fix(kernel): stabilize apt update under sustained I/O by fslongjin · Pull Request #2157 · DragonOS-Community/DragonOS

fslongjin · 2026-07-23T08:44:37Z

Summary

replace duplicate-prone NAPI scheduling with an exact-once SCHED/MISSED state machine and bounded, fair polling
move virtio-net to Q64 raw queues with preallocated DMA slots, asynchronous TX completion, and race-closing EVENT_IDX notification handling
prevent the same task from running or being enqueued on two CPUs during remote wakeups, stop/resume, affinity changes, and migration tails
batch ext4 sequential range allocation and orphan extent reclamation while preserving journal credits, checksums, and initialized-before-published block ordering
avoid read-before-write I/O for complete ext4 blocks and handle disappearing proc fd tables without panicking
pin the merged virtio-drivers queue fixes from the DragonOS mirror

Root cause

The observed apt update stalls were produced by several interacting bottlenecks and correctness gaps:

virtio-net used two-descriptor RX and TX rings, while each TX packet synchronously waited for completion
NAPI could enqueue the same instance more than once and did not close the callback-enable race
remote network wakeups could enqueue a task before its previous CPU completed switch-out
apt package-store writes caused ext4 to allocate and journal missing blocks one at a time, and interrupted-download cleanup committed each tail extent separately
proc fd traversal could race process exit and unwrap an already removed descriptor table

Together these issues amplified interrupt, scheduler, and synchronous metadata costs into low throughput, long cleanup delays, and occasional kernel instability.

Implementation

Network data path

add a host-testable atomic NAPI transition crate
use FIFO scheduling with packet, poll-call, and elapsed-time budgets
keep RX/TX callback state paired with NAPI ownership
preallocate page-backed RX/TX DMA slots and recycle completions asynchronously
fail-stop on queue-accounting or malformed completion invariants

Scheduler

track actual CPU execution ownership independently from runqueue placement
wait for switch-out completion before remote enqueue
serialize stop wakeups and affinity placement under the scheduler placement locks
revalidate pending migration state, affinity, and runqueue ownership in the switch tail

Filesystem and procfs

allocate contiguous append ranges in one ext4 transaction
support journaled appends to the right-most external extent leaf
reserve transaction credits before batched tail reclamation
batch same-group orphan extent cleanup
write complete blocks without a preceding read
treat missing fd/fdinfo tables as normal process-exit races

Validation

cargo test -p napi-state: 7 passed
cargo test -p another_ext4: 131 passed
virtio-drivers: 27 unit tests and 8 doctests passed before pinning the merged revision
make kernel: passed against the pinned mirror revision
QEMU fresh apt update: completed a 31.6 MB refresh without a hash mismatch
interrupted partial download: resumed from the previous offset
fragmented partial-file cleanup: reduced from minutes to a fraction of a second

Remaining scope

This change removes the confirmed correctness failures and synchronous-I/O amplification. Full Linux-style ext4 delayed allocation and writeback-time allocation remain separate future performance work.

Rework NAPI scheduling around an exact-once SCHED/MISSED state machine with bounded, fair polling. Move virtio-net to deeper raw queues with preallocated DMA buffers, asynchronous TX completion, and race-closing EVENT_IDX callback handling. Track actual cross-CPU task ownership independently from runqueue placement and serialize stop wakeups, affinity changes, and migration tails so a task cannot execute or enqueue twice during network wakeups. Batch ext4 sequential range allocation and orphan extent reclamation, preserve journal credit and checksum invariants, and avoid read-before-write for complete blocks. This prevents apt package-store writes and interrupted-download cleanup from synchronously amplifying metadata I/O. Treat disappearing proc fd and fdinfo tables as normal exit races, refresh the pinned virtio-drivers revision, and add host-testable NAPI transition coverage. Validated with cargo test -p napi-state, cargo test -p another_ext4, and make kernel. Signed-off-by: longjin <longjin@dragonos.org>

fslongjin · 2026-07-23T08:44:56Z

@codex review

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 732d525c08

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

Refresh whole-disk capacity from the backing block device so loop devices expose their configured size after LOOP_SET_FD. Keep the static start LBA on I/O hot paths, safely fall back after device teardown, and report only complete loop sectors. Revalidate cached proc fd and fdinfo entries against the live descriptor table. Return ENOENT across zombie and reaped-task paths, including reads from an already-open fdinfo file, matching Linux 6.6 semantics. Add a dunitest covering cached fd entries across process exit. Signed-off-by: longjin <longjin@dragonos.org>

