Improve packet validation, receive buffer handling, and processing efficiency by AlanJAS · Pull Request #111 · PlatformLab/HomaModule

AlanJAS · 2026-10-03T21:29:33Z

This PR fixes several correctness issues and reduces repeated work in transmission, reception, qdisc, and trace analysis.

Validate GRANT/RESEND priorities, normalize configured priority counts, and check GRO headers before accessing packet fields.
Handle GSO segmentation failures and roll back partially registered offload callbacks.
Bound partial receive copies and add a dedicated buffer-release ioctl so releasing buffers cannot dequeue another message.
Cache transmit fragment positions, grow fragment arrays geometrically, and stop ordered gap scans when no overlap is possible.
Avoid redundant queue locking in qdisc and free detached packets after releasing the shared lock.
Replace repeated front removals in tthoma.py with deque operations.
Correct per-CPU GRANT aggregation and elapsed-time reporting in ttmerge.py.

Validation: All 851 kernel-mock unit tests pass with ASan using local compatibility adaptations for the available headers. Differential checks preserve fragment lookup, packet accounting, and trace-analysis results.
Isolated CPU microbenchmarks show approximately 30× faster sequential fragment lookup and 4.9× faster fragment-array allocation for 1 MB messages with 4 KiB pages. Small-message paths show minor additional overhead. These measurements use mocked kernel dependencies; end-to-end throughput and RPC latency have not been measured.

Validate wire priorities before mutating RPC state or indexing priority_map. Add regression tests for priorities 8 and 255.

Copy only the requested bytes across discontiguous buffer pages. Avoid unsigned offset wraparound in copy_out, get, and contiguous.

Add HOMAIOCRELEASE using the existing receive-argument layout. Return unreleased tokens after a partial failure and wake buffer waiters. Use it in receiver::release and document the required module support.

Preserve skb_segment error pointers and handle an empty result before accessing IP headers. Add allocation-failure injection and a regression test.

Stop on linearization failure and check common and type-specific lengths. Share the existing header-length table with GRO. Add regression tests for truncated headers, an invalid type, and a failed pull.

Extend the existing upper clamp with a lower bound of one before walking the cutoff array. Cover zero, minus one, and INT_MIN in regression tests.

Do not attempt IPv6 registration if IPv4 fails. Unregister IPv4 if IPv6 fails, so a failed module load cannot leave callbacks registered. Add fault injection, registration tracking, and retry regression tests.

Cache the starting fragment and its byte offset under the RPC lock. Sequential sends avoid repeated prefix scans; earlier retransmissions restart at zero. Add reuse and backward-offset regression coverage.

Avoid repeated prefix copies when high-order allocation falls back to small pages. If the larger allocation fails, retry the original minimum capacity before reporting ENOMEM. Cover fallback growth and retry.

Use the existing sorted gap invariant to reject early duplicates after the first nonoverlapping gap. Preserve duplicate handling and metrics. Add a regression test that checks only one gap is examined.

Use deques for GRO handoffs and sorted clock synchronization samples to make front removals O(1), preserving FIFO order and analysis results.

Use __skb_dequeue under the existing qdev lock to avoid redundant queue locking. Detach deferred packets while holding the lock, then free them after releasing it to reduce contention.

johnousterhout · 2026-10-05T15:50:47Z

Thanks very much for this PR. I'm pretty backlogged right now so it may be a couple of weeks before I get to this, but I wanted to let you know that it's now on my radar.

AlanJAS added 14 commits October 3, 2026 02:15

Reject invalid GRANT and RESEND priorities

54ca67b

Validate wire priorities before mutating RPC state or indexing priority_map. Add regression tests for priorities 8 and 255.

Bound partial copies and guard offset arithmetic

915f394

Copy only the requested bytes across discontiguous buffer pages. Avoid unsigned offset wraparound in copy_out, get, and contiguous.

Release receive buffers without dequeuing a message

0978062

Add HOMAIOCRELEASE using the existing receive-argument layout. Return unreleased tokens after a partial failure and wake buffer waiters. Use it in receiver::release and document the required module support.

Propagate GSO segmentation errors before rereferencing

9887b5d

Preserve skb_segment error pointers and handle an empty result before accessing IP headers. Add allocation-failure injection and a regression test.

Validate GRO headers before accessing packet fields

60b00ae

Stop on linearization failure and check common and type-specific lengths. Share the existing header-length table with GRO. Add regression tests for truncated headers, an invalid type, and a failed pull.

Normalize priority counts to the supported range

68052c2

Extend the existing upper clamp with a lower bound of one before walking the cutoff array. Cover zero, minus one, and INT_MIN in regression tests.

Roll back partially registered offload callbacks

1b4a677

Do not attempt IPv6 registration if IPv4 fails. Unregister IPv4 if IPv6 fails, so a failed module load cannot leave callbacks registered. Add fault injection, registration tracking, and retry regression tests.

Reuse transmit fragment position across sequential skbs

32561fc

Cache the starting fragment and its byte offset under the RPC lock. Sequential sends avoid repeated prefix scans; earlier retransmissions restart at zero. Add reuse and backward-offset regression coverage.

Grow transmit fragment arrays geometrically

3f3ca69

Avoid repeated prefix copies when high-order allocation falls back to small pages. If the larger allocation fails, retry the original minimum capacity before reporting ENOMEM. Cover fallback growth and retry.

Stop ordered gap scans when no overlap is possible

ac25f60

Use the existing sorted gap invariant to reject early duplicates after the first nonoverlapping gap. Preserve duplicate handling and metrics. Add a regression test that checks only one gap is examined.

sum the GRANT metrics for each CPU

9a54e19

Correctly calculate the intervals in ttmerge.py

6a2a185

Replace repeated pop(0) calls in tthoma.py with deque.popleft()

ead06f5

Use deques for GRO handoffs and sorted clock synchronization samples to make front removals O(1), preserving FIFO order and analysis results.

Optimize qdisc dequeue and RPC cleanup

f446bc9

Use __skb_dequeue under the existing qdev lock to avoid redundant queue locking. Detach deferred packets while holding the lock, then free them after releasing it to reduce contention.

Navigation Menu

Repository navigation

Improve packet validation, receive buffer handling, and processing efficiency - #111

Improve packet validation, receive buffer handling, and processing efficiency#111
AlanJAS wants to merge 14 commits into
PlatformLab:mainfrom
AlanJAS:main

AlanJAS commented Oct 3, 2026

Uh oh!

johnousterhout commented Oct 5, 2026 •

edited

Loading

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

AlanJAS commented Oct 3, 2026

Uh oh!

johnousterhout commented Oct 5, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

johnousterhout commented Oct 5, 2026 •

edited

Loading