Branch: refs/heads/misc-bugfixes Home: https://github.com/kronosnet/kronosnet Commit: 76dd6e1cbe8dfd46ba6957750718bc0cf4a6b289 https://github.com/kronosnet/kronosnet/commit/76dd6e1cbe8dfd46ba6957750718bc... Author: Fabio M. Di Nitto fdinitto@redhat.com Date: 2026-05-22 (Fri, 22 May 2026)
Changed paths: M libknet/host.c
Log Message: ----------- libknet: fix defragmentation buffer reclamation logic
Fix two bugs in the defragmentation buffer reclamation logic that prevent efficient memory management:
1. Window calculation was using the old received sequence number (dst_seq_num) instead of the current packet's sequence number (seq_num) to calculate the valid buffer window. This caused buffers to be reclaimed based on stale sequence information, potentially freeing buffers that should still be valid or keeping buffers that should be reclaimed.
2. Window size calculation incorrectly used defrag_bufs_max (configuration limit, default 1024) instead of allocated_defrag_bufs (actual allocated count, typically 32) when allocated_defrag_bufs < defrag_bufs_max. This created an excessively large window (1025 sequence numbers for 32 buffer slots), causing integer wraparound in the tail calculation and preventing proper buffer reclamation.
Example: With seq_num=150, allocated_defrag_bufs=32, defrag_bufs_max=1024: - Buggy: tail = 150 - 1025 = -875 → wraps to 64661 (prevents reclamation) - Fixed: tail = 150 - 33 = 117 (correct reclamation window)
The window size must always match the actual allocated buffer capacity to ensure proper reclamation of stale defragmentation buffers.
Signed-off-by: Fabio M. Di Nitto fdinitto@redhat.com Co-Authored-By: Claude Sonnet 4.5 noreply@anthropic.com
Commit: 81c8dfbba77e9abe16aaac0b3e5958a2a00dce88 https://github.com/kronosnet/kronosnet/commit/81c8dfbba77e9abe16aaac0b3e5958... Author: Fabio M. Di Nitto fdinitto@redhat.com Date: 2026-05-22 (Fri, 22 May 2026)
Changed paths: M libknet/host.c
Log Message: ----------- libknet: fix sequence number wraparound calculation in defragmentation
The distance calculation between sequence numbers was incorrect when wraparound occurred. The formula was backwards: it subtracted in the wrong direction and didn't account for the +1 needed for modular arithmetic.
This caused incorrect buffer reclamation decisions near the SEQ_MAX boundary.
Practical examples (SEQ_MAX = 65535):
Example 1 - Normal case (no wraparound): Last packet: dst_seq_num = 1000 New packet: seq_num = 1005 Expected distance: 5
BEFORE (wrong): seq_dist = dst_seq_num - seq_num = 1000 - 1005 = -5 (negative!)
AFTER (correct): seq_dist = seq_num - dst_seq_num = 1005 - 1000 = 5
Example 2 - Wraparound case: Last packet: dst_seq_num = 65534 New packet: seq_num = 3 (wrapped around) Expected distance: 5 (65534→65535→0→1→2→3)
BEFORE (wrong): seq_dist = (SEQ_MAX - seq_num) + dst_seq_num seq_dist = (65535 - 3) + 65534 = 131066 (huge wrong number!)
AFTER (correct): seq_dist = (SEQ_MAX - dst_seq_num) + seq_num + 1 seq_dist = (65535 - 65534) + 3 + 1 = 5
The +1 accounts for the transition from 65535→0 being one step, not zero.
Verification that circular buffer cleaning is not broken:
The seq_dist value is used to determine whether the new packet is: a) Within the circular buffer window (seq_dist < KNET_CBUFFER_SIZE) b) Far enough to require full buffer clear (seq_dist > threshold) c) Should trigger incremental cleaning (fall through case)
Test case 1 - Normal sequential packet: dst_seq_num = 1000, seq_num = 1005, expected distance = 5
BEFORE: seq_dist = -5 (unsigned overflow ~65530) → Incorrectly clears entire buffer for normal sequential packets!
AFTER: seq_dist = 5 → Correctly identifies packet as within buffer window, no clearing needed
Test case 2 - Wraparound (close distance): dst_seq_num = 65534, seq_num = 3, expected distance = 5
BEFORE: seq_dist = 131066 → Falls through to circular buffer cleaning code incorrectly
AFTER: seq_dist = 5 → Correctly identifies packet as within buffer window
Test case 3 - Large jump requiring buffer clear: dst_seq_num = 1000, seq_num = 50000, expected distance = 49000
BEFORE: seq_dist = -49000 (unsigned ~16536) → Clears buffer (correct by accident)
AFTER: seq_dist = 49000 → Clears buffer (correct by design)
The circular buffer cleaning code (lines 673-684) uses seq_num and dst_seq_num directly via modulo operations to find buffer positions. It does not use seq_dist for position calculations, only for the threshold check to determine whether to run. The fix corrects the threshold logic without affecting the position calculations.
Signed-off-by: Fabio M. Di Nitto fdinitto@redhat.com Co-Authored-By: Claude Sonnet 4.5 noreply@anthropic.com
Commit: 81162ed5d2ff6ebb8a944584f0cd2bb986772e6f https://github.com/kronosnet/kronosnet/commit/81162ed5d2ff6ebb8a944584f0cd2b... Author: Fabio M. Di Nitto fdinitto@redhat.com Date: 2026-05-22 (Fri, 22 May 2026)
Changed paths: M libknet/tests/Makefile.am A libknet/tests/int_seq_wraparound_stress.c
Log Message: ----------- [tests] add comprehensive sequence number wraparound stress test
Add new int_seq_wraparound_stress test that exercises sequence number wraparound and defragmentation buffer management with realistic packet loss scenarios.
Test scenarios:
1. Normal sequential packets with occasional loss 2. Wraparound boundary (65535->0) with packet loss 3. Large sequence jumps (>KNET_CBUFFER_SIZE) triggering buffer clearing 4. Out-of-order fragment delivery within same sequence number 5. Out-of-order complete packet delivery (different sequence numbers) 6. Extreme packet loss beyond receive window 7. Wraparound with extreme loss (>KNET_CBUFFER_SIZE gap) 8. Wraparound stress with multiple cycles and duplicate detection 9. Fragment corruption across wraparound (historical bug regression test) 10. Defrag buffer reclamation window calculation validation
Signed-off-by: Fabio M. Di Nitto fabbione@kronosnet.org Co-Authored-By: Claude Sonnet 4.5 noreply@anthropic.com
Commit: a541323887eb5d5c1a335ffb348b52b80441bf2f https://github.com/kronosnet/kronosnet/commit/a541323887eb5d5c1a335ffb348b52... Author: Fabio M. Di Nitto fdinitto@redhat.com Date: 2026-05-22 (Fri, 22 May 2026)
Changed paths: M libknet/tests/Makefile.am A libknet/tests/int_defrag_edge_cases.c
Log Message: ----------- [tests] add defragmentation edge cases test suite
Add new int_defrag_edge_cases_test that validates critical defragmentation buffer management scenarios not covered by existing tests.
Test scenarios:
1. Last fragment arriving first - validates special buffer positioning at KNET_MAX_PACKET_SIZE - len offset, last_first flag handling, and fragment relocation when packet completes. Tests asymmetric MTU with fragments in order [3/3, 1/3, 2/3] and verifies data integrity across all fragments (280 bytes total with distinct payloads 'A', 'B', 'C').
2. Buffer exhaustion and reclamation - validates window-based reclamation when all 32 defrag buffers are filled with incomplete packets. Sends new complete packet with distant sequence number (9000 vs 2000-2031) triggering reclamation of all out-of-window buffers. Verifies data integrity with no corruption from old buffer data in new packet.
3. Fragment data overwrite protection - validates correct assembly of large fragmented packets without buffer overflow. Test case 1 sends 100 fragments × 500 bytes (50,000 bytes total), test case 2 sends 50 fragments × 1,300 bytes (65,000 bytes approaching KNET_MAX_PACKET_SIZE). Each fragment filled with its index number and verified on assembly.
Signed-off-by: Fabio M. Di Nitto fabbione@kronosnet.org Co-Authored-By: Claude Sonnet 4.5 noreply@anthropic.com
Compare: https://github.com/kronosnet/kronosnet/compare/a63f9768576d...a541323887eb
To unsubscribe from these emails, change your notification settings at https://github.com/kronosnet/kronosnet/settings/notifications