Research window: 2026-11-17 08:00 ~ 2026-11-18 08:00 (Beijing time, 24 hours)

Sources: RT-Thread official news API (club.rt-thread.org, category rt_org), GitHub main repository commits / PR / Issue / releases / tags, all repositories in the RT-Thread org checked in push order, package index repository packages, Windows tool repository env-windows, topic:rt-thread active repository search, Google News Chinese and English queries (via proxy, including three-day and seven-day windows), Hacker News Algolia, Tavily search


Index of This Issue

  • I. Kernel and security fixes
    • BLK_PARTITION ioctl: user pointer used directly as kernel address (11-17)
    • FinSH path completion: a one-line truncation after more than fifty days (11-17)
    • STM32 input capture: one timer split into four devices (11-17)
  • II. Newly reported defect: DMA ping buffer and cache line
    • serial_v2 misalignment: a 44-byte header pushes the ping buffer off the 32-byte boundary (11-17)
  • III. BSP and vendor adaptation
    • ArtInChip D213: from under review to merged in three days (11-17)
  • IV. Official and community activities
    • Grand prize project of the Embedded Competition open-sourced: a six-layer Agent monitoring system (11-17)
    • Official news API has no new items for the eighth consecutive issue (11-18)
  • V. Review queue
    • Three maintenance patches queued in the early morning (11-18)
    • Two long-running proposals still on the periphery (11-17)
    • One duplicate proposal withdrawn by its author (11-17)
  • VI. Summary and trend observations
    • Four items cleared in one day; the security line is converging
    • Delay is almost entirely in the segment from report to proposal
    • The evidentiary standard for defect reports is rising
    • The three-day cycle on the ecosystem side

I. Kernel and Security Fixes

1.1 BLK_PARTITION ioctl: User Pointer Used Directly as Kernel Address (11-17)

Date: 2026-11-17 14:50 (merge time) Source: PR #11981, Issue #11952, Commit 265116d7

The issue was filed on October 29, with the reported range listed as master (v5.4.0-dev) and v5.2.0 through v5.3.0; the introducing commit is c424cb818695 in the common block layer. The reproduction environment was QEMU vexpress-a9, RT-Smart, ARMv7-A, 32-bit, with an SD image, where the kernel creates /dev/sd0.

The shape of the problem is straightforward: RT_DEVICE_CTRL_BLK_PARTITION treats a raw pointer passed in from user mode as the destination address for rt_memcpy, so an unprivileged process can designate an arbitrary kernel address as the write target. Worse, rt_memcpy is not followed by setting err to success, so even if the ioctl ultimately returns -1, the contents of the target address have already been overwritten—the error code contradicts the facts, and the “failure” the caller receives is an illusion.

The fix takes two paths. Under RT_USING_LWP, lwp_self() is first used to determine whether execution is in a user-mode context; if so, the partition data is first copied to a temporary variable on the kernel stack and then safely copied back with lwp_put_to_user(). When the pointer is invalid, lwp_put_to_user() returns -EFAULT and writes no bytes. Internal kernel calls continue to use the original rt_memcpy path, with the missing err = RT_EOK added. The change is 294 lines added and 59 lines removed across five files, all under components/drivers/block/: blk.c, blk_dev.c, blk_dev.h, blk_dfs.c, and blk_dfs.h.

The PR was opened on November 3 and merged on the afternoon of November 17—14 days in between. The issue itself was not closed along with the PR and remains open; whether commands such as RT_DEVICE_CTRL_BLK_GETGEOME mentioned in passing by the reporter are structurally similar has no conclusion in this issue.

1.2 FinSH Path Completion: A One-Line Truncation After Fifty-Plus Days (11-17)

Date: 2026-11-17 11:33 (merge time) Source: PR #12057, Bug #11839, Commit 4c823281

Bug #11839 was filed on September 25, with the title explicitly citing CWE-122 heap out-of-bounds write: msh_auto_complete_path does not guard the fixed-length buffer (RT_PATH_MAX or a hardcoded 256 bytes) when concatenating the current working directory with user input, so pressing Tab to trigger completion on an overlong path in msh is enough to write past the boundary. The reported environment was stm32f407-atk-explorer, GCC, v5.3.0.

