Research window: 2026-11-19 08:00 ~ 2026-11-20 08:00 (Beijing time, 24 hours)

Sources: Official news API, main repository commits and PRs, releases and tags, org-wide repository and packages pushes, topic searches, Google News (via proxy), HN, EET-China column.


In This Issue

  • I. Official and Community Activities
    • Two-city technical salons scheduled for December 12 and December 19 (11-19)
    • AI Kit toolchain to be released on-site at the salon (11-19)
    • Official news API sees zero new items for the tenth consecutive issue (11-20)
  • II. Kernel and Security
    • Mailbox handle not validated, allowing read/write to any kernel (11-19)
    • Fix proposal touches ten files (11-19)
  • III. Device Layer and Build Tools
    • Device lifecycle lock made into an option disabled by default (11-19)
    • dfs open failure now reports positive errno (11-19)
    • Group library’s local compile flags applied twice (11-19)
    • Pico documentation adds toolchain path override syntax (11-19)
  • IV. Trade-offs of Serial V1
    • Timeout event has no branch in V1’s switch (11-19)
    • A five-line fix and six hundred lines of formatting (11-19)
    • Maintainer’s stance: V1 kept for compatibility, new BSPs guided to V2 (11-20)
  • V. Packages and Ecosystem
    • siot enters the index as a configurable MQTT client (11-19)
    • NFC card reader library opens rtt branch and first alpha tag (11-19)
  • VI. Review Queue and Collaboration
    • Zero merges in a day, queue grows from 138 to 142 (11-20)
    • RH850 port stuck on CI credentials (11-19)
    • Maintainer performs batch review at 19:05 (11-19)
  • VII. Summary and Trend Observations
    • All increments are on the proposal side
    • An assertion is not an ownership check
    • Revisiting three old issues
    • Two events in December, AI Kit debut

I. Official and Community Activities

1.1 Two-city technical salons scheduled for December 12 and December 19 (11-19)

Date: 2026-11-19 17:30 (published). Source: EET-China.

The Beijing session is on December 12, at the Entrepreneur Reception Hall, 22nd Floor, Tower DH3-A, Dinghao, No. 3 Haidian Street, Haidian District. The Shenzhen session is on December 19, at the Multi-function Hall, 5th Floor, Tower A, Fintech Building, Shenzhen Science and Technology Industrial Park.

The article claims a gathering of eight leading chip vendors, listing Infineon, Renesas, ST, NXP, GigaDevice, Fullhan, NOVOSENSE, Arm, and XuanTie. The topics are described as edge AI, machine learning deployment, new product matrices, and open ecosystem evolution.

On-site perks are listed as development board blind boxes, check-in blind boxes, and multiple rounds of raffles, with the top prize being a DJI Pocket 3 worth 2,000 yuan. Registration is first-come, first-served, with limited seats.

This is the only new item from official channels within the window. The same column also published a preview of the HPMicro Ecosystem Day on 11-18, which fell within the previous issue’s window.

1.2 AI Kit toolchain to be released on-site at the salon (11-19)

Date: 2026-11-19 (article published). Source: Salon announcement · same article.

The article calls RT-Thread AI Kit the “biggest Easter egg” of this salon, positioned for edge intelligence development, bringing AI capabilities into embedded devices. An experience session will be held on-site.

The original text gives no version number, supported chip list, or interface form. These will have to wait until the December 12 Beijing session.

Assessment: Placing the toolchain debut at an offline salon rather than a standalone release suggests the release cadence follows events. The basis is that the article only says “released on-site” and “verify upon arrival,” with no version number or download entry.

1.3 Official news API sees zero new items for the tenth consecutive issue (11-20)

Date: 2026-11-20 (API check). Source: API.

The API returns five items. The latest is still “From OS Foundation to AI Implementation: RT-Thread’s Full-Stack Capabilities Anchor Industrial Intelligent Upgrades” from November 6 at 15:54. Counting from November 7, this is the tenth consecutive issue with zero new items.

The official website news page is client-rendered and cannot be scraped for entries, so the API is authoritative.

II. Kernel and Security

2.1 Mailbox handles are not validated, allowing read/write to arbitrary kernel memory (11-19)

Date: 2026-11-19 03:36 (fix proposal submitted). Source: #11960.

