# session 77 (r683/r684) — vkCreateDevice PASSES: the 0x7E event validation and the readiness-word completion

## The two-part unblock

1. **r682 fix validated** (`e3c1cdc5`, loaded on r683/r684): the
   nvkms/Vulkan path allocates its completion event as
   NV01_EVENT_KERNEL_CALLBACK_EX (class 0x7E); the emulation now
   registers it alongside 0x79. Live proof (r683, sema_dbg=1):

       GSP_RM_ALLOC class=0x7e paramsSize=24 hClient=0xc1e00011
           hParent=0xcaf00001 hObject=0xcaf00003
       event registered hEvent=0xcaf00003 hSrc=0x0
           notifyIndex=1006632972 (off 8, slot 13)
           params=1100e0c1000000007e0000000c00003c0000000000000000
       EVPOST all ev[13] cl=0xc1e00011 ev=0xcaf00003 nidx=0x3c00000c rc=0

   (notifyIndex 0x3c00000c = subdevice 0x3c, index 0xc =
   NV2080_NOTIFIERS_GRAPHICS). The guest consumes the records
   (`guest_read == our_write` in the rmstate dump), but the guest's
   event dispatch does NOT write the event's notification word — the
   residual delivery gap.

2. **The completion word found and driven (the actual unblock)**. The
   main-thread vkCreateDevice wait (eglcore 0xdceb80, session 76) reads
   a u32 readiness word: `should-wait = (word != 1)`; the word lives at
   `+0x80` of a **16 KB rw-s /dev/nvidiactl mapping** (the event
   notification buffer; holder chain sub -> obj2 -> holder, resolved
   live by t79.py). The host can write it directly in the guest RAM
   file:

       PA = pagemap PFN << 12  (+0x80);  file offset = PA - 0x80000000
       (above-4G RAM; below-4G would be PA itself per the cycle's
       guest-RAM registration: gpa 0x0/0x100000000 -> off 0x0/0x80000000)

   Writing 1 flipped the wait and **vkCreateDevice returned success**
   (`VKDEVICE ok family=0`) — reproduced three times:
   * r683: manual write after a t79 probe;
   * r684: nudger (all 16 KB mapping words) + slotfill restart;
   * r684d: probe -> single-word write (the reliable recipe).

## The reliable recipe (per run)

1. r682+ module; `sema_dbg=1 gsp_rm_ev_log=300` optional.
2. Launch `vk_frame` + `slotfill2.sh` (EGL completion-slot fills; the
   slots re-arm, so re-run slotfill if the client is stuck in the
   sched_yield spin before the word-wait).
3. Wait for the main thread in `do_poll` (the word-wait).
4. `t79.py` (gdb) -> `WAIT word=0x...`; guest pagemap -> PFN -> PA;
   host write `01 00 00 00` at `PA-0x80000000`.
5. `VKDEVICE ok` follows within ~10 s. The client then proceeds to
   image/memory/pipeline setup; its next wait is NOT the same word
   (the word stays 1) — next session: probe the follow-on state
   (attach fails while the client spins in the post-device phase;
   use the /proc-only checks + a longer observation window).

## Cautions learned

* Writing ALL 16 KB mapping words (the multi-page nudger) can trip the
  -4 give-up path (an RM error flag word gets written): restrict nudges
  to the single probed word.
* Attaching gdb to the client while it is in the tight post-device spin
  can hang the attach (leave a `timeout`-wrapped gdb, then kill that
  gdb by PID and `kill -CONT` the client — the pending SIGSTOP from a
  dead tracer otherwise wedges it).
* zsh truncates 64-bit hex literals in `$(( ))`: compute PFN/PA in
  python, not zsh.

## Tools

* `t79.py` — resolves the wait chain (sub -> obj2 -> holder -> word VA)
  and prints the client's nvidiactl rw-s mapping sizes.
* `guest_words.sh` — prints VA/PA of the +0x80 word for every 16 KB
  nvidiactl mapping (pagemap; run in the guest as root).
* `nudge.sh` — host-side loop: poll guest_words.sh and write 1 to any
  zero word (multi-page variant; prefer the single-word recipe).

## ADDENDUM (r685, session 78) — POST-DEVICE WAIT IDENTIFIED: the NvOgl (GL ucode) engine init; push windows now READABLE (resolver6)

* The post-device spin is the **GL module bring-up**: the main thread's
  A8-layer retry loop (eglcore 0xa85d24) works on a work item whose
  embedded name is ASCII **"NvOgl_m0"** (+0xf04: "NvOgl_m40"); the D6
  submit layer (0xd624e0) ORs **0x200000** into its pending word (the
  same engine-pending bit SESSION 75/76 watched on E+0x150), then the
  A8 loop yields via the B284B0 spin helper (sched_yield) waiting for
  the GL ucode init to complete on the GR engine.
* **The push windows ARE now readable** (first time): the resolver6
  recipe (window-record pair {push_va,0x20000}+{ring_va,0x40000} ->
  same-page 0x20000 memdesc PA -> `sema_map_add`) injected
  `0x4480000 -> 0x120cfb000` (content `2001c00d 70000000 2001c00d
  68000000` = the 2-method semaphore-like push) and `0x4ca0000 ->
  0x12e51e000` (the 1432-byte GR init stream: SET_OBJECT 0xc5c0/0xc5b5,
  `0x200121ed`, `0x200120a9`, bulk methods, tail `2001c00d
  70/68`). The service now decodes the streams (`sema-push ... tail=0dc0
  012000000070...`) where SESSION 73-74 saw "unread".
  CAUTION: the resolver injects every candidate; a zero-page candidate
  injected last replaces the good mapping for that VA (re-inject the
  good page after the scan).
* **Remaining gate**: the service still reports `have_rel=0` for every
  GR push (the CE-shaped release matcher) and — decisively — the
  GR-forward machinery is NOT engaging: `drain held=4 starved=149
  forwarded=0 ... gr_nocap=4`. The GL ucode init work is never
  forwarded/executed, so the completion (and the 0x200000 bit clear)
  never arrives. Next: the GR forward/capture path for these channels
  (operand capture + release decode for the GR shape), then the client
  proceeds to VKFRAME.

## ADDENDUM (r686, session 79) — the low-VA release fix and the corrected injection recipe

* The GR/GL init releases decode only with the low-aperture acceptance
  (`SEMA_VA_LO` bound relaxed to >= 0x4000000 in both release forms,
  commit cb8b75c5): live `rel=0x4b19b10 pay=1` (REPORT_SEM group
  `200426c0 00000000 04b19b10 00000001 10000000`), `rel=0x4b17bb0
  pay=3` and `rel=0x4b17e30 pay=1` (0x5c-form `2005c017 04923060 ...
  02000001`); `have_rel=1`, `applied` rising, un-translated releases
  queue in svc_pending (417 observed).
* Corrected pair-specific push injections for r686:
  `0x4480000 -> 0x125692000` (record page 0x1163c9000, live desc),
  `0x4ca0000 -> 0x131649000` (page 0x122f20000), plus
  `0x4b16000 -> 0x12a76a000` (the {0x4b16000,0x1000} window).
  CAUTION: resolver7 injects each page's descs for BOTH pairs (union
  loop) - fix it to keep per-pair pages or inject manually from the
  per-pair scan. The window-record layout is {tag, VA, SIZE, 0,
  "tail", ...} stride 0x40 (search the VA u64 with size at +8 and
  "tail" at +0x10).
* Remaining: the semaphore VAs (0x4b17bb0/0x4b17e30/0x4b19b10) have no
  window records in the known record pages; the service cannot resolve
  them (targeted rm_mem miss) so the releases stay pending and the
  NvOgl wait spins. Find their PAs (client mapping pages / RM
  VA-tracker / live poll-watch), inject, and the pending releases
  should apply.

## ADDENDUM (r686, session 80) — ALL low-aperture window records found (the "liat" tag fix); the semaphore windows identified and their releases applying

* **The record scan fix**: the RM window record is 0x40 bytes
  {u64 tag; u64 VA; u64 SIZE; u64 0; u32 0x7461696c; u32 0}. The tag is
  stored as the u32 0x7461696c - on disk the BYTES are "liat", NOT
  "tail" (the hex display `7461696c` reads as "tail" only as a number).
  Searching for b'tail' finds nothing; `wall2_recscan.py` searches for
  the byte pattern 6c 69 61 74 and decodes {tag,VA,SIZE} at
  -0x20/-0x18/-0x10. It found **all 75 records** in one pass.
* **The GR init semaphore windows** (r686):
  `va=0x4b17000 size=0x2000 -> pa=0x12a76f000` (verified by content:
  +0xbb0=3, +0xe30=1 - exactly the rel=0x4b17bb0 pay=3 /
  rel=0x4b17e30 pay=1 payloads) and
  `va=0x4b19000 size=0x1000 -> pa=0x122f9b000` (after injection the
  release applied there: +0xb10 became 2). Injections:
  `echo 0x4b17000,0x12a76f000,0x2000,1 > sema_map_add` and
  `echo 0x4b19000,0x122f9b000,0x1000,1 > sema_map_add`.
* **State after the injections**: the low-VA releases decode
  (`have_rel=1`) and apply to the semaphore pages; `pending` hovers at
  ~418 (the client re-kicks continuously); the NvOgl wait still spins
  (the A8 loop compares per-entry counter tables r14+0x5e8 vs the
  wanted set - the completion is not just the semaphore value). Next:
  the A8 counter tables' source (what fills r14+0x5e8 / the "wanted"
  table - likely the GR channel's own progress counters) or the GR
  forward/ucode execution (`forwarded=0 gr_nocap=7` still).

## ADDENDUM (r686, session 81) — the counter-autoreg gap analysed: the counter page 0x20400f000 is unresolvable; the manual `sema_counter` registration is the bypass

* `svc_counter_autoreg` (per-pass for eng==0 channels, n_gr=20) needs a
  translation for the rank slot VA `0x20400fec0 + (19-r)*0x10`
  (r = gr_index = rank by ascending handle among the guest's GR
  channels): either an ipin pin (`svc_ipin_lookup`, 9275 entries live
  from the guest pin stream) or `svc_counter_derive` (the
  vaspaces/discovered/targeted walk chain, `sema_counter_derive=1`
  default). Both miss for the counter page -> no "counter auto" log,
  `pub_calls=0` -> the A8 loop's per-entry counter tables never fill and
  the NvOgl wait spins.
* The client does NOT map the counter pool: `/proc/<pid>/pagemap` for
  0x20400f000/0x20400ffd0/0x20400ffc0/0x20400ffb0 is all-zero (not
  present) and no VMA covers it - it is a UVM kernel-side object, so the
  pin stream cannot resolve it through the client's page tables. The
  stream's log shows `sem=0` (zero semaphore resolutions) even though
  the want-list (`/tmp/guest_want_vas.txt`, host-written) contains 44
  `20400f` entries; the stream's "channel-anchored expansion" for the
  UVM tracking sems (guest_pin_stream.py ~line 950-1010) is the
  machinery that must find them (it is written for the CUDA second-client
  stall - "the only live uvmt event is the spinning sem").
* Bypass available now: the `sema_counter` parameter registers a channel
  slot manually - `echo <h_object>,<pa> > .../parameters/sema_counter`.
  Once the counter page's PA is known (via the guest kcore/UVM walk),
  this wires the publish path without a kernel rebuild.
* The rank slots for the GL client's channels (handles 0xbeec0100,
  0xbeec0104, 0xbeee0100 -> ranks 0,1,2): 0x20400ffd0 / 0x20400ffc0 /
  0x20400ffb0.

## ADDENDUM (r686, session 82) — the counter-page PA hunt: pagemap route confirmed dead, vidmem-window hypothesis tested and not supported

* The client's own pagemap does NOT contain the counter pool: the
  semfix_uvmt tool's resolver (`guest_pmap_win.py <pid> 0x20400f000
  0x204010000 64`) returns NOTHING, as does the UVM pool range
  (`0x121000000 0x121100000`). The tool's `pin_counter_page` has
  therefore never succeeded for these clients (the STALL WATCHDOG's
  "re-pin counter" attempts across sessions produced no "counter:
  pinned" line).
* VIDMEM-window hypothesis tested: the counter VA 0x20400f000 is in the
  emulated vidmem window (>= SVC_VIDMEM_VA_BASE 0x200080000), so its
  window offset is 0x3F87000 (0x20400f000 - 0x200080000). The shadow
  there is reachable and all-zero (`sema_vidmem_dump 0x3f87000,64`);
  poking the rank slots (+0xfd0/+0xfc0/+0xfb0 = ranks 0/1/2) with 6 and
  then 0x40 via `poke_vidmem` changed the shadow (verified by dump) but
  did NOT move the client's wait. The vidmem route is not (or not only)
  the completion path; the A8 loop's tables have another source.
* gdb-attach hazard (again): attaches to the client in its post-device
  spin hang with no output; two gdb chains had to be killed by PID
  (never a pattern) and the client SIGCONT'd. Probe only while the
  client is in do_poll.

## ADDENDUM (r687, session 83) — THE A8 COUNTER SOURCE DECODED: a client readback window the host can write (first live read-back confirmed)

* The A8 loop calls `a82f30(obj)` per entry; a82f30 calls the object's
  method pointer at ***(obj+0x19b8)** = eglcore+0xb250f0, which returns
  `(hi<<32)|count`. The count comes from **b1bf00**:
  - fast path (obj+0x224 != 0): `*(u32*)(obj+0x1078) - 1`;
  - slow path (observed: obj+0x224 == 0): **b27c90(obj, 0x800000bb/0xbc)**
    which reads `*(u32*)(*(*(obj+0x1480)+0x58) + idx*16)` - a **client
    READBACK WINDOW** (idx = 0xbb..0xbb+n, n = *(u32*)(obj+0x22efc));
    the loop takes the MIN of those reads.
* Live (r687, object handle 0xbeec0100): the window = VA
  0x7827e6c12000 (a /dev/nvidiactl mapping), all zeros; obj+0x224=0
  (slow path active), obj+0x1078=7 (fast-path field unused).
* **The host CAN write it**: resolving the window's PA via the client's
  pagemap and writing 0x100 into the window (offsets 0xbb0..0x1ff0, via
  the RAM file) made the client read it - **obj+0x1070 = 0x100** (the
  method stores the returned count there) and the A8 "actual" table
  entry became `0xffffffff00000100`. FIRST LIVE PROOF that the client's
  completion counter can be driven from the host.
* The wait still spins: the A8 loop's first scan (compare
  a85e1d: actual vs want, want observed 0) passes with our value, but
  the loop continues into a second scan (compare a85ec1, driven by
  b29dc0) and the `*(rsp+0x38)` branch (a85e60). NEXT: decode those
  (the second scan's actual source and the rsp+0x38 flag), and whether
  the other objects (entries 0-2 of the A8 table) need the same window
  writes.
* The other objects' windows were not observed (only 0xbeec0100 called
  a82f30 in the sampled window); the nvidia0-side window (obj+0x1478)
  is NOT present in the pagemap (VM_IO; host cannot write it via the
  RAM file).

## ADDENDUM (r687, session 84) — the a86038 want-scan mapped: ALL WANTS PASS with the driven window value; the residual gate is the A8 loop's internal callback/retry machinery

* Live dump of the a86038 scan (BP at 0xa8608e, per-index):
    i=0 want=0x0 actual=0x0 pend=no
    i=1 want=0x0 actual=0x0 pend=no
    i=2 want=0x0 actual=0x0 pend=no
    i=3 want=0x6 actual=0xffffffff00000100 pend=no
  (want from a82fd0(job_obj, i) = the job object's +0x98 array maxed
  against global+0x600; actual from r14+0x5e8 = our driven counter).
  NO pending entries -> the scan completes.
* Writing 0x100 to the ENTIRE readback window (offsets 0x0..0x1ff0 of VA
  0x7827e6c12000) did not clear the wait either.
* The loop still cycles: the bt stays at the a85f08-a85f1a yield call
  (b284b0(*(rsp+0x10), 0xb) - a bounded pause/yield spin of ~0x14
  units, hence the slow cadence), and the exit-attempt path a85f30
  decrements *(rbp) (w0=1) while firing the global callbacks
  (0x206f790/0x206f7a0 vtables: *(%rax), *0x10, *0x18) with args from
  rbp+0x30/0x38/0x40. The remaining gate is this callback/retry
  machinery (or another wait object), not the readback values.
* NEXT: resolve the global callback targets (0x206f790/0x206f7a0
  vtables at runtime) and trace the a85f30 exit path (what the
  callbacks do and what makes *(rbp) reach 0); or find how the real GPU
  writes the readback window (the op-4 methods 0x800000bb/0xbc the
  service skips) and emulate that write so the client's own machinery
  completes.

## ADDENDUM (r687, session 85) — **THE NvOgl WAIT CLEARED** (the session-count gate) and the client advanced to vkAllocateMemory (new wall: -2 OUT_OF_DEVICE_MEMORY)

* The A8 loop's exit requires the session count `*(u32*)(r14+0x5bc)`
  (= lib_base + 0x2070b7c, the global at +0x20705c0+0x5bc) to reach 0;
  the loop head increments `*(rbp)` and the entry scans pass with the
  driven window value, but with count=4 the exit checks
  (a85b48/a85e78: `if (r12d == 0)`) are never taken - the loop waits for
  the engine sessions to complete (count -> 0).
* **DIRECT TEST (succeeded)**: `set {unsigned int}0x<base+0x2070b7c> = 0`
  via gdb -> the loop exited, the NvOgl init completed, and the client
  ADVANCED past the post-device wait for the first time:

      VKDEVICE ok family=0
      VKFAIL vkAllocateMemory(dev, &ma, NULL, &img_mem) -> -2 (line 204)

* **NEW WALL**: `vkAllocateMemory` for the image (DEVICE_LOCAL) returns
  -2 (VK_ERROR_OUT_OF_DEVICE_MEMORY) at vk_frame.c:204 - the guest
  driver's heap/type selection fails before any alloc RPC (no
  GSP_RM_ALLOC in dmesg). Next: the memory-heap reporting
  (FB_GET_INFO_V2 0x20801303 is handled; check the heap-size/type
  queries the driver uses for the DEVICE_LOCAL type, and the emulated
  VRAM size the driver sees).
* Recipe for the next run: slotfill + word nudge -> VKDEVICE ok; then
  write 0 to lib_base+0x2070b7c (gdb; the harness) once the NvOgl wait
  starts (the A8 loop's count gate); then the memory-alloc wall.

## ADDENDUM (r689, session 86) — **THE DEVICE_LOCAL -2 WALL CRACKED**: the 512 MiB emulated VRAM was the gate; the "nvidia-nvidia-vgpu-turing-2g" mdev type (2 GiB) lets the allocation proceed

* Evidence chain: the ICD hooks (t91.py) captured the client's allocation
  `ALLOC sType=5 size=0x4000 typeIndex=1` (a valid 16 KB DEVICE_LOCAL
  request); the standalone memory-properties dump (vk_memprops.py)
  showed device-local heap[0] = 512 MiB with a separate 2936 MiB
  non-local heap. With 512 MiB the driver failed the allocation
  (VK_ERROR_OUT_OF_DEVICE_MEMORY, no RM RPC).
* **FIX TESTED**: the cycle with `VGPU_MDEV_TYPE=nvidia-nvidia-vgpu-turing-2g`
  (start_vm_keep.sh honours the env var) -> `heap[0] size=0x80000000
  (2048 MiB)` (vk_memprops.py) -> the SAME client with the SAME recipe
  (slotfill + word nudge + count auto-clear) reaches `VKDEVICE ok` and
  the allocation NO LONGER fails -2: the client proceeds into
  `os_schedule` (a deeper wait).
* REMAINING in the new state: the client's later pushes (va=0x44852d4 /
  0x4ca52d4, then +0x586c) extend beyond the first window page and read
  `pa=0x0`/zeros (the sema injections covered page 1 only; the window's
  page list is scattered - the memdesc carries only the FIRST page).
  NEXT: extend the window injections to the full scattered page list
  (from the memdesc's page array) or accept 1-page pushes; then the
  allocation completes and the client proceeds to the buffer/pipeline.
* Tools: vk_memprops.py (standalone memory-properties dump; safe
  alongside other clients), vk_alloc_test.py (standalone allocation
  reproducer - NOTE: it hits vkCreateDevice -4 without the harness
  (word nudge + count clear), so it cannot replace the vk_frame run),
  t91.py (auto-clear + ICD memprops/alloc hooks).