The fix is only 5 lines added and 0 lines removed, in a single file, components/finsh/msh.c. The PR description is equally blunt: truncate overlong input paths directly rather than letting the stack blow up in place. It was opened on November 14 and merged on the morning of November 17, with the bug closed in the same second.

In yesterday’s report this was still one of the “four small community patches still queued on the periphery”; today it is the third of the four to land. 53 days from report to fix, 3 days from PR to merge—the delay is almost entirely concentrated in the stretch after the report, before someone takes action.

1.3 STM32 Input Capture: One Timer Split into Four Devices (11-17)

Date: 2026-11-17 08:59 (merge time) Source: PR #12052, Commit 0dffe478

The original driver’s granularity was the physical timer: one TIM registered one device, only one capture channel could be selected, and the remaining channels of the same timer sat idle; there was also no PWM input or dual-edge capture based on CH1/CH2 pairing.

The new version lowers the granularity to the channel. TIM1 through TIM4 can each select CH1 through CH4, and different channels of the same physical timer are each registered as independent Input Capture devices, sharing the counter, prescaler, and HAL handle while maintaining their own capture state. Two exclusive modes are added: PWM Input pairs CH1 direct capture with CH2 indirect capture, using the timer’s Slave Reset to compute the period and the high/low level durations; DUAL_EDGE uses CH1/CH2 to capture the two edges of the same signal separately, with the counter running freely and the interval between adjacent edges computed from timestamps in the interrupt. Both exclusive modes require that the entire TIM not be occupied by other independent EDGE devices.

Kconfig gains mode and channel selection options for TIM1 through TIM4, and board-level parameters (device name, digital filtering, capture polarity, ARR) go into two inputcapture_config.h files for f1 and h7. The change is 1540 lines added, 285 lines removed, across five files. The PR was opened on November 12 and merged 5 days later.

II. Newly Reported Defect: DMA ping Buffer and cache line

2.1 serial_v2 Misalignment: 44-Byte Header Pushes ping Buffer Off the 32-Byte Boundary (11-17)

Date: 2026-11-17 08:33 (issue time) Source: Defect #12063

At 8:33 a.m., a defect report with hardware reproduction landed in the issue tracker. The title was long, but it captured the full causal chain: the DMA receive ping buffer in serial_v2 is not aligned to the cache line, the D-cache invalidation operation overruns into the adjacent heap block, and on STM32H7 this manifests as the tx fifo ring buffer head rolling back and the next write triggering an assertion.

The allocation works like this: rt_serial_rx_enable() performs a single rt_malloc that grabs three contiguous regions — the rx fifo struct, the rx ring buffer, and the DMA ping buffer — with the ping buffer starting at sizeof(struct rt_serial_rx_fifo) + rx_bufsz. On 32-bit Cortex-M7 this struct is 44 bytes (two 16-byte rt_ringbuffer plus rx_cpt, rx_cpt_index, rx_timeout). So whenever rx_bufsz is a multiple of 32, the ping buffer start is always 44 + rx_bufsz ≡ 12 (mod 32) — it never lands on a 32-byte boundary, and the block end is likewise misaligned.

The trigger is in the STM32 serial_v2 driver: on every DMA receive idle event it calls the CMSIS D-cache invalidate routine on this ping buffer. Cortex-M7 cache lines are 32 bytes, and invalidation operates on whole lines, so both ends go wrong. Downward, up to 31 bytes before the ping buffer (the tail of the rx ring buffer, data the CPU just wrote) are invalidated along with it, dirty lines are discarded, and the data reverts to the stale contents in DDR — silent corruption of received data. Upward, the CMSIS version in the STM32Cube package rounds the length up by the start offset (op_size = dsize + (addr & 31)), so the loop runs up to 31 bytes past the end of the ping buffer, invalidating the next heap block.

The consequence observed on hardware is very concrete: the heap block immediately following the rx fifo is the tx fifo (allocated separately by rt_serial_tx_enable()), and the buffer_ptr of its first field rb sits at offset 0, right inside the over-invalidated line. The pointer written by rt_ringbuffer_init() is discarded before it is written back, memory re-reads as the pre-allocation zero value, buffer_ptr becomes NULL, and the next rt_device_write() asserts outright. In one sentence: enable serial_v2 DMA receive on an M7 board with D-cache on, receive a few bytes and then send, and the transmit side dies.

