALL_MME_MACRO(34) 0x3910 as ONE_INCR (op 5, 0xa00a0e44), while the
guard only accepted the op=1 INC form. Accept ops 1/3/4/5 and match the
method, which flips has_draw to 1 and lets the persisted admission reach
the replay lane.
|
| |
|
|
|
|
|
|
| |
The receipt prints state_push/rt_valid/rt0_first/rtp/mask/pchan/has_draw
at the beef draw-synth reject, so the next boot names which of the five
conjoined admission predicates failed. Runner: wait at most GZ_SECS+120
for mission-done (the mission itself writes it by GZ_SECS+30 when the
guest is alive; a dead boot no longer costs 7.5 minutes).
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
c918-c921 had no RT-PERSIST and every beef draw was rejected on the
missing colour-target bind, while c910 had the bind on the X/2D channel.
RT-SCAN logs the last A/B/W/H candidates and seen flags for every beef
channel push (bounded to 32) so the next boot answers whether the bind is
absent from the push, filtered by the surface-class checks, or never
authored.
The batch runner also streams the kernel log with 'dmesg -w' per cycle
and merges it into the per-cycle evidence; the ring wraps in under four
minutes at this debug volume and had truncated c921's key receipts.
|
| |
|
|
|
|
|
|
|
| |
c910 proved the persisted bind arrives on the X/2D channel
(RT-PERSIST h=0x1000013 B=0x20400000 W=0x780 H=0x438) while the beef
draw window is on 0xbeef0100, so the eng89c channel equality refused the
exact window it was meant to admit. The stored channel is now
diagnostic only; the draw-marker guard, the surface-class validity and
the per-session clear remain.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Four fixes from a code audit of eng89/eng89b, each of which could have
made the maps=6 experiment unobservable or harmful:
- Draw-move anchor: use the FIRST 0x3810 run only. The 247-word window
also carries its own 0x3810 later; the old code moved the draw to the
last one, i.e. behind the setup calls it must precede. Removed the
post-scan promotion.
- Prelude/stage budget: the raster-enable words (8) pushed the prelude
toward the 96-word cap and could skip the whole stage-bind block
(SPH-BIND-STAGE-SKIP reason=budget), hiding the GS bind. Array and
gates grown to 128.
- RT persistence: require B/W/H only (A optional, matching the existing
first-valid check), store the channel and require it for admission,
require a real draw marker (0x1618/0xd78/0x3910) in the window, and
clear the state in svc_session_flush so a recycled channel cannot
inherit a dead session's target.
- GS live map: moved out of the SPH-AUTHOR copy branch and derived from
the family page's actual VTG image at +0x900 (not the source record),
so an overlap skip no longer makes maps=6 inert. New gs_addr=0
parameter set (GS at 0x780) keeps the GS header clear of the VTG copy.
Test matrix: params-r1624-maps6-780.txt (maps=6, gs_addr=0).
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
- GZCAP_SYNC=1: synchronous readback path for software GL (the async
PBO+fence path never signals under Mesa llvmpipe, so 0 frames were
recorded; the GPU path keeps the async path).
- GZCAP_READBUF=back: read GL_BACK on a double-buffered X11 window
(Xvfb); the default GL_FRONT is correct only for the single-buffered
GPU pbuffer. Without it every Xvfb frame was black.
- File gate: only an SDL renderer process opens the capture, and an
existing non-empty capture is appended, never truncated. The old
'wb' open in every LD_PRELOAD child turned a 288-frame capture into a
4108-byte file; a helper without SDL also opened it via the
glReadPixels fallback.
- Log the read buffer in the async arm line.
Validated by the 2026-11-16 GZDoom video runs: 2200-frame captures,
clean_end=1, corrupt=0, 30 fps encodes, delivered via files.vgpu.io.
|
| |
|
|
|
|
|
|
|
|
|
|
| |
Most beef draw pushes carry no colour-target bind of their own (the
client binds the target once per context/target switch in an earlier
state push), so the per-window admission refused the real frame pushes
on most boots ('beef draw-synth: no valid last colour-target bind', the
shallow-boot lottery). Track the last complete, plausible bind seen on a
trusted push (same surface-class validity as the first-bind check; an
invalid later write never overwrites a valid state) and admit a beef
draw window on it: RT-PERSIST + '3D mark admitted (persisted
colour-target bind)' receipts. Knob sema_gl_win_rt_persist, default 1.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
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.
