[mickeyl]

Stacked on #448. Keep this draft until that correctness fix lands, then rebase. The lifecycle-only changes can be reviewed here: mickeyl/esp-nimble-cpp@f47ba10...7e5f086

Concurrent writers, early credit completions and disconnect/deinit can leave send waiters or callback work referring to an old connection. Serialize whole writes, submit native operations on the host, retain completion state across early notifications, abort incomplete writes and drain pending work before channel reuse/destruction. Reclaim the shared callback worker/queue after the last channel, and delete the CoC server before the native NPL runtime is released. (The early return on a failed nimble_port_stop() that this PR originally added is already on master.)

The pool now allocates configurable full-SDU buffers, including mbuf and packet-header headroom. This changes memory use: five buffers by default, rather than subdividing about three MTUs into 250-byte blocks.

Observable callback contract:

• onConnect/onRead remain on the host by default; optional read offload uses a shared worker.
• onDisconnect always runs on that worker, after deferred reads.
• Blocking writes belong on application workers. Owners stop acceptance, disconnect, wait for isIdle and drain their own workers before deleting a channel; no deletion from its callback.
• Channel construction/destruction belongs to one application management task.

Validation (local, not part of this PR):

• Rebased onto the cleaned-up fix(l2cap): propagate send failures without replaying partial SDUs #448 and current master; tests and CI removed as requested on build: support ESP32-S31 and servers without broadcaster role #447/fix(l2cap): propagate send failures without replaying partial SDUs #448, to return in the separate test-suite PR.
• 16 native concurrency/error/ownership/pool/lifecycle tests pass, also with AddressSanitizer/UBSan and ThreadSanitizer.
• CoC-enabled compile/link probes pass on ESP32-S3 with ESP-IDF 5.5 and 6.1, both inline and deferred callback modes, without library warnings. These are compile/link checks, not RF tests.

The host-task policy overlaps conceptually with #419 and should be coordinated during review. HCI resource windows, private header diagnostics, MPS alignment and the standalone measurement harness are outside this PR. Earlier hardware measurements informed the design; they are not presented as RF acceptance of every extracted commit.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

Share

• X
• Mastodon
• Reddit
• LinkedIn