The issue was filed by XueDugu on October 31, titled around an unvalidated fake object enabling arbitrary kernel read/write. The issue cites ecf2d8215961 (2022-12-03) as the introducing commit, with v5.0.0 as the first affected version; v5.3.0 and master remain affected. It is tagged in progress, with five comments.

The mechanism has three layers. sys_mb_send() passes the user-supplied mb directly to rt_mb_send(). sys_mb_recv() validates the output pointer value but not mb. The RT_ASSERT in the mailbox implementation only reads the type field of the struct.

LWP already has an object ownership table, maintained by lwp_user_object_add/delete. These IPC system calls bypass that table.

The path is therefore direct. A process maps a page and fills it with a fake struct rt_mailbox. It sets type to RT_Object_Class_MailBox, points msg_pool at an arbitrary kernel address, and then calls the public mailbox system calls.

The PoC runs on qemu-virt64-aarch64. The configuration is GICv2, single core, 128 MiB. The switches are RT_USING_SMART=y, RT_USING_MAILBOX=y, RT_USING_DEBUG=n.

The result is that the target qword changes from 0x74003e2f2068736d to 0x1122334455667788, and sys_mb_recv can read the same qword back.

The issue also includes board-level validation on a Raspberry Pi Pico 2 W (RP2350). The board runs single-address-space RT-Thread, validating the kernel mailbox API primitive; the user/kernel boundary is demonstrated by the QEMU full-system run.

2.2 The fix proposal touches ten files (11-19)

Date: 2026-11-19 11:36 (submitted). Source: #12015.

The proposal was opened by cms19859230182-lang on November 10 and pushed as commit f76c0ef4 at 11:36 a.m. on November 19. The scale is 10 files, 1440 lines added, 126 lines removed.

The largest changes are in lwp_pid.c (+746/-95) and syscall_rt_ipc.c (+331/-23). The other eight are src/ipc.c, src/timer.c, src/thread.c, src/object.c, src/defunct.c, and three header files.

The proposal body contains only three sentences. Handles must be matched against the object table by type and a reference taken; timer-related parameters must be rechecked; closing and enqueueing must share a single lock.

Assessment: This is not a patch for a single system call; it replaces how RT-Smart resolves handles wholesale. The basis is that the changes land simultaneously in five places—lwp, ipc, object, thread, and timer—while the issue points at only the mailbox path.

III. Device Layer and Build Tools

3.1 Device lifecycle lock made an option that is off by default (11-19)

Date: 2026-11-19 10:44 (proposal). Source: #12068.

The proposal responds to #11234. That issue was raised on March 3 and described how multiple threads opening and closing the same device concurrently causes ref_count to be miscounted. The proposal has 5 commits, 5 files, 366 lines added, and 48 lines removed.

The approach adds a new Kconfig option, RT_USING_DEVICE_LIFECYCLE_LOCK, which depends on device and mutex and is off by default. When enabled, each device uses its own priority-inheritance mutex to serialize initialization, open-mode decisions, reference-count updates, and driver callbacks.

Callbacks may sleep. Calls from different threads wait for the current transition to finish. Recursive transitions on the same device return -RT_EBUSY. Nested calls across devices require the caller to guarantee lock ordering.

The defect was reproduced on vexpress-a9 with a counting driver. In the baseline, an independent device receives two driver opens and has two active references; after the final close and reopen overlap, only one reference remains, open_flag drops to zero, and the driver does not reopen. It reproduces on both single-core and dual-core SMP.

The test case is 229 lines and runs under kernel_basic.cfg, which already enables this option.

Assessment: Being off by default means existing behavior is unchanged; only explicit enablement alters scheduling and latency. This is based on the option depending on device and mutex, and on the proposal carving out a separate exception for the single-threaded board-init case that opens the console.

3.2 dfs open failures now report positive errno (11-19)

Date: 2026-11-19 08:54 (proposal). Source: #12067.

The proposal relates to #10558. That issue was raised on July 28, 2025 and described the inconsistent sign of error codes from rt_set_errno(). The proposal has 4 commits, 6 files, 108 lines added, and 71 lines removed.