|
| |
|
|
|
|
|
|
| |
CONT.549/c890: MME-PRELUDE pack-part decoded the packed program's
length word to 0x3D2F5F73-like garbage (1602/1026586419, expected
/2799) and the carrier armed a continuation on it. gsp_phys_gr_mme_
carry_arm() now refuses maxa==0 or maxa>64K, so a corrupt length can
never arm a bogus pack continuation.
|
| |
|
|
|
|
|
|
|
| |
CONT.548/c886: a clean 247 window (VS-BIND base=0x450c1f4) reached
MME-CARRY-FIRE gen=4 and was immediately cancelled with reason=
request-busy: svc_gr_carrier_request() claimed the armed carrier (the
take clears the arm) before taking the forwarding slot, so the lost
slot lock left the pack uncompleted. Probe SVC_FWD_INFLIGHT before the
take so the retryable busy case returns without consuming the arm.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
| |
CONT.545/c880: the armed pack carrier (MME-CARRY-ARM gen=3 off=1506/2799)
never fired because the pump needs a matching CLIENT beef candidate and
the 2-frame mission ended before one re-appeared; it expired with no
FIRE/CANCEL/WAIT receipt. The pump's carry block now, when there is no
candidate but the channel's golden pack is still contiguously loaded
(gsp_phys_gr_mme_carry_ctx_alive: pack_load_h==carry_h && golden ready),
submits svc_gr_carrier_request() as its own synthetic candidate
(MME-CARRY-FIRE reason=self-candidate). A truly gone context is
reported MME-CARRY-CANCEL reason=ctx-gone, never silent. Release stays
beef-only; the eng68 chain still runs after the carrier fire (pack
complete -> release replay -> mac=0x22).
|
| | |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
(sema_gl_mme_pack_carry_client)
CONT.539: c858's split pack landed on the client channel h=0x1000013
and the continuation never armed because gsp_phys_gr_mme_carry_arm()
required (h & 0xfffff000)==0xbeef0000. c848 only worked because the
same bytes landed on h=0xbeef0100.
New knob sema_gl_mme_pack_carry_client (default 0) relaxes the arm and
take gates to the split's own owner channel; the synthetic windowless
LOAD-ONLY launch still goes through that owner's forwarding candidate
(the pump locates it by h_object), and the builder reads the global
pack cursor so the continuation bytes are identical. The saved-window
release carrier stays beef-only; its MME-CARRY-ARM mode=release fires
once the pack becomes resident (the owner completion clears the
pack-not-resident gate on the next pass). Params-r1624 enables it.
|
| |
|
|
|
|
|
| |
The previous commit appended a duplicate; the inline line-1 params
already carried sema_gl_raster_prelude=1, so c848 ran WITH the raster
prelude and still saw the SPH18 DS with no stores - the scissor/depth
is not the remaining gate; the GS SPH18 is.
|
| |
|
|
|
|
|
|
|
|
|
| |
Plan B from CONT.538: the 247 window carries 0x12cc/0x12e8/0x1380=0
(SET_DEPTH_TEST / SET_DEPTH_WRITE / SET_STENCIL_TEST, per the code
comment at the knob) and no SET_CT_WRITE / SET_SCISSOR_ENABLE - those
are client channel state. On a replayed 3D DRAW_IAB the replay
channel keeps hardware defaults, so every fragment can be discarded
(CT_WRITE=0). Enable the existing sema_gl_raster_prelude remedy
(CT_WRITE(0..7)=0x1111, depth/stencil disable, SCISSOR_ENABLE(0))
alongside gs_img=1/maps=2/TCS-off.
|
| |
|
|
|
|
|
|
| |
c850-c853 (maps=0, pass-through): the 247-window-class boot did not
reproduce in these cycles (c850 no release, c851 poison 0xbadf1201,
c852 a 730-page upload burst with no DRAW_IAB release lines), so the
SPH18 read needs a maps=0 247 boot that did not occur. Step 2 of the
A/B per the plan: maps=2 (IMAP PS-exact) with gs_img=1 and TCS off.
|
| |
|
|
|
|
|
|
| |
c848: with the correct VTG VB base the DRAW_IAB runs and the hardware
DS now flags SPH18 (GS check2): DS-FAST ds 80000000->81040000. The
header-only GS SPH (gs_img=1, family+0xe80) with maps=4 (the r1569
variant) still fails the GS attribute-map check. Step 1 of the A/B:
maps=0 (pass-through IMAP/OMAP). Keep TCS off (stage_disable=0x04).
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
c846: the 150-window release fired with VS0=0 because the held window
was saved before the client's VERTEX_STREAM CALL(2) was captured
(0x3810 lives in a later push); the dispatch-time sticky payload
(end=0x450d100 base=0x450b5dc) arrived after the save. The release
fire happens seconds later, so build the state in
gsp_phys_gr_mme_release_fill: with the current 5-word payload, prepend
[0x1c00 fmt=0x1014, base] + [0x600 size=end-base] + [0x3810 the full
5-word call] to the saved window, skipping anything it already
carries. MME-VS-FILL receipt; the normal patch relocates the pair and
the 0x3810 range.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
c836 proved the client's VERTEX_STREAM CALL(2) payload is
[stream, end_hi, end_lo, base_hi, base_lo] (MME-VSTREAM:
end=0x1e02100 base=0x1e001f4, MME-VSTREAM-LINE base=0x450c1f4
end=0x450e100 pages=3 first8=...3f800000 3e4ccccd...). eng79 wrongly
took d[2] (the END 0x450e100, a zero page) as the stream base; the base
d[4]=0x450c1f4 carries the float2 vertices. The Doris slot-2 sim with
the full 5-word payload emits LOCATION 0x1c08<-base and SIZE
0x604<-end-base.
- service: svc_gl_vs_call[5] (getter out[5]) - capture v0..v4, log 5
words; eng82 page-set collection adds the end (v2) and base (v4).
- booter eng79: slot-2 base = vcall[4] (require vcall[1]==0,
vcall[3]==0, aligned), end = vcall[2]; size pair = end-base; receipt
base=.. end=..; VS-BLOB2 dumps at the base.
- booter eng77: emit the FULL 5-word 0x3810 call (op=5 cnt=5) even when
the window carries its own call (the slot-34 reorder moves the draw
ahead of it); the 0x3810 MME-VSTREAM patch maps [base,end] into one
contiguous shadow line.
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
c840: VS-BLOB2 va=0x450e100 unresolved - the slot-2 VERTEX_STREAM
payload's VB page (0x450e000) was never part of the forwarded launch's
page set, so the sema resolver cannot map it and the synthesized
stream-location pair cannot be relocated (the engine would fetch an
unmapped guest VA).
svc_gl_field now handles 0x3810: every plausible address word
(>=0x100000, <2^40, 4-byte aligned) in the CALL(2) payload is added
via svc_gl_add_ptr, so the candidate/forward page set includes the VB
page and the release's pair relocation can resolve it.
|
| |
|
|
|
|
|
|
| |
The slot-2 VERTEX_STREAM VA (0x450e100) is now bound as the stream
location; the release still stores nothing. Dump the first 64 B of the
client VB (resolved HPA) at save time (VS-BLOB2) so the 24x20 B float2
vertex payload is verifiable/rejectable, and compare the relocated
shadow (VS0) against pagebase+client_offset (the suspected 0xf4 skew).
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
c835: the 151-word DRAW window (slot 34, no 0x3810 call) was saved and
released; the client's VERTEX_STREAM_TURING CALL(2) was captured 11 s
later from the 247-word state push, too late for that save. The
247-word window carries BOTH slot 34 and the 0x3810 call (relocated
d2=0x1e22100).
release_save now supersedes a saved record without a 0x3810 run when
the new window carries one, so the release uses the window that has
the VTG stream state (and eng79's slot-2 VA synth pair can bind the
real VB base).
|
| |
|
|
|
|
|
|
|
|
|
|
| |
The client's VERTEX_STREAM_TURING CALL(2) payload [0,0,0x0450e100]
carries the real VTG vertex buffer VA; the slot-2 macro itself emits
location=0 / size=<VA> (Mesa slot-2 sim), i.e. the base is never
written to 0x1c04/0x1c08. Synthesize the stream-location pair with the
CALL(2) VA as the base instead: release_save takes gsp_sema_gl_vs_call()
and, for stream 0, prefers payload[2] (masked) over the scanned CB
candidate, emitting the usual [fmt=0x1014, upper, lower] pair + the
size pair. The patch relocates the VA like any 0x1c00 pair.
MME-VS-BIND receipt gains src=slot2|scan|sticky.
|
| |
|
|
|
|
|
| |
The 16-hit cap hid later CALL slots (the VARRAY_STAGING_BUF_TURING
slot 1 / 0x3808 base binder) in the 728-byte beef push. 32 hits and a
1600-byte line keep the whole 0x100..0x1ff + 0x3800..0x39fc set in
one receipt.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The beef 728-byte push carries CALL_MME_MACRO(2) (method 0x3810,
golden slot 2 = VERTEX_STREAM_TURING) with payload [0,0,0x0450e100]:
the Mesa mme_tu104_sim of golden slot 2 with those args emits
SET_VERTEX_STREAM_A_LOCATION 0x1c04<-0, 0x1c08<-0, and the Turing size
pair 0x600<-0, 0x604<-<the VA> - the VTG stream state the per-frame
DRAW window slices omit. The census (eng76, 0x100..0x1ff) finds it at
0x148 of the 728 push; the 247-word window carries the relocated call
at 0x2e0 (d2=0x1e22100), the 151-word DRAW window does not.
- service: sticky svc_gl_vs_call[3] from every 0x3810 run (VS-CALL
receipt), getter gsp_sema_gl_vs_call().
- booter release_save: when the window has no 0x3810 run, prepend the
captured ONE_INCR run (0x3810<-payload) ahead of the DRAW window;
the normal patch relocates the VA payload (sema_gl_mme_reloc).
- vs_pre buffer 12->24 words. Census range stays 0x100..0x1ff (eng76).
|
| |
|
|
|
|
|
|
| |
The beef pushes carry 0x114/0x11c/0x124 runs; extend sema_mme_census
from 0x114..0x124 to the whole 0x100..0x1ff page so every setup
method in that family (3 data words/hit) is visible in the beef
windows (128/151/247/324/491 words) and the X 221-word window. The
existing 0x1c00..0x1cf8/0x600..0x61c ranges stay in the same line.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
| |
c824/c825: the release RT0/RT1 target is 0xa000000 (the data-rec
bounce base from gsp_phys_gr_surf_launch_target's rec-over-alias
preference) while the aliased 1920x1080 X-screen surface is 0x15d0000
(2025 pages). The changed pages (0x120700000/0x120704000/0x120705000)
are inside the screen range but the draw's RT offsets are computed
against the bounce base.
sema_gl_surf_alias_pref (default 0; params=1): after the SURF prelude
patch, rewrite its target words to gsp_phys_gr_surf[].alias so the
prelude and the RT0 copy (r1483) both name the contiguous surface
alias. SURF-ALIAS-PREF receipt.
|
| |
|
|
|
|
|
| |
The raster-kill A/B never reached a release: c794 and c829 both hit the
badf1201 held-window poison before the release. Revert the param to the
known-good TCS-off + vs_autobind configuration; the raster_extra
hypothesis stays open for a clean-boot A/B.
|
| |
|
|
|
|
|
|
| |
The release now binds VS0/SZ0 (scanned float2 vertex data), viewport and
RT, runs the DRAW_IAB, but no surface page changes. The neutral
blend/alpha/cull/point/ZCULL state is the untested fragment-kill lever
(c794 poisoned before the release); re-enable it in params for the
c828+ A/B.
|
| |
|
|
| |
end (c788); TCS-off DS clean; remaining blockers are the vertex buffer (eng74 VS-BLOB) and render-target store/alias visibility (eng74 alias WIT); stores still 0
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
| |
1. VS-BLOB: on a VS-SCAN candidate, print the 16 words at the winning
block and the DRAW range (starts 0,3,..,21 * stride 20 + 8), so the
CB-page heuristic candidate can be validated/rejected against the
Mesa DRAW_IAB shape.
2. Surface alias WIT: gsp_phys_gr_surface_alias_locked records every RT
guest page (VA+HPA) it maps in a persistent table; the launch WIT
capture notes them, so a render-target store through the 0xa/0xc
alias becomes a GUEST-WIT changed line. This fixes the eng73 gap
(the surface alias does not use the pair alias cache).
params: sema_gl_wit_shown=40 so the RT page is not hidden by the 4-line
budget.
|