NGX mapping truncate-before-write during OTA refresh can cause UpdateFeature 0xbad00002
Nobody has claimed this yet.
Assessment
- Difficulty
- 5/5
- Estimated time
- Over a week
- Newbie friendliness
- 15/100
Research direction
Start by reading Streamline's production OTA path in source/core/sl.plugin-manager/pluginManager.cpp and source/core/sl.ota/ota.cpp, then trace the NGX mapping reader and updater publication path described in the report. Reproduce the empty-file window with a reader trace that captures the bytes and coordination results in the same run. Done means a supported publication or coordination protocol keeps readers on complete validated configuration generations and addresses consistency across files.
Written by the indexing model from the issue text.
Description
NGX shared configuration is exposed as an empty file during normal OTA refresh
Environment: Windows11 x64 build26300.9711 (26H2), RTX5070TiLaptop, driver617.42 (32.0.16.1742), public NGX/DLSS310.9.16 and Streamline2.14.16.b742680bf PRODUCTION, D3D12, Unreal5.8.4 / engine identity5.8. Actual loaded Streamline plugins: sl.common, sl.reflex, sl.dlss_g and sl.pcl. NGX feature snippets loaded for dlss, dlssg, deepdvc and dlssd report310.9.16. OTA and the application's valid ProjectID remain enabled and unchanged; the identifier is redacted.
This report follows the environment/reproduction/log information requested in Streamline's contribution guidance. It contains selected sanitized log excerpts only; no application assets, account details, source archive or full process trace is attached.
Reproduction
- Initialize NGX with a valid application ProjectID using the public D3D12 API, then obtain capability parameters. Normal NGX/Streamline startup launches signed nvngx_update.exe descendants.
- Observe, without modifying or locking it, the normal OTA configuration path:
ProgramData/NVIDIA/NGX/models/config/versions/1/files/nvngx_mapping.json. - During a naturally occurring zero-length refresh window, issue one public NVSDK_NGX_UpdateFeature request for SuperSampling(feature1), using the same valid ProjectID/engine identity.
- In the captured minimal host, initialization, capability acquisition and an earlier update succeeded. The targeted update logged LoadMappingFileData:222 “Member file in JSON file is not array” at05:22:08.9903383UTC and returned0xBAD00002 at05:22:08.9903617UTC; call duration0.135ms. The diagnostic did not edit the cache, prevent OTA writes, alter return values or change identity.
Expected: a reader sees a complete previous or complete new validated mapping generation throughout OTA refresh.
Observed: normal vendor OTA refresh temporarily exposes an empty original configuration file. A public API reader can then treat that transient state as invalid mapping data and fail.
Writer and reader evidence
Independent descendant file traces identify the signed updater as a writer of the original mapping path, including a real UE-startup SDK-bundle updater with the host as parent. The native call stack aligns with DownloadAppIndependentFiles → NGXUpdaterCurl::downloadFile → CRT file open/write. The installed617.42 downloadFile passes its supplied destination directly to a wide CRT file-open path with mode wb and sharing0x40, exposing a truncate-before-content window. The three shared configuration files each showed21 empty states in a startup with21 updater descendants; this count alone does not prove simultaneous writers.
The inspected static normal path attempts to create/wait on ngx_updater_running before calling apiUpdate and subsequently calls ReleaseMutex when a handle exists; successful runtime acquisition/ownership was not captured. The issue must therefore be assessed as a shared publication/reader transaction, not simply assumed to be missing updater-to-updater locking. Exact failed SDK read bytes and lock-acquisition results have not yet been captured in the same failing run; an incomplete tracer run is not presented as an NVIDIA runtime failure.
Public UpdateFeature in the installed core dispatches CreateProcessW and returns process-creation success without waiting for the update to finish. Repeating that call or observing success is not a download-completion barrier.
A separate ordinary UE startup tested the host OTA switches without adopting them in the application configuration. Logs confirmed Streamline flags0x85 (both OTA preference bits cleared) and explicit DLSS model OTA disabled, yet21 unique signed updater children were observed, including a non-optional SDK-bundle bootstrap/update. This is evidence of updater invocation and transient configuration empty states, not21 completed upgrades or a same-run failing SDK read. The public Streamline2.14.16 plugin manager calls checkForOTA on the production path and converts eAllowOTA into requestOptionalUpdates; the OTA implementation still invokes the SDK updater when that argument is false, omitting -optional. The release history also distinguishes optional and mandatory OTA loading. These host preferences do not establish a global no-write mode.
Natural startup recurrence, 2026-10-22 UTC
A subsequent D3D12 application startup reproduced the failure on the normal Streamline path, without synchronizing its API call to an observed empty-file window or modifying the cache. Selected sanitized lines:
2026-10-22T16:08:25.364Z commonEntry.cpp:1368 [ngxLog] LoadMappingFileData:222 Member file in JSON file is not array, invalid data
2026-10-22T16:08:25.364Z commonEntry.cpp:961 [updateNGXFeature] NVSDK_NGX_UpdateFeature(&applicationId, feature) failed 0xbad00002
2026-10-22T16:08:41.605Z LogDLSSNGXRHI: DLSS model OTA update enabled
The first two lines are local2026-10-23 00:08:25.364 (UTC+8). The later DLSS host's OTA-enabled/initialization stage begins about16s after the Streamline failure. The third line reports enabled policy, not the exact execution time of a later UpdateFeature call. Moving only that later application-side call would therefore leave the earlier observed Streamline failure uncovered.
A fresh read-only check matched all35 files in the accepted engine/driver/plugin dependency baseline; this inventory does not freeze global OTA caches or models. No same-run failed reader payload is claimed. The traceable writer mechanism, earlier API negative case and this natural recurrence remain separate observations.
Additional reader-buffer evidence remains unavailable: the installed portable CDB did not pass its ordinary-file capture prerequisite, so no real SDK capture was attempted. No successful reader capture or repair is claimed; the discontinued custom process-tree tracer remains excluded.
Checked released candidate
The official621.01 notebook package was downloaded once and verified, with only core/updater/NVI/catalog extracted. Package Authenticode, catalog trust and both PE catalog memberships passed; explicit cache-only WinVerifyTrust checks also passed. Both components report32.0.16.2101.
Complete bounded disassembly comparisons show the same downloadFile write/truncate path and matching CRT open/CreateFileW/write helpers. Core mapping-path reader, JSON parser, ifstream-open, MapProjectId and UpdateFeature retain the same normalized instructions/control flow. Two updater dispatch functions differ only in inspected debug-source-line constants. A new read-only downloaded-module signature-validation open is separate from configuration publication. This does not establish621.01 as a repair; no621.01 component was executed or installed.
Requested correction
Please preserve a complete last-known-good configuration while refreshing: download to a unique temporary file, validate content/schema, finish/flush it, then publish through a supported atomic replacement on the same volume. Readers must either retain a complete generation or participate in a supported coordination protocol. If multiple configurations must be coherent, define their generation boundary.
Please clarify three supported options while preserving application identity and existing rendering features:
- Is there a supported maintenance-only download/publication mode followed by runtime use of complete validated generations without updater writes, covering both Streamline production mandatory/bootstrap SDK requests and NGX feature updates? Please give the exact public API or setting ID/value, scope, minimum version, activation timing, completion guarantees and behavior when required content is unavailable, including other NGX applications.
- Do any published NVAPI DRS settings or R615 NGX session/runtime-parameter APIs provide that control? The pinned public DRS header exposes NGX/SL rendering, model and preset overrides; we have not identified an OTA publication/completion setting. The public runtime-parameter declarations expose session/Set/Get entry points but name the ten NV_NGX_RUNTIME_KEY values RESERVED0 through RESERVED9. If a supported control exists, please provide its documented key semantics and compatibility contract, rather than requiring applications to infer reserved values or treat per-PID feature-override state as an OTA transaction API.
- Which exact released signed component/version fixes final-path truncate-before-write publication, or supplies a completion/generation protocol covering all publishers and readers? Please provide the bug ID, release note or documented protocol, including cross-file generation consistency. A general stability statement would not establish this specific repair.
Changing IDs, suppressing errors, or relying on private event/mutex names is not an acceptable application repair. The host-switch experiment above was diagnostic only; the formal OTA and graphics settings remain unchanged.
Supporting hashes, exact RVAs, HTTP/CRC/signature receipts and selected local excerpts are retained separately. Application identity, personal directories and account information are intentionally omitted from this draft.
- Dominant language
- C
- Stars
- 1.4k
- Forks
- 174
- PR merge metrics
- No merged PRs in 30d
Getting set up
- No Dockerfile or Docker Compose file
- No pull request template
- Read the contributing guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
More from NVIDIA-RTX/Streamline
-
Title: Vulkan: VK_FORMAT_R8_UNORM missing in Vulkan::getFormatPossibly taken A pull request linked to this issue is open or already merged. Open
Difficulty 2/5 1-3 hours Newbie friendliness 74/100
NVIDIA-RTX/Streamline#133 ·
-
sl.log folder makes it hard to integrate, as many developer has *.log in the exclude filter for git/p4.Possibly taken @Dragon31337 claimed this 72 days ago. Open
Difficulty 2/5 1-3 hours Newbie friendliness 64/100
NVIDIA-RTX/Streamline#121 ·
-
Difficulty 2/5 1-3 hours Newbie friendliness 86/100
NVIDIA-RTX/Streamline#120 ·
-
C++23 issueOpenbug triaged
Difficulty 2/5 1-3 hours Newbie friendliness 72/100
NVIDIA-RTX/Streamline#111 · 1 comment ·
-
Difficulty 4/5 3-5 days Newbie friendliness 35/100
NVIDIA-RTX/Streamline#136 · 1 comment · 1 reaction ·
All issues in NVIDIA-RTX/Streamline
Similar issues
-
ecs_ptr_from_json does not free the buffer of a string longer than 256 bytes on two error pathsOpen
Difficulty 2/5 1-3 hours Newbie friendliness 72/100
SanderMertens/flecs#2444 ·
Maintainers usually reply within 1 day
-
Status: Opened
Difficulty 2/5 1-3 hours Newbie friendliness 68/100
espressif/esp-eth-drivers#201 ·
-
Difficulty 1/5 Under an hour Newbie friendliness 74/100
libsdl-org/SDL#16788 ·
Maintainers usually reply within 1 day
-
Difficulty 2/5 1-3 hours Newbie friendliness 68/100
-
Difficulty 2/5 1-3 hours Newbie friendliness 74/100
cisco/ChezScheme#1110 ·
Maintainers usually reply within 1 day