[PATCH 08/12] cachefiles: never get a new anon fd if ondemand_id is valid

[Baokun Li]

Hi Jingbo,

On 2024/5/6 11:09, Jingbo Xu wrote:
>
> On 4/24/24 11:39 AM, libaokun at huaweicloud.com wrote:
>> From: Baokun Li <libaokun1 at huawei.com>
>>
>> Now every time the daemon reads an open request, it requests a new anon fd
>> and ondemand_id. With the introduction of "restore", it is possible to read
>> the same open request more than once, and therefore have multiple anon fd's
>> for the same object.
>>
>> To avoid this, allocate a new anon fd only if no anon fd has been allocated
>> (ondemand_id == 0) or if the previously allocated anon fd has been closed
>> (ondemand_id == -1). Returns an error if ondemand_id is valid, letting the
>> daemon know that the current userland restore logic is abnormal and needs
>> to be checked.
> I have no obvious preference on strengthening this on kernel side or
> not.  Could you explain more about what will happen if the daemon gets
> several distinct anon fd corresponding to one same object?  IMHO the
> daemon should expect the side effect if it issues a 'restore' command
> when the daemon doesn't crash.  IOW, it's something that shall be fixed
> or managed either on the kernel side, or on the daemon side.
If the anon_fd is not unique, the daemon will only close the anon_fd
corresponding to the newest object_id during drop_object, and the
other anon_fds will not be closed until the daemon exits.

However, the anon_fd holds the reference count of the object, so the
object will not be freed, and the cookie will also not be freed. So
mounting a same-named image at this point will cause a hung task
in fscache_hash_cookie() by waiting for the cookie to unhash.

The object_id and anon_fd of an object are supposed to be unique
under normal circumstances, this patch just provides that guarantee
even in the case of an exception.

Thank you very much for the review!

Regards,
Baokun
>> ---
>>   fs/cachefiles/ondemand.c | 34 ++++++++++++++++++++++++++++------
>>   1 file changed, 28 insertions(+), 6 deletions(-)
>>
>> diff --git a/fs/cachefiles/ondemand.c b/fs/cachefiles/ondemand.c
>> index b5e6a851ef04..0cf63bfedc9e 100644
>> --- a/fs/cachefiles/ondemand.c
>> +++ b/fs/cachefiles/ondemand.c
>> @@ -14,11 +14,18 @@ static int cachefiles_ondemand_fd_release(struct inode *inode,
>>   					  struct file *file)
>>   {
>>   	struct cachefiles_object *object = file->private_data;
>> -	struct cachefiles_cache *cache = object->volume->cache;
>> -	struct cachefiles_ondemand_info *info = object->ondemand;
>> +	struct cachefiles_cache *cache;
>> +	struct cachefiles_ondemand_info *info;
>>   	int object_id;
>>   	struct cachefiles_req *req;
>> -	XA_STATE(xas, &cache->reqs, 0);
>> +	XA_STATE(xas, NULL, 0);
>> +
>> +	if (!object)
>> +		return 0;
>> +
>> +	info = object->ondemand;
>> +	cache = object->volume->cache;
>> +	xas.xa = &cache->reqs;
>>   
>>   	xa_lock(&cache->reqs);
>>   	spin_lock(&info->lock);
>> @@ -269,22 +276,39 @@ static int cachefiles_ondemand_get_fd(struct cachefiles_req *req)
>>   		goto err_put_fd;
>>   	}
>>   
>> +	spin_lock(&object->ondemand->lock);
>> +	if (object->ondemand->ondemand_id > 0) {
>> +		spin_unlock(&object->ondemand->lock);
>> +		ret = -EEXIST;
>> +		/* Avoid performing cachefiles_ondemand_fd_release(). */
>> +		file->private_data = NULL;
>> +		goto err_put_file;
>> +	}
>> +
>>   	file->f_mode |= FMODE_PWRITE | FMODE_LSEEK;
>>   	fd_install(fd, file);
>>   
>>   	load = (void *)req->msg.data;
>>   	load->fd = fd;
>>   	object->ondemand->ondemand_id = object_id;
>> +	spin_unlock(&object->ondemand->lock);
>>   
>>   	cachefiles_get_unbind_pincount(cache);
>>   	trace_cachefiles_ondemand_open(object, &req->msg, load);
>>   	return 0;
>>   
>> +err_put_file:
>> +	fput(file);
>>   err_put_fd:
>>   	put_unused_fd(fd);
>>   err_free_id:
>>   	xa_erase(&cache->ondemand_ids, object_id);
>>   err:
>> +	spin_lock(&object->ondemand->lock);
>> +	/* Avoid marking an opened object as closed. */
>> +	if (object->ondemand->ondemand_id <= 0)
>> +		cachefiles_ondemand_set_object_close(object);
>> +	spin_unlock(&object->ondemand->lock);
>>   	cachefiles_put_object(object, cachefiles_obj_put_ondemand_fd);
>>   	return ret;
>>   }
>> @@ -367,10 +391,8 @@ ssize_t cachefiles_ondemand_daemon_read(struct cachefiles_cache *cache,
>>   
>>   	if (msg->opcode == CACHEFILES_OP_OPEN) {
>>   		ret = cachefiles_ondemand_get_fd(req);
>> -		if (ret) {
>> -			cachefiles_ondemand_set_object_close(req->object);
>> +		if (ret)
>>   			goto out;
>> -		}
>>   	}
>>   
>>   	msg->msg_id = xas.xa_index;