There are three symptoms. Opening a nonexistent file on a mounted filesystem returns -1, but errno is negative. When the descriptor table is full, DFS v1 reports -ENOMEM in errno, while DFS v2 returns 0, and the caller treats 0 as a valid descriptor. On RT-Smart, ARM processes receive this false success from both the raw syscall and the musl side.

The fix carries the allocation-failure reason out of the private slot helper. Reaching the descriptor limit reports EMFILE; failure to grow the table or allocate a file object reports ENOMEM; the shared allocator publishes the error after unlocking. A failed POSIX open() returns -1 with the corresponding positive errno, and a successful open may still return descriptor 0.

The test case is added to tc_posix_api.c, 60 new lines, covering both the missing-file and full-table cases. The A9 DFS CI configuration is switched to the current Kconfig symbols and selects the board-level sd device.

The validation environment is bsp/qemu-vexpress-a9, QEMU 7.2.22. Both DFS v1 and v2 are run, the filesystem is tmpfs, and DFS_FD_MAX=12.

The proposal body attributes the symptom to commit 6fdfe700. In the repo, that hash is the mem.cfg commit, which only touches memheap test cases.

Assessment: This reference appears to be a typo. This is based on 6fdfe700’s commit message being unrelated to dfs, and on none of the three commits merged on 11-18 containing dfs changes.

3.3 Local compile flags for grouped libraries applied twice (11-19)

Date: 2026-11-19 22:03 (proposal). Source: #12070.

The proposal fixes #4417. That issue was raised on March 4, 2021 and labeled enhancement. The change touches only tools/building.py, with 27 lines added and 21 lines removed.

The symptom is that a group declared with LIBRARY=True that also carries local compile flags registers two different SCons actions for the same object. One goes through the library builder, the other is generated when the application processes the local flags.

The fix compiles the library group’s objects with local flags the same way as an ordinary group, then archives them. During a normal application build, these library sources are no longer added to the separate local-object traversal.

Validation used SCons 4.11.1 and a minimal fixture with scons -Q -n. No firmware build or board-level test was run.

3.4 Pico docs add toolchain path override (11-19)

Date: 2026-11-19 19:25 (proposal). Source: #12069.

The proposal responds to #11206. That issue was raised on February 18, 2026 and described how Pico’s generate_headers.py cannot find arm-none-eabi-gcc on Linux and WSL.

After #11635, the SDK sub-build already inherits the toolchain directory from the configuration, but the RP2040 README did not document how to configure it. This change adds the RTT_EXEC_PATH override and a command to verify the selected compiler.

The change touches only bsp/raspberry-pico/RP2040/README.md, with 12 lines added. It does not touch build logic or BSP configuration. The author notes that local validation confirmed rtconfig.py reads the override and SConstruct passes it to the Pico SDK subprocess.

IV. Trade-offs of Serial V1

4.1 The timeout event has no branch in V1’s switch (11-19)

Date: 2026-11-19 22:54 (discussion update). Source: #11909.

RT_SERIAL_EVENT_RX_TIMEOUT has the value 0x05. It is defined in dev_serial.h, but the switch in rt_hw_serial_isr() only handles 0x01 through 0x04, and there is no default branch.

When a driver sends this event, the framework does nothing. The event is silently dropped.

Two drivers in the repo rely on it to split data frames. The ch579m BSP sends it from drv_uart.c when UART_II_RECV_TOUT arrives.

The SWM341 driver in the synwit libraries sends the same event from drv_uart.c when UART_IT_RX_TOUT is set.

Both BSPs use V1, so the “frame received” notification never reaches the application layer. Frame boundaries are lost, and the only fallback is reporting via RX_IND based on the FIFO watermark.

The situation differs in V2. dev_serial_v2.h only defines 0x01 through 0x04, without this macro, so a driver referencing it would fail to compile outright. The author notes that adding a case in V2’s isr would raise 'RT_SERIAL_EVENT_RX_TIMEOUT' undeclared. If V2 needs this semantics, it should go through the read-side RT_SERIAL_CTRL_SET_RX_TIMEOUT.

4.2 A five-line fix and six hundred lines of formatting (11-19)

Date: 2026-11-19 (scope of the proposed changes). Source: 34d2a002.

The proposal spans 4 files in total, adding 868 lines and deleting 504. The number looks large, but it becomes clear once broken down.

