memcg: avoid dead loop when setting memory.max

A softlockup issue was found with stress test:
 watchdog: BUG: soft lockup - CPU#27 stuck for 26s! [migration/27:181]
 CPU: 27 UID: 0 PID: 181 Comm: migration/27 6.14.0-rc2-next-20250210 #1
 Stopper: multi_cpu_stop <- stop_machine_from_inactive_cpu
 RIP: 0010:stop_machine_yield+0x2/0x10
 RSP: 0000:ff4a0dcecd19be48 EFLAGS: 00000246
 RAX: ffffffff89c0108f RBX: ff4a0dcec03afe44 RCX: 0000000000000000
 RDX: ff1cdaaf6eba5808 RSI: 0000000000000282 RDI: ff1cda80c1775a40
 RBP: 0000000000000001 R08: 00000011620096c6 R09: 7fffffffffffffff
 R10: 0000000000000001 R11: 0000000000000100 R12: ff1cda80c1775a40
 R13: 0000000000000000 R14: 0000000000000001 R15: ff4a0dcec03afe20
 FS:  0000000000000000(0000) GS:ff1cdaaf6eb80000(0000)
 CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
 CR2: 0000000000000000 CR3: 00000025e2c2a001 CR4: 0000000000773ef0
 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
 PKRU: 55555554
 Call Trace:
  multi_cpu_stop+0x8f/0x100
  cpu_stopper_thread+0x90/0x140
  smpboot_thread_fn+0xad/0x150
  kthread+0xc2/0x100
  ret_from_fork+0x2d/0x50

The stress test involves CPU hotplug operations and memory control group
(memcg) operations. The scenario can be described as follows:

 echo xx > memory.max 	cache_ap_online			oom_reaper
 (CPU23)						(CPU50)
 xx < usage		stop_machine_from_inactive_cpu
 for(;;)			// all active cpus
 trigger OOM		queue_stop_cpus_work
 // waiting oom_reaper
 			multi_cpu_stop(migration/xx)
 			// sync all active cpus ack
 			// waiting cpu23 ack
 			// CPU50 loops in multi_cpu_stop
 							waiting cpu50

Detailed explanation:
1. When the usage is larger than xx, an OOM may be triggered. If the
   process does not handle with ths kill signal immediately, it will loop
   in the memory_max_write.
2. When cache_ap_online is triggered, the multi_cpu_stop is queued to the
   active cpus. Within the multi_cpu_stop function,  it attempts to
   synchronize the CPU states. However, the CPU23 didn't acknowledge
   because it is stuck in a loop within the for(;;).
3. The oom_reaper process is blocked because CPU50 is in a loop, waiting
   for CPU23 to acknowledge the synchronization request.
4. Finally, it formed cyclic dependency and lead to softlockup and dead
   loop.

To fix this issue, add cond_resched() in the memory_max_write, so that it
will not block migration task.

Link: https://lkml.kernel.org/r/[email protected]
Fixes: b6e6edc ("mm: memcontrol: reclaim and OOM kill when shrinking memory.max below usage")
Signed-off-by: Chen Ridong <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Cc: Roman Gushchin <[email protected]>
Cc: Johannes Weiner <[email protected]>
Cc: Shakeel Butt <[email protected]>
Cc: Muchun Song <[email protected]>
Cc: Wang Weiyang <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>

mm/memcontrol.c

@@ -4166,6 +4166,7 @@ static ssize_t memory_max_write(struct kernfs_open_file *of,
 		memcg_memory_event(memcg, MEMCG_OOM);
 		if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
 			break;
+		cond_resched();
 	}
 
 	memcg_wb_domain_size_changed(memcg);