scheduling while atomic in z3fold

From: Oleksandr Natalenko <oleksandr@natalenko.name>
To: linux-kernel@vger.kernel.org
Cc: linux-mm@kvack.org, Andrew Morton <akpm@linux-foundation.org>,
	Sebastian Andrzej Siewior <bigeasy@linutronix.de>,
	Steven Rostedt <rostedt@goodmis.org>,
	Mike Galbraith <efault@gmx.de>,
	Thomas Gleixner <tglx@linutronix.de>
Subject: scheduling while atomic in z3fold
Date: Sat, 28 Nov 2020 15:05:23 +0100	[thread overview]
Message-ID: <20201128140523.ovmqon5fjetvpby4@spock.localdomain> (raw)

Hi.

While running v5.10-rc5-rt11 I bumped into the following:

```
BUG: scheduling while atomic: git/18695/0x00000002
Preemption disabled at:
[<ffffffffbb93fcb3>] z3fold_zpool_malloc+0x463/0x6e0
…
Call Trace:
 dump_stack+0x6d/0x88
 __schedule_bug.cold+0x88/0x96
 __schedule+0x69e/0x8c0
 preempt_schedule_lock+0x51/0x150
 rt_spin_lock_slowlock_locked+0x117/0x2c0
 rt_spin_lock_slowlock+0x58/0x80
 rt_spin_lock+0x2a/0x40
 z3fold_zpool_malloc+0x4c1/0x6e0
 zswap_frontswap_store+0x39c/0x980
 __frontswap_store+0x6e/0xf0
 swap_writepage+0x39/0x70
 shmem_writepage+0x31b/0x490
 pageout+0xf4/0x350
 shrink_page_list+0xa28/0xcc0
 shrink_inactive_list+0x300/0x690
 shrink_lruvec+0x59a/0x770
 shrink_node+0x2d6/0x8d0
 do_try_to_free_pages+0xda/0x530
 try_to_free_pages+0xff/0x260
 __alloc_pages_slowpath.constprop.0+0x3d5/0x1230
 __alloc_pages_nodemask+0x2f6/0x350
 allocate_slab+0x3da/0x660
 ___slab_alloc+0x4ff/0x760
 __slab_alloc.constprop.0+0x7a/0x100
 kmem_cache_alloc+0x27b/0x2c0
 __d_alloc+0x22/0x230
 d_alloc_parallel+0x67/0x5e0
 __lookup_slow+0x5c/0x150
 path_lookupat+0x2ea/0x4d0
 filename_lookup+0xbf/0x210
 vfs_statx.constprop.0+0x4d/0x110
 __do_sys_newlstat+0x3d/0x80
 do_syscall_64+0x33/0x40
 entry_SYSCALL_64_after_hwframe+0x44/0xa9
```

The preemption seems to be disabled here:

```
$ scripts/faddr2line mm/z3fold.o z3fold_zpool_malloc+0x463
z3fold_zpool_malloc+0x463/0x6e0:
add_to_unbuddied at mm/z3fold.c:645
(inlined by) z3fold_alloc at mm/z3fold.c:1195
(inlined by) z3fold_zpool_malloc at mm/z3fold.c:1737
```

The call to the rt_spin_lock() seems to be here:

```
$ scripts/faddr2line mm/z3fold.o z3fold_zpool_malloc+0x4c1
z3fold_zpool_malloc+0x4c1/0x6e0:
add_to_unbuddied at mm/z3fold.c:649
(inlined by) z3fold_alloc at mm/z3fold.c:1195
(inlined by) z3fold_zpool_malloc at mm/z3fold.c:1737
```

Or, in source code:

```
 639 /* Add to the appropriate unbuddied list */
 640 static inline void add_to_unbuddied(struct z3fold_pool *pool,
 641                 struct z3fold_header *zhdr)
 642 {
 643     if (zhdr->first_chunks == 0 || zhdr->last_chunks == 0 ||
 644             zhdr->middle_chunks == 0) {
 645         struct list_head *unbuddied = get_cpu_ptr(pool->unbuddied);
 646
 647         int freechunks = num_free_chunks(zhdr);
 648         spin_lock(&pool->lock);
 649         list_add(&zhdr->buddy, &unbuddied[freechunks]);
 650         spin_unlock(&pool->lock);
 651         zhdr->cpu = smp_processor_id();
 652         put_cpu_ptr(pool->unbuddied);
 653     }
 654 }
```

Shouldn't the list manipulation be protected with
local_lock+this_cpu_ptr instead of get_cpu_ptr+spin_lock?

Thanks.

-- 
  Oleksandr Natalenko (post-factum)