Re: [fuse] 6b2fb79963: WARNING:at_fs/fuse/file.c:#tree_insert[fuse]

[Vasily Averin]

On 6/16/20 3:35 PM, Miklos Szeredi wrote:
> On Mon, Jun 15, 2020 at 7:59 AM Vasily Averin <vvs@xxxxxxxxxxxxx> wrote:
>>
>> On 6/15/20 3:50 AM, kernel test robot wrote:
>>> FYI, we noticed the following commit (built with gcc-9):
>>>
>>> commit: 6b2fb79963fbed7db3ef850926d913518fd5c62f ("fuse: optimize writepages search")
>>> https://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git master
>>
>>> [ 1030.995703] ------------[ cut here ]------------
>>> [ 1030.997563] WARNING: CPU: 1 PID: 17211 at fs/fuse/file.c:1728 tree_insert+0xab/0xc0 [fuse]
>>> [ 1031.021943] RIP: 0010:tree_insert+0xab/0xc0 [fuse]
>>> [ 1031.057802] Call Trace:
>>> [ 1031.060015]  fuse_writepages_fill+0x5da/0x6a0 [fuse]
>>> [ 1031.062536]  write_cache_pages+0x171/0x470
>>> [ 1031.064933]  ? fuse_writepages+0x100/0x100 [fuse]
>>> [ 1031.067419]  ? terminate_walk+0xd3/0xf0
>>> [ 1031.069707]  ? _cond_resched+0x19/0x30
>>> [ 1031.072140]  ? __kmalloc+0x274/0x280
>>> [ 1031.074407]  fuse_writepages+0x8a/0x100 [fuse]
>>> [ 1031.076599]  do_writepages+0x43/0xe0
>>
>> It is  WARN_ON(!wpa->ia.ap.num_pages);
>> however tree_insert() was called from fuse_writepages_fill() with ap->num_pages = 0;
>> In submitted PATCH RFC we have used
>>
>> +static int tree_insert(struct rb_root *root, struct fuse_req *ins_req)
>> ...
>> +       pgoff_t idx_to   = idx_from + (ins_req->num_pages ?
>> +                               ins_req->num_pages - 1 : 0);
>>
>> Though committed patch version have
>>
>> +static void tree_insert(struct rb_root *root, struct fuse_writepage_args *wpa)
>> ...
>> +       pgoff_t idx_to = idx_from + wpa->ia.ap.num_pages - 1;
>> ...
>> +       WARN_ON(!wpa->ia.ap.num_pages);
>>
>> Miklos,
>> do you have any objections if I return to our initial proposal?
>> Am I missed something and it is not allowed now?
> 
> No objections, but you need to explain with a comment why that special
> casing of num_pages == 0 is needed.  I don't understand it yet.

fuse_writepages_fill() handles few different cases,
in some of them data->wpa should be re-allocated and re-initialized,
another cases requires just add new page into existing wpa and increments num_pages.

In current implementation fuse_writepages_fill() tries to share the code,
by this way for new wpa it calls tree_insert() with num_pages = 0 
then switches to common code used non-modified num_pages and increments it at the very end.

I've prepared patch that calls tree_insert() with num_pages = 1,
It have few other minor optimizations therefore I think it is better
than remove incorrect WARN_ON from tree_insert proposed earlier.

During investigations I've noticed few other stranges in neighborhood:
1) fuse_writepages() ignores some errors taken from fuse_writepages_fill()
I believe it is a bug: if .writepages is called with WB_SYNC_ALL 
it should either guarantee that all data was successfully saved or return error.

2) fuse_writepages_fill uses following construction:
if (wpa && ap->num_pages &&
    (A || B || C)) {
	action;
} else if (wpa && ap->num_pages == data->max_pages) {
	if (D) {
		the same action;
	}
}

- ap->num_pages check is always true and can be removed
- "if" and "else if" calls the same action and can be merged.

I'm going to sent separate patches to resolve it.

Thank you,
	Vasily Averin