Research window: 2026-11-11 08:00 ~ 2026-11-12 08:00 (Beijing time, 24 hours)

Sources: RT-Thread official news API and official website news page, GitHub main repository commits / PR / Issue / releases / tags, RT-Thread org repository push list (35 repositories checked in order of push), Google News Chinese and English queries (via proxy, including three-day window and bare query comparison), HN Algolia, Electronic Engineering Album RT-Thread column, club.rt-thread.org forum


Index

  • I. Kernel and RISC-V Interrupts
    • PLIC: Dispatch the handler first, then complete the interrupt (11-11)
  • II. rt-link Long Frames: Issue and Fix One Hour and Forty-Three Minutes Apart
    • Issue: A single frame triggers heap overflow, present since v4.1.0 (11-11)
    • Proposal: Record the allocated length in the record structure (11-11)
  • III. Seven Commits Overnight: Audit List Turns to Three-Month-Old Debts
    • GPT and CFI: Two old integer-boundary problems (11-12)
    • 9PFS and serial ioctl: Trust boundaries between server and user space (11-12)
    • tty null LWP and pin legacy interrupts (11-12)
    • Legacy USB fix closed by policy, resubmitted the next day (11-12)
  • IV. BSP and Test Infrastructure
    • K230 PWM readback: Compensation value mistaken for high-level count (11-11)
    • utest multi-round runs start reporting debts: Atomic test case leaks 25 KB per round (11-11)
    • packages and version lines: Still no pushes this period (11-12)
  • V. Official and Community Events
    • STM32 seminar four-city tour, RT-Thread sets up booth on site (11-11)
    • Official news API has no new items for the fourth consecutive period (11-12)
    • Google News three-day window yields one hit (11-11)
  • VI. Summary and Trend Observations
    • From CVEs to in-house issues: Auditing enters its second phase
    • Quality tiers of fixes: Board-tested versus static-check-only
    • Pending review queue: 158 proposals, 365 issues
    • Tug-of-war between maintenance boundaries and willingness to fix

I. Kernel and RISC-V Interrupts

1.1 PLIC: Dispatch the handler first, then complete the interrupt (11-11)

Date: 2026-11-11 09:44 (merge time) Source: Proposal #12011, Defect #12010

The previous issue covered APLIC’s EOI timing; this issue merges the same timing issue in PLIC, submitted independently by different authors.

handler_plic_irq() originally had the order reversed: it first wrote claim to PLIC to declare completion, then dispatched the handler. If the interrupt source is level-triggered and remains asserted while the handler executes, the PLIC gateway will forward the request again, causing the same interrupt to be dispatched twice. The fix appears to be just swapping two lines—execute handler_irq() first, then write claim to complete.

A direct swap introduces a second problem: the handler may disable its own interrupt source, and PLIC ignores completion writes for “not enabled” interrupts, so this claim is never returned and the interrupt never arrives again. The proposal references Linux’s SiFive PLIC driver and temporarily sets the enable register before and after completion—first reads enable and mask; if enabled, writes claim directly; if not enabled, holds enable_lock for a second confirmation, temporarily sets the bit, writes claim, then clears the bit.

Two files, 35 lines added, 2 lines deleted. libcpu/risc-v/common/interrupt.c only changes the order; libcpu/risc-v/common/pic/pic-sifive-plic.c carries all the logic for the completion path. Author boa-z, submitted November 9, merged two days later, Defect #12010 closed the same day.

Viewed alongside #11966: one is APLIC writing setipnum_le when it shouldn’t, the other is PLIC completing before it should have handled—both fall under “completion action disconnected from the actual state of the interrupt source.” The two interrupt controller paths on the RISC-V side are now each aligned with common practice.

2.1 Issue: Single-frame heap overflow, present since v4.1.0 (11-11)

Date: 2026-11-11 13:32 (issue) Source: Defect #12030

When reassembling long frames, the rt-link receive path independently trusts three quantities — the buffer allocation size, the in-frame offset, and the current frame length — all of which can be constructed by the link peer. A single malformed frame is enough to make reassembly write past the allocated dataspace.

The reporter provided two reproductions: first, QEMU’s qemu-vexpress-a9, with the rt-link port wired to the board’s second PL011 and the peer injecting frames on the wire; second, a Raspberry Pi Pico 2 W with a trigger firmware that feeds the exact same bytes into rt_link_hw_write_cb() via the USB-CDC console — the standard board-level port entry point given in the component documentation. Both reach the same overflow point.

