https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=295992
Bug ID: 295992
Summary: bnxt: System hangs when nstat and sysctl -a run
concurrently
Product: Base System
Version: CURRENT
Hardware: Any
OS: Any
Status: New
Severity: Affects Only Me
Priority: ---
Component: kern
Assignee: [email protected]
Reporter: [email protected]
While running the heavy traffic on Broadcom 400G NIC(Thor2), I observed a
temporary system hang which causes traffic to come to halt for some time when
nstat (Network traffic monitoring tool) is running and "sysctl -a" is run in
parallel. The issue is intermittent. Below are the traces of both threads with
our analysis:
-------------
nstat:
# procstat -kk 39693
PID TID COMM TDNAME KSTACK
39693 101860 nstat - mi_switch+0x172
__mtx_lock_sleep+0x1c1 __mtx_lock_flags+0xdd sysctl_root_handler_locked+0x8c
sysctl_root+0x22f userland_sysctl+0x1b6 kern___sysctlbyname+0x226
sys___sysctlbyname+0x2d amd64_syscall+0x169 fast_syscall_common+0xf8
sysctl -a:
# procstat -kk 40586
PID TID COMM TDNAME KSTACK
40586 102075 sysctl - mi_switch+0x172
sched_ule_bind+0x8a cpu_est_clockrate+0x81 hwpstate_get_cppc+0x48
cf_get_method+0xf0 cpufreq_curr_sysctl+0x68 sysctl_root_handler_locked+0x9c
sysctl_root+0x22f userland_sysctl+0x1b6 sys___sysctl+0x65 amd64_syscall+0x169
fast_syscall_common+0xf8
It's a timing-related temporarily system hang when the traffic monitoring tool
nstat and sysctl -a are queried concurrently. This conflict results in both
threads hanging temporarily, causing a CPU halt that leads nstat to report zero
Tx/Rx packets and NaN CPU utilization once the hang resolves.
The root cause is a Giant mutex convoy triggered by the cpufreq sysctl handler.
Specifically, the sysctl -a process acquires the Giant mutex and then attempts
to migrate to a specific CPU to read clock rates, where it remains blocked if
that target CPU is saturated with NIC traffic.
Trace Analysis:
- sysctl -a (The Blocker): Holds the Giant mutex while executing
cpufreq_curr_sysctl. The handler calls hwpstate_get_cppc, which invokes
cpu_est_clockrate. Inside cpu_est_clockrate, the thread calls sched_bind to
migrate to the target CPU. Because the target CPU is busy, mi_switch is called,
and the thread sleeps while still holding the Giant mutex.
- nstat (The Victim): Attempts to execute a sysctl call but becomes blocked
at __mtx_lock_sleep while waiting for the Giant mutex held by the sysctl
thread.
Technical Details:
- Giant Acquisition: kern_sysctl.c (Lines 22-23) acquires the mutex because
the OID lacks the CTLFLAG_MPSAFE flag.
- CPU Migration: cpu_machdep.c (Lines 444-450) forces the thread to bind to
the target cpu_id via sched_bind.
- Thread State: The sysctl thread remains in mi_switch within sched_ule.c
(Line 3023) until the target CPU can schedule it, effectively locking out any
other processes (like nstat) that require the Giant mutex.
------------------
--
You are receiving this mail because:
You are the assignee for the bug.
