[Intel-gfx] [PATCH v3 6/6] freezer, sched: Rewrite core freezer logic

[Peter Zijlstra]

On Wed, Oct 26, 2022 at 01:43:00PM +0200, Peter Zijlstra wrote:
> On Wed, Oct 26, 2022 at 01:32:31PM +0300, Ville Syrjälä wrote:
> > Short form looks to be this:
> > <4>[  355.437846] 1 lock held by rs:main Q:Reg/359:
> > <4>[  355.438418]  #0: ffff88844693b758 (&rq->__lock){-.-.}-{2:2}, at: raw_spin_rq_lock_nested+0x1b/0x30
> > <4>[  355.438432] rs:main Q:Reg/359 holding locks while freezing
> 
> > <4>[  355.438429] ------------[ cut here ]------------
> > <4>[  355.438432] rs:main Q:Reg/359 holding locks while freezing
> > <4>[  355.438439] WARNING: CPU: 0 PID: 6211 at kernel/freezer.c:134 __set_task_frozen+0x86/0xb0
> > <4>[  355.438447] Modules linked in: snd_hda_intel i915 mei_hdcp mei_pxp drm_display_helper drm_kms_helper vgem drm_shmem_helper snd_hda_codec_hdmi snd_hda_codec_realtek snd_hda_codec_generic ledtrig_audio snd_intel_dspcfg snd_hda_codec snd_hwdep snd_hda_core snd_pcm prime_numbers ttm drm_buddy syscopyarea sysfillrect sysimgblt fb_sys_fops fuse x86_pkg_temp_thermal coretemp kvm_intel btusb btrtl btbcm btintel kvm irqbypass bluetooth crct10dif_pclmul crc32_pclmul ecdh_generic ghash_clmulni_intel ecc e1000e mei_me i2c_i801 ptp mei i2c_smbus pps_core lpc_ich video wmi [last unloaded: drm_kms_helper]
> > <4>[  355.438521] CPU: 0 PID: 6211 Comm: rtcwake Tainted: G     U             6.1.0-rc2-CI_DRM_12295-g3844a56a0922+ #1
> > <4>[  355.438526] Hardware name:  /NUC5i7RYB, BIOS RYBDWi35.86A.0385.2020.0519.1558 05/19/2020
> > <4>[  355.438530] RIP: 0010:__set_task_frozen+0x86/0xb0
> > <4>[  355.438536] Code: 83 60 09 00 00 85 c0 74 2a 48 89 df e8 ac 02 9b 00 8b 93 38 05 00 00 48 8d b3 48 07 00 00 48 c7 c7 a0 62 2b 82 e8 ee c1 9a 00 <0f> 0b c6 05 51 75 e3 02 01 c7 43 18 00 80 00 00 b8 00 80 00 00 5b
> > <4>[  355.438541] RSP: 0018:ffffc900012cbcf0 EFLAGS: 00010086
> > <4>[  355.438546] RAX: 0000000000000000 RBX: ffff88810d090040 RCX: 0000000000000004
> > <4>[  355.438550] RDX: 0000000000000004 RSI: 00000000fffff5de RDI: 00000000ffffffff
> > <4>[  355.438553] RBP: 0000000000000000 R08: 0000000000000000 R09: c0000000fffff5de
> > <4>[  355.438557] R10: 00000000002335f8 R11: ffffc900012cbb88 R12: 0000000000000246
> > <4>[  355.438561] R13: ffffffff81165430 R14: 0000000000000000 R15: ffff88810d090040
> > <4>[  355.438565] FS:  00007fcfa43c7740(0000) GS:ffff888446800000(0000) knlGS:0000000000000000
> > <4>[  355.438569] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> > <4>[  355.438582] CR2: 00007fceb380f6b8 CR3: 0000000117c5c004 CR4: 00000000003706f0
> > <4>[  355.438586] Call Trace:
> > <4>[  355.438589]  <TASK>
> > <4>[  355.438592]  task_call_func+0xc4/0xe0
> > <4>[  355.438600]  freeze_task+0x84/0xe0
> > <4>[  355.438607]  try_to_freeze_tasks+0xac/0x260
> > <4>[  355.438616]  freeze_processes+0x56/0xb0
> > <4>[  355.438622]  pm_suspend.cold.7+0x1d9/0x31c
> > <4>[  355.438629]  state_store+0x7b/0xe0
> > <4>[  355.438637]  kernfs_fop_write_iter+0x124/0x1c0
> > <4>[  355.438644]  vfs_write+0x34f/0x4e0
> > <4>[  355.438655]  ksys_write+0x57/0xd0
> > <4>[  355.438663]  do_syscall_64+0x3a/0x90
> > <4>[  355.438670]  entry_SYSCALL_64_after_hwframe+0x63/0xcd
> 
> Oh I think I see what's going on.
> 
> It's a very narrow race between schedule() and task_call_func().
> 
>   CPU0						CPU1
> 
>   __schedule()
>     rq_lock();
>     prev_state = READ_ONCE(prev->__state);
>     if (... && prev_state) {
>       deactivate_tasl(rq, prev, ...)
>         prev->on_rq = 0;
> 
> 						task_call_func()
> 						  raw_spin_lock_irqsave(p->pi_lock);
> 						  state = READ_ONCE(p->__state);
> 						  smp_rmb();
> 						  if (... || p->on_rq) // false!!!
> 						    rq = __task_rq_lock()
> 
> 						  ret = func();
> 
>     next = pick_next_task();
>     rq = context_switch(prev, next)
>       prepare_lock_switch()
>         spin_release(&__rq_lockp(rq)->dep_map...)
> 
> 
> 
> So while the task is on it's way out, it still holds rq->lock for a
> little while, and right then task_call_func() comes in and figures it
> doesn't need rq->lock anymore (because the task is already dequeued --
> but still running there) and then the __set_task_frozen() thing observes
> it's holding rq->lock and yells murder.
> 
> Could you please give the below a spin?

Urgh.. that'll narrow the race more, but won't solve it, that
prepare_lock_switch() is after we clear ->on_cpu.

Let me ponder this a wee bit more..