The affected versions are stated precisely: present since v4.1.0, verified tag by tag on v4.1.0, v5.0.0 through v5.3.0, and master; v4.0.x is unaffected. Code location is components/utilities/rt-link/, config option RT_USING_RT_LINK; the reporter notes the affected surface is firmware that “enables this component and connects to an external WiFi module through the standard board-level port,” and points out the typical configuration is RT_LINK_SERVICE_WIFI.

2.2 Proposal: Record the allocation length in the record structure (11-11)

Date: 2026-11-11 15:15 (proposal) Source: Proposal #12031

One hour and forty-three minutes after the issue was filed, the same reporter submitted a fix: two files, 84 lines added, 12 lines removed.

The core change adds a dataspace_size field to struct rt_link_record. Previously the structure held only the dataspace pointer, with no indication of how large it was — so during reassembly there was no baseline against which to check whether “offset plus length” went out of bounds. With this field added, the bounds check becomes possible.

Four checks follow: the first frame’s extend.parameter must be non-zero and no greater than RT_LINK_FRAMES_MAX * RT_LINK_MAX_DATA_LENGTH, with both invalid values and allocation failures exiting cleanly via _stop_recv_long(); subsequent frames validate the data_len upper bound, that index is less than the frame count and the total frame count, and that offset + data_len does not exceed dataspace_size using overflow-safe arithmetic, only then calling rt_link_hw_copy(); all release and reset paths clear dataspace_size; and rt_link_frame_check() gains rejection of an over-limit head.length, aligning the receive path with the constraints of rt_link_send().

An adjacent hazard is handled along the way: when the short-frame handler entry or a new long-frame first frame arrives while a previous long frame has not yet finished, _stop_recv_long() is called first, preventing the dataspace from being overwritten or leaked.

III. Seven Patches Overnight: The Audit List Turns to March’s Old Accounts

3.1 GPT and CFI: Two Old Integer-Boundary Problems (11-12)

Date: 2026-11-12 00:04 and 00:49 (proposals) Source: Proposal #12034, Defect #11260, Proposal #12038, Defect #11272

Between 00:04 and 00:56 on November 12, the same reporter submitted seven patches in a row, no longer targeting 2024 CVEs but the issues they themselves had opened between March and September of this year.

#12034 handles the GPT partition table (issue #11260, 2026-03-17): heap memory is allocated as num_partition_entries * sizeof_partition_entry, a product that can wrap on 32-bit, after which the code still iterates ptes[] using the on-disk original entries_nr, causing a heap out-of-bounds read. The fix validates entry size and count before allocation, rejects product overflow, and changes the iteration upper bound to the smaller of “the number of entries actually allocated” and max_partitions.

#12038 handles the CFI flash sector table (issue #11272, 2026-03-19), and this one deserves separate mention because it fills a gap left by the previous fix. The November 5 #11558 (commit e2e507c4) had already changed sect_count > ARRAY_SIZE to >=, but only plugged half the hole: when sect_count exactly equals CFI_FLASH_SECT_MAX, the inner loop breaks but the outer num_erase_regions continues running and repeatedly reports errors, and probing still returns RT_EOK and registers a truncated sector table. This patch logs correctly on a full table and jumps to _fail with -RT_EINVAL to abort probing; it further adds checks for division by zero and unsigned underflow when sect_count == 0, an upper bound on the dev_size shift, an upper bound on the size_ratio multiplication, plus a check of the rt_ioremap() return value and zeroing on the _fail path to prevent a double unmap.

3.2 9PFS and Serial ioctl: Trust Boundaries Between Server and User Space (11-12)

Date: 2026-11-12 00:13 and 00:40 (proposals) Source: Proposal #12035, Defect #11289, Proposal #12037, Defect #11429

#12035 handles 9PFS (issue #11289, 2026-03-23). getdents truncates the 16-bit name_len when writing it into the 8-bit dirp->d_namlen, and only then validates the record boundary—values greater than or equal to 256 wrap around and slip past the check; the copy is also not clamped to the local buffer limit of d_name. In addition, an exaggerated Rread.count is silently truncated rather than rejected, and p9_transaction() lacks basic validation of rx_size and msg_size. The fix moves length validation earlier to the full 16-bit width, caps the copy at the local name buffer, and tightens both the transaction layer and the read path.

#12037 handles the RT-Smart serial ioctl (issue #11429, 2026-06-02). The earlier commit c5ead825 had added user-space marshalling for common commands such as RT_DEVICE_CTRL_CONFIG, but F_SETFL, FIONWRITE, and kernel-internal or unknown commands still pass through the arg_size == 0 branch, handing raw user pointers directly to rt_device_control() and serial->ops->control(). This patch switches to a whitelist: F_SETFL updates fd->flags and returns directly; user space is only allowed the already-marshalled struct commands and scalar commands explicitly known to carry no pointers, with everything else returning -RT_ENOSYS; it also rejects baud_rate == 0 and adds null checks for each command against null pointers and ops->configure.

3.3 tty Null LWP and pin legacy Interrupts (11-12)

Date: 2026-11-12 00:53 and 00:56 (proposals) Source: Proposal #12039, Defect #11717, Proposal #12040, Defect #11784

#12039 handles RT-Smart’s tty (issue #11717, 2026-08-18): when a non-LWP thread reads the tty, tty_wait_background() dereferences pgrp without checking td->lwp == NULL, so the first read(stdin) in finsh can fault. Three files, 23 lines; when lwp is null, the background job wait is skipped entirely.

#12040 handles the GPIO legacy interrupt path (issue #11784, 2026-09-07): when a device has no pin_attach_irq, the old path is taken, and pin_api_attach_irq publishes the hdr/args pair without disabling interrupts, so the ISR may read a half-updated pair. One file, 8 lines; publishing and clearing are moved into a critical section.

Both are on the scale of “one missing check,” and both were validated only by static checks, without board testing.

3.4 The legacy USB Fix Closed by Policy, Resubmitted the Next Day (11-12)

Date: 2026-11-12 00:17 (proposal) Source: Proposal #12036, Defect #12026, Proposal #12027

The previous issue covered this: defect #12026 reported that the legacy USB host stack treats the device-reported bLength as the read length for the full device descriptor, allowing an out-of-bounds write; the maintainer submitted #12027 to add length validation, which was promptly closed with “legacy code, will no longer be continuously fixed or maintained.”

Early this morning, #12036 resubmitted the same issue by a different reporter, with broader coverage than the closed patch: besides the bLength check, it also tightens the hub descriptor length and the upper bounds for writing hub/HID wMaxPacketSize into small buffers—five files, 71 lines added, 13 removed. The proposal text states the background—#11288 (2026-03-23) had already tightened bNumInterfaces, while bLength had long lacked an equivalent check. Issue #12026 remains open.

Within 24 hours, the same code went through “issue, eight-line fix, closed as no longer maintained, broader fix resubmitted.” This is not duplicated effort: the closure statement defines the maintenance scope, while the resubmission defines the cost—five files and 71 added lines to close off a physically reachable attack surface.

IV. BSP and Test Infrastructure

4.1 K230 PWM readback: compensation value mistaken for high-level count (11-11)

Date: 2026-11-11 17:20 (proposal) Source: Proposal #12032

When kd_pwm_set() writes the output compare register, it stores period_count - high_count—the K230’s PWM uses “period minus high level” as the compare value; but kd_pwm_get() treats this register directly as high_count. The result is that the pulse width returned by rt_pwm_get() is exactly the complementary value: set 25%, read back 75%.

The fix is to read pwmcmp0 once as the period count and share it between both conversions; the high-level count is derived back from period_count - compare, and when the shared period is short enough that the other channel’s compare value exceeds pwmcmp0, a 0 pulse width is returned (a legal 0% state), still returning RT_EOK and keeping pulse no greater than period.

The verification is more solid than the average for this batch: on the host, real driver functions were paired with simulated registers, and on the 01Studio CanMV K230 it was tested with GPIO52/PWM4 (pwm1 channel 1). The original five test points still pass, and cross-channel readback goes from 50000 and 4294942296 nanoseconds to 50000 and 0 nanoseconds—the former number is itself evidence of wraparound. The test cases were also extended to 0%, 25%, 75%, and 100% duty cycles, periods requiring a non-zero divider, and scenarios with cross-channel compare value underflow.

4.2 utest multi-round runs start to show the bill: atomic test case leaks 25 KB per round (11-11)

Date: 2026-11-11 19:55 (proposal) Source: Proposal #12033, Proposal #11983

The utest multi-round runs and resource sampling merged on November 9 (#11983) are starting to pay off: they let the same test case run multiple rounds within a single boot and sample free memory and the thread table between rounds, so the test case’s own resource leaks have nowhere to hide.

This commit handles core.atomic. test_atomic_add() creates a new semaphore and three dynamic threads each round, and neither is reclaimed. After a thread exits, its TCB and stack are only released once rt_defunct_execute() runs in the idle thread, but the test case keeps occupying runnable threads, so the idle thread never gets a chance to be scheduled.

Measured data (bsp/qemu-vexpress-a9, multi-round switch set to 3 rounds): when running only core.atomic, used heap goes 63608 → 88848 → 114088 bytes, up 25240 per round; ps shows three residual t1/t2/t3 each, in close state. After the fix, three rounds stay constant at 38368, with zero residuals. The semaphore part is more subtle: when run together with the event and atomic test cases, it grows only 64 bytes per round, exactly the size of one rt_semaphore.

The fix is thirteen lines: delay 1 millisecond before returning so the defunct queue gets processed, then delete the semaphore. This is consistent with the existing wait convention in hooklist_tc.c.

4.3 packages and version lines: still no pushes this issue (11-12)

Date: 2026-11-12 (verification) Source: RT-Thread org repositories, releases

Sorting the 35 organization repositories by push time, only the main repository had a push within the window (the PLIC fix on 11-11 09:44). The last push to packages is still 11-09 10:27, env is stuck at 11-06, sdk at 11-05, and rtthread-nano, smart-build, pkgs-test, toolchains-ci, and rtthread-skills all show no activity. On the package index side, this is the second consecutive issue with zero changes.

There are no new releases or tags; the latest is still v5.3.0 (2026-09-10).

V. Official and Community Activities

5.1 STM32 Seminar Tours Four Cities, RT-Thread Exhibits On-Site (11-11)

Date: 2026-11-11 11:47 (published) Source: EET-China RT-Thread Column, ST Official Event Page

The 2026 STM32 Seminar kicks off this month across four stops: Suzhou on November 12, Beijing on November 19, Chengdu on December 3, and Shenzhen on December 10. According to the ST official event page, the scale is 2 summits plus 5 forums, 8 interactive talks, over 35 technology demos, and 25 partners, with the exhibition area covering robotics, graphics applications, smart door locks, wireless audio, and edge AI.

RT-Thread will exhibit on-site with its latest products. The article also includes a repost giveaway: keep the post on your Moments for at least 24 hours, screenshot it, and send it to Xiaoshidi; 3 winners will each receive a NUCLEO-G031K8 development board. Deadline is December 10.

The first stop, Suzhou, is today. This article was published in RT-Thread’s WeChat official account column, and the official website news API has not yet indexed it—the external communication cadence currently has WeChat leading and the official website lagging.

5.2 Official News API Shows No New Items for the Fourth Consecutive Issue (11-12)

Date: 2026-11-12 (verified) Source: Official News List

The API still returns the same five articles from November 6, with the latest timestamp at 11-06 15:54: elexcon full-stack capabilities, Vinaro industrial PLC salon, Southeast Asian university industry-education integration, Lingang automotive software ecosystem salon, and the CIIF AI car Workshop. This batch was fully covered in the 11-09 issue (covering the 72-hour weekend window), and there were no new additions on 11-10, 11-11, or in this issue.

The official website news page list is client-rendered, so the API is authoritative. Notably, there is a new article on the WeChat side today but none on the API side—the two are not the same publishing pipeline.

5.3 Google News Three-Day Window Yields One Hit (11-11)

Date: 2026-11-11 (verified) Source: Google News

This issue differs from previous ones: RT-Thread when:3d returned one result in the Chinese channel, namely the STM32 Seminar article above. All three results in the English channel were irrelevant (two AMD EPYC processor reviews and one HackerNoon coffee project roundup) and were removed based on title semantics.

睿擎派 OR 睿擎 when:7d returned zero hits; RT-Thread 嵌入式 when:7d likewise returned only this one article. The bare query RT-Thread returned 42 results, with the latest still being this article, followed by the November 3 Wuzhen Summit piece—the channel is functioning normally, and there really was only this one new article within the three-day window.