The actual fix is 34d2a002, which adds 5 lines and deletes 0 in dev_serial.c. The fix merges the timeout event into the RX_IND branch: first drain the hardware FIFO, then invoke rx_indicate with the length already buffered in the software FIFO.

Another commit, 87b851b4, is clang-format, adding 622 lines and deleting 505 in the same file. The formatting is what inflates the diff.

The remaining three files are test cases. uart_rx_timeout.c is 226 lines, Kconfig 5 lines, SConscript 11 lines. The test cases do not depend on hardware and can run without a board.

4.3 The maintainer’s stance: keep V1 for compatibility, steer new BSPs to V2 (11-20)

Date: 2026-11-20 06:54 (comment). Source: #11909 discussion thread.

The controversy surfaced as early as October 22. Rbb666’s comment was: V2 has been out for a long time, and maintaining V1 and V2 in parallel over the long term is a trap; personally, I think V1 should be frozen for historical compatibility only, with serial’s focus shifting to V2.

The same comment contains two more facts. RTT’s summer camp and introductory training still use V1. Many BSPs only use serial for MSH, with very shallow requirements, a legacy matter. He believes subsequent BSPs should proactively guide contributors to prioritize V2, and BSPs such as Renesas have already adapted to V2.

Early this morning Yunlk replied, building on this positioning. Since V1 will continue in service, the events it silently drops are worth fixing. This proposal only adds one case, touches neither V2 nor any driver interface, and includes test cases that do not depend on hardware.

Assessment: This discussion defines V1’s nature: not frozen, but low-maintenance. The basis is that the maintainer said “BSPs that only want to use it for MSH can keep V1,” not “V1 stops accepting fixes.”

V. Packages and Ecosystem

5.1 siot enters the index as a configurable MQTT client (11-19)

Date: 2026-11-19 14:05 (proposal update). Source: packages #2059.

The proposal was raised by yzy726 on November 18 and updated within the window to November 19, 14:05. It changes 3 files, adding 183 lines. iot/Kconfig 1 line, iot/siot/Kconfig 128 lines, iot/siot/package.json 54 lines.

The package itself is an MQTT 3.1.1 client supporting QoS 1 and bounded I/O, optional TLS with CA and hostname verification, and depending on Mbed TLS 2.x. The version is pinned to v1.1.1 (commit c4956bee), with v1.1.0, v1.0.1, v1.0.0 and latest retained in the index; older versions hide the newly added channel options.

The configurable items are Host, port, Client ID, account and Topic. The proposal notes it contains no personal server credentials and is not tied to any board, sensor, or course application.

There is also an optional layer: command routing at the single-device, group, and deployment levels, kept separate by device state and acknowledgment topics, with the routing layer off by default. It comes with a standalone Python teacher-side web controller, including feature toggles, a roster, an action allowlist, an auth API, and Mosquitto 2.x ACL export.

5.2 The NFC reader library opens an rtt branch and its first alpha tag (11-19)

Date: 2026-11-19 17:21 (branch commit). Source: rtt branch.

The repo is an RT-Thread port of NXP Reader Library 07.18.00, self-described as providing OSAL/BAL/DAL and Kconfig. On November 19 the rtt branch was created, and the tag v0.1.0-alpha.1 was placed on the same commit 2828285.

The branch has two commits within the window. The 15:16 commit adds an RC663 SPI port to NXP Reader Library. The 17:21 commit adds bilingual documentation and a CI workflow.

The latest commit on main stops at November 18, 10:18; development has moved to the rtt branch.

Assessment: This is a personal project with 0 stars, and its maturity is not comparable to packages in the index. It is listed because it produced verifiable output that day and is a rare NFC-direction port in the RT-Thread ecosystem.

VI. Review Queue and Collaboration

6.1 Zero merges in a day, queue grows from 138 to 142 (11-20)

Date: 2026-11-20 (queue count). Source: Proposal list.

During the window: 0 merges, 4 new proposals, 0 new issues. Open proposals: 142 (previous period: 138); open issues: 341 (unchanged from previous period). 138 plus 4 new proposals equals exactly 142.

Zero commits to the main repo; the latest master commit remains 6fdfe700 (11-18 17:29). None of the org’s 16 repos received any pushes during the window.