fslongjin · 2026-07-23T11:10:40Z

@codex review

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: c285a71b9c

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

Consume completed RX buffers independently of TX queue availability so NAPI can continue making receive progress under transmit pressure. Use a lazy response token for receive-side replies and reserve DMA capacity only when smoltcp actually emits a frame. Keep explicitly reserved tokens for standalone transmit calls and split RX/TX token types to preserve their ownership invariants. Signed-off-by: longjin <longjin@dragonos.org>

fslongjin · 2026-07-23T11:43:07Z

@codex review

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: d65ff73d07

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

Keep the smoltcp Ethernet frame limit separate from the user-visible IP MTU so rtnetlink reports 1500 bytes instead of 1514. Run the existing DHCP acquisition window in a kernel worker, allowing SystemState::Running and userspace startup to proceed when no DHCP server is available. Add a strict rtnetlink MTU regression and teach dunitest to parse singular GoogleTest summaries so the one-case suite is enforced by CI. Signed-off-by: longjin <longjin@dragonos.org>

fslongjin · 2026-07-23T15:12:34Z

@codex review

chatgpt-codex-connector · 2026-07-23T15:22:23Z

Codex Review: Didn't find any major issues. Chef's kiss.

Reviewed commit: 3ba7c4babc

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

Signed-off-by: longjin <longjin@dragonos.org>

fslongjin · 2026-07-23T18:50:48Z

Root cause confirmed: FAT cluster allocation and release were not serialized across different inodes. Concurrent page-cache writeback could select the same free FAT entry, corrupt a cluster chain, and eventually make FATFile::ensure_len() reach an impossible missing cluster.

The fix adds a filesystem-wide FAT metadata lock, aligned with Linux 6.6 msdos_sb_info::fat_lock, around allocation/free chain transitions. Data-cluster zeroing remains outside that lock after the cluster is reserved and linked. A new Dunitest concurrently grows and fsyncs eight independent FAT files, then verifies every block; it also checks the filesystem type, capacity, cleanup paths, and interrupted/short reads.

Validation: make kernel, cross-build of the new Dunitest, git diff --check, and an 8-role adversarial review all passed.

fslongjin · 2026-07-23T18:50:59Z

@codex review

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 36c3be8b74

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

Signed-off-by: longjin <longjin@dragonos.org>

fslongjin · 2026-07-24T04:22:48Z

@codex review

chatgpt-codex-connector · 2026-07-24T04:33:42Z

Codex Review: Didn't find any major issues. 🎉

Reviewed commit: 5e3693d178

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

github-actions Bot added the Bug fix A bug is fixed in this pull request label Jul 23, 2026

chatgpt-codex-connector Bot reviewed Jul 23, 2026

View reviewed changes

Comment thread kernel/src/filesystem/procfs/pid/fd.rs Outdated

Comment thread kernel/src/filesystem/procfs/pid/fdinfo.rs Outdated

chatgpt-codex-connector Bot reviewed Jul 23, 2026

View reviewed changes

Comment thread kernel/src/driver/net/virtio_net.rs Outdated

chatgpt-codex-connector Bot reviewed Jul 23, 2026

View reviewed changes

Comment thread kernel/src/driver/net/virtio_net.rs Outdated

Comment thread kernel/src/net/net_core.rs

fslongjin added 2 commits July 23, 2026 17:41

Merge origin/master into codex/fix-apt-update-data-path

cd0aad1

Signed-off-by: longjin <longjin@dragonos.org>

fix(fat): serialize concurrent cluster allocation

36c3be8

Signed-off-by: longjin <longjin@dragonos.org>

github-actions Bot added the test Unitest/User space test label Jul 23, 2026

chatgpt-codex-connector Bot reviewed Jul 23, 2026

View reviewed changes

Comment thread kernel/crates/another_ext4/src/ext4/alloc.rs Outdated

fix(ext4): bound transactional range probes

5e3693d

Signed-off-by: longjin <longjin@dragonos.org>

fslongjin merged commit e374b45 into DragonOS-Community:master Jul 24, 2026
15 of 16 checks passed

Uh oh!

Uh oh!

Conversation

fslongjin commented Jul 23, 2026

Summary

Root cause

Implementation

Network data path

Scheduler

Filesystem and procfs

Validation

Remaining scope

Uh oh!

fslongjin commented Jul 23, 2026

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

Uh oh!

fslongjin commented Jul 23, 2026

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

fslongjin commented Jul 23, 2026

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

Uh oh!

fslongjin commented Jul 23, 2026

Uh oh!

chatgpt-codex-connector Bot commented Jul 23, 2026

Uh oh!

fslongjin commented Jul 23, 2026

Uh oh!

fslongjin commented Jul 23, 2026

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

fslongjin commented Jul 24, 2026

Uh oh!

chatgpt-codex-connector Bot commented Jul 24, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant