Return-Path: linux-nfs-owner@vger.kernel.org Received: from mail-pa0-f45.google.com ([209.85.220.45]:54146 "EHLO mail-pa0-f45.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S965219AbbBBWkR (ORCPT ); Mon, 2 Feb 2015 17:40:17 -0500 Received: by mail-pa0-f45.google.com with SMTP id et14so87953682pad.4 for ; Mon, 02 Feb 2015 14:40:16 -0800 (PST) From: Tom Haynes To: Trond Myklebust Cc: Linux NFS Mailing list Subject: [PATCH v6 21/53] nfs41: serialize first layoutget of a file Date: Mon, 2 Feb 2015 14:38:35 -0800 Message-Id: <1422916747-86649-22-git-send-email-loghyr@primarydata.com> In-Reply-To: <1422916747-86649-1-git-send-email-loghyr@primarydata.com> References: <1422916747-86649-1-git-send-email-loghyr@primarydata.com> Sender: linux-nfs-owner@vger.kernel.org List-ID: From: Peng Tao Per RFC 5661 Errata 3208: | A client MAY always forget its layout state and associated | layout stateid at any time (See also section 12.5.5.1). | In such case, the client MUST use a non-layout stateid for the next | LAYOUTGET operation. This will signal the server that the client has | no more layouts on the file and its respective layout state can be | released before issuing a new layout in response to LAYOUTGET. In order to make such a signal unique to server, client needs to serialize all layoutgets using non-layout stateid. We implement this by serializing layoutgets when client has no layout segments at hand. Signed-off-by: Peng Tao Signed-off-by: Tom Haynes --- fs/nfs/pnfs.c | 35 +++++++++++++++++++++++++++++++---- fs/nfs/pnfs.h | 1 + 2 files changed, 32 insertions(+), 4 deletions(-) diff --git a/fs/nfs/pnfs.c b/fs/nfs/pnfs.c index fa00b56..7e1bac1 100644 --- a/fs/nfs/pnfs.c +++ b/fs/nfs/pnfs.c @@ -1288,6 +1288,7 @@ pnfs_update_layout(struct inode *ino, struct nfs_client *clp = server->nfs_client; struct pnfs_layout_hdr *lo; struct pnfs_layout_segment *lseg = NULL; + bool first; if (!pnfs_enabled_sb(NFS_SERVER(ino))) goto out; @@ -1295,6 +1296,8 @@ pnfs_update_layout(struct inode *ino, if (pnfs_within_mdsthreshold(ctx, ino, iomode)) goto out; +lookup_again: + first = false; spin_lock(&ino->i_lock); lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags); if (lo == NULL) { @@ -1312,10 +1315,27 @@ pnfs_update_layout(struct inode *ino, if (pnfs_layout_io_test_failed(lo, iomode)) goto out_unlock; - /* Check to see if the layout for the given range already exists */ - lseg = pnfs_find_lseg(lo, &arg); - if (lseg) - goto out_unlock; + first = list_empty(&lo->plh_segs); + if (first) { + /* The first layoutget for the file. Need to serialize per + * RFC 5661 Errata 3208. + */ + if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET, + &lo->plh_flags)) { + spin_unlock(&ino->i_lock); + wait_on_bit(&lo->plh_flags, NFS_LAYOUT_FIRST_LAYOUTGET, + TASK_UNINTERRUPTIBLE); + pnfs_put_layout_hdr(lo); + goto lookup_again; + } + } else { + /* Check to see if the layout for the given range + * already exists + */ + lseg = pnfs_find_lseg(lo, &arg); + if (lseg) + goto out_unlock; + } if (pnfs_layoutgets_blocked(lo, 0)) goto out_unlock; @@ -1343,6 +1363,13 @@ pnfs_update_layout(struct inode *ino, lseg = send_layoutget(lo, ctx, &arg, gfp_flags); atomic_dec(&lo->plh_outstanding); out_put_layout_hdr: + if (first) { + unsigned long *bitlock = &lo->plh_flags; + + clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock); + smp_mb__after_atomic(); + wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET); + } pnfs_put_layout_hdr(lo); out: dprintk("%s: inode %s/%llu pNFS layout segment %s for " diff --git a/fs/nfs/pnfs.h b/fs/nfs/pnfs.h index 9ccaa50..4993ba6 100644 --- a/fs/nfs/pnfs.h +++ b/fs/nfs/pnfs.h @@ -95,6 +95,7 @@ enum { NFS_LAYOUT_ROC, /* some lseg had roc bit set */ NFS_LAYOUT_RETURN, /* Return this layout ASAP */ NFS_LAYOUT_INVALID_STID, /* layout stateid id is invalid */ + NFS_LAYOUT_FIRST_LAYOUTGET, /* Serialize first layoutget */ }; enum layoutdriver_policy_flags { -- 1.9.3