summaryrefslogtreecommitdiffstats
path: root/kernel
Commit message (Collapse)AuthorAge
* sema: r391f take 2 - once-only GRCE inline writes keyed by (push_va, push ↵Arthur Hanson Rasmusson10 hours
| | | | | | | | | | | | | | | | hash, dst, bytes) The first take gated on !replay, which broke pushes whose first pass was untrusted (their bytes are only authored on the later trusted pass) and hung the second client of the boot. Key the inline write by (push_va, push content hash, dst, bytes): a replay of the same push is a genuine duplicate and is skipped, a new push (its release payload makes the hash unique) always runs. Table is session-scoped (cleared by svc_session_flush), 64 slots. Root cause retained: victory_reduce's cudaMemset(d_sum,0,8) is a GRCE LOAD_INLINE_DATA write; every re-decode of that push re-ran it and re-zeroed the accumulator after the launch's copy-out (the sum=0 class).
* sema: r391f once-only GRCE inline writes - the sum=0 class root causeArthur Hanson Rasmusson10 hours
| | | | | | | | | | | | | | | | | | A REPLAYED ring entry must not re-run its LOAD_INLINE_DATA writes. The fill/copy ops were already gated by allow_ops (cleared for entry_replay), but the inline-write gate was 'trusted || allow_ops' and a replay is trusted by construction, so every re-decode of an already-processed push re-wrote its inline bytes into the vidmem shadow. victory_reduce's cudaMemset(d_sum, 0, 8) is exactly such an inline write. Observed on r396: 'sema-grce: inline write dst=0x7845ef000000 -> pa=0x7c00000 ap=vidmem bytes=8 data8=0000000000000000' fired 86 us before the D2H readback of the same VA - zeroing the accumulator the launch had just written (OUT-PAGE-POST val8=0000008000080000 = the correct sum) so the client read zeros (sum=0, bad=1). The same replay rewrote the page repeatedly (482 inline writes/boot). The r383 replay-capture path only needs to CAPTURE the launch; the inline bytes were authored on the first (non-replay) pass.
* sema/boot: r391e print the shadow base pointer in the copy-out and bigcopy ↵Arthur Hanson Rasmusson10 hours
| | | | | | | | | | witnesses Rule in/out a stale or duplicate vgpu->bars[1] shadow between the launch copy-out (svc_fwd_req.vgpu) and the D2H reader (svc_copy's vgpu): the sum=0 runs show the write landing (OUT-PAGE-POST val8 = the sum) and the reader of the same ledger page reading zeros 8 ms later with no logged writer in between. Diagnostic only.
* boot/gr: r391d post-write witness - read the VA's landing page immediately ↵Arthur Hanson Rasmusson10 hours
| | | | | | | | | | | | | after the copy-out write The sum=0 class: OUT-PAGE-WIT shows the changed page's bytes in the bounce (the correct sum) and the ledger resolution (pa=0x7c00000 ap=FB), the copy-out writes hpa=0x16f07a000, and the D2H read of the same ledger page 8-34 ms later reads zeros. sema-fbhpa proves __va(hpa) == base+pa at submit time. This witness reads the page right after the memcpy through the same path the D2H reader uses, so the next failing run names whether the write target and the reader's page differ (aperture/staleness) or the write vanished.
* sema/boot: r391c OUT-PAGE-WIT resolution witness for the sum=0 readback classArthur Hanson Rasmusson11 hours
| | | | | | | | | | The GR copy-out now prints, per changed page, the landing-table resolution of the VA (the exact source the D2H reader's ledger-first branch uses) and its first 8 bytes beside the write target hpa - a launch/readback disagreement (write landed on another page/aperture than the reader resolves) becomes one line instead of a forensics session. Diagnostic only (bounded by the existing copywit budget).
* sema: r391b fix svc_op_translate polarity in the per-page fill resolverArthur Hanson Rasmusson11 hours
| | | | | | | | | svc_op_translate returns 0 on success; the new svc_fill_dst_page_pa tested it without negation, so every page answerable only by the walk/scan returned -ENOENT and the fill truncated (r391 boot: FILL-TRUNC for the page-table writes at va=0x...a00200 -> the client stalled at init). Also demand the unresolved page through the pin channel before failing (stalled VA row -> the userspace stream pushes it to the guest scanner).
* sema: r391 per-page destination resolution for virtual fills - the r390 ↵Arthur Hanson Rasmusson11 hours
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | guest panic root cause The 383620-byte context-init zero fill (FILL-NOREMAP class) resolved only its FIRST page through the pin (va=0x7c36c7400000 -> pa=0x131b63000, the pin stream's first run = 29 pages) and then wrote the whole 375 KiB contiguously past it. The pin stream shows the VA's pages are independent guest allocations with split runs (0x131b63000 / 0x131a80000 / 0x128340000), so the contiguous write covered unrelated memory - on r390 exactly the guest kernel's page-table pages 0x131ba1000/0x131ba2000/0x131bab000. The direct map grew three 4 KiB holes and the guest died with 'Fatal exception in interrupt' (unmap_page_range / cpuacct_account_field / psi_task_change faults on physmap addresses). svc_copy has always translated virtual operands per 4 KB page ('a linear PA walk past the first page reads the wrong memory'); svc_fill never did. New svc_fill_dst_page_pa() resolves every page of a virtual fill through the same authority order (fresh pin wins and repairs the landing table, then the landing table, then the resolver chain); the SYS write loop clamps each chunk to the page boundary and writes at that page's own landing; a page with no resolution fails the fill loudly (FILL-TRUNC) instead of writing through a contiguous guess. The >= 64 KiB landing-table pre-store now records only the contiguous VID (shadow) case; SYS landings are stored per page as they are written. Also extends the r384 page-table-shape refusal to every page of a virtual zero fill (the r390 fill's first page was the client's buffer, so the first-page-only check could not see the PTE pages it later covered). Evidence: /tmp/opencode/r390-live.log (FILL-EXEC + sema-opadd lines), /tmp/opencode/r390-guestram.raw (three zeroed PTE pages, CR3 walk), pin log 12-pin_gr_stream-r390.log line 789 (the split runs).
