From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: with ECARTIS (v1.0.0; list xfs); Mon, 19 May 2008 23:36:03 -0700 (PDT) Received: from cuda.sgi.com (cuda2.sgi.com [192.48.168.29]) by oss.sgi.com (8.12.11.20060308/8.12.11/SuSE Linux 0.7) with ESMTP id m4K6Zwr1031525 for ; Mon, 19 May 2008 23:35:58 -0700 Received: from verein.lst.de (localhost [127.0.0.1]) by cuda.sgi.com (Spam Firewall) with ESMTP id 21E1B18D2D2 for ; Mon, 19 May 2008 23:36:47 -0700 (PDT) Received: from verein.lst.de (verein.lst.de [213.95.11.210]) by cuda.sgi.com with ESMTP id 2HnVGaEz6wuHHMeu for ; Mon, 19 May 2008 23:36:47 -0700 (PDT) Received: from verein.lst.de (localhost [127.0.0.1]) by verein.lst.de (8.12.3/8.12.3/Debian-7.1) with ESMTP id m4K6adF3008954 (version=TLSv1/SSLv3 cipher=EDH-RSA-DES-CBC3-SHA bits=168 verify=NO) for ; Tue, 20 May 2008 08:36:39 +0200 Received: (from hch@localhost) by verein.lst.de (8.12.3/8.12.3/Debian-6.6) id m4K6adwN008952 for xfs@oss.sgi.com; Tue, 20 May 2008 08:36:39 +0200 Date: Tue, 20 May 2008 08:36:39 +0200 From: Christoph Hellwig Subject: Re: [PATCH 2/2] kill xfs_lock_dir_and_entry Message-ID: <20080520063639.GC8869@lst.de> References: <20080502105803.GC17870@lst.de> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <20080502105803.GC17870@lst.de> Sender: xfs-bounce@oss.sgi.com Errors-to: xfs-bounce@oss.sgi.com List-Id: xfs To: xfs@oss.sgi.com ping? On Fri, May 02, 2008 at 12:58:03PM +0200, Christoph Hellwig wrote: > When multiple inodes are locked in XFS it happens in order of the inode > number, with the everything but the first inode trylocked if any of > the previous inodes is in the AIL. > > Except for the sorting of the inodes this logic is implemented in > xfs_lock_inodes, but also partially duplicated in xfs_lock_dir_and_entry > in a particularly stupid way adds a lock roundtrip if the inode ordering > is not optimal. > > This patch adds a new helper xfs_lock_two_inodes that takes two inodes > and locks them in the most optimal way according to the above locking > protocol and uses it for all places that want to lock two inodes. > > The only caller of xfs_lock_inodes is xfs_rename which might lock up to > four inodes. > > > Signed-off-by: Christoph Hellwig > > Index: linux-2.6-xfs/fs/xfs/xfs_vnodeops.c > =================================================================== > --- linux-2.6-xfs.orig/fs/xfs/xfs_vnodeops.c 2008-05-02 08:30:24.000000000 +0200 > +++ linux-2.6-xfs/fs/xfs/xfs_vnodeops.c 2008-05-02 08:30:30.000000000 +0200 > @@ -1897,111 +1897,6 @@ std_return: > } > > #ifdef DEBUG > -/* > - * Some counters to see if (and how often) we are hitting some deadlock > - * prevention code paths. > - */ > - > -int xfs_rm_locks; > -int xfs_rm_lock_delays; > -int xfs_rm_attempts; > -#endif > - > -/* > - * The following routine will lock the inodes associated with the > - * directory and the named entry in the directory. The locks are > - * acquired in increasing inode number. > - * > - * If the entry is "..", then only the directory is locked. The > - * vnode ref count will still include that from the .. entry in > - * this case. > - * > - * There is a deadlock we need to worry about. If the locked directory is > - * in the AIL, it might be blocking up the log. The next inode we lock > - * could be already locked by another thread waiting for log space (e.g > - * a permanent log reservation with a long running transaction (see > - * xfs_itruncate_finish)). To solve this, we must check if the directory > - * is in the ail and use lock_nowait. If we can't lock, we need to > - * drop the inode lock on the directory and try again. xfs_iunlock will > - * potentially push the tail if we were holding up the log. > - */ > -STATIC int > -xfs_lock_dir_and_entry( > - xfs_inode_t *dp, > - xfs_inode_t *ip) /* inode of entry 'name' */ > -{ > - int attempts; > - xfs_ino_t e_inum; > - xfs_inode_t *ips[2]; > - xfs_log_item_t *lp; > - > -#ifdef DEBUG > - xfs_rm_locks++; > -#endif > - attempts = 0; > - > -again: > - xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT); > - > - e_inum = ip->i_ino; > - > - xfs_itrace_ref(ip); > - > - /* > - * We want to lock in increasing inum. Since we've already > - * acquired the lock on the directory, we may need to release > - * if if the inum of the entry turns out to be less. > - */ > - if (e_inum > dp->i_ino) { > - /* > - * We are already in the right order, so just > - * lock on the inode of the entry. > - * We need to use nowait if dp is in the AIL. > - */ > - > - lp = (xfs_log_item_t *)dp->i_itemp; > - if (lp && (lp->li_flags & XFS_LI_IN_AIL)) { > - if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL)) { > - attempts++; > -#ifdef DEBUG > - xfs_rm_attempts++; > -#endif > - > - /* > - * Unlock dp and try again. > - * xfs_iunlock will try to push the tail > - * if the inode is in the AIL. > - */ > - > - xfs_iunlock(dp, XFS_ILOCK_EXCL); > - > - if ((attempts % 5) == 0) { > - delay(1); /* Don't just spin the CPU */ > -#ifdef DEBUG > - xfs_rm_lock_delays++; > -#endif > - } > - goto again; > - } > - } else { > - xfs_ilock(ip, XFS_ILOCK_EXCL); > - } > - } else if (e_inum < dp->i_ino) { > - xfs_iunlock(dp, XFS_ILOCK_EXCL); > - > - ips[0] = ip; > - ips[1] = dp; > - xfs_lock_inodes(ips, 2, XFS_ILOCK_EXCL); > - } > - /* else e_inum == dp->i_ino */ > - /* This can happen if we're asked to lock /x/.. > - * the entry is "..", which is also the parent directory. > - */ > - > - return 0; > -} > - > -#ifdef DEBUG > int xfs_locked_n; > int xfs_small_retries; > int xfs_middle_retries; > @@ -2135,6 +2030,45 @@ again: > #endif > } > > +void > +xfs_lock_two_inodes( > + xfs_inode_t *ip0, > + xfs_inode_t *ip1, > + uint lock_mode) > +{ > + xfs_inode_t *temp; > + int attempts = 0; > + xfs_log_item_t *lp; > + > + ASSERT(ip0->i_ino != ip1->i_ino); > + > + if (ip0->i_ino > ip1->i_ino) { > + temp = ip0; > + ip0 = ip1; > + ip1 = temp; > + } > + > + again: > + xfs_ilock(ip0, xfs_lock_inumorder(lock_mode, 0)); > + > + /* > + * If the first lock we have locked is in the AIL, we must TRY to get > + * the second lock. If we can't get it, we must release the first one > + * and try again. > + */ > + lp = (xfs_log_item_t *)ip0->i_itemp; > + if (lp && (lp->li_flags & XFS_LI_IN_AIL)) { > + if (!xfs_ilock_nowait(ip1, xfs_lock_inumorder(lock_mode, 1))) { > + xfs_iunlock(ip0, lock_mode); > + if ((++attempts % 5) == 0) > + delay(1); /* Don't just spin the CPU */ > + goto again; > + } > + } else { > + xfs_ilock(ip1, xfs_lock_inumorder(lock_mode, 1)); > + } > +} > + > int > xfs_remove( > xfs_inode_t *dp, > @@ -2210,9 +2144,7 @@ xfs_remove( > goto out_trans_cancel; > } > > - error = xfs_lock_dir_and_entry(dp, ip); > - if (error) > - goto out_trans_cancel; > + xfs_lock_two_inodes(dp, ip, XFS_ILOCK_EXCL); > > /* > * At this point, we've gotten both the directory and the entry > @@ -2239,9 +2171,6 @@ xfs_remove( > } > } > > - /* > - * Entry must exist since we did a lookup in xfs_lock_dir_and_entry. > - */ > XFS_BMAP_INIT(&free_list, &first_block); > error = xfs_dir_removename(tp, dp, name, ip->i_ino, > &first_block, &free_list, resblks); > @@ -2347,7 +2276,6 @@ xfs_link( > { > xfs_mount_t *mp = tdp->i_mount; > xfs_trans_t *tp; > - xfs_inode_t *ips[2]; > int error; > xfs_bmap_free_t free_list; > xfs_fsblock_t first_block; > @@ -2395,15 +2323,7 @@ xfs_link( > goto error_return; > } > > - if (sip->i_ino < tdp->i_ino) { > - ips[0] = sip; > - ips[1] = tdp; > - } else { > - ips[0] = tdp; > - ips[1] = sip; > - } > - > - xfs_lock_inodes(ips, 2, XFS_ILOCK_EXCL); > + xfs_lock_two_inodes(sip, tdp, XFS_ILOCK_EXCL); > > /* > * Increment vnode ref counts since xfs_trans_commit & > Index: linux-2.6-xfs/fs/xfs/xfs_dfrag.c > =================================================================== > --- linux-2.6-xfs.orig/fs/xfs/xfs_dfrag.c 2008-04-26 17:43:14.000000000 +0200 > +++ linux-2.6-xfs/fs/xfs/xfs_dfrag.c 2008-05-02 08:30:30.000000000 +0200 > @@ -128,7 +128,6 @@ xfs_swap_extents( > xfs_swapext_t *sxp) > { > xfs_mount_t *mp; > - xfs_inode_t *ips[2]; > xfs_trans_t *tp; > xfs_bstat_t *sbp = &sxp->sx_stat; > bhv_vnode_t *vp, *tvp; > @@ -153,16 +152,7 @@ xfs_swap_extents( > vp = XFS_ITOV(ip); > tvp = XFS_ITOV(tip); > > - /* Lock in i_ino order */ > - if (ip->i_ino < tip->i_ino) { > - ips[0] = ip; > - ips[1] = tip; > - } else { > - ips[0] = tip; > - ips[1] = ip; > - } > - > - xfs_lock_inodes(ips, 2, lock_flags); > + xfs_lock_two_inodes(ip, tip, lock_flags); > locked = 1; > > /* Verify that both files have the same format */ > @@ -265,7 +255,7 @@ xfs_swap_extents( > locked = 0; > goto error0; > } > - xfs_lock_inodes(ips, 2, XFS_ILOCK_EXCL); > + xfs_lock_two_inodes(ip, tip, XFS_ILOCK_EXCL); > > /* > * Count the number of extended attribute blocks > Index: linux-2.6-xfs/fs/xfs/xfs_inode.h > =================================================================== > --- linux-2.6-xfs.orig/fs/xfs/xfs_inode.h 2008-05-01 22:56:57.000000000 +0200 > +++ linux-2.6-xfs/fs/xfs/xfs_inode.h 2008-05-02 08:30:30.000000000 +0200 > @@ -522,6 +522,7 @@ void xfs_iflush_all(struct xfs_mount *) > void xfs_ichgtime(xfs_inode_t *, int); > xfs_fsize_t xfs_file_last_byte(xfs_inode_t *); > void xfs_lock_inodes(xfs_inode_t **, int, uint); > +void xfs_lock_two_inodes(xfs_inode_t *, xfs_inode_t *, uint); > > void xfs_synchronize_atime(xfs_inode_t *); > void xfs_mark_inode_dirty_sync(xfs_inode_t *); ---end quoted text---