rcu-tasks: Fix race in schedule and flush work
authorPadmanabha Srinivasaiah <treasure4paddy@gmail.com>
Thu, 17 Feb 2022 15:25:19 +0000 (16:25 +0100)
committerPaul E. McKenney <paulmck@kernel.org>
Tue, 12 Apr 2022 00:06:42 +0000 (17:06 -0700)
commitf75fd4b9221d93177c50dcfde671b2e907f53e86
treeb39a812b34a8fff44241530580cd06a9c401e55b
parent3123109284176b1532874591f7c81f3837bbdc17
rcu-tasks: Fix race in schedule and flush work

While booting secondary CPUs, cpus_read_[lock/unlock] is not keeping
online cpumask stable. The transient online mask results in below
calltrace.

[    0.324121] CPU1: Booted secondary processor 0x0000000001 [0x410fd083]
[    0.346652] Detected PIPT I-cache on CPU2
[    0.347212] CPU2: Booted secondary processor 0x0000000002 [0x410fd083]
[    0.377255] Detected PIPT I-cache on CPU3
[    0.377823] CPU3: Booted secondary processor 0x0000000003 [0x410fd083]
[    0.379040] ------------[ cut here ]------------
[    0.383662] WARNING: CPU: 0 PID: 10 at kernel/workqueue.c:3084 __flush_work+0x12c/0x138
[    0.384850] Modules linked in:
[    0.385403] CPU: 0 PID: 10 Comm: rcu_tasks_rude_ Not tainted 5.17.0-rc3-v8+ #13
[    0.386473] Hardware name: Raspberry Pi 4 Model B Rev 1.4 (DT)
[    0.387289] pstate: 20000005 (nzCv daif -PAN -UAO -TCO -DIT -SSBS BTYPE=--)
[    0.388308] pc : __flush_work+0x12c/0x138
[    0.388970] lr : __flush_work+0x80/0x138
[    0.389620] sp : ffffffc00aaf3c60
[    0.390139] x29: ffffffc00aaf3d20 x28: ffffffc009c16af0 x27: ffffff80f761df48
[    0.391316] x26: 0000000000000004 x25: 0000000000000003 x24: 0000000000000100
[    0.392493] x23: ffffffffffffffff x22: ffffffc009c16b10 x21: ffffffc009c16b28
[    0.393668] x20: ffffffc009e53861 x19: ffffff80f77fbf40 x18: 00000000d744fcc9
[    0.394842] x17: 000000000000000b x16: 00000000000001c2 x15: ffffffc009e57550
[    0.396016] x14: 0000000000000000 x13: ffffffffffffffff x12: 0000000100000000
[    0.397190] x11: 0000000000000462 x10: ffffff8040258008 x9 : 0000000100000000
[    0.398364] x8 : 0000000000000000 x7 : ffffffc0093c8bf4 x6 : 0000000000000000
[    0.399538] x5 : 0000000000000000 x4 : ffffffc00a976e40 x3 : ffffffc00810444c
[    0.400711] x2 : 0000000000000004 x1 : 0000000000000000 x0 : 0000000000000000
[    0.401886] Call trace:
[    0.402309]  __flush_work+0x12c/0x138
[    0.402941]  schedule_on_each_cpu+0x228/0x278
[    0.403693]  rcu_tasks_rude_wait_gp+0x130/0x144
[    0.404502]  rcu_tasks_kthread+0x220/0x254
[    0.405264]  kthread+0x174/0x1ac
[    0.405837]  ret_from_fork+0x10/0x20
[    0.406456] irq event stamp: 102
[    0.406966] hardirqs last  enabled at (101): [<ffffffc0093c8468>] _raw_spin_unlock_irq+0x78/0xb4
[    0.408304] hardirqs last disabled at (102): [<ffffffc0093b8270>] el1_dbg+0x24/0x5c
[    0.409410] softirqs last  enabled at (54): [<ffffffc0081b80c8>] local_bh_enable+0xc/0x2c
[    0.410645] softirqs last disabled at (50): [<ffffffc0081b809c>] local_bh_disable+0xc/0x2c
[    0.411890] ---[ end trace 0000000000000000 ]---
[    0.413000] smp: Brought up 1 node, 4 CPUs
[    0.413762] SMP: Total of 4 processors activated.
[    0.414566] CPU features: detected: 32-bit EL0 Support
[    0.415414] CPU features: detected: 32-bit EL1 Support
[    0.416278] CPU features: detected: CRC32 instructions
[    0.447021] Callback from call_rcu_tasks_rude() invoked.
[    0.506693] Callback from call_rcu_tasks() invoked.

This commit therefore fixes this issue by applying a single-CPU
optimization to the RCU Tasks Rude grace-period process.  The key point
here is that the purpose of this RCU flavor is to force a schedule on
each online CPU since some past event.  But the rcu_tasks_rude_wait_gp()
function runs in the context of the RCU Tasks Rude's grace-period kthread,
so there must already have been a context switch on the current CPU since
the call to either synchronize_rcu_tasks_rude() or call_rcu_tasks_rude().
So if there is only a single CPU online, RCU Tasks Rude's grace-period
kthread does not need to anything at all.

It turns out that the rcu_tasks_rude_wait_gp() function's call to
schedule_on_each_cpu() causes problems during early boot.  During that
time, there is only one online CPU, namely the boot CPU.  Therefore,
applying this single-CPU optimization fixes early-boot instances of
this problem.

Link: https://lore.kernel.org/lkml/20220210184319.25009-1-treasure4paddy@gmail.com/T/
Suggested-by: Paul E. McKenney <paulmck@kernel.org>
Signed-off-by: Padmanabha Srinivasaiah <treasure4paddy@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
kernel/rcu/tasks.h