* sema: r384b narrow the fill pin-demand hold to SYS (guest RAM) resolutions + ↵Arthur Hanson Rasmusson14 hours
| | | | | | | | | | | | | | | FILL-EXEC witness - The r384 boot held a large fill whose walk resolution was VID (the emulated vidmem shadow) - the hold was pointless overhead there and the re-decode re-resolves (the resolver is non-deterministic), so the hold now applies only when the fill would land in guest RAM (SYS aperture), the class that zeroed the guest's page-table pages. Shadow resolutions stay immediate. - FILL-EXEC witness (bounded 64/boot): va, pa, aperture, provenance and pattern of every >= 4 KiB virtual fill execution - the decisive line for the boot-corruption forensics; the large zero fills were invisible until now.
* sema: r384 fill-destination guards - pin-demand hold for large unproven ↵Arthur Hanson Rasmusson15 hours
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | fills + page-table-shape refusal for zero fills (the boot-corruption lottery root-caused to zero writes landing on guest page-table pages) Forensics from the r383 boot (guest panic at t=216 during victory_strided): the guest kernel's direct map had TWO 2 MiB holes (phys 0x105000000 and 0x131800000) because their PTE pages at gpa 0x12d3c2000/0x3000 - two adjacent 4 KiB page-table pages - were found all-zero. Kernel objects (the kprobe nodes UVM registers at module load) lived in those regions; the first access after the zeroing faulted (Oops #1 get_kprobe from uvm_init_entry), cascading to the t=230 panic. The zeroing is an 8 KiB all-zero write, the exact shape of the guest's window page-table clear (LINE_LENGTH=8192, value=0), and the emulation's virtual fill resolver is demonstrably non-deterministic: four executions of that same 8 KiB fill resolved to four different guest pages (0x126af8000, 0x12d8e4000, 0x12f20c000, 0x1187fb000 - every one refused by the below-floor guard). The below-floor guard covers window VAs only; a >= 1 TiB VA fill that resolves through the walk/scan had no guard at all. Two nets in svc_fill, both default-on (sema_fill_guard=1) and both fail-open/histed: - net 1: a >= 64 KiB virtual fill whose only authority is the walk/scan (no pin, no landing-table entry) is held through a bounded 800 ms window (sema_fill_guard_ms) while its VA is demanded through the pin channel (the userspace pin stream injects the guest-pagemap truth), so the retried decode resolves through a real pin. Mirrors the m169/r378 copy guards; releases with the historical behaviour after the bound. - net 2: a ZERO fill about to land in guest RAM on a page whose first 64 qwords are all present+NX+in-range PTEs is REFUSED loudly (FILL-PT-REFUSED) - a page table page is never a legitimate zero-fill destination. Pin-backed resolutions are exempt (guest pagemap truth). New witnesses: FILL-HOLD (held fills) and FILL-PT-REFUSED (refused page-table targets), both bounded.
* sema/boot: r383 kill the blanket bounce copy-out (alias+staleness write ↵Arthur Hanson Rasmusson15 hours
| | | | | | | | | surface) + write guards; REPLAY-CAPTURE closes the silent launch-drop hole; CE operand misses feed the pin-demand channel - bounce copy-out now records staged_hpa[bi] per launch and skips: (a) recs whose hpa != staged slot content (alias/leftover), (b) pages the engine did not change (~955 of 968 per boot launch); the remaining changed pages re-validate via gsp_sema_gr_write_guard before the 4 KiB memcpy. Removes the blanket 968-page write-back that was the remaining boot-corruption lottery vector (guest slab freelist clobbered by a foreign page write). - new bounded WGUARD-SKIP guards: gr_write_guard (pin-first re-resolution) on copy-out, release-word zero/restore (rel_va plumbed through the submit), stale-recover-zero, counter release; svc_write_guard pin-agreement on 4/8-byte SYS/FB service writes; svc_copy SYS-dst hold on a fresh-pin contradiction. - svc_qmd_accum replay hole: a trusted replay of an unseen push now captures (REPLAY-CAPTURE) instead of silently dropping; known push -> LAUNCH-SKIP witness; capture history ring prevents double-capture. - svc_op_translate miss path emits the stalled-VA demand signal for CE operands.
* sema: r377 pin-demand for unresolved launch operand pages - every dropped ↵Arthur Hanson Rasmusson17 hours
| | | | data-fill page is logged as a stalled VA (the userspace pin stream watches dmesg for that line and pushes the VA to the in-guest scanner's demand file), so the bounded launch hold can resume on real pins instead of waiting out its bound (the r345 wedge)
* sema: r376 pin-agreement on the copy ledger-first branches - a fresh ↵Arthur Hanson Rasmusson17 hours
| | | | disagreeing pin beats a recorded landing on both dst (write) and src (read) at top level and per page, repairing the entry; closes the stale-entry write (guest-corruption) and read (wrong-result) classes
* sema: r375b drop the resolver-answer ledger store - a stale resolver page ↵Arthur Hanson Rasmusson17 hours
| | | | recorded as the write authority poisons every later write of the VA (r375 boot slab GPF in kmem_cache_alloc at guest t=77s before any client); the table is fed by writes and fresh pins only
* sema: r375 read-after-write authority - the landing table wins for device-VA ↵Arthur Hanson Rasmusson17 hours
| | | | copy sources (top-level and per page), and a resolver answer is recorded so launch/read paths converge; the engine output can no longer be read through a different page/aperture than it was written
* sema: two-path identity check for the FB shadow hpa (__va(hpa) vs base+pa) - ↵Arthur Hanson Rasmusson17 hours
| | | | the sum=0 class shows the engine output written through one path and read through the other
* sema: page-note resolution diagnostic (va/pa/ap/branch) for the launch-page ↵Arthur Hanson Rasmusson17 hours
| | | | vs readback disagreement (sum=0 class); budget 512
* sema: pin-wins wave - a fresh pin re-homes the landing table ↵Arthur Hanson Rasmusson18 hours
| | | | | | | | | | | | | | | | (svc_handle_map_add) and wins over the ledger in svc_gr_page_note, svc_gr_data_fill and svc_fill; gate the displaced-pin fan-out on a current pin match Root cause class (sub-agent audit + r371 log forensics): the landing table (ledger) moved only on host writes, so a VA re-pinned to a NEW page (CUDA reuses VA ranges across processes) kept resolving to the OLD, freed page. Every later write through that entry landed on memory the guest kernel had since reused - the r371 mail client's guest oops (NULL page array in nv_free_system_pages, 8 zero bytes) is this class. Fixes: (1) svc_handle_ map_add repairs a disagreeing ledger entry from the fresh pin (PIN-REPAIR); (2) svc_gr_page_note and svc_gr_data_fill keep a live pin over the ledger (PIN-WINS); (3) svc_fill checks the pin before the ledger fallback; (4) the pool alt fan-out writes the displaced page only while the current pin still names it (alt_skip). New diag: pinrep= alt_skip=.
* sema: store the landing table for every launch page in svc_gr_page_note - ↵Arthur Hanson Rasmusson19 hours
| | | | indirect/bank operands (the kernel output buffer) resolved there while the client D2H readback used an older ledger landing, read stale zeros and the kernel looked wrong (victory_reduce sum=0 with the correct sum proven in the bounce page)
* sema: refuse below-floor fills that resolve outside vidmem (LOWFILL-REFUSED) ↵Arthur Hanson Rasmusson19 hours
| | | | - the 0x204020000 engine-window clear resolved to a different guest-RAM page on every execution and zeroed 8 KB of guest kernel memory (gmain NULL deref in __fdget, task_struct corruption)
* sema: opdef and GR hold bounds are never re-armed while active - the ↵Arthur Hanson Rasmusson19 hours
| | | | arm-on-VA-change rule moved the deadline on every held entry (streams2 wedge: opdef=955/13/0, drain held climbing, expired 0); entries held while a bound is active ride it and flush together
* sema: TABLE-COPY witness - dump carried bytes of page-table page/entry ↵Arthur Hanson Rasmusson20 hours
| | | | copies (bounded) so zero-walking tables trace to their building copy or prove the source was zero
* sema: CE fill byte-exact semantics per the driver's own memset push - ↵Arthur Hanson Rasmusson20 hours
| | | | LINE_LENGTH_IN is the size in BYTES (count=line_length/elem; the old count=line_length wrote elem x too much), value from DST_X mapping (CONST_A/CONST_B), r367
* sema: keep count=1 remapconst parts as flags only (NOT have_remap) - r365 ↵Arthur Hanson Rasmusson20 hours
| | | | proved below-floor window fills executing via unproven resolution zeroed guest RAM (slab GPF -> guest panic); add LOWFILL-WIT resolution witness for below-floor fills
* sema: decode count=1 SET_REMAP_CONST A/B/COMPONENTS (the CE page-table fill ↵Arthur Hanson Rasmusson20 hours
| | | | constants) + opt-in vidmem-window fill execution (gsp_sema_winfill, off) + FILL-RAW raw-push diagnostic; remapconst log budget 256
* sema: opdef absolute per-launch hold bound (GRFIX lesson) + physical ↵Arthur Hanson Rasmusson21 hours
| | | | | | | | | | | | | | | | guest-RAM fill branch + diagnostics The opdef hold used a slot-keyed deadline that a pin-batch rewind re-arms on every replay (observed opdef=3301/32/0 with the guest wedged); the bound is now keyed to the launch push VA and written back through the pump snapshot, like the GR hold bound (r213). The fill op builder also accepts physical SYS destinations below the VA floor (guest-RAM page-table clears; bounded by SEMA_PA_MAX). The vidmem-window zero-fill experiment (r362) was reverted: executing those fills with elem=1 wrote 4-byte partial zeros and broke results (no drops but sum=0); the window fills need their true element size and value semantics decoded first (see handoff).
* sema: revert the capture-site launch hold (opdef holds accumulated without ↵Arthur Hanson Rasmusson22 hours
| | | | expiry and wedged client init; keep the content-hash witness and the bound-expiry retry-forward)
* gr/sema: content-hash witness for the once-only op gate (rewritten fills now ↵Arthur Hanson Rasmusson22 hours
| | | | run - stale GPU page tables class); launch hold without capture trust + bound-expiry retry-forward (no silent launch drop); guarded zero-fill for client device VAs; walk/rehold diagnostics
* gr: record the launch's resolved operand pages in the ledger - fixes the D2H ↵Arthur Hanson Rasmusson23 hours
| | | | readback reading a stale page (victory_reduce sum=0 class); add GMMU walk trace + fwd-hpa diagnostics
* gr: single scheduling edge (pcas default 0) - fixes double engine execution; ↵Arthur Hanson Rasmusson24 hours
| | | | retire-wait tracked-total fix; sema acquire wait gate per-entry + C56F op semantics (opt-in); fill-shape decode diagnostic
* vgsp: C56F acquire wait gate as an opt-in experiment (default off)Arthur Hanson Rasmusson28 hours
| | | | | | | | | | | | | | | | | | | | | | The streams2 readback-before-scale class is root-caused: core CUDA starts compute-channel pushes with a C56F SEM_EXECUTE acquire (ACQ_CIRC_GEQ on a tracking semaphore; e.g. the D2H readback push waits on 0x20400fef0 >= its sequence number) and the decoder ignored it, so the copy ran ahead of the scale launch it waits for. The gate implemented here (decode the acquire; svc_run_ops holds the push's work until the semaphore satisfies, with the driver's release(N)+acquire(N) same-push pattern treated as satisfied; op-pending bounded hold; kill switch at 2048 holds; diagnostics sema-wait held rows with a per-session budget) OVER-HOLDS on boots where a channel's first pushes wait on counter values only their own withheld release would install (wait counts reach thousands, the client stalls; r337/r338/r339). The tracking-semaphore value semantics (what the guest word should read vs the driver's sequence numbers) need a dedicated pass before this can be enabled. Default OFF so behaviour is the pre-r336 one; enable per-boot with sema_wait_gate=1 for the experiment. Everything else (pin-wins, vector_add, the battery) is unaffected.
* vgsp: honor the C56F host-semaphore acquire (fix the streams2 ↵Arthur Hanson Rasmusson28 hours
| | | | | | | | | | | | | | | | | | | | | | readback-before-scale class) Core CUDA starts each compute-channel push with a C56F SEM_EXECUTE acquire method (five words [ADDR_LO][ADDR_HI][PAYLOAD_LO][PAYLOAD_HI] [EXECUTE]; op 0/2/3/4/5 = acquire/equal, strict GEQ, circular GEQ, AND, NOR). The decoder handled only op 1 (release) and 6 (reduction), so every wait was ignored and a push's work could run ahead of the dependency it waits on. Observed on streams2: the D2H readback push carries ACQ_CIRC_GEQ on the channel progress counter (0x20400ffd0 >= 4, the 4th push = the scale launch); the readback executed before the scale, and the client verified the pre-scale buffer (bad=32767, c2 = a2+b2 without the x2). Now the decoder records the acquire and svc_run_ops honors it before running any of the push's work: satisfied (read + circular/equal/GEQ compare) -> proceed; unsatisfied -> set op_tr_pending, so the entry is held and its release withheld (bounded by SVC_OPDEF_HOLD_MS, loud) and the re-decode runs the work once the counter advances. Unreadable semaphore: fail-open (pre-r336 behaviour). New stat wait=held/pass.
* vgsp: pin-backed translation wins the write-side ledger divergence (r336)Arthur Hanson Rasmusson29 hours
| | | | | | | | | | | | | | | | | | | The m180 unification redirected a divergent write to the ledger's page (where an earlier write landed) even when the current translation was pin-backed (guest pagemap truth, op_tr_src == 1). When the guest remaps/reuses a VA (free + reallocate, or migration), the ledger page is freed and can be reallocated by the guest kernel: the redirected write then smears client copy data over kernel memory. Observed on the r335 vector_add run: an RM map/AVL node's pointer field held client c-data (floats 551.25/552.0), the free-path traversal faulted (mapFindLEQ_IMPL/_mapInsertBase), and the guest panicked (Attempted to kill init). 4448 write-side divergences in that run. Now a pin-backed translation wins: the write goes to the fresh page and the ledger store after the write re-homes the entry, so readers converge on the same page. Non-pin translations (2..5) keep the m180 redirect - they can carry stale PTEs (m169) and the write history stays the authority there. New counter sema-svc ledger pinwins.
* vgsp: read the status-queue consumption cursor from the command queue (SWAP_RX)Arthur Hanson Rasmusson33 hours
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | Under MSGQ_FLAGS_SWAP_RX (both sides set it; flags=1 in both headers) the guest driver's msgqRxMarkConsumed writes its STATUS readPtr to pOurRxHdr->readPtr, and the driver's pOurRxHdr is the COMMAND-queue buffer's RX header (msgq.c msgqRxLink swap branch: pReadOutgoing = &pOurRxHdr->readPtr with pOurRxHdr from msgqTxCreate's pBackingStore = pCommandQueue). The status queue's own +rxHdrOff is the mirror we publish - our command-queue read position. The status free-space check read the mirror, so it tracked our own writes: with the guest mid-drain the ring was declared full (free=0, peer=0, write=62/63), replies were written into it anyway, unread records were overwritten, and the guest RM read a newer record - GspMsgQueueReceiveStatus: Bad sequence number. Expected 447 got 510. Possible memory corruption. - after which every RPC failed (status 0x25) and the first client's cudaMalloc returned device busy or unavailable. That is the boot lottery: on a lost boot the first client (mail) FAILs and the boot is poisoned for every later client. Evidence (r330 vs r332): r330 mail FAILed at the first cudaMalloc and only one channel unregistered in the whole boot (the guest RM never tore its channels down after the status desync); with the fix r332 mail + matmul pass and every client's channels unregister (79 unregisters, one session flush per client). Also adds bounded diagnostics: sema-d2hmap (per-page destination resolution of >=1MB host-dest copies) and a raised DIVERGE witness budget.