All push timestamps predate the window. rt-thread last pushed at 11-18 17:29, packages at 11-18 09:45, env-windows at 11-16 17:14.

In verifiable records, the main repo landed between 2 and 24 commits per day from 11-16 to 11-19. Today: zero.

6.2 RH850 port blocked on CI credentials (11-19)

Date: 2026-11-19 19:05 (comment). Source: #11998.

The Renesas RH850 G4KH port and qemu-rh850-u2c board package still stands at 52 files, +4122 lines. The latest commit is 6309b2b4 (11-18 22:38), with no content changes.

What changed is the review note. At 19:05, BernardXiong wrote that the failure fingerprint is identical to before. The only failing step is Install riscv32-esp-elf ToolChains for ESP32C6 under RT-Thread BSP Static Build Check. That step returns 401 Bad credentials on curl -H "Authorization: ***".

His conclusion: this is a fork-PR credential and environment issue, unrelated to this proposal. A direct rerun will most likely hit the same empty token; it needs to wait for the CI credentials to be fixed before rerunning, or the job can be ignored at merge time.

6.3 Maintainer batch review at 19:05 (11-19)

Date: 2026-11-19 19:05 (event time). Source: #11842.

Around 19:05, the update timestamps of #11998, #11842, #12065, and #12069 all refreshed simultaneously. The event types are mentioned and subscribed — i.e., the maintainer named the relevant people in comments.

No new commits. #11842 still stands at 4 files, +314 lines, -8 lines. #12065 still stands at 1 file, +8 lines, -5 lines.

VII. Summary and Trend Observations

7.1 All the increment is on the proposal side

No commits and no merges during the window. The four new proposals total 13 files, +513 lines, -140 lines.

Assessment: The weight of this day lies in proposals, not in landings. The basis: zero commits to the main repo, zero merges, while each of the four new proposals is tied to an issue — the earliest filed in March 2021.

7.2 An assertion is not an ownership check

RT_ASSERT reads the type field in the struct. In a fake object, that field is filled in by the caller.

Assessment: This kind of gap cannot be stopped by an assertion. The basis: the PoC passes the assertion simply by writing type as RT_Object_Class_MailBox, whereas a real ownership check must rely on the object table.

7.3 Revisiting three old issues

#4417 was filed on March 4, 2021; #8143 on October 17, 2023; #10558 on July 28, 2025. Today’s three proposals correspond to them respectively.

Assessment: What gets old issues moving is not agreement but someone taking action. The basis: all three issues had been open for a long time with few comments, and the proposals arrived with commits and test cases.

7.4 Two events in December, AI Kit debut

December 12 in Beijing, December 19 in Shenzhen. Both events run a full afternoon, with the AI Kit released on site and a hands-on session.

Assessment: The AI Kit currently has only a description — no version number, chip list, or interface. The debut is offline, so details will have to wait for the Beijing event. The basis: the article only says “released on site” and “verify on arrival.”

Appendix: Source Verification Table

Source Verification Result
RT-Thread official news API Five items, latest 11-06 15:54, tenth consecutive issue with zero new additions
Official website news page Client-side rendered, items cannot be scraped; API is authoritative
Main repo commits Zero commits in window, master latest 6fdfe700 (11-18 17:29)
PRs merged in window 0
PRs opened in window 4: #12067, #12068, #12069, #12070
Issues opened in window 0
Open PRs / issues 142 / 341 (previous issue 138 / 341)
releases and tags Latest remains v5.3.0 (2026-09-10), no new additions in window
RT-Thread org repos 16 repos, zero pushes in window
Package index repo packages Zero commits; pending PR #2059 updated 11-19 14:05
Windows tools repo env-windows Last push 11-16 17:14, zero commits in window
topic:rt-thread active repos 2 repos: NXP_NfcRdLib created rtt branch and tagged alpha; oxide-compiler is a Rust compiler, unrelated
Google News Chinese 1 item in window (Shuangcheng salon, 11-19 17:30); the 11-18 and 11-17 items predate the window
Google News English Zero hits in 3-day and 7-day windows; bare query works normally, so channel failure can be ruled out
HN Algolia Both candidates were literal “thread” false matches, removed
Security issue #11960 is RT-Smart mailbox unauthorized read/write, tagged in progress, fix in #12015

Full Source List