All seven candidates from HN Algolia were literal “thread” false matches (agent session sync tools, Slack thread management, Apple Support Community shutdown, free-threaded Python discussion), unrelated to RT-Thread, and were all removed.

VI. Summary and Trend Observations

6.1 From CVEs to In-House Issues: Audit Enters Its Second Phase

Date: 2026-11-12 Source: Pull request list

The eleven PRs covered in the previous issue were tied to 2024 CVEs, with corresponding issue numbers between #8283 and #8291. The seven PRs filed in the early hours of today use a different set of numbers: #11260, #11272, and #11289 date from March 2026, #11429 from June, #11717 from August, #11784 from September, and #12026 from yesterday.

This shows that the audit is not a matter of “finish the CVE list and call it done,” but rather follows a longer list. For the project maintainers, the pressure on the review queue has not eased as the CVEs are closed out: ten new PRs arrived during the window, all requiring manual review, and most touch deep subsystem semantics—GPT tables, the 9P protocol, VFS, and serial ioctl—changes that cannot be judged at a glance.

6.2 Quality Tiers of Fixes: Board-Tested vs. Static-Check Only

Date: 2026-11-12 Source: PR #12032, PR #12033

Among the ten new PRs in this issue, the verification sections fall into two categories. The K230 PWM PR ran on real hardware and provided specific before-and-after readings; the utest atomic test case PR provided three rounds of memory and thread-count sampling, and further distinguished leak amounts between “running only core.atomic” and “running three test cases consecutively.”

The other seven security PRs uniformly state “statically confirmed, not board-tested, no PoC constructed,” with two explicitly noting “no malicious 9P server set up” and “no malicious CFI geometry constructed.” This honesty is worth noting, but it also means these fixes currently rest only on compilation and code review. The practical implication for users is that they should not be treated as verified protections before they land in a release.

6.3 Review Queue: 158 PRs, 365 Issues

Date: 2026-11-12 (verification) Source: Pull request list, Issue list

Compared with the previous issue’s 149 PRs and 365 issues, PRs increased by nine and issues remained flat. Breakdown: one merge during the window (#12011), ten new PRs, one new issue (#12030), and one issue closed (#12010).

None of the PRs called out in the previous issue have merge records this period, though the CAN driver refactor (#11779) was updated during the window. The tty ioctl user-pointer validation thread is now carried by #12037, but that too is still under review—the same topic has gone from an issue to a PR without moving much further up the queue.

6.4 Tension Between Maintenance Boundaries and Willingness to Fix

Date: 2026-11-12 Source: PR #12036

The legacy USB code went through four steps within 24 hours: issue filed, eight-line fix, closed as no-longer-maintained, and a broader fix resubmitted. The closure statement set out the maintenance scope; the resubmission set out the cost—both coexisted on the same issue, and neither side convinced the other.

An RTOS that has been around for twenty years inevitably carries legacy components, and declaring them no longer maintained is a reasonable resource decision. But for products still using the old RT_USING_USB_HOST stack, that declaration amounts to transferring security responsibility to the integrator. If the project could list “no longer maintained but still in-tree” components, it would be more useful than having that information scattered across individual PR comments—a point raised in the previous issue and validated once again this period.

Appendix: Source Verification Table

Source Verification Result
RT-Thread official news API Five items, latest 11-06 15:54, fourth consecutive issue with zero new items
Official news page List is client-rendered; API is authoritative and consistent with it
GitHub main repo commits One commit in window (PLIC completion ordering), checked by committer time
GitHub PRs and Issues 1 merged, 10 new proposals, 1 new issue, 1 issue closed; 158 proposals and 365 issues pending review
RT-Thread org repositories 35 repositories checked sorted by push; only the main repo had pushes in the window
packages repository No pushes; last at 11-09 10:27
releases and tags No additions; latest remains v5.3.0 (09-10)
Google News Chinese One hit in the three-day window (STM32 seminar); bare query returned 42 results, confirming the channel is working
Google News English Three results, all irrelevant, removed
HN Algolia All seven candidates were literal false matches on “thread”, removed
EET-China RT-Thread column Site unreachable both directly and via proxy; body text taken from the page content returned by the search service
club.rt-thread.org forum Returned WAF block page (HTTP 468), could not be verified

Full Source List