* vgsp: fail4 - route pending-pass releases through opdef withholding ↵Arthur Hanson Rasmusson38 hours
| | | | | | | | | | | | | | | | | | (cold-run gap) The fwd-defer arm took priority over svc_opdef_release, so a pass whose op was still pending recorded its fence release in svc_fwd_rel (applied at the hold break / fallback) instead of the opdef slot - the fence fired while the copy was still owed, the driver refilled the ping-pong staging, and the late capture/read got refilled bytes (cold-run vector_add gap: one to four chunks per cold boot). svc_fwd_defer_arm() is now skipped when svc.op_tr_pending, so those releases go through svc_opdef_release and apply only when the retried pass has no pending op (the op has run by then). Measured cold-run vector_add bad across this iteration: 193536 -> 100352 (r323) -> 479232 (r324 wrong queue-pending experiment, reverted) -> 66560 (r325, ~1 chunk). All other acceptance items green on r325 (mail OK, smi OK, matmul OK, transformer OK).
* vgsp: fail4 - carry the channel into pending releases; hold the fence on ↵Arthur Hanson Rasmusson39 hours
| | | | | | | | | | | | | | | | | | | | | | | | | | | | denied passes Two fixes for the cold-boot classes (CONT.60): 1) Pending releases starved forever: svc_apply_pending resolved the pool semaphore with h=0, but the verified pool resolver is channel-keyed - it never matched the per-channel pool entries while the very same lookup at decode time succeeded (matmul r322: 'pool-rel cand ... rrc=0 cur=0' then 'pending release starved va=0x121000010 payload=1 tries=4/64...' forever, the client hung and the boot died). Pending entries now record the channel they were decoded for (struct svc_pending_release.h, set from svc.cur_ch which svc_decode_stream stamps at its top) and svc_apply_pending resolves with it. Also applies to the 'completion hint rc=-2' class. 2) The fence escaped on denied passes: in the P6b wave-2 overflow branch the release was queued only when n_ops>0; a denied pass (n_ops==0) but op-pending (a copy still owed) applied the release inline, letting the driver refill the ping-pong staging before the copy ran - the cold-run capture gap (vector_add run 1 bad=157696/193536). The queue condition is now n_ops>0 || svc.op_tr_pending. Measured (r323 fresh boot, full sequence): mail OK, nvidia-smi OK, matmul OK (no death - the r322 killer), transformer OK (no stall), vector_add cold-run bad=100352 (two chunks, down from 157696-193536; warm runs remain bad=0).
* vgsp: sync gsp_sema_service.c to fail4 after cherry-pick conflict ↵Arthur Hanson Rasmusson39 hours
| | | | (wd_rewind_off_until struct member repair)
* vgsp: fail4 - LAND ALL H2D CHUNKS: consume the staging snapshot; capture on ↵Arthur Hanson Rasmusson39 hours
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | first decode vector_add was deterministic-wrong (bad=786335 checksum=359549561 identical across boots) because the 1MB H2D chunk copies always read the driver's ping-pong staging AFTER it had been refilled (the fence release applied long before the late op execution; 60+ rc=-2 retry storms per chunk). Two wiring fixes (r314-r315): 1) The decode-time snapshot was NEVER CONSUMED: svc_copy looked it up (svc_copy_snap_lookup at the top) and freed it on success, but no code read from it - every copy fell through to the live per-page source resolution and read refilled staging. The source read loop now takes the bytes from the snapshot when present (memcpy from snap + o). The copy also no longer needs a live source translation, so it executes on its first allowed pass instead of storming retries. 2) Capture on the DENIED pass too: the old capture gate lived in the op-build block (allowed passes only - by then the fence fired and staging was refilled). The denied device-dst copy branch now captures immediately (pipeline-order moment), same translation-probe guard. Measured r315 (fresh boot): vector_add first run bad=157696 (cold-run capture coverage gap), second run bad=0 checksum=392344090.5 STATUS OK with ALL EIGHT chunks clean and no retry storms (~75 s, guest survives). Previously 786335 bad / identical checksum every boot.
* vgsp: fail4 - cure the stale-wedge republish loop; enable staging snapshots ↵Arthur Hanson Rasmusson39 hours
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | for the sw-copy client Two fixes, both demanded by live evidence: 1) Stale-wedge republish loop (r311-r312): the pump's per-channel write-back did not persist gr_pub_armed, so the residual watchdog re-armed every pass; its recovery kept zeroing the completion word and re-publishing (pub-adv/ zero/stale-wedge 3s cycles, exit: pub-adv pa=... val=104 cur=0 last=0). The word oscillated under a live client that then never completed (r311 vector_add wedged; guest sshd died). Persist gr_pub_armed/counter_gen in the write-back and keep them (plus slot_pend_tries) across same-generation re-registration. Also guard the residue zero: never zero a completion word authored by the live session generation (r312 instrumentation: pub-adv-wrc + stale-recover- zero + skip witness). Verified r313: stale-wedge rows 0 (was 66+), client completes, guest survives. 2) Staging snapshot capture for the sw-copy client (r311): the decode-time H2D capture gate required src_aperture == CE_APERTURE_SYS, but the driver sends no SET_SRC_PHYS_MODE for the 1MB chunk copies, so src_ap defaulted to VID and the gate never matched - zero captures on every run r302-r310 while executions read the (driver-refilled) staging late. Gate judge is now the translation: capture when the first source page resolves through SYS. First verified captures (r311: 2-3 per run, real client bytes in first4) - partial coverage only; the remaining pipeline-order gap is the open lead (see handoff CONT.57). Diagnostics kept: snapgate witness, pub-adv-wrc, stale-recover-zero[-skip].
* vgsp: fail4 - denied device-dst copies hold+retry (fix the chunk drop)Arthur Hanson Rasmusson39 hours
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Root cause of the vector_add chunk drop (CONT.55 chase): the sw-copy client's GR-classified GRCE channel (eng=0, h=0x5c000059) decodes its 1MB H2D chunk copies as CE transfers, but ops on that channel run only once the pin/trust stream (injected) reaches each push page. The first pass over a chunk push lands before that; with allow_ops=0 the copy was never built and the entry was consumed silently. All later passes classify as replays (slot_ops_complete=1), so the entry was never retried: 6 of 8 chunks skipped, kernel computed on stale VRAM (vector_add r302-r307: 2-5 of 8 chunk ops ever existed, the rest never ran; the +2MB 'phase' signature is stale residue). Fix: a denied launch that WOULD have built a type=1 device-dst copy (no remap, both offsets, out >= SVC_VAS_FLOOR) with a trusted push now marks the pass op-pending. op_tr_pending un-completes the slot, so the next pass is treated fresh and the copy runs as soon as trust (injected with allow=1) or the operand translation arrives - the existing m163/opdef hold+retry machinery. Replay passes are deliberately eligible (that is what makes the retry possible); a per-slot slot_pend_tries cap (4 per put epoch) prevents a slot whose trust never arrives from wedging the channel in 10s hold cycles. Measured (r308, sema_dbg): all 8 consecutive chunk dsts build ops (0x...cc..d3) and retry to success - 28 CEAUDIT rc=0 soft copies, 38 rc=0 vs 20 rc=-2 executions; previously max 5 dsts and only the last 2-3 ran. Regression checks on r309 with the fix: probe RAWA2+PROBE2B bad=0, victory_mail OK, victory_matmul OK (one transient first-run fail, clean re-run), transformer OK. vector_add now advances through the full H2D set but still does not complete (retry storms stretch the copy phase; the run later wedges in the GR stale-wedge recovery loop - named as the next lead). Diagnostics kept (dbg-gated, bounded): sema-pend set/hold witness, sema-consume consumed-slot witness, sema-entry de-gated for eng=0 channels, dmatrace extended with replay/eng/pend_would, budgets raised under sema_dbg so boot init cannot spend the trace before the run.
* vgsp: fail4 r297-r300 - decode-time staging snapshot (H2D shapes only) + ↵Arthur Hanson Rasmusson39 hours
| | | | | | | | | | | | | | | | | | | | | ops-order release guard 1. CONT.53 step (1): device-dst copies from host sources now snapshot the source at op-build (bounded pool 16 x 4 MB, 30 s TTL) and execute from the snapshot, so the staging read happens at the push's processing position. 2. The capture gate is aperture-based (dst VID, src SYS) after the r300 captures showed a naive VA-floor gate also caught D2H readbacks and froze pre-kernel zeros. 3. Ops-order release guard: a pass that built ops queues its releases through the pending pass instead of applying inline, so a fence never clears before the pass's copy ops read their source. A/B (r297/r299/r300): snapshots fire (6-7 per run) but the +2 MB phase persists (win@0/1M = exp[+2M]); the r300 capture contents show the H2D op sources already carry late-chunk data at decode - the scrambling is at the driver's staging hand-off into the push stream, before our service sees the op (next instrument: correlate the client's PTRS with per-capture buffer identity).
* vgsp: fail4 - pipeline-order source snapshot for held staged copiesArthur Hanson Rasmusson39 hours
| | | | | | | | | | | | | | | | | | | | The m169 device-dst guard defers a staged H2D up to 3 s; the driver's ping-pong staging buffers are overwritten meanwhile, so a deferred read returned a later chunk's content (the +2 MB vector_add phase = the 2x1 MB staging stride). Held copies now snapshot their source at hold time (bounded: 8 slots x 4 MB, 30 s TTL) and the retried execution copies from the snapshot, matching the hardware's in-order staging read. r296 A/B: the snapshot never fires for vector_add - its H2D dsts resolve with pins promptly and execute without the hold; the +2 MB persists identically. So the staging overrun happens WITHOUT the guard: the driver's 2-buffer staging discipline relies on the copy's completion fence (a semaphore applied when the copy executes) to pace the restaging; if our completion for chunk i arrives before/independent of the staging read, the driver restages and our later decode reads chunk i+2. Next lead: the staging fence's increment timing vs the copy's execution (the release-before-work class extended to the H2D staging fences), not the m169 hold.
* vgsp: fail4 - definitive rec-content witness (data[0..2] first8 at the rec hpa)Arthur Hanson Rasmusson39 hours
| | | | | | | | | | | | r295 measurement: rec[0] va=0x72a93cc00000 hpa=0x119155000 first8=43100000 43108000 = floats 144.0 144.5 = a[524288..524289] - i.e. the first A-rec's backing page already holds the +2 MB data, before copy-in. Chained with the r285 upload witnesses (the H2D op sources read the other buffer's chunk-2/3 data), the +2 MB phase originates on the HOST-BUFFER side of the uploads (the pinned/staged source resolution for the client's host buffers), not in the GMMU walk (verified correct against the driver's own format: PD3[48:47]/PD2[46:38]/ PD1[37:29]/PD0[28:21]/PT[20:12]).
* vgsp: fail4 - device-survival chain for the large-buffer sequence + CE ↵Arthur Hanson Rasmusson42 hours
| | | | | | | | | | | | | | | | | | | | | | | | | | | | both-side op advance Survival (the death cascade is gone for the full user sequence): 1. pool residue re-seed applies ABSOLUTE (cur=0) - the UVM assert (uvm_channel.c:204) that took the device down no longer fires (committed earlier in 67148bd8). 2. completion-hint window 10s->1.5s (67148bd8). 3. status-ring free-space discipline in gsp_rm_send_flags: the guest's consumption cursor (msgqRxHeader.readPtr at stat+rxHdrOff, the msgqTxGetFreeSpace source) gates EVENT posts; events drop freely when short, replies wait (bounded) and proceed. A guard rejects a cursor the guest cannot be behind (peer > our writePtr - the observed mirror tracks writePtr+1 and gating on it silenced every wake-up and wedged the clients). Verified live: mail OK, nvidia-smi OK, victory_matmul OK, vector_add completes (content still wrong) and victory_transformer STATUS OK AFTER it - no ISR panic, no guest death, guest ssh alive. Content: CE linear auto-advance now applied to BOTH offset registers after each executed op (the driver re-sets one side per op and the engine advances the other; H2D ops showed an explicit advancing dst with a fixed src base: the device a held h_a[0..1MB) replicated). vector_add content still FAIL (~66% bad) - the H2D operand stream under retry/replay passes is the next lead (opadd rows show five dst x two src combos with heavy retry multiplicity).
* vgsp: fail4 r259-r262 - capture ring (4 slots) + diagnostic budgets; CE ↵Arthur Hanson Rasmusson46 hours
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | auto-advance tried and reverted 1. svc_gr_cand is now a 4-slot ring addressed by an accessor macro (svc_gr_cand -> ring[pos]); captures that would overwrite a DIFFERENT launch's unconsumed candidate move to a free slot, and svc_gr_request selects this channel's oldest unconsumed candidate before consuming; the defer-arm gate checks any unconsumed candidate. Struct tag renamed svc_gr_cand_rec to keep the macro safe. (The streams2 cross survives this - it is not (only) the single slot; see the new evidence in the handoff.) 2. Diagnostic budgets raised for the triage (opadd/op_trace 2048, memsets rc-log 2048). These exposed the real memset shape: the 8 MB cudaMemset is SIXTEEN grid=1024/block=512 kernel launches whose win0 param bank is fully correct (dst patched to the reloc, w0x168=0x5a5a5a5a, launches OK) - yet the D2H readback returns zeros, so the loss is between the engine's stores and the readback pages (the rec hpa/landing class, GRVAUDIT unlanded mass), not the params. 3. A CE-style per-op dst auto-advance was implemented, evaluated (no memset improvement) and reverted - re-sets are decoded and the risk to the green clients was not justified. Verified on r262 (boot 1 of the acceptance round): ctaid_probe RAWA 0 1 2 3 4 5 6 7 + PROBE2 bad=0; transformer STATUS OK; streams2 bad=65536 (red); memcpy_roundtrip wedged the guest (flaky: green on r254, wedges on r260/r262).
* vgsp: fail4 r254-r257 - small-fill hold + copy landing ledger + per-launch ↵Arthur Hanson Rasmusson48 hours
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | reloc half Three fixes on top of 54cf842c, targeting the big-buffer clients: 1. m163 parity for fills: a small failing FILL op with a virtual device destination is now held/retried like the small copies (-ENOENT with the pins still landing was silently counted nomap and the release applied - the 8 MB cudaMemset landed only partially). Verified: memcpy_roundtrip now reaches STATUS OK (RT1 bad=0 csum exact, MSET bad=0, h_m all 0x5A). 2. svc_copy landing ledger parity with svc_fill: a large virtual-dst copy records its landing pages so the GR rec build (m171 ledger preference) follows the page the copy actually wrote. streams2's c1 moved from a stale-page mix ([28.0, 28.625, 29.25]) to the correct-page values ([0.0, 0.125, 0.25]). 3. Per-launch relocated-data base: concurrent forwarded launches (streams2's vadd pair) shared one base and their bounce windows overlapped; the base now alternates across the grown region's two halves keyed to the capture push (retries keep their half), cap halved accordingly. Also raise the per-page data attribution budget (16 -> 192 pages) for span forensics. Remaining on streams2/victory_multikernel: the single-slot svc_gr_cand capture is overwritten by the second concurrent launch before the work item consumes it - ctaid-probe evidence: c1 == a2+b2 exactly (the stream-2 result in the stream-1 output) with the in-tree candidate-overwrite witness present. Next: make the capture a small FIFO drained in order by the gsp-phys-work item.
* vgsp: fail4 r249-r252 - GRCE slot-run rule + arg-rec order fix; probe fully ↵Arthur Hanson Rasmusson2 days
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | green Two fixes on top of c66a0172: 1. Slot rule for LOAD_INLINE_DATA runs: in the UVM upload pushes each run fills a whole 32 KB slot - its data runs to the next run header (or to the next 0x8000 boundary for the last run), regardless of the count field (observed slot starts every 0x8000: cnt=4095 and cnt=2497/2753 variants). Compact pushes keep the count*4 reading. The old reading truncated each slot run to half its data and walked the tail as methods. 2. Argument-ring order: the bridge's arg-block search must see the run's START as the inline-write destination; the new cursor advance must not leak into rec->dst_va (an end-of-run dst_va made the QMD param patch miss and the engine never consumed the launch - regression found and bisected against the pre-session ko via victory_mail). Verified one boot (r251/r252, standard params): ctaid_probe RAWA 0 1 2 3 4 5 6 7 + last8 16376..16383, PROBE1/PROBE2/PROBE2B bad=0; transformer STATUS OK; victory_mail STATUS OK. Residual (distinct class, NOT the H2D delivery): streams2 bad=65534, memcpy_roundtrip bad=8388608, victory_multikernel bad=2883066. Their big spans collapse to a single rec/bank entry in the launch tables ('bank 0 hit @0x168: VA -> reloc' for one page of the span) and the engine write-back replicates one page across the span (OUT-ALIAS-WIT partners) - the original span-aliasing class on the large-buffer paths.
* vgsp: fail4 r242-r248 - GRCE LOAD_INLINE_DATA cursor advance + H2D decode ↵Arthur Hanson Rasmusson2 days
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | instruments The probe/mallocecho H2D (cudaMemcpy H2D of a 64 KB device buffer) arrives as LOAD_INLINE_DATA method runs (op=3, subch=1, addr 0x1b4) carrying the payload inline in the push, split across several runs (and, depending on pool state, across several pushes of one sequence). The inline-write handler wrote every run at st.offset_out without advancing, so a multi-run upload kept only the last run's words and the readback page held deterministic junk (RAWA 16382 16383 8193...). Fix: a LOAD_INLINE_DATA run advances the destination cursor by the bytes it wrote (hardware FIFO semantics); consecutive runs of one upload now tile from the first destination. Verified (fresh boot, standard params): mallocecho d-afterH2D head 0 1 2 3 4 5 6 7 (was junk); ctaid_probe RAWA first8 0 1 2 3 4 5 6 7 (was 16382 16383 8193 8194...). Residual: the upload spans multiple runs/pushes and only the first 4095-word run lands today (bad=12289, tail zeros) - the remaining runs are still lost to the decoder's run-boundary handling (see sema-next-hdr captures). Instruments added (all gsp_sema_dbg-gated, bounded, session-reset): sema-pushall (de-windowed push decode trace with page_pa), sema-opadd, sema-copyent, sema-dmatrace (denied transfers), sema-bigrun (count>=1024 runs on len>=8192 pushes with cursor state), sema-lid-hdr, sema-skiprun, sema-mystery, sema-next-hdr (header after each big LID run). LDA/xfer-raw/stream dumps de-gated from ch->h_object!=0 so unidentified channels are visible.
* vgsp: fail4 r239 - widen per-recp content witnesses to n_data<=1024 (probe ↵Arthur Hanson Rasmusson2 days
| | | | | | | | | | | | | | | | | | | | | | class is ~950 recs) The probe's launches carry ~950-970 recs (its 2 MB VA arena is one span), so the <=64-rec gates on A-TRACE pt=P3 (engine-read source content) and READ-DIVERGE-WIT (fill vs copy-path vs ledger per rec) never fired for the probe - exactly the launches they were built for. Raise both gates to 1024; budgets unchanged (128 lines / first 8 recs), so the transformer's 1418-rec launches stay out. r239 findings this rides on: the probe's H2D (h_A->d_A) leaves ZERO host trace (no copy, no launch, no audit in a continuous 6.6k-line dmesg stream) and A's vidmem reads as deterministic junk from the very first post-upload read; the driver's 512-thread memset kernel zero-writes with fe_esr=0x80000000 and the 3x fresh-attempt retry path (fresh qmdwin bank + win0 bank + rebuilt PB) does NOT recover it. A minimal new repro (blk512) shows the band bug without any probe machinery: grid(8,1)x512 writes 4/4 buffers perfectly while grid(16,4)x(16,16) 'out[i]=i' is deterministically bad=12288 (rows >= 16), with the engine-side scan reporting 31/31 changed pages - the by>=1 stores exist somewhere the client does not read.
* vgsp: newbind holds are residual-only and short (r237 T2 3 s GP_GET delay)Arthur Hanson Rasmusson2 days
| | | | | | | | | | | | | | | | | | r237-2: T2's GR channel carried T1's residual put=115; my 3 s empty-decode hold delayed the host's GP_GET during the residual walk, and the guest unwound at binding+3.46 s (zeroed PROMOTE_CTX, then full FREE) - exactly when the host finally consumed the residual. The guest's ctx-init patience is ~3 s, so the hold must never delay the walk that long. - New SVC_GR_NEWBIND_HOLD_MS 1200u and per-channel wd_newbind_hold_until: the hold windows (empty-decode and unmapped entry) now apply only while it is fresh (<1.2 s; the rebuild lands in ~0.8 s) and only when put == get (residual context, no producer work pending). Any pending producer work is walked without delay, so GP_GET always advances promptly. - wd_newbind_until (3 s) still governs retry eligibility; the opdef drop, chanswitch log, pending-map/rebuild path are unchanged. make kernel green.
* vgsp: fail4 r235 - A-TRACE dir tag + 64KB copies exempt from bigcopy budgetArthur Hanson Rasmusson3 days
| | | | | | | | | | | | | | | | | | | | | | r235: only 4 of the probe's ~9 64KB copies reached the ring (the sema_bigcopy_log budget had absorbed the rest - the A H2D is missing entirely), and the events lack direction/buffer identity. Now: - every total==65536 copy bypasses the bigcopy budget (its own 128-line A-TRACE budget still bounds volume) so the probe's H2D and all D2H readbacks always appear; - A-TRACE carries dir=H2D/D2H (device=VID low-pa, host=SYS high-pa) plus pt=P1/P2 (upload read/landed) or P4/P4R (readback source/result). r235 decode so far: the A upload lands ([0 1 2 3] in device pa 0x7800000 at t=.94) and the SAME device page is later rewritten to [16382 16383 8193 8194] - h_A's own values at shifted offsets - with the pattern repeated across consecutive device pages (0x7800000/0x7810000/ 0x7820000), i.e. a repeated-page/offset-wrapped write inside the emulated copy or engine path, not client-side. The trace fix makes the next capture name the exact path (missing H2D included). (cherry picked from commit 9b35145136c949405c7fb4d22b3405c174ecb055)
generated by cgit v1.2.3 (git 2.39.1) at 2026-10-21 14:27:49 +0000