The performance difference comes from the OS abstraction layer we plan to implement. Here's the design:
The profile code will sit in libc, but the system API isolation layer beneath it will provide two backends: a POSIX-compatible backend and a kernel-native backend. When building in flat mode, the isolation layer switches to the kernel API path — direct calls into the kernel without going through the POSIX syscall layer. For low-performance IoT chips, eliminating that overhead on every timer operation, mutex acquisition, and event delivery is a significant win. If we were to place the profiles in apps/ instead, we'd be in an awkward position: the code in apps/ would need to call kernel APIs directly in flat mode to achieve the same performance, which breaks the convention that apps/ code should be pure POSIX. The architectural boundary should be clean — kernel APIs belong in the kernel or libc, and apps/ stays POSIX-only. Putting the profile code in libc keeps that boundary intact while still allowing the flat-mode optimization path. Xiang Xiao <[email protected]> 于2026年10月16日周五 16:12写道： > On Fri, Oct 16, 2026 at 3:49 PM chao an <[email protected]> wrote: > > > So the real reason you've been objecting to my proposal is that your team > > has been working on something similar, correct? > > > > If you can give me a clear timeline for when your implementation will be > > submitted upstream, I'm happy to withdraw my PR and let your code serve > as > > the baseline. I'm not here to fight over ownership of the Bluetooth > stack. > > My only motivation is that NuttX's lack of a capable Bluetooth stack is > > driving every developer toward Zephyr — and as a PMC member, I think you > > should be asking yourselves why this gap has remained unfilled for so > long. > > > > > I object to landing the whole bluetooth stack into kernel space directly > either yours and ours . > This is the same feeling from raiden00pl too, let's see other people's > opinions here. > But it's fine for me to host the whole code in nuttx-apps.git and refactor > components to kernel > space step by step. > > > > As for the current PR, I'll remove the shell components from the stack > and > > move them into apps/ where they belong. The base profiles (A2DP, BLE) > will > > stay in libc — > > > The architecture is wrong to host any bluetooth profile in libc. > > in flat build mode this gives a significant performance > > boost to the overall stack, which is the primary deployment model for > most > > NuttX devices. > > > > > Performance should remain the same whether the base profile is hosted in > libc or apps, > could you point out any technical reason which impacts the performance in > the flat mode? > > > > A few additional data points regarding our timeline: our product is > > scheduled for release in December, and Bluetooth SIG certification is > > already in progress — we expect to receive it around November. If the > > baseline were to switch to your team's stack with its different > > architecture, that certification cycle would need to be repeated, which > is > > not trivial given the time and cost involved. > > > > Xiang Xiao <[email protected]> 于2026年10月16日周五 15:08写道： > > > > > On Fri, Oct 16, 2026 at 2:44 PM chao an <[email protected]> wrote: > > > > > > > Hi raiden00pl, > > > > > > > > Regarding your point about the NimBLE port using the POSIX API — yes, > > it > > > > uses POSIX, but there are real issues with the current implementation > > > that > > > > go beyond just kernel/protected mode support. > > > > 1. ble_npl_hw_enter_critical is not a real critical section > > > > > > > > Look at the NuttX port's implementation: > > > > > > > > > > > > > > https://github.com/apache/mynewt-nimble/blob/master/porting/npl/nuttx/src/os_atomic.c > > > > > > > > static pthread_mutex_t s_mutex = > > PTHREAD_RECURSIVE_MUTEX_INITIALIZER_NP; > > > > > > > > uint32_t ble_npl_hw_enter_critical(void) > > > > { > > > > pthread_mutex_lock(&s_mutex); > > > > return 0; > > > > } > > > > > > > > This is a mutex, not a critical section. It does not disable > interrupts > > > or > > > > preemption — an ISR can still fire and corrupt data that the > "critical > > > > section" was supposed to protect. On top of that, using a mutex opens > > the > > > > door to ABBA deadlocks when combined with any other lock held inside > > the > > > > critical region: > > > > > > > > - Thread A: enter_critical → lock(s_mutex) → try lock(mutex M) ← > > > blocked > > > > by B > > > > - Thread B: lock(M) → enter_critical → lock(s_mutex) ← blocked by > A > > > > > > > > Compare this with the FreeRTOS port, which correctly uses > > > > vPortEnterCritical() / vPortExitCritical() (i.e., actually disabling > > > > scheduling/preemption). The current NuttX port simply does not meet > the > > > > semantic contract that the NimBLE stack expects from > > > > ble_npl_hw_enter_critical. > > > > 2. Time subsystem uses clock_gettime instead of kernel tick APIs > > > > > > > > > > > > > > > > > > https://github.com/apache/mynewt-nimble/blob/master/porting/npl/nuttx/src/os_time.c > > > > > > > > ble_npl_time_get() is backed by clock_gettime(CLOCK_MONOTONIC, ...), > > > which > > > > involves a syscall and floating-point arithmetic on every call. In a > > BLE > > > > stack where timestamps and timing are on the hot path, this adds > > > measurable > > > > overhead compared to accessing the kernel tick counter directly. > > > > 3. Others are already using the kernel approach — and for good reason > > > > > > > > A concrete example: Xiaomi's Bluetooth stack (zblue) for > openvela/NuttX > > > > uses kernel APIs directly, bypassing POSIX for the > performance-critical > > > > paths: > > > > > > https://github.com/sprhawk/zblue-port-openvela/tree/yaotepai/port/kernel > > > > > > > > This isn't an academic exercise — it's a production choice. They > > > recognized > > > > that a BLE stack needs tight integration with the OS timer subsystem, > > > > direct access to critical section primitives, and efficient event > > > delivery, > > > > none of which POSIX provides well in this context. > > > > My question > > > > > > > > If other organizations are already using a more performant, > > > architecturally > > > > correct kernel-level integration for their BLE stacks on NuttX, why > > > should > > > > we be told that we must use the slower, less correct POSIX path? > > NuttX's > > > > POSIX compliance is valuable, but it shouldn't come at the cost of > > > shipping > > > > a BLE stack with a broken critical section primitive. There should be > > > room > > > > for both approaches, and the decision should be driven by technical > > > merit, > > > > not by treating POSIX as the only acceptable answer. > > > > > > > > > > > I don't object to moving the key bluetooth stack to kernel space like > how > > > the current bluetooth stack is done, > > > what I object to is that the current state of > > > https://github.com/apache/nuttx/pull/20686 is far to move entirely > > > into kernel space. > > > If I low my quality bar, I have requested my teammate to upstream our > > > implementation: > > > > https://github.com/sprhawk/zblue-port-openvela/tree/yaotepai/port/kernel > > > since much work needs to be done before it can achieve the quality > > > upstream. > > > BTW, many zephyr class BT features are developed by my team and > upstream > > to > > > Zeprhy community. > > > If considering the quality, our version is much better since: > > > > > > 1. Product on smart speaker last year > > > 2. Pass Bluetooth SIG BT/BLE 6.2 certification at 2026.06 > > > 3. Product on wear device by end of this year > > > > > >