The reporter gave a framework-level fix: in rt_serial_rx_enable(), align the ping buffer start upward to 32 bytes (RT_ALIGN) and round the length up to a multiple of 32, leaving about 32 bytes of extra headroom at allocation so both line-rounding operations stay within their own block; zero cost on cores without D-cache. The impact scope is stated clearly: both v5.3.0 (the first version with the DMA ping design) and master are affected, introduced by ac8fba012 (author date 2024-09-04, the serial_v2 rework merged 2025-05-16); the same buffer is also handed to the serial_v2 drivers for gd32, hc32, n32, at32, wch, and imxrt. Those cores currently have no D-cache, but imxrt’s drv_uart_v2.c does no cache maintenance at all, which the reporter noted separately. Self-assessed severity is medium, CVSS 3.1 roughly 4.6 to 5.3; triggering requires physical or serial-line access, and the consequence is kernel heap integrity damage plus a deterministic assertion.

The issue was claimed by liangzhan5 four and a half hours later — the same maintainer who merged the block device change that morning. Labels: in progress and component: drivers.

III. BSP and Vendor Adaptation

3.1 ArtInChip D213: From Under Review to Merged in Three Days (11-17)

Date: 2026-11-17 11:19 (merge time) Source: PR #12058, commit 9ddb0249

D213ECV-DEMO-V4-0 is ArtInChip’s evaluation board, built around a T-Head C906, RV64GC, 128 MiB DDR3. Proposed on November 14, merged on the morning of November 17 — review completed in three days, faster than most proposals in the same period.

Thirty files, 1783 lines added, zero deletions. The content is a complete board package: it reuses the RISC-V common port layer (M-mode startup, context switching, FPU state, PLIC, system tick), adds D213 platform initialization, link layout, GPIO and UART0 clock and reset, and serial-v1 UART0 adaptation; the external loader first initializes DDR, PLL, and memory and cache attributes, then loads RT-Thread at 0x40000000. UART0 transmits by polling, the receive thread feeds serial-v1’s FIFO, the receive thread has higher priority than FinSH, the default receive buffer is 512 bytes, transmit and configuration waits both have limits, and the divisor write requires a successful LCR/DLAB readback.

In the report the day before yesterday it was still a pending proposal “sharing an origin with that floating-point fix”; today it is a BSP directory in mainline. For observers on the RISC-V side, what is worth noting is the consistency of this thread: the D213 porter first hit an incorrect fs1 after a thread switch on their own BSP, reproduced it on the official qemu-virt64-riscv, and that is what produced the floating-point context fix; by the time the board package came in, that fix was already in mainline.

IV. Official and Community Activities

4.1 Grand Prize Project from Embedded Contest Goes Open Source: A Six-Layer Agent Monitoring System (11-17)

Date: 2026-11-17 17:30 (published) Source: EE Times China column article

The 2026 RT-Thread Embedded Contest, whose rankings were announced on November 9, had its grand prize entry technically deconstructed yesterday evening in a column on the official WeChat account. The work, titled “Multi-Sensor Device Information Fusion Monitoring System Based on an On-Device Agent Kernel,” is by author Cao Xiansong; the hardware is the contest-provided Edgi-Talk (Infineon PSoC E84: Cortex-M55 with Helium, M33, Ethos-U55 NPU, 6 MB SRAM), and the software runs on RT-Thread 5.0.2.

The architecture has six layers: application layer (seven LVGL pages), Agent kernel, tools and channels, network protocol stack, kernel and BSP, and external services. The Agent kernel is taken from the upstream standalone software package Agent-latest, imported read-only; a single AgentLoop thread processes sessions serially, with a 10 KB stack and priority 10, and input and output mailbox groups each with depth 10, dropping messages and logging when a queue is full. A single conversation round sends at most 6 requests to the large model, executes the same tool at most 3 times, and skips calls whose parameter signatures are completely identical; the context is trimmed to 16 entries per round, with the system prompt resident throughout.

The tool layer registers 10 tools: three examples (add, mul, compare) plus seven MQTT business tools (mqtt_connect, mqtt_disconnect, mqtt_publish, mqtt_subscribe, mqtt_receive, mqtt_rule, mqtt_history); the subscription table has 13 slots and 4 threshold rules, each topic retains its last 6 history entries for trend summaries, and a rolling window cache of depth 8 holds messages that cannot be analyzed immediately. The network stack is wifi-host-driver (CYW55500 Wi-Fi 6), SAL and netdev, lwIP 2.1.2, paho-mqtt, webclient 2.2.0, and mbed TLS 2.28.1; the file system uses DFS and FAL.

The article itself spells out the gaps: the broker currently uses plaintext TCP with username/password authentication, and TLS or account isolation must be added before multi-node access at scale; TTS, NPU inference, and service-layer management and visualization remain reserved items. The code is on Gitee (cxs_123/muti_device_fusion) and can be imported directly with RT-Thread Studio.

The point of this piece is not “another grand prize winner” but that it draws the Agent kernel’s boundary clearly: intent understanding, fusion assessment, and alert copy go to the remote large model, while the device side only does message routing, tool scheduling, and result feedback. For a real-time kernel, this is currently the most compute-efficient and least error-prone split.

4.2 Official News Feed Has No New Items for the Eighth Consecutive Issue (11-18)

The feed still holds the five old articles uploaded on November 6: the full-stack capabilities piece on industrial intelligent upgrading, the PLC piece on the Vinaro tech salon, the Southeast Asia industry-education integration piece, the automotive software industry ecosystem salon piece, and the AI car piece from the Industrial Fair. Checking upload times one by one, all fall between 10:53 and 15:54 on November 6, with zero new items in this issue’s window. The official news page is client-rendered and its entries cannot be scraped, so the feed remains the source of record.

V. Pending Review Queue

5.1 Three Maintenance Patches Queue Up in the Early Morning (11-18)

Date: 2026-11-18 03:11 to 03:13 (filing time) Source: PR #12064, PR #12065, PR #12066

Between 3:11 and 3:13 a.m., parthitce filed three patches in a row: [libc][common] makes sys/signal.h adopt picolibc’s type definitions (#12064), [utest][klibc] confines TC_rt_memmove_too_long’s accesses to its own buffer (#12065), and [ci][utest] selects RT_USING_MEMHEAP in mem.cfg (#12066). None of the three changes functionality; they are test and header-consistency cleanups, all small in scope.

5.2 Two Long-Running Proposals Remain on the Periphery (11-17)

Date: 2026-11-17 (latest update) Source: PR #11998, PR #11912, PR #11842, PR #11964

The RH850 G4KH port and qemu-rh850-u2c board package (#11998) has been open since November 6, with 48 files and 3,921 lines added, and was still updated early yesterday morning; OctaFlash support (#11912), open since October 21, has 34 files with 4,550 additions and 1,302 deletions, the largest in the queue; the HUST CSE fix for a stack overflow in the devtmpfs long-path component (#11842) has been queued from September 27 to today, with 4 files and 314 additions and 8 deletions; and the signal test case (#11964) has been open since November 1. What all four have in common is that they are waiting on review, not on the author.

5.3 One Duplicate Proposal Withdrawn by Its Author (11-17)

Date: 2026-11-17 09:16 (close time) Source: PR #12055, PR #12029

The ptyfs unregister-release proposal (#12055), filed on November 14, was closed at 9:16 a.m. on November 17 without being merged—the same issue had already been merged the day before via #12029. This is a common ending on the audit line this round: the same defect is spotted by two groups, the earlier one lands in mainline, and the later one withdraws itself.

VI. Summary and Trend Observations

6.1 Four Cleared in a Day; the Security Line Is Converging

Four merges landed in the main repository during the window: an arbitrary kernel write in block device partitions, a FinSH path completion overflow, STM32 input capture extension, and the ArtInChip D213 board package. The first two are security, the latter two features. Open proposals: 141 (143 last issue); open issues: 341 (unchanged from last issue).

For the audit line, the significance of this day is settling accounts: the block device issue reported on October 29 was merged today, and the FinSH overflow reported on September 25 was closed today. Together with ptyfs, ivshmem, and iso9660 from earlier issues, the backlog of RT-Smart and file-system defects reported since late July is now largely cleared from mainline.

6.2 The Delay Is Almost Entirely in the Report-to-Proposal Stage

Both are merges, but the time structure of the two security fixes is completely different. The block device one was reported October 29, proposed November 3, and merged November 17—5 days from report to proposal, 14 days from proposal to merge; the FinSH one was reported September 25, proposed November 14, and merged November 17—50 days from report to proposal, 3 days from proposal to merge. Yesterday’s iso9660 one took 17 minutes.

The three samples point to the same conclusion: review is not slow; what is slow is finding someone willing to do the work. However detailed a defect report is, without a proposal it is just a line in the ledger. That is also why #12063 deserves separate attention—it was claimed within 4.5 hours, while #11839 waited 50 days.

6.3 The Evidence Standard for Defect Reports Is Rising

#12063 is the highest-quality report in recent memory: it gives the affected version range (since v5.3.0, unfixed on master), the introducing commit (with both author date and merge date), a byte-level struct size derivation (44 + rx_bufsz ≡ 12 mod 32), the behavioral difference between two CMSIS variants, the concrete hardware symptom (tx fifo header zeroed, assertion on the next write), a self-assessed CVSS, and a statement that no duplicate was found on search. This is not “I found a bug” but a work order that can be followed directly.

For maintainers, the value of such a report is that it saves reproduction; for the project, it puts a cross-module issue—the contract between framework and driver—on the table: the problem is not in any one driver, but in serial_v2 allocating three blocks of memory contiguously without stipulating alignment.

6.4 The Three-Day Cycle on the Ecosystem Side

The D213 board package was proposed November 14 and merged on the 17th. Three days is fast for a new BSP of thirty files and nearly 1,800 lines. Board-level support on the RISC-V side is clearly accelerating this year: the average time from C906 single-board to mainline directory is shrinking, and most proposals come with external loader and memory layout descriptions, giving reviewers something to work from.

The official side remains quiet: the news feed has zero new items for the eighth consecutive issue, with the last entry still the November 6 event recap. The only official item with substantive content that day was the evening deconstruction of the contest grand prize project—it is more technical evangelism than product launch, but for anyone trying to understand how an on-device Agent actually lands on a real-time kernel, it is more useful than any promotional piece.

Appendix: Source Verification Table

Source Verification Result
RT-Thread official news API Five items, all old event posts uploaded on 11-06; eighth consecutive period with zero new items
Official news page Client-side rendered, items cannot be scraped; API results take precedence
Main repo commits 4 in the window, 3,622 lines added, 344 lines deleted, 41 file changes
PRs merged in the window 4: #11981, #12057, #12058, #12052
PRs opened in the window 3: #12064, #12065, #12066
Issues opened in the window 1: #12063
Issues closed in the window 1: #11839 (closed with the merge of #12057)
PRs closed without merging in the window 1: #12055 (duplicate of #12029)
Open PRs / issues 141 / 341 (previous period: 143 / 341)
releases and tags Latest remains v5.3.0 (2026-09-10); no new items in the window
Package index repo packages Last push 11-16 11:22; zero commits in the window
Windows tools repo env-windows Last push 11-16 17:14; zero commits in the window
Other repos in the RT-Thread org No pushes in the window; top five by push recency: rt-thread, env-windows, packages, smart-build, env
topic:rt-thread active repos Only zlgopen/awtk (11-17 09:06), general GUI improvements, no specific RT-Thread association
Google News Chinese Over three days, one EE Times column (competition project sharing, already included) and four CSDN roundups; zero hits for 睿赛德 and 睿擎派 over three days
Google News English Zero hits in both three-day and seven-day windows; bare queries work normally, ruling out channel failure
HN Algolia All four candidates were literal “thread” false matches (gilguard, green threads, etc.) and were excluded
Security advisories No RT-Thread entries in the GitHub security advisory API; no new IDs this period

Full Source List