From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org Received: from vger.kernel.org (vger.kernel.org [23.128.96.18]) by smtp.lore.kernel.org (Postfix) with ESMTP id 36F1BEB64DD for ; Thu, 13 Jul 2023 04:40:18 +0000 (UTC) Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S233675AbjGMEkR (ORCPT ); Thu, 13 Jul 2023 00:40:17 -0400 Received: from lindbergh.monkeyblade.net ([23.128.96.19]:42478 "EHLO lindbergh.monkeyblade.net" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S233446AbjGMEkQ (ORCPT ); Thu, 13 Jul 2023 00:40:16 -0400 Received: from out-10.mta1.migadu.com (out-10.mta1.migadu.com [95.215.58.10]) by lindbergh.monkeyblade.net (Postfix) with ESMTPS id 281F51BC6 for ; Wed, 12 Jul 2023 21:40:15 -0700 (PDT) Message-ID: <077f4874-015b-a534-4a29-de877b735e38@linux.dev> DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=linux.dev; s=key1; t=1689223213; h=from:from:reply-to:subject:subject:date:date:message-id:message-id: to:to:cc:cc:mime-version:mime-version:content-type:content-type: content-transfer-encoding:content-transfer-encoding: in-reply-to:in-reply-to:references:references; bh=git7KZMGpcoyXjLF+BIDp31/p7iLby/qgJrULfG9e6g=; b=OCKSYom9t8/EedSo3Iw1M4pc/i6LMRJTyoRVH3c8eoln0nk1J1rhXO4XRSzDy+Iou1COCk b1rSvr0B3Fm+jVKYifMHG7m+AR0Go3gq3SIjSsRsJdMaSoFF2FQQs5CemEETc3eptXRCCD bgF2a0objtV0MepM3RB3bNtcMyBoklA= Date: Thu, 13 Jul 2023 12:40:05 +0800 MIME-Version: 1.0 Subject: Re: [PATCH 3/3] io_uring: add support for getdents Content-Language: en-US X-Report-Abuse: Please report any abuse attempt to abuse@migadu.com and include these headers. From: Hao Xu To: Dominique Martinet Cc: io-uring@vger.kernel.org, Jens Axboe , Pavel Begunkov , Christian Brauner , Alexander Viro , Stefan Roesch , Clay Harris , Dave Chinner , linux-fsdevel@vger.kernel.org, Wanpeng Li References: <20230711114027.59945-1-hao.xu@linux.dev> <20230711114027.59945-4-hao.xu@linux.dev> <858c3f16-ffb3-217e-b5d6-fcc63ef9c401@linux.dev> In-Reply-To: Content-Type: text/plain; charset=UTF-8; format=flowed Content-Transfer-Encoding: 8bit X-Migadu-Flow: FLOW_OUT Precedence: bulk List-ID: X-Mailing-List: io-uring@vger.kernel.org On 7/13/23 12:05, Hao Xu wrote: > > On 7/13/23 00:10, Dominique Martinet wrote: >> Hao Xu wrote on Wed, Jul 12, 2023 at 03:53:24PM +0800: >>>>> +        if (file_count(file) > 1) >>>> I was curious about this so I found it's basically what __fdget_pos >>>> does >>>> before deciding it should take the f_pos_lock, and as such this is >>>> probably correct... But if someone can chime in here: what guarantees >>>> someone else won't __fdget_pos (or equivalent through this) the file >>>> again between this and the vfs_getdents call? >>>> That second get would make file_count > 1 and it would lock, but lock >>>> hadn't been taken here so the other call could get the lock without >>>> waiting and both would process getdents or seek or whatever in >>>> parallel. >>>> >>> This file_count(file) is atomic_read, so I believe no race condition >>> here. >> I don't see how that helps in the presence of another thread getting the >> lock after we possibly issued a getdents without the lock, e.g. >> >> t1 call io_uring getdents here >> t1 sees file_count(file) == 1 and skips getting lock >> t1 starts issuing vfs_getdents [... processing] >> t2 calls either io_uring getdents or getdents64 syscall >> t2 gets the lock, since it wasn't taken by t1 it can be obtained >> t2 issues another vfs_getdents >> >> Christian raised the same issue so I'll leave this to his part of the >> thread for reply, but I hope that clarified my concern. > > > Hi Dominique, > > Ah, I misunderstood your question, sorry. The thing is f_count is > init-ed to be 1, > > and normal uring requests do fdget first, so I think it's ok for normal > requests. > > What Christian points out is issue with fixed file, that is indeed a > problem I think. After re-think of it, I think there is no race in fixed file case as well, because the f_count is always >1 > > >> >> ----- >> >> BTW I forgot to point out: this dropped the REWIND bit from my patch; I >> believe some form of "seek" is necessary for real applications to make >> use of this (for example, a web server could keep the fd open in a LRU >> and keep issuing readdir over and over again everytime it gets an >> indexing request); not having rewind means it'd need to close and >> re-open the fd everytime which doesn't seem optimal. >> >> A previous iteration discussed that real seek is difficult and not >> necessarily needed to I settled for rewind, but was there a reason you >> decided to stop handling that? >> >> My very egoistical personal use case won't require it, so I can just say >> I don't care here, but it would be nice to have a reason explained at >> some point > > > Yes, like Al pointed out, getdents with an offset is not the right way > to do it, > > So a way to do seek is a must. But like what I said in the cover-letter, > I do think the right thing is to > > import lseek/llseek to io_uring, not increment the complex of getdents. > > > Thanks, > > Hao > >