Date:	Mon, 9 Jun 2014 13:11:06 +0200
From:	John Muir <john@...ir.com>
To:	Maxim Patlasov <mpatlasov@...allels.com>
Cc:	Miklos Szeredi <miklos@...redi.hu>,
	fuse-devel <fuse-devel@...ts.sourceforge.net>,
	Linux List <linux-kernel@...r.kernel.org>
Subject: Re: [fuse-devel] [PATCH 0/5] fuse: close file synchronously (v2)

On 2014.06.09, at 12:46 , Maxim Patlasov <mpatlasov@...allels.com> wrote:

> On 06/09/2014 01:26 PM, John Muir wrote:
>> On 2014.06.09, at 9:50 , Maxim Patlasov <mpatlasov@...allels.com> wrote:
>> 
>>> On 06/06/2014 05:51 PM, John Muir wrote:
>>>> On 2014.06.06, at 15:27 , Maxim Patlasov <mpatlasov@...allels.com> wrote:
>>>> 
>>>>> The patch-set resolves the problem by making fuse_release synchronous:
>>>>> wait for ACK from userspace for FUSE_RELEASE if the feature is ON.
>>>> Why not make this feature per-file with a new flag bit in struct fuse_file_info rather than as a file-system global?
>>> I don't expect a great demand for such a granularity. File-system global "close_wait" conveys a general user expectation about filesystem behaviour in distributed environment: if you stopped using a file on given node, whether it means that the file is immediately accessible from another node.
>>> 
>> By user do you mean the end-user, or the implementor of the file-system? It seems to me that the end-user doesn't care, and just wants the file-system to work as expected. I don't think we're really talking about the end-user.
> 
> No, this is exactly about end-user expectations. Imagine a complicated heavy-loaded shared storage where handling FUSE_RELEASE in userspace may take a few minutes. In close_wait=0 case, an end-user who has just called close(2) has no idea when it's safe to access the file from another node or even when it's OK to umount filesystem.

I think we're saying the same thing here from different perspectives. The end-user wants the file-system to operate with the semantics you describe, but I don't think it makes sense to give the end-user control over those semantics. The file-system itself should be implemented that way, or not, or per-file

If it's a read-only file, then does this not add the overhead of having the kernel wait for the user-space file-system process to respond before closing it. In my experience, there is actually significant cost to the kernel to user-space messaging in FUSE when manipulating thousands of files.

> 
>> 
>> The implementor of a file-system, on the other hand, might want the semantics for close_wait on some files, but not on others. Won't there be a performance impact? Some distributed file-systems might want this on specific files only. Implementing it as a flag on the struct fuse_file_info gives the flexibility to the file-system implementor.
> 
> fuse_file_info is an userspace structure, in-kernel fuse knows nothing about it. In close_wait=1 case, nothing prevents a file-system implementation from ACK-ing FUSE_RELEASE request immediately (for specific files) and schedule actual handling for future processing.

Of course you know I meant that you'd add another flag to both fuse_file_info, and in the kernel equivalent for those flags which is struct fuse_open_out -> open_flags. This is where other such per file options are specified such as whether or not to keep the in-kernal cache for a file, whether or not to allow direct-io, and whether or not to allow seek.

Anyway, I guess you're the one doing all the work on this and if you have a particular implementation that doesn't require such fine-grained control, and no one else does then it's up to you. I'm just trying to show an alternative implementation that gives the file-system implementor more control while keeping the ability to meet user expectations.

Regards,
 
John.

--
John Muir - john@...ir.com
+32 491 64 22 76

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists