This bug is awaiting verification that the linux-gcp-tcpx/5.15.0-1003.3 kernel in -proposed solves the problem. Please test the kernel and update this bug with the results. If the problem is solved, change the tag 'verification-needed-focal-linux-gcp-tcpx' to 'verification-done- focal-linux-gcp-tcpx'. If the problem still exists, change the tag 'verification-needed-focal-linux-gcp-tcpx' to 'verification-failed- focal-linux-gcp-tcpx'.
If verification is not done by 5 working days from today, this fix will be dropped from the source code, and this bug will be closed. See https://wiki.ubuntu.com/Testing/EnableProposed for documentation how to enable and use -proposed. Thank you! ** Tags added: kernel-spammed-focal-linux-gcp-tcpx-v2 verification-needed-focal-linux-gcp-tcpx -- You received this bug notification because you are a member of Kernel Packages, which is subscribed to linux in Ubuntu. https://bugs.launchpad.net/bugs/2039575 Title: SMC stats: Wrong bucket calculation for payload of exactly 4096 bytes Status in Ubuntu on IBM z Systems: Fix Committed Status in linux package in Ubuntu: New Status in linux source package in Jammy: Fix Committed Status in linux source package in Lunar: Fix Committed Status in linux source package in Mantic: Fix Committed Bug description: SRU Justification: [Impact] * There is a wrong bucket calculation for payload of exactly 4096 bytes in SMC stats counters. * SMC_STAT_PAYLOAD_SUB and SMC_STAT_RMB_SIZE_SUB have these issues. * The impact is that a system silently updates an incorrect stats counter in case the underlying kernel is not UBSAN enabled. (Difficult to detect.) * But if a kernel is UBSAN enabled, one will see a UBSAN 'array-index-out-of-bounds' warning every time this occurs, like: [ 26.335381] UBSAN: array-index-out-of-bounds in /build/linux-O6Qi7m/linux-5.15.0/net/smc/af_smc.c:2402:3 [ 26.335385] index -1 is out of range for type 'u64 [9]' [ 26.335388] CPU: 0 PID: 274 Comm: iperf3 Tainted: G E 5.15.0-79-generic #86-Ubuntu [ 26.335391] Hardware name: IBM 8561 T01 772 (KVM/Linux) [ 26.335393] Call Trace: [ 26.335397] [<00000000cd92e63a>] dump_stack_lvl+0x62/0x80 [ 26.335404] [<00000000cd92e36c>] ubsan_epilogue+0x1c/0x48 [ 26.335406] [<00000000cd52d3c4>] __ubsan_handle_out_of_bounds+0x94/0xa0 [ 26.335411] [<000003ff8033f9da>] smc_sendmsg+0x2aa/0x2d0 [smc] [ 26.335425] [<00000000cd6a79a4>] sock_sendmsg+0x64/0x80 [ 26.335431] [<00000000cd6a7a32>] sock_write_iter+0x72/0xa0 [ 26.335433] [<00000000cd1d4000>] new_sync_write+0x100/0x190 [ 26.335438] [<00000000cd1d4bb8>] vfs_write+0x1e8/0x280 [ 26.335440] [<00000000cd1d7014>] ksys_write+0xb4/0x100 [ 26.335442] [<00000000cd932c7c>] __do_syscall+0x1bc/0x1f0 [ 26.335446] [<00000000cd940148>] system_call+0x78/0xa0 [Fix] * a950a5921db4 a950a5921db450c74212327f69950ff03419483a "net/smc: Fix pos miscalculation in statistics" [Test Plan] * Since this got identified while the livepatch for jammy patches got added, one could run a simiar (or the same) test like mentioned at LP#1639924 (but only jammy comes with official livepatch support). * Alternatively a dedicated SMC stats counters test with a payload of exactly 4096 bytes could be done (which is probably easier): * Install uperf (https://uperf.org/ - https://github.com/uperf/uperf). (Wondering if it makes sense to pick this up as Debian/Ubuntu package ?!) * Reset SMC-D stats on client and server side. * Start uperf at the server side using: # uperf -vs * Update profile with remote IP (server IP) and start uperf at client: # uperf -vai 5 -m rr1c-4kx4k---1.xml with uperf profile: # cat rr1c-4kx4k---1.xml <?xml version="1.0"?> <profile name="TCP_RR"> <group nprocs="1"> <!--group nthreads="1"--> <!-- if we want to run processes --> <!--group nprocs="1"--> <transaction iterations="1"> <flowop type="connect" options="remotehost=<remote IP> protocol=tcp tcp_nodelay" /> </transaction> <transaction iterations="1"> <flowop type="write" options="size=4096"/> <flowop type="read" options="size=4096"/> </transaction> <transaction iterations="1"> <flowop type="disconnect" /> </transaction> </group> </profile> * The smcd stats output is: # smcd -d stats reset SMC-D Connections Summary Total connections handled 2 SMC connections 2 (client 2, server 0) v1 0 v2 2 Handshake errors 0 (client 0, server 0) Avg requests per SMC conn 14.0 TCP fallback 0 (client 0, server 0) RX Stats Data transmitted (Bytes) 5796 (5.796K) Total requests 9 Buffer full 0 (0.00%) Buffer downgrades 0 Buffer reuses 0 8KB 16KB 32KB 64KB 128KB 256KB 512KB >512KB Bufs 0 0 0 2 0 0 0 1 Reqs 8 0 0 0 0 0 0 0 TX Stats Data transmitted (Bytes) 9960 (9.960K) Total requests 19 Buffer full 0 (0.00%) Buffer full (remote) 0 (0.00%) Buffer too small 0 (0.00%) Buffer too small (remote) 0 (0.00%) Buffer downgrades 0 Buffer reuses 0 8KB 16KB 32KB 64KB 128KB 256KB 512KB >512KB Bufs 0 2 0 0 0 0 0 0 Reqs 18 0 0 0 0 0 0 1 Extras Special socket calls 0 cork 0 nodelay 0 sendpage 0 splice 0 urgent data 0 * Instead of including the payload in the wrong >512 KB buckets, the output should be to have 19 reqs in the 8 KB buckets for TX stats and 9 reqs in the 8 KB bucket for RX stats. [Where problems could occur] * The changes are in common code /net/smc/smc_stats.h hence any potential negativ impact is not limited to a specific platform. but limited to statistics for shared memory communication (SMC) hardware statistics. * But limited to the functions SMC_STAT_PAYLOAD_SUB and SMC_STAT_RMB_SIZE_SUB. * The bucket (aka pos) calculation is not that trivial, issues here could cause other calculation error, that may lead to slightly different, but still wrong numbers. * Further issues can happen in case variables in use may overflow. * Issues can also happen if the ternary conditional operators are not correctly used, which again may lead to wrong calculations. [Other Info] * This patch fixes upstream e0e4b8fa5338 ("net/smc: Add SMC statistics support"). * Since the fix got upstream accepted with v6.6-rc6, not only 22.04, but also 23.04 and 23.10 are affected. (22.10 already reached it EOS). * Will not affected the current development release (N-release), since it will at least be based on a 6.6 kernel. __________ Bug description by Nils H.: ------------ Overview: ------------ The following line in the SMC stats code (net/smc/smc_stats.h) caught my attention when using a payload of exactly 4096 bytes: #define SMC_STAT_PAYLOAD_SUB(_smc_stats, _tech, key, _len, _rc) \ do { \ typeof(_smc_stats) stats = (_smc_stats); \ typeof(_tech) t = (_tech); \ typeof(_len) l = (_len); \ int _pos = fls64((l) >> 13); \ typeof(_rc) r = (_rc); \ int m = SMC_BUF_MAX - 1; \ this_cpu_inc((*stats).smc[t].key ## _cnt); \ if (r <= 0) \ break; \ _pos = (_pos < m) ? ((l == 1 << (_pos + 12)) ? _pos - 1 : _pos) : m; \ <--- this_cpu_inc((*stats).smc[t].key ## _pd.buf[_pos]); \ this_cpu_add((*stats).smc[t].key ## _bytes, r); \ } \ while (0) With l = 4096, _pos evaluates to -1. Checking with the following uperf profile: # cat rr1c-4kx4k---1.xml <?xml version="1.0"?> <profile name="TCP_RR"> <group nprocs="1"> <!--group nthreads="1"--> <!-- if we want to run processes --> <!--group nprocs="1"--> <transaction iterations="1"> <flowop type="connect" options="remotehost=<remote IP> protocol=tcp tcp_nodelay" /> </transaction> <transaction iterations="1"> <flowop type="write" options="size=4096"/> <flowop type="read" options="size=4096"/> </transaction> <transaction iterations="1"> <flowop type="disconnect" /> </transaction> </group> </profile> smcd stats output: # smcd -d stats reset SMC-D Connections Summary Total connections handled 2 SMC connections 2 (client 2, server 0) v1 0 v2 2 Handshake errors 0 (client 0, server 0) Avg requests per SMC conn 14.0 TCP fallback 0 (client 0, server 0) RX Stats Data transmitted (Bytes) 5796 (5.796K) Total requests 9 Buffer full 0 (0.00%) Buffer downgrades 0 Buffer reuses 0 8KB 16KB 32KB 64KB 128KB 256KB 512KB >512KB Bufs 0 0 0 2 0 0 0 1 Reqs 8 0 0 0 0 0 0 0 TX Stats Data transmitted (Bytes) 9960 (9.960K) Total requests 19 Buffer full 0 (0.00%) Buffer full (remote) 0 (0.00%) Buffer too small 0 (0.00%) Buffer too small (remote) 0 (0.00%) Buffer downgrades 0 Buffer reuses 0 8KB 16KB 32KB 64KB 128KB 256KB 512KB >512KB Bufs 0 2 0 0 0 0 0 0 Reqs 18 0 0 0 0 0 0 1 Extras Special socket calls 0 cork 0 nodelay 0 sendpage 0 splice 0 urgent data 0 Instead of including the payload in the wrong >512KB buckets, output should be to have 19 reqs in the 8KB buckets for TX stats and 9 reqs in the 8KB bucket for RX stats. -------- Repro: -------- 0. Install uperf. 1. Reset SMC-D stats on client and server. 2. Start uperf at server side: "uperf -vs". 3. Update profile with remote IP (server IP) and start uperf at client: "uperf -vai 5 -m rr1c-4kx4k---1.xml" (uperf profile, see above) To manage notifications about this bug go to: https://bugs.launchpad.net/ubuntu-z-systems/+bug/2039575/+subscriptions -- Mailing list: https://launchpad.net/~kernel-packages Post to : kernel-packages@lists.launchpad.net Unsubscribe : https://launchpad.net/~kernel-packages More help : https://help.launchpad.net